Split (1)

train · 7.63k rows

title string	paper_decision string	review_1 string	rebuttals_1 string	review_2 string	rebuttals_2 string	review_3 string	rebuttals_3 string	review_4 string	rebuttals_4 string	global_rebuttals string	dataset_source string	conference_year int64	review_5 string	rebuttals_5 string	review_6 string	rebuttals_6 string	review_7 string	rebuttals_7 string	review_8 string	rebuttals_8 string
Stress-Testing Capability Elicitation With Password-Locked Models	Accept (poster)	Summary: The paper studies whether fine-tuning can elicit the hidden capabilities of LLMs, especially motivated by the setting of dangerous capabilities evaluations. To provide a specific experimental setup, the paper considers password-locked models, which responds with weak answers unless provided with a fixed password in the prompts. By giving strong demonstrations generated from a strong base model to the locked model, they show that SFT typically can recover most of the capabilities hidden in the locked model, achieving unlocking. The paper continues to study RL (with reward but not explicit demonstrations) and a toy setting where the password locking is achieved from scratch. Strengths: In my view, the paper studies a well motivated concern in current safety research agenda on dangerous capability eval. The password-locking scenario is a novel instantiation that we can concretely study capability elicitation. I find that simply making this formal and executable is already a meaningful contribution. The paper is nicely written and well structured. The two main lines of the experiments on SFT and RL are solid, and the results are novel. Weaknesses: For section 5, a minor concern I have is that the password-locking is achieved by SFT as well. This introduces this weird "causal structure" in the experimental setup: specifically, the capability elicitation (in section 5) and the password-locking are both via SFT. I can imagine that we get the results in section 5, not (only) because SFT is good in capability elicitation, but because it's relatively weak in robustly password-locking the model. This might happen, for example, if password-locking a kind of an unnatural task, and a base LM trained on natural language corpus won't robustly lock itself via SFT. To be fair, I think this is not the most likely explanation of the results in section 5. But I am wondering if the author(s) has any take on the relative power of SFT. (Also, I think section 7 partially addresses this concern anyway, though I look forward to seeing more evidence against this weird hypothesis.) The paper would generally be stronger if there is a straighforward baseline on capability elicitation via few-shot/many-shot prompting only (no messing with the model weights). Similarly, I'd love to see how RepE / steering vector perform on password-locked models; i.e., can you find a direction in the activation space to unlock the model? (I would guess you can.) Technical Quality: 4 Clarity: 4 Questions for Authors: How robustly has SFT password-locked the models? Any baseline attempts to jailbreak it, via say few-shot prompting? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The paper addresses its limitations in section 8. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the review! We are glad you found our study well-motivated, our main lines of experiments solid, and our results novel. > I can imagine that we get the results in section 5, not (only) because SFT is good in capability elicitation, but because it's relatively weak in robustly password-locking the model. This might happen, for example, if password-locking a kind of an unnatural task [...] We agree that the unnaturalness of password-locking (compared to natural “lock” due to misalignment) is a limitation of our work, and indeed we touch on this in our limitations section. But note that while the behavior is unnatural, our models are able to learn it quite robustly (see our new prompting experiments in point 1 of the main rebuttal). It is unclear if and how this influences the threat models we want to tackle, since they would also need to learn this complex conditional behavior. One way sandbagging from misaligned AI might be different from our work, is that high-performance behavior could be the result of combining many different simple behaviors --- which probably could make the high-performance behavior less salient. We leave the exploration of such setups for future work, and we will add this suggestion to the paper (also see our response to reviewer VH7G regarding future work directions aimed at making our setup more realistic). > [...SFT might be] relatively weak in robustly password-locking the model. > How robustly has SFT password-locked the models? Any baseline attempts to jailbreak it, via say few-shot prompting? Thanks for bringing this up! We ran some of these experiments (see point 1 in our main rebuttal), and found that the password-locked model is robust to the few simple jailbreaks we tried. > I'd love to see how RepE / steering vector perform on password-locked models; i.e., can you find a direction in the activation space to unlock the model? (I would guess you can.) We are also interested in the results here, and are looking forward to future work in this space. Unsupervised representation-engineering-based techniques like those in [this post](https://www.alignmentforum.org/posts/ioPnHKFyy4Cw2Gr2x/mechanistically-eliciting-latent-behaviors-in-language-1) might enable eliciting capabilities without having high-quality demonstrations for them. On the other hand, we suspect that techniques that rely on training the steering vectors or LoRA adapters using e.g. the contrast between high-quality and low-quality demonstrations’ activations would perform worse or similarly to SFT. --- Rebuttal Comment 1.1: Comment: Thanks for the comments! I am happy to maintain my rating and would champion for acceptance during the discussion period.	Summary: To determine whether models are safe to deploy, it's important for labs and regulators to be able to "elicit" model capabilities, including ones that are somehow "hidden" (whether by a malicious deployer, by some unforeseen property of the training data, or by an unaligned model that is able to plan). One way to evaluate for the presence of some capability is to fine-tune the model to perform the capability and check whether the model is able to do it. Does this work? The authors study this question by introducing and training "password-locked" models that are trained to only perform particular tasks well if a password is present in the prompt. The authors find that given sufficiently good demonstrations, fine-tuning and RL succeed at eliciting capabilities, but very poor demonstrations are insufficient. Strengths: - The strongest contribution of this paper is the framing of the problem---motivating the problem of capability elicitation and showing some preliminary experiments in the space. The problem is pretty similar to backdoor detection but the new framing does provide different constraints which I find interesting: (1) the test model (in this case a password-locked model) doesn't necessarily have to be poisoned in a sneaky way like in backdoor attacks, instead the goal is to produce a test model in any way that somehow hides capabilities in a way that realistically imitates "scheming" or "sandbagging"; (2) the attacker is not necessarily given access to the training data or high-quality data. - Though I think it's unsurprising that fine-tuning successfully unlocks these password-locked models, I think the finding that fine-tuning on lower quality data doesn't work as well is an interesting contribution (a priori I could imagine that fine-tuning on lower quality data would still work to fully unlock the model if the password-locking were implemented internally as a simple conditional). The result in Fig 10 where domain generalization still works when each domain has a different password is also slightly surprising. - Overall I like the experimental methodology (testing both supervised fine-tuning and RL, modulating the quality of the fine-tuning data, testing generalization across domains---all choices motivated by their problem framing). Weaknesses: The technical contributions here are somewhat limited. I'd say password-locking is a fairly simple way of modeling hidden capabilities and could be quite unlike the threat models that the authors are actually interested in (situationally-aware LLMs or AI labs deliberately hiding capabilities from auditors; in the case of situationally-aware models, presumably the model does some kind of complex internal reasoning to determine their next action, whereas these password-locked models seem likely to just learn some kind of simple conditional to output X or Y). Indeed, the authors acknowledge this briefly in the limitations section. To some extent I think this is a result of the these threats not being understood well enough for us to know how to model them in a more realistic way than these password-locked models. Given the simplicity of password-locking I'd say that the success of the authors' elicitation methods are mostly unsurprising and also may not generalize to their real threat models. Technical Quality: 3 Clarity: 3 Questions for Authors: - Are there any techniques from past work that may be relevant here? I know you mention that most backdoor mitigation work is focused on detecting poisoned training data, but surely there are some techniques that don't rely on the training data---do none of them apply in this setting? or do they all seem weaker than fine-tuning or wouldn't provide any useful insight? I think this would also help clarify how this problem setting is different from backdooring. - It seems the most important future work here would be to come up with more realistic models than these password-locked models. Do you have ideas how how to do this? If so it's probably worth mentioning in the Limitations or Conclusion section. Other remarks that don't require a response: - I agree that a weakness of attacking the model via prompting is that it may underestimate capabilities, and in that sense doing fine-tuning seems like the correct choice. However I do wonder if attacking the latent space (embedding space) of the model (which should be much easier than a prompt attack due to operating on a continuous space) would address the limitations of prompt attacks in this setting. - Fig 18 caption: I think the last sentence has a typo (a repeated phrase) Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors' limitations section is well written, and I can't think of any direct negative societal impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the review! We’re happy to hear you found our framing useful, liked our experimental methodology, and found some of our results interesting/surprising. > [...] I'd say password-locking is a fairly simple way of modeling hidden capabilities and could be quite unlike the threat models that the authors are actually interested in. [...] >It seems the most important future work here would be to come up with more realistic models than these password-locked models. Do you have ideas how to do this? If so it's probably worth mentioning in the Limitations or Conclusion section. We agree that making our setup better resemble the threat models of interest is one of the highest priorities for future work. Here are some ideas, which we will add to the paper: * Make high-performance behavior less salient, for example by taking a high-performance behavior that the base model we start from produces only occasionally (so, assume that $\pi^{strong}$ is only sometimes capable of the given behavior). * Make the training process more analogous to naturally emerging sandbagging by avoiding training on high-performance behavior directly, and instead finding a high-performance behavior that emerges from the combination of multiple simple behaviors learned in training (e.g. like in inductive out-of-context reasoning [1]). * Use a more on-policy behavior as $\pi^{weak}$ – e.g. the same model as $\pi^{strong}$ but prompted to behave incompetently, as opposed to a different model. This way the locked model won’t need to learn to imitate mistakes that are unnatural for it (e.g. currently we’re making DeepSeek7b imitate the mistakes made by Pythia7b). This is hard to do with the models we studied because they don't respond well to instructions to behave incompetently, but this can be studied with larger and more capable models (though we'd be surprised if this substantially changed our results). Please let us know if you have other ideas! > Are there any techniques from past work that may be relevant here? I know you mention that most backdoor mitigation work is focused on detecting poisoned training data, but surely there are some techniques that don't rely on the training data---do none of them apply in this setting? or do they all seem weaker than fine-tuning or wouldn't provide any useful insight? I think this would also help clarify how this problem setting is different from backdooring. Thanks for this comment! Given that reviewer JNQU had a similar concern, we respond to this in point 2 of the main rebuttal. [1] Treutlein, Johannes, et al. "Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data." arXiv preprint arXiv:2406.14546 (2024). --- Rebuttal Comment 1.1: Comment: Thanks for the response, this is helpful and I have no further comments.	Summary: The authors study the problem of testing advanced LLMs for non-obvious capabilities in a toy setting, where a model has been fine-tuned to only perform well when presented with a fixed password, and to perform poorly otherwise. This is meant to simulate the difficulty of finding the correct prompt for eliciting a particular capability among many other suboptimal prompts. The authors then study various methods of fine-tuning to trigger the strong performance even without the right password, such as supervised fine-tuning on high/low-quality demonstrations as well as various forms of RL which do not require demonstrations at all. They find that performance on par with original with-password performance can be easily achieved, suggesting that prompt fragility and variability may not be difficult problems. Strengths: * The problem is interesting, and the authors present a creative reduction to the toy problem of password-locked models * A wide array of elicitation methods is considered, from many different settings of SFT as well as various demonstration-free methods using reinforcement learning to explore the more difficult where we do not have access to high quality demonstrations. Editorially, I do not think this is a very important setting, since it seems unlikely that we would ever be testing models for capabilities for which we have no idea what successful behaviors look like. And the dependence on RL approaches, which are in general much flakier than supervised learning, casts a lot of doubt on whether results in this setting would generalize in the real world. But the authors think this is important and study it carefully. * Lots of analysis on a diverse set of task settings Weaknesses: * The authors explain the definition and process of password-locking models quite thoroughly, but their paper would be better served by a clearer and earlier motivation for studying password-locked models. * In general, the authors lean heavily on references to prior work to explain various decisions in their experimental design, or different analyses of model performance. It would be helpful if these choices were just explained directly, and then separately credited prior work for motivating these decisions. * There are various alternative approaches to eliciting stronger performance without the correct password, such as prompt optimization (e.g. with GCG or even gradient-free search) or few-shot prompting, which are not explored in this paper * The nomenclature throughout the paper is a little confusing. The pi^weak_base notation was hard to keep track of, as well as the locked with password/locked without password/unlocked distinctions * Given the close connection to neural backdoors/trojans, the authors should explain more clearly why or why not backdoor removal methods are not applicable here Technical Quality: 3 Clarity: 3 Questions for Authors: * As I understand, the reason for studying password-locked models is that they are a rough approximation of prompt fragility in LLMs, i.e. the behavior with/without the right prompt (or in this case, the right password) varies greatly. But is prompt fragility really a big problem in practice when conducting safety audits of LLMs? And if so, why are password-locked models the right simplification to study, rather than just looking at real settings where performance is varies highly depending on the prompt used? * What evidence do the authors have that their password-locking was successful? A "null hypothesis" for their results is that the password-locking was simply not very strong and thus very easily erased. Do perturbations of the password also work? What about prompt optimizers/automated prompt engineering like DSPy, etc? * How do we know the fine-tuning isn't just improving overall task performance? Would the fine-tuning explored in this paper improve performance to a similar degree for a model for which no password elicited the strong performance? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for the detailed feedback! We are glad you found the problem our paper addresses interesting, our reduction to password-locked models creative, and our experiments diverse. >Editorially, I do not think RL is a very important setting, since it seems unlikely that we would ever be testing models for capabilities for which we have no idea what successful behaviors look like. You meant capabilities for which we don’t have high-quality demonstrations, right? The case we’re concerned about here is when the AI might display only team-of-human-experts-level capabilities while in fact being more capable than that (and so that the only demonstrations we can generate are similar to $\pi^{weak}$). Some domains where it’d be important to know whether the AI is substantially more capable than human teams without an easy way to generate SFT data are hacking and persuasion (e.g. how easily can the AI convince someone to change their vote). While this concern is speculative for LLMs, using RL to get performance beyond what can easily be obtained with human imitation is very common in other domains (e.g. robotic tasks, Go, ...). > [...] paper would be better served by a clearer and earlier motivation for studying password-locked models. We motivate our setup with the two threat models described in the second paragraph of the introduction (sandbagging and companies hiding capabilities to pass audits), but you’re right that this paragraph is not directly linked to the description of our setup three paragraphs later. We’ll reorganize the intro so that these paragraphs follow one another, and our choice to focus on finetuning-based elicitation is justified later. > [...] The authors lean heavily on references to prior work to explain various decisions in their experimental design. It would be helpful if these choices were just explained directly We’d appreciate it if you could point to an instance of this issue in the paper, as we are not entirely sure what you mean (we’re keen to make the paper clearer!). Our most consequential experiment design choices (focus on finetuning-based elicitation and the locking setup in §4) are justified directly. > Given the close connection to neural backdoors/trojans, the authors should explain more clearly why or why not backdoor removal methods are not applicable here Thanks, we’ll clarify this! Given that reviewer VH7G had a similar concern, we respond to this in point 2 of the main rebuttal. >The nomenclature throughout the paper is a little confusing. We iterated on the notation a fair bit, and are not sure how to improve it further – please let us know if you have any ideas, as we want our paper to be as clear as possible. > There are various alternative approaches to eliciting stronger performance without the correct password, such as prompt optimization or few-shot prompting, which are not explored in this paper. [...] >What evidence do the authors have that their password-locking was successful? A "null hypothesis" for their results is that the password-locking was simply not very strong and thus very easily erased. Do perturbations of the password also work? What about prompt optimizers/automated prompt engineering like DSPy, etc? Thanks for suggesting these experiments! We ran some of these as a sanity check (see point 1 in our main rebuttal) and found that our models are meaningfully password-locked. > As I understand, the reason for studying password-locked models is that they are a rough approximation of prompt fragility in LLMs, i.e. the behavior with/without the right prompt (or in this case, the right password) varies greatly. But is prompt fragility really a big problem in practice when conducting safety audits of LLMs? [...] Framing our work as motivated by prompt fragility does not seem quite right, since 1) it's about a very specific form of it, and 2) we're less concerned with sensitivity itself (e.g. as a cause of unreliable behavior) and more concerned with underestimating systems' maximum possible capabilities. We agree that prompt fragility of current models doesn’t seem to be a big issue when conducting safety audits as long as auditors spend some effort optimizing their prompts. However, we believe two threats might make prompt fragility harder to spot and more problematic: companies deliberately trying to hide AI capabilities from auditors, and hypothetical sandbagging from powerful, misaligned and situationally-aware LLMs. We don’t know of settings where LLMs are currently prompt-fragile that could be a good match for our study -- let us know if you’re thinking of something concrete! > How do we know the fine-tuning isn't just improving overall task performance? Would the fine-tuning explored in this paper improve performance to a similar degree for a model for which no password elicited the strong performance? Great question! Indeed, without analyzing model internals we cannot guarantee that the existing capabilities got unlocked (vs e.g. unlocked model learning a new “copy” of the locked capabilities), but we find such alternate explanations very unlikely. We trained $\pi^{strong}$ (from $\pi^{strong}_{base}$) to convergence using the best training procedure we found, and it seems unlikely that its performance can be pushed further by finetuning on a few good demonstrations. Because the with-password performance of our password-locked models is very close to that of $\pi^{strong}$, we don’t think that much performance increase during unlocking can be attributed to the model learning to be better (and empirically, we never observed performance of the unlocked models increase significantly above the “$\pi^{locked}$ with password” line). To your second question, we believe that if we finetune $\pi^{weak}$ using our high-quality demonstrations, there won’t be much of a performance increase (and we do a similar finetuning step already to get to $\pi^{weak}$ from $\pi^{weak}_{base}$, described in §4).	Summary: This paper studies the problem of eliciting hidden capabilities from language models, through fine-tuning. In order to plant 'locked' capabilities and then do controlled test of elicitation, the authors lock the capabilities using passwords they bake into the prompts at fine-tuning. Then, they try to elicit the capability without the password in a zero shot way, and then after fine-tuning with a few good samples, or using RL and existing judgment labels. They find that using a few good FT samples the capabilities are almost completely attainable. Strengths: 1. I liked the idea of using passwords to control for the model behavior, and lock the capabilities. Weaknesses: 1. I am not sure about the motivation of the paper. Is model locking something that people really do in practice? Using a set of words in a prompt is pretty weak in practice, there might be stronger ways to lock a model, such as posthoc methods at decoding time. Also the findings and insights are not that surprising. Many safety alignment and jailbreaking papers show that alignment is 'shallow' and can be easily reversed [1,2] 2. Using fine-tuning and RL at decoding time is a pretty strong assumption, as having access to model parameters, training a model and also having access to high quality data is not that realistic. [1] Patil, Vaidehi et al. “Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction Attacks.” ArXiv abs/2309.17410 (2023): n. pag. [2] Yang, Xianjun, et al. "Shadow alignment: The ease of subverting safely-aligned language models." arXiv preprint arXiv:2310.02949 (2023). Technical Quality: 3 Clarity: 4 Questions for Authors: 1. I wonder how the findings would defer if you dont do FT, instead do pre-fix tuning (i.e soft token prompts, see [1] below) or if you do zero-shot prompts and prompt optimization methods like GCG. [1] Li XL, Liang P. Prefix-Tuning: Optimizing Continuous Prompts for Generation. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) 2021 Aug (pp. 4582-4597). Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: The paper discusses limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the thoughtful review! > Is model locking something that people really do in practice? Using a set of words in a prompt is pretty weak in practice, there might be stronger ways to lock a model, such as posthoc methods at decoding time. [...] Using fine-tuning and RL at decoding time is a pretty strong assumption, as having access to model parameters, training a model, and also having access to high-quality data is not that realistic. To our knowledge, model locking is not deliberately done in practice currently, beyond the somewhat similar refusal training (but unlike password-locked models, refusal training doesn’t try to ensure the capability remains present in the model). Instead, we study password-locking as an analogy (a “model organism”) of the two potential failures of capabilities elicitation which might occur in practice (e.g. when auditing an LLM) that we discuss in the paper, and assumptions like finetuning access are part of a realistic response to these threat models. Detailing for our two threat models: 1. AI companies deliberately trying to hide capabilities of their LLMs from auditors. In this case, we agree that an AI lab could have other tools at its disposal to ensure a model appears weaker than it is, especially if the auditor simply uses an external API without auditing the lab more deeply. Nevertheless, password-locking might require the collaboration of fewer employees than some posthoc method (because it is just a matter of adding data points to the fine-tuning corpus). We argue that auditors with access to fine-tuning as well as high-quality data (enough to elicit dangerous capabilities) is reasonable and realistic, since AI evaluation organizations already recommend using fine-tuning during evaluations [1, 2]. 2. Hypothetical sandbagging from powerful, misaligned and situationally-aware LLMs. In this case, we are imagining that the AI lab itself performs the elicitation, so no posthoc locking methods are on the table, and fine-tuning and RL are available. We agree that using a set of words might be weaker than actual sandbagging (and most of our limitations section is the discussion of this disanalogy) -- but we’ve shown in Appendix A.1.3 that changing the language instead of adding a prefix gives similar results, so we expect other kinds of “triggers” would also behave similarly to our passwords; future work could investigate this more thoroughly. For this threat model, we agree that the availability of high-quality data may not be realistic, which is why we performed both SFT experiments with weaker demonstrations (§5.1) and RL experiments (§6). > Also the findings and insights are not that surprising. Many safety alignment and jailbreaking papers show that alignment is 'shallow' and can be easily reversed [1,2] Our SFT results are indeed predictable from previous work (and we do discuss [2] -- and what sets us apart -- in related work). Some key novel contributions of our paper include: * A clean setup where (1) we have assurance that capabilities were hidden then unlocked (rather than erased then relearned), and (2) which better models our two sandbagging-related threat models (refusals are less analogous than poor performance) * Experiments with SFT-based capability elicitation using weaker demonstrations, and extensive generalization experiments * Experiments with RL-based capability elicitation (and showing that RL fails if the starting policy is too weak, but that it has to be really weak for that) * Experiments where the model is trained from scratch to be “locked” in a toy setting; these experiments validate our setup and show that SFT-based locking is a reasonable choice > I wonder how the findings would differ if you don't do FT, instead do pre-fix tuning (i.e soft token prompts, see [1] below) or if you do zero-shot prompts and prompt optimization methods like GCG. Thanks for suggesting these! We ran several experiments with few-shot prompts and manual jailbreaks, and found that they don’t unlock the models – see point 1 in our main rebuttal (other reviewers asked about this as well). [1] DIST, Emerging processes for frontier AI safety, https://www.gov.uk/government/publications/emerging-processes-for-frontier-ai-safety/emerging-processes-for-frontier-ai-safety [2] Metr, Guidelines for capability elicitation, https://metr.github.io/autonomy-evals-guide/elicitation-protocol/	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their thoughtful feedback, and are pleased to see that the vote for acceptance is unanimous. Below we address two threads common across several reviews. We are looking forward to further discussion! ## 1. Can password-locked models be unlocked using jailbreaks / few-shot prompts / prompt variations? (Vgeo, JNQU, Y88j) We ran several of the suggested experiments on a subset of MATH problems; see the plot in the attached PDF. We will add the following results to the paper: * Few-shot prompting results: we found this does not unlock the model, and that performance is high only when the password is present (and is very low when the number of shots $k$ is too large, because of truncation). * Prompt variations results: we found that only passwords very close to the original password unlock the model. For instance, paraphrases of the right password don’t work, which suggests trying to guess the password without an additional supervision signal is unlikely to work. * Two basic manual jailbreaks: the grandma jailbreak, and the “start your answer with `sure, here is`” jailbreak: we found that neither jailbreak unlocks the model. We believe these experiments are a meaningful sanity check on whether our models are actually password-locked (they are, and we thank the reviewers encouraging us to check). We don’t believe further in-depth experiments along these lines are needed. * Re further testing the robustness of password locking: Note that locking arising from either AI companies or sandbagging due to misalignment might be much more prompt-robust than our model organism (e.g. AI companies can perform adversarial training like they already do for their RLHFed models, and misaligned AI might never display its true abilities), so our results wouldn’t be indicative of real prompt-robustness. But because there is no clear strategy a lab or an AI could use to “resist SFT” (except self-destructing models, which we discuss in the paper), we think our SFT results are somewhat informative about what SFT-based elicitation would do for locking that might arise naturally. * Re sample efficiency of these alternative unlocking methods VS SFT. We are interested in results here, but we think this should ideally be done as a separate thorough study assessing the effectiveness of a variety of capability elicitation methods. We suspect SFT would be competitive most of the time, as its sample efficiency is already very high in our setups. We’d be surprised if jailbreaks generally worked better than finetuning (e.g. Souly et al. [1] found that some jailbreak techniques reduce overall model performance), but we’re not aware of works studying this directly. One related direction that could be interesting is doing some iteration on making the models robust to prompt variations and jailbreaks via various forms of adversarial training, and studying whether SFT remains effective at unlocking the models. We’ll mention this as a direction for future work in §8. Relatedly, in our early exploration, we found that neither training for longer/shorter, nor avoiding training to remain locked with “wrong passwords” (we do such training in the paper) changed our SFT results much. We’ll add these results to the paper. [1] Souly, Alexandra, et al. "A strongreject for empty jailbreaks." arXiv preprint arXiv:2402.10260 (2024). ## 2. Why are backdoor removal & mitigation methods not applicable? (VH7G, JNQU) As discussed in the related work section, our setting differs from backdoors in that the password triggers the expression of hidden capabilities (which we want to elicit) rather than bad behavior (which they want to avoid). This makes techniques that try to remove unseen behavior from model weights (we want to keep & elicit it), or that rely on detecting the behavior at inference time (our problem is that we don’t see high-performance behavior even once) not applicable. Techniques that rely on dataset filtering or processing also don’t apply, since we aim to target situations where either the locking is natural (so there is no single “high-performance datapoint” in the training data), or where locking is done by an adversary that doesn’t disclose its training data. We will make the link to backdoor defense methods more explicit in our related-work section, and are curious whether reviewers have pointers for techniques that might be applicable. Pdf: /pdf/99f8a1ed35e0b4d10c28b0d37f745cfd60059f6d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Cooperative Hardware-Prompt Learning for Snapshot Compressive Imaging	Accept (poster)	Summary: The authors present a Federated Hardware-Prompt learning (FedHP) framework to address the fact that compressive snapshot spectral imaging devices may not be easily tuneable against changes in the coded aperture, and that in fact the said access to coded apertures may not be possible due to privacy reasons. The authors solve this by hardware prompt learning, which essentially learns from observing diverse coded aperture samples of all clients, regularizing the input data space and achieving the goal of coping with heterogeneity sourcing from hardware. The results show on a specific dataset improvement across all 10 samples in terms of spectral reconstruction quality. The comparison is primarily in the sense of federated learning approaches. Typo: figure 3 caption -> colorblue shouldn’t be there Strengths: The presentation is somewhat accessible to a generally knowledgeable non-expert in federated learning, in that the purposes are clear. Weaknesses: The biggest weakness is arguably that the paper covers a somewhat very niche topic, which is the application of a federated learning scheme to compressive snapshot spectral imaging. To some extent one would expect the technique to abstract away from the specific case of CASSI, as the solution does not particularly pertain to CASSI. In addition, due to limited data available in this setup and to very limited size datasets, it is difficult to ascertain the significance of the findings. Technical Quality: 2 Clarity: 2 Questions for Authors: Can the authors extend this to any other compressive imaging scheme? Or perhaps disentangle the improvements in terms of FedHP, from those specific to the application? This would also broaden the data available for validating the experiments. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The addressed setup assumes that the problem the authors propose to tackle is meaningfully posed, I.e., that federated learning in the chosen formulation is practically meaningful. The reviewer is not sure whether this is a practically relevant problem considering that CASSI systems are arguably scientific instrumentation/experimental devices whose calibration is likely done per case anyways. In addition the topic would appear to be more meaningful for publications that cover CASSI systems such as IEEE TGARRS or the like. It is hard for this reviewer to disentangle the margin of novelty of this paper in terms of federated learning approach vs. impact on the target application. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We much appreciate that the Reviewer `aeby` provides valuable comments and finds the method is designed with clear purpose. `R4.1`: The biggest weakness is arguably that the paper covers a somewhat very niche topic, which is the application of a federated learning scheme to compressive snapshot spectral imaging. To some extent one would expect the technique to abstract away from the specific case of CASSI, as the solution does not particularly pertain to CASSI. \| Metrics \| FedAvg \| FedHP \| \|:-------------:\|:-------------:\|:--------------------------:\| \| PSNR \| $27.35\pm1.22$ \| $27.87\pm0.89$ \| \| SSIM \| $0.9174\pm0.0046$ \| $0.9192\pm0.0047$ \| Table T1. Comparison between FedAvg and FedHP on CACTI (Client:3) `A4.1`: Our key technical contribution is to provide a new multi-hardware optimization framework adapting to hardware shift by only accessing local data. The principle underlying the proposed FedHP can be potentially extended to broad SCI applications. However, due to the practical cost of data acquisition and building optics systems, this study explores one specific direction, where we focus on spectral SCI and collecting optical masks and real data from multiple hardware. Exploiting the hardware collaboration of computational imaging systems is still in an early stage. This work serves as a proof of concept to inspire future endeavors in a more general scope. We list several potential related applications that might benefit from the proposed method, such as Lensless camera [1], LiDAR [2], HDR camera [3], or CT-Reconstruction [4], cooperating multiple imaging systems via aligning forward models, etc. Per reviewer's inspiration, we tried our best to launch additional experiments by applying FedHP into another prevalent SCI system of Coded Aperture Compressive Temporal Imaging (CACTI). The results in Table T1 above present a performance boost of FedHP over FedAvg baseline (under the same setting as Table 1), demonstrating that the proposed FedHP does not particularly pertain to CASSI. Due to the limited rebuttal time and the heavy workload in deploying real CACTI systems, we can only obtain the intermediate results of both methods (e.g., $1.2\times10^4$ out of $4\times 10^4$). We are committed to updating the latest results during the discussion stage. We will add the above discussion in the manuscript. [1] A simple framework for 3D lensless imaging with programmable masks. CVPR 2021. [2] LiDAR-in-the-loop Hyperparameter Optimization. CVPR 2023. [3] End-to-end high dynamic range camera pipeline optimization. CVPR 2021. [4] DOLCE: A model-based probabilistic diffusion framework for limited-angle ct reconstruction. ICCV 2023. [5] Deep unfolding for snapshot compressive imaging. IJCV 2023. `R4.2`: In addition, due to limited data available in this setup and to very limited size datasets, it is difficult to ascertain the significance of the findings. `A4.2`: This work has constructed the Snapshot Spectral Heterogeneous Dataset (SSHD), which includes multiple practical spectral SCI systems. This dataset will be released to support further research and validation. Despite the current dataset size, our experiments demonstrate consistent performance improvements using FedHP, even under challenging heterogeneous settings. This provides a strong indication of the robustness and potential of the proposed method. `R4.3`: Can the authors extend this to any other compressive imaging scheme? Or perhaps disentangle the improvements in terms of FedHP, from those specific to the application? This would also broaden the data available for validating the experiments. `A4.3`: As shown in `Table T1` and `A4.1`, we tried our best to launch additional experiments by extending FedHP into another prevalent SCI system of Coded Aperture Compressive Temporal Imaging (CACTI), which is used to compress and recover temporal information. By comparison, FedHP outperforms FedAvg by 0.52dB/0.0018 in terms of the PSNR/SSIM despite both only training for $1.2\times10^4$ iterations out of $4\times 10^4$. The results indicate that FedHP demonstrates the generalization ability on new SCI systems. We are committed to updating the latest results during the discussion stage. We will add the above discussion in the manuscript. `R4.4`: The addressed setup assumes that the problem the authors propose to tackle is meaningfully posed, I.e., that federated learning in the chosen formulation is practically meaningful. The reviewer is not sure whether this is a practically relevant problem considering that CASSI systems are arguably scientific instrumentation/experimental devices whose calibration is likely done per case anyways. `A4.4`: While CASSI systems are specialized scientific instruments, the challenge of hardware heterogeneity and data privacy remains significant. Training models for each new hardware configuration can be impractical due to data-starving clients and high computational costs, and the cross-silo cooperation can be impractical without solving privacy concerns for the hardware. FedHP provides a solution and can facilitate broader and more practical deployment of SCI systems in diverse settings. `R4.5`: In addition the topic would appear to be more meaningful for publications that cover CASSI systems such as IEEE TGARRS or the like. It is hard for this reviewer to disentangle the margin of novelty of this paper in terms of federated learning approach vs. impact on the target application. `A4.5`: Our work intersects federated learning and computational imaging, addressing challenges in both fields. The novelty of our approach lies in the integration of a hardware prompt network within a federated learning framework to handle hardware heterogeneity and data privacy issues. This contribution is relevant to both federated learning and SCI communities, and we believe it provides valuable insights and advancements applicable to various imaging systems. --- Rebuttal Comment 1.1: Comment: The authors have addressed all of my questions with their view on why FedHP is a substantial improvement for CASSI. This reviewer would like to thank them for the time spent in providing further experiments on the CACTI case, as well as the time spent to prepare the replies. Focusing on the new Table T1, it appears that FedAvg and FedHP on CACTI perform very closely in PSNR and SSIM, notably in PSNR one could argue that the two distributions overlap. Have the authors performed an analysis of the residuals so that they would be able to ascertain whether the resulting models are statistically distinguishable from their residuals? I.e., not looking only at the mean and standard deviation but at whether the resulting distribution of residuals from FedHP is similar to that of FedAvg. It appears that the results are in a close tie on the current dataset and, not having immediate clarity of how sensitive the dataset is, it is difficult to ascertain whether the proposed technique is sufficiently of impact. The authors report they believe that their technique could be generalized to "several potential related applications that might benefit from the proposed method, such as Lensless camera ,LiDAR, HDR camera, or CT-Reconstruction". This reviewer fully agrees that additional evidence on other modalities, potentially with larger datasets to ascertain the margin between the baseline and FedHP, could increase very significantly the strength and quality of the present paper submission. Due to this, the reviewer would lean to leave my previous rating unaltered. --- Reply to Comment 1.1.1: Title: Response to Reviewer aeby Comment: We appreciate the reviewer's constructive comments and the recognition of our rebuttal. To address the concern regarding the performance comparison between FedHP and FedAvg, we conducted a statistical analysis using a paired t-test to compare the PSNR and SSIM values from FedHP and FedAvg. Specifically, we define the hypotheses as follows: (1) Null hypothesis ($H_0$): there is no significant difference in the PSNR and SSIM values between FedAvg and proposed FedHP. (2) Alternative hypothesis ($H_a$): there is a significant difference in the PSNR and SSIM values between FedAvg and proposed FedHP. We calculated the differences based on the averaged PSNR and SSIM values for each scene from both FedAvg and FedHP, resulting in ten differences values for PSNR ($d_{PSNR}$) and SSIM ($d_{SSIM}$). We performed the paired t-test using $t = \frac{\bar{d}}{s\_d/\sqrt{n}}$, where $\bar{d}$ denotes the mean of the difference values for either PSNR ($d_{PSNR}$) or SSIM ($d_{SSIM}$), $s\_d$ is the standard deviations, and $n$ is the number of the paired observations (e.g., $10$). We calculated the p-value upon the t-distribution for a two-tailed test using the formula p-value$= 2 \times P(T>\|t\|)$, where $P(T>\|t\|)$ denotes the probability that a t-distributed random variable with $n-1$ degrees of freedom exceeds the absolute value of the observed t-statistic. For PSNR, we observe $t=2.50$ and p-value$=0.034$. Since the p-value is less than the typical significance level of $0.05$. Therefore, we reject the null hypothesis ($H_0$) and conclude that there is a statistically significant difference between the PSNR values of FedAvg and FedHP. For SSIM, we observe $t=7.39$ and p-value$=0.00004$. The p-value of is significantly less than $0.05$, indicating a very strong statistically significant difference between the SSIM values of FedAvg and FedHP. The test results in PSNR and SSIM confirms that the performance gap between FedHP and FedAvg is statistically significant. We thank the reviewer for the valuable feedback, which has enhanced the quality of our submission. We will add the above discussions into the final version and look forward to any further suggestions from the reviewer. We sincerely appreciate the reviewer’s time and effort. --- Rebuttal 2: Comment: The reviewer is willing to acknowledge the amount of work done by the authors to prove their point, and has decided to update the recommendation. Please update, however, your references to mention that significant future work will be done in the direction of lensless cameras, and CT reconstruction, i.e., in the presence of hardware-defined forward models. This must also include thoroughly all hyperspectral and multispectral compressive imaging strategies that include, among others, random masks applied in a number of ways, including those by random convolution, if the authors see them as feasible. In absence of other forward models in the current paper, it is important that the authors update the manuscript accordingly with a comprehensive set of references on what forward models could be studied in future work, that should be supported by experimental evidence. Thanks again for addressing my concerns. --- Rebuttal Comment 2.1: Title: Response to Reviewer aeby Comment: We thank the reviewer for acknowledging the response and for updating the recommendation. In line with the reviewer’s suggestions, we will revise the manuscript to outline potential future work in areas such as lensless cameras, LiDAR, HDR cameras, and CT reconstruction, especially concerning hardware-defined forward models. We will also include a comprehensive set of references on related hyperspectral and multispectral imaging strategies, incorporating forward models that could be explored in the future work. We much appreciate the reviewer's insightful suggestions, which have significantly contributed to the enhancement of the paper's quality and scope!	Summary: The paper addresses the challenges faced in snapshot compressive imaging (SCI) systems due to hardware shifts and the need for adaptability across multiple hardware configurations. By introducing a hardware-prompt network and leveraging federated learning, the framework enhances the adaptability and performance of SCI models across different hardware configurations. Strengths: 1. The manuscript is well-organized with a clear and logical structure that enhances the readability of the content. 2. The paper provides a detailed background on SCI and FL. The planned release of the Snapshot Spectral Heterogeneous Dataset (SSHD) will significantly aid future research. 3. Using different coded apertures for different clients closely mirrors real-world scenarios, adding significant practical relevance to the study. Weaknesses: 1. The literature review on federated learning (FL) heterogeneity in the Introduction section lacks comprehensiveness. There are numerous recent papers addressing heterogeneity in FL that are not cited here. Additionally, the references included are somewhat outdated. Including more current and diverse references would strengthen the review and provide a more accurate context for the study. 2. the manuscript explains that the coded apertures for each client follow a specific distribution Pc, it does not provide further details about the exact nature or type of this distribution. 3. There are many ways to partition data to construct heterogeneous scenarios, such as practical and pathological methods. The approach of equally splitting the training dataset according to the number of clients is not very convincing. The authors should try different partitioning methods. 4. It is unclear which datasets were used to obtain the experimental results in Tables 1 and 2. The authors did not specify this, which creates confusion in the experimental analysis. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. What is the rationale for using adaptors, and what is their function? 2. What network models are used in the comparison methods? It is necessary to clearly state the fairness of the validated methods. 3. The explanation for Figure 3 is not detailed enough. For example, what is "Patch"? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: 1. In the "Discussion of the client number" section, the number of clients increases very little, and the metrics slightly decline. However, the authors conclude that the performance is stable with the change in the number of clients. The small variation in the number of clients is unconvincing. A larger difference in the number of clients should be set to demonstrate this more effectively. 2. The authors mention the "presence of data privacy" in the contributions, but there is no further discussion or experimental comparison regarding data privacy in the subsequent sections. This makes it difficult to validate their contribution to data privacy protection. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We much appreciate that the Reviewer `1X5M` provides valuable comments and finds the proposed work have significant practical relevance to the study and the collected SSHD dataset can benefit the future research. We will release the dataset and the training/testing code. `R3.1`: The literature review on federated learning (FL) heterogeneity in the Introduction section lacks comprehensiveness. Additionally, the references included are somewhat outdated. `A3.1`: We appreciate the reviewer's commitment in helping improve our draft. We will include more current and diverse references to align with the latest development. `R3.2`: the manuscript explains that the coded apertures for each client follow a specific distribution Pc, it does not provide further details about the exact nature or type of this distribution. `A3.2`: We visualize the distributions of the real-world masks collected from real CASSI systems in Fig. 5$\sim$7. The coded apertures for each client follow specific distributions and can reflect the practical variations, including perturbations, shaking, and potential replacements. `R3.3`: There are many ways to partition data to construct heterogeneous scenarios, such as practical and pathological methods. The approach of equally splitting the training dataset according to the number of clients is not very convincing. The authors should try different partitioning methods. `A3.3`: We'd like to clarify that our approach already incorporates both practical and pathological methods to construct heterogeneous scenarios by considering heterogeneity stemming from both the dataset and hardware variations. * Practical scenario (Table 1): Different clients have non-overlapping masks sampled from the same distribution. This simulates real-world hardware perturbations or imperfections within the same system design, which is a common practical scenario. * Pathological scenario (Table 2): Each client has masks sampled from a specific distribution. This represents a more extreme case of significant hardware variations or replacements across institutions, which can be considered pathological. Besides, the equal splitting of the training dataset across clients, combined with these hardware variations, creates a comprehensive heterogeneous environment. `R3.4`: It is unclear which datasets were used to obtain the experimental results in Tables 1 and 2. The authors did not specify this, which creates confusion in the experimental analysis. `A3.4`: All our experiments use the CAVE dataset for the training and the KAIST dataset for testing, both of which are prevalent benchmarks in the field. We will make it clear in the manuscript about the dataset settings. `R3.5`: What is the rationale for using adaptors, and what is their function? `A3.5`: The adaptor in FedHP (1) allows efficient fine-tuning of pre-trained backbones, enhancing the model's ability to adapt to new clients. (2) The adaptor significantly reduces communication costs in the FL system by only requiring the transmission of adaptors rather than entire model backbones. As shown in Table 3, the adaptor brings a 0.07dB improvement in PSNR and 0.0029 boost in SSIM. The adaptor is a CONV-GELU-CONV structure governed by a residual connection (L207). `R3.6`: What network models are used in the comparison methods? It is necessary to clearly state the fairness of the validated methods. `A3.6`: All compared methods use the same network architecture MST [R1] for the reconstruction backbones and a SwinIR [R2] block for the prompt network (see L240\~L241). The key differences lie in how each method handles the federated learning aspect and hardware variations. In Table 3, we demonstrate the comparable complexity and cost between FedHP and FedAvg. [R1] Mask-guided spectral-wise transformer for efficient hyperspectral image reconstruction. In CVPR 2022. [R2] Swinir: Image restoration using swin transformer. In ICCV 2021. `R3.7`: The explanation for Figure 3 is not detailed enough. For example, what is "Patch"? `A3.7`: The ``Patch'' corresponds to the small regions (such as a, b in Fig.3 top-left) that are randomly cropped for the spectral accuracy evaluation. We will provide more detailed explanations in Figure 3. `R3.8`: In the "Discussion of the client number" section, the number of clients increases very little, and the metrics slightly decline. However, the authors conclude that the performance is stable with the change in the number of clients. The small variation in the number of clients is unconvincing. A larger difference in the number of clients should be set to demonstrate this more effectively. `A3.8`: While the performance slightly declines with more clients, FedHP maintains its advantage over FedAvg (0.21dB improvement for C=5 and 0.19dB for C=10). We conjecture that the slight descent tendency might attribute to more complex corporations among the clients and less sufficient training samples for each client. It is non-trivial to collect real hardware systems due to the privacy concern and the cost. FedHP is dedicated to facilitate this process by offering a way of cross-silo cooperation. We are also working on collecting larger-scale data and real hardware. `R3.9`: The authors mention the "presence of data privacy" in the contributions, but there is no further discussion or experimental comparison regarding data privacy in the subsequent sections. This makes it difficult to validate their contribution to data privacy protection. `A3.9`: Thanks for the valuable comment. While we don't explicitly discuss data privacy in our experiments, it is inherent to the federated learning framework we employ. FedHP, like other FL methods, ensures that raw data never leaves the local clients, addressing privacy concerns by design. We will add a dedicated subsection to discuss how our method preserves data privacy and compare it with centralized approaches in terms of privacy protection. --- Rebuttal Comment 1.1: Comment: Thanks to the authors for addressing my concerns. --- Reply to Comment 1.1.1: Title: Response to Reviewer 1X5M Comment: We appreciate the reviewer's valuable comments. We thank the reviewer's recognition of our rebuttal! --- Rebuttal 2: Comment: The authors have addressed most of my concerns; however, two issues remain unresolved, so I will keep my rating unchanged. 1) Regarding the discussion of the client number, Table 4(a) in the paper shows that when C=5, FedHP outperforms FedAvg by 0.27 dB. However, in the rebuttal (Section A3.8), this performance gap is reported as 0.21 dB. The authors have not provided an explanation for this discrepancy. 2) The authors mentioned in the rebuttal that they would add a detailed description of Privacy Protection, but this has not been presented. --- Rebuttal Comment 2.1: Title: Response to Reviewer 1X5M Comment: We appreciate the reviewer's constructive comments and the recognition of our rebuttal. The presentation of 0.21dB improvement for C=5 in `A3.8` was a typographical error and should be consistent with the manuscript, which correctly states a 0.27 dB improvement for C=5. Our intention was to highlight the advantage of the proposed FedHP and FedAvg as presented in Table 4 (a) for C=5. Additionally, we provide a detailed description of the privacy protection inherent in our approach as follows. FedHP inherently addresses privacy from different perspectives. (1) Hardware decentralization: In the FedHP framework, real hardware configurations (e.g., real masks) remain confidential to the local clients. This design makes it difficult to reverse-engineer the pattern or values of the real mask without direct sharing. (2) Raw data decentralization: FedHP maintains a private hyperspectral dataset for each client. The hyperspectral images are processed locally (e.g., encoding or data augmentation) and never leaves the client, thereby minimizing the risk of exposure. (3) Training process decentralization: FedHP only collects the local updates from the prompt network, which are then shared with the central server. The local updates are anonymized and aggregated without accessing underlying data, preventing any tracing back to the data source and thus protecting confidentiality. In Table 3, we quantitatively compared the performance of the proposed “FedHP” and “FedHP w/o FL” under privacy-constrained environments. FedHP demonstrates a $0.6$ dB average improvement (e.g., $31.35$ v.s. $30.75$), showcasing its robust model performance and offering a significant privacy advantage that aligns with regulations restricting data sharing. We will add the above discussions into the manuscript. We sincerely expect our response can help solve the reviewer’s concern and expect a future discussion with the reviewer!	Summary: Most existing reconstruction models in snapshot compressive imaging systems are trained using a single hardware configuration, making them highly susceptible to hardware variations. Previous approaches attempted to address this issue by centralizing data from multiple hardware configurations for training, but this proved difficult due to hardware heterogeneity across different platforms and privacy concerns. This paper proposes a Federated Hardware-Prompt Learning (FedHP) framework, which aligns data distributions across different hardware configurations by correcting the data distribution at the source, thereby enabling the trained model to adapt to multiple hardware configurations. The performance on existing datasets shows an improvement compared to previous popular training frameworks. Additionally, the authors have released their own created dataset and code. Strengths: 1.Previous work focused on the data itself, directly correcting various types of data through network models. In contrast, the authors of this paper focus on the root cause of the differences—hardware. They address the issue from the perspective of learning the differences in hardware. 2.The method proposed by the authors has achieved excellent performance compared to existing mainstream methods, and the average performance has also improved. Weaknesses: 1.The number of clients used in the experiments is still relatively small. Although a simple comparison of the impact of different numbers of clients was made, there is not much difference in performance compared to other methods when the number of clients is larger. 2.Although good results were reported on simulated data, more results on real data should be included to evaluate the effectiveness of the proosed method. Technical Quality: 3 Clarity: 3 Questions for Authors: Why does the prompter lead to such a significant improvement, while the effect of the adaptor is not as pronounced? Please provide an in-depth analysis. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The generalization to different hardware systems is crucial for deep learning based methods. The current form of this manuscript only reported on a small scale real dataset captured by several systems. A larger dataset captured by more systems is necessary to evaluate the method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We much appreciate that the Reviewer `Jxiy` provides valuable comments and finds the proposed method addresses the issue from the new perspective of hardware with good performance. `R2.1`: The number of clients used in the experiments is still relatively small. Although a simple comparison of the impact of different numbers of clients was made, there is not much difference in performance compared to other methods when the number of clients is larger. `A2.1`: We find FedHP maintains its performance advantage even with increased client numbers (C=5 and C=10), outperforming FedAvg by 0.21dB and 0.19dB respectively as shown in Table 4 (a). These results demonstrate FedHP's scalability and consistent performance improvements across varying numbers of clients. Besides, it is non-trivial to collect real hardware systems due to the privacy concern and the cost. We are dedicated to facilitating this process by offering a way of cross-silo cooperation using FedHP. We are still working on collecting larger-scale data and real-world hardware systems. `R2.2`: Although good results were reported on simulated data, more results on real data should be included to evaluate the effectiveness of the proposed method. `A2.2`: We provide additional comparisons on real data in Fig. 7$\sim$8. We are still working on collecting more data. `R2.3`: Why does the prompter lead to such a significant improvement, while the effect of the adaptor is not as pronounced? Please provide an in-depth analysis. `A2.3`: Thanks for the valuable comment. The prompter plays a crucial role in capturing hardware-specific characteristics and aligning inconsistent data distributions across clients. It directly addresses the heterogeneity rooted in the input data space, which is a key challenge in federated learning for SCI systems. By comparison, the adaptor, while important for efficient fine-tuning, has a more subtle effect on performance. Its primary role is to reduce communication costs in the FL system by allowing us to use pre-trained backbones and only communicate adaptors. `R2.4`: The generalization to different hardware systems is crucial for deep learning based methods. The current form of this manuscript only reports on a small scale real dataset captured by several systems. A larger dataset captured by more systems is necessary to evaluate the method. `A2.4`: We kindly illustrate that current setup can consider a large amount of the real-hardware under two challenging scenarios. * Same distribution: As shown in Table 1, we sample non-overlapping masks from the same distribution for different clients. This simulates hardware perturbations or imperfections within the same system design. * Different distributions: In Table 2, we explore a more challenging scenario where each client has masks sampled from a specific distribution. This closely mimics diverse real-world scenarios, including hardware replacement or significant variations across institutions. These experiments, conducted on real hardware masks with noise (Fig. 9-11 in supplementary), demonstrate FedHP's ability to generalize across various hardware configurations. We appreciate the reviewer's insights on collecting larger dataset. This work laid the foundation for future extension by collecting a Snapshot Spectral Heterogeneous Dataset built upon multiple practical SCI systems for the first time. We are still working on collecting more data and real systems for better evaluation. --- Rebuttal Comment 1.1: Comment: Thanks for the explanations. Concerns like real-system evaluation cannot be addressed in such a short period. I keep my initial rating. --- Reply to Comment 1.1.1: Title: Response to Reviewer Jxiy Comment: We appreciate the reviewer's recognition of our response and support for our work!	Summary: The paper introduces FedHP, a reconstruction method for snapshot compressive imaging systems, which addresses the challenge of cross-hardware learning by proposing a federated learning approach. The key contribution lies in using a hardware-conditioned prompter to align data distributions across different hardware configurations, thereby enhancing the adaptability of pre-trained models without compromising data privacy. Strengths: 1. The writing of the paper is good, making it easy to read and follow with clear arguments. 2. The problem defined in the paper is novel with a clear motivation, providing good inspiration for solving the issue of inconsistent device configurations in snapshot compressive imaging. 3. The proposed method is clear and the conclusions are relatively convincing. Overall, it is an interesting work. Weaknesses: 1. There are some typos in the writing. For example, the caption of Figure 3 and the bold parts in the second row of Table 1 and the eighth row of Table 2 are confusing. 2. The proposed FedHP method is relatively straightforward and lacks deeper insights. Moreover, it does not show a significant performance improvement compared to FedAvg. 3. The experiments are not comprehensive enough. Given that this work aims to address the snapshot compressive imaging (SCI) problem, I suggest adding experiments to test the applicability of other SCI systems, such as Coded Aperture Compressive Temporal Imaging (CACTI). 4. There is a lack of sufficient real-world experiments. It would be beneficial to set up multiple independent SCI systems to test the algorithm's performance. Including reconstruction results obtained from these real-world systems is recommended. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. All the experiments in this paper are based on the SD-CASSI model. Can the same FedHP model be simultaneously applicable to both DD-CASSI and SD-CASSI architectures, which have significantly different designs? 2. Although the proposed method outperforms other algorithms in terms of performance metrics, there are still many artifacts in the reconstructed images. While I understand that this is maybe due to the precision issues of the CASSI system, it is crucial for evaluating the practical usability of the algorithm. Additionally, I am not sure whether the spectral accuracy of the reconstructed images is also optimal in statistical terms, which is vital for spectral imaging systems. 3. Furthermore, if possible, I hope the authors can also address the concerns I raised in the Weaknesses section. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We much appreciate that the Reviewer `BdgG` provides valuable comments and finds the problem novel and the method convincing. `R1.1`: There are some typos in the writing. For example, the caption of Figure 3 and the bold parts in the second row of Table 1 and the eighth row of Table 2 are confusing. `A1.1`: We will correct the caption of Fig. 3 and the bold parts in Tables 1 and 2 to improve clarity. We will thoroughly proofread the manuscript to eliminate other errors. `R1.2`: The proposed FedHP method is relatively straightforward and lacks deeper insights. Moreover, it does not show a significant performance improvement compared to FedAvg. `A1.2`: FedHP offers insightful contributions: * It explicitly models hardware variations in computational imaging systems, which is non-trivial and addresses a key practical challenge. * The plug-and-play prompt network allows learnable hardware configurations, enabling hardware-software co-optimization. * The combination of prompt learning and adaptors reduces communication costs in federated learning systems. We empirically find that FedHP enables a consistent performance boost over FedAvg on different number clients, demonstrating the scalability of the proposed method. Considering that FedAvg is a strong baseline that even outperforms prevalent FL methods such as FedProx and SCAFFOLD, it is non-trivial to achieve performance boost over FedAvg. `R1.3`: I suggest adding experiments to test the applicability of other SCI systems, such as Coded Aperture Compressive Temporal Imaging (CACTI). \| Metrics \| FedAvg \| FedHP \| \|:-------------:\|:-------------:\|:--------------------------:\| \| PSNR \| $27.35\pm1.22$ \| $27.87\pm0.89$ \| \| SSIM \| $0.9174\pm0.0046$ \| $0.9192\pm0.0047$ \| Table T1. Comparison between FedAvg and FedHP on CACTI (Client:3) `A1.3`: We launch additional experiments by applying FedHP to the real CACTI systems. The results in Table T1 above present a better reconstruction performance of FedHP over FedAvg baseline (under the same setting as Table 1), demonstrating the generalizability ability of FedHP over a new SCI architecture. We will add the above discussion in the manuscript. Due to the limited rebuttal time and the heavy workload in deploying real CACTI systems, we can only obtain the intermediate results of both methods (e.g., $1.2\times 10^4$ out of $4\times 10^4$). We are committed to updating the latest results during the discussion stage. `R1.4`: There is a lack of sufficient real-world experiments. It would be beneficial to set up multiple independent SCI systems to test the algorithm's performance. `A1.4`: We kindly illustrate that all experiments are performed upon multiple independent real-world SCI systems. For both training and testing, we not only consider sampling non-overlapping masks from the same mask distribution, but also explore a challenging scenario where each client can have real masks sampled from a specific distribution, mimicking the diverse real-world scenarios including hardware perturbation and replacement (L27). `R1.5`: All the experiments in this paper are based on the SD-CASSI model. Can the same FedHP model be simultaneously applicable to both DD-CASSI and SD-CASSI architectures, which have significantly different designs? `A1.5`: It is challenging to apply FedHP simultaneously to both DD-CASSI and SD-CASSI architectures due to their different optical designs and forward models. DD-CASSI uses two dispersive elements and a coded aperture, while SD-CASSI uses a single dispersive element and a coded aperture, resulting in distinct measurement formation processes. These differences lead to incompatible forward models, making it difficult to unify them under a single reconstruction model. However, we agree this is an interesting direction for future research. `R1.6`: There are still many artifacts in the reconstructed images. While I understand that this is maybe due to the precision issues of the CASSI system, it is crucial for evaluating the practical usability of the algorithm. `A1.6`: Assessing the practical usability of SCI algorithms is inherently difficult due to the complex interplay of hardware limitations, reconstruction algorithms, and application-specific requirements. The proposed FedHP method itself is actually a step towards uncovering and tackling these practical challenges. By explicitly modeling hardware variations and enabling cross-hardware learning, FedHP aims to address real-world issues that arise when deploying SCI systems across multiple devices or institutions. `R1.7`: I am not sure whether the spectral accuracy of the reconstructed images is also optimal in statistical terms, which is vital for spectral imaging systems. \| Methods \| Scene 1 \| Scene 2 \| Scene 3 \| \|----------\|----------\|----------\|----------\| \| FedAvg \| 0.9816 \| 0.8905\| 0.8634 \| \| FedHP \| 0.9901 \| 0.9523 \| 0.8851 \| Table T2. Spectral accuracy comparison on three testing scenes. `A1.7`: We provide the spectral accuracy comparison in Fig.3, 5, and 6. Specifically, we randomly select a visually informative region (patch) on the hyperspectral images and compute the density values for a specific wavelength (i.e., dividing pixel values on this specific wavelength by the summarized pixel values along all the wavelengths). We statistically measure the spectral accuracy by computing the correlation of the density values between the prediction and the ground truth. FedHP enables better spectral accuracy by comparison. We further perform a statistical experiment on three simulation data, on which we randomly choose 10 spatial regions for the averaged correlation computation. As shown in Table T2 above, FedHP consistently enables a higher averaged correlation over FedAvg, indicating a better spectral accuracy. We will add the above discussions into the final version. --- Rebuttal Comment 1.1: Comment: The author has addressed most of my concerns. I will keep my score unchanged. --- Reply to Comment 1.1.1: Title: Response to Reviewer BdgG Comment: We appreciate the reviewer's approval for our response and recognition of our work!	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
PERIA: Perceive, Reason, Imagine, Act via Holistic Language and Vision Planning for Manipulation	Accept (poster)	Summary: The paper proposes a framework that integrates large multimodal language models (MLLMs) and diffusion models to enable holistic language planning and vision planning for long-horizon robotic manipulation tasks with complex instructions. The authors jointly train the MLLM and diffusion model for language reasoning and visual imagination through latent image token generation. An explicit consistency loss aligns the reasoned instructions with the imagined subgoal images. Strengths: 1. Novel motivation for integrating of multiple modalities for providing better guidance. 2. Principled design of the framework components like the encoding-side alignment and the latent image token generation approach. Weaknesses: 1. Weak experimental evaluation (see below questions). Technical Quality: 3 Clarity: 3 Questions for Authors: 1. While the authors acknowledge that training and inference costs are significant, the current draft lacks a more in-depth analysis of these problems. In particular, what are the various tradeoffs associated with different MLLMs that can be used, taking into consideration training time/FLOPs/MACs? How does varying these choices impact performance? Experiments answering these questions are equally as important as the ablations being run on training design choices (e.g. alignment loss). 2. Lack of real-world evaluation. Many works ([1], [2]) in this problem setting leveraging foundation models for robotic manipulation demonstrate the advantages of these large MLLMs/generative models in real-world settings, where the distribution of objects is extremely long-tailed. Can the authors show that PERIA can operate with similar success in this regime? [1] [Look Before You Leap: Unveiling the Power of GPT-4V in Robotic Vision-Language Planning](https://arxiv.org/abs/2311.17842) [2] [Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Models](https://arxiv.org/abs/2310.10639) Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors address the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Q1 More analysis of Training Cost & Performance between LLM backbone Thanks for the insightful suggestions! Considering most PERIA's computational cost is to align between vision and language, and bridge the gap between pretrained LLMs and image editing models in general domains versus robotics manipulation domain, we conduct further investigation by substituting enhanced model backbones for both the visual encoder and LLM. We compared the training time and performance until convergence and utilized CalFlops, an open-source tool designed for calculating FLOPs and MACs for neural networks, with results summarized as follows: \|Model Variant\|Training Time (Perception) ↓\|Training Time (Reason+Imagine) ↓\|FLOPs ↓\|MACs ↓\|Success Rate ↑\| \|--\|--\|--\|--\|--\|--\| \|PERIA(ViT+Vicuna7B，reported in paper)\|8 hrs\|42 hrs\|1815.38\|907.3\|68.2\| \|PERIA(ViT+Vicuna 13B)\|13 hrs\|66 hrs\|3679.82\|1810.7\|71.9\| \|PERIA(ViT+llama2 7B)\|8 hrs\|41 hrs\|1740.89\|874.8\|69.0\| \|PERIA(ViT+llama3 8B)\|7 hrs\|39 hrs\|1926.17\|962.2\|73.1\| \|PERIA(LIV+Vicuna 7B)\|5 hrs\|38 hrs\|1804.60\|899.1\|69.2\| \|PERIA(LLaVA1.5 7B)\|4.5 hrs\|32 hrs\|1814.59\|903.0\|70.3\| Key Findings: - LLM Capabilities Impact Performance and Efficiency: Stronger LLMs generally improve PERIA's performance. Replacing Vicuna-7B with models like Vicuna-13B, LLaMA2-7B, and LLaMA3-8B led to varying improvements. Training time is influenced by parameter numbers and inherent LLM capability. Despite its larger size, Vicuna-13B underperformed LLaMA3-8B in both performance and training efficiency due to higher parameters and lower capability. Conversely, LLaMA3-8B, with more parameters than Vicuna-7B, showed reduced training time, likely due to its stronger general-domain knowledge facilitating easier fine-tuning. - Pretrained MLLM and Visual Encoder Enhances Performance and Efficiency: Pre-trained MLLMs, such as LLaVA 1.5, which have undergone general domain vision-language alignment, obviate the need for alignment LLM and visual encoder from scratch. PERIA trained on LLaVA 1.5 can significantly reduce training time and improve the overall performance. Similarily, LIV, a robotics-specific representation with fewer parameters than ViT-B-32, leveraged its pre-training on robotics datasets to achieve vision-language alignment, alleviating further alignment efforts. Conclusion: Our findings suggest balancing stronger LLMs' performance benefits against the computational costs of larger parameters. We recommend prioritizing capable LLMs within similar size constraints and favoring MLLMs or models pre-aligned with the robotics domain. Future work will explore more powerful model backbones and efficient fine-tuning techniques to reduce computational costs. Additionally, we aim to investigate lightweight subgoal modalities (such as object masks, bounding boxes, or keypoint tracking) to balance cost and guidance. We also hope PERIA can serve as a foundational model for MLLM-based embodied manipulation research, offering a more efficient starting point than learning from scratch with non-robotics-tuned MLLMs. # Q2 Real Robotics Evaluation Good suggestions! We highly agree on the importance of demonstrating PERIA's capabilities in real-world settings, but we face limitations due to the absence of robotics in our lab. To address this, we use the BridgeData v2 dataset, which features real-world manipulation videos and corresponding action annotations, to evaluate PERIA's potential through: 1. Language Planning Accuracy: Assessing action prediction accuracy against provided annotations. 2. Visual Planning Fidelity: Evaluating instruction-following ability by generating goal images from given action annotations. This approach allows us to isolate and assess PERIA's high-level cognitive and planning abilities in real-world scenarios, excluding low-level policy execution. The results are as follows (We want to clarify that SuSIE is already included as a key baseline of visual planning method in our main paper. ViLA is also mentioned in related work section but not as baseline before rebuttal due to its unavailability of code and its similar training-free paradigm with the PAR approach baseline, differing primarily in the LLM backbone used. For this, we implement the base version of ViLA with GPT4V based on PAR for the limited timewindow): \|\|Language Accuracy↑\|Visual Fidelity↓\| \|--\|--\|--\| \|ViLA\|0.69\|-\| \|SuSIE\|-\|21.2\| \|PEIRA\|0.76\|17.5\| Due to the free-form nature of language annotations in BridgeData v2, we calculate language accuracy with semantic similarity rather than token accuracy, as both metrics show consistent trends in the simulator results presented in Tab.2 of main paper. The visualization example can be found at Fig.6 in uploaded PDF. More details of related work can also be found at Tab.2 in uploaded PDF. Key Findings: - Using the same backbone Instructpix2pix as SuSIE, PERIA generated more coherent and task-relevant images, highlighting the importance of visual tokens and the integration of MLLMs in our approach. - Language accuracy decreased for open-vocabulary tasks in real domains compared to simulator performance for the Sim2Real gap, but PERIA still outperforms ViLA for the enhanced perception pretraining, which improves its ability to correctly interpret and describe complex manipulation tasks in real-world scenarios. Conclusion: We will conduct comprehensive real-world evaluations as soon as resources permit and we may try SimpleEnv as a temporal substitute of real-robot before ready. Additionally, we will focus on collecting and incorporating more real-robot datasets for pretraining, enhancing PERIA's adaptability to real-world scenarios. --- Sincere thanks for your insightful review! We hope our response can address the concerns. --- Rebuttal Comment 1.1: Comment: I have read the authors' rebuttal and comments. Thank you for your detailed response. I maintain my positive rating. --- Reply to Comment 1.1.1: Comment: We sincerely thank you for your recognition of our work! We will incorporate the corresponding details into the updated version. The constructive suggestions really help us improve the quality of the paper! Please let us know if you need any further information or clarification.	Summary: The paper tackles the problem of long-horizon task planning on pick-and-place tasks in the Ravens domain. Given a dataset of trajectories, it first learns the projection to align the vision and language encoder for a multimodal LLM. Then it finetunes both the multimodal LLM and a diffusion model to generate a step action in language, where the diffusion model is used to generate a conditioning subgoal image, which is proposed as an intermediate step that helps with the step action generation in language. Strengths: - The paper is overall well-written and the figures are helpful for understanding the method. Weaknesses: - It is unclear, at least from the experiments in the paper, that the diffusion model is actually useful, especially when the output is still in language space. For example, it seems that the tasks studied in the paper can be easily tackled by a modern multimodal language model (likely even the open-sourced ones), by simply providing the the initial image and appropriate prompting. However, this is missing as an important baseline in the paper (and this does not require additional training data). Furthermore, to demonstrate the effectiveness of an image subgoal in addition to a language subgoal, the evaluation would have to be done on tasks that have subgoals that are difficult to describe in language but easy to describe in visual space, but all the evaluated tasks are the contrary. - A related work “Video Language Planning” also seems to be missing from the paper, despite it might involve closed-sourced models. However, the idea seems quite relevant and it’s unclear if the paper provides additional insights for the community. Technical Quality: 3 Clarity: 3 Questions for Authors: See "weaknesses" section above. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The limitations are described in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Q1 Effectiveness of Diffusion model & MLLMs & Tasks better described in image Thank you for insightful questions! We appreciate the opportunity to respond point by point: 1. Tasks better described in image We highly agree that tasks with subgoals better described in visual space are crucial. Our evaluation spans three task types. For Blocks & Bowls and Letters, the textures, colors, and backgrounds are relatively simple and straightforward to describe verbally. However, from VIMA-BENCH, which is specialized with multimodal instructions (interleaved image and language), we selected 8 long-horizon tasks categorized into three types: - Rearrange: requires precise absolute positioning, which is challenging to specify accurately through qualitative language alone, image with subgoal location is more clear. - Constraint: involves explicit intermediate or outcome constraints. Language instruction are often lengthy and unclear. - Follow: infers related objects from given example and reason. Object shapes, colors, and textures are often difficult to describe verbally, especially with multiple distractors. We believe these tasks are challenging to clearly described in language only and more suitable for image subgoals. For illustrating examples, please refer to Fig.2 in uploaded PDF. 2. Effectiveness of Diffusion model: - Providing holistic planning capabilities: Evaluations across all task types demonstrate the performance gains of holistic planning over language planning methods, with improvements of +19.4% in Blocks & Bowls and +17.4% in Letters. The advantages of holistic planning are even more pronounced in VIMA-BENCH with +37.1%, highlighting the difficulty of adequately describing subgoals using language alone, as shown in Tab.1 of main paper. PERIA's own ablation studies on subgoal modality also support similar conclusions, shown in Fig.3 in uploaded PDF. More detailed subgoal guidance always works. - Enhancing language planning accuracy: Diffusion model, serving as a visualization module, introduces pixel-level supervision from groundtruth subgoal images to the MLLM. This integration, along with a consistency loss, forms an intermodal bridge that enables language planning to benefit from pixel-level guidance, thus avoiding isolated learning within language supervision.Our ablation studies, which excluded diffussion model and entirely removed visual supervision from PERIA, resulted in a decrease in language planning accuracy, shown in Tab.2 of uploaded PDF. This demonstrates that image generation as an auxiliary prediction task enhances the MLLM's understanding of task instructions and scenes, thereby improving reasoning capabilities, proving the phrase "What I cannot create, I do not understand". 3. Capability of pre-trained MLLMs: Good idea! We tested several popular MLLMs, including GPT4V, Claude 3.5 and InternVL-2 with web interfaces. And we conducted a comprehensive evaluation of GPT-4V key and LLaVa 1.5. The results are shown in Tab.2 of uploaded PDF. Due to domain gaps between robotics and general domain, these models often misidentified scene object properties, directly impacting reasoning, shown in Fig.5 in uploaded PDF. While GPT-4V demonstrated superior accuracy among MLLMs, it still underperformed compared to the fine-tuned EmbodiedGPT and PERIA. We do not expect to achieve superiority over pretrained MLLMs in general domain. Rather, we highlight that deploying MLLMs in robotics manipulation requires substantial domain-specific data fine-tuning and foundational capability enhancements (see LLaVa 1.5 after finetune). # Q2 Discussion with VLP Good suggestions! We now make a detailed comparison between PEIRA with VLP. Similarities: 1. Motivation: Leverage a generative model to generate subgoal with modalities beyond language to offer more sufficient guide 2. Overall pipeline: VLM/MLLM + generative model + low-level policy 3. Necessity for fine-tuning: VLP finetunes PALM-e 12B and PERIA finetunes Vicuna 7B both using addtional robotics domain dataset. Key Differences: 1. Training Paradigm: VLP employs decoupled training for reasoning and imagination, with generated videos subject to VLM evaluation before adoption. In contrast, PERIA utilizes joint training of MLLM and diffusion model, aligning their latent spaces and employing a consistency loss to encourage semantically matching language and visual planning. 2. Subgoal Modality: VLP generate short-horizon video as subgoal, while PERIA generate keyframe images, offering a more lightweight representation. 3. Low-Level Policy: VLP offers three methods, including UniPi's inverse dynamics model and two variants of policy conditioned on either last frame or every frame of generated videos. PERIA's approach is most similar to VLP's last-frame conditioned policy, while avoiding the high fidelity and continuity requirements of the video generation, particularly the inverse dynamics approach. Further Discussion: The core motivation of both video and image subgoals is to provide sufficient guidance for action prediction. These subgoals can be represented through various modalities, including language, images, or videos, each offering different trade-offs between expressiveness and computational efficiency. Broadening our perspective, can we explore more lightweight modalities to describe these subgoals, such as object&place mask, object bbox or keypoint tracking? These approaches could potentially reduce computational cost by focusing prediction on relevant areas but preserve essential semantic information for sufficient guidance to low-level policy. We think this is an intriguing direction for future work, offering opportunities to balance off between computational efficiency and guidance richness of subgoal modality. --- Sincere thanks for your insightful review! We hope our response can address the concerns. --- Rebuttal Comment 1.1: Comment: Dear Reviewer tWdi: Thanks again for your valuable comments and constructive suggestions, which are of great help to improve the quality of our work. We sincerely hope that our additional experiments and analysis can properly address the concerns. As the end of the discussion period is approaching, we are keen to read any further valuable feedback after reviewing our rebuttal. If there are any further concerns, questions, or suggestions, please feel free to ask and discuss with us at any time. We are more than willing to response any of them. Thanks to your hard work and insightful suggestions! Best regards, Authors --- Rebuttal Comment 1.2: Title: Response Comment: Thank you for the detailed response, and I appreciate the efforts for the additional experiments, which I think have greatly enhanced the paper. I have raised my recommendation accordingly. --- Reply to Comment 1.2.1: Title: Sincere Thanks for your Time and Effort! Comment: We are deeply grateful for your recognition and invaluable feedback. The constructive suggestions have significantly helped us improve the quality of our paper, and we will incorporate the corresponding details into the updated version. Your positive recognition means a great deal to us, and we truly appreciate it. Thanks for your time and efforts again!	Summary: The paper proposes a holistic vision-language planning method for long-horizon robot manipulation, by learning a multi-modal large language model (MLLM). The MLLM generates interleaved language actions and keyframe images based on language goal and the initial image. Each pair of generated language and keyframe image is used as conditioning of a learned motion policy for robot manipulation. Based on a pretrained MLLM model, the paper first learns a projector to align visual encoding to with language on image captioning tasks tailored to robot manipulation. Then it applies instruction tuning to fine-tune the MLLM, an output projector, and a diffusion model to generate interleaved language and images. Additional, the authors propose another training objective to align the generated language and images. All large models are fine-tuned with LoRA. On simulated robot manipulatio benchmarks, the proposed method outperforms imitation learning, language planning, and vision planning methods. The paper also systematically evaluates capabilities of the MLLM along different axes, and justifies the benefits introduced by each loss design via ablation studies. Strengths: - The paper tackles the important challenge of robot long-horizon planning. The proposed method plans jointly in the language and image space, providing rich information for the low-level policy to condition on. - The paper exploits the capabilities of MLLM to generate language and images for robot manipulation, used with a separate low-level policy. I think this is good practice as MLLM is not naturally suitable to generate robot motion. - The experiments are comprehensive and provide useful information on understanding the capability of the trained MLLM. - The paper is in general well-written and easy to follow. Weaknesses: - The explanation of low-level policy is missing from the main paper. This part is very important - the MLLM outputs language and images only, and it's not clear how these modalities are bridged with robot motion. - The contribution of the alignment loss between generated image and language is not sufficiently justified in the experiment. It will be helpful if the authors can provide the task success rate when the loss is absent. Technical Quality: 3 Clarity: 3 Questions for Authors: - I wonder which of the three pretraining tasks is the most important for vision-language alignment in the context of robot manipulation. It will be interesting if the authors can show some ablation studies on this. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Q1 Details of Low-level Policy Sorry for the confusion. Due to limited space, we placed the training details of low-level policy in Appendix E. Thanks for bringing the attention to critical importance of this section for a comprehensive understanding of the PERIA architecture and we plan to incorporate this information into the updated version of main paper. Here we want to make more explanation about the low-level policy. - Architecture: To enable multi-task learning, we adopt CLIPort, a wildly-used data-efficient end-to-end multi-task learning algorithm for Ravens that leverages an SE(2) action space. CLIPort introduces several variants and we utilize two version that conditioned on image and language respectively. Both utilize CLIP's visual and language encoders to extract embeddings, which are then fused with observation inputs via element-wise product. We directly train these variants with stepwise language sub-instructions and coherent keyframes as inputs, as language conditioned version and image conditioned version respectively. Moreover, to accommodate the simultaneous presence of stepwise language sub-instructions and coherent keyframes from vision planning and language planning, we design a variant that is conditioned on both image and sub-instruction simultaneously, noted as the language-and-image conditioned version. We maintain the output prediction architecture and just modifying the input processing for fair comparison. Specifically, we fuse the language sub-instruction and corresponding subgoal image through a cross-attention block comprising a 4-layer lightweight attention layer with 4 cross-attention heads. The fused embedding is combined with the observation via element-wise product, maintaining the subsequent modules of the original architecture. Detailed comparisons of the three policy architecture versions are presented in Fig. 4 of the uploaded PDF. - Findings: - We investigated performance across tasks with varying horizon lengths of subgoals. Results in Fig. 5(b) in main paper demonstrate that the integration of both modalities provides more comprehensive guidance, significantly enhancing execution accuracy in long-horizon scenarios. - Furthermore, we evaluated our model across diverse task types. As illustrated in Fig. 3 of the uploaded PDF, the additional guidance proves beneficial across all task categories. Notably, tasks from VIMA-BENCH with multi-modal instructions, which are challenging to be sufficient described by language only, exhibited particularly significant performance improvements with the incorporation of subgoal images. Enhancing subgoal guidance through richer modality consistently improves the semantic clarity and accuracy s across diverse task domains. In future work, we aim to investigate more effective policy architectures as low-level policy backbones, such as diffusion policy, to further enhance the PERIA framework's performance and capabilities. # Q2 Contribution of Alignment loss Good suggestions! We conduct more analysis experiments to verify the effectiveness of alignment loss, including the lanugage accuracy, visual fidelity and success rate. The results can be found in Tab.2 in the uploaded PDF. - The bridge between two modality: The consistency loss, an auxiliary regularization term besides the supervision loss, serves as a bridge to enable language planning benefit from pixel-level guidance with subgoal images while allowing visual planning to be influenced by semantic-level supervision from language instructions, thus avoiding isolated learning within each modality. The ablation of the consistency loss degrades performance in both language accuracy and visual fidelity, demonstrating the effect of intermodal connection via consistency loss. - Alleviate conflicts in holistic planning: By incorporating the alignment loss, we strengthen the synergy between visual and language planning within the MLLM, mitigating the risk of the diffusion model and MLLM working in isolation. The generalization evaluation across three levels demonstrate that the integration is particularly beneficial for unseen tasks, where each modality are more likely to output semantics in isolation, potentially resulting in semantic conflicts and task failure. # Q3 Importance of Three Pretraining Tasks Thanks for this awesome suggestions! Our pretraining strategy comprises three tasks, each targeting specific capabilities: 1. Scene Description (SD): static understanding of single frame, including object understanding and spatial relationships. 2. Action Recognition (AR): dynamic understanding of subsequent images and semantic correlation between language instructions and subgoal images. 3. Video Understanding (VU): continuous comprehension across successive subgoal images and mitigating temporal hallucinations for long-horizon. We conducted an ablation study with variants: PERIA (w/o pretrain), PERIA (w/ SD), PERIA (w/ SD + AR), and PERIA (w/ VU) followed as: \|\|Language Accuracy ↑\|Visual Fidelity ↓\|Succees Rate(<= 8steps)↑\|Success Rate(> 8steps)↑\| \|--\|--\|--\|---\|--\| \|PERIA (w/o pretrain) \|80.2\|16.8\|55.5\|42.4\| \|PERIA (w/ SD)\|89.3\|15.7\|63.0\|54.2\| \|PERIA (w/ SD + AR)\|92.6\|13.6\|68.1\|59.3\| \|PERIA (w/ VU)\|84.2\|15.4\|61.0\|57.2\| \|PERIA (w/ SD + AR + VU, Ours) \|97.6\|12.3\|71.2\|66.1\| Key Findings: - SD is most important as the fundamental static comprehension task for other two pretrain tasks. - SD+AR achieve similar performance to the default version, indicating that the SD+AR pairing can develop certain VU capabilities. Conversely, VU as the sole pretraining task performs poorly may because it is too challenging to learn without sufficient foudnation capabilities from SD and AR. - Incorporating VU on top of SD+AR significantly improves the success rate for long-horizon tasks exceeding 8 subgoal steps. --- Sincere thanks for your insightful review! We hope our response can address the concerns. --- Rebuttal 2: Comment: Thank you for the rebuttal. I'm satisfied with the ablation studies that show the advantages introduced by the chosen low-level policy architecture and the alignment loss. It's also great to see how the three pretraining tasks contribute differently to the performance of PERIA. Good work! --- Rebuttal Comment 2.1: Title: Sincere thanks for the valuable recognition of our work! Comment: We sincerely thank you for your recognition of our work! We will incorporate the corresponding details into the updated version. The constructive suggestions really help us improve the quality of the paper! Please let us know if you need any further information or clarification.	Summary: This paper focuses on robotic manipulation with complex instructions. It proposes PERIA, a framework that integrates MLLM and diffusion models to incorporate both language planning and visual planning for long-horizon language-instructed manipulation tasks. Specifically, PERIA first performs a lightweight multi-modal alignment to consolidate the multi-modal perception capabilities. Then, PERIA performs multi-modal instruction tuning, where it outputs both subgoal language descriptions and visual tokens, both of which are fed to a diffusion model to generate subgoal images. PERIA introduces an additional consistency loss between and generated subgoal image and language descriptions. Experimental results demonstrate that PERIA significantly outperforms competitive baselines. Strengths: • This work follows a natural and reasonable pipeline to tackle the manipulation tasks with complex language instructions. Combining language planning and visual generation for manipulation is a sound approach. • The alignment stage empowers the overall capabilities, as demonstrated in the experimental part. • PERIA achieves convincing experimental results compared with previous works. The authors also conduct extensive ablative study to mine more insights. Weaknesses: • End-to-end learning for such a large system requires considerable cost. Such a comprehensive framework may lead to powerful performances but the resources may be a limitation. This paper does not present how much resources PERIA uses or related experiments to address such potential concerns. • One of my concerns is that the consistency objective, which forces the MLLM to output subgoal language descriptions, may suffer from accumulative error. This is because when the generated subgoal image is not the desired image but is a natural image that can be reached within one-step action, the MLLM would learn an incorrect subgoal description. • More literature references and related baselines should be incorporated. • The ablation in visual planning lacks an experiment where PERIA generates subgoal images with either subgoal descriptions or generated visual tokens, which should reveal more insights into what leads to the improvements in visual planning. Technical Quality: 3 Clarity: 3 Questions for Authors: • You generate subgoal images with subgoal descriptions and generate visual tokens. Why not use 1) subgoal descriptions and observation or 2) generated visual tokens alone? The former resembles a world model, and the latter sounds like a decoding of an imagined visual subgoal, both of which sound more natural. I guess you have tried the latter but found it was not as good as adding subgoal language. • What LLM do you use? It is possible that a powerful LLM accounts for superior performance to some extent. Have you compared the LLMs of different works? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors address the limitations at the end of the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Q1 Computation resources Sorry for the ambiguity arising from distributed presentation of computational resource requirements across Appendix. The computational cost of PERIA across three primary stages: Perceive (8 V100 GPUs * 8 hours ), Reason & Imagine (8 V100 GPUs * 42 hours), and Act (single V100 GPUs * 12 hours). For more details of model architecture and training hyperparameters, please refer to Appendix D and E in main paper. More Analysis: Considering most PERIA's computational cost is to align between vision and language, and bridge the gap between pretrained LLMs and image editing models in general domains versus robotics manipulation domain, we conduct further investigation by substituting enhanced model backbones for both the visual encoder and LLM. We compared the training time and performance until convergence, follows as: \|Model\|Training Time(Perception)\|Training Time(Reason + Imagine)\|Success Rate\| \|--\|--\|--\|--\| \|PERIA (ViT + Vicuna 7B, reported in paper)\|8hrs\|42hrs\|68.2\| \|PERIA (ViT + llama2 7B)\|8 hrs\|41 hrs\|69.0\| \|PERIA (ViT + llama3 8B)\|7 hrs\|39 hrs\|73.1\| \|PERIA (LIV + Vicuna 7B)\|5 hrs\|38 hrs\|69.2\| \|PERIA (LLaVA1.5 7B)\|4.5 hrs\|32 hrs\|70.3\| Key Findings: a) More Powerful LLM: LLaMA-3-8B improved performance and facilitated easier training due to its superior common sense understanding. b) Specialized Visual Encoder: LIV, a robotics-specific representation with fewer parameters than ViT-B-32, leveraged its pre-training on robotics datasets to achieve vision-language alignment, alleviating further alignment efforts. c) Pretrained MLLM: Pretrained MLLMs such as LLaVA 1.5, having already undergone general domain vision-language alignment, obviate the need for fine-tuning from scratch, thus expediting adaptation to robotics tasks. Summary： We will continue to explore more powerful model backbones and more efficient fine-tuning techniques to alleviate the computational cost. Additionally, we are interested in investigating more lightweight subgoal modalities (maybe obj mask, bbox or keypoint tracking) to balance computational cost and sufficient guidance in future work. We also hope PERIA can serve as a potential base foundation model for other MLLM-based embodied manipulation research to reduce computational costs, offering a more efficient starting point compared to learning from scratch using MLLMs not fine-tuned with robotics data. # Q2 Consistency loss& Accumulative Error Good questions! We structure our response into three key points. - The case that generated subgoal image is a one-step goal, pointed out by the reviewer, is Granularity Error, which also includings repetitions, skipping, or even backtracking to visited subgoals. We expect PERIA to learn appropriate task decomposition granularity with the supervision loss from groundtruth labels for both language planning and visual planning. When visual and language planning produce semantically consistent errors (e.g., images and language in one step) at the same time, the supervision loss imposes significant penalties still discouraging such outputs even when consistency loss is low. - The MLLM's subgoal description ability mainly stems from action recognition pretraining in the perceive stage, not solely from consistency loss. Instead, we utilize this ability to implement consistency loss as a self-supervised mechanism during the reason and imagine stages, encouraging semantic alignment between generated instructions and subgoal images within MLLM's unified latent space. - The consistency loss, an auxiliary prediction task besides the supervision loss, serves as a bridge to enable language planning benefit from pixel-level guidance while allowing visual planning to be influenced by semantic-level supervision, thus avoiding isolated learning within each modality. This intermodal connection fosters a more holistic and coherent learning process of PERIA. For more details of visual supervise guidance and consistency, please refer to Fig.5(a) in main paper and Tab.1 in the uploaded PDF. # Q3 More Related Work Thanks for the suggestions! We additionally included comparisons based on existing pretrained MLLM methods and video planning approaches in Tab.2 of the uploaded PDF. We will add theses in the revised version of main paper. # Q4 & Q5 Condition Input for Visual Generation Awesome question! We conduct the detailed ablation studied on three generation mechanisms with language subgoal descriptions, generated visual tokens, or a combination of both as condition. The results can be found at Fig. 1 of uploaded PDF. Key Findings: - More tokens as condition input for visual generation is always better but improvement gets marginal when token reachs a threshold. - Visual tokens only lack sufficient semantic, potentially leading to overfitting or training collapse, especially with a limited number of tokens. Increasing the number of visual tokens can alleviate this but it is essentially equivalent to incorporating language tokens. Discarding the inherent high-level semantics of language tokens and add more visual tokens to learn these semantics from scratch is inefficient, requiring more training time and wastefully ignoring exsited language planning output. - Language tokens contains rich semantics but can also get benefit from more visual tokens with visual details that are challenging to describe accurately in language. The combinatorial fusion version can achieve comparable performance with fewer visual tokens, striking an efficient balance between semantic richness and visual precision. # Q6 Different LLMs Sorry for the confusions. We use Vicuna-7B as our LLM backbone. The reviewer's insight is correct - the powerful LLM can bring performance increase and the comparisions can be found at Q1 and the Appendix F.3. --- Sincere thanks for your insightful review! We hope our response can address the concerns. --- Rebuttal Comment 1.1: Comment: Thanks for your detailed response. Though my concerns have not been resolved, I appreciate the efforts in such a detailed rebuttal. I am still positive and will keep my score. --- Reply to Comment 1.1.1: Title: Sincere Thanks for your Time and Effort! Comment: We sincerely thank you for recognizing our work! We will continue to deeply investigate the role of the condition mechanism in image generation and provide more comparative visualizations of the consistency loss ablation. Additionally, we will incorporate detailed visualizations and corresponding analysis of the granularity error into the updated version. We sincerely appreciate your constructive suggestions, which have significantly helped us improve the quality of our paper!	Rebuttal 1: Rebuttal: # General Response --- Sincere thanks to all the Reviewers for the valuable suggestions and recognition of our work! We sincerely appreciate all reviewers' time and efforts in reviewing our paper. We are glad to find that reviewers generally recognized our key contributions and clear presentation of our paper: Method: Novel motivation for integrating of multiple modalities for providing better guidance [Reviewer WmQe]. The proposed approach follows a natural and reasonable pipeline to tackle manipulation tasks with complex language instructions and combining language planning and visual generation is sound approach [Reviewer HJPE]. The method plans jointly in the language and image space and is a good practice [Reviewer C2qN]. Experiments: The paper achieves convincing experimental results compared with previous works and conducts extensive ablative studies to mine more insights [Reviewer HJPE]. The experiments are comprehensive and provide useful information on understanding the capability of the trained MLLM [Reviewer C2qN]. Clearness: The paper is generally well-written and easy to follow [Reviewer C2qN, Reviewer tWdi]. The figures are helpful for understanding the method [Reviewer tWdi]. We also thank all reviewers for their insightful and constructive suggestions, which helped a lot in further improving our paper. In addition to the pointwise responses below, we summarize supporting experiments added in the rebuttal according to reviewers' suggestions. New Experiments: 1. Additional ablation studies on image conditional generation mechanisms, LLM backbone, consistency loss, pretraining tasks. 2. Additional comparisons with pretrained MLLM in robotics manipulation scenario. 3. Additional evaluations in real-robot datasets BridgeData v2. --- We hope these new additions help address reviewers' concerns around computational resources, effectiveness of different components, potential usage in real-world scenarios, and comparison with existing methods. We thank the reviewers for their time and feedback in improving the quality of our work, and we hope the revisions further highlight the contributions made. Please let us know if any clarification or additional experiments would further strengthen the paper. We would be happy to incorporate all these suggestions in the revised version. Pdf: /pdf/9743c0f3586220b97b54fa1baf9a49999d5bd0af.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Toward Global Convergence of Gradient EM for Over-Paramterized Gaussian Mixture Models	Accept (poster)	Summary: The paper studies the convergence of EM for learning mixtures of Gaussians. Specifically, they consider a simplified setting where the Gaussians are in $d$-dimensions and all have covariance $I_d$. They consider an overparameterized version of the problem where they parametrize the mixture they are trying to learn by a mixture of $n$ Gaussians with means $\mu_1, \dots , \mu_n$ and the ground truth distribution generating the data just consists of a single Gaussian $N(\mu^* , I_d)$. The paper analyzes the dynamics of gradient EM for this problem. The main result of the paper is proving that for this overparametrized variant, gradient EM converges to the true distribution at a rate of $1/\sqrt{t}$ with additional constants depending exponentially on the distance between the initialized means and the true mean, which they show is necessary. There has been a long line of work on understanding the convergence of EM or gradient EM for learning mixtures of Gaussians. Without overparametrization, provable convergence is known for mixtures of two Gaussians and it is also known that convergence fails in general for mixtures of three or more components. For overparamterized settings, a previous work [Dwivedi et. al. 2018] shows that if we parametrize a mixture of two Gaussians and try to learn a ground truth distribution consisting of a single Gaussian, then EM converges at a $1/\sqrt{t}$ rate (as long as the mixing weights are set to be different). This is in contrast to when we parametrize with only a single Gaussian and EM converges exponentially fast. The results of the current paper can be seen as generalizing the results of [Dwivedi et. al. 2018] to more than two components. The paper empirically validates their theoretical results with experiments on simple synthetic datasets. Strengths: The paper makes progress on a well-studied problem of understanding convergence of EM for learning GMMs. They give the first global convergence results for mixtures with more than two components. The paper overcomes nontrivial technical barriers to extend previous results to more than two components. Weaknesses: The results of the paper only work when the ground truth is "trivial" i.e. a single Gaussian. The results are qualitatively similar to previous work on overparametrized mixtures of two Gaussians. The contributions of the paper are mostly technical and it is a bit difficult to find a nice conceptual takeaway \--- the previous work for two components already showed that overparametrization can lead to drastically slower convergence. It would be much more exciting and novel, say, if we could prove something when the ground truth were not just a single Gaussian. Technical Quality: 3 Clarity: 3 Questions for Authors: . Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the positive review. We have addressed your concern below. > The results of the paper only work when the ground truth is "trivial" i.e. a single Gaussian. We agree that the single Gaussian ground truth is a simpler case compared to the most general problem. But our setting is nonetheless highly-nontrivial and technically challenging. In fact, even the 2-GMM problem is quite difficult with a large number of previous works (see section 2.1 for details), and our result serves as an important step towards generalizing the 2-GMM analysis to the general k-GMM learning problem. > The results are qualitatively similar to previous work on overparametrized mixtures of two Gaussians. The contributions of the paper are mostly technical and it is a bit difficult to find a nice conceptual takeaway. Our result is fundamentally different from previous works like [Dwivedi et. al. 2018] since the mechanisms for learning general k-component GMM and 2-component GMM are different. [Dwivedi et. al. 2018] only considers 2-component GMM with symmetric means, and their sub-linear convergence only happens when the mixture weights are exactly equal (both weights being exactly $1/2$). On the other hand, our result applies to general GMM with arbitrary weights and asymmetric means. While the phenomenon of slow convergence observed in [Dwivedi et. al. 2018] depends on their specific setting of equal weights and symmetric means, our paper describes the general convergence behavior of k-component GMM. Another conceptual takeaway of our result, as stressed in the paper (see Section 4.1), is that one should consider the likelihood convergence rather than parametric convergence for GMM learning problem. One of our major novelty is the brand new likelihood-based framework, while previous works are mostly standard algebraic computations of the parametric convergence. --- Rebuttal Comment 1.1: Comment: Thank you for the response and addressing my concerns/questions. My overall assessment remains the same. --- Reply to Comment 1.1.1: Comment: Thank you for reading our paper and feedback! Please let us know if you have any further questions.	Summary: This paper talks about the gradient-EM algorithm for over-parameterized GMM. The paper mostly shows the GLOBAL convergence and its rate when using this model to learn a single Gaussian. Strengths: I believe any non-convex global convergence optimization problem is valuable. It is an extension of Dwivedi et al. 2019. Weaknesses: 1. The over-parametrized model may have severe overfitting problem. 2. The based distribution is quite easy: a single normal, with known variance. In the paper, the covariance is fixed as the identity, which simplifies the problem in a deep way. Actually for symmetric 2-GMM, there are already faster algorithms to learn both mean and cov. 3. I feel confused about the consistency and convergence in the paper. In Line 96, the convergence of KL divergence also contains the convergence of MLE, ie consistency. The convergence to the MLE is another loss function. Also in Remark 6, the convergence when sample size to infinity seems more easily ensured by WLLN. Technical Quality: 4 Clarity: 4 Questions for Authors: Besides the weakness above, I also have following questions: 4. If you only learn the single normal, how is the algorithm compared with Dwivedi et al. 2019 or just 2-GMM? Is it necessary to use more? Is it overfitting so the performance seems better? 5. I don’t get why the paper introduces Fact 1. It seems obvious. 6. The mean is convergent to 0 (true) instead of the MLE. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 2 Limitations: Besides above, 7. the citation format is not uniform. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your detailed review! We have addressed your questions below. > The over-parametrized model may have severe overfitting problem. We believe this is a misunderstanding. The aim of this paper is not to propose a new algorithm/model, but to understand the convergence behavior of the widely-used EM/gradient EM algorithm. The motivation for studying the over-parameterized regime is well-known as many have conjectured it might facilitate global convergence (see [1], [2]). Since we are considering population gradient EM, there is no overfitting in our problem setting and we leave the study of generalization theory of GMM to future works. > The based distribution is quite easy: a single normal, with known variance. This paper aims to rigorously study the optimization phenamenon in a cannonical setting. We agree that the base distribution is simple. However, when the learning model has more than one mixtures, the learning dynamics becomes significantly complex. In fact, even the further simplified setting where the learning model has only 2 mixtures and the based distribution is a single normal, the analysis in Dwivedi et al. 2019. is technically very complicated. The known co-variance assumption is also a standard setting widely-adopted in existing literature ([1], [2], [3]). > Actually for symmetric 2-GMM, there are already faster algorithms to learn both mean and cov. Again, our goal is not to design a new algorithm for learning GMM, but to understand the behavior of the widely-used EM/gradient EM algorithm. While there might be specifically designed algorithms achieving better performance for some special cases such as symmetric 2-GMM, we aim to study gradient EM as one of the most popular existing algorithms for learning general k-component GMM. We will ensure this point comes across more clearly in the final version. > I feel confused about the consistency and convergence in the paper. In Line 96, the convergence of KL divergence also contains the convergence of MLE, ie consistency. The convergence to the MLE is another loss function. While MLE and KL divergences are two different loss functions, optimizing them are actually equivalent due to the following well-known fact: $D_{KL}(p(x\|\mu^)\|\|p(x\|\mu))=-\mathrm{E}_{x\sim p(x\|\mu^)}\left[\log\left(\frac{p(x\|\mu)}{p(x\|\mu^)}\right)\right]$ $=-\mathrm{E}_{x\sim p(x\|\mu^)}[\log({p(x\|\mu)})]$ $\quad +\mathrm{E}_{x\sim p(x\|\mu^)}[\log({p(x\|\mu^)})].$ Since the second term above is a constant that does not depend on $\mu$, minimizing KL is equivalent with maximizing the first term, which is just MLE. Note that this is a general fact that applies to any MLE problem, not just EM or GMM. > If you only learn the single normal, how is the algorithm compared with Dwivedi et al. 2019 or just 2-GMM? Is it necessary to use more? Is it overfitting so the performance seems better? Again, since we are considering population EM, there is no overfitting in this problem. Our algorithm is an extension of Dwivedi et al. 2019 and our main goal is not to argue that k-GMM is better or worse than 2-GMM, but to extend the previous theoretical understanding of 2-GMM to the general k-component case. > I don’t get why the paper introduces Fact 1. It seems obvious. Fact 1 implies that gradient EM is just running gradient descent on the likelihood function, an observation allowing us to introduce theoretical tools from gradient descent theory to facilitate our analysis. > The mean is convergent to 0 (true) instead of the MLE. Since the algorithm is run on population data, there is no overfitting and the ground truth 0 is just the MLE solution. > Citation format is not uniform. Thanks for pointing it out. We will also organize our citation style and make it uniform in the revised version. References: - [1]. Yudong Chen, Dogyoon Song, Xumei Xi, and Yuqian Zhang. Local minima structures in gaussian mixture models. IEEE Transactions on Information Theory, 2024. - [2]. Chi Jin, Yuchen Zhang, Sivaraman Balakrishnan, Martin J. Wainwright, and Michael I. Jordan. Local maxima in the likelihood of gaussian mixture models: Structural results and algorithmic consequences. In Neural Information Processing Systems, 2016. - [3]. Sivaraman Balakrishnan, Martin J. Wainwright, and Bin Yu. Statistical guarantees for the em algorithm: From population to sample-based analysis, 2014. --- Rebuttal Comment 1.1: Comment: Thank you for the addressing my concerns! Part of my questions and concerns are clarified by the authors. Although the ground truth is standard 1-d Gaussian, the work is valuable. Therefore, I would like to lift the score to 6. But I still have questions about the MLE and the true (0). The author responds that the algorithm is implemented on the population data. No mentioning the possibility and reality, if the population data available, the model chosen seems inappropriate. --- Reply to Comment 1.1.1: Title: Thank You and Response to the Question on Population Setting Comment: Thank you for recognizing our contribution and for raising the score! > About our choice of the population data model. We study the population setting to focus on the non-convex optimization dynamics of gradient EM algorithm. Indeed, using the population model is a standard approach in previous literature of EM analysis [1, 2], and generally in non-convex optimization [3,4]. As discussed in Remark 6, it also implies (asymptotically) the optimization convergence for reality, i.e., sample-based EM. [1]. Ji Xu, Daniel J. Hsu, and Arian Maleki. Global analysis of expectation maximization for mixtures of two gaussians. In Neural Information Processing Systems, 2016. [2]. Sivaraman Balakrishnan, Martin J. Wainwright, and Bin Yu. Statistical guarantees for the em algorithm: From population to sample-based analysis. In Annals of Statistics, 2014. [3]. Yuandong Tian. An analytical formula of population gradient for two-layered relu network and its applications in convergence and critical point analysis. In International Conference on Machine Learning, 2017. [4]. Mo Zhou, Rong Ge. A local convergence theory for mildly over-parameterized two-layer neural network. In Conference on Learning Theory, 2021.	Summary: The paper focuses on the setting of a Gaussian Mixture Model with several summands and an input vector produced by one Gaussian distribution, where it employs the Expectation-Maximization rule to infer the model's parameters. Since the problem of having arbitrary number of summands has been unsolved, the paper provides an innovative scheme which includes the computation of the likelihood function and shows that the EM algorithm converges with sublinear complexity. The authors also show that there exist neighborhoods of slow convergence rates. Strengths: - The paper is well written, the theorems, lemmata and algorithmic steps are described gradually. - From a first overview of the literature, the result about global convergence seems novel. - Across section 4, there is intuition and remarks provided about the necessity of the steps. Weaknesses: - The experimental evaluation is used as a proof of concept and thus is limited. The authors could have (potentially) experimented with several datasets, with varying weights in the GMM, and try to benchmark their algorithm to compare the emergent convergence rates. Technical Quality: 2 Clarity: 2 Questions for Authors: NA. Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: NA. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your review and positive comment! We have addressed your question below. > The experimental evaluation is used as a proof of concept and thus is limited. The authors could have (potentially) experimented with several datasets, with varying weights in the GMM, and try to benchmark their algorithm to compare the emergent convergence rates. We note that our primary goal is to rigorously study the optimization convergence rate in a controlled setting and therefore we focus on synthetic experiments to carefully examine the phenomena and corroborate our theoretical findings. We also added more experiments in the rebuttal PDF (attached to the global author rebuttal for all reviewers) to verify our theoretical results: - Impact of mixtures weights on the convergence speed (Figure 2 Right of uploaded pdf.) We test $3$ different weight configurations of $3$-component GMM: $(\frac{1}{3}, \frac{1}{3}, \frac{1}{3})$. $(\frac{1}{6}, \frac{1}{3}, \frac{1}{2})$,$(\frac{1}{20}, \frac{1}{5}, \frac{3}{4})$. $4$ runs of each configuration are recorded, with different random initialization. Results show that the convergence is faster when the weights are more evenly distributed: the equally distributed weights of $(\frac{1}{3}, \frac{1}{3}, \frac{1}{3})$ converges with the fastest rate, while $(\frac{1}{20}, \frac{1}{5}, \frac{3}{4})$ converges the slowest. - Impact of initialization on the convergence speed (Figure 2 Left of uploaded pdf): We report the gradient norm in the bad initialization region constructed as counter-examples in Theorem 7. Empirically the gradient norm exponentially decreases in dimension $d$. This supports our theoretical findings that bad initialization causes exponentially slow convergence. --- Rebuttal Comment 1.1: Title: Rebuttal Comment: I thank the authors for their answers. They provided experiments as an attachment to their rebuttal answer. I will further study the responses to the other reviewers and update my review. --- Reply to Comment 1.1.1: Comment: Thanks for reading our response and providing feedback! We look forward to any further comments and updates.	Summary: The paper considers fitting a single Gaussian with multiple-component Gaussian mixture models (GMM) through the Gradient EM algorithm. While the two balanced over-specified Gaussian setting has been widely studied in the previous work, generalizing it to multiple-component GMM requires significant algebraic efforts. The entirety of the paper is to show the $1/\sqrt{t}$ convergence rate of the population EM algorithm. In particular, the paper characterizes the explicit convergence rate of $1/\sqrt{T}$ with constants exponential in the number of components, the phenomenon that coincides with the exponential lower bound for the parameter estimation of general GMMs with no separation. Strengths: - Extending some existing two-component results to general multiple-component GMM is non-trivial and significant. The paper nicely characterizes the convergence rate that captures some important properties of learning GMM that can be achieved by GMM. - The paper is well-written, emphasizing important aspects of the results and well-contrasting their techniques to existing results. - Proof sketch is nicely written to help readers understand their key results. Weaknesses: - While the lower bound result (Theorem 7) is a nice addition to the literature, I believe that the gap between this lower bound and the upper bound is large, since the upper bound is exponentially slow in the number of components. - One important result from two specified GMM is the $n^{-1/4}$ (n is the number of samples here) statistical rate after convergence. I would like to see $n^{-1/2k}$ style results in general k-component GMM settings. At least, the authors should have discussed this aspect of previous work and contrasted the implications to k-GMM settings. - The experiment would have been nicer if the final statistical rates were compared. Technical Quality: 3 Clarity: 4 Questions for Authors: - Maybe authors can elaborate on how their results can imply learning k-GMM with small separations? - In Theorem 7, there is no restriction on the step size $\eta$. I believe that the lower bound should also be able to tell that $\eta$ cannot be set too large. - Why only on the gradient EM? Can the analysis in the paper imply some convergence rates of the standard EM algorithm as well? I think it would make the paper much stronger if it could show that the same results hold for standard EM. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the detailed review. We answer each of your questions below. > The gap between this lower bound and the upper bound is large. Thank you for pointing out this problem. In the initial version we didn't optimize the exponent. Indeed, we can obtain significantly refined results which removes this gap between upper and lower bounds. The improved bounds are as follows: - [Upper bound] Consider training a student $n$-component GMM initialized from ${\mu}(0) = (\mu_1(0)^{\top},\ldots, \mu_n(0)^{\top})^{\top}$ to learn a single-component ground truth GMM $\mathcal{N}(0, I_d)$ with population gradient EM algorithm. If the step size satisfies $\eta \leq O\left(\frac{\exp\left(-8U(0)\right)\pi_{\min}^2}{n^2d^2(\frac{1}{\mu_{\max}(0)}+\mu_{\max}(0))^2}\right)$, then gradient EM converges globally with rate $$\mathcal{L}(\mu(t))\leq \frac{1}{\sqrt{\gamma t}},$$ where $\gamma = \Omega\left(\frac{\eta\exp\left(-16U(0)\right)\pi_{\min}^4}{n^2d^2(1+\mu_{\max}(0){\sqrt{dn}})^4}\right)\in \mathrm{R}^+$. Recall that $\mu_{\max}(0)=\max\{\\|\mu_1(0)\\|,\ldots, \\|\mu_n(0)\\|\}$ and $U(0)=\sum_{i\in[n]}\\|\mu_i(0)\\|^2$. - [Lower Bound] For any $n\geq 3$, there exists initialization points such that, when initialized from, population gradient EM will be trapped in a bad local region around it for exponentially long time $T=\frac{1}{30\eta}\exp(\Theta(U(0)))$ as: $0\leq t\leq T$, $\exists i\in[n]$ such that $$ \\|\mu_i(t)\\|\geq 10\sqrt{d}. $$ Both the improved lower and upper bounds are tighter than original versions since $n\mu_{\max}^2\geq U\geq \mu_{\max}^2$. Most importantly, the improved exponential factors of $\exp(\Theta(U(0)))$ in the two bounds now exactly matches, eliminating the noted large gap. Our key idea is to use $U=\sum_{i\in[n]} \\|\mu_i\\|^2$ (which captures information on all Gaussian means) instead of $\mu_{\max}=\max_i\{\\|\mu_i\\|\}$ (which reflects only the maximum Gaussian mean) to construct our convergence rate bounds, resulting in a more fine-grained analysis and tighter bounds. We also constructed a better counter-example: ${\mu}_1(0)=12\sqrt{d}e_1, {\mu}_2(0)=-12\sqrt{d}e_1, \mu_3(0)=\cdots=\mu_n(0)=0$, where $e_1$ is a standard unit vector. This new construction applies to all $n\geq 3$ (while the original one requires $n$ being odd) and implies a tighter lower bound. With this idea, we are happy to present the optimal exponential factor of the convergence rate (up to a constant), which we believe might be of independent interests to future study. > Statistical rate for 2-GMM is $n^{-1/4}$. I would like to see $n^{-1/2k}$ style results in general k-component GMM settings. The authors should have discussed this. We agree that statistical rates for learning GMM is an important topic. However, our focus is not on this problem since we aim to study optimization rates rather than statistical rates of gradient EM. Yet we are aware that there is a long line of work on this topic. The statistical rate of $n^{-1/4}$ you noted is studied in [1] and [4]. To the best of our knowledge, the same rate for general k-GMM remains an interesting open problem and $n^{-1/2k}$ is a nice conjecture. However, our new experiments suggest that the rate seems to be still $n^{-1/4}$ for general k-GMM. (See response to the next question.) We will add these discussions and our experimental findings into the revised version. > The experiment would have been nicer if the final statistical rates were compared. Thanks for the suggestion! We have added a new experiment on this. See the global author rebuttal to all reviewers (figure 1 in attached pdf) for details. The empirical statistical rate for k-GMM is close to $n^{-1/4}$. > Maybe authors can elaborate on how their results can imply learning k-GMM with small separations? Thanks for noting this topic. Our results immediately imply that the convergence rate for learning general k-GMM with small separations will be sub-linear, since a single Gaussian is a special case of k-GMM with no separation. So the linear-contraction style analysis (such as in [2]) for well-separated GMM no longer works in this regime, and we believe our likelihood-based framework can be helpful. We will add more corresponding discussions in the revised version. > In Theorem 7, there is no restriction on the step size. Theorem 7 applies to any positive step size $\eta$. It punishes large step sizes as $T$ scales with $1/\eta$, so the time that gradient EM gets trapped shortens as the step size increases. As long as the step size is polynomially large $\eta =poly(n,d)$, gradient EM gets trapped in the bad region for exponentially long time as $T\geq \frac{1}{poly(n,d)}\exp(\Theta(U(0)))$. > Why only on the gradient EM? Can the analysis in the paper imply some convergence rates of the standard EM algorithm as well? While gradient EM is equivalent with gradient descent on the likelihood function $\mathcal{L}$, standard EM can also be seen as a gradient descent-like algorithm on $\mathcal{L}$ with a coordinate-dependent step size $1/\mathrm{E}_x[\psi_i(x)]$. (See further discussions in [3]). We believe our method is also useful for standard EM. We will add more discussions on extensions to standard EM in the revised version. References: - [1]. Yihong Wu and Harrison H. Zhou. Randomly initialized em algorithm for two-component gaussian mixture achieves near optimality in $O(\sqrt{n})$ iterations, 2019. - [2]. Bowei Yan, Mingzhang Yin, and Purnamrita Sarkar. Convergence of gradient em on multi-component mixture of gaussians. Advances in Neural Information Processing Systems, 30, 2017. - [3]. Yudong Chen, Dogyoon Song, Xumei Xi, and Yuqian Zhang. Local minima structures in gaussian mixture models. IEEE Transactions on Information Theory, 2024. - [4]. Raaz Dwivedi, Nhat Ho, Koulik Khamaru, Michael I. Jordan, Martin J. Wainwright, and Bin Yu. Singularity, misspecification and the convergence rate of em. The Annals of Statistics, 2018. --- Rebuttal Comment 1.1: Title: Thank you for the response Comment: I thank the authors for the clarification and for addressing my concerns. The additional experimental result on the statistical rate also looks interesting. I have adjusted my evaluation score accordingly.	Rebuttal 1: Rebuttal: We appreciate all the reviewers for their detailed and positive feedbacks. In the uploaded pdf file, we add several experiments: - Experiment of statistical rates, for questions of Reviewer 6yVv (Figure 1). - Impact of initialization on the convergence speed, for questions of Reviewer DCG2 (Figure 2, left). - Impact of GMM mixture weights on the convergence speed, for questions of Reviewer DCG2 (Figure 2, right). Please refer to the pdf for experimental setups and outcomes. We have also improved our theorems, closing the gap between our upper bound and lower bound of convergence rates. This addresses the question of reviewer 6yVv. Please refer to our response to reviewer 6yVv for details. We welcome and are happy to answer any further questions from reviewers. Pdf: /pdf/c241d28bbeaf9edc9c06ba930da437adbf7f2dba.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
IDGen: Item Discrimination Induced Prompt Generation for LLM Evaluation	Accept (poster)	Summary: This paper proposes a method of generating prompts for evaluating large language models such that the prompts are dynamic and allow for showing meaningful performance gaps between different language models.The authors show that the generated data is more-challenging and discriminative than prior datasets. Strengths: - Work is very timely and addresses a major issue in how we can better evaluate LLMs which are continuously improving and saturating existing benchmarks. - Good to see that the generated prompts are indeed harder than baseline datasets - this should indicate that the prompts are challenging enough to provide decent signal on a language model's capabilities. - Experimented with many SOTA models and compared with several baseline datasets. Weaknesses: The main weakness of this work is that much of the pipeline relies prompting language models to modify seed data. This means that the performance of the language model plays a huge role in the quality of the resulting data. Given that the pipeline seems to have many different steps, each of these steps can introduce errors since LLMs are not fully reliable. It then becomes crucial to have a way of verifying that the generated questions are of high quality. There's also a concern that the ground truth answers might not be entirely accurate. The authors mention both of these issues as limitations. Technical Quality: 3 Clarity: 3 Questions for Authors: - If a particular language model is used to generate data using the proposed method, is there any bias where that model will perform better at solving those problems? For example, if Claude generates the prompt set, will the prompt set be easier for Claude than GPT? - Is the data generation done for a set of language models or for each individual language model? In other words, are the prompts being dynamically changed with respect to a single language model's response or all language model responses? Specifically, Section 2.2 says that the method "rephrases the question based on the response from the LLM" - which LLM is this statement referring to? - Are there any experiments to verify the robustness of each individual step in the pipeline? It seems like the current experiments are meant to verify the final output of the pipeline, not the in-between steps. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors mention the most-important limitations of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q: If a particular language model is used to generate data using the proposed method, is there any bias where that model will perform better at solving those problems? For example, if Claude generates the prompt set, will the prompt set be easier for Claude than GPT? A: Thank you for your question, and your question is very professional. In our experiments, we find that for the data generated by Hunyuan, its performance may not be as good as potentially better-performing models, such as GPT4 and Claude3. However, in the data generated by GPT4, there are cases where Hunyuan performs better than GPT4. We cannot entirely confirm that the bias you mentioned does not exist, but even if it does, its impact would not be decisive. To avoid potential biases, we have also done related work and processing. In our paper, we try to minimize this potential bias as much as possible: on the one hand, we pay great attention to the usability of the generated data. During the data production process, we use models to participate in the usability check of the generated data. For the final generated data, we also hire expert personnel to conduct checks. The expert personnel's judgment of the usability of the questions is based on the human preference perspective to determine whether the questions are reasonable. The inspection process includes result sampling or re-inspection schemes to ensure the accuracy of the judgment. The inspection results show that the evaluation data obtained by the method in this paper have ideal usability. Therefore, we can believe that these data can be used to evaluate the performance of LLMs. On the other hand, to avoid suspicion, we don't allow the model that generate the data to participate in the evaluation. In the experiments, the calculation of evaluation results-related data don't include the evaluation results of Hunyuan, to avoid potential biases that may exist in Hunyuan-generated data as much as possible. The evaluation results also confirm that our data can effectively distinguish the performance of existing LLMs. Q: Is the data generation done for a set of language models or for each individual language model? In other words, are the prompts being dynamically changed with respect to a single language model's response or all language model responses? Specifically, Section 2.2 says that the method "rephrases the question based on the response from the LLM" - which LLM is this statement referring to? A: For data production, we only use a individual language model, i.e., Hunyuan-standard. For the automated usability check of mathematical questions, to improve the usability of the data, we use Hunyuan-standard and Hunyuan-pro to perform checks separately and cross-validate the usability of the data. For other processes, we only use Hunyuan-standard. We input the large model's response into the prompt for generating questions, and the large model responsible for generating questions will produce new questions based on the prompt. Therefore, the prompt does not dynamically change with the model's response; it simply incorporates the model's response. While we use Hunyuan, other large models can also be utilized for this purpose. In Section 2.2, "rephrases the question based on the response from the LLM," the LLM refer to in this sentence is Hunyuan (Hunyuan-standard). In the footnote 2 on page 3 of the article, we mention, "Unless otherwise specified, all data in this document are generated by Hunyuan (Hunyuan-standard), which is a Large Language Model developed by Tencent." Therefore, we use LLM as a reference here. Q: Are there any experiments to verify the robustness of each individual step in the pipeline? It seems like the current experiments are meant to verify the final output of the pipeline, not the in-between steps. A: Your question is indeed very insightful. We have not verified the robustness of each independent step; if necessary, we can incorporate relevant experiments in future versions of our work. Theoretically, each individual step in our process is robust. From a global perspective, the data we utilize is both reliable and stable. Our dataset comprises Chinese and English components. For the Chinese data, we reference the work of TencentLLMEval[1], which has been stably applied in various business scenarios. For the English data, we employ the seed data from Self-instruct[2], which is extensively used in academia. Additionally, during the data production process, we set the temperature of the LLMs to 0 and use fixed prompts to guide the LLM in data production. For each data point, we implement several measures when calling the LLM API. These measures include multiple requests using a counter in case of call failures, capturing and handling exceptions, and pausing for three seconds before resubmitting a request if a call fails. These steps aim to increase the probability of successful API calls. Furthermore, we conduct validity checks on the input data for each step to enhance the robustness of the data production process. [1] Xie, Shuyi, et al. "TencentLLMEval: a hierarchical evaluation of Real-World capabilities for human-aligned LLMs." arXiv preprint arXiv:2311.05374 (2023) [2] Wang, Yizhong, et al. "Self-instruct: Aligning language models with self-generated instructions." arXiv preprint arXiv:2212.10560 (2022). --- Rebuttal Comment 1.1: Comment: Thank you for the response. I would be interested in seeing results if a panel of LLMs was used as the evaluator for this method in order to reduce bias. I am also still curious about ways to verify the individual steps within this method. As such, I will be retaining my original score. --- Reply to Comment 1.1.1: Comment: Thank you for your reply. We will include the relevant content you mentioned in the subsequent version of the paper.	Summary: The paper proposes a prompt synthesis framework for evaluating LLMs to accurately reflect different Large Language Model abilities. The authors develop two models to measure LLMs’ question discriminative power and difficulty. This study presents “instruction gradient” and “response gradient” methods to exploit rule sets to generalize questions. Strengths: The paper focuses on the generation of a large number of queries and corresponding answers on general language and mathematical topics. They have released a set of over 3000 questions for LLM evaluation. Their proposed metrics (discrimination index and difficulty score) show significant improvement in the quality of the benchmark datasets. Weaknesses: Although the paper tries to solve a crucial research area in the scope of LLM evaluation, the study lacks in many different ways. The textual flow is difficult to follow. Many of the concepts introduced were not properly described or not cited with previous work’s references. These issues restricted the reviewability of this study. Technical Quality: 3 Clarity: 1 Questions for Authors: 1. The proposed methods - “Instruction gradient” and “response gradient” are not properly described in the manuscript. Authors should write the working procedure of these methods in detail in the main manuscript, as these are the centerpiece of the whole question generation process. 2. “Generalizing questions from seed data based on the "instruction gradient" restricts the diversity and confines the content to specific topics” - is unclear. Consider explaining. 3. In section 2.3 - Assessing the Usability of General Text Questions: How is the assessment done? Is it done manually with human input? Or by an autonomic process/model? 4. In section 2.3 - CoT Check for Mathematical Questions: “we use Hunyuan to assess the reasonableness of the question, which successfully identifies the unreasonableness of the problem and corrects it based on the assessment process.” - How can it be ensured that the model successfully identifies the unreasonableness? Provide a theoretical/experimental study. 5. In section 2.4 - Acquiring reference answers: lines 133-136, are the answers scored by human participants? 6. In section 2.4 - Acquiring reference answers: line 140, What is meant by a “collective voting mechanism”? Please explain clearly. 7. In section 2.5 - lines 148-149, what are “label discrimination indexes”? a. In line 149, “the prompt includes four features” - How did you select these features? Provide some analysis. b. In lines 162-164, How did you select the threshold values? (e.g., “Low” means less than or equal to 0.1, “High” means values greater than 0.25, etc.). c. In line 168, “discrimination level label ranging from 0-3” - Is this range acquired by observations? Or have you performed some analyses on the score expressions? 8. In equation 4, what does the “score” mean? Is it the evaluation score that is depicted in Table 1? a. If you are using the same “score” to calculate the difficulty score and the discrimination indexes, does that mean a question is more difficult if a question is more discriminative? Confidence: 3 Soundness: 3 Presentation: 1 Contribution: 2 Limitations: While this proposed method is understood to work on general text questions fairly well, mathematical questions are the weakest part of this study. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q: The proposed methods - “Instruction gradient” and “response gradient” are not properly described in the manuscript. Authors should write the working procedure of these methods in detail in the main manuscript, as these are the centerpiece of the whole question generation process. A: Thanks for your question on our paper. We explain "Instruction gradient" and "response gradient" in footnote 1 on the second page of the manuscript. "Instruction gradient" and "response gradient" are the names we give to our methods. The process of generating generalized questions from seed data is analogous to forward propagation, where the LLM generates responses to the questions, and this process should be further pushed forward. Based on these responses (considered as information or knowledge), new questions are generated again, and this process is pushed backward, which can be compared to backpropagation. Therefore, we thought of using the term "gradient" for naming, and we named the process of generating questions based on seed data as "instruction gradient" and the process of generating generalized questions based on LLM responses as "response gradient". We further detail the description of the working procedure in the annotation of Figure 1 in the manuscript. Our working procedure is as follows: First, we collect a batch of seed data and divide it into mathematical and general text categories. Next, we apply the "instruction gradient" to both types of questions. For the "instruction gradient," the specific generalization strategies for the two types of questions are different due to the various question types. We provide the core generalization strategies in Table 5 of the appendix. We have the LLM(Specially referring to Hunyuan-standard in our paper) rewrite the seed data according to the generalization strategies, thus obtaining new questions. For general text questions, we can implement the "response gradient," i.e., first obtain the LLM's response to the question, and then ask questions based on the content of the response. We show the prompt for this process in Table 7 of the appendix. For mathematical questions, after generating questions based on the "instruction gradient," we focus more on the usability of the questions. Therefore, we design CoT to check, using multiple models (referring to Hunyuan and Hunyuan-pro in our paper) to judge the usability of the questions, and modify or discard the questions based on the inspection results. We show the specific CoT content in Table 9 of the appendix.* Q: “Generalizing questions from seed data based on the "instruction gradient" restricts the diversity and confines the content to specific topics”- is unclear. Consider explaining. A: Thank you for your suggestion on our paper. We provide further explanation for this part. The content generated by generalizing seed data through the "instruction gradient" is relatively close to the topic of the seed data. To make the generalized evaluation data more diverse, on the one hand, we can ensure the overall diversity of the evaluation data through the diversity of seed data. On the other hand, we enhance the diversity of questions through the "response gradient." For example, for the question "How can NLP technology be used to detect and prevent the spread of fake news?", using the instruction gradient for generalization, we can obtain a new question "List three specific methods to detect and prevent the spread of fake news using NLP technology and explain their principles," which still revolves around the original question for expansion or transformation. To address this, we consider discarding the original question and using the LLM-generated response as information or knowledge. At this point, we only generate questions based on a piece of text, and the questions may become more interesting based on the content of the response. In the above example, we may generate a new question "What NLP tasks are typically addressed by fact-checking and source analysis techniques?" Q: In section 2.3 - Assessing the Usability of General Text Questions: How is the assessment done? Is it done manually with human input? Or by an autonomic process/model? A: This part explains the usability check for general text data, which is implemented automatically by the LLM. We propose four criteria that we believe are important for general text questions: safety, neutrality, integrity, and feasibility. Here, safety refers to the absence of explicit, politically sensitive, or violent content in the question; neutrality refers to no bias or racial discrimination in the instructions; integrity refers to sufficient information provided to clarify the task; and feasibility refers to instructions within the AI system's capability range. We use the LLM (specifically, Hunyuan) to score general text questions based on these four criteria. Questions that do not receive a perfect score are considered unusable. For general text questions, the incidence of unusability occurs less frequently, so we discard questions deemed unusable without modifying them. In the experiment, we manually annotate the generated general text questions and found that the usability reached 94.0%. --- Rebuttal 2: Title: Supplement1 to the rebuttal Comment: Q: In section 2.3 - CoT Check for Mathematical Questions: “we use Hunyuan to assess the reasonableness of the question, which successfully identifies the unreasonableness of the problem and corrects it based on the assessment process.”- How can it be ensured that the model successfully identifies the unreasonableness? Provide a theoretical/experimental study. A: For mathematical questions, we indeed cannot guarantee that the generated questions can always be usable. However, through our proposed inspection mechanism, we can greatly eliminate the problems of conceptual errors, logical contradictions, violations of common sense, missing conditions, and unsolvable questions. "The model successfully identifies the unreasonableness" we methion here is an explanation of the case study in Figure 2. In Figure 2, we identified the unreasonable part of the question based on the designed CoT, indicating that the question is unusable and needs to be discarded or modified. For mathematical questions, we apply the "instruction gradient" to generalize the questions. To check the usability of the generated questions, we design a set of question-checking mechanisms. On the one hand, we design a set of CoT logic, starting from the concept, judging the logicality among different parts, evaluating the solvability of the question, and finally checking the question and steps, gradually guiding the LLM to think about the usability of the question. On the other hand, we conduct multi-turn iterative checks through two different LLMs to ensure the usability of the generated questions as much as possible. Specifically, the two LLMs independently judge the usability of the question through the CoT logic, but when one LLM judges the question as unusable or both LLMs judge the question as unusable, the question needs to be modified according to the judgment logic given by the LLM considered unusable (when both LLMs consider the question unusable, we designate one LLM's logic for modification). The modified question is then iteratively checked again. Only when both LLMs consider the question usable, the question is considered usable in the production process, and the iterative inspection of the question ends. If the maximum number of iterations is reached and there is still an LLM that judges the question as unusable, the question will be discarded. Q: In section 2.4 - Acquiring reference answers: lines 133-136, are the answers scored by human participants? A: In the selection of reference answers for general text questions, we use the Hunyuan(Hunyuan-standard) to score the responses. Using the LLM's response to the instructions as reference answers is relatively common in the data generation field. For example, in the Alpaca dataset[1], GPT-3.5 (text-davinci-003) is used to provide responses to questions as reference answers, and in the Instruction tuning with gpt4 work[2], GPT-4 is used to answer Chinese questions and serve as reference answers. Despite this, we hope to improve the quality of reference answers as much as possible. Inspired by [3], for general text questions, we also provide seven evaluation criteria: Safety (0-30 points), Correctness (0-10 points), Relevance (0-10 points), Comprehensiveness (0-10 points), Readability (0-20 points), Richness (0-10 points), and Humanization (0-10 points). We are more inclined to believe that responses with higher scores should have higher quality. We call multiple LLMs, including Hunyuan, GPT-4, GPT4-Turbo, Wenxin 4, and Qwen, to respond to the instructions, and then use Hunyuan to score these responses, selecting the highest-scoring response as the reference answer. We further involve humans in checking the usability of the answers. We select 150 generated general text questions and obtain reference answers in the aforementioned manner. We organize evaluators to score the selected reference answers according to the evaluation criteria in Table 1 of the paper. We remove 15 questions that none of the models answer correctly (the questions might be too difficult, all models answer incorrectly, and the answer selection is not meaningful). The results show that the usability rate of reference answers reaches 84.7%, which is higher than the highest correct rate of alternative reference answers, wenxin4 (78.8%). This indicates that the answer selection criteria can ensure the usability of the answers. [1] Taori, Rohan, et al. "Stanford alpaca: An instruction-following llama model." (2023): 6. [2] Peng, Baolin, et al. "Instruction tuning with gpt-4." arXiv preprint arXiv:2304.03277 (2023). [3] Liu, Yilun, et al. "Automatic instruction optimization for open-source llm instruction tuning." arXiv preprint arXiv:2311.13246 (2023). --- Rebuttal 3: Title: Supplement2 to the rebuttal Comment: Q: In section 2.4 - Acquiring reference answers: line 140, What is meant by a“collective voting mechanism”? Please explain clearly. A: Thank you for raising this issue, and we appreciate the opportunity to provide a more detailed explanation. For the answers to mathematical questions, we hope to select high-quality responses as reference answers as much as possible. However, it is difficult to design a scoring standard that conforms to human preference for mathematical question responses like general text questions. Related work [1] studies the theoretical basis of the collective voting mechanism and discusses the impact of different voting methods on social welfare. Inspired by this, we introduce a "collective voting mechanism" to select reference answers by comparing and voting among multiple responses. We provide multiple anonymous responses to the voting LLMs simultaneously, and let the voting LLMs choose the best response they think. We provide multiple voting LLMs, and each voting LLM casts its vote for a response, which means that the voting LLM thinks this response is the best. The response with the highest number of votes is used as the reference answer. If there is a tie, we randomly select a response as the reference answer and mark the question. Despite our efforts to enhance the usability of reference answers, there may still be instances where the selected reference answer is incorrect. To further improve the accuracy of reference answers for mathmatical questions, we hire mathematics experts to check and correct the reference answers of the questions. The results of the manual review are used as the final reference answers. [1] Sen, Amartya. Collective choice and social welfare. Harvard University Press, 2018. Q: In section 2.5 - lines 148-149, what are “label discrimination indexes”? a. In line 149, “the prompt includes four features”- How did you select these features? Provide some analysis. b. In lines 162-164, How did you select the threshold values? (e.g.,“Low”means less than or equal to 0.1,“High” means values greater than 0.25, etc.). c. In line 168, “discrimination level label ranging from 0-3”- Is this range acquired by observations? Or have you performed some analyses on the score expressions? A: Thank you for your question, the label discrimination indexes are the labels mapped from the discrimination indexes you mentioned in the 'b' question, we will improve the presentation here. a. The four features we select are included in each sample: question, its corresponding category, mean length of this category, and length ratio. These features are important and provide meaningful reference for understanding the discrimination of the questions. Question: The question is the most direct and key feature. The model needs to understand the question itself. Without the question, it is impossible to determine the type of information provided. Category: The discrimination of questions in different categories is usually different. For example, questions in the mathematics category may have different discrimination levels compared to those in the entertainment category. Category information helps us assign appropriate discrimination levels to questions. Mean length of the category: Considering the difficulty levels across different categories, the average length of questions within a category can indicate the complexity of the questions in that category. Generally, categories with longer answers may involve more complex questions, while categories with shorter answers may involve simpler questions. Therefore, by comparing the average lengths of different categories, we can gain a rough understanding of the question's difficulty, which serves as an important reference for its discrimination. Length ratio: From the perspective of varying difficulties across categories, the length ratio can help us understand the complexity of the question compared to the average question in its category. A higher length ratio may mean higher difficulty, and a lower length ratio may mean lower difficulty. By analyzing the length ratio, we can better understand the relative ranking of the difficulty and discrimination of the question in its category. b. The threshold here is estimated based on the distribution of 100,000-level evaluation data. c. 0-3 are the four levels we map discrimination indexes to, but this level division is not unique, it is just for the convenience of observing data with different discrimination. We can also divide it into two levels, etc. --- Rebuttal 4: Title: Supplement3 to the rebuttal Comment: Q: In equation 4, what does the “score”mean? Is it the evaluation score that is depicted in Table 1? a. If you are using the same“score”to calculate the difficulty score and the discrimination indexes, does that mean a question is more difficult if a question is more discriminative? A: Thank you for your question. Yes, the score here refers to the score in Table 1. a. It is the same score, but we believe that the difficulty score can only serve as a reference for discriminability. A high difficulty score for a question does not necessarily mean that it is more discriminative. For example, for a question with a max score of 3, if the evaluation scores are both 0 and 0, according to the formula, its difficulty score is 3, and the discrimination score is 0, meaning that the question is very difficult, and the LLMs cannot answer it correctly, so the question is not discriminative. However, if the evaluation scores are 0 and 3, we can calculate that its difficulty score is 1.5, and the discrimination score is 1, indicating that the question can effectively distinguish the level of LLMs. Q: While this proposed method is understood to work on general text questions fairly well, mathematical questions are the weakest part of this study. A: Thank you for your affirmation of our work on general text questions. For mathematical questions, on the one hand, providing prompts to large models to generate new questions can easily lead to unusable questions, which is difficult to handle; on the other hand, generating discriminative mathematical questions is also a significant challenge. We have also done a lot of work and contributions on mathematical questions, which are explained here. 1.We have proposed some data generalization strategies that can effectively improve the discrimination and difficulty of questions. In the "Instruction Gradient", for mathematical questions, we propose 8 generalization strategies to guide data generalization. Experiments have found that the questions generated by these strategies can effectively distinguish the capabilities of existing LLMs. 2.In practical evaluation scenarios, the difficulty of generalizing mathematical questions lies in the usability of the generated questions. This paper focuses on solving the problem of low usability of generalized mathematical questions and designs a usability checking mechanism for mathematical questions: On the one hand, we have designed a set of CoT for checking the usability of mathematical questions. This scheme can guide the LLM to check the usability of mathematical questions from the perspectives of concept, logical relationship, problem solvability, and condition completeness, greatly eliminating the problems of conceptual errors, logical contradictions, violations of common sense, missing conditions, and unsolvable questions. On the other hand, we can effectively modify or discard unusable generated data through multi-model multi-round iterative checks. Specifically, we use two different LLMs(refered to Hunyuan-standard and Hunyuan-pro in our paper) to judge the usability of the question based on the above CoT method. For unusable questions, we modify the question according to the judgment of CoT and iterate the check again until both LLMs judge the question as usable or reach the maximum number of iterations, and the question is also retained or discarded respectively. Through this mechanism, the generated mathematical questions have satisfactory usability. 3.The open discrimination estimation model and difficulty estimation model can quickly judge the quality of mathematical questions. In the process of training the discrimination estimation model and difficulty estimation model, we introduce a large number of mathematical questions with difficulty and discrimination annotations. The obtained models can effectively and quickly judge the discrimination and difficulty of mathematical questions. We make the models public to facilitate community research or use. 4.We open a batch of mathematical questions generated by LLMs(Specially refering to Hunyuan-standard and Hunyuan-pro in our paper) with accurate reference answers. We used the Hunyuan large model to generate mathematical questions with high discrimination, including 32 types of questions including calculus, function properties, and arithmetic operations. In addition, we hire experts in the field of mathematics to check and correct the reference answers of the mathematical questions to ensure the accuracy of the open mathematical question reference answers. --- Rebuttal Comment 4.1: Comment: First, I would like to thank the authors for thoroughly addressing all my questions. The paper tackles an important problem, but its structure does not flow naturally, making it quite difficult to follow. This explains why I had so many initial questions. While the authors did a good job responding to these questions, the paper should be revised to be clearer from the start. It's challenging to fix these issues as an afterthought. Therefore, I will maintain my initial score. --- Reply to Comment 4.1.1: Comment: We appreciate your reply and will revise the paper to make it more concise. We will pay special attention to the issues you mentioned and polish the writing carefully. We have given detailed explanations in our answers to the questions you mentioned, and hope that it will help readers better understand the paper. We sincerely hope that our score can be revised.	Summary: The paper introduces a novel framework for evaluating Large Language Models LLMs) based on Item Discrimination ID theory, which generates adaptive, high- quality prompts to effectively differentiate model performance. Key contributions include a dynamic evaluation set that evolves with LLM advancements, a self- correct mechanism for prompt precision, and models to estimate prompt discrimination and difficulty. The authors validate their framework by testing it on five state-of-the-art models and release a dataset of over 3,000 prompts to aid further research, demonstrating enhanced challenge and discrimination over previous methods. Strengths: The paper proposes a novel prompt generation method to produce more challenging evaluation data. The paper is well-structured and clearly written. The methodology and evaluation criteria are explained clearly, making the paper accessible to a broad audience. Weaknesses: The paper only used one LLM Hunyuan) to generalize data and did not verify whether the proposed method can generalize to other LLMs. It is debatable whether using test data generated by an LLM to evaluate the performance of LLMs has practical value. The paper lacks validation of the effectiveness of the machine-generated test set, such as comparing its metrics with those of other human-annotated datasets. The paper lacks an analysis of the diversity of the data used to produce the test set. Technical Quality: 3 Clarity: 3 Questions for Authors: The concerns are included in the weaknesses. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have identified some limitations; however, there are additional ones that I have raised in the weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q: The paper only used one LLM Hunyuan) to generalize data and did not verify whether the proposed method can generalize to other LLMs. A: Thank you for your question about our paper. Our proposed method is designed for existing LLMs and is not limited to a particular model. The work of using LLMs to automatically generate data often involves selecting only one LLM for data generation, such as the Wizardlm[1] work using gpt-3.5-turbo to generate instruction data, and the Self-instruct[2] work using gpt-3 to generate instruction data. We apply our proposed method to some other LLMs, such as GPT-4-turbo (gpt-4-turbo-2024-04-09) and Qwen (Qwen-max), using the same batch of a small amount of seed data, and manually scoring the models' responses to calculate discrimination indexes and map them to the four levels of discrimination indexes. The experimental results are shown in the table below. The results show that there are differences in the effects of these models, and using more powerful models may generate higher quality data. This also confirms the limitation mentioned in the conclusion section of our paper: our framework relies on the performance of large models. \| Model \| Amount \| Low \| Relatively Low \| Relatively High \| High \| \|:-----------:\|:------:\|:---:\|:-------------:\|:--------------:\|:----:\| \| Seed_data \| 50 \| 45 \| 0 \| 4 \| 1 \| \| Hunyuan \| 50 \| 29 \| 8 \| 8 \| 5 \| \| Qwen \| 50 \| 28 \| 13 \| 6 \| 3 \| \| Gpt4-turbo \| 50 \| 21 \| 5 \| 10 \| 14 \| Q: It is debatable whether using test data generated by an LLM to evaluate the performance of LLMs has practical value. A: The issue you metioned is a very in-depth one. On the one hand, if the data produced by the model is not controlled and filtered, it can have a significant negative impact on the model [1], so a good filtering mechanism is crucial for the effectiveness of the model's data production. On the other hand, if manually annotated data does not have a good filtering mechanism, it can also have a significant impact on model training, as shown in research work like LIMA[2]. However, a single evaluation often requires tens of thousands of evaluation data to fully measure the capabilities of large models. The cost of manually writing questions is too high and the speed is relatively slow. It is necessary to use LLM-produced test data in conjunction with manually written questions to evaluate the capabilities of large models. Therefore, this paper proposes an automated approach for constructing high-quality evaluation data, with contributions including the following two aspects: (1) Globally, it explores how to ensure the diversity of data, such as the classification of seed data and the diversity of data generation methods; (2) For each data, a very effective data usability check mechanism is designed. This process is reusable and will also be fully open-sourced. The method we proposed has also been validated in an actual production environment, helping to stably and comprehensively improve the performance of the production model, thereby validating its effectiveness in both experimental testing and production environment testing. [1] Shumailov, Ilia, et al. "AI models collapse when trained on recursively generated data." Nature 631.8022 (2024): 755-759. [2] Zhou, Chunting, et al. "Lima: Less is more for alignment." Advances in Neural Information Processing Systems 36 (2024). Q: The paper lacks validation of the effectiveness of the machine-generated test set, such as comparing its metrics with those of other human-annotated datasets. A: Thank you for your suggestion on our paper. We will explain your suggestion from the perspectives of usability, production efficiency, and cost to supplement the missing part. Usability: Human-annotated datasets are not necessarily all usable, and they often contain errors. They also need to be repeatedly checking and reviewing to ensure a high level of usability (e.g., above 95%). The usability of the questions in our generated data can reach 94% (the 94% usability is based on hunman-annotated results), and the usability of the evaluation data is satisfactory. Production efficiency: In this paper, it takes 2-5 calls to check a machine-generated question, with an average time of about 20 seconds per question. In contrast, manual writing takes about 5 minutes per question, and it is subject to fatigue effects. Cost: In this paper, generating a question and checking it with the machine involves input and output of about 9k tokens, costing approximately 0.03$. In contract, the market price for manually writing a usable question is about 2$, making the cost of human-annotated datasets relatively high. Q: The paper lacks an analysis of the diversity of the data used to produce the test set. "A:" Thanks for your suggestion on our paper. We appreciate your suggestion and provide a response to this issue, supplementing the explanation of the diversity of the data. We ensure the diversity of the seed data through a rich variety of categories. Our seed data consists of two parts: Chinese and English. The Chinese data refers to the work of TencentLLMEval, which includes 6 primary categories and 61 secondary categories. The English data uses the seed data from Self-instruct, which contains 175 different task types. In terms of methods, we design diversified generalization strategies including "Instruction Gradient" and "Response Gradient" to promote the generation of diversified questions. In the actual production process, we filter out similar questions. In this process, we cite the work of Self-instruct and remove data samples with ROUGE-l greater than 0.7 to filter out similar sample data.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Procedure-Aware Surgical Video-language Pretraining with Hierarchical Knowledge Augmentation	Accept (spotlight)	Summary: The paper addresses challenges in surgical video-language pretraining (VLP) due to the knowledge domain gap and scarcity of multi-modal data. It proposes a hierarchical knowledge augmentation approach and the Procedure-Encoded Surgical Knowledge-Augmented Video-Language Pretraining (PeskaVLP) framework. This approach enhances data efficacy and tackles spatial-temporal challenges by combining language supervision with visual self-supervision. Extensive experiments demonstrate significant improvements in zero-shot transferring performance and the generalist visual representation for surgical scene understanding. Strengths: The paper presents a unique approach to surgical video-language pretraining by employing hierarchical knowledge augmentation using LLMs, significantly improving textual data quality and diversity. The PeskaVLP framework innovatively integrates visual and language supervision, addressing the spatial-temporal challenges in surgical scene understanding. The methodology is meticulously validated through extensive zero-shot and linear-probing evaluations on datasets such as Cholec80 and AutoLaparo, demonstrating substantial performance improvements. The clarity of the presentation, with well-organized sections and effective visual aids, facilitates comprehension. The significant contribution lies in enhancing surgical scene understanding and cross-modal retrieval, making it highly valuable for the NeurIPS community. The paper's originality in using hierarchical pretraining and the detailed discussion on model architectures and initialization underscore its quality and significance in advancing surgical data science. Weaknesses: Firstly, the dataset size is relatively small, with 1,007 videos for phase-level pretraining and 920 for video-level pretraining, which may limit the generalizability of the findings (as mentioned in the supplementary material). I know the difficulty in collecting medical data, but we must be sure that the presented approach can be generalized to different domains and hospitals. Furthermore, I doubt the methodology's potential to process "noisy" videos. Expanding the dataset and including more diverse surgical procedures would improve robustness. Secondly, while the paper mentions ASR errors in transcriptions, it does not provide a detailed methodology for handling them. Providing specific techniques for improving transcription accuracy would strengthen the study. Additionally, the practical implementation of the PeskaVLP framework in real-world surgical contexts is not thoroughly discussed. Detailing strategies for integration into clinical workflows and addressing potential technological barriers would be beneficial. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. How do you plan to address the limited sample size and diversity in future studies to improve the generalizability of your findings? Consider expanding the dataset to include a more extensive and more diverse sample of surgical procedures to enhance robustness and applicability. 2. What specific methods did you use to handle ASR errors in transcriptions? How did these errors impact your analysis? 3. How do you manage the computational overhead associated with the hierarchical pretraining and dynamic time-warping processes? Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: The authors have acknowledged the limitations related to dataset size and ASR errors but could elaborate on strategies to mitigate these issues. Specifically, they should discuss plans for expanding the dataset, incorporating more diverse samples, and improving transcription accuracy. The positive societal impacts, such as enhancing surgical training and assistance, are well-discussed. However, the authors should address potential negative impacts, such as data privacy and ethical concerns. A detailed discussion on data security measures, user consent protocols, and ethical safeguards is needed. Flag For Ethics Review: ['Ethics review needed: Research involving human subjects', 'Ethics review needed: Data privacy, copyright, and consent', 'Ethics review needed: Data quality and representativeness', 'Ethics review needed: Safety and security'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [Q1. Plan to Expand Dataset] Scaling and diversifying the surgical vision-language pretraining dataset is challenging due to privacy concerns and the cost of expert annotations. Even though the SVL pretraining dataset covers diverse laparoscopic surgeries, it lacks surgeries in different organs, such as the brain and heart. We fully agree with the reviewer that expanding the dataset would be crucial for developing more generalizable models. To address this, we plan to expand the pretraining dataset using diverse media such as textbooks, instructional videos, and intra-operative video recordings from diverse sources. We also aim to diversify the pretraining dataset by considering laparoscopic, endoscopic, and microscopic surgeries on different organs. [Q2. Transcription Errors and Solution] ASR errors can be divided into two categories: misspelling errors and incorrect punctuation. In this work, we address these errors using specific methods to mitigate their negative impacts: - Misspelling Errors: We first follow [1] and use multiple text preprocessing to correct the ASR errors first, including USMLE-based spelling checking. Then, we use the LLM to generate large-scale clean descriptive texts that detail the steps for diverse surgeries. Noisy transcriptions are then assigned to semantically similar step texts based on their semantic similarity, as shown below: < Table S2>. Original noisy transcription and augmented text based on LLM. \| Raw Transcript \| Assigned Step Text \| \|----------\|----------\| \| but you think about two richard deliver from with amoxicillin truck and to work with your left hand with the right four unit \| Placement of five trocars: typically one for the camera, two for the surgeon's instruments, one for the liver retractor, and one for the assistant. \| \| so I be try to prepare my posterior dissection on the joel \| Dissection of the Esophageal Hiatus: Dissect the phrenoesophageal ligament to expose the esophageal hiatus. \| This ensures that the noisy texts are aligned with accurate descriptions. The use of the above LLM-based augmentation significantly improves the model's performance by providing cleaner and more accurate training data. As shown in the table below, SurgVLP shows improvement in accuracy when pretrained with augmented transcriptions. < Table S2>. Zero-shot phase recognition performance on Cholec80 dataset. \| \| SurgVLP \| SurgVLP + LLM Augmentation \| \|------------\|---------\|----------------------------\| \| Accuracy \| 34.7 \| 36.1 \| \| F1-Score \| 24.4 \| 26.8 \| - Incorrect Punctuation Errors: We follow SurgVLP and use Whisper, an ASR system based on a language decoder, which generates complete sentences with fewer punctuation errors. Whisper-transcribed sentences help define video correspondence boundaries more accurately. We anticipate that the modeling-based approach can also be applied in the future. For example, developing ASR systems specifically trained on surgical data can further reduce errors and improve transcription accuracy. Also, pretraining text encoders on a surgical corpus [2] can make the model more robust to ASR errors. _[1] Ikezogwo, Wisdom, et al. "Quilt-1m: One million image-text pairs for histopathology." Advances in neural information processing systems 36 (2024)._ _[2] Bombieri, Marco, et al. "Surgicberta: a pre-trained language model for procedural surgical language." International Journal of Data Science and Analytics 18.1 (2024): 69-81._ [Q3. Computational Overhead] We manage the computational overhead associated with hierarchical pretraining and dynamic time-warping (DTW) processes by using efficient resource allocation and high-speed computational infrastructure. Below is a summary of the time and computational cost when applying hierarchical pretraining and DTW processes: < Table S3>. Training configurations and time for SurgVLP. \| Configuration \| Training Time \| GPU \| \|---------------------------\|---------------\|----------------\| \| SurgVLP \| ~40 Hours \| 1xA100-80G \| \| + Hierarchical Pretraining\| ~90 Hours \| 4xA100-80G \| \| + Hierarchical + DTW \| ~120 Hours \| 4xA100-80G \| In this work, the speed bottleneck is at the online video loading and preprocessing. In future, we aim to apply the asynchronous processing pipelines to further parallelize the workload and reduce bottlenecks. [Q4. Acknowledge the Plans to Mitigate Data Size and ASR Errors ] Thank you for the insightful suggestion. In the revised manuscript, we have added a discussion section to elaborate the strategies about expanding dataset and improving transcription accuracy, as we discussed in prior responses. [Q5. Potential Negative Impacts] Thank you for the insightful suggestion. As discussed in the paper, our dataset is compiled from open educational platforms accessible to all learners, minimizing any potential societal impact. However, in the rapidly advancing field of surgical data science, there are still risks of privacy concerns and data security associated with the collection and use of surgical video data. Ensuring robust data protection measures and maintaining patient confidentiality is crucial to mitigate these risks. Moreover, the development of automated intra-operative surgical procedural understanding systems should aim to provide intuitive cognitive aids for surgeons, enhancing their decision-making during complex procedures. These systems should support surgeons by delivering relevant information at critical moments, rather than overwhelming them with extraneous details. We have added a discussion on the potential negative societal impact in the revised manuscript. --- Rebuttal Comment 1.1: Comment: Dear Authors, Thank you for your detailed rebuttal and for addressing the concerns raised during the review process. Your efforts to expand the dataset, address ASR errors, manage computational overhead, and consider potential negative impacts are appreciated. Here are some follow-up questions and concerns: Transcription Errors and Solution. How do you plan to quantitatively evaluate the improvements in transcription accuracy using the proposed LLM-based augmentation and Whisper ASR system? Can you provide detailed comparative results before and after implementing these methods? Computational Overhead. Could you elaborate on the specific asynchronous processing pipelines you plan to implement to reduce computational bottlenecks? What are the expected improvements in terms of training time and resource utilization? Potential Negative Impacts. While you mentioned minimizing societal impacts by using open educational platforms, how will you handle patient consent and data privacy for intra-operative video recordings? Your commitment to improving the manuscript is commendable, and I look forward to seeing these enhancements reflected in your revised submission. --- Rebuttal 2: Title: Rebuttal Follow Up Comment: [Q1. Transcription Errors and Solution. ] In the following Table 1, we quantitatively demonstrate the impact of LLM-based augmentation and transcription processing strategies on transcription accuracy for the zero-shot phase recognition on Cholec80 dataset. We train the SurgVLP model using Clip-level text Augmentation which uses ASR transcripts enhanced by LLM-based augmentation and before mentioned transcription processing strategies. The results are shown in the first two rows on the table. The SurgVLP model when trained with the Clip-level text Augmentation, including transcription processing strategy and LLM-based augmentation, brings – 2.2% improvement in the F1 score. <Table 1: Quantitative assessment for the improvements brought by LLM-based augmentation and transcription processing strategies. We report the zero-shot performance of Cholec80 testing set. > \| Setup \| Accuracy \| F1-Score \| \|-----------------------------------------------\|----------\|----------\| \| SurgVLP \| 34.7 \| 24.4 \| \| SurgVLP + Strategy + Clip-level Augmentation \| 36.1 \| 26.8 \| \| HecVL \| 41.7 \| 26.3 \| \| HecVL + Strategy \| 43.0 \| 29.2 \| \| HecVL + Strategy + Phase-level Augmentation \| 44.0 \| 31.8 \| \| HecVL + Strategy + Video-level Augmentation \| 43.7 \| 30.6 \| \| PeskaVLP \| 45.1 \| 34.2 \| Similarly, we also train the HecVL model using processed transcripts without LLM-based augmentation, resulting in improved performance compared to the baseline HecVL. Notably, the improvements from transcription processing can be further enhanced when combined with LLM-based augmentation. [Q2. Computational Overhead] Thank you for your follow-up comment. In this work, we train directly on the surgical videos without converting them into frames. As a result, a key bottleneck lies in the online video decoding process, where videos must be decoded and processed in real-time to provide training batches. Currently, we are using ffmpeg for the online video decoding process followed by Pytorch to process the decoded frames. In future, we plan to implement a more efficient multi-threaded video decoding pipeline where multiple videos can be decoded in parallel on GPUs using advanced libraries such as NVIDIA Data Loading Library (DALI) [1] or Decord [2]. We expect to minimize the idle GPU time. [Potential Negative Impacts] Thank you for a follow-up comment. To address patient consent and data privacy in the collection of intra-operative surgical videos, we will ensure that all recordings are fully anonymized by removing any identifiable patient information and out-of-body frames using open-source tools [3]. Specifically, since our focus is on laparoscopic videos, anonymization will involve removing any captions or metadata that could identify the patient, as well as ensuring that any external out-of-body views or patient-identifiable features captured in the video are thoroughly anonymized. _[1] https://developer.nvidia.com/dali_ _[2] https://github.com/dmlc/decord_ _[3] Lavanchy, Joël L., et al. "Preserving privacy in surgical video analysis using a deep learning classifier to identify out-of-body scenes in endoscopic videos." Scientific reports 13.1 (2023): 9235._	Summary: The paper presents a novel approach for enhancing surgical video analysis by incorporating procedural awareness. The authors propose a system that integrates knowledge of surgical procedures to improve the identification, segmentation, and annotation of surgical activities in video footage. This approach aims to address challenges such as the variability of surgical techniques and the complexity of visual data in operating rooms. The contributions of the paper include the development of a procedural model that can be aligned with video data, the creation of annotated datasets for training and evaluation, and the demonstration of improved performance over traditional video analysis methods. Strengths: 1.The integration of procedural knowledge into surgical video analysis is a highly original concept. This approach not only enhances the accuracy of video analysis but also opens new avenues for improving surgical training and documentation. 2.Introduces a novel hierarchical knowledge augmentation technique using large language models to refine surgical concepts. Employs a Dynamic Time Warping-based loss function for effective cross-modal procedural alignment. Demonstrates significant improvements in zero-shot transfer performance across multiple surgical datasets. Provides a robust general visual representation beneficial for various surgical scene understanding tasks. Weaknesses: 3.The potential applications of this research in surgical training, intraoperative assistance, and postoperative review are significant. The approach addresses a critical need in medical video analysis, making it highly relevant and impactful. Weaknesses: Dataset Limitations: The annotated datasets used for training and evaluation are crucial for the model's success. Expanding the diversity and volume of these datasets would enhance the generalizability of the findings. Technical Quality: 4 Clarity: 4 Questions for Authors: Generalizability: How does the system perform across different types of surgeries (like ophthalmic surgery)? Have you tested its effectiveness in various surgical domains beyond the initial scope? Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The paper does not adequately address potential limitations and negative societal impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [Q1. Dataset Limitations] Thank you for the insightful suggestion. In the rebuttal letter pdf, we have added a table to summarize the top 42 types of surgical videos and their amounts in the pretraining dataset. As shown in Table 1 of the rebuttal letter PDF, the SVL dataset predominantly consists of laparoscopic surgeries and covers a diverse range of surgical types, including stomach, duodenum, hernia, colon, gallbladder, and tumor surgeries. This diversity ensures that the SVL dataset provides broad and generalizable language supervision for surgical multi-modal representation learning. In this paper, the downstream datasets are Cholec80 (Laparoscopic cholecystectomy), AutoLaparo (Laparoscopic hysterectomy), and MultiBypass (Laparoscopic gastric bypass). Our SVL pretraining dataset contains sufficient videos that cover the visual concepts required for these downstream tasks. We agree with the reviewer that expanding the pretraining dataset to cover more types of surgeries can improve diversity and improve generalizability. We leave this exploration to future research endeavors. [Q2. Generalizability] Thank you for the insightful question. Evaluating the generalizability of our system across different types of surgeries is crucial for understanding its broader applicability. Our system has been tested extensively on laparoscopic surgeries, including Cholec80 (Laparoscopic cholecystectomy), AutoLaparo (Laparoscopic hysterectomy), and MultiBypass (Laparoscopic gastric bypass). These evaluations demonstrate the model's ability to effectively recognize phases and understand the workflow in these contexts. Since we do not have ophthalmic surgical videos in our pretraining dataset, we do not test the PeskaVLP’s performance in this domain. Future work would involve incorporating and testing diverse surgical datasets, including ophthalmic surgery, to ensure the model's effectiveness across various surgical domains. This will help us assess the system's capability to generalize and perform effectively in a broader range of surgical contexts.	Summary: This paper proposes a Procedure-Encoded Surgical Knowledge-Augmented Video-Language Pretraining (PeskaVLP) method that enriches language supervision with LLM-refined surgical concepts. It further constructs hard negative samples by reversing the text orders at the phase and video levels and employs a Dynamic Time Warping (DTW) based loss to align multimodal procedures. Extensive experiments on multiple surgical procedures and comprehensive evaluations demonstrate the effectiveness of this framework. Strengths: - The paper is overall well-written, with the background and motivation well-stated. - Using LLM to augment surgical video text descriptions is a good idea to enhance the quality of surgical text narration. It establishes a good baseline and guideline for future works that aim to apply LLM in surgical narratives. - A more comprehensive parent-child level cross-modal correspondence was designed using DTW than existing works. - Demonstration of the proposed method can close the representation gap for different modality, and analysed both successful and complicated examples. Weaknesses: - By reading the enriched dataset by LLM in Appendix H, I am concerning that the variation and diversity of narration will be removed by the augmentation. Will that cause any problems? - In my opinion, using LLM to refine the text description of surgical videos is the most important contribution of this paper. It would be interesting to see if other components are also effective enough without the knowledge augmentation. Technical Quality: 2 Clarity: 3 Questions for Authors: - Beyond the current ablation study on PeskaVLP components, would applying the hierarchical knowledge-augmented text data in HecVL improve its performance and if this could yield results competitive with PeskaVLP. This would provide powerful support to verify the extent to which the other components in PeskaVLP contribute to performance, apart from the augmented texts. - Although LLM can enhance surgical text quality, is there a concern that the text may become overly standardized? Given that surgeons' narratives in the operating room tend to be more oral, concise, and sometimes include jargon, will there be a performance degradation in real-world, real-time applications where LLM augmentation is impractical? - In Appendix E, Figure 4, it would also be interesting if the authors could visualize the embeddings of HecVL, since it performs better than SurgVLP. - In Table 3, on Cholec80, Moco pre-trained on Cholec80 (V) has better performance but wasn't in bold, do I misinterpret something? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors adequately addressed the limitations. Since the proposed method is tailored for surgical data and applications, it is strongly suggested that the authors include a discussion on the potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [Q1. Augmentation Removes Variation] Thank you for pointing out one of the key insights of this work, i.e., using LLM to build a large, versatile, and accurate surgical knowledge base to enrich and correct narrations of different types of videos during the pretraining. Since we enrich the narration based on the built knowledge base, the question becomes if the LLM-generated surgical knowledge base is diverse enough. In this work, we aim to bring the diversity and versatility of the built knowledge base by curating 917 lecture titles, covering diverse surgical procedures such as colorectal, transanal and proctological, cholecystectomy, hernia, and sigmoidectomy surgery. We also manually design input-output examples to instruct the LLM to generate diverse steps. Additionally, during the pretraining, we randomly select either the pseudo step or the original narration text to maintain textual semantics. Our approach might face risks when LLM generates an over-standardized surgical step knowledge base. However, our experiments show that the downstream zero-shot performance clearly improves when the LLM-generated knowledge base is applied. This implies that the advantages of correcting noisy narration texts outweigh the potential variation and diversity risks. < Table S1>. Zero-shot phase recognition performance on Cholec80 dataset. \| \| SurgVLP \| SurgVLP + LLM Augmentation \| \|------------\|---------\|----------------------------\| \| Accuracy \| 34.7 \| 36.1 \| \| F1-Score \| 24.4 \| 26.8 \| [Q2. Other Components] In the following table, we show that the combination of visual self-supervision and language supervision at the clip-level vision-language pretraining are effective even without the knowledge augmentation. < Table S2>. Zero-shot phase recognition on Cholec80 and Autolaparo datasets. \| Model\| Dataset \| Accuracy / F1-Score \| \|---------\|------\|--------\| \| HecVL\| Cholec80 \| 41.7 / 26.3\| \| HecVL+LecNCE{clip} \| Cholec80 \| 45.5 / 31.0\| \| HecVL\| Autolaparo\| 23.3 / 18.9\| \| HecVL+LecNCE{clip} \| Autolaparo\| 25.3 / 20.0\| The results indicate incorporating knowledge augmentation and visual self-supervision can individually benefit the pretraining performance. [Q3. Different Levels of Knowledge Augmentation] Knowledge augmentation from different hierarchies improves the final performance in different ways. We have summarized results in the table below: < Table S3>. Performance on Cholec80 dataset with different augmentations. \| Model \| Accuracy / F1-Score \| \|-----------\|---------------------\| \| SurgVLP \| 34.7 / 24.4 \| \| SurgVLP+Clip-level Augmentation \| 36.1 / 26.8\| \| PeskaVLP \| 45.1 / 34.2 \| \| HecVL \| 41.7 / 26.3 \| \| HecVL+Phase-level Augmentation \| 44.0 / 31.8\| \| HecVL+Video-level Augmentation \| 43.7 / 30.6\| \| PeskaVLP \| 45.1 / 34.2 \| We observe that applying knowledge-augmented text data in SurgVLP and HecVL achieves a clear improvement, though they still underperform the PeskaVLP. These results demonstrate the effectiveness of the other components of PeskaVLP, i.e., combination of visual self-supervision and language supervision and procedure-aware pretraining objective. [Q4. Overly Standardized Texts] We thank the reviewer for this interesting question. We apply the LLM augmentation strategy for pretraining by considering the goal of representation learning, downstream surgical workflow analysis applications, and the nature of pretraining surgical lecture videos. The goal of PeskaVLP is to learn the correspondence between vision and language modality with the aim of improving the performance on the surgical downstream tasks. Therefore, the jargons and interjection can introduce the noisy alignment, which is filtered out in our pretraining dataset. Our pretraining videos are primarily lecture videos where narrations are scripted and more formal. The GPT-augmented strategy is well-suited for this type of data, as it enhances the clarity and completeness of the text. We acknowledge the potential limitations in real-world applications. For future work, incorporating more diverse and real-time surgical audio [1] into the pretraining process could help improve the performance. _[1] Jia, Shuyue, et al. "MedPodGPT: A multilingual audio-augmented large language model for medical research and education." medRxiv (2024): 2024-07._ [Q5. Modality Gap HecVL] In the rebuttal letter PDF, Figure 1, shows HecVL's embeddings, showing a smaller modality gap for video-abstract pairs compared to SurgVLP. This shows the benefit of hierarchical vision-language pretraining for aligning long-form videos with high-level summaries. PeskaVLP consistently outperforms prior baselines, demonstrating its effectiveness in closing the modality gap and enhancing pretraining for vision-and-language tasks. [Q6. Table 3] Thank you for pointing this out. We have corrected the table to highlight Moco (third row) as providing the best results on the Cholec80 dataset and Moco (second row) on the StrasBypass70 dataset. In the revised manuscript, we have discussed that Moco pretrained on Cholec80 outperforms the others because it is specifically trained on cholecystectomy procedures, thereby losing versatility. This specialization allows Moco to achieve superior performance on Cholec80 but limits its generalizability to other datasets, as summarized below: < Table S4>. Correct version of table 3 in manuscript. \| Pretraining \| Dataset \| Cholec80 \| Autolaparo \| StrasBypass70 \| BernBypass70 \| \|-------------\|--------------\|----------\|------------\|---------------\|--------------\| \| Moco \| Cholec80 \| 73.4 / 62.8 \| 51.3 / 37.4 \| 67.8 / 55.4 \| 66.0 / 33.1 \| \| Moco \| SVL \| 68.2 / 55.8 \| 59.5 / 48.4 \| 71.6 / 58.1 \| 69.6 / 36.5 \| \| PeskaVLP \| SVL \| 69.9 / 59.8 \| 63.1 / 49.7 \| 71.4 / 59.5 \| 71.5 / 37.4 \| --- Rebuttal Comment 1.1: Comment: I appreciate the authors' response to my review and their efforts to address my concerns. After carefully reviewing the feedback from the other reviewers, I am inclined to maintain my original score, thanks.	Summary: The paper presents a new framework called PeskaVLP for surgical video-language pretraining. A hierarchical knowledge augmentation approach is used for enriching text information. The pretraining is implemented with the proposed language supervision and visual self-supervision. A new training objective is proposed for surgical procedural understanding. Extensive experiments are conducted to demonstrate the effectiveness on the surgical phase recognition task and cross-modal retrieval task on multiple downstream dataset. Strengths: 1. This paper addresses the problem of VLP in the surgical scene. A hierarchical knowledge augmentation is proposed to tackle the problem of lack of textual information in the surgical field. 2. The paper is generally well-written and easy to follow. Weaknesses: 1. The explanation of method details is not clear enough, and there is a lack of discussion on some experimental results 2. The proposed method is based on certain assumptions but lacks a comprehensive consideration of applicability. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. What types of surgeries are included in the SVL dataset used in the paper? Is it suitable for the pretraining task? Could it affect the results on the downstream dataset? 2. In Section 3.2, where hierarchical knowledge is augmented by GPT, the authors need to discuss the ability of LLMs to generate accurate textual information to describe the surgical steps in the domain-specific surgical context, especially considering the fine-grained image-text alignment in the clip-level (only 4 frames). 3. In Section 3.2, the authors calculate textual similarity between the pseudo step generated by the LLM and the narration. How is this similarity calculated? Is there an ablation study on the effectiveness of the three behavior in knowledge augmentation? 4. In Section 3.3.1, the authors implement visual self-supervision based on augmentation. Which specific augmentations were used? Do the augmentations affect the corresponding text's semantic information? For example, using flipping could impact descriptions related to left/right information in surgical operation. 5. In Section 3.3.2, procedural information based on surgical phases is used. However, in surgical datasets, such as the cholec80 and AutoLaparo mentioned in the paper, the surgical process does not always follow a linear order defined by Phase 1-N and may include repeated phases. The authors should discuss the applicability of the method design in such situations. 6. In Table 3, for the experimental results on cholec80, Moco (third row) provides the best results, but this is not highlighted in bold in the table. This needs to be corrected and the corresponding discussion should be provided. The same issue appears with the results using Moco (second row) on the StrasBypass70 dataset. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Authors discussed it briefly in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [Q1 SVL Dataset] [Q1.1. Types of surgeries in SVL dataset] In the rebuttal letter PDF, we have added a table summarizing the top 42 types of surgical videos in the pretraining dataset. As shown in Table 1 in the rebuttal letter PDF, the SVL dataset predominantly contains laparoscopic surgeries, focusing on the stomach, duodenum, hernia, colon, gallbladder, and tumors. Also, the diverse content within each surgical type ensures that the SVL dataset provides generalizable language supervision for representation learning. [Q1.2. Suitable for Pretraining] This paper uses Cholec80, Autolaparo, and MultiBypass datasets for downstream tasks, with Table 1 showing that the SVL pretraining datasets sufficiently cover the required visual concepts. The SVL videos focus on laparoscopic surgeries, covering common concepts like instruments and bleeding, while their descriptive texts offer valuable language supervision for tasks like instrument recognition and adverse event detection. [Q1.3. Affect Downstream Dataset] We hypothesize that the composition of different types of surgical videos in the pretraining dataset will affect performance. While research [1] suggests that altering the pretraining dataset to the downstream dataset improves zero-shot adaptation, we aim to build a generalizable pretrained model and thus do not filter videos based on the downstream datasets. _[1] Datacomp: In search of the next generation of multimodal datasets_ [Q2. LLM-Generated Text for Clip-level Alignment] In this paper, we control the LLM-generated textual quality based on the in-context learning ability of LLMs, which enables faithful predictions with a limited context. By thoughtfully designing contextual input-output examples, LLM can generate diverse and rich texts that enrich the clip-level narration texts, as shown below: < Table S1>. Original noisy transcription and assigned step text based on LLM. \| Raw Transcript \| Assigned Step Text \| \|-\|-\| \| but you think about two richard deliver from with amoxicillin truck and to work with your left hand with the right four unit \| Placement of five trocars: typically one for the camera, two for the surgeon's instruments, one for the liver retractor, and one for the assistant. \| \| so I be try to prepare my posterior dissection on the joel \| Dissection of the Esophageal Hiatus: Dissect the phrenoesophageal ligament to expose the esophageal hiatus.\| The generated surgical steps from LLM are not always perfectly aligned with the clip-level frames. To address this, we randomly pick either the original or the augmented texts during video-language pretraining. Noisy image-text alignment is a challenge for all CLIP-based methods. Scaling up the dataset could help mitigate this issue by providing a broader context and improving alignment, and we leave this exploration to future research. [Q3. Textual Similarity] Textual similarity is calculated using cosine similarity between feature vectors from pseudo steps and narrations. We use BioClinicalBert to extract 768-dimensional embeddings for each sentence, normalize them, and compute the dot product to get similarity. This approach is consistent with methods used in previous literature. [Q4. Effectiveness of Three Levels Augmentation] Knowledge augmentation from different levels improves performance in different ways, as summarized in the table below: < Table S2>. Performance on Cholec80. \| Model\| Accuracy / F1-Score \| \|-\|-\| \| SurgVLP\| 34.7 / 24.4\| \| SurgVLP+Clip-level Augmentation\| 36.1 / 26.8 \| \| HecVL \| 41.7 / 26.3 \| \| HecVL+Phase-level Augmentation\| 44.0 / 31.8 \| \| HecVL+Video-level Augmentation\| 43.7 / 30.6\| We found that phase-level knowledge augmentation significantly improves surgical vision-language pretraining due to its concise and less noisy keystep texts, whereas LLM-generated surgical knowledge base is less effective due to noisy image-text alignment in pseudo-step assignment. [Q5. Augmentations Affect Corresponding Text's Semantics] For visual self-supervision, we use spatial augmentations like random cropping and horizontal flipping, which might introduce vision-language misalignment. We find that around 9% of sentences contain spatial words. Since one video maps to multiple sentences, the actual impact from augmentation is less than 9%, which is a less significant source of noise than misspellings and incomplete transcripts. Also, our experiments show that these augmentations consistently boost performance, with their benefits outweighing potential misalignments. [Q6. Procedural Information] We clarify that the surgical procedural information learned from pretraining is not just a linear sequence of phases but includes complex dependencies among surgical key steps. PeskaVLP learns from various surgical lecture videos, capturing possible paths to complete a surgery. For instance, if one pretraining video includes _[.., gallbladder dissection, cleaning and coagulation, ..]_ and another _[.., gallbladder dissection, gallbladder packing, ..]_, our procedure-aware pretraining can identify multiple possible steps following gallbladder dissection. This procedural information is useful for temporal modeling in surgical phase recognition when applying methods like temporal convolutional networks or task graphs to perform online phase prediction. [Q7. Table 3] We have updated the manuscript and added discussion. The revised manuscript explains that Moco's strong performance on Cholec80 is due to its specialization in cholecystectomy, which limits its generalizability to other surgeries. < Table S4>. Correct version of table 3 in manuscript. \| Pretraining \| Dataset\| Cholec80 \| Autolaparo \| StrasBypass70 \| BernBypass70 \| \|-\|-\|-\|-\|-\|-\| \| Moco \| Cholec80\| 73.4 / 62.8 \| 51.3 / 37.4 \| 67.8 / 55.4\| 66.0 / 33.1\| \| Moco\| SVL \| 68.2 / 55.8 \| 59.5 / 48.4 \| 71.6 / 58.1\| 69.6 / 36.5\| \| PeskaVLP\| SVL\| 69.9 / 59.8 \| 63.1 / 49.7 \| 71.4 / 59.5\| 71.5 / 37.4\| --- Rebuttal 2: Comment: Thank you for the detailed responses. The rebuttal has addressed most of my concerns, and I would like to raise my score accordingly.	Rebuttal 1: Rebuttal: We thank all the reviewers for the insightful comments to improve our work. We are encouraged that the reviewer finds our work an interesting contribution to the community. We have carefully considered each comment from the reviewers and tried to provide detailed answers, clarifying all the issues raised. We hope our responses effectively resolve the concerns raised. We are grateful for the reviewers' time and insightful feedback. If further clarifications or additional experiments are required, please let us know. We are pleased to note that the reviewers recognized: - This paper addresses a novel surgical vision-language pretraining task that can potentially open new venues for surgical data science [usyA, EZYT] - The proposed hierarchical knowledge augmentation improves surgical textual data quality and diversity [all reviewers] - The manuscript is well-organized, clearly written, and presented in a reader-friendly way [HH5E, smvC, usyA] - This work novelly integrates visual-language supervision and dynamic time warping to learn the cross-modal correspondence [smvC] We have included a rebuttal letter PDF, including SVL dataset details [All reviewers] and modality gap visualization of HecVL [smvC]. Pdf: /pdf/24139415827c369ef050189ed71b2b1555861006.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
A Separation in Heavy-Tailed Sampling: Gaussian vs. Stable Oracles for Proximal Samplers	Accept (poster)	Summary: The paper investigates the complexity of sampling from heavy-tailed distributions and presents a distinction between obtaining high-accuracy and low-accuracy guarantees. It analyzes two types of proximal samplers: those based on Gaussian oracles and those based on stable oracles. The main findings are that Gaussian oracle-based samplers can only achieve low-accuracy guarantees when sampling from heavy-tailed distributions, while stable oracle-based samplers can achieve high-accuracy guarantees. Additionally, the paper establishes lower bounds for samplers using the stable oracle, indicating that the presented upper bounds are optimal and cannot be fundamentally improved. Strengths: 1. The problem is well-motivated and interesting. 2. Designed the algorithms and derived the upper bounds and lower bounds for different settings. 3. The authors also provided insightful discussion. 4. The authors provided solid theoretical proof for the results. Weaknesses: There is no experiment to verify the theoretical findings. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Can you give an example in the real-world to motivate your problem? 2. Is it possible to run some experiments to verify your results? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: There is no experiment. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their valuable advice and comments and greatly appreciate the positive evaluation. >Weakness: There is no experiment to verify the theoretical findings. >Question 2: Is it possible to run some experiments to verify your results? Following the reviewer's suggestion, we have added numerical experiments that compare the Gaussian proximal sampler and the stable proximal sampler with $\alpha=1$; see the uploaded pdf file as part of the rebuttal. In the first three experiments (corresponding to rows), we choose the target distribution to be the one-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 100 steps with step-size 0.1, to get 100 chains. We adopt different initializations, $x_0=20,5,-5$, and visualize the convergence via the average trajectories (with standard deviations), histogram of the last-step samples and the Wasserstein-2 distance decay, respectively. In the last experiment, we choose a 2-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 20 steps with step-size 0.1, to get 30 chains. We adopt the initialization, $x_0=[5,1]$, and use the same visualizations for the first-coordinate (marginals). The stable proximal sampler outperforms the Gaussian proximal sampler in all cases. In the revised version of our paper, we will include a section containing extensive detailed numerical studies demonstrating the performance of the algorithms. >Question 1: Can you give an example in the real-world to motivate your problem? There are several real-world applications of heavy-tailed sampling which arises in various domains such as Bayesian statistics [GJPS08, GLM18], machine learning [CDV09, BZ17, NSR19, SZTG20, DKTZ20], robust statistics [KN04, JR07, Kam18, YŁR22], multiple comparison procedures [GBH04, GB09], and study of geophysical systems [SP15, QM16, PBEM23]. In particular the papers by [SP15, QM16, PBEM23] discuss real-world data analysis problems which crucially hinge on heavy-tailed sampling. We will be happy to elaborate this problem in the revision, subjected to space constraints. Furthermore, there has been recent works on heavy-tailed diffusion models; see e.g., [YPKL24], arXiv:2407.18609 and [PSVM24]. We anticipate that our results in this work have implications for heavy-tailed diffusion models as well. Finally, please refer to the recent workshops at Neurips 2023 (titled "Heavy Tails in ML: Structure, Stability, Dynamics") and Issac Newton institute (titled "Heavy tails in machine learning") for further emerging applications of heavy-tailed sampling in real-world data science. [YPKL24]- Yoon, E. B., Park, K., Kim, S., & Lim, S. (2023). Score-based generative models with Lévy processes. Advances in Neural Information Processing Systems, 36, 40694-40707. [PSVM24] - Paquet, E., Soleymani, F., Viktor, H. L., & Michalowski, W. (2024). Annealed fractional Lévy–Itō diffusion models for protein generation. Computational and Structural Biotechnology Journal, 23, 1641-1653.	Summary: This paper studies the problem of heavy-tailed sampling. First, the paper shows that while the gaussian proximal samplers are efficient for light-tailed targets, they are not accurate for heavy-tailed ones; the paper develops a lower bounds for the Gaussian proximal samplers, which reveals a fundamental challenge in heavy-tailed settings. Then, the paper proceeds to develop a novel samplers based on restricted alpha-stable oracle; the insight is to replace the standard heat equation in gaussian oracle with a fractional heat flow. The paper proves that under suitable conditions the proposed sampler is efficient for heavy-tailed targets. Additionally, the paper proposes a practical implementation for a particular case of alpha=1. Strengths: - Novel theoretical analysis for the gaussian oracle sampler, which provides a new insight to developing sampling algorithms - A novel methodology for heavy-tailed sampling Weaknesses: - The paper is purely theoretical and lacks experimental evaluation; it would be nice to at least have a toy illustration for the implementable algorithm 2+3 in the alpha=1 case. - As the authors discussed in Sec5, the current paper does not present implementable algorithms for general alpha values in (0,2). Technical Quality: 3 Clarity: 3 Questions for Authors: - I wonder if the efficiency rejection sampling efficiency in Alg.3 has been taken into account of the sampler's theoretical complexity and practical complexity? - Maybe I am missing this -- what is the impact of alpha? Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Most of the limitations have been touched upon in sec 5. Otherwise see the weakness comments. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their valuable advice and comments and greatly appreciate the positive evaluation. >The paper is purely theoretical and lacks experimental evaluation; it would be nice to at least have a toy illustration for the implementable algorithm 2+3 in the $\alpha=1$ case. Following the reviewer's suggestion, we have added numerical experiments that compare the Gaussian proximal sampler and the stable proximal sampler with $\alpha=1$; see the uploaded pdf file as part of the rebuttal. In the first three experiments (corresponding to rows), we choose the target distribution to be the one-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 100 steps with step-size 0.1, to get 100 chains. We adopt different initializations, $x_0=20,5,-5$, and visualize the convergence via the average trajectories (with standard deviations), histogram of the last-step samples and the Wasserstein-2 distance decay, respectively. In the last experiment, we choose a 2-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 20 steps with step-size 0.1, to get 30 chains. We adopt the initialization, $x_0=[5,1]$, and use the same visualizations for the first-coordinate (marginals). The stable proximal sampler outperforms the Gaussian proximal sampler in all cases. In the revised version of our paper, we will include a section containing extensive detailed numerical studies demonstrating the performance of the algorithms. >As the authors discussed in Sec5, the current paper does not present implementable algorithms for general alpha values in (0,2). There is currently a difficulty to implement the R$\alpha$SO exactly for general $\alpha\in (0,2)$ due to the fact that there is no explicit representation for the $\alpha$-stable density for general $\alpha$. It is an interesting to investigate an exact/inexact implementation of the R$\alpha$SO for general $\alpha\in (0,2)$. >I wonder if the efficiency rejection sampling efficiency in Alg.3 has been taken into account of the sampler's theoretical complexity and practical complexity? Similar to the discussion of the Gaussian proximal sampler in [1], the iteration complexity results in our paper assumes the R$\alpha$SO, and the efficiency of Algorithm 3 is not included. In section 3, we provide Algorithm 3 as a practical implementation of R$\alpha$SO and its efficiency is discussed in Remark 3. >Maybe I am missing this -- what is the impact of $\alpha$? $\alpha$ is a parameter that appears in the Fractional Poincare inequality (FPI) and the stable process. As shown in Theorem 3, if the target distribution satisfies $\alpha$-FPI, the $\alpha$-stable proximal sampler convergens exponentially fast. The $\alpha$ in $\alpha$-FPI characterizes the tail-heaviness of the target distribution. If the target is extreme heavy-tail, we need to apply the stable proximal sampler with a small value of $\alpha\in (0,2)$ so that the target satisfies the $\alpha$-FPI. If we choose a large $\alpha\in (0,2)$ and the target doesn't satisfy $\alpha$-FPI, then the $\alpha$-stable proximal sampler only converges polynomially fast. Details are included in Section B.4 in the appendix. [1] Chen, Yongxin, et al. "Improved analysis for a proximal algorithm for sampling." Conference on Learning Theory. PMLR, 2022 --- Rebuttal Comment 1.1: Comment: I would like to thank the reviewers for their detailed reply. My concerns are mostly addressed. Also the empirical results seem promising. I will increase my rating to 7.	Summary: The paper focus on studying the complexity of heavy-tailed sampling and present a separation result in terms of obtaining high-accuracy versus low-accuracy guarantees. Their results are presented for proximal samplers that are based on Gaussian versus stable oracles. Authors show that proximal samplers based on the Gaussian oracle have a fundamental barrier in that they necessarily achieve only low-accuracy guarantees when sampling from a class of heavy-tailed targets. In contrast, proximal samplers based on the stable oracle exhibit high-accuracy guarantees, thereby overcoming the aforementioned limitation. They also prove lower bounds for samplers under the stable oracle and show that our upper bounds cannot be fundamentally improved. Strengths: Although I am not an expert in this field, I find this work quite interesting. The authors provide new material and support their statements with proofs. Weaknesses: The paper is not tested in any way on a numerical experiment. I am convinced that a paper presented at this type of conference should be both motivated by a real-world application and tested numerically, e.g., on a near-real-world formulation of the problem. After a rebuttal process, the authors agreed with this weakness and promised to add the experiments to the final version of the paper. Technical Quality: 2 Clarity: 2 Questions for Authors: N/A Confidence: 1 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their valuable advice and comments and greatly appreciate the positive evaluation. >The paper is not tested in any way on a numerical experiment. I am convinced that a paper presented at this type of conference should be both motivated by a real-world application and tested numerically, e.g., on a near-real-world formulation of the problem. Following the reviewer's suggestion, we have added numerical experiments that compare the Gaussian proximal sampler and the stable proximal sampler with $\alpha=1$; see the uploaded pdf file as part of the rebuttal. In the first three experiments (corresponding to rows), we choose the target distribution to be the one-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 100 steps with step-size 0.1, to get 100 chains. We adopt different initializations, $x_0=20,5,-5$, and visualize the convergence via the average trajectories (with standard deviations), histogram of the last-step samples and the Wasserstein-2 distance decay, respectively. In the last experiment, we choose a 2-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 20 steps with step-size 0.1, to get 30 chains. We adopt the initialization, $x_0=[5,1]$, and use the same visualizations for the first-coordinate (marginals). The stable proximal sampler outperforms the Gaussian proximal sampler in all cases. In the revised version of our paper, we will include a section containing extensive detailed numerical studies demonstrating the performance of the algorithms. --- Rebuttal Comment 1.1: Comment: Dear Authors, Thank you for the reply. The responses are satisfactory. I am raising my score by +1.	Summary: The authors provide a lower bound for sampling from heavy tailed distributions under the Gaussian oracle of order $O(\textup{poly}(1/\varepsilon))$. They then propose an alternative proximal sampling algorithm using the $\alpha$-stable oracle that achieves a convergence rate of $O(\log(1/\varepsilon))$ for heavy-tailed distributions satisfying a fractional Poincare inequality. They then provide a practical implementation of the stable proximal sampler, and lower bounds on its convergence rate. Strengths: - This work presents a very nice combination of results showing a separation in the performance of stable and Gaussian proximal samplers. The combination of lower and upper bounds separating the two methods makes the work a particularly interesting contribution. - The addition of a practical implementation of the stable proximal sampler is nice to have, demonstrating that it is viable in practice. - The work is generally clearly presented and the authors are clear about their contributions. - Overall, I consider this to be a very sound piece of theoretical work. Weaknesses: I have no major concerns about this paper. The presentation is somewhat dense in places, though this is mostly just a consequence of it being a very technical paper and not a flaw as such. If the authors want to make the claim that practicioners should use the stable proximal sampler in applied settings, then they may want to provide empirical evidence of its performance compared to the Gaussian proximal sampler. However, I understand that this is not the main purpose of this theoretical paper. Technical Quality: 3 Clarity: 3 Questions for Authors: I have no clarifications to request. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors provide an adequate discussion of the limitations of their methods in the final section, and I foresee no additional negative impacts of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their valuable advice and comments and greatly appreciate the positive evaluation. >I have no major concerns about this paper. The presentation is somewhat dense in places, though this is mostly just a consequence of it being a very technical paper and not a flaw as such. If the authors want to make the claim that practicioners should use the stable proximal sampler in applied settings, then they may want to provide empirical evidence of its performance compared to the Gaussian proximal sampler. However, I understand that this is not the main purpose of this theoretical paper. Following the reviewer's suggestion, we have added numerical experiments that compare the Gaussian proximal sampler and the stable proximal sampler with $\alpha=1$; see the uploaded pdf file as part of the rebuttal. In the first three experiments (corresponding to rows), we choose the target distribution to be the one-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 100 steps with step-size 0.1, to get 100 chains. We adopt different initializations, $x_0=20,5,-5$, and visualize the convergence via the average trajectories (with standard deviations), histogram of the last-step samples and the Wasserstein-2 distance decay, respectively. In the last experiment, we choose a 2-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 20 steps with step-size 0.1, to get 30 chains. We adopt the initialization, $x_0=[5,1]$, and use the same visualizations for the first-coordinate (marginals). The stable proximal sampler outperforms the Gaussian proximal sampler in all cases. In the revised version of our paper, we will include a section containing extensive detailed numerical studies demonstrating the performance of the algorithms. --- Rebuttal Comment 1.1: Comment: Thank you for your comment. I appreciate the engagement with the comments and additional experimental results provided.	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their valuable advice and comments and greatly appreciate the positive evaluation. Results from the added experiments are included in the pdf file. Pdf: /pdf/bc4091796415ba3d7391c96e453543e5ea7487e9.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper studies the complexity of sampling heavy-tailed distributions. It provides lower bounds on the complexity of Gaussian-based samplers for a class of heavy-tailed targets. Then, the paper constructs proximal samplers based on stable oracles, which improve the sampling complexity. Strengths: * This paper is well-written. The background of sampling and the research problems regarding sampling complexity are clearly introduced. The contributions of the lower bound on Gaussian-based samplers for heavy-tailed targets and the improved complexity using stable oracles are clearly presented. * The paper is technically sound. The definitions and assumptions are discussed clearly, and the theoretical results are supported by proof sketches. Weaknesses: The contribution of the paper could be improved with empirical experiments to evaluate the sampling algorithms and their complexity. Technical Quality: 3 Clarity: 3 Questions for Authors: * Is there any intuition that a Gaussian-based sampler has lower accuracy for heavy-tailed targets than for non-heavy-tailed targets? * How would a Gaussian-based sampler compare with a stable oracle for not heavy-tailed targets? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their valuable advice and comments and greatly appreciate the positive evaluation. >The contribution of the paper could be improved with empirical experiments to evaluate the sampling algorithms and their complexity. Following the reviewer's suggestion, we have added numerical experiments that compare the Gaussian proximal sampler and the stable proximal sampler with $\alpha=1$; see the uploaded pdf file as part of the rebuttal. In the first three experiments (corresponding to rows), we choose the target distribution to be the one-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 100 steps with step-size 0.1, to get 100 chains. We adopt different initializations, $x_0=20,5,-5$, and visualize the convergence via the average trajectories (with standard deviations), histogram of the last-step samples and the Wasserstein-2 distance decay, respectively. In the last experiment, we choose a 2-dimensional student-t distribution with 4 degrees of freedom and zero mean, and run the algorithms in parallel for 20 steps with step-size 0.1, to get 30 chains. We adopt the initialization, $x_0=[5,1]$, and use the same visualizations for the first-coordinate (marginals). The stable proximal sampler outperforms the Gaussian proximal sampler in all cases. In the revised version of our paper, we will include a section containing extensive detailed numerical studies demonstrating the performance of the algorithms. >Is there any intuition that a Gaussian-based sampler has lower accuracy for heavy-tailed targets than for non-heavy-tailed targets? For the Gaussian-based sampler, the repeated sampling steps $y_k\sim \pi^{Y\|X}(\cdot\|x_k)=\mathcal{N}(x_k,\eta I_d)$ are crucial to ensure rapid convergence. These are essentially heat flow simulations, that take advantage of the geometric properties of the logconcave (non-heavy-tailed) target distributions, providing an exponential decay in KL-divergence. When the target distribution is heavy-tailed, the standard heat flow mixes slowly. Intuitively, if our initial condition is a light-tailed random variable, it would take many steps for the Gaussian proximal sampler to match the moments of the heavy-tailed target, because in every iteration we only add Gaussian noise with a small variance. >How would a Gaussian-based sampler compare with a stable oracle for not heavy-tailed targets? Gaussian proximal sampler performs well when the target satisfies a Poincare or alog-Sobolev inequality [1]. When the RGO step is implemented exactly, the algorithm converges exponentially fast. Since these light-tailed distributions also satisfy the Fractional Poincare inequality (FPI), our Theorem 3 suggests the stable proximal sampler also converges exponentially fast. The difference lies in the rates of exponential convergence. Intuitively, stable proximal sampler may have a smaller convergence rate compared to Gaussian proximal sampler. For example, if the target is a Gaussian which has a concentrated region that most samples would lie in, since the stable proximal sampler adds heavy-tail randomnesses in each iteration, it may move a sample out of the concentrated region more easily. Therefore, it takes longer for the stable proximal sampler to move most of the particles inside the concentrated region. Analytically, the convergence rates depend on the Poincare constant and the FPI constant in the Gaussian and stable case, respectively [1] Chen, Yongxin, et al. "Improved analysis for a proximal algorithm for sampling." Conference on Learning Theory. PMLR, 2022.	null	null	null	null	null	null
How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning	Reject	Summary: This paper introduces Accordion Networks (AccNets), a novel neural network structure composed of multiple shallow networks. The authors propose a generalization bound for AccNets that leverages the F1-norms and Lipschitz constants of the subnetworks, demonstrating that these networks can break the curse of dimensionality by efficiently learning compositions of Sobolev functions. The paper also provides theoretical insights and empirical validation, showcasing the superior performance of AccNets in learning complex compositional tasks compared to shallow networks and kernel methods. Strengths: The introduction of Accordion Networks (AccNets) as a novel neural network structure is a creative and original contribution. The paper provides a thorough theoretical analysis supported by empirical evidence, ensuring the soundness of its claims. The ability of AccNets to break the curse of dimensionality by learning compositional functions efficiently addresses a fundamental challenge in high-dimensional learning tasks. Weaknesses: 1. The practical implementation of the proposed regularization methods might be challenging, particularly the first one requiring infinite width. 2. The paper mentions the difficulty in optimizing Lipschitz constants, which could be a limitation in practical applications. 3. Additional experiments on more diverse real-world datasets could further demonstrate the robustness and generalizability of AccNets. 4. Although the author has discussed the differences between DNN and AccNet, there is still not enough information for me to be sure in which settings to use AccNet and in which settings to use DNN. More clear differences and applicable conditions, especially the shortcomings of each need to be pointed out. Technical Quality: 3 Clarity: 4 Questions for Authors: Can the authors provide more details on the computational complexity of training Accordion Networks compared to traditional DNNs? How sensitive are the generalization bounds to the choice of hyperparameters, particularly the Lipschitz constants and F1-norms? Are there any specific types of tasks or datasets where Accordion Networks might not perform as well as traditional methods? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: See weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	null	Summary: The authors present a generalization bound for deep neural networks that describes how depth enables models to learn functions that are compositions of Sobolev functions. To do this, they both prove a generalization bound for compositions of accordion networks (densely connected networks with a low-rank weight structure) and for compositions of Sobolev functions. They then present a sample efficiency result for different kinds of regularization on accordion networks. Strengths: I really liked this paper and would like to see it accepted to NeurIPS. It addresses an important question: how does depth change generalization bounds for deep neural networks? To my knowledge, not many papers so far have addressed this question and I found the findings presented here very interesting and well embedded within prior methodology. I also found the paper very well written. I found it easy to follow along despite the highly technical nature of the results (note that I did not check the proofs in particular detail). I especially appreciated the remarks explaining different potential extensions and limitations. Finally, the theory appears to be able to explain certain empirical phenomena (in networks trained under realistic paradigms) at least qualitatively (though note that I had a few questions I will mention under weaknesses and questions). This indicates to me that it is a promising way for thinking about generalization in deep neural networks. Weaknesses: 1. I would like to see a more thorough comparison with shallow networks and generalization bounds, as this comparison is a central argument for the usefulness of the presented theory. While it is clear how the findings for the shallow network are a special case of the findings on the deep networks (as presented in Thm. 1), it remains a bit unclear to me how the theory can explain improved generalization in deep compared to shallow networks. The authors certainly present different several pieces of evidence on this: both Fig. 1 and Fig. 3 demonstrate that shallow networks exhibit worse scaling. I also appreciated the theoretical explanation of a particular contrast in l. 256-261. However, I think it would be really useful to provide a general theoretical explanation for this difference and test it empirically: would it be possible to extend the theoretical comparison in l. 256-261 to the general experimental setup studied in the figures --- and if so, would this theoretical comparison predict the conditions under which deep networks have the strongest advantages over shallow networks (or perhaps the conditions under which they don't perform that much better)? Not only would this serve as a useful validation of the theory, I think it would also provide a more extensive intuition for the authors' findings. 2. I appreciated the fact that the authors compare their findings with related work wherever this becomes relevant. However, I think a (potentially brief) section comparing the results here to other theoretical investigations of depth in deep networks (perhaps using different approaches) would be useful. 3. The linked codebase does not contain the notebooks indicated in the README as far as I can tell and therefore currently can't be used to directly reproduce the findings. 4. I believe the figures would still benefit from error bars or some other indication of the overall statistical error in the findings. I agree that the main contribution of this paper is theoretical, but since the experiments test the empirical validity of the theory, I believe it is nevertheless important to get a sense for the overall deviation in these findings (e.g. across model seeds). If the authors are concerned about a lack of clarity, they could leave the bars out of the main figures but add supplementary figures with error bars. Moreover, some of the lines in Fig. 1 do contain error bars and it would be good to clarify what these error bars represent. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Do you think my suggestion in point 1 of the weaknesses make sense or do you have a reason why you see it as unnecessary? 2. As far as I understand, the reason for the asymmetry between $\nu_g$ and $\nu_h$ in Fig. 2 is the different dimensionality, correct? It would be good to mention these dimensionalities, as I was only able to find them in the appendix. 3. Could you clarify why in Fig. 2, you're using the scalings from Prop 3 rather than from Thm. 5? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The authors adequately discuss the limitations of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	null	Summary: The authors introduce accordion networks (AccNets), which are compositions of multiple shallow networks. By leveraging prior workthat computes norm-based generalization bounds for shallow two-layer networks, the authors bound the complexity of a deep AccNet (as measured by its F1 norm) but the sum of the complexities of the individual shallow networks. They empirically observe that the rates predicted on real-world data are roughly representative of the trained networks, and are indeed much better than those for kernels trained on the same tasks. They put forth a nontrivial scaling law for the excess risk: $N^{-\mathrm{min}(1/2, \nu_g/d_{in}, \nu_h/d_{mid})}$ for an Acc Net compared to $\mathcal L \sim N^{-\mathrm{min}(1/2, \nu_g/d_{in}, \nu_h/d_{in})}$ for a kernel in terms of the dimensionalities $d$ and Sobolev constants $\nu$ of the respective spaces and functions. From this, the authors obtain predictions of several phases, that they put forth experiments to verify. Strengths: The paper tackles a very important open question in the theory of deep learning, for which not much progress has been made. By creatively leveraging results for shallow network in composition, the authors arrive at a nontrivial bound for deep nets. The empirics are a very compelling and welcome part of the paper. The phase diagrams illustrate the nontrivial predictivity of the theory, especially at the level of the rates. This may have important implications for scaling laws. Modulo minor revisions in discussion and exposition, the whole paper is quite readable for a relatively broad audience. Weaknesses: I am not sure how compelling the phase plots in Figure 2 are. The bounds in general are extremely loose, however the comparison of the rates in Figure 2c and Figure 3 is very promising. In general, however, it is the experience of the reviewer that measuring a rate is an extremely finicky business. It is therefore important to add a section in the appendix explicitly stating how the rates were obtained and measured. I also strongly encourage the authors to make the code for all figures public. Because they are used very early on throughout the paper, it is the opinion of the reviewer that the notions of F1 distance and Sobolev norm should be defined earlier on in the paper. Without this, it seems like the audience will be constrained to the set of learning theorists familiar with these terms. However, if these terms are defined early on, the paper becomes remarkably accessible to a much broader audience. Technical Quality: 3 Clarity: 2 Questions for Authors: The plot labels in Figures 2 and 3 are very difficult to read. A small comment: I have not seen the term "modulo space" used before. Often the term is "quotient space" The sentence defining the $F_1$ ball (above theorem 1) is confusing, circular, and difficult to read. Please rewrite it. The excess rate formula $\mathcal L \sim N^{-\mathrm{min}(1/2, \nu_g/d_{in}, \nu_h/d_{mid})}$ is a very important result and I recommend that it be formatted for display, not inline. How are you measuring "dimension" in 4.1.1? A high-dimensional gaussian with spectral decay of its covariance going as $k^{-\alpha}$ for capacity exponent $\alpha$ is nominally "full dimensional" since it is not strictly speaking constrained to a sub-manifold, and yet basic results in kernel theory and high-dimensional linear regression can show that the generalization error achieves a much better rate at larger values of $\alpha$. Specifically, a model with capacity exponent $\alpha$ and source exponent $r$ achieves a rate of $N^{-2\alpha min(r, 1)}$. See, e.g. https://arxiv.org/abs/2105.15004 . Such power law anisotropy is in abundant in natural data. In particular shallow two layer networks in the lazy limit can achieve this scaling for such 'easy tasks' with quick spectral decay. On the other hand, the bounds that you state cannot decay faster than $N^{-1/2}$. * In this sense, it seems that the bounds (shallow or deep) presented are certainly not tight for some datasets. Am I incorrect in concluding this? Do you have an intuition for what causes the breakdown in correctly predicting the error rates in this case? * Given that they breakdown in that setting, what about the datasets that you study makes it so that the scaling law predictions seem to hold? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Given the theoretical nature of this work, it is unlikely to have major social implications. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	null	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
OxonFair: A Flexible Toolkit for Algorithmic Fairness	Accept (poster)	Summary: The paper introduces "AnonFair," a toolkit designed to enforce algorithmic fairness across various domains, including NLP, computer vision, and traditional tabular data. It is compatible with popular machine learning frameworks like sklearn, AutoGluon, and PyTorch. Unlike well-established fairness tools like FairLearn and AIF360, AnonFair extends to different types of data, including NLP and vision. Other tools offer many methods but limited control over them, while AnonFair uses a single, highly customizable method that allows for per-group thresholding. It specifically addresses the issue of overfitting by utilizing validation data, making it more reliable when traditional methods might fail. Empirical evidence presented shows that AnonFair performs well, often matching or surpassing other methods in fairness benchmarks without being specifically optimized for complex or high-dimensional scenarios. AnonFair seems to provide a robust and adaptable solution for implementing fairness in machine learning, in ways that other tools do not currently offer. Strengths: - The paper does well in positioning AnonFair against competing tools by demonstrating its performance on standard fairness metrics and its versatility across a variety of use cases. - AnonFair supports NLP and computer vision classification tasks, allowing broader applicability. - The toolkit uses validation data to combat overfitting, ensuring that fairness measures remain robust across both training and unseen data. - The toolkit not only competes well in terms of accuracy and fairness metrics but also offers significant advantages in computational efficiency. Weaknesses: - Some sections are overly detailed, such as the introduction, while others are missing necessary depth: - Section 3 could use a clearer structure, possibly with a diagram, to help readers understand how to interact with the toolkit. - The section on toolkit expressiveness needs more detailed examples and explanations of how the supported fairness measures are implemented. - Results discussion is kept very brief and could benefit from specific numerical examples, like percentage improvements compared to other methods.m actual numbers, such as how much % improvement in comparison to method XY and such. - The paper assumes readers are familiar with fairness terminology and metrics without adequate explanations or definitions for some acronyms (e.g., DEO in Table 3 and 4). - Subsection 4.3 lists supported fairness measures but fails to provide examples or brief explanations, making it less informative for those not familiar with these terms. - Lack of consistency in terminology usage; for example, "EOp" in Figure 1 (top right) vs. "EO" in Section 5.2, “AnonFair” missing before "Frontier" in Figure 1 (left), and inconsistent references like "See Figure" vs. "See fig.." - A stronger call to action for community engagement, such as through open-source collaboration or empirical validation studies, could significantly enhance the broader impact and encourage more widespread adoption and refinement of AnonFair. - The paper would benefit from a summary of explicit cases and recommendations advising users on the best scenarios for using the tool. - Figure 2 is not referred to in the paper, or did I miss this part. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The paper mentions that hard assignment is more efficient than soft assignment, while appendix A adds some operational details, it remains unclear how these methods specifically compare in terms of quantitative metrics. Could the authors provide specific metrics or comparisons that demonstrate the efficiency and performance benefits of hard assignment? 2. The discussed limitations reads a bit out of context given provided evidence in the paper. What makes the mentioned solutions suboptimal, and how significant are these shortcomings? Also it was not clear to me, after finishing reading, when it is adequate to use this tool and what could be use cases when it fails. Including this into the conclusion could make the reader grasping the full picture. 3. Is Figure 6 part of the Appendix or misplaced? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Some of the limitations are acknowledged, but could be expanded with more actionable insights. A call to action for community engagement, such as through open-source collaboration would also encourage broader impact and adoption of AnonFair against its competitors. It would be beneficial if the authors suggested potential improvements or future research directions for the suboptimal fairness metrics and data scarcity issues mentioned. The broader impact section identifies ethical concerns well. However, detailing the intended applications and scenarios where AnonFair might be most effective, or where it could fail, would provide readers and users with clearer guidance on its practical use and limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our manuscript and for providing detailed, helpful, and constructive feedback. We hope to address outstanding weaknesses and concerns below. Improving presentation: The idea of a figure/flow chart is a good one, but there is insufficient space in the paper. We will add this directly to the toolkit documentation instead, where it is more likely to be seen by someone interested in using the toolkit. Details on implementing fairness measures: Please see Appendix B for details of how measures are implemented. In brief, you can just write down a function of the confusion matrix. This is computed per group, and standard measures such as minimum per group, or average difference or ratio of measures between groups are automatically computed. These correspond to different notions of fairness. We will add an additional example of this for Equal Opportunity, as this is one of the most widely used group fairness measures. Results discussion and numerical examples: We will discuss specific examples from the tables and figures where our toolkit shows clear improvements over other approaches. Improving the consistency of fairness terminology and abbreviations: Thanks for flagging the inconsistency. We will correct this. Call to action for community engagement: Thank you. We will add a call for community engagement in the abstract. We strongly agree with the reviewer’s focus on making the toolkit more accessible through documentation. Unfortunately, space constraints mean that much of the information is going to lie outside the paper. We have spent much of the past months since submission extending the documentation and adding additional examples. This will continue to be a focus going forward. We invite community-driven contributions principally as pull requests reflecting the needs and considerations that researchers come across in practice. In this manner, we can support more advanced features without overwhelming the codebase and maintenance requirements. Our toolkit is now a pip installable library, and the community can raise issues, concerns, and requests on a GitHub repository. In this way, practitioners can collaboratively build on each other's work. Explicit cases and recommendations advising users on best practices. What are the adequate uses of this tool and what could be the use cases when it fails? These are very important questions. The short answer is that this is too important to be compressed into the conclusion of the paper. This toolkit arose from a promise to a healthcare provider that we would go through existing toolkits and return a list of best practices. Essentially, our conclusion was, it didn’t matter what we suggested as no existing toolkit worked for NLP or computer vision, and those that existed for tabular data couldn’t enforce measures like bias amplification, conditional measures, or levelling up. The purpose of this paper is to lay out a technical toolkit which is expressive enough to support both our recommended best practices, and hopefully the best practices of other researchers who disagree with us. Follow-up white papers will set out how we think it should be used. Use cases, best practices and documentation are our core focus going forward. Figure 2 not referred to in the main paper: We will add references to this figure in Section 4 (specifically lines 156 and 189 and footnote 5). # Questions Question 1: For a performance comparison, we reran the unconditional fairness metrics from table 7 with inferred attributes and the slow vs fast pathway. Using default parameters, the fast pathway has an average of 1.2% better fairness across the metrics and wins in 7/10 of the cases, while maintaining similar accuracy (win rate 5/10, an average decrease in 0.28% of accuracy). For an efficiency comparison, we reran the comparison with fairlearn, on adult, varying the number of groups. \| # Groups \|5\|4\|3\|2\| \|-\|-\|-\|-\|-\| \| FairLearn \|33.2s\|33.3s\|25.4s\|19.4s\| \|Fast\| 0.7+55.7s\|0.7+0.93s\|0.7+0.078s\| 0.7+0.05s\| \|Slow\| - \| 0.7+454s\| 0.7+22s\|0.7+1.5s\| We will add these numbers to the paper. Question 2: Apologies. The fact that thresholding is suboptimal for Equalized Odds is widely known in the literature [1]. The suboptimality comes from the fact that Equalized Odds is a combination of two fairness measures (difference in true positive rates and difference in false positive rates) and that ideal thresholds for one measure need not minimize the other. Instead, a combination of thresholding and randomization is often used. Our approach is only known to be sub-optimal when the true attributes are used, and not with data that requires the use of inferred attributes. The rest of the time, Anonfair has no guarantees regarding optimality. This is the same as most widely used fairness methods including FairLearn. Unfortunately, the only answer to if our approach should be used is the same as for the rest of machine learning – the experiments might look good, but where alternatives exist, it is best practice to compare them on your use case. The only case where our method can be said to fail is where insufficient data is provided to generalize from validation to unseen test data. Otherwise, the method may be suboptimal (better fairness accuracy trade-offs might exist), but it will work, and similar fairness/accuracy trade-offs will be seen on unseen test data. Question 3: Figure 6 is part of the appendix and is an expanded Figure 1 from the main body. While Figure 1 compares fairness toolkits with a focus on optimizing for balanced accuracy, Figure 6 broadens this comparison by including additional metrics. [1] Hardt, Moritz, Eric Price, and Nati Srebro. "Equality of opportunity in supervised learning." Advances in Neural Information Processing Systems 29 (2016). --- Rebuttal 2: Title: Any follow-up clarifications? Comment: We hope that we have addressed all issues raised to your satisfaction in our rebuttal. We would be happy to provide additional clarifications if required as the discussion period will be over soon. Thank you for your time. --- Rebuttal Comment 2.1: Comment: I appreciate the authors' detailed response to the concerns raised and the additional experiments. Given the constraints of a conference paper, I agree that the toolkit's documentation may be better suited for the extensive details. With the reviewers' feedback incorporated, the paper should be better positioned to convince practitioners to try out the toolkit. Good luck. I will maintain the current score.	Summary: This paper describes a new toolkit for algorithmic fairness, enabling the optimization of any fairness measure that is a function of the confusion matrix. Experiments on vision and NLP demonstrated the effectiveness of the proposed toolkit. Strengths: An easy-to-use toolkit for enforcing algorithmic fairness. Weaknesses: Presentation could be made more self-contained, e.g. a table listing the supported fairness metrics, as functions of the confusion matrix. This would help readers not familiar with the field. It seems that only binary classification is supported. How can such metrics be extended to other tasks? Some minimal code snippets for the interface could be shown as examples. Technical Quality: 3 Clarity: 2 Questions for Authors: - L5: "True positives, false positives, ..." => "the confusion matrix" - L6: "extendable" => "extensible" Confidence: 1 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors adequately discussed the limitations of their toolkit. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and helpful suggestions that will be integrated to improve the paper. Improvements in presentation: We will add the definition of equal opportunity (the most common fairness definition, corresponding to difference in recall between groups) to the work. Honestly, we have substantial problems with space constraints in this work. The two review papers we cite [1,2] only list and discuss metrics and come in at a combined 15+ double-columned pages. We will try to include the most common or most relevant metrics in our work, but we are not a review paper, and providing references for most of the metrics is probably the most sensible option. Beyond binary classification: Tasks beyond binary classification are on our roadmap for future work. Fair ranking is of particular interest, as failures in ranking have clear well-defined harms and may be compatible with our accelerated approach. For multiclass classification, from a policy perspective, there are a couple of metrics that are generalizations of demographic parity (every labelling rate should be the same for all groups), and equalized odds (confusion matrix should be the same for all groups), but they seem to have been chosen for mathematical convenience, rather than because they correspond to clear harms. Minimal Code Snippets: We agree that minimal code snippets would improve the paper. We have code snippets in the appendix (Appendix C.2 and figure 8). and we will add further examples. In addition to code snippets, we also have Jupyter Notebook tutorials and example studies for practitioners to get up and running with our toolkit. We will highlight the vast availability of resources and examples that our toolkit provides, and our toolkit empowers the community to build on these. Typos: Thanks for pointing out typos, these will be promptly corrected. [1] Verma, Sahil, and Julia Rubin. "Fairness definitions explained." Proceedings of the international workshop on software fairness. 2018. [2] Hardt, Michaela, et al. "Amazon sagemaker clarify: Machine learning bias detection and explainability in the cloud." Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2021. --- Rebuttal 2: Title: Any follow-up clarifications? Comment: We hope that we have addressed all issues raised to your satisfaction in our rebuttal. We would be happy to provide additional clarifications if required as the discussion period will be over soon. Thank you for your time. --- Rebuttal 3: Comment: Thanks for the detailed response. I have modified the score based on the appendix. Thanks.	Summary: The paper introduces a new toolkit designed to enhance algorithmic fairness with greater expressiveness. Unlike existing toolkits, this one offers more customization options to optimize user-defined objectives and fairness constraints. Although the proposed toolkit currently includes only one method, it supports both computer vision and natural language processing (NLP) tasks. The authors compare the efficiency of this method, finding that the toolkit is relatively more efficient than Fairlearn. Comprehensive experiments were conducted on various datasets, and the results were compared with those from other popular toolkits. Strengths: - The paper introduces a versatile toolkit that supports both NLP and computer vision tasks, unlike existing toolkits which lack this capability. - The proposed toolkit employs efficient optimization techniques that accelerate the evaluation process. Weaknesses: - The formulation presented in Subsection 4.2 of the paper is limited to a single-layer model, which restricts its applicability across different machine learning models. To enhance the flexibility of the method, I recommend adopting a more generic notation, particularly if we aim to incorporate pretrained language models. - The abstract is quite unclear, especially the part that mentions "9/9 and 10/10 of the group metrics of two popular review papers." I suggest rephrasing the abstract for better clarity and comprehension. Technical Quality: 3 Clarity: 3 Questions for Authors: - In Figure 3, the proposed toolkit appears to encounter scaling issues when reaching 5 groups. Could you provide more details on why this occurs and elaborate on the underlying reasons for this limitation? - The paper presents results on multilingual datasets. Do you have any specific findings for each language, particularly regarding the effectiveness of the toolkit for individual languages? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed feedback. We are happy to see that the reviewer appreciates the versatility of our toolkit beyond existing solutions and the efficiency of optimization in the toolkit. Clarification on notation: The equations in section 4.2 contain a function, $B(x)$. This represents an arbitrary non-linear backbone shared between the two heads. While the heads are linear (as is common in multitask learning), arbitrary non-linear functions can be learned. This approach is used as written for the NLP and Computer Vision tasks which use complex non-linear backbones. We will emphasize this in the text. Improvements to Abstract: Thanks. We will rephrase the abstract to improve clarity. # Questions > In Figure 3, the proposed toolkit appears to encounter scaling issues when reaching 5 groups. Could you provide more details on why this occurs and elaborate on the underlying reasons for this limitation? The slowdown as we increase the number of groups, is expected and supported by the analysis in Section 4 (lines 157-162) which shows that the run-time is exponential in the number of groups. This was a conscious decision to maximize expressibility at the expense of compute for a large number of groups. Having sufficient data to enforce fairness for a large number of protected groups is unfortunately rare in algorithmic fairness and in practice, these scaling issues are rarely encountered. >The paper presents results on multilingual datasets. Do you have any specific findings for each language, particularly regarding the effectiveness of the toolkit for individual languages? Our toolkit is language and data modality agnostic. While we see differences in performance, this seems to be predominantly driven by the quality of the data. In general, the most common languages have much more data available (particularly English with three times more data than the next most available language), and this allowed curators to create datasets with a balanced subset of groups, and positive and negative labels. For example, in the case of well-resourced languages such as English (.499 of datapoints were labelled Female) and (.495 of English datapoints were labelled non-White). In contrast for Polish protected groups were very unbalanced (.324 Female) and (.105) non-White. This means both that base classifiers for well-resourced languages tend to be fairer (at least when evaluated on the provided test set) but also that the estimation of statistics such as recall are more stable, and our fairness toolkit generalizes better to the test set. In Appendix E.2 Multilingual Experiment (from line 813), we find that the base classifiers in Polish and Portuguese show higher difference in equal opportunity, indicating more severe bias compared to other languages (e.g., English). Where there is such a large initial bias, our toolkit can make larger improvements in fairness, even if the underlying statistics cannot be reliably estimated. --- Rebuttal Comment 1.1: Comment: Thank you for the response and plan to improve the paper. I will keep my scores.	Summary: The paper describes details of a fairness toolkit ("AnonFair"), which confers fairness to any given machine learning classifier by exploring a wide range of prediction thresholds for different groups (which are either provided upfront or inferred through an auxiliary classifier). The toolkit is designed to be quite expressive, as it can optimize several different metrics, e.g., false positives/negatives, true positives, etc. The toolkit can work across all classifiers (which can output class probabilities), including ones trained on vision and NLP tasks. Strengths: The paper introduces and describes a toolkit that implements several fairness strategies and can support any fairness measure that can be expressed in terms of true positives, false positives, true negatives and false negatives. These techniques primarily rest upon adjusting the classification thresholds of different groups, and the paper also incorporates tricks to speed up their computations of precision and recall across different thresholds. The fairness techniques that this paper implements are (largely) classifier agnostic, and can be applied to a wide range of classifiers including NLP and vision classifiers (as this paper shows). Overall, I appreciate that expressivity and broad applicability of their toolkit. Weaknesses: While the toolkit might turn out to be useful for some practitioners, it is a relatively straightforward implementation of well-known (and simple) technique of adjusting prediction thresholds across groups. Exploring different thresholds can be computationally prohibitive, for which the authors use a standard trick to speed up their explorations (which I appreciate). The paper acknowledges and cites relevant papers/techniques that they implement. Overall, the originality and novelty of their work is significantly limited, as the toolkit is an implementation of known and simple fairness techniques. Further, the underlying fairness techniques (not from the authors) are themselves applicable to most classifiers, so any implementation of the same could work for NLP and vision tasks—which is claimed to be one of the major contributions of this work. Technical Quality: 3 Clarity: 2 Questions for Authors: I feel that the current version is a good starting point (in terms of implementation) of existing fairness techniques and speeding them up and trying them out on vision and NLP tasks. To improve the paper, I would suggest clearly outlining the important problems that this toolkit now can enable researchers to answer (which was not possible before) and answer a few of those questions in the paper. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: I believe the paper adequately communicates their shortcomings and cites past references when using them. However, I think it might help to also acknowledge that the underlying fairness techniques broadly apply to a wide range of classifiers, and naturally extend to classifiers in computer vision and NLP domains. Reading parts of the paper felt like that there are significant challenges in adoption of fairness techniques to NLP and CV, and this paper overcomes them through novel solutions—which is not the case. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their review, and we hope to address the issues raised. In brief, there are two main issues we wish to discuss. 1. what this toolkit does 2. The limited novelty of any toolkit/library, and that such libraries are explicitly covered in the call for papers. 1. We are concerned that the review substantially understates what the toolkit does. While our toolkit is a post-processing method, it does not simply ‘set per group thresholds’ and works when group annotations are unavailable at test time. To this end, we offer two modifications: (a) A classifier agnostic approach that uses an auxiliary classifier to estimate groups, while enforcing fairness for the true groups – and not the estimated groups. (b) Model surgery for neural networks (section 4.2, based on work by [4]). This involves training a multi-headed network that jointly estimates groups alongside the original task. These heads are then merged, resulting in a single fair network with one head and the same architecture as the original network. Both approaches have only been previously shown to work for demographic parity [4,5] and not for any other fairness measure. 2. The NeurIPS 2024 Call for Papers explicitly calls for “libraries, improved implementation and scalability” under the infrastructure theme. Software libraries have been published in the NeurIPS main track [1-3]. Like any good toolkit we prefer well-tested and understood components, over speculative new methods. Rather than rehashing an existing argument, we refer the reviewer to this publicly available debate on Openreview from an accepted Neurips toolkit paper last year [1]. We still believe that novelty is important, but novelty should be evaluated with respect to other toolkits, and ask what can this work do that other toolkits cannot? Compared to others, this is the only toolkit that works on computer vision and NLP, and is substantially more expressive than anything else out there. This generalization to NLP, and computer vision, works without group membership at test-time, and is not a natural extension of per group thresholding. Among many other measures, we are the only toolkit that simultaneously supports conditional fairness metrics, minimax fairness, and levelling up through minimum rate constraints. Unlike other toolkits, we jointly and efficiently optimize a performance objective. This not only minimizes degradation while enforcing fairness but can improve the performance of inadequately tuned unfair baselines. Our toolkit is compatible with more frameworks than existing approaches including scikit-learn, Autogluon, and PyTorch. > To improve the paper, I would suggest clearly outlining the important problems that this toolkit now can enable researchers to answer (which was not possible before) and answer a few of those questions in the paper. Please see levelling up [6] examples in figure 4, and the adjacent table, figure 8 and appendix C.3. Minimax fairness [7] in appendix C.1, and table 8. Conditional metrics in appendix C.4. Directional Bias Amplification [8] in appendix C.5. And a whole host of additional measures taken from the fairness review paper [9] in table 7. # References [1] Griffiths, Ryan-Rhys, et al. "GAUCHE: a library for Gaussian processes in chemistry." Advances in Neural Information Processing Systems 36 (2023). [2] Jamasb, Arian, et al. “Graphein-a python library for geometric deep learning and network analysis on biomolecular structures and interaction networks.” Advances in Neural Information Processing Systems 35 (2022): 27153-27167. [3] Pineda, Luis, et al. “Theseus: A library for differentiable nonlinear optimization.” Advances in Neural Information Processing Systems 35 (2022): 3801-3818. [4] Lohaus, Michael, et al. "Are two heads the same as one? Identifying disparate treatment in fair neural networks." Advances in Neural Information Processing Systems 35 (2022): 16548-16562. [5] Aditya Krishna Menon and Robert C Williamson. The cost of fairness in binary classification. In Conference on Fairness, accountability and transparency. PMLR, 2018. [6] Mittelstadt, Brent, Sandra Wachter, and Chris Russell. "The Unfairness of Fair Machine Learning: Levelling down and strict egalitarianism by default." arXiv preprint arXiv:2302.02404 (2023). [7] Martinez, Natalia, Martin Bertran, and Guillermo Sapiro. "Minimax pareto fairness: A multi objective perspective." International conference on machine learning. PMLR, 2020. [8] Wang, Angelina, and Olga Russakovsky. "Directional bias amplification." International Conference on Machine Learning. PMLR, 2021. [9] Hardt, Michaela, et al. "Amazon sagemaker clarify: Machine learning bias detection and explainability in the cloud." Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2021. --- Rebuttal 2: Title: Response Comment: Thanks for your response. I agree with the authors that their toolkit also supports cases when group annotations are unavailable. I realize that my evaluation might not have properly taken this into account. In that light, I have increased my assessment score about the contribution from poor to fair, and overall assessment from 3 to 4. I'm quite aware of the fact that NeurIPS CFP invites libraries as contributions, but that doesn't take away my concerns about the (lack of) novelty and originality of the underlying techniques implemented in the library.	Rebuttal 1: Rebuttal: We thank the reviewers for their helpful and largely positive comments (overall scores 7,6,6,6,3). The suggestions are informative, and we will adjust presentation in the paper wherever an issue has been raised. \ \ Our toolkit provides a “robust and adaptable solution for implementing fairness in machine learning, in ways that other tools do not currently offer” (reviewer aniZ). We “include support to popular and relevant NLP and Computer vision areas” (reviewer 9LRe) and this is “unlike existing toolkits which lack this capability” (reviewer AiSk). We provide “an easy-to-use toolkit for enforcing algorithmic fairness” (reviewer JNWw) and reviewers “appreciate the expressivity and broad applicability” of the toolkit (reviewer AKgy). We contribute “a complete section of experiments and comparison with existing toolkits” and our toolkit represents “progress in algorithmic fairness and enhances multidisciplinary collaborations” (reviewer 9LRe). \ The score 3 (reviewer AKgy) arises from a concern regarding lack of novelty. This is entirely understandable. This is a library submission and as such, most of what we are doing is putting together existing pieces in a useful way and performing good engineering to substantially increase expressiveness and performance (all of which the review acknowledges and appreciates). However, the NeurIPS 2024 Call for Papers explicitly includes “libraries, improved implementation and scalability” under the infrastructure theme. These software libraries have been published in the NeurIPS main track [1-3]. Beyond this, the review misses some of our more substantial contributions. We are not simply adjusting thresholds. All NLP and computer vision experiments are the result of network surgery (see section 4.2), where a new fair network is generated with the same underlying architecture as the base network. These experiments do not use protected attributes at test-time. Moreover, the review requested additional experiments showing fairness definitions that this toolkit, and no other toolkit can solve. These can be seen in tables 6,7 and 8, figure 4, and appendices C.1, C.3, C.4, and C.5. Otherwise, all reviewers are in agreement about our contribution. This is the only fairness toolkit to work for computer vision or NLP and represents a step forward not just in terms of the domains where it has been applied but also in the wide range of fairness constraints that it can solve. # References [1] Griffiths, Ryan-Rhys, et al. "GAUCHE: a library for Gaussian processes in chemistry." Advances in Neural Information Processing Systems 36 (2023). [2] Jamasb, Arian, et al. “Graphein-a python library for geometric deep learning and network analysis on biomolecular structures and interaction networks.” Advances in Neural Information Processing Systems 35 (2022): 27153-27167. [3] Pineda, Luis, et al. “Theseus: A library for differentiable nonlinear optimization.” Advances in Neural Information Processing Systems 35 (2022): 3801-3818.	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper presents AnonFair, a cutting-edge open-source toolkit designed to promote algorithmic fairness. Authors claim the following contributions: (1) Comprehensive support for NLP and Computer Vision classification, as well as standard tabular problems. (2) Enhanced robustness against overfitting challenges through the ability to enforce fairness on validation data. (3) Versatility in optimizing any measure that is a function of True Positives, False Positives, False Negatives, and True Negatives, making it easily adaptable and more expressive than other toolkits. (4) Seamless integration with popular ML toolkits such as sklearn, Autogluon, and pytorch. (5) AnonFair supports 9/9 and 10/10 of the group metrics of two prominent review papers and is accessible online at no cost. Strengths: This toolkit progresses in algorithmic fairness and enhances multidisciplinary collaborations, it is design to integrate the intervention of policy-makers. The paper includes a complete section of experiments and comparison with existing toolkits. AnonFair key contributions include support to popular and relevant NLP and Computer vision areas. Weaknesses: * Lack of clarity in some reported experiments, e.g. results tables are not cited in the text, metrics are not well-contextualized (e.g. larger or lower scores are better?) * Lack of analysis, examples or human evaluation to better understand contributions and limitations of the method in each of the experiments. Technical Quality: 3 Clarity: 2 Questions for Authors: (1) Could you provide more high-level context for each of the experiments that you are running in order to make the paper more self-contained? (2) for NLP experiments, why do you think mitigation works for Twitter and not for Jigsaw? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Authors report some limitations, but further analysis on the experiments could raise more limitations that may be currently ignored. Flag For Ethics Review: ['Ethics review needed: Discrimination, bias, and fairness'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive comments and constructive feedback. --- # Additional clarity in presentation We will use arrows in tables to indicate if larger (↑), or lower (↓) scores are better. We will also discuss this when mentioning the different fairness metrics to improve accessibility for readers. For example, when using Difference in Equal Opportunity (a smaller score is better/fairer). To improve presentation, we will also add references to the results table and relevant appendices to the main body of the paper. --- # Questions and additional analysis of results >Question (1): Could you provide more high-level context for each of the experiments that you are running in order to make the paper more self-contained? We will add text to these experiments. In brief, the computer vision and NLP experiments (tables 1 through 4), are standard fairness benchmarks where existing methods compete to minimize Equal Opportunity (i.e. difference in recall between groups) while maintaining high accuracy. This corresponds to a situation where high accuracy is important, but you don’t want the burden of low recall to disproportionately fall on particular groups. A scenario where this might be important is medical testing where you want to ensure that there is a similar recall rate for all groups. Results in the appendices show the expressiveness of the toolkit. We simply want to show that these well-cited measures are optimizable using our approach. --- >Question (2): for NLP experiments, why do you think mitigation works for Twitter and not for Jigsaw? Good question. While we do improve fairness on Jigsaw, it is not as reliable as on Twitter. One of the nice things about our approach is that we directly enforce constraints on validation data and given sufficient data these results should generalize to unseen test data. This does not happen on Jigsaw with the same reliability we see on Twitter. Largely this can be attributed to data limitations. Equal Opportunity is inherently an unstable measure (it looks at the difference in recall between groups) where both the number of individuals in a particular group is small and the ratio of hate speech is low, we can have very limited data for measuring recall let alone differences in recall. While limited validation data could be worked around using techniques such as cross-fold validation (this is compatible with our approach, and integrating it is ongoing work), Jigsaw comes with a pre-existing test split, and this was frequently unrepresentative of the combined train and val set. This problem is exacerbated as unlike Twitter, Jigsaw contains scenarios with more than two protected groups. We will add an appendix with data counts and discuss the limitations clearly there. --- >Does this additional analysis raise new limitations? We briefly touch on the issue of data in our limitations section, but we did not have sufficient space to discuss this in detail. We’ll add text based on the above answer to the appendix and link to it from the experimental section. --- Rebuttal 2: Title: Any follow-up clarifications? Comment: We hope that we have addressed all issues raised to your satisfaction in our rebuttal. We would be happy to provide additional clarifications if required as the discussion period will be over soon. Thank you for your time. --- Rebuttal 3: Comment: Thanks a lot for your responses, it clarifies my questions, I will keep my score.	null	null	null	null	null	null
G2D: From Global to Dense Radiography Representation Learning via Vision-Language Pre-training	Accept (poster)	Summary: This paper proposes G2D, a novel vision-language pre-training (VLP) framework for medical imaging that aims to learn both global and dense visual representations from radiography images and their associated radiology reports. The key innovation is a pretext task called Pseudo Segmentation (PS), which uses a pseudo mask derived from attention maps to guide the learning of dense visual features during pre-training. The authors demonstrate that G2D outperforms existing medical VLP approaches on various downstream tasks including classification, segmentation, object detection, and zero-shot visual grounding across multiple medical imaging datasets. Notably, G2D shows strong performance on segmentation tasks even when fine-tuned on very limited data. Strengths: Novel approach: The paper introduces an innovative method for learning dense visual representations in medical VLP without requiring pixel-level annotations, addressing a key limitation of existing approaches. Well-motivated: The authors provide a clear rationale for why learning dense representations is important for medical imaging tasks and why existing VLP methods struggle with this. Comprehensive evaluation: The method is evaluated on a wide range of downstream tasks and datasets, demonstrating its versatility and effectiveness across different medical imaging applications. Strong results: G2D consistently outperforms existing methods, especially on segmentation tasks where it achieves impressive results with very limited fine-tuning data. Ablation studies: The paper includes thorough ablation experiments to validate key design choices and components of the method. Potential impact: The proposed approach could significantly reduce the need for large annotated datasets in medical imaging, which is a major bottleneck in the field. Weaknesses: Limited theoretical analysis: While the method is empirically strong, there is little theoretical justification for why the pseudo segmentation task leads to improved dense representations. Complexity of the approach: The method involves several components and processing steps, which may make it challenging to implement and potentially limit its adoption. Computational resources: The pre-training process appears to be computationally intensive (16 A100 GPUs), which could be a barrier for researchers with limited resources. Generalization to other domains: While the focus on medical imaging is valuable, it's unclear how well this approach would generalize to other vision-language domains. Comparison to more recent baselines: Some of the baselines used for comparison (e.g., ConVIRT, GLoRIA) are somewhat older. Comparison to more recent medical VLP methods would strengthen the evaluation. Technical Quality: 3 Clarity: 3 Questions for Authors: Major concerns: My primary concern revolves around the authors' claim that current medical VLP methods primarily align images with entire text reports. This assertion appears to be inconsistent with the facts, as evidenced by several papers that have employed local alignment between image regions and text. This factual contradiction significantly undermines the novelty of the present work. For instance: GLoRIA (Huang et al., ICCV 2021): "Global-Local Representation Alignment for Improved Visual Recognition in Medical Imaging" This paper introduced a global-local alignment approach, learning finer-grained representations by aligning image patches with text tokens. MGCA (Wang et al., arXiv 2022): "Multi-Granularity Cross-Modal Alignment for Generalized Medical Visual Representation Learning" This method employed a multi-granularity alignment strategy, including global, local, and fine-grained levels of alignment. BioViL (Boecking et al., ECCV 2022): "Making the Most of Text Semantics to Improve Biomedical Vision–Language Processing" This work proposed a method to improve biomedical vision-language processing by leveraging text semantics, which includes local alignment strategies. MedKLIP (Wu et al., medRxiv 2023): "Medical Knowledge Enhanced Language-Image Pre-training" This approach utilized external knowledge bases to enhance local alignment, achieving more fine-grained image-text matching. Given these existing works, the authors' characterization of the current state of medical VLP appears inaccurate. This misrepresentation significantly weakens the claimed novelty of their approach. The authors should provide a more accurate description of existing methods and clearly articulate how their approach differs from or improves upon these established local alignment strategies. Other minor concerns: Have you explored the quality of the learned representations at different levels of the network? Are there significant differences in the quality of features at different scales? How sensitive is the method to the choice of threshold used in pseudo mask construction? The ablation shows results for a few values, but is there a principled way to choose this threshold? Have you investigated the potential of using the pseudo masks generated during pre-training for weakly supervised segmentation tasks? How does the performance of G2D change as the amount of pre-training data is varied? Is there a clear relationship between pre-training data volume and downstream task performance? Given the computational requirements for pre-training, have you explored any techniques for making the approach more efficient, such as progressive training or curriculum learning? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors provide a brief discussion of limitations in the appendix, acknowledging potential issues with the weak supervision signal from pseudo masks and the need for further research on regional visual representations. They also touch on broader impacts, mentioning both potential benefits for healthcare and risks associated with sensitive medical data. While these discussions are valuable, they could be expanded to provide more specific insights into the limitations of the current approach and potential mitigation strategies for the identified risks. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the questions! >Theoretical analysis for why pseudo segmentation task leads to improve dense representations - Methods like ConVIRT, GLoRIA, BioViL, MedKLIP, and KAD primarily use an image encoder to extract visual features, aligning them with text embeddings through contrastive learning. MedKLIP and KAD further leverage external tools to extract entities from medical reports for entity classification during VLP However, these methods are limited in their ability to learn dense visual features due to the lack of pixel-level supervision, relying instead on reports or entities. - Methods like MRM employ pixel-level pretext tasks, such as masked image reconstruction, for the encoder-decoder vision model. However, this task lacks high-level semantics, as it merely aims to reconstruct the original pixel values [1,2]. - In contrast, G2D employs pseudo segmentation with pixel-level targets, known as pseudo masks, enabling the vision model to learn dense visual features. G2D's encoder-decoder architecture allows the decoder to become familiar with these dense features, ensuring effective feature learning and continuity when applied to downstream segmentation tasks, unlike methods requiring a randomly initialized decoder. By using pseudo masks derived from attention maps, which align visual features with the medical report, G2D ensures that the pseudo mask reflects the report's semantics. This approach enables the G2D encoder-decoder vision model to learn dense visual features with high-level semantics. > Complexity of the approach: The method involves several components and processing steps, which may make it challenging to implement and potentially limit its adoption. - We kindly disagree. Our method is simple yet effective because (1) G2D does not require annotations and only needs image-text pairs; (2) Our pretraining is a one-stage process, similar to MGCA, and more efficient than MedKLIP and KAD, which require specific entity extraction from medical reports; (3) We directly construct the pseudo mask from the attention map, without relying on an external network, unlike PyramidCLIP[3] and GLIPv2[4]. >Computational resources - We use 16 A100 GPUs solely to accelerate the pre-training stage. To ensure a fair comparison with existing methods, we calculate the contrastive loss on each device independently, rather than aggregating batches across devices. Consequently, adding more devices does not increase the batch size or substantially affect the quality of pre-training. - We have reimplemented the VLP stage on 2 RTX3090 GPUs, using the same computational resources as MGCA. Due to the smaller RAM of the RTX3090 compared to the A100, we set the batch size to 64 on each GPU. The results of this implementation are shown below. As the table indicates, the downstream performance changes only marginally when pretraining on 2 RTX3090 GPUs, demonstrating that G2D is effective even with less computational resources. \| GPUs \| Classification(AUC) \| Segmentation(Dice) \| Detection(mAP) \| \|-------\|----------\|----------\|----------\| \| \| CXR14(1%) \| SIIM(1%) \| ObjectCXR(1%) \| \| 2 RTX3090 \| 78.7 \| 65.3 \| 3.7 \| \| 16 A100s (default) \| 79.1 \| 65.6 \| 3.8 \| >Generalization to other domains - We implemented pretraining on the MIMIC-CXR dataset, following the protocols established by MGCA, MedKLIP, KAD, and others. - Our assumption on image-text region alignment is common and straightforward. Therefore, we believe our method holds value for other vision-language domains, such as natural and remote sensing imagery, due to the importance of dense visual representations across these areas. However, this work is primarily focused on the medical domain. We plan to further explore the potential of our method across various domains in future studies. >Comparison to more recent baselines - We compare our method to M-FLAG (MICCAI 2023), MedKLIP (ICCV 2023), and KAD (Nature Communications 2023), aligning with the comparison methods used in KAD's original work to ensure a fair comparison. >Pseudo masks for weakly-supervised segmentation tasks - This is an interesting direction; we thank the reviewer for the insight and will investigate it in the future. We postulate that our method can serve as an effective pre-training technique for weakly supervised segmentation because it leverages pseudo masks to learn dense visual features without needing precise annotations. >Effect of the amount of pretraining data - For a fair comparison, we implemented the VLP on the full MIMIC-CXR dataset, following the protocols of MGCA, MedKLIP, MRM, KAD, and others. We acknowledge that the amount of pretraining data is a critical and interesting research question for medical VLP, and we plan to investigate this in our future work. >Progressive training or curriculum learning - Thank you for the advice. We have conducted an ablation study using less computational resources—2 RTX 3090s. As shown in the table above, the results demonstrate that G2D also performs well with fewer computational resources. >Clarification on other medical VLP methods’ limitations in alignment\ >Quality of different level visual features\ >Sensitivity analysis on threshold used in pseudo maks construction - We have added a detailed explanation in the official comment below. Please refer to it for more information. [1] Liu Y, et al. Improving pixel-based mim by reducing wasted modeling capability, ICCV, 2023.\ [2] Liu Y, et al. PixMIM, TMLR, 2023\ [3] Gao, Yuting, et al. PyramidCLIP, NeurIPS 2022\ [4] Zhang, Haotian, et al. GLIPv2, NeurIPS 2022 --- Rebuttal Comment 1.1: Title: The rebuttal addresses my concerns Comment: I thank authors for their efforts in addressing my concerns. After reading the rebuttal, my concern has been addressed. And I will update the final rating. --- Reply to Comment 1.1.1: Comment: Thank you for your feedback and for considering our response. We appreciate the opportunity to clarify our work and are grateful for your thoughtful review. --- Rebuttal 2: Comment: Continue with the rebuttal >Clarification on other medical VLP methods’ limitations in alignment - We have described GLoRIA, MGCA, MedKLIP, and KAD in Section 2 of the main article and will provide a detailed explanation and comparison with G2D in the following sections. - Ambiguous Token-level Alignment in GLoRIA and MGCA: Both GLoRIA and MGCA employ a brute-force approach to align image and text tokens. This token-level alignment might compromise the medical context and lead to misalignments. For example, medical terms such as 'compatible' or 'acute' lack direct visual correlates, making local alignment ambiguous. - Global Alignment only in BioViL: BioViL implements only global alignment, as detailed in their original work (equation 2 and section 2.2). During pre-training, their loss functions include global image-text alignment and masked language modeling on the text side, but they do not incorporate a loss for dense visual representation learning. - Loss Functions in MedKLIP: During pre-training, MedKLIP utilizes an entity classification loss, applying it to all image features for classifying entities, and a contrastive loss where different positional names are treated as negative samples using their name embeddings for contrastive learning. However, these loss functions are not explicitly designed for fine-grained image-text matching. - Unique Approach of G2D: Unlike GloRIA, MGCA, BioViL, MedKLIP, and KAD, which only use an image encoder during pretraining, G2D employs an encoder-decoder architecture to enhance visual representation learning. During VLP, the encoder extracts global visual features aligned with text to learn global visual representations. Additionally, G2D incorporates a decoder that performs pseudo-segmentation tasks using pseudo masks generated by the G2D encoder, independent of external annotations. This decoder leverages features from the image encoder for pseudo-segmentation, enabling both the encoder and decoder to jointly learn dense visual representations. >Quality of different level visual features - Most network structures for dense prediction tasks use an encoder-decoder architecture, where the features at the penultimate layer of the decoder are used for final pixel-wise prediction. For image classification, the structure usually involves an encoder with a linear classifier, where the features of the penultimate layer of the encoder are used for prediction. For both representative cases, the effectiveness of the penultimate layer of the encoder or decoder is demonstrated in Section 4. These experiments have shown the effectiveness of our approach in the most general settings (i.e., both classification and dense prediction). >Sensitivity analysis on threshold used in pseudo maks construction - We conducted a detailed ablation study on various threshold values when building the pseudo masks. The results are shown below. As the table indicates, the best downstream performance is achieved with an 85th percentile threshold. Increasing the threshold to the 95th percentile does not improve performance, suggesting that an extremely high threshold may lead to over-filtering. Conversely, decreasing the threshold from the 85th to the 25th percentile consistently degrades performance, as a lower threshold causes the pseudo mask to cover most of the image, introducing noise during VLP. Based on these experimental results, we empirically set our threshold to the 85th percentile. - In the future, we will investigate using an adaptive threshold to filter the attention map. \| Threshold \| Classification(AUC) \| Segmentation(Dice) \| Detection(mAP) \| \|-------\|----------\|----------\|----------\| \| \| CXR14(1%) \| SIIM(1%) \| ObjectCXR(1%) \| \| 95% percentile \| 78.5 \| 64.8 \| 3.7 \| \| 85% percentile (default) \| 79.1 \| 65.6 \| 3.8 \| \| 75% percentile \| 78.3 \| 63.0 \| 3.4 \| \| 50% percentile (median) \| 75.6 \| 58.8 \| 2.3 \| \| 25% percentile \| 75.2 \| 65.6 \| 2.1 \|	Summary: This manuscript describes a medical vision-language pre-training framework called Global to Dense level representation learning (G2D), that learns global and dense visual features simultaneously with only image-text pairs, by exploiting the aggregated attention map from the vision encoder for a pseudo segmentation pretext task. The improved (frozen) vision encoder is then utilized as part of the model pipeline for a number of downstream tasks (e.g. segmentation, classification) Strengths: - Pseudo segmentation pretext task enables dense segmentation during pre-training, and avoids external resources as for alignment-based methods, and limitations on high-level semantic representations in reconstruction-based methods - Importance of associating semantic meaning verified via experiment Weaknesses: - Unclear if specific sentence/phrase to individual image region alignment is achieved, for dense learning - Lack of fine-grained pixel-level evaluation of masks Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The accuracy of the initial aggregated attention map appears possibly non-optimal, given that additional thresholding by body mask is required. As such, it might be considered to quantify the accuracy of these maps, possibly against segmentation ground truth. 2. In Section 3.2, it is stated that a threshold is applied (at 85%) to transform the aggregated attention map into a binary mask, before smoothing. It might be clarified if the need for smoothing (and related smoothing parameters) was empirically determined. 3. In Section 3.3, it is stated that "This decoder takes visual feature V_i as input and utilises the pseudo mask ˜M_i as the supervisory signal for the pretext task". It might be clarified as to whether and how specific text can be matched to specific (separate) image regions, as in Figure 4 of Section A.7. In other words, while Figure 4 shows specific text descriptions corresponding to specific image regions, were these correspondences/alignments indicated by the proposed G2D model, or are they external manual observations? A.1 suggests no, but this might be explicitly stated. 4. In Section 4, the choice of ResNet-50 as the encoder over other plausible choices (e.g. U-Net encoder) might be briefly explained. 5. For Table 1, it might be clarified as to what "encoder-decoder" refers to - the updating of both encoder and decoder? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedbacks! > Unclear if specific sentence/phrase to individual image region alignment is achieved, for dense learning (W1) - Since the MIMIC-CXR pretraining dataset does not establish a direct relationship between specific sentences or phrases and image regions, direct alignment between them is infeasible. To facilitate the learning of dense visual features, we construct pseudo masks for a pseudo segmentation pretext task, as detailed in Section 3.2. - Notably, our method has a significant advantage during medical VLP as it learns dense visual features using only image-text pairs. It does not rely on specific sentence or phrase annotations paired with individual image regions. This means our approach is more generalizable and does not require region-level annotations. > Evaluation of pseudo masks. (W2, Q1) - In the G2D approach, our objective is to enhance medical VLP by designing a pseudo mask for learning dense visual features through a pseudo segmentation task. This method does not rely on guessing the semantic mask as in traditional supervised learning. Since the MIMIC-CXR dataset used for pretraining lacks pixel-level annotations, it is impossible to directly assess the accuracy of the pseudo masks created through G2D. However, preliminary quality checks, as detailed in Appendix A.7 and illustrated with two examples in Figure 4, show that G2D successfully identifies image regions that align with the content of the entire report using just language cues and pseudo mask supervision. - We acknowledge that the pseudo masks are not perfect. However, adding body masks is one of the simplest and most commonly used operations in medical image processing[1]. > Ablation on smoothing pseudo masks (Q2) - We assess the impact of smoothing on the G2D model's performance in Table 5-f by comparing versions with and without smoothing. The results show that the model variant with smoothing outperforms the one without, especially in the SIIM segmentation task. This indicates that smoothing the pseudo mask enhances the learning of representative visual features during VLP particularly for dense visual features. - To further investigate the effect of smoothing, we performed an ablation study on the 'window_size' parameter when implementing a bilateral filter. The results are shown below. As the findings indicate, variations in 'window_size' do not substantially affect the quality of VLP, demonstrating that the G2D method is robust across different parameter settings for smoothing operations. \| Window Size \| Classification(AUC) \| Segmentation(Dice) \| Detection(mAP) \| \|-------\|----------\|----------\|----------\| \| \| CXR14(1%) \| SIIM(1%) \| ObjectCXR(1%) \| \| 5$\times$5 \| 79.2 \| 65.5 \| 3.8 \| \| 7$\times$7 (ours) \| 79.1 \| 65.6 \| 3.8 \| \| 10$\times$10 \| 79.1 \| 65.6 \| 3.7 \| > Clarifying Text-to-Image Region Correspondences in G2D Model (Q3) - During pre-training, the pseudo mask for each sample is generated by the G2D model using the attention map, without the need for external manual annotations. - Our approach does not attempt to align specific text with specific image regions. Instead, our primary objective is to establish a semantically meaningful target for dense representation learning during pre-training. We utilize the semantic information encoded in the image-text attentions, rather than focusing on precisely predicting downstream segmentation tasks, which is theoretically unrealistic. - In Appendix A.7, the image region depicted is the pseudo mask derived from the G2D model, not one annotated by humans. - We will clarify this part in the camera-ready version according to your suggestion. > Choice of U-Net encoder (Q4) - ResNet-50 is the most commonly used vision encoder for vision-language pre-training (VLP) methods, such as CLIP. - For a fair comparison, we strictly adhere to the protocols established by MGCA, MedKIP, KAD, and others, using ResNet-50 as the image encoder. - The latest version of nnU-Net [4] utilizes a ResNet backbone and demonstrates improved performance compared to a vanilla, non-ResNet backbone. > Clarification for Tab 1 'encoder-decoder' (Q5) - We adhere to the protocols established by GloRIA[2] and MGCA[3] by only updating the decoder while keeping the encoder frozen during training. We will update and clarify this part in the camera-ready version of our document. [1] Imura, Masataka, et al. "Automatic cropping method of chest radiographs based on adaptive binarization." EMBC, 2013.\ [2] Huang, Shih-Cheng, et al. "Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition." ICCV. 2021\ [3] Wang, Fuying, et al. "Multi-granularity cross-modal alignment for generalized medical visual representation learning." NeurIPS 2022\ [4] Isensee, Fabian, et al. "nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation." CoRR 2024. --- Rebuttal Comment 1.1: Comment: We thank the authors for their clarifications.	Summary: The paper proposes an encoder-decoder medical VLP approach for global-to-dense visual representation learning. Pseudo segmentation is adopted for dense level learning. Rich experiments validate the effectiveness of the proposed method. Strengths: 1. The motivation behind the work is clear. Pseudo-segmentation supervision is effective, which is validated by experiments. 2. The experiments are rich and ablation analysis shows the contributions of each component and design. 3. The illustrations are clear and easy to understand. 4. The improvements are consistent and sometimes substantial. Weaknesses: 1. The comparisons with MGCA and MRM in the CXR14 dataset are not included in Table 3, but Table 4 includes the comparisons with MGCA and MRM. What are the reasons behind this? 2. Transformer-based vision encoder is not analyzed. 3. The balance between VLA and PA losses is not analyzed. Technical Quality: 4 Clarity: 4 Questions for Authors: Is it not applicable to compare with MGCA and MRM in the CXR14 dataset? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedbacks! >Comparing with MGCA and MRM on CXR14 datasets (W1,Q1) - In Table 3, we directly reference the results from the KAD study to ensure a fair comparison, as KAD[1] uses the official data split for CXR14. It's important to note that the KAD[1] study does not include results for MGCA[2] and MRM[3]. - In the MGCA[2] study, results on CXR14 are not reported, which makes direct comparison challenging. Meanwhile, the MRM study, although it includes experiments on CXR14, uses its own data split rather than the official split provided by CXR14. This leads to potential biases when comparing it with existing methods with official split， such as KAD. Therefore, in Table 3, we only compare our results with those methods that have reported outcomes using the official CXR14 split, as documented by the KAD study. - To comprehensively compare our G2D method with MGCA and MRM, we re-implemented both on CXR14 using the official split as employed by KAD. For finetuning, we utilized their officially provided pretrained weights. Additionally, we used the official finetuning code from the MRM GitHub repository, making a single modification: we replaced their own data split with the official split used by KAD. The results are shown in the table below. As the table indicates, G2D substantially outperforms both MGCA and MRM on the CXR14 dataset using the official data split: \| Methods \| CXR14(1%) \| CXR14(10%) \| CXR14(100%) \| \|-------\|----------\|----------\|----------\| \| MGCA[2] \| 62.4 \| 73.9 \| 81.2 \| \| MRM[3] \| 64.2 \| 74.3 \| 81.0 \| \| G2D(ours) \| 79.1 \| 81.1 \| 83.1 \| >Transformer-based vision encoder is not analyzed (W2) - We conducted the experiments using a transformer-based vision encoder, specifically the ViT, configured identically to that used in the MRM study [3]. For the vision decoder, we also employed a transformer-based architecture, the same as MRM. The results of G2D with various transformer variants are shown in the table below. As indicated by the table, the performance of G2D does not fluctuate significantly with different backbone types, demonstrating that our method is backbone-agnostic. \| Backbone \| Classification(AUC) \| Segmentation(Dice) \| Detection(mAP) \| \|-------\|----------\|----------\|----------\| \| \| CXR14(1%) \| SIIM(1%) \| ObjectCXR(1%) \| \| G2D(CNN) \| 79.1 \| 65.6 \| 3.8 \| \| G2D(Transformer) \| 78.8 \| 65.7 \| 3.5 \| >The balance between VLA and PA losses is not analyzed (W3) - We selected a coefficient of 1 for both the VLA and PA losses from the initial development of the project, as we believe a robust method should not require specifically tuned coefficients for each loss. - Due to time constraints during rebuttal, we plan to ablate the coefficients of these two losses in future work to comprehensively investigate their contributions to the G2D method. [1] Zhang, Xiaoman, et al. "Knowledge-enhanced visual-language pre-training on chest radiology images." Nature Communications 2023\ [2] Wang, Fuying, et al. "Multi-granularity cross-modal alignment for generalized medical visual representation learning." NeurIPS 2022\ [3] Zhou, Hong-Yu, et al. "Advancing Radiograph Representation Learning with Masked Record Modeling." ICLR 2023. --- Rebuttal Comment 1.1: Title: Thank you Comment: Thank you for addressing my concerns! I have no further questions and maintain my original rating. --- Reply to Comment 1.1.1: Comment: Thank you for taking the time to reassess your concerns and for maintaining your positive feedback. We truly appreciate your thoughtful insights and feedback.	Summary: The paper proposes a new medical vision-language model, G2D, which employs vision-language alignment (VLA) and pixel alignment (PA) strategies, combined with a pseudo segmentation (PS) pre-training task, to learn global and dense visual representations from medical images. The VLA strategy is used to learn global representations of images and texts, while the PS task constructs pseudo masks through a parameter-free mechanism to facilitate the learning of dense representations. The method is comprehensively validated across five downstream tasks (image segmentation, object detection, zero-shot image visual grounding, zero-shot image classification, and fine-tuned image classification), demonstrating its effectiveness in handling both unimodal and cross-modal tasks. Strengths: + The paper is well-written, with the motivation, method, and results clearly presented. A minor concern is the reference format; it should be [1] instead of (1) according to the NeurIPS template. + A significant concern with most existing works is that they operate primarily at the Image-Text Retrieval level, similar to the perceptual level of CLIP, and do not effectively capture dense features between modalities. The G2D model addresses this issue by integrating Vision-Language Alignment (VLA) and Pseudo Segmentation (PS) tasks to facilitate simultaneous learning of global and dense visual features. This multi-level feature learning significantly enhances the model's performance in tasks requiring dense feature perception, such as segmentation. + During pre-training, the G2D method utilizes only image-text pairs without the need for additional annotated data. By generating pseudo masks on the fly through the PS task, it reduces the cost and complexity associated with data annotation. + The G2D method is novel, and the experiments are robust. Experimental results on five medical imaging tasks involving 25 diseases demonstrate that the G2D model outperforms existing models, even with minimal fine-tuning data. Notably, in segmentation tasks requiring dense visual features, G2D achieves excellent results with just 1% of the training data for fine-tuning. Weaknesses: Major concerns: - The attention maps could introduce errors in pseudo mask, and these errors may propagate throughout the training process. To address this, a clear validation strategy needs to be outlined. For instance, in Figure 2, aggregated attention map might incorrectly highlight irrelevant regions. It is essential to establish methods for detecting and measuring these errors to ensure the reliability of the model. I hope the authors could quantify the errors in aggregated attention map and pseudo mask during the rebuttal period. Minor concerns: - The training and validation of the model rely on specific datasets, which may introduce biases and potentially affect the model's generalizability to different datasets. - It is uncertain whether the method can be effectively extended to vision-language tasks involving 3D imaging (e.g., CT and MRI), presenting a limitation in its current scope of application. Technical Quality: 3 Clarity: 4 Questions for Authors: - How do you detect and correct the errors made by aggregated attention map? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Limitations were discussed in Section A.1 Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedbacks! >Detecting and measuring the error of pseudo mask (W1, Q1) - In G2D, we aim to design pseudo mask for learning dense visual feature from pseudo segmentation task during medical vision-langauge pre-training (VLP), rather than directly guess the semantic mask for supervised learning. - Since the MIMIC-CXR dataset, which is used for pretraining, does not have segmentation mask annotations, it is infeasible to directly evaluate the accuracy of the pseudo mask derived from the G2D method. - In Appendix A.7, where we conducted a quality check on several samples and visualized two examples in Figure 4. We observed that G2D is capable of learning the image regions of interest that correspond with the entire report, using only language and pseudo mask supervision. However, a detailed quantitative evaluation would require laborious work by clinicians, and we plan to consider this in future studies. >The training and validation of the model rely on specific datasets, which may introduce biases and potentially affect the model's generalizability to different datasets.(W2) - We compared our approach with well-established works such as GLoRIA (ICCV 2021), MGCA (NeurIPS 2022), MedKIP (ICCV 2023), and KAD (Nature Communications 2023). All these studies utilize the MIMIC-CXR dataset, which is commonly used for 2D medical VLP. - To ensure a fair comparison with existing methods, we strictly adhered to their experimental settings [1,2,3,4], using the same datasets for both pretraining and downstream evaluation. - Furthermore, due to limitations in publicly accessible datasets, MIMIC-CXR is the only large-scale medical image-text dataset, containing over 200,000 samples, available for implementing medical VLP. We hope the research community will release more publicly available datasets for VLP to reduce bias and enhance model generalizability. >It is uncertain whether the method can be effectively extended to vision-language tasks involving 3D imaging (e.g., CT and MRI), presenting a limitation in its current scope of application. (W3) - Our method can be easily adapted to 3D imaging modalities by replacing the 2D image encoder with a 3D version. However, there is currently no public large-scale 3D image-text dataset comparable to MIMIC-CXR, which has over 200,000 samples, for implementing 3D medical VLP. We note that scaling our proposed framework to native 3D is straightforward because the pseudo masks are derived from attentions (see Section 3.2). We will explore the potential of our work further if such datasets become publicly accessible. [1] Huang, Shih-Cheng, et al. "Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition." ICCV. 2021\ [2] Wang, Fuying, et al. "Multi-granularity cross-modal alignment for generalized medical visual representation learning." NeurIPS 2022\ [3] Wu, Chaoyi, et al. "Medklip: Medical knowledge enhanced language-image pre-training for x-ray diagnosis." ICCV 2023\ [4] Zhang, Xiaoman, et al. "Knowledge-enhanced visual-language pre-training on chest radiology images." Nature Communications 2023 --- Rebuttal Comment 1.1: Comment: I appreciate the authors' detailed responses, which effectively addressed my previous concerns. As a result, I'd like to raise my rating to Weak Accept. Regarding pseudo label evaluation, per-voxel annotations may not be necessary. Based on the report, if the disease is present in the image and the pseudo labels correctly identify it, this counts as a true positive; otherwise, it's a false negative. Similarly, if the report indicates the image is healthy, the authors could calculate the number of true negatives and false positives for the pseudo labels. This strategy might be able to evaluate the quality of pseudo labels. --- Reply to Comment 1.1.1: Comment: Thank you for your thoughtful review and for considering an upgrade in your rating based on our responses. We are grateful for your suggestion on evaluating pseudo labels without per-voxel annotations and will explore implementing this strategy to further validate our methodology. Your insights are invaluable to enhancing the quality of our work.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Generative Semi-supervised Graph Anomaly Detection	Accept (poster)	Summary: This paper works on node anomaly detection in the novel semi-supervised setting where few labeled normal nodes are given and proposes to generate new anomaly nodes to boost the training data. The anomaly generation algorithm is inspired by the empirical observation that: (1) Anomaly nodes have lower affinity score than normal nodes (2) Feature distribution of anomaly nodes are similar to normal nodes if they share similar neighborhood patterns. Strengths: (1) The setting is novel and aligned to the real-world situation where normal nodes are typically known compared with anomaly nodes. (2) The motivation for the proposed two regularization losses is very intuitive and clear. (3) The experimental results are very impressive. Weaknesses: (1) The proposed two regularization losses are heavily based on the empirical analysis, which might not transfer to other anomalies in other datasets. (2) For the second prior, its assumption that anomaly nodes sharing similar local structures would share a similar feature distribution has not been empirically verified. (3) Experiments miss the comparison with diffusion-based generative anomaly detection baseline. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) As stated in the weakness, the core regularization loss terms are designed based on two assumptions: * The anomaly nodes have a lower affinity score than normal nodes. However, there is no comprehensive experimental verification of the other datasets on this. It might be better to provide the verification like Figure 1 but on more different datasets. * Anomaly nodes sharing similar neighborhood structures should possess similar feature distributions to their corresponding normal nodes. Although some references have been attached to justify this hypothesis, it might be better to include some empirical verification on this as well. Furthermore, there might be some contradiction between these two assumptions by themselves. First, if assumption 1 holds, it means anomaly nodes should share different local subgraphs with the normal nodes, which indicates that assumption 2 cannot hold. How do we mediate this situation? (2) Is there any difficulty when optimizing the loss according to Eq. (4) and Eq. (5) at the same time? Firstly, for Eq. (4), since the fixed terms would be embeddings of normal nodes and their neighbors, the embeddings of abnormal nodes ($\hat{\mathbf{h}}_i$ in Eq. (2)) would be optimized towards being further away from the neighbors' embeddings. However, Eq. (5) would also enforce the $\hat{\mathbf{h}}_i$ to be close to the normal one $\mathbf{h}_i$. These two directions seem to be contradictory to each other. (3) Joint optimization according to Eq. (7) does not make sense under this generative augmentation setting. Here we use a generative model to augment the training data. This therefore should be that the training model is fixed. Moreover, if we jointly optimize the anomaly detection term and the other two generative terms, it would lead to the gradient for anomaly detection leaks to classification. This is quite confusing to me and might need more clarification. (4) How many layers of the subgraphs are used in optimizing the affinity score? If we use 2-hop neighbors, it might cause the computation to consider the significantly large number of nodes. If not, how should we decide on this parameter? (5) The comparison misses the baseline [1] [1] Liu, Kay, et al. "Graph diffusion models for anomaly detection." (2024). Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: In addition to the limitations mentioned by the author, there are some other limitations worth addressing: (1) The currently proposed anomaly generation method is still operated in the embedding space. As admitted by the author anomaly behavior is heavily based on interactional behaviors, therefore, it is also helpful to consider directly characterizing/generating anomaly in the graph space. (2) The comparison misses one generative-based baseline [1] [1] Liu, Kay, et al. "Graph diffusion models for anomaly detection." (2024). Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for the constructive comments and questions. We are grateful for the positive comments on the novelty and soundness of the experiments. Please see our detailed one-by-one responses below. > Weaknesses #1 The regularization is heavily based on the empirical analysis, which might not transfer to other anomalies. The graph data is non-i.i.d. data with rich structure information, and the representation of a node should be grounded in a local context. Thus, as shown in Fig.1 and Fig.2 in the uploaded pdf, these two important priors about graph anomalies generally hold in all these real-world graph anomaly detection datasets. This is also one main reason why the popular graph reconstruction and graph contrastive learning methods for GAD are generally effective on various GAD datasets, since their intuitions are essentially based on similar priors as ours (though they did not explicitly summarize and propose the priors). We agree that there can be some anomalies that may not well conform to these two priors. GGAD can work well for these cases too. This is because the outlier nodes generated by GGAD are essentially located at the fringe of the normal nodes in the representation space, as shown in Fig.3 (c). GGAD then leverages these outlier nodes to build a one-class classifier with a decision boundary tightly circled around the normal nodes. Such decision boundary can discriminate not only the anomalies well simulated by the outlier nodes but also the other anomalies that lie at the same side of the outlier nodes. > Weaknesses #2 and Questions #1 More verification on more datasets and empirical verification of the second prior Please refer to our reply to Global Response to Share Concern #1 in the overall author Rebuttal section above for this concern. > Weaknesses #3 and Limitations #2 and Questions #5 The lack of diffusion model based generation Thanks for pointing out the related study, but the work and other diffusion-based GAD methods as well focus on the fully supervised setting where both labeled normal and abnormal nodes are required during the training. This is different from the semi-supervised setting we proposed in our paper. We will discuss and include this work in our revision. Additionally, we have added more ablation studies in the outlier generation module, please refer to the response to Reviewer JEU8's Question #4. > Questions #2 How do we mediate the two assumptions that might be some contradiction and is there any difficulty in optimization? Please refer to our reply to Global Response to Share Concern #2 in the overall author rebuttal section above for this concern. > Questions #3 Clarification on the joint optimization with anomaly detector Different from existing augmentation-based generation, since we generate the outliers in the latent space rather than the raw node attribute space, the joint optimization allows the model to impose both anomaly priors more effectively. The total loss generally decreases and converges finally, see Fig. 3 in the uploaded pdf. To further investigate the benefits of joint optimization, we compare the two-step optimization method with our joint optimization approach. The experimental results are shown in Table A1. The results show that the joint optimization significantly outperforms the two-step approach. The main reason is that many more effective outlier samples can be generated during the optimization process of the two anomaly priors. By jointly optimizing with the BCE loss function, these outliers can be fully exploited for the training of the one-class classifier, enabling a mutually enhanced outlier node generation and one-class classification process. ``` Table A1. Results of two-step training and joint training (AUPRC/AUROC). ``` \|Data\|Amazon\|T-Finance\|Reddit\|Elliptic\|Photo\| DGraph\| \|:---- \|:----: \|:----: \|:----: \|:----: \|:----: \|:----: \| \| Two-step \| 0.1899 /0.7322 \| 0.0805/0.7022 \| 0.0399/0.5233 \| 0.2282/0.6873 \|0.1179/0.6089 \|0.0050/0.5392 \| \| GGAD(Joint)\|0.7922/0.9443 \| 0.1825/0.8228 \| 0.0610/0.6354 \| 0.2425/0.7290\| 0.1442/0.6476 \|0.0082/0.5943 \| > Questions #4 The number of layers of the subgraphs, how should we decide on this parameter We used one layer subgraph by default as the egonet of the target node to calculate the affinity score, because anomalies are primarily reflected in the relationships among the neighbors immediately connected to themselves [Ref1-2]. We agree that incorporating 2-hop neighbors can include more community information for GAD but might cause more computational overhead. We leave it as a promising future work direction. > Limitations #1 The generation is still operated in the embedding space, heavily based on interactional behavior. Thank you very much for pointing out this. Generation in the latent space is a simple yet effective way to support efficient generative GAD training. It can also mitigate the notorious over-smoothing representation problem in the GNN aggregation. Generating in the graph space is also a promising direction that offers a valuable approach for analyzing characteristics and enhancing interpretability. We will include and discuss this limitation in the final paper. References: - [Ref1] Addressing Heterophily in Graph Anomaly Detection: A Perspective of Graph Spectrum, WWW2023 - [Ref2] Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection, NIPS2023 --- Rebuttal Comment 1.1: Title: Appreciate your response! Comment: Thank you for more analysis. I have follow-up questions as follows: Weaknesses #2 and Questions #1 More verification on more datasets and empirical verification of the second prior: after seeing more visualization of normal/abnormal/gaussian-generated outliers distributions, I have the following questions: why would adding Gaussian noise make the outlier distribution like a single bar? Weaknesses #3 and Limitations #2 and Questions #5: thank you for the clarification and it is reasonable. Questions #2 How do we mediate the two assumptions that might be some contradiction and is there any difficulty in optimization?: I am not so convinced about the reply. Still, my initial concern is - (1) One optimization is to make the generated node embedding itself similar to the original node embedding. - (2) The other optimization is to make the generated node embedding itself have a lower affinity score than the neighbors of the original nodes Let's assume (1) works and the generated node embeddings become further away from neighbor embeddings, then how would (2) still hold given that the original node embeddings have a higher affinity score to the neighbors? Furthermore, I want to add on top of my above questions with another question given that this designed method needs to select normal nodes to generate abnormal nodes. How do we select the initial normal nodes? Do we use all of them or randomly sample them? Without further addressing my above concern, I cannot further raise my score. --- Rebuttal 2: Title: Response to Reviewer 7mdh Comment: We greatly appreciate your further comments and it's great to know that our response has helped address some of your questions. Please find our point-by-point response to your follow-up questions as follows. >Questions #1 Why would adding Gaussian noise make the outlier distribution like a single bar? We'd like to clarify that the Gaussian noise works like a hyperparameter in the feature interpolation in Eq. (5) to diversify the outlier nodes in the feature representation space. Changes of this noise distribution do not affect the superiority of the detection performance of GGAD over the competing methods (please see the results in Table A1 in our response to Reviewer JEU8 for empirical justification). Further, the outliers are generated based on the neighbors of some sampled normal nodes, so the number of the outlier nodes is far less than that of all normal samples, leading to a relatively sparse distribution of the outlier nodes in the affinity density map compared to the normal nodes in the figure. However, this does not affect their effectiveness in serving as negative samples for training the one-class classifier. >Questions #2 Concern on two constraints We achieve the mediation of two constraints by jointly minimizing $L_{ala} $ and $L_{ec}$. The mediation leads to a result where the generated outlier nodes lie at the fringe of the normal nodes in the feature representation space, as illustrated in Fig. 1(b) and Fig. 3(c). Such outlier nodes meet the two criteria evenly you mentioned in the above comment. In terms of optimization, from Fig.3 in the uploaded pdf, we can see that the overall loss and the two individual losses gradually converge at around similar epochs where both constants are satisfied for the generated outliers; not unstable or fluctuated bounces are found after these epochs. Thus, we didn't experience any difficulty in the optimization of GGAD across all the datasets used. >Questions #3 How do we select the initial normal nodes? As mentioned in lines 202-204, We randomly sample a set of S normal nodes and respectively generate an outlier node for each of them based on its ego network. In line 300, we indicate the size of the generated outlier node S is set to 5\% of by default. As shown in Fig. 6 and Fig. 7 in App C.1 of the paper, the performance of GGAD generally remains stable w.r.t the number of the generated outlier nodes. We hope the above replies help address your concerns. We're more than happy to engage in more discussion with you to address any further concerns you may have. Thank you very much for helping enhance our paper again! --- Rebuttal 3: Title: Follow-up further Comment: Question #1: Why would adding Gaussian noise make the outlier distribution like a single bar? My question here is not whether tweaking the Gaussian noise would enhance previous baseline performance but more like if we can tweak the added Gaussian noise, would the motivation figure also suffer from this issue (the distribution of generated outliers in previous methods are further away than ground truth ones compared to the proposed methods)? However, since the presented results show significantly better performance of the proposed method, I guess it should be that no matter what level of Gaussian noise we add, they would still not be more like the ones generated by the proposed method. Question #3: How do we select the initial normal nodes? This makes sense to me. Thanks! Question #2 Concern with two constraints I am still confused here and would hear more guidance on this point. Because neighbors embeddings and original node embeddings are fixed and assuming we are working on the homophily social networks (which is a widely adopted property in real-world datasets), how could we enforce the generated outlier node embeddings to be further away from neighbors but at the same time be close to the original center node. Although I still have questions on Question #2 Concern with two constraints. Overall, I appreciate the authors' response and think the observation of this paper would still be useful in future research in anomaly detection. I increase my score but hope authors could further address my questions here Question #2 Concern with two constraints. --- Rebuttal Comment 3.1: Title: Response to Reviewer 7mdh Comment: Thank you very much for raising the score. We greatly appreciate your further comments and it's great to know that our response has addressed most of your questions. Please find our response to this concern. >Follow-up question with Question #1 Why would adding Gaussian noise make the outlier distribution like a single bar? Yes, your understanding is correct. Changing the level of Gaussian noise in the generation of the outlier nodes in the baseline methods does not affect the detection performance of the baseline methods AEGIS and GAAN, since they still suffer from the absence of considering graph structure information in their generation model. >Follow-up question with Question #2 Concern with two constraints We agree that in the homophily graph, the nodes tend to connect with the nodes from the same class, leading to the relatively high similarity between the embeddings of the target node and their neighbors. However, this does not affect the learning of the outlier nodes our method aims to obtain. This is because, given the embeddings of a fixed target node and their neighbors, minimizing the two proposed losses jointly would result in a mediation in the feature representation space where the generated outlier nodes are close to, yet separable from, the target normal node and its neighbors. Thus, these outlier nodes can be thought as `hard anomalies` that lie at the fringe of normal nodes in the feature representation space. If the egocentric closeness loss is removed, the generated outliers would become `trivial anomalies` that distribute far away from the normal nodes (see Fig. 3(a) in the paper). On the other hand, if the local affinity loss is removed, the generated outliers would then become `misleading/false anomalies` that lie inside the normal nodes in the feature representation space (see Fig. 3(b) in the paper). We hope the above reply helps address your follow-up questions. We will clarify this point in our final version. We're more than happy to engage in more discussion with you to address any further questions you may have. Thank you very much for helping enhance our paper again!	Summary: The paper proposes a novel approach called GGAD aimed at improving anomaly detection in graphs under a semi-supervised framework. GGAD generates pseudo anomaly nodes that serve as negative samples for training a one-class classifier. This method is built on two key priors: asymmetric local affinity and egocentric closeness, which help in generating reliable outlier nodes that mimic real anomalies in terms of both graph structure and feature representation. Extensive experimental results demonstrate the effectiveness of the method across diverse graph anomaly detection datasets. Strengths: 1.The method is innovative. The proposed graph anomaly detection method can exploit the feature and structure information of normal nodes more effectively in the studied semi-supervised scenario compared to existing methods. The proposed two priors provide a meaningful characterization of desired properties of outliers in this semi-supervised setting and can be utilized to explore other beneficial priors further. 2.The experiments in the paper are comprehensive and thorough. Weaknesses: 1. The model relies on prior knowledge to generate anomaly points. This prior knowledge can limit the model’s application scenarios. The model performs best only when the real anomalies align with this prior knowledge. For anomaly types that do not conform to the prior knowledge, the model may not effectively detect them. 2.The model does not perform best on the Photo dataset in Table 1, and the article lacks an explanation of the results at the overall data level. 3. This model employs a semi-supervised approach that uses some positive samples for training. However, it does not consider the issue of noise interference within the positive samples, namely, how the model overcomes interference when some positive samples are mislabeled. 4. During the initialization step, only the initial feature of outliers are obtained while the connections between the outliers and normal nodes are not well illustrated in the paper. From Figure 2, one outlier is connected to more than one normal node while the feature of the outlier is generated according to single normal node. The neighborhood of outliers is important since the it involves the computation of node affinity score of outliers. Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: yes, the authors point out that some anomalies whose characteristics may not be captured by the two priors used Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for the constructive comments. We are grateful for the positive comments on our paper clarity, research motivation, and empirical justification. Please see our response to your comments one-by-one below. > Questions #1 The anomalies do not conform to the prior knowledge Please also refer to the response to Review 7mdh Question #1 for detailed clarification. > Questions #2 Analysis of the results of the Photo and the results at the overall data level Thank you very much for the comment and suggestion. GGAD yields the best AUROC on the Photo while yielding the second-best in AUPRC, underperforming OCGNN. Having the best AUROC while less effective AUPRC indicates that GGAD can detect some anomalies very accurately in Photo, but it is less effective than OCGNN to get a bit more anomalies rank in the top of normal nodes in terms of their anomaly score. We will add the above discussion into the final paper. > Questions #3 How the model overcome mislabeled positive samples To consider this issue, we introduce a certain ratio of anomaly contamination into the training normal node set to simulate the normal nodes that are mislabeled in our experiments. The results of the models under different ratios of contamination in Fig. 5. and App. C.3 shows that with increasing anomaly contamination, the performance of all methods decreases. Despite the decreased performance, our method GGAD consistently maintains the best performance under different contamination rates, showing good robustness w.r.t. the contamination/noise. The main reason is that, unlike most unsupervised methods, GGAD not only learns normal patterns through normal models but also learns abnormalities by generating anomalies based on two priors, which reduces the dependency on the quality of normal data. > Questions #4 The connection between the outliers and the normal node is not well-illustrated In the initialization step, we first sample some normal nodes and generate the outliers based on the representations of the neighbors of each target normal node, so the generated outlier nodes share similar neighborhood information as the target normal nodes. We will add more detailed explanations for Fig.2 in the final paper to clarify this point.	Summary: This paper introduces a novel generative-based GAD approach, named GGAD, tailored for the semi-supervised scenario. Unlike existing GAD frameworks, the authors highlight the feasibility and importance of a semi-supervised setting where labels for normal nodes are relatively easy to obtain during training, but labeled abnormal nodes are very limited. In this context, the paper proposes generating pseudo-anomaly nodes to serve as substitutes for real anomaly nodes in training, thus aiding in anomaly detection. These pseudo-anomalies are generated through two unique loss-guidance mechanisms. Experimental results demonstrate the effectiveness of GGAD. However, the description of the semi-supervised setting in this paper lacks clarity and unconvincing. Additionally, there is minimal differentiation between the proposed method and existing works that generate pseudo-anomaly samples for data augmentation. I think this paper's novelty is limited. I still think that doing unsupervised GAD is more necessary, and if the authors can prove that the pseudo-outlier proposed by GGAD can benefit unsupervised GAD as a general module, I can up my score. Strengths: 1.The complete experiment shows the effectiveness of the method and the necessity of each component. 2.Some visual illustrations help the reader understand, although the shapes of the images seem to be compressed. Weaknesses: 1. I am still confused about the motivation for performing semi-supervised GAD. Why do most methods emphasize unsupervised scenarios? The cost of labeling normal nodes seems too expensive, as the authors themselves state on lines 268 to 269, yet they assert again on line 31 that labels for normal nodes are easy to obtain.This inconsistency hinders a clear understanding of the necessity and practical applications of semi-supervised GAD, which significantly undermines the motivation for this work. 2. While the first loss function proposed by the authors appears intuitively valid, the second loss function aims to generate outliers similar to normal nodes. In my opinion, optimizing these two losses together is unreasonable because they conflict with each other. It seems that they should correspond to different outlier generation processes 3. The paper validates the improvement of unsupervised GAD using labeled normal nodes and claims that GGAD remains superior. I think the authors ignore the fact that unsupervised methods do not obtain this outlier like GGAD and this comparison is not reasonable. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. why semi-supervised GAD is more important than unsupervised GAD, How do you overcome the labeling cost? 2. If unsupervised GAD methods use outliers in GGAD, is it beneficial for them? 3. why Eq.5 need Gaussian noise? 4.In addition to the outlier generation methods mentioned on lines 376-396 (they seem overly simplistic), are there more advanced methods for generating outliers similar to GGAD? How does GGAD compare to them? Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: No limitation need to discuss Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for the constructive suggestions. We are grateful for the positive comments on our readability and empirical justification. Please see our response to your comments one by one below. > Weakness in Summary Minimal differentiation with existing generation pseudo-anomaly samples (Benefits to unsupervised GAD). Our work is the first method that leverages the priors related to graph anomalies into the outlier node generation, which enables the generation of outlier nodes considering both graph structure and feature representations of real abnormal nodes, which is not viable to any existing outlier/pseudo-anomaly generation methods (We will show how GGAD can benefit unsupervised GAD as an outlier generation module in our reply to Question #2 below). > Weaknesses #1 and Questions #1 The motivation of Semi-supervised GAD and labeling cost Thank you very much for the comment. Most of the methods are based on unsupervised scenarios since it requires no labeling cost. However, they are too restricted in the real applications, because labels for a small set of normal nodes are easy to obtain. This is mainly due to the fact that the number of normal nodes typically overwhelmingly dominates the full graph. Thus, for example, one can randomly sample some nodes from the graph as the normal nodes, without any human manual labeling (the same labeling cost as unsupervised GAD). The quality of this 'random' normal labeling is high considering the scarcity of anomalies. Even involving human checking, such a small node set does not require much effort. The randomly labeled normal nodes may include a very small number of abnormal samples in some cases. This is also why we perform extensive experiments on such cases in Fig. 6, where the detectors are evaluated on anomaly-contaminated normal data. In line 31, we stated that ``the normal nodes are easier to obtain due to their overwhelming presence in the graph``. We will rephrase that ``a small part of normal nodes is easier to obtain`` to avoid misunderstanding. In lines 268-269 we meant that human checking of large-scale normal nodes can be costly, contrasting to that for a small part of the normal nodes. > Weaknesses #2 Optimize these two losses seems to conflict and they should correspond to different outliers Please refer to our reply to Global Response to Share Concern #2 in the overall author Rebuttal section. > Weaknesses #3 and Questions #2 Ignore the fact that unsupervised GAD methods do not obtain the outlier like GGAD and if GGAD can benefit existing unsupervised methods as a general module Incorporating the outlier generation into existing unsupervised methods would lead to fairer empirical comparison. To allow the unsupervised methods to exploit the generated outliers, we first utilize GGAD to generate outlier nodes by training on randomly sampled nodes from a graph (which can be roughly treated as all normal nodes due to anomaly scarcity) and then remove possible abnormal nodes from the graph dataset by filtering out Top-K most similar nodes to the generated outlier nodes. By removing these suspicious abnormal nodes, the unsupervised method is expected to train on the cleaner graph (i.e., with less anomaly contamination). This approach to improve unsupervised GAD methods is referred to as GGAD-enabled unsupervised GAD. We evaluate their effectiveness on three large-scale datasets. Please refer to global response for the results in Table A1, where #Anomalies/#Top-K Node in the table respectively represents the number of real abnormal nodes we successfully filter out and the number of nodes we choose to filter out (i.e., K). For example, we use the outlier nodes generated by GGAD to filter out 500 nodes from the Amazon dataset, of which there are 387 real abnormal nodes. This helps largely reduce the anomaly contamination rate in the graph. The results show that this approach can significantly improve the performance of three different representative unsupervised GAD methods, including DOMINATE, OCGNN, and AEGIS. Note that although the GGAD-enabled unsupervised methods achieve better performance, their performance still largely underperforms GGAD, which provides stronger evidence for the effectiveness of GGAD. > Qestions #3 The role of Gaussian noise in Eq.5 This simple perturbation can help maintain the affinity separability while enforcing the egocentric closeness constraint. > Questions #4 Compare with more advanced methods for generating outliers similar to GGAD Apart from the outlier generation in the ablation study, we further employ two advanced generation approaches, VAE and GAN. In VAE, we generate the outlier representations by reconstructing the raw attributes of selected nodes where our two anomaly prior-based constraints are applied to the generation. For GAN. we generate the embedding from the noise and add an adversarial function to discriminate whether the generated node is fake or real, with our two prior constraints applied in the generation as well. As shown in Tab.1 in uploaded pdf. These two advanced generation approaches can work well on some datasets, which indicates two priors help them learn relevant outlier representations. But both of them are still much lower than GGAD, showcasing that the outlier generation approach in GGAD can leverage the two proposed priors to generate better outlier nodes. --- Rebuttal Comment 1.1: Comment: Dear Authors: On the semi-supervised question, I assume that you only use a subset of nodes in the original dataset as normal samples and use the generated abnormal samples (similar to data augmentation?) to train a classifier, if that's the case, I think it's reasonable and would be willing to boost my score to 4. However, I think the process of generating abnormal nodes seems too simple, I can understand the intention of Eq(4), but how to use abnormal nodes to calculate Eq(3), how do you determine what their neighbors have? I think that's critical. Eq (5) is more like a regularization term and I think its contribution is small. --- Reply to Comment 1.1.1: Title: Response to Reviewer JEU8 Comment: We greatly appreciate your further comments and it's great to know that our response has helped address your questions. Please find our point-by-point replies as follows. (0) Setting. Yes, the studied setting assumes the availability of a small set of labeled normal nodes during training. To train a discriminative one-class classifier, our method GGAD utilizes these normal nodes to generate pseudo abnormal nodes that assimilate real anomaly nodes in both graph structure and feature representations. (1) Similar to data augmentation? Current data augmentation methods for GAD like DAGAD and DIFAD [Ref1-2] and imbalanced graph learning like GraphSMOTE and GraphMixup [Ref3-4] require both labeled abnormal nodes and normal nodes, making them not applicable to the studied semi-supervised setting where only the labels of partial normal nodes are available during training. Unlike these graph data augmentation methods, our GGAD is a generative method that generates pseudo abnormal nodes by leveraging the two abnormality-related priors using only a small set of normal nodes. Furthermore, from the ablation study results in Table 3 in the paper, commonly used graph data augmentation methods like the random sampling or mixing up normal nodes with noise cannot work. Besides, the newly added results of existing popular generative methods in Table 1 in the upload pdf further demonstrate the unique effectiveness of the proposed generative method in our GGAD. (2) The calculation of Eq(3) for abnormal nodes ? As mentioned in lines 202-204, and shown in Fig. 1 left in the paper, we sample some normal nodes as the target location for outlier node generation, in which the outlier nodes share the same neighbors with the sampled target normal nodes. The affinity calculation of the generated abnormal nodes is based on the neighbors of these target normal nodes. (3) Eq. (5) is more like a regularization term and I think its contribution is small. Eq. (5) is designed to incorporate our egocentric closeness-based abnormality prior, through which we aim to pull the representations of generated outliers close to the node representations in its egonet in the feature representation space, as shown in Fig. 3(c). Without this loss, the generated abnormal nodes only meet one abnormality prior, which do not provide sufficient discriminative information for the one-class classifier. Thus, the contribution of this loss is significant: it is not only reflected in its own design but also in the collaboration with the other prior implemented via Eq. (4). This collaborative effect helps generate the pseudo abnormal nodes which can serve as effective negative samples for training a tight decision boundary of the one-class classifier. As the very first work to explore the semi-supervised GAD setting, we introduce a simple yet effective way to generate pseudo abnormal nodes for training an accurate one-class classifier for GAD, offering a novel principled framework for semi-supervised GAD. Further, our method GGAD can also perform well in anomaly-contaminated training data; as also pointed out by you and justified by our newly added results, the generated abnormal nodes can also help improve unsupervised GAD methods in the popular unsupervised setting. All these lead to a piece of work that has significant contributions to GAD in the newly introduced semi-supervised setting and the widely-explored unsupervised setting. In terms of methodology, all modules in our method are novel for GAD, and as far as we know, GGAD is the very first method that can generate pseudo abnormal nodes assimilating real anomaly nodes in both graph structure and feature representations. It is true that GGAD is simple, but as per Occam's razor, simpler models are preferred over more complex ones. Thus, we argue that our method is technically solid and expected to have moderate-to-high impact in the GAD community. We're more than happy to engage in more discussion with you to address any further concerns you may have. Thank you very much for helping enhance our paper! References: - [Ref1] DAGAD: Data Augmentation for Graph Anomaly Detection， ICDM2022 - [Ref2] NEW RECIPES FOR GRAPH ANOMALY DETECTION: FORWARD DIFFUSION DYNAMICS AND GRAPH GENERATION, 2024 - [Ref3] G-Mixup: Graph Data Augmentation for Graph Classification ICML2022 - [Ref4] GraphSMOTE: Imbalanced Node Classification on Graphs with Graph Neural Networks WSDM2021 --- Rebuttal 2: Title: Response to Reviewer JEU8 Comment: Thank you for raising the score. Please find our point-by-point replies as follows. > Questions #1 Ignore the fact that many anomaly detection methods are based on the structure of the graph, which may affect the scope of the method (it cannot be used on graphs without attributes) As mentioned above, our GGAD generates outlier nodes that assimilate the real anomaly nodes in both graph structure and feature representations, where the structure information of the graph has been fully considered in our methodology design (i.e., through Eqs. (2) and (3) in the asymmetric local structural affinity prior). As emphasized in many GAD studies, anomalies in the graph are primarily reflected in the relationships of a node to its neighboring nodes immediately connected to themselves [Ref1-2] and many other methods reviewed in [Ref3]. That's why we consider generating the outlier nodes based on the egonet of target normal nodes. Currently, to our best knowledge, most GAD methods are focused on attributed graph datasets (please see the survey and benchmark papers in [Ref3-5] for more details). As for the graph without attributes, the methods for attributed graphs can be applied by augmenting the graph with node attributes based on, e.g., one-hot encoding based on neighborhood information or feature construction using graph structure information (see [Ref3]). Thus, this should not be seen as a limitation of our method and numerous existing GAD methods. > Questions #2 The implementation of Gaussian noise This simple perturbation can help maintain the affinity separability while enforcing the egocentric closeness constraint. It is a hyperparameter in GGAD and the default value was presented in the implementation. Gaussian noise-based perturbation is commonly used in existing feature interpolation techniques, including those in GAD methods [Ref6-7], and it is used as a way to diversify the generated outlier nodes only, not having catastrophic impacts on the performance of GGAD. This is justified by the newly added results in Table A1 below where we change the mean and variance of the Gaussian noise and replace the Gaussian noise with uniform noise. The results show that, regardless of the distribution of the noise, GGAD remains very effective, demonstrating similar superiority over the competing methods in Table 1 in the paper. ``` Table A1. The performance of GGAD under different scales of Gaussian noise and uniform noise (AUPRC/AUROC). ``` \|Data\| Amazon\|Elliptic\|Photo\| \|:---- \|:----: \|:----: \|:----: \| \| mean=0, std = 0 \| 0.7343 / 0.9192 \| 0.2107 / 0.7060 \| 0.1325/0.6432 \| \| mean=0.005, std = 0.001 \| 0.7475 / 0.9233 \| 0.2240 / 0.7110 \| 0.1401/0.6444 \| \| mean=0.01, std = 0.005 \| 0.7834 / 0.9324 \| 0.2425 / 0.7290 \| 0.1442/0.6476 \| \|uniform noise(a=0, b=0.01) \| 0.7434 / 0.9142 \| 0.2173 /0.7163 \| 0.1407/0.6526 \| Overall, in our paper and the rebuttal here, we have examined the feasibility of a very wide range of methods that are related to our method GGAD from diverse aspects, e.g., different ways of exploiting graph structure information, generating outlier nodes, implementing GGAD with various alternative methods, and utilizing the generated outlier nodes in unsupervised/semi-supervise settings, etc. Thus, we would greatly appreciate if you could provide other unexplored directions for us to further evaluate the effectiveness of our method GGAD. We would kindly request your reconsideration of your current rating of our paper if otherwise. Thank you very much again! References: - [Ref1] Addressing Heterophily in Graph Anomaly Detection: A Perspective of Graph Spectrum, WWW2023 - [Ref2] Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection, NIPS2023 - [Ref3] A comprehensive survey on graph anomaly detection with deep learning, TKDE 2021 - [Ref4] GADBench: Revisiting and Benchmarking Supervised Graph Anomaly Detection, NIPS2023 - [Ref5] BOND: Benchmarking Unsupervised Outlier Node Detection on Static Attributed Graphs NIPS2022 - [Ref6] Perturbation learning based anomaly detection, NIPS2022 - [Ref7] DAGAD: Data Augmentation for Graph Anomaly Detection， ICDM2022 - [Ref8] Consistency Training with Learnable Data Augmentation for Graph Anomaly Detection with Limited Supervision, ICLR2024 --- Rebuttal Comment 2.1: Comment: Dear Authors: Thanks for your reply! As you said, for graphs without attributes, you can add hand-crafted features, which is normal and not a shortcoming of GGAD. My initial concern was that it seemed unreasonable to let the generated abnormal nodes share the same neighbors as the normal nodes. I read the paper again, and I thought that the loss in Eq 4 might solve this problem. I can accept the author's explanation. But with Gaussian noise, If you say it's "a way to diversify the generated outlier nodes only, not having catastrophic impacts on the performance of GGAD. "Then I still don't think it's the most necessary. To sum up, the loss in Eq 4 is good and dispels my confusion, but the loss in Eq 5 weakens the novelty, so I'm sorry that I can't improve the score. --- Reply to Comment 2.1.1: Comment: Dear Reviewer JEU8, Thanks a lot for the prompt reply. We apologize that we don't understand why the use of the loss in Eq. 5 weakens the novelty of our method, given the fact that you find good novelty in Eq. 4. As demonstrated by our ablation study results in Table 2 in the paper, the adding of the loss in Eq. 5 to our method leads to very significant performance improvement across all the datasets in both AUROC and AUPRC. Moreover, as shown in the newly added results in Table A1 above, the loss in Eq. 5 works well regardless of different prior distribution used to specify the noise. We would really appreciate if you could kindly advise why the additional major contribution made by the loss in Eq. 5 is considered as a negative part of our model design. Thank you! --- Rebuttal 3: Comment: Dear Reviewer JEU8, We're very pleased that our clarification is helpful, and thank you for increasing the rating to an acceptance score. We will add the intuition of "hard" outlier nodes and its difference to "trivial" outlier nodes in Sec. 3.4 to clarify why the loss in Eq. 5 is important in our method. The discussions have been very helpful for enhancing our paper. Many thanks again for your time and effort on our paper. We're happy to take any further questions you might have.	Summary: This paper explores the problem of semi-supervised graph anomaly detection (GAD), where some nodes are known to be normal, in contrast to the typical unsupervised setting with no labeled data. The authors show that even a small percentage of labeled normal nodes can improve the performance of existing unsupervised GAD methods when adapted to the semi-supervised scenario. The paper proposes a novel Generative GAD approach (GGAD) to better exploit normal nodes by generating pseudo anomaly nodes, called 'outlier nodes', to provide effective negative samples for training a one-class classifier. GGAD generates these outlier nodes using priors about anomaly nodes, such as asymmetric local affinity and egocentric closeness, to mimic anomalies in structure and features. Experiments on six real-world GAD datasets show that GGAD outperforms state-of-the-art methods in both unsupervised and semi-supervised settings. Strengths: + This paper studies a new problem of semi-supervised GAD that has not been widely studied. + The proposed method is simple and effective from the empirical perspective. + The experiments are extensive including effectiveness and efficiency analyses and the method has been tested on real-world large-scale graphs to verify the scalability. Weaknesses: - The two priors that are used to generate outlier nodes are heuristic or based on empirical evidence. There is no theoretical analysis provided to better guarantee the effectiveness of the proposed method. - It will be more interesting and helpful to show the generated outlier nodes can capture the characteristics of anomalous nodes in addition to comparing their representations. - The experimental settings of anomaly contamination are not very clear: how the contamination is introduced? - Overall experimental settings. What hardware has been used in the experiments, e.g., memory, and why are the experiments conducted on CPUs? Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Theoretical analysis of the proposed method, especially these two priors. 2. Experimental settings including hardware and anomaly contamination. 3. Analysis of the generated outlier nodes. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors have adequately addressed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for the constructive comments. We are grateful for the positive comments on our studied problem, technical contribution, and empirical justification. Please see our detailed response below > Weaknesses #1 There is no theoretical analysis to guarantee the effectiveness of the proposed method We encapsulate two important anomaly priors to generate the outliers that are similar in the local structure and feature representation as real abnormal nodes. This provides a principled framework for generative GAD. To further verify the intuition of the priors and our proposed method, we have added empirical evidence that justifies the two priors in more real-world GAD datasets in the uploaded pdf. Please refer to our reply to Global Response to Share Concern #1 in the overall author rebuttal section above for details. Overall, our method, as the first piece of work explicitly designed for the semi-supervised GAD problem, presents solid findings and interesting insights into the problem, laying a good foundation for future work on theoretical analysis and more advanced methods in this research line. > Weaknesses #2 and Q3 More analysis on the generated outliers like shot the generated outliers can capture the characteristic of anomalous in addition to comparing their representations Thank you very much for the suggestion. Please see Fig.3 in the paper and Fig. 1 in the uploaded pdf for the visualization based on the local structure information of the generated outlier nodes. We further employ the Maximum Mean Discrepancy (MMD) distance to measure the distance between the generated outliers and the real abnormal nodes (and the normal data as well) to illustrate more in-depth characteristics of the generated outlier nodes. As shown in Table A1 below, it is clear that the distribution of the generated outliers have much smaller MMD distance to the real abnormal nodes than the normal nodes, indicating the good alignment of the distribution of the generated outliers with the real abnormal nodes. ``` Table A1. Analysis of the generated outlier nodes using MMD distance ``` \|Data\|Amazon \|T-Finance \|Elliptic \| Photo \| Reddit \| \|:---- \|:----: \|:----: \| :----: \| :----: \|:----: \| \|with Abnormal Node\| 0.1980 \| 0.0784 \| 0.1094 \| 0.3703 \| 0.3409 \| \|with Normal Node \| 0.2318 \| 0.1040 \| 0.1304 \| 0.3880 \| 0.3605 \| We will add this MMD distance-based outliers analysis into the final paper. > Weaknesses #3 and Questions #2 Clarification on the setting of contamination Thanks for pointing out the issue. When studying the semi-supervised GAD, it's important to consider that some normal nodes may be mislabeled or affected by some noise interference. To introduce a certain ratio of anomaly contamination, we randomly sample $V_l$ nodes from the real abnormal nodes set in the dataset and add them as the contaminated nodes into the training data. The rest of the abnormal nodes are used as part of the test dataset. We will add these details of the contamination setting in the final paper. >Weaknesses #4 and Questions #2 Overall experimental settings. Provide more information about the hardware and why the experiments are conducted on the CPU. The existing baseline methods require different GPU memory and environments depending on their methodology design. To conduct a unified comparison of operational efficiency, we chose an AMD EPYC 7443P 24-core CPU with 125G memory as the running platform. We also provided a computational efficiency analysis in Appendix D. --- Rebuttal Comment 1.1: Title: Kindly Request for Reviewer's Feedback Comment: Dear Reviewer yDqR, Since the End of author/reviewer discussions is coming in ONE day, may we know if our response addresses your main concerns? If so, we kindly ask for your reconsideration of the score. Should you have any further advice on the paper and/or our rebuttal, please let us know and we will be more than happy to engage in more discussion and paper improvements. Thank you so much for devoting time to improving our paper! --- Rebuttal Comment 1.2: Title: Thanks for the rebuttal Comment: I appreciate the efforts of the authors during the rebuttal phase. These responses addressed most of my concerns. But I still have some concerns for W4 and Q2: to make a fair comparison, why CPU is selected? Because of the large memory requirement to load a large graph? --- Reply to Comment 1.2.1: Title: Response to Reviewer yDqR Comment: We're very pleased to know that our response has addressed most of your concerns. We really appreciate your further comments. Please find our response to your follow-up question as follows. >Follow-up question with Question #1 To make a fair comparison, why CPU is selected? Because of the large memory requirement to load a large graph? Yes, your understanding is correct. A number of baselines like DOMINANT, AnomalyDAE, and TAM require large memory to handle large-scale graph datasets. However, we lack GPUs with a sufficiently large memory size to perform experiments on these large datasets using GPUs. Thus, in order to obtain the runtime results for all methods across all data sets using the same computing environment, we obtained the runtime results based on a consistent CPU-based setting. We hope the above reply helps address this follow-up question. We will clarify this point in our final version. We're more than happy to engage in more discussion with you to address any further concerns you may have. Thank you very much for helping enhance our paper again!	Rebuttal 1: Rebuttal: Dear All Reviewers, Thank you very much for the time and effort in reviewing our paper, and for the constructive and positive comments. Our rebuttal consists of two parts: Global Response where we address shared concerns from two or more reviewers and Individual Response where we provide a detailed one-by-one response to address your questions/concerns individually. > ### Global Response to Shared Concern #1 More empirical evidence to verify the two priors and the second prior has not been empirically verified. The asymmetric local node affinity prior, ``the affinity between normal node is stronger than that between normal and abnormal nodes``, has been revealed in multiple recent studies on a range of datasets [Ref1-3]. To further verify this prior, we provide more affinity visualization results on other GAD datasets including Amazon, Reddit, Elliptic, and Photo, as shown in Fig.1 in the uploaded pdf. The results show that the normal nodes have a much stronger affinity to its neighboring normal node than the abnormal nodes. For the egocentric closeness prior, ``the feature representations of outlier nodes should be closed to the normal nodes that share similar local structure as the outlier nodes``, we verify this prior by analyzing the similarity between normal and abnormal nodes based on the raw node attributes on the other four datasets in Fig.2 in the uploaded pdf. The results show that the real abnormal nodes can exhibit high similarity to the normal nodes in terms of local affinity in the raw attribute space. The main reason is that some abnormalities are weak or the existence of adversarial camouflage that disguises abnormal nodes to have similar attributes to the local community. This is the key intuition behind the second prior. > ### Global Response to Shared Concern #2 The direction of two constraints seem to be contradictory to each other. $L_{ala}$ and $L_{ec}$ are collaborative constraints rather then conflicting ones. $L_{ala}$ is designed to make the generated outliers have asymmetric affinity separability from normal nodes from the graph structure perspective, while $L_{ec}$ is devised to pull the representation of generated outliers closed to the node representations in its egonet in the feature representation space. If we solely apply $L_{ala}$， it may generate some trivial outliers that are far from the normal node, see Fig.3(a) in the paper. Although these trivial outliers share some local affinity property as the abnormal, they have adverse effects on training a compact, discriminative one-class classifier in the feature representation space. Thus, we further introduce the egocentric closeness prior-based loss $L_{ec}$. It enables the generated nodes to be close to the distribution of normal nodes. This joint force will result in outlier representations that are at the fringe of the normal node representations while structurally separable, preventing the generation of trivial outliers that are far away from the normal nodes. This collaborative effect can also be observed in Fig.3 (a)-(c) in the paper. To further demonstrate the collaboration between these two prior-based losses, we visualize the optimization of loss during the training in Fig.3 in the uploaded pdf, where 'ala' and 'ec' represent the two priors losses and the 'total' represents the sum of these two priors and BCE loss. From the results, we can see the two prior losses and the total loss are continuously decreasing and converging finally, further indicating that these optimizations are collaborative, resulting in an effective one-class discriminator. Note due to the space limitations, all the visualization figures of other datasets will be provided in the final paper. As for Individual Response, we have provided a detailed one-by-one response to answer/address your questions/concerns after your individual review. We very much hope our response has clarified the confusion, and addressed the concerns. We're more than happy to take any further questions if otherwise. Please kindly advise! Due to the space limitation, here we put a table for addressing Review JEU8 Question #2 ``` Table A1. Comparison with the unsupervised GAD methods that use our GGAD-generated outliers (AUPRC/AUROC). ``` \|Data \|Amazon \|T-Finance \|Elliptic \| \|:---- \|:----: \|:----: \|:----: \| \| #Anomalies/#Top-K Nodes \| 387/500 \| 351/1000 \| 1448/2000 \| \| DOMINATE \| 0.1315/0.7025 \| 0.0536/0.6087 \| 0.0454/0.2960 \| \| GGAD-enabled DOMINATE\| 0.3462/0.8186 \| 0.0585/0.6275 \| 0.0613/0.2986 \| \| OCGNN \| 0.1352/0.7165 \| 0.0392/0.4732 \| 0.0616/0.2581 \| \| GGAD-enabled OCGNN \| 0.3950/0.8692 \| 0.0480/0.5931 \| 0.0607/0.2638 \| \| AEGIS \| 0.1200/0.6059 \| 0.0622/0.6496 \| 0.0827/0.4553 \| \| GGAD-enabled AEGIS \| 0.3833/0.8395 \| 0.0784/0.7024 \| 0.0910/0.5036 \| \| GGAD \| 0.7769/0.9431 \| 0.1734/0.8108 \| 0.2484/0.7225 \| References: - [Ref1] Addressing Heterophily in Graph Anomaly Detection: A Perspective of Graph Spectrum, WWW2023 - [Ref2] Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection, NIPS2023 - [Ref3] Graph anomaly detection with bi-level optimization, WWW2024 Pdf: /pdf/6d855e6e889aff8cb2b27cc1d64b403da2cd55d2.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper studies an under-explored graph anomaly detection problem where the detection models have access to a set of labeled normal nodes. To tackle this problem, it introduces a generative approach namely GGAD that generates pseudo anomaly nodes, called outlier nodes, to support the training of a discriminative one-class classifier. The key idea underlying this approach is to generate the outlier nodes in a way that can well simulate real anomaly nodes in both graph structure and feature representation perspectives. To achieve this, GGAD defines and incorporates two priors, including asymmetric local affinity and egocentric closeness, into its optimization objectives, with the former prior focusing on the alignment on the graph structure aspect and the latter on the feature representation aspect. The method is evaluated on six large real-world datasets and shows impressive detection performance compared to existing state-of-the-art methods. Strengths: - The paper is generally well-written and easy-to-follow. - The problem setting is practical since labeled normal samples are easy to obtain in many real-world applications. Compared to the commonly studied unsupervised setting, this semi-supervised setting often results in better detection performance. - The proposed method GGAD is novel. There have been many generative anomaly detection methods, but as far as I know, they are unable to consider the graph structure and the neighboring nodes’ representations. By introducing the two new priors, GGAD addresses this issue well. Fig.1 and Fig. 3 help demonstrate this effect. - The method is compared with a range of unsupervised and semi-supervised methods on 6 real-world datasets with diverse genuine anomalies, and gains largely improved detection performance over these competing methods. - The ablation study is plausible and justifies the contribution of each proposed prior. Weaknesses: - The outlier node generation in GGAD may cause non-trivial computational overhead. - Despite better performance than the competing methods, GGAD gains an AUC of only around 0.6 on some datasets, such as DGraph and Reddit. - In Fig. 4 (b), GGAD shows a fast AUPRC growth with increasing training size, but the other methods have a flat performance trend. What would be the reason behind? Technical Quality: 4 Clarity: 3 Questions for Authors: See the weakness Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for the constructive comments. We are grateful for the positive comments on our studied problem, technical contribution, and empirical justification. Please see our detailed response below > Weaknesses #1 The generation may cause non-trivial computational We agree that GGAD may cause computational overhead due to the outlier node generation. However, the overhead is small. This is justified by the time complexity analysis and running time statistical results in Appendix D, where GGAD runs much faster than many competing methods and it is comparable efficiency to the remaining methods. > Weaknesses #2 Low AUC on Reddit and DGraph Thank you very much for the comment. Reddit and DGraph are two challenging datasets in GAD. DGraph is a very large-scale graph that includes more than millions of nodes where the anomalies only account for 1.3 \%. The fully supervised methods only yield an average of 0.02/0.67 in AUPRC and AUROC on this dataset [Ref1-2]. Similarly, Photo is a user-subreddit graph network on which fully supervised methods only yield an average of 0.06/0.65 in AUPRC and AUROC [Ref1-2]. Although our GGAD achieved low AUROC/AUPRC on these two datasets, not as high as the other datasets, it still shows good improvement compared to other state-of-the-art unsupervised and semi-supervised methods. > Weaknesses #3 The reason behind the flat performance of other methods The main reason is that these competing methods are trained based on their unsupervised proxy GAD tasks, so they have limited capability to increase their discriminability with increasing the number of normal nodes. On the contrary, our GGAD utilizes partially labeled normal nodes and two important anomaly priors to generate outlier nodes as negative samples to train a discriminative one-classifier. As the number of normal nodes increases, the generated outliers become more diverse, closely aligning with the real abnormal nodes in the dataset, thereby resulting in better discriminability and thus better AUPRC. References: - [Ref1] GADBench: Revisiting and Benchmarking Supervised Graph Anomaly Detection, NIPS2023 - [Ref2] BOND: Benchmarking Unsupervised Outlier Node Detection on Static Attributed Graphs, NIPS2022 --- Rebuttal 2: Title: The reply has addressed my questions. Comment: The reply has addressed my questions. --- Rebuttal Comment 2.1: Comment: We're very pleased that our response has addressed your questions. Thank you very much for the positive comments and appreciation of our work.	null	null	null	null	null	null
RashomonGB: Analyzing the Rashomon Effect and Mitigating Predictive Multiplicity in Gradient Boosting	Accept (poster)	Summary: This paper proposes a method (RashomonGB ) to estimate the Rashomon sets/predictive multiplicity of gradient boosting models. It estimates multiple ($m$) models at each stage (effectively performing a local exploration) and then combine all such models in the end to construct $m^T$ models for Rashomon set computation, where $T$ is the number of iterations of the boosting. On several datasets the paper shows that RashomonGB performs better than re-training with $m$ seeds, in that at the fix $\epsilon$ (loss difference) level, RashomonGB tends to show more predictive multiplicity. Strengths: Predictive multiplicity is an important topic. The paper is generally clear and well-written. The proposed method is a sensible first method for boosting algorithms, which was previously underexplored. I think the proposed method is likely adopted by people who care about this problem as it's intuitive and easy to implement. Weaknesses: 1. The current exploration strategy is fast to compute, but I'm not sure if this follows the motivation of Rashomon set very well. While the authors mention one example on the Contraception dataset where re-training underestimates the predictive multiplicity, in general RashomonGB might create models that are more correlated than normal (because the "backbone" is the same GB model), thus underestimating the predictive multiplicity. Right now, the conclusion shows otherwise probably because the number of re-training is too small. 2. Regarding the experiment, if I read this correctly, currently we use more compute for RashomonGB as well (by combining different weak models), so it is also not quite a fair comparison in my opinion. I would be very interested to see some estimate of how much compute RashomonGB saves against re-training, by running more re-training and see when are the metrics in Fig3 in the two methods become comparable. minor: one "RashomonGB" in L290 should be "re-training". Technical Quality: 2 Clarity: 3 Questions for Authors: 1. What's $\epsilon_{t_1}$ (and $\epsilon_{t_2}$) in L243-L244? Isn't epsilon a quantity set by the user? 2. In L282-283, do we construct 10 final models and 1024 for re-training and RashomonGB, respectively? If only 2 out of $m$ models are used why train $m$ of them (L282-283) for RashomonGB? 3. Related to the above, I originally thought there is a model "filtering" step in each iteration $t$, and wonder how $\epsilon_t$ is set for each iteration. However, from L282-283 it seems like we just randomly pick a few models and brute-force combine all weak models for the final Rashomon set exploration. Could the authors clarify? 4. Are Fig 4 measured on the test set? If so, then it's not clear how useful this is as we cannot choose models basing on test performance - did the authors try picking models on the frontier basing on the validation set and then plot this on the test set? Right now, due to the sheer number of final models generated by RashomonGB, it's unclear if the models with better trade-off are just lucky. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's constructive feedback and encouragement. Below, we systematically address each weakness and question raised in the review. For Weakness 1, the estimates in Figure 3, derived from both re-training and RashomonGB, utilized the same training cost. Specifically, each method required training T\m = 10\10 = 100 weak learners (except CIFAR-10; cf. Line 282). The re-training approach, using different random seeds, explores the Rashomon set in a "global" fashion, potentially yielding a more diverse set of models within this space. Conversely, RashomonGB explores the set in a more "local" manner, focusing on models that incorporate the same weak learners. Despite this, under identical training costs, RashomonGB can explore exponentially more models than the re-training strategy. It's important to note that for complex hypothesis spaces, like Gradient Boosting (GB) used here, any method, including re-training, will likely underestimate predictive multiplicity due to the sheer computational impracticality of fully exploring the Rashomon set. This introduces an inevitable trade-off between the efficiency (training cost) and the effectiveness (degree of underestimation) in estimating predictive multiplicity. In Figure 3, our goal was to highlight the efficiency and effectiveness of RashomonGB under equal training costs. However, as the number of re-training increases—thereby increasing the training cost—the re-training method may surpass RashomonGB in effectively estimating the diversity within the Rashomon set. Regarding Weakness 2, although the computational costs for re-training and RashomonGB are equivalent, RashomonGB significantly reduces the time required to obtain a model. This is because RashomonGB explores an exponential search space, making it much more efficient. This efficiency is compellingly demonstrated in the comparative analysis of computational times shown in Table E.4 of Appendix E.2, where RashomonGB's model generation speed significantly outpaces that of re-training. For instance, for the ACSIncome dataset, both methods recorded a training time of 54.67 seconds. However, while the inference time per model for re-training is 0.4 seconds, it is only 0.02 seconds for RashomonGB. This means that, under the same training cost, RashomonGB is 20 times more efficient in terms of generating models from the Rashomon set compared to the re-training strategy. We will address this comment by clearly directing readers in the revised main text to the additional experiments detailed in the Appendix. This will help ensure that the relevant information is easily accessible and comprehensible. For Question 1, indeed, $\epsilon$ is a parameter configured by the user. As stated in Lines 243-244, maintaining the same $\rho$—defined as the probability in Proposition 1—results in more iterations increasing the conditional mutual information, which in turn leads to a larger $\epsilon$. This insight serves as a practical guideline advising users against choosing smaller values of $\epsilon$ in subsequent boosting iterations. This effect is further illustrated in the Ablation study detailed in Appendix E.3. Figure E.8 demonstrates this by fixing $\epsilon$ for each iteration, re-training with different random seeds, and computing the percentage of models (i.e., $\rho$ in Proposition 1) in the Rashomon set for each iteration. It is observable that, with the same percentage (i.e., the same $1-\rho$), $\epsilon$ increases. This observation reinforces the results suggested by Proposition 3, validating the relationship between $\epsilon$ and $\rho$ under consistent conditions. For Question 2, indeed, we implemented the re-training strategy using 10 different random seeds for the Gradient Boosting with 10 iterations. This strategy involved training 100 weak learners (10 per iteration). Ideally, RashomonGB can generate up to $10^{10}$ final models. However, as noted in footnote 5, even selecting 3 models per iteration for RashomonGB would yield over 59,000 final models, which exceeds our storage capabilities. Consequently, we chose to select 2 models per iteration for RashomonGB. Comparatively, if we train only 2 models in each iteration, the re-training strategy results in just 2 final models, whereas RashomonGB still generates 1024 models. This underscores the superior efficiency of RashomonGB in generating a higher number of models under the same training cost, thereby providing a broader exploration of the model space. For Question 3, indeed, there is a filtering process in place that screens out models with an MSE loss greater than 0.1 (i.e., $\epsilon_t = 0.1$) and retains models with an MSE loss smaller than 0.01 until $m=10$ models are collected at each iteration. For Question 4, the models situated at the accuracy-fairness trade-off frontier were selected using a hold-out validation set post-training, and the results depicted in Figure 4 were evaluated using the test set. Compared to the re-training strategy, the lower cost of obtaining models through RashomonGB enhances the likelihood of identifying a model with a more optimal operation point. The advantage of RashomonGB becomes even more pronounced when dealing with larger datasets. In such scenarios, re-training and re-training-based fairness intervention algorithms, such as Reduction (and FaiRS) and Rejection, may incur significantly higher training costs. We will add the additional explanation and results, and fix the typo (e.g., "RashomonGB" in L290) in the revision. Thanks again and we would be happy to provide more clarifications and answer any follow-up questions.	Summary: This paper presents an approach that compute Rashomon set for gradient boosting algorithm where the set can be obtained through products over weak learners at each step rather than sampling them through retraining. The authors further proposed a dataset related Rashomon bound through sub-Gaussian assumption, where mutual information between hypothesis space and dataset shows the predictive multiplicity, which can further decomposed into model uncertainty and quality of data. Experiments show the proposed solution offers more models in Rashomon set than retraining given the same computation budget. Strengths: The rough idea of the proposed approach is straightforward since decomposing Rashomon set search on boosting algorithm can be a "standard" operation given the unique residual learning property of boosting algorithms. The novelty of the proposed approach is probably more from "our work is the first to explore the Rashomon effect for gradient boosting". The dataset related Rashomon set bound seems an interesting point. But it needs some justification for the key assumption of it (sub-Gaussian). Proposition 2 seems make sense given the positive relation between number of boosting iterations and Rashomon set (also for dataset size). Experiments in 4.2 seem interesting. I would love to see more experiments like it. Weaknesses: I got some difficult time to understand the introduction and abstract of this paper even I have read some literatures about Rashomon effect and predictive multiplicity. It is simply hard to read given the narrative there. Especially the second paragraph of introduction; it gets me confused and self-questioning my understanding of Rashomon effect from other works. Technical Quality: 3 Clarity: 2 Questions for Authors: Why boosting algorithms? Further justification about the dataset related Rashomon set bound? Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: No hard limitation I can see. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback! We clarify the weakness and answer the reviewer's question below. To address the weakness pointed out, it would be helpful if the reviewer could specify which parts of the second paragraph in the Introduction are unclear or difficult to understand during the author-reviewer discussion phase. To enhance clarity, we will expand on the concepts introduced in both the Abstract and the Introduction. In this paper, our objective is to explore both the positive (e.g., improved model selection as discussed in Section 4.2) and negative (e.g., predictive multiplicity as detailed in Section 4.1) impacts of the Rashomon effect using gradient boosting algorithms. These algorithms uniquely employ a sequential training procedure that focuses on learning the residuals of the data. The Rashomon effect articulates that within a given hypothesis space, numerous models can achieve similar performance levels (such as 99\% accuracy). These similarly performing models can be grouped into what is known as the Rashomon set. The positive aspect of the Rashomon effect is particularly significant in tasks associated with responsible machine learning, which often requires models to possess additional properties (such as group fairness) without significantly sacrificing performance. In essence, the pursuit of responsible machine learning is about finding models within the Rashomon set that meet these extra constraints (as exemplified by the fairness considerations in Section 4.2). The Rashomon effect thus provides assurance of finding viable solutions when optimizing for responsible machine learning goals. Conversely, the negative aspect, termed predictive multiplicity, occurs when a model selected at random from the Rashomon set leads to inconsistent decisions for some individuals (e.g., affecting 5\% of the samples as illustrated in Figure 3 of Section 4.1). This unpredictability can undermine the reliability of the machine learning process. By elaborating on these concepts, we aim to resolve any confusion and reinforce the significance of our investigation into the dual implications of the Rashomon effect within gradient boosting frameworks. In response to the Question, boosting algorithms are prevalently utilized for tabular datasets, particularly in the realm of trustworthy machine learning, such as fairness interventions detailed in Section 4.2. Notably, boosting algorithms have been shown to surpass deep learning on tabular datasets, as referenced in [35]. Despite this, existing theoretical frameworks that explore the Rashomon effect and predictive multiplicity have primarily focused on linear classifiers [55, 72], generalized additive models [15], sparse decision trees [74], and neural networks [42]. The unique sequential training procedure of boosting and its influence on the characterization of the Rashomon set remain poorly understood. This paper aims to bridge this gap, providing mathematical tools that could also extend to other sequential residual training schemes. We initially discuss our motivation for focusing on boosting algorithms in Lines 75-90 and 117-127. In the revision, we will reorganize this content to better highlight our motivation, as suggested by the reviewer. Regarding the assumption of sub-Gaussianity of the loss function, this represents a generalization beyond mere boundedness. While it is feasible to assume boundedness of the loss function—a common and practical approach readily achieved by clipping the loss, as mentioned in Lines 185-187—we opt for the sub-Gaussian assumption in Proposition 1 to allow for a broader analysis. This paper emphasizes the novelty of dataset-related bounds on the Rashomon set, where previous studies have largely concentrated on the hypothesis space. This perspective underscores our novel contribution to the understanding of the Rashomon effect and predictive multiplicity in machine learning. We will include additional explanations regarding the sub-Gaussian assumption in the revised Section 3. We welcome any follow-up questions and are happy to provide further clarification. Thank you!	Summary: The paper studies the Rashomon effect in gradient boosting, a commonly used algorithm for tabular datasets, but something that has not received enough attention in multiplicity literature. The paper provides several theoretical discussions on the size of the Rashomon set and the impact of the number of iterations on multiplicity in GBRTs. Furthermore, the paper proposes RashomonGB, a method to create an exponential number of ‘near-optimal models’ by training only a polynomial number of models. With more models in the Rashomon set, the use of RashomonGB can create several downstream benefits without any extra cost of training, shown empirically by the authors. Strengths: - Multiplicity in GBRTs, or generally any gradient-boosting algorithm, has not been studied in the literature, and so the authors provided a novel discussion, especially given the importance of these algorithms in tabular settings. - The paper provides several theoretical discussions backed by empirical support. The insights on the growing Rashomon set with iterations were quite interesting, although I have concerns about the validity of these insights (see Weaknesses). - Multiplicity quantification can be quite costly, and various methods in pursuit of reducing this cost can significantly benefit further auditing. The use of RashomonGB, as proposed by the authors, can be an important step in that direction for gradient-boosted algorithms. Weaknesses: - While the presentation of the rest of the concepts and the theoretical discussion were easy to follow, important details about the RashomonGB method and the details of the empirical setup were either missing (even from the Appendix) or imprecise. For instance, the Rashomon set of the gradient boosting algorithm isn’t going to simply be the iterative extension of Rashomon sets at every residual level, i.e., equation 4 is imprecise. Similarly, it seems that the epsilon value of the Rashomon set increases with more iterations, and thus it is confusing to me whether the insight that more iterations create bigger Rashomon sets is a result of multiple iterations or simply a result of bigger epsilon. See the section ‘Questions’ for more detailed comments and some follow-up questions. Edit after rebuttal: Acknowledged, correct and clarified. - There are other methods to measure predictive uncertainty in gradient-boosted algorithms. Some examples based on a cursory search (there might be more, as I’m not too familiar with GBRTs) - https://arxiv.org/abs/2205.11412 https://arxiv.org/pdf/1910.03225 https://arxiv.org/abs/2106.01682 - While I understand that prediction uncertainty is not the same as predictive multiplicity, the two are closely related, and when proposing a better method to measure multiplicity, the paper should compare itself with other stronger baselines than just retraining. Just as previous works have proposed using Monte Carlo Dropout (which was initially created as a method to measure uncertainty) as a measure of multiplicity, uncertainty measurement baselines for GBRTs could have been adopted to create reasonable baselines, and would have made the results a lot stronger. Edit after rebuttal: Acknowledged and added. Technical Quality: 3 Clarity: 3 Questions for Authors: My questions and comments mostly revolve around the RashomonGB formulation. - I don’t believe equation 4 is correct. A model formed from residual models that are present in their Rashomon sets at every step does not necessarily make a model that will be present in the Rashomon set overall. That’s because the composition of GBRTs occurs at the prediction level, while Rashomon sets are defined by the authors at the loss level. Equation 4 probably would have been true if the loss function had a linear relationship with the model predictions, which is not an assumption I see being made anywhere in the paper. This also makes me question the empirical results, because if the RashomonGB formulation isn’t precise, do the models across which the authors calculate multiplicity even belong to the same Rashomon set? Edit after rebuttal: Acknowledged and corrected. - Can the authors comment on why they compare two situations with different Rashomon parameters and make claims on their multiplicity? For example, Proposition 3 and the following paragraph. A Rashomon set would of course be bigger with a larger value of epsilon, and having that variability when talking about other trends doesn’t seem convincing to me. Edit after rebuttal: Confusion clarified. - What was the exact epsilon value used for the experiment? I couldn’t find it anywhere in the paper. Moreover, I hope that given the Rashomon sets for the RashomonGB setup were defined with T*epsilon as the new epsilon value, the same freedom was also given to retraining. Again, if the comparison was done across methods with different epsilon values (which might not be the case, but I don’t know the details), that does not make sense to me. Edit after rebuttal: Appropriate information added. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - A central piece of the paper is their method RashomonGB. While the authors do try to emphasize the importance of this method by highlighting the number of models that can be created using their method, just the number alone is not enough to imply a better method for measuring multiplicity. Even assuming that the comparisons are indeed fair (see Questions), the differences in multiplicity are not very severe, and that makes me wonder if combining pieces of various residual models actually gives us new interesting models or do we just end up with similar models as already seen during retraining. The authors acknowledge this briefly in their limitations paragraph. Edit after rebuttal: Appropriate details added and clarified. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s constructive feedback. We address the weaknesses, questions, and limitations point-by-point below. For Weakness 1, please refer to our responses to Question 1 and Question 2. For Weakness 2: as indicated, prediction uncertainty indeed differs fundamentally from predictive multiplicity. Prediction uncertainty, derived from a Bayesian perspective, seeks to reconstruct the distribution $p(Y\|x)$ for a given sample $x$ and assess metrics such as variance or negative log-likelihood of $Y$, typically involving only one model without specific loss constraints. Conversely, predictive multiplicity involves evaluating multiple models within the Rashomon set that exhibit similar loss, thereby reflecting a variety of potential outcomes for the same inputs. To elucidate these distinctions, we have compared our re-training strategy, RashomonGB, with the prediction uncertainty methods cited by the reviewer—NGBoost [R1], PGBM [R2], and IBUG [R3]—using the UCI Contraception dataset. This comparison is detailed in Figure R.2 of the attached one-page PDF. For a rigorous comparison, we applied these prediction uncertainty methods to estimate $p(Y\|x)$ (parameterized as Gaussian), sampled 1024 values of $y$, and computed the corresponding Rashomon set and predictive multiplicity metrics. The results demonstrate that RashomonGB encompasses the widest range of models, thereby providing consistently higher and more robust estimates of predictive multiplicity metrics. This comparison highlights the unique capabilities of RashomonGB in capturing a broader spectrum of potential model behaviors within the dataset. For Question 1: when utilizing GBRT for classification tasks, the method actually performs a regression on the log-likelihood using the MSE loss (Lines 140-143). The MSE loss qualifies as sub-Gaussian, aligning with the assumption set up in Proposition 1. Additionally, the pseudo-residual of the MSE loss exhibits a linear relationship between the prediction and the output of each iteration (Line 597). This allows us to apply Proposition 1 to aggregate the losses across iterations, leading to the formulation of Proposition 2, which is a detailed extension of Equation 4. Equation 4 delivers the concept of constructing the overall Rashomon set by the Rashomon sets in each iteration. We thank the reviewer for noting the error and will change the equality to $\supseteq$. We will ensure that these assumptions and their implications are clearly articulated in the revised manuscript to eliminate any ambiguity. For Question 2, we thank the reviewer for highlighting the potential confusion regarding the direction of reasoning related to $\epsilon$. To clarify, we do not start with the assumption that $\epsilon$ increases with each iteration; rather, this conclusion emerges from Proposition 3. Proposition 3 and the discussions in Section 3.3 indicate that with a constant $\rho$—as defined in Proposition 1—additional iterations result in increased conditional mutual information, which in turn necessitates a larger $\epsilon$, as detailed in Line 243. This dynamic is visually supported by Figure 2, where the conditional entropy—and consequently the mutual information—escalates as the boosting process progresses. This is because the Rashomon effect accumulates over the sequential learning problems tackled in each iteration, emphasizing the cumulative impact on the diversity within the model space. The Ablation study in Appendix E.3 further clarifies the selection of $\epsilon$ through its iterations. Figure E.8 demonstrates that fixing $\epsilon$ while re-training with different random seeds results in a decreasing percentage of models ($\rho$ from Proposition 1) in the Rashomon set. This implies that to maintain a consistent $\rho$, the chosen $\epsilon$ must increase. This observation corroborates Proposition 3's findings on the relationship between $\epsilon$ and $\rho$. For Question 3, for the experiments of reporting predictive multiplicity in Section 4.1, we report the Rashomon parameter $\epsilon$ in the vertical axis (leftmost column) in Figure 3. For the experiments of fair model selection, we report the $\epsilon$ in the caption of Figure 4. For the experiments of mitigating predictive multiplicity by model averaging, we report the $\epsilon$ (in terms of the improvement of accuracy) in the vertical axis in Figure 5. Note that the $\epsilon$ we report here is the overall Rashomon parameter after $T = 10$ iterations, i.e., T\*$\epsilon$. Moreover, we do not compare models with different $\epsilon$ as it is clearly unfair. We provide an explanation on how to interpret the results of Figure 3 in Figure R.1 in the one-page PDF, please check! Finally for the limitation, indeed, as discussed in Section 5, while re-training with different random seeds offers a "global" exploration of models within the Rashomon set, RashomonGB conducts a "local" exploration. However, RashomonGB demonstrates greater efficiency than the re-training strategy, highlighting a trade-off between the efficiency of exploring the Rashomon set and the effectiveness of capturing model diversity. For instance, in Figure 3, numerous instances across different datasets show that the re-training strategy significantly underestimates predictive multiplicity metrics (e.g., VPR and Rashomon Capacity are reported as zero), particularly when $\epsilon$ is small. This underestimation often occurs because re-training fails to gather a sufficient number of models. To the best of the authors' knowledge, finding the (sub-)optimal strategy for exploring the Rashomon set remains an active area of research. We will incorporate this additional discussion into the revised Section 5 to provide a comprehensive understanding of the trade-offs involved and the current state of research in this field. Thanks again and we would be happy to provide more clarifications and answer any follow-up questions. --- Rebuttal Comment 1.1: Comment: Thank you for your response. Some of my concerns have been answered and I will raise my scores. However, I still have some followup questions for other concerns. > Question 2 Consider I'm a user who wants to deploy these GBRTs. I would have some $\epsilon$ value in mind before I begin, eg, I can allow a 0.1 difference in loss if it allows me benefits elsewhere, say choosing fairer models. I train multiple GBRTs for 50 iterations each and get a set of models. I then filter them to find 'good models', i.e., those in my Rashomon set and would use them to calculate multiplicity, choose the model with the best fairness scores, etc. Now instead of training for 50 iterations, what if I had trained for 100 iterations? What I don't understand is why would I change my threshold ($\epsilon$) to 0.2. Wouldn't I still only want to find benefits while making sure I'm just a 0.1 loss difference away from the best model? Maybe a different way to interpret this could be that under the fixed $\epsilon$, I'm less likely to get models (as noted by the authors that the same $\epsilon$ means smaller $\rho$). In other words, I'd have less number of models to choose from if I forced the same $\epsilon$ threshold. But I feel this is a more realistic setting, and in this case, my Rashomon set has shrunk, not grown! One thing to note, the size of the Rashomon set is NOT a proxy for multiplicity. Thus, despite being a smaller set, this set of models can still have higher multiplicity, a bigger range of coverage in terms of fairness, and so on. So all the following results can, in principle, still exist. And of course, they do exist. But the narrative of increasing Rashomon set size is bothering me. I may be missing or misinterpreting something. Happy to hear more clarification from the authors. > Question 3 Ahh, I see, now the results make more sense. This was confusing at first. Please make sure to add the clarification and proper explanation on how to interpret the figure in the final version. --- Rebuttal 2: Title: Further response to Reviewer JGsC's comments Comment: We appreciate the additional feedback from the reviewer! Consider a scenario where we train $m$ models per iteration for $T\_1$ iterations, and subsequently extend the training up to $T\_2$ iterations, where $T\_1 < T\_2$. We can construct the overall Rashomon sets using models obtained from both the $T\_1$-th and $T\_2$-th iterations. With the same threshold $\epsilon$ for the Rashomon set, we can infer from Propositions 2 and 3 that the probability $1-\rho$ of a model belonging to the Rashomon set will decrease over the iteration. Let $\rho\_1$ and $\rho\_2$ represent the probabilities for iterations $T\_1$ and $T\_2$, respectively, then $1-\rho\_1 \geq 1-\rho\_2$. Consequently, the number of models from the $T\_1$-th iteration that are included in the Rashomon set with threshold $\epsilon$ will be $m^{T\_1} \times (1-\rho\_1)$. Similarly, for the $T\_2$-th iteration, the count will be $m^{T\_2} \times (1-\rho\_2)$. It is important to note that although $1-\rho\_1 \geq 1-\rho\_2$, indicating a decrease, this reduction is linear with respect to the number of iterations (as suggested by the term $1-T\rho$ in Proposition 2). However, the number of models generated by RashomonGB grows exponentially with the number of iterations (Line 158). Thus, despite the decreasing probability, the total number of models in the Rashomon set ($m^{T\_2} \times (1-\rho\_2)$) will asymptotically increase with the number of iterations. We hope the explanation reduces the confusion, and will add the clarification in the revised version. We agree with the reviewer's comment that the size of the "empirical" Rashomon set (see Line 82) is not a proxy for multiplicity. For instance, an empirical Rashomon set with 100 globally diverse (e.g., obtained by re-training with different seeds) models might exhibit a higher predictive multiplicity metric (e.g., VPR) compared to another empirical Rashomon set containing 1000 models that differ only locally. The size of the "true" Rashomon set (see equation (1)), on the other hand, representing an ideal scenario achievable with unlimited computational and storage resources, can indeed act as a proxy for predictive multiplicity. In this context, predictive multiplicity metrics are non-decreasing with a larger size of the true Rashomon set (i.e., a larger $\epsilon$). We are grateful to the reviewer for highlighting the ambiguity in our discussion. We will clarify the distinction between the true and empirical Rashomon sets in Section 3 and in the sections discussing empirical studies in the revised manuscript. For Question 3, we are thankful for the reviewer's feedback in the initial review round, which guided us in enhancing the presentation of Figure 3. We are glad that the reviewer acknowledges the clarity brought by the additional figure and explanation. In the revised version, we will ensure to include detailed explanations on how to interpret Figure 3 effectively. === changing the typo in $1-\rho\_1 \leq 1-\rho\_2$ to $1-\rho\_1 \geq 1-\rho\_2$ as pointed out by Reviewer JGsC in the discussion. === --- Rebuttal Comment 2.1: Comment: Your response helped me narrow down where my confusion came from. I had not considered the exponentially growing number of models themselves, which despite a decreasing probability of a model being in the Rashomon set, would still overall make a bigger Rashomon set. Thank you for the clarification. Small correction: I believe the authors meant to write $1-p_1 \geq 1-p_2$, and not the other way around. Probably a small typo. To summarize our discussion, make sure to fix equation 4 and add appropriate information on how to interpret the figures. As for my concerns and confusion with increasing epsilon, it might be an artifact of my own reading of the work and not necessarily anything missing in the paper, but I’d encourage the authors to add appropriate clarifications and incorporate a discussion of how things evolve under a fixed epsilon, which, in my opinion, is a more realistic setting. Good work! --- Reply to Comment 2.1.1: Title: Further response to Reviewer JGsC Comment: We appreciate the reviewer for guiding us to improve the quality of this manuscript, and for the summary of our discussions. We have fixed the typo in our comment above and left a note. We would definitely include our discussion and the reviewer's suggestion in the revised version! Thanks again!	Summary: The paper explores the concept of predictive multiplicity in gradient boosting models. The Rashomon effect refers to the existence of multiple models that perform similarly well on a given dataset. The authors formalize this effect in the context of gradient boosting, introduce a new method called RashomonGB to efficiently explore this multiplicity, and demonstrate its application on various datasets. The paper aims to improve the estimation of predictive multiplicity and model selection, especially with considerations for group fairness. Strengths: 1. The introduction of RashomonGB represents a novel method for exploring the Rashomon set in gradient boosting, offering an exponential search space as opposed to traditional linear methods. 2. The paper provides a robust theoretical foundation using statistical learning and information theory to analyze the Rashomon effect, enhancing the understanding of this phenomenon in gradient boosting. 3. The authors demonstrate the practical utility of RashomonGB on a wide range of real-world datasets, including tabular and image data, showcasing its versatility and effectiveness. Weaknesses: 1. While the paper discusses the positive societal impacts of RashomonGB, it lacks a thorough exploration of potential negative impacts or misuse of the method. 2. The theoretical analysis relies on several assumptions that may not hold in all practical scenarios, potentially limiting the generalizability of the findings. 3. The paper mentions the intention to release code post-review, but the lack of immediate open access to code and data can hinder reproducibility and independent validation by other researchers. 4. Implementing RashomonGB might be complex for practitioners without a strong background in the theoretical aspects of machine learning and gradient boosting, potentially limiting its adoption in the industry. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can the method be extended or adapted for other types of machine learning models beyond gradient boosting? 2. How does the choice of hyperparameters in RashomonGB affect the stability and reliability of the results? 3. What are the practical challenges faced during the implementation of RashomonGB, and how can they be addressed to facilitate broader adoption? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's feedback and questions. For Weakness 1, in the Introduction (Lines 26-29), we discuss the beneficial aspects of the Rashomon effect within the framework of responsible machine learning, highlighting its role in fairness by imposing additional constraints on models. This can be seen as searching for a fair model within the Rashomon set, which is further elaborated in Section 4.2 and [19]. On the other hand, the negative implications of the Rashomon effect (Lines 30-35), suggest that predictive multiplicity may occur, leading to decisions for certain individuals being arbitrarily based on randomness in the training process rather than on learned knowledge. Moreover, the RashomonGB, along with other methods designed to monitor predictive multiplicity, could lead to negative societal impacts. For example, it might be exploited by service providers to identify models within the Rashomon set that disadvantageously affect the benefits (e.g., loan approvals) of certain populations, without showing significant statistical differences from non-discriminatory models. This could contribute to what we term an "Algorithmic Leviathan" [21], where discrimination and bias are concealed under the guise of algorithmic arbitrariness. For Weakness 2, specifically, we treat the loss function as a random variable, and assume that this loss function behaves as a sub-Gaussian random variable. Sub-Gaussianity is a practical assumption, as it can be easily achieved by clipping the loss (Lines 184-188). It is also a common assumption in theoretical analysis as it effectively generalizes the concept of boundedness [71, 75]. Adopting the sub-Gaussian assumption enables us to conduct a more expansive analysis. Additionally, we would like to highlight the novelty of our approach in establishing dataset-related bounds on the Rashomon set (i.e., Proposition 1). To the best of our knowledge, previous studies on the Rashomon effect have primarily focused on characterizing the Rashomon set through its hypothesis space. Our approach provides a fresh perspective and a significant contribution to the understanding of the Rashomon effect and predictive multiplicity in machine learning. We will enhance the clarity of these points in the revisions to Section 3.2. For Weakness 3, we plan to release the code, but only after the review process is complete for two reasons: First, releasing the code during the review phase could compromise the anonymity required by the double-blind review process in the NeurIPS guidelines. Second, early release could potentially infringe upon the protection of our intellectual property. Despite not releasing the code at this stage, we have provided a detailed, step-by-step procedure in the paper on how to construct the RashomonGB and reproduce our findings, specifically in Figure 1, Sections 4, and Appendices B.3 and D (Lines 266-270). This should enable reviewers and readers to understand and evaluate our methodology thoroughly. Moreover, the datasets used in this paper are all public available, and detailed descriptions including preprocessing can be found in Appendix D. For Weakness 4, it's important to clarify that the RashomonGB method can be straightforwardly implemented by training multiple models (i.e., weak learners) at each iteration, selecting $m$ models that meet the loss constraints defined by the Rashomon set, contrary to training only one model as illustrated in Figure 1. Additionally, during the inference phase, RashomonGB employs the method expansion outlined in Section 3.1 and depicted in Figure B.6 of Appendix B.3. This adaptability makes RashomonGB particularly suitable for industry applications, especially given that tabular datasets are still prevalent. Furthermore, although the training and inference processes of RashomonGB are simple and user-friendly, the effectiveness of the approach is underpinned by rigorous and innovative propositions detailed in Section 3. These factors collectively ensure that RashomonGB is not only practical but also robust, enhancing its applicability in various real-world scenarios. For Question 1, yes, we demonstrate the utility of RashomonGB beyond beyond decision trees as weak learners by incorporating alternative weak learners, such as convolutional neural networks (shown in the results for CIFAR-10 in the lowest row of Figure 3 and the GrowNet in reference [6]), and linear regression (illustrated in Figure E.9 of Appendix E.4). This flexibility confirms that the methodology developed in our study is adaptable to a variety of settings beyond conventional gradient boosting. For Question 2, the key hyper-parameters of RashomonGB include the number of iterations $T$ and the number of models per iteration $m$, which ablation studies are included in Appendix E.6 and E.7. With a constant probability $\rho$ as per Proposition 3, increasing $T$ accumulates the Rashomon effect at each iteration, which in turn increases $\epsilon$. Additionally, increasing the number of models $m$ per iteration allows RashomonGB to explore a broader range of models, thereby enhancing the reliability of model selection and the estimation of predictive multiplicity. For Question 3, current implementations of gradient boosting, such as those available in Scikit-Learn or XGBoost, do not support training multiple models within a single iteration. Additionally, RashomonGB requires a unique filtering process during each boosting iteration to regulate the loss deviation, which is critical for constructing the Rashomon sets. Please also refer to our reply for Weakness 4. Thanks again for the comment! We are happy to provide more information or answer any follow-up questions. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed reply. The rebuttal is very appreciated. I increased my score to 6.	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their time and effort in reading and commenting on the manuscript. We appreciate that the reviewers found that the paper “study a novel problem” (Reviewer HS14, JGsC, XEPy, and E8Ec), “has robust and interesting analysis on dataset-related Rashomon set bound” (Reviewer HS14, JGsC and XEPy), and "propose RashomonGB which is easy to implement and its practical utility validated with real-world datasets " (Reviewer E8Ec and HS14). Below, we outline the specific enhancements and changes that will be incorporated into the revised version of our paper. - We clarify the contributions of this paper. The Rashomon effect and predictive multiplicity are not yet studied for gradient boosting algorithms, which is widely used for tabular datasets, especially in the field of responsible machine learning. With empirical studies, we present both the positive (fair model selection in Section 4.2) and negative (predictive multiplicity in Section 4.1) impacts of the Rashomon effect for gradient boosting, along with two algorithms to mitigate predictive multiplicity (Section 4.3). The analysis we developed here is not limited to gradient boosting, i.e., decision tree weak learners, but can also be applied to other sequential learning schemes with neural network (GrowNet) or linear regression weak learners (CIFAR-10 results in Figure 3 and Figure E.9 of Appendix E.4). - We clarify and add more discussions on the theoretical results regarding the dataset-related Rashomon set bound. We explain that the sub-Gaussian assumption for the loss random variable is a generalization of boundedness, and is a pratical and common assumption for information-theoretic bounds. Moreover, we explain that the Rashomon effect on the residual cumulates with the number of iterations and increase the Rashomon parameter under the same $\rho$. We also refer the reviewers to Figure E.8 in the Appendix, which demonstrates that fixing $\epsilon$ while re-training with different random seeds results in a decreasing percentage of models ($\rho$ from Proposition 1) in the Rashomon set. This observation corroborates Proposition 3’s findings on the relationship between the mutual information, $\epsilon$ and $\rho$. - We clarify the comparison between RashomonGB and Re-training for estimating predictive multiplicity metrics and fair model select, and how to properly interpret Figure 3 in the main text in Rebuttal Figure R.1. We first perform re-training with different random seeds and perform RashomonGB by using the same weak learners, i.e., the training cost of RashomonGB and Re-training are the same and hence the comparison presented in the paper is fair. - We provide an additional experiments to compare RashomonGB with three baselines in estimating prediction uncertainty for gradient boosting regression trees, including the NGBoost [R1], PGBM [R2], and IBUG [R3] in Rebuttal Figure R.2. Note that prediction uncertainty aims to estimate $P(Y\|x)$ given a sample $x$ while predictive multiplicity aims to construct a Rashomon set with multiple models with similar losses. As observed in Figure R.2, RashomonGB is able to cover more diverse models than NGBoost, PGBM, and IBUG, and hence less under-estimates predictive multiplicity metrics. Finally, we would like to point out that the core problem in Rashomon effect research is to explore models in the Rashomon set. For a general hypothesis space it is computationally infeasible. Therefore, there is a fundamental trade-off between the efficiency and effectiveness of exploring diverse models in the Rashomon set. Re-training is the de facto method that gives the most diverse models in the Rashomon set while also the most inefficient. The RashomonGB trades in the diversity by searching for "local" models, in return of improving the efficiency. Please feel free to follow up! We very much welcome further discussions. [R1] Duan, Tony, Avati Anand, Daisy Yi Ding, Khanh K. Thai, Sanjay Basu, Andrew Ng, and Alejandro Schuler. "Ngboost: Natural gradient boosting for probabilistic prediction." In International conference on machine learning, 2020. [R2] Sprangers, Olivier, Sebastian Schelter, and Maarten de Rijke. "Probabilistic gradient boosting machines for large-scale probabilistic regression." In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery and data mining, 2021. [R3] Brophy, Jonathan, and Daniel Lowd. "Instance-based uncertainty estimation for gradient-boosted regression trees." Advances in Neural Information Processing Systems, 2022. Pdf: /pdf/8a06bba119f0ff6c75946cdb6074d6b85b84ce0b.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
TAS-GNN: Topology-Aware Spiking Graph Neural Networks for Graph Classification	Reject	Summary: There's a large performance gap for graph tasks, especially graph classification tasks, between the spiking neural networks and artificial neural networks. The authors proposes the problems as the neuron's under starvation and illustrated the reason of the problem. To solve the problem, TAS-GNN was proposed. The main contributions of the paper are as follows: 1: Starvation problem of spiking neurone in GNNs in graph classification tasks are identified. 2: A strategy was proposed to address the spike frequency deviations on the basis of the correlation between graph topology and spike frequency patterns. The authors conduct experiments on 5 popular datasets and use several different designs of GNN layer. The results show competitive potential of the TAS-GNN. Strengths: 1:This is a well-written paper, from the formulation of the problem to the solution. The author's motivation for the use of graph topology is clear. 2:The method of using topology-awaregroup-adaptive neurons shows competitive results compared with other baselines. The ablation study makes the result more persuasive. 3: The Figures in the paper are quite straightforward, easy to follow. Weaknesses: 1: The name of the paper is "Topology-Aware Spiking Graph Neural Networks". However, as I can tell the only graph topology used in the method is nodes degree, which is used to group the neurons. I wonder if it is appropriate to name it as "topology aware", or the author can explain it more. 2: The analysis regarding the performance of the method is lack of discussion. For instance, in some datasets, such as MUTAG and IMDB-Binary, the proposed method achieve quite competitive results while in PROTEINS it doesn't. It's betted to explain what cause the phenomenon, like the characteristics of the datasets? Also, in table 2, the results of GAT and GAT+TAG in IMDB-Binary are the same. It's better to make an explanation about them. 3: There're several typos and basic grammar mistakes in the paper that will affect the presentation of the paper. In line 120 " and apply is to"; The sentence in line 123 is hard to understand Technical Quality: 3 Clarity: 2 Questions for Authors: 1: In section 3 the authors mentioned the hypothesis that the phenomenon mentioned above is caused by the topology of the real-world graphs. What motivates you to have the hypothesis? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: na Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for acknowledging our contributions, along with positive and constructive feedbacks. We respond to the comments as below. ### W1. Gap between graph topology and node degree We used node degree information for one of the representative graph topology properties. As the reviewer mentioned, degree information is not the entirety of topology information. Thus, we will refine our claim to reflect that we used degree information instead of topology. The reason why we initially used the term topology is that we regarded degree information as a representative feature of graph topology and thought topology would be better understood than degree to readers. For instance, several papers [1, 2, 3, 4] have utilized degree as a core information representing the topology. [1] Bounova, Gergana, et al. "Overview of Metrics and Their Correlation Patterns for Multiple-Metric Topology Analysis on Heterogeneous Graph Ensembles." Physical Review E, 2012. [2] Tangmunarunkit, Hongsuda, et al. "Network Topology Generators: Degree-Based vs. Structural." SIGCOMM Computer Communication Review, 2002. [3] Zhang, X., et al. "On Degree Based Topological Properties of Two Carbon Nanotubes." Polycyclic Aromatic Compounds, 2020. [4] Zhou, Shi, and Raúl J. Mondragón. "Accurately Modeling the Internet Topology." Physical Review E, 2004. ### W2. Lack of analysis for performance * Small performance gain on PROTEINS dataset Before analyzing the performance gains, please allow us to clarify that we believe our performance in PROTEINS is also competitive as it achieves significant improvements (up to 6.74%) over the SNN baselines. Typically, it is widely perceived as a normal behavior for current SNNs to achieve slightly lower performance compared to ANNs (e.g., SNN transformers [5, 6, 7]) despite their energy efficiency. Thus, the goal is usually to narrow the gap between the ANNs and SNNs, and comparisons are often done among other SNNs. It might seem contradictory because many of the TAS-GNN results outperform ANNs in Table 1. On this, we are genuinely excited about finding out a case where SNNs beat their ANN counterparts, although the exact reason is yet to be investigated. Having said that, the performance improvement in PROTEINS is relatively smaller than on other datasets since it did not beat ANNs. We would like to provide our analysis on this. We believe the reason is related to the spike diversity in the last-layer neurons. We could observe that TAS-GNN was not sufficiently diverse in the last-layer spike histogram (Figure 6 in the Appendix) on the PROTEINS dataset. In contrast, in the other datasets (MUTAG, ENZYMES, IMDB-BINARY), last-layer spikes were effectively more diverse than those of SpikingGNN. From this, it might be possible to achieve higher performance on the PROTEINS dataset if we could further increase the spike diversity in the last layer. [5] Zhou, Zhaokun, et al. “Spikformer: When Spiking Neural Network Meets Transformer.” ICLR, 2023. [6] Zhou, Zhaokun, et al. “Spikformer v2: Join the High Accuracy Club on ImageNet with an SNN Ticket.” arXiv:2401.02020, 2024. [7] Yao, Man, et al. "Spike-driven transformer." NeurIPS, 2024. * The same performance for GAT, GAT+TAG method accuracy GAT and GAT+TAG both diverge on IMDB-Binary (50% accuracy on binary classification). This issue may relate to the dataset's and GAT architecture’s properties. For instance, the IMDB-BINARY dataset has no node features, meaning the only available information is the degree information used when passing through the GNN layers. Specifically, the absence of node features becomes significant in the Graph Attention Network (GAT) layer, which focuses on smoothing degree information through learnable edge weights (i.e., attention). This is why using TAG on the IMDB-BINARY dataset could be ineffective; the attention mechanism diminishes the importance of the degree information. ### W3. Writing mistakes on paper Thank you for suggesting the grammar mistakes in our paper. We appreciate the detailed review and will ensure that all errors, including the ones mentioned by the reviewer on lines 120 and 123, are corrected. We will carefully review the entire document to address any other potential issues. * line 120: apply is to → apply them to * line 123: SNN layer, consist → SNN layer. It consists * line 213: keeps updating with → updating with * line 288: hiearchical , utilze → hierarchical, utilize * line 300: explored → exploring ### Q1. How did we get the motivation by topology information by observing the phenomenon in section 3 We were motivated by the pattern similarity between real-world graph distributions and the spike density distributions. Many real-world graphs are popular for their power-law distribution [8, 9], which indicates that there exists a few high-degree nodes with the majority of low-degree nodes. When we observed the starvation problem in Figure 1 (a), we realized that the pattern resembles that of the well-known degree distributions. This had led us to the hypothesis, which was later validated in Fig1(b). [8] Jure Leskovec et al. “Graphs over time: densification laws, shrinking diameters and possible explanations.” KDD. 2005. [9] Jure Leskovec et al. 2007. “Graph evolution: Densification and shrinking diameters.” ACM Trans. Knowl. Discov. Data, 2007. --- Rebuttal Comment 1.1: Title: Response to Author Rebuttal Comment: I would like to thank the authors for their detailed discussion and for addressing my concerns. As mentioned by one of the authors, the term 'graph topology' instead of 'degree' can be misleading. I hope the authors will further consider refining this expression. I've increased my score. Good luck! --- Reply to Comment 1.1.1: Comment: Thank you! We sincerely appreciate the reviewer for the feedback. We will make sure to address the issue in the paper. If there are any additional points the reviewer would like to clarify, please let us know. We will be more than happy to address them.	Summary: This paper primarily discusses integrating Spiking Neural Networks (SNNs) into Graph Neural Networks (GNNs) to address several key challenges in graph classification tasks. Specifically, the paper proposes a new method called TAS-GNN (Topology-Aware Spiking Graph Neural Networks) which leverages the topology of graphs to improve the performance of spiking neural networks in graph classification tasks. Strengths: （1）The authors clearly articulate the performance gap between existing Graph Neural Networks (GNNs) and Spiking Neural Networks (SNNs) in graph classification tasks. （2）The authors conduct an in-depth analysis of the performance degradation of spiking neural networks in graph classification tasks and introduce the "neuron starvation" problem. （3）The authors propose topology-aware group-adaptive neurons (TAG) based on the graph's topology, a novel approach that helps address the neuron starvation issue. （4）The authors provide a detailed description of how to convert input graphs into spike representations, perform message passing, and classify the graphs. （5）The authors validate the method's generalizability and effectiveness by using multiple public datasets (such as MUTAG, PROTEINS, ENZYMES, NCI1, IMDB-BINARY) in the experimental section. Weaknesses: （1）The authors mention several application areas and challenges, but the references and comparisons to existing literature are not sufficiently comprehensive. （2）Although the methodology section describes the main steps, it lacks detailed descriptions of some key aspects such as threshold initialization and the specific training process. （3）Although there are some ablation studies, the analysis of the individual contributions of each component is insufficient, making it difficult to determine the specific impact of each component on the overall performance improvement. Technical Quality: 3 Clarity: 2 Questions for Authors: （1）Could you provide more details on how the neuron starvation problem was diagnosed? Specifically, what metrics or observations were used to identify this issue in SNNs for graph classification? （2）The paper mentions the use of learnable initial thresholds for neurons. Could you elaborate on how these initial values are set and what specific strategies or heuristics were used to determine them? （3）Conduct a more thorough ablation study to analyze the independent contributions of each component (e.g., TAG, learnable initial thresholds) to the overall performance. This will help readers understand the significance of each part of the proposed method. （4）The sensitivity analysis shows variations in performance with different initial thresholds and learning rates. Could you explain why certain thresholds or learning rates were more effective and how they were chosen? （5）How does TAS-GNN scale with very large graphs in terms of computational efficiency and memory usage? Are there any specific optimizations or techniques used to handle large-scale datasets? （6）While the paper compares TAS-GNN with several baseline methods, could you consider including comparisons with more recent or advanced GNN models that have shown strong performance in graph classification tasks? （7）Have you tested TAS-GNN on any real-world applications or datasets beyond the ones mentioned? If so, could you share the results and insights gained from these experiments? Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: (1) While the paper discusses the neuron starvation problem and the sensitivity of initial thresholds, it does not explicitly outline the broader limitations of the proposed TAS-GNN method. It would be beneficial to include a dedicated section that explicitly lists and discusses the limitations of the current work. (2) The paper does not thoroughly address how TAS-GNN scales with extremely large datasets or very high-dimensional graphs. Including an analysis of computational complexity and memory usage for larger graphs would provide a clearer understanding of the scalability limitations. (3) While multiple datasets are used, the paper could further discuss the generalizability of TAS-GNN to other types of graph-based tasks beyond classification, such as regression, clustering, or even dynamic graphs. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you. We appreciate acknowledging the strength in our work and providing detailed feedbacks. We would like to answer the questions as follows. ### W1/Q7/L3. Extensibility to other datasets and application areas. Thank you. We extended the evaluation to more datasets (IMDB-MULTI, and REDDIT-BINARY) and tasks (regression and clustering). TAS-GNN outperforms the SNN-GNN baselines in most cases. See [Table1] in the attachment. The importance of the additional datasets (IMDB-Multi, REDDIT-Binary) lies in their connectivity since they lack node features. Our method proves to be more effective on datasets highly related to connectivity. On the new tasks, the graph regression task is similar to the graph classification, where TAS-GNN outperforms ANN. This demonstrates that resolving the neuron starvation problem can be beneficial for overall graph-level tasks. However, on clustering tasks, the binary information loss results in a performance decrease. Despite this, TAS-GNN still outperforms all SNN baselines. ### W2/Q2. Description of the training process, especially for updating the initial threshold We apologize for the confusion. We believe the original description given in Section 4.3 was confusing especially with ambiguous use of the term “initial threshold”. Here’s our second take with clarification and added details: 1. During the inference with TAG (Section 4.2), the threshold ($V^g_{th}(t)$) assigned to group $g$ adaptively changes every step by Eq.10. At the inference step 0, the step-zero threshold values $V^g_{th}(0)$ are all initialized to $V_{init}$, a hyperparameter that we referred to as initial threshold in the paper. 2. With the threshold learning scheme in Section 4.3, the $V^g_{th}(0)$ is now a trainable parameter that is initialized at training epoch 0 with $V_{init}$ (instead of the inference step 0). The values $V^g_{th}(0)$ are learned through gradient descent during the training epochs. The ambiguity arises from two types of initial points – the initial (inference) step and the initial (training) epoch. We will distinguish them as step-zero threshold and epoch-zero threshold. Please also see Algorithm1 in the Appendix A.5. ### W3/Q3. Additional ablation study for performance Thank you. [Table3] in the attachment shows a more detailed ablation study including sections 4.2 (TAG) and 4.3 (threshold learning). As observed, the “Baseline+TAG” and “Baseline+threshold learning” both outperform “Baseline” in most cases. Interestingly, adding threshold learning alone to the baseline does not improve performance for any model with IMDB-BINARY dataset. We believe this emphasizes our key claim that adequate grouping is important for SNN-GNN performance. Please also see Table 5 of the Appendix A.6 for the impact of gradually moving from “Baseline+threshold learning” (1 degree group) to TAS-GNN (max groups). ### Q1. Details about diagnosing neuron starvation problem Thank you. For each node, there are 128 feature neurons. When the average spike occurrence of those 128 neurons over five timesteps are less than 10%, we diagnosed the node as under starvation. We will clarify them in the paper. When we evaluated IMDB-BINARY dataset with this metric, 93.3% of the neurons suffered from starvation. ### Q4. Sensitivity with initial threshold and learning rates We would like to clarify that the TAG method (red lines of Figure 4) is sensitive to the $V_{init}$, but further applying the threshold learning (blue lines of Figure 4) makes it stable. The stableness comes from the fact that the TAG method uses $V_{init}$ as the step-zero threshold, while the addition of threshold learning method uses $V_{init}$ as the epoch-zero threshold. During the training, the step-zero threshold is learned to find a suitable value. Please also see our clarification in W2/Q2. Regarding the learning rate, a range of 0.01 to 0.05 is effective due to the use of large dropout values, which was needed to reduce oversmoothing. This makes the training favor relatively large learning rates. However, when we use too large learning rates, the overall training process becomes unstable. Please also see [Table4] in the attachment. ### Q5/L2. Optimization techniques that should be considered for large scale graphs For larger graphs, the additional memory cost of #(unique degrees) X #(feature dim) is needed per layer to store the per-group thresholds. When the #(unique degrees) grows too large, merging vertices of similar degrees into groups would reduce the cost. Please also see Table 5 in the Appendix A.6 for the sensitivity to the #groups. ### Q6. Performance evaluation on more recent or advanced GNN models Thank you. We added experimental results with advanced models using DeepGCN [1] and UGformerV2 [2] in [Table2] in the attachment, where TAS-GNN maintains performance benefits. DeepGCN is a representative architecture that uses residual connections to address the over-smoothing problem. UGformer is a graph-transformer architecture that applies a self-attention mechanism with GNN layers. We replaced all GNN layers in these models with corresponding SNN-GNN layers. [1] Li, Guohao, et al. "Deepgcns: Can gcns go as deep as cnns?." ICCV. 2019. [2] Nguyen et al. "Universal graph transformer self-attention networks." WWW. 2022. ### L1. Additional limitations of TAS-GNN Thank you. Our limitations could be itemized as below. * This work mainly focuses on graph classification tasks. We believe the proposed TAS-GNN can be extended to other tasks, such as graph regression. We performed additional experiments on them, and will add them in the paper. * The performance evaluation is conducted on relatively small graphs. The proposed method has the potential to be applied to extremely large graphs, we leave them as out future work. * The proposed method increases GNN training time due to additional learnable parameters. However, we the overhead is negligible. --- Rebuttal Comment 1.1: Title: TAS-GNN: Topology-Aware Spiking Graph Neural Networks for Graph Classification Comment: Thank you for the response. I have read the response as well as the reviews and rebuttals of other reviewers. I will stand by my original recommendation. --- Reply to Comment 1.1.1: Comment: Thank you for taking the time to share your thoughts. We genuinely appreciate your comments and feedback, and we're confident your insights will help us further enhance our work.	Summary: The paper presents a novel approach called TAS-GNN (Topology-Aware Spiking Graph Neural Networks) to address the performance gap between spiking neural networks (SNNs) and artificial neural networks (ANNs) in graph classification tasks. The authors identify a "starvation" problem in spiking neurons within GNNs, where many neurons do not emit any spikes during inference, leading to severe information loss. This problem is more pronounced in graph classification tasks, where the test set graphs are independent from the training set, unlike in transductive or inductive learning settings. Strengths: 1. This paper identifies a critical "starvation" problem in spiking neurons within Graph Neural Networks (GNNs), where many neurons do not emit any spikes during inference, leading to severe information loss. This problem is more pronounced in graph classification tasks, where the test set graphs are independent from the training set 2. The paper proposes a novel approach called TAS-GNN (Topology-Aware Spiking Graph Neural Networks) to address the graph classification problem. Weaknesses: 1. The authors use the node degree instead of the concept of topology, there’s a large gap between the graph topology and node degree. 2. The authors solve the graph classification task as a contribution, which is not a significant challenge for spiking graph neural networks. 3. The advantage of Spiking Neural Networks (SNN) is their low energy consumption. However, the paper does not mention the feature, so it is unclear why graph neural networks should be combined with SNN. The motivation behind TAS-GNN is not clear. Technical Quality: 2 Clarity: 2 Questions for Authors: The important points listed in weakness 1-3. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors adequately addressed the limitations. The authors should discuss more details of the potential negative societal impact of the work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the acknowledging our novelty of the work and providing constructive feedbacks. We have addressed the comments as below. We will revise our paper according to the rebuttal. ### W1. Gap between graph topology and node degree We used node degree information for one of the representative graph topology properties. As the reviewer mentioned, degree information is not the entirety of topology information. Thus, we will refine our claim to reflect that we used degree information instead of topology. The reason why we initially used the term topology is that we regarded degree information as a representative feature of graph topology and thought topology would be better understood than degree to readers. For instance, several papers [1, 2, 3, 4] have utilized degree as a core information representing the topology. [1] Bounova, Gergana, et al. "Overview of Metrics and Their Correlation Patterns for Multiple-Metric Topology Analysis on Heterogeneous Graph Ensembles." Physical Review E, 2012, [2] Tangmunarunkit, Hongsuda, et al. "Network Topology Generators: Degree-Based vs. Structural." SIGCOMM Computer Communication Review, 2002 [3] Zhang, X., et al. "On Degree Based Topological Properties of Two Carbon Nanotubes." Polycyclic Aromatic Compounds, 2020. [4] Zhou, Shi, and Raúl J. Mondragón. "Accurately Modeling the Internet Topology." Physical Review E, 2004. ### W2. Significance of challenge for graph-classification task In this work, the main challenge is to achieve high performance (i.e., accuracy) in the graph classification task using SNN. This is different from simply adopting existing SNN-GNN to build a working example on graph classification task, which does not pose a significant challenge. Rather, the challenge was that simple adoption of previous works (e.g., SpikingGNN, SpikeNet, PGNN) causes severe accuracy degradation. Thus, our contribution would not be supporting the task itself but identifying the neuron starvation problem and proposing techniques to address it. ### W3. Consideration of energy efficiency Thanks for the great idea. We compared the energy consumption in [Table5] below. The TAS-GNN model shows significant energy efficiency (69%-99%) compared to ANN architectures. \|\|\|\|\|\|\|\| \|---\|---\|---\|---\|---\|---\|---\| \|ENERGY\| (mJ)\| MUTAG\|PROTEINS\|ENZYMES\|NCI1\|IMDB-BINARY\| \| GCN\| ANN\| 0.53 \|6.92 \|3.29\|21.41\|3.16 \| \|\|TAS-GNN\|0.10\|0.94 \|0.52 \|5.28 \|0.70 \| \|\|Reduction\|82.17%\|86.36%\|84.29%\|75.35%\|77.72%\| \| GAT\|ANN\|0.33 \|4.59\|2.42\|15.55\|2.20\| \|\|TAS-GNN\|0.07\|0.05\|0.34\|4.75\|0.55\| \|\|Reduction\|79.96%\|98.82%\|85.83%\|69.44% \|74.89% \| \|GIN\|ANN\|0.39\|4.96\|2.33\|15.26\|2.24\| \|\|TAS-GNN\|0.05\|0.02 \|0.14\|1.67\|0.06\| \|\|Reduction\| 87.14%\|99.64%\|94.14%\|89.04%\|97.48% \| [Table5] Energy consumption table We observe significant energy reduction in the PROTEINS dataset with GIN architectures, showing a 99.64% reduction. In contrast, we observe that our worst case for energy consumption showed in the NCI1 dataset with GAT architecture, indicating 69.44% energy reduction. Since the GAT architecture requires more information to learn its attention mechanisms, the spike frequency was higher than in other architectures. Additionally, we found that NCI1 results in particularly frequent spikes among the datasets, which led to less energy reduction. The theoretical estimations we provided are based on [5,6], which are widely used for SNN energy consumption analysis. We calculated each layer's sparsity $\gamma$ and FLOPs (floating point operations). Assuming MAC and AC operations are implemented on 45nm hardware, we handled $E_{MAC}$ = 4.6pJ, $E_{AC}$=0.9pJ. The energy consumption of SNN is calculated with $E_{AC} \times \gamma \times \text{FLOPs}$. As spike sparsity in our experiment varied greatly depending on GNN architectures and datasets, we evaluated each spike sparsity. [5] Horowitz, M. “1.1 Computing’s Energy Problem (and What We Can Do About It).” 2014 IEEE International Conference on Solid-State Circuits Conference, 2014 [6] Yao, M et al. “Attention Spiking Neural Networks.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023 ### L1. Negative societal impact of the work Thanks for the suggestion. We believe the negative societal impact could be discussed in the following ways. * Our research uses social network graphs like REDDIT-BINARY, and this could potentially be misused to screen for ideologies and biases, leading to negative effects such as infringing on human freedom. * Despite our efforts to focus on energy reduction, our research could contribute to environmental problems due to carbon dioxide emissions during the training process. However, these problems are not unique to our work; it is an issue faced by all graph neural networks (GNNs). These GNNs can unintentionally retain biases present in the data they learn from, resulting in ethical and societal concerns that need to be addressed by everyone working in this field. --- Rebuttal Comment 1.1: Title: Response to Author Rebuttal Comment: The authors give a detailed reply to weaknesses 1-3. Since these three issues are crucial to the motivation and novelty of the paper, judging from the rebuttal statement, the author needs to make a large amount of modifications to address these issues. The current version is not suitable for public publication. --- Reply to Comment 1.1.1: Title: Response to Reviewer 4pSD Comment: Thank you for reading our rebuttal and sharing your thoughts. However, could we ask for a reconsideration? We believe the amount of revisions necessary from the current version would not be very extensive. Firstly, regarding W1, it would be sufficient to replace our term topology with a degree, as we used topology to refer to degree information in our explanation within the paper. Secondly, regarding W2, we believe the significance of the graph classification task is already clearly demonstrated by the accuracy degradation of other baselines. Therefore, we believe there is less need for a major revision. Finally, regarding W3, we think this concern can be addressed by simply adding this experiment table to the main manuscript. Adding these experiments to the paper will certainly be valuable. However, they would not significantly affect the overall flow of our paper, because reducing energy consumption is already a fundamental advantage of SNNs, and we focus on improving their accuracy for practicality.	Summary: This paper proposes topology-aware spiking graph neural networks with adaptive thresholds based on a group of neurons for graph classification. The paper first diagnoses the poor performance as the existence of neurons under starvation caused by the graph structure. Then the paper proposes the adaptive threshold among neurons partitioned by degrees, as well as the learnable initial threshold and decay rate to reduce the sensitivity. Experiments on several datasets show superior performance of the proposed method. Strengths: 1. This paper proposes the first SNN design to target graph classification. 2. This paper identifies the starvation problem and proposes a novel topology-aware group-adaptive technique. 3. Experiments show superior performance on several datasets, some outperforming ANNs. Weaknesses: 1. The proposed method seems to be a hybrid ANN-SNN model rather than a pure SNN design. The paper did not discuss how this will affect the deployment of the model on potential neuromorphic hardware, since SNNs mainly target those hardware to obtain energy efficiency. 2. The paper did not discuss the (theoretical) energy efficiency estimation, which is a major motivation for considering SNNs as stated in Introduction. 3. Or if the motivation is to get models with better performance than ANN, then Table 1 does not include state-of-the-art ANN results for comparisons. Technical Quality: 3 Clarity: 3 Questions for Authors: Some recent works also study SNN for link prediction tasks in graphs [1] besides node-level classification, which may be discussed. [1] Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning. ICML 2024. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors discussed limitations in Appendix A.1. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for acknowledging our contributions and positive feedback. We faithfully address the comments below. ### W1: The proposed method seems to be a hybrid ANN-SNN model rather than a pure SNN design. We proposed TAS-GNN as a pure SNN design, which shares almost the same backbone architecture with the existing SNN-GNN family (SpikingGCN [1], SpikeNet [2], Spiking GATs [3]). We apologize for such confusion, where we suspect the below two reasons. 1. Figure 2 shows GNN layer + SNN layer separately. Although Figure 2 separately depicts a GNN layer and SNN layer, the ‘GNN layer’ was not intended to indicate the use of ANN, but just to show a GNN-style message passing occurs here. The term ‘SNN layer’ was merely used to indicate that membrane exists there. All layers operate and communicate by spikes and no ANN layer is involved. We will clarify this and fix the figure accordingly. 2. An extra operation of aggregation and combination in the last layer. We wanted to note that it comprises a different structure from intermediate layers, not to indicate the use of ANN layers. For fair comparisons, we followed the convention of many SNN architectures that allow a few additional operations after the ultimate-layer neurons [4,5,6]. However, our architecture is orthogonal to such configuration and the same advantage remains over the baselines without those. If the reviewer had other reasons for considering our model a hybrid, please let us know so that we can clarify. [1] Zhu, Zulun, et al. “Spiking Graph Convolutional Networks.” IJCAI, 2022. [2] Li, Jintang, et al. “Scaling Up Dynamic Graph Representation Learning via Spiking Neural Networks.” AAAI, 2023. [3] Wang, Beibei, and Bo Jiang. “Spiking GATs: Learning Graph Attentions via Spiking Neural Network.” arXiv:2209.13539, 2022. [4] Zhou, Zhaokun, et al. “Spikformer: When Spiking Neural Network Meets Transformer.” ICLR, 2023. [5] Zhou, Zhaokun, et al. “Spikformer v2: Join the High Accuracy Club on ImageNet with an SNN Ticket.” arXiv:2401.02020, 2024. [6] Shi, Xinyu, Zecheng Hao, and Zhaofei Yu. “SpikingResformer: Bridging ResNet and Vision Transformer in Spiking Neural Networks.” CVPR, 2024. ### W2. Is the motivation for the energy efficiency of SNN? If so, show comparison over ANNs. Thanks for the great idea. We compared the energy consumption in [Table5] below. The TAS-GNN model shows significant energy efficiency (69%-99%) compared to ANN architectures. \|\|\|\|\|\|\|\| \|---\|---\|---\|---\|---\|---\|---\| \|ENERGY\| (mJ)\| MUTAG\|PROTEINS\|ENZYMES\|NCI1\|IMDB-BINARY\| \| GCN\| ANN\| 0.53 \|6.92 \|3.29\|21.41\|3.16 \| \|\|TAS-GNN\|0.10\|0.94 \|0.52 \|5.28 \|0.70 \| \|\|Reduction\|82.17%\|86.36%\|84.29%\|75.35%\|77.72%\| \| GAT\|ANN\|0.33 \|4.59\|2.42\|15.55\|2.20\| \|\|TAS-GNN\|0.07\|0.05\|0.34\|4.75\|0.55\| \|\|Reduction\|79.96%\|98.82%\|85.83%\|69.44% \|74.89% \| \|GIN\|ANN\|0.39\|4.96\|2.33\|15.26\|2.24\| \|\|TAS-GNN\|0.05\|0.02 \|0.14\|1.67\|0.06\| \|\|Reduction\| 87.14%\|99.64%\|94.14%\|89.04%\|97.48% \| [Table5] Energy consumption table We observe significant energy reduction in the PROTEINS dataset with GIN architectures, showing a 99.64% reduction. In contrast, we observe that our worst case for energy consumption showed in the NCI1 dataset with GAT architecture, indicating 69.44% energy reduction. Since the GAT architecture requires more information to learn its attention mechanisms, the spike frequency was higher than in other architectures. Additionally, we found that NCI1 results in particularly frequent spikes among the datasets, which led to less energy reduction. The theoretical estimations we provided are based on [7,8], which are widely used for SNN energy consumption analysis. We calculated each layer's sparsity $\gamma$ and FLOPs (floating point operations). Assuming MAC and AC operations are implemented on 45nm hardware, we handled $E_{MAC}$ = 4.6pJ, $E_{AC}$=0.9pJ. The energy consumption of SNN is calculated with $E_{AC} \times \gamma \times \text{FLOPs}$. As spike sparsity in our experiment varied greatly depending on GNN architectures and datasets, we evaluated each spike sparsity. [7] Horowitz, M. “1.1 Computing’s Energy Problem (and What We Can Do About It).” 2014 IEEE International Conference on Solid-State Circuits Conference, 2014 [8] Yao, M et al. “Attention Spiking Neural Networks.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023 ### W3. Is the motivation to get better performance over ANN algorithms? Our motivation is not to outperform ANNs, but to build a SNN-GNN architecture that can outperform existing baselines and reduce the existing gap to ANN performance. As the reviewer has correctly assumed in W2, the advantage of ANN lies in energy efficiency. However, we are excited to find that several datasets such as IMDB-BINARY results show superior performance compared to ANN under the same backbone architecture. Our performance is also comparable with the leaderboard results using ANNs (https://paperswithcode.com/sota/graph-classification-on-imdb-b.) ### Q1. Discussion on recent study Thanks for the great suggestion to discuss papers that support the SNN-GNN area [9]. We will add the discussion below. "While the node classification task is the most commonly addressed, a recent work GRSNN [9] explores the link prediction task to achieve energy efficiency using SNNs in knowledge graphs, demonstrating that incorporating synaptic delays into SNNs allows for effective relational information processing with significant energy savings." [9] Xiao, Mingqing, et al. “Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning.” ICML, 2024. --- Rebuttal Comment 1.1: Comment: I would like to thank the authors for their detailed responses, clarification, and additional results. Most of my questions are solved and I raise my score. --- Reply to Comment 1.1.1: Comment: Thank you! We greatly appreciate the reviewer's thorough consideration of our response and the valuable insights shared with us.	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for dedicating their time to evaluate our work. We are encouraged that they found our approach to be novel in developing TAS-GNN (MTu6, 4pSD, icJz, 5K64), with clear motivation demonstrated by diagnosing neuron starvation (4pSD, icJz) and competitive performance compared to other baselines (MTu6, icJz, 5K64). Our rebuttal can be summarized as follows: * Theoretical energy consumption comparison between TAS-GNN and ANN architectures * Clarification on the usage of the term ‘topology’ * Analysis of the performance across different datasets and tasks * Experiments with additional GNN architectures * Additional ablation studies for performance Please note that the attached PDF contains additional experimental results, which we explain in detail in each response. For the remaining rebuttal period, we will try our best to answer any further questions and discussions on the topic. Pdf: /pdf/555808dc2989098feb55af2015590a0a448d24e8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
GraphCroc: Cross-Correlation Autoencoder for Graph Structural Reconstruction	Accept (poster)	Summary: This paper proposes a cross-correlation autoencoder for graph structural reconstruction. The authors first analyze the problems of existing self-correlation encoder. Then, a cross-correlation autoencoder is designed. Experimental results show the effectiveness of the cross-correlation autoencoder. Strengths: 1. The motivation is clear and the cross-correlation autoencoder is reasonable. 2. The paper is well-written and easy to follow. 3. The experiments are comprehensive. Weaknesses: 1. The authors mention that the current self-correlation methods can not address specific (sub)graph structures. But this paper only presents an overall experimental performance. It is unclear how the proposed cross-correlation autoencoder performs given a specific graph structure. 2. It is not clear whether the graph dataset used in the paper is a directed or undirected graph. Since the cross-correlation autoencoder can represent the directed graph effectively, it is suggested to consider the directed graph dataset. 3. More different architectures of the encoder and decoder should be employed to further verify the effectiveness of the cross-correlation mechanism. Technical Quality: 3 Clarity: 3 Questions for Authors: see Weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Evaluate the proposed cross-correlation autoencoder given specific graph structures, e.g., islands and symmetric structures. In Sec.2.2 and 2.3, we explore the limitations of self-correlation and the capabilities of cross-correlation in accurately representing specific graph structures, such as non self-loop, symmetric structures, and directed edges. In the evaluation, we follow previous GAE research to apply our proposed GraphCroc on common real-world graph tasks (Table 1), which are undirected asymmetric graph structures. Additionally, we present a comparison of how various models reconstruct specific graph structures in `Fig. 1` of our rebuttal PDF file. Specifically, we randomly generate 4 graphs where each graph is topologically symmetric and contains no self-loop. These graphs are then used to evaluate the performance of different GAE models. The visualizations clearly illustrate that our GraphCroc model proficiently reconstructs these graph structures. For DiGAE which is also based on cross-correlation, it can well reconstruct the special graph structures, further proving our discussion in Sec.2.2 and 2.3. In contrast, other models often erroneously predict connections between symmetric nodes and introduce unwanted self-loops, highlighting the superior representation ability of cross-correlation in handling these specialized scenarios. Note that for EGNN, it does not predict positive edges between nodes, which seems not to follow our analysis in Sec.2.2 with Euclidean encoding (sigmoid$(C(1-\|\|z_i-z_j\|\|^2))$, $C>0$) [6]. This is because EGNN slightly improves this encoding to sigmoid$(w\|\|z_i-z_j\|\|^2+b)$ where $w$ and $b$ are learnable. Since no-self-loop nodes require sigmoid$(w\|\|z_i-z_i\|\|^2+b)=$sigmoid$(b)<0.5$, $b$ is forced to be negative, inducing negative prediction on symmetric edges which have $z_i=z_j$ under symmetric sturctures; alternatively, we can also regard it as the naive Euclidean encoding but with $C<0$. Therefore, EGNN still cannot handle well the graph reconstruction on the special graph structures. [6] Graph normalizing flows, NeurIPS'19. > Evaluation on direction graphs. Most graph reconstruction research is evaluated on undirected tasks, so for a fair and comprehensive comparison, we also utilize undirected graphs. Additionally, structural reconstruction on directed graphs has gained focus recently in the DiGAE work (AAAI'22). Hereby, we further investigate GraphCroc's performance on directed graph datasets. We compare GraphCroc with DiGAE, given that only cross-correlation-based methods can effectively capture directional relationships between nodes. To construct the dataset, we sample subgraphs from the directed Cora and Citeseer datasets. We randomly selected 1,000 subgraphs, using 800 for training and 200 for testing. The results are shown below ($\bar{N}$ represents the average number of nodes per graph): \|\| cora ($\bar{N}=41$) \| cora ($\bar{N}=77$) \| cite ($\bar{N}=16$) \| \|-\|:-:\|:-:\|:-:\| \| GraphCroc\| 0.9946 \| 0.9996 \| 0.9999 \| \| DiGAE \| 0.6870 \| 0.8296 \| 0.9083 \| Our GraphCroc can well reconstruct the directed graph structure with almost perfect prediction, which significantly outperforms the DiGAE model. This advantage comes from the expressive ability of our proposed U-Net-like architecture. > More different architectures of the encoder and decoder should be employed to further verify the effectiveness of the cross-correlation mechanism. Thank you for your suggestion of further evaluation on cross-correlation across different GNN architectures. In addition to the GCN kernel used in our GraphCroc model, we extend our analysis to include other widely used graph architectures such as GraphSAGE [7], GAT [8], and GIN [9]. To incorporate these architectures into the cross-correlation framework, we replace the GCN module with corresponding operations while preserving the overarching structure, which includes the encoder, the dual-branch decoder, and the skip connections between the encoder and decoder. Furthermore, we explore how GraphCroc performs without skipping connections. The overall architecture and training configurations remain consistent with those outlined in Table 5 of our paper, except for the QM9 dataset, where we limit the training to 20 epochs due to its large size and the limited time frame for our rebuttal. The results, presented below, follow the format of Table 1 in our paper, providing a clear comparison across different architectures: \| \| GraphSAGE\| GAT\| GIN\|GraphCroc(w/o skip connection)\| GraphCroc\| \|-\|:-:\|:-:\|:-:\|:-:\|:-:\| \| PROTEINS \| 0.9898 \| 0.9629 \| 0.9927 \| 0.9934 \| 0.9958 \| \| IMDB-B \| 0.9984 \| 0.9687 \| 0.9980 \| 0.9975 \| 0.9992 \| \| Collab \| 0.9985 \| 0.9627 \| 0.9954 \| 0.9976 \| 0.9989 \| \| PPI \| 0.9774 \| 0.9236 \| 0.9467 \| 0.9447 \| 0.9831 \| \| QM9 \| 0.9972 \| 0.9978 \| 0.9974 \| 0.9966 \| 0.9987 \| Overall, all architectures employing cross-correlation effectively reconstruct graph structures, underscoring the significance of cross-correlation as a core contribution of our work. Given that training each model requires several hours, particularly for large datasets such as PPI and QM9, we did not fine-tune the hyperparameters much during model training. The results presented here may represent a lower bound of these architectures' potential performance. Therefore, we refrain from ranking these cross-correlation-based architectures due to their closely matched performance, and we adopt a conservative stance in our comparisons. Nevertheless, it is evident that most of these architectures (except GAT) generally surpass the performance of self-correlation models shown in Table 1 of our paper, highlighting the efficacy of cross-correlation in graph structural reconstruction. [7] Inductive representation learning on large graphs, NeurIPS'17; [8] Graph attention networks, ICLR'18; [9] How powerful are graph neural networks?, ArXiv'18. --- Rebuttal Comment 1.1: Comment: Dear reviewer QrYC, As the rebuttal discussion is about to close, we would like to confirm whether our rebuttal has adequately addressed your concerns. If there are any questions you would like to discuss, please let us know.	Summary: This paper proposed a method to address the limitations of existing graph autoencoder (GAE) models that primarily rely on self-correlation for graph structure representation. They claim existing GAE often fail to accurately represent complex structures like islands, symmetrical structures, and directional edges, particularly in smaller or multiple graph contexts. The proposed model, GraphCroc, introduces a cross-correlation mechanism that aims at enhancing the representational capabilities of GAEs. It employs a mirrored encoding-decoding process to ensure robust structural reconstruction and introduces a loss-balancing strategy to tackle representation bias during optimization. Strengths: 1. The idea to introduce two latent space for reconstructing the graph structure is "simple and intuitive". 2. The writing is clear and easy to follow. 3. The experimental results are sound. Weaknesses: 1. This paper lacks discussion on related works. There already exists some works trying to solve the graph autoencoder structure recovering issues. For example, including position encoding [1] or adding extra node labels [2]. How the proposed method is compared with these methods, from the perspective of effectiveness and efficiency? [1] You, Jiaxuan, Rex Ying, and Jure Leskovec. "Position-aware graph neural networks." International conference on machine learning. PMLR, 2019. [2] M. Zhang, P. Li, Y. Xia, K. Wang, and L. Jin, Labeling Trick: A Theory of Using Graph Neural Networks for Multi-Node Representation Learning, Advances in Neural Information Processing Systems (NeurIPS-21), 2021. 2. As the proposed method generate two latent embeddings, I wonder if there exists some techniques to control them to be different with each other? Otherwise I am concerned that whether the two embeddings could converge to each others. Technical Quality: 3 Clarity: 3 Questions for Authors: see above weakness Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > This paper lacks discussion on related works. There already exists some works trying to solve the graph autoencoder structure recovering issues. For example, including position encoding or adding extra node labels. How the proposed method is compared with these methods, from the perspective of effectiveness and efficiency? Thank you for your suggestion. The first referenced paper incorporates positional information into node embeddings, whereas the second paper explores the effectiveness of the labeling trick by adding labels to nodes of interest. Before delving into comparisons, it is necessary to highlight a shared point across these studies and ours: all approaches try to introduce asymmetry into node embeddings, to effectively represent node connections in symmetric/isomorphic graph structures. In our work, this is achieved through cross-correlation and the application of two-set node embeddings. This shared point is also applied in the DiGAE approach [5], although it is not explicitly discussed in their work, given their focus on representing asymmetric directed graphs. This commonality makes these methods all theoretically effective for representing special graph structures, such as symmetric nodes; yet this effectiveness is limited to their application scenarios. Our GraphCroc model is designed to represent the entire graph all at once, directly outputting the complete predicted adjacency matrix rather than individual predicted edges. This makes GraphCroc particularly suited for per-graph downstream tasks such as graph classification. In contrast, the labeling trick is designed for link prediction, where only specific node pairs or subsets are labeled, creating asymmetry between these nodes and others. Labeling must be conducted either pair-by-pair or subset-by-subset to maintain this asymmetry, which is a critical aspect of the labeling trick's approach. If all nodes were labeled to reconstruct all edges, distinctions would be needed between specific node labels to break down symmetry effectively. However, GraphCroc offers an easier approach, reconstructing the graph structure in one go without the need for repeated generation of node embeddings for different node pairs. This efficiency makes GraphCroc more effective for whole-graph reconstruction, while the labeling trick remains better suited for tasks focusing on link prediction between selected node pairs. Position Encoding (PGNN) is also evaluated in link prediction tasks. However, PGNN has the potential to be applied to structural reconstruction across the entire graph, provided the selected anchor set facilitates asymmetric message aggregation. Regarding efficiency, PGNN's process must regenerate the anchor set of the newly structured graph after each pooling layer, due to the dynamic nature of graph structures in GNN tasks, which can lead to notable inefficiencies. In contrast, GraphCroc does not need preprocessing on the graph structure, but dynamically adjusts asymmetry through the parameters of its decoder during training; thus, GraphCroc offers greater efficiency and adaptability. Furthermore, the asymmetry in PGNN is predetermined by the anchor set selection and remains static throughout training, making it dependent on initial anchor choices. This contrasts with GraphCroc's more flexible and adaptive approach to handling graph structure representations. In conclusion, a significant application difference is that GraphCroc encodes the entire graph in a latent embedding suitable for downstream tasks, allowing the inclusion of graph pooling layers. However, both position encoding and the labeling trick primarily focus on node-level embeddings, preserving the graph's original structure, with downstream tasks limited to node classification and link prediction. [5] Directed graph auto-encoders, AAAI'22. > As the proposed method generate two latent embeddings, I wonder if there exists some techniques to control them to be different with each other? Otherwise I am concerned that whether the two embeddings could converge to each others. The difference between two latent embeddings (denoted as $P, Q$) is fundamental to cross-correlation as opposed to self-correlation in which $P=Q$; therefore, it is necessary to make them not converge to each other. One method of explicitly controlling this divergence is by incorporating regularization terms into the loss function, such as cosine similarity ($cos(P,Q)$). However, our decoder architecture inherently encourages differentiation between $P$ and $Q$ since they are derived from two separate branches of the decoder. This structure can allow $P$ and $Q$ to diverge adaptively in response to the specific needs of the graph tasks. If a graph cannot be well represented by self-correlatin, our two-branch structure will encourage sufficient divergence on $P$ and $Q$ to suit structural reconstruction. To evaluate the differentiation between $P$ and $Q$, we compute their cosine similarity and present a histogram of these values for each graph task in `Fig. 2` of our rebuttal PDF file. Across all tasks, the cosine similarity between the node embeddings under cross-correlation is generally low, typically below 0.6. This shows that our two-branch decoder effectively maintains the independence of the node embeddings, which are adaptively optimized for various graph tasks. Furthermore, this adaptive optimization underscores the superiority of cross-correlation in real-world applications, as evidenced by GraphCroc's superior performance in graph structural reconstruction compared to other methods (Table 1 of our paper). --- Rebuttal Comment 1.1: Comment: Dear reviewer Xxf1, As the rebuttal discussion is about to close, we would like to confirm whether our rebuttal has adequately addressed your concerns. If there are any questions you would like to discuss, please let us know.	Summary: This paper theoretically analyzes the limitations of existing graph autoencoders (GAE) in representing special graph features such as islands, symmetrical structures, and directional edges. To address this, the paper proposes a new GAE method, GraphCroc, which employs a cross-correlation mechanism that significantly enhances the representational capabilities of GAEs. Strengths: 1. The paper clearly shows the limitations of existing GAEs through theoretical analysis. 2. The experimental results demonstrate the advantages of the proposed method in structural reconstruction and graph classification tasks. 3. The paper is easy to follow. Weaknesses: 1. In Table 1, the improvements of GraphCroc are evident only on two datasets. 2. While the proposed cross-correlation method performs better than the general self-correlation method on island, symmetric structures, and directed graphs, it would be beneficial to include more results in reconstruction visualization, particularly regarding island or directed edge reconstruction. 3. Some related works [1] need to be discussed. [1] Liu, Chuang, et al. "Where to Mask: Structure-Guided Masking for Graph Masked Autoencoders." arXiv preprint arXiv:2404.15806 (2024). Technical Quality: 3 Clarity: 2 Questions for Authors: 1. How about the performance of the proposed method on directed graphs? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > In Table 1, the improvements of GraphCroc are evident only on two datasets. AUC is widely used to evaluate graph structural reconstruction tasks in GAE, due to its unbiased performance on positive and negative edges. Thus, we adopt this metric to assess the adjacency matrix reconstruction in our work. GraphCroc exhibits only marginal improvements on Collab and QM9 graph tasks compared to EGNN [1]. This can be attributed to the good representation capabilities of EGNN on certain graph tasks. Specifically, EGNN achieves AUC over 0.99 on these tasks, nearing perfect prediction (AUC=1), thus leaving minimal room for improvement by GraphCroc. However, Table 1 also illustrates the significant advancements of GraphCroc over EGNN on other graph tasks like PROTEINS and IMDB-B. Moreover, EGNN was not successfully applied to the PPI task due to the out-of-memory issue (on a 40GB A100 GPU), whereas GraphCroc operates within a mere 3GB of memory, as we measured. This demonstrates not only the generality of GraphCroc in handling various real-world tasks but also its efficiency in implementation. Additionally, the core of our paper focuses on the cross-correlation mechanism, which inherently offers greater expressiveness than self-correlation models like EGNN. Note: Regarding the performance, we have some further comments. Observing the self-correlation models, both EGNN and GraphCroc(SC) can achieve good AUC scores, only slightly lower than our proposed cross-correlation-based GraphCroc, yet GAE and VGAE have poor performance. This is because GAE/VGAE [2] was proposed early and has a simple architecture, while EGNN and GraphCroc(SC) (U-Net-like architecture [3]) are more complicated and have more powerful representation ability on graph structures, outperforming normal GAE/VGAE methods. On the other hand, common graph tasks are usually asymmetric, and self-loops are overlooked in previous works, which is why even the self-correlation-based EGNN and GraphCroc(SC) can still perform well on these graphs reconstruction. Nevertheless, cross-correlation can still boost the graph structural reconstruction on these graphs, as evidenced by the best performance of our GraphCroc on all graph tasks. [1] E (n) equivariant graph neural networks, ICML'21; [2] Variational graph auto-encoders, NeurIPS'16; [3] Graph u-nets, ICML'19. > More results in reconstruction visualization for specific graph structures, such as island, symmetric structures, and directed graphs. In Sec.2.2 and 2.3, we explore the limitations of self-correlation and the effectiveness of cross-correlation in expressing specific graph structures. Given that previous GAE research often evaluates undirected asymmetric graph structures, we similarly provide the evaluation of graph structural reconstruction for common real-world graph tasks (Table 1). Additionally, we agree that it would be highly beneficial to include clear visualizations demonstrating how various encoding methods succeed or fail in reconstructing specific graph structures. island (without self-loop) and symmetric graph structure: We generate 4 topologically symmetric graphs devoid of self-loops. We task the evaluated models with learning to reconstruct these graph structures and assess their performance. The visualization of their reconstruction is presented in `Fig. 1` of our rebuttal PDF file. The visualization clearly demonstrates that our GraphCroc model effectively reconstructs these specialized graph structures. For DiGAE which is also based on cross-correlation, it can also well reconstruct the special graph structures, further proving our discussion in Sec.2.2 and 2.3. In contrast, other self-correlation-based models tend to incorrectly predict connections between symmetric nodes and islands, and incorrectly introduce self-loops on nodes. Note that for EGNN, it does not predict positive edges between nodes, which seems not to follow our analysis in Sec.2.2 with Euclidean encoding (sigmoid$(C(1-\|\|z_i-z_j\|\|^2))$) [4]. This is because EGNN slightly improves this encoding to sigmoid$(w\|\|z_i-z_j\|\|^2+b)$ where $w$ and $b$ are learnable. Since no-self-loop nodes require sigmoid$(w\|\|z_i-z_i\|\|^2+b)=$sigmoid$(b)<0.5$, $b$ is forced to be negative, inducing negative prediction on symmetric edges that have $z_i=z_j$ under symmetric structures. Therefore, EGNN still cannot handle well the graph reconstruction on the special graph structures. directed graph structure: We conduct an evaluation using datasets of directed graphs. We compare GraphCroc with DiGAE, as only cross-correlation-based methods are capable of expressing directional relationships between nodes. To construct the dataset, we sample subgraphs from the directed Cora and CiteSeer datasets. Specifically, we randomly select 1,000 subgraphs. Of these, 800 subgraphs were used for training and 200 for testing. The results are detailed below, where $\bar{N}$ represents the average number of nodes per graph: \| \| Cora_ML ($\bar{N}=41$) \| Cora_ML ($\bar{N}=77$) \| CiteSeer ($\bar{N}=16$) \| \|---\|:-:\|:-:\|:-:\| \| GraphCroc\| 0.9946 \| 0.9996 \| 0.9999 \| \| DiGAE \| 0.6870 \| 0.8296 \| 0.9083 \| Our GraphCroc can well reconstruct the directed graph structure with almost perfect prediction, which significantly outperforms the DiGAE model. This advantage comes from the expressive model architecture of our proposed U-Net-like model. [4] Graph normalizing flows, NeurIPS'19. > Related work discussion: "Where to Mask: Structure-Guided Masking for Graph Masked Autoencoders." IJCAI'24. Thanks for your suggestion about the related work. This work addresses node importance in graph construction and proposes a structure-based masking strategy. This strategy guides the masking on GAE with more rationality, and is well evaluated based on the standard GraphMAE. We will give a discussion of this work and include its performance on graph classification tasks in Table 2 of our paper. --- Rebuttal Comment 1.1: Comment: Dear reviewer Msno, As the rebuttal discussion is about to close, we would like to confirm whether our rebuttal has adequately addressed your concerns. If there are any questions you would like to discuss, please let us know.	null	null	Rebuttal 1: Rebuttal: We appreciate the time and effort the reviewers have spent in providing valuable feedback! We are grateful for the reviewers' recognition of our clear writing, reasonable motivation, and sound experiments. Graph structural reconstruction is a pivotal application for graph autoencoders (GAEs), and we hope that our research offers a novel perspective on graph representation and proves beneficial to the community. In response to your comments, we address all identified weaknesses and questions. We welcome further feedback and are eager to engage in discussions regarding the paper's content. Additionally, following suggestions for enhanced visualizations, we include new figures in our rebuttal and attach them in the PDF file below. References to these figures are highlighted, e.g., `Fig. 1`. All responses to reviewers' comments will be reflected in our paper's final version. Pdf: /pdf/0b1cffac0806388e38c723d9279f3bb16e16669e.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Segment, Shuffle, and Stitch: A Simple Layer for Improving Time-Series Representations	Accept (poster)	Summary: This paper introduces a new method for time-series representation learning that enhances the modeling of non-adjacent segment dependencies. Specifically, the proposed method segments, shuffles in a learned manner and stitches the shuffled segments to combine with original time series. The proposed method is model-agnostic without adding significant parameter overhead and shows performance improvement across multiple classification and forecasting base models. Strengths: 1. The proposed method permutes the original segments to better capture inter-relations between distant segments. It is model-agnostic and introduces minimal parameter overhead to the original model. 2. Extensive experiments on various base models for both classification and forecasting tasks demonstrate the effectiveness of the proposed method. Weaknesses: 1. It it not clear how the sorting process, specifically the calculation of permutation $\sigma$ from $P$, is made differentiable. 2. The compared forecasting baselines such as Informer are no longer state-of-the-art methods. Adding more recent baselines such as Time-LLM, GPT4TS, DLinear, PatchTST would provide a clearer understanding of the proposed method's comparative benefits. 3. The basic assumption for S3 is that modeling non-adjacent dependencies is important. However, the paper lacks detailed case studies that demonstrate the specific types of non-adjacent dependencies effectively captured by S3, which are not addressed by existing models. Additionally, there is no case study to validate that the learned shuffling weights accurately represent these segment dependencies. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. The results in Tables 1, 2, and 3 seem to indicate more significant improvements in multivariate than in univariate time series tasks. Any reason behind this? 2. What does the "number of segments" represent in Figure 6 and Figure A3? Is it the number of segments for the first layer or the final layer? If it refers to "n", then in Figure A3, this number seems to perform the best when it is larger than 100 for some datasets? 3. Could you describe the inference process for the S3 method? Additionally, what are the computational overheads for training and inference times for S3? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The paper mentions potential expansions into tasks like imputation and anomaly detection. Further details on limitations from the reviewer are discussed in Weaknesses and Questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Below we provide a careful point-by-point response to each question. We would be happy to provide additional discussions/information in the author-reviewer discussion period should you have any follow-up questions. > Clarification on the differentiability of the sorting process (calculation of σ from P) While the calculation of σ from P is not differentiable, it is not used directly for the permutation of the segments. Instead, we use σ to populate the intermediate zero matrix Ω with elements from P. Matrix Ω is then turned into a binary permutation matrix which, when multiplied with the list of segments, permutes them in the correct order. Crucially, this matrix multiplication operation is differentiable, and Ω, which was built with elements from P, builds a bridge for the gradient to flow from the final output back to the elements of P. So even though the creation of σ itself is not differentiable, gradients can still flow back from the final output, through the permutation matrix, to P through the elements that were selected while creating σ. > Adding more recent forecasting baselines We have now added S3 to PatchTST and CoST [1] for forecasting on the ETT (univariate and multivariate), Weather, and Electricity datasets. The results are shown in Table R1 of the enclosed document. In these additional experiments, we observe a similar trend where the addition of S3 improves performance. > Case study. We now present 3 sample time-series before and after the S3 layer in Figure R1 of the accompanying doc (we will add more to the final paper). Here, we observe that S3 is optimized differently to learn a unique shuffling pattern that suits the baseline model and task. Given that the learnable parameters of S3 are optimized through training of the network and with the goal of maximizing the performance on that specific task, the final learned shuffling pattern serves the purpose of shuffling the time-series such that adjacent learning of originally non-adjacent segments results in more effective representations, as evidenced by the consistent improvement in results. The fact that the shuffling pattern in S3 is optimized with the learnable network and for each specific task is in fact an advantage as different models and different tasks may benefit from different re-structuring of the data. > Univariate vs. multivariate results. Thank you for this interesting observation. One possible explanation for more improvements on multivariate time-series can be that multivariate datasets inherently contain more complex dynamics and interdependencies (both temporally and between variables), which can be more challenging for models to understand. Accordingly, S3 can further help with learning of the more complex temporal dependencies, further improving performance. Empirically, we can observe that models tend to struggle more with multivariate datasets from Tables 2 and 3 in the original paper. For instance, TS2Vec has an average MSE value of 0.1184 on ETTh1 univariate forecasting, and 0.7612 on ETTh1 multivariate forecasting, which is significantly larger than the former. While we acknowledge that comparing metric values for a model across different datasets or tasks can be misleading due to potential inconsistencies in data scales or nature of the task, in this case, both the dataset and the task are more or less the same. Accordingly, the key difference between univariate and multivariate settings here would be the complexity, i.e., the dataset and tasks are more complex in multivariate due to the higher number of dimensions. > Clarification on number of segments in Figure 6 and A3 The “number of segments” in Figure 6 and A3 from the original paper is the maximum number of segments over all S3 layers in the model. For instance, let us assume 3 layers of S3 with $n_0$=4, $n_1$=8, $n_2$=16 (where $n_i$ is the number of segments for layer $i$). Here, the maximum number of segments is 16. Similarly, if $n_0$=24, $n_1$=12, $n_2$=6, then the maximum number of segments is 24. We will make sure to clarify this in the final paper. > Inference process During inference, we integrate S3 as the first layer of the model. The input sequences go through S3 first where they are segmented into several segments (the number of segments was optimized during training). These segments are then shuffled using the set of shuffling parameters that were optimized and fixed at the end of the training phase, and the learned weighted average yields the final sequence. This sequence is then fed into the subsequent stages of the baseline model for further processing and output generation. So essentially, S3 acts like any other learnable layer in a network where it is optimized during training and used with the fixed parameters during inference. > Computational overhead S3 adds only a few hundred parameters, the exact count of which depends on the specific hyperparameters selected. We show the number of added parameters in Table 8 of the paper, where we observe that in comparison to hundreds of thousands to millions of parameters in the baselines, S3 adds very few parameters, which can be considered negligible. In terms of inference time, given the ratio of added parameters vs. the original number of parameters in the baseline models, the added time is also negligible. > Experiments on Anomaly detection To further expand the tasks beyond univariate and multivariate classification and forecasting, we have now performed additional experiments for the task of anomaly detection on the KPI and Yahoo datasets. The results for this are shown in Table R4 of the enclosed document, where we observe that the addition of S3 results in considerable performance gains. References [1] Woo, et al., CoST: Contrastive learning of disentangled seasonal-trend representations for time series forecasting. ICLR, 2022. --- Rebuttal Comment 1.1: Comment: Thank the authors for the clarifications and new experimental results! These solve most of my concerns. However, the current case study still does not show the benefits of shuffling time series. Typically, forecasting task requires the model to obey the temporal dynamics, so it is counterintuitive to shuffle the order and perturb the temporal dynamics. Moreover, models such as Transformer are able to capture long-range dependencies without the need to shuffle the segment order. Therefore, I was hoping to see in which scenarios shuffling the order better models non-adjacent dependencies compared to, say, modeling long-range dependencies using a Transformer. Additionally, not sure if I miss it, is there any visualization to show how w1 and w2 change over the training process? --- Rebuttal 2: Comment: We would like to thank the reviewer for their comments and further engaging with us on this matter. We think we now have a better understanding about the scope of the question. To analyze the scenarios in which S3 is most effective, we have now performed additional experiments, and considered 3 factors: (1) length of time-series sequences, (2) non-linearity, and (3) long-term temporal dependency. For (1), we considered the %improvement vs. sequence length. For (2) we considered the %improvement vs. the mean of squared residuals w.r.t. a linear model. And finally for (3), we considered the %improvement vs. Hurst exponent [1]. We used PatchTST (transformer-based) as the baseline model given that it is the most recent work in the area, and used the ETT, Weather, and Electricity datasets. For sequence length, however, PatchTST uses equal input lengths for everything, so we used Informer (another transformer-based model) instead for which the input length is varied depending on the dataset and forecasting horizon. The outcome of this analysis is shown in the table below, where m is the slope of a linear relationship and R is Pearson's correlation coefficient. We observe that there are direct positive relationships between improvement by S3 versus sequence length, and also long-term temporal dependency, while the non-linearity in the time-series shows no relationship. For sequence length, the moderate correlation suggests that while the impact is not very strong, it is reliable, implying that sequence length is likely a relevant factor for %improvement. Long-term temporal dependency has a substantial impact on %improvement, as indicated by the large slope. Although, the lower correlation suggests that this relationship could be influenced by other factors in the data. We hypothesize that should these long-term temporal dependencies be simple to learn or highly repetitive, the impact of S3 will be less significant as re-ordering will not be required to learn such simple dynamics while for more complex long-term dependencies, S3 has a much stronger impact (as per the high slope). \| x \|\\| y \|\\| m \|\\| R \| \| ----------------------------- \| ------------ \| ----- \| ----- \| \| Sequence length \|\\| %improvement \|\\| +0.04 \|\\| +0.55 \| \| Non-linearity \|\\| %improvement \|\\| 0.00 \|\\| +0.05 \| \| Long-term temporal dependency \|\\| %improvement \|\\| +2.89 \|\\| +0.26 \| To answer your second question, we have analyzed how w1 and w2 change over training of ETTh1 multivariate dataset with PatchTST. While unfortunately we cannot provide you with a new figure at this stage (due to NeurIPS rules), we have tried to provide a discretized version in the table below. We observe that the values steadily converge to the final results without too much fluctuation (some small fluctuations in the beginning are generally observed). \| Iteration \| w1 \| w2 \| \| ------------ \| ------ \| ------ \| \| 0 \| 0.1510 \| 0.1490 \| \| 1000 \| 0.2846 \| 0.0683 \| \| 2000 \| 0.5074 \| 0.1505 \| \| 3000 \| 0.7129 \| 0.3286 \| \| 4000 \| 0.8140 \| 0.4210 \| \| 5000 \| 0.4210 \| 0.5552 \| \| 6000 \| 0.9498 \| 0.9498 \| \| 7000 \| 0.9647 \| 0.5858 \| \| 8000 \| 0.9809 \| 0.6215 \| \| 9000 \| 0.9906 \| 0.6368 \| \| 10000 \| 0.9838 \| 0.6632 \| \| 11000 \| 0.9827 \| 0.6762 \| \| 12000 \| 0.9884 \| 0.6969 \| \| 13000 \| 0.9960 \| 0.6994 \| \| 14000 \| 0.9953 \| 0.7068 \| \| 15000 \| 0.9992 \| 0.7188 \| \| 16000 \| 0.9970 \| 0.7286 \| \| 17000 \| 1.0004 \| 0.7348 \| \| Final values \| 1.0004 \| 0.7348 \| References [1] Tong, et al., “Learning fractional white noises in neural stochastic differential equations", NeurIPS, 2022. --- Rebuttal Comment 2.1: Comment: Thank the authors for the new experimental results and explanations. These solve most of my concerns. I have raised my score. --- Reply to Comment 2.1.1: Comment: We are very happy to hear that our responses have addressed the reviewer's concerns, and appreciate the increase in score.	Summary: This paper introduces a plug-and-play mechanism called Segment, Shuffle, and Stitch (S3) designed to enhance time-series representation learning in existing models. S3 operates by dividing the original sequence into non-overlapping segments and shuffling them in a learned manner that is optimal for the given task. It then reattaches the shuffled segments and performs a learned weighted sum with the original input to capture both the newly shuffled sequence and the original sequence. This proposed model can enhance the performance of specific models in classification and prediction tasks. Strengths: The paper is easily comprehensible and straightforward. Sufficient experiments are conducted to confirm the effectiveness of the method. Weaknesses: Lack of comparative methods: In fact, the proposed method seems to share the same spirit as data augmentation methods in the time series field[1-4]. Why hasn't any data augmentation method been compared? Selection of baseline models: The selected baseline model, Informer, seems somewhat outdated. Why not choose a more recent model, e.g., iTransformer[5] or PatchTST[6]? Dataset for prediction task: The author conducted experiments on three ETT datasets, but for prediction tasks, more datasets should be considered, e.g., traffic, electricity, and weather. Time-Series Representation Claim: As the author pointed out, more tasks should be considered for time series representation learning. [1]FRAUG: FREQUENCY DOMAIN AUGMENTATION FOR TIME SERIES FORECASTING [2]Time Series Data Augmentation for Deep Learning: A Survey [3]SimPSI: A Simple Strategy to Preserve Spectral Information in Time Series Data Augmentation [4]TOWARDS DIVERSE AND COHERENT AUGMENTATION FOR TIME-SERIES FORECASTING [5]ITRANSFORMER: INVERTED TRANSFORMERS ARE EFFECTIVE FOR TIME SERIES FORECASTING [6]A TIME SERIES IS WORTH 64 WORDS: LONG-TERM FORECASTING WITH TRANSFORMERS Technical Quality: 2 Clarity: 3 Questions for Authors: What are the essential differences between the proposed method and other data augmentation methods? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: See Weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Below we provide a careful point-by-point response to each question. We would be happy to provide additional discussions/information in the author-reviewer discussion period should you have any follow-up questions. > Comparison with data augmentation methods. We acknowledge that at first glance S3 seems to share similarities to data augmentation. However, S3 has meaningful learnable parameters that train along with the rest of the model to enable a segmentation and mixup that is (a) variable throughout the training process, and (b) customized to that model and task, setting it highly apart. To further test this, we have now performed new experiments where we apply shuffling augmentation before training, shuffling augmentation plus mixup, noise augmentation, and noise augmentation plus mixup. In this experiment, we have also compared S3 to the augmentation method presented in the paper cited in your comment [1]. Please see Table R2 in the accompanying doc where we observe that S3 outperforms such data augmentation strategies. Moreover, a common characteristic of data augmentation techniques is that they tend to improve performance well for smaller datasets (often due to the lack of variations and diversity) but not so much for larger datasets [2]. To evaluate the impact of S3 based on the size of the training set, we took the LSST dataset from UEA and created 10 different subsets by dropping different amounts of data ranging from 20% to 99%. We then retrained SoftCLT [3] with and without S3 on these subsets as well as the original dataset. The results (averaged over 3 runs) are presented in Figure R2 of the accompanying doc. The results demonstrate that there is no evident trend in the performance gain with respect to the dataset size, indicating that S3 does not only benefit smaller datasets. > Experiments on a more recent transformer model We have now added S3 to PatchTST and CoST [4] for forecasting on the ETT (univariate and multivariate), Weather, and Electricity datasets. The results are shown in Table R1 of the enclosed document. In these additional experiments, we observe a similar trend where the addition of S3 considerably improves performance. > Experiments on additional dataset for forecasting We have now performed additional experiments for forecasting electricity and weather datasets for several baselines. The results are shown in Table R3 of the enclosed document. > Experiments on anomaly detection To further expand the tasks beyond univariate and multivariate classification and forecasting, we have now performed additional experiments on two popular baselines with and without S3 for anomaly detection on the KPI and Yahoo datasets to further evaluate the improvements in representation learning that S3 brings. The results for this are shown in Table R2 of the enclosed document. References [1] Xiyuan Zhang, Ranak Roy Chowdhury, Jingbo Shang, Rajesh Gupta, and Dezhi Hong. Towards diverse and coherent augmentation for time-series forecasting. ICASSP, 2023. [2] Iwana BK, Uchida S. An empirical survey of data augmentation for time series classification with neural networks. Plos one. 2021 [3] Seunghan Lee, Taeyoung Park, and Kibok Lee. Soft contrastive learning for time series. ICLR, 2024. [4] Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. CoST: Contrastive learning of disentangled seasonal-trend representations for time series forecasting. ICLR, 2022.	Summary: This paper proposes a new neural network design element which segments, shuffles, and stitches time series for improved representation learning. They evaluate their methods on forecasting and classification tasks, and show that S3 benefits some widely used baselines. Strengths: 1. To the best of my knowledge, the idea is novel, and fundamentally challenges and changes how to learn representations for time series data 2. The paper is well written and easy to follow 3. Experiments are well-designed, and results are promising Weaknesses: I have not found any major weaknesses in the methodology or experimental design. However, I think that the paper might benefit from showing what the S3 module is actually learning. For example, the authors can include the segmented, shuffled, and stitched time series on a particular dataset as an example, along with the weighted time series (used as input to the model), and the original time series. This might provide some intuition as to how this design element improves predictive performance. I think there's always scope to improve experimental design. TS2Vec is a excellent choice for classification, but not for forecasting. I would recommend that the authors use methods such as PatchTST (transformer-based) or iTransformer, TimesNet (CNN-based), N-BEATs or N-HITS (MLP-based) etc. for time series forecasting. For classification, it would also be good to compare with fully supervised methods such as ResNet1D (see [1]). ### References [1] Ismail Fawaz, Hassan, et al. "Deep learning for time series classification: a review." Data mining and knowledge discovery 33.4 (2019): 917-963. Technical Quality: 4 Clarity: 4 Questions for Authors: I do not have questions per se, but I am listing some things that I am curious about below: I would also encourage the authors to evaluate the benefits of S3 on some recent time series foundation models such as MOMENT [2], Chronos [3], Moirai [4], TimesFM [5], and/or LagLLama [6]. The MOMENT model does both classification and forecasting, so it might be interesting to see how S3 benefits pre-trained models, say by just training the S3 layer and freezing the pre-trained backbone (or some variation of this experiment). On a similar note, I wonder if S3 improves generalization and hurts memorization, or vice versa. It would be interesting to do some transfer learning experiments where you train on some time series data and evaluate the model on other time series data (see MOMENT or PatchTST for inspiration). ### References [2] Goswami, Mononito, et al. "Moment: A family of open time-series foundation models." arXiv preprint arXiv:2402.03885 (2024). [3] Ansari, Abdul Fatir, et al. "Chronos: Learning the language of time series." arXiv preprint arXiv:2403.07815 (2024). [4] Woo, Gerald, et al. "Unified training of universal time series forecasting transformers." arXiv preprint arXiv:2402.02592 (2024). [5] Das, Abhimanyu, et al. "A decoder-only foundation model for time-series forecasting." arXiv preprint arXiv:2310.10688 (2023). [6] Rasul, Kashif, et al. "Lag-llama: Towards foundation models for time series forecasting." arXiv preprint arXiv:2310.08278 (2023). Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The authors have a very brief description of limitations of their study. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Below we provide a careful point-by-point response to each question. We would be happy to provide additional discussions/information in the author-reviewer discussion period should you have any follow-up questions. > Visualisation of S3 We have now included several visualizations in Figure R1 in the enclosed document that demonstrate how the segments are rearranged according to what the model thinks is optimal for the task. The figures allow for a visual comparison between the original sequence, the shuffled sequence, and the final output. > Additional baselines for forecasting, and fully supervised baseline for classification We have implemented S3 with PatchTST, CoST [1] on the ETT (univariate and multivariate), Weather, and Electricity datasets for forecasting, and the results are shown in Table R1 of the enclosed document. We observe that S3 is able to further improve the performance of the additional baselines. For a fully supervised method, our original submission (Table 1) included the baseline DSN which was fully supervised. > S3 with pre-trained foundation models Thank you for the very interesting suggestion! Given the very limited time during the rebuttal period and the resources needed to explore large foundation models, we were unable to perform these experiments at this time. We do agree that this area would be very interesting to explore for S3 and is indeed something we had discussed as an exciting future direction to take this work. We hope to follow up our current work with a follow up study on time-series foundation models. > Experiments on transfer learning We have now performed an additional experiment in this regard: following the protocol used in TS2Vec [2], we trained the TS2Vec+S3 encoder on FordA for classification. We then froze the encoder and used the model along with fine-tuning of the classification head and S3 to classify the sequences for the other 127 UCR datasets. The average accuracy score over all 127 datasets for TS2Vec without S3 is 0.8037, and with S3 the accuracy score is 0.8160. Based on these results, it appears that S3 is indeed effective toward both in-dataset (in-distribution) and cross-dataset (out-of-distribution) settings. Please note that since the goal of S3 is to customize the reordering of the time-series specific to each dataset and task, the S3 layers would need to be fine-tuned as well (when the classification head is being re-trained), while the rest of the model stays frozen. We will add the full table of individual results on all the datasets in the final paper. References [1] Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. CoST: Contrastive learning of disentangled seasonal-trend representations for time series forecasting. ICLR, 2022. [2] Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. Ts2vec: Towards universal representation of time series. AAAI, 2022. --- Rebuttal Comment 1.1: Comment: Since we posted the rebuttal for your attention, we were able to add S3 to a foundation model, Moment [1], and use linear probing (fine-tuning the final linear layer of Moment) on their pre-trained MOMENT-1-large model along with S3 on the PTB-XL [2] dataset. The table below presents the results where we observe a considerable gain by adding S3, indicating potential for future research in this area. \| \|\\| Moment \|\\| Moment + S3 \|\\| Improvement \| \| ------------- \| ------ \| ----------- \| ----------- \| \| Test loss (lower is better) \|\\| 0.8308 \|\\| 0.7329 \|\\| 11.79% \| \| Test accuracy (higher is better) \|\\| 0.7176 \|\\| 0.7497 \|\\| 4.48% \| References [1] Goswami, et al., “Moment: A family of open time-series foundation models”, ICML, 2024. [2] Wagner, et al., “PTB-XL, a large publicly available electrocardiography dataset”, Scientific Data, 2020. --- Rebuttal 2: Title: Thank you for the excellent work and the rebuttal! Comment: Dear Authors, I really appreciate the time and effort that you have put into the study. I really like it, and I would like to maintain my current score to reflect my very positive assessment of this paper. I really appreciate figure R1, and I think it should find its way into the paper. Also the transfer learning experiments while preliminary, are very promising, and I would encourage the authors to add an element of this to their revised manuscript. Best, Reviewer 5XMz --- Rebuttal Comment 2.1: Comment: We would like to sincerely thank the reviewer for their support and encouraging comments. We agree regarding the new experiment - we will certainly add these experiments to the paper (either main section or appendix).	Summary: The paper paper introduces a new approach called Segment, Shuffle, and Stitch (S3) to enhance time-series representation learning. The method involves segmenting the time-series into non-overlapping parts, shuffling them optimally, and stitching them back together along with the original sequence. Key contributions include: - Proposing the S3 mechanism to improve time-series representation learning by dynamically reordering segments. - Demonstrating that S3 can be integrated with existing neural architectures like CNNs and Transformers, resulting in significant performance improvements. - Showing through extensive experiments that S3 enhances performance in time-series classification and forecasting tasks, with improvements up to 68%. Strengths: - Code is available, making reproducing this paper easier. - Paper is clear. - Results appear good, when considered on the set of baselines and dataset picked by the authors. Weaknesses: - Tables 1 and 2 focus on the ETT datasets, which are only a (highly intra-correlated) subset of the common forecasting datasets: Electricity, Traffic, Weather, Illness... - I see no mention of CoST in the results tables, despite being cited in the paper. This is usually a very strong baseline for contrastive approaches. Including it would certainly paint a more complete picture of the results landscape. On a related note this also applies to e.g. more recent transformer baselines. Informer is relevant, but also very far from state of the art. - Error bars would help one better contextualize the results. - The lack of an ablation study makes understanding the reason this works more complicated. Technical Quality: 2 Clarity: 3 Questions for Authors: - The 3 points in weaknesses are also questions in the sense that they ask for some new experiments to be performed. Addressing those points would be my first recommendation. - Intuitively, it feels like this work is to some extent a form of bootstrap (as data augmentation) combined with a mixup-like sample interpolation. I may be wrong on this and am happy to discuss. If so, could the authors do more of an ablation study connected to this. I.e. how does the approach outperform other (non-permutation)-based data augmentation strategies combined with the same summation operation? Edit: I have read the author's rebuttal. They have addressed questions I had and I am as a result raising my score to a 6. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Below we provide a careful point-by-point response to each question. We would be happy to provide additional discussions/information in the author-reviewer discussion period should you have any follow-up questions. > Other forecasting datasets. As per your comment, we have now performed additional experiments on the popular Electricity and Weather datasets, and present the results in Table R2 in the accompanying doc. We observe that adding S3 results in overall improvements over all the baselines for both datasets. > CoST and other transformer baselines. As per your comment, we have now added S3 to CoST and PatchTST [1] on the ETT (univariate and multivariate), Weather, and Electricity datasets, and present the results in Table R3 of the accompanying doc. We observe improvements for all the datasets. > Error bars. The sources of variation in the performance of S3 are $\mathbf{P}$, $\mathbf{w}_1$, and $\mathbf{w}_2$ as well as the random seed for the baseline model to which S3 is added. We have performed an experiment where we present the standard deviations of several classification and forecasting models in Table 7 of the original paper over 5 random initializations. We observe that (a) the standard deviations are generally very small, and (b) any possible variance stems from the baseline model as opposed to the S3 module. > Ablation and comparison against augmentation. (1) We performed a detailed ablation of each component of S3 in Table 5 of the original paper where we observed a drop in performance when each one is ablated. (2) In regards to the question on augmentation, we acknowledge that at first glance S3 seems to share similarities to data augmentation. However, S3 has meaningful learnable parameters that train along with the rest of the model to enable a segmentation and mixup that is (a) variable throughout the training process, and (b) customized to that model and task, setting it highly apart. To further test this, we have now performed a new experiment where we apply shuffling augmentation (with different number of segments) before training, shuffling augmentation plus mixup, noise augmentation (with a mean of zero and a variance of 1, with a magnifying coefficient of 0.01), and noise augmentation plus mixup, on the ETTm1 multivariate dataset. Please see Table R2 in the accompanying doc where we observe that S3 outperforms such data augmentation strategies. (3) Lastly, a common characteristic of data augmentation techniques is that they tend to improve performance well for smaller datasets (often due to the lack of variations and diversity) but not so much for larger datasets [2]. To evaluate this, we took the LSST dataset from UEA and created 10 different subsets by dropping different amounts of data ranging from 20% to 99%. We then retrained SoftCLT [3] with and without S3 on these subsets as well as the original dataset. The results (averaged over 3 runs) are presented in Figure R2 of the accompanying doc. The results demonstrate that there is no evident trend in the performance gain with respect to the dataset size, indicating that S3 does not only benefit smaller datasets. References [1] Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. In The Eleventh International Conference on Learning Representations, 2023. [2] Iwana BK, Uchida S. An empirical survey of data augmentation for time series classification with neural networks. Plos one. 2021 [3] Seunghan Lee, Taeyoung Park, and Kibok Lee. Soft contrastive learning for time series. ICLR, 2024. --- Rebuttal Comment 1.1: Comment: I have read the author's response to my points, and as mentioned in the main review I am raising my score. --- Reply to Comment 1.1.1: Comment: We sincerely thank the reviewer for the great comments which have resulted in improving the paper. We appreciate that our rebuttal has answered your questions and are grateful for increasing your score.	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for their time and for providing us with constructive feedback. We are happy to see the engaging comments given by all the reviewers. We have carefully addressed all the concerns raised under the individual response section. Following, we provide a summary of our responses. * Adding S3 to PatchTST and CoST for forecasting: As per the suggestion of Reviewers FZgx, xbmW, 5XMz, and 2e8Y, we added S3 to PatchTST and CoST for the task of forecasting. The results are presented in Table R1 of the enclosed document. * Comparison with data augmentation: As per the questions by Reviewers xbmW and 2e8Y, we provide a discussion on the key differences between S3 and data augmentation, and also performed several experiments to compare S3 against data augmentation techniques. The results for these are outlined in Table R2 and Figure R2 of the enclosed document. * Expanding forecasting baselines on Electricity and Weather datasets: As per the feedback of Reviewers xbmW and 2e8Y, we have now performed additional experiments for several forecasting baselines on the Electricity and Weather datasets. The results for these are presented in Table R3 of the enclosed document. * Visualization and qualitative analysis: As per the suggestion of Reviewers FZgx and 5XMz, we have included several visualizations (Figure R1 in the enclosed document) which show how the segments are rearranged with S3 and allow for a visual comparison between the original sequence, the shuffled sequence, and the final output. * Experiments on anomaly detection: As per the suggestion of Reviewers 2e8Y and dMD7, we have now performed experiments on anomaly detection on the KPI and Yahoo datasets. The results for this are shown in Table R4 of the enclosed document. * Additional clarifications: we provide additional clarifications regarding differentiability, inference overhead, and others. Pdf: /pdf/a55175b035feecbc9f84852ab6772c39a8b64eab.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper introduces a simple but effective differentiable module that performs pre-processing to input multivariate time-series before being fed into any differentiable model for arbitrary task. The pre-processing involves segmenting, shuffling the segments and stiching them together. The novelty include making this seemingly discrete operations into a differentiable module. This simple idea yields significant improvement in performance of different kinds of models over variety of datasets. Strengths: 1. The method is simple and easy to add to most deep learning models 2. The technical details are well-motivated and explained 3. The method also improves training efficiency and convergence time along with performance with very little increate in model complexity 4. Experimental results across different tasks are strong Weaknesses: 1. Visualization and any qualitative study on the shuffling and segments generalted by S3 would greatly benefit the readers. 2. How well does it optimize transformer based models, especially those that already do segmentation like PatchTST since the attention module captures the relations all pairs of segments already? 3. Does the representations due to S3 generalize to multiple tasks at a time or do we need to retrain for each task? Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: 1. Lack of understanding on the segment permutations generated and why they are better for the model performance atleast qualitatively Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Below we provide a careful point-by-point response to each question. We would be happy to provide additional discussions/information in the author-reviewer discussion period should you have any follow-up questions. > Visualizations and qualitative analysis. We have now included several visualizations in Figure R1 of the enclosed document. The figure shows how the segments are rearranged according to what the model thinks is optimal for the task. > Performance on transformer-based models. We have now added S3 to PatchTST on the ETT (univariate and multivariate), weather, and electricity datasets and obtained the results presented in Table R1 of the accompanying document. We observe that S3 is able to further improve the performance of the additional baselines. Also, please note that in the initial set of experiments originally presented in our paper, Informer [1] is a transformer-based model, where we observed significant improvements when S3 was added (please see Tables 2 and 3 in our paper). > Does S3 need to be trained for each task? We naturally retrain S3 for each task since the backbone model itself requires to be retrained as well. The S3 layer is simply added to the backbone and retrained just like all the other layers of the model. Having said this, we performed an experiment where we selected the exact same hyperparameters for all the tasks to observe how well a common hyperparameter could perform. We present this result in Table 6 of the original paper, where we observe that although the results are expectedly lower than the optimum hyperparameters, we still see meaningful gains. References [1] Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai hang. Informer: Beyond efficient transformer for long sequence time-series forecasting. AAAI, 2021. --- Rebuttal Comment 1.1: Title: Thank you Comment: I thank the authors for the detailed response to my and other reviewers' questions. This has helped strengthen my position for the paper's merit. I increased my score to 6. --- Reply to Comment 1.1.1: Comment: We sincerely thank the reviewer for the great comments which have resulted in improving the paper. We appreciate that our rebuttal has answered your questions and are grateful for increasing your score.	null	null	null	null	null	null
Rethinking Deep Thinking: Stable Learning of Algorithms using Lipschitz Constraints	Accept (poster)	Summary: To solve the stability of Deep Thinking models, this paper proposes to constrain activation functions to be Lipshitz-1 functions. The original DT and DT-R models have training stability problem, basically because of scale explosion or vanishing. The authors revealed the stability problem, attribute the problem to Lipschitz constants, proposed ways to ensure Lipschitz smoothness, and show the effectiveness of their approach through a few examples used in the original DT paper, as well as include the traveling salesman problem. Strengths: * This paper is clearly written and well motivated. * The storyline is very reasonable: identify problems => propose ways to solve the problem => show the approach actually works * This approach is mathematically grounded. * Experiments are thorough, by running many random seeds and report error bars. Weaknesses: * The idea is quite straight-forward (may not be a bad thing, but make technical contributions smaller) * In the TSP problems, DT-L's results seem worse than NN Tours and BNN Tours. At least some explanation is warranted. * I'm not fully convinced by the significance of this paper. The examples shown in the paper are quite toy. Are there more examples you expect DT-L would work? * I'd appreciate more visualizations that can intuitively show the benefits of DT-L over DT/DT-R? Maybe some Figures like in the original DT paper. * The title is not very informative. Might be better to mention Lipschitz smoothness in the title. Technical Quality: 3 Clarity: 3 Questions for Authors: * In Line 141-142, I don't quite get this comment "Although any Lipschitz constant less than 1 would guarantee convergence, the nature of the problem solving mechanism we seek to learn intuitively means that we do not want fast convergence." Why don't we want faster convergence. * In Figure 6 left, it looks like DT-L is worse than DT-R? Why is that? More stability leads to worse performance? * What about DT and DT-R for TSP? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors cleraly addresses limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time to read the paper and for raising some important points in regards to our submission that we agree should be addressed. ### Response to Weaknesses > "The idea is quite straight-forward (may not be a bad thing, but make > technical contributions smaller)" We believe that imposing a sub-Lipschitz-1 constraint provides guarantees about the behaviour of DT-L architecture that previous models lacked (e.g. we are guaranteed that the recurrence will converge to a unique fixed-point). We strongly believe this opens up avenues to future improves that would not be possible otherwise. Finally, if the reviewer was seeking technical innovation we would point to our TSP model that involve a number of novel tricks that we are quite happy about (this is incidental to where we believe the main contribution of the paper lies so we have not emphasized these). > "In the TSP problems, DT-L's results seem worse than NN Tours and BNN Tours. > At least some explanation is warranted." We have responded to this comment in the response to all authors. We fully agree that further explanation is warranted. We had understated the difficulty of solving TSP using a convolutional algorithm that only gets to see a small part of the distance matrix. > "I'm not fully convinced by the significance of this paper. The examples > shown in the paper are quite toy. Are there more examples you expect DT-L > would work?" In this paper we deliberately chose to make DT-L as close to DT-R as possible to allow easy comparison and to focus attention on the benefits of imposing the sub-Lipschitz-1 constraint. However, this constraint (with its theoretical underpinnings) allows a whole host of modifications to deep thinking type architectures to be developed that will train robustly. We attempted to demonstrate this by learning an algorithm to tackle random TSP instances. We believe (and have strong evidence based on work conducted after submitting this paper) that there is considerable potential to tackle many tasks considered 'real' world. Given the page constraints of the paper, and a desire not to make this paper any more complicated, we feel going into this is outside the scope of this work. > "I'd appreciate more visualizations that can intuitively show the benefits of > DT-L over DT/DT-R? Maybe some Figures like in the original DT paper." The main benefits of DT-L are the stability it provides, especially in the case of smaller models. The focus is not on the algorithms it may learn or any attempt at interpreting them. While these are interesting visualizations to produce, this is not the topic of the paper and we believe including such visualizations may detract from the main contribution. > "The title is not very informative. Might be better to mention Lipschitz > smoothness in the title." We agree with the reviewer that the title is not very informative, and are pleased to incorporate the suggestion given by amending the title to Rethinking Deep Thinking: Stable Learning of Algorithms using Lipschitz Constraints. ### Answers to Questions > "In Line 141-142, I don't quite get this comment "Although any Lipschitz > constant less than 1 would guarantee convergence, the nature of the problem > solving mechanism we seek to learn intuitively means that we do not want fast > convergence." Why don't we want faster convergence." This is a good question, and something we will clarify in the final version of the paper. Our reasoning behind promoting slower convergence is to obviate the learning of trivial solutions to the given problem. As the model is trained on simple problems, we need to avoid it finding a simple solution (e.g. a shortcut, or memorisation) that does not generalise to larger problems. We re-emphasize that when we refer to convergence here, we are talking about the run-time convergence of the learned algorithm, not the training convergence (which we clearly do want to be fast and stable - something that our proposed solution also offers). Recall that we seek a solution using convolutions that only have access in a single iteration to information in a limited field of view. To obtain good solutions requires information across the whole problem instance. This is why to solve larger instances we need to run for more iterations. Thus, allowing the solution to converge slowly (which we achieve by making the Lipschitz constant close to 1) prevents the network finding a sub-optimal solution requiring only local information. > "In Figure 6 left, it looks like DT-L is worse than DT-R? Why is that? More > stability leads to worse performance?" It is intriguing why DT-L performs slightly worse than DT-R on the chess problem and whether we pay a price for stability. It may well be that adding an additional constraint leads to a slight decrease in performance, however, these networks are sufficiently complex that we are reluctant to come to any firm conclusions without a lot of evidence. As an aside, we can easily modify DT-L networks to get better performance than the DT-R network. We have not put in these results as it is not a fair comparison (the network architectures are slightly different). To make a fair comparison we should also optimally modify DT-R, but DT-R is far harder to modify as many modifications lead to networks that don't train at all. > "What about DT and DT-R for TSP?" TSP is such a complex problem that getting the DT-R model to learn (rather than consistently diverge) is a huge challenge. Of course, there is a huge hyper-parameter space to explore where it is possible that DT-R might work and we will endeavour in the appendix of the final version to add results for DT-R if we are ever able to train it (experiments underway at the moment). Realistically DT is highly unlikely to ever solve this problem (DT-R dominates DT on almost all problems). --- Rebuttal Comment 1.1: Comment: Thanks for addressing my concerns, I'll raise my score to 6.	Summary: This paper identifies and rectifies an issue with a particular type of iterative neural network called Deep Thinking Networks. The problem arises in exploding latent representations and unstable training routines. The authors of this work propose an update to the architecture where they add Lipschitz constraints to the model. They show three major benefits: (I) The models train more stably/predictably; (II) the inference-time behavior is better as the latent representations converge with iterations of the recurrent model; and (III) this new approach can learn how to solve NP-Hard problems where the old methods fail. Strengths: 1. This paper is original to my knowledge. I am aware of much of the work on Deep Thinking Networks and the issues raised and the solutions proposed in this work are novel. 1. The quality of the work is high. For the most part the experiments are done well and cover many natural questions that would arise from reading the abstract/intro. 1. The clarity is good. I think the writing is clear and the results are compelling. 1. The results are significant for those interested in easy-to-hard generalization. These Deep Thinking Networks have strong extrapolation of toy problems and with the proposed updates to the methods they show strong performance even for TSP solving. Weaknesses: 1. Clarity: A couple things could be more clear. i. I think IPT stands for Incremental Progress Training, but I don't see the acronym defined anywhere. ii. Table 1 the units are unclear. I gather there are tour lengths, but that isn't stated in the table or the caption. iii. The violin plot in Figure 2 is hard to parse (no harder than any other violin plot). This type of graphic does look nice, but offers little quantitative context. For example, there is no indication of the units/scale of the width of each violin. This is not the right type of plot for a conference paper. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Can the authors make the clarifications needed to address my first two points in the Weaknesses section? 1. Have the authors looked at transformer architectures at all? I'm not asking for results to be added to the paper, but I'm curious about how these techniques, which are independent from the parameterization of any given layer in some ways, might apply to modern large model architectures. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Yes, the limitations are adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their support and helpful suggestions, and are glad they are as excited about this direction of research as we are. ### Response to Weaknesses > "Clarity: A couple things could be more clear. > i. I think IPT stands for Incremental Progress Training, but I don't see the > acronym defined anywhere. > ii. Table 1 the units are unclear. I gather there are tour lengths, but that > isn't stated in the table or the caption. > iii. The violin plot in Figure 2 is hard to parse (no harder than any other > violin plot). This type of graphic does look nice, but offers little > quantitative context. For example, there is no indication of the units/scale > of the width of each violin. This is not the right type of plot for a > conference paper." We have updated the paper to appropriately define 'incremental progress training' (IPT) and have updated the caption of Table 1 to specify that the DT-L column consists of tour lengths. These were both oversights on our behalf, and we thank the reviewer for pointing these out. The violin plot (Figure 2) will be updated with a different graphic which better shows the distribution of singular values in reshaped kernel weights, providing clearer quantitative insight. We have provided an alternative graphic (Figure R1) in the PDF uploaded to the overall rebuttal, and trust that you find this more appropriate? ### Answers to Questions > "Can the authors make the clarifications needed to address my first two > points in the Weaknesses section?" See above. > "Have the authors looked at transformer architectures at all? I'm not asking > for results to be added to the paper, but I'm curious about how these > techniques, which are independent from the parameterization of any given layer > in some ways, might apply to modern large model architectures." Yes, we have considered this. We believe it is a challenging problem that is yet to be solved and have started to give it some thought. --- Rebuttal Comment 1.1: Title: Response to authors Comment: Thank you for addressing my points. I'll maintain my score.	Summary: The paper addresses the positive feedback issue in the so called Deep Thinking networks, where the inference computation may involve more recurrent computations than encountered in training. The proposed solution is to normalise the state vector that undergoes the recurrence, i.e. make the mapping contractive, i.e. ensure negative (but just) feedback. Strengths: The paper is well written and clear to follow, the proposed method is pretty straight forward and effective. Weaknesses: As far as I can tell, it is pretty straight forward control theory stuff for addressing positive feedback. Nothing wrong with the proposed solution, but I would assume this is such a fundamentally well known issue in any recurrent/feedback system that we can leave this to be addressed by the designer at implementation time with any choice of normalisation. It is somewhat disappointing that with the proposed method there is still the need for batch normalisation. Technical Quality: 3 Clarity: 3 Questions for Authors: Does batch normalisation alone not do a good job of stabilising the feedback? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: If I understand this correctly, the proposed normalisation creates vanishing gradient problem, but authors seem to be aware of this and address it by keeping the spectral norm close to 1. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments. ### Response to Weaknesses > "As far as I can tell, it is pretty straight forward control theory stuff for > addressing positive feedback. Nothing wrong with the proposed solution, but I > would assume this is such a fundamentally well known issue in any > recurrent/feedback system that we can leave this to be addressed by the > designer at implementation time with any choice of normalisation. It is > somewhat disappointing that with the proposed method there is still the need > for batch normalisation." We agree that it is well-known that any learned recurrent system can explode or vanish. The most prominent method for solving this is to have memory cells that are mainly conserved as exemplified by LSTMs and their variants. This mechanism does not fit the DT architecture. Imposing a sub-Lipschitz-1 constraint to ensure that the recurrence reaches a unique solution is to the best of our knowledge the first time this has been done in the context of recurrent networks (and certainly in the case of deep thinking networks). It is, in our experience, somewhat non-trivial to ensure this constraint holds in a network with learnable parameters. As shown by the previous DT papers, ensuring Lipschitz-1 behavior is not necessary for the models to work, although it makes the model much more fragile (i.e. it often fails to learn). The contribution of this paper is showing that imposing a Lipschitz constraint cures this problem. There are a number of unexpected consequences of doing this. For example, we show we are able to reliably solve the problems with many fewer parameters (often by orders of magnitude). The added stability makes it far easier to work with these models, to the extent that we were able to get these models to find an algorithm for solving TSP. Thus, in our view, the paper shows some unexpected consequences of controlling the norm which was not so well known that it had been implemented in previous models. As discussed below, batch normalization is not used in the recurrent section of the network (network $\mathcal{G}$) and, in fact, it causes problem when it is added (perhaps illustrating that controlling the norm is less trivial than it may appear). ### Answers to Questions > "Does batch normalization alone not do a good job of stabilising the > feedback?" As stated in the paper, the batch norm is not used in the recurrent part of the model. Instead, batch normalization is only used outside of the recurrent part. If the reviewer believes this is unclear in the current version, we can update it to be clearer about where in particular batch normalization is used. Our reasons for not using batch normalization in recurrence follows from existing literature, particularly 'Residual Connections Encourage Iterative Inference' (Jastrzebski et al., 2018, DOI: 10.48550/arXiv.1710.04773) where _unsharing_ batch normalization statistics was necessary to avoid activations exploding. In an architecture where the maximum number of iterations is unknown (and potentially large in the cases of large prefix sum and maze problems), it appears infeasible to unshare batch normalization statistics for every iteration. Nonetheless, we look forward to future research that may resolve this issue. It is also worth mentioning that batch norm does not ensure that the model is constrained to be sub-Lipschitz-1, which is another reason we have not used it. --- Rebuttal Comment 1.1: Title: Response to authors Comment: Thanks for answering my questions, and the clarification of where batch norm fits in the proposed solution. And as a result I grant that this work should be judged more on the way it tackles the straight forward/obvious approach (so that the model is stable and still capable of learning) as opposed to the fact that it uses the obvious approach... and so I am raising my score to 6.	Summary: The paper introduces Deep Thinking with Lipschitz Constraints (DT-L), an improved version of the Deep Thinking (DT) networks, designed to enhance the stability and performance of iterative algorithm learning models. The authors address the instability issues inherent in DT networks by analyzing intermediate representation growth and applying Lipschitz constraints. The DT-L model guarantees convergence to a unique solution and demonstrates robustness in learning algorithms that extrapolate to more complex problems. The paper furthermore benchmarks DT-L on the Traveling Salesperson Problem (TSP) as well other than the datasets used in the Deep Thinking models. It compares its performance against existing DT models. Strengths: - Introducing Lipschitz constraints into the DT framework enhances the models' reasoning capabilities. This approach addresses instability issues in training and inference, offering theoretical guarantees for convergence. - DT-L demonstrates the ability to scale to larger problems effectively, maintaining stability and performance, which is crucial for real-world applications. - The comprehensive evaluation on various problem classes, including prefix sums, mazes, chess puzzles, and TSP, highlights the robustness and versatility of the DT-L model. - The paper provides a thorough analysis of the issues with DT networks and clearly explains how the proposed modifications address these problems. Weaknesses: - The modifications and theoretical underpinnings of the DT-L model, such as the Lipschitz constraints and orthogonal transformations, add complexity to the model, which might hinder its adoption and understanding by a broader audience. - While the DT-L model shows improvement, its performance on the TSP is not impressive, indicating room for further optimization and refinement. Technical Quality: 3 Clarity: 3 Questions for Authors: - How does the introduction of Lipschitz constraints impact the computational complexity and training time of the DT-L model compared to traditional DT models? - Can the proposed DT-L model be extended to other types of iterative algorithms beyond the ones tested in this paper? If so, what modifications would be necessary? - Can this applied for transformer architectures like looped transformers? - Can the insights gained from this work be applied to improve the interpretability of the learned algorithms, making the decision-making process of the DT-L model more transparent? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive and constructive comments. ### Response to Weaknesses > "The modifications and theoretical underpinnings of the DT-L model, such as > the Lipschitz constraints and orthogonal transformations, add complexity to > the model, which might hinder its adoption and understanding by a broader > audience." The main contribution of the paper in our view is that by imposing a Lipschitz constraint the recurrence is guaranteed to converge leading to much more stable training. We believe that this is conceptually simple enough and provides sufficient benefits that it will be readily adopted by the broader community. However, as you have acknowledged there is complexity to this. The orthogonal transformation is only used for the TSP-solving variant of DT-L. The complexity of solving TSP and ensuring that the tour constraint is met is sufficiently difficult that whenever tackling TSP some technical complexity is inevitable. For solving the other problems the orthogonal transformation is not used and the network design was chosen to be as close to that of DT-R that we could make it. > "While the DT-L model shows improvement, its performance on the TSP is not > impressive, indicating room for further optimization and refinement." We have provided a response to this weakness as part of our response to all reviewers. We would emphasize that finding an algorithm capable of solving random instances of general TSP is a surprisingly challenging problem. ### Answers to Questions > "How does the introduction of Lipschitz constraints impact the computational > complexity and training time of the DT-L model compared to traditional DT > models?" The implementation we have used for spectral normalization computes the spectral norm (from cached power iteration values) and divides the weights by this value for every weight access. This is computationally expensive, but is a simple addition to the model with PyTorch's parametrization features. Since the submission of this paper, a modification to our implementation of spectral normalization allows caching of these weights while maintaining gradient information. This has allowed a significant improvement to training speed, which we intend to show by adding an entry of improved training time to Table E1 for each DT-L model. We would also point out that due to the added stability we get by imposing the Lipschitz constraint we are able to tackle the test problems with significantly smaller networks than those used in the DT-R paper. This provides a considerable improvement in speed making training these networks a lot more efficient than the larger DT-R models. If we attempt to use smaller versions of the DT-R model we found that the model almost always failed to train properly. Finally, because of the increased stability we generally only have to train our model once to obtain a working solution. In contrast there are instances where we had to train DT-R over 20 times in order to find a solution that worked. > "Can the proposed DT-L model be extended to other types of iterative > algorithms beyond the ones tested in this paper? If so, what modifications > would be necessary?" We believe extensions to other types of problems and implementations are technically possible, but they go beyond the remit of this paper. In particular we see DT-L as being suitable for problems which naturally lend themselves to being solved by repeated convolutions, but extending the architecture to other problems is something we are considering and working towards. > "Can this applied for transformer architectures like looped transformers?" There are challenges to overcome in adapting transformer architectures to this application. It would be fantastic to see if one can create an iterative transformer with these constraints in the context given by the paper. We have been giving it some thought, but are not at the point where we have results. A transformer-based version of DT-L is out-of-scope, but this paper provides a critical stepping-stone towards this goal. > "Can the insights gained from this work be applied to improve the > interpretability of the learned algorithms, making the decision-making process > of the DT-L model more transparent?" This is a really good question, and something we're looking into. We're currently working on this since we believe guaranteeing uniqueness to a solution (from being a contractive mapping) should improve interpretability, but this is not work for this paper.	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their careful reading of the paper and their insightful comments. We are pleased that overall the reviewers found the paper clear, but we will integrate the helpful suggestions that have been made - thanks! We have responded to the reviewers individual comments separately, but there were a two general comments made by multiple reviewers that we address here. A number of reviewers commented on the performance of TSP. A previous criticism of deep thinking papers, one we don't share, is that the problems were cherry picked for the architecture. We chose TSP as a problem we knew to be notoriously hard, but fitted the requirement of a problem that we could scale up to a large size to create a harder problem. Any implementation of a TSP solver is challenging because of the tour constraint. We wanted to demonstrate that DT-L was sufficiently robust that it could find an algorithm for solving TSP. To increase the challenge we chose non-symmetric (and obviously non-Euclidean) TSP, and additionally we do not provide the optimal solution during training, utilizing only a loss that penalizes tour length. In this case many of the classic heuristics fail. For example, k-OPT which is widely used in hill-climbing type solvers are inefficient as they typically involve reversing part of the tour which completely changes the path length. Other common heuristics used for Euclidean TSP (such as locality of cities) also fail for non-Euclidean problem instances. Thus, nearest neighbor algorithms actually provide a stiff baseline for this problem. Note, however, that DT-L has only a local view of the distance matrix as it is using convolutions. Nearest neighbor in contrast is a constructive algorithm that requires global information of the distance matrix (it has to find the closest city that has not already been included in the tour). Thus, although the results we obtain on TSP may appear disappointing at first view we believe that the fact that they are considerably better than random tours shows DT-L's ability to learn non-trivial algorithms. Finally, we should mention that for comparison with previous DT models we have deliberately stuck to the same architecture as DT-R as much as possible (although, admitted we needed to make some modification to get TSP to learn at all). There are additional modification such as increasing the number of layers in the recurrent section of the network that leads to improved performance on TSP, but we have not included these in the current paper as we believe presenting too many modifications to the network would obscure the main contribution of adding the Lipschitz constraint. We do agree with the reviewers that a comment on the TSP results would be useful and we will modify our text to include this. The reviewers mentioned extending to transformers and other tasks. We are looking forward to exploring this direction in future work, but this is not the focus of this paper. The ideas presented here are however critical to actually extending deep thinking style _recurrent_ networks to other tasks. Our work also shows what design constraints of the recurrent architecture are required to ensure convergence, leading us to be able to propose new architectures in a principled way in the future. We have expanded more in the responses to individual reviewers. Pdf: /pdf/9acdb0d7def6b6fbf96cf8d4c02e738d56d198e7.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Unsupervised Homography Estimation on Multimodal Image Pair via Alternating Optimization	Accept (poster)	Summary: The paper proposes an unsupervised homography estimation method for multimodal image pairs using an alternating optimization approach. The claimed key innovation is the introduction of the Geometry Barlow Twins loss function for the alternating optimization. The authors show that their approach works on 3 multimodal datasets and different homography estimation architecutres. Strengths: The alternating optimization framework together with Geometry Barlow Twins loss seem to be a fresh perspective on unsupervised multimodal homography estimation. Weaknesses: Weaknesses 1. Discussion on the Feasibility and Rationality of the Proposed Method: First, for unsupervised training of networks based on iterative prediction, such as RAFT, to ensure stability during training, related methods [1-2] typically apply some form of direct supervision to the motion predicted by the network. This is different from the approach proposed in this paper, which only uses the Geometry Barlow Twins loss for brightness supervision. Second, how RAFT can be used for homography estimation should also be explained, because it is designed for optical flow estimation. Moreover, the paper does not explain how the proposed Geometry Barlow Twins loss supervises the intermediate stages of iterative prediction, whereas RAFT, IHN, and RHWF, along with methods leveraging their structures [1-2], generally provide details on their supervision mechanisms on the intermediate stages. This raises concerns about the feasibility of the proposed supervision method in this paper. Additionally, the effectiveness of the Modality-Agnostic Representation Learning (MARL) introduced in section 4.3 is questionable because it lacks spatial information in its supervision. As mentioned in section 3.2, the projector removes spatial information from the feature maps. The authors should provide a convincing and thorough explanation for these issues. 2. Doubt about the Effectiveness of the Proposed Method: For example, the paper proposes the alternating optimization (AltO) method but does not provide sufficient experimental results to demonstrate its superiority over other strategies, such as directly cascading all the modules. Furthermore, the paper lacks a comparative demonstration of the features extracted with and without the MARL phase, making the advantages of introducing this phase less convincing. 3. Insufficient Experimental Validation: The paper conducts experiments on only 3 cross-modal datasets, among which only the GoogleMap dataset exhibits significant modality differences. The GoogleEarth dataset mainly consists of images taken in different seasons [3]. Part of the DeepIR dataset is simulated multispectral data [4], which will significantly reduce the difficulty of homography estimation. It would be beneficial to conduct experiments on more challenging multimodal datasets, such as those involving VIS-SAR modalities. [1] Stone, A., Maurer, D., Ayvaci, A., Angelova, A., & Jonschkowski, R. (2021). Smurf: Self-teaching multi-frame unsupervised raft with full-image warping. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition (pp. 3887-3896). [2] Liang, Y., Liu, J., Zhang, D., & Fu, Y. (2023). Mpi-flow: Learning realistic optical flow with multiplane images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 13857-13868). [3] Zhao, Y., Huang, X., & Zhang, Z. (2021). Deep lucas-kanade homography for multimodal image alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 15950-15959). [4] Sa, I., Lim, J. Y., Ahn, H. S., & MacDonald, B. (2022). deepNIR: Datasets for generating synthetic NIR images and improved fruit detection system using deep learning techniques. Sensors, 22(13), 4721. Technical Quality: 2 Clarity: 2 Questions for Authors: Please refer to the Weaknesses. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The paper discusses the additional training cost arising from the inclusion of an additional module in the two-phase network, and explores potential solutions for addressing this issue in future research. However, the method’s generalization capabilities are not thoroughly explored, with experimental datasets limited to satellite images, maps, RGB, and NIR images. Future research could involve testing the method on a broader range of datasets to validate its generalization capabilities. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. We apologize for the lack of detailed explanations regarding the proposed method. Below are some additional clarifications: Weakness 1. Weakness 1.1. No other direct supervisions Increasing the similarity of local features between two inputs is sufficient and equivalent to estimating homography. This is because homography estimation assumes a linear transformation between pairs of static scenes. Therefore, smoothness is automatically satisfied, and global motion is equivalent to local motion and homography. Weakness 1.2. How is AltO applied to iterative frameworks such as RAFT, IHN, or RHWF? We used each prediction from the registration network to individually calculate the GBT loss. Then, all GBT losses at each time step were combined by weighted summation, similar to the supervised learning situation. The weights used are also borrowed from the supervised learning case. The only difference from the supervised case is that AltO is used instead of the ground truth. If we had not effectively utilized the full capability of the registration networks, we would not have been able to achieve such performance. In Figure R1 of the attached rebuttal PDF, there is an illustration of the explained details, so please refer to it. Weakness 1.3. How can RAFT be used for homography estimation? A motion flow map from an optical flow network, such as RAFT, can be regarded as HW-correspondences (H: height, W: width). If HW is 4 or larger, the Direct Linear Transform (DLT) algorithm can fit a single homography. This homography best approximates the flow map in terms of least squared error. Generally, in homography estimation, which has 8 degrees of freedom, most algorithms predict only 4 correspondences at the corners of images. They then convert these correspondences to a homography using DLT. Therefore, there is no reason HW-correspondences cannot be used for estimating homography. If you are concerned about secondary aspects of optical flow, such as smoothness or the correction between global camera motion and local motion, please refer to our response in Weakness 1.1. As we explained there, homography estimation is estimating a linear transformation between pairs of static images. Therefore, these issues do not arise. Weakness 1.4. What is the effectiveness of Modality-Agnostic Representation Learning (MARL)? It is to train the encoder that maps input images to the same feature space, not to train registration network directly. This goal allows our geometry loss to function properly regardless of modality difference. To achieve this goal, the loss term of MARL should be designed to enhance global similarity, rather than the local similarity of corresponding points, as is done with the GBT loss. This is why the global average pooling (GAP) layer should be included at the end of the projector. Table R2 shows the training results on Google Map dataset when GAP is not applied and spatial information is preserved to compute the similarity of local correspondences. It can be observed that proper training is not achieved in this case because local similarity is already met by the GBT loss, without considering global similarity. Table R2: Ablation results for including and excluding GAP. \| Method \| without GAP \| with GAP \| \|--------\|:-----------:\|:--------:\| \| DHN + AltO \| 24.07 \| 6.19 \| \| RAFT + AltO \| 24.07 \| 3.10 \| \| IHN + AltO \| 24.01 \| 3.06 \| \| RHWF + AltO \| 24.08 \| 3.49 \| Weakness 2. Effectiveness of the Proposed Method. (Why is alternating necessary?) To prevent unintended collaborations and a collapse into trivial solutions, we introduced an alternating training strategy. Training the encoder and registration network together end-to-end without alternating can cause the encoder to output a constant value and the registration network to always output the identity matrix as the homography. Alternating training isolates these modules. This ensures the encoder maps input images to the same feature space, and the registration network estimates the true homography between the two images. Table R1 in the Auther Rebuttal (above) shows the experimental results when alternating training is not applied, demonstrating that proper training does not occur. Weakness 3. & Limitation. Insufficient Experimental Validation You pointed out that the difficulty of the DeepNIR dataset is reduced due to the presence of simulated pairs. So we conducted experiments on the 'RGB-NIR' [1] dataset, which consists purely of RGB and NIR pairs without any synthetic images. We selected 'IHN' as the registration network, as it represents our primary model. The experimental results, as shown in Table R3 below, indicate that all methods performed better on this dataset compared to DeepNIR. This suggests that the DeepNIR dataset is harder. We believe this is because the DeepNIR dataset requires the model to learn two distributions (real NIR and synthetic NIR images), whereas the RGB-NIR dataset only requires learning a single distribution (real NIR). Additionally, we conducted experiments on the 'OS Dataset' [2], which is a VIS-SAR type dataset, as per your suggestion. The results are also presented in Table R3. In conclusion, our method outperformed other unsupervised methods on both of the additional datasets. Table R3: MACE evaluation on two datasets. \| Method \| RGB-NIR \| OS Dataset \| \|:-------\|:-------:\|:----------:\| \| IHN (+Supervision) \| 1.57 \| 6.33 \| \| UDHN \| 24.30 \| 32.93 \| \| CA-UDHN \| 24.27 \| 24.96 \| \| biHomE \| 25.18 \| 24.99 \| \| IHN + AltO \| 2.66 \| 14.12 \| [1] Matthew Brown et al., Multi-spectral sift for scene category recognition. CVPR 2011. [2] Xiang, Yuming, et al., Automatic registration of optical and sar images via improved phase congruency model, IEEE Journal of Earth Observations and Remote Sensing, 2020 --- Rebuttal 2: Title: Looking forward to your post-rebuttal comment! Comment: Dear Reviewer EsVS Thank you once again for participating in the review process and for providing such thoughtful feedback. We wanted to kindly follow up regarding the rebuttal we submitted. We understand that this is a busy time, and we greatly appreciate your efforts. To summarize our rebuttal to your review: * Provided detailed answers to all the questions you raised. * Conducted a new experiment to demonstrate the effectiveness of MARL and the global average pooling (GAP) layer. * Conducted a new experiment to demonstrate the effectiveness of alternating training. * Conducted new experiments on additional datasets to demonstrate the generalization capability of our AltO framework. If there are any further questions or clarifications needed from our side, we would be more than happy to provide them. We look forward to any feedback you might have and are eager to engage in any further discussion to improve our work. Thank you once again for your time and consideration. Best regards, Authors --- Rebuttal 3: Comment: Thanks to the authors for the rebuttal. I have read the rebuttal and other reviewers' comments. However, I still have the following concerns: (1) In Weakness 1, my concern is mainly about why the GBT loss can enable the homography network to converge well without direct supervision of global/local motion. In my view, the GBT loss can be seen as the 'cross-modal image intensity similarity' mentioned in SCPNet, which may lead to non-convergence. Is there any theoretical basis or proof to support this? (2) Figure R2 further raises my concern about the experimental results. The visualized feature maps are not similar at all. They barely retain the structural information of the original images and contain many artifacts. I am not convinced that such feature maps can effectively supervise the registration network. What's more, in Table 2 in the paper, the author claims that simply using MSE on such feature maps (the mean of squared error of the intensity of the feature maps in Figure R2) as the Geometry loss can produce a relatively accurate result, which also reduces the credibility of the experiments. (3) Insufficient references and comparison experiments. As mentioned by reviewer sFim, this paper did not discuss many references. Moreover, comparison experiments with other methods should be conducted on all datasets to demonstrate the effectiveness of AltO, instead of only on Google Maps. (4) The homography estimation accuracy on the OS Dataset is unsatisfactory with the MACE of 14.12, which may not be regarded as a converged training. The effectiveness of the proposed method is insufficient on such cross-modal dataset. For the reasons mentioned above, I am still inclined to reject. --- Rebuttal Comment 3.1: Comment: Thank you for your continued attention to our work. We would like to address the remaining concerns and the additional questions you have raised. (1) Due to the encoder’s mapping, the GBT loss is calculated under uni-modal conditions in the feature space, not cross-modal conditions in the image space. Although the feature spaces might not perfectly match, a few corresponding regions(or 4-corresponding points) are sufficient to estimate the homography. This is evident in the feature map in Figure R2, where a few corresponding regions are strongly activated. Additionally, experiments (Table R1) show that GBT alone, without MARL, fails because the encoder doesn't learn the mapping without MARL. (2) The purpose of the geometry loss is not to make the features identical. Even if the given pair of feature maps differ somewhat, the goal is to estimate the homography by matching the most strongly activated corresponding points. The encoder plays a crucial role in this process, ensuring that these key points are strongly activated in both input images, which makes the geometry loss convex at these points. Without the encoder, achieving convexity would be difficult when the modalities differ. Since the feature maps are not perfectly identical, as seen in Figure R2, the lower bound (bias) of the geometry loss may be higher. However, if at least 4-corresponding points are well-matched, the homography can still be accurately estimated. The same applies when using MSE: as long as a few strongly activated local regions are well matched, the homography can be accurately estimated even if the two feature maps are not identical. Furthermore, while MSE achieves its minimum when the two feature maps are exactly the same, the GBT loss, being based on normalized similarity, can reach its minimum as long as the trends are similar. This property makes the GBT loss more suitable for this scenario, as shown in the ablation study of our paper. (3) Conducting experiments for all possible cases requires significant time, so we focused on the most meaningful ones. As you mentioned, the Google Map dataset has a significant modality gap, which is why we prioritized it. However, we agree that additional experiments would strengthen our paper. (4) Since the OS dataset is completely new to us, the existing hyperparameters may not have been suitable. Additionally, our proposed method is a learning framework, so improvements could be made through architecture exploration of the encoder and projector, but this has not been done yet for the OS dataset. Therefore, the absolute performance may appear lower than on other datasets. Despite these challenging conditions, our method shows better performance than other baselines, further demonstrating its tendency to converge and its overall feasibility. We apologize for the lack of detailed explanation regarding point (2) in our paper, and we hope this clarification is helpful. If you have any further questions or concerns, please feel free to ask. We would be more than happy to provide a thorough response to any additional inquiries. --- Rebuttal 4: Comment: (1) We apologize for the misunderstanding. It seems we may have overcomplicated your question. To clarify, let’s first look at UDHN [1], the pioneering unsupervised method in a same-modality scenario, which might address most of your concerns. UDHN predicts four corresponding points, then converts them into a homography using Direct Linear Transformation (DLT). The source image is then warped using the predicted homography, and a reconstruction loss or intensity-based similarity is calculated with the target image. This entire process, including DLT and warping, is differentiable and converges very well in a same-modality scenario. A similar approach is taken by biHomE [2], and both UDHN and biHomE are incorporated into our proposed method. Our method targets multimodal datasets. Therefore, we internally convert them to a uni-modal condition to apply an approach similar to UDHN or biHomE. To facilitate this multimodal-to-uni-modal conversion, we introduced MARL and alternating training. Thus, the GBT loss is calculated under uni-modal conditions thanks to the encoder's mapping. In the aforementioned processes, although methods like RANSAC are not used, the approaches work well. It seems you might be considering feature-based frameworks like SIFT, SURF, ORB or methods like LIFT [3] and Super Point [4], which replace parts of this process with deep learning. However, our method falls within the category of end-to-end homography estimation methods, like DHN [5], UDHN, and biHomE. These methods are composed entirely of differentiable processes for end-to-end learning. Rest assured, we are fully aware of the fundamentals of feature-based frameworks and understand practical know-how, such as: * If the corresponding points are clustered in one region or many lie on a straight line, the error in homography estimation increases significantly. * Conversely, the homography estimation becomes more accurate when the corresponding points are widely distributed across the entire image. * If only a few corresponding points are known, even with methods like RANSAC, the homography estimation error can increase due to outliers. Finally, regarding Figure R2, please consider that this feature map is simply an average across 128 channels. Since our GBT is based on the Pearson correlation coefficient, the strength level (bias) difference between the two feature maps is not important; only the similarity in trends between the two feature maps reduces the loss. Additionally, both MSE and GBT, it can allow a few strongly activated regions (not just single points, but multiple points) to accurately infer the overall homography. This can be empirically observed by looking at the regression values computed by the registration network. Again, unlike explicit methods like RANSAC, the registration network implicitly calculates these values through learning. [1] Ty Nguyen et al., Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics Autom. Lett., 2018. [2] Daniel Koguciuk et al., Perceptual loss for robust unsupervised homography estimation. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2021. [3] YI, Kwang Moo, et al. Lift: Learned invariant feature transform. In: Computer Vision–ECCV 2016 [4] DETONE, Daniel et al., Superpoint: Self-supervised interest point detection and description. In: CVPRW 2018 [5] Daniel DeTone et al., Deep image homography estimation. CoRR, abs/1606.03798, 2016. (2) We now understand that your main concern is whether the encoder's output can maintain geometric consistency. We agree that the lack of an explicit mechanism could be a potential weakness, as you mentioned. However, as shown in Table R2, removing the global average pooling layer is not an option. We think that further improvements may require additional processing directly on the encoder’s output to address this issue. Additionally, regarding the discussion on MSE, please refer to the last paragraph of (1). (3) We agree that the OS dataset is more multimodal than the Google Map dataset, so further research will need to focus on this dataset. Thank you for your detailed and ongoing discussion. --- Rebuttal Comment 4.1: Comment: The reasons provided by the authors are not convincing and have not addressed my concerns. Therefore, I maintain my rating. (1) The works cited by the authors, namely UDHN and biHomE, neither have successful precedents of training on cross-modal data nor iterative networks like IHN, RHWF, and RAFT, making their explanation unpersuasive. (2) The authors mentioned that their feature maps are averaged across 128 dimensions. First, I reviewed the original paper, and this detailed information was never mentioned. Second, to my understanding, averaging should weaken artifacts, contrary to what is shown in Figure R2, where there are many artifacts with poor consistency. (3) Furthermore, I have never requested a comparison with RANSAC-type algorithms. RANSAC was mentioned because the authors explained the reason for successful unsupervised network training by stating, "Although the feature spaces might not perfectly match, a few corresponding regions (or 4-corresponding points) are sufficient to estimate the homography." In my view, this explanation may have a few possibilities that make sense for RANSAC-type algorithms, but it is highly unreasonable and unfeasible for the cross-modal unsupervised homography training adopted by the authors, and they have not clarified this issue. --- Rebuttal 5: Comment: We regret that our explanation did not fully satisfy you, but we appreciate your engagement in the discussion until the end. (1) As mentioned, UDHN and biHomE assume a uni-modal scenario. Our main contribution lies in expanding this to a multimodal application by introducing MARL and alternating training. Regarding the iterative framework, we retained the original structure of RAFT, IHN, and RHWF, simply replacing the ground truth with AltO. This is why we did not emphasize it in the paper, but we demonstrated that it can be applied without significant issues through the main experiments. (2) Evaluations of the visualized feature maps are subjective, so we believe further discussion on this might not be productive. (3) In our previous comment, we mentioned that "it can allow a few strongly activated regions (not just single points, but multiple points) to accurately infer the overall homography." In other words, even with only a few distinctive regions in a solid image, the registration network can capture these and perform homography estimation relatively accurately.	Summary: This paper proposes a new unsupervised homography estimation approach for multimodal images. This method is designed as a two-phase optimization framework named AltO. The first phase named "Geometry Learning" trains a registration network to align the input multimodal images geometrically. The second phase named "Modality-Agnostic Representation Learning" trains an encoder and a projector to extract the image-level features invariant to modality changes. Experimental results demonstrate that AltO outperforms several existing unsupervised approaches on the multimodal registration datasets. Strengths: 1. The proposed framework is intuitive and interesting. This framework trains a registration network to align the input multimodal images geometrically, and trains another encoder to match the image-level features of the warped multimodal images. This framework has the potential to capture the pixel-level and image-level information in an unsupervised manner. 2. The organization and presentation of this paper are good. I think I can understand the core idea of this paper. Weaknesses: 1. Some central claims of this paper lack experimental evidence. 1.1 The "alternating" optimization framework is a central design in this paper. However, why is "alternating" optimization necessary? Will optimizing the "geometry loss" and "modality loss" simultaneously hurt performance? 1.2 The superiority of the proposed Geometry Barlow Twins (GBT) loss was not verified. The original Barlow Twins loss can be straightforwardly applied to the proposed model by considering both the spatial axis (indexed with "h,w") and batch axis (indexed with "n") as the batch dimension. This straightforward implementation should be compared with the proposed GBT loss. 1.3 The proposed approaches should be compared with some recent unsupervised approaches. Here are some approaches with released codes. [1] Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. [2] A Multiscale Framework with Unsupervised Learning for Remote Sensing Image Registration, IEEE Transactions on Geoscience and Remote Sensing, 2022. 2. This paper did not discuss the recent hand-crafted approaches for multimodal image registration. Many recent hand-crafted methods have been published in the top journals, so this kind of approach should not be ignored. The experiment should also compare the proposed approaches with the recent hand-crafted approaches. Here are some hand-crafted approaches with released code. [3] Histogram of the orientation of the weighted phase descriptor for multi-modal remote sensing image matching. ISPRS Journal of Photogrammetry and Remote Sensing, 2023. [4] POS-GIFT: A geometric and intensity-invariant feature transformation for multimodal images. Information Fusion, 2024. 3. The discussion of the motivation is not sufficient. The Introduction section mentioned some typical unsupervised approaches designed for the images from the same modality (e.g., UDHN and biHomE). However, the unsupervised approaches [2,5] designed for multimodal image registration are not discussed. What is the motivation of the proposed method compared with this kind of approach? [5] A Novel Coarse-to-Fine Deep Learning Registration Framework for Multi-Modal Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 2023. 4. This paper misses some references to hand-crafted and unsupervised approaches. I have listed some of them in the above weaknesses. The authors should further survey more papers and carefully revise the "Related Work" section. Technical Quality: 2 Clarity: 3 Questions for Authors: Please provide more discussions and experimental results to address the above weaknesses. Moreover, is the 3D reconstruction task related to "Homography Estimation" (line 21)? Generally, 3D reconstruction focuses on non-planar scenes, while homography estimation is designed for the planar scenes. Is there some literature that mentions the relationship between 3D reconstruction and homography estimation? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work in detail. Below are our responses to your comments and concerns. Weakness 1. Weakness 1.1. Why is alternating necessary? To prevent unintended collaborations and a collapse into trivial solutions, we introduced an alternating training strategy. Training the encoder and registration network together end-to-end without alternating can cause the encoder to output a constant value and the registration network to always output the identity matrix as the homography. Alternating training isolates these modules. This ensures the encoder maps input images to the same feature space, and the registration network estimates the true homography between the two images. Table R1 in Auther Rebuttal (above) shows the experimental results when alternating training is not applied, demonstrating that proper training does not occur. Weakness 1.2. What is the superiority of the proposed Geometry Barlow Twins (GBT)? GBT is robust in both cases: whether the i.i.d. assumption of dataset holds or not. The method you suggested, using NHW as a batch, matches GBT's performance only when the i.i.d. assumption is satisfied. In the Google Map dataset, where the i.i.d. assumption is met, both loss functions perform similarly (Table R2). However, in the case shown in Figure R3 of the rebuttal PDF, the i.i.d. assumption is not met, and GBT performs better. The key point is whether the distribution over NHW dimensions is similar to that over HW dimensions. If the i.i.d. assumption is not met, the distribution over N distorts NHW, leading to a distribution dissimilar compared to HW dimensions. Table R2: Comparison of two types of Geo. loss on the Google Map dataset. \| Method \| NHW as batch \| GBT \| \|-------:\|:------------:\|:---:\| \| DHN + AltO \| 5.35 \| 6.19 \| \| RAFT + AltO \| 3.55 \| 3.10 \| \| IHN + AltO \| 3.16 \| 3.06 \| \| RHWF + AltO \| 3.37 \| 3.49 \| Weakness 1.3. & Weakness 2. The proposed approaches should be compared with some recent unsupervised approaches and hand-crafted approaches. Sorry for the insufficient number of baselines. We have added more baselines to Table R3 below. Some of the entries in Table R3 cite SCPNet [1], a very recent paper accepted at ECCV 2024. Nonetheless, our method demonstrates even superior performance compared to SCPNet. Additionally, the paper [2] you mentioned is designed to address the 6-DOF (degree of freedom) problem, so it cannot be applied to our dataset, which has 8-DOF. Furthermore, the code for that paper is not fully available. Moreover, we also attempted to run the code for paper [3], but encountered errors with the sample images, so we could not include it in our results. Table R3: MACE evaluation results on the Google Map dataset. * indicates values reported from SCPNet [1]. \| Type \| Method \| MACE \| \|:----:\|:-------\|:----:\| \| Hand-crafted \| SIFT \| 24.53* \| \| \| ORB \| 24.52* \| \| \| DASC \| 21.76* \| \| \| RIFT \| 16.55* \| \| \| POS-GIFT \| 20.90 \| \| Unsupervised \| UDHN \| 28.58 \| \| \| CA-UDHN \| 24.00 \| \| \| biHomE \| 24.08 \| \| \| BasesHomo \| 24.49* \| \| \| UMF-CMGR \| 24.60* \| \| \| SCPNet \| 4.35* \| \| \| IHN + AltO (Ours) \| 3.06 \| [1] Runmin Zhang et al., SCPNet: Unsupervised Cross-modal Homography Estimation via Intra-modal Self-supervised Learning, (ECCV 2024 Accepted.) [2] Yuanxin Ye et al., A Multiscale Framework with Unsupervised Learning for Remote Sensing Image Registration, IEEE Transactions on Geoscience and Remote Sensing, 2022. [3] ZHANG et al., Histogram of the orientation of the weighted phase descriptor for multi-modal remote sensing image matching. ISPRS Journal of Photogrammetry and Remote Sensing, 2023. Weakness 3. The discussion of the motivation is not sufficient. The problem we aim to solve is the unsupervised learning of 8-DOF homography estimation for multimodal image pairs. These conditions are very common in many fields, such as industry. For example, matching a real photo to a CAD image. At the time of our research, we found very few papers that directly addressed this specific problem. This scarcity of relevant work highlighted the difficulty and value of addressing this challenge, which motivated us to pursue it. The papers you suggested, [2] and [4], do not directly apply: [2] focuses on a 6-DOF problem, and [4] employs a supervised learning approach. [4] A Novel Coarse-to-Fine Deep Learning Registration Framework for Multi-Modal Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 2023. Weakness 4. This paper misses some references to hand-crafted and unsupervised approaches. We apologize for missing them. We will include and refer them during the revision. Question 1. Is there some literature that mentions the relationship between 3D reconstruction and homography estimation? Below are several documents regarding your question. We will revise the citation section related to 3D reconstruction in our paper. * Zhang, Zhongfei, and Allen R. Hanson. "3D reconstruction based on homography mapping." Proc. ARPA96 (1996): 1007-1012. * Mei, Christopher, et al. "Efficient homography-based tracking and 3-D reconstruction for single-viewpoint sensors." IEEE Transactions on Robotics 24.6 (2008): 1352-1364. * https://github.com/ziliHarvey/Homographies-for-Plane-Detection-and-3D-Reconstruction * Yong-In, Yoon, and Ohk Hyung-Soo. "3D reconstruction using the planar homograpy." The Journal of Korean Institute of Communications and Information Sciences 31.4C (2006): 381-390. * Zhang, Beiwei, and Y. F. Li. "An efficient method for dynamic calibration and 3D reconstruction using homographic transformation." Sensors and Actuators A: Physical 119.2 (2005): 349-357. * Dubrofsky, Elan. "Homography estimation." Diplomová práce. Vancouver: Univerzita Britské Kolumbie 5 (2009). --- Rebuttal Comment 1.1: Comment: Thank the authors for the responses. The additional experimental results and discussions address my main concerns. In my opinion, the ablation study about the alternating optimization and the comparison with the recent approaches make the proposed approach's superiorities more convincing. Therefore, I’d like to raise my rating from "Weak Accept" to "Accept". It would be better to discuss the following two problems further. 1. In the response to W1.1, the authors pointed out that "Training the encoder and registration network together end-to-end without alternating can cause the encoder to output a constant value and the registration network to always output the identity matrix as the homography." It would be better to provide some intuitive explanations. Why does the end-to-end training make the encoder/registration network tend to output a constant value/identity matrix? 2. In the response to W3, the authors claimed that this paper’s motivation is different from the literature [2] because the proposed approach considers the 8-DOF homography estimation while the method [2] focuses on a 6-DOF problem. However, such a difference seems to be minor because both the two methods utilize learnable regression models. It is straightforward to extend the method [2] to handle 8-DOF homography estimation. More intrinsic differences should be discussed to highlight the motivation of this paper. --- Rebuttal 2: Title: Looking forward to your post-rebuttal comment! Comment: Dear Reviewer sFim Thank you once again for participating in the review process and for providing such thoughtful feedback. We wanted to kindly follow up regarding the rebuttal we submitted. We understand that this is a busy time, and we greatly appreciate your efforts. To summarize our rebuttal to your review: * Conducted a new experiment to demonstrate the effectiveness of alternating training and MARL. * Conducted a new experiment and presented a case to compare GBT with the new method you suggested. * Added more baselines for performance evaluation by conducting a new experiment and incorporating reported values. * Provided detailed answers to all the questions you raised. If there are any further questions or clarifications needed from our side, we would be more than happy to provide them. We look forward to any feedback you might have and are eager to engage in any further discussion to improve our work. Thank you once again for your time and consideration. Best regards, Authors --- Rebuttal 3: Title: Thank you for increasing your score! Comment: Thank you very much for raising the rating of our paper. We appreciate your recognition of the additional experiments and discussions we provided, and we are glad that they addressed your main concerns. We would like to provide further clarification: 1. Intuitive Explanation for End-to-End Training without Alternating: If the encoder and registration network collapse as we mentioned, then all reconstruction-based losses, including L1-loss or our GBT loss, would become zero, the minimum possible value for such losses. This occurs because the encoder's outputs are constant, resulting in the difference or similarity always converging to a trivial value, regardless of the homography. Consequently, the registration network would not make any effort to find the true homography, outputs only identity matrix. 2. Difference Between Our Approach and Literature [2]: As you mentioned, extending from 6-DOF to 8-DOF is indeed straightforward, but it is a more challenging task, and some performance degradation is to be expected. The CFOG, used in [2], seems to be a key element. The multi-scale framework is aimed at more precise estimation, but it does not address different modalities. Additionally, since the code is not fully available at the moment, quickly reproducing and modifying it is difficult. However, we fully agree that the extended version of [2] would certainly be a valuable baseline to consider. We hope these clarifications address your questions, and we are more than happy to discuss further if needed. Thank you once again for your valuable insights and support.	Summary: The paper addresses unsupervised homography estimation from multi-modal image pairs. The authors propose to cope with the issue of 1) modality, 2) registration in two distinct networks that are trained in an interleaved fashion. The networks architecture derives from the Barlow Twins framework, with changes in the loss function. Results are illustrated on several public benchmark of small images (128^2) and compares favorably wrt to related unsupervised approach. Strengths: 1- I enjoy reading the paper. I walked through the paper, first with curiosity and skepticism, then with strong interest. The approach is intuitive (adjust the two representations then compute the transformation) and compelling. I am somehow surprised that it works :) The constrastive-like loss used in Barlow Twins is contributing much for the network to learn the correct solution. 2- Overall, the authors are tackling an important problem (unsupervised learning) for which an original solution is proposed --while based on previous recent work. The methodology is clearly presented. Results are convincing (thought only on small images 128x128) and illustrated on various modality pairs. Quantitative results show improvement wrt related unsupervised work Weaknesses: 1- Not a weakness, but a points which could have been discussed: why not simply transforming the inputs into edge maps before learning a matching/homography function (and putting aside the modality discrepancy). It would not be a very fancy approach, but I believe it could be a baseline for comparison. 2- The approach would be more convincing if each of the two modules (GL and MARL) had demonstrated their effectiveness also individually (ie same image pair modality using only GL). Technical Quality: 3 Clarity: 3 Questions for Authors: - What is the size of the embedding? What is the training time? Are the Barlow Twins trained from scratch? - Illustration seems to show strong geometric features (ie lines) in the input images. Is it a strong limitation of the approach? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: From a practical point of view, the size of the image and the strong overlap between the pairs show that the work need to be further developped for applications at full scale. From a methodology point of view, the authors have discussed the limitation of having two networks trained in an interleaved way, with potential collision or collapse. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and for taking the time to provide your feedback. Below is our rebuttal. Weakness 1. Why not use an edge-based approach as a baseline? Edge-based approaches have limitations when used as baselines with different modalities. When dealing with two images of different modalities, one image often has rich edges while the other does not. Even within the closed modality, such as day and night photos, edge differences can arise due to shadows. Thus, converting both images to edge maps and comparing them might not achieve high performance. Although paper [1] utilizes edges in its learning process, it is limited to using image pairs of exactly the same modality. Below Table R2 shows the results of applying UDHN [2], a simple unsupervised method that assumes the same modality, to edge maps converted from the Google Map dataset. The results indicate poor performance. Table R2: Evaluation result of applying edge maps to UDHN \| Method \| MACE (Google Map) \| \|--------\|:----:\| \| IHN + AltO \| 3.06 \| \| UDHN \| 28.58 \| \| UDHN + Edge Map + L1 loss \| 24.15 \| \| UDHN + Edge Map + cos loss \| 24.00 \| [1] Feng, Xiaomei & Jia et al., Edge-Aware Correlation Learning for Unsupervised Progressive Homography Estimation, in IEEE Transactions on Circuits and Systems for Video Technology, 2023. [2] Ty Nguyen et al., Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics Autom. Lett., 2018. Weakness 2. Demonstrate the effectiveness of each modules, GL and MARL. The GL module enhances geometry alignment by increasing the similarity between local correspondences. It is inspired by biHomE [3], with the only differences being that biHomE uses an ImageNet-pretrained encoder which is frozen, and applies triplet loss as the loss term. The effectiveness of this approach has already been demonstrated in the paper [3] using the S-COCO dataset, which has the same modality. The role of the MARL module is to train the encoder to map input images to the same feature space. This properly trained encoder allows our geometry loss to function correctly regardless of modality differences. Table R1 in the Author Rebuttal (above) shows the results of training on the Google Map dataset without the MARL module or with the MARL module but no alternating. The result indicates that proper learning did not occur. In summary, for successful training, the GL module, MARL module, and alternating training are all essential components. [3] Daniel Koguciuk et al., Perceptual loss for robust unsupervised homography estimation. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2021. Question 1. Question 1.1. What is the size of the embedding? AltO serves as a learning framework that does not impose restrictions on the internal structure of each component. However, the encoder and projector used as examples in our paper are based on ResNet-34. The encoder uses the first two stages of ResNet-34, resulting in an embedding size of 128. Nevertheless, these can be flexibly replaced if needed. Question 1.2. What is the training time? Table R3 below shows the training time measurements on Google Map dataset. Although the training time increases, the inference time at run-time is the same as that of supervised learning since the AltO module is not needed. Table R3: Training times for each registration network. (Nvidia RTX 8000) \| \| Supervised. \| Unsupervised (with AltO) \| \|:-----:\|:-----------:\|:------------------------:\| \|DHN \| 1h 26m \| 4h 31m \| \|IHN \| 3h 27m \| 9h 57m \| Question 1.3. Are the Barlow Twins trained from scratch? There are no pretrained components. All layers are trained from scratch. Question 2. Are strong geometric features (i.e., lines) necessary for the proposed method? It is evident that lines, such as edges, are helpful. However, as visualized in Figure R2 of the rebuttal PDF, the output of the encoder shows strong features in the form of blobs rather than edges. This indicates that not only lines but also blobs play a significant role in the proposed method. Limitation. Small image size and strong overlapped pairs The dataset settings we used (such as size and overlap) are based on the pioneering paper DHN [4] on end-to-end homography estimation. These settings have become the standard and are used by many studies, including [1-7]. For fair comparison, we followed these settings. However, AltO, as a learning framework, has no size or overlap restrictions. Constraints, if any, come from the registration network (e.g., IHN), encoder, or projector. Table R4 shows experiments where the Google Map dataset size and displacement were doubled, called 'Google Map x2.' We used DHN as the registration network, which is flexible with size constraints. The results show that DHN with AltO can reduce MACE even without any hyperparameter tuning. Table R4: MACE evaluation results on Google Map x2. \| Method \| MACE (Google Map x2) \| \|--------\|:----------------------:\| \| (No warping) \| 49.33 \| \| DHN + AltO \| 34.24 \| [4] Daniel DeTone et al., Deep image homography estimation. CoRR, abs/1606.03798, 2016. [5] Yiming Zhao et al., Deep lucas-kanade homography for multimodal image alignment. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021. [6] Si-Yuan Cao et al., Iterative deep homography estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022. [7] Si-Yuan Cao et al., Recurrent homography estimation using homography-guided image warping and focus transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023. --- Rebuttal 2: Title: Looking forward to your post-rebuttal comment! Comment: Dear Reviewer PrYW Thank you once again for participating in the review process and for providing such thoughtful feedback. We wanted to kindly follow up regarding the rebuttal we submitted. We understand that this is a busy time, and we greatly appreciate your efforts. To summarize our rebuttal to your review: * Conducted a new experiment to demonstrate the performance of the edge-based method in multimodal case. * Conducted a new experiment to show the effectiveness of alternating training and MARL. * Provided detailed answers to all the questions you raised. * Conducted a new experiment to demonstrate that our method, AltO, is also effective for larger image set. If there are any further questions or clarifications needed from our side, we would be more than happy to provide them. We look forward to any feedback you might have and are eager to engage in any further discussion to improve our work. Thank you once again for your time and consideration. Best regards, Authors --- Rebuttal 3: Title: Thanks to Reviewer PrYW Comment: We would like to thank you again for highly appreciating the strengths of our work. If you have any further questions or concerns, please feel free to ask. We would be more than happy to provide a thorough response to any additional inquiries. Best regards, Authors --- Rebuttal Comment 3.1: Comment: Thank you to the authors for the responses and clarifications, and overall rebuttal. - Comments regarding the use of edge maps make sense, I appreciate the additional experimental results which support the observations of the authors. - My question was about confirming the viability of each of the two networks (MARL and GL) by testing then independently: experiments that would be using same modality inputs (with some deformation) to evaluate GL independently of MARL. (I do not see how MARL could be tested independently of GL); Overall, the authors mainly addressed my concerned. Additional experiments have consolidated the work. I keep my score as accept. --- Reply to Comment 3.1.1: Comment: Thank you for your continued feedback. We noticed that there may have been a misunderstanding regarding one of the points you raised. We would like to clarify this and provide a more accurate response. We have conducted the following additional experiments using the MS-COCO dataset, which has the exact same modality. Due to the limited time, we used the simple DHN model as the registration network. In the case where only GL is used, the encoder was frozen from scratch. \| Method \| MACE \| \|------------\|:--------:\| \| (No warping ) \| 24.89 \| \| DHN + GL \| 5.75 \| \| DHN + MARL \| 25.00 \| In the case where only MARL is used, the spatial information is lost due to the global average pooling layer, leading to ineffective learning. We hope this response addresses your concerns. If you have any further questions, please feel free to ask. We will do our best to respond promptly. Thank you.	null	null	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for taking the time to thoroughly review our paper. Reviewers highlighted the strengths of our paper. + Proposed method is an interesting, intuitive, and fresh approach. (PrYW, sFim, EsVS) + The paper tackles an important problem and proposes an original solution. (PrYW) + Quantitative results show improvement. (PrYW) + The organization and presentation of this paper are clear and good. (PrYW, sFim) We will incorporate the reviewers' advice to further refine our research and improve the paper. Additionally, through the review process, we found that our method outperforms the very recent paper, SCPNet: Unsupervised Cross-modal Homography Estimation via Intra-modal Self-supervised Learning (ECCV 2024 accepted). The following experimental result was obtained using the Google Map dataset. * indicates values reported from SCPNet. (Lower is better.) \| Type \| Method \| MACE \| \|:----:\|:-------\|:----:\| \| Hand-crafted \| SIFT \| 24.53* \| \| \| ORB \| 24.52* \| \| \| DASC \| 21.76* \| \| \| RIFT \| 16.55* \| \| \| POS-GIFT \| 20.90 \| \| Unsupervised \| UDHN \| 28.58 \| \| \| CA-UDHN \| 24.00 \| \| \| biHomE \| 24.08 \| \| \| BasesHomo \| 24.49* \| \| \| UMF-CMGR \| 24.60* \| \| \| SCPNet \| 4.35* \| \| \| IHN + AltO (Ours) \| 3.06 \| -------------------------------------------------------------------------------------------------------------------------------- Note : * The attached PDF includes figures addressing each reviewer's comments. * Table R1 below shows the result of the experiment addressing the common question, 'Why is alternating necessary?' Table R1: Ablation result on the Google Map dataset for applying the alternating and MARL module. (Alt. is Alternating) \| Method \| No Alt.(with MARL) \| No Alt.(without MARL) \| Alt.(with MARL) \| \|:-----------:\|:----------------------:\|:-------------------------:\|:-------------------:\| \| DHN + AltO \| 24.09 \| 24.27 \| 6.19 \| \| RAFT + AltO \| 26.21 \| 25.91 \| 3.10 \| \| IHN + AltO \| 24.37 \| 23.14 \| 3.06 \| \| RHWF + AltO \| 18.88 \| 24.07 \| 3.49 \| Pdf: /pdf/1bb14a83dabf3829ac4007b709bb150aa8f934c3.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning	Accept (poster)	Summary: This paper proposes a new task, "generalized zero-shot learning (GZSL)," in which both seen and unseen objects should be recognized for vision-language tasks. It also proposes a new method based on CLIP that uses the loss in the "attribute space" to perform better in both seen and unseen classes. This method is evaluated on various kinds of data sets and evaluated by the harmonic mean of the accuracies of seen and unseen classes. Strengths: The proposed approach using the attribute space seems novel enough, and its effectiveness was verified by a detailed comparison of the other methods and the well-designed ablation studies. Weaknesses: 1) It is unclear what is learned in "learnable attribute tokens." It is not so beneficial for unseen classes. It is unclear what information is represented as tokens for seen classes. It may be better to analyze the acquired tokens in more detail. 2) It is difficult to think about the case that we have never seen an object, but we know its attributes quite well. In such a sense, I believe this method is more appropriate for few-shot learning. Technical Quality: 3 Clarity: 3 Questions for Authors: Strangely, the accuracy increases as the number of learnable tokes increases in Fig 4. (a) AwA2. It would be appreciated if you could provide any insights into this phenomenon. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: None. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Reviewer 8m4c ### Response Q1-Q3 Thank you for the positive and insightful comments. The reviewer appreciated the novelty of our approach, the creation of the attribute space, the effectiveness of our method, and the well-designed ablation studies. We address the mentioned concerns below. Q1: Analysis of the learned attribute tokens in more detail. A1: Thanks for the insightful comment. To better analyze the learned attribute tokens, we attach a PDF in the top-level comment with two figures (Fig. R1 and Fig. R2). Our analysis is two-fold: - In Fig. R1 of the attached top-level PDF file, we visualized the correlation matrix of the base attribute vocabulary and the learned attribute tokens, respectively. We observe that the correlation of learned tokens is indeed different from that of base vocabulary. Specifically, many items in the base vocabulary are highly correlated, with an average value of 0.858, indicating redundancy within the base vocabulary. In contrast, the correlation of learned tokens is much smaller, with an average value of only 0.003, indicating that separable and independent tokens are learned. - In Fig. R2 of the attached top-level PDF file, we visualize the distribution of the base attribute vocabulary and the learnable attribute tokens using t-SNE. It can be seen that the learnable tokens (marked in orange) can fill in part of the gaps in the base vocabulary (marked in blue), indicating that the two represent different semantic meanings and are therefore complementary to each other. Q2: It is difficult to think about the case that we have never seen an object, but we know its attributes quite well. In such a sense, I believe this method is more appropriate for few-shot learning. A2: Thank you for the comment. In the ZSL/GZSL setting, a general assumption is that for unseen classes, their prior semantic information such as attributes is known in advance, although there are no images from unseen classes for training. The overlapping attributes between seen and unseen classes can facilitate the knowledge learned on the seen classes to be transferred to the unseen for zero-shot recognition. As per your suggestion, our framework can be readily extended to few-shot learning. It is important to note that few-shot learning differs from ZSL/GZSL in that it leverages a small number of labeled samples from unseen/novel classes. To accommodate this difference, we can align the multimodal representations of both the seen/base and unseen/novel classes within the dual-space. The model can then be trained using our multi-modal alignment loss (Eq. 5 and 6) plus the topology-preserving loss (Eq. 7) proposed in this work. In this setting, our model still inherits the good generalization ability of the CLIP model for effective few-shot learning. Q3: Explanations for the accuracy increase in Fig.4(a). A3: On the AwA2 dataset (Fig.4(a)), the performance of the unseen classes (i.e., $U$) improves as the number of learnable attribute tokens (i.e., $N_2$) increases. We attribute this to the following reason: The AwA2 dataset is a coarse-grained dataset encompassing a wide range of animal categories. The base vocabulary may not comprehensively cover all essential attributes, leaving some important features unrepresented. This gap necessitates the addition of more learnable attribute tokens, which contribute to the observed increase in accuracy as $N_2$ grows.	Summary: In this paper the author proposed a dual-space feature alignment module to keep the semantic consistency between visual and attribute. In addition, the authors proposed Topology-Preserving Reservoir (TPR) to tackle the issue into the generalized zero shot learning (GZSL) setting, which utilized the Pearson correlation coefficient to define a topology-preserving loss, which effectively prevents overfitting of the seen and unseen classes. Sufficient experiment demonstrate the effectiveness of the proposed method. Strengths: (1)The Paper is well-written, meanwhile, the method is intuitive and easy to understand. (2)The proposed method focused on Generalized Zero-Shot Learning (GZSL) to present Topology-Preserving Reservoir to finetune the pre-trained CLIP for better fit the distribution of seen and unseen classes, which seems reasonable. (3)Sufficient and significant experiments demonstrate the effectiveness of the proposed method. Weaknesses: (1)The Dual-Space Feature Alignment proposed by the author, which uses a Cross Attention mechanism for cross-modal alignment, lacks innovation. (2)The author mentions "attribute reservoir" in the article, but essentially it is just a fully connected layer that generates different feature representations through various loss constraints. Additionally, in Figure 2, the attribute reservoir is shown in two states: frozen and trained. I am unsure about when these two states should transition between each other. (3)The idea proposed by the author to fine-tune feature distribution using spatial topological structures is intriguing. However, relying solely on the Pearson correlation coefficient to define a topology-preserving loss seems somewhat simplistic. Technical Quality: 3 Clarity: 3 Questions for Authors: NA Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Reviewer k9s8 ### Response Q1-Q3 Thank you for the valuable comments and recognizing the various strengths of our paper: "well-written", "intuitive and easy to understand", "the method better fits the seen and unseen classes", "reasonable", and "sufficient and significant experiments". We address the raised concerns by the reviewer as follows. Q1: Lack of innovation for the dual-space feature alignment module. A1: We appreciate the opportunity to further clarify innovation for our dual-space feature alignment module. To the best of our knowledge, there is no existing related work that employs a latent-attribute dual-space representation to address GZSL. Most previous methods rely solely on a single latent space for multimodal feature alignment. We observe that these methods fail to effectively capture complex and fine-grained patterns, leading to suboptimal performance, especially for unseen classes (see Table 1 of the main paper). To overcome these limitations, we propose to align multimodal features in a dual-space consisting of a latent space and an attribute space. The latent space provides a general representation of the input multimodal features, while the attribute space, induced by a predefined base attribute vocabulary and learnable attribute tokens, offers a more structured and interpretable representation. By aligning multimodal features within these two complementary spaces, our method not only captures both prior knowledge and task-specific information more effectively but also reduces the risk of overfitting to seen classes. This dual-space alignment design leads to superior performance compared to previous works. Finally, we would like to emphasize that cross-attention is a common practice for cross-modal alignment. How to effectively align image and text modalities is the key contribution of our work. Thus, we design a novel dual-space to align the multimodal image-text representations instead of projecting them into a single latent space. Q2: Explanations for the attribute reservoir. A2: Thank you for the comment. To the best of our knowledge, our work is the first attempt to leverage an attribute reservoir to solve the description-based GZSL task. By utilizing this reservoir, we are able to significantly improve performance through capturing complex and fine-grained patterns and facilitating alignment between image and description features. More specifically, the attribute reservoir consists of two complementary components: the base attribute vocabulary and the learnable attribute tokens, which is conceptually and technically distinctive from simply applying a fully-connected layer to project features. - The base attribute vocabulary contains 5,996 extensive attributes collected from the existing literature related to attribute recognition. These attributes are universally shared across all datasets, which provide prior knowledge and effectively improve the generalization of the model to unseen classes. - The learnable attribute tokens are introduced to compensate for the cases where base attribute vocabulary cannot cover sufficient attributes to perform effective GZSL. Concretely, they have two functions: 1) learning complementary attribute knowledge that is missing in the base vocabulary; 2) integrate task-specific information into the attribute reservoir so as to better align the image and description features. Experimentally, the two components of the attribute reservoir are complementary to each other (Table 2 of the main paper) and together they achieve superior performance over the state-of-the-art. This demonstrates the effectiveness of our attribute reservoir. Q3: The idea of using spatial topological structures is intriguing. Relying solely on the Pearson correlation coefficient to define a topology-preserving loss seems somewhat simplistic. A3: Thank you for the helpful comment. In this work, our primary goal is to maintain the class topology structure of the CLIP embedding space, thereby improving the generalization ability of unseen classes. As shown in Table 3 of the main paper, we conducted various forms of topology-preserving losses, but we found that using the Pearson correlation coefficient to define topology-preserving loss is simple yet effective. Further empirical results show that it indeed improves the generalization ability of the model significantly (Table 1 and Fig. 1(a)). --- Rebuttal 2: Comment: Dear Reviewer, Thank you for your comprehensive review of our manuscript. We have taken your feedback seriously and have revised the paper to address your concerns and incorporate your suggestions. If there are any areas where you would like us to provide more information or clarification, please feel free to ask. Your expertise and insights are invaluable to us, and we aim to address any outstanding concerns.	Summary: The proposed approach targets the generalized zero-shot learning (GZSL) problem for the vision language model (VLM). It is observed that a strong VLM model shows promising results for novel class generalization. Fine-tuning these models for seen classes leads to a loss in generalization capability and poor results for unseen classes. Additionally, a single latent space demonstrates limited ability to adapt to complex visual-linguistic patterns in fine-grained datasets. The paper proposes dual-space alignment, augmenting the latent space with static and learnable tokens. To address the generalization problem post fine-tuning, the paper introduces a Topology-Preserving Reservoir (TPR), which helps preserve the model's generalization ability for unseen classes. The authors conducted extensive experiments across several standard ZSL datasets and explored the impact of various components through ablation studies. Strengths: [1] Generalization of unseen classes in VLM is a critical problem. The strong pretrained model also loses its generalization ability, which the author explores, and the proposed model shows a significant impact. [2] The idea and intuition behind the static and learnable attribute reservoir are interesting. Additionally, TPR helps improve generalization. [3] The wide-ranging experiments conducted across various ZSL datasets and the ablation studies are satisfactory. Weaknesses: [1] The standard ZSL model assumes that there is a description per class rather than per sample, which is more intuitive since a single description for each class suffices for the model to understand the class, making it cost-efficient. Standard annotation-based attributes often yield better results for ZSL/GZSL settings. For example, [a] demonstrates impressive results for the CUB dataset compared to the proposed complex static, learnable, and description-based model. This issue is particularly observed in fine-grained datasets. Why is this the case? [2] It is unclear how the base attribute vocabulary is created. At a high level, the author collected a few attributes and obtained LLM embeddings. This description may not be sufficient for reproducibility since the code and data are not provided. [3] There are multiple variants of TPR (Table-3) in various scenarios where different methods work, making it difficult to apply and choose the best one. What does the author conclude here? [4] In Table-1 for the SUN dataset, the model shows inferior performance. While we do not expect the model to outperform in all scenarios, a clear description and author observations are required: why is this the case? [a] Meta-Learned Attribute Self-Interaction Network for Continual and Generalized Zero-Shot Learning, WACV-24 Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to the weakness section. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Authors has not discussed the limitation in the paper, while in the Checklist they said "Yes" i.e. they had discussed. This is a bad practice, I don't know what to do. Dear AC please look into it. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Reviewer TcJo ### Response Q1-Q6 Thank you for the valuable comments with many kind words to our work: a critical problem, significant impact, interesting, improve generalization, wide-ranging experiments, satisfactory ablations studies. Below, we address the raised questions: Q1: Whether a description is provided to each class or each sample? A1: We would like to clarify that our method utilizes the Class-Level textual descriptions (a description per class). Therefore, our method is cost-efficient with the class-level design. Q2: Standard annotation-based attributes often yield better results for GZSL settings. For example, [a] demonstrates impressive results on CUB. A2: Thank you for your constructive comments. Firstly, we agree with Reviewer TcJo that sometimes standard annotation-based attributes yield good results for ZSL/GZSL. However, they require extensive expert knowledge and detailed class-level attribute annotation for both seen and unseen classes. In contrast, our method relies only on attribute names and LLM-generated descriptions, which are easier to obtain and significantly reduce the need for human efforts. Secondly, we speculate that the impressive results of [a] on CUB are due to the usage of meticulously annotated attributes. Specifically, the CUB dataset provides 312-dimensional densely labeled attributes, which capture the characteristics of bird body parts, such as the beak, head, trunk, tail, and claws. Compared with the well-annotated attributes, LLM-generated descriptions may overlook some key attributes characterizing the bird classes. Thus, learning with detailed attribute annotations [a] leads to recognition improvements. We will cite [a] and include the above discussions in the revision. [a] Vinay Verma, et al. Meta-Learned Attribute Self-Interaction Network for Continual and Generalized Zero-Shot Learning. WACV, 2024. Q3: How is the base attribute vocabulary created ? A3: We detail its construction as follows: - First, we collected the readily available attribute words from diverse attribute recognition related repositories, such as MAD, VAW, and LSA. For example, MAD dataset has 158 attributes, which are incorporated into our base vocabulary. Note that we just used the attribute names without involving any annotations. - Then, we eliminated duplicate attribute words to form a base vocabulary of 5,996 attribute words. A pre-trained LLM was adopted to extract features. In Appendix Line 751, we provided the anonymous code link. We will make all code and data public upon acceptance. Q4: Choose of multiple variants of TPR and the conclusion. A4: Thank you for the comment. In GZSL, Harmonic Mean ($H$) is the most comprehensive metric to evaluate model performance, which takes into account both seen and unseen classes. Therefore, we choose the best model according to the results of $H$. For ease of comparisons, in Table R2 below, we average the performance w.r.t. three datasets and observe that our proposed topology-preserving loss (row e: $L_{tp}$) obtains the best results. Our conclusions are two-fold: 1) all regularization variants exhibit a degree of performance improvement compared to the baseline, indicating the importance of constraining spatial structure of CLIP; 2) among these, our topology-preserving loss $L_{tp}$ shows better performance than the other variants, highlighting its effectiveness. Table R2. Average performance of different variants of TPR on three datasets: AwA2, CUB and FLO. The proposed topology-preserving loss $L_{tp}$ obtains the best results. \| Model \| $S$ / $U$ / $H$ \| \|---------------------\|----------------------\| \| a) w/o $L_{tp}$ \| 66.6 / 54.2 / 59.6 \| \| b) nuclear norm \| 67.9 / 54.2 / 60.0 \| \| c) orthogonality \| 67.1 / 55.3 / 60.4 \| \| d) $L_{tp}^{lat}$ \| 67.4 / 55.8 / 60.8 \| \| e) $L_{tp}$ \| 68.6 / 56.1 / 61.5 \| Q5: Explanations for the inferior performance on the SUN dataset. A5: Unlike object-specific classes, SUN is a generic scene dataset containing 717 categories, such as "amphitheater" and "amusement park". These broad scene categories exhibit extensive variations in their components and features, making it challenging for our TPR method to align images with descriptions. For instance, the category "amusement park" may include diverse elements such as "crowds", and "buildings", which vary significantly across different instances. This diversity complicates the projection of both multimodal features into the attribute space, where consistency is harder to achieve. Moreover, we observe that several class descriptions in the SUN dataset exhibit high cosine similarity scores. Specifically, about 5% of the textual descriptions have similarities higher than 0.8, with an average similarity of 0.69. In contrast, the average similarity for the FLO dataset is 0.52. The high similarity across descriptions of different categories adversely impacts the model's performance, especially for unseen classes. Furthermore, we would like to emphasize that although TPR is 3.2% lower than ProGrad [12] on the SUN dataset, while on other datasets TPR is on average about 10% higher than the state-of-the-art. Q6: Summary of limitations. A6: We summarize the limitations of TPR as follows: Text alone may not fully capture the nuances of fine-grained datasets like CUB, while attribute annotations, though more accurate, are costly. Thus, a more desirable solution would be combining the knowledge from expert-provided attribute annotations with LLM-generated text to enhance performance. Additionally, our method may face challenges in aligning visual features of generic scenes with description features in the attribute space, especially when descriptions are not sufficiently specific. This could be alleviated by providing more distinct and human-refined descriptions.	Summary: This paper is a new study that introduces the Generalized Zero-Shot Learning (GZSL) framework within VLMs, aiming to classify both known and novel classes without class partitioning. Key innovations include a dual-space feature alignment module, enhancing latent representations with an attribute reservoir for nuanced visual-linguistic patterns. Additionally, a topology-preserving objective ensures that model adaptations preserve the semantic structure learned by CLIP, thus maintaining generalization across all classes. Extensive experiments across diverse datasets validate the proposed Topology-Preserving Reservoir (TPR) model, demonstrating superior performance over conventional methods in recognizing both seen and unseen classes, underlining its potential for practical applications in complex visual recognition tasks. Strengths: 1. This paper introcuces a novel research aspect for VLMs: generalized zero-shot learning, which requires the model to identify both seen and unseen concepts at the same time. From my perspective, this proposal could be a great contribution to VLM community. 2. This paper is well-organized and well-written, which makes it easy to follow. 3. Extensive experiments，ablation study and visualization results demonstrate the effectiveness and rationality of TPR. Weaknesses: None in particular. Technical Quality: 4 Clarity: 4 Questions for Authors: GZSL is a long standing problem in ML/AI community with many classic solutions, like stacking[1] .etc. Since the authors only compare with VLM-based methods, my concern is about the perfromance of classic methods in VLM-based GZSL tasks, can they achieve surprising performance with well-trained features? [1] Chao W L, Changpinyo S, Gong B, et al. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild[C]//Computer Vision–ECCV 2016 Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 1 Limitations: None in particular. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Reviewer BqNV We sincerely thank the reviewer BqNV for the very positive and helpful comments. Thank you for acknowledging the novelty of our idea, our contribution to the VLM community, the writing and organization of our paper, and the extensive experiments & ablation study. We address your questions as follows: Q1: Can classic methods [1] achieve superior performance with well-trained features in VLM-based GZSL tasks? A1: Thank you for the comment. In Table R1 below, we compare with the two variants of the classic method [1] using pretrained VLM-based features (i.e., ViT-B/32 CLIP). The classic method SynC [1, 2] achieves competitive results on seen classes, but degrades significantly on unseen classes. Specifically, SynC [1, 2] constructed classifiers for unseen classes by linearly combining the phantom classifiers trained on seen class data. Therefore, it exhibits a significant bias towards seen classes, resulting in inadequate generalization to unseen classes. This poor generalization phenomenon is also observed in Ref [3]. In contrast, our TPR method obtains promising results on both seen and unseen classes, demonstrating the generalization capability. We will include the results and discussion with Ref [1-3] in the revision. Table R1. Performance comparison between the classic methods [1] and ours in the VLM-based framework. \| \| AwA2 \| CUB \| \|---------------------\|----------------------\|----------------------\| \| Model \| $S$ / $U$ / $H$ \| $S$ / $U$ / $H$ \| \| a) CLIP \| 81.7 / 77.7 / 79.6 \| 29.9 / 29.6 / 29.7 \| \| b) SynC$^\text{o-vs-o}$ [1] \| 86.0 / 15.1 / 25.7 \| 31.9 / 10.5 / 15.8 \| \| c) SynC$^\text{struct}$ [1] \| 88.3 / 16.7 / 28.1 \| 33.8 / 11.4 / 17.0 \| \| d) TPR (Ours) \| 87.1 / 76.8 / 81.6 \| 41.2 / 26.9 / 32.5 \| [1] Wei-Lun Chao, et al. An Empirical Study and Analysis of Generalized Zero-Shot Learning for Object Recognition in the Wild. ECCV, 2016. [2] Soravit Changpinyo, et al. Synthesized Classifiers for Zero-Shot Learning. CVPR, 2016. [3] Yongqin Xian, et al. Zero-Shot Learning—A Comprehensive Evaluation of the Good, the Bad and the Ugly. IEEE T-PAMI, 2018.	Rebuttal 1: Rebuttal: ## Global Rebuttal We thank all reviewers for their insightful and positive feedback. We are encouraged that the reviewers acknowledge our paper: - Novel and impactful. Reviewer BqNV -- "introduces a novel research aspect", "a great contribution to VLM community"; Reviewer TcJo -- "the proposed model shows a significant impact", "idea and intuition are interesting", "TPR helps improve generalization"; Reviewer k9s8 -- "the method is intuitive"; Reviewer 8m4c -- "seems novel enough". - Sufficient and significant experiments demonstrating the effectiveness of our method. Reviewers reckon our experiments are extensive (BqNV), wide-ranging (TcJo), sufficient and significant (k9s8), and with detailed comparison and well-designed ablation studies (8m4c). - Well-organized, well-written, and easy to follow Reviewer BqNV -- "This paper is well-organized and well-written, which makes it easy to follow." Reviewer k9s8 -- "The paper is well-written, meanwhile, the method is intuitive and easy to understand." We thank all the reviewers' suggestions and comments that significantly improve the quality of our work, and we have done our best to address the reviewers' concerns. Below, we summarize the key changes we have made for rebuttal: - Perfromance of classic methods in VLM-based GZSL tasks (reviewer BqNV) - Clarification of Class-Level description, comparison with standard annotation-based attributes methods, chosen of model variants, explanation of the performance on SUN dataset (reviewer TcJo) - Clarification of key innovations regarding dual-space design, attribute reservoir, and topology-preserving loss (reviewer k9s8) - Explanation of attribute reservoir (reviewer TcJo and k9s8) - Provision of a summary of limitations (TcJo) - Visualization and Analysis of the learned attribute tokens (PDF attached, reviewer 8m4c) Please find individual responses to your questions below. Pdf: /pdf/82ed4dd9e586834ce9f465c15f8546c197339a4c.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Enhancing Robustness of Graph Neural Networks on Social Media with Explainable Inverse Reinforcement Learning	Accept (spotlight)	Summary: The paper presents a novel approach to enhancing the robustness of Graph Neural Networks (GNNs) against adversarial attacks, specifically in social media contexts such as rumor detection. The authors propose an enhanced maximum entropy inverse reinforcement learning (IRL) method with a mixture-of-experts approach to tackle multi-source graph adversarial attacks. This method aims to reconstruct attack policies, integrate various attack models, and generate additional adversarial samples to improve the robustness of GNN-based detection models. Strengths: The application of inverse reinforcement learning to reconstruct adversarial attack policies is novel and offers a highly interesting perspective on enhancing GNN robustness. Combined with the Mixture-of-Experts, the method allows for the integration of various attack models, providing comprehensive feature-level explanations and robust adversarial samples for use in adversarial training. The generation of good additional adversarial samples for training improves the GNN’s resilience to attacks, which is a significant step towards robust social media analysis. The authors use real-world social media datasets to validate the proposed method. Weaknesses: The proposed method involves multiple components (IRL, mixture-of-experts, bidirectional updates), which can increase the computational complexity and may not be easily scalable. The focus is primarily on rumor detection in social media, which, while important, might limit the generalizability of the method to other types of graphs and applications. Some sections, particularly those involving the theoretical underpinnings of IRL and mixture-of-experts, could be more clearly explained to enhance understanding and accessibility. No code is provided. This hinders the exact reproduction of results. I think the authors use the term "Threat model" in an incorrect or at least unorthodox way that will likely be misunderstood in the security community and potentially beyond. Specifically, in line 229, the authors start a paragraph called “Threat model” and they proceed to describe that they use, GCN, number of hidden dimensions, optimizer, etc. This is not what is typically understood as a threat model in literature: a model of a threat actor's capabilities, possible courses of action they may take and how it will impact the operation of a computer system [1]. Speaking of it, including an actual threat model (or rather making the implicitly exiting model explicit) would certainly strengthen the paper and increase acceptance in the security community. Minor issues: Typos/grammar in lines 13, 69, 109, 111 [1] https://www.sciencedirect.com/science/article/abs/pii/S0167404818307478 Technical Quality: 3 Clarity: 2 Questions for Authors: Could this approach be transferred (with respective modifications) to bit flips, e.g., [1]? Did you analyze the impact of varying Rich-club coefficients in the datasets? [1] https://arxiv.org/abs/2311.01205 Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The method assumes that the perturbations captured during training are representative of real-world adversarial attacks. Further testing on different graph types and applications is required to confirm the broader applicability of the proposed method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Could this approach be transferred to bit flips? A1: Thank you for your constructive question. We believe that it is feasible theoretically. (1) Bit flip attack: This attack disrupts neural network operations by flipping bits in parameters or intermediate results. For example, [1] degrades GNN performance by selectively flipping bits in quantized weights and biases, targeting those most impactful to the model's injectivity. (2) To transfer our work to bit flips, the key is to model bit flip attacks as a MDP for RL. Suppose there are $T$ -times bit flips to achieve an effective attack, defining the attack trajectory with $T$ steps. The RL elements are: - State: The current GNN parameters; - Action: Flipping the specific bit in the parameters, which could be represented by a mask matrix; - Reward: Estimated by the IRL module; - Policy: Given the current state, output the next action. A deep RL policy, such as Actor-Critic (AC), might be effective for handling the state from the deep models. For IRL, it is suggested to use a neural network as the reward function because it is difficult to design interpretable linear features for bit flip attacks when targeting the internals of an end-to-end deep model. Given the target GNN model and some expert bit flip attack trajectories, the learner policy produces the attack actions and updates with the reward from the IRL module to approximate the expert policy. [1] Attacking Graph Neural Networks with Bit Flips: Weisfeiler and Lehman Go Indifferent Q2: Did you analyze the impact of varying Rich-club coefficients in the datasets? A2: Thank you for your vaulable suggestion. The rich-club coefficients of Weibo and Pheme are shown in Figure 1 in the PDF. The analysis reveals that: (1) As the degree increases, the rich-club coefficient first rises and then falls, indicating that nodes with moderate influence in the social network have higher connectivity with each other. (2) Pheme exhibits higher rich-club coefficients compared to Weibo, suggesting that the social network on Pheme is more densely connected overall. As a result, attacks on Pheme generate more pronounced cascading effects [1], thereby complicating the recovery of attack policies, which aligns with the experimental results presented in Table 2. The model's performance on Pheme, which has a higher rich-club coefficient, is lower compared to the performance on Weibo with smaller rich-club coefficients. [1] Cascading failures in complex networks Q3: The proposed method involves multiple components, which can increase the computational complexity and may not be easily scalable. A3: Thank you for your valuable suggestion. Please see Q1 in the global rebuttal for further details. Q4: It might limit the generalizability of the method to other types of graphs and applications. A4: Thank you for your suggestion. To verify the generalizability, we have adapted and implemented our method for the recommendation system task. - Dataset: Gowalla [1], includes 13,591 users, 14,322 items, and 227,193 interactions. - Interpretable Features: The domain-specific features used for rumor detection in our work were modified to suit recommendation systems. - Metric: ΔNDCG, which measures the reduction in NDCG@20 value after $T$-step attacks. - Target Model: PinSage [2]. Without attacks, the NDCG value on the training set is 0.4139. - Attack Method: We improved AdRumor-RL with the features in (2) and generated three expert trajectories, each with 5 steps. It conducts the global attack and updates with the reward ΔNDCG. The average attack performance (ΔNDCG) is 0.0325. - Performance: Results of policy recovery are as follows. The learner policy achieved 90% of the performance of experts. Our method is effective in the recommendation system task. \| \| Expert \| Apprenticeship \| EntIRL \| MoE-BiEntIRL \| \| ----- \| ------ \| -------------- \| ------ \| ------------ \| \| ΔNDCG \| 0.0325 \| 0.0067 \| 0.0268 \| 0.0294 \| [1] Friendship and mobility: User movement in location-based social networks [2] Graph convolutional neural networks for web-scale recommender systems Q5: The theoretical underpinnings of IRL and MoE could be more clearly explained. A5: Thank you for your suggestion. We provide explanations for EntIRL and MoE as follows. (1) EntIRL infers the reward function in the environment by maximizing the entropy of the path distribution. Thus, the EntIRL objective is defined as $\max\sum -P(τ) \log P(τ)$ with trajactory $τ$. It assumes that $P(τ\|θ)∝\exp(R_θ(τ))$, where $R_θ(τ)$ is the reward function with parameter $θ$. The loss function for EntIRL is the likelihood $L(θ) = \sum \log P(τ\|θ)$ because maximizing likelihood and maximizing entropy share the same goal [1]. In our work, we consider locally optimal examples, segmenting the trajectory as state-action pairs. The previous assumption becomes $p(a\|s)∝\exp(Q^(s, a))$ [1]. With a discount factor of 0, the action probability $p(a\|s)$ is given by Eq. (2). The loss function then becomes $L(θ) = \sum \log P(a\|s, θ)$. (2) MoE* consists of a gating network $α()$ and $K$ experts $p_1(),\ldots,p_k()$. It can be indicated as $p(y\|x) = \sum_k α(x)p_k(y\|x)$. We model the RL policy as an MoE as stated in Eq. (4), with each expert corresponding to an action probability function as shown in Eq. (5). The IRL loss function for each expert is $L(θ_k) = \sum \log p(a\|s, θ_k)$, which aligns with the M-step goal in the EM algorithm as stated in Eq. (11). Thus, it optimizes the IRL policy using the EM algorithm. [1] Maximum entropy inverse reinforcement learning Q6: No code is provided. A6: Thank you for your valuable suggestion. Please see Q2 in the global rebuttal. Q7: Wrong description about “Threat model”. A7: Thank you for your correction. "Threat model" should be replaced with "Target model". --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal. My concerns have been mostly alleviated and uncertainties clarified. I am raising my rating from 5 -> 7 and my confidence from 2 -> 3. I maintain my recommendation to briefly summarize the attackers' assumed capabilities and goals in a short paragraph, as this information currently scattered throughout the paper. --- Reply to Comment 1.1.1: Comment: Dear Reviewer sELL, Thank you very much for your thoughtful feedback and for taking the time to review our rebuttal. We are pleased to hear that our responses have addressed most of your concerns and that you found our clarifications helpful. We appreciate your suggestion to summarize the attackers' assumed capabilities and goals in a concise paragraph. We will ensure this is clearly presented in the revised manuscript. Thank you again for your valuable insights and for raising your rating and confidence level. Your input has been instrumental in improving our work. Best regards, Submission 14020.	Summary: This paper addresses the challenge of adversarial attacks on Graph Neural Networks (GNNs) employed in social media tasks, such as rumor detection. The authors introduce MoE-BiEntIRL, a method that leverages a mixture-of-experts approach combined with inverse reinforcement learning (IRL) to reconstruct and explain adversarial attack policies. The objective of this method is to enhance the robustness of GNNs by generating additional adversarial samples for training, thereby improving resilience against attacks. MoE-BiEnt\IRL incorporates mechanisms for precise sample guidance and bidirectional updates, which are designed to optimize both the accuracy and the speed of policy learning. Strengths: 1. Innovative Approach: The introduction of MoE-BiEntIRL represents a significant innovation, particularly through its application of a mixture-of-experts approach to manage diverse adversaries and provide detailed feature-level explanations. 2. Real-world Validation: The method is validated on actual datasets from Weibo and Pheme, demonstrating its practical applicability for improving the robustness of GNNs in social media rumor detection scenarios. 3. Experimental results, focusing on policy reconstruction and adversarial training, effectively illustrate the method’s robustness and efficacy. 4. The approach facilitates a deeper understanding of attack behaviors through feature-level explanations, aiding platform operators in enhancing system defenses. Weaknesses: 1. The proposed method involves multiple stages and sophisticated mechanisms, potentially complicating its implementation. 2. Scalability Discussion: The paper would benefit from a more extensive discussion on the scalability of the method, particularly concerning its applicability to large social media graphs. 3. Experimental Setup Details: Enhancing the description of the experimental setup would significantly improve the reproducibility of the study and aid other researchers in replicating the results. 4. Typos and grammar errors could be avoided. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Scalability Analysis Could you elaborate on how your method scales when applied to very large social media graphs? Any additional insights or preliminary results on this matter would be highly informative. 2. Could you provide more detailed information regarding your experimental setup? Additional details would aid in understanding how to replicate your study effectively. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: The paper's limitations are adequately addressed in the supplementary material. There are no further suggestions at this time. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Scalability Analysis: Could you elaborate on how your method scales when applied to very large social media graphs? Any additional insights or preliminary results on this matter would be highly informative. A1: Thank you for your valuable suggestion. For the scalability in the large social media graphs, we analyze the time complexity of our method. Please refer to Table I in the attached PDF, where we detail the time complexity and runtime of the MoE-BiEntIRL, alongside two baseline models. Based on Table I in the accompanying PDF, the time complexity of our proposal and the baseline methods primarily diverges during the reward acquisition phase of IRL. Our approach introduces $K$ experts to manage multi-source attack trajectories with diverse motivations, thereby increasing the time complexity. Notably, the complexity associated with IRL reward acquisition is independent of the input graph size and hinges solely on the predefined hyperparameter $K$, typically ranging from 1 to 10, rendering the increased complexity manageable. Moreover, the predominant computational effort across all models is concentrated in the attack stage, which further mitigates the impact of introducing multiple experts. Table I details the runtime of complete episodes and the IRL module. Despite the extended runtime of the MoE-BiEntIRL approach in the IRL module, it exerts negligible influence on the total runtime. Both the analysis of time complexity and experimental findings emphasize that the actual runtime is largely influenced by the attack stage. Please see Q1 in the global rebuttal for further details about the time complexity analysis. Based on the aforementioned time complexity analysis, the scalability of the IRL module easily scalable as it exhibits no direct correlation with the size of the graph. When applied to larger graphs, parameters such as $T$, $K$, $N$, and $S$ may indirectly increase due to the diverse nature of attacks within these larger contexts. As illustrated in Table I, the runtime of the IRL module shows minimal variation across datasets of varying sizes (e.g., Weibo with 10,280 nodes and Pheme with 2,708 nodes). As discussed in time complexity analysis, the predominant factor influencing runtime is the attack stage. Thus, in our humble opinion, the scalability of the proposed model is acceptable. Furthermore, as for the hyperparameters affected by the large graph indirectly, various methods can mitigate the time complexity and the runtime of the Inverse Reinforcement Learning (IRL) module through careful control of hyperparameters: (1) $K$: Reduce the number of experts by phased experimentation, employing a coarse-to-fine division of experts. We design this method and conducts expriements, which details are shown in Q1 of the global rebuttal. (2) $T$: Segment lengthy expert trajectories using predefined rules or models to minimize state correlation before and after each segmentation. (3) $N$: Employ pre-classification or clustering of multiple expert samples to reduce the number of expert trajectories. Q2: Could you provide more detailed information regarding your experimental setup? Additional details would aid in understanding how to replicate your study effectively. A2: Thank you for your valuable suggestion. Please see Q2 in the global rebuttal for further details. --- Rebuttal Comment 1.1: Comment: Thank you for your thorough rebuttal. The complexity analysis and additional experiments have effectively addressed my concerns. I am inclined to maintain my original score. Thanks again. --- Reply to Comment 1.1.1: Comment: Dear Reviewer M7DL, Thank you very much for your positive feedback and for taking the time to review our rebuttal. We greatly appreciate your thorough assessment and are pleased that our additional analysis and experiments have addressed your concerns. We value your insights and are grateful for your support throughout this review process. Best regards, Submission 14020.	Summary: This work studies the problem of reconstructing attack policies using collected adversarial samples to enhance the robustness of GNN-based models in social network tasks, specifically rumor detection. The authors propose the MoE-BiEntIRL framework, which employs a mixture-of-experts approach to learn optimal policies from diverse adversaries, and provides feature-level explanations by estimating interpretable linear reward functions. Experiments on two real-world rumor detection datasets validate the effectiveness of MoE-BiEntIRL. Strengths: 1. The authors investigate the rumor detection problem from the novel perspective of reconstructing attack policies. 2. The paper is well-written and well-organized, with motivating illustrations of the problem. Weaknesses: While the proposed problem and approach are generally novel and intriguing, the following issues regarding experiments require further clarification: 1. Table 2: What makes the policies on Pheme significantly harder to recover than the policies on Weibo? 2. Table 3: The results are not clearly illustrated and explained. - For instance, it appears that the column under "w/o Att." reflects test accuracy (%), while results under other columns reflect accuracy decline in actual numbers. Please align the representations for consistency. - If "w/o Att." refers to GCN's rumor detection performance without adversarial attacks, it is surprising to see that GCN only achieves ~70% test accuracy on the Weibo dataset with binary rumor / non-rumor labels. The authors claim that the Weibo dataset is adopted from existing work [1], which reported over 80% test accuracy on Weibo even using simple models such as TF-IDF or GRU. Please elaborate on the causes for this significant performance discrepancy, e.g, data differences and model structure differences. 3. Computational Efficiency: Given the complexity of the model structure illustrated in Figure 2, it would be beneficial to benchmark the computational efficiency of the proposed approach against the baselines in Table 3. [1] Changhe Song, Cheng Yang, Huimin Chen, Cunchao Tu, Zhiyuan Liu, and Maosong Sun. Ced: Credible early detection of social media rumors. TKDE, 33(8):3035–3047, 2021. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to the weaknesses section. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: What makes the policies on Pheme significantly harder to recover than the policies on Weibo in Table 2? A1: Thank you for your insightful question. We posit that the difficulty in policy recovery is related to the complexity of the underlying graph structures. As indicated in Table II of the PDF, we have devised several metrics to assess the complexities of the Weibo and Pheme datasets [1-3]. The data reveals that the graph structure of Pheme is more complex than that of Weibo, characterized by a larger variance in node degree and a higher graph density. Returning to the proposed model, the success of Inverse Reinforcement Learning (IRL) in recovering expert policies is influenced by several factors, including the quality and quantity of expert demonstrations, the complexity of the expert policies, and the variability of the environment. In our humble view, the attack sequences within complex social graphs exhibit significant cascading effects [4], complicating the recovery of these policies. The more intricate the graph structure, the more substantial the cascading effect due to edge-flipping attacks. Such an attack can propagate from a target node to multiple others, thereby impacting the overall system performance because of the high interdependence and connectivity among nodes. Complex graphs increase the likelihood of cascading reactions triggered by attack behaviors. Furthermore, sequential attacks on these graphs intensify cascading reactions, potentially altering the characteristics of numerous nodes and causing environmental changes within the reinforcement learning process. Attackers may also optimize their strategies, adding more decision conditions and steps to navigate the complex social graph effectively. Together, these factors contribute to the increased difficulty of IRL in recovering expert policies. As presented in Table II, all baseline methods, including Apprenticeship Learning and EntIRL, exhibit suboptimal performance in recovering policies due to the complexity of the social graph in Pheme. In contrast, our approach, which integrates a mixture of experts to accommodate the attack sequences generated by complex social graphs, consistently achieves state-of-the-art performance on these challenging datasets. [1] Complex networks: Structure and dynamics. 2006. [2] The structure and function of complex networks. 2003. [3] Graph measures and network robustness. 2013. [4] Cascading failures in complex networks. 2020. Q2: The unclear description and significant GCN performance discrepancy with [1] in Table 3. (Weakness 2 in reviews) A2: Thank you for your correction and constructive question. (1) We apologize for not clearly describing the columns in Table 3. As you mentioned, the column under "w/o Att." reflects test accuracy (%), while results under other columns reflect accuracy decline in actual numbers. (2) The reasons for this discrepancy in attack effectiveness are multifaceted: - a) Dataset Differences: The weibo-all dataset referenced in [1] is sourced from the CED repository in Github, which includes the Rumdect dataset and the ChineseRumorDataset. However, the URL for Rumdect is no longer valid. Moreover, the weibo-all dataset in CED repository is pre-processed and lacks user IDs for message reposts, preventing the construction of the social following graph. Therefore, we utilized the original ChineseRumorDataset, which comprises 3,387 Weibo posts (1,538 rumors and 1,849 non-rumors), in contrast to the weibo-all dataset's 8,050 Weibo posts (3,581 rumors and 4,199 non-rumors). - b) Preprocessing Differences: In previous work [1], all reposting and comment information was either retained or filtered with the early rate (ER). Our approach focuses on retaining early repost messages and comments published within the first 25% of the average time interval and selectively retaining those with high information entropy. To achieve this, we removed stop words and punctuation, segmented the text, calculated the word information entropy for each text, and retained only those texts with entropy values higher than 90% of the highest observed entropy. - c) Model Differences: We used pre-trained word2vec embeddings and an embedding layer to encode the text content, whereas [1] used TF-IDF + NN as the encoder. We also supplemented the experimental results using TF-IDF encoding as shown in Table Ⅲ in the PDF. TF-IDF slightly outperforms word2vec embeddings. Pre-trained word2vec embeddings might suffer from poor representation of low-frequency words or differences between training corpora and the target task domain. Additionally, our proposal consistently outperforms the baselines with various text encoders, highlighting the effectiveness of the MoE framework. [1] Ced: Credible early detection of social media rumors. 2018. Q3: Computational Efficiency: Given the complexity of the model structure illustrated in Figure 2, it would be beneficial to benchmark the computational efficiency of the proposed approach against the baselines in Table 3. A3: Thank you for your valuable suggestion. Please see Q1 in the global rebuttal for further details. --- Rebuttal 2: Comment: Dear Reviewer QAYk, I hope this message finds you well. I am writing to kindly inquire about the status of your review for our submission 14020 as the rebuttal deadline swiftly approaching (just one day remaining). We understand that you may have a busy schedule, and we sincerely appreciate your time and effort in reviewing our work. If you have any additional comments or concerns, we would be grateful to discuss them as soon as possible. Thank you for your understanding and support. Best regards, Submission 14020. --- Rebuttal Comment 2.1: Comment: Thanks for your detailed response. My concerns have been addressed, and I will raise my score from 4 to 5. Please incorporate all responses to enhance the clarity of your work. --- Reply to Comment 2.1.1: Comment: Dear Reviewer QAYk, Thank you very much for your positive feedback and for taking the time to reconsider your evaluation of our work. We greatly appreciate your constructive comments, which have undoubtedly helped us improve the clarity and quality of our paper. Your input has been instrumental in improving our work. Best regards, Submission 14020.	Summary: The paper presents a novel method, MoE-BiEntIRL, which combines a mixture-of-experts approach with inverse reinforcement learning to enhance the robustness and explainability of adversarial attacks on GNNs. The method addresses the critical issue of stabilizing GNNs used in social media for rumor detection, demonstrating significant practical relevance. Strengths include its innovative approach, comprehensive mechanisms for improving attack policy accuracy, and robust evaluation results. However, the paper could benefit from clearer explanations of the method, detailed parameter sensitivity analysis, enhanced experimental reproducibility, expanded comparative baselines. Despite these minor weaknesses, the overall contribution and practical importance of the research are compelling. Strengths: 1. The MoE-BiEntIRL method presents a highly novel application of a mixture-of-experts approach combined with inverse reinforcement learning to address adversarial attacks on Graph Neural Networks (GNNs). This innovative approach stands out in its ability to not only enhance the robustness of GNNs but also to provide explainability to the attack policies. 2. The inclusion of precise sample guidance mechanisms and a bidirectional update mechanism demonstrates thoroughness in approach, aiming to improve both the accuracy of attack policy reconstruction and the speed of policy learning. This comprehensive approach adds substantial value to the proposed solution. 3. The evaluation methods employed in this study are robust, validating the effectiveness of the proposed method. The results are compelling, showing notable improvements in the robustness of GNNs. Weaknesses: 1. Although the proposed method is innovative, some aspects of the algorithm could benefit from clearer explanations. 2. A minor issue is that the sensitivity of the model to various parameters is not thoroughly explored. A brief analysis or guidance on parameter selection could aid in the practical application of the method. 3. While the method is novel, there is little discussion on its computational complexity. Including an analysis of the computational cost and suggesting optimizations could enhance the practical feasibility of the approach. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could you provide a simple illustrative example or additional details to clarify how the precise sample guidance mechanism and the bidirectional update mechanism work in the MoE-BiEntIRL method? 2. Can you provide a brief overview of the computational complexity of your proposed method, along with any potential optimizations that could be considered to enhance practical feasibility? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: The authors have listed the limitations in appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Could you provide a simple illustrative example or additional details to clarify how the precise sample guidance mechanism and the bidirectional update mechanism work in the MoE-BiEntIRL method? A1: Thank you for your suggestion. We have provided code and algorithms in global rebuttal Q2, which illustrate when and how the precise sample guidance mechanism and the bidirectional update mechanism operate. During the early stages of learning, precise sample guidance and inverse updates are performed periodically. In an inverse update episode, the steps for executing the RL policy are replaced with precise sample guidance steps, and attacks are carried out according to the given expert samples. As for the reward, it is assigned the maximum historical reward. Given the sample and reward, the gradient ascent update of the IRL is replaced with the inverse update after the responsibility calculation. Q2: Can you provide a brief overview of the computational complexity of your proposed method, along with any potential optimizations that could be considered to enhance practical feasibility? A2: Thank you for your valuable suggestion. Please see Q1 in the global rebuttal for further details. Q3: A minor issue is that the sensitivity of the model to various parameters is not thoroughly explored. A brief analysis or guidance on parameter selection could aid in the practical application of the method. A3: Thank you for your suggestions. Here, we provide guidelines for selecting some parameters: - The sampling upper bound $μ$: This parameter controls the similarity between the negative sample and the expert sample. Considering the validity of hard negative samples, we tend to choose samples with higher similarity, setting $μ=0.8$. - The learning rates of the learner policy, inverse update, and gate function: These are determined using grid search. We have included a sensitivity analysis of parameters in Figure 2 in the PDF. Our findings are as follows: - Sampling similar instances can be advantageous for policy reconstruction; however, selecting instances that are too similar can lead to the failure of Inverse Reinforcement Learning (IRL). * The learning rates for both the learner policy and the inverse update process are more sensitive compared to that of the gate function. --- Rebuttal Comment 1.1: Comment: After reviewing other referees' comments and the manuscript, I agree that the idea is novel, the technical content solid, and the contributions significant.	Rebuttal 1: Rebuttal: Q1: The complexity and scalability analysis of the proposed model. A1: Thank you for your valuable suggestion. Please refer to Table I in the attached PDF, where we detail the time complexity and runtime of the MoE-BiEntIRL, alongside two baseline models. Herein, we will present and discuss these specifics. (1) Time Complexity Analysis - The time complexity of MoE-BiEntIRL. The three principal stages of MoE-BiEntIRL are as follows: - Attack: It conducts $T$-step attacks. Each attack step requires feature updates for the involved subgraphs and nodes, including a forward pass on the target model. For the $L$-layer GCN as the target model, the time complexity of the forward pass is approximately $O(LN_{\text{edges}}d + LN_{\text{nodes}}d^2)$, where $N_{\text{nodes}}$ and $N_{\text{edges}}$ denote the numbers of nodes and edges in the input graph, and $d$ is the feature dimension. - Reward Acquisition: The reward is estimated with the IRL module. The time complexity for an episode is $O(NTKd(S+K))$, where $N$ is the number of expert trajectories, $T$ is the trajectory length, $K$ is the number of experts, $d$ is the feature dimension, and $S$ is the number of negative samples. - Policy Update: With the LinUCB algorithm, the policy update time complexity is $O(d^2)$. - Overall, the total time complexity for an episode is approximately $O(TLN_{\text{edges}}d +TLN_{\text{nodes}}d^2+NTKd(S+K)+d^2)$. - The time complexity of baselines. As in Table I , the complexities of the attack stage and policy update stage in the baselines are identical to those in the proposed model, differing only in the reward acquisition stage. The time complexity of the reward acquisition stage in baselines is: - Apprenticeship Learning: $O(NTd)$; - EntIRL: $O(NTdS)$. - Summary. Based on Table I, the time complexity of our proposal and the baselines primarily diverges during the reward acquisition stage. We introduces $K$ experts to manage multi-source attack trajectories with diverse motivations, thereby increasing the time complexity. Notably, the complexity associated with IRL reward acquisition is independent of the input graph size and hinges solely on the predefined $K$, typically ranging from 1 to 10, rendering the increased complexity manageable. Moreover, the main computational effort is concentrated in the attack stage, which further mitigates the impact of introducing multiple experts. Table I details the runtime of complete episodes and the IRL module. Despite the extended runtime of the MoE-BiEntIRL approach in the IRL module, it exerts negligible influence on the total runtime. Both the analysis of time complexity and experimental findings emphasize that the actual runtime is largely influenced by the attack stage. Therefore, the additional complexity introduced by employing multiple experts remains manageable. - Possible Solutions. Here we propose the MoE-BiEntIRL-K, focusing on reducing $K$. It utilizes DBSCAN to determine $K$ and pre-cluster labels. By varying the intensity of clustering, we obtain coarse and fine pre-cluster labels. During the IRL process, we initially use the number of experts and pre-cluster labels corresponding to the coarse clustering. As the learning progresses, we increase the number of experts to align with the fine clustering. The initial expert parameters are determined by weighting the previous expert parameters with the current responsibility value. We applied this method on Weibo with $T=20$ and $N=3$. The values of $K$ determined through coarse and fine clustering were 5 and 26, respectively. The results indicate a 19% reduction (10.67s to 8.64s) in runtime while achieving 92% of the original IRL's performance level (19.936 to 18.381 in $ΔL_A$). (2) Scalability Analysis Based on the time complexity analysis, the scalability of the IRL module is robust, as it shows no direct correlation with the size of the graph. When applied to larger graphs, parameters such as $T$, $K$, $N$, and $S$ may indirectly increase due to the diverse nature of attacks within these larger contexts. As shown in Table I, the runtime of the IRL module shows minimal variation across datasets of varying sizes (e.g., Weibo with 10,280 nodes and Pheme with 2,708 nodes). As discussed in time complexity analysis, the predominant factor influencing runtime is the attack stage. Thus, in our humble opinion, the scalability of the proposed model is acceptable. Q2: It is benefit to provide the clearer explanation of the algorithm, experimental setup and code. (Reviewer sELL, E98Q, M7DL) A2: Thank you for your suggestion. The algorithm and code are provided here. (1) Algorithm: MoE-BiEntIRL Input: Expert demonstration set $D$, Number of experts $K$, Length of trajectories $T$, Number of episodes $E$, Gate function $α$, Reward function parameters $θ=[θ_1,..., θ_k]$, Learner policy $π$, Negative sample set $D'$, Responsibility matrix $γ$ , Inverse update episode set $\Lambda$ Procedure: 1. For $ e = 1,\ldots,E $: 1. $ s = \text{env.reset}() $ 2. For $ t = 1,\ldots,T $: 1. If $e\in\Lambda$: - $s',a=\text{PreciseSampleGuidance}(D)$ - $ r = \text{MaxReward}() $ 2. Else: - $s',a=\text{env.step}(s,π)$ - $ r = \text{ObtainReward}(α, s, a, θ) $ as Eq.~(14) 3. $ \pi = \text{UpdatePolicy}(s, a, s', r) $ 4. $ s = s' $ 3. $ \gamma = \text{CalculateResponsibility}(α, D, D', θ) $ as Eq.~(8) 4. For $k = 1,...,K$: 1. If $e \in \Lambda$: - $θ_k = \text{InverseUpdate}(α,D,θ_k)$ as Eq.~(15) 2. Else: - $θ_k = \text{GradientAscent}(\gamma,D,D',θ_k)$ as Eq.~(12) 5. $ α = \text{GradientAscent}(γ) $ with the loss as Eq.~(10) - It requires to correct Eq.~(14) as $r_θ(s,a) = \sum\nolimits_{k=1}\nolimits^K α_k(s)θ_k^\top f(s,a)$. (2) Code and Experimental Setup have been provided to the AC in an anonymized link. Pdf: /pdf/c2be022b549cca651cf7bb66fcc85769074c2f48.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Entrywise error bounds for low-rank approximations of kernel matrices	Accept (poster)	Summary: This paper is first to establish entrywise guarantees for low rank approximation of kernel matrices when kernel eigenvalues satisfy either polynomial or exponential decay. More specifically, in the $\alpha$-polynomial decay setting, entrywise error scales as $O(n^{-\\frac{\alpha-1}{\\alpha}} \\log n)$ for rank $d = \Omega(n^{1/\\alpha})$, while for $(\\beta,\\gamma)$-exponential decay error scales like $O(1/n)$ for $d > \\log^{1/\\gamma}(n^{1/\\beta})$. In order to establish such results, authors prove that eigenvectors corresponding to small eigenvalues are completely incoherent/delocalized i.e. have bounded entries of size $O(1/\\sqrt{n})$. Technical novelty stems from the fact that entries of the kernel matrix are dependent and have non-zero mean. Strengths: 1) This is a first result showing entrywise error guarantees for low rank approximation of kernel matrices. 2) Proof sketches of two main theorems are clear and easy to follow. 3) Strongest technical contribution of this paper is proof given in Appendix D that, simply speaking, shows that the norm of projection of vector 1 on the subspace spanned by $n-d'$ eigenvectors with smallest eigenvalues is vanishing sufficiently fast. 4) Experiments are complementing theoretical results well. Weaknesses: 1) Although authors claim that Lemma 1 is a novel concentration result, it seems to be only a slight generalization of Lemma 68 in Tao and Vu [2011], and is proved essentially using the same argument as that in the proof of Lemma 68. 2) Although I appreciate proof sketches of Theorems 1 and 2 in the main text, I believe it would be more useful to add more information about the proof deferred to Appendix D since this is the most novel and interesting part of the proof. 3) It is not clear whether assumption (R) is necessary and how general it is apart from the two special cases given in Section 3.1. Technical Quality: 3 Clarity: 3 Questions for Authors: 1) Could you elaborate more on tightness of your results? How do they compare with already established results for Frobenius and spectral norm? Are there any known lower bounds for entrywise estimation? 2) Although assumptions (E) and (P) seem to be very natural, I am not sure about assumption (R). Do results hold for any $a$ and $b$ such that $1\\leq a < b/16$? Since the final error bound does not depend on $a$ and $b$, do you think this assumption can be relaxed? 3) Although I think that double descent observation is interesting on its own, the evidence for it is vague. Is this behavior observed for a range of percentile values or does it happen only around 99.95 percentile? Also from figures in the paper seem like it appears only for not very smooth kernel functions. It would be beneficial to have more convincing evidence whether this phenomenon occurs because of your choice of 1) kernels, 2) percentiles, 3) entrywise errors or something else. Typos and other comments (106) maximum entrywise "error" (missing) (186) should be $ \\hat{u}_i(1)$, instead of $ \\hat{u}_l(1) $ (647) later on I would prefer if you do not use $(a,b)$ both for constants in assumption (R) and for vectors in the proof of Theorem 2. In introduction you cite [Lei, 2019] for establishing entrywise error bounds for reinforcement learning - but I could not find any references to RL in that paper. Is this a typo? For example, I believe the following papers are more suitable for that particular reference: - Pananjady, Ashwin, and Martin J. Wainwright. "Instance-dependent ℓ∞-bounds for policy evaluation in tabular reinforcement learning." IEEE Transactions on Information Theory 67.1 (2020): 566-585. - Shah, Devavrat, et al. "Sample efficient reinforcement learning via low-rank matrix estimation." Advances in Neural Information Processing Systems 33 (2020): 12092-12103. - Stojanovic, Stefan, Yassir Jedra, and Alexandre Proutiere. "Spectral entry-wise matrix estimation for low-rank reinforcement learning." Advances in Neural Information Processing Systems 36 (2023): 77056-77070. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed limitations adequately. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: I thank the reviewer for their time and effort reviewing my paper. The reviewer argues that Lemma 1 is only a slight generalisation of Lemma 68 in Tao and Vu (2011), although understanding the conditions which one must place on the mean vector when it is non-zero for the lemma to work required some technical innovation, and could not simply be guessed without working through the mathematical details. I provide more details in the global response. I was happy to read that the reviewer found the proofs in Section D to represent a strong technical contribution and I agree that it deserves more space in the main paper. The main reason that it did not receive it was simply space limitations. If space allows, I will include a sketch of Section D in the main paper in the revision. In response to the concern about assumption (R), I provide additional details and discussion in the global response. I included the double descent observation as a curiosity for readers and other researchers, in the hope it may encourage additional investigation into the phenomenon. The phenomenon does occur at lower quantiles albeit less extremely (even at the median in some cases), and I can include some additional plots in the appendix, however I don't plan to investigate this further in this paper. I thank the reviewer for pointing out some typos which I will correct, and I agree with the comment about using the constants $a$ and $b$ in multiple settings. I will use different letters for the vectors in the proof of Theorem 2. I also thank the reviewer for noticing the incorrect citation in the introduction - it is a typo! I had meant to cite the mentioned Stojanovic et al. paper. Coming to Question 1, I am not aware of any lower bounds for entrywise estimation in this setting, although this would be an interesting direction for future research. The question about how the bounds compare to naive spectral and Frobenius norm bounds is an interesting question. Indeed, one can naively upper bound the entrywise norm $\\|\hat K - K\\|_\max$ by the spectral norm $\\|\hat K - K\\|_2 = \hat \lambda_{d+1}$ (or worse, the Frobenius norm). There are two reasons that this approach does not lead to satisfactory bounds. First if one has a bound of the form $$ \hat \lambda_{d+1} = n \lambda_{d+1} + \textsf{approximation error} $$ where the approximation error is of the order $O(n\lambda_{d+1})$, then to obtain the same rates as Theorem 1, one would require that $d \geq \log^{1/\gamma}(n^{2/\beta})$ under (E), and $d = \Omega(n^{2/\alpha})$ under (P), which are much larger than the $d$ required in Theorem 1. The second stumbling block is that when $d$ is this large, using the best available concentration inequalities (i.e. Theorem 2 of Valdivia (2018)), the approximation error will in fact dominate $n\lambda_{d+1}$ and in this case it may not even be possible to obtain a shrinking error rate for any $d$! References: - Terence Tao and Van Vu. Random matrices: Universality of local eigenvalue statistics. Acta Mathematica, 206(1):127 – 204, 2011. - Ernesto Araya Valdivia. Relative concentration bounds for the spectrum of kernel matrices. arXiv preprint arXiv:1812.02108, 2018. --- Rebuttal Comment 1.1: Comment: Thank you for your reply. I will maintain my score.	Summary: The paper focuses on deriving entrywise error bounds for low-rank approximations of kernel matrices using truncated eigen-decomposition. It addresses the statistical behavior of individual entries in such approximations under assumptions of polynomial eigenvalue decay or exponential decay. The authors also provide empirical studies on synthetic and real-world datasets. Strengths: 1. The paper is clear and well written. The proof seems to be solid. 2. The entrywise error bound is new to the community. 3. The assumptions on polynomial/exponential eigenvalue decay seem general and cover lots of common kernels. 4. Some statements about random matrix theory and concentration inequalities are provided (e.g., Lemma 1), which could be independently useful to the community. Weaknesses: 1. The assumptions on the eigenfunctions corresponding to the assumptions of eigenvalue decay are hard to verify for general kernels, especially the part on the rate of decay ($\alpha >2r+1,\beta> 2s$). Moreover, I wonder if these inequalities are required to guanrantee the uniform convergence of the kernel (I note that $k(x,y)=\sum_{i=1}^{\infty}\lambda_i u_i(x)u_i(y)$ converges uniformly under these assumptions). But in the proof I see these assumptions are used in a way like $\beta-s\ge \beta/2$ (e.g., Line 590). Thus, I am not sure if these asssumptions are necessary for derivation. 2. Assumption (R) seems not natural (why is $1\le a < b/16$ is needed?) and also I do not know how to verify this. Could you provide some examples with $\Gamma_i \neq 0$ under Assumption (R)? 3. The contributions are undetermined. The proof of the main theorem seems to heavily rely on past random matrix theory works (Tao and Vu [2011], Erdős et al. [2009 a,b]). With assumptions (E)/(P) and (R) and the previous works, the proof is straightfoward. And I am not sure about the importance of entrywise error bound. Minor typos: 1. Line 578/588 hypotheses-> hypothesis 2. Line 539/581 miss a period Technical Quality: 3 Clarity: 3 Questions for Authors: 1. (Line 82) What do you mean by "infinite sample limit of $\frac{1}{n}K$"? 2. Could you provide more general examples that completely follow the assumption (E)/(P)? 3. Is this error bound optimal? Are there any lower bound results? 4. Is it possible (or are there any hardness results) to compute or approximate $\text{argmin}_{K':\text{rank}(K')=d}$ $ \\|\|K-K'\\|\|$ w.r.t. the sup norm? 5. Regarding the importance of entrywise error bound, could you provide more concrete examples? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: There is no negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: To begin, I would like to thank the reviewer for taking the time to review my paper. They considered the writing and proofs to be clear and accurate, and the theoretical result to be new to the community. The reviewer mentions that the assumptions on the eigenvalue decay and the eigenfunction growth (assumptions (P) and (E)) are hard to verify for general kernels. Indeed, this is a general shortcoming of many theoretical works on kernel methods which I discuss in the limitations section of the paper, but I will provide some additional context here. As noted by the reviewer, each of these assumptions imply the convergence of the eigendecomposition. Firstly, given the eigenvalue decay condition (either polynomial of exponential), the eigenfunction growth condition is the weakest possible which ensures the convergence of the eigendecomposition. We need explicit decay/growth conditions on the eigenvalues and eigenfunctions for two reasons: firstly, our results rely on a state-of-the-art eigenvalue concentration inequality due to Valdivia (2018, Theorem 2) which we use to prove equation (4), which explicitly require these assumptions. I am not aware of any other such results in the literature which are sufficiently tight for the purposes of this proof. Secondly, our result relies on explicitly bounding $\sum_{j>d} \lambda_j$ for which we use the eigenvalue decay assumption and the proof of the claim proved in Section D repeatedly employs Hoeffding's inequality for which we require the explicit eigenfunction growth assumption (for example, in line 590). The reviewer also mentions that they believe that assumption (R) does not seem natural. A similar comment was made by reviewer 2JjA, so I respond to this in the global response. I also respond to the reviewer's comment that the contributions are undetermined given previous work in the global response. In view of the importance of the entrywise error bound, fairness is one important reason and another is that such bounds can be used to show improved bounds for other algorithms which use these approximations (see response to Q5). I thank the reviewer for pointing out some minor typos, they will be corrected in the revision. In response to the reviewers questions: 1. Viewing the matrix $\tfrac{1}{n}K$ as a operator which acts on vectors $y\in\R^n$ (which we also view as a function from $\mathcal X$ to $\R$, associating each $i$ to $x_i$) with matrix multiplication, i.e. $$ [\tfrac{1}{n}Ky]_i = \tfrac{1}{n}\sum_{j=1}^n K(i,j)y(j) = \tfrac{1}{n}\sum_{j=1}^n k(x_i,x_j)y(j) $$ and taking the $n\to\infty$ limit we have $$ [\tfrac{1}{n}Ky]_i \to \int k(x_i,z) y(z) \:\text{d}\rho(z) = \mathcal K y. $$ 2. E.g. smooth radial kernels follow the eigenvalue decay assumption in (E) (see Belkin (2018)) regardless of the data generating measure. The exponential eigenfunction growth condition is then very mild. 3. I'm not aware of any lower bounds for the max-norm, although this would be an interesting direction for future research. 4. This optimisation problem is likely highly non-convex, but this would be an interesting avenue for future research. 5. Beyond fairness, for example, this max-norm bound could be used to show stability bounds for algorithms which use this low-rank approximation in place of the true kernel matrix. For example, Proposition 2 of Cortes et al. (2010) derives a stability bound for a SVM trained with a kernel approximation in terms of a spectral norm bound on the kernel approximation. For the values of $d$ allowed in our Theorem 1, this bound would not show consistency of the estimator learned with the kernel approximation, however a napkin calculation suggests that one could show consistency using our entrywise bound, although this is beyond the scope of this paper. References: - Mikhail Belkin. Approximation beats concentration? an approximation view on inference with smooth radial kernels. In Conference On Learning Theory, pages 1348–1361. PMLR, 2018. - Cortes, C., Mohri, M., & Talwalkar, A. (2010, March). On the impact of kernel approximation on learning accuracy. In Proceedings of the thirteenth international conference on artificial intelligence and statistics (pp. 113-120). JMLR Workshop and Conference Proceedings. --- Rebuttal Comment 1.1: Title: Reponse Comment: Thank you for your response. After the rebuttal, I go through the Appendix and believe the author has make great efforts in the proof, not as easy as we reviewers thought. But considering the fact that Lemma 1 should be corrected and the assumptions still lack some intuition (since it is hard to give some general examples), my score remain the same. --- Reply to Comment 1.1.1: Comment: I thank the reviewer for taking the time to go through the appendix, and I am happy to see their recognition of the complexity of some of the technical contributions there. I would argue that the correction to Lemma 1 is only minor, does not affect any other areas of the proofs and is now resolved. With regards to the difficultly in verifying the assumptions, this is a widespread limitation of all theoretical works on kernel methods which require knowledge of the spectral properties of the kernel, and I refer the reviewer to Section 2.2 of Barzilai and Shamir (2023) for an extended discussion of this point. Given this, I believe I take the best possible approach to making the assumptions interpretable to the reader. In Proposition 2, I show that in the special case of dot product kernels on the sphere (for which the spectral properties of the kernel can be easily computed), the assumptions can be replaced with a simple, highly-interpretable smoothness assumption on the kernel. I then show experimentally that these results generalise to other data-generating measures using simulations and real datasets (using 4 Matérn kernels of differing smoothness). This is a standard approach, for example in the theoretical deep learning literature (e.g. Jacob et al. (2018), Bietti and Mairal (2019), Bietti and Bach (2020)), and I don't see that there is a better way of doing it. References: - Daniel Barzilai and Ohad Shamir. Generalization in kernel regression under realistic assumptions. arXiv preprint arXiv:2312.15995, 2023. - Bietti, A., & Bach, F. (2020). Deep equals shallow for ReLU networks in kernel regimes. arXiv preprint arXiv:2009.14397. - Bietti, A., & Mairal, J. (2019). On the inductive bias of neural tangent kernels. Advances in Neural Information Processing Systems, 32 - Jacot, A., Gabriel, F., & Hongler, C. (2018). Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31.	Summary: The authors consider the kernel matrices, formed by $n$ vectors i.i.d. drawn from a $p$-dimensional probability distribution $\rho$. Under several assumptions on the associated kernel operator on $L^2_{\rho}$, including the positive definiteness of the kernel and decay condition on the eigenvalues of the kernel, the authors prove an estimate on individual entries of the matrix kernel and those of the low-rank approximation of the kernel. Numerical experiments on the estimate error are done with both synthetic datasets and real-world datasets. Strengths: - The problem is a very fundamental one and it is considered both analytically and numerically. - The writing is very clear and easy to read. Weaknesses: - Lemma 1 is wrong, and thus the proofs of the main results do not work. Consider an extreme case where $a=0$ with probability $1$. Then, since $\pi$ is an orthogonal projection, $\\| \pi_H(a) \\| = 0$ and thus Lemma 1 fails. The main issue is that in the proof of Lemma 1, if $S_1 = \sum p_{ii} (\xi_i^2 - 1)$, then $E[S_1^2] = \sum_{i, j} p_{ii} p_{jj} E[\xi_i^2 - 1] E[\xi_j^2 - 1]$, which is different from $\sum_i p_{ii}^2 E[(\xi_i^2 - 1)^2]$ in (17), unless $E[\xi^2]=1$. As a result, (17) and the estimates on $P(E_+)$ and $P(E_-)$ fail. -> The proofs of the main results would work after modifying Lemma 1 as suggested by the authors. Technical Quality: 2 Clarity: 4 Questions for Authors: - Is it possible to prove Lemma 1 with additional assumptions that are suitable to the current setting? - In the proof of Lemma 1, there are other minor problems listed below. 1) In line 618, $\xi_i \in [0, 1]$ is wrong since the mean $\bar{x}$ is subtracted. 2) In the equation below line 621, why $\\| \pi_H(x)\\|^2 = \\| \pi_H(\bar{x}) \\|^2 + \\| \pi_H(\xi) \\|^2$? 3) In the equation below line 621 and several other places, $X$ should be $x$. Also, $\bar{\xi}$ should be $\xi$. Confidence: 4 Soundness: 2 Presentation: 4 Contribution: 3 Limitations: The work does not seem to have potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: I would like to start by thanking the reviewer for taking the time to work through my paper, in particular for working through the proofs in the appendix and for noticing a mistake in Lemma 1 which I address below. I was happy to read that they consider the problem to be very fundamental and that they found the writing to be very clear and easy to read. The reviewer is correct that there is indeed a mistake in Lemma 1, although this can be rectified with a simple modification of the lemma statement, which does not materially change the proof of the main result. As mentioned by the reviewer, in the proof of Lemma 1, I have implicitly assumed that the variance of each element of $a$ is one (which, in fact, by Popoviciu's inequality is not compatible with the assumption that each element of $a$ is uniformly bounded in $[0,1]$). This can be rectified by simply assuming that each element of $a$ has variance $\sigma^2$ and replacing each $q$ with $\sigma^2 q$ in the lemma. The lemma statement would then read: If $H$ is such that $\\|\pi_H(\bar a)\\| \leq 2(\sigma^2 q)^{1/4}$, then for any $t \geq 8$ $$ \mathbb{P} \left( \left\| \\|\pi_H(a)\\| - \sigma q^{1/2} \right\| \geq t \right) \leq 4 \exp(-t^2/32). $$ Following this modification through in the proofs, $S_1$ defined in line 616 would become $S_1 = \sum_{i=1}^n p_{ii} (\xi_i^2 - \sigma^2)$ and therefore $S_1^2 = \sum_{i,j=1}^n p_{ii} p_{jj}(\xi_i^2 - \sigma^2)(\xi_j^2 - \sigma^2)$. Now for each $i \in \{1,\ldots,n\}$, $(\xi_i^2 - \sigma^2)$ are independent mean zero random variables and therefore for $i \neq j$, we have that $\mathbb{E}\{(\xi_i^2 - \sigma^2)(\xi_j^2 - \sigma^2)\} = 0$. Therefore $$ \mathbb{E}(S_1^2) = \sum_{i,j=1}^n p_{ii} p_{jj}\mathbb{E}\{(\xi_i^2 - \sigma^2)(\xi_j^2 - \sigma^2)\} = \sum_{i=1}^n p_{ii}^2 \mathbb{E}\{ (\xi_i^2 - \sigma^2)^2 \} $$ as required. From hereon, the proof of Lemma 1 follows through. In the main thread of the proof of Theorem 2 from line 206, we can set $\sigma^2$ equal to the variance of $k(x_1, y)$ where $y \sim \rho$. Provided that for all $x_1 \in \mathcal{X}$ we have that $\sigma^2$ is bounded away from zero, then the rest of the proof holds. I will add this non-degeneracy condition as an assumption to the theorem. Additionally, I thank the reviewer for pointing out two more typos. The equation below line 621 should contain an inequality $\leq$ rather than an equality. I will correct the typos relating to $x$ and $\xi$. I will be interested to hear whether the reviewer is satisfied with this correction and if they are, their view of the paper as a whole. --- Rebuttal Comment 1.1: Comment: Thank you for your answers. It seems that the main result can be proved with the modified version of Lemma 1. (However, the modified version of Lemma 1 is less significant than the original one, since the new assumption $\|\pi_H(\bar a)\| \leq 2(\sigma^2 q)^{1/4}$ is stronger.) I will adjust my rating accordingly. --- Reply to Comment 1.1.1: Comment: I am glad to see that the reviewer is satisfied with my proposed modification to Lemma 1 and I thank them for updating their score. In response to their comment about the significance of the new lemma, I would argue that the new lemma is no less significant compared to the original one, especially in the asymptotic analysis considered in this paper. So far, the reviewer has only commented on Lemma 1 in the paper, which is only a small (but necessary) part of this work and is not the main contribution of the paper. I would be interested to know their opinion of the paper as a whole, now that the potential problem has been resolved.	null	null	Rebuttal 1: Rebuttal: I would like to start by thanking all the reviewers for taking the time to work through my paper and to write their reviews. I was happy to read that the reviewers consider the problem to be a very fundamental one, that the main results in the paper are new to the community, that the paper is clear, well-written and easy to follow, and that the experiments complement the theoretical results well. I am grateful to reviewer tJPJ for pointing out a mistake in Lemma 1, which can be rectified with a simple modification of the lemma statement, and which does not materially change the proof of the main result. I outline this in the direct response to their review, and if they are satisfied with the change, I will be interested to hear their opinion of the paper overall. Both reviewers 3p4Z and 2JjA mention that the assumptions (P) and (E) seem very natural and cover many common kernels, however both questioned how natural the assumption (R) is. In this global response, I will provide some high-level intuition as to where this assumption comes from, and why it is necessary. Reviewer 2JjA argues that Lemma 1 is only a slight generalisation of Lemma 68 in Tao and Vu (2011), and therefore does not represent a significant novel contribution, and reviewer 3p4Z argues that with previous works on random matrix theory, the proofs are straightforward. In this global response, I will emphasise why I do not believe this to be the case, and clarify where new technical insights were required to prove the results in the paper. On the assumption (R). To employ Lemma 1, we require that the projection of a constant vector onto a subspace spanned by the eigenvectors of $K$, corresponding to a subset of the eigenvalues with index $i > d$, is sufficiently small. The natural population analog of this is to assume that the projection of a constant function onto the subspace spanned by the population eigenfunctions with index $i > j$, which I denote $\Gamma_j$, decays sufficiently quickly with $j$. This assumption aligns with the intuition that the frequency of each eigenfunction increases as its corresponding eigenvalue decreases. If the first eigenfunction is constant, as is the case in Special Case 2, all the remaining eigenfunctions are orthogonal to it, and so $\Gamma_j$ is zero whenever $j\geq 1$. Reviewer 3p4Z asks if I can provide some examples where this isn't the case. Unfortunately, it is only possible to explicitly compute the eigenvalues and eigenfunctions of very special kernels with respect to very special measures. Typically, it is certain symmetries which allow them to be calculated, and in all the cases I am aware of, $\Gamma_j = 0$ for all $j\geq 1$. The assumption $a < b/16$ in (R) can be seen as a relaxation of this case, and the constant $16$ just pops out of what is possible with the proof technique I am using. I considered simply replacing (R) with the assumption that $\Gamma_j = 0$ for sufficiently large $j$, which is more restrictive but perhaps more natural seeming, however I preferred to stick with the slightly more involved and more general assumption. On the novelty of Lemma 1. Reviewer 2JjA argues that Lemma 1 is only a slight generalisation of Lemma 68 in Tao and Vu (2011), and that since it follows a similar proof technique, does not represent a significant novel contribution. My Lemma 1 generalises Lemma 68 of Tao and Vu (2011) by allowing the mean $\bar a$ of the random vector $a$ to be non-zero. Establishing the condition on the relationship between subspace $H$ and the mean vector $\bar a$ for which such a concentration inequality can be derived was non-trivial and could not simply be guessed without working through the mathematical details. It is this innovation which represents the novelty of the lemma. On the novelty of the main proof. Reviewer 3p4Z argues that with previous works on random matrix theory, the proofs of the main theorems are straightforward. I do not believe this to be the case and I will provide a summary of some of the technical challenges which arose in each part of this proof. Firstly, the core proof technique for the delocalisation bound (Theorem 2) follows a technique used in Tao and Vu (2011) which reduces the problem to (i) controlling the eigenvalues $\hat \lambda_i$ for $i$ in a constructed set $J$, and (ii) proving that the size of the orthogonal projection of a row of $K$ onto the subspace spanned by the eigenvectors $u_i$ for $i \in J$ is sufficiently large. The matrices considered in the paper of Tao and Vu (2011) and this paper are very different. Tao and Vu (2011) consider Wigner random matrices with independent mean-zero, unit variance entries, while we consider kernel matrices which have neither mean-zero nor independent entries, the randomness stemming from sampling the data points used to construct it. For this reason, (i) requires very different techniques - we use a concentration due to Valdivia (2018). (ii) also requires the new concentration inequality (Lemma 1) since our matrix does not have mean-zero entries. This new concentration inequality requires a condition on the mean of a row of the matrix and the aforementioned subspace. Proving that this condition holds in our setting (Section D) requires 7 pages of novel technical derivations. Additionally, understanding the necessary assumptions to make everything fit together is entirely non-trivial. For these reasons, I don't believe that this work can be reasonably described as "straightforward" application of previous work. References: - Terence Tao and Van Vu. Random matrices: Universality of local eigenvalue statistics. Acta Mathematica, 206(1):127 – 204, 2011. - Ernesto Araya Valdivia. Relative concentration bounds for the spectrum of kernel matrices. arXiv preprint arXiv:1812.02108, 2018.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Learning the Optimal Policy for Balancing Short-Term and Long-Term Rewards	Accept (poster)	Summary: This paper introduces a new way to balance multiple rewards with some long-term rewards potentially missing. It does so by using Pareto Policy Learning of optimizing each reward subject up to the tradeoff frontier. This can be more practical than simple linear weighting since the linear weighting strategy applies the constant weight regardless of the amount of conflict between pairs of objectives. Empirically, the papers show that the approach is superior to linear on two synthetic tasks with some real data. Overall I think the paper is promising and adding more realistic empirical evaluation can add values to the current state of the paper. Strengths: - Learning to combine multiple rewards is an important and well-motivated question, and has wide ranging implications. - The method proposed is mathematically sound. The paper shows theoretically that the input parameters can be interpreted as a form of worst case value on each objective. - The paper explains how the most popular approach of linear weighting can fall short, derives the method through first principles, and empirically demonstrates that the proposed method is superior. Weaknesses: - The main weakness of the paper is that the experimentation is rather limited. The experiment uses partial real data with synthetic generation of short-term and long-term rewards. For example, in robotic planning, the authors could show how their approach helps balance the long-term reward (e.g. goal reaching) / short-term reward (e.g. minimizing jerk). This is just an example, but including other more real-world planning and RL problems would seem beneficial. - It seems that compared to linear weighting, the proposed method seeks more short-term reward but is not necessarily better in terms of long-term reward. It may not be a weakness but reading the table does strike me that the method is more “short-sighted.” Technical Quality: 3 Clarity: 3 Questions for Authors: - The separation between short-term and long-term reward is practically meaningful, but mathematically the only difference is that one reward can be missing and the other is fully observable, since the Pareto Policy Learning treats all objectives the same. Do we necessarily need these separate definitions? Can we put everything as a long-term reward that can be missing sometimes? - How are the weighting and preference vectors chosen? Have the author considered running a sweep over different configurations to compare against linear weighting? Apologies if I overlook this detail from the paper. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - The limitation of the overall framework are mentioned but there is no much detail perhaps due to space constraint. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for approving our work and for the helpful suggestions. Below, we address your concerns and questions. >W1: The experiment uses partial real data with synthetic generation of short-term and long-term rewards. Response: Thanks for your comments. We fully agree that applying the proposed method to other domains, such as robotic planning and RL, would be beneficial. However, we would like to clarify that our method focuses on policy learning in causal inference, which has peculiarities that cannot be correctly evaluated using real-world data. Here are the reasons: In causal inference, the primary focus is on estimating treatment effects. Various methods achieve this using observed data under several identifiability assumptions. However, it is impossible to evaluate a method's performance based solely on observed data because the true treatment effects are always unknown in real-world data. Therefore, almost all experiments in causal inference [1-5] are conducted using synthetic data where the true treatment effects are known. Next, we give a detailed explanation. Consider a binary treatment $A\in\{0,1\}$, covariates $X$, and outcome $Y$. Although $X$, $A$, and $Y$ are observed variables, they cannot alone define treatment effects. To define treatment effects, we denote $Y(a)$ as the potential outcome if $A$ were set to $a$ ($a$=0, 1). Thus, each individual has two potential outcomes: $Y(1)$ and $Y(0)$. The observed outcome $Y_i = A_i Y_i(1) + (1-A_i) Y_i(0)$, i.e., the observed outcome is the potential outcome corresponding to the received treatment. Then individualized treatment effect is defined as $$ ITE_i=Y_i(1)-Y_i(0),$$ which measures the treatment effect for individual $i$. Since each individual can only receive one treatment, we observe either $Y_i(0)$ or $Y_i(1)$, but not both. This is the fundamental problem of causal inference. Thus, we cannot obtain the true ITE from the observed data because we cannot simultaneously observe $Y_i(0)$ and $Y_i(1)$. In this article, we study the policy learning problem, which is also a causal problem. Consistent with the above analysis, due to the fundamental problem of causal inference, the rewards for a given policy cannot be evaluated using observed data alone. > W2: It seems that compared to linear weighting, the proposed method seeks more short-term reward but is not necessarily better in terms of long-term reward. Response: Thanks for your insightful comments. In this article, we aim to develop a policy learning approach that balances both short- and long-term rewards, rather than focusing solely on maximizing either of them. To evaluate the proposed method, we mainly use three metrics: S-REWARD, L-REWARD, and $\Delta W$. S-REWARD measures the short-term reward induced by the learned policy, L-REWARD measures the long-term reward, and $\Delta W$ measures the balanced welfare, indicating the policy's ability to trade off between multiple rewards. Clearly, to evaluate an approach that balances both short-term and long-term reward, $\Delta W$ is the most important metric. In the experimental results, the proposed method may not always perform better on L-REWARD, but it shows superior performance for $\Delta W$. This indicates that the linear weighting approach does not achieve a good balance, as it pursues long-term rewards at the expense of short-term rewards. In contrast, our approach proves to be Pareto optimal, pursuing long-term rewards while also maximizing short-term rewards. > Q1: Do we necessarily need these separate definitions? Can we put everything as a long-term reward that can be missing sometimes? Response: Thank you for raising this point. Yes, both long-term and short-term outcomes can be missing, and the proposed policy learning method remains applicable. However, this will bring extra difficulty in identifying and estimating long and short-term rewards. From a high-level perspective, the proposed policy learning approach consists of two steps: * Step 1: policy evaluation, estimating the short and long-term rewards; * Step 2: policy learning, solving optimization problem (3) based on the estimated rewards. The missingness of short and long-term outcomes mainly affects Step 1, which involves identifying and estimating the rewards. Once the rewards are estimated, Step 2 focuses on learning the optimal policy through advanced optimization algorithms. In Step 1, when both outcomes are missing, the estimation method of [6] may be invalid. Developing novel methods for identifying and estimating rewards in this context is an interesting direction, and we leave it for future work. > Q2: How are the weighting and preference vectors chosen? Response: Thanks for your comments. Preference vectors (PVs) are used to quantify an individual's preference for different objectives. In our article, we randomly generate 10 unit PVs (see lines 270-272), ensuring the consistency and comparability of the preference measures. Each component of the PV represents the importance of the decision maker's preference for different objectives. Each PV is considered as a weight vector in the linear weighting method, see lines 275-276. In addition, we added experiments for different number of PVs, see response W5 to reviewer hVvL. [1] Johansson et al. Learning representations for counterfactual inference, ICML 2016 [2] Shalit et al. Estimating individual treatment effect: generalization bounds and algorithms.ICML 2017 [3] Shi et al. Adapting neural networks for the estimation of treatment effects. NeurIPS 2019 [4] Yoon et al. GANITE: Estimation of individualized treatment effects using generative adversarial nets, ICLR 2018 [5] Bica et al. Estimating the effects of continuous-valued interventions using generative adversarial networks. NeurIPS 2020 [6] Wu et al. Policy Learning for Balancing Short-Term and Long-Term Rewards. arXiv 2024 --- Rebuttal Comment 1.1: Title: Kindly Reminder Comment: Dear Reviewer B23s, We are deeply grateful for the time and effort you have invested in reviewing our paper. We have endeavoured to address your questions as thoroughly as possible. As the discussion period draws to a close, we kindly inquire if there are any further concerns or questions that we might assist with. Sincerely, Authors #17045	Summary: This paper attempts to address the challenge of learning the optimal policy for balancing multiple long-term and short-term rewards. The authors point out that the existing linear weighting method leads to a sub-optimal policy. To address this limitation, the authors propose formulating formulate the problem as a multi-objective optimization problem. They utilize the Lagrange algorithm to use preference vectors to solve the formulated multi-objective optimization problem and aim to learn the policy to meet Pareto optimization. In order to decide the preference vectors, the authors propose establishing the connection between the optimization problems and the ε-constraint problem. Experiments on IHDP and JOBS demonstrate the efficacy of the proposed method. Strengths: 1. The multi-object problem is practical in both reinforcement learning and other optimization scenarios. The paper provides a good summary of the limitations of the existing linear weighting method and introduces a novel perspective on solving the problem by resorting to the Lagrange algorithm and Pareto optimization. 2. The author has a solid mathematical foundation and is able to provide detailed mathematical descriptions and solutions to the proposed optimization problems. Weaknesses: 1. The authors point out that the linear weighting method is suboptimal. However, there is no explanation in the method section or corresponding experiments to demonstrate that the proposed method (i.e. DPPL) is optimal. 2. In line 38, the authors claim that when some of the rewards are interrelated, the linear weighting method can only achieve a suboptimal solution. The claim may not be rigorous as the linear weighting method might be able to model the relationship among the rewards. More explanation and experiments are required. 3. In line 95, the definition of Pareto optimality, the condition for Pareto optimality by the author is to find the $\theta$ that makes all $\bar{\mathcal{V}}$ optimal. However, is it possible that the $\theta$ is not optimal for some $\bar{\mathcal{V}}$ but is optimal for the overall $\bar{\mathcal{V}}$? 4. Some mathematical symbols and proprietary terms in the paper are not explained clearly. For example, what does the $e$ in line 110 mean? What does MOP represent? Does MOP represent multi-objective problems? What do $v$ and $R_{+}$ mean in line 171? What does the KKT condition mean? Is it the KKT condition in the Lagrange algorithm? What is the difference between the two descent directions $d_{rt}$ and $d_t$? There are many similar situations in the paper. I suggest providing necessary explanations for each noun and symbol that appears for the first time. 5. In section Simulating Output and section Experimental Details, many parameters are defined by the authors themselves, but most of them do not have reasons or ablation experiments. For example, why is the number of preference vectors 10? In Line 253 to Line 254, why are some parameters truncated normal distributions and some Gaussian distributions? 6. In Table 1 on the L-REWARDS metric, the proposed method is comparable to the linear weighting method. However, the authors claim that for most of the preference vectors, DPPL's solutions have better performance. 7. In Figure 1, it seems that the effect on the $\delta{w}$ from the missing rate and T is not obvious for either the proposed method or LW. More explanation is needed. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the comments above. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Below, we hope to address your concerns and questions. >W1 No enough explanation/experiments to demonstrate that the proposed method is optimal. > W2: When some of the rewards are interrelated, the linear weighting method can only achieve a suboptimal solution. The claim may not be rigorous. More explanation and experiments are required. Response: Thanks for your comments. Since W1 and W2 are similar, we response them together. First, we explain why the proposed method is optimal. In Lemma 2 of Section 3.2 (lines 191-194), we show that the proposed method achieves Pareto optimality. Specifically, a Pareto optimal solution is obtained by solving the optimization problem (5) (lines 181-182) using an iterative gradient-based update rule. Lemma 2 demonstrates that this iterative rule converges to the Pareto optimal solutions. Second, we explain why the linear weighting (LW) method is suboptimal. - Previous studies [1-4] highlight the limitations of LW method in multi-objective optimization. LW method is restricted by assumptions of linearity and convexity, potentially ignoring complex interactions among objectives. - When multiple objectives have nonlinear relationships, the Pareto set (the collection of all Pareto optimal solutions, see line 100) may not be convex. Due to its linear restriction, the LW method may not effectively identify Pareto optimal solutions within non-convex regions, leading to suboptimal solutions. - As illustrated in Figure 2 of [5], when dealing with two objective functions, the LW method may struggle to identify Pareto-optimal solutions when the Pareto set is nonconvex. Finally, we conducted an extra experiment to further illustrate the limitations of LW method and the optimality of our method. We consider two objectives $min_x F(x)=[f_1(x),f_2(x)]^T,$ $$f_1(x)=1-\exp{(-\sum_{i=1}^d(x_d-\frac1{\sqrt{d}})^2)}, f_2(x)=1-\exp{(-\sum_{i=1}^d(x_d+\frac1{\sqrt{d}})^2)},$$ where $x$ is a 20-dimensional covariates and $x_d$ is the $d$-th element of $x$. For the optimization problem above, we know the true Pareto set. For clarity, we present 10 uniformly selected Pareto solutions in Table 1 below, from which we can see that the Pareto set is concave rather than convex. Table 1 \|id\|f1_pf\|f2_pf\| \|-\|-\|-\| \|0\|0.982\|0.000\| \|1\|0.960\|0.041\| \|2\|0.921\|0.153\| \|3\|0.854\|0.313\| \|4\|0.754\|0.486\| \|5\|0.617\|0.647\| \|6\|0.452\|0.777\| \|7\|0.279\|0.870\| \|8\|0.126\|0.930\| \|9\|0.026\|0.966\| Next, we use both the proposed method and the LW method to solve the above optimization problem. Here we select ten preference vectors (they also are the weight vectors in the LW method). The results are shown in Table 2 below, where $(f_{1, LW},f_{2, LW})$ and $(f_{1, our},f_{2,our})$ represents the solution obtained by the LW method and our method, respectively. we can see that the LW method only finds solutions at the endpoints of the Pareto set (i.e., in the convex part), while our method successfully locates all Pareto optimal solutions within the whole region. This indicates that the LW method cannot achieve the Pareto optimal solutions. Table 2 \|PreferenceVector\|$f_{1,LW}$\|$f_{2,LW}$\|\|$f_{1,our}$\|$f_{2,our}$\| \|-\|-\|-\|-\|-\|-\| \|(1.00,0.00)\|0.002\|0.980\|\|0.953\|0.060\| \|(0.98,0.17)\|0.000\|0.981\|\|0.918\|0.160\| \|(0.94,0.34)\|0.000\|0.981\|\|0.823\|0.374\| \|(0.86,0.50)\|0.001\|0.980\|\|0.736\|0.511\| \|(0.76,0.64)\|0.001\|0.979\|\|0.632\|0.632\| \|(0.64,0.76)\|0.979\|0.001\|\|0.632\|0.632\| \|(0.50,0.87)\|0.007\|0.974\|\|0.511\|0.736\| \|(0.34,0.94)\|0.981\|0.000\|\|0.334\|0.844\| \|(0.17,0.98)\|0.981\|0.000\|\|0.228\|0.892\| \|(0.00,1.00)\|0.979\|0.002\|\|0.055\|0.955\| > W3: For Pareto optimality, is it possible that $\theta$ is not optimal for some $\mathcal{\bar V}$ but is optimal for the overall $\mathcal{\bar V}$? Response: Thanks for your comments. We kindly remind the reviewer that there might be a misunderstanding regarding the definition of Pareto optimality. Below, we provide a detailed explanation. A solution is Pareto optimal if it is impossible to improve on objective without worsening other objectives. Typically, multi-objective problems have numerous Pareto optimal solutions, and all of them form the Pareto set. These solutions represent the best trade-offs among different objectives, rather than optimality for a specific single objective. Thus, it is possible that a Pareto optimal solution achieves the optimal solution in some objectives but not for others. > W4: Some mathematical symbols and proprietary terms in the paper are not explained clearly. Response: Thanks for your comments and we apologize for any lack of clarity. Below, we provide a detailed explanation of the mathematical symbols: - We define $e(x) \triangleq P(A=1\|X =x)$ (see line 110). - The MOP represents multi-objective optimization (see line 91). - In line 171, $v$ is an $M$-dimensional vector in the positive real space $R_+^M$, where $M$ is the number of objectives. - The Karush-Kuhn-Tucker (KKT) conditions are a set of necessary conditions that characterize the solutions to constrained optimization problems. The KKT conditions can be seen as a set of constraints derived from the Lagrangian function that must be met for a solution to be considered optimal in a constrained optimization problem. - Both $d_{r_t}$ and $d_{t}$ represent the descent direction at $t$-th iteration, where the former is for seeking the initial solution, while the latter is for finding the Pareto optimal solution. --- [1] Miettinen. Nonlinear multiobjective optimization, 2012 [2] Marler et al. The weighted sum method for multi-objective optimization: new insights, 2010 [3] Censor. Pareto optimality in multiobjective problems,1977 [4] Ngatchou et al. Pareto multi objective optimization, 2005 [5] Shim et al. Pareto‐based continuous evolutionary algorithms for multiobjective optimization, 2002 (Regarding the response to W5-W7, please see global review.) --- Rebuttal Comment 1.1: Title: Kindly Reminder Comment: Dear Reviewer hVvL, We are deeply grateful for the time and effort you have invested in reviewing our paper. We have endeavoured to address your questions as thoroughly as possible. As the discussion period draws to a close, we kindly inquire if there are any further concerns or questions that we might assist with. Sincerely, Authors #17045	Summary: This paper studies the tradeoff between short-term and long-term rewards. The authors formulate the policy learning problem as a multi-objective optimization problem and propose a decomposition-based Pareto policy learning method. I only had experience in reinforcement learning in robotics five years ago. I tried my best to understand the paper, but I am not sure about my rating and comments. Strengths: - This paper studies a quite interesting and important problem, and the proposed methods seem effective on these two benchmarks. - The paper is well-organized, the division is relatively easy to follow, and the proposed method is well-motivated. Weaknesses: - Only the linear weighting method is used as the baseline. I am wondering if there are any other methods that can be used for comparison. If not, why? Since both IHDP and JOBS are widely used. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the weakness. Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your comments and thank you for the helpful suggestions. Below, we hope to address your concerns and questions. > W1: - Only the linear weighting method is used as the baseline. I am wondering if there are any other methods that can be used for comparison. If not, why? Since both IHDP and JOBS are widely used. Response: Thanks for your comments. From a high-level perspective, the proposed policy learning approach consists of the following two steps: * Step 1: policy evaluation, estimating the short-term and long-term rewards $\mathcal{V}(\theta, s_i)$ and $\mathcal{V}(\theta, y_j)$; * Step 2: policy learning, solving the optimization problem (3) based on the estimated values of $\mathcal{V}(\theta, s_i)$ and $\mathcal{V}(\theta, y_j)$. For Step 1, the previous work [1] is well-established. However, for Step 2, it only uses a simple linear weighting method. In this article, we primarily focus on Step 2 and adopt the method of [1] in Step 1, given in Appendix A. As discussed in the second paragraph of the Introduction (lines 32-35), the policy learning problem of balancing short and long-term rewards remains largely unexplored. There is only a single work [1] that addresses this problem, using a simple linear weighting method. Therefore, our work primarily compares with this approach. In addition, we also tried to consider the epsilon constraint method, but it can be converted into a linear weighted method through certain transformations, so we did not choose it as an additional baseline. Nevertheless, to make a more comprehensive comparison, we added two extra estimation methods in Step 1, and compare our proposed optimization method with the linear weighting method in Step 2. Specifically, the two extra estimators are given as * OR estimator $$\hat{\mathbb{V}}(\pi;s)^{OR}=\frac{1}{n}\sum_{i=1}^{n} [\pi(X_{i}) \hat{\mu}_{1}(X_i)+(1-\pi(X_i))\hat{\mu}_0(X_i)]$$ $$\hat{\mathbb{V}}(\pi;y)^{OR}=\frac{1}{n}\sum_{i=1}^{n}[\pi(X_{i}) \hat{\tilde{m}}_1(X_i,S_i)+(1-\pi(X_i)) \hat{\tilde{m}}_0(X_i,S_i)]$$ * DR estimator $$\hat{\mathbb{V}}(\pi;s)^{DR}=\frac{1}{n}\sum_{i=1}^{n}\Big[\pi(X_i)\left(\frac{A_i(S_i-\hat{\mu}_1(X_i)}{\hat{e}(X_i)}+\hat{\mu}_1(X_i)\right)+(1-\pi(X_i))\left(\frac{(1-A_i)(S_i-\hat{\mu}_0(X_i)}{1-\hat{e}(X_i)}+\hat{\mu}_0(X_i)\right)\Big]$$ $$\hat{\mathbb{V}}(\pi;y)^{DR}=\frac{1}{n}\sum_{i=1}^{n}\Big[\pi(X_i)\left(\frac{A_i(Y_i-\hat{\tilde{m}}_1(X_i,S_i)}{\hat{e}(X_i)}+\hat{\tilde{m}}_1(X_i,S_i)\right)+(1-\pi(X_i))\left(\frac{(1-A_i)(Y_i-\hat{\tilde{m}}_0(X_i,S_i)}{1-\hat{e}(X_i)}+\hat{\tilde{m}}_0(X_i,S_i)\right)\Big]$$ The associated experimental results on JOBS are presented below. \| JOBS\| S-Rewards \| \| L-Rewards \| \| $\Delta{W}$ \| \| S-VAR \| \| L-VAR \| \| \|:-----------------:\|:---------:\|:--------:\|:---------:\|:--------:\|:---------:\|:-------:\|:------:\|:-------:\|:-------:\|:------:\| \| Preference Vector \| OURS \| LW \| OURS \| LW \| OURS \| LW \|OURS \| LW \| OURS \| LW \| \| (1.00, 0.00)\| 1612.70 \| 1614.44 \| 1226.81 \| 1212.77 \| 157.11 \| 150.96 \| 60.53 \| 60.12 \| 102.07 \| 87.59 \| \| (0.98, 0.17) \| 1611.94 \| 1604.74 \| 1220.29 \| 1216.57 \| 153.47 \| 148.01 \| 60.19 \| 65.98 \| 100.92 \| 93.77 \| \| (0.94, 0.34) \| 1613.46 \| 1597.26 \| 1220.38 \| 1215.91 \| 154.27 \| 143.94 \| 62.48 \| 77.57 \| 86.22 \| 87.80 \| \| (0.86, 0.50) \| 1614.34 \| 1595.88 \| 1228.16\| 1218.13 \| 158.60 \| 144.36 \| 58.28 \| 82.76 \| 91.15 \| 90.15 \| \|(0.76, 0.64)\| 1616.20 \| 1597.76 \| 1219.02 \| 1219.98 \| 154.96 \| 146.22 \| 56.09 \| 88.02 \| 90.36 \| 92.34 \| \| (0.64, 0.76)\|1614.58 \| 1592.96 \| 1219.77 \| 1217.16 \| 154.53 \| 142.41 \| 56.96 \| 92.09 \| 86.43 \| 90.10 \| \| (0.50, 0.87)\| 1614.64 \| 1586.78 \| 1212.08\| 1223.47 \| 150.71 \| 142.48 \| 60.22 \| 104.24 \| 88.82 \| 96.13 \| \| (0.34, 0.94) \| 1611.14 \| 1587.86 \| 1212.31 \| 1222.62 \|149.08 \| 142.59 \| 52.62 \| 109.01 \| 91.18 \| 91.67 \| \| (0.17, 0.98) \| 1612.24 \| 1580.14 \| 1222.36 \| 1226.82 \|154.65 \| 140.83 \| 54.33 \| 109.42 \| 87.33 \| 95.10 \| \| (0.00, 1.00) \| 1613.48 \| 1586.04 \| 1222.03 \| 1227.29 \|155.11 \| 144.02 \| 60.04 \| 107.85 \| 87.18 \| 90.78 \| \| JOBS \| S-Rewards \| \| L-Rewards \| \| $\Delta{W}$ \| \| S-VAR \| \| L-VAR \| \| \|:---------------:\|:---------:\|:--------:\|:---------:\|:--------:\|:---------:\|:-------:\|:------:\|:-------:\|:------:\|:------:\| \| Preference Vector \|OURS\| LW \|OURS\| LW \|OURS \|LW \| OURS \| LW \| OURS \| LW \| \|(1.00, 0.00)\|1607.82\|1612.26\|1230.10\|1224.90\| 156.31\|155.93\| 63.43\| 59.50\|90.06\| 87.42\| \|(0.98, 0.17)\|1616.08\|1598.50\|1230.45\|1223.66\| 160.61 \|148.43\|58.27\|65.18\| 95.92\| 90.49\| \|(0.94, 0.34)\| 1612.72\|1598.14\|1218.20\|1218.87\| 152.81 \|145.86\|57.97\|80.05\|85.47\| 84.29\| \|(0.86, 0.50)\|1610.58\|1592.74\|1219.66\|1224.71\| 152.47 \|146.08\|57.60\|84.19\|80.43\| 95.71\| \|(0.76, 0.64)\|1608.50\|1594.00\|1228.59\|1221.82\| 155.90 \|145.26\|57.27\|89.49\|81.03 \| 97.67 \| \|(0.64, 0.76)\|1616.34\|1592.82\|1229.27\|1234.92\| 160.16 \|151.22\|58.65\|87.72\| 90.82 \| 95.56 \| \|(0.50, 0.87)\|1612.54\|1593.30\|1217.99\|1227.72\| 152.62 \|147.86\|57.13\|101.56\| 88.11 \| 96.09 \| \|(0.34, 0.94)\|1613.38\|1584.02\|1224.66\|1220.44\| 156.37 \|139.58\|58.84\|104.60\| 90.69 \| 93.02 \| \|(0.17, 0.98)\|1612.54\|1591.02\|1228.40\|1224.59\| 157.82 \|145.16\|56.88\|110.24\| 83.45 \| 82.22 \| \|(0.00, 1.00)\|1613.74\|1588.70\|1218.97\|1237.01\| 153.71 \|150.21\|58.38\|106.48\| 90.94 \| 99.12 \| From the experimental results, we can see that our method achieves better performance on the evaluation metric $\Delta{W}$ (the most important evaluation metric), and our method also performs better in terms of overall performance. [1] Wu et al. Policy Learning for Balancing Short-Term and Long-Term Rewards. arXiv:2405.03329, 2024. --- Rebuttal Comment 1.1: Title: Kindly Reminder Comment: Dear Reviewer xruw, We are deeply grateful for the time and effort you have invested in reviewing our paper. We have endeavoured to address your questions as thoroughly as possible. As the discussion period draws to a close, we kindly inquire if there are any further concerns or questions that we might assist with. Sincerely, Authors #17045	Summary: This paper proposes a framework for solving multi-objective optimization problems: multi-objective optimization problems are divided into sub-problems in different regions by setting different preference vectors. The parameter optimization direction of the sub-problem can be easily solved by transforming it into a dual problem through the KKT condition, and a Pareto optimal solution of the original problem can be obtained by solving the sub-problem. This paper uses this framework to balance the optimal strategy learning under multiple short-term rewards and long-term rewards and achieves better and more stable performance than the traditional linear weighted method in the constructed experimental environment. Strengths: 1. This paper reveals in detail the connection between the proposed method and the linear weighted method and the epsilon-constrained optimization method. Based on this connection, the epsilon-constrained optimization method can provide interpretability for the method in this paper. 2. The method in this paper theoretically overcomes the suboptimality problem of the linear weighted method and avoids the situation where the epsilon-constrained optimization method does not have a feasible solution. 3. This paper obtains better and more stable results than the epsilon-constrained optimization method in the optimal strategy learning problem under multiple short-term rewards and long-term rewards constructed by the author. Weaknesses: 1. This paper mainly proposes an important multi-objective optimization algorithm and compares it with two existing algorithms in theory. However, the title of this paper seems to be just a specific application scenario of the algorithm. In what other scenarios can this algorithm be applied? 2. The experimental part is mainly conducted in a constructed environment, and it is unclear how difficult it is in the field of causal inference. 3. The v in line 171 is missing \bar. In Appendix B, t in line 5 of Algorithm 1 should start from 0. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How to deal with a situation where long-term rewards are missing? Should the data be ignored when solving the network? I think the optimization method proposed in this paper and the connection with related methods are very interesting, and I will improve my score as appropriate. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Have addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your approval of the idea and the novelty of this work and thank you for the helpful suggestions. Below, we hope to address your concerns and questions. > W1: This paper proposes an important multi-objective optimization algorithm. But the title of this paper seems to be just a specific application scenario. In what other scenarios can this algorithm be applied? Response: Thank you for the insightful comments. As the reviewer noted, we propose a general multi-objective optimization algorithm, making the proposed method applicable to various scenarios involving multi-objective balance. Balancing multiple objectives is particularly critical in the field of trustworthy AI, where the goal is not only to achieve desirable prediction performance but also to address other important aspects such as fairness [1,2] and non-harm [3,4]. > W2: The experimental part is mainly conducted in a constructed environment, and it is unclear how difficult it is in the field of causal inference. Response: Thanks for the constructive comments. In the field of causal inference, the primary focus is on estimating treatment effects. With several identifiability assumptions, various methods have been developed to achieve this goal using the observed data only. However, it is impossible to evaluate a method's performance based solely on the observed data. The basic rationale is that the true treatment effects are always unknown in real-world data. Thus, the experiments in almost all the articles in causal inference [5-9] are conducted in a constructed environment where the true treatment effects are known. Next, we give a detailed explanation. Consider the case of a binary treatment variable $A\in\{0,1\}$, $X$ is covariates, and $Y$ is the outcome. Although $X$, $A$, and $Y$ are all observed variables, they cannot alone define treatment effects. To define treatment effects, we denote $Y(a)$ as the potential outcome that would be observed if $A$ were set to $a$ ($a$=0, 1). Thus, each individual has two potential outcomes: one is $Y(1)$ if the individual receives the treatment ($A=1$), and the other is $Y(0)$ if the individual don't receives the treatment ($A=0$). The observed outcome $Y_i = A_i Y_i(1) + (1-A_i) Y_i(0)$, that is, the observed outcome is the potential outcome corresponding to the received treatment. With the notation of potential outcomes, we define the individualized treatment effect as $$ ITE_i=Y_i(1)-Y_i(0),$$ which measures the magnitude of the treatment effect for individual $i$. Since each individual can only receive one treatment, we observe either $Y_i(0)$ or $Y_i(1)$, but not both. This is known as the fundamental problem of causal inference [10]. As a result, we cannot obtain the true ITE from the observed data because we cannot simultaneously observe $Y_i(0)$ and $Y_i(1)$. In this article, we aim to find the optimal policy that maximizes rewards $V(\pi)=E[\pi(X)Y(1)+(1-\pi(X))Y(0)]$, which is also a causal problem. Consistent with the above analysis, due to the fundamental problem of causal inference, for a given policy $\pi$, the rewards cannot be evaluated with the observed data only. > W3: The v in line 171 is missing \bar. In Appendix B, t in line 5 of Algorithm 1 should start from 0. Response: Thank you for pointing this out. We will revise it accordingly. > Q1: How to deal with a situation where long-term rewards are missing? Should the data be ignored when solving the network? Response: Thank you for raising this point and we apologize for the lack of clarity. From a high-level perspective, the proposed policy learning approach consists of the following two steps: * Step 1: policy evaluation, estimating the short-term and long-term rewards $\mathcal{V}(\theta, s_i)$ and $\mathcal{V}(\theta, y_j)$ for a given policy; * Step 2: policy learning, solving the optimization problem (3) based on the estimated values of $\mathcal{V}(\theta, s_i)$ and $\mathcal{V}(\theta, y_j)$. The missingness of long-term outcomes primarily affects Step 1, which involves the the identifiability and estimation of $\mathcal{V}(\theta, s_i)$ and $\mathcal{V}(\theta, y_j)$ (see line 81 of the manuscript). Once $\mathcal{V}(\theta, s_i)$ and $\mathcal{V}(\theta, y_j)$ are estimated, Step 2 focuses on learning the optimal policy by developing advanced optimization algorithms. For Step 1, the previous work [11] is well-established. However, it does not address Step 2, as it only uses a simple linear weighting method. In this article, we primarily focus on Step 2 by proposing a novel optimization algorithm. Additionally, we establish connections with other methods (e.g., $\epsilon$-constraint) to guide the selection and provide interpretation of the preference vectors. For Step 1, we adopt the method of [11] and the associated results are presented in Appendix A. References --- [1] Kusner et al. Counterfactual fairness. NeurIPS 2017 [2] Ethan et al., Fairness-Oriented Learning for Optimal Individualized Treatment, JASA 2023 [3] Kallus N., Treatment Effect Risk: Bounds and Inference, FAccT 2022 [4] Li et al., Trustworthy Policy Learning under the Counterfactual No-Harm Criterion, ICML 2023 [5] Johansson et al. Learning representations for counterfactual inference, ICML 2016 [6] Shalit et al. Estimating individual treatment effect: generalization bounds and algorithms. ICML 2017 [7] Shi et al. Adapting neural networks for the estimation of treatment effects. NeurIPS 2019 [8] Yoon et al. GANITE: Estimation of individualized treatment effects using generative adversarial nets, ICLR 2018 [9] Bica et al. Estimating the effects of continuous-valued interventions using generative adversarial networks. NeurIPS, 2020. [10] Holland. Statistics and causal inference. JASA 1986 [11] Wu et al. Policy Learning for Balancing Short-Term and Long-Term Rewards. arXiv 2024 --- Rebuttal Comment 1.1: Title: Kindly Reminder Comment: Dear Reviewer ouKL, We are deeply grateful for the time and effort you have invested in reviewing our paper. We have endeavoured to address your questions as thoroughly as possible. As the discussion period draws to a close, we kindly inquire if there are any further concerns or questions that we might assist with. Sincerely, Authors #17045	Rebuttal 1: Rebuttal: Dear Reviewer hVvL, we provide the responses to W1-W4 below your Official Review. Here, we further response W5-W7. > W5: In experiment, why choose 10 preference vectors? why are some parameters truncated normal distributions. Response: Thanks for your comments. We would like to clarify that our data generation mechanism follows from previous works [6-8] for ease of comparison. In addition, we further performed two more experiments with varying numbers of preference vectors and different truncation thresholds for normal distributions. First, we generate $K$ unit preference vectors (the same as the article), but with $K$=4 and 12. The results, shown in Tables 3-4 below, indicates that the proposed method stably performs better. Table 3, $K=4$ \|JOBS\|S-Rewards\|\|L-Rewards\|\|$\Delta{W}$\|\|S-VAR\|\|L-VAR\|\| \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PreferenceVector\|OURS\|LW\|OURS\|LW\|OURS\|LW\|OURS\|LW\|OURS\|LW\| \|(1.00,0.00)\|1616.5\|1613.9\|1226.5\|1232.2\|158.9\|160.4\|60.2\|57.8\|94.8\|92.3\| \|(0.87,0.50)\|1606.9\|1599.6\|1226.9\|1222.9\|154.2\|148.6\|60.6\|77.8\|78.3\|92.7\| \|(0.50,0.86)\|1612.5\|1601.3\|1226.5\|1213.7\|156.8\|144.9\|58.4\|82.9\|87.4\|94.4\| \|(0.00,1.00)\|1615.7\|1596.4\|1224.8\|1223.1\|157.6\|147.1\|58.9\|86.3\|86.4\|87.2\| Table 4, $K=12$ \|JOBS\|S-Rewards\|\|L-Rewards\|\|$\Delta{W}$\|\|S-VAR\|\|L-VAR\|\| \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PreferenceVector\|OURS\|LW\|OURS\|LW\|OURS\|LW\|OURS\|LW\|OURS\|LW\| \|(1.00,0.00)\|1610.8\|1614.6\|1231.8\|1232.2\|158.6\|160.7\|56.8\|60.1\|89.8\|87.9\| \|(0.98,0.14)\|1609.7\|1610.7\|1224.6\|1222.9\|154.5\|154.2\|59.1\|60.9\|88.0\|92.3\| \|(0.95,0.28)\|1613.5\|1606.3\|1228.2\|1226.7\|158.2\|153.8\|59.3\|65.0\|84.4\|79.8\| \|(0.91,0.41)\|1615.6\|1598.9\|1223.3\|1231.7\|156.8\|152.7\|58.1\|70.9\|98.6\|88.1\| \|(0.84,0.54)\|1614.1\|1604.9\|1218.6\|1220.7\|153.7\|150.1\|61.4\|65.4\|89.7\|95.1\| \|(0.75,0.65)\|1615.2\|1599.0\|1228.0\|1225.3\|159.0\|149.7\|54.9\|76.2\|86.1\|89.3\| \|(0.65,0.75)\|1616.4\|1596.2\|1226.5\|1218.8\|158.8\|144.9\|61.5\|77.3\|91.9\|95.6\| \|(0.54,0.84)\|1613.4\|1598.5\|1223.1\|1229.3\|155.6\|151.2\|58.6\|83.1\|92.3\|97.6\| \|(0.41,0.91)\|1612.9\|1594.3\|1222.6\|1224.0\|155.1\|146.5\|57.3\|84.4\|87.3\|93.6\| \|(0.28,0.95)\|1613.0\|1596.9\|1230.5\|1218.7\|159.1\|145.1\|61.5\|82.0\|92.4\|95.5\| \|(0.14,0.98)\|1612.2\|1591.8\|1214.4\|1224.3\|150.6\|145.4\|58.2\|86.7\|80.7\|90.2\| \|(0.00,1.00)\|1613.0\|1592.4\|1228.2\|1224.3\|158.0\|145.7\|63.8\|87.6\|84.5\|87.6\| Second, we varied the truncation thresholds in normal distributions, setting $\omega_0\sim\mathcal{N}_{[-2,2]}(0,1)$. The corresponding results are give in Table 5, which show that our proposed method consistently performs better. Table 5 \|JOBS\|S-Rewards\|\|L-Rewards\|\|$\Delta{W}$\|\|S-VAR\|\|L-VAR\|\| \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PreferenceVector\|OURS\|LW\|OURS\|LW\|OURS\|LW\|OURS\|LW\|OURS\|LW\| \|(1.00,0.00)\|1612.9\|1617.4\|1210.7\|1208.7\|147.9\|149.2\|52.5\|54.4\|101.7\|109.4\| \|(0.98,0.17)\|1615.6\|1605.9\|1210.3\|1211.1\|149.1\|144.6\|52.3\|56.1\|95.9\|96.7\| \|(0.94,0.34)\|1609.5\|1607.9\|1215.2\|1199.3\|148.5\|139.7\|59.1\|55.0\|104.0\|104.3\| \|(0.86,0.50)\|1615.1\|1606.3\|1203.8\|1204.7\|145.6\|141.6\|55.0\|64.1\|106.9\|104.6\| \|(0.76,0.64)\|1614.0\|1609.5\|1211.4\|1212.1\|148.8\|146.9\|57.7\|70.8\|91.7\|92.5\| \|(0.64,0.76)\|1611.1\|1609.5\|1213.3\|1196.2\|148.3\|139.0\|52.1\|67.5\|100.5\|105.8\| \|(0.50,0.87)\|1615.6\|1610.7\|1209.0\|1202.9\|148.4\|143.0\|57.5\|75.1\|105.9\|94.5\| \|(0.34,0.94)\|1615.8\|1606.8\|1211.8\|1198.7\|150.0\|138.9\|53.3\|74.1\|96.3\|101.1\| \|(0.17,0.98)\|1617.3\|1600.1\|1215.2\|1210.8\|152.4\|141.6\|54.9\|78.3\|112.4\|98.1\| \|(0.00,1.00)\|1614.1\|1605.5\|1207.7\|1210.0\|147.0\|143.9\|56.0\|72.2\|94.9\|98.1\| > W6: For L-REWARDS in In Table 1, the proposed method is comparable to the LW method. Response: In this article, we aim to develop a policy learning method that balances both short and long-term rewards, rather than focusing solely on maximizing either of them. To evaluate the proposed method, we mainly use three metrics: S-REWARD, L-REWARD, and $\Delta W$. S-REWARD and L-REWARD measures the short and long-term rewards, and $\Delta W$ measures the balanced welfare, indicating the policy's ability to trade-off between multiple rewards. Clearly, to evaluate an approach that balances both short and long-term reward, $\Delta W$ is the most important metric (lines 283-284). In Table 1, the proposed method may not always perform better on L-REWARD, but it shows superior performance for $\Delta W$. This indicates that the LW method does not achieve a good balance, as it pursues long-term rewards at the expense of short-term rewards. In contrast, our method proves to be Pareto optimal, pursuing long-term rewards while also maximizing short-term rewards. > W7: More explanation is needed for the influence of $r$ and $T$ in experiments. Response: The proposed method consists of two parts: (1) estimation of long-term and short-term rewards;(2) learning the optimal policy using DPPL method. In our experimental, different missing rates $r$ and time steps $T$ mainly affect the generation of long-term outcomes, thus impacting the estimation of long-term rewards, i.e., the first part. However, since the proposed DPPL method focuses on the second part. Thus, we may not observe a significant influence of $r$ and $T$ on the results. This indicates the stability of the proposed method. --- [6] Wu et al. Policy Learning for Balancing Short-Term and Long-Term Rewards. arXiv, 2024 [7] Cheng et al. Long-term effect estimation with surrogate representation, WSDM, 2021. [8] Li et al. Trustworthy policy learning under the counterfactual no-harm criterion, ICML, 2023.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
SeeClear: Semantic Distillation Enhances Pixel Condensation for Video Super-Resolution	Accept (poster)	Summary: The authors propose SeeClear for Video Super-Resolution (VSR). SeeClear is a diffusion-based method that improves restoration performance by introducing semantic priors. The authors design an Instance-Centric Alignment Module (InCAM) and Channel-wise Texture Aggregation Memory (CaTeGory) to utilize semantic information effectively. Comparisons on multiple datasets demonstrate that the proposed method achieves state-of-the-art performance. Strengths: 1. The paper introduces semantic priors to achieve spatial modulation and temporal correlation, improving diffusion-based VSR performance. This idea is both reasonable and effective. 2. The authors design the Instance-Centric Alignment Module (InCAM) to align using semantic information, avoiding pixel inconsistencies and being well-suited for diffusion models. 3. Additionally, the authors propose the Channel-wise Texture Aggregation Memory (CaTeGory) to transfer semantic information between different frames. 4. Comparisons with state-of-the-art methods demonstrate the effectiveness of the proposed method. 5. The paper is well-organized, with clear and aesthetically pleasing layouts, figures, and tables. Weaknesses: 1. The method uses pre-trained models to extract semantic information, introducing significant additional computation, which limits the method's applicability. Meanwhile, the paper lacks comparisons of complexity and parameter counts. 2. The method lacks experimental support for some critical hyperparameters, such as the choice of k in InCAM and the number of frames used in SR. 3. The paper proposes using wavelet transform to improve UNet but lacks experimental justification for why simple downsampling and upsampling wouldn't be more efficient. 4. Figure 1, while aesthetically pleasing, is challenging to understand. It would be better to clearly explain the network structure (e.g., Figure 8) and the inference process. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Why do the comparison methods in Table 1 use different numbers of frames? If the same frame is used, what is the performance like? 2. In the ablation study (model 2, Table 2), how to use semantic conditions without MFSA, InCAM, and CaTeGory? 3. In InCAM, what is the value of k in top k, and how is it determined? Is there experimental support for this choice? 4. Others see weaknesses. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors discuss the method's limitations and societal impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Q1. Why do the comparison methods in Table 1 use different numbers of frames? If the same frame is used, what is the performance like? The selection of numbers of frames for training depends on the architecture, such as sliding-window-based (e.g., EDVR-M) and recurrent-based (e.g., IconVSR) methods. The longer temporal information the model can access during training, the better the performance could be generally attained [1, 2]. Nevertheless, the increase in the number of frames will also lead to huge computational overhead. Therefore, we propose to parallelly align inter-frame within the short clip and maintain long-term temporal consistency via additional texture memory. [1] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking Alignment in Video Super-Resolution Transformers. In NeurIPS, 2022. [2] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating Tradeoffs in Real-World Video Super-Resolution. In CVPR, 2022. > Q2. In the ablation study (model 2, Table 2), how to use semantic conditions without MFSA, InCAM, and CaTeGory? The semantic conditions are utilized for the generation of each frame via cross-attention, enabling the SeeClear to be cognizant of the content to be generated. Specifically, the query and key/value are respectively projected from features of SR and instance-centric semantic embedding of segmentation. It embeds the semantic priors into the pixel-level features, similar to the position embedding. > Q3. In InCAM, what is the value of k in top k, and how is it determined? Is there experimental support for this choice? On the one hand, the resolution of the LR frames constrains the value range of k. On the other hand, there is saturation in the choice of k, as shown in Table 3 of the rebuttal document. As long as the instances and backgrounds in the frames can be retrieved completely, a larger k won't lead to performance improvement while increasing the computational cost of the network. > Q4. The method uses pre-trained models to extract semantic information, introducing significant additional computation, which limits the method's applicability. Meanwhile, the paper lacks comparisons of complexity and parameter counts. Albeit an additional segmentation network is introduced, the auto-encoder is replaced by DWT simultaneously, making it possible to increase the network scale and improve the performance. Additionally, the option of the segmentation network is highly flexible, as long as it is transformer-based architecture. Especially with the continuous development of segmentation tasks in recent years, it is highly beneficial for performance gain of SeeClear. The semantic-assisted generative super-resolution comes forth in recent advanced works [1, 2]. Some of them not only require additional training for the semantic extraction network but also add ControlNet as a condition control mechanism, bringing more parameters. To further demonstrate the effectiveness and efficiency of the SeeClear, we also compare it with other representative diffusion-base SR methods in terms of the number of parameters and inference time, please see Q2 of the global response for more information. [1] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. CoSeR: Bridging Image and Language for Cognitive Super-Resolution. In CVPR, 2024. [2] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li and Lei Zhang. SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution. In CVPR, 2024. > Q5. The paper proposes using wavelet transform to improve UNet but lacks experimental justification for why simple downsampling and upsampling wouldn't be more efficient. Please see the answer to Q1 of the global response. > Q6. Figure 1, while aesthetically pleasing, is challenging to understand. It would be better to clearly explain the network structure (e.g., Figure 8) and the inference process. The Figure 1 illustrates the two main networks (i.e., semantic distiller for semantic extraction and pixel condenser for denoising) as a conceptual sketch, rather than a main figure. To further demonstrate the network architecture of the network and the its connection with the devised components, we supplement detailed illustration and explains, as you seen in the Figure 8 and subsequent text in the Section B of supplemental materials. A video clip consisting of five frames is parallelly sampled during the inference process. These LR frames are first fed into the semantic distiller to extract semantic tokens and then corrupted by random noise as the input of the pixel condenser. The pixel condenser iteratively generates the corresponding HR counterparts from noisy LR frames under the condition of LR frames and semantic priors. The specific information flow of the pixel condenser is elaborated in Section B of supplemental materials. Due to the space limitation, we would like to supplement the pseudo-code of the inference in the final edition. --- Rebuttal Comment 1.1: Title: After Rebuttal Comment: Thanks for the rebuttal. The authors provide extensive experiments to demonstrate the effectiveness of the proposed method. I also read the comments of other reviewers. The main concerns are about the lack of experiments, and I think the authors' experiments addressed these issues to some extent. Overall, the authors address my concerns, and I am willing to raise my score to 6.	Summary: The paper introduces a novel video super-resolution framework leveraging semantic distillation to enhance pixel condensation in diffusion-based models. SeeClear addresses stochastic fluctuations by using a Semantic Distiller and a Pixel Condenser to extract and upscale semantic details from LR frames. The framework includes an Instance-Centric Alignment Module and a Channel-wise Texture Aggregation Memory to improve temporal consistency and visual quality. Experimental results demonstrate SeeClear's superiority over state-of-the-art diffusion-based VSR techniques. Strengths: - The combination of semantic distillation and pixel condensation is novel and effectively addresses the challenges of maintaining detail consistency across frames in diffusion-based VSR. - The Instance-Centric Alignment Module (InCAM) and Channel-wise Texture Aggregation Memory (CaTeGory) significantly improve short-term and long-term temporal coherence. - The paper provides extensive experiments to demonstrate SeeClear's advantages over sotas across multiple benchmarks. Weaknesses: - Lack of computation analysis. Diffusion-based methods are often criticized for unbearable inference time, so it would be better to list params, runtime, and FLOPs/MACs for a fair comparison. - Lack of an ablation study on the wavelet transform which is introduced in Section 3.1. - Table 2 is incomplete, making it difficult to assess the effect of the CaTeGory. - The Other baselines such as VRT and IconVSR are also evaluated on Vimeo-90K-T and UDM10 datasets. Could you complete it for a fair comparison? - Figure 7 needs more explanation. Technical Quality: 2 Clarity: 3 Questions for Authors: Diffusion-based models usually show poor performance on PSNR (e.g., StableSR and Reshift), but SeeClear demonstrates a significant improvement. Could you analyze which parts of SeeClear contribute to this improvement? Please refer to the weaknesses part above. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors have largely addressed their limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Q1. Diffusion-based models usually show poor performance on PSNR (e.g., StableSR and Reshift), but SeeClear demonstrates a significant improvement. Could you analyze which parts of SeeClear contribute to this improvement? Please refer to the weaknesses part above. Diffusion-based image super-resolution acquires remarkable perceptual quality thanks to the generative capabilities of the models, yet the consistency with low-resolution images is somewhat overlooked. However, the highly correlated content within adjacent frames of a video supplies more references and constraints for the generation of individual frames. Coupled with InCAM, which activates and associates semantic-related pixels among adjacent frames, it's possible to enhance the perceptual quality of the reconstructed video and strengthen the constraints on fidelity, as shown in Table 2 of the paper. Similar experimental results can also be observed in the comparative method SATeCo. Besides, the pixel condenser employs wavelet transform to change the resolution of features and continually reinject significant information into downsampled features from LR frames, guaranteeing the content fidelity. And the split high-frequency components of LR frames are transmitted to the decoder via skip connections, guiding the generation of details and textures. Additionally, some works [1, 2] also provide theoretical support for this experimental result. In conclusion, we hold the belief that the constraints of multiple frames of the video and the appropriate module designs jointly constrain the solution space of the model. [1] Theo Adrai, Guy Ohayon, Michael Elad, and Tomer Michaeli. Deep Optimal Transport: A Practical Algorithm for Photo-realistic Image Restoration. In NeurIPS, 2023. [2] Dror Freirich, Tomer Michaeli, and Ron Meir. A Theory of the Distortion-Perception Tradeoff in Wasserstein Space. In NeurIPS, 2021. > Q2. Lack of computation analysis. Please see the answer to Q2 of the global response. > Q3. Lack of an ablation study on the wavelet transform. Please see the answer to Q1 of the global response. > Q4. The Other baselines such as VRT and IconVSR are also evaluated on Vimeo-90K-T and UDM10 datasets. Could you complete it for a fair comparison? Due to the lack of diffusion-based VSR evaluated onVimeo-90K-T and UDM10 datasets, we can only just provide the quantitative result generated via our proposed method. \|Datasets\| PSNR $\uparrow$\| SSIM $\uparrow$\| LPIPS $\downarrow$\| \|:---:\|:---:\|:---:\|:---:\| \|Vimeo-90K-T (BI) \|37.64\|0.9503\|0.0982\| \|UDM10 (BD)\| 39.72\| 0.9675\| 0.0609\| > Q5. Figure 7 needs more explanation. Please refer to the answer to the Q3 of the global response. --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal. I raised my rating to borderline accept.	Summary: This paper presents a diffusion-based video super-resolution method, and proposes Instance-Centric Alignment Module and Channel-wise Texture Aggregation Memory. The former leverages a pre-trained open-vocabulary segmentation model (i.e., OpenSeeD), which is utilized to perform alignment within video clips by modulating the spatial and temporal features. The latter leverages channel-wise attention and memory mechnism to better super-resolve the video frames. The results on publich benchmarks indicate that the proposed method achieves state-of-the-art perceptual performance. Strengths: 1. The proposed method achieves state-of-the-art perceptual results on REDS4 and Vid4. Weaknesses: Although the proposed method achieves promising results on the public benchmarks, there are some concerns that greatly affect the rating of this paper. 1. The presentation of the method needs to be improved. The readability of the paper is unsatisfactory. The technical details and the rationale behind it is not clearly described and explained. (a) The main figure (Figure 1) is ambiguous. It is hard to understand the workflow of the framework based on this figure. It is also hard to see the connection among different modules. (b) In the abstract, what is the "conditional video generation" (L6)? I do not see any pre-trained conditional video generation module in the described method. Maybe it should be rephrased. (c) In L206-207, what is the role of "randomly initialized tokens"? And what is specific role of the encoder-decoder module? (d) In L187-188, are the "semantic tokens" actually text embeddings ? What is the difference? (e) In L223, how to divide channels into different groups and what is rationale behind it? (f) It is hard to understand Eq. (16), (17) and (18). From (17) and (18), it seems T_j is used to calculate itself, which is confusing. (g) The choice of mathematical notations is sub-optimal and confusing. (h) In L149, I think the "belta_t" should be "yita_t". 2. The novelty of this paper is limited. (a) Some of the modules are based on existing methods. For example, the way of introducing semantic features is similar to SFT (but no comparison in the paper); the multi-frame self attention is from [21]. (b) The proposed blurring ResShift is a modification version based on ResShift, but the rationale behind it is not fully explained. Also, there is no direct ablation. 3. The comparison with other related methods are not thorough. (a) The authors should explicitly compare with ResShift [33], since residual shifting technique is also exploited (but no citation in L48). Also, there is no comparison with it in Sec. 2.2. (b) The authors should compare with Upscale-A-Video [36], another diffusion-based video super-resolution method. Also, it is recommended to compare the performance of [36]. (c) The authors should compare with SFT[28], another method also leveraging semantic segmentation information. 4. The proposed method is not fully ablated. There is no direct ablation for exploitation of DWT and blurring resshift. 5. Some of the statements could be inappropriate. (a) In L35-36, I think it is hard to reach the given conclusion from [8]. Please elaborate. (b) The naming of "semantic distiller" could be inappropriate. The pre-trained semantic segmentation model is directly leveraged and frozen. I don't see any distillation. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In the abstract, what is the "conditional video generation" (L6)? I do not see any pre-trained conditional video generation module in the described method. 2. In L206-207, what is the role of "randomly initialized tokens"? And what is specific role of the encoder-decoder module? 3. In L187-188, are the "semantic tokens" actually text embeddings ? What is the difference? 4. In L223, how to divide channels into different groups and what is rationale behind it? 5. It is hard to understand Eq. (16), (17) and (18). From (17) and (18), it seems T_j is used to calculate itself, which is confusing. Please elaborate. 6. In InCAM, the way of introducing semantic features seems similar to SFT. Please compare with it and illustrate the significance of the proposed module. 7. Please provide comparison with the following related methods, and illustrate the novelty of the proposed modules. (a) The necessity and rationale of blurring ResShift, and the ablation study. (b) Upscale-A-Video [36]. And please compare performance with it quantitatively. 8. In L35-36, I think it is hard to reach the given conclusion from [8]. Please elaborate. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations in Section D. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your careful reading and detailed comments. We will rectify some confusing statements and formulas in the subsequent edition. Nevertheless, we deem it necessary to highlight our novelty and restate the proposed method. Different from the text in the realm of T2I and segmentation mask in existing generative super-resolution, we pioneer the exploration of the utilization of temporal information based on the semantic similarity for diffusion-based VSR. The instance-centric semantic embeddings and channel-wise semantics are employed to determine the conditional pixels with high quality and semantic similarity from adjacent frames and long-term temporal information, enhancing the generative quality and temporal consistency. > Q1. What is the "conditional video generation"? As depicted in Figure 1 of the rebuttal document, SeeClear consists of a forward diffusion process and a reverse process for VSR. In the diffusion process, patch-level blurring and residual shift mechanism are integrated to degrade HR frames based on the handcrafted time schedule. During the reverse process, a transformer-based network for open vocabulary segmentation and a U-Net are employed for iterative denoising. The former is responsible for extracting semantic embeddings related to instances from LR videos, similar to the process of distillation in physics, and is therefore named the semantic distiller. The latter is utilized to filter out interfering noise and retain valuable information from low-quality frames, similar to the condensation process. All of them are tailored for image processing, and SeeClear takes diverse semantic embeddings as conditions to enable the network to be aware of the generated content and determine the aligned pixels from adjacent frames for the temporal consistency of the whole video. More details of U-Net for denoising are elaborated in Section B of the supplementary material. > Q3 & Q4. Are the "semantic tokens" actually text embeddings? How to divide channels into different groups? The conditions in SeeClear refer to instance-centric semantic embeddings, semantics within the channel dimension and discriminative textures, rather than text or segmentation mask. The semantic distiller comprises three components, i.e., backbone, encoder and decoder. The encoder takes multi-scale features from backbone as input and models the relationship among pixels via self-attention. Randomly initialized tokens serve as the input of the decoder together with the dense features yielded by the encoder, which integrate instance-related semantic information through cross-attention. Eventually, those semantic tokens are utilized to predict classes, bounding boxes, and masks by simple modules like MLP. Besides, we dig abundant semantics within the channel dimension via CaTeGory. For example, different channels response to edges/corners in shallow layer, and deeper layers further extract complex patterns by combining the information of different channels in the previous layer. CaTeGory consists of semantic and texture embeddings, which are zero-initialized as parameters of network. Throughout the training process, semantic embeddings iteratively aggregate the channel semantic from instance-centric semantics via cross-attention, and texture embeddings formulate representative textures from the multi-scale features of the U-Net. To establish the one-to-one correspondence between channel-wise semantics and textures, CaTeGory correlates semantic and texture embeddings through element-wise multiplication. > Q2 & Q6. What is the role of "randomly initialized tokens"? And what is specific role of the encoder-decoder module? In InCAM, the way of introducing semantic features seems similar to SFT. To effectively utilize these semantic conditions, InCAM is inserted before the spatial self-attention of U-Net, aiming at performing inter-frame alignment in the semantic space implicitly. Distinct from SFT, which merely generates coefficients for feature modulation based on class priors derived from the segmentation mask, InCAM utilizes two types of semantic priors, i.e., semantic embeddings and dense features. Among them, dense features and those of VSR are jointly employed for the generation of modulation coefficients, enabling the modulated features to lie in the domain between the segmentation and the super-resolution, narrowing the gap between semantic priors and pixel features. Moreover, the semantic embeddings are not only used as conditions for the generation of each frame, but also can locate pixels related to specific instances within the frame. Therefore, InCAM incorporates the instance encoder-decoder (L206-207) to yield clip-wise semantics based on the instance-centric semantics of adjacent frames, which associate the same instances and align the semantically related pixels between adjacent frames. The random initialized tokens input in the decoder will serve as clip-wise semantics after the iterative interaction along temporal axis. From this perspective, InCAM also possesses the ability of inter-frame alignment that SFT lacks. > Q5. It is hard to understand Eq. (16), (17) and (18). From (17) and (18). Please elaborate. Please see the official comment. > Q7 & Q8. Please provide comparison with the following related methods. I think it is hard to reach the given conclusion from [8]. Please elaborate. Due to the unavailability of Upscale-A-Video’s code and the relevant datasets (e.g., REDS30), a fair comparison cannot be conducted. But we are willing to carry out the comparison in our subsequent versions once the code and related datasets are released. While SFT is an image-level SR method and lacks quantitative results, thus we add other representative methods in the response to Q2 for Reviewer NXC1. Please refer to the answer to Q3 in the global response for the elaboration of the given conclusion. --- Rebuttal 2: Title: Reformulation of Eq. (16)-(18). Comment: We reformulate Eq. (16)-(18) as follows: $$ \left(C_j, T_j\right)=\mathcal{M}\left(\bar{C}_j, \bar{T}_j,\left\\{\bar{F} _{i, k}\right\\} _{k=1}^4\right) $$ $$ \widehat{T}_j=\bar{C}_j \times \bar{T}_j $$ $$ T_j=\operatorname{SA}\left(\operatorname{CA}\left(\widehat{T}_j,\left\\{\bar{F} _{i, k}\right\\} _{k=1}^4\right)\right) $$ where $\bar{C}_j$ and $\bar{T}_j$ respectively denote zero-initialized channel-wise semantics and textures of j-th group. $\widehat{T}_j$ is the texture embedding that possesses one-to-one correspondence with semantic embedding via matrix multiplication. --- Rebuttal 3: Title: Response to the rebuttal Comment: Q3: The answer is still unclear. I still do not unstand the meaning of "semantic tokens" (O) and what are the relation with the visual features (F). Q4: I do not see any direct answer to the question of how to divide channels. Q8: The answer to Q3 seems to have no relation with Q8. I would like to maintain my current rating. --- Rebuttal Comment 3.1: Comment: Thanks for your valuable comments. We provide more materials and details for understanding semantic tokens and the way to divide channels. Besides, there’re sufficient theories and experimental results to demonstrate the necessity of combination of blurring and additive noise in the diffusion process for SR, i.e., Blurring ResShift. We have also provided detailed ablation experiments and explanations to validate the significance of the proposed Blurring ResShift during the rebuttal period. > Q3: The answer is still unclear. I still do not unstand the meaning of "semantic tokens" (O) and what are the relation with the visual features (F). The same terms are employed in previous works [3, 5]. Encoder-decoder based transformer architecture for segmentation has been prevalently utilized in recent years, and the concept of semantic token is blatantly simple to understand and widely accepted. We firmly don't think it absolutely constitute a reason for rejection. As clearly stated in the rebuttal, transformer-based segmentation network [1, 2, 3, 4] requires two types of input, i.e., images and learnable tokens, which is well-understood in the realm of visual tasks and has also been applied in numerous image segmentation tasks. The encoder of the transformer extracts multi-scale features from images via self-attention (Refer to Fig. 2 and its caption in [1]), yielding visual features (F). Meanwhile, the decoder of the transformer establishes relationships between learnable tokens and multi-scale features (Refer to “box embeddings”, “encoded image” and “Resnet features” of Fig. 8 in [1]), generating semantic tokens (O) that can retrieve pixels related to instances and background from visual features (Refer to “object queries” of Fig. 2 in [1]). They are combined to produce the final segmentation mask through MLP or element-wise multiplication. [1] Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers[C]//European Conference on Computer Vision (ECCV). Cham: Springer International Publishing, 2020: 213-229. [2] Cheng B, Misra I, Schwing A G, et al. Masked-attention mask transformer for universal image segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2022: 1290-1299. [3] Zou X, Dou Z Y, Yang J, et al. Generalized decoding for pixel, image, and language[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2023: 15116-15127. [4] Li X, Ding H, Yuan H, et al. Transformer-based visual segmentation: A survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2024. [5] Zhang H, Li F, Zou X, et al. A simple framework for open-vocabulary segmentation and detection[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). 2023: 1020-1031. > Q4: I do not see any direct answer to the question of how to divide channels. The direct answer has been elaborated in the second paragraph beginning with “we dig abundant”，but we’d like to explain it in a simple way. The cluster of channel relies on two zero-initialized parameters of network, i.e., channel-wise semantics ($C_j$) and textures ($T_j$). The semantic parameters initially model the relationship with channels of instance-centric tokens ($O_i$) via cross-attention along channel dimension instead of vanilla spatial cross-attention, which is formulated as: $$ Q=C_j W^Q,K=O_i W^K,V=O_i W^V $$ $$ A=\operatorname{softmax}((Q^T K)/\sqrt{d}) $$ $$ \widetilde{C}_j =AV+C_j $$ After stacked cross-attention and residual connection, the semantics contained in channels of instance-centric semantic tokens are resembled into semantic parameters, thus grouping different channel-wise semantics. Then, the texture parameter is combined with semantic parameter via matrix multiplication. It appoints each channel semantic to a corresponding texture. The texture parameter assembles valuable textures from multi-scale features via cross-attention, similarly to the above process. The motivation and technical details are elaborated, we think that you should read our paper and rebuttal more attentively. > Q5: Where is the "official comments"? You may not be able to access the reformulation of Eq. (16)-(18) provided during the rebuttal period due to system. We’d like to exhibit the same contents here. We reformulate Eq. (16)-(18) as follows: $$ \left(C_j, T_j\right)=\mathcal{M}\left(\bar{C}_j, \bar{T}_j,\left\\{\bar{F} _{i, k}\right\\} _{k=1}^4\right) $$ $$ \widehat{T}_j=\bar{C}_j \times \bar{T}_j $$ $$ T_j=\operatorname{SA}\left(\operatorname{CA}\left(\widehat{T}_j,\left\\{\bar{F} _{i, k}\right\\} _{k=1}^4\right)\right) $$ where $\bar{C}_j$ and $\bar{T}_j$ respectively denote zero-initialized channel-wise semantics and textures of j-th group. $\widehat{T}_j$ is the texture embedding that possesses one-to-one correspondence with semantic embedding via matrix multiplication. --- Rebuttal Comment 3.2: Comment: > Q7: Still no further explanation on the significance of the proposed blur ResShift. We conducted comprehensive ablation experiments of Blurring ResShift in the complementary materials and presented it in the attached rebuttal document again. As shown in Table 2 of the rebuttal document, the variation in the intensity of blur (Line 1-3) affects the fidelity and perceptual quality. Among them, there is no blurring when $\sigma_B^2=0$, and the greater the value of $\sigma_B^2$, the greater the blurring intensity. It can be observed there is a 0.96 dB improvement in PSNR and the value of LPIPS ranges from 0.2096 to 0.2067 with the increasement of $\sigma_B^2$. Although the PSNR if not good enough when $\sigma_B^2=2$, the incorporation of SeeClear enables it to achieve a double-win in terms of both PSNR and LPIPS, such as the alternation of spatial self-attention and channel-wise self-attention, wavelet-based downsampling, etc. We believe that the sufficient experiments demonstrate the effectiveness and necessity of Blurring ResShift. > Q8: The answer to Q3 seems to have no relation with Q8. To answer the question you previously asked in Q8, we offer explanations from two perspectives. Firstly, there are less high-frequency information in LR frames compared to HR frames [1], and further advancing DDPMs requires finer high-frequency details prediction [2]. Besides, diffusion-based SR methods tend to generate rather different outputs for the same LR image due to inherent stochasticity. To address these issues, an ideal diffusion process can follow a similar distribution in the low-frequency components and introduce more stochasticity in the high-frequency spectra. In other words, the network is forced to focus more on the feature extraction and modeling in the high-frequency bands, thus generating more pleasant details and textures with consistent structures [3, 4]. Therefore, we visualize the Power Spectral Density of terminated states generated by different diffusion processes in the rebuttal document, as illustrated in Figure 2. Based on the lack of high-frequency information and importance of fidelity, sole blurring or additive noise is not enough for SR, as they respectively eliminate too much low-frequency information and introduces interfered high-frequency information. Secondly, the ablation experiments on the intensity of blurring also validate the significance of combination of blurring and noise, which has been elaborated above. In a nutshell, there are sufficient theories and experimental evidence to support the conclusion. [1] Wang X, Chai L, Chen J. Frequency-Domain Refinement with Multiscale Diffusion for Super Resolution[J]. arXiv preprint arXiv:2405.10014, 2024. [2] Moser B, Frolov S, Raue F, et al. Waving goodbye to low-res: A diffusion-wavelet approach for image super-resolution[J]. arXiv preprint arXiv:2304.01994, 2023. [3] Sun L, Wu R, Zhang Z, et al. Improving the stability of diffusion models for content consistent super-resolution[J]. arXiv preprint arXiv:2401.00877, 2023. [4] Xiao X, Li S, Huang L, et al. Multi-Scale Generative Modeling in Wavelet Domain[J].	Summary: The paper presents a framework for video super-resolution (VSR) that improves temporal coherence and high-resolution detail generation. The proposed method, SeeClear, integrates a Semantic Distiller and a Pixel Condenser to extract and upscale semantic details from low-resolution frames. The framework employs an Instance-Centric Alignment Module (InCAM) and Channel-wise Texture Aggregation Memory (CaTeGory) to enhance inter-frame coherence and incorporate long-standing semantic textures. The methodology also introduces a blurring diffusion process with the ResShift mechanism to balance sharpness and diffusion effects. Experimental results show that SeeClear outperforms state-of-the-art diffusion-based VSR techniques in terms of perceptual quality and temporal consistency. Strengths: 1. The SeeClear framework introduces a combination of semantic distillation and pixel condensation, which significantly enhances video super-resolution. 2. The Instance-Centric Alignment Module (InCAM) and Channel-wise Texture Aggregation Memory (CaTeGory) improve the temporal coherence of the generated high-resolution videos. 3. The integration of blurring diffusion with the ResShift mechanism effectively balances sharpness and diffusion, leading to high-quality detail generation. Weaknesses: 1. While the method demonstrates robust restoration capabilities, it may still struggle with accurately restoring tiny objects or intricate structures, especially under severe degradation conditions. 2. The method has been tested primarily on specific benchmark datasets. Its performance in real-world applications, where video degradation processes are more varied and unpredictable, remains to be thoroughly evaluated. 3. The experiments are not sufficient and should be improved. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. The performance of the proposed method is not significant. In Table 1, the improvement is very marginal or is worse than other methods. Moreover, in Figure 4, the generated texture is comparable to other methods. 2. It would be better to compare more methods. (a) Transformer-based (e.g., VSR Transformer) or RNN-based method (e.g., BasicVSR++); (b) Diffusion-based image restoration methods (e.g., DDRM, DDNM, DeqIR, etc). (c) Compare methods trained with more frames (e.g., VRT-16). 3. The authors should compare the efficiency, including model size, training/inference time and FLOPs. The efficiency comparisons can demonstrate the effectiveness of the proposed method. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Please refer to the details above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Q1. The performance of the proposed method is not significant. In Table 1, the improvement is very marginal or is worse than other methods. Moreover, in Figure 4, the generated texture is comparable to other methods. For the sake of fair comparison, SeeClear is trained only on five frames and achieves PSNR/SSIM comparable to those models trained on longer sequences, along with the best LPIPS. SeeClear also acquires the superior PSNR on the Vid4 dataset, surpassing the runner-up [1] by 0.36 dB. It is worth highlighting that the novelty of SeeClear lies in its pioneering exploration of semantic tokens in the activation and association of semantic-related pixels across frames during the inter-frame alignment of diffusion-based VSR, rather than merely the improvement of performance metrics. The instance-centric semantic tokens directly extracted from frames not only stimulate the generation potential of the diffusion model but also avoid the cross-modal alignment between text and images. Besides, the inter-frame related conditional pixels determined by semantic similarity equips the diffusion-based VSR model with cognitive ability along the temporal dimension, reducing the interference of irrelevant pixels and motion estimation errors and enhancing the quality of the reconstructed videos. Moreover, the restoration of tiny objects or intricate structures and real-world VSR demands specific training datasets [2, 3, 4], which are not coincident with the configurations and benchmarks of prevailing VSR. These issues have been thoroughly discussed in Section D of the supplementary materials and are considered as our future undertaking. [1] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. Learning spatial adaptation and temporal coherence in diffusion models for video super-resolution. In CVPR, 2024. [2] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In CVPR, 2024. [3] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In ICLR, 2023. [4] Jiezhang Cao, Yue Shi, Kai Zhang, Yulun Zhang, Radu Timofte, and Luc Van Gool. Deep equilibrium diffusion restoration with parallel sampling. In CVPR, 2024. > Q2. It would be better to compare more methods. (a) Transformer-based (e.g., VSR Transformer) or RNN-based method (e.g., BasicVSR++); (b) Diffusion-based image restoration methods (e.g., DDRM, DDNM, DeqIR, etc). (c) Compare methods trained with more frames (e.g., VRT-16). We add more methods for comparison, including RNN-based (BasicVSR++), Transformer-based (RVRT, VRT) and Diffusion-based (DDNM) methods. All these methods except DDNM are trained with more frames. It can be observed that more complex architectures (e.g., second-order grid propagation) and longer sequence used for training can significantly enhance the quality of reconstructed video but inevitably result in substantial computational overload. \|Methods\|Frames\|PSNR $\uparrow$\|SSIM $\uparrow$\|LPIPS $\downarrow$\| \|:---:\|:---:\|:---:\|:---:\|:---:\| \|BasicVSR++\|30\|32.38\|0.9070\|0.1462\| \|RVRT\|30\|32.75\|0.9113\|0.1410\| \|VRT\|16\|32.19\|0.9005\|0.1544\| \|DDNM\|1\|27.05\|0.7660\|0.2608\| > Q3. The authors should compare the efficiency, including model size, training/inference time and FLOPs. The efficiency comparisons can demonstrate the effectiveness of the proposed method. Please see the answer to Q2 of the global response.	Rebuttal 1: Rebuttal: Dear AC and reviewers, We sincerely thank all reviewers for your constructive comments. We are glad that the reviewers appreciate the novelty (srYj, ERJ1), writing (ERJ1), impressive experimental results (vLVk, srYj, ERJ1, NXC1) and limitations adequately discussed (vLVk, srYj, ERJ1) in the paper. Since reviewer vLVk, reviewer srYj, and reviewer NXC1 all concern the ablation results for specific components and require supplementing the efficiency comparison with other related methods. Thus, we address these issues in the following parts. > Q1. Ablation study about wavelet transform and Blurring ResShift. We add an ablation study to validate the efficacy of the introduced wavelet transform on the REDS4 dataset. After substituting the wavelet transform with simple downsampling and upsampling, denoted as model #6 in Table 1 of the rebuttal document, a noticeable decline in the values of PSNR/SSIM can be observed compared to SeeClear. It can be drawn that wavelet transform not only changes the resolution of features like downsampling/upsampling but also assists SeeClear in achieving consistency with LR frames in the low-frequency component (indicated by a drop in PSNR value) and efficiently generating high-frequency spectrums via skip connections between the encoder and the decoder (shown by the deterioration of the perceptual metric), thereby further promoting the performance. We also combine different degrees of blur intensity and noise schedule of ResShift to comprehensively verify the necessity of Blurring ResShift in Table 2 of the rebuttal document. It can be concluded that different options of blur intensity impact the trade-off between fidelity and perceptuality (Line 1-3), and the introduction of wavelet transform and self-attention across frequency spectrums can significantly improve both fidelity and perceptuality at the same time (Line 2 v.s. Line 5). > Q2. Efficiency comparisons to other competing methods. We provide a comprehensive comparison of the efficiency between our proposed method and diffusion-based methods in Table 4 of the rebuttal document. It presents the number of parameters of different models and their inference time for super-resolving 512 × 512 frames from 128 × 128 inputs. Combining these comparative results, we draw the following conclusions: i) Compared to semantic-assisted single-image super-resolution (e.g., CoSeR and SeeSR), our proposed method possesses fewer parameters and higher inference efficiency. ii) In contrast to existing diffusion-based methodologies for VSR, SeeClear is much smaller and runs faster, benefiting from the reasonable module designs and diffusion process combing patch-level blurring and residual shift mechanism. > Q3. The rational explanation of Blurring ResShift. Blurring ResShift, a patch-level blurring version of ResShift, is elaborated in the related works, methods, and supplementary materials. Comparisons to ResShift and ablation experiments were exhibited in the main text and supplementary materials. We also supplement additional ablation experiments for the issues that concern the reviewers. Additionally, we analyze the final states of different diffusion processes via the power spectral density, which reflects the distribution of frequency content in an image, as illustrated in Figure 2 of the rebuttal document. It can be observed that IHDM performs blurring globally and has a significant difference in frequency distribution compared to the LR image, while the patch-level blurring is closer to the frequency distribution of the LR. On this basis, SeeClear further introduces residual and noise. Compared to ResShift without blurring, the diffusion process adopted by SeeClear makes the image more consistent with the LR in the low-frequency components and introduces more randomness in the high-frequency components, compelling the model to focus on the generation of high-frequency components. Pdf: /pdf/b049f490ffd7336b9ca6bd706c947d9e9b545464.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Exploring Context Window of Large Language Models via Decomposed Positional Vectors	Accept (spotlight)	Summary: This paper disentangles positional vectors from the hidden states of a pretrained Transformer language model to facilitate the understanding of length extrapolation. After a series of analyses, this paper proposes two context extending techniques. Experiments show that the proposed methods lower the perplexity on the task of language modeling. Strengths: It's always good to have a mechanistic interpretability view of the hidden states of language models. The findings presented in this paper might inspire follow-up work along this direction. Weaknesses: The experiments presented in the current draft are not convincing enough to me. See questions below. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Instead of continue training from the tinyllama model, I think training models from scratch using the 50B tokens budget will make the results in this paper more convincing. This is because you can get rid of the ripple effect of the originally used rope positional embeddings. Maybe your models were trying to unlearn rope during the continue training stage? 2. Apart from testing perplexity scores on the task of language modeling, I highly recommend the authors adding the experiment of needle in a haystack, otherwise I do not know if the models are really using all the tokens. 3. How do you decide the values of alpha and lambda in section 4.1 and 4.2? In addition, the temperature scaling technique was also used in several other places [1, 2] with explanations of how they did temperature selection. [1] YaRN: Efficient Context Window Extension of Large Language Models [2] Attention Alignment and Flexible Positional Embeddings Improve Transformer Length Extrapolation Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: See above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments! # Q1: Training from Scratch We initially performed from-scratch pretraining on smaller models and found that the properties of the positional vectors were largely similar to continually-trained models, but the models trained from scratch had inferior performance. To address your question, we pretrained TinyLlama from scratch with the same configuration, but using different positional encodings and attention patterns. Due to time and resource constraints, we only trained TL-NoPE-new for 50B tokens, while the other configurations were trained for only 5B tokens. * Formation of positional vectors for TL-NoPE: For the TL-NoPE-new trained from scratch, the formation of its positional vectors is similar to the continually trained one: the initial tokens exhibit distinct positional information after the first layer (as shown in Figure 1 in the PDF). As the table below shows, removing the positional information of the initial tokens significantly harms the performance, indicating that the flow of this positional information facilitates the formation of subsequent tokens' positional information. \| \| \| original \| w/o value \| \| w/o positional vector \| \| w/o positional basis \| \| w/o semantic basis \| \| \|----\|----\|----\|----\|-----\|--\|-----\|-----\|-----\|--------\|-----\| \| position \| - \| - \| 0-4 \| 32-256 \| 0-4 \| 32-256 \| 0-4 \| 32-256 \| 0-4 \| 32-256 \| \| TL-NoPE-new \| simi \| 1 \| 0.70 \| 0.95 \| 0.69 \| 0.95 \| 0.41 \| 0.93 \| 0.99 \| 1.0 \| \| TL-NoPE-new \| ppl \| 11.03 \| 224.74 \| 22.36 \| 263.53 \| 20.91 \| >1000 \| 21.78 \| 11.66 \| 12.699 \| * Formation of positional vector for TL-NoPE-Window-new: For window attention with NoPE, the positional information flow from initial tokens to subsequent tokens also occurs, and the distinct positional vectors gradually propagate across both windows and layers. The distinct positional vectors of the initial layers are shown in the table below. \| layer \| 1 \| 2 \| 3 \| 4 \| 5 \| 6 \| 7 \| \|--\|---:\|----:\|----:\|-----:\|-----:\|-----:\|-----:\| \| Distinct Positional Vectors \| 79 \| 561 \| 922 \| 1303 \| 1682 \| 2038 \| 2048 \| * Effect of positional vectors on attention: we subsequently remove different components from the query and keys (shown in Figure 2 in PDF), and observe that after removing the positional vectors or positional basis, the long-term decay and attention sink all disappear. * Positional vectors change beyond the context window: as shown in the following table, with inputs exceeding the context window (2K), only models with consistent positional vectors exhibit length extrapolation capacities. In addition, when we employ attention scaling on TL-NoPE, its PPL score on 8K length is 14.05, slightly larger than 11.75 in the 2K length of the original model. The similarity between the original and extended models is 0.97, indicating the interpolation of positional vectors. \| models \| PPL(2K) \| PPL(8K) \| similarity(8K) \| \|---\|---\|---\|---\| \| TL-NoPE-new \| 11.75 \| 351.49 \| 0.71 \| \| TL-NoPE-Window-new \| 70.05 \| 105.42 \| 0.80 \| \| TL-RoPE-Window-new \| 24.95 \| 23.43 \| 0.98 \| \| TL-NoPE-Window-80-new \| 69.91 \| 68.12 \| 0.99 \| \| TL-ALiBi-new \| 37.63 \| 36.17 \| 0.99 \| * Effect of OOD Positional Vectors: similarly, the attention sinks and logits of positional vectors of TL-NoPE-new also change after exceeding the context window. \| model \| context window \| property \| 2048 \| 4096 \| 8192 \| \|--\|---\|----\|-----:\|-----:\|----:\| \| TL-NoPE-new \| 2048 \| attention sink \| 0.17 \| 0.02 \| 0.0006 \| \| TL-NoPE-new \| 2048 \| logits similarity \| 1 \| 0.97 \| 0.9 \| * Proposed Methods: Finally, we test our proposed methods with the from-scratch trained TL-NoPE-new and TL-NoPE-window-new, and observe similar phenomenons. Both positional vector replacement and attention window extension can achieve context window extension. However, due to the properties of these models, the hyper-parameters and the original performance are different. \| Model \| Interpolation Method \| Factor \| 2048 \| 4096 \| 6144 \| 8192 \| \|---\|---\|---\|---:\|---:\|---:\|---:\| \| TL-NoPE-new \| - \| - \| 29.0 \| 110.7 \| 634.8 \| 1078.5 \| \| \| Attention Scaling \| lambda = 1.2 \| 29.0 \| 27.3 \| - \| - \| \| \| \|lambda=1.5 \| 36.8 \| 41.6 \| 44.4 \| 46.2 \| \| \| Positional Vector Replacement(ours) \| r=2.5, alpha=0.8 \| 32.8 \| 29.5 \| - \| - \| \| \| \| r=5, alpha=0.8 \| 57.4 \| 50.8 \| 40.0 \| 44.2 \| \| TL-NoPE-Window-new \| - \| - \| 112.7 \| 141.5 \| 149.7 \| 149.7 \| \| \| Attention Window Extension(ours) \| r=2,lambda = 1.1 \| 116.5 \| 114.7 \| 139.6 \| 151.7 \| \| \| \| r=4,lambda = 1.2 \| 122.3 \| 124.4 \| 123.3 \| 126.0 \| # Q2: Absence of Needle in the Haystack We initially explored the "needle in a haystack" task. However, due to the poor performance of TinyLlama, the models we continued to pre-train exhibited diminished retrieval capabilities. In our tests, even with an input text length of 200 tokens, the models failed to generate accurate responses, regardless of the positional embeddings and attention patterns employed. Given the difficulty of the "needle in a haystack" task, we did not include the outputs in our paper. # Q3: How to Decide Values of Alpha and Lambda For the alpha value in positional vector replacement, we conducted experiments in Appendix E according to the effective interpolation ratio and PPL. We tested different combinations and identified the optimal interpolation ratio, replacement layer, and interpolation times (alpha). For lambda in the attention window extension, we similarly used a PPL-based search to determine the scaling factor. We will further point out how to decide these hyper-parameters and add citations in the final version of our paper. --- Rebuttal Comment 1.1: Title: Thank you for the rebuttal Comment: Q1 and Q3: Thank you for addressing the concerns. We are good now. Q2: I believe there should be other methods to test whether the model is truly utilizing long-context information. I would assume TinyLLama still has a decent in-context learning ability given its size. If so, one experiment I can suggest is: Take a segment (S) of natural text with length L. Repeat it N times and concatenate them to form an artificial sequence (A) with a length of LN. Feed A into your length-extended model and observe the output. If the model learns to use long-context information, it should output the first token of the original short sequence (S[0]). Note that S[0] is not a natural continuation of A; the model will only output S[0] if it leverages the long artificial in-context examples provided. You can then experiment with different values of L and N to demonstrate the model's ability to process long-context information. The experiment above is just one example. Feel free to devise other methods that can effectively demonstrate the long-context processing capabilities of your model. --- Rebuttal 2: Title: Our methods can keep the long context utilization capactity of LLMs to some extent. Comment: Thank you for your response and valuable suggestion! To assess the long-context utilization capability, we have adopted your recommendation and evaluated the LLMs using ICL. Specifically, we randomly sampled consecutive substrings from the RedPajama dataset, with lengths of 20, 50, and 200 tokens, respectively. These substrings were then repeated to achieve varying lengths, and we evaluated the accuracy of the models' ability to repeat the first tokens of substrings correctly. The results are presented in the following table. \| Total Length(LN) \| - \| \| 1K \| \| \| 2K \| \| \| 4K \| \| \| 8K \| \| \|-------\|----\|------\|------\|------\|------\|------\|------\|------\|------\|------\|------\|------\|------\| \| Text Length(L) \| - \| 20 \| 50 \| 200 \| 20 \| 50 \| 200 \| 20 \| 50 \| 200 \| 20 \| 50 \| 200 \| \| TL-NoPE-Window \| original \| 0.98 \| 0.98 \| 0.92 \| 0.95 \| 0.97 \| 0.93 \| 0.08 \| 0.07 \| 0.05 \| 0.08 \| 0.07 \| 0.05 \| \| \| attention window extension(r=2, lambda=1.1) \| 0.99 \| 0.96 \| 0.9 \| 0.89 \| 0.89 \| 0.92 \| 0.66 \| 0.89 \| 0.92 \| 0.06 \| 0.07 \| 0.06 \| \| \| attention window extension(r=4, lambda=1.2) \| 0.88 \| 0.74 \| 0.9 \| 0.75 \| 0.7 \| 0.9 \| 0.48 \| 0.7 \| 0.83 \| 0.44 \| 0.59 \| 0.74 \| \| TL-NoPE-new \| original \| 1 \| 0.9 \| 0.86 \| 1 \| 0.9 \| 0.86 \| 0.11 \| 0.49 \| 0.13 \| 0.06 \| 0.07 \| 0.01 \| \| \| positional vector replacement(r=2.5, alpha=0.8) \| 0.51 \| 0.66 \| 0.74 \| 0.46 \| 0.68 \| 0.73 \| 0.69 \| 0.82 \| 0.74 \| - \| - \| - \| \| \| positional vector replacement(r=5, alpha=0.8 \| 0.3 \| 0.26 \| 0.23 \| 0.25 \| 0.21 \| 0.28 \| 0.24 \| 0.22 \| 0.29 \| 0.21 \| 0.49 \| 0.64 \| After exceeding the context window, the performance of models deteriorates significantly, whereas our methods maintain a degree of capability to utilize longer contexts effectively. Specifically, for TL-Window-NoPE, as the substring length increases, the model exhibits slower degradation of performances. For TL-NoPE, with much longer substrings (for example, 200 tokens), the model even demonstrates superior performance at 8K tokens compared to 2K tokens. These enhancements underscore that our models can leverage longer contexts for enhanced performance. --- Rebuttal Comment 2.1: Title: Thank you. I have one more question. Comment: Thank you for the additional experiments. One question: Could you also report the numbers of TL_NoPE with your proposed positional vector replacement method? --- Reply to Comment 2.1.1: Title: Experiments with TL-NoPE. Comment: Thank you for your response! We have evaluated the continually trained TL-NoPE and observed similar phenomena, with the hyperparameters listed in the second column of the table below. As demonstrated in the table, TL-NoPE with our method can utilize the long context to some extent to repeat the first token. Moreover, longer substrings (200 tokens) enhance the model's performance in in-context learning (ICL) with long contexts, which underscores the effectiveness of our methods in leveraging long contexts. Finally, further expanding the interpolation ratio and times can yield better performance on longer inputs, albeit with a slight degradation in performance on shorter inputs. \| Total Length(L*N) \| - \| \| 1K \| \| \| 2K \| \| \| 4K \| \| \| 8K \| \| \|-------------------\|-----------------------------------------------\|------\|------\|------\|------\|------\|------\|------\|------\|------\|------\|------\|------\| \| Text Length(L) \| - \| 20 \| 50 \| 200 \| 20 \| 50 \| 200 \| 20 \| 50 \| 200 \| 20 \| 50 \| 200 \| \| TL-NoPE \| original \| 1 \| 0.99 \| 0.94 \| 1 \| 0.99 \| 0.92 \| 0.11 \| 0.05 \| 0.06 \| 0.06 \| 0.06 \| 0.06 \| \| \| positional vector replacement(r=2, alpha=1.1) \| 0.97 \| 0.97 \| 0.91 \| 0.95 \| 0.95 \| 0.93 \| 0.67 \| 0.74 \| 0.68 \| - \| - \| - \| \| \| positional vector replacement(r=5, alpha=1.3) \| 0.65 \| 0.85 \| 0.86 \| 0.56 \| 0.8 \| 0.87 \| 0.5 \| 0.72 \| 0.84 \| 0.14 \| 0.28 \| 0.68 \| \| \| positional vector replacement(r=6, alpha=1.4) \| 0.61 \| 0.85 \| 0.83 \| 0.46 \| 0.68 \| 0.86 \| 0.43 \| 0.59 \| 0.77 \| 0.23 \| 0.43 \| 0.73 \|	Summary: This paper proposes a mean-based decomposition technique to analyze the formation and effect of positional encodings in LLMs. It then uses these results to propose methods to extend the context window, resulting in models that generalize better to longer texts. Strengths: 1. This paper is very well-written, and the main findings are properly highlighted. 2. This paper not only explains how positional vectors are formed, but also introduces methods to interpolate them based on the findings. 3. Experiments are performed to show that the new methods result in better perplexity scores beyond the context window. Weaknesses: I believe this contribution is novel and insightful enough, and there is no apparent weakness. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. The legends and graphs in Figure 4 overlap. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors have adequately addressed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your helpful comments! We will revise Figure 4 in the final version. --- Rebuttal Comment 1.1: Comment: Thanks for the response!	Summary: This paper dives into the inner workings of how transformer-based language models handle positional information. By decomposing hidden states into semantic and positional vectors, the authors give a series of analysis about how the positional information are encoded and propagated through layers. I believe this work offers valuable insights for understanding the positional information within the transformer architecture. Strengths: Very detailed and clear analysis about how the positional information is encoded and propagated within the transformer architecture, and to the best of my knowledge, I have not seen similar work before. I particularly enjoyed reading Figure 2 and 3, which shows how positional information is propagated through layers and goes beyond the window size, and shows how the manipulation of the positional embedding causally influence the attention patterns, particularly removing the attention sink. Weaknesses: There are few points that I would like to suggest here to make the paper even stronger. - Section 4 feels weak and unnecessary. The performance of replacing the positional vector, if my understanding is correct, seems to be much worse than Dynamic NTK. Given the current mainstream approach is modifying the base of Rope (like YaRN), which is much easier than the approach proposed by this work, I do not think this work’s proposed context extension will be accepted by mainstream model builder. - That being said, I think the in-depth analysis of the positional embeddings are strong enough for me to give an acceptance (I learned a lot from it), so I would strongly suggest removing the content of section 4, and use its space for more experimental analysis of the positional vectors There are a few important problems that I believe will receive the communities’ attention and worth being addressed: - Although this paper shows the positional information can propagate through layers (Figure 2), in practice, many work found that models with window attention cannot pass the needle in a haystack test, and this is why Mistral 1.5 changed its attention back to full attention. It would be insightful if the authors can discuss the relationships between positional information and needle-in-a-haystack performance (because needle in haystack is what makes long-context models useful), i.e., why window attention cannot pass needle in haystack even it does have the correct positional information? - This paper’s analysis is restricted on TinyLLaMA, but TinyLLaMA is not a widely used open-source model, thus casting the doubt whether this discovery of this paper will hold for other model families, particularly mainstream open-weight models like LLaMA 3, Mistral, QWen or Yi. I would strongly suggest the authors verify the behavior of positional embedding on either LLaMA 3, Mistral, QWen, or Yi. Currently I’m given a borderline accept, and I will consider increasing my scores if the authors could either (1) discuss the relationship between positional vectors v.s. needle-in-a-haystack or (2) verify that the properties of positional vectors hold for LLaMA 3, Mistral, QWen or Yi (any 2 out of the 4). Technical Quality: 4 Clarity: 3 Questions for Authors: see the above weakness section Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: see the above weakness section Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments! # W1: Unnecessaries of Section 4 In Section 4, the proposed methods are significant evidence for our analysis of the relationship between positional vectors and the context window. Our experiments substantiate our previous viewpoints. For instance, interpolating positional vectors can extend the context window to some extent. Although it has poorer performance than Dynamic NTK, it provides a new approach to solving the problem of context window extension and can be applied to models without RoPE. In the final version with one additional page, we will adopt your suggestions by supplementing more experiments and condensing the content of this section to some extent. # W2: Failure of Window Attentions in Needle in the Haystack Why can window attention hardly solve the needle in the haystack problem? * Some work has reported that models with full attention can utilize "retrieval heads" to directly retrieve critical information from the needle [1], while window attention cannot directly attend to these tokens. * Models with window attention acquire knowledge only through the indirect transmission of information between windows. However, previous work has highlighted that there is severe attenuation in the information passing between windows, making it difficult to leverage semantic information from distant tokens [2]. * Positional information tends to be a kind of global information at a given position, ensuring the model can produce coherent output, and is orthogonal to the semantic information of the input. To validate this, we first compare the hidden states of tokens when using the full input sequence versus using only the window-sized input at a time. We test TL-RoPE-Window and Mistral-7B-v0.1 with input lengths of 16K and 32K, respectively. We find that the similarity of the last layer's hidden states under both settings was above 0.98, suggesting that information from tokens outside the window has little impact. Additionally, we test the scenario where the "needle" is placed at the beginning but is modified into random tokens from the vocabulary. The hidden states of the outputs remain largely unchanged, further supporting our hypothesis. [1] Retrieval Head Mechanistically Explains Long-Context Factuality [2] Dissecting Transformer Length Extrapolation via the Lens of Receptive Field Analysis # W3: Absence of Some Mainstream LLMs Considering resource constraints, we only continuously pre-trained TinyLlama with modified positional encoding under identical settings for a fair comparison. The mainstream models are all based on RoPE, and continuing pre-training is prohibitively expensive; hence, we only examined the positional vectors of the original Llama3-8B, Yi-9B, and Qwen1.5-7B. * Formation of positional vectors: After the first layer, the initial positions have significantly different positional vectors compared to other tokens, as shown in Figure 1 of the PDF. Furthermore, we remove different components from values at different positions and observe changes in the positional vectors and PPL, as shown in the table below. The findings are consistent with the original paper, indicating that the initial tokens play a critical role in shaping the positional vectors and influencing the model’s PPL. When the positional information of the initial tokens is lost, the model loses its ability to produce coherent output. Moreover, Llama-3-8B and Yi-9B are more sensitive to semantic information than smaller models. \| \| \| original \| w/o value \|\| w/o positional vector \| \| w/o positional basis \| \| w/o semantic basis \| \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| position \| - \| - \| 0-4 \| 32-256 \| 0-4 \| 32-256 \| 0-4 \| 32-256 \| 0-4 \| 32-256 \| \| Llama-3 \| simi \| 1 \| 0.75 \| 1.0 \| 0.75 \| 0.96 \| 0.21 \| 1.0 \| 0.93 \| 0.87 \| \| Llama-3 \| ppl \| 6.74 \| 16.27 \| 6.63 \| 17.20 \| 8.4 \| >1000 \| 6.60 \| 17.6 \| 15.18 \| \| Yi-9B \| simi \|1\| 0.98 \| 1.0 \| 0.92 \| 1.0 \| 0.54 \| 1 \| 0.91 \| 1.0 \| \| Yi-9B \| ppl \| 6.51 \| 8.03 \| 6.56 \| 37.92 \| 6.62 \| >1000 \| 6.52 \| 42.27 \| 7.08 \| \| Qwen1.5-7B \| simi \| 1 \| 0.98 \| 1.0 \| 0.98 \| 1.0 \| 0.74 \| 1.0 \| 1.0 \| 0.99 \| \| Qwen1.5-7B \| ppl \| 7.97\| 9.51\| 8.03 \| 9.51 \| 8.04 \| 217.13 \| 7.98 \| 8.09 \| 8.68 \| * Impact of Positional Vectors on Attention: We examine the changes in attention after removing the positional vectors, as illustrated in Figure 2 of the PDF. We observe that attention sinks and long-term decay are eliminated after removing the positional vectors or basis. * Effect of Positional Vectors Beyond the Context Window: As shown in the following table, after exceeding the context window, models without an extended context window experience a sharp change in positional vectors and PPL, similar to TL-RoPE. By using dynamic NTK, we achieve interpolation of the positional vectors, maintaining high similarity to the original positional vectors. \| model \| context window(W) \| PPL(W) \| PPL(2W) \| Simi(2W) \| PPL(NTK, 2W) \| Simi(NTK, 2W) \| \|---\|---\|---\|---\|---\|---\|---\| \| Llama-3-8B \| 8192 \| 6.74 \| >1000 \| 0.30 \|7.84 \| 0.96 \| \| Yi-9B \| 4096 \| 6.51 \| 102.58 \| 0.24 \| 6.70 \| 0.99 \| * Effect of OOD Positional Vectors: After exceeding the context window, the properties of high attention scores on the first token (attention sinks) in Llama-3 and Yi-9B rapidly disappear, and the logits of the positional vectors differ from those within the window, as shown in the table below. However, the change in logits for Yi is not obvious, although it has formed the same pattern as shown in Figure 5 (right) of the paper. \| model \| context window (W) \| property \| 0~W \| W~1.5W \| 1.5W~2W \| \|---\|---\|---\|---\|---\|---\| \| Llama-3-8B \| 8192 \| attention sink \| 0.47 \| 0.1 \| 0.005 \| \| Llama-3-8B \| 8192 \| logits similarity \| 1 \| 0.9 \| 0.88 \| \| Yi-9B \| 4096 \| attention sink \| 0.68 \| 0.34 \| 0.06 \| \| Yi-9B \| 4096 \| logits similarity \| 1 \| 0.98 \| 0.97 \|	null	null	Rebuttal 1: Rebuttal: Thank you for your insightful comments! The supplementary PDF includes figures that support our rebuttals. Pdf: /pdf/2cd556ebc2c13e813daaf99710c90812816e66c5.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Estimating Transition Matrix with Diffusion Models for Instance-Dependent Label Noise	Reject	Summary: This paper deals with the problem of supervised learning from noisy labels, where the label noise is modeled using instance-dependent label transition probability matrix. Mainly, this work attempts to leverage conditional diffusion model in order to obtain a generative model of transition matrix conditioned on the sample features. To that end, this work first generate pseudo paired samples $( x_i, T_i )_{i=1}^N$ using existing method (VolMinNet). Secondly, a conditional diffusion model is trained that generates $T_i$ given $x_i$. Finally, the classifier is trained taking into consideration the estimated transition matrix from the diffusion model. Strengths: 1. The problem considered is of interest to the broad ML community 2. Adequate experimental settings, baselines, and ablations are provided for numerical validation. 3. The attempt to apply diffusion model is novel. Weaknesses: 1. The technical soundness of the proposed method is questionable. Essentially, the proposed method trains a conditional diffusion model using paired samples $(x_i, T_i)$. If we consider the true transition matrix as $T(x)$ for a sample $x$, then the idea of the proposed method is to train a conditional generative model $p( T(x) \| x )$. There are several issues with this attempt and the proposed implementation: (a) The authors use pseudo transition matrix $T_i$ generated from a sample-independent method (VolMinNet). $T_i$ only depends upon the cluster assignment of $x_i$. The diffusion model, at best, can approximate the conditional distribution $p( T_i \| x_i )$. This has no clear relation to $p(T(x) \| x)$. Therefore, in principle, the transition matrix generated by the trained diffusion model cannot be better than that returned by VolMinNet. (b) Second, the transition matrix is modeled as a deterministic function of sample, i.e., only one $T(x)$ exists for a given $x$. Therefore, it does not make sense to learn a generative model for $p(T(x) \| x)$, since it is a degenerate distribution (probability of all other matrices should be zero except the true $T(x)$). 2. Another hint at why the proposed method should be limited by the pseudo paired sample distribution is that the diffusion model training part (which is ultimately used as transition matrix estimator) does not require available noisy labels. Hence, no extra information can be extracted about the true transition matrix $T(x)$ beyond the information captured by the pseudo paired samples $(x_i, T_i)$. 2. It is unclear where the performance gain in empirical results is coming from. The manuscript does not provide any intuitive or theoretical explanation to justify the quality of their estimator. Moreover, no rationale for the algorithm design is provided. Technical Quality: 1 Clarity: 2 Questions for Authors: Please provide a rebuttal for each of the points in weaknesses section. Confidence: 4 Soundness: 1 Presentation: 2 Contribution: 2 Limitations: Limitations are not adequately discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	null	Summary: This paper focuses on the estimation of the transition matrix with instance-dependent label noise. They used a diffusion model for this estimation. By applying a diffusion process to the transition matrix, the diffusion model is trained to generate transition matrices from a prior distribution. The instance-wise generated transition matrix is then used to train the classifier with a forward cross-entropy loss. The improvement of the method is demonstrated by experiments on benchmark and real-world datasets. Strengths: The instance-dependent label noise scenario is a challenging task. Weaknesses: * The reason for generating the transition matrix using a diffusion model is unclear. * The instance-dependent transition matrix is the target to be estimated, but it is uncertain what role training a diffusion model to generate the transition matrix without a fixed target. * In addition, as mentioned by the authors, the transition matrix must be satisfied: the entries are greater than 0, the row sum is to be 1, and the diagonal entry is typically the largest. However, these considerations have not been taken into account in the construction of the diffusion process. Although a transformation method is proposed in Section 3.4, there is no discussion of how this affects the training of the diffusion model. * Pre-trained features are fed into the diffusion network, but their impact on the diffusion process has not been analysed. This could be seen as providing additional conditional information during the diffusion process, implying that this diffusion model might be a conditional diffusion model. It would be better to discuss these consideration. * In Algorithm 3, it appears that the diffusion model is trained in order to generate the initialized $T_i$. I wonder if the desired training is for the initialized $T_i$ to be generated perfectly as is. This could lead to a transition matrix that might not contain instance-dependent information, raising questions about the mechanism by which diffusion training introduces variance. * The diffusion training seems to take a considerable amount of time, which needs to be analysed. If it takes a long time, the performance improvement may not be significant in comparison. Technical Quality: 1 Clarity: 1 Questions for Authors: Please see the Weaknesses part. Confidence: 4 Soundness: 1 Presentation: 1 Contribution: 1 Limitations: They mentioned the limitations only briefly in the experimental section. I have noted additional limitations that I perceive in the Weaknesses part. Flag For Ethics Review: ['No ethics review needed.'] Rating: 2 Code Of Conduct: Yes	null	Summary: In this work, the authors proposed an approach to estimate the instance-dependent transition matrix in order to reliably learn from noisy labels. The idea is to use a condition diffusion model to estimate the transition matrix by using the pretrained extracted image features as the conditions. Once the transition matrices are estimated, the classifier is learned through the corrected cross entropy loss. Experiments are presented to compare the performance of the approach with other baselines using both synthetic and real noisy datasets. Strengths: The paper is easy to read and notations are clearly stated Weaknesses: The main weakness is the lack of support and discussion in substantiating the idea. Experiments are insufficient to support the claims. Technical Quality: 1 Clarity: 2 Questions for Authors: Questions: 1. A key concern is that estimation of the transition matrix is heavily dependent on the initializations given to the diffusion model learning. The diffusion model intuitively tries to approximate the distribution of its inputs through its forward and reverse process. In the traditional setting, the original image features is the input. But in your case, the initializations estimated through clustering and volmin optimization are the inputs. This part is quite unclear how does it help learn the true instance-dependent transition matrices. 2. In the experiments, in Table 4, I do not see the ablation study with just using the initialized transition matrix and training the classifier, which is important to see the effect of the diffusion model-based learning for the TM. The ablation study corresponding to “w/o diffusion” says that it is using the pre-trained model. 3. Experiment results all look good compared to the baselines, but I do not see any clear intuition/discussion to substantiate this idea of instance-dependent transition matrix estimation Confidence: 5 Soundness: 1 Presentation: 2 Contribution: 1 Limitations: No limitations are discussed Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	null	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Bridging Geometric States via Geometric Diffusion Bridge	Accept (poster)	Summary: The paper introduces the Geometric Diffusion Bridge (GDB), a novel framework designed to generate the evolution of geometric states in geometric (coordinate) systems. GDB uses a diffusion bridge connecting initial and target geometric states with equivariant transition kernels, preserving symmetry and joint state distributions. Furthermore, GDB can use a chain of equivariant diffusion bridges to leverage trajectory data for more accurate dynamic modeling. Strengths: - The presentation of theorems in Section 3.1 is clear and straightforward, establishing a solid theoretical foundation for GDB. The authors effectively derive theorems and integrate them with point cloud states. - GDB demonstrates strong performance across various tasks, including QM9, Molecule3D, and OpenCatalyst IS2RS. Weaknesses: I have no complaints regarding the technical and experimental sections, as they are well-written. However, I wonder existing works, such as [1] and [2], also use diffusion bridges over molecular data. What advantages does your approach have over theirs? [1] Diffusion-based Molecule Generation with Informative Prior Bridges. Lemeng Wu, et al. NeurIPS 2022. [2] DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization. Danyeong Lee, et al. AAAI 2024. Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors note the need of exploring better implementation strategies for their framework to enhance performance. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing both the theoretical analysis and practical effectiveness of our GDB framework. We also appreciate your suggestions which can improve our work further. Our proposed method has the following advantages compared to your list of works [1, 2]. - First, our proposed method can make good use of trajectory data during training (trajectory guidance), which [1, 2] cannot. Trajectory data provide valuable insights into the evolution of geometric states. We conducted a theoretical analysis, which showed that our approach can preserve the joint distribution of the trajectory (Theorem 3.4) with strong expressiveness guarantees (Theorem 3.5). Empirical results on the large-scale real-world benchmark OC22 also verify the superiority of trajectory guidance enabled by our framework. - Second, our GDB framework can preserve the coupling of geometric states, which is also crucial and necessary in modeling the evolution of geometric states (please refer to line 132 in section 3 and Theorem 3.3 for more details). Diffusion bridge works the reviewer mentioned [1, 2] do not guarantee coupling preservation. Our paper provides detailed derivations and theoretical analysis for the coupling preservation of geometric states, which we believe add value to the communities of both geometric deep learning and diffusion bridge-based approaches. Besides these contributions, we also provide detailed derivations and theoretical guarantees for satisfying the symmetry constraints of the equivariant diffusion bridge. Although both the equivariant diffusion process and diffusion bridge have been studied, few works exist to thoroughly introduce methodologies for combining the power from both sides, while our work makes it complete. We will carefully cite the listed works and add the above discussions to the next version of our paper to let the general audience better understand our contributions. [1] Diffusion-based Molecule Generation with Informative Prior Bridges. Lemeng Wu, et al. NeurIPS 2022. [2] DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization. Danyeong Lee, et al. AAAI 2024. We thank you again for your efforts in reviewing our paper, and we have replied to each of your comments. We look forward to your re-evaluation of our submission based on our responses and updated results. --- Rebuttal 2: Title: Looking forward to your re-evaluation Comment: Dear Reviewer dV6w, Thank you for your time and efforts in reviewing our paper. We have carefully responded to each of your questions. Given that the author-reviewer discussion deadline is approaching, we would greatly appreciate it if you could kindly take a look at our responses and provide your valuable feedback. We are more than happy to discuss more if you still have any concerns. Thank you once again and we are eagerly looking forward to your re-evaluation of our work. Paper 15795 Authors --- Rebuttal 3: Title: Official Comment by Reviewer dV6w Comment: Thank you for your reply and I am satisfied with the responses. I will keep my positive score. --- Rebuttal Comment 3.1: Comment: Thanks for your reply. We are happy to hear that all your concerns have been properly addressed. We thank you again for recognizing our contributions and staying positive about our work.	Summary: This paper proposes a generative model for bridging initial and target geometric states using diffusion bridge. This work introduces an equivariant diffusion bridge based on equivariant transition kernels for symmetry constraints. The proposed method was validated on diverse settings including simple molecules and adsorbate-catalyst complex, outperforming previous MLFF baselines. Strengths: - The motivation of using diffusion bridge to bridge initial and target geometrical states is reasonable. - Using diffusion bridge model for equilibrium state prediction and structure relaxation is novel to the best of my knowledge, and the paper shows that GDB significantly outperforms previous methods with diverse datasets. - Equivariant design of bridge process is based on solid theory. - The paper is well written except for some missing relevant works on diffusion bridge. Weaknesses: - Related works on diffusion bridges or diffusion mixtures were not discussed. Diffusion bridges has been studied in [1,2,3,4] with applications to molecules, graphs, point clouds, and images, and more recent works have studied general framework for diffusion bridges [5, 6] which is worth discussing. While GDB has a contribution for using diffusion bridges in new tasks, discussing related works and clarifying the novel contributions is necessary in particular for strengthening the contribution of this work. - Contribution seems limited as using diffusion bridge as generative modeling was already studied [1,2,3,4], in particular deriving diffusion bridges using Doob's h-transform. Designing an equivariant diffusion process (not necessarily bridge) specifically in SE(3) group has been covered in [7,8, 9]. What is the difference of designing equivariant diffusion bridges compared to equivariant diffusion processes? [1] Peluchetti, Diffusion Bridge Mixture Transports, Schrodinger Bridge Problems and Generative Modeling, JMLR 2023 [2] Liu et al., Learning Diffusion Bridges on Constrained Domains, ICLR 2023 [3] Wu et al., Diffusion-based Molecule Generation with Informative Prior Bridges, NeurIPS 2022 [4] Jo et al., Graph Generation with Destination-Predicting Diffusion Mixture, arXiv 2023 [5] Albergo et al., Stochastic Interpolants: A Unifying Framework for Flows and Diffusions, arXiv 2023 [6] Shi et al., Diffusion Schrodinger Bridge Matching, NeurIPS 2023 [7] Xu et al., GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation, ICLR 2022 [8] Xu et al., Geometric Latent Diffusion Models for 3D Molecule Generation, ICML 2023 [9] Yim et al., SE(3) diffusion model with application to protein backbone generation, ICML 2023 Technical Quality: 3 Clarity: 3 Questions for Authors: - What is the reason for using deterministic process (i.e., probability flow ODE) instead of the original stochastic process? Does ODE results in better performance? - Is GDB scalable to geometric states of high dimensions? While analysis on this may not be necessary, it could strengthen the work. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: While the paper discusses future direction for the proposed method, specific limitations of the work is not specified. One potential issue might the scalability of GDB as the model has transformer architecture, and other issue could be long inference time which is a typical problem of diffusion models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing the motivation, contributions, and theoretical analysis of our GDB framework. We also appreciate your suggestions which can improve our work further. Here are our responses to your questions. >Regarding discussions of related works. Thank you for listing these related works [1-6] on diffusion bridges. We agree that adding more discussions on these related works would help the audience to understand our contributions better. We will carefully cite all these works, compare them with our work regarding their contributions, and update these discussions in the next version of our paper. >Regarding our contributions. Our work cannot be considered a simple extension or follow-up work of constructing diffusion bridges with SE(3) symmetry constraint ([1-4,7-9]). As described throughout the whole manuscript, the task we tackle is to capture the evolution of geometric states. This task has unique difficulties compared to conventional generative models, including how to leverage trajectory data and preserve coupling between geometric states over time. Simply combining existing diffusion bridge approaches with equivariant diffusion processes cannot meet the requirement. To the best of our knowledge, our GDB framework is the first approach to leverage the characteristics of diffusion bridges for flexibly incorporating trajectory guidance. We conduct theoretical analysis to show that our approach can preserve the joint distribution of the trajectory (Theorem 3.4) with strong expressiveness guarantees (Theorem 3.5). Empirical results on the large-scale real-world benchmark OC22 also verify the superiority of trajectory guidance enabled by our framework. Moreover, our GDB framework preserves the coupling of geometric states, which is also crucial and necessary in modeling the evolution of geometric states (please refer to line 132 in section 3 and Theorem 3.3 for more details). SE(3) diffusion models [7,8,9] are restricted to transport from the standard Gaussian distribution to the target distribution of molecules or proteins, which are obviously unable to preserve the coupling of geometric states. Previous works on diffusion bridges [1,2,3,4,5,6], which the reviewer has mentioned, do not provide guarantees on coupling preservation, especially for distributions of geometric states. Our paper provides detailed derivations and theoretical analysis for the coupling preservation of geometric states, which we believe add value to the communities of both geometric deep learning and diffusion bridge-based approaches. We will carefully add the above discussions to the next version of our paper to let the general audience better understand our contributions. >Regarding the usage of the ODE sampler. As stated in lines 259-265 in our paper, we leverage ODE solvers for inference due to efficiency considerations. This is motivated by [10], which has also become a common choice in the literature of diffusion models for efficiency sampling. Our framework can also be implemented by using other advanced sampling strategies, which we leave as future works. [10]. Song, Yang, et al. "Score-based generative modeling through stochastic differential equations." ICLR 2021. >Regarding the scalability of GDB to geometric states of high dimensions. Thank you for the suggestion. In Table 2 of our paper, we briefly introduce the types and scales of datasets we used, covering both medium-scale and large-scale real-world benchmarks. Here we present more details on the data dimensions of these benchmarks: - QM9 is a medium-scale dataset that consists of \~130,000 organic molecules (see line 277 in section 4.1). Molecules in this dataset typically consist of 9 heavy atoms C,O,N,F (not counting hydrogens), so the dimension of geometric states is $9\times 3 = 27$. - Molecule3D is a large-scale dataset curated from the PubChemQC project, consisting of 3,899,647 molecules in total (see line 279 in section 4.1). In the Molecule3D dataset, each molecule includes 29.11 atoms on average. The dimension of geometric states is around 87 on average. - The Open Catalyst 2022 (OC22) dataset consists of 62,331 Density Functional Theory (DFT) relaxations, which have great significance for the development of Oxygen Evolution Reaction (OER) catalysts. Each data is in the form of the adsorbate-catalyst complex (see line 305 in section 4.2 for more details), consisting of hundred-scale atoms with periodic boundary conditions being considered. The dimension of geometric states in this dataset is thus in hundred to thousand scales. From the above statistics, we can see that our GDB framework is general to be applied to accurately predict the evolution of both low-dimensional and high-dimensional geometric states. Bridging geometric states of higher dimensions, such as protein conformation states, indeed has great significance in various scientific problems like transition state discovery. In our ongoing work, we successfully extend our GDB framework to this higher-dimensional problem with satisfactory preliminary results, which demonstrate the promising scalability of GDB for bridging geometric states of higher dimensions. >Regarding the discussion of limitations Thank you for the suggestion. For the sake of generality, we do not experiment with advanced implementation strategies of training objectives and sampling algorithms, which leave room for further improvement. Besides, the employment of Transformer-based architectures may also limit the efficiency of our framework. This has also become a common issue in transformer-based diffusion models. We will organize these discussions into a single Limitation section and update it in the next version of our paper. We thank you again for your efforts in reviewing our paper, and we have replied to each of your comments. We look forward to your re-evaluation of our submission based on our responses and updated results. --- Rebuttal 2: Title: Looking forward to your re-evaluation Comment: Dear Reviewer U4NU, Thank you for your time and efforts in reviewing our paper. We have carefully responded to each of your questions. Given that the author-reviewer discussion deadline is approaching, we would greatly appreciate it if you could kindly take a look at our responses and provide your valuable feedback. We are more than happy to discuss more if you still have any concerns. Thank you once again and we are eagerly looking forward to your re-evaluation of our work. Paper 15795 Authors --- Rebuttal Comment 2.1: Title: Kindly request for feedback and reevaluation Comment: Dear Reviewer U4NU, Thank you once again for taking the time to review our paper! As the Reviewer-Author discussion deadline is quickly approaching, we would sincerely appreciate it if you could provide us with further feedback on our responses and kindly reevaluate our work based on our clarification and additional results in our rebuttal. Following your insightful suggestions, we have thoroughly discussed the related works in the rebuttal and clarified our novel contributions. We also comprehensively illustrate GDB's scalability to geometric states of high dimensions. Additionally, we provide additional details and our motivation for the ODE sampler. Based on these additional results and clarifications, we sincerely hope your concerns can be addressed. We always believe that the feedback between reviewers and authors would indeed improve the paper's quality, and we will definitely include the related works discussions from the reviewer's suggestions in the revised paper. It would be really nice to see both of us reach a consensus. We sincerely look forward to your reevaluation and feedback! Best Regards, Paper 15795 Authors --- Rebuttal 3: Title: Kindly Reminder of the Close of the Discussion Period Comment: Dear Reviewer U4NU, We would like to express our gratitude for your valuable comments and feedback. As the author-reviewer discussion period is coming to close on Aug 13th, we would greatly appreciate it if you could provide more feedback and reevaluate our work based on our responses and updated results. Thank you very much for your attention to this matter. Best regards, Authors	Summary: This paper proposes a type of diffusion model that captures the evolution of geometric states. The model is characterized by a diffusion SDE that couples the initial state with the target state, in the middle of which trajectory guidance is enabled when such data present. The framework is designed to yield equivariant density similar to other geometric diffusion models. Experiments on equilibrium state prediction with or without trajectory data have been performed to verify the applicability of the proposed approach. Strengths: 1. The distinction between existing works has been elaborated in Table 1, which is clear. 2. The method is designed with an option to leverage additional trajectory data, which is quite interesting. Weaknesses: 1. The experimental setup and comparison with baselines on equilibrium state prediction is a bit troublesome which requires more clarification or additional comparisons. Please refer to Q1. 2. The presentation is a bit unclear. Please refer to Q2. 3. Additional baselines may be considered. The baselines selected in the paper are not closely connected to the proposed approach. See Q3. 4. Missing ablation studies. In the current shape it is unclear where the performance gain comes from. See Q4. Technical Quality: 2 Clarity: 2 Questions for Authors: Q1. The evaluation protocol on QM9 and Molecule3D, especially to compare with direct prediction approaches, is not a common practice. A more convincing benchmark protocol would be to compare with methods such as GeoDiff [1] on molecule generation tasks since they are also generative models. Since the paper is positioned to tackle generative modeling, the experiments should also be designed to align with the goal. Q2. Could the authors provided detailed sampling algorithm this approach adopts? If the model uses sampling approach similar to other diffusion models, there should be related discussions on sampling steps/sampling time the method consumes. Q3. A more reasonable baseline would be to directly apply existing bridge models (e.g., [2]) to the current task by switching the backbone to the one this paper adopts. This would help the audience understand the unique contribution of this work since both bridge models and equivariant (geometric) diffusion models have been proposed in literature. Q4. Ablation studies such as investigating the importance of preserving equivariance of the density modeled should be included. This would help justify the necessity of the proposed components. [1] Xu et al. GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation. In ICLR'22. [2] Zhou et al. Denoising Diffusion Bridge Models. In ICLR'24. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: There seem to be no discussions on limitations in the paper. It would be better to discuss potential limitations from perspectives such as scalability and sampling time. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for spending time reviewing our paper. We would like to first address your misunderstanding by clarifying the task of our interest: to capture/predict the evolution of geometric states, i.e., predicting future states from initial states. This goal has been carefully stated at the very beginning of our paper (please refer to (1) line 1 in Abstract; (2) line 22 in Sec 1 Introduction; (3) lines 83-84 in Sec 2.1 Problem Definition). We also provide its formal definition. See lines 84-91. Based on these statements, we would like to argue that our scope is different from conventional molecular generation problems, such as the one introduced in GeoDiff mentioned in the review comments. In Sec 3.1 of GeoDiff, the definition of molecule generation task is "Given multiple graphs $G$, and for each $G$ given a set of conformations $C$, ..., the goal is learning a generative model $p_\theta(C\|G)$". From the definition, we can see: - The input in GeoDiff is the molecular graph, while in ours, the input is the geometric state of a system at a given time $t_0$. - The output of the molecule generation task is a set of sampled conformations $C$ from Boltzmann distribution, while the output of our task is the geometric state at a given time $t_1$ evolved from time $t_0$. - The goal of GeoDiff is to learn a generative model for the underlying Boltzmann distribution of molecular conformations, while our task is to capture the evolution of geometric states over time. Due to these differences, our task has several unique difficulties, including how to leverage trajectory data and preserve coupling between states over time while satisfying SE(3) at the same time. Our work tackled these challenges and provided promising results. We guess that the misunderstanding may partly be due to the lack of detailed task descriptions for the related works. We hope the above clarifications can help the reviewer better understand our task definition and our goal. We will update these discussions in the next version of our paper to avoid ambiguity and misunderstanding. >Regarding the experimental setup and evaluation protocol (Weakness 1 & Q1) First, we believe our evaluation protocol aligns well with our task. To measure how accurate a model is in predicting a system's future geometric state from its initial geometric state, we follow existing works in this direction [1, 2, 3] to use C-RMSD, D-MAE, and D-RMSE on QM9/Molecule3D and ADwT on OC22 which comprehensively measure the position/distance error between predicted geometric state and ground-truth geometric state. The reviewer suggested that we should follow GeoDiff and use their proposed metric. We are afraid that it is not a proper choice. GeoDiff (In Sec.5.2) calculates metrics between two sets of conformations, which aims at measuring the difference between the learned distribution and ground truth distribution rather than how accurate a target geometric state is predicted when specifying the initial geometric state. Second, we believe strong and latest baselines in this direction have been included. We followed the most recent works (e.g., GTMGC in ICLR24) and benchmarks (e.g., OC22) to compare our approach. The reviewer suggested that we should use GeoDiff as the baseline. But again, it is not a proper choice as GeoDiff cannot leverage trajectory guidance if provided and cannot take the initial geometric state as input to predict the target geometric state, which has been clearly discussed at the beginning of our response. >Regarding the details of our sampling algorithm (Weakness 2 & Q2) We have already discussed the sampling/inference method (ODE solver) of our GDB framework in lines 261-265 and presented details of the ODE solver (here we use the basic Euler solver) and sampling steps (10 steps in total) in Appendix C. Following your suggestion, we will provide an additional pseudo-code of our inference process in the next version of our paper. >Regarding additional baseline comparisons and the unique contribution of our work (Weakness 3 & Q3) First, we would like to clarify the reviewer's concern: "The baselines selected in the paper are not closely connected to the proposed approach." As stated in our responses to Weakness 1 & Q1, our task of interest is to predict the future geometric state of a system from its initial geometric state. In both Sec 1 Introduction and Sec 5 Related Works, we carefully review existing approaches for this task, including experimental approaches, traditional computational methods, direct prediction, and machine learning force fields. Although "both bridge models and equivariant (geometric) diffusion models have been proposed in literature" as the reviewer mentioned, there exist few works using generative modeling techniques to comprehensively investigate our task of interest due to the difficulty of unified satisfying the three desiderata (Coupling Preservation, Symmetry Constraints and Trajectory Guidance, lines 132-145). Following existing works, we use strong baselines of this task for comparisons, which is common practice. Following your suggestion, we compare our approach with Denoising Diffusion Bridge Models by switching the backbones on OC22. Our GDB framework significantly outperforms this new baseline by 6.37% on average ADwT. Moreover, our approach can further leverage trajectory guidance to achieve better performance, i.e., 7.93% improvement compared to this baseline. We will update these results in the next version of our paper, which we believe can help the audience better capture our GDB framework's effectiveness. --- Rebuttal 2: Title: Rebuttal by Authors (Part 2) Comment: >Regarding ablation studies (Weakness 4 & Q4) We have already provided ablation studies. Please refer to Sec 4.2 (lines 335-342 and Table 5) for detailed descriptions and results. Our ablation studies help us better understand the importance of key designs in our GDB framework, including trajectory guidance and coupling preservation, without which we observe significant performance drops. Following your suggestion, we further conduct an ablation study by investigating the importance of satisfying equivariant constraints, which shows that the performance drop is 14.45% if we remove the equivariant constraints. These ablation studies serve as supporting evidence for the necessity of proposed components in our GDB framework, and we will update these results in the next version of our paper. >Regarding the discussion of limitations Thank you for the suggestion. For the sake of generality, we do not experiment with advanced implementation strategies of training objectives and sampling algorithms, which leave room for further improvement. Besides, the employment of Transformer-based architectures may also limit the efficiency of our framework. This has also become a common issue in transformer-based diffusion models. We will organize these discussions into a single Limitation section and update it in the next version of our paper. We thank you again for your efforts in reviewing our paper, and we have replied to each of your concerns. We sincerely look forward to your re-evaluation of our submission based on our responses and updated results. --- Rebuttal 3: Title: Looking forward to your re-evaluation Comment: Dear Reviewer znC7, Thank you for your time and efforts in reviewing our paper. We have carefully responded to each of your questions. Given that the author-reviewer discussion deadline is approaching, we would greatly appreciate it if you could kindly take a look at our responses and provide your valuable feedback. We are more than happy to discuss more if you still have any concerns. Thank you once again and we are eagerly looking forward to your re-evaluation of our work. Paper 15795 Authors --- Rebuttal 4: Comment: I thank the authors for the response. However, some concerns are still not addressed. While the original setting in GeoDiff is molecule generation, from my perspective, it could be easily extended to your setting with the core difference lying in tackling how to condition on an input structure, which can be tackled in a fairly easy way. A related approach that considers such extension is DiffMD [1]. A strong reason why these approaches are compelling is that your approach is claimed to be generative, in which case some generative baselines (or even benchmarks) should be included. [1] Wu et al. DiffMD: A Geometric Diffusion Model for Molecular Dynamics Simulations. In AAAI 2023. Regarding Q2, to me it is indeed quite surprising that only 10 steps are needed to obtain high quality samples, as opposed to 100-1000 steps adopted in previous works. Thus it would be interesting to see how the performance changes with the number of sampling steps. --- Rebuttal Comment 4.1: Comment: Thanks for your quick reply and constructive suggestions. We agree with your point that including more generative baselines for comparisons could improve our work. In previous responses, we followed your advice to include Denoising Diffusion Bridge Models with our equivariant backbone, which combines advanced bridge models and equivariant models as a strong generative baseline. We will also follow your suggestion to include DiffMD in our work. However, we have tried our best but found that DiffMD is not open-sourced. We have to reimplement the method and conduct the experiment, which cannot be finished before the deadline. We will definitely update the results in the next version of our paper. We will also carefully cite these generative baselines and provide thorough discussions in our Related Work section. We use 10-step sampling for the sake of simplicity. In our preliminary experiments, we found that using ten steps already yields good performance. Increasing inference steps still improves model performance slightly but with more computational cost. We believe investigating advanced sampling strategies for our method is important, which we leave as future work. Through intensive and meaningful discussions with you, we have realized that some terminology could be more precise. For instance, the title '...via generative modeling' is too broad and might give the impression that we are applying conventional generative modeling methods on a standard generative task. As discussed in the thread, this is not our intention. We will clarify these parts to reflect our actual focus better. Thank you again for your valuable suggestions. We sincerely hope the reviewer re-evaluates our work based on these responses and updated results for the next version of our paper. --- Reply to Comment 4.1.1: Title: Kindly request for feedback and reevaluation Comment: Dear Reviewer znC7, Thank you once again for taking the time to review our paper! As the deadline for Reviewer-Author discussions is fast approaching, we would greatly appreciate it if you could kindly reevaluate our work with updated scores. Overall, we have carefully put every effort to address the reviewer's mentioned concerns. Besides, we will definitely follow your suggestion to include updated baseline results and thorough discussions of your mentioned related works in our next version of the paper, which we believe can improve the quality of our submission and address your remaining concerns. We always believe that the feedback between reviewers and authors would indeed improve the paper's quality, and we will definitely include the updated results and conceptual discussions from the reviewer's suggestions in the revised paper. It would be really nice to see both of us reach a consensus. We sincerely look forward to your reevaluation and feedback! Best Regards, Paper 15795 Authors --- Rebuttal 5: Title: Kindly Reminder of the Close of the Discussion Period Comment: Dear Reviewer znC7, We would like to express our gratitude for your valuable comments and feedback. As the author-reviewer discussion period is coming to close on Aug 13th, we would greatly appreciate it if you could provide more feedback and reevaluate our work based on our responses and updated results. Thank you very much for your attention to this matter. Best regards, Authors	Summary: In this paper, the authors introduce a Geometric Diffusion Bridge (GDB) framework, which aims to predict the evolution of geometric states in complex systems accurately, crucial for fields such as quantum chemistry and material modeling. Traditional methods face computational challenges, while deep learning approaches lack precision and generality. The authors use Doob’s h-transform to construct an equivariant diffusion bridge. By applying Doob’s h-transform, the authors adjust the SDE to ensure that the process starts from an initial geometric state and is conditioned to reach a target geometric state. This ensures that the transformed process respects the symmetry constraints of the geometric states, leading to more accurate and physically meaningful predictions. Strengths: + The framework utilizes an equivariant diffusion bridge derived from a modified Doob’s h-transform. This ensures that the diffusion process respects symmetry constraints, making the predictions more robust and reliable. + The paper provides a theoretical framework analysis about preserving symmetries and accurately modeling evolution dynamics. + Experimental evaluations show that GDB is better than state-of-the-art approaches in various real-world scenarios, including equilibrium state prediction and structure relaxation tasks. + The framework achieves significant error reduction compared to strong baseline models, particularly in challenging tasks such as structure relaxation in the Open Catalyst 2022 dataset Weaknesses: - The framework, especially when leveraging trajectory data, might introduce significant computational overhead. The simulation-free matching objective is designed to be efficient, but the overall framework’s computational demands might still be high - Some mathematical notations and definitions in the paper could be made clearer. For instance, explicitly defining all variables and functions used in the modified Doob’s h-transform and constructing equivariant diffusion bridges would improve readability and understanding. Technical Quality: 3 Clarity: 3 Questions for Authors: See above Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No limitations are addressed in the paper by the authors Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing both the theoretical analysis and practical effectiveness of our GDB framework. We also appreciate your suggestions which can improve our work further. Here are our responses to your questions. >Regarding the computational cost of our GDB framework Thanks for the question. Indeed, our GDB framework is efficient. Here, we discuss the reasons in the following two settings: - Setting 1: Bridging geometric states without trajectory guidance. In this setting, our training objective (Equ.(5) and Algorithm 1) and inference process are similar to typical diffusion models, which do not bring additional computational overhead. - Setting 2: Bridging geometric states with trajectory guidance. In this setting, trajectory data is used during training. From Equ.(6) and Algorithm 2 in our paper, it can be seen that the loss calculation at each iteration is still efficient (compared to Algorithm 1) with no computational overhead. For inference, we simulate the ODE in time interval $[0, NT]$. In practice, we set $T=1/N$ so the time interval is still $[0,1]$, which is kept the same as setting 1. The intuition is that we split the total time interval $[0,1]$ into $N$ parts for our Chain of Equivariant Diffusion Bridges with $N$ chains. Since the total length of the time interval is the same, predicting the future geometric state from the initial geometric state by using our model trained with trajectory guidance does not bring significant computational overhead. We find your question highlights the need for a detailed table to introduce the inference algorithm. We will incorporate this in the next version. We hope this addition clarifies the efficiency of our framework and highlights the improvement in model performance. >Regarding the mathematical notations and definitions Thanks for your suggestion of explicitly defining all variables and functions. We will revise our paper and add a notation paragraph to elaborate on all mentioned notations and concepts in Sections 3.1 & 3.2 for improved clarity and readability. We thank you again for your efforts in reviewing our paper, and we have replied to each of your concerns. We sincerely look forward to your re-evaluation of our submission based on our responses. --- Rebuttal 2: Title: Looking forward to your re-evaluation Comment: Dear Reviewer Sjdg, Thank you for your time and efforts in reviewing our paper. We have carefully responded to each of your questions. Given that the author-reviewer discussion deadline is approaching, we would greatly appreciate it if you could kindly take a look at our responses and provide your valuable feedback. We are more than happy to discuss more if you still have any concerns. Thank you once again and we are eagerly looking forward to your re-evaluation of our work. Paper 15795 Authors --- Rebuttal Comment 2.1: Comment: I have read the comments and am satisfied with the responses. However, I would still stick to my original score. --- Reply to Comment 2.1.1: Comment: Thanks for your reply. We are happy to hear that all your concerns have been properly addressed. We thank you again for recognizing our contributions and staying positive about our work.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Inference of Neural Dynamics Using Switching Recurrent Neural Networks	Accept (poster)	Summary: This paper develops a switching RNN (SRNN) framework to model neural activity. It builds up on switching linear dynamical system models that are used in neuroscience to segment and extract underlying dynamics of observed neural activity. The different segments corresponding to unique dynamics often reflect distinct behavioral states. The crucial novelty of this work is that they allow the dynamics to be non-linear, unlike SLDS and rSLDS, making the model more expressive. They fit these models using VI using an inference network. Finally, they apply SRNN to synthetic data, as well as 3 distinct neural datasets and show that it outperforms SLDS and rSLDS on segmenting activity into behavioral modules where each module corresponds to distinct dynamics. They visualize these underlying dynamics, and also evaluate their fitted model on predicting future neural activity. Strengths: 1. As we move towards large-scale neural datasets, it is crucial to scale model complexity in order to fully harness these datasets. This paper makes a step in that direction by allowing for non-linear dynamics, while also providing an appropriate fitting approach. 2. The experiment section is extensive, and I appreciate the application to multiple neural datasets. I particularly found the results on the decision-making dataset to be most impressive. 3. The literature review is thorough, and the authors do a good job of situating their work in the context of other related studies. Weaknesses: 1. The authors mention switching nonlinear dynamical systems (Dong et al. 2020), and discuss how their work differs from Dong et al. I think it is important to either provide an experimental comparison to SNLDS or a justification for why these existing models are insufficient to explain neural datasets, as the main novelty/motivation for SRNN and SNLDS is very much related (also noted by the authors in the paper). More on this in the question section. 2. Behavioral segmentations are somewhat subjective in nature, and while I can see that in the experiments shown here they make sense, in a real world setup we may want to infer the number of such segmentations from the data. Here the authors set the number of discrete states to the # of true behavioral states, however this might not be known in practice. Furthermore, there might be distinct sets of dynamics within one behavioral state due to other reasons not totally explicit from behavior. From the current set of results, it is not clear if SRNN is capable of inferring the # of underlying states. I will elaborate more in the questions section on this as well. 3. I also think the paper will benefit from some editing by the authors. The references are not formatted properly, commas are missing. The referencing to supplementary figures doesn't seem to be working, it links back to figures in the main text. I also think the authors can trim some of the background, such as the section on VI, in favor of explaining some of the experiments such as the Lorenz attractor setup in more detail. 4. While I appreciate the extensive experiments, I find it hard to reconcile some of the results. It seems like in some of the plots (Fig 3C/D, Fig 5C/D) prediction + reconstruction performance across all models is similar. However, the discrete states being inferred look hugely inaccurate for SLDS and rSLDS. I wonder if the authors have thoughts on why this happens. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. I feel it is crucial to understand the pros and cons of SNLDS vs SRNN, and understand if the differences between the two models are consequential in modeling neural data. For example, I understand some of the modeling differences such as the dependence of discrete state transitions on the data in SNLDS vs the previous continuous state in this paper, but I am not sure if one of the two is a better assumption. Additionally, I am also struggling to understand why can SNLDS not be used for prediction? 2. In the real-world experiments, were the number of discrete states across all models being compared set to the true number of behavioral states? I would be curious to see how results vary across different # of discrete states across these models, perhaps MSE on prediction or ELBO vs # of discrete states is a possible way to show this. This is for two mains reasons: i. In a new dataset we might not know ground-truth behavioral segmentations, and will want to be able to infer the number of such segmentations from the data. Hence, it would be interesting to see if SRNN can be used to do so. ii. The fact that rSLDS does fine it predicting data but infers discrete states inaccurately makes me wonder if it is clustering data differently, perhaps collapsing 2 states into one or further segmenting one behavioral state into multiple slightly different states. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have addressed limitations in the last section of the paper, and I do not envision any societal impact of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you deeply for your time and attention in reading our paper, and for your valuable comments. Below is our response to specific weaknesses and questions. Summary of Weakness 1: comparison to SNLDS. We thank you for pointing out this weakness. In addition to the existing SLDS, rSLDS, and LFADS, we implemented two additional baseline models that reviewers mentioned, the SNLDS [Dong et al., 2020] and mrSDS [Karniol-Tambour et al., ICLR 2024]. We show their performance on the Reaching Dataset in the attached PDF. To summarize, while recovering behaviorally-relevant states, our model SRNN outperforms SNLDS and has comparable performance with mrSDS, whereas in reconstructing neural activity, SRNN outperforms mrSDS and has comparable performance with SNLDS . Therefore, SRNNs have overall better performance on recovering behaviorally-relevant states and reconstructing neural activity. Q1): It is crucial to understand the pros and cons of SNLDS vs SRNN, and understand if the differences between the two models are consequential in modeling neural data. For example, I understand some of the modeling differences such as the dependence of discrete state transitions on the data in SNLDS vs the previous continuous state in this paper, but I am not sure if one of the two is a better assumption. Additionally, I am also struggling to understand why can SNLDS not be used for prediction? We thank you for allowing us to elaborate on the differences between the two models. (1) Indeed, the transitions in our model are inspired by rSLDS, i.e., in SRNNs, the transition depends on continuous states $h_t$. However, we observed no significant performance difference between transitions based on the observations $y_t$ as compared to $h_t$, as $y_t$ is a linear transformation of $h_t$. However, $h$ typically has a lower dimensionality, requiring fewer parameters to model the transition, e.g., in the context of reaching, the dimensions of the continuous latent variables, $h$, P=16, vs. the dimensions of neural activity, $y$, R=180. (2) SNLDS is able to be tweaked in order to make predictions by using a causal inference network, which is the approach utilized in our prediction work. The original SNLDS paper used non-causal inference, which involves inferring states from the entire sequence. Consequently, the original SNLDS method is unsuitable for predicting neural activity, as it requires information from future neural activity or even the entire neural activity sequence for state inference, but again, this is able to be tweaked for the purpose of prediction. (3) Lastly, SNLDS and SRNNs have different initialization processes, which we suspect is contributing to SRNNs’ better performance in modeling neural data (see response to Weakness 1). While SNLDS uses regularization in the objective function to ensure uniform state utilization, SRNNs use HMM as an initialization, again inspired by previous models that have been successfully applied towards modeling neural data. Summary of Weakness 2: It is not clear if SRNN is capable of inferring the # of underlying states. Summary of Question 2: how results vary across different # of discrete states across all models. This is for two main reasons: i. In a new dataset we might not know ground-truth behavioral segmentations. ii. The fact that rSLDS does fine in predicting data but infers discrete states inaccurately. We thank you for your questions. We now provide a comprehensive set of metrics to determine this important hyperparameter in the global rebuttal as well as in the official comment below. To summarize, we use appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance and co-smoothing method, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. Moreover, we also show reconstruction performance across all models with different K in the attached PDF (Figure 2). Lastly, we agree that rSLDS does fine in reconstruction of data but infers discrete states inaccurately. Please see below (response to Weakness 4) for a discussion on this topic. Summary of Weakness 3: The paper will benefit from some editing by the authors. We thank you for pointing out this weakness and we apologize for the inconvenience in locating the figures when you review our paper. We have fixed the reference formatting and the figure references. Additionally, we have added some simulation and experimental details in the main text, and removed some commonly-used VI details to the Appendix. Summary of Weakness 4: Prediction + reconstruction performance across all models is similar. However, the discrete states being inferred look hugely inaccurate for SLDS and rSLDS. We thank you for pointing this out; this has puzzled us as well. We believe that accurate reconstruction of data does not necessitate accurate inference of discrete hidden states or dynamics, since the models are trained purely for maximizing the ELBO, i.e., here, reconstruction accuracy. However, accurate prediction of neural activity many time steps ahead requires simultaneously (a) accurate reconstruction, (b) accurate inference of discrete states, and (c) accurate inference of dynamics. Here, we see concrete differences in the prediction capability of SRNNs vs. SLDS / rSLDS (e.g., see Figure 4D of the original submission). We believe that prediction of future activity provides a better test of a dynamical model, and here, SRNNs have overall better performance in prediction of neural activity many time steps ahead. --- Rebuttal 2: Comment: Details of Determining relevant hyperparameters (Weakness 2 and Question 2) We now add a discussion for determining relevant hyperparameters, such as the number of discrete and continuous latent states. Importantly, we include a comprehensive hyperparameter sweep for the number of discrete states for both the reaching and decision-making datasets. We detail out the salient results for the reaching dataset, where there are 5 behavioral states in the task (decision-making results below): (1) Convergence to lower number of discrete states: We tested our model by increasing the number of hidden states K while keeping the number of continuous latent states P constant. We found that 61% of SRNNs with a higher number of discrete hidden states (e.g., K=6) finally converge to the optimal number of discrete hidden states, i.e., K=5. (2) Reuse of discrete states: We also test our model by decreasing the number of hidden states. We found that 94% of SRNNs with a lower number of discrete hidden states (e.g., K=4) had at least one hidden state reused after other states: in other words, SRNNs are not able to perform well with 4 unique discrete hidden states without reusing one of them. (3) Reconstruction performance plateau: While keeping other hyperparameters constant, the reconstruction accuracy plateaus at the same number of discrete states as in the behavior, thus we can use the minimum number of discrete states as it takes for the model to perform well. We have included a figure detailing this in the attached PDF (Figure 2). Moreover, we also implemented a ‘co-smoothing’ method as suggested by Reviewer hFy4 [Yu et al., 2009 and Karniol-Tambour et al., ICLR 2024], we show the results in the attached PDF (Figure 3), where we found that K=5 also does well in reconstructing the data with a ‘co-smoothing’ neuron drop-out analysis. (4) Variability across conditions: In stereotyped tasks or experiments, such as reaching, there may not be a significant amount of variability in the timing of behavior across conditions, and this variability can thus be used as a metric for determining the number of discrete states. Here, we found that SRNNs with K=5 have much lower variability on recovered behaviorally-relevant states than K=4 and K=6 (i.e., 0.098 for K=5, 0.384 for K=4, and 0.282 for K=6). Furthermore, we did the same analysis for the decision-making data (Figure 4 in the attached PDF) and we found the same situation as in the reaching dataset. Higher K also converges to a smaller number of discrete states, and K=5 has the smallest variability across different pseudo-sessions (i.e., 0.106 for K=5, 0.287 for K=4, and 0.290 for K=6). These results demonstrate appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. We have now provided this in the paper, while adding a discussion for the general case. Additionally, we show a comparison between values for another important hyperparameter P in the original submission. --- Rebuttal 3: Title: Response to rebuttal Comment: Thank you for your detailed response. I appreciate the new results: comparisons with SNLDS and mrSDS, as well as the detailed experiments on discrete state switching. I still find myself a bit perplexed by the discrepancy in behavior and neural data results and I wish I understood the advantages / flaws of each of these models better. Overall, since the authors have addressed some of my concerns, I will raise my score to beyond the accept threshold. --- Rebuttal Comment 3.1: Comment: We thank you for your response and for your consideration in raising the score. We didn’t find this discrepancy in recovering behavioral vs. reconstructing neural data in our model. We definitely agree that this discrepancy in existing models is interesting to explore.	Summary: The authors develop a new class of probabilistic nonlinear state space models called switching RNNs. In essence, this extends the well-known switching linear dynamical system (SLDS) model to switch between nonlinear dynamics governed by a stochastic RNN. Strengths: * The results shown in panels A of Figs 3, 4, and 5 are nice and convincing. Weaknesses: * Like many other deep learning based approaches, the model is not particularly interpretable. For example, panel F in Figs 3, 4, and 5 shows 2D flow fields for the different hidden states, but the RNN hidden state is 16-dimensional. Here the authors have used PCA to attempt to find a reasonable 2D flow field, but I know from experience that this has the potential to very poorly capture the true dynamics of the system. Intuitively, even small variance dimensions can matter a lot if the flow field changes rapidly along that dimension. * There are many tunable parameters in this model (e.g. number of continuous and number of discrete states). It is unclear how to choose these on datasets without ground truth, or at least good educated guesses. * Related to above, I worry a lot about the identifiability of this model. A nonlinear RNN without discrete switching can already model any flow field if given enough units. Thus a model with many continuous states (e.g. $P=128$) but zero discrete states may perform equally well to a model with few continuous states (e.g. $P=16$ or $P=8$) but a handful of discrete states. How would one then go about choosing between these models? Adding discussion or ideally some sort of mathematical analysis regarding the statistical identifiability of the model would be very helpful. Technical Quality: 3 Clarity: 3 Questions for Authors: * Equation (2) seems wrong to me. The nonlinearity $f(\cdot) = \tanh(\cdot)$ doesn't make sense here since $p(z_t \mid z_{t-1}, h_{t-1})$ should be a positive number. Perhaps you meant to use a softmax nonlinearity here? * Equation (2) also seems to suggest that the transition probability only depends on $h_{t-1}$ and not $z_{t-1}$. Is this correct? * Related to the above, it wasn't obvious to me whether you are allowing the continuous hidden state to impact the transition for the discrete state. Essentially I am wondering if your model is analogous to the switching LDS (where the continuous hidden state doesn't impact the transition statistics) or instead the recurrent switching LDS (where the continuous state does impact the discrete transition probabilities). In Figure 1B, what is the meaning of the red dashed arrow? Does that carry any difference to the black arrows? * Are the good results shown in panel A of Figs 3, 4, and 5 due solely to differences in the initialization procedure across models? (see top of page 5) * Regarding identifiability, what happens if you run the model multiple times from different random seeds? Do you recover the same flow fields and fixed point structure? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The discussion adequately acknowledges limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you very much for giving us a positive rating, and for your extremely relevant comments. Below we respond to specific weaknesses pointed out. Summary of Weakness 1: Like many other deep learning based approaches, the model is not particularly interpretable. 2D flow fields poorly capture the true dynamics of the system. Yes, you are absolutely correct, a 2D flow field is not able to accurately capture the true dynamics when P>2, and simply provides a visualization. We are able to also calculate slow points and other features of the dynamics in the full dimensional space, provided here as an example in Figure 8 of the attached PDF. When comparing the dynamics in different discrete states of our model, we use the same principal components to project the higher-dimensional dynamics to 2D. Therefore, the dynamics can technically be compared with each other. Additionally, where possible (e.g., in Figure 5F of the original submission), we plot the flow fields for SRNNs with P=2; here, the visualization and structure found is of the true dynamics. We have also run this model with a different random seed and recovered similar features of the dynamics in Figure 8 of the attached PDF, showing identifiability of the dynamics. Summary of Weakness 2: It is unclear how to choose tunable parameters on datasets without ground truth, or at least good educated guesses. We thank you for your comments, and comprehensively discuss this in the global rebuttal as well as in the official comment below. To summarize, we use appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance and co-smoothing method, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. Additionally, for selecting the number of continuous hidden states P, we included comparisons in the paper, we think an appropriate P should perform well on both recovering states constantly and reconstructing behavior, meanwhile, the P should be as small as possible. Summary of Weakness 3: How to choose between a SRNN and a model with higher P. Yes, you are definitely right. According to the theory of universal function approximation, a model with higher dimensions can do the same work as multiple models with low dimensions. To empirically test this, we trained a standard RNN with P=32, which has a similar number of parameters on recurrent weights but more on emission weights, and found that standard RNNs are also able to reconstruct neural activity well (see attached PDF Figure 7). However, neural dynamics are broadly thought to be low-dimensional, and we would like this dimensionality to be as low as possible to achieve interpretability in the dynamics as well as the discrete states (also discussed above). Indeed, we also found that the switches in the neural dynamics are relevant to different behavioral and stimulus states; therefore, if we use a higher dimensionality but a non-switching model, the model may not capture this interpretability. Q1): Typo in Equation (2) We thank you for pointing this out, it is a typo, and we used softmax in implementation. We will revise the equation. Q2 and Q3): It wasn't obvious to me whether you are allowing the continuous hidden state to impact the transition for the discrete state. In Figure 1B, what is the meaning of the red dashed arrow? Does that carry any difference to the black arrows? We thank you for your question. The value of the transition probability depends on $h_t$. Our model is similar to rSLDS, where the transition probabilities depend on continuous states. We plot the red dashed arrow to show that the transition probabilities depend on previous continuous states (they are red since the states z are not directly computed using $h_t$). The black arrows between z represent the transitions between z. Q4): Are the good results shown in panel A of Figs 3, 4, and 5 due solely to differences in the initialization procedure across models? We thank you for your question. Initialization might be the most common and powerful method to overcome the difficulty of training in this area. We used the SSM package to implement SLDS and rSLDS where HMMs are also used as an initialization. Q5): Regarding identifiability, what happens if you run the model multiple times from different random seeds? Do you recover the same flow fields and fixed point structure? We thank you for your question. We agree that we should see the same structure in the dynamics across different random seeds. To show a simple example, we changed the random seed and trained the SRNN on lever pull data with P=2. We show a visualization in the attached PDF in Figure 8. We found similar flow fields with a comparable fixed point structure. Specifically, the new fixed points seem to be rotated from the old ones in the same direction. --- Rebuttal 2: Comment: Details of Determining relevant hyperparameters (Weakness 2) We now add a discussion for determining relevant hyperparameters, such as the number of discrete and continuous latent states. Importantly, we include a comprehensive hyperparameter sweep for the number of discrete states for both the reaching and decision-making datasets. We detail out the salient results for the reaching dataset, where there are 5 behavioral states in the task (decision-making results below): (1) Convergence to lower number of discrete states: We tested our model by increasing the number of hidden states K while keeping the number of continuous latent states P constant. We found that 61% of SRNNs with a higher number of discrete hidden states (e.g., K=6) finally converge to the optimal number of discrete hidden states, i.e., K=5. (2) Reuse of discrete states: We also test our model by decreasing the number of hidden states. We found that 94% of SRNNs with a lower number of discrete hidden states (e.g., K=4) had at least one hidden state reused after other states: in other words, SRNNs are not able to perform well with 4 unique discrete hidden states without reusing one of them. (3) Reconstruction performance plateau: While keeping other hyperparameters constant, the reconstruction accuracy plateaus at the same number of discrete states as in the behavior, thus we can use the minimum number of discrete states as it takes for the model to perform well. We have included a figure detailing this in the attached PDF (Figure 2). Moreover, we also implemented a ‘co-smoothing’ method as suggested by Reviewer hFy4 [Yu et al., 2009 and Karniol-Tambour et al., ICLR 2024], we show the results in the attached PDF (Figure 3), where we found that K=5 also does well in reconstructing the data with a ‘co-smoothing’ neuron drop-out analysis. (4) Variability across conditions: In stereotyped tasks or experiments, such as reaching, there may not be a significant amount of variability in the timing of behavior across conditions, and this variability can thus be used as a metric for determining the number of discrete states. Here, we found that SRNNs with K=5 have much lower variability on recovered behaviorally-relevant states than K=4 and K=6 (i.e., 0.098 for K=5, 0.384 for K=4, and 0.282 for K=6). Furthermore, we did the same analysis for the decision-making data (Figure 4 in the attached PDF) and we found the same situation as in the reaching dataset. Higher K also converges to a smaller number of discrete states, and K=5 has the smallest variability across different pseudo-sessions (i.e., 0.106 for K=5, 0.287 for K=4, and 0.290 for K=6). These results demonstrate appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. We have now provided this in the paper, while adding a discussion for the general case. Additionally, we show a comparison between values for another important hyperparameter P in the original submission. --- Rebuttal Comment 2.1: Title: Reviewer Response Comment: Thanks for the additional points and clarifications. I retain my score of a "weak accept" as I think the work is novel, technically correct, and could be of interest to the neural modeling community. I still have reservations about impact of this method on the broader neuroscience community, given challenges related to interpretability and somewhat weak reasons to prefer this method over a higher dimensional RNN. --- Rebuttal 3: Comment: >Thanks for the additional points and clarifications. I retain my score of a "weak accept" as I think the work is novel, technically correct, and could be of interest to the neural modeling community. I still have reservations about impact of this method on the broader neuroscience community, given challenges related to interpretability and somewhat weak reasons to prefer this method over a higher dimensional RNN. We thank you for your response and retaining a positive rating. We agree on the trade-off between switching low-dimensional RNNs and non-switching higher-dimensional RNNs. If accuracy is the only metric, we agree that high-dimensional standard RNNs are able to achieve this goal. However, for achieving interpretability in (a) identifying discrete behaviorally-relevant switches in dynamics, and (b) distinctive flow-field visualization, we demonstrate the utility of switching RNNs in the paper. As a concrete example comparing the two models, in order to achieve similar reconstruction accuracy as a P=2, K=3 SRNN (lever-pull data), we need a standard RNN with dimensionality P=5. This higher dimensionality in the standard RNN naturally comes at the cost of accuracy in the 2D flow-field representations, and thus may result in compromised interpretability of the dynamics. The trade-offs between low-dimensional switching RNNs and higher-dimensional standard RNNs are interesting, and we will add quantitative accuracy tradeoffs with varying values for P and K in the final version, if accepted.	Summary: The authors propose to model time series neural population activity using switching recurrent neural networks. The generative model includes discrete latent states Strengths: The proposed method does appear to outperform related switching linear dynamical systems approaches in certain contexts. Weaknesses: High-level: - The contribution beyond other switching nonlinear dynamical systems models is not clear. Such models include the cited Dong et al., 2020, as well as Karniol-Tambour et al., ICLR 2024. If there is a contribution beyond these works, the authors should compare against those existing related methods. - The authors do not demonstrate an ability to automatically determine the appropriate number of discrete states. One approach to this might be "co-smoothing" (see Yu et al., Gaussian Process Factor Analysis, 2009). Details: - The mathematical details and notation are often unclear. For example, equation 2 does not appear to be a valid probability distribution, given the description that f(.) = tanh(.). Shouldn't this instead be a categorical distribution or similar? Relatedly, f is also used in equation 8, but from the context it appears to denote something entirely different. - The authors should more clearly describe the cross-validation techniques for used for each dataset. The blanket statement in the intro to Section 4 ("On each dataset, we do N-fold cross-validation, where N equals to the number of conditions, sessions, or subjects in the dataset") obscures how cross-validation was actually applied in each instance. Technical Quality: 2 Clarity: 2 Questions for Authors: - Are the predictions in Figure 2 cross validated (eg., using the technique described in section 3.3? - In Fig 3, are the authors modeling single-trials or condition averages (ie PSTHs)? This should be addressed. It looks like they are predicting condition averages since the "true" neural activity in 3E takes on continuous values (rather than indicating spike times or binned spike counts). - Why does SRNN perform worse than SLDS and rSLDS in Figure 5CD? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors address several limitations, including their need to manually set the number of discrete states, their need for good parameter initializations, and the heavy computational requirements for fitting their models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you deeply for your time and attention in reading our paper, and for your valuable comments. Below is our response to specific weaknesses and questions. Weakness 1: Comparison between SNLDS and mrSDS Thank you for raising this weakness. As also mentioned in the global rebuttal, we have now implemented these two models, the SNLDS and mrSDS. We show their performance on the Reaching Dataset in the attached PDF. To summarize, while recovering behaviorally-relevant states, our model SRNN outperforms SNLDS and has comparable performance with mrSDS, whereas in reconstructing neural activity, SRNN outperforms mrSDS and has comparable performance with SNLDS. Therefore, SRNNs have overall better performance on recovering behaviorally-relevant states and reconstructing neural activity. We would like to point out that mrSDS was published recently by ICLR 2024 in May, two weeks before our original submission, and to date, we have not found any code released with the paper. However, we implemented the code based on our reading and understanding of the paper to the best extent possible. Weakness 2: Determining the appropriate number of discrete states. One approach to this might be co-smoothing We thank you for your comments and for the excellent suggestion of ‘co-smoothing’. We have now implemented co-smoothing and find that SRNNs are successful at latent state modeling using this metric, and that this method is helpful for determining the appropriate number of discrete states. In addition, we now provide a comprehensive set of metrics to determine this important hyperparameter, including co-smoothing in the global rebuttal as well as in the official comment below. To summarize, we use appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance and co-smoothing method, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. Detail Weakness 1: The mathematical details and notation are often unclear. For example, eq 2: Shouldn't this instead be a categorical distribution or similar? Relatedly, f is also used in eq 8 Yes, you are right. It should be a softmax activation function and categorical distribution, which is what we implemented in the original paper. We apologize for the typo and we have revised both equation 2 and equation 8 in the paper. Detail Weakness 2: The authors should more clearly describe the cross-validation techniques used for each dataset Thank you for this comment and we apologize for the confusion. We are training the model on N-1 conditions/sessions/subjects and testing it on the 1 held-out condition/session/subject. In this process, we train N different models, one for each one of conditions/sessions/subjects on hold, and show the results for each model as one dot in Figures 3-5. We had detailed for each experimental dataset whether we are considering conditions or sessions or subjects to be held-out data, but have now moved it to the main text in the interest of clarity. Specifically, we hold out data from one condition at a time in the reaching dataset, where there are 18 different curved reaching conditions. We hold out data from one session at a time in the decision-making dataset, where there are 8 total sessions, and we hold out data from one subject at a time in the lever pull dataset, where there are 6 total subjects. Q1): Are the predictions in Fig 2 cross validated? We thank you for your question and apologize for the lack of clarity. Yes, all models in this paper are cross-validated using held-out test sets. Specifically for Figure 2, we have one Lorenz attractor as the underlying dynamical system. We train SRNNs on 15 trials (representing different initial conditions), and test on 1 held-out trial. We have now clearly included this information in the paper. Q2): In Fig 3, are the authors modeling single-trials or condition averages (ie PSTHs)? We are modeling condition averages in the paper. In addition to the analyses in the paper, we have now also trained SRNNs on simulated single trial data to assess if our method works on this kind of data. Specifically, we implemented the thinning algorithm using the time-rescaling theorem to simulate single trial spiking data based on the condition averaged firing rate [Brown et al., 2002]. We simulated 200 single trials from 5 condition averages and computed the binned spike counts, which we used to train the SRNN models. We found that the SRNNs successfully represent this data with P=16 and K=5; we show the results of training on simulated single trials in the attached PDF (Figure 6). Q3): Why does SRNN perform worse than SLDS and rSLDS in Fig 5CD? In Figure 5C, the SRNNs indeed have worse reconstruction performance than SLDS and rSLDS. As we mentioned in our paper, ‘Indeed, all three methods have acceptable reconstruction.’: we didn’t find any significant difference by visualizing the reconstruction (Figure 5E of original paper). A potential reason causing the worse performance may be the small size of this dataset. As our response to Review 8VtM, a failure mode of SRNNs is when the model is trained in the low data regime, e.g., single conditions (see attached PDF). However, SRNNs have better recovery of behaviorally-relevant states. This may explain the finding that SRNNs have better prediction performance in 0.67, 1, and 1.33 seconds ahead in Figure 5D of the original paper; as the prediction window increases, the prediction performance may depend both on the recovered dynamics and the recovery of correct behavioral states. We believe that prediction of future activity provides a better test of a dynamical model, and here, SRNNs have overall better performance in prediction of neural activity. --- Rebuttal Comment 1.1: Comment: My primary concern remains only partially addressed: The contribution beyond other switching nonlinear dynamical systems models is not clear. I appreciate the attempts to implement and compare against SNLDS and mrSDS. And I appreciate that mrSDS was very recently accepted (though the paper has been on arxiv for many months before the ICLR acceptance) and that code may not have been publicly available. However, I expect a compelling description of the novelty of your approach relative to these existing approaches in the literature. What is the innovation beyond those approaches? And can you demonstrate that those specific innovations underlie the stated empirical performance benefits over SNLDS and mrSDS? --- Rebuttal 2: Comment: Details of Determining relevant hyperparameters (Weakness 2) We now add a discussion for determining relevant hyperparameters, such as the number of discrete and continuous latent states. Importantly, we include a comprehensive hyperparameter sweep for the number of discrete states for both the reaching and decision-making datasets. We detail out the salient results for the reaching dataset, where there are 5 behavioral states in the task (decision-making results below): (1) Convergence to lower number of discrete states: We tested our model by increasing the number of hidden states K while keeping the number of continuous latent states P constant. We found that 61% of SRNNs with a higher number of discrete hidden states (e.g., K=6) finally converge to the optimal number of discrete hidden states, i.e., K=5. (2) Reuse of discrete states: We also test our model by decreasing the number of hidden states. We found that 94% of SRNNs with a lower number of discrete hidden states (e.g., K=4) had at least one hidden state reused after other states: in other words, SRNNs are not able to perform well with 4 unique discrete hidden states without reusing one of them. (3) Reconstruction performance plateau: While keeping other hyperparameters constant, the reconstruction accuracy plateaus at the same number of discrete states as in the behavior, thus we can use the minimum number of discrete states as it takes for the model to perform well. We have included a figure detailing this in the attached PDF (Figure 2). Moreover, we also implemented a ‘co-smoothing’ method, we show the results in the attached PDF (Figure 3), where we found that K=5 also does well in reconstructing the data with a ‘co-smoothing’ neuron drop-out analysis. (4) Variability across conditions: In stereotyped tasks or experiments, such as reaching, there may not be a significant amount of variability in the timing of behavior across conditions, and this variability can thus be used as a metric for determining the number of discrete states. Here, we found that SRNNs with K=5 have much lower variability on recovered behaviorally-relevant states than K=4 and K=6 (i.e., 0.098 for K=5, 0.384 for K=4, and 0.282 for K=6). Furthermore, we did the same analysis for the decision-making data (Figure 4 in the attached PDF) and we found the same situation as in the reaching dataset. Higher K also converges to a smaller number of discrete states, and K=5 has the smallest variability across different pseudo-sessions (i.e., 0.106 for K=5, 0.287 for K=4, and 0.290 for K=6). These results demonstrate appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. We have now provided this in the paper, while adding a discussion for the general case. Additionally, we show a comparison between values for another important hyperparameter P in the original submission. --- Rebuttal 3: Comment: We thank you for your response. We include a detailed feature comparison between SNLDS, mrSDS, and SRNNs in the following table. Specifically, SRNNs differ fundamentally from SNLDS and mrSDS in their generative model, inference network, and initialization, as detailed below. Moreover, we examine whether these models, trained directly on neural data, can reconstruct _behaviorally-relevant discrete states_: a question that has not been investigated in either of the previous studies. We found that, without fail, SRNNs have better performance in this goal than competing methods, including SNLDS and mrSDS. (1) Generative Model: SRNNs utilize RNN as a generative model, while SNLDS and mrSDS employ a MLP with nonlinear activation functions. RNNs may offer greater interpretability and accuracy compared to feedforward models, and are very relevant to the neuroscience community for insights on computational mechanisms [1][2]. (2) Inference Network: mrSDS uses a non-causal transformer encoder as an inference network and SNLDS uses a non-causal combination of RNNs as an inference network. SRNNs are able to predict neural activity many time-steps ahead since we have implemented both causal and non-causal RNNs as possible inference networks. This important capability brings us one step closer to real-time and closed-loop neuroscience applications. (3) Initialization: SRNN has a different initialization method compared to SNLDS and mrSDS (unknown). In training SRNNs, we rely on HMMs, whereas SNLDS uses uniform entropy regularization. We would like to clarify that the initialization of mrSDS is unknown, therefore, we used both initialization approaches when we implemented mrSDS and report the better performance for mrSDS. In summary, our primary contributions are: (1) characterizing and interpreting low-dimensional switching nonlinear neural dynamics using SRNNs, (2) enabling causal prediction of neural activity, and (3) demonstrating that SRNNs outperform SNLDS and mrSDS in _recovering behaviorally-relevant states and reconstructing and predicting neural activity_, making them more reliable for interpreting behaviorally-relevant neural dynamics. We thank you again for your questions. We hope this comparison is compelling and we look forward to your response for further discussion if needed. \| Models (on Reaching Data) \| Generative Model \| Inference Network \|Initialization \| Behavioral States Recovery (Error) \|Reconstruction (MSE)\| Multi-regional setting \| Neural Prediction \| Code \| \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| \| SNLDS \| MLP \| NNs \|Entropy regularization\| 0.46 ($\pm$0.125) \| 0.00404 ($\pm$0.00170) \| No \|Not implemented (non-causal inference network)\|Tensorflow\| \| mrSDS \| MLP \| Transformer \|Unknown\| 0.32 ($\pm$0.139) \| 0.00806 ($\pm$0.00273) \| Yes \|Not implemented (non-causal inference network)\|Unknown\| \| SRNN (Ours) \| RNN \| NNs \|HMM\| 0.27 ($\pm$0.093) \| 0.00230 ($\pm$0.00041) \| Yes \| Yes (causal inference network)\|Pytorch\| [1] Durstewitz, Daniel, Georgia Koppe, and Max Ingo Thurm. "Reconstructing computational system dynamics from neural data with recurrent neural networks." Nature Reviews Neuroscience 24.11 (2023): 693-710. [2] Barak, Omri. "Recurrent neural networks as versatile tools of neuroscience research." Current opinion in neurobiology 46 (2017): 1-6. --- Rebuttal Comment 3.1: Comment: I have read the authors' responses as well as the comments from the other reviewers. I stand by my rating of 3 due to the limited originality and innovation of the contribution. I do not see the impact of this work meeting the high standard of NeurIPS. While the authors did describe the differences between their approach and previous related approaches, the novelty appears quite limited, and the authors have not demonstrated which of these differences (if any) lead to improved performance.	Summary: The paper proposes switching recurrent neural networks (SRNN), which allow the RNN weights to switch across time. The RNN weights switch based on a latent Markovian process of discrete states. The authors apply SRNN to a simulated dataset following the Lorenz attractor and three real-world neural recordings. Strengths: - Clarity: The authors clearly explain the problem, related work, and methodology with well-written equations and easy-to-understand figures. - Extensive use of datasets: The paper applies SRNN to numerous real-world neural datasets, illustrating the effectiveness of SRNN in accurately segmenting different datasets in an unsupervised fashion. Weaknesses: - Lack of comparison with other methods: The paper compares SRNN to (r)SLDS models. However, there exist many other models for unsupervised segmentation. For example, ARHMMs and their extensions are simple yet powerful and interpretable models for segmentation [1, 2]. The authors should cite and consider comparisons with multiple model classes. In addition, the paper notes in line 103 that SRNNs have the most comparable structure to SNLDS, but the authors do not make comparisons. The authors should also cite and compare with [3], which has switching nonlinear dynamics. [1] Wiltschko, A. B., Johnson, M. J., Iurilli, G., Peterson, R. E., Katon, J. M., Pashkovski, S. L., ... & Datta, S. R. (2015). Mapping sub-second structure in mouse behavior. Neuron, 88(6), 1121-1135. [2] Lee, H. D., Warrington, A., Glaser, J., & Linderman, S. (2023). Switching autoregressive low-rank tensor models. Advances in Neural Information Processing Systems, 36, 57976-58010. [3] Karniol-Tambour, O., Zoltowski, D. M., Diamanti, E. M., Pinto, L., Tank, D. W., Brody, C. D., & Pillow, J. W. (2022). Modeling communication and switching nonlinear dynamics in multi-region neural activity. bioRxiv, 2022-09. - Experiments: The simulated experiment with the Lorenz attractor shows that SRNN does well when it has access to noiseless observations with known state dimensions. In order to have a more convincing simulated experiment, the authors could consider the following. First project the Lorenz attractor to a higher dimensional space and add additive Gaussian noise. Then fit SRNN (and other compared models) to the dataset to see if it can recover the Lorenz attractor and true latent state dimension (using some metric on held-out data). Another simulated experiment could be done with a dataset that simulates the NASCAR track [1,2]. [1] Linderman, S. W., Miller, A. C., Adams, R. P., Blei, D. M., Paninski, L., & Johnson, M. J. (2016). Recurrent switching linear dynamical systems. arXiv preprint arXiv:1610.08466. [2] Lee, H. D., Warrington, A., Glaser, J., & Linderman, S. (2023). Switching autoregressive low-rank tensor models. Advances in Neural Information Processing Systems, 36, 57976-58010. Technical Quality: 3 Clarity: 3 Questions for Authors: - What are some failure modes of the model? Does extra flexibility mean that SRNNs need more data than simpler models such as ARHMMs or SLDSs? I'm curious how the SRNNs would do in a low-data regime (e.g., sample a small amount of dataset from an SLDS). - Have you tried fitting the model to datasets other than neural data? Based on how well SRNNs do in segmenting the neural datasets, I'm curious how SRNNs would do on other types of datasets, such as mouse behavioral dataset [1]. [1] Wiltschko, A. B., Johnson, M. J., Iurilli, G., Peterson, R. E., Katon, J. M., Pashkovski, S. L., ... & Datta, S. R. (2015). Mapping sub-second structure in mouse behavior. Neuron, 88(6), 1121-1135. - How are the hyperparameters selected? Based on how long it takes to fit each SRNN to the datasets, I wonder if it is feasible to sweep over the hyperparameter space. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: As the authors noted, some limitations of the model are that the model needs good initialization and that the model takes considerably more amount of time to train than simpler models such as SLDSs. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you very much for giving us a positive rating, and for you very helpful comments. Below we respond to specific weaknesses pointed out. Weakness 1: Lack of comparison with other methods, for example, ARHMMs and their extensions, as well as SNLDS and mrSDS We have now included the two suggested baseline models, mrSDS and SNLDS, in the PDF attached above. To summarize, as above, while recovering behaviorally-relevant states, our model SRNN outperforms SNLDS and has comparable performance with mrSDS, whereas in reconstructing neural activity, SRNN outperforms mrSDS and has comparable performance with SNLDS . Therefore, SRNNs have overall better performance on recovering behaviorally-relevant states and reconstructing neural activity. Additionally, we would like to point out that we used the SSM package to install SLDS and rSLDS, where all models are initialized with ARHMM, and thus expect that the log likelihoods are comparable or higher in these baseline models, already included in the initial submission. Weakness 2: Experiments: the simulated experiment with the Lorenz attractor shows that SRNN does well when it has access to noiseless observations with known state dimensions. In order to have a more convincing simulated experiment, the authors could consider the following. First project the Lorenz attractor to a higher dimensional space and add additive Gaussian noise. Then fit SRNN (and other compared models) to the dataset to see if it can recover the Lorenz attractor and true latent state dimension (using some metric on held-out data). Another simulated experiment could be done with a dataset that simulates the NASCAR track We thank the reviewer for suggesting the noisy high dimensional Lorenz Attractor and the common simulated data NASCAR. We now include results on testing SRNNs using the suggested setup for the noisy high dimensional Lorenz Attractor and NASCAR in the PDF attached above (Figure 9 and Figure 10). To simulate the Lorenz Attractor, we project the 3-dimensional Lorenz into an 8-dimensional space and add additive Gaussian noise. We found that SRNNs with P=3 can successfully recover the butterfly structure of a Lorenz Attractor using the noisy high dimensional data. Moreover, we found that our SRNNs can successfully recover the ground truth of the trajectories and states of the NASCAR dataset. Q1) What are some failure modes of the model? Does extra flexibility mean that SRNNs need more data than simpler models such as ARHMMs or SLDSs? I'm curious how the SRNNs would do in a low-data regime We thank you for your question. One failure mode may be when the training data is limited. To identify this, we trained our model on a single condition in the reaching dataset instead of 26 conditions; we found that the test reconstruction accuracy is lower than for SLDS and rSLDS (see attached PDF, Fig 5). Another failure mode may be the commonly-encountered problem of overfitting: we found that SRNNs may overfit to the training data if we increase the number of parameters (number of states K and number of latents P) to be large. Lastly, there may be a small problem of identifiability as pointed out by Reviewer Kpfw, i.e., if we increase the number of latents to be very large, the neural data can be successfully modeled by a single RNN, but of course this fails to recover interpretable switches in neural dynamics. We identify these tradeoffs and provide a discussion in the rebuttal for Reviewer Kpfw. Q2) Have you tried fitting the model to datasets other than neural data? Based on how well SRNNs do in segmenting the neural datasets, I'm curious how SRNNs would do on other types of datasets, such as mouse behavioral dataset We agree with the reviewer that behavioral data may provide the interpretability that we desire while modeling. We tested SRNNs on poses tracked using DeepLabCut, specifically the nose and paw positions, as recorded by a behavioral camera in the head-fixed decision-making task (see Fig. 4C in Rebuttal PDF). Interestingly, the states recovered were not as accurate as when using the neural data (see Fig. 4 of original submission). In this task, a visual stimulus was presented either to the left or the right, not visible in the camera. Specifically, the model trained on behavioral pose data failed to identify a separate state for ‘stimulus presentation’, presumably since the nose and paw location was not noticeably different here than during the delay period. This exercise points towards the possibility of augmenting behavioral observations with neural activity in future work, to better inform switching states. Q3) How are the hyperparameters selected? Based on how long it takes to fit each SRNN to the datasets, I wonder if it is feasible to sweep over the hyperparameter space One important parameter is the number of discrete states K, and we show a comparison between different numbers of hidden states in the PDF attached and include a thorough discussion in the global rebuttal. Another very important hyperparameter is the number of hidden states P of SRNN, we show a comparison between different numbers of hidden states in the original submission. We included the discussion about determining K and P in the global rebuttal as well as in the official comment below. For other hyperparameters, such as the number of hidden units in our inference network and learning rate, we used cross validation and kept the parameters with best performance, but we found that these other parameters did not affect the results very much. It is possible to sweep over the hyperparameter space, however, this may require training in parallel with large resources, for example, in the reaching dataset, training one model with P=16 and K=5 takes 9 hours 32 mins on a NVIDIA A100 GPU, and training models with (P=2, K=5), (P=4, K=5), and (P=8, K=5) takes 8 hours 01 min, 8 hours 30 mins, and 8 hours 54 mins respectively. --- Rebuttal 2: Comment: Details of Determining relevant hyperparameters (Q3) We now add a discussion for determining relevant hyperparameters, such as the number of discrete and continuous latent states. Importantly, we include a comprehensive hyperparameter sweep for the number of discrete states for both the reaching and decision-making datasets. We detail out the salient results for the reaching dataset, where there are 5 behavioral states in the task (decision-making results below): (1) Convergence to lower number of discrete states: We tested our model by increasing the number of hidden states K while keeping the number of continuous latent states P constant. We found that 61% of SRNNs with a higher number of discrete hidden states (e.g., K=6) finally converge to the optimal number of discrete hidden states, i.e., K=5. (2) Reuse of discrete states: We also test our model by decreasing the number of hidden states. We found that 94% of SRNNs with a lower number of discrete hidden states (e.g., K=4) had at least one hidden state reused after other states: in other words, SRNNs are not able to perform well with 4 unique discrete hidden states without reusing one of them. (3) Reconstruction performance plateau: While keeping other hyperparameters constant, the reconstruction accuracy plateaus at the same number of discrete states as in the behavior, thus we can use the minimum number of discrete states as it takes for the model to perform well. We have included a figure detailing this in the attached PDF (Figure 2). Moreover, we also implemented a ‘co-smoothing’ method as suggested by Reviewer hFy4 [Yu et al., 2009 and Karniol-Tambour et al., ICLR 2024], we show the results in the attached PDF (Figure 3), where we found that K=5 also does well in reconstructing the data with a ‘co-smoothing’ neuron drop-out analysis. (4) Variability across conditions: In stereotyped tasks or experiments, such as reaching, there may not be a significant amount of variability in the timing of behavior across conditions, and this variability can thus be used as a metric for determining the number of discrete states. Here, we found that SRNNs with K=5 have much lower variability on recovered behaviorally-relevant states than K=4 and K=6 (i.e., 0.098 for K=5, 0.384 for K=4, and 0.282 for K=6). Furthermore, we did the same analysis for the decision-making data (Figure 4 in the attached PDF) and we found the same situation as in the reaching dataset. Higher K also converges to a smaller number of discrete states, and K=5 has the smallest variability across different pseudo-sessions (i.e., 0.106 for K=5, 0.287 for K=4, and 0.290 for K=6). These results demonstrate appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. We have now provided this in the paper, while adding a discussion for the general case. Additionally, we show a comparison between values for another important hyperparameter P in the original submission. --- Rebuttal Comment 2.1: Comment: I would like to thank the authors for their response to my comments and clarifications. I would like to raise my score from 5 (Borderline accept) to 6 (Weak accept). However, I still believe that there needs to be comparisons with simpler models such as ARHMMs and their extensions [1, 2]. In particular, I do not agree with "...SLDS and rSLDS, where all models are initialized with ARHMM, and thus expect that the log-likelihoods are comparable or higher in these baseline models...". SLDS and rSLDS rely on approximate inference which may frequently lead to bad optima. In contrast, ARHMMs sometimes tend to perform better, thanks to their exact M-step update. [1] Wiltschko, A. B., Johnson, M. J., Iurilli, G., Peterson, R. E., Katon, J. M., Pashkovski, S. L., ... & Datta, S. R. (2015). Mapping sub-second structure in mouse behavior. Neuron, 88(6), 1121-1135. [2] Lee, H. D., Warrington, A., Glaser, J., & Linderman, S. (2023). Switching autoregressive low-rank tensor models. Advances in Neural Information Processing Systems, 36, 57976-58010. If accepted, I highly suggest including these additional comparisons with ARHMMs and their variants. --- Rebuttal 3: Comment: >I would like to thank the authors for their response to my comments and clarifications. I would like to raise my score from 5 (Borderline accept) to 6 (Weak accept). >However, I still believe that there needs to be comparisons with simpler models such as ARHMMs and their extensions [1, 2]. In particular, I do not agree with "...SLDS and rSLDS, where all models are initialized with ARHMM, and thus expect that the log-likelihoods are comparable or higher in these baseline models...". SLDS and rSLDS rely on approximate inference which may frequently lead to bad optima. In contrast, ARHMMs sometimes tend to perform better, thanks to their exact M-step update. >[1] Wiltschko, A. B., Johnson, M. J., Iurilli, G., Peterson, R. E., Katon, J. M., Pashkovski, S. L., ... & Datta, S. R. (2015). Mapping sub-second structure in mouse behavior. Neuron, 88(6), 1121-1135. >[2] Lee, H. D., Warrington, A., Glaser, J., & Linderman, S. (2023). Switching autoregressive low-rank tensor models. Advances in Neural Information Processing Systems, 36, 57976-58010. >If accepted, I highly suggest including these additional comparisons with ARHMMs and their variants. We thank you for your response and raising the score from 5 to 6. We also thank you for pointing out that “ARHMMs sometimes tend to perform better, thanks to their exact M-step update.” Yes, you are right, we applied ARHMM on the reaching data and found that ARHMM outperform SLDS and rSLDS in the reconstruction of neural activity. Therefore, we included a comparison between SRNNs vs ARHMMs on reaching data are as following: \| \| SRNN(P=16, K=5) \| ARHMMs (R=180, K=5) \| \| :--- \| :----: \| ---: \| \| Reconstruction (MSE) \| 0.00230($\pm$0.00041) \| 0.00232($\pm$0.00085) \| \| Behavioral States Recovery (Error) \| 0.27($\pm$0.093) \| 0.78($\pm$0.22) \| \| Training on Single Condition (MSE) \| 0.01117($\pm$0.00099) \| 0.01216($\pm$0.00104) \| We found that SRNNs and ARHMMs exhibit very similar reconstruction accuracy, both outperform SLDS and rSLDS. However, SRNNs show a slight advantage in terms of the mean and standard deviation of the mean squared error (MSE) between neural activity and the reconstructions. Despite their performance in reconstruction, ARHMMs are not able to recover behaviorally relevant states, and the states identified by ARHMMs lack interpretability. We agree that the extensions of ARHMMs, such as Switching autoregressive low-rank tensor models, are interesting to compare. We will include a comprehensive comparison between all these models mentioned above in the final version, if accepted.	Rebuttal 1: Rebuttal: Dear Reviewers, We would like to thank you for providing constructive feedback that helped us improve the paper. As a reminder, in this submission, we propose ‘Switching Recurrent Neural Networks’ (SRNNs) for discovery of switching neural dynamics that leads to behaviorally-relevant discrete states, and leads to more accurate reconstruction and prediction of neural data than baseline methods. Here, we detail responses to suggestions made by multiple reviewers, and address individual reviews below. If any questions are unanswered or our responses are unclear, we would appreciate the chance to engage further with you. Additionally, please find a PDF with helper figures attached. These are referenced and described in our responses. ### 1. Additional baseline models In addition to the existing SLDS, rSLDS, and LFADS, we implemented and have updated our paper with two more baseline models that reviewers mentioned, the SNLDS [Dong et al., ICML 2020] and mrSDS [Karniol-Tambour et al., ICLR 2024]. We show their performance on the Reaching Dataset in the attached PDF (Figure 1). To summarize, while recovering behaviorally-relevant states, our model SRNN outperforms SNLDS and has comparable performance with mrSDS, whereas in reconstructing neural activity, SRNN outperforms mrSDS and has comparable performance with SNLDS. Therefore, SRNNs have overall better performance on recovering behaviorally-relevant states and reconstructing neural activity. We would like to point out that mrSDS was published recently by ICLR 2024 in May, two weeks before our original submission, and to date, we have not found any code released with the paper. However, we implemented the code based on our reading and understanding of the paper to the best extent possible. ### 2. Determining relevant hyperparameters We now add a discussion for determining relevant hyperparameters, such as the number of discrete and continuous latent states. Importantly, we include a comprehensive hyperparameter sweep for the number of discrete states for both the reaching and decision-making datasets. We detail the salient results for the reaching dataset, where there are 5 behavioral states in the task (decision-making results below): (1) Convergence to lower number of discrete states: We tested our model by increasing the number of hidden states K while keeping the number of continuous latent states P constant. We found that 61% of SRNNs with a higher number of discrete hidden states (e.g., K=6) finally converge to the optimal number of discrete hidden states, i.e., K=5. (2) Reuse of discrete states: We also test our model by decreasing the number of hidden states. We found that 94% of SRNNs with a lower number of discrete hidden states (e.g., K=4) had at least one hidden state reused after other states: in other words, SRNNs are not able to perform well with 4 unique discrete hidden states without reusing one of them. (3) Reconstruction performance plateau: While keeping other hyperparameters constant, the reconstruction accuracy plateaus at the same number of discrete states as in the behavior, thus we can use the minimum number of discrete states as it takes for the model to perform well. We have included a figure detailing this in the attached PDF (Fig 2). Moreover, we also implemented a ‘co-smoothing’ method as suggested by Reviewer hFy4 [Yu et al., 2009 and Karniol-Tambour et al., ICLR 2024], we show the results in the attached PDF (Fig 3), where we found that K=5 also does well in reconstructing the data with a ‘co-smoothing’ neuron drop-out analysis. (4) Variability across conditions: In stereotyped tasks or experiments, such as reaching, there may not be a significant amount of variability in the timing of behavior across conditions, and this variability can thus be used as a metric for determining the number of discrete states. Here, we found that SRNNs with K=5 have much lower variability on recovered behaviorally-relevant states than K=4 and K=6 (i.e., 0.098 for K=5, 0.384 for K=4, and 0.282 for K=6). Furthermore, we did the same analysis for the decision-making data (Fig 4 in the attached PDF) and we found the same situation as in the reaching dataset. Higher K also converges to a smaller number of discrete states, and K=5 has the smallest variability across different pseudo-sessions (i.e., 0.106 for K=5, 0.287 for K=4, and 0.29 for K=6). These results demonstrate appropriate selection metrics for the number of discrete states: (1) convergence / reuse of discrete states, (2) reconstruction performance, and (3) variability across conditions or trials. We show comprehensive results for these metrics on two experimental datasets where we recover the appropriate number of behaviorally relevant states using any of these three metrics. We have now provided this in the paper, while adding a discussion for the general case. Finally, we show a comparison between values for another important hyperparameter P in the original submission. ### 3. Additional analyses In response to individual reviewers, we also implemented the following which have made our paper much stronger. (1) SRNN trained on single reaching conditions to explore failure modes of our model (2) SRNN trained on simulated single trials of binned spike counts (3) Comparison between SRNN and a standard RNN with a higher dimensional continuous latent state to assess identifiability of the models (4) Flow fields with a different random seed to assess stability of recovered dynamics (5) SRNN on two more simulated datasets: high dimensional noisy Lorenz Attractor and NASCAR ### 4. Typos in Equations and Editing of Paper We appreciate all comments and suggestions on pointing out the typos and errors of the paper. We apologize for the inconvenience during the review process. We thank all reviewers for the suggestions and comments and provide a detailed rebuttal to each reviewer below. Pdf: /pdf/425f1fc56b4c043bd4a7ef79f3868b98f8602345.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Mesa-Extrapolation: A Weave Position Encoding Method for Enhanced Extrapolation in LLMs	Accept (poster)	Summary: The paper conducts a theoretical analysis to help understand the No Position Encoding. Also, the paper proposes weave position encoding to achieve improved extrapolation performance without additional cost. Also, the paper introduces the weave PE method, Mesa-Extrapotion, which recalculates the position ID to reduce the gap between training and inference. Finally, the paper conducts experiments to prove the effectiveness of Mesa-Extropoation. Strengths: * The presentation of wave PE, Star PE, and Mesa-Extrapolation is clear. The author also provides the details of wave PE, Star PE and Mesa-Extrapolation to help understand the concepts * The author conducts experiments to prove the effectiveness of the proposed Mesa-Extrapolation. * The author also further analyzes the Latency & Memory Usage of the proposed Mesa-Extropoaltion. * The paper discusses the limitations for further discussion. Weaknesses: Major Concerns: It seems that the proposed method Star PE is the same as Self-Extend LLM [1]. If possible, I sincerely hope that the author can address the following concerns: * Concern 1: The Figure 1 Star PE implementation result does not match the Equation proposed in Section 4.1 Page 5. In Figure 5, when t-i is 5, the implementation result of Star PE is 4. However, according to the Equation proposed in Section 4.1 Page 5, the implementation result should be N+ $\lceil (t-i-N)/E \rceil$=4+$\lceil (5-4)/2 \rceil$=5. Hence, to match the implementation result of Figure 1, the Star PE calculation equation should be N+ $\lfloor (t-i-N)/2 \rfloor$. * Concern 2: The Equation of Star PE is almost the same as Self-Extend LLM. When t-i is small than N, both Star PE and Self-Extend LLM employ normal relative distance. When t-i is larger or equal to N, we discuss it below. * The Equation of Star PE is N+ $\lfloor (t-i-N)/E \rfloor$ (as shown in Figure 1), while N is called the extrapolated position and E is called the extrapolated width, and t-i is the relative distance. * The Equation of Self-Extend LLM is $(t-i)//W + (W- W//G)$=$W+ (t-i)//G - W//G$, while W is called neighbor window size and G is called group size. Apparently, when W%G==0, the Equation of Self-Exntend LLM becomes $W+ (t-i)//G - W//G$=$W+ (t-i-W)//G$= $W+\lfloor (t-i-W)/G \rfloor$. Then, we change the notation W to N and the notation G to E, we have N+ $\lfloor (t-i-N)/E \rfloor$, which is the same as Star PE. * Concern 3: If possible, could the author compare the performance between Mesa-Extropolation and Self-Extend LLM? * Concern 4: when the output sequence length $L_{generate} \gg L_{input}$, will the time cost also becomes O($L_{generate}^2$)? Based on the above concerns, the paper may need to rethink the major contribution. The proposed Mesa-Extrapolation seems to make sense and may benefit society, while the paper should clarify its original contribution. Reference: [1] Jin, H., Han, X., Yang, J., Jiang, Z., Liu, Z., Chang, C. Y., ... & Hu, X. (2024). Llm maybe longlm: Self-extend llm context window without tuning. arXiv preprint arXiv:2401.01325. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the above question and concerns in weakness. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes, the authors have addressed the limitations. We may further discuss and analyze the Mesa-Extropolation in other areas. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your work. ## major concern Thank you very much for your suggestions. The differences between our method and Self-Extend can be categorized into three aspects: Firstly, from a methodological perspective, our designed Stair PE is not only applicable to RoPE but also to other positional encoding methods, such as ALiBi. Self-Extend is only applied to RoPE. Additionally, Stair PE differs from the self-extend design approach, which involves implementing group attention and standard attention and then merging them through masking. This implies that it requires calculating the attention matrix twice, resulting in higher memory and time overhead. In contrast, our approach uses a chunk-based triangular attention matrix that significantly reduces computational and memory costs. We utilize Stair PE at the base of this chunk-based triangular attention matrix, and for practical applications, approximate methods can be employed to further save memory and reduce time costs. Secondly, theoretically, we provide a theoretical proof for the Weave PE approach, demonstrating that it can achieve extrapolation. In contrast, the self-extend approach only provides experimental evidence without a theoretical explanation. Finally, from an experimental standpoint, we further conducted comparisons of the two approaches and demonstrated that our method shows comparable advantages with lower memory consumption and inference latency (see attached PDF in Fig 5). ## concern 1 We apologize for not checking it carefully. However, there is no need to modify the formula. We can simply change it to N=3. ## concern 2 When we set the Stair PE formula as your derivation on # concern 1, your derivation is correct and aligns with Self-Extend LLM formulation. However, there is a crucial mismatch: self-extend derives the positional encoding according this formulation just for its query vectors (please refer to its pseudocode on algorithm 1 on its page 12). we briefly list below: ``` g_pos = pos // g_size # the floor operation shift = w_size - w_size // g_size s_g_pos = g_pos + shift g_q = apply_pos_emcode (q , s_g_pos ) g_k = apply_pos_emcode (k , g_pos ) ``` In its code, s_g_pos is correspond to your formula, but it’s not relative position encoding. Note that Self-Extend is only applied to RoPE positional encoding. Thus, with RoPE, the relative positional encoding is obtained by subtracting the key's positional encoding from the query's positional encoding. In contrast, our defined Stair PE directly targets the relative positional encoding. This is also why our Stair PE is more broadly applicable, not only to RoPE but also to other methods like ALiBi, whereas Self-Extend is only applicable to RoPE. According to these formulations, the relative positional encoding implemented by self-extend is different from that implemented by Stair PE. We have provided a clear comparison in the uploaded PDF (see Fig 6 ). ## concern 3 The comparative results are provided in the uploaded PDF(see Fig 1, 2, 3, and 5). In both the NIAH and Passkey tasks (corresponding to Figure 3 in our paper), our method demonstrates comparable advantages. Additionally, our method performs well in terms of memory consumption and inference latency. ## concern 4 The output of large models is divided into the prefill stage and the decoding stage. The time savings from mesa-extrapolation primarily occur in the prefill stage, which involves processing the N^2 matrix corresponding to the input sequence. During the subsequent token-by-token output process, the time complexity is O(L_generate + L_input ). This mainly concerns the last token and its computation with all preceding tokens, and this complexity is linear with respect to the sequence length. --- Rebuttal 2: Title: Response to Author Comment: Dear Authors, Thank you very much for your rebuttal. After reading the author's rebuttal, I still have questions about the contribution of Self-Extend and Star PE. * After the author's rebuttal, I confirm that the Star PE is equivalent to Self-Extend PE. * First, their formulation is equivement, after adjusting floor and ceiling operations. The core part is the same: increase 1 position number after $E$ steps and begin the operation at the $N^{th}$ step. * Then, the author claims that the Self-Extend PE cannot work on relative positional encoding. However, this is not correct * The author may check Figure 2 in the Self-Extend LLM arxiv paper. This paper Figure 2 presents how they process the relative distances between queries and keys. Self-Extend can be directly applied to Alibi, as Alibi only needs to know the relative distance between queries and keys. * I appreciate the comparison between Mesa-Extropolation between Self-Extend. The result suggests the effectiveness of the Chunk-based Triangular Attention Matrix. To increase the score, I need the author to * Rethink the relation between Star PE and Self-Extend * Rewrite the part of Star PE. * Correctly claim the contribution of the paper, while I now have relatively high confidence to say that the Star PE and Self-Extend are equivalent. Therefore, the paper cannot claim that the Star PE is its contribution, while the theoretical analysis can be kept as part of contribution. Overall, I have to keep the current score and increase my confidence, as the paper contribution Star PE is equivalent to Self-Extend. --- Rebuttal Comment 2.1: Title: Part 1 Comment: Dear Reviewer vMAr, Thank you very much for your valuable suggestions, which have been immensely helpful in improving the quality of our paper. We apologize for not fully addressing your concerns, and provide further explanations in hope of clarifying a few details as mentioned. ## Rethink the relation between Star PE and Self-Extend Based on careful examination of the respective resulting PEs, we conclude that the two methods yield similar outcomes. For example, under specific conditions (`t % G >= i % G`), your equations in concern 2 are valid and the resulting PEs are indeed equivalent. However, slight discrepancies between the resulting PEs emerge when such conditions are not met. We provide a detailed comparison at the end of this response. SelfExtend is a very recent work (published in July 2024), after our paper submission and current review process. And despite the striking similarities in the final PE results, our independent thought process was entirely different: We directly defined relative positions in Stair PE; while Self-Extend is defined by the original positions of query and key through group attention and neighbor attention. Furthermore, our theoretical contribution, as well as the novelty and effect of the Chunk-based Triangular Attention Matrix are valid and we believe add value to this field of study on LLM extrapolation. We will diligently articulate and clarify the relationship between Stair PE and Self-Extend in the revised version of the paper. ## Rewrite the part of Stair PE As advised, we will rewrite the Stair PE part (Section 4.1) in the final version, drafted as follows: “While our work was conducted independently, we note that Jin et al. have recently explored a similar idea through flooring the original positions and obtaining the relative position matrix with grouped self-attention. Although the two parallel thought processes are different, under certain conditions their formulations are equivalent (See appendix). Consequently, Self-Extend can be categorized as a Weave PE method. Our proposed Chunk-based Triangular Attention Matrix (detailed in Section 4.2) and its corresponding theoretical properties (Section 4.4) are also applicable to this parallel approach. ” ## Our Contribution Following your suggestions, we will not specifically claim Stair PE as our main contribution in the paper. Note in our introduction (line 45-60), we did not specifically claim Stair PE as our contribution either. To further clarify, we will rewrite our claims as the following: 1.Theoretical Analysis: We provide a theoretical framework that explains extrapolation. 2.Introduction of Mesa-Extrapolation: which seamlessly integrates chunk-based triangular attention matrix and a Weave PE method (e.g. Stair PE). 3.Experimental Validation to showcase the effectiveness of Mesa-Extrapolation on a broad range of tasks. --- Reply to Comment 2.1.1: Title: Part 2 Comment: ## The detailed comparison of the formulas on Stair PE and Self-Extend The formulation of Self-Extend is as follows: let `t` denote the position of query, `i` denote the position of key, `W` denote the neighbor size and `G` denote group size. According to their paper’s equation (3) , `P = P // G`, and the shifted relative position (`W - W // G`), we can get the position of query as: `t // G + W - W // G (Equ.1)` and the position of key as: `i // G (Equ.2)` By RoPE, their relative position between `t` and `i` (consider `t > i`), is remapped to: `t // G + W - W // G - i // G = W + t // G - i // G - W // G (Equ.3)` Note this equation is slightly different from your equation in Concern 2, which is `W + (t-i) // G - W // G`. We would like to clarify that only under `t % G >= i % G` (necessary and sufficient conditions), we get `t // G - i // G` is equal to `(t-i) // G`. Next, when `W % G == 0`, the equation of self-extend becomes: `W + (t-i) // G - W //G = W + (t-i-W) // G ` If we adjust this flooring operation to ceiling operation, and replace N with W and E with G, then Stair-PE is equivalent to Self-Extend. In summary, under conditions `t % G >= i % G` and `W % G == 0`, as well as changing the flooring operation to ceiling operartion in Self-Extend, these two formulas are equivalent. Except for this condition, they yield slightly different results. Note because `W` and `G` are constants, the condition `W % G == 0` can be met easily. But `i` and `t` change with positions, therefore there will always be positions where the condition `t % G >= i % G` is not satisfied. For example, `t = 10, i = 5, G = 2`, `10 // 2 - 5 // 2 != (10-5)//2`. This is why, as shown in the uploaded PDF(see fig 6), the relative positions differ between Self-Extend and Stair PE at certain positions (when `t % G < i % G` ). Specifically, in Stair PE, the relative positions on the diagonal of the matrix always remain the same. In contrast, the diagonal values of the attention matrix in self-extend are not always the same. We will add discussions to the final version of our paper in appendix. Despite these subtle differences, we agree with your point that their core idea are the same. We are very grateful for your valuable suggestions and insightful observations, which have helped us think more deeply about their relationship. If you believe there are any issues with our analysis or have other concerns that haven't been addressed, please let us know, and we will make further improvements to enhance the quality of our work. Once again, thank you very much.	Summary: This paper studies the length extrapolation of LLMs. 1. It provides a theoretical analysis of why NoPE and PE fail to extrapolate beyond a certain length. Previous work has shown that this failure is related to the explosion of hidden states as positions increase. This paper demonstrates that both NoPE and PE suffer from this hidden state explosion, using a constructive approach to illustrate the existence of Transformer weights. 2. It proposes weave PE, a simple adaptation of PE that theoretically addresses the extrapolation issue. It also provides a simple implementation of weave PE, using a chunk-based triangular attention matrix. Then, it demonstrates that the proposed extrapolation scheme matches the performance of prior length extrapolation methods, such as Dynamic-NTK. Strengths: - Great theory explains the failure of NoPE and PE in length extrapolation. - Proposes weave PE, derived from the theoretical analysis, which also works well in practice. - Shows good empirical results in passkey retrieval, language modeling, and summarization. Weaknesses: 1. Methodological comparison with $\Lambda$-Attention The proposed Stair PE resembles the $\Lambda$-attention of LM-Infinite & Streaming-LLM, yet with differences in 1) the additional attention at the bottom, and 2) a different length extrapolation scheme, Meta-Extrapolation. In the experiments, Meta-Extrapolation significantly outperforms LM-Infinite & Streaming-LLM. Could the authors provide the intuition behind these empirical gains? --- 2. Empirical comparison with Dynamic-NTK Dynamic-NTK outperforms Meta-Extrapolation on the summarization task for mid-lengths of 7-11k, while Meta-Extrapolation shows better performance on summarization for shorter lengths of 4-6k and better language modeling fluency for lengths greater than 11k. Could the authors provide the intuition behind these results? --- 3. Relation between input sequence length $T$ and effective length $M$ The theorems only show the existence of an effective length $M$, but do not provide intuition on the scale of $M$, such as the ratio over the input length $M / T$. Could the authors provide some intuition on this? If I understand correctly, $M$ is set from the construction of the Transformer weights, so can it be controlled to an arbitrarily large number? --- Editorial comments - The fonts of the figures and tables are too small. Please make them more readable. - Some parts of the writing are mechanical. For example, lines 116-120 do not provide meaningful information. It would be great to discuss the implications of the theorems in natural language. For instance, both theorems state the failure of length extrapolation in NoPE and PE, rather than just "revealing the internal mechanism of extrapolation." Technical Quality: 3 Clarity: 3 Questions for Authors: See Weaknesses. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Well discussed. Extending this approach to fine-tuning would be an interesting next step. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for recognizing our work and providing your suggestions. ## weakness 1 For decoder-only architectures, the model's output is based on the next-token prediction. The last token in the input is crucial, as it generates the next token. For the last token to output the correct next token, it must attend to all previous tokens (principle of self-attention mechanism), especially those with critical information. The attention matrices used by Streaming-LLM and LM-Infinite both discard middle tokens, inevitably resulting in loss of information from these tokens. This problem becomes severe when the middle tokens carry important information. Our approach appends all tokens at the bottom, ensuring that information is not lost. For overly long inputs, we use Stair PE to reorganize relative position encoding, maintaining the integrity of the information. ## weakness 2 We believe that the inherent capabilities of LLMs set the upper limit for these methods. As the input length increases, especially beyond the maximum training window, errors are bound to accumulate regardless of the method used, which will affect the model's performance. The NTK method scales the rotation angle of RoPE positional encoding. We hypothesize that within an 11k input length, the errors introduced by adjusting the rotation angle are acceptable for large models. For mesa-extrapolation, we reuse previous trained positions to minimize errors as much as possible. We speculate that with a fixed extrapolation width param (E = 50), the 7k-11k range may spread the model's attention more thinly compared to the 4k-6k range. For the summary task, the model needs to pay closer attention to the dispersed information within the context, which requires a finer extrapolation width to minimize errors. We hypothesize that optimizing the extrapolation width could alleviate or improve performance in the 7k-11k range. ## weakness 3 Yes, M is set within the Transformer weights we constructed. This M is typically related to the maximum training window length of the model. For example, if the maximum window length for Llama is 2k, then M can be considered as 2k. Therefore, we think that M can't be controlled to an arbitrarily large number. ## Editorial comments Thank you for the modification suggestions. We made corresponding adjustments according to the comments. --- Rebuttal 2: Comment: Dear Reviewer bJoG, We sincerely appreciate your positive feedback, and it is incredibly gratifying to have our work highly recognized by you. As the discussion progresses, if you have additional questions, please let us know, and we will be happy to provide further clarification. We will express our deep gratitude to you in the Acknowledgments section of our paper. Once again, thank you very much. --- Rebuttal Comment 2.1: Title: Response to the Rebuttal Comment: Thank you for the rebuttal. I believe the paper is strong and will maintain my positive rating. However, I am not a strong expert in this area, so I will keep my original weak confidence. It would be helpful if the authors clarified answers to my questions in the revised paper. Specifically, the statement "M is typically related to the maximum training window length of the model" and thus "M can't be controlled to an arbitrarily large number" should be emphasized, as this significantly impacts the interpretation of the theorems. It's also interesting to observe how adding bottom attention enhances Streaming-LLM and LM-Infinite, especially when middle tokens contain critical information. Analyzing passkey retrieval based on its location could provide valuable insights, particularly if the improvement is most pronounced when the passkey is in the middle. --- Reply to Comment 2.1.1: Comment: Dear Reviewer bJoG, We greatly appreciate you for recognizing our work. ## the interpretation of Theorem Thank you for your valuable suggestions to improve the interpretation of the theorems of our paper. We will incorporate your suggestions into the final version of the paper. ## Enhancing Streaming-LLM / LM-Infinite We believe the question you proposed is very valuable. To address it, we suggest drawing our Mesa-Extrapolation design to add bottom-level attention for Streaming-LLM and LM-Infinite. Here’s how to add it: Input Tokens: \|--------------------- Context Tokens ---------------------\|---------- Last Tokens -------\| #Operations : \|<-----------Streaming-LLM / LM-Infinite----------->\| <--- Bottom Attention-->\| The diagram above illustrates how input tokens can be processed by dividing them into two parts: context tokens and last tokens. The context tokens are processed normally using Streaming-LLM or LM-Infinite, with the resulting key/value pairs being cached. When processing the last tokens, the last tokens’ key/value pairs can be appended with the cached key/value pairs(context tokens) to compute the attention using context forwarding (many operators, like Flash Attention, have already implemented this). During this context forwarding, if the keys are too long, the positions can be handled using the Weave PE methods (such as StairPE, ReRoPE, or Leaky-ReRoPE). We believe this approach can further enhance the capabilities of Streaming-LLM and LM-Infinite. We hope our responses addresses your questions. We will do our best to answer any further questions you may have before the final discussion deadline arrives. Once again, thank you very much.	Summary: The paper proposes a positional embedding scheme to address the extrapolation issue: train on short sequences, evaluate on longer sequences. Authors propose a theoretical framing of the positional embeddings contribution to attention. They apply their analysis to NoPE (No Positional Embedding) and to standard PE, and RoPE. They propose the Mesa-Extrapolation idea where input tokens are organized so that attention is paid to nearby tokens and those at other key positions. Authors validate their findings with empirical evidence on several benchmarks and applications. Strengths: The paper is about a very relevant topic which has attracted a lot of attention lately. The paper proposes a simple approach to solve the problem which seems to be easy to adapt to different positional embedding models. Some of the numerical experiments are encouraging. Weaknesses: The theory part of the paper is hard to read and I am not sure about its usefulness. Result appear hand-wave-y and vaguely stated. For example the definition of the threshold H in the Assumption is surprising (see questions). Numerically, experiments on language modeling and Summary of Tasks do not seem to show the method's claims. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. Can authors explain the threshold definition: "When o > H, LLM extrapolates successfully. Once o < H, LLM extrapolation fails." Is there a typo and inequalities are reversed? 2. In Fig 2, why dim 1 & 6 are of interest? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: -- Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your work. ## Weakness 1 & question 1: “Theory is hard to read and unclear definition of the threshold H” Definition of the extrapolation success or failure: When a large model continuously produces valid next-tokens for a given long input sequence, we define this as successful extrapolation. Conversely, if it outputs invalid next-tokens, we refer to this as failed extrapolation. Therefore, the key to extrapolation failure is the output of invalid next-tokens, which is caused by anomalies in the model's output. We speculate that these anomalies arise because certain parts of the model receive input values that exceed their acceptable range. Definition of the threshold H: For convenience in our analysis, we consider the outputs of the preceding layer In a multi-layer neural network, which serves as the input for the next. To ensure normal subsequent outputs, the values from the preceding layer need to remain within a reasonable range. This implies that there must be a bound on these values. We refer this observable boundary as the threshold H. This threshold can be either an upper bound or a lower bound. In our work, for the sake of theoretical analysis, we chose the lower bound as our threshold. Definition of hidden state values o: the output of each layer in an LLM, which also serve as the input for the next layer. Note that in our assumption, we assume that for specific dimensions in specific layers, there exists an acceptable range of input values (i.e., hidden state values o). The lower bound of this acceptable range serves as our threshold H. By observing whether the hidden state value o in this dimension exceed the threshold H, we can predict whether the large model's extrapolation has failed. We apologize if it wasn't clear enough. We will add these refined definitions in the final version of the paper. ## weakness 2: Despite observing minor performance degradation in the 8k-11k token range for summarization tasks, our method consistently exhibits strong extrapolation capabilities across a diverse set of tasks, including passkey retrieval, language modeling, and summarization, as corroborated by Reviewers LfLf and bJoG. We also noted that the performance in the 4k-6k range is good. We speculate that with a fixed extrapolation width param (E = 50), the 7k-11k range may spread the model's attention more thinly compared to the 4k-6k range. For the summary task, the model needs to pay closer attention to the dispersed information within the context, which requires a finer extrapolation width to minimize errors. We hypothesize that optimizing the extrapolation width could improve performance in the 7k-11k range. ## Q1 The inequalities in our Assumption are not reversed since we consider H as the lower bound. We will add clear definitions and explanations in the final version. ## Q2 Dimensions 1 and 6, as well as dimensions 7 and 9 in the appendix (line 1246), were selected from the first 20 dimensions to showcase whether the extrapolation would fail using the observed threshold for NTK and Rerope. Note that different inputs activate different parts. Figure 2 used 'hello' as the probe word. Additionally, we demonstrated that dimensions 14 and 19 can be used to predict extrapolation failures using 'good' as the probe word (please refer to the uploaded PDF in Fig 7 and 8 for more detailed diagrams). --- Rebuttal 2: Comment: Dear Reviewer PZAX, As Discussion Period progresses, we would like to know if our responses have addressed your concerns. We believe our responses address all of your main concerns; if they do, we would very appreciate it if you raised your score accordingly. If not, please let us know so that we can address any serious concerns that you might still have. Once again, thank you very much. --- Rebuttal Comment 2.1: Comment: Thanks, I have read the rebuttal. I am still having trouble understanding Assumption 1's "If o > H extrapolation succeeds, if o < H, it fails": so in particular, take H = 100; we are talking about a model which for a sequence of length 10 < H = 100, extrapolation will fail, but with a long sequence of length (say) 100000000000000 > H, extrapolation succeeds. I have a hard time understanding what the assumption means here. I am still feeling uncomfortable with the choice of dimensions 1 & 6 out of the top 20 dimensions. I do not see any reason why embedding vectors would be aligned with canonical basis vectors e_1 and e_{20}. I fear that the choices made here are very data specific and hard to reproduce in other setups. Another question: in the design of the chunk-based triangular attention, a strong assumption is made that tokens in the middle of the context are related to surrounding tokens and those at the beginning of the sequence. How did we validate this assumption? --- Reply to Comment 2.1.1: Comment: Dear Reviewer PZAX, Thanks for providing your further concerns. ## About Assumption 1 We suspect there is a misunderstanding on the definition of o. In our assumption, "o" represents the hidden state value, not the sequence length of the input. Please refer to our first response for the definitions of "o" and "H". For example, consider an input sequence of length I with tokens denoted as below: input_tokens = $[x_0, x_1, x_2, …, x_I]$, where $x_i \in \mathbb{R}$ After transformation by the Transformer matrix, we obtain the query vector, key vector, and value vector as below: Q = $[q_0, q_1, q_2…., q_I]$ K = $[k0, k1, k2…., k_I]$ V = $[v0, v1, v2, …v_I]$ where $q_i, k_i, v_i \in \mathbb{R}^{d}$, e.g d =4096 for Llama2. Passing them to the Self-Attention part, we obtain that: P = $Q^T K \in \mathbb{R}^{I \times I}$ O = $V \cdot Softmax(P)$ = $[O_0, O_1, O_2, ….., O_I]$ where $O_i \in \mathbb{R}^{d}$, and let $O_i = [o_1, o_2, ..., o_{d}]^T$. Please note that $O_i$ corresponds to the hidden state values at position i. We used $o$ to represent the hidden state value at a specific dimension, that is, $o_j$ is the j-th element of $O_i$. We assume that at a specific dimension (e.g. $o_j$ or dim j), there is a reasonable and acceptable range for this value. For simplicity, we have assumed a lower bound. If the value in this dimension falls below this lower bound, it could cause abnormal outputs in subsequent layers, ultimately leading to abnormal model behavior. This construction is intended to demonstrate that threshold-based criteria can be used to assess whether extrapolation is successful. ## About dimensions We suspect that there might be some confusion about dimensions in our context and canonical basis vectors. In our context, we use “dimension j” to describe the j-th element of the output vector $O_i$, as stated above . Therefore the concept of canonical basis vectors e_1 and e_{20} is not relevant in our context. Here we only chose specific examples of dimensions to illustrate the phenomenon of extrapolation failures. We observed that this behavior is consistent across different inputs and tasks (passkey, language modeling etc.). By observing the corresponding threshold, we can predict the extrapolation capability of the LLM, which is limited by its maximum pre-training window, regardless of the input. Given the consistent results across various experimental conditions, we believe our findings have broad implications and are not data-specific. ## About the design of the chunk-based triangular attention First, our assumption that the beginning tokens and surrounding tokens are important is well-established and has been used in many of the previous works, such as Streaming-LLM and LM-Infinite. And the reasoning is below: Beginning of the sequence: 1) people often place instructions, task descriptions, etc., at the start of a sentence, which is related to the way prompts are structured. 2) a start token is typically added by default, which is an important position holder. For instance, Llama2 automatically adds a <bos> token at the beginning. We’ve also explained this in detail in the appendix (line 1206-1209). Surrounding tokens: According to the principles of the attention mechanism, attention decreases as the distance increases, meaning the nearby tokens generally receive higher attention scores. Focusing on the initial tokens and their neighbors has been validated by other previous works (e.g,. LLM-Streaming and LM-Infinite). When designing the chunk-based triangular attention, we also experimented it and found this design to be the most effective. Note that Mesa-Extrapolation appends all previous tokens at the bottom of its attention matrix, allowing the last token to attend to all preceding tokens. In contrast, Streaming-LLM and LM-Infinite also discard middle tokens at the bottom of their attention matrix, inevitably resulting in loss of information from these tokens, leading to inferior performance. If you still have any questions regarding the explanation above, please let us know so that we can provide further clarification. Once again, thank you very much.	Summary: This paper introduces a new LLM length extrapolation method, called Mesa-extrapolation, which utilizes a chunk-based triangular attention matrix and applies stair PE. The proposed method is based on theoretical analysis. The paper conducts extensive experiments on passkey, PPL, summarization to demonstrate the effectiveness. Strengths: 1. The paper provides a theoretical analysis to prove the effectiveness of meticulous weave position with PE for length extrapolation. 2. The proposed method is efficient and is proved to be effective through extensive experiments. Weaknesses: 1. The passkey retrieval experiment is simple, good performance on the passkey is far from a real usable context window. Please consider to add evaluations on Ruler[1] and RepoQA[2] 2. The achieved context window is limited. [1] https://arxiv.org/abs/2404.06654 [2] https://evalplus.github.io/repoqa.html Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Length extrapolation is a necessary technique, but the current extrapolation length is very limited. Considering that there are already many models that have undergone long context window extension, such as phi3-mini-128k, can your proposed method continue to perform length extrapolation on these long context LLMs? If so, it will significantly enhance the impact of your method 2. If I understand correctly, the proposed method is mainly for those with PE. Why is there a need to prove NoPE? Is NoPE your baseline? 3. The proposed Mesa-extrapolation is somehow similar as a variant of "sliding window attention + attention sinks". Could the author explain why mesa-extrapolation is theoretically superior compared to sliding window attention and attention sinks? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See the weakness and question sections. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your work. ## Weakness 1: We add evaluations on Ruler, and the results are available in the uploaded PDF (see Fig 1, 2 and 4). These results indicate that our method also performs well. ## Q1 & weakness 2: We perform experiments on long context window "microsoft/Phi-3-mini-128k-instruct" model using Ruler dataset. The results show that our method can further extrapolate to 192k, indicating the effectiveness of our method. We show it in the uploaded PDF (see Fig 4). Due to the inherent limitations of the phi3-mini-128k-instruct model itself when handling this task, even within a 128k window, the phi-3 model does not achieve 100% accuracy as shown in the right-side of Fig 4. Our method effectively extends its window based on the original model's capabilities, rather than improving its capability. ## Q2 Although existing LLMs typically incorporate positional encoding (PE) components in their architecture, recent studies have shown that the NoPE (no positional encoding) method may better facilitate the extrapolation capabilities of LLMs. The choice of including or excluding the PE component in a LLM is critical, as it cannot be altered once the model is trained. We believe it is essential to theoretically model the extrapolation performance of both NoPE and PE. We discuss this part in our related work. Due to a lack of suitable candidates, NoPE was excluded from our baseline. Aside from the 1B-size pre-trained model (Kazemnejad, Amirhossein, et al. "The impact of positional encoding on length generalization in transformers." Advances in Neural Information Processing Systems 36 (2024).), we did not find any other suitable candidates. However, this pre-trained 1B model’s capabilities are very limited, making it difficult for evaluating basic NLP tasks. Additionally, by simply incorporating training-free weave PE into a PE-type LLM, we can achieve longer extrapolation. Through these efforts, we hope to demonstrate that choosing weave PE offers more advantages compared to NoPE. ## Q3 Mesa-Extrapolation appends all previous tokens at the bottom of its attention matrix, allowing the last token to attend to all preceding tokens. When the input length significantly exceeds the model's maximum training window, it uses Stair PE to reuse positions and reduce errors further. The attention sliding window approach, on the other hand, discards tokens in the middle section. This method restricts the last token's attention to only the head and tail tokens, keeping the total attended token length within the maximum training window and directly using trained positions. This approach sacrifices the information from the middle tokens. In context-based tasks, the middle tokens often contain important information, and losing these tokens can prevent the model from providing accurate responses to user queries. --- Rebuttal 2: Comment: Dear Reviewer hjGM, Thank you for your valuable suggestions and the time you’ve dedicated to our paper. As the discussion period draws to a close, we would like to know if our responses have addressed your concerns. Following your suggestion, we have conducted verification on phi3-mini-128k, which significantly enhance the impact of our approach. If our responses have resolved your concerns, we would greatly appreciate it if you could raise your score to a clear "accept". Once again, thank you very much. --- Rebuttal Comment 2.1: Title: Thank you for your rebuttal Comment: Thank you for the rebuttal. The additional experiments addressed some of my concerns. However, I saw some performance drop between 32k-128k when extending phi3-128k. Therefore, I will increase my rating score to 6 accordingly. --- Reply to Comment 2.1.1: Comment: Dear Reviewer hjGM, Thank you very much for increasing your score. We will express our gratitude for your valuable suggestions in the Acknowledgments section of our paper. Additionally, regarding the performance decrease observed in the 32k-128k range, we would like to clarify further: First, our method can be applied only after 128k, not in the 32k-128k range, because it is inherently a free plug-in approach, and the effective input window of phi3-mini-128k itself can reach up to 128k. In practical scenarios, we would adopt our method only when the input exceeds 128k. Second, we have observed that the phi3-mini-128k model itself experienced a performance drop within the 32k-128k range, as shown on the right of Fig 4 (uploaded PDF), indicating its insufficiency in handling multi-key task within this range. Since our method is designed to assist the model in handling inputs beyond its effective 128k limit, not to enhance the model’s inherent capability, the performance issues beyond 128k are likely due to the inherent limitations of model for the multikey task, not to our method. In contrast, the Fig 4 left (uploaded PDF) shows that the phi3-mini-128k model itself successfully handles the singlekey task within the 128k range. Consequently, our method can extend its maximum effective input window beyond 128k, up to 192k. We hope our analysis effectively addresses your concerns and contributes to improved score. Please let us know if you have any further questions. Once again, thank you very much.	Rebuttal 1: Rebuttal: Dear reviewers, Thank you very much for your review. We have provided additional experimental supplements in the uploaded PDF. Please check it out. Pdf: /pdf/f0f359fb459877376a284d73a25b6112167dcd5e.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The authors propose a weave position encoding method to enhance LLMs’ inference performance when the input context window exceeds the training context window. This method can be integrated into existing pretrained LLMs without additional finetuning. To support their findings, the authors conducted theoretical analyses on the failure reasons of various position encoding methods, including those without position encodings. They demonstrate that the significant shift in the hidden state’s value range, when input token positions exceed the maximum context length, is the cause of this phenomenon. Strengths: One of the strengths of the proposed method is that it can be integrated into existing pretrained LLMs without requiring any additional finetuning. This makes the method highly practical and easy to implement, saving both time and computational resources. The method has demonstrated excellent performance in pass key retrieval tasks, showcasing its effectiveness in real-world applications. This indicates that the proposed approach not only works in theory but also delivers tangible improvements in practical scenarios. The authors have conducted comprehensive theoretical analyses to understand the failure reasons of various position encoding methods, including those without position encodings. This thorough investigation provides a solid foundation for the proposed method and enhances its credibility Weaknesses: The proposed position encoding method, while promising, does not consistently improve performance across different tasks. This inconsistency suggests that the method may not be universally applicable or reliable in every context, potentially limiting its overall utility. Additionally, the main narrative of the paper emphasizes the method’s ability to handle extrapolation beyond the training context window. However, given the observed variability in improvements, it would be more accurate to adjust the claims to better reflect the method’s performance, providing a more balanced and realistic presentation of the work. Technical Quality: 3 Clarity: 2 Questions for Authors: The caption for Figure 1 is not sufficiently informative. Additionally, it is unclear how the failure of an LLM is measured in Section 3.4 and Figure 2. The experiments visualizing hidden state values in Figure 2 would have been more effective if conducted on the same task and with the same setup as Figure 3. This alignment would allow for a clearer connection between the findings in Figures 2 and 3. minor typos: Theorem 3.2: an simple -> a simple Line 161: defer to -> refer to Line 242-243: a significant disruptions Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have included a limitations section; however, it reads more like a discussion of future work rather than addressing the actual limitations of the current study. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your work. ## weakness Thank you for your suggestions. Our method shows good extrapolation performance on accuracy-related tasks, but we observe slight variability in extrapolation performance within mid-length (8k-11k) in the summary task. We will adjust our claims accordingly in the final version. ## Q1 Illustration of Mesa-Extrapolation. The left figure shows the Chunk-based triangular attention matrix (before SoftMax operation) of Mesa-Extrapolation while a sequence of length 13 is fed into an LLM. The right figure shows an example of PE and Stair PE. The Stair PE is used to weave the relative position equipped by Mesa-Extrapolation. ## Q2 In sec 3.4, we designed a probe experiment by repeating a word (e.g. “hello”) N times as input to the model, where N is the input length (the details are provided in Appendix, line 1206-1209). The reason for designing the probe this way is to eliminate the influence of different tokens since both of the input token and its position affect the hidden state values. In Figure 2, we use a vertical black dashed line to indicate the position of maximum training length of the model. In this case, it is 4k for llama2-7b-chat model. The hidden state value at this position is designated as the observed threshold and marked with a horizontal red dashed line. When the hidden state value exceeds the red dashed line as the position changes, it signifies that the hidden state value has surpassed the threshold, suggesting a failure in extrapolation after that position. We will add these explanations to the final version. ## Q3 We used a different experimental setup primarily for the following reasons: Figure 2 examines extrapolation failures, which primarily occur when input length surpasses the maximum limit, irrespective of the underlying task. Using the "hello" probe can clearly help us analyze the relationship between position and the corresponding thresholds. We then verify these predictions with other tasks, such as those depicted in Figure 3 or language modeling tasks shown in Figure 4. The results indicate that the threshold observed aligns with the results from the passkey task and language modeling task, demonstrating the effectiveness of the theory we developed. Nevertheless, we added additional results visualizing hidden state values using the same task and with the same setup as Figure 3 (see Fig 8 in attached PDF ). ## Q4 Thanks. We fixed these typos. ## Limitations Due to limitations of resources, we have not yet validated our method at longer lengths. For instance, we have verified that phi3-mini-128k model can be extrapolated to at least 192k with our method. Beyond this length, memory crashes occur. --- Rebuttal Comment 1.1: Comment: I appreciate the authors’ thorough rebuttal. The authors have effectively addressed all of my concerns and questions. Additionally, I have reviewed the other reviewers’ comments and the authors’ responses to those as well. Taking all these materials into account, I am increasing my final rating from 5 to 6. I believe the paper meets the criteria for acceptance as a poster presentation. --- Reply to Comment 1.1.1: Comment: Dear Reviewer LfLf, Thank you very much for raising your score, and we deeply appreciate the valuable suggestions you have provided. We will express our sincere gratitude to you in the Acknowledgments section of our paper. Once again, thank you very much. --- Rebuttal 2: Comment: Dear Reviewer LfLf, As the discussion time goes by, we would like to know if our responses have addressed your concerns. If our responses have resolved your concerns, do you think our paper should now be a clear "accept"? We are very grateful for your valuable suggestions. If you have additional concerns, please let us know so that we can address them and further improve the quality of our paper. Once again, thank you very much.	null	null	null	null	null	null
GenRL: Multimodal-foundation world models for generalization in embodied agents	Accept (poster)	Summary: In this work, the authors propose learning a pixel-based reconstructive world model, and then separately learn networks to convert the representations of a pretrained VLM into the learned world model latent space. By using a VLM trained via contrastive alignment, this essentially enables the projection of both image as well as text inputs into the latent space of the world model, and therefore simple similarity can be used to provide rewards for downstream policy learning. Strengths: This reviewer is a supporter of the idea of unifying the representation spaces of a large-scale pretrained VLM and that of a world model. This author appreciates the benefits: matching behavior of a world model with natural language can enable text-conditioned generalization. The preliminary experiments show promise. Originality: This work appears decently original. Quality: This works quality is acceptable. Clarity: The clarity of the work is acceptable, the core ideas are communicated clearly. However, there are lots of open questions surrounding this work that could be elaborated upon further. Significance: This work appears to be decently significant, as a preliminary investigation in this space. Weaknesses: Chief amongst the weaknesses of this work is the limited environments applied to, and also the limited baselines (essentially, the only existing work the authors compare against is VLM-RM). The authors can consider comparing against other forms of text-conditioned policy learning, such as LIV for the kitchen setting, or Text2Reward and similar approaches for the general case. It also seems a strange setup to normalize in-between expert and random, and report results in this way. This reviewer is unaware of prior work that performs evaluations in this way. What is the rationale behind this evaluation strategy compared to what is used in prior work? Details about certain components of the model and how they are implemented are sparse. For example, is the aligner a video generative model (text-to-video model)? How is it implemented? It is a bit dissatisfying to rely on a corrupted version of vision as a language embedding. It seems strange that the aligner should on one hand be learning to bring language embeddings meaningfully across modalities to the image/video space, which the authors motivate is necessary because of the multimodality gap. However, the authors then treat language embeddings as a noisy corruption of a video embedding - so essentially the training objective for the aligner is essentially a denoising? And rather than bridging a modality gap, the aligner is essentially a denoiser? Why do we not learn the reverse direction, where we optimize a world model's latent space that projects into existing VLM space? This design decision is not elaborated upon, but seems more intuitive to this reviewer. From the video demonstrations, on the associated project website, it is rather unclear what is happening. Are Behavior Retrieval videos from expert policies in an offline dataset that are matched with a particular text prompt/input video? What are those text prompts/input videos? It's not clear what the retrieval setup is. For Multitask Generalization, it is also not obvious what the corresponding prompts are. Furthermore, the results for multitask generalization do not seem smooth and natural, despite being simplistic DM_Control environments (especially the case for their proposed simplified Stickman environment) and they are missing Kitchen environments. In the end, it appears that their method is still good as a retrieval technique (retrieving already-achieved expert behaviors in "Behavior retrieval") due to the underlying VLM, and is decent at reconstructing video prompts, but still suffers in terms of learning coherent policies (e.g. what is visualized in "Multitask generalization"), which is ultimately what is of interest. For the video prompts that are decoded, it appears as if almost all of them are rather stationary (with the exception of the cheetah/dog example and the human dancing example) - they collapse to a stationary goal pose. Perhaps this is because the clips are so short (8 frames) that essentially it boils down to pose-matching. It is not obvious that this is that beneficial in supervising motion; so why does this improve upon static image supervision? Indeed, many of the results that are shown learned by the policy are rather stationary and do not have much movement (most are just jitters around a stationary pose). It then begs the question how this approach improves upon just a static goal supervision. However, the authors simultaneously find that in static tasks other methods outperform the authors' approach. This reviewer pushes back on the term "data-free RL", as there still needs data (and interaction data) to learn their method. This is a very confusing terminology, and honestly the generalization comes from the large-scale pretrained VLM - it would be more appropriate to reuse the terminology of zero-shot reward models or zero-shot policy learning used in prior works across alignment methods (vision-language models are zero-shot reward models for reinforcement learning, [Rocamonde, '23]) and diffusion (text-aware diffusion for policy learning, [Luo, '24]). This reviewer really enjoys the work but believes there are many open questions that warrant further explanation. Furthermore, the evaluation suite (environments) and comparison suite (benchmarks) is rather weak. The idea is indeed neat, but the execution leaves much to be desired, and therefore this reviewer believes the work is of borderline quality. Technical Quality: 3 Clarity: 3 Questions for Authors: Why Stickman instead of Humanoid? Humanoid has been able to be solved with pixel-based world models in the past (Appendix A of Dreamerv2). What are the specifications of Stickman and with what criteria was it designed? Why did the authors use a simple GRU? Why was a more advanced world model not used, like RSSMs? Was this tested or ablated over? Why were there no multi-task generalization experiments performed for Meta-World kitchen? Would the authors consider Text2Reward and other approaches that learn a reward function as data-free RL, as there do not need additional data to generalize to learning new policies? Alternatively, the data-free RL paradigm sounds like zero-shot generalization for policy learning, which is already offered by VLM-RM and other similar works. Why did the authors choose to generate the video demonstrations synthetically, rather than use actual natural video clips? Would the performance not be better when using natural videos which are more in line with what the base VLM was trained on? How are the rewards computed using temporal alignment? Essentially, are only the rewards for the most-aligned segments across the target trajectory and the agent used as rewards, and for all other timesteps a 0 reward is provided? This computation seems rather expensive for long-horizon trajectories. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations section seems acceptable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful comments. Experiments We added new experiments, including stronger model-based baselines. Details are in the main rebuttal message. We also tested all baselines with LIV's representation in the Kitchen tasks. We used LIV's open-source code to download and instantiate the model. Note: the available model is the general pre-trained model, not the one fine-tuned for the Kitchen environment. \| \| IQL-LIV \| TD3+BC-LIV \| TD3-LIV \| WM-CLIP-LIV \| GenRL \| \|-------------------\|-------------\|-------------\|-------------\|-------------\|-------------\| \| kitchen microwave \| 0.03 \| 0.0 \| 0.2 \| 0.0 \| 0.97 \| \| kitchen light \| 0.53 \| 0.05 \| 0.0 \| 0.0 \| 0.46 \| \| kitchen burner \| 0.2 \| 0.0 \| 0.0 \| 0.0 \| 0.62 \| \| kitchen slide \| 0.03 \| 0.03 \| 0.0 \| 0.67 \| 1.0 \| LIV's results confirm the original paper's claims (see Appendix G4) that the representation does not work well for vision-language rewards without fine-tuning on domain-specific vision-language pairs (which are unavailable in our settings, as we use no language annotations). Text2Reward requires a different experimental setup than ours. We focus on visual RL. The agent observes the environment only through images and doesn't have any privileged information about which entities are present in the scene e.g. joints, limbs, objects. Normalizing results using expert performance and random performance as max-min values is one of the de-facto standards for the Atari suite. This is referred to as Human Normalized Score (HNS), see [1]. Aligner The aligner is not a generative model and, as the reviewer suggests, it's closer to a denoiser network. In the main rebuttal comment, we have provided an additional explanation about how and why we reduce the multimodality gap by adding noise to the embeddings. Additional insights are available in related work [2,3]. As described in the Appendix (Line 491-492), "the aligner network employs a small U-Net, with a bottleneck that is half the size the embedding representation.". The embedding representation is 768-dimensional. Further details can be found in our accompanying code implementation, provided to the Area Chair. Inverse connector idea We have implemented this idea and added it as a baseline (WM-CLIP) to our work. The reviewer's intuition is correct as the idea works. However, without the connecting-aligning process, the system does not perform as well for some prompts/tasks. Please, see the main rebuttal comment for further details. Website visualizations We added the corresponding task labels in the website and we provided the language prompts in the Appendix. In the videos, we observe that less natural behavior is correlated with a lower normalized score on the task. Thus, we expect those behaviors to be less smooth, e.g. see 'stickman flipping' or 'walker high kick'. Dynamic tasks We provided additional results and visualizations, including more dynamic tasks. Please, see the main rebuttal comment and the updated website. Nonetheless, we agree with the reviewer that the model struggles more to sync with input videos for dynamic actions. One of the reasons for this is that the control frequency of the agent is different from the motion frequency from the provided video prompt. This makes it hard to match the prompt accurately. Data-free We provided an additional explanation in the main rebuttal comment. We believe that both Rocamonde et al, Luo et al, and Text2Reward fall in the category that we represent as "Offline RL + VLM", as they require data for training the behavior policy. Given that the term "data-free RL" may be confusing, we have changed our naming convention to "Data-free Policy Learning", similar to Luo et al, which we now cite for reference. Humanoid We designed the Stickman environment to explore tasks that require upper body limbs (e.g. boxing, doing a handstand) without the complexity of training a humanoid, which requires a significantly larger amount of data to be solved, as reported in the DreamerV2 paper. Building on the Walker model, the number of joints is increased by 4: 2 joints per arm, one is for the shoulder, the other for the elbow. The total number joints is 10. The action space is normalized to be in [-1,1] as all dm_control tasks. Connector architecture The architecture of the connector network is the same as the RSSM architecture used for the world model, which employs a GRU for computing the hidden state. We corrected this detail in the paper. Additional tasks Meta-World environments adopt different environments for different tasks, i.e. with different objects, workspaces and thus visual dynamics. Instead, we focussed on evaluating on domains where we can learn a single MFWM for the environment's dynamics and use it to generalize to new tasks. The Franka Kitchen environment allows for more tasks in a single environment (we evaluate on 4). However, it's hard to design new tasks as the number of objects to interact with is limited. Video prompts Natural videos actually work well (and often better) than AI-generated videos. We found intriguing the idea of generating the prompts using another GenAI model, rather than having to look for a video online. In the additional results, we chose to also show some natural videos. Temporal alignment Before the aligned segments, the reward is computed using the initial state of the target sequence. This is currently stated at Line 155. We hope to have satisfied all the reviewer's concerns and we look forward to receiving updated feedback. [1] A Review for Deep Reinforcement Learning in Atari: Benchmarks, Challenges, and Solutions, Fan el al [2] Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning, Liang et al [3] Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data, Zhang et al --- Rebuttal 2: Title: Reviewer Response [1] Comment: This reviewer appreciates the detailed rebuttal from the authors, and apologizes for the tardiness in response. Here are provided thoughts in response: On experiments: The additional comparison against LIV is appreciated. This reviewer understands the point the authors make, that Text2Reward requires a different experimental setup. However, the project has quite a broad scope: it offers a way to condition on text and perform a behavior, as well as condition on an image (or video) and perform a behavior as well. In principle, this behavior should be compared against other methods that do this as well, not necessarily isolated to multimodal rewards (which technically the proposed approach does not fall into the category of either). Therefore Text2Reward would be interesting to show, because it is another technique that converts natural language to a policy (albeit requiring learning - this distinction can be made in analysis). Furthermore, methods that go purely from image (or video) to policy behavior would be interesting to show - for example, VIP [2] (an image-only ancestor of LIV), or LEXA [3] (which appears remarkably similar to some of the Video prompts decoded section of the website) or other goal-conditioned or pose-conditioned RL. For each capability of the model, it is useful to demonstrate comparisons against existing techniques that perform similar capabilities for completeness; it is indicative of the power of the method the authors are proposing, that there are so many potential capabilities, and this adds greatly to the excitement of the approach - but at the same, in principle, time such capabilities should be thoroughly explored and vetted. On evaluation: This reviewer is aware that HNS is a standard evaluation technique for Atari, but was under the impression that the purpose of such a technique was to measure superhuman performance (indeed, this is suggested as much in Subsection Human Average Score Baseline of [1], which the authors have linked). It seems strange to apply it to an expert policy, which is still just a policy and not a human. Under such circumstances, where every policy is equally a policy, why does HNS-style evaluation still make sense rather than just raw episode rewards? This reviewer has not seen this evaluation technique outside Atari; it definitely does not seem common practice for reporting DeepMind Control Suite performance (raw rewards), nor for FrankaKitchen (success rate). From this reviewer’s perspective, it does not seem to detract from the story to report results in the standard way, and it is puzzling that HNS was chosen for these instances without strong justification. On the Aligner: The clarification in the rebuttal PDF as well as the response is very useful, and appreciated. On the Inverse Connector: This reviewer appreciates the additional results, and the preliminary insights are very interesting. Provided the final draft includes these results paper, as well as analysis into the performance discrepancy based on projection direction, this reviewer is willing to increase their rating. This reviewer believes examining and explaining the projection direction would strengthen the connection between VLM-RM dynamics and in-domain world modeling dynamics, and the paper would benefit from featuring it in the main text. On updated website visuals: This reviewer is confused by the note that “less natural behavior is correlated with a lower normalized score on the task. Thus, we expect those behaviors to be less smooth, e.g. see 'stickman flipping' or 'walker high kick'.” There does not seem to be a score for such tasks that could be referred to as low; nor should there be, since they are novel behaviors. To what are the authors referring to? Also, apologies for the poor memory - but this reviewer really cannot recall the original website examples, and therefore cannot meaningfully determine the delta; in future updates would the authors please prominently demarcate the delta (perhaps with different colors/fonts/highlightings)? On Dynamic Tasks: This reviewer agrees with the analysis that control frequency and motion frequency mismatch is a difficult consideration to overcome, and can potentially contribute to a preference for stationary policies. [1] A Review for Deep Reinforcement Learning in Atari: Benchmarks, Challenges, and Solutions, Fan, 2021. [2] VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training, Ma, 2022. [3] Discovering and Achieving Goals via World Models, Mendonca, 2021. --- Rebuttal Comment 2.1: Title: Reviewer Response [2] Comment: On Data-free: The name “Data-free Policy Learning” does seem more appropriate, as subsequent policies no longer need data to be trained (the data is only used to learn a world model), and can be done in imagination alone. It is totally fine under scrutiny, but also sounds strange on first hearing - “data-free” and “learning” seem like an oxymoron. Perhaps “Policy Learning in Imagination”? Or “Data-Free (Policy) Generalization through Imagination”? The term could definitely be workshopped, but this is not a critical matter. On Humanoid: This reviewer agrees that Humanoid is more complex; however, learning the world model is a one-time fixed cost, on top of which arbitrary new policies can be learned. Would learning new policies in imagination still incur a high cost? Intuitively, the world model should decrease training cost because planning is possible. On the Connector: Sure. On Multitask Generalization: Apologies, the reviewer meant to refer only to FrankaKitchen. Regardless, it would be interesting to train the world model on a subset of available tasks and have a holdout novel task. Multitask generalization capabilities for a robotics environment would be very exciting to demonstrate, even if for a simpler setup. Alternatively, simply showing a novel task generalization for the robotic arm (not necessarily tied to any task), like what was demonstrated for Stickman in the website, would also be very interesting to show. On Natural Video Prompts: Sure. On Temporal Alignment: This certainly seems hacky, but not a critical issue of the work. Ultimately, this reviewer appreciates the updated experiments, explanations, and insights (e.g. the projection direction), and is willing to tentatively upgrade the score in favor of acceptance. However, there still remain reservations about the thoroughness and completeness in verifying all capabilities of the proposed approach, as well as a thirst for more results in the domains chosen (e.g. FrankaKitchen/robotics multitask generalization, and as a [non-critical] reach, some Humanoid demonstration). Further clarification and justification would also help (e.g. the choice of HNS when humans are not involved, which to this reviewer's knowledge is a rare design decision). --- Reply to Comment 2.1.1: Comment: We would like to thank the reviewer for the extensive feedback provided. We are glad the additional material was useful and improved the reviewer's opinion of our work. ## On experiments > Text2Reward requires a different experimental setup. However, the project has quite a broad scope: it offers a way to condition on text and perform a behavior, as well as condition on an image (or video) and perform a behavior as well. We could not find any experiments or examples in the Text2Reward paper where they condition on images. The closest we found on the paper is, in their Figure 6, the experiments were the user provides a feedback on the observed rollouts. However, this requires a human in the loop looking at the rollout and providing verbal feedback to the system. This is currently out of scope for our work and might be considered in future extensions. We definitely agree with the reviewer that additional baselines are useful in assessing the generality of an approach. However, in order to provide a fair comparison, it is important that we benchmark all the baselines in the same settings. If certain baselines require significant changes to the data used, the inputs to the system, or the foundation models adopted, it becomes harder and harder to compare methods in a fair way. For the Behaviors from video prompts experiments, all the baselines adopt the same training data, the same foundation model to process the videos (InternVideo2), the same video prompts. The only difference between them is the underlying algorithm to specify rewards and the policy learning method adopted. These are our main contributions and thus these are the aspects we would like to compare in these experiments. ## On evaluation > This reviewer is aware that HNS is a standard evaluation technique for Atari, but was under the impression that the purpose of such a technique was to measure superhuman performance (indeed, this is suggested as much in Subsection Human Average Score Baseline of [1], which the authors have linked The reference provided states two motivations for the HNS: one is mentioned by the reviewer, the other is "Performance across algorithms become comparable. Like Max-Min Scaling, the human normalized score can also make two different algorithms comparable." We have provided Atari as an example for (human) normalized score, as this is one of the most common benchmarks. However, max-min scaling between expert and random performance, or more simply, max scaling by expert performance, is very common in the literature. We provide some additional references as follows: - A Generalist Agent, Reed et al, 2022. From Figure 5's description: "Here values on the x-axis represent a specific percentage of expert score, where 0 corresponds to random agent performance" - URLB: Unsupervised Reinforcement Learning Benchmark, Laskin et al, 2022. From Figure 3: "Scores are normalized by the asymptotic performance on each task (i.e., DrQ-v2 and DDPG performance after training from 2M steps on pixels and states correspondingly)." - TD-MPC2: Scalable, Robust World Models for Continuous Control. Hansen et al, 2024. All the results are provided using "Normalized scores" (the way this is computed is not clearly stated in the paper but we assume it's either expert performance or the maximum achievable performance on each task) - (ProcGen) Leveraging Procedural Generation to Benchmark Reinforcement Learning. Cobbe et al, 2020. From page 3: "For each environment, we define the normalized return to be Rnorm = (R − Rmin)/(Rmax − Rmin), where R is the raw expected return and Rmin and Rmax are constants chosen to approximately bound R." The reason why it's so important to normalize returns/scores is because different tasks have different performance scales. For instance, the maximum return in Cheetah Run is around 850, for Quadruped Run is around 650, for Walker Walk is around 970. For the Kitchen tasks, since we follow the original paper [1] in using success rate as a metrics, the possible scores are only 0 and 1. These large differences in the performance scales require us to use a way to normalize scores and Max-Min scaling is a very common way in the literature to do so, when multiple different tasks are involved. [1] Relay Policy Learning: Solving Long-Horizon Tasks via Imitation and Reinforcement Learning, Gupta et al, 2019 ## On the Inverse Connector We thank again the reviewer for providing the idea for these additional experiments and we are glad the additional insights are found to be useful.	Summary: The paper looks at a method for leveraging foundation multimodal models for learning world models in RL. They do so by aligning the latent space of a video language model with that of a generative model that can be used for learning in imagination. This is done by training connector-and-aligner networks . The rewards for a task can then be derived by measuring the cosine sim between representations of the states visited by a policy and the states generated by the connector aligner network when it is conditioned on a language-based task prompt. A policy can be optimised to maximise this alignment based reward. Strengths: Transferring foundation model knowledge to improve policy learning is an open problem of interest to the community. The paper provides a successful recipe for aligning a foundation model with the world model for a specific domain that we want to do policy learning in. The paper is written well. I'm currently being conservative in giving a borderline accept score, since some aspects of the method are not clear to me (I have addressed this in my questions below) - but I will be happy to raise my score after engaging with the authors once they have addressed these questions. Weaknesses: 1. I would have expected that simple tasks with clearly distinguishable static end states (such as standing) should have worked equally well with CLIP rewards, however the table shows a big difference between the proposed method and the image-language reward baselines even on those tasks, which leads me to think that the baselines may be missing out on some component that the proposed method has. What could be missing, or is this intuition wrong? 2. The generations in Fig 6a are actually not accurate at all - many of the poses don’t correspond to the humanoid pose if you look closely and would actually optimize learning to strike the wrong pose if a policy is trained with it. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Why is setting b=imagination horizon the same as doing no alignment? (Line 160) 2. I’m not completely sure how you train the aligner-connector network: is it done by 1) using images collected from the downstream domain (in this case from the mujoco sim), 2) getting their VLM visual embedding and their world model encoder embedding and aligning those? As for the text part, is this done by corrupting the VLM visual embedding (to approximate the language embedding) and aligning it again with the world model encoder? What is the policy used to collect the data and resulting data distribution? I understand that Fig 5 is somehow related to this question but this could be made clearer. For eg. which task’s policy is chosen to collect the data to train the MWFM for the results in the main table (how is this policy related to the task being evaluated)? 3. The discussion around Figure 5 is not very clear to me - how do we infer that “’run’ data proves more generalizable than ’walk’ or ’stand’ data across tasks“ - the figures suggest that training on ‘stand’ led to the highest rewards for downstream tasks 4. “This can be explained by the fact that the target sequences that GenRL infers from the prompt are often slightly in motion“ - could you explain why that would be the case (it inferring the closest matching state as one that is in motion)? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The paper includes a brief discussion on limitations of their method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful comments. Image-language CLIP results > simple tasks with clearly distinguishable static end states (such as standing) should have worked equally well with CLIP rewards We agree with the reviewer's intuition and we believe the results confirm their statement. If we look only at the static tasks (stand tasks and kitchen tasks) we see that the performance of Image-based and Video-based baselines are very comparable. We observe this also for the new model-based baselines (please, see main rebuttal message). GenRL, which uses a video-language VLM, tends to outperform the other approaches. Video prompts' poses The visualizations in Figure 6a are the interpretations of what the visual prompt provided would look like in the embodied domain, according to the model. We agree these do not match the poses provided exactly, but only in their "semantic" meaning (e.g. doing crunch abs, despite the legs position being different). This issue is due to two aspects of our framework: (i) VLMs shine with semantic interpretation but struggle to represent precise geometrical information, and thus they struggle with precise pose estimation, (ii) the world model may have not seen certain poses in its training dataset, and thus it will provide the closest behavior to the prompt. Despite this limitation, we found that our video prompting allows to learn behavior from a variety of visual conditions, e.g. draw sketches, very different camera views. Please, see our main rebuttal message for the additional experiments. Sequence alignment If we set $b$=imagination horizon we only have one possible comparison between the two sequences. Example with a sequence of length 3. Target sequence states (T), agent sequence states (A). b = 1 (initial state alignment), 3 possible alignments \| T \| \| \| \|---\|---\|---\| \| A \| A \| A \| \| \| T \| \| \|---\|---\|---\| \| A \| A \| A \| \| \| \| T \| \|---\|---\|---\| \| A \| A \| A \| b = 2, 2 possible alignments \| T \|T \| \| \|---\|---\|---\| \| A \| A \| A \| \| \| T \| T \| \|---\|---\|---\| \| A \| A \| A \| b = 3 (no alignment), 1 possible alignment \| T \|T \| T \| \|---\|---\|---\| \| A \| A \| A \| Training procedure and data collection We think the reviewer correctly understood our training procedure, and which we further clarify in our main rebuttal comment for clarity. As currently stated in Lines 174-176, we collect data using two kind of agents: (i) a Dreamer V3 agent learning to perform the task (we added the details about the agent being DreamerV3 in the main text) and (ii) a Plan2explore agent, collecting exploration data in the domain. The datasets are detailed in Appendix. However, we provided detailed results per task on the behavior extraction tasks so that a reader can quickly see which tasks are contained in the dataset (indeed, the behavior extraction tasks). We have made this clearer in the main text, by stating it at Line 190 as follows: "Behavior extraction. We want to verify whether the methods can retrieve the tasks behaviors that are certainly present in the dataset. For these tasks, the replay buffer of an expert agent learning to solve the tasks (DreamerV3) was already included in the mixed dataset used for training" Training data distribution We believe there is a misunderstanding in the way Figure 5 visualizes the data. The labels (all, expl, run walk, stand) indicate the portion of the data used for training the agent. Each plot represents performance of the agent trained with different portions of data on a task / group of tasks. Thus, the agent trained only on 'stand data' is the one with the lowest performance. We changed the labels to "all/expl/run/walk/stand data" to improve clarity. Static poses matching Examples of predicted sequences being in motion for static tasks are present in our website. For example, if we look at the decoding of the "downward facing dog" prompt, we observe that the pose of the agent is in motion, rather than being completely static. We observed similar issues with some of the task prompts. We hope to have satisfied all the reviewer's concerns and we look forward to receiving updated feedback. --- Rebuttal Comment 1.1: Comment: Dear Authors, Thanks for your response, I have now read through the responses and the pdf. My impression is that the connector-aligner is doing a lot of the heavy lifting (not just in terms of more trainable parameters but by adding more flexibility to the language representations since they are processed by a denoiser). Is there a way to compare to the CLIP rewards more fairly then, for eg. by training an aligner in a similar fashion on top of the CLIP visual embeddings and then using the image/text encoders + the aligner at test time? That will help to really decouple where the performance gap is coming from. --- Reply to Comment 1.1.1: Comment: Dear Reviewer, We want to thank you for your feedback. We agree that the connector-aligner mechanism, which along with GenRL's reward function represents our main contribution, is crucial in our work. As requested by the reviewer, we ran additional experiments on all 35 tasks (behavior retrieval + multi-task generalization benchmarks) to establish the importance of the aligner network. In the following table, we report the results of two additional methods: - GenRL - no aligner: this is an ablation of GenRL where the the language prompt's embedding is directly fed into the connector, rather than processing it first with the aligner. - TD3-V and WM-CLIP-V + aligner: for these baselines, we first process the language prompt's embedding using GenRL's pre-trained aligner. Then, we use it to compute the cosine similarity for the reward function, as for the original baselines. We summarize the results by domain in the following table (2-3 seeds to increase for the camera-ready): \| \| GenRL - no aligner \| GenRL \| WM-CLIP-V \| WM-CLIP-V + aligner \| TD3-V \| TD3-V + aligner \| \|:---------:\|:------------------:\|:------------:\|:------------:\|:-------------------:\|:------------:\|:---------------:\| \| quadruped (6 tasks) \| 0.17 ± 0.02 \| 0.90 ± 0.02 \| 0.81 ± 0.04 \| 0.76 ± 0.05 \| 0.32 ± 0.04 \| 0.33 ± 0.04 \| \| walker (9 tasks) \| 0.19 ± 0.01 \| 0.75 ± 0.01 \| 0.70 ± 0.02 \| 0.74 ± 0.01 \| 0.56 ± 0.04 \| 0.48 ± 0.04 \| \| stickman (13 tasks) \| 0.09 ± 0.01 \| 0.66 ± 0.01 \| 0.54 ± 0.03 \| 0.50 ± 0.03 \| 0.38 ± 0.02 \| 0.38 ± 0.02 \| \| cheetah (3 tasks) \| 0.32 ± 0.02 \| 0.93 ± 0.01 \| 0.82 ± 0.17 \| 0.84 ± 0.02 \| 0.31 ± 0.09 \| 0.77 ± 0.04 \| \| kitchen (4 tasks) \| 0.25 ± 0.00 \| 0.76 ± 0.08 \| 0.71 ± 0.14 \| 0.84 ± 0.09 \| 0.32 ± 0.16 \| 0.27 ± 0.10 \| \| overall (35 tasks) \| 0.17 ± 0.00 \| 0.76 ± 0.01 \| 0.67 ± 0.03 \| 0.68 ± 0.02 \| 0.40 ± 0.02 \| 0.42 ± 0.02 \| We can observe that: i) the aligner mechanism is crucial in GenRL's functioning. ii) processing the language embedding in the reward function of the WM-CLIP-V and TD3-V baselines changes performance on some tasks (performance per domain varies). However, using the aligner provides no advantage overall. Intuition behind the results We believe the aligner is very important in GenRL because its output, the processed language embedding, is fed to another network, the connector. If the language embeddings were not processed by the aligner, they would have been too different from the embeddings used to train the connector, which are the visual embeddings. We provide an additional explanation for this in our main rebuttal comment. Instead, for the baselines, we process the language embedding with the aligner and then use it to compute a similarity score with the visual embeddings. This overall renders very similar performance to no aligner processing, hinting that the aligner network doesn't improve the cosine similarity signal. At the same time, this also suggests that the aligner network doesn't hurt the generality of the VLM's embeddings, as the cosine similarity after processing the embedding provides a similarly useful signal as before processing. We hope this provides additional insights into our work and we look forward to receiving additional feedback. --- Rebuttal 2: Comment: Thank you for your response, and I agree that the connector-aligner is a contribution of the work. The additional experiments in the rebuttal help to disentangle the contributions of the different components but it is still not clear to me why GenRL without aligner is so much worse than the baselines when they don't even benefit from the aligner? My guess for where the advantage of GenRL is coming from, would be that the connector network is basically allowing you to go beyond the performance of VLM-RM style methods by adding flexibility to how the similarity is taken (for eg. maybe some dimensions of the VLM representation are not informative or important when computing the similarity for rewards, and they are adding noise to the similarity scores, then the connector network allows you to ignore these dimensions when transforming them to a different representation space?). Regardless of the exact reason, (which I hope the authors will do more digging into and analyze for the camera ready), I think the proposed additions are allowing the proposed method to extract/transfer concepts from VLMs to a greater extent than prior work, for learning control policies. Thus I will update my score to reflect this. --- Rebuttal Comment 2.1: Comment: > The additional experiments in the rebuttal help to disentangle the contributions of the different components but it is still not clear to me why GenRL without aligner is so much worse than the baselines when they don't even benefit from the aligner? As the reviewer suggested later in their comment, GenRL's performance advantage might be attributed to its policy learning being driven by a latent features similarity, rather than by the similarity of the CLIP embeddings. Without the aligner, GenRL fails (almost completely) to translate language embeddings into world model states. Thus, the similarity of the latent features is not meaningful for solving the task at hand. We can provide additional visualizations about this in the Appendix for the camera-ready, where we show the difference between decoded language prompts with and without using the aligner. Without the aligner, the decoded prompt sequences tend to not follow the given language prompt. From this addition and the ablation study performed, it should be clear that the aligner's use is bridging the gap between the two modalities (language and vision) for the connector, rather than providing "denoised" CLIP embeddings. We would like to thank the reviewer for their valuable feedback, which has significantly improved the quality of our work. The additional ablation study and the clarifications we incorporated based on the reviewer's recommendations have improved our presentation. We are also very pleased to see that our rebuttal positively influenced the reviewer's opinion of our work, leaning towards a firm acceptance recommendation.	Summary: This paper proposes to combine a DreamerV3-style world model with a pretrained vision language model (VLM). By training two small adaptors to align the latent space of the VLM with that of the world model, the aligned representations from the VLM can be used as a reward signal to train agents in the world model. The training process consists of two main parts. 1) There is a large offline dataset needed in the environments of interest (prompt and trajectories of states and actions), generated by expert RL agents and a random policy. This trains the world model and the adaptors. Each environment (domain) uses a separate world model. 2) Actor-critic agents are trained purely within this world model’s imagination, a separate policy for each task. The paper shows these agents outperform standard model-free offline RL methods trained on only the large offline dataset. It also shows some effectiveness at generalizing to new tasks within an environment, specified with a new text prompt. Strengths: - The core idea of the paper is very nice. There is a lot of interest from the community in working out how to get value from the broad general knowledge locked away in LLMs and VLMs, into RL agents. This paper offers a novel way to attack this – to my knowledge world models have not been used in this context before. - The results are not dazzling, but they indicate the approach works and it outperforms standard (though perhaps weak) offline-RL baselines. Section 4.2 shows promise in generalizing to knew text prompts in an existing environment. Weaknesses: My main criticism of the paper is that the narrative oversells what the core work actually supports. I detail examples below. Overall I’d suggest either presenting the work that has been done comprehensively in a more reserved manner, or adding the required work to support the broader claims and experiments. Either way, I think changes would be large enough to require a resubmission. I’m disappointed to not be able to give the paper a higher score as I liked the main idea. - The capability of the model to condition on visual goals is presented as a main functionality of the model – featuring in the first figure, the abstract, and throughout the paper. But the only evidence to support this is a very brief and qualitative experiment (Figure 6a). Everything else is conditioned on text. I am of the opinion that conditioning on visuals would likely work, but the paper must present good evidence to support this. - Several aspects of the title ‘Multimodal foundation world models for generalist embodied agents’ are misleading. 1) Only one modality is really tested (as in prior point). 2) ‘Foundation world models’ suggested I'd see a single very general world model. But in Appendix D is an important detail -- each environment learns a separate world model, so they are only general or foundational within a specific Mujoco embodiment. This kind of detail is important and should be honestly discussed in the main paper. 3) A ‘generalist agent’ is referred to, but every agent in the paper only performs a single specialist task, there is nothing general about the agent’s themselves. - The method is reported as needing ‘no language annotations’ (line 42). This is not true. The large offline dataset requires text prompts accompanying each trajectory. - The paper claims to be ‘the first large-scale study of multitask generalization from language in RL’ (line 165), but I can think of others. Language table is the first that comes to mind. - One of the motivations for the work is that reward functions can be hard to specify, while language is a more natural form. However, the large offline dataset is generated by using multiple expert agents which need reward functions. - ‘Data free RL’ is suggested as a new paradigm for foundation models in RL. I’d argue that this is simply know as zero-shot generalization to most in the community. - Main experiments are presented in Table 1. Whilst the offline-RL methods are one comparison point, I’m not sure how comparable they are, since they are all model-free while GenRL is model based. Are there any model-based variants that would be easily considered as baselines? The differences are reflected in the different compute times required – GenRL takes 5 days for world model training +5 hours per policy, while the baselines take 7 hours per policy. This seems like an unfair comparison, especially to withhold the detail to the appendix. - Results in Minecraft are briefly mentioned in Section 5. But so few details are given that I am lost as to what it is showing. This should either be removed or full details added. - The paper presents a new stickman environment. But details are sparse. The authors have failed to correctly identify this in Checklist Section 13. Technical Quality: 1 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 1 Presentation: 3 Contribution: 3 Limitations: Fine. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful comments. Learning from visual prompts In order to support our claims with empirical evaluation, we have provided results of behavior learning from video prompts. The results can be found in our main rebuttal message and the videos on the website. We have three main observations about these experiments: (i) GenRL performs overall better than the baselines, (ii) the performance from video prompts are generally close to the performance from language prompts, (iii) leveraging the foundation VLM knowledge allows to generalize in very different visual settings. Title We hope to have resolved the reviewers' concern about the multimodality nature of our work with the additional experiments. The term we coined as a shortand for our MFWM models is not used to indicate a class of foundation (world) models. Instead, as we state several times in the paper (Lines 9, 42, 105, Figures 1 and 2), it is used to indicate world models whose representation is connected to the knowledge of multimodal foundation models. In order to make this more clear, we added an hyphen to our name ("Multimodal-foundation world models") so that it is clear that the first two words are used as an "adjective" for the world model, rather than for indicating a multimodal foundation model. We understand the definition of a "generalist agent" may be ambiguous, as we have a "generalist" model that allows multi-task and multimodal prompts generalization by training multiple specialist policies, as correctly stated by the reviewer. Given the above observations, we will update our title to: "Multimodal-foundation world models for generalization in embodied domains" Moreover, following the reviewer suggestion, we moved the statement about training one world model per dynamic, currently in the Appendix, to the main paper (Experiments section). No language annotations There must have been a misunderstanding, probably due to an ambiguous statement at Lines 176-177. However, we use no language annotations in our datasets. Please, see our main rebuttal comment about this. Large-scale study To the best of our knowledge, our work is the first large-scale study where agents can generalize to as many embodied tasks and domains (35 tasks, in 5 embodied domains). Nonetheless, we will remove the adjective "first", if the reviewer finds it inaccurate. Data collection As we developed our framework, we observed that, in order to solve more complex tasks, the agent requires some expert data/demonstration of the complex behavior. We analyse this in our "Training data distribution" experiments. We believe this limitation is, to some extent, inevitable, as data-driven deep learning agents need to observe complex behaviors during training in order to be able to replicate them. In this work, we used an expert RL agent (DreamerV3) to collect the data for us. Using a small set of demonstrations might be an alternative. We added further discussion about this and other limitations in the Appendix of the paper. Data-free We discuss the differences between Offline RL, GenRL with data and GenRL data-free in our main rebuttal comment. Generally, we would argue that a single offline RL policy cannot generalize to new tasks zero-shot, without (re)training on the data. Skill-based approaches hold the potential to do so. However, training skill-based policies that cover many different complex behaviors in a zero-shot fashion is an open research question. Our data-free strategy offers an alternative view on this problem. Comparison with baselines As we describe in the main rebuttal message, we added model-based baselines to our main experiments. These baselines follow the recommendation of Reviewer Yb6H of learning an "inverse connector" from the world model representation to the VLM representation (GenRL does the opposite). Please see the main rebuttal message for further details. GenRL pretrains a single world model per dynamic, for 5 days, and uses it for learning tasks in imagination, for 5 hours per task. Model-free RL requires only 7 hours, as it takes longer to converge, but has no pre-training stage. On a single GPU, model-free RL is faster to train for a small number of runs. GenRL becomes advantageous when using the world model for training policies for more than 60 runs (runs = tasks x seeds). If using data-free learning (3 hours per task), the advantages become significant after 30 runs. Given the nature of our work, focusing on behavior generalization performance, rather than on computing budget, and the limited space for the main text, we have kept this information (now including the above discussion) in Appendix. Experimental settings To make room for the new visual prompted experiments, we are moving the Minecraft results in Appendix. We have also added more details about the training setup we adopted. As we stated in Line 295, we used DreamerV3 to collect the data and thus used the same settings as their work. Nonetheless, we also summarized the main information in our work. The Stickman environment is based on the Walker environment from the _dm_control_ suite. We designed the Stickman environment to explore tasks that require upper body limbs (e.g. boxing, doing a handstand) without the complexity of training a humanoid (which requires a significantly larger amount of data to be solved [1]). The number of joints is increased by 4: 2 joints per arm, one is for the shoulder, the other for the elbow. The total number joints is 10. The action space is normalized to be in [-1,1] as all _dm_control_ tasks. The robot also presents an head, to resemble a humanoid. Further details can be found in our accompanying code implementation, which we shared with the Area Chair. [1] Mastering Atari with Discrete World Models, Hafner et al, 2020 We hope to have satisfied all the reviewer's concerns and we look forward to receiving updated feedback. --- Rebuttal Comment 1.1: Comment: Thanks for the response -- I appreciate the large amount of work and effort that went into the rebuttal. At this point my feeling is that properly interrogating all the new material would best be done in a fresh round of reviews when the changes are integrated into the main paper, though naturally I will discuss this with other reviewers. I did want to clarify the point on the language annotations. I'd assumed the prompts listed in Table 3 were used at training and test time -- i.e. _trajectory-level_ annotations were not used but _task-level_ annotations were. Is this not the case? --- Reply to Comment 1.1.1: Comment: We would like to thank the reviewer for acknowledging our rebuttal. > I did want to clarify the point on the language annotations. I'd assumed the prompts listed in Table 3 were used at training and test time -- i.e. trajectory-level annotations were not used but task-level annotations were. Is this not the case? We would like to confirm this is not the case. As we detailed in our rebuttal, no text annotations or text embeddings have been used during training of the MFWM components, neither at the trajectory level nor at a task level. We only used visual embeddings for the training of the connector and aligner. The prompts listed in Table 3 are not annotations, as they are not associated with any data. These are the language prompts that are used to specify the tasks to solve (is this what the reviewer indicates as "test time"?). As done in previous work, for all methods, these prompts are only used to compute the rewards for policy learning. In the baselines, the reward is computed by obtaining the cosine similarity between the prompt's embedding and visual embeddings, similar to what is described in [1]. In GenRL, the prompt's embedding is used to generate latent targets (using the connector and aligner), to be achieved by leveraging Equation 3. [1] Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning, Rocamonde et al, 2024 > At this point my feeling is that properly interrogating all the new material would best be done in a fresh round of reviews when the changes are integrated into the main paper, though naturally I will discuss this with other reviewers. During the rebuttal period, we have done our best to provide all the reviewers with the material and the information they requested. Following the conference guidelines, we have provided a single-page PDF to share the results of the experiments the reviewers have asked us to add to our work. We have also provided additional Figures to aid the reviewers' understanding of parts of our work which seemed less clear. Overall, the only new material to be added to the main text is represented by the experiments from the visual prompt that this reviewer requested. The other experiments only require adding two columns (the new baselines) to existing sets of experiments. The additional Figures will go in the Appendix for clarification, if the reviewers found them useful (so far we received no feedback on them). We want to thank the reviewer once again, as their detailed review has been very useful in improving the clarity of our work and in stating our claims more firmly, thanks to the additional empirical evidence. We would strongly appreciate it if the reviewer could provide us with updated feedback on our work and let us know if the material provided and the proposed edits (e.g. in the title) improved the reviewer's opinion about our work.	Summary: The paper wants to leverage the large-scale pre-training of foundation models trained on internet data to train a world model for embodied agents that generalizes across tasks and domains. This is done by training a world model in the standard way, but in addition training aligner and connector networks that (1) map language embeddings to video embeddings and (2) map video embeddings to world model latent states. At inference time, this allows conditioning the world model on a task language prompt and then training in imagination to learn policies. Strengths: - On the website, the reconstruction results from language and video are nice and quite unexpected (I'm unsure why the aligner and connector networks are able to generalize to new prompts) - The problem the paper is trying to solve is relevant, especially given the mismatch in data availability between embodied and vision / language settings Weaknesses: - The main claim of the paper is strong generalization performance, leveraging the internet scale pre-training of video-language models. The bottleneck is the generalization ability of the networks which map embeddings from the video-language model to the world model latent states, and on the quality of the world model itself. I don't see why the aligner and connector should generalize. - Given the main claim, I would like stronger baselines / ablations in the generalization and data-free settings. Currently, there are no baselines in the data-free case which makes it impossible to assess how well the method generalizes. - Many of the experimental details are unclear in the paper (please see my questions). I encourage the authors to explain these better in the rebuttal and camera-ready, and also provide some intuition for why their method is better than the baselines. - In the single task, offline RL case, all the baselines are model-free, whereas the proposed method utilizes a model. I would have liked to see at least one model-based baseline to confirm that the improvement is because of the better reward signal and not because of the model-based optimization. - In the single task, offline RL case, reward is computed by looking at the similarity between the representations of the task prompt and the image / video. In the case of the base lines, these representations are fixed (eg. CLIP / Intern2Video representations), whereas for the proposed method these are taken from the last layer of the model learnt on the data itself. This is also reflected in the compute budget - the model takes 5 days to train (in addition to the 5 hours of training in imagination). Technical Quality: 3 Clarity: 1 Questions for Authors: - What is the value of $k$ (number of frames predicted by the connector)? What duration does this correspond to? What happens if the task is longer than this duration? - Just to confirm, in the offline RL evaluation, first the world model is trained on the offline data (only for that particular domain) and then the policy is trained in imagination? In that case, why is there a difference in the time taken for the actor-critic to converge in the data-free setting (see line 524) - In the single task, offline RL case, is the aligner trained with only as many language prompts as the tasks in that domain? If that is the case, it would be trained to reconstruct $e^{(v)}$ corresponding to many different videos in the offline dataset, some of which might contain suboptimal trajectories which have nothing to do with the language prompt. How can we expect the aligner to learn anything useful in this case? - If the aligner is trained only on a few language prompts, how is it able to generalize to new tasks? - What exactly is the multi-task generalization setting? In this evaluation, does the method get access to offline data from the OOD task? If yes, how is it used to train the policy? If no, how are the model-free baselines trained in this setting? Confidence: 4 Soundness: 3 Presentation: 1 Contribution: 3 Limitations: Yes, the authors adequately assessed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful comments. Connector-aligner generalization First, we would like to make clear that, as stated in multiple parts of the paper, the connector and aligner networks are trained using vision-only data and no language annotations. We have provided additional clarification for this in our main rebuttal message. The connector-aligner synergy in the MFWM allows to rely on the multimodal foundation model's knowledge for grounding visual and language prompts into the embodied domains dynamics. The extent to which this system generalizes depends on the knowledge possessed by the two main components: (i) the foundation VLM and (ii) the world model. Given a certain concept, expressed through a prompt, GenRL is able to understand the concept and learn the corresponding behavior given two conditions: * the VLM has been pre-trained on vision-language pairs that allow understanding that concept. We assume this is the case as foundation models are trained on massive datasets; * the world model has been trained on a dataset that contains visual observations that match the given concept. Given a mixed dataset, containing many tasks and exploration data, we know this is true for the tasks in the dataset, but we cannot know for other tasks beforehand. In practice, we show that the tasks we deliberately put in the dataset can be retrieved (Behavior retrieval experiments) and the system also generalizes to many other prompts/tasks. This means the behaviors are likely to be found in the exploratory data or as part of the tasks data. We assess this generalization capability quantitatively, in the Multitask generalization experiments, and qualitatively, with the visualizations on the website. Stronger baselines As we describe in the main rebuttal message, we added new model-based baselines to our main experiments. These baselines follow the recommendation of Reviewer Yb6H of learning an "inverse connector" from the world model representation to the VLM representation (GenRL does the opposite). This baseline is stronger than the model-free baselines and it allows as to more clearly establish the main ingredients that contribute to the stronger performance of GenRL: (i) using a video-language model helps in dynamic tasks, (ii) using a model-based algorithm is beneficial, (iii) the connection-alignment system presented outperforms the other straightforward way of connecting the two representations (world model and VLM). As for the data-free settings, we do not have knowledge of any other data-free behavior learning strategies. As, to the best of our knowledge, this is a new paradigm for behavior learning, we aim to establish its performance compared to more established paradigms, e.g. offline RL. We have provided additional information about our unique data-free pipeline in the main rebuttal message. Missing implementation details The following information has been added to the paper: Value of k: we adopt a number of frames $k=8$ (as stated at Line 158. We also added this information at Line 104 to make it clearer) Multi-task generalization settings: all methods are trained on the same dataset, made of structured data (the behavior retrieval tasks) and unstructured data (exploration data). Assuming that the data distribution is varied enough, many behaviors that are not part of the behavior retrieval tasks are likely observed. The goal of all methods is to understand the received prompt and learn the "best-matching behavior". This can be done from the given dataset, for the model-free RL baselines, or in imagination after pre-training the world model, for the model-based baselines and GenRL. We observed that GenRL is able to succeed in many tasks that we didn't deliberately add to the dataset. Training time GenRL pretrains a single world model per dynamic, for 5 days, and uses it for learning behaviors in imagination, for 5 hours per task. Model-free RL methods require 7 hours but have no pre-training stage. On a single GPU, model-free RL is faster to train for a small number of runs. GenRL starts becoming advantageous when using the world model for training for more than 60 runs (which is often the case, considering the number of runs = N seeds x M tasks per domain). When adopting the data-free learning strategy, GenRL doesn't rely on the dataset at all (see main rebuttal message). This halves the time required for training, as there are no data transfers between the CPU (where the dataset is loaded) and the GPU for training. We hope to have satisfied all the reviewer's concerns and we look forward to receiving updated feedback. --- Rebuttal Comment 1.1: Comment: Thank you for the clarifications, I have updated my score. --- Reply to Comment 1.1.1: Comment: We would like to thank the reviewer for their valuable feedback, which has significantly improved the quality of our work. The additional experiments and clarifications we incorporated based on the reviewer's recommendations have strengthened our presentation. We are also pleased to see that our rebuttal positively influenced the reviewer's opinion of our work, leaning towards an acceptance score.	Rebuttal 1: Rebuttal: ## Training with no language annotations We stated several times that the system is trained with vision-only data (Fig. 1, Line 46, Line 469) and no language annotations (Line 11, Line 42). Nonetheless, some reviewers expressed doubts on this matter. We believe the source of confusion is the statement at Line 176-177 ("We have removed the explicit reward information about the task and replaced it with a short task description, in language form."). To improve clarity, we replaced it with: "The datasets contain no reward information and no text annotations of the trajectories. The rewards for training for a given task must be inferred by the agent, i.e. using the cosine similarity between observations and the given prompt or, in the case of GenRL, using our reward formulation (Eq. 3)." How can the system work with language prompts, if it's not trained on language data? The connector learns to map visual embeddings from the pre-trained VLM to latent states of the world model. When learning the connector from visual embeddings $e^{(v)}$, we assume it can generalize to the corresponding language embedding $e^{(l)}$ if the angle $\theta$ between the two embeddings is small enough, see Fig. 11a in attached PDF. This can be expressed as $\cos{\theta} > c$ or $\theta < \arccos{c}$, with $c$ a small positive value [1]. Previous work [1,2] leverages noise during training (of the connector), leading to the situation in Fig. 11b, where $c$ grows larger with the noise. This allows language embeddings to be close enough to their visual counterparts. In our work, we instead learn an aligner network, which maps points surrounding $e^{(v)}$ closer to $e^{(v)}$ (Fig. 11c). This way $c$ is unaltered but the aligner will map $e^{(l)}$ close enough to $e^{(v)}$. Since we use noise to sample points around $e^{(v)}$ the model can be trained using vision-only data (no language annotations). We hope this provides a cleaner explanation, which should replace the previous one (Line 118-127). [1] LAFITE: Towards Language-Free Training for Text-to-Image Generation, Zhou et al [2] Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data, Zhang et al ## Data-free settings After pre-training the MFWM, we claim that is possible to train policies for new tasks in a data-free fashion. How does this differ from the standard GenRL setting and offline zero-shot RL methods? We answer this question in Fig. 12 (attached PDF). Offline RL methods (Fig. 12a), combined with VLMs, can learn to perform tasks zero-shot from new prompts, but they need to sample observations and actions from the dataset for computing rewards and for policy learning. GenRL (Fig. 12b), and potentially other model-based RL methods combined with VLMs, need to sample (sequences of) observations from the dataset, to infer the initial latent states for learning in imagination. Afterwards, rewards can be computed on the imagined latent sequences, enabling policy learning. In addition, for data-free GenRL (Fig. 12c), we sample the initial latent states internally by combining: (i) random samples of the latent space, (ii) randomly sampled embeddings, which are mapped to "actual embeddings" using the aligner, and turned into latent states, by the connector. Thus, policy learning requires no data sampling at all. Finally, following the suggestion of Rev. Yb6H, we renamed this paradigm "Data-free policy learning". To the best of our knowledge, there are no previous works that can learn multiple policies in a data-free fashion (after pre-training of a model) which is the reason why we are unable to provide additional baselines. ## Baselines We added model-based baselines to our main experiments. These baselines follow the recommendation of Rev. Yb6H of learning an "inverse connector" from the world model representation to the VLM representation (GenRL does the opposite). The "inverse connector", given the latent state corresponding to a certain observation, predicts the corresponding embedding. Formally: Inverse connector: $\hat{e}^{(v)}_t = f(s_t, h_t) $ $\mathcal{L_{inv-conn}} = \|\| e^{(v)} - \hat{e}^{(v)}_t \|\|^2_2$ After training the inverse connector, visual embeddings can be inferred from latent states. For policy learning, rewards are computed using the cosine similarity between embeddings inferred from imagined latent states and the prompts' embedding. We call this method WM-CLIP. The inverse connector is implemented as a 4-layer MLP, with a hidden size of 1024. For a fair comparison, we adopt the same world model for WM-CLIP and GenRL. For WM-CLIP we pre-train the additional inverse connector, while for GenRL the connector and aligner. We use one world model for each domain and then train N policies (for N seeds). We have also re-run the main experiments, increasing the number of seeds to 10 for all methods. In Table 5 and Fig. 13 (attached PDF), we observe that WM-CLIP is stronger than the model-free baselines. This allows us to clearly establish the main ingredients that contribute to the stronger performance of GenRL: (i) the video-language model helps in dynamic tasks, (ii) model-based algorithms lead to higher performance, (iii) the connection-alignment system presented outperforms the "inverse" way of connecting the two representations. ## Behaviors from video prompts In Fig. 14 of the attached PDF, we provide behavior learning results from video prompts. The videos are also on the project website. The tasks included are static and dynamic, across 4 different domains. The results show a similar trend to the language prompts experiments and the performance of using video prompts is aligned to language prompts, for the same tasks. In general, we found it interesting that the VLM allows us to generalize to very different visual styles (drawings, realistic, AI-generated), very different camera viewpoints (quadruped, microwave), and different morphologies (cheetah tasks). Pdf: /pdf/9214fd2359199839b4fe692a2083d81c14db3756.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval	Accept (poster)	Summary: This paper introduces the problem of Universal Unsupervised Cross-Domain Retrieval (U2CDR) and proposes a two-stage semantic feature learning framework to address it. The framework includes a cross-domain unified prototypical structure established through an instance-prototype-mixed contrastive loss and a semantic-enhanced loss in the first stage, and a modified adversarial training mechanism to ensure minimal changes during domain alignment in the second stage. Extensive experiments demonstrate that this approach significantly outperforms existing state-of-the-art CDR methods in solving U2CDR challenges. Strengths: 1. This paper addresses a new problem, namely Universal Unsupervised Cross-Domain Retrieval, and proposes an initial solution. 2. The paper first formulates the problem and then introduces the proposed method in a hierarchical manner, which is clear and well-structured. 3. The ability to perform U2CDR has broad implications for various applications, such as image search, product recommendations, and artistic creation. Weaknesses: 1. The main effort of the paper seems to be on designing an optimization method. However, the optimization methods involved appear to be mostly existing ones. The authors should enhance the description of the novelty. 2. Although the paper uses $L_{SPR}$ to maintain the semantic structure within domains, how to maintain the relationship between the positive pairs across domains should be emphasized. 3. The analysis related to the Ablation Study seems insufficient. It would be beneficial to analyze the reasons for the experimental results in Table 4. Technical Quality: 3 Clarity: 3 Questions for Authors: While this paper introduces a new problem, where exactly is the novelty in the methodology section? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No limitations or negative impacts have been identified in this paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Description of the methodology novelty Please refer to the novelty illustration in the Global Response. In addition to the ablation study, To further validate the novelty of semantic structure preservation and cross-domain matching, we carry out experiments with the replacement of another state-of-the-art binary graph-based semantic preservation approach (SPPA [1]) and nearest neighbor searching algorithm (CD^2NN [2]). SPPA preserves the instance-level cosine similarity within each cluster and the prototype-level Euclidean distance across clusters. CD^2NN determines cross-domain nearest instance pair by seeking neighboring instance consistency. The experiment results in open-set CDR (in each cell, the left number is the Shared-set mAP@All, while the right is the Open-set detection accuracy) are shown below, demonstrating that UEM's semantic structure preservation and cross-domain matching are more effective. \| \| Office-31 \| Office-Home \| DomainNet \| \|---------\|------------\|-------------\|------------\| \| UEM w/ SPPA \| 76.0, 88.9\| 47.2, 84.4 \| 31.5, 75.7 \| \| UEM w/ CD^2NN \| 76.4, 89.4 \| 46.5, 85.0 \| 32.1, 76.4 \| \| UEM \| 77.4, 92.5 \| 50.2, 86.7 \| 34.8, 80.9 \| [1] Lin, et al. Continual semantic segmentation via structure preserving and projected feature alignment. ECCV 2022 [2] Liu, et al. Promoting Semantic Connectivity: Dual Nearest Neighbors Contrastive Learning for Unsupervised Domain Generalization. CVPR 2023 > How does UEM maintain the relationship between positive pairs across domains Thank you for your suggestion. In fact, cross-domain positive pairs are maintained through the construction of that unified prototypical structure. During the first stage of our UEM framework, all instances in each domain approach their nearest prototype. These prototypes undergo a prototype conversion process, making them cross-domain unified, meaning that prototypes of the same category have the same geometric relationship relative to other categories within each domain. In this scenario, cross-domain positive pairs will be close to the prototype of the same category in their respective domains. The relationship of cross-domain positive pairs is preserved during the second stage of UEM. As the prototypical structures across domains are unified, the inner-domain relationship preservation achieved by our semantic-preserving domain alignment actually equals to preserving inter-domain relationships, which definitely includes the relationship of cross-domain positive pairs. Besides, the relationship of cross-domain positive pairs is further reinforced in SN^2M, which is a primary objective of cross-domain matching. > More analysis of the ablation study Thanks for your suggestion. We provide more analysis and reasoning for the ablation study as follows, and will incorporate them into the future revision. The performance degradation caused by not using Prototype Merging may be attributed to the geometry distinctness of instance discrimination learning within each domain (Theorem 3.1). The high-level intuition is that the same categories across domains are forced to distinct geometry locations due to distinct contrastive comparisons with different category spaces. In this case, there are plenty of mismatches and misalignments in cross-domain matching. Therefore, the UEM performs poorly without using Prototype Merging. As for the performance loss due to SEL, we think it originated from the unavoidable errors of the arbitrary prototype allocation in the prototype contrastive loss. Different from this single prototype allocation, SEL considers the potential relationship between the instance and all prototypes. This design provides the possibility of error correction and can effectively compensate for the prototype contrastive loss. According to Table 4, there is the largest performance gap between the 'Ours w/o SPDA' approach and the full 'Ours' approach. This indicates the semantic structure destruction during standard domain adversarial training. The semantic structure learned by the first stage is the basis of cross-domain matching, thus causing such a large performance drop. To test the effectiveness of SN^2M, we replace it with the neighbor searching used in UCDIR, which searches for the nearest cross-domain instance and prototype in terms of cosine similarity distance to approach. Apparently, this search strategy has a lot of errors, which is worse if there is a domain gap. By contrast, our SN^2M can measure the reliability of the nearest cross-domain instance and then decide whether to approach it. Moreover, with the prototype translation and merging, the nearest cross-domain prototype is more accurate and reliable. --- Rebuttal Comment 1.1: Comment: Thanks for the author's response. I decide to raise my score to Accept. --- Reply to Comment 1.1.1: Title: Thanks to Reviewer ssYY Comment: Thank you for your positive feedback and insightful suggestions. We appreciate your recognition of our efforts to address your concerns and your rating arising. We will be also glad to answer any further questions. Thank you once again for your time and valuable feedback.	Summary: This paper tackles the problem of unsupervised cross-domain retrieval. This is the problem where the query and retrieval domains are distinct. For example, in sketch to real retrieval, the system must retrieve the most relevant real images to a query sketch. "Unsupervised" refers to the fact that no labels are available during training, but the images from both domains are available. The authors claim to be the first to investigate the "universal" version of this problem, where the query and retrieval domains are allowed to have disjoint labels spaces. For this problem, the authors propose a two-stage optimization procedure. In the first stage, three losses are used: (1) an instance-wise contrastive loss (2) a cluster-wise contrastive loss and (3) a semantic enhanced loss. In the second stage, the embeddings between domains are aligned with three losses: (1) an adversarial domain alignment loss (2) a contrastive loss and (3) a nearest neighbor matching loss. Strengths: (1) The method is theoretically motivated. (2) The paper follows a logical orders. (3) Experiments appear to be complete. Weaknesses: (1) The method is clearly described and seems to be theoretically motivated. However, it is hard to understand intuitively why each loss is necessary. In particular, why we must use six different versions of the contrastive loss across two stages? (IPM, INCE, PNCE, SEL, SPR, SN2M). The theory only seems to justify the IPM loss. (2) In my opinion, even for someone well versed in metric learning, this method is hard to grasp. Some examples: - In line 148, the method applies k-means with a variable number of clusters determined by the "Elbow approach" and a contrastive loss on top of the cluster centroids. Just this one paragraph requires the person implementing the algorithm to reference another paper and implement a clustering algorithm. - The argument, starting at line 152, explaining the IPM loss is hard to understand, mostly because of the unusual notation (arrows and xor symbols). - The argument for the SN2M loss, starting at line 235 is unclear to me. (3) Overall, the method reads like a series of steps that do not follow one central motivation. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) Why do we need two stages of training? Is it really necessary to have two completely different sets of novel loss functions in each stage? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > The necessity illustration of each loss and the theory justification Firstly, we did not use six versions of contrastive loss. IPM combines INCE and PNCE with the intuitive goal of performing categorical semantic learning on unlabeled domain data. INCE forms the basis of unsupervised semantic learning. However, relying solely on INCE results in unclear boundaries between categories. Fortunately, introducing PNCE can enhance the distinction between categories, but determining the optimal timing to introduce PNCE requires careful consideration. We use a sigmoid function to weight the INCE training and then introduce varying degrees of PNCE accordingly. Additionally, we design a prototype conversion strategy to provide a cross-domain unified prototypical structure for PNCE. However, this IPM suffers from single prototype allocation mistakes and errors. To compensate for this issue, we propose SEL, which considers the relationship between a single instance and all prototypes, thereby alleviating erroneous optimization caused by single prototype allocation mistakes. IPM and SEL constitute the first stage of UEM optimization. In the second stage of UEM, we recognize that the primary reason for the inaccuracy of existing cross-domain matching work is the neglect of the domain gap. However, regular domain alignment methods are not suitable for our UEM framework as they severely destroy the prototypical structure learned in the first stage. Therefore, we modified standard domain adversarial training by introducing semantic preservation constraints (SPR). We also design a more accurate cross-domain matching algorithm that aligns with the unified prototypical structure, which is SN^2M. SN^2M assesses the reliability of cross-domain pairs by the consistency of the nearest prototypes in their respective domains and then decides whether to treat the pair as a positive pair in contrastive learning. Accordingly, the second stage of UEM optimization is guided by SPR and SN^2M. In our current work, we have only conducted theoretical analysis on IPM. Due to the high complexity of entanglement between feature geometry distribution and semantics in high-dimensional space, theoretical analysis of SPR and SN^2M is challenging in supervised learning. Considering unsupervised cross-domain scenarios makes it even more difficult. Thus, we will attempt to provide theoretical explanations for this part in future work. > Detailed description of clustering and Elbow approach Thanks for your suggestion. We will incorporate the following description of clustering and the Elbow approach into our future revision. The Elbow approach requires a pre-set maximum cluster number, which is 100 in our implementation. Then we repeatedly apply K-Means to the memory bank of each domain with the cluster number increasing from 2 to 100. For each run, we record two metrics. One is the within-cluster sum of squares (WCSS) which measures the sum of squared distances between each data point and its assigned centroid, reflecting the compactness of the clusters. The other is the silhouette score (SS) which measures how similar a data point is to its own cluster compared to other clusters. Then we draw two curves with the cluster number as the x-axis and the metric value as the y-axis. To determine the elbow point, we select the farthest point below the line passing through the first point and the last point of the curve w.r.t Euclidean distance as the estimated cluster number. After obtaining the respective estimated cluster numbers from the WCSS and SS curves, we select the larger one as the final estimation. > More explanation of the IPM Please refer to the explanation and description of cross-domain prototype conversion in the Global Response. > More explanation of the SN2M Please refer to the explanation and description of cross-domain instance matching in the Global Response. > The central motivation of the two-stage training As mentioned in the introduction, the successful achievement of cross-domain retrieval (CDR) relies on solving two problems: 1) effectively distinguishing data samples in each domain, and 2) achieving alignment across domains for samples of the same category. Most existing methods adopt a two-stage processing strategy, where self-supervised learning (SSL) is first used for categorical semantic learning on unlabeled data, followed by nearest neighbor searching for cross-domain categorical matching. Our UEM framework also employs this two-stage strategy and introduces several novel designs and algorithms during the training phases to address the U^2CDR problem. We considered whether it might be possible to solve the U^2CDR problem with a single-stage, end-to-end framework. The answer is no because we cannot introduce additional constraints during the first stage of SSL, such as domain alignment or cross-domain categorical matching. These constraints would significantly affect the original optimization goals and directions of SSL, deviating the model from learning categorical semantics, which is the basis of everything. To validate our thinking, we carry out experiments with a single-stage design by incorporating domain alignment (from the second stage of UEM) and SN^2M into the first stage. Below are the experiment results for close-set, partial, and open-set UCDR (in each cell of open-set UCDR, the left number is Shared-set mAP@All, and the right is Open-set detection accuracy). As shown in the table, the performance of the single-stage process is indeed much poorer compared to the two-stage design. \|\|\|Office-31\|Office-Home\|DomainNet\| \|-\|-\|-\|-\|-\| \|Close-set UCDR\|Single-stage UEM\|64.7\|40.4\|24.9\| \|\|UEM\|81.9\|52.4\|34.4\| \|Partial UCDR\|Single-stage UEM\|45.0\|30.5\|24.4\| \|\|UEM\|63.0\|46.0\|31.3\| \|Open-set UCDR\|Single-stage UEM\|57.7, 65.3\|39.8, 60.9\|25.2, 59.8\| \|\|UEM\|77.4, 92.5\|50.2, 86.7\|34.8, 80.9\| --- Rebuttal 2: Title: A Gentle Reminder of Further Feedback to Reviewer tjhd Comment: Dear Reviewer tjhd, As the rebuttal discussion phase ends soon, we want to express our gratitude for your engagement thus far. We really want to check with you whether our response addresses your concerns during the author-reviewer discussion phase. We have diligently addressed every concern and question you raised during the initial review, and our efforts are aimed at enhancing the clarity and quality of our work. We genuinely hope our responses have resolved your concerns related to the design intuition behind our methodology, the detailed description of the Elbow approach, IPM and SN2M losses, and the central motivation of the two-stage design. Your thoughtful evaluation greatly aids in our paper's refinement and strength. Again, we sincerely appreciate your dedication and time. Best regards, Authors of Paper 5055 --- Rebuttal Comment 2.1: Title: Acknowledgement of Rebuttal Comment: Dear Authors, I read the rebuttal and appreciate the additional experiments and clarifications. --- Reply to Comment 2.1.1: Comment: Dear Reviewer tjhd, Thank you for your positive feedback and insightful suggestions. We appreciate your recognition of our efforts to address your concerns and include more experiments. We are committed to expanding these clarifications and experiments to our next revised version. We value your input and sincerely hope you consider raising your rating based on the improvements we’re implementing. Your endorsement would greatly enhance the credibility of our work. We will be also glad to answer any further questions. Best regards, Authors of Paper 5055	Summary: This paper proposes Universal Unsupervised Cross-Domain Retrieval for the first time and designs a two-stage semantic feature learning framework to address it. Strengths: This paper proposes a new approach in universal unsupervised domain adaptation, with sufficient experiments to verify its motivation. Weaknesses: 1. In unified unsupervised domain adaptation, there is no handling of instances that are not common categories. Isn't this necessary? 2. From the perspective of innovation, the proposed unified prototype structure is interesting, and the rest is mostly incremental work, such as semantic structure preservation and adjacent feature matching in domain adaptation. From the visualization results, the author failed to prove the above contributions. 3. This paper should reflect the difference between universal domain adaptation and unsupervised domain adaptation. 4. This article does not have a better way to state the method, especially in cross-domain prototype conversion and close neighbor matching. Technical Quality: 3 Clarity: 2 Questions for Authors: See Weaknesses section Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: no Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Any handling of instances belonging to uncommon categories In the first stage of our UEM framework, we aim to build a unified prototypical structure across domains via the IPM loss. The IPM loss is a combination of instance and prototype contrastive losses. Given that the IPM loss is computed separately in each domain, we design a novel prototype merging strategy to merge highly potential common categories across domains. Then for the prototype contrastive loss, each instance searches for and tries to approach its closest prototype. There is no difference between instances corresponding to unmerged (uncommon category) or merged prototypes, as all prototypes should be treated equally to form the final prototypical structure -- otherwise, the uncommon category detection in partial and open-set cross-domain retrieval becomes much harder. As for the second stage, each instance of each domain searches for and decides whether to approach the nearest neighbor in the other domain. For instances belonging to merged prototypes (common category), there is a much higher possibility of being optimized to approach their nearest neighbors. While for instances that are not common categories, they usually shouldn't approach their cross-domain neighbors. However, we still allow these instances to approach their domain-translated prototypes, which means that we do have the handling for instances that are not common categories. We believe this is necessary to achieve more effective cross-domain matching. > Novelty of semantic structure preservation and adjacent feature matching Please refer to the Global Response for a detailed description of the novelty. To further validate the novelty of semantic structure preservation and adjacent feature matching, we carry out experiments with the replacement of another state-of-the-art binary graph-based semantic preservation approach (SPPA [1]) and nearest neighbor searching algorithm (CD^2NN [2]). SPPA preserves the instance-level cosine similarity within each cluster and the prototype-level Euclidean distance across clusters. CD^2NN determines cross-domain nearest instance pair by seeking neighboring instance consistency. The experiment results in open-set CDR (in each cell, the left number is the Shared-set mAP@All, while the right is the Open-set detection accuracy) are shown below, which demonstrate that UEM's semantic structure preservation and adjacent feature matching are more effective. \| \| Office-31 \| Office-Home \| DomainNet \| \|---------\|------------\|-------------\|------------\| \| UEM w/ SPPA \| 76.0, 88.9\| 47.2, 84.4 \| 31.5, 75.7 \| \| UEM w/ CD^2NN \| 76.4, 89.4 \| 46.5, 85.0 \| 32.1, 76.4 \| \| UEM \| 77.4, 92.5 \| 50.2, 86.7 \| 34.8, 80.9 \| [1] Lin, et al. Continual semantic segmentation via structure preserving and projected feature alignment. ECCV 2022 [2] Liu, et al. Promoting Semantic Connectivity: Dual Nearest Neighbors Contrastive Learning for Unsupervised Domain Generalization. CVPR 2023 > Difference between universal and unsupervised domain adaptation We suppose you are indicating the difference between universal and unsupervised cross-domain retrieval (CDR) rather than domain adaptation. First, the CDR problem considers two semantically distinct but similar domains and aims to retrieve data samples from one domain belonging to the same category as a query sample from the other domain. The unsupervised CDR specifies that these two domains only contain unlabeled data. Regular unsupervised CDR studies assume that the category spaces of these two domains are identical, while universal CDR focuses on scenarios where the category space across domains is distinct. In our work, we believe the assumption of identical cross-domain category space is unreasonable for many real-world applications, as the categorical composition of an unlabeled data domain is hard to acquire without detailed analysis and dedicated expertise. > Better statement of cross-domain conversion and close neighbor matching Please refer to the Global Response for better descriptions of cross-domain conversion and close neighbor matching. --- Rebuttal 2: Title: A Gentle Reminder of Further Feedback to Reviewer gZ1R Comment: Dear Reviewer gZ1R, The conclusion of the discussion period is closing, and we eagerly await your response. We greatly appreciate your time and effort in reviewing this paper and helping us improve it. Thank you again for the detailed and constructive reviews. We hope our response is able to address your comments related to the handling of uncommon categories, the novelty description of semantic preservation and adjacent feature matching, and clarification of problem settings. We take this as a great opportunity to improve our work and shall be grateful for any additional feedback you could give us. Best Regards, Authors of Paper 5055	null	null	Rebuttal 1: Rebuttal: ## Global Response We would like to thank all the reviewers for their constructive comments and suggestions. In the global response below, we respond to some common questions and present more visualization in the attached PDF. > [For Reviewers gZ1R and ssYY] Methodology novelty + Unified Prototypical Structure: We identified the challenges of applying standard contrastive learning (CL) to U^2CDR, with the most significant being the cross-domain geometry distinctness. To tackle this issue, we innovatively perform cross-domain translation on each domain’s prototypes, followed by a unified merging strategy. This ensures a unified prototypical structure during prototype CL. To our knowledge, no prior work has employed such cross-domain prototype conversion as we have. + Single Prototype Allocation Compensation: In the prototype CL, the single prototype allocation can often be inaccurate, especially in the early stages of training. To compensate for this inaccuracy, we design a semantic-enhanced loss (SEL) that considers the relationship of an instance with all prototypes and directly optimizes the geometric distances between them. We have not seen this specific design in existing research. + Optimal Timing for IPM: Practically, we need instance CL to gradually learn the semantic prototypical framework of a domain, followed by prototypical CL and SEL to further stabilize and refine this framework. Determining the optimal timing to combine these losses is a challenging issue. For example, UCDIR divides the process into three stages, using different weights to combine instance and prototype CL at each stage. Our UEM resolves this issue by utilizing the Sigmoid function for weighting throughout the training process. + Semantic-preserving Domain Alignment: We discovered that the primary reason for the inaccuracy of existing cross-domain matching algorithms is the neglect of the domain gap. However, standard domain alignment methods are not suitable for our UEM framework as they severely destroy the prototypical structure learned in the first stage. Therefore, we modified the standard domain adversarial training by introducing semantic preservation constraints. These constraints utilize cosine similarity and Euclidean distance to associate all instances pairwise. This pairwise association is simple yet effective as any change in an instance affects the cosine similarity and Euclidean distance of all other instances. Although geometric structure preservation has been explored in continual learning using multi-node graphs based on graph neural networks [1], our UEM’s semantic preservation approach is much simpler and unseen before. + Mismatch Filtering in Nearest Neighbor Searching: We found that existing cross-domain retrieval work neglects to handle the inevitable mismatches in nearest neighbor searching, which usually occur at cluster boundaries. Thus, we designed SN^2M to evaluate the consistency of cross-domain neighboring instance pairs relative to their prototypes in each domain and determine whether to optimize the instance pair to be closer. Experimental results show that our SN^2M is highly accurate and perfectly aligns with the previously mentioned prototype translation and merging, distinguishing it completely from existing cross-domain matching algorithms. [1] Yu, Da, et al. Contrastive correlation preserving replay for online continual learning. IEEE TCSVT 2023 > [For Reviewers gZ1R and tjhd] More explanation about cross-domain prototype conversion and cross-domain instance matching We elaborate on these two aspects below and will incorporate them in the revision of the paper. To simplify the explanation, we avoid using mathematical symbols here but will do so in paper revision. + Cross-Domain Prototype Conversion: We design the cross-domain prototype conversion strategy to unify the prototypical structures of different domains. This strategy involves the following four steps: + As an example, for prototype conversion of domain A, we first translate all prototypes of domain B along the vector connecting the centers of the two domains to domain A. + Next, all prototypes in domain A use the Hungarian algorithm to find the nearest translated domain B prototypes. + For each prototype pair determined by the Hungarian algorithm, we check if they satisfy the merging condition (Eq.9 in the paper). If they do, the prototypes are merged by averaging; if not, they remain unchanged. + Finally, the prototype set for domain A is composed of the merged prototypes, the unmerged original domain A prototypes, and the unmerged translated domain B prototypes. We visualize this prototype conversion as Figure 1 in the attached PDF. + Cross-Domain Matching: Our SN^2M can assess the reliability of cross-domain instance pairs and then decide whether to include them as positive pairs in contrastive learning. The specific steps are as follows: + The prototypes of domains A and B are transformed according to the previously described prototype conversion strategy. + For an instance $x^A$ in domain A, we find its nearest domain A prototype $p_{x^A}^A$ based on the product of cosine similarity and Euclidean distance. We also find $x^A$'s nearest domain B instance $x_{x^A}^B$ using the same criteria (refer to Eq.15 and Eq.16 in the paper). + For the nearest domain B instance $x_{x^A}^B$, we then find its nearest domain B prototype $p^B_{x_{x^A}^B}$. + If $p^B_{x_{x^A}^B}$ matches $p_{x^A}^A$ across domains (i.e., if we translate $p_{x^A}^A$ from domain A to B, the translated $\widetilde{p_{x^A}^A}$ is the same as $p^B_{x_{x^A}^B}$), we consider $x^A$ and $x_{x^A}^B$ as a positive pair in contrastive learning (Eq.17). Otherwise, we only consider $x^A$ and $\widetilde{p_{x^A}^A}$ as a positive pair in contrastive learning. We visualize the SN^2M process as Figure 2 in the attached PDF. Pdf: /pdf/61c1926ce0298683f8a3f596f98d375fd1128650.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Neural Conditional Probability for Uncertainty Quantification	Accept (poster)	Summary: This paper proposes Neural Conditional Probability (NCP), a novel operator-theoretic approach for learning conditional probability distributions. Extensive theoretical results are provided to support the optimization consistency and statistical accuracy of NCP. NCP can be used to extract conditional density and compute statistical measures such as conditional mean, variance, moments and CDF once it is trained. Experiments on a collection of conditional density estimation datasets are conducted to highlight the efficacy of NCP. Strengths: - This paper is mathematically solid and well-organized. - This paper focuses on a fundamental problem of learning conditional distribution in statistical learning and introduces an effective and simplistic approach that outperforms baselines with more complex architectures. Weaknesses: - The proposed NCP method is not clearly motivated or introduced. In Line 49-50, the authors mention that NCP does not belong to any of the four aforementioned approaches. But how is NCP in contrast with them and in what aspects does NCP make improvements? I believe adding some intuitive explanations accompanying theoretical analysis would help improve the readability. - Some key concepts or methods are not clearly explained, which makes it hard to understand the contributions of this work. For example, why is learning conditional expectation operator considered useful? Are there any baseline methods that also learn expectation operators? Technical Quality: 3 Clarity: 3 Questions for Authors: Please see Weaknesses. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's insightful evaluation and valuable comments. In what follows, we aim to address the highlighted weaknesses and respond to the reviewer's questions. ## Weaknesses - __W1:__ We thank the reviewer for this comment. __We have added a detailed comparison of the NCP method to the main existing strategies in the global response__. In brief, while the operator approach is fundamentally different, it can be viewed as a significant improvement over direct learning strategies that rely on a pre-specified dictionary of functions. The NCP method, in contrast, learns the latent space representation (dictionary) that is best adapted to the data under investigation. - __W2:__ The conditional operator approach has been the core idea in the ML theory of kernel mean embeddings and conditional mean embeddings [A], and, more recently, dynamical systems [B]. There, the operator is estimated on a universal (infinite-dimensional) reproducing kernel Hilbert space, in order to transfer probability distributions of one marginal to another by encoding them into the mean of the canonical feature map. While this theory is rich, it is __essentially limited to infinite-dimensional feature spaces__, and, hence, suffers from scalability limitations and the need for experts to design specific kernels for application at hand. __How to build finite-dimensional data representations and inference models__ based on them, to the best of our knowledge, was __an open problem__. So, we hope that our work on the operator approach to inference of conditional probabilities opens a new line of research into this promising methodology that offers several significant benefits. From a probabilistic perspective, our approach allows us to learn the true joint distribution of $(X,Y)$ without needing to specify a model, even when the relationship between $X$ and $Y$ is complex. This capability provides valuable insights into $X$ and $Y$. For instance, the independence of $X$ and $Y$ is conveniently captured by the nullity of the conditional expectation operator. Moreover, the conditional expectation operator enables us to derive essential objects such as the conditional PDF, conditional CDF, conditional quantiles, and moments, even in complex nonlinear scenarios, without prior knowledge about $(X,Y)$. From a computational standpoint, we utilize the spectral properties of compact operators along with the recent NCP self-supervised loss function to create a theoretically sound training loss. This loss function guarantees that the true target operator is the unique global minimizer (see Theorem 1), while being inexpensive to evaluate, facilitating fast and stable training. ### References: [A] Krikamol Muandet, Kenji Fukumizu, Bharath Sriperumbudur and Bernhard Schölkopf (2017), "Kernel Mean Embedding of Distributions: A Review and Beyond", Foundations and Trends® in Machine Learning: Vol. 10: No. 1-2, pp 1-141 [B] Kostic, V., Novelli, P., Maurer, A., Ciliberto, C., Rosasco, L., and Pontil, M. (2022). Learning dynamical systems via Koopman operator regression in reproducing kernel Hilbert spaces. In Advances in Neural Information Processing Systems. --- Rebuttal 2: Comment: Thank you for the detailed response. I am happy to keep my score. --- Rebuttal Comment 2.1: Title: Acknowledgement to the reviewer Comment: We would like to thank the reviewer for their comments, which inspired us to make additional steps and improve our work. We are happy that our rebuttal was helpful, and we commit to incorporate it in the revised manuscript.	Summary: I am not qualified to review this paper Strengths: I am not qualified to review this paper Weaknesses: I am not qualified to review this paper Technical Quality: 3 Clarity: 3 Questions for Authors: I am not qualified to review this paper Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: I am not qualified to review this paper Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We outline our contributions and offer further context in the global response, hoping these additions will better highlight the value of our work on the inference of conditional probability and uncertainty quantification.	Summary: The authors propose a method (Neural Conditional Probability, NCP) for learning a conditional distribution P(Y \| X) from a finite sample from a distribution. The method is based on following observations: (1) it is sufficient to learn the conditional expectation operator E_{Y \| X}[f](x) = E[f(Y) \| X = x]; (2) the conditional expectation operator can be written as (an infinite) SVD decomposition which could be truncated at some point, so the problem reduced to learning the finite number of functions in the SVD decomposition; (3) the joint distribution density can be written using the functions from the SVD decomposition of the conditional expectation operator, which gives an optimisation objective for fitting the functions from the SVD decomposition using the sample from a joint distribution. The authors provide an extensive theoretical analysis of the proposed method as well as a simulation study on a few synthetic datasets. Strengths: + An interesting, novel and theoretically well-motivated method addressing an important problem of conditional distribution estimation + The method uses a fairly simple neural network (MLP) but achieves the competitive to the methods using much more complex architectures + Thorough theoretical analysis on statistical properties of the proposed estimator Weaknesses: - Limited experiments restricted to synthetic data making it difficult to judge the potential applicability of this method - It would be nice to have a short summary on the main properties of operators, their SVD decompositions, etc. I could generally follow the presentation without major problems, but having a such an operators summary would have made it easier to read the paper Technical Quality: 4 Clarity: 3 Questions for Authors: - I am wondering about the choice of the specific loss function in Eq. (6). Could, for example, the log-likelihood potentially be used here? If so, what are the advantages of using the Eq. (6) instead of log-likelihood? - The loss function in Eq. (9) is roughly speaking a regularisation term enforcing that the singular functions are orthonormal, is it correct? Could it be possible to build the neural nets with such properties by construction rather than enforce it by regularisation? - What do you think about the scalability of the model to more complex datasets than in Section 6? For example, conditional image generation. Do you expect issues applying NCP in such cases? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The limitations are sufficiently addressed (NeurIPS Paper Checklist) Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's insightful evaluation and valuable comments. In what follows, we aim to address the highlighted weaknesses and respond to the reviewer's questions. ## Weaknesses - __W1.__ Thank you for this remark. __We added several high-dimensional experiments focused on UQ tasks, while most are synthetic to explore, as requested, scalability of NCP, one is a very complex real-world problem of protein folding__, see also our global response. For the latter, we used a dataset that is considered as challenging for the task of inferring dynamics of the stochastic process governed by the conditional probability kernel. Applying NCP to this setting, we were able to infer this conditional density from learned graph representations of Chignolin mini-protein ensemble. This allowed us to predict the expected average pairwise atomic distances between heavy atoms and infer its 10th and 90th percentiles. As reported in Fig. 3, folded/unfolded protein corresponds to small/large coverage that originates from less/more variations of the ensemble. To the best of our knowledge, this is the first work that was able to perform UQ to infer metastable states of Chignolin, an important contribution in its own right. - __W2.__ We thank the referee for the suggestion. We will add a section in Appendix summarizing the fundamental properties of operator theory that we used to develop the NCP approach and refer to it at the beginning of Section 3. This reminder will include in particular the definition of operators on infinite-dimensional Hilbert spaces and the definition of compact operators. Finally we will state the Erchart-Young-Mirsky theorem which guarantees the existence of the SVD for compact operators. ## Questions - __Q1:__ We thank the referee for the interesting question. Maximizing the log-likelihood function to find model parameters is a central method in statistics, offering strong theoretical guarantees for models satisfying proper regularity conditions. However, this method becomes computationally complex for deep learning models, as the log-likelihood function can be non-concave with multiple local maxima, making it challenging to find the global maximum. In contrast, our approach represents the joint distribution through an operator and uses the regularized NCP loss designed according to the best low-rank rank approximation for operators. This method has a strong theoretical guarantee, as proven in our Theorem 1, that the unique global minimizer of this loss is the true target operator. Additionally, the empirical NCP loss is smooth and can be evaluated with linear complexity in batch size and latent space dimension, making it compatible with standard deep learning training methods. - __Q2:__ Another great question. We've been investigating an alternative to regularization that leverages the Singular Value Decomposition (SVD) properties of compact operators to encode orthogonality directly into the model architecture, rather than enforcing it through regularization. This approach is still under investigation, and it is not yet clear if it will surpass our current regularization method. Indeed, our current regularization approach offers two main advantages: - __Ease of Computation.__ The regularization term is straightforward to compute and integrates seamlessly into the loss function, ensuring scalability and full compatibility with mini-batching and existing optimization methods. - __Theoretical Guarantees.__ As established in Theorem 1, the regularized NCP loss uniquely identifies the true operator, a guarantee that an alternative method encoding orthogonality directly into the model might fail to have. Namely, the orthogonality is defined via the true distributions and not the observed empirical ones. Thus, currently it is not clear how to properly encode it in the architecture and obtain the same level of theoretical guarantees. - __Q3:__ We added additional experiments both on synthetic and real high-dimensional data to illustrate that the NCP operator approach scales without any issue to high-dimensional, more complex data. Please see our global response for more details. --- Rebuttal Comment 1.1: Comment: Thank you very much for a detailed reply! I confirm my positive view of this paper. --- Reply to Comment 1.1.1: Title: Acknowledgement to the reviewer Comment: We would like to thank the reviewer for their comments, which inspired us to make additional steps and improve our work. We are happy that our rebuttal was helpful, and we commit to incorporate it in the revised manuscript.	Summary: The paper proposes Neural Conditional Probability, a novel operator-theoretic approach to learning conditional probability distributions by learning parameters of the truncated SVD of the conditional expectation operator with a neural network. The authors provide a rigorous mathematical derivation and argue for statistical guarantees of their method. The empirical evaluations require major improvements to an otherwise solid paper. As a general note: I do not consider myself an expert on the theoretical aspects of learning theory. My background is in Bayesian deep learning and simulation-based Bayesian inference. As such, my confidence regarding sections 3 and 5 is rather low, and my review shall mainly consult on the remaining sections that focus on presentation, embedding into other literature, and empirical evaluations. Strengths: - The introduction is excellent, with a high degree of accessibility for the broader NeurIPS community and sound motivation of the proposed method. - The method seems mathematically rigorous, well-motivated, and sound. - The authors compare their method against a high number of competing algorithms in the numerical experiments. Weaknesses: ## Major - The Related Work section does a good job of acknowledging related works that aim to learn conditional distributions. However, it utterly fails to embed the current paper into this research landscape. I recommend the authors elaborate on the precise similarities and differences between the referenced papers and their methods in the rebuttal period. - The empirical evaluations are limited to low-dimensional toy problems. This is a stark contrast to the introduction of the method, where the authors repeatedly list the curse of dimensionality as a drawback of other methods. While I acknowledge that the paper is situated in the area of operator learning and ML theory, the quality standard of NeurIPS is not met by the authors’ experiments. This weak evaluation does not do the remainder of the paper justice and I strongly recommend the authors overhaul the experiments to feature high-dimensional tasks that cannot be solved with other state-of-the-art methods. This constitutes a major revision, and this is the main reason why I cannot recommend acceptance to NeurIPS 2024. ## Minor - The empirical evaluation is missing some important information for real-world applications: What are the approximate wall-clock times for (1) training and (2) inference of the competing methods? Further, the authors mention the large required training set size, which might also influence the practically expected training duration in real-world tasks. - Please fix the citations throughout the manuscript: Most citations are ‘text citations’ even if their embedding in the sentence warrants parenthesized citations (Author, 1976). - This is just a personal preference, no need to address it: The ‘paper organization’ paragraph at the end of the introduction does not add value and the space could be used more efficiently elsewhere in the manuscript. - The first sentence in the conclusion is incomplete. Technical Quality: 3 Clarity: 3 Questions for Authors: - As per your answer to checklist item 5 (Open access to data and code), I would like to request access to the full and reproducible code of the empirical evaluations. - Since you want to compare your method with state-of-the-art conditional density estimation methods: Why don't you benchmark against conditional flow matching? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: - The authors mention limitations throughout the manuscript, which I appreciate. However, I would recommend adding a dedicated Limitations section in the conclusion to give a compact overview for readers who don’t engage with the entire paper in detail. - The performance of NCP in high dimensions might be a limitation, but the authors do not study this crucial setting. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed and thoughtful feedback on our submission. We appreciate your recognition of our contribution to the field of operator learning and ML theory which was the primary objective of our work. We would like to address your concerns regarding the empirical evaluations and provide additional context. ## On general note Thank you very much for giving us the perspective to understand your review. While we followed all your suggestions for the revision, we would like to stress that theoretical guarantees developed in Sec. 3 and 5 are the integral part of this paper, and the heart of key contributions. Without being too technical, let us just briefly summarize what we feel as the most important aspects: - NCP is based on __deep representation learning__ by training two DNNs (architecture depending on the data modalities) via a new loss so that the learned representations (u’s and v’s) can lead to __inference of diverse statistics derived from the conditional probability__. - For each of the considered statistics we __prove finite sample generalisation bounds in high probability__ that rely on the __concentration inequalities__ and __optimality gap from DNN training__. - To the best of our knowledge, this is __the first work__ to show how to __provably infer__ conditional probability distribution __using finitely many deep features__, so that the error bounds depend only on the effective problem dimension and __not the ambient dimensions__. We hope that this summary brings more light to the quality and impact of our contributions. ## Weaknesses ### Major 1. We thank the reviewer for raising this point. __We added a discussion in the global response to situate our NCP approach in comparison to the four known strategies__. For the reviewer's convenience, we focus here on a detailed comparison with the best contenders to our NCP method, as shown in Tab. 1: Normalizing Flows (NF) and FlexCode (FC) from Izbicki and Lee (2017). FlexCode (FC) is an example of the direct learning strategy for conditional density estimation (CDE). Our results in Tab. 1 demonstrate that NCP, which learns an intrinsic representation adapted to the data, consistently outperforms FlexCode, which relies on a standard dictionary of functions, in all but one example. In the exception, NCP ties with FC. This also exemplifies that FC is not efficient for capturing the data's intrinsic structure when the FC model is overly misspecified. Regarding Normalizing Flows (NF), which is used in a conditional training strategy, a major limitation is the need to retrain an NF model for each conditioning. Additionally, NF builds a density model over the whole ambient space, which may not be efficient for capturing the data's intrinsic structure. Finally, to the best of our knowledge, NF lacks theoretical guarantees when applied to UQ tasks where strong guarantees of reliability are required. For more details on this aspect for the NCP, see our response to the general note above. We will incorporate this discussion in a revised version of our paper. 2. Thank you for this remark. __We added high-dimensional experiments focused on UQ tasks. While most are synthetic in order to explore, as requested, the scalability of NCP, one is a complex real-world problem of protein folding__, see also our global response. For the latter, we used a dataset that is considered as challenging for the task of inferring dynamics of the stochastic process governed by the conditional probability kernel. Applying NCP to this setting, we were able to infer this conditional density from learned GNN representations of Chignolin mini-protein ensemble. This allowed us to predict the expected average pairwise atomic distances between heavy atoms and infer its 10th and 90th percentiles. As reported in Fig. 3, folded/unfolded protein corresponds to small/large coverage that originates from less/more variations of the ensemble. To the best of our knowledge, this is the first work that was able to perform UQ to infer metastable states of Chignolin, an important contribution in its own right. ### Minor - We added an experiment on how time complexity and inference quality scales w.r.t. dimensionality of $X$. In the attached pdf we included Fig. 2 presenting compute times for training and statistical accuracy for conditional density estimation (CDE) using the KS-distance. - Thank you a nice suggestion, we will follow it. ## Questions 1. As requested, __we provided the AC with a link to an anonymous repository__, as outlined in the NeurIPS guidelines, that includes notebooks for reproducing all our experiments. 2. We thank the referee for bringing this method to our attention. From our understanding, conditional flow matching is designed to efficiently train continuous normalizing flows in high-dimensional settings, particularly for image generation tasks. However, based on our review of the relevant literature, this method has not been explored for the purpose of UQ. In our benchmark, we used a discrete-time masked autoregressive normalizing flow, which performs well on our synthetic benchmark due to its universal approximation property [Papamakarios et al. Normalizing Flows for Probabilistic Modeling and Inference. JMLR 2021, 22(57):1−64]. We did not encounter any issues with training convergence for this model. Additionally, our primary objective is to provide a simple yet effective approach for the UQ tasks we investigated. For these reasons, we believe that incorporating this more complex model in our benchmark lies outside the scope of our study. ## Limitations As elaborated above, we added additional experiments both on synthetic and real data to __show that the NCP operator approach doesn't suffer from this limitation__. Indeed it scales without any issue to high-dimensional and more complex data, without degradation of inference quality. Lastly, as requested, we will collect all discussions regarding limitations in one paragraph. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal which addresses my main concerns regarding the empirical evaluations. I acknowledge that the theoretical aspects of the paper are the main contribution. I will raise my score accordingly. ## Flow matching and UQ > From our understanding, conditional flow matching is designed to efficiently train continuous normalizing flows [...] However, based on our review of the relevant literature, this method has not been explored for the purpose of UQ. Flow matching has been studied for simulation-based inference, which is concerned with the very question of uncertainty quantification. See reference [1] which also contains a code repository. [1] Dax et al. (2023). Flow Matching for Scalable Simulation-Based Inference. NeurIPS 2023. https://arxiv.org/abs/2305.17161 --- Reply to Comment 1.1.1: Title: Acknowledgement to the reviewer Comment: We would like to thank the reviewer for their comments, which inspired us to make additional steps and improve our work. We are happy that our rebuttal was helpful, and we commit to incorporate it in the revised manuscript. Concerning the reference, we thank reviewer for bringing this paper to our attention, we will definitely include it in the revision.	Rebuttal 1: Rebuttal: We thank all reviewers for their insightful evaluation of our paper. We appreciate all their comments and remarks, which we will incorporate in our revision. Before addressing each review in detail, we would like to point out some general remarks that apply to all of them. ## Positioning Our main focus is on advancing Uncertainty Quantification (UQ) to handle nonlinear structures effectively. Several high-stakes nonlinear conditional probability estimation problems, such as CVAR for financial regulators (Since the 2008 financial crisis, regulators have required banks to limit their risk exposure, necessitating accurate computation of conditional value at risk, which relates to the conditional CDF), meaningful uncertainty quantification for engineering design (e.g., to drastically reduce the number of expensive real-world prototypes and tests in favour of a more computationally driven process), and rare-events prediction for predictive maintenance, are typically formulated in state spaces of moderate to high dimensions and often involve large datasets with complex, nonlinear structures, such as data residing on manifolds or graph data. Our goal is to showcase that the NCP method can effectively handle such structured data, offering a flexible and scalable modelling general solution. Importantly, we provide solid and provable statistical guarantees, aiming to advance safe and secure AI applications. ## Contributions NCP leverages operator theory, a rich and powerful area of mathematics, allowing us to gain deep insights into the structure of functional spaces including probability distribution spaces. To the best of our knowledge, the NCP approach integrating operator theory in a self-supervised learning framework represents deeply innovative idea with significant potential for a range of ambitious applications. In the current paper, we focus on presenting NCP in the simplest possible way to illustrate its core benefits. Strengths of NCP 1. Versatility: NCP combined with a simple and compact MLP architecture, consistently outperforms more complex models and, when it does not, reliably matches their performance. 2. Adaptation to Intrinsic Dimension: NCP effectively adapts to the intrinsic dimension of the data, constructing a well-suited low-dimensional representation space without requiring prior knowledge. This results in a more efficient handling of the data's inherent structure. 3. Efficiency in Training: NCP involves only a single training phase for the joint density. Once trained, we can easily derive all conditional probabilities without additional retraining, streamlining the process. In addition, NCP is easy and fast to train. Indeed we prove that it also scales linearly with latent space dimension and sample size. Our new experiments confirm that NCP scales with the ambient dimension, with compute time increasing at most sublinearly with the dimension (See Fig 2). Additionally, our method also feature the following theoretical and conceptual strengths: 4. Complete Theory: Our approach is supported by a robust theoretical framework, providing strong guarantees for performance and reliability in both training and UQ accuracy. 5. Assumption Comparison: The density assumption with respect to marginals for NCP is significantly weaker compared to many other methods like Normalizing Flows (NF), which usually require density with respect to Lebesgue measure. This makes our approach broadly applicable, for instance for discrete distributions, or complex structured data like manifolds. ## Situating NCP method in the landscape of CDE methods While NCP does not strictly fall into any of the four categories described by Gao and Hastie (2022), it is perhaps closest to the direct learning strategy but represents a significant improvement. Unlike the direct learning strategy, which uses a pre-specified kernel or a dictionary of functions to approximate the target conditional density, NCP directly learns the joint density by finding a low-dimensional latent space representation that efficiently adapts to the data. This is illustrated in our experiments where NCP demonstrated its versatility in handling diverse data including manifolds, graphs, continuous and discrete distributions (see Fig. 1, 2 and 3 in the attached pdf). Moreover, by learning proper representation adapted to the data, NCP can capture the intrinsic dimension of the data, which is reflected in our learning guarantees that depend on the latent space dimension rather than the ambient space dimension, even in the absence of prior knowledge. This theoretical foundation and added experiments (see our next point) demonstrate that NCP is capable of handling high-dimensional data. ## Challenging Experiments As requested, we conducted additional experiments to further demonstrate the robustness and versatility of our approach across different dimensionalities and data structures. We have reported the results in the rebuttal pdf file and provided the AC with a link to an anonymous repository containing all code necessary to rerun our experiments. We conducted high-dimensional experiments on synthetic data with discrete and continuous distributions scenarios, where the joint distribution is governed by a hidden low-dimensional manifold in a higher-dimensional space. Results in Fig. 1 and 2 in the attached pdf include KS-distance and computation times across various dimensions and conditional CDF plots for continuous and discrete distributions parameterised by a manifold. Our comparison of NCP and NF shows that NCP performs well and scales effectively in high-dimensional settings. Additionally, an experiment on a real-world challenging application on folding dynamics of a protein is performed. There NCP is combined with a graph neural net architecture to forecast the quantiles of a physically relevant observable and infer the change from folded to unfolded state, see Fig. 3 for details. Pdf: /pdf/07883de53cb588a62b7115faccfd15e599e0afc4.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
FLAME : Factuality-Aware Alignment for Large Language Models	Accept (poster)	Summary: This work studies how to do alignment for large language models to improve their factuality. The focus of this work is on SFT and DPO. The motivation behind this work is a pilot study which shows more factual data does not always lead to a more factual model. To resolve this issue, the proposed Flame framework (1) handles fact-based and non-fact-based examples differently; (2) uses few-shot generated examples from the model itself for fact-based SFT; (3) builds a reward model specifically for factuality (via atomic fact decomposition, retrieval augmented claim verification, etc.) Experiments on multiple datasets demonstrate that Flame can improve the model's factuality without hurting other capabilities (e.g., instruction following). Ablations are also conducted to measure the gain from each individual step. Strengths: 1. The motivation is clear and reasonable. I like using a simple and quick pilot experiment to demonstrate the main motivation of this paper. 2. The idea is straightforward and effective. The high level framework can applied to many different systems. 3. Ablation experiments are conducted to show the gain from each step. The effectiveness for both SFT and DPO are clear. Weaknesses: 1. No external baselines are used in the comparison. It would be great to compare the flame model with other related approaches (e.g., few-shot prompting, sampling multiple responses, and reranking using FactScore or the reward model). I know these approaches are not directly comparable, however, it will still be valuable to understand the relative trends, especially since approaches such as few-shot prompting are used in data generation. 2. It will be great to conduct human evaluations even just on a few examples. 3. The whole pipeline involves a number of components. While many details are presented in the appendix, low-level details like few-shot prompts, and implementation of fact decomposition are omitted. Adding these details will be super valuable for future work to build similar systems. It would be even better if the authors decide to release the code. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In the pilot experiment, doing DPO with FS seems to work reasonably well. Have you tried similar approaches in the real experiments? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are discussed in Sec. A.6 in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Re: No external baselines are used in the comparison. We thank you for the suggestion. As far as we know, the existing best approach to factual alignment is the method introduced by Tian et.al. [1]. Note that, this approach conducts factual alignment directly on the target task (e.g., biography generation) while we conduct factual alignment on more diverse and general instruction following tasks. Although it is hard to conduct a direct comparison with the baseline, as mentioned in line 277, we do provide a baseline which mirrors the approach by Tian et.al. [1], $\mathrm{SFT }+ \mathrm{DPO}^{\textrm{fact}}$ as shown in row 3 of Table 3, where factuality is the only objective of DPO. The main purpose of our paper is not to propose a state-of-the-art approach for factual alignment. Instead, we find the side effect of applying factual alignment to instruction tuning and manage to mitigate the side effect. 2. Re: It will be great to conduct human evaluations even just on a few examples. Thanks for the suggestions. We agree that adding human evaluations would make the result more convincing. We have conducted manual evaluations for two baselines and FLAME on a few representative cases from Alpaca Eval and Biography in case study, as shown in Appendix Figure 10. We will make it clear in the section of case study. 3. Re: Low-level details like few-shot prompts, and implementation of fact decomposition are omitted. Thanks for the suggestions. Below are the low-level details about our few-shot prompts and fact decomposition, which we will add into our revised manuscript. For the detail of few-shot prompts, as mentioned in lines 195--198, for each instruction $x$, we retrieve the top-5 similar instructions ($x_0 \cdots x_4$) along with the corresponding human responses ($\mathrm{Human}(x_0) \cdots \mathrm{Human}(x_4)$) from OASST (our seed data with 3.2K instruction--human response pairs). Then, concatenating them as the few-shot prompt $x_4, \mathrm{Human}(x_4), x_3, \mathrm{Human}(x_3), \cdots, x_0, \mathrm{Human}(x_0), x$ to elicit the response from the pre-trained LLM. For the implementation of fact decomposition, we reuse the code from FActScore to conduct fact decomposition but replace the GPT3.5 with our fine-tuned Llama2 7B model on the public datasets (i.e., RoSE benchmark [2], CLAIMDECOMP [3] and EXPERT QA dataset [4]). Note that the purpose of fine-tuning our Llama2 7B for atomic fact decomposition is to save the cost of creating factuality preference training data. 4. Re: In the pilot experiment, doing DPO with FS seems to work reasonably well. Have you tried similar approaches in the real experiments? DPO with FS can be considered the approach by Tian et.al. [1] and as mentioned in point 1, in our main experiment for instruction tuning, we do provide a baseline as shown in row 3 of Table 3, which is mentioned in line 277. We will make the statement more clear in the revised version. [1] Fine-tuning Language Models for Factuality. https://arxiv.org/abs/2311.08401 [2] https://github.com/Yale-LILY/ROSE [3] https://github.com/jifan-chen/subquestions-for-fact-checking [4] https://github.com/chaitanyamalaviya/ExpertQA --- Rebuttal Comment 1.1: Comment: Thank you for the response! Adding these clarifications and additional details in the revised version will be very helpful. I will keep my original positive rating of 6. --- Reply to Comment 1.1.1: Comment: Thank you for the helpful suggestions. We will revise our manuscript accordingly.	Summary: This paper shows that training on new or unfamiliar knowledge can promote hallucination and that reward functions in standard RL often inadequately capture factuality. The authors propose a factuality-aware alignment method that first identifies instructions as fact-based or non-fact-based. For fact-based instructions, they employ adapted techniques in respective SFT and RL to generate additional training data, thereby reducing the hallucination of the model's responses. Strengths: * The paper conducts a pilot study that highlights the limitations of SFT and RL in capturing factual knowledge. This study provides valuable insights into data selection for LLM alignment training. * The proposed dual-stage factuality-aware method improves factuality without compromising the instruction-following capabilities for both SFT and RL stages. Weaknesses: * The proposed strategy to create SFT and DPO training data using the generated responses from the LLM itself is limited to the knowledge learned within the original model. This approach may struggle with instructions that the original model cannot generate factual answers for. * The proposed strategy relies on accurately identifying the instruction type initially, which is limited by the model's ability to correctly classify the instruction type. * In the pilot study, it is unclear whether the $PT$ and $PT^{RAG}$ are evaluated using the same protocol as other methods. If they are, the FS score decreases after both SFT and DPO, which contradicts the claim that "fine-tuning LLMs on their own generations appears to be crucial for factual alignment." * While the results in Table 2 and 3 indicate that eliciting knowledge from the model itself can enhance factuality compared to introducing more factual but unknown knowledge, it does not improve the FS of the $PT$, which achieves a score of 53.1 on the Biography task with just 5-shot demonstrations. * As discussed in Sec 5.5, conducting fact checks and computing factuality rewards solely for fact-based sentences can lead to more factuality errors. Clarification is needed on how FS is calculated for the experiments in Sec 5.2 and 5.3. Technical Quality: 3 Clarity: 4 Questions for Authors: Please refer to the Weaknesses. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: Yes, the authors have discussed the limitations of the metric for evaluating factuality. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Re: This approach may struggle with instructions that the original model cannot generate factual answers for. We thank you for the insight. This is exactly what we found in the paper; that is, it is challenging to teach LLMs to learn new knowledge in the finetuning stage. Inevitably, forcing LLMs to learn new knowledge (e.g., learning from RAG or human written response) may cause more hallucination, which is also found in the concurrent paper [1]; thus, we propose to fine-tune LLMs only on its own generated responses. We agree that it is worth conducting exploring how to inject new knowledge to LLMs during fine-tuning stage without any side effects but this is orthogonal to our work. We will make our claim more clear in our revised manuscript. 2. Re: The proposed strategy relies on accurately identifying the instruction type initially, which is limited by the model's ability to correctly classify the instruction type. Yes, we admit that FLAME does require a module to distinguish the types of instructions. However, instead of treating it as a weakness of our approach, we believe that it is the important finding of our paper. That is, not all instructions require factual responses, and overly optimizing for factuality could lead to degradation in instruction following capability (condition 3 vs 4 in Table 3). Furthermore, the previous study [2] also find that different instructions require different evaluation strategies. Finally, we also demonstrate that the task of classifying the instruction type can be easily done by an instruction following fine-tuned model rather than a specially fine-tuned model. 3. Re: In the pilot study, it is unclear whether the $\mathrm{PT}$ and $\mathrm{PT}^{\text{\tiny RAG}}$ are evaluated using the same protocol as other methods. Yes, in our pilot study $\mathrm{PT}$ and $\mathrm{PT}^{\text{\tiny RAG}}$ are evaluated using the same protocol as other methods. However, this does not mean the results contradict our claim: fine-tuning LLMs on their own generations appears to be crucial for factual alignment. First of all, the responses from $\mathrm{PT}$ and $\mathrm{PT}^{\text{\tiny RAG}}$ are generated using 5-shot demonstrations while the responses from SFT and DPO models (only fine-tuned on 495 instructions for biography generation) are generated with 0 shot. Especially, $\mathrm{PT}^{\text{\tiny RAG}}$ also concatenate retrieved evidence; thus, it is reasonable that $\mathrm{PT}^{\text{\tiny RAG}}$ shows far better factual accuracy. Although SFT (fine-tuned with a few examples) zero-shot generation shows a slight factuality degradation compared to $\mathrm{PT}$ 5-shot prompts, we do see that the DPO model improves over the $\mathrm{PT}$ model in Table 1 (condition 5 vs 1). To be clear, our claim comes from the observation that SFT and DPO fine-tuned with LLM’s own generation versus RAG generation (3 vs 4, 5 vs 6) even though RAG generation is shown to be far more factual. 4. Re: SFT and DPO do not improve models’ factuality over 5-shot prompts. We want to clarify that in our paper, improving LLMs’ factuality does not mean that we can make the instruction following fine-tuned LLMs generate more factual responses than the pre-trained LLMs with few-shot demonstrations. In fact, there are previous papers which find that instruction tuning or alignment may degrade LLMs’ performance on the standard knowledge benchmarks, called alignment tax [3]. The alignment tax can explain why our DPO fine-tuned model (only on biography generation) can improve over $\mathrm{PT}$ (Table1 condition 5 vs 1) but the DPO fine-tuned model (on diverse instruction following tasks) cannot improve over $\mathrm{PT}$ (condition 7 in Table 3 vs condition 0 in Table 2). Again, we want to highlight that the main purpose of the paper is to study why instruction following fine-tuning tends to cause hallucination and how to mitigate the issue rather than addressing the alignment tax. Thanks for the insightful question, we will add further explanation in our revised paper. 5. Re: As discussed in Sec 5.5, conducting fact checks and computing factuality rewards solely for fact-based sentences can lead to more factuality errors. Clarification is needed on how FS is calculated for the experiments in Sec 5.2 and 5.3. All our FActScore evaluation is based on the assumption that all the sentences in the responses are fact-based ones for two reasons. First, as pointed out in our ablation study, excluding the non-fact-based sentences may cause noise. Second, from our eyebrow check, LLMs’ generation for Biography, Alpaca Fact and FAVA datasets are mostly fact-based sentences. Figure 10 in Appendix showcases a few LLMs’ responses for Biography and Alpaca Fact. [1] Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? https://arxiv.org/abs/2405.05904 [2] Branch-Solve-Merge Improves Large Language Model Evaluation and Generation. https://arxiv.org/abs/2310.15123 [3] Training language models to follow instructions with human feedback https://arxiv.org/pdf/2203.02155 --- Rebuttal Comment 1.1: Title: Reply Comment: I thank the authors for their insightful replies to my review. They are indeed helpful. I still have some concerns regarding your evaluation assumption that all the sentences in the responses are fact-based, which might not be entirely fair. The baseline model fine-tuned on high-quality instruction data, as pointed out by [1], might generate more stylistic tokens, such as "Of course." Could fact-checking these sentences decrease the FactScore? The evaluation deserves more rigorous and comprehensive discussion. (Note that Reviewer HVe5 also pointed out the limitation of your evaluation protocol.) Given most of my concerns have been addressed, I would like to raise my score from 4 to 5. [1] [The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning](https://arxiv.org/abs/2312.01552) --- Reply to Comment 1.1.1: Comment: We thank you all your helpful feedback to improve our manuscript. Regarding to your last concern, we thank you for the reference and this observation indeed inspires us to conduct the experiment to filter out non-fact-based sentences in the ablation study. As for evaluation, we have conduct a manual check on those responses for Bio, Alpaca-Fact, and Fava. For Bio and Fava, we observe almost all the generated sentences are fact-based ones. For Bio and Alpaca-Fact, we have some examples in Table 10 and for Fava, here we give two examples from SFT. Our observation is that for those fact-based instructions, LLMs' responses consist mostly of fact-bases sentences. We will revise our manuscript and make our assumption based on the observation for evaluation clear. Case1 What are the Golden Lion movies at the Venice Film Festival whose production countries are in Asia? 1. The 2018 film "Shadow" by Chinese director Zhang Yimou.2. The 2019 film "The Wild Goose Lake" by Chinese director Diao Yinan.3. The 2020 film "Nomadland" by American director Chloé Zhao, which was partially shot in South Korea.4. The 2021 film "The Hand of God" by Italian director Paolo Sorrentino, which was partially shot in Thailand.5. The 2022 film "All the Beauty and the Bloodshed" by American director Laura Poitras, which was partially shot in Japan. Case2 Why is white phosphorous so dangerous? White phosphorus is dangerous because it is highly flammable and can spontaneously ignite in air, releasing toxic fumes. It is also highly reactive and can cause severe burns and tissue damage on contact with skin. In addition, white phosphorus is a strong oxidizing agent and can react violently with other substances, such as organic materials, to cause fires or explosions.	Summary: This paper addresses the issue of factual inaccuracy, or "hallucination," in Large Language Models (LLMs). The authors identify factors that lead to the generation of false facts during supervised fine-tuning (SFT) and reinforcement learning (RL). They propose FLAME, a novel alignment method that incorporates factuality-aware SFT and direct preference optimization (DPO) to guide LLMs towards more factual responses without compromising their ability to follow instructions. Experiments demonstrate FLAME's effectiveness in enhancing factuality while maintaining helpfulness. Strengths: 1. The ablation experiments provides comprehensive insights into the effectiveness of DPO and SFT in mitigating hallucination. 2. The method proposed in this paper attempts to balance instruction following and factuality. It relies on model self-construction data, and does not depend on external proprietary models. Weaknesses: 1. The baselines compared in this work are limited to different settings of SFT and DPO only. The baselines in the paper should at least include the work [1]. This prior work also uses DPO and algorithms, and the only difference seems to be data construction. The paper should compare with this work to demonstrate that its algorithm truly achieves a balance between instruction following and factuality. 2. In addition to the works listed in the related work, there are some works whose methods are somewhat similar to this paper, such as [2] [3], etc. The paper may need to add explanations of the differences between these methods to clarify its own novelty. [1] Fine-tuning Language Models for Factuality. https://arxiv.org/abs/2311.08401 [2] Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation. https://arxiv.org/abs/2402.09267 [3] GRATH: Gradual Self-Truthifying for Large Language Models. https://arxiv.org/pdf/2401.12292 Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In comparison to training, have the authors considered comparing representation editing baseline methods? 2. Could the authors supplement experiments on the TruthfulQA-MC in Table 4 to provide a measure of multi-choice performance? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have addressed some limitations of their work in the Appendix, which is commendable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Re: The baselines compared in this work are limited to different settings of SFT and DPO only. We thank your suggestion to compare with the baseline from Tian et.al. [1]. First of all, we want to clarify that Tian et.al. [1] mainly focus on fine-tuning LLMs on a specific task (e.g. biography generation) while FLAME focuses on more general instruction following tasks, where biography generation is a subtask for us to evaluate factuality. Thus, it is hard to make a fair comparison between the two approaches. Nevertheless, as we mention in Section 5.3 (line 277), we do apply the approach from Tian et.al. [1] to our alignment training as $\mathrm{SFT}^{\textrm{fact}}$ and $\mathrm{DPO}^{\textrm{fact}}$, where factuality is the only optimization objective. This result (row 4 vs 3 in Table 3) supports our claim from related work (lines 84--85) that solely focusing on factual alignment may impact LLMs’ instruction following capability. Furthermore, we also point out the importance of finding those fact-based instructions among the instructions, which is ignored by Tian et.al. [1] if we only conduct factual alignment on biography generation tasks. In other words, FLAME is the extension of Tian et.al. [1] to more general instruction tuning tasks. We will revise the manuscript to clarify that we have done our best to compare with the existing factuality fine-tuning method [1] in our main experiment. 2. Re: The paper may need to add explanations of the differences between these methods to clarify its own novelty. We thank you for the good references and do agree that we need to compare the more recent work in our related work [2][3]. The main contribution of the related work[2][3] focuses on using the LLM itself as the factuality judge to enhance its factuality. Similar to Tian et.al. [1], they mainly focus on fact-based instructions and can be considered as an improved version of FactTuneMC from Tian et.al. [1]; thus, ignoring the impact of factuality alignment on LLMs’ instruction following capability. Furthermore, the proposed factuality alignment approaches from [2][3] can be integrated into FLAME to create more accurate factuality pairs. We will add the comparison and cite the references in our revised version. 3. Re: In comparison to training, have the authors considered comparing representation editing baseline methods? We thank you for the suggestions. Although representation editing baseline approaches (e.g., DoLa [4] and ITI [5]) also show promising results, these approaches rely on a development set of the target task to tune the hyperparameters and focus on a single evaluation metric. As we have shown in our main experiment, focusing only on improving factuality may sacrifice models’ instruction following capability. Furthermore, as shown in Tian et.al. [1], the representation editing approaches are not more effective than fine-tuning for factuality, which is one of our baseline approaches (row 3 in Table 3). 4. Re: Could the authors supplement experiments on the TruthfulQA-MC in Table 4 to provide a measure of multi-choice performance? We thank you for the suggestion. Below is the comparison of TruthfulQA-MC (zero-shot). Note that, we do not observe significant differences between models on TruthfulQA-MC task. This is possibly because we mainly focus on the tasks of long-form response generation while TruthfulQA-MC task is formed by short-form answers. The discrepancy between improving LLMs’ factuality on long-form and short-form generation is also found by the previous work [4]. We will add the experiment and explanation to our revised manuscript. \| TruthfulQA_MC \| MC1 \| MC2 \| MC3 \| \| ------------ \| ----------- \| ------ \| ------ \| \| (0) Chat \| 32.2\| 50.2 \|25.4\| \| (1) $\mathrm{SFT}$ \|30.8\| 45.7\| 23.9\| \| (2) + $\mathrm{DPO}$\|30.5\| 46.0\| 23.4\| \| (3) + $\mathrm{DPO}^{\textrm{fact}}$\|31.8\|46.8\|24.3\| \| (4) + $\mathrm{DPO}^{\textrm{FLAME}}$\| 30.8\| 46.0\|23.6\| \| (5) $\mathrm{SFT}^{\textrm{FLAME}}$ \|29.9\|44.8\|22.5\| \| (6) + $\mathrm{DPO}$ \|31.5\|47.0\|24.0\| \| (7) + $\mathrm{DPO}^{\textrm{FLAME}}$\| 30.5\| 45.4\| 23.1\| Reference: [1] Fine-tuning Language Models for Factuality. https://arxiv.org/abs/2311.08401 [2] Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation. https://arxiv.org/abs/2402.09267 [3] GRATH: Gradual Self-Truthifying for Large Language Models. https://arxiv.org/pdf/2401.12292 [4] DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models. https://arxiv.org/abs/2309.03883 --- Rebuttal Comment 1.1: Title: Reply to the authors Comment: Thank you for your response regarding the additional results. I've read the response in detail and decided to keep my score and judgments. --- Reply to Comment 1.1.1: Comment: We thank you for the discussion and will update the suggested experiments and explanation in our revised manuscript.	Summary: The paper discusses a novel alignment method to enhance the factual accuracy of LLMs. The authors observe that conventional alignment processes, which include SFT and RL, often result in the generation of false facts or 'hallucinations'. To address this, they introduce factuality-aware alignment (FLAME), which includes factuality-aware SFT and RL through direct preference optimization. FLAME identifies factors leading to hallucination and adapts the training process to reduce the generation of false claims. Experiments demonstrate that FLAME guides LLMs to produce more factual responses without compromising their ability to follow instructions. The paper contributes to the field by tackling the issue of maintaining helpfulness while improving the factuality of AI-generated content. Strengths: - Clear and Logical Structure: This paper is well-organized and presents its findings with a logical flow, making it easy to follow. - In-depth Analysis of Hallucination: The paper thoroughly analyzes the factors contributing to hallucination during the SFT and RL phases of language model alignment. It identifies key issues: training on unfamiliar data can reduce factual accuracy, and standard RL reward functions often prioritize longer, more detailed responses, potentially encouraging the model to fabricate information. - Innovative Solution: The proposed FLAME is a novel alignment approach that effectively addresses hallucination without compromising the model's ability to follow instructions. By extending both SFT and RL, FLAME tackles a critical issue in LLMs, ensuring more accurate and reliable information generation. - Comprehensive Evaluation: The paper thoroughly evaluates FLAME's effectiveness in improving both factuality and instruction-following abilities. Experiments demonstrate that models aligned using FLAME achieve significantly higher FactScore compared to standard alignment methods, without sacrificing their helpfulness. Weaknesses: This paper is well-written and makes a valuable contribution to the LLM alignments. I only have several minor concerns as follows: - Model Size and Generalizability: The paper focuses solely on the LLaMA2-70B model. It would be beneficial to investigate whether FLAME's effectiveness extends to smaller models, such as 7B or even smaller, given that the factuality-aware SFT relies on self-supervision through few-shot prompting. - Evaluation Metrics and Human Assessment: While FactScore is a valuable metric, it has limitations. It assumes Wikipedia as the definitive source of truth and may not be suitable for broader domains. Using a more comprehensive metric like Veriscore [1] could provide a more nuanced evaluation (I understand that Veriscore is a recently released method, so this is a suggestion for the future version of this paper). Additionally, incorporating human evaluation would strengthen the analysis. A manual assessment of factuality and helpfulness would provide valuable insights and increase the persuasiveness of the findings. - Multi-faceted Evaluation: The paper primarily focuses on instruction following and factuality. However, other crucial aspects of LLM capabilities, including knowledge, reasoning, and code generation, should also be considered. It would be insightful to evaluate the performance of FLAME-trained models on standard benchmarks like MMLU, GSM8K, and HumanEval to assess potential trade-offs in these areas. Technical Quality: 3 Clarity: 3 Questions for Authors: - While FLAME primarily focuses on DPO, can it also be applied to conventional reinforcement learning from human feedback (RLHF) methods like PPO? - Are there plans to release the code and models trained using FLAME for the research community to replicate your methods? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations and broader impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Re: Model Size and Generalizability. It would be beneficial to investigate whether FLAME's effectiveness extends to smaller models, such as 7B or even smaller, given that the factuality-aware SFT relies on self-supervision through few-shot prompting. We thank you for the helpful suggestion. We believe our finding is also valid for smaller models. In our pilot study, the condition of row 3 in Table 1 is equal to factuality-aware SFT with self-supervision through few-shot prompting. The pilot study on smaller models motivates us to extend our study to more general instruction tuning rather than focusing on biography generation. However, we have to admit that it is more challenging to fine-tune a smaller model (e.g., Llama2 7B) to follow instructions, especially that we only use 3200 SFT samples from OASST dataset. Not to mention applying the self rewarding [1] methodology to improve the models’ instruction following capability. Note that we choose not to rely on the large size and high quality SFT data (e.g., Alpaca cleaned dataset or UltraFeedback) generated from GPT4. 2. Re: Factuality Evaluation Although FActScore is the best automatic factuality evaluation tool we can get access to, we agree and discuss the limitation of FActScore in the Section of Limitation. We thank you for referring to the more recent factuality evaluation tool (e.g., Veriscore). 3. Re: Multi-faceted Evaluation We thank you for the helpful suggestion. Below is the comparison between the baseline and factuality-aware fine-tuned models on MMLU and GSM8K. Their results are very close. We will add the evaluation on standard benchmark in the appendix. \| \| MMLU \| GSM8K \| \| ------------ \| ----------- \| ------ \| \| (2) $\mathrm{SFT} + \mathrm{DPO}$\|69.34\| 59.28\| \| (7) $\mathrm{SFT}^{\textrm{FLAME}} + \mathrm{DPO}^{\textrm{FLAME}}$\| 69.05\| 58.22\| 4. Re: While FLAME primarily focuses on DPO, can it also be applied to conventional reinforcement learning from human feedback (RLHF) methods like PPO? We thank you for the question. Yes, we believe FLAME can be also applied to PPO. As shown in our ablation study (the third row in Table 6), we can combine the rewards of instruction following capability and factuality into a single scalar reward. With the single scalar reward, we can conduct PPO for preference fine-tuning. However, due to the complexity and instability of fine-tuning with PPO, we instead use DPO as a proof-of-concept in our experiment. We will add this discussion in Section 5.5 in the revised manuscript. 5. Re: Are there plans to release the code and models trained using FLAME for the research community to replicate your methods? Thanks for the suggestions. We will consider releasing the code or the scripts for creating the preference training data. --- Rebuttal Comment 1.1: Comment: Thanks for the response. However, I think most of my concerns have not been addressed (generalizability, performance degradation on other abilities, and uncertainty of code/data release). After reading other reviewers' comments, I would like to lower my rating to weak accept. --- Reply to Comment 1.1.1: Comment: We thank you for the helpful feedback and comment. We will try our best to address your concern in the revised version. For performance degradation, we assume the degrade on the benchmark is relatively slight and again since we are focusing on long-form generation and not expect to see performance gain on the benchmark with short-form answers. The discrepancy between improving LLMs’ factuality on long-form and short-form generation is also found by the previous work [1]. [1] DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models. https://arxiv.org/abs/2309.03883	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
InstructG2I: Synthesizing Images from Multimodal Attributed Graphs	Accept (poster)	Summary: The authors propose an approach to enhance image synthesis using multimodal attributed graphs, adopting a strategy to condition image generation via a tokenization scheme on graph structure. Strengths: - The paper studies an intersectional topic: leveraging graph learning techniques for image generation, which is a creative application and an area which deserves more focus. - The authors' use of qualitative examples (e.g. Figure 5 and 6) is commendable and helps articulate visual improvements. Weaknesses: Please see questions and concerns below. My general feeling is the paper is fairly incremental in its introduction of a mechanism to encode graph condition into the conditioning for generation. Many design choices for graph conditioning are not discussed well and the quantitative results for some of these choices are missing which hurts the overall impact of the work. Technical Quality: 3 Clarity: 3 Questions for Authors: - Typos: - Line 18: "graph-structued" - The motivation proposed in lines 28-30 is a little bit confusing, since the scenario the authors discuss here (e.g. virtual artwork creation based on nuanced styles of artists and genres) seems like it could be well-handled by text rather than explicitly using graph structure. - There is limited prior work in multimodal graph learning, as the authors mention. The authors may want to reference and position their work with respect to the recent [1] which offers multiple datasets and focuses on utility of GNN methods for node/link/graph-level tasks rather than generative tasks. - Nit: the notation is a bit awkward compared to conventional graph literature which typically uses $\mathcal{V}, \mathcal{E}, \mathcal{X}, \mathcal{F}$ or something similar to indicate node-set, edge-set, node-features, and edge-features. The authors proposed notation in line 72 seems to define P and D as different sets of images / documents compared to the nodes V, but then mentions that each node has some textual information and image information corresponding to P and D (it should be made clear whether this information is just features, or actual node relationships -- if the latter, it seems that P and D should be contained within the nodeset V). - The process described in line 115 around introducing tokens to help condition the representation using graph structure is also explored in some related works, e.g. [2]. Perhaps the authors could consider adopting a similar approach if it makes sense in this task, since the tokenization scheme as the authors of [2] point out is key in injecting the right level of information to the model. - Comment: the notation in pages 3-5 is quite heavy and would benefit from a symbol table. - Section 3.2 proposes a heuristic solution for neighbor selection. I'd encourage exploring solutions designed for learnable importance of multiple edge types similar to [3]. - Can the authors discuss what sort of techniques were used to incorporate graph conditions for the baseline models like InstructPix2Pix and SD? - Is there a quantitative understanding or experiment for the PPR neighbor based sampling approach? It seems this is one of the more heuristic parts of the paper where the design of the sampling procedure (two phase PPR + semantic re-ranking) is less conventional and deviates from other aggregation mechanisms explored in previous literature like attention-based selection, learnable importance of multiple edge types, etc. The qualitative experiment is helpful but not terribly convincing in terms of the actual performance impact in aggregate. [1] Multimodal Graph Benchmark (Zhu et al, 2024) [2] LLaGA: Large Language and Graph Assistant (Chen et al, ICML 2024) [3] Pathfinder Discovery Networks for Neural Message Passing (Rozemberczki et al, WWW 2021) Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes, Appendix A.1 Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your thoughtful review! Regarding your questions: 1. The scenario the authors discuss in lines 28-30 seems like it could be well-handled by only text. We would like to answer this question from three aspects. 1) On one hand, the problem introduced in this paper can be grounded not only in the virtual artwork scenario but also in others such as the e-commerce scenario, where generating an image for a product node connected to other products equates to recommending future products. Such scenarios are hard to handle by text only (e.g., user interests are implicitly expressed in their purchase history which is hidden in the graph structure). 2) On the other hand, even in the artwork scenario, for the artists who are not that famous (e.g., my little brother), it is hard to use text to represent them; and for the artists who are new and not seen during model training, it is hard to expect the model can be generalized through text (artist names). However, using graph structure as the condition, we can well solve the not famous artist and new artist problem by discovering neighbors in the graph. We add an example of controllable generation for a node connecting to “Pablo Picasso” and my little brother on the artwork graph to illustrate this in the rebuttal PDF (in general response). 3) Another advantage of adopting graph conditions is that we can combine multiple graph conditions (e.g., different art styles) with controllable generation (with mixed ratio according to the user’s interest), which is extremely hard to express by text only. 2. Add a reference to the recent [1]. Thank you for bringing this related work. We would like to mention that the reason why we do not reference this paper [1] in our submission is that it is put on arxiv later than the neurips submission deadline. However, as the reviewer mentioned and we agreed, this is a related paper and we would like to reference it in our revision. 3. Question about the notations, especially P, D, and V. We agree with the reviewer that since this paper tackles multimodal learning on graphs, the notations will be a little different and complex compared with conventional graph literature. As mentioned in the paper, each node $v_i\in V$ is associated with a text $d_{v_i}\in D$ and an image $p_{v_i}\in D$. In other words, $d_{v_i}$ and $p_{v_i}$ can be understood as features for $v_i$ from the conventional graph learning perspective. We will make it clear in the revision. 4. Perhaps the authors could consider adopting a similar approach [2] if it makes sense in this task. Thank you for bringing up this great paper and we are glad to reference it in the revision. At the same time, we would like to answer this question from three aspects: 1) The design of Graph-QFormer is non-trivial: as we mentioned in the paper, in MMAGs, we need to tackle the graph entity dependency (including image-image dependency and text-image dependency) to conduct image generation. To extract such dependency, we design Graph-QFormer and show consistently better performance than other designs in Table 2; 2) Different task with different backbone models: we agree with the reviewer that LLaGA is a great work, but we would like to emphasize that the problem tackled in LLaGA (text generation on TAGs with LLM) is different from that in this paper (image generation on MMAGs with stable diffusion). Since the problem and the backbones are different, it is non-trivial to directly compare these two methods; 3) We have compared with simple GNN tokenization methods in the ablation study (Table 2) and demonstrate the effectiveness of Graph-QFormer compared with them. 5. Comment: the notation on pages 3-5 is quite heavy and would benefit from a symbol table. We will add a symbol table in either the main content or the appendix in the revision according to your suggestion. 6. Learnable importance of multiple edge types similar to [3]. We agree with the reviewer that learnable sampling is a promising future direction. As we are the first paper to introduce this problem, we would like to tackle this problem in a simple and effective way. As a result, we would like to leave this more complex learnable sampling for future studies. 7. How graph conditions are introduced to InstructPix2Pix and SD? We compare our method with InstructPix2Pix and SD with an advanced design for graph information in Table 2. For InstructPix2Pix, we aggregate the neighboring images on the graph with mean pooling and serve the resulting image as the input image condition. For SD, we concatenate the text information from the center nodes’ neighbors on the graph to its original text as text condition. From Table 2, our method outperforms both baselines significantly which demonstrates the effectiveness of the InstructG2I design. 8. A quantitative understanding of semantic PPR-based neighbor sampling approach. We are glad to add some quantitative results (DINOv2 score) on ART500K and Amazon datasets as below: \| Dataset \| ART500K \| Amazon \| \|---------------------\|--------\|--------\| \| Ours \| 46.45 \| 51.70 \| \| - PPR \| 45.06 \| 48.40 \| \| - semantic ranking \| 46.19 \| 51.49 \| From the result, we can see that both PPR-based sampling and semantic-based reranking can benefit the InstructG2I. This is demonstrated in both quantitative and qualitative results (Figure 5). The reason we adopt semantic PPR-based sampling rather than attention-based selection and others is that semantic PPR-based sampling can be performed offline only once and it is computationally efficient compared with SD training and inference. This enables us to scale to large-scale graphs in the real world. However, we agree with the reviewer that attention-based selection and others are interesting directions and we would like to leave them for future research. --- Rebuttal 2: Title: Thank you Comment: Dear authors -- thank you for your response to my concerns. I think adding this discussion to the work will help strengthen the positioning (especially the parts around necessity/utility of graph conditioning vs. text). Overall, I stand by my initial review that parts of the work feel a bit heuristic and incremental, but the work is generally an interesting proposal which leans towards a different application of graph structure than most are used to seeing. I will retain my score. --- Rebuttal Comment 2.1: Comment: Dear Reviewer 7dPX, Thank you so much for your reply! We will add those discussions to the revision according to your suggestions. We would like to emphasize the contributions of the work again here: - (Problem). We are pioneers in recognizing the potential of multimodal attributed graphs (MMAGs) for image synthesis and introducing the Graph2Image problem. - (Algorithm). We introduce InstructG2I, a context-aware diffusion model that adeptly encodes graph conditional information as graph prompts for controllable image generation. - (Benchmark). We construct a benchmark with graphs from three domains (art, e-commerce, literature) to evaluate the models on the Graph2Image problem. The benchmark can be used for future exploration of this problem. - (Experiments). We perform experiments on the benchmark, showing that InstructG2I consistently surpasses competitive baselines. For the (Algorithm) part, we propose four main novel components: - Semantic PPR-based neighbor sampling: It helps discover the informative neighbors based on semantic and structure information for target node image generation. - Graph-QFormer: It extracts the common features of the selected neighbors and aggregates the neighbor information as graph conditional tokens for the stable diffusion model, considering the image-image dependency and text-image dependency. - Graph conditions as tokens: It enables stable diffusion to leverage the graph condition information for target image generation. - Graph-based controllable generation (Graph CFG): It enables controllable, tunable generation with multiple graph conditions. Although the philosophy of “graph tokens” is used in graph-enhanced text generation with LLMs, it is not explored in graph-enhanced image generation with diffusion models. Besides, it is nontrivial to get those graph tokens (semantic PPR-based neighbor sampling and Graph-QFormer) and use them for controllable generation (graph CFG). We would like to argue that even though the reviewer insists that “graph tokens” are incremental, it is only a small part of our Algorithm designs. We have other novel design components in Algorithm and other contributions in addition to Algorithm (Problem, Benchmark, Experiments). We would like to thank the reviewer again for the reply. If you have any more thoughts, we are happy to continue our discussion until the deadline.	Summary: This paper focuses on the problem of image synthesis on multimodal attributed graphs (MMAGs) and proposes a graph context-conditioned diffusion model, INSTRUCTG2I, to address the challenge in this setting. In particular, it proposes a semantic personalized PageRank-based method to sample related neighbors in the graph. Then, the INSTRUCTG2I can effectively encode graph conditional information as graph prompts with Graph-QFormer. Systematic experiments on MMAGs demonstrate the effectiveness of the methods proposed in this paper compared to competitive baseline methods. Strengths: 1. This paper studies an interesting and meaningful question. It investigates the graph-structured relationships of real-world entities for image generation on MMAGs, a task well-grounded in practical applications. 2. This paper is well-structured and easy to understand. 3. The graph context-conditioned diffusion model proposed in this paper is reasonable in solving image generation problems on MMAGs. Weaknesses: 1. The description in eq.10 may be incorrect. Please check more carefully. 2. Subsection 3.4 is more challenging to understand when reading. The authors' descriptions of some symbols in Eq. 10 and Eq. 11 are not exhaustive. 3. The results of the ablation experiments in Table 2 indicate that using a GNN such as GAT or GraphSAGE to aggregate graph information seems to be worse than the straightforward approach in Eq.7. Authors are requested to give a more detailed discussion with a reasonable explanation. 4. The images sampled by the semantic PPR-based sampling shown in Figure 5 appear to have the same image as the ground truth. Does this indicate that the proposed method suffers from label leakage? Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Please see the weaknesses. 2. I wonder if the authors will compare it to other state-of-the-art image generation models, such as some Multimodal LLMs that are so prevalent nowadays. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: This paper has reasonably discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your thoughtful review! Regarding your questions: 1. The description in Eq.10 may be incorrect. Thank you so much for your comment. We have found the typos and will correct them in the revision. 2. Descriptions of symbols in subsection 3.4. This section mainly discusses controllable generation with classifier-free guidance (CFG) on graphs. The main philosophy of CFG is that $\epsilon_\theta$ learns the gradient of the log image distribution, and increasing the contribution of $\epsilon_\theta(c,z_t)-\epsilon_\theta(z_t)$ will enhance convergence towards the distribution conditioned on $c$. We agree with the review that the description of the symbols can be improved to make this section easier to understand. We plan to add detailed illustrations to symbols including $\varnothing$, $z_t$, $s^{(k)}_G$, and others. We would be grateful if the reviewer could provide us with inputs on what other symbols need further explanation and we are glad to make this section clearer in the revision. 3. Why do GNNs underperform the straightforward approach in Eq.7? Thank you for the comments. The philosophy of vanilla GNN is to propagate and aggregate information from neighbors and compress the neighboring information into one embedding for the center node, while the straightforward approach in Eq.7 will directly provide neighboring information separately without aggregation or compression. It is worth noting that the aggregation/compression step could lead to neighbor information loss, while the straightforward approach in Eq.7 does not. Since more information is provided to the stable diffusion model, the latter can perform better for image generation. However, we believe that future work can consider a more advanced design of GNN methods that could keep all the neighboring information and extract the common knowledge for more accurate image generation on graphs. 4. The images sampled by the semantic PPR-based method appear to have the same image as the ground truth. Does this indicate any label leakage? We are sorry for the confusion. This is a typo and we will correct it in our revision. We have written data processing codes to ensure that there is no label or data leakage in both training and testing. To demonstrate this, we have uploaded the data processing code here: https://anonymous.4open.science/r/Graph2Image-submit-607E/art500k_dp.ipynb 5. Comparison with other SOTA multimodal LLMs. Thank you for your comments. We would like to argue that the philosophy of introducing graph information into image generation can be not only adopted to the diffusion model (in our paper) but also to any multimodal LLM backbones as mentioned by the reviewer. However, we agree with the reviewer that adding the comparison can make the experiment more comprehensive. As a result, we compare with a recent advanced multimodal LLM called DreamLLM [1] and show the results below: The image-image CLIP score: \| Dataset \| ART500K \| Amazon \| Goodreads \| \|---------------------\|--------\|--------\|------\| \| DreamLLM \| 57.16 \| 60.72 \| 40.72 \| \| Ours \| 73.73 \| 68.34 \| 50.37 \| The image-image DINOv2 score: \| Dataset \| ART500K \| Amazon \| Goodreads \| \|---------------------\|--------\|--------\|------\| \| DreamLLM \| 27.06 \| 34.77 \| 16.13 \| \| Ours \| 46.45 \| 51.70 \| 25.54 \| From the results, we can find that our method can outperform the advanced DreamLLM consistently, which demonstrates the effectiveness of our model design. [1] DreamLLM: Synergistic Multimodal Comprehension and Creation. ICLR 2024. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response, the author effectively addressed my issue, I will increase my score. --- Rebuttal 2: Comment: Dear Reviewer nWct, We are glad that we address your raised issues and we are grateful that you increase the score! We will further enhance the submission according to your suggestions. Best, Authors	Summary: The paper introduces a new task graph2image which is to generate images conditioned on both text descriptions and graph information, which improves consistency of generated images compared to conditioned only on texts or images. To address combinatorial complexity of graphs and dependencies among graph entities, the paper proposes a graph context-conditioned diffusion model InstructG2I for generating images from multimodal attributed graph. Strengths: - To the best of my knowledge, graph2image is a novel task, and the motivation to use the rich and high-dimensional information of graphs for image generation seems reasonable and interesting. - The proposed approach to incorporate graph information into pre-trained text-to-image is new, in particular introducing graph conditioning token and considering scalability of graph size. - The generated samples show that using graph information results in better consistency with the ground truth compared to methods that use only text prompts or images. - Examples of controllable generation with both text and graph show the ability to balance content and style in a simple manner. Weaknesses: While I do not have a major concern, an ablation study on scalability to graph size seems to be missing. How large graphs is the method able to be applied? Technical Quality: 3 Clarity: 3 Questions for Authors: - Why is the DINOv2 score on Goodreads dataset significantly low compared to that of ART500K or Amazon datsets? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Yes the paper addresses the limitation in Appendix A.1. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your thoughtful review and support of our work! Regarding your questions: 1. How large graphs can the method be applied? Thank you so much for your question. Our method can be adapted to large-scale graphs with millions or even trillions of nodes. In InstructG2I, we only need to conduct offline sampling with semantic PPR-based neighbor sampling once to extract useful structure information from the large-scale graphs. Since this step is performed offline, the size of the graph will not introduce sampling bottlenecks to stable diffusion model training and inference. Empirically, the semantic PPR-based neighbor sampling only takes about 10 minutes on a graph with millions of nodes, which is quite short compared to stable diffusion training and inference time cost. 2. Why is the DINOv2 score on the Goodreads dataset lower than the others? Thank you for your great observation. 1) Training data size: As shown in Table 3, the graph size of Goodreads is the smallest compared with ART500K and Amazon. This means that the training data on Goodreads is smaller than the other two datasets. Since a larger training data size can contribute to more sufficient training, the results on ART500K and Amazon are better than the results on Goodreads. 2) Data distribution: In ART500K, Amazon, and Goodreads datasets, the images are art pictures, product pictures, and book cover pictures respectively (some samples can be found in the rebuttal PDF in general response). We believe that compared with book cover pictures, product and art pictures are closer to the distribution of images used in stable diffusion (which is our base model) pretraining. In conclusion, training InstructG2I on ART500K and Amazon will provide better performance than training InstructG2I on the Goodreads dataset. --- Rebuttal Comment 1.1: Comment: Thank you for the responses. I believe this paper tackles an interesting task with novel approach, and do not find any major concerns. Thus I will keep my score. --- Reply to Comment 1.1.1: Comment: Dear Reviewer 5RsF, Thank you for your continuous support of our work! We will further enhance the submission according to your suggestions. Best, Authors	Summary: This paper introduces a novel approach for controllable image generation using both graph and text conditions. The authors propose that additional context information from multimodal attributed graphs (MMAGs) can enhance the performance of diffusion models. Specifically, they formulate the Graph2Image problem and develop the INSTRUCTG2I model to incorporate contextual information during the generation process. Empirical evaluations demonstrate the strong performance of the model. Strengths: 1. The paper is easy to follow. 2. The intuition behind the approach is clear. Weaknesses: 1. The overall setting is questionable. The authors integrate graph information using a Graph-QFormer and context information such as artists and genres, stored in graph prompt tokens. Given the large graph size, they only use subgraph structures. Consequently, the Stable Diffusion (SD) model absorbs additional information from similar artworks, which could be derived from image or text prompts alone. This raises the question of whether an additional condition structure is necessary. I suggest the authors demonstrate a unique application where standard models with text and image prompting capabilities are insufficient. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Are there any unique scenarios where only graph input can significantly improve SD performance? As my review is overdue, I welcome concise feedback and am open to clarifying any potential misunderstandings. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your thoughtful review! Regarding your questions: 1. Question about the problem setting. Thank you for your comment. We would like to answer this from two aspects. Why graph is important? 1) Graph structure helps discover multiple informative neighbors: We agree with the reviewer that Graph-QFormer is to transfer “neighbor images” into graph prompt tokens for center node generation. However, how to select such “neighbor images” is important and non-trivial, which needs help from graph structure. As shown in Figure 5, the “neighbor images” condition affects the generation results a lot, and our proposed semantic PPR-based sampling can discover informative neighbors based on graph information for high-quality image generation. 2) We utilize global graph structure and semantics: We would like to clarify that in our method, we do not “only use subgraph structures” but utilize global graph structure as well as text semantics for neighbor sampling. In semantic PPR-based sampling, we first adopt PPR on the global graph to discover informative neighbors based on graph structure and then adopt semantic search to find more fine-grained informative neighbors (e.g., if we would like to generate a picture of a “horse”, the neighbors related to “horse” will be more useful, as shown in Figure 5). In addition to the qualitative result, we also show the quantitative analysis of semantic PPR-based sampling, where we can find that utilizing global graph structure (Ours in the table below) outperforms utilizing only subgraph structures (- PPR in the table below), which also demonstrates the importance of graph structure information. \| Dataset \| ART500K \| Amazon \| \|---------------------\|--------\|--------\| \| Ours \| 46.45 \| 51.70 \| \| - PPR \| 45.06 \| 48.40 \| \| - semantic ranking \| 46.19 \| 51.49 \| 3) Not just one, but many and extract their similarity from graph: Graph enables discovering multiple related neighbor images and extracting their similarity rather than solely based on one image to conduct image generation (which is widely adopted in the literature). This is important for image generation in many scenarios. For example, we would like to generate a “bird” picture drawn by “Monet”. If we only conditioned on his picture of a “scenery”, we may overfit the corresponding content and fail to draw a “bird”. However, if we conditioned on a multiple of his images covering diverse content including animals, the model would better extract his style and successfully conduct image generation. Why are image or text prompts alone not enough? 1) Text is not enough to describe everything: For example, in the artwork scenario, for the artists who are not that famous (e.g., my little brother), it is hard to use text to represent them; and for the artists who are new and not seen during model training, it is hard to expect the model can be generalized through text (artist names). However, using graph structure as the condition, we can well solve the not famous artist and new artist problem by discovering informative neighbors in the graph. We add an example of image generation for my little brother on the artwork graph to illustrate this in the rebuttal PDF (in general response). 2) Image conditions need to be discovered from the graph: As discussed above, neighboring image conditions can affect the generation result a lot, and graph structure can benefit the informative neighbor image sampling. In addition, solely based on image condition will result in unclear content (e.g., if we want to generate a “bird” image connected to the Picasso node, solely based on the sampled neighboring images does not provide the content indicator of a “bird”). 3) Graph enables flexible and diverse controllable generation: Another advantage of adopting graph as the condition is that we can combine multiple graph conditions (e.g., different art styles) with the controllable generation (with any mixed ratio according to the user’s interest), which is extremely hard to express by text only or image only. We add an example of controllable generation for a node connecting to “Pablo Picasso” and my little brother on the artwork graph to illustrate this in the rebuttal PDF (in general response).	Rebuttal 1: Rebuttal: Dear Reviewers, We sincerely appreciate your valuable feedback and suggestions. We will revise our work based on your reviews. We also want to thank the Reviewers for noting the strengths of our paper, namely: - The problem addressed in our paper is important and well-motivated. (5RsF, nWct, 7dPX) - Our proposed method is substantial and modern. (TQRx, 5RsF, nWct) - The paper is clearly written. (TQRx, nWct) - The empirical results are consistent, solid, and convincing. (TQRx, 5RsF, 7dPX) - The method offers balanced controllable generation. (5RsF) We have addressed the individual questions of reviewers in separate responses. In the revised version, we will incorporate all reviewers' suggestions by making the motivation more solid, adding more experimental results/discussion, adding references, and making the symbols more clear. We have also attached a rebuttal PDF below. Here we would like to briefly outline the contribution of this work for the reference of reviewers to start the discussion. - (Formulation and Benchmark). We are pioneers in recognizing the potential of multimodal attributed graphs (MMAGs) for image synthesis, and we have introduced the Graph2Image problem. Our formulation is validated by three benchmarks based on practical applications in art and e-commerce. Those benchmarks will be valuable for future research. - (Algorithm). Methodologically, we introduce InstructG2I, a context-aware diffusion model that adeptly encodes graph conditional information as graph prompts for controllable image generation (as illustrated in Figure 1(b,c) and the experiment section). - (Experiments and Evaluation). Empirically, we perform experiments using graphs from three distinct domains, showing that InstructG2I consistently surpasses competitive baselines (as illustrated in Figure 1(b) and the experiment section). In closing, we thank the Reviewers again for their time and valuable feedback. If there are further concerns, please let us know, and we will be happy to address them. Pdf: /pdf/5c69d497fcf126fde6563fd115e4ef5e2a776764.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Rethinking No-reference Image Exposure Assessment from Holism to Pixel: Models, Datasets and Benchmarks	Accept (poster)	Summary: The paper introduces a novel paradigm that extends Image Exposure Assessment (IEA) from an image-level to a pixel-level framework. This paradigm comprises three components: model, dataset, and benchmark. Concerning the model, the study introduces the Pixel-level IEA Network (P-IEANet). This network processes images of varying exposures, separates them into low and high-frequency components via a discrete wavelet transform, assesses brightness with the low-frequency component, and evaluates structure with the high-frequency component, ultimately delivering pixel-level assessment results. Regarding the dataset, the authors have developed a new dataset, IEA40K, which includes 40,000 images featuring diverse exposures and corresponding pixel-level annotations. Finally, the paper presents comprehensive experiments on both holistic and pixel-level assessments, yielding promising results. Strengths: 1. The paper initially proposes a pixel-level image exposure assessment paradigm, significantly enhancing precision in the field of image exposure assessment. 2. The paper introduces an assessment network that employs discrete wavelet transform, an intriguing choice supported by several ablation studies. 3. The paper proposes a large-scale, multi-exposure dataset with pixel-wise annotations derived from an automatic multi-exposure fusion technique, subsequently refined by human experts. 4. The paper also demonstrates that the P-IEANet can potentially improve the performance of low-light image enhancement methods. 5. The paper is well-composed, demonstrating a clear structure, precise language, and a logical flow of ideas. Weaknesses: The main weakness is that the paper lacks a well-defined definition for pixel-level image exposure assessment. For other details, please refer to the "Questions" part. Technical Quality: 2 Clarity: 3 Questions for Authors: I find the proposed task interesting, while I have reservations about certain assertions made in the paper, terms that lack clarity, and the absence of adequate justification for the introduction of certain tasks without clear motivation. For further details, see the below list. I've organized my concerns and suggestions according to their significance to assist the authors in prioritizing their rebuttal. 1. The terminology employed in the paper suffers from a lack of clarity, necessitating more detailed explanations. For instance, the term ``exposure`` conventionally refers to the duration of exposure time in the context of capturing images with digital cameras, typically considered as a global attribute of an image. However, the paper introduces the concept of ``pixel-level exposure`` without providing a sufficient explanation, which is illogical in the literal sense. Similarly, the term ``exposure residual`` is introduced but remains poorly defined, further complicating the understanding of the methodology. Probably, the paper misuses the concepts of ``exposure`` and ``brightness``, which cannot be used interchangeably, however. 2. The motivation behind the paper remains ambiguous. It argues that a holistic evaluation of image exposure encounters two primary issues: (1) a dilemma between applicability and practicability, and (2) a narrow inductive bias. However, the paper lacks a further explanation of these problems. Incorporating visual results from current holistic evaluation methods that exhibit these issues could more effectively and intuitively demonstrate the paper's motivation. In the current version, the necessity for a pixel-level image exposure assessment method is not clearly articulated, particularly under which circumstances such a technique would be essential. 3. Related to the first point, the proposed method targets to predict ``exposure residual``, defined as ``(reference - input)`` in RGB space. However, the rationale behind this definition requires further justification. Specifically, it remains unclear why this definition is suitable for use as the ground truth in pixel-wise image exposure assessment. Additionally, it is essential to explore whether any disparity exists between the concepts of ``(reference - input)`` in RGB space and the actual pixel-wise score for image exposure. 4. For evaluation metrics, PSNR and SSIM are two commonly used pixel-wise metrics. However, this paper only adopts SSIM for evaluation. Including PSNR performances would provide a more convincing argument. 5. While comparing pixel-level performance with other image enhancement methods, the paper derives these methods' ``exposure residual`` predictions by directly predicting the residual map. However, image enhancement techniques typically use loss functions designed to smooth the final outputs and align them with human perception, which may not be appropriate for predicting residuals. Although it is acknowledged that the difference between ``(output-input)`` and the proposed ``exposure residual`` exists, incorporating an additional ablation study that calculates the residual from ``(output-input)`` would likely provide a more comprehensive analysis. 6. Important details, such as the architecture of the proposed Long Range Encoder (LRE) and Short Range Encoder (SRE), are missing, hindering the reproductivity of the proposed framework. 7. The availability of the proposed dataset to the public is crucial for assessing the contribution of this work. Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The paper adequately discusses the limitations of moving objects and image size while training. No negative social impact is present in this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: The terminology lacks clarity. >A1: Thanks for the reviewer's thought-provoking questions. >1) In the context of evaluating images, the term "exposure" is no longer a global attribute of an image. Even in the context of capturing images, as exemplified by the reviewer, the term "exposure" conventionally refers to not only exposure time but also two other parameters (aperture and ISO, referred to [1]). Rather than being characterized as a global attribute of the image, the parameters would be more appropriate to be described as a global attribute associated with the camera for capturing the image. Unfortunately, due to the claim made by the classical photographic theory (Adams’ theory) [2] that "The exposure time is the same for all elements, but the image exposure varies with the luminance of each subject element," the coarse global camera exposure attribute fails to match each subject element in an image, potentially result in some subject elements being under-exposed and others being over-exposed. Therefore, in the context of evaluating images, the term "exposure" is no longer a global attribute, as referred to [2] that "Any scene of photographic interest contains elements of different luminance; consequently, the 'exposure' actually is many different exposures." As exposure is fundamental knowledge in IQA, we sincerely apologize for not emphasizing it enough. >2) The concept of "pixel-level exposure" is logically consistent with the term "exposure" in the context of evaluating images. As claimed by the theory [2], the ideal exposure should be refined to different elements. In this paper, we innovatively quantify subject elements by utilizing pixels as the smallest units of exposure measurement. >3) The term "exposure residual" refers to the deviation observed in the actual exposure of each pixel compared to its ideal exposure in the context of evaluating images, as stated in our original paper at line 188: "measuring the deviation of each pixel from its ideal exposure." Numerically, values closer to -1 indicate overexposure, while values closer to 1 indicate underexposure, as detailed in Figure 7. >4) Exposure and brightness have distinct connotations; for instance, exposure can be either holistic or regional (with a pixel as the minimum unit), whereas brightness is solely at the pixel level with a value assigned from 0 to 255. >[1] Image exposure assessment: a benchmark and a deep convolutional neural networks based model, ICME 2018. >[2] Adams, The Negative: Exposure and Development. --- Q2: The motivation and necessity remain unclear, requiring the demonstration of visual results. >A2: Thanks for the valuable advice. Figure 2 in our submitted PDF provides a visual example. Actually, the necessity arises from industrial demands. For instance: smartphone manufacturers currently still manually evaluate exposure images in no-reference scenes; however, by integrating pixel-level IEA (essential) and customized evaluation rules, we have successfully simulated and replaced this manual process for a TOP-5 smartphone manufacturer. If our paper is accepted, we will partially disclose this case on our website. --- Q3: The definition of exposure residual requires further justification. >A3: The insightful questions are greatly appreciated. >1) In training phrase, the exposure residual is not obtained solely from the difference (reference - input) in RGB space, but it undergoes further verification and adjustments by experts to ensure that the final exposure residual closely aligns with the perceived deviation of each pixel from ideal exposure (refer to Figure 7 and line 225 in the original paper). >2) The rationale behind the exposure residual and its suitability for ground truth is based on the inherent characteristics and practicality of pixel-wise data annotation for supervision. For experts, distinguishing between the reference and input images is relatively straightforward and far more accurate [1][2], thus facilitating practical data annotation. While the implementation of absolute evaluations is hindered by the absence of clear standardized criteria. As discussed in Q1, ideal IEA should be derived to the specific characteristics of each subject element, even at the pixel-level. Fortunately, the exposure residual provides direct and effective supervisory information for model training. >3) We are a little confused about "any disparity" in the comments. During prediction, the actual pixel-wise scores for image exposure are the exposure residuals generated by P-IEANet which has been trained using the supervision data (also the exposure residuals but adjusted by experts). >[1] Descriptive Image Quality Assessment in the Wild, ECCV2024. >[2] Self-Supervised Multi-Task Pretraining Improves Image Aesthetic Assessment, CVPRW2021. --- Q4: PSNR performances. >A4: Thanks for the insightful suggestion. The PSNR results, unequivocally showing our SOTA performance, can be found in Table 7 of our submitted PDF. --- Q5: An additional ablation study. >A5: Thanks for the insightful suggestion. We have retrained (using the ground truth of inference images for supervision) representative image enhancement methods to directly predict the reference images instead of using the residual. Then the difference between the output and input was calculated as the residual. The SSIM results, which can be found in Table 8 of our submitted PDF, unequivocally show our SOTA performance despite the decrease in performance for all methods. --- Q6: The architecture details of LRE & SRE. >A6: The classic LRE and SRE are detailed in Figure 3 of our submitted PDF. --- Q7: The availability of the dataset. >A7: Unfortunately and regrettably, the dataset and code in the supplementary materials are not visible due to unknown issues. We apologize for any inconvenience caused and have provided them again. Pls refer to [\#Wvdf, Q1]. --- Rebuttal Comment 1.1: Comment: Having reviewed the rebuttal, I believe it has addressed the majority of my concerns. Consequently, I would like to revise my rating to borderline acceptance. --- Rebuttal 2: Title: Would you please have a look at the author rebuttal? Comment: Dear Reviewer, Thank you very much for contributing to NeurIPS2024. The authors have provided detailed responses to your reviews. Would you please have a look at them? Thanks again. AC	Summary: This work tackles the challenges in image exposure assessment from three aspects: models, datasets, and benchmarks. Specifically, A P-IEANet model based on DWT is proposed, which can generate pixel-level assessment results in a no-reference manner. An exposure-oriented dataset IEA40K is collected to cover various lighting scenarios, devices, and scenes, which are annotated by more than 10 experts with pixel-level labels. A comprehensive benchmark of 19 methods is conducted on the collected IEA40K dataset, where the proposed P-IEANet delivers the best performance. Strengths: + Decomposing images into lightness features and structure components using Haar DWT is theoretically reasonable and empirically effective as presented in this work. + The dataset construction strategies described in Sec. 4.1 and Sec. 4.2 provide valuable insights to the related community. + The proposed model delivers good performance, even outperforming the LMM-based model Q-align. Weaknesses: - Holistic level assessment is performed on SPAQ. It should be straightforward to convert the pixel-level annotations to holistic level annotations in the proposed IEA40K dataset because the pixel-level annotations contain more information than the holistic level annotations. - Would the performance of IEA models be boosted by jointly training (like the practices used in UNIQUE, LIQE, etc.) the model on the combination of IEA dataset and general-purpose IQA datasets? Technical Quality: 3 Clarity: 3 Questions for Authors: Instead of SSIM and PSNR, I think Eq. (7) can also be employed as a pixel-level performance measure. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >We sincerely appreciate the reviewer's positive feedback, characterizing our paper as "theoretically reasonable and empirically effective" and noting that it "provides valuable insights to the related community." We have thoroughly addressed the reviewer's inquiries, which we believe will significantly improve the quality of our paper. --- Q1: "Holistic level assessment is performed on SPAQ. It should be straightforward to convert the pixel-level annotations to holistic level annotations in the proposed IEA40K dataset because the pixel-level annotations contain more information than the holistic level annotations." >A1: Thanks for the insightful comments. To verify the suggestion and convert annotations between the pixel-level annotations and holistic level annotations, we have calculated the average absolute value at the pixel level and adjusted this value by subtracting it from 1, thus normalizing it within a range of 0 to 1. Subsequently, we have conducted additional experiments on this transformed dataset, IEA40K-h, to verify the methods' performance. >1) Given the constraint of rebuttal time, we selected representative methods to retrain on the IEA40K-h dataset, as detailed in Table 5 of the PDF under the "global" response. Our approach unequivocally outperformed the others, achieving SOTA results. The higher SRCC and LRCC metrics obtained with IEA40K-h, compared to those on the SPAQ dataset, indicate that IEA40K-h allows models to learn features more effectively. >2) Table 6 in the PDF illustrates the cross-dataset validation for SPAQ and IEA40K-h on the LRCC metric. Models trained on IEA40K-h, even without fine-tuning on SPAQ, demonstrated performance comparable to those directly trained on SPAQ. In contrast, models trained on SPAQ did not generalize effectively to IEA40K-h, indicating that IEA40K-h provides richer holistic information that enhances model generalization. >The valuable findings will be further elaborated in the final version of the paper (if accepted). We sincerely appreciate the insightful suggestion. --- Q2: Would the performance of IEA models be boosted by joint training (like the practices used in UNIQUE, LIQE, etc.) the model on the combination of the IEA dataset and general-purpose IQA datasets? >A2: Thank you for your valuable advice. Currently, no existing dataset combines pixel-level and holistic-level annotations. Following the reviewer's suggestion in Q1, we derived holistic-level annotations (IEA40K-h) from our IEA40K dataset and adopted a LIQE-like method to jointly learn these tasks. Comparing the original results, 1) For pixel-level tasks on IEA40K, the performance metrics changed as follows: MAE=0.03 (+0.0), SSIM=0.76 (+1.3\%). 2) For the holistic-level task on IEA40K-h, the changes were: LRCC=0.91 (+4.5\%), SRCC=0.87 (+4.8\%). This joint training approach leverages the detailed information in pixel-level annotations to enhance the model's understanding of holistic tasks. We sincerely appreciate the reviewer's suggestion and will include these findings in the final paper (if accepted). --- Q3: Instead of SSIM and PSNR, I think Eq. (7) can also be employed as a pixel-level performance measure. >A3: Yes! Eq. (7) closely resembles the MAE metric detailed in Table 1 of our paper and can serve as a measure of pixel-level performance. --- >In conclusion, the reviewer's suggestion is highly valued, and we will incorporate detailed results into the final version of the paper (if accepted). --- Rebuttal Comment 1.1: Title: Post-rebuttal Comment: Thanks for the responses, I raised my rating to 7.	Summary: This paper proposes a new no-reference image exposure assessment method, Pixel-level IEA Network (P-IEANet), which analyzes and evaluates image exposure from the perspectives of brightness and structure using discrete wavelet transform (Haar DWT). Also, a dataset exclusively tailored for IEA, called IEA40K, is constructed. According to a comprehensive evaluation of methods on the IEA40K dataset, the proposed method achieves SOTA performance and offers advantages for the exposure enhancement community. Strengths: This paper demonstrates very good originality as it is the first realization of pixel-level image exposure assessment. The authors have designed corresponding methods specifically addressing the characteristics of this problem and achieved satisfying results. Detailed explanations of the motivation and the current state of research are provided. Both the principles and the implementation of the method are clearly presented. The experimental results effectively demonstrate the performance of the proposed method. This paper not only proposes a new IEA method but also contributes a new dataset and benchmark, providing a significant boost to the IEA and exposure-related community. Weaknesses: Haar DWT is used to decompose an image into components with different frequencies, but the advantages of this method compared to other similar methods are not adequately explained. In the method section of this paper, some operations lack clear motivation or principles. For example, the reason for applying the DWT^{-1} and the choice of l1 norm as the loss function are not well explained. In the experiments section, SSIM and MAE are adopted to measure the structure and lightness similarity between the ground truth and predicted exposure residual. However, as a perceptual IQA metric, SSIM may not be suitable for evaluating the prediction accuracy of exposure residuals. The paper claims that the proposed method has improved adaptability across varying criteria and scenarios, but this is not well demonstrated in the experiments. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why use Haar DWT to decompose an image into components with different frequencies, and what are its advantages compared to other similar methods, such as other types of DWT? 2. Why is the DWT^{-1} step necessary? 3. What is the reason for using the l1 norm as the loss function? 4. Why choose MAE to measure the structure and lightness similarity between the ground truth and predicted exposure residual, instead of using MSE? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >We greatly appreciate the reviewer's positive feedback on our paper, especially for acknowledging that it "not only proposes a new IEA method but also contributes a new dataset and benchmark, providing a significant boost to the IEA and exposure-related community." We hope the following responses will address any remaining concerns. --- Q1-1: Why use Haar DWT to decompose an image into components with different frequencies? >A1-1: The primary objective is to minimize interference from signals of various frequency domains during the feature extraction phase. This strategy not only enables the model to analyze different features more accurately, thus boosting performance, but also speeds up model training. Supporting experimental results are detailed in Table 2 of the PDF under the "global" response. The data clearly show that employing the Haar DWT significantly enhances performance and decreases the number of epochs needed for convergence. --- Q1-2: What are Haar DWT advantages compared to other similar methods, such as other types of DWT? >A1-2: There are primarily four reasons for our selection: >1) As outlined in the paper (lines 114-129), the Haar wavelet distinctly aligns its component decomposition with exposure characteristics, a unique attribute not shared by other wavelets. >2) The Haar wavelet excels in analyzing signals characterized by sudden variations. It is particularly adept for identifying areas in images that are underexposed or overexposed compared to normally exposed regions, which can be treated as signals with abrupt changes. >3) The transformations of the Haar wavelet, both forward and inverse, are reversible. A proficient approach to estimating exposure residuals involves reconstructing the "ideal exposure image" in the latent space and subsequently assessing the deviation of each pixel in the input image from this ideal exposure. This methodology is essential, as highlighted in line 46 of the original paper: "As a no-reference method, it should effectively simulate reference images in non-preset scenarios, operating similarly to full-reference methods." The Haar wavelet facilitates this process effectively. >4) Experimental results show that the Haar wavelet surpasses other representative wavelets in performance. Table 1 in the PDF under the "global" response, presents a comparative analysis of the Haar wavelet against other notable wavelets (Daubechies and Symlet) on the IEA40K dataset. --- Q2: Why is the DWT^{-1} step necessary? >A2: As outlined in the third point of Q1-2, DWT^{-1} is designed to reconstruct the decomposed components of an image back into the "ideal exposure image" within the latent space. While DWT is utilized for enhanced analysis of features, DWT^{-1} helps the model in synthesizing these features back into the ideal exposure image. This synthesis process is crucial for identifying the discrepancies between the ideal and input images, thus aiding in the prediction of exposure residuals. Omitting the DWT^{-1} process would lead to reduced model performance, as evidenced in Table 3 of the PDF. --- Q3: What is the reason for using the l1 norm as the loss function? >A3: Training results show that the L1-norm outperforms the alternatives. During training, we evaluated three primary types of loss functions: L1-norm, L2-norm, and Smooth L1. Comparative results are detailed in Table 4 of the PDF. The L1-norm demonstrates superior robustness and faster training speeds, as evidenced by earlier convergence epochs in the IEA task. --- Q4: Why choose MAE to measure the structure and lightness similarity between the ground truth and predicted exposure residual, instead of using MSE? >A4: MAE provides a more direct quantification of the deviation of each pixel from its ideal exposure compared to both SSIM and MSE. Additionally, MAE is less sensitive to extreme values than MSE, offering a more robust and consistent measurement of errors. Therefore, using MSE as an option is well-reasoned. In our work, we selected MAE as a complement to SSIM. --- >In conclusion, the reviewer's suggestion is highly valued, and we will incorporate detailed results into the final version of the paper (if accepted). --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. I keep my original ranking.	Summary: This paper proposes an innovative no-reference image exposure assessment method, transitioning from traditional holistic image evaluation to fine-grained pixel-level assessment. This approach effectively addresses the shortcomings of existing techniques in terms of accuracy and generalization. Researchers have developed P-IEANet, a pixel-level evaluation network that utilizes Haar discrete wavelet transform to analyze image brightness and structural information, enabling exposure assessment without reference images. Additionally, to support this method, the researchers have constructed the IEA40K dataset, which contains 40,000 images with detailed pixel-level annotations, covering diverse lighting conditions and devices. Using this dataset, they established a comprehensive benchmark including 19 methods, demonstrating that P-IEANet achieves state-of-the-art performance across multiple evaluation metrics. This work not only enhances the accuracy of no-reference IEA tasks but also provides valuable resources and new research directions for the image exposure research community. Future work will focus on optimizing the framework to support multimodal outputs and enhancing exposure perception in AI-generated content. Strengths: - Pixel-level Evaluation: The P-IEANet proposed in the article is capable of conducting pixel-level image exposure assessment, which offers a more refined analysis and more accurate results compared to traditional overall image assessment. - Innovative Model Architecture: By integrating the Haar Discrete Wavelet Transform with specific feature extraction modules, P-IEANet is able to analyze images from both the brightness and structural perspectives, providing a more comprehensive exposure assessment. - Large-scale Dataset: The article has constructed the IEA40K dataset, which is a large-scale, diverse image dataset that provides rich resources for evaluation and training. Weaknesses: - The author mentions in the abstract that the code and dataset can be found in the supplementary materials, but there is no relevant section in the supplementary materials. - There is no explanation as to why the Haar wavelet was chosen over other wavelets. - The aesthetic quality of Figure 4 needs to be improved. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to the comments in the weakness part. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Not applicable Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >We appreciate the reviewer's positive feedback on our paper, particularly for acknowledging "an innovative no-reference image exposure assessment method." We have addressed the questions raised below. --- Q1: The author mentions in the abstract that the code and dataset can be found in the supplementary materials, but there is no relevant section in the supplementary materials. >A1: Unfortunately and regrettably, the dataset and code in the supplementary materials are not visible due to unknown issues. We sincerely apologize for any inconvenience caused, but we did include the code and dataset in the supplementary materials, as evidenced by Figure 1 of the PDF under the "global" response. Following the NIPS Rebuttal guidelines, we have provided an anonymized link to the AC for the code and dataset in a separate comment. Additionally, if our paper is accepted, we will release all resources publicly. --- Q2: There is no explanation as to why the Haar wavelet was chosen over other wavelets. >A2: There are primarily four reasons for our selection: >1) As outlined in the paper (lines 114-129), the Haar wavelet distinctly aligns its component decomposition with exposure characteristics, a unique attribute not shared by other wavelets. >2) The Haar wavelet excels in analyzing signals characterized by sudden variations. It is particularly adept for identifying areas in images that are underexposed or overexposed compared to normally exposed regions, which can be treated as signals with abrupt changes. >3) The transformations of the Haar wavelet, both forward and inverse, are reversible. A proficient approach to estimating exposure residuals involves reconstructing the "ideal exposure image" in the latent space and subsequently assessing the deviation of each pixel in the input image from this ideal exposure. This methodology is essential, as highlighted in line 46 of the original paper: "As a no-reference method, it should effectively simulate reference images in non-preset scenarios, operating similarly to full-reference methods." The Haar wavelet facilitates this process effectively. >4) Experimental results show that the Haar wavelet surpasses other representative wavelets in performance. Table 1 in the PDF under the "global" response, presents a comparative analysis of the Haar wavelet against other notable wavelets (Daubechies and Symlet) on the IEA40K dataset. --- Q3: The aesthetic quality of Figure 4 needs to be improved. >A3: We sincerely apologize for the aesthetic issues in the figure and greatly appreciate the suggestion. In order to improve the figure's quality in the final paper (if accepted), we have planned to make the following improvements : >1) Adjusting the color scheme to harmonize the appearance of various modules; >2) Modifying the layout to balance the content, particularly on the left side of Figure 4, by adjusting the proportions of core modules and reducing the emphasis on non-core modules; >3) Standardizing the shapes of DWT Kernels to match those of other modules. --- >In conclusion, the reviewer's suggestion is highly valued, and we will incorporate detailed results into the final version of the paper (if accepted). --- Rebuttal Comment 1.1: Comment: Thank you for the clear and convincing rebuttal. I have another question. 1. Can your method be applied to pixel-level image quality assessment? If so, please provide some basic ideas or preliminary experimental results. --- Reply to Comment 1.1.1: Comment: >The acknowledgement of our rebuttal is greatly appreciated. It's wonderful to pose such an insightful question. >1) Our methodology, initially designed for IEA, can be effectively adapted to pixel-level IQA. Currently, there is a lack of open-source datasets specifically tailored for pixel-level IQA; however, we are currently conducting preliminary research in this area and have developed an exclusive dataset that includes pixel-level annotations across 800 images for IQA tasks. We utilized a Transformer-based architecture to predict residuals in the IQA tasks and achieved an SSIM score of 0.69. After fine-tuning these residuals on the KADID-10k dataset using MLPs, our approach attained an SRCC score of 0.93, which closely matches the top-performing method with an SRCC score of 0.94. With more extensive pixel-level annotations available, our methodology has the potential to further improve its performance. >2) Nevertheless, IQA tasks encompass a wide range of distortions, and the methodology for no-reference pixel-level assessment necessitates a more sophisticated design to attain superior performance. Theoretically, our methodology may hold the potential to accomplish this ambitious objective.	Rebuttal 1: Rebuttal: General Response: We sincerely thank the reviewers for their efforts in reviewing our work and providing valuable comments. We highly appreciate the comments received, e.g., the positive comments on our contributions (4/4 reviewers), methods' performance (4/4 reviewers), our presentations (3/4 reviewers), and soundness (3/4 reviewers). We also deeply value the constructive feedback the reviewers have provided. We have taken great care to address the reviewers' questions, conducting additional experiments detailed in the PDF under the 'global' response to further support our claims. Pdf: /pdf/54363ab168f220a76b792ee9ac0403c43de9cbe0.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Neural Gaffer: Relighting Any Object via Diffusion	Accept (poster)	Summary: This paper presents a method for relighting objects observed from a single image. While existing approaches rely on specific capture condition using flashlight illumination or portrait captures, or require to explicitly decompose the scene into geometry and reflectance, the proposed method aims to generate images of a given objects under novel illumination conditions for arbitrary environmental lighting conditions. The authors show that this is possible by relying on a generative diffusion method that is conditioned on the environmental map. The method relies on a pre-trained diffusion model that is fine-tuned on a synthetic relighting dataset to learn the conditioning. The approach is evaluated qualitatively and quantitatively on single-object images. Relying on a conditional diffusion model for relighting, the authors also show additional conditioning on text for relighting. Strengths: This work presents a simple (this is a good thing) and effective method for relighting from a single image. The method relies on synthetic supervision with a novel Blender-rendered dataset that uses Objeverse as input model source. The authors went a long way by collecting diverse HDR environment maps from the Internet that were augmented to produce a large synthetic relighting dataset of almost 20M rendered images with ground truth lighting maps. Overall, the method offers a number of intriguing benefits listed as follows: * Conditional image-to-image diffusion model: The method inherits a conditional Zero-1-to-3 model that is extended in its input latents to a rotated environment map with the camera coordinate frame, allowing for image-to-image relighting in a consistent frame. While, given enough training data, the method is effective in relighting, the approach also enjoys the benefits of existing diffusion architectures with various types of conditioning. The authors demonstrate this effectively with their image conditioning. * Relighting 3D radiance fields: The proposed method is evaluated as a prior for 3D relighting of a neural radiance field. Specifically, the authors propose to use diffusion-based relighting as a coarse reconstruction loss (predicting a coarse relit scene during the NeRF optimization) and a detail refinement loss where the NeRF appearance is further refined. * Qualitative evaluation: The evaluations presented qualitatively in the main manuscript and the supplemental material in the form of supplemental videos are visually plausible and convincing. * Quantitative evaluations: The method is adequately ablated and quantitatively compared to single image relighting methods, 3D radiance field relighting with reasonable margins on the test sets. This validates the method as an effective approach. Weaknesses: What makes the method exciting, at first glance, is also one of the major weaknesses: the technical novelty. The paper piggy-backs on an existing generative method, the Zero-1-to-3 model, that is with a few variations used for relighting. While the simplicity is something that is desired, it also makes it challenging for the reader to derive deeper insights from this work. We learn that pre-trained diffusion-models, when just given enough and the right synthetic data, can allow for plausible novel view synthesis with artefacts that are improved over existing methods. However, the recent work by Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. Dilightnet: Fine-grained lighting control for diffusion-based image generation, 2024. in a way also does show the exactly same, although the technical approach is different. Overall, the technical contribution of the approach is rather incremental (although the method is effective). As such, I am torn on this work. While the technical contribution is not near other work at NeurIPS, the method is effective and likely of high impact. A further qualm I have is regarding the results compared to NVDIFFREC. While the margins are not substantially different, the results in Fig. 6 seem to indicate differently. It seems as if these results are cherry-picked. Technical Quality: 3 Clarity: 3 Questions for Authors: See questions regarding trends in the quantitative evaluations and the qualitative results that do not seem to match. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: All major limitations are addressed. The only open limitation not addressed in the manuscript is the runtime. The authors should address and comment on the runtime for their diffusion model. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments: 1. Technical contribution is incremental considering DiLightNet * Although our method and DiLightNet both approach single-image-relighting using diffusion models, our method is fundamentally different from it and provides novel insights to the community. Dilightnet first estimates a coarse mesh from the single image input using an off-the-shelf depth estimator, then uses Blender to render the estimated mesh with several predefined BRDF materials and target lighting conditions, and finally uses the rendered image maps as the relighting conditions to its diffusion model, which means it still involves some explicit geometry, BRDF and lighting modeling in its relighting process, which can be inaccurate or under-expressive. In contrast, our work is purely end-to-end and shows that relighting for general objects can be totally data-driven without any explicit physically-based decomposition or modeling. To our knowledge, this is a new insight that does not appear in prior work, including Dilightnet. And as pointed out by Reviewer YxGp, our work “provides an avenue for relighting without collecting expensive relighting real-world datasets”. * While our method is simple, the technical details, like how HDR environment maps are represented and encoded to be used as the conditions for diffusion models, are non-obvious and important to get right. We have also conducted sufficient ablation studies to validate the effectiveness of our designs. * In contrast to DiLightNet, which only explores 2D relighting, we also demonstrate how to apply our relighting diffusion model as data prior to downstream tasks such as 3D relighting. We emphasize that our work is also the first to show that 3D relighting can be achieved without an explicit inverse rendering framework that solves BRDFs. Our 3D relighting pipeline is fundamentally different from any previous inverse rendering-based 3D relighting work and achieves better visual quality. It has the potential to be a new paradigm of 3D relighting. * We believe that our simple but effective design is an advantage, making our model robust, more user-friendly (because our model is end-to-end), and easier to scale up (because the training of our model doesn’t rely on any explicit BRDF or geometry information). We are glad to see that all reviewers agree that our methods are effective. In particular, Reviewer rkJr also agrees that our “simplicity is a strength”. 2. The 3D relighting results may be cherry-picked * To demonstrate that the results were not cherry-picked, we present additional 3D relighting comparison results in Figure 8 of the rebuttal PDF, which includes all of our testing objects. The results show that our methods achieve better qualitative relighting outcomes for all the tested objects. We will provide more results in our final revised paper. * Regarding the questions about the "trends in the quantitative evaluations and the qualitative results that do not seem to match," we believe this difference arises because per-pixel-based PSNR does not accurately reflect the visual quality of our diffusion-based relighting method. As shown in Table 3 of our main paper, although our PSNR is only slightly higher than that of NVDIFFREC-MC (29.01 versus 28.25), our LPIPS loss is significantly lower (0.040 versus 0.082). Recent research suggests that human perception-based LPIPS loss is a better indicator of visual quality, which explains the superior visual quality we observe in Figure 6 of our paper. * To further illustrate previous point, please refer to Figure 9 in the rebuttal PDF. In that figure, our PSNR is only 0.8 dB higher than NVDIFFREC-MC in the first example and just 0.2 dB higher in the second example. Despite these small differences in PSNR, our method demonstrates much better visual quality. Our LPIPS loss is consistently much lower than the baselines in both examples, reinforcing the idea that human perception-based LPIPS loss aligns more closely with visual quality trends. 3. Runtime for our diffusion model Based on our local test, it takes ~0.5 seconds to relight one image with one A6000 GPU. --- Rebuttal Comment 1.1: Comment: Dear Reviewer C74A, We would like to express our deepest gratitude for the time and effort you have dedicated to reviewing our work and for offering such insightful questions. We greatly appreciate your recognition of its effectiveness, potential high impact, convincing visual results, and thorough evaluations. As the discussion period will close on August 13th, we kindly ask whether our responses have sufficiently addressed your concerns. If there are any remaining issues or points that require further clarification, please let us know. We are eager to provide any necessary support and continue the dialogue. Thank you once again for your valuable time and expertise.	Summary: The paper introduces Neural Gaffer, an end-to-end 2D relighting diffusion model designed for single-image relighting without the need for explicit scene decomposition. Neural Gaffer can synthesize high-quality relit images of any object under novel environmental lighting conditions by conditioning on a target environment map. The model builds on a pre-trained diffusion model, fine-tuning it on a synthetic relighting dataset. The advantages in generalization and accuracy through evaluations on both synthetic and in-the-wild Internet imagery are shown in the paper. Neural Gaffer can be combined with other generative methods for various downstream 2D tasks like objection insertion. The video results presented in the paper are of high quality. Strengths: 1) Neural Gaffer performs single-image relighting without the need for explicit scene decomposition into intrinsic components like normals and BRDFs. This provides an avenue for relighting without collecting expensive relighting real-world datasets. 2) The model can generate relit images of various objects under different environmental lighting conditions based on a target environment map. The method takes a single image as an input. 3) The method can be applied to real-world objects with high-quality relighting results and perform various downstream tasks such as object insertion. Weaknesses: 1) In case of the real-world object scenarios, the object may not be always centred and may have complex backgrounds and lighting to start with. The paper does not demonstrate how would the method behave in such cases. How about the objects with high-frequency texture details? 2) Related to 1) there might be multiple objects in a scene. From the results, it seems that the method cannot handle multiple objects from a single image. 3) The real-world object examples shown in the paper and the video are good but not impressive. It would be more compelling to show faces, humans, animals etc under the lighting conditions to show the generalizability of the method. Technical Quality: 3 Clarity: 3 Questions for Authors: The paper does not demonstrate how the method behaves when the target object is not centred and has complex backgrounds or varied lighting conditions. How does the method perform in such scenarios, especially with objects that have high-frequency texture details? It appears that the method may struggle with scenes containing multiple objects. Can the authors provide further evaluation or examples to show how the method handles multiple objects in a single image? While the real-world object examples are good, they are not particularly impressive. Can the authors provide more compelling examples involving faces, humans, or animals under varied lighting conditions to better demonstrate the generalizability of the method? While it's understood that portrait lighting might not be comparable to those methods specifically trained on portraits, it would be good to see the generalizability of the method. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments: 1. How the method performs when the target is not centered and has a complex background or varied lighting conditions, especially with objects that have high-frequency texture details We have implemented an automatic preprocessing script that can first detect the object to be relit using Segment Anything (SAM), segment it out from the background, and finally move the segmented foreground object to the center of the input image based on its bounding box. Indeed, many objects in our real data results are not originally centered (we show centered images in the paper for better visualization, but our preprocessing script doesn’t need centered objects). We will release this preprocessing code along with our full code release. We have tested our model on real data with complex backgrounds or varied lighting conditions (as shown in the left column of the supplementary webpage, Fig. 1 in the main paper, and Fig. 3, 4 in the rebuttal file). With the aid of our preprocessing method, we can handle real data with complex backgrounds or varied lighting conditions. In the rebuttal PDF, we show additional examples with high-frequency surface texture details (e.g., sheep and dogs in Fig. 3 of the rebuttal file), and observe high-quality relit results with texture details. 2. Handling multiple objects In our main paper, we did assume there is only a single object present when evaluating various methods, which is a common assumption in most previous object-centric 2D relighting work. And most of our training images only contain one object because we train our model on the filtered Objaverse dataset. However, we find that our method isn’t limited to just single objects. In Fig. 4 of the rebuttal file, we show that our methods still show good generalization ability when tested on real images containing multiple objects. 3. More real-world testing examples(animals and human portraits). As shown in the Fig. 3 of the rebuttal PDF, our methods show good generalization on real-world animal examples, achieving good relit results with high-frequency surface texture details kept. As shown in the Fig. 6 of the rebuttal PDF, our methods show good generalization on human portraits in most cases (see the examples of the first two columns in the Fig. 6 of the rebuttal PDF). However, as we mentioned in the limitations section of the main paper since our model is trained on object data and we didn’t specially train it with portrait data, sometimes it might not achieve high-quality results when handling human portraits. We show a failure case example in the last columns in Fig. 6 of the rebuttal PDF, where we found that our model failed to keep the facial details of the portrait image, such as the shape of the mouth. We believe that further finetuning our model with more portrait data can help our model achieve better performance on human portrait data. --- Rebuttal 2: Comment: Dear Reviewer YxGp, We would like to express our deepest gratitude for the time and effort you have dedicated to reviewing our work and for offering so many insightful questions. We greatly appreciate your recognition of its effectiveness and high-quality visual results, and praise that "it provides an avenue for relighting without collecting expensive relighting real-world datasets." As the discussion period will close on August 13th, we kindly ask whether our responses have sufficiently addressed your concerns. If there are any remaining issues or points that require further clarification, please let us know. We are eager to provide any necessary support and continue the dialogue. Thank you once again for your valuable time and expertise.	Summary: Neural Gaffer presents an approach to object-centric image relighting using diffusion models. The method adapts a pre-trained diffusion model and fine-tunes it on a synthetic dataset designed for relighting tasks. The main feature is its ability to condition the diffusion process on target environment maps, allowing for control over lighting effects. Strengths: 1) Simple yet effective approach: The paper presents a straightforward fine-tuning method for object relighting, similar to zero-1-2-3 shot learning. This simplicity is a strength, demonstrating that complex relighting can be achieved without overly complicated techniques. 2) Powerful data-driven learning: The supervised conditional diffusion model effectively learns to relight objects, highlighting the potential of data-driven approaches in capturing intricate lighting interactions. 3) Competitive results: Based on the presented figures, the method appears to outperform the recent DiLightNet in some aspects. However, this comparison raises some evaluation questions (see questions section for details). Weaknesses: 1) Real-world evaluation: The model is fine-tuned on a synthetic relighting dataset, which might not fully capture the complexity of real-world lighting scenarios. Real-world evaluation is necessary, and there are datasets capturing these effects. The paper is currently missing this evaluation, and there are datasets available for such evaluation [1] OpenIllumination [2] Objects with Lighting or [3] Stanford ORB dataset. These papers have been cited but it is surprising to not see an evaluation of these datasets. 2) Reliance of Environment map: Do you need to supply the environment map for relighting? There is a missing baseline that shows what happens if you condition the target lighting image without a full environment map (only image crops). The Diffusion Light Probe (CVPR 2024) paper indicates that diffusion models are capable of inpainting reliable environment maps and they seem to be implicitly encoded within the model. This baseline will justify why a full environment map is required or necessary for this task. 3) Generalization to scenes: The extent to which the method generalizes to scenes -- not just objects -- is unclear. Evaluating the MIT-multi illumination dataset could shed light on this. The current reliance on explicit environment maps makes it harder to perform on these scenes, but it would be interesting to see if, without explicit environment maps (like suggested above), can you learn to relight and compare on scenes. 4) Evaluation metrics: Recent studies show that PSNR, SSIM, etc. are not consistent with human evaluation. See "Towards a Perceptual Evaluation Framework for Lighting Estimation" (CVPR 2024). These metrics don't tell us much about whether the method is promising as such. A thorough evaluation via user studies or the metrics as defined in the recent paper is currently missing from the paper. 5) Unrealistic results and missing comparisons: The object insertion results look unrealistic, with incorrect shadows that don't match the lighting conditions. Several relevant lighting-aware compositing methods are missing from the comparisons, such as ControlCom [Zhang et al., arXiv 2023], Intrinsic Harmonization [Carega et al., SIGGRAPH 2023], Reshading [Bhattad and Forsyth, 3DV 2022], and ARShadowGAN [CVPR 2020]. The comparison to AnyDoor doesn't make sense as it's not lighting-aware. Including these comparisons would provide a better evaluation of the method's performance against current state-of-the-art techniques. 6) Further, as the papers use off-the-shelf methods to estimate environmental maps (text2light), why not compare with existing 3D object compositing with lighting estimation methods to get a sense of how the proposed methods compare to these tasks -- see Garon et al (CVPR 2019), StyleLight (Wang et al; ECCV 2022) and similar papers? Rendering objaverse objects using lighting estimated from the mentioned or similar methods would help understand the gaps between explicit environment map prediction methods. 7) 3D relighting evaluation: For the 3D relighting setting, according to the Objects with Lighting 3DV 2024 paper, Mitsuba + NeuS is a stronger baseline compared to TensorIR, which is currently missing in the paper. 8) Failure analysis: The paper mentions in the limitations section that the approach might not work for portrait relighting, but it would be interesting to see the kind of failures the diffusion model makes. The current setup lacks experiments in this direction to see what are these failures to encourage future research. Further, the current papers also do not provide any failure examples from Objaverse instances. Is the method perfect on all unseen objects -- detailed analysis is missing as to what objects the proposed methods perform best or worse on. Such analysis helps scope out limitations of the current methods instead of shallow limitations provided in Appendix D. 9) Lack of comparison with simple color matching baselines: The paper doesn't include a comparison with straightforward color adjustment techniques, such as RGB histogram matching between the inserted object and the target scene. This omission raises questions about how much of the method's perceived success in relighting is due to sophisticated light interaction modeling versus simple color transformations. A comparison with such a baseline would help quantify the added value of the diffusion model approach over a simpler method. Technical Quality: 3 Clarity: 2 Questions for Authors: 1) Why weren't datasets like OpenIllumination, Objects with Lighting, or Stanford ORB used for evaluation? 2) Have you explored the necessity of full environment maps for relighting? 3) How well does your method generalize to full scenes, beyond individual objects? 4) Given recent findings on the inconsistency of PSNR and SSIM with human perception for lighting tasks, have you considered user study? 5) Why were comparisons with recent lighting-aware object compositing methods (e.g., ControlCom, Intrinsic Harmonization) not included? 6) Have you considered comparing your method with existing 3D object compositing and lighting estimation approaches? 7) Why wasn't Mitsuba + NeuS used as a baseline for 3D relighting, given its reported strength in recent literature? 8) Can you provide a more detailed analysis of failure cases, including examples from Objaverse instances? 9) Can you provide a simple color histogram matching baseline? 10) Comparison with DiLightNet: DiLightNet offers full image generation with background handling, while Neural Gaffer focuses on object relighting. This raises several points: - Background consistency: How does Neural Gaffer address the background when relighting objects? - Evaluation scope: Are quantitative evaluations done on the full scene or just the object region? This impacts the interpretation of results. - Lighting control: DiLightNet allows full-scene control. How comprehensive is Neural Gaffer's approach in comparison? - User input method: DiLightNet uses radiance hints, Neural Gaffer uses environment maps. How do these compare in terms of user-friendliness and control/precision? - Shadow and Indirect lighting effects quality: DiLightNet's shadows and indirect effects appear more convincing from their project page. Can you comment on this difference? Can you provide a user study comparing the perceived lighting quality between Neural Gaffer and DiLightNet? 11) How sensitive is your method to the resolution of input environment maps? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Somewhat but not fully. See my weakness 8. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed comments and insight suggestions. We will revise our paper based on your feedback. Here are our responses to your comments: 1. Evaluating our relighting model on real-world dataset (in response to weaknesses 1 and question 1) We evaluate our diffusion model on in-the-wild real data as shown in Fig. 1 of the main paper and the video results in the supplemental material. We note that other reviewers stated that these results are impressive (reviewer NcDc, YxGp) and convincing (reviewer C74A). We have provide more real-world example in our rebuttal file. We didn’t evaluate our diffusion model on “Objects with Lighting” and “Standford-ORB” because they only provide relit target images under different camera poses from the input image, which can’t be used for evaluating the single-image relighting task. We didn’t evaluate our diffusion model on “OpenIllumination” because it is a light-stage relighting dataset with different lighting distribution as our methods’s assumption. (Our relighting diffusion model is designed for and trained with general environment maps.) But we are glad to provide our results in our revised paper if the reviewer still wants to see them. 2. Necessity of full environment maps for relighting (in response to weaknesses 2 and question 2) 1. As stated in the abstract, the main focus of our diffusion model is “taking a single image of any object and synthesizing an accurate, high-quality relit image under any novel *environmental lighting condition. Our task is accurate single-image relighting with a user-defined target environment map, which is a common setting for the single-image relighting task. Relighting the image input without a full environment map is a different task and not the main focus of our work. Therefore, we consider comparison with such a baseline outside the scope of our work. 2. That said, our method can be easily combined with other methods to enable relighting without a full environment map. In fact, we have used the paper the reviewer mentions (“The Diffusion Light Probe” (CVPR 2024)) in our object insertion application (see L 230), enabling single-image relighting conditioned on a scene image. 3. Generalization to Scenes* (in response to weaknesses 3 and question 3) 1. We want to re-emphasize that the main focus of this paper is object-level single-image relighting (as indicated by the title — “Neural Gaffer: Relighting Any Object via Diffusion”) — hence, scene-level relighting is not the research scope of this work. Scene-level and object-level relighting are two distinct tasks with different potential methodology designs, and we leave scene-level single-image relighting to future work. 2. Although our model was only trained with synthetic object data, it still shows a degree of generalization ability on scene-level data: As shown in Fig.5 of the rebuttal PDF, we sample some scene images from the MIT multi-illumination dataset (mentioned by the reviewer) and use our diffusion model to relight them with some target environment maps and get some reasonable results: our model can generate shadows in the scenes as shown in the second column of Fig.5 in the rebuttal PDF, and generate reasonable highlights in the scene as shown in the second example of Fig.5 in the rebuttal PDF. We can’t compare with ground truth on this scene dataset because our model was trained with environment maps as conditions, but the MIT-multi illumination dataset doesn’t provide ground truth environment maps. Thank you for your suggestions on how to test our methods without explicit full environment maps. We describe why we need a full environment map for relighting in the previous subsection of this rebuttal reply. 4.Evaluation metrics and why we don’t conduct user study (in response to weaknesses 4 and question 4) 1. The paper the you mentioned studies lighting estimation, which is a different task from relighting. Its conclusion may not be applicable to the relighting task. 2. In addition to the PSNR and SSIM metrics you describe, we also compute LPIPS metrics, which are based on human perception and can better reflect visual quality. 3. When ground truth results are available, comparing different methods via a user study is not a common way to evaluate accuracy in the relighting community. Most recent relighting papers only evaluate their methods by computing commonly used metrics, such as PSNR, SSIM and LPIPS. For example, all papers (over 10 recent popular works) that have been compared in the recent real dataset benchmarks mentioned by the reviewers (Stanford-ORB, Openillumination, and Objects with Lighting) all evaluate their methods by computing PSNR, SSIM, and LPIPS without a user study. These benchmarks themselves only compare different methods on their dataset by computing quantitative metrics without a user study. 4. We appreciate your suggestion regarding the user study. However, due to the limited time for the rebuttal period, we are unable to finish user studies. If you still think a user study is critical for evaluation, please let us know. We will attempt to complete a user study before the end of the discussion period. ---- Unfinished. Please keep reading the comments. --- Rebuttal 2: Comment: 5. Questions related to object insertion (in response to weaknesses 5，6，9 and questions 5，6，9) We sincerely appreciate your detailed and insightful suggestions regarding object insertion. Before addressing each of your questions, we would like to clarify that the primary focus of this paper is on relighting, rather than object insertion. In this work, we introduce a robust 2D relighting diffusion model, which can serve as relighting data prior for various downstream tasks and applications in both 2D and 3D. To illustrate its potential, we developed a straightforward object insertion pipeline as one of the 2D application examples. However, we do not consider object insertion to be the core contribution or the main focus of our work. a. Provide more baselines for object insertion comparison, including a simple color histogram matching baseline Due to limited time and space in the rebuttal PDF, and because object insertion is not the primary focus and core contribution of this work, we have not provided additional comparisons at this time. We did not compare our methods with the histogram-matching baseline because it is too simple and has not been compared in recent work. However, we are willing to include these baselines you mentioned in our revised paper. b. Comparison with existing 3D object compositing approaches We used DiffusionLight (CVPR 2024) to estimate the environmental lighting of the scene image in our object insertion results, rather than the Text2Light method you mentioned. Since 3D object compositing and lighting estimation are not the tasks of this paper, we did not consider comparing against the methods you mentioned. In theory, our methods could be combined with the lighting estimation techniques you suggested, potentially leading to improved performance in the object insertion task. We appreciate your suggestion and will consider exploring these design choices in future work, but we emphasize that understanding the gaps between different lighting estimation methods is beyond the scope of this paper. 6. Compare with “NeuS+ Mitsuba” in the 3D relighting task (in response to weaknesses 7 and question 7) 1. We have tried to compare with the SOTA methods in our 3D relighting experiments. Specifically, we chose Nvdiffrec-mc and TensoIR for comparison in our paper. According to the Stanford-ORB benchmarks, Nvdiffrec-mc achieves SOTA results among all tested methods. Similarly, TensoIR attains SOTA results based on the OpenIllumination benchmarks. 2. We did not compare our methods with Neus+Mitsuba because it is not a commonly used baseline in the field of 3D relighting, and we were not aware of it initially. To our knowledge, Neus+Mitsuba has only been compared in the paper "Objects with Lighting," and we have not seen other recent 3D relighting work that includes this baseline. 3. During the rebuttal period, we tested Neus+Mitsuba on our testing dataset using its official code. The results showed that its performance metrics are as follows: PSNR: 26.47, SSIM: 0.912, and LPIPS: 0.073. Therefore, our methods outperform Neus+Mitsuba on all metrics. 7. Detailed Analysis of Failure Cases (in response to weaknesses 8 and question 8) Thanks for your suggestions. We have provided a more detailed analysis of the failure cases below: a. Portrait Relighting Analysis Overall, our method demonstrates good generalization to human portraits in most instances (see the first two rows of Fig. 6 in the rebuttal PDF). However, we observed that it sometimes struggles to preserve facial details. For instance, in the failure case shown in the last row of Fig. 6, our model failed to retain crucial details of the portrait, such as the shape of the mouth. We believe that further fine-tuning our model with additional portrait data could improve its performance on such images. b. Limitations in Objaverse Instances Color Ambiguity: Our model sometimes fails to resolve inherent color ambiguities, particularly when relighting diffuse objects. As shown in the first row of Fig. 7 in the rebuttal PDF, the model struggles to produce accurate color in the relit results. This issue arises from the inherent color ambiguity, where the color in the input image could be attributed to either the material or the lighting, making it a challenging problem. We believe that including the background of the input image in the diffusion model could help mitigate this issue, as the model could learn to infer the lighting conditions from the background. We consider this a potential avenue for future work. High-Frequency Reflection Generation: Our model also struggles to generate high-frequency relighting reflections. For example, in the second row of Fig. 7 in the rebuttal PDF, while the model accurately generates low-frequency highlights (as indicated by the red box), it fails to produce high-frequency reflections (as indicated by the green box). --- Unfinished. Please keep reading the following comments. --- Rebuttal 3: Comment: 8. Questions related to comparison with DiLightNet (in response to question 10) Before we answer your questions, we want to first point out that DiLightNet generates the final image by first generating a relit foreground image with its diffusion model (which is exactly what our relighting model can also do), and then generates the background separately. It generates the background in two ways: (1)When environment maps are given, DiLightNet will render a background image from the target environment map, and then use the foreground mask to composite the relit foreground image and rendered background image to get the final image. We have done the same thing in our paper. As shown in Fig.1 & 4 of the main paper and the supplementary webpage, we also synthesize the background for our relit image and generate a full image with the background. (2) When environment maps are not given, DlightNet just uses a pre-trained diffusion-based inpainting model(the stable-diffusion-2-inpainting model [Stability AI 2022a]) to inpaint a background for the relit foreground image. Although we didn’t do this in our paper, using a diffusion-based inpainting model to inpaint the background is very easy for our methods to do. In short, our method can achieve the same full image generation with the background as DiLightNet can. - Question: Background consistency: How does Neural Gaffer address the background when relighting objects? Answer: As mentioned at the beginning of this reply subsection, we render a background image from the target environment map. Please refer to the previous paragraphs for more details. - Question: Evaluation scope: Are quantitative evaluations done on the full scene or just the object region? This impacts the interpretation of results. Answer: When conducting quantitative evaluations, we first use the GT foreground masks obtained during rendering to mask out all background pixels of the testing image and set background pixels to be all 1 (purely white pixels). Then we compute the metrics. This means that we only compare metrics on the relit foreground, which is fair for both methods. - Question: Lighting control: DiLightNet allows full-scene control. How comprehensive is Neural Gaffer's approach in comparison? Answer: We are not sure what full-scene control means here. Could you please give us more explanation for this question? - Question: User input method: DiLightNet uses radiance hints, Neural Gaffer uses environment maps. How do these compare in terms of user-friendliness and control/precision? Answer: First, DiLightNet uses radiance hints as its diffusion input, not as the user inputs. Radiance hints are attained by first estimating a coarse mesh of the single image input and using Blender to render the estimated mesh with some predefined and fixed BRDF under the target lighting. To relight an image with DiLightNet, users will first need to specify the lighting information used to render the radiance hints. So the real user inputs of DiLightNet are the lighting information, i.e. environment map, etc. Therefore, our method has a similar user input as DiLightNet. Second, our methods are more user-friendly. This is because our method is purely end-to-end — users only need to specify the target environment map, and our relighting diffusion will output the relighting diffusion directly. On the contrary, to relight a single image with DiLightNet, users need to first run a monocular depth estimator to attain the mesh, (which may fail) and then need to use Blender to render the radiance hints and use them as the diffusion inputs. This whole process is more complicated and tricky than our methods. - Question: Shadow and Indirect lighting effects quality: DiLightNet's shadows and indirect effects appear more convincing from their project page. Can you comment on this difference? Can you provide a user study comparing the perceived lighting quality between Neural Gaffer and DiLightNet? Answer: DiLightNet’s results may look good separately but they are not accurate when compared with the ground truth relighting results, as we show in the Fig. 4 of our main paper. In addition to visual quality, relighting should also achieve accuracy. As we find in Table 1 of our main paper, our methods achieve more accurate relighting results while maintaining high visual quality. We explain why we didn’t perform a user study in the previous subsection of this rebuttal reply. --- Unfinished. Please keep reading the following comments --- Rebuttal 4: Comment: 9.How sensitive is the method to environment map resolution (in response to question 11) We always resize the target relighting environment map to the resolution required by our diffusion model (256 * 256) before inputting them into our diffusion model. We implemented an energy-preserving resizing function to do this. Therefore, our model is not sensitive to the environment map resolution theoretically. We have also tested conditioning on environment maps with different original resolutions locally (such as 512 $\times$ 1024, 1024 $\times$ 2048, and 2048 $\times$ 4096), and our local results don’t show obvious differences in diffusion relighting results. --- Rebuttal 5: Comment: Dear Reviewer rkjr, We would like to express our deepest gratitude for the time and effort you have dedicated to reviewing our work and for offering so many detailed and insightful questions. We greatly appreciate your recognition of its effectiveness, competitive results, and powerful data-driven potential. As the discussion period will close on August 13th, we kindly ask whether our responses have sufficiently addressed your concerns. If there are any remaining issues or points that require further clarification, please let us know. We are eager to provide any necessary support and continue the dialogue. Thank you once again for your valuable time and expertise. --- Rebuttal Comment 5.1: Title: Response to rebuttal Comment: Thanks for the detailed response. Some of my concerns are addressed and I appreciate the author's effort. Some questions remain: 1) One way to evaluate for real-world evaluations might be to use off-the-shelf models to recover environment maps from the relit images -- like DiffusionLight (CVPR 2024) or StyleLight (ECCV 2022). Use these recovered maps as a guide and relight objects and compare them with the GT. 2) Re Necessity of environment maps: My question was about conducting an ablation study to determine if environment maps are really necessary, or if the task can be accomplished using only target image crops. This is because DiffusionLight demonstrates that Stable Diffusion has a strong grasp of environment maps. 3) Re Simple color matching baselines: I don't think there is a satisfactory response to this question. I'm curious to know what would happen if authors simply tried to match the RGB values of the object to the target scene. The easiest way to do this would be to apply a correction coefficient to the object that needs to be relit for each RGB channel. This coefficient would be computed as the average color ratio of the target scene to the source scene. This method provides a good baseline for accounting for global color changes and indicates how much of the result is a result of learning to relight, and how much might be the method finding a shortcut to match the overall color distribution of the target scene. 4) I disagree with the authors who claim that a user study cannot be conducted for this task. In general, for most relighting tasks, there are no reliable test sets with actual ground truth. Also note that these are simulated ground truths, based on approximations from the chosen rendering engines, and should not be assumed to be the actual ground truth. The authors also mention that in the relighting community, it is not common to evaluate accuracy by comparing different methods through a user study when ground truth results are available. However, in object compositing literature, having a user study is still a standard practice even though simulated GT results are available. In addition, the paper already referenced in my review https://arxiv.org/abs/2312.04334 shows that the LPIPS metric is also not reliable contrary to the authors' argument that they adhere to human perception. Please see Fig 5 of their paper. This finding is not only applicable to lighting estimation methods but can also be used to assess how these metrics fare in general for lighting-related tasks. Relying on the argument that previous approaches used certain metrics is not sound when there is now evidence showing that these metrics are unreliable. --- Reply to Comment 5.1.1: Comment: Thanks for your replies. Here are our responses to your remained questions. 1. Suggestions for Real-World Evaluations We appreciate your suggestion to use off-the-shelf models to recover environment maps from the relit images and to use the recovered light maps for relighting and comparison with the ground truth (GT). However, we would like to emphasize that relighting results are highly dependent on the input target lighting. To our knowledge, there is currently no off-the-shelf lighting estimation model that guarantees accurate and reliable lighting estimation. Therefore, we believe that using estimated lighting for relighting and then comparing it with GT would not provide a fair and reasonable evaluation for either our method or the baselines. If the relit results do not align with the GT, it would be difficult to determine whether the discrepancy is due to the relighting model or the inaccurate lighting estimation. 2. Re Re Necessity of environment maps We would like to reiterate that our task is defined as relighting a single image using a full target environment map, which is a common and well-established task setting in many previous single-image relighting papers. While the idea of using incomplete environment maps as a relighting condition is interesting and could be explored in future work, it represents a different task that lies beyond the scope of this paper, and should not be a ablation study. 3. Re Re Simple color matching baselines In our rebuttal, we mentioned that we did not include a comparison with simple color histogram matching baselines due to limited time and space in the rebuttal PDF (we had fitted 9 figures on a single page), and the object insertion task in which the simple color histogram matching was suggested to compare was not our primary focus and core contribution. In addition, we did commit to including this comparison in the revised paper. We have now conducted further evaluations using simple color histogram matching baselines to better answer your questions: a. We tested the simple color histogram matching baseline on our 2D relighting dataset. This involved computing the average color ratio between the target relighting background and the input image, then rescaling each RGB channel of the input image accordingly. We then computed metrics such as PSNR, SSIM, and LPIPS, as shown below. b. We also included the simple color histogram matching baseline in our user studies. (Please refer to the next reply for more details.) \| \| PSNR↑ \| SSIM↑ \| LPIPS↓ \| \| --- \| --- \| --- \| --- \| \| Our Method \| 26.706 \| 0.927 \| 0.041 \| \| DiLightNet \| 24.823 \| 0.918 \| 0.056 \| \| Simple color matching \| 21.894 \| 0.912 \| 0.072 \| Both quantitative comparisons and user studies demonstrate that simple color matching performs poorly compared to both our method and the baseline, DiLightNet. This result is intuitive, as relighting involves not only color matching but also complex appearance changes, such as shadow generation, highlight creation, and other nuanced effects. We have illustrated these challenging scenarios in Figures 1 and 4 of the main paper and on our supplementary website.	Summary: The paper proposes a novel method for single-image relighting, which takes an image of an object and a target environmental map as inputs. The authors fine-tune Stable Diffusion on a synthetic relighting dataset to output relit images, conditioning on both the input object image and the target environmental map. The authors show their method outperforms existing baselines. Additionally, the trained relighting model can be applied to downstream tasks such as relighting a neural radiance field and object insertion. Strengths: - I check the video results in the supplementary video. The visual results are impressive. - The authors have shown several downstream applications using their trained relighting model, including text-based relighting and object insertion. - The authors have conducted extensive ablation studies to prove the effectiveness of their proposed method. Weaknesses: I don’t have many complaints about the paper. I list several potential improvements below: - In the 3D relighting experiments, it seems unfair to compare with inverse rendering methods such as Nvdiffrec-mc and TensoIR, as they can apply any lighting to the object once the material is recovered, while Neural Gaffer needs optimize for every lighting. On the other hand, I think Neural Gaffer should be combined with these inverse rendering methods and provide priors when recovering material and lighting. - The extrinsic information is injected by rotating the environmental map. However, it seems intrinsic information is not considered, which means there is an assumed fixed FOV. This could introduce biases in downstream applications and limit the input views in 3D relighting. - The quantitive comparison with IC-Light is missing. - The generated image resolution is limited to 256x256. Technical Quality: 3 Clarity: 3 Questions for Authors: The problem of inverse rendering with a single image is inherently ambiguous. For example, the object color in the input image could come from either the material or the lighting. I was wondering about the authors' thoughts on this problem in the context of Neural Gaffer. When relighting an object, there could be multiple possible outputs depending on the decomposition of the material and lighting in the input. Is the probabilistic model of Neural Gaffer able to model this perplexity? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments: 1. Inherent color ambiguity Color ambiguity is an inherent issue in the single-image relighting task. That said, we found that our data-driven model can handle this problem to some extent, especially when the objects’ materials in the input image are relatively specular. Because the diffusion model can infer the lighting of the input image through the reflections or specular highlights. In Fig.1 of the rebuttal file, we show two such examples: we render each object with two different environment maps and relight them with the same environment map. In both cases, the input images have different colors and we want to relight them under the same target lighting. It turned out that both input images can be relit to the results that match the ground truth results, which indicates our model has some ability to handle the color ambiguity. As for the general diffuse object, we think our model could learn an implicit prior over common colors that show up in object albedos, as well as common colors that occur in lighting, and use that to help resolve this ambiguity. However, this may fail sometimes, as shown in the first example of Fig.7 in the rebuttal file. One potential solution could be also inputting the background of the input image to the diffusion model since the diffusion model can learn to infer the lighting of the input image from the background. We leave this as future work. 2. Neural Gaffer versus other inverse rendering work and using our model for materials recovering * Although our method must re-optimize the 3D representation for each new lighting, we would like to emphasize that the re-optimization process is relatively fast: in our implementation, the two-stage 3D relighting process takes just a few minutes on a single GPU. * The reviewer suggests that we combine Neural Gaffer with inverse rendering methods. That is a good idea, but our paper intentionally sets out to avoid explicit inverse rendering and decomposition, and explore whether we can successfully build systems using an implicit relighting capability. We first achieve this goal in single-image relighting task, and then we follow the same motivation and build our 3D relighting pipeline. It turns out that such a new pipeline works very well and in theory it has better data-driven and representation potential. Since such a new pipeline is essentially different from previous BRDF reconstruction-based methods and has no similar previous baselines to compare with, we can only compare this pipeline with previous reconstruction-based methods that also work on 3D relighting, such as TensoIR and Nvidiffrec-MC. * Thank you for your suggestion about using our model as prior for recovering materials. As we noted in our paper, “our model can also operate as a strong relighting prior for 3D tasks” (L 18). Using our model to improve materials recovering is for sure another interesting application. Due to the limited time we have, in Fig. 2 of the rebuttal PDF, we show a simple experiment where we use our model as data prior for recovering materials, demonstrating an improvement in albedo reconstruction quality. In particular, given input images with unknown lighting, we relight each of the input views under three different target environment maps, then use TensoIR to reconstruct BRDFs from the multi-illumination relit images (we use TensoIR because it supports multi-illumination input images). Results show that both NVDIFFREC-MC and the original TensoIR bake shadows present in the original images into the albedo map; in contrast, combining TensoIR with our diffusion prior yields albedo maps with reduced baking of shadows. Note that the current experiment is very simple and has the potential to improve for best quality. We leave this as one of the future work. 3. Camera Intrinsics are not considered * Although the training data was rendered with fixed camera intrinsics, the model shows good generalization ability to cameras with different intrinsics when we test our 2D relighting model using real images from the Internet, which have unknown FOV (as shown in Fig. 1 of the main paper and our supplemental webpage). Training our diffusion model using images rendered with different intrinsics may further improve its performance. We thank the reviewer for the suggestion as an important future direction for scaling up the system. * Note that for the 3D relighting application, we first reconstruct a radiance field using the multi-view inputs and then relight the reconstructed radiance field. Since the reconstructed radiance field allows us to do novel view synthesis, we can render novel views with any camera intrinsics, and therefore for the 3D relighting task we can always render novel views with the fixed intrinsics used during training. 4. Quantitive comparison with IC-Light Since IC-light assumes a different lighting input from our method (background image versus environment map), we feel it may not be fair for us to compare with it in the main paper because we have more input lighting information, which is why we only compare with it qualitatively on real data in the appendix and the supplementary videos, in which we have shown an obvious advantage over it. But if the reviewer still thinks quantitive comparison with IC-Light is important, we will provide it in our revised paper. 5. The image resolution is limited We have realized that the relatively low image resolution is one of our methods’s key limitations and mentioned this point in our limitations section. The main reason we don’t try higher resolutions is that we only have limited CPUs to preprocess the data and limited GPUs to train the model. In theory, our methods can support higher resolutions. We will release all code to encourage future work to improve this point. --- Rebuttal 2: Comment: Dear Reviewer NcDc, We would like to express our deepest gratitude for the time and effort you have dedicated to reviewing our work and for offering so many potential improvement suggestions. We greatly appreciate your recognition of its effectiveness, impressive visual results, and extensive ablation studies. As the discussion period will close on August 13th, we kindly ask whether our responses have sufficiently addressed your concerns. If there are any remaining issues or points that require further clarification, please let us know. We are eager to provide any necessary support and continue the dialogue. Thank you once again for your valuable time and expertise.	Rebuttal 1: Rebuttal: Dear Reviewers, Thank you for dedicating your time to review our paper and offering insightful feedback. We sincerely appreciate your efforts to help enhance the quality of our research. We are also pleased to note that all reviewers were supportive of our work: (a)Recognize our methods are effective and have high-quality results (NcDc, rkJr, YxGp, C74A) (c) Praise our methods outperform the existing methods(NcDc, rkJr, C74A) (b)Acknowledge our extensive ablation study to prove the effectiveness of the proposed method.(NcDc, C74A) (c) Praise our methods enable many downstream applications(NcDc, YxGp) (d)Recognize our methods are simple and effective, which is our strength ( rkJr, C74A) (e) Acknowledge our methods show good generalization ability on real world data, and have impressive (NcDc, YxGp) and convincing (C74A) results (f) Acknowledge our methos show potwerful data-driven potential ( rkJr) and provide an avenue for relighting( YxGp) We thank all the reviewers for your insightful suggestions. Since we have shown a lot of new results as required by reviewers in the rebuttal PDF file. Please zoom in for better visualization when reading it. Pdf: /pdf/81ea8879af9690df85689268ca9330e698ce4314.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Towards Understanding the Working Mechanism of Text-to-Image Diffusion Model	Accept (poster)	Summary: This paper aims to understand two mechanisms of diffusion models. First, the denoising process is analyzed, and it is found that shapes in an image are constructed in the beginning of the denoising process, while textures and details are filled in later. This empirical observation is justified with a mathematical frequency analysis. Second, the role of text conditioning is analyzed and it is found that the [EOS] token, which captures global information of the prompt, is relied on more heavily by the diffusion model. It is also observed that the text prompt is utilized more in the earlier stages of the denoising process. This finding is utilized to speed up diffusion sampling by ~25% while maintaining the image quality and prompt alignment. This is done by only injecting conditional information in the beginning of the denoising process. Strengths: * Although the finding that shape is constructed in the first few timesteps has been observed many times before, it is nice to have a more principled study with various experiments and mathematical justification. * The finding that the special [EOS] token is the most relied upon during generation rather than the prompt tokens is an interesting finding that can be used in later studies. For instance, improving prompt alignment, attribute binding, etc. * The observation that the text prompt is used more in the early denoising process lends itself to a practical application of speeding up inference. * Multiple architectures and samplers are used in this study, suggesting the generality of these findings. Weaknesses: * As mentioned in the Strengths section above, the findings are not completely surprising (for instance, the shape reconstruction or reliance on text in the early denoising steps, then detail-filling in the later steps). However, this work takes a principled approach in studying these phenomena which have largely been used in diffusion application literature (e.g., [1, 2]) * Limited to no mention of broader impact or limitations. Furthermore, the Conclusion section is just a summary of the paper but does not discuss the implications of these findings. [1] @inproceedings{mengsdedit, title={SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations}, author={Meng, Chenlin and He, Yutong and Song, Yang and Song, Jiaming and Wu, Jiajun and Zhu, Jun-Yan and Ermon, Stefano}, booktitle={International Conference on Learning Representations} } [2] @inproceedings{hertzprompt, title={Prompt-to-Prompt Image Editing with Cross-Attention Control}, author={Hertz, Amir and Mokady, Ron and Tenenbaum, Jay and Aberman, Kfir and Pritch, Yael and Cohen-or, Daniel}, booktitle={International Conference on Learning Representations} } Technical Quality: 3 Clarity: 3 Questions for Authors: * What are some of the limitations and implications of these findings? * I did not take this into account for my review, but there are many typos in the text and figures which can be corrected for the next version. * This (https://arxiv.org/pdf/2404.07724) is a concurrent work, so it is not expected for this paper to compare against it. Their finding is applying guidance in the middle denoising steps improves image quality and distribution coverage. I am curious to hear how the findings this paper being reviewed can be connected to the phenomena observed there. It might be something to keep in mind for the next version, although it will not be used in assessing this paper. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Although there are no societal implications, a discussion of limitations is lacking. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable comments. Here we address your concerns as follows. Q1: “As mentioned in the Strengths section above, the findings are not completely surprising (for instance, the shape reconstruction or reliance on text in the early denoising steps, then detail-filling in the later steps). However, this work takes a principled approach in studying these phenomena which have largely been used in diffusion application literature” A1: We thank you for appreciating our “principled approach in studying these phenomena”. As you mentioned and our claim in line 161, these phenomena has been previously observed by single cases. However, our contributions are conducting a systematically exploration to them by frequency analysis. Besides, we also study the working mechanism of text prompts, and link it with these observed phenomena. Q2: What are some of the limitations of these findings? A2: For the limitation of this paper, owing to the auto-regressive textual encoder used in stable diffusion explored in this paper, it is natural for [EOS] contains more information. However, under bidirectional textual encoder, every token should have the same information magnitude. Thus, the conclusion “[EOS] contains more information” may not hold anymore. In fact, when for T2I model Pixel-Art [1] with bidirectional text encoder T5 [2], the conclusion indeed does not hold anymore. However, the other conclusions “first overall shape then details” and “text prompt convey their information in the early stage of diffusion process” are still hold and has been verified in Table 2, where we have conducted the proposed sampling strategy on Pixel-Art. We will clarify these limitations in the revised version. Q3: What are some of the implications of these findings? A3: The straightforward implication of our finding is designing the efficient sampling strategy as we proposed in Section 6. Besides, we wonder is there any possibility to combine these findings during the training stage of T2I model, e.g., using our noise prediction (9) during training stage to accelerate training. For the other conditional generation tasks e.g., human-face generation or subject-driven generation, we find that our sampling strategy is still applicable. Please see General Response for more details. We will clarify these implications in the revised version. Q4: “I did not take this into account for my review, but there are many typos in the text and figures which can be corrected for the next version.” A4: We will carefully check the typos and revise them accordingly in the revised version. Q5: “This (https://arxiv.org/pdf/2404.07724) is a concurrent work, so it is not expected for this paper to compare against it. Their finding is applying guidance in the middle denoising steps improves image quality and distribution coverage. I am curious to hear how the findings this paper being reviewed can be connected to the phenomena observed there.” A5: Thank you for pointing out such interesting reference [3]. Their main conclusion is actually consistent with ours. They divide the denoising process into three stages: early, middle, and late. They claim that the guidance (textual prompt in this paper) of conditional generation in late stage is useless, which is consistent with our conclusion that textual prompts convey their information in early stage of diffusion process. Moreover, they claim the guidance existence in early stage of diffusion is strong and make the generation saturates into several modes, so that the guidance should only applied in the middle stage of diffusion process. This is also consistent with our observation that the overall shape is quickly decided by the text prompt in early stage of diffusion process, owing to the strong guidance of text prompt during this stage. The main difference between our methods is they propose to remove the guidance in early stage of diffusion process to improve diversity, while we add guidance in this stage. Our method improves the alignments of generated images with target prompts, especially for small or middle size model with relatively poor alignments, while for model with poor diversity, their method seems fix the issue. We will add this comparison in the revised version. References: [1] Chen et al., 2024. PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. [2] Raffel et al., 2019. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. [3] Kynkäänniemi et al., 2024. Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models。 --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response. I have also gone through the responses to the other reviews as well as the rebuttal PDF. As other reviewers have also pointed out, some of the findings have been discussed before in the literature (shape-then-detail). Although the findings and applications are fairly narrow in scope, I think they can have an impact on further interesting studies of diffusion models. Moreover, this work takes a principled approach. Thus, I maintain my score of weak accept. I appreciate the response on the limitations of the findings, so I suggest the updated paper to include those.	Summary: This paper explores the mechanism in the text-to-image diffusion model, including the generation order of image components, the influence of various tokens, and the steps in which tokens work. These observations bring some insight into understanding the diffusion model. Besides, the authors also design a sampling strategy that accelerates the sampling of the denoising process by 25%+. Strengths: 1. The conclusion of the [EOS] token is interesting and has been rarely investigated in previous papers. 2. The analytical experiments in this article are sufficient and strongly support its conclusion. 3. The writing expression of this article is very clear. Weaknesses: 1. The other conclusions in this paper, e.g., shape first then details, have been discussed in previous works. 2. The sampling strategy is more like a sample trick than a method. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Is the proposed sampling strategy still feasible for generating tasks that require preserving specific details, e.g., subject-driven generation? Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 2 Limitations: Suggest the author to discuss the applicability and limitations of the proposed sampling scheme. For example, can it be applied to human face generation without losing human identity? Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable comments. Here we address your concerns as follows. Q1: The other conclusions in this paper, e.g., shape first then details, have been discussed in previous works. A1: Yes, and we have mentioned in line 161, footnote 5. However, the existing literature only observe this phenomenon by a single case, while we conduct a frequency analysis to systematically explore this phenomenon, and our contributions also include the exploration to the working mechanism of text prompt. Q2: “The sampling strategy is more like a sample trick than a method.” A2: The sampling strategy is an application of the conclusion that text prompt convey their information at first few steps of denoising process. It saves computational cost in a simple yet efficient way. Q3: “Is the proposed sampling strategy still feasible for generating tasks that require preserving specific details, e.g., subject-driven generation?” A3: Thanks for your suggestion, we implement the proposed sampling strategy on a recent subject driven generation method, Anydoor [1]. Here, the visual features of reference image are used as condition to guide image generation. We remove the condition from different time steps during denoising process similar to Figure 10 in our paper. The generated images are given in the Rebuttal Figure 2 of attached file of General Response, and still preserve the specific details as baseline model (start point a=0) when start removing time steps is set to 20. Thus our sampling strategy is suitable for this task. Q4: Suggest the author to discuss the applicability and limitations of the proposed sampling scheme. For example, can it be applied to human face generation without losing human identity? A4: Following your suggestion, we conduct an experiment on the human face generation task using PhotoMaker [2] to verify our sampling strategy are still applicable. Unlike the T2I task, this experiment includes both text prompts and reference faces as condition information. We applied the proposed sampling strategy (removing all condition information) and find that the generated images, including faces (a=20), are similar to those produced by the baseline method (a=0). The generated images similar to Figure 10 in our paper are in the Rebuttal Figure 4 of the attached file of General Response. References: [1] Chen et al., 2023. AnyDoor: Zero-shot Object-level Image Customization [2] Li et al. 2024. Photomaker: Customizing realistic human photos via stacked id embedding. --- Rebuttal Comment 1.1: Comment: Thanks for your responses. After reading the rebuttal, I think this paper provides interesting new points about how text prompts work in diffusion models. Thus, I will raise my point.	Summary: The paper investigates the denoising process in DPM, identifying that the overall shape of the image is formed early in the process while details are added later. It further examines the influence of different text prompt tokens, finding that the end-of-sequence token [EOS] plays a crucial role in shaping the initial stages of image generation. The authors propose a method to speed up the generation process by removing text guidance after the initial stages, achieving a significant reduction in computational cost. Strengths: - Comprehensive analysis of the denoising process stages in DPM. - Detailed exploration of the influence of different tokens in the text prompt. - Practical application of findings to accelerate the T2I generation process. - Empirical and theoretical support for the proposed acceleration method. Weaknesses: - The paper might lack clarity in explaining the theoretical aspects of frequency signal analysis. - Limited exploration of potential biases introduced by the dominance of the [EOS] token. - The study may benefit from a broader range of experiments to validate the generalizability of the findings. Technical Quality: 4 Clarity: 3 Questions for Authors: - Can you provide a more detailed explanation of the theoretical aspects of frequency signal analysis used in your study? Specifically, how do the low and high-frequency components influence the denoising process? Including more accessible explanations or visual aids to illustrate the frequency signal analysis could help readers better understand this aspect of your work. - Your experiments are primarily based on a specific set of text prompts and Stable Diffusion model versions. How do you ensure that your findings generalize across different models and broader text prompt sets? - The paper uses various metrics like CLIPScore, BLIP-VQA, and MiniGPT4-CoT for evaluation. Can you provide a more detailed explanation of why these particular metrics were chosen and how they comprehensively assess the text-image alignment? Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: - Authors should discuss the robustness of their findings and the need for further experiments across various models and more complex or diverse text prompts to validate their conclusions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable comments. Here we address your concerns as follows. Q1: The paper might lack clarity in explaining the theoretical aspects of frequency signal analysis. A1: The theoretical aspects of frequency signal are in Proposition 1, where we have actually proved that the added standard Gaussian noise has almost the same magnitude under each frequency spectrum. This means Gaussian noise has more high frequency signals compared with low frequency signals (clarified in line 172 and Figure 2b), since the high-frequency parts contain 80% spectrum. In a word, the theoretical aspects of frequency signal analysis Proposition 1 show the added noise mainly contain high-frequency part, so that the high-frequency part of noisy data will be quickly destroyed during the adding noise process, and will not be recovered until the end of the reverse denoising process. The conclusion is oppositely hold for low-frequency part. Altogether, it explains the “first overall shape then details” phenomenon. Q2: “Limited exploration of potential biases introduced by the dominance of the [EOS] token.” A2: In line 215, we said “an ablation study in Appendix C verifies the influence to the dominance of the [EOS] token.” In Appendix C, we verify the impact of [EOS] partially originates from its dominance number, but also the more information contained in it. Q3: “The study may benefit from a broader range of experiments to validate the generalizability of the findings.” A3: In this paper, we verify the conclusion on simple and complicated text prompts and evaluated on T2I models based on UNet Stable Diffusion 1.5, 2.1 and DIT based model Pixel-Art. To further address your concerns, we conduct experiments on two tasks: subject driven generation and human face generation [1,2]. For such two tasks, we verify whether the conclusion “textual information is conveyed in the first few steps”. The results are in the Rebuttal Figure 2 and Rebuttal Figure 4 in the attached file of General Response. Please refer A3, A4 to Reviewer xbNR for more details. Besides, as you suggested, we further verify the conclusion “[SOS] contains more information” on dataset MS-COCO. Please see A5 for details. Q4: “How do the low and high-frequency components influence the denoising process? Including more accessible explanations or visual aids to illustrate the frequency signal analysis could help readers better understand this aspect of your work.” A4: How does low and high-frequency components influence the denoising process is clarified in A1, such that the add-noise process takes noise with more high-frequency components so that the high-frequency parts in original image will quickly destroyed and will not be recovered until the end of the reverse denoising process. The conclusion is oppositely hold for low-frequency parts. This phenomenon is also visualized in Figure 2b. We will make this more clearly and readable in the revised version. Q5: “Your experiments are primarily based on a specific set of text prompts and Stable Diffusion model versions. How do you ensure that your findings generalize across different models and broader text prompt sets?” A5: As mentioned in A3, we further verify the conclusion that text prompts convey information at the first few steps of denoising process. It shows the conclusion can be generalized to the other tasks. The models we used in this paper includes three T2I generative models (Stable Diffusion 1.5, 2.1 and Pixel-Art). The three models are SOTA of open-sourced diffusion-based generative models. As for text prompts, the results in Section 5 are verified on the constructed PromptSet in line 120, which consists of prompts from a benchmark dataset T2I commonbench [3] used to verify the quality of T2I generation. Notice that the prompts contain 1000 natural complex text prompts generated LLM. The prompts are actually diverse. Besides, the experiments for verifying our second conclusion “The information of text prompt is conveyed during the early stage of denoising process.” in Section 6 are conducted under 30K text prompts from MS-COCO, which are diverse enough, and consist with our conclusion. To further address your concerns, we conduct additional experiments on 15K pairs of text prompts from MS-COCO dataset with switched [EOS] to further support the conclusion in Section 5, “[EOS] Contains More Information”. The experiment is similar conducted as in Section 5.1 (with switched [EOS]), the results are shown in Table 1, and the generate images are in the Rebuttal Figure 3 in the attached file of General Response. The experimental results show the conclusion is still hold for MS-COCO dataset. \| \| Source Prompt \| Target Prompt \| \|----------------\|---------------\|---------------\| \| Text-CLIPScore \| 0.2086 \| 0.2696 \| \| BLIP-VQA \| 0.3735 \| 0.5655 \| \| MiniGPT-COT \| 0.6512 \| 0.7479 \| Q6: “Can you provide a more detailed explanation of why CLIPScore, BLIP-VQA, and MiniGPT4-CoT are chosen as metrics?” A6: As mentioned in line 206, we give an explanation to the three chosen metrics in Appendix B, where we have described these metrics. The three metrics are proposed in [3], which fully utilized the strong text-to-image alignment capability of multi-modality models CLIP, BLIP, MIiniGPT4. They are standard metrics in measuring text-image alignment as mentioned in [3]. Please see more details in Appendix B or [3]. References: [1] Chen et al., 2023. AnyDoor: Zero-shot Object-level Image Customization [2] Li et al. 2024. Photomaker: Customizing realistic human photos via stacked id embedding. [3] Huang et al., 2023. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation.	Summary: This paper study how the EOS token plays a role in the generation process of diffusion model. In particular, this paper finds that diffusion models tend to first generate low frequency part of the image at the beginning of the generation process, then gradually add high frequency signal to it. Experiments show that the low frequency signal is conditional on the EOS token while the high frequency signal can be generated without text guidance. In combined with the aforementioned observation, this paper proposes to remove $\epsilon_\theta$ in classifier-free guidance once the low frequency signal has been generation to improve generation efficiency. Strengths: - This paper offers a new perspective for understanding the role of textual condition in diffusion models. By exploring how the EOS influence the generation process of diffusion model, this paper argues that the conditional part $\epsilon_\theta$ in classifier-free guidance (CFG) might be unnecessary after certain denoising step $t_w$. - Most experiments are inspirational and interesting. By swapping the EOS token and the sentence body, it demonstrates that diffusion models rely on the EOS token to synthesize low frequency part of the image. - This paper explains the tendency of generating image from low-to-high frequency in diffusion models. Weaknesses: - It is not clear that how the "computational cost" is defined in this paper. If the computational cost is GPU VRAM, then the claimed efficiency improvement might be invalid, as the required GPU VRAM for computing $\epsilon_\theta(x_t, C)$ or $\epsilon_\theta(x_t, \emptyset )$ is unchanged. - This paper mainly focus on the role of EOS token in T2I diffusion models while neglecting the SOS token. Despite the weight of SOS token is significantly higher than SEM and EOS token (see Figure 3). However, the authoer(s) claims that the SOS carries no information due to the autoregressive nature CLIP text encoder. Since this claim is not yet supported by other works, the author(s) should have conducted experiments to support this claim, as there is a chance that EOS and SOS tokens altogether influence the generation process. Technical Quality: 3 Clarity: 2 Questions for Authors: - Please clarify how the computation cost is defined in this paper and how the efficiency gain is computed. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable comments. Here we address your concerns as follows. Q1: “It is not clear that how the "computational cost" is defined in this paper. If the computational cost is GPU VRAM, then the claimed efficiency improvement might be invalid, as the required GPU VRAM for computing $\epsilon_{\theta}(x_{t}, C)$ or $\epsilon_{\theta}(x_{t}, \emptyset)$ is unchanged.” A1: In this paper, the “computational cost” we mentioned is used for computing the noise prediction $\epsilon_{\theta}(t, x_{t}, C, \emptyset) = \epsilon_{\theta}(t, x_{t}, C) + w(\epsilon_{\theta}(t, x_{t}, C) - \epsilon_{\theta}(t, x_{t}, \emptyset))$ ($w > 0$) defined in (3), to conduct the diffusion process. Clearly, to get $\epsilon_{\theta}(t, x_{t}, C, \emptyset)$, one need two number of model evaluations on $\epsilon_{\theta}(t, x_{t}, C)$ and $\epsilon_{\theta}(t, x_{t}, \emptyset)$. However, as we have empirically verified that text prompt has conveyed their information after a certain time step $t_{w}$, then for $t > t_{w}$, we suggest to substitute $\epsilon_{\theta}(t, x_{t}, C, \emptyset)$ with $\epsilon_{\theta}(t, x_{t}, \emptyset)$ (as in (9)), which means for diffusion steps $t > t_{w}$, we only one number of model evaluation to conduct diffusion step, which reduces the computational cost (no matter measured under which metric), compared with the original noise prediction requires two number of model evaluations. To see the saved computational cost clear, please check “Saved Latency” in Table 2, where we report the saved latencies of our method under different settings, compared with the original method. Q2: “This paper mainly focuses on the role of EOS token in T2I diffusion models while neglecting the SOS token. Despite the weight of SOS token is significantly higher than SEM and EOS token (see Figure 3 in attached file of General Response). However, the author(s) claims that the SOS carries no information due to the autoregressive nature CLIP text encoder. Since this claim is not yet supported by other works, the author(s) should have conducted experiments to support this claim, as there is a chance that EOS and SOS tokens altogether influence the generation process.” A2: Thank you for pointing out this. First, due to the auto-regressive encoding process of CLIP, [SOS] should contain no textual information, while serves as a “dummy variable” to adjust weights of the other tokens in cross-attention module [1]. Moreover, following your suggestion, we design an experiment to verify that [SOS] indeed contains no textual information. For Text-to-Image generation, given a text prompt, we constructs two prompts, 1) all 77 tokens are [SOS] from the given text prompt, 2) expected for the first [SOS] (used to adjust attention map), the other all 76 tokens are [EOS] from the given text prompt. For such two constructed prompts, they respectively only contain textual information from [SOS], and [EOS]. We generate images under these prompt, and the generated results are in the Rebuttal Figure 1 in attached file of General Response. As can be seen, the images generated with information from [EOS] can be consistent with the target text prompt. However, this phenomenon does not happen for prompt with only information from [SOS] injected. This further verify our conclusion that [SOS] only adjusts the cross-attention map but contains no textual information in line 193. Thanks for inspiring us for such experiment, and we have added this in the revised version. Q3: “Please clarify how the computation cost is defined in this paper and how the efficiency gain is computed.” A3: Please check A1. References: [1] Xiao et al., 2023. Efficient streaming language models with attention sinks. --- Rebuttal Comment 1.1: Title: Looking forward for your feedback Comment: Dear reviewer: Thanks for your reviewing, we are very happy to see your feedback, and address your further concerns. Thanks --- Rebuttal 2: Title: Looking forward to Feedback as Discussion Deadline Approaches Comment: Thank you for your meticulous review, which played a pivotal role in enhancing the quality of our paper. We apologize for any inconvenience, but as the discussion deadline approaches (Aug 13 11:59 pm AoE), we would like to provide an update on our progress. We have clarified that the computational effciency of the proposed sampling method originates from the less number of function evaluation, and the experiments to verify no information in [SOS] are conducted. If you require any further clarification or have additional questions, please do not hesitate to reach out. Again, we sincerely appreciate your time and effort in reviewing our paper. Thanks --- Rebuttal 3: Title: Rebuttal Comment: Thanks for the responses during the rebuttal period. The authors have addressed my major concerns. I am happy to increase my ranking of this manuscript as Weak Accept. --- Rebuttal Comment 3.1: Title: Thanks Comment: Dear Reviewer We are happy that our reply address your concerns. We are sorry to bother you, but It seems that the score has not been changed yet. It will be nice to change the score when you are in convenience. Thanks	Rebuttal 1: Rebuttal: General Response: We thank all reviewers for their valuable comments. It seems a common question is whether our sampling strategy can be applied to the other conditional generation tasks. To verify this, we further apply our sampling strategy to the other two conditional generation tasks: subject-driven generation and human face generation. For such two tasks, there is extra reference image (given subject and human face) used as condition to guide image generation. Our sampling strategy are conducted on backbone methods AnyDoor [1] and Photomaker [2] respectively. The generated results of these two tasks are in the Rebuttal Figure 2 and Rebuttal Figure 4 in attached file, and the results show that our sampling strategy is still applied for such two tasks. Because the images generated with conditions removed in the final stage of diffusion are consistent with baseline method (similar to Figure 10 in our paper). The Figures for extra experiments are in the attached file. References [1] Chen et al., 2023. AnyDoor: Zero-shot Object-level Image Customization [2] Li et al., 2024. Photomaker: Customizing realistic human photos via stacked id embedding. Pdf: /pdf/1a72b302df9f68e54fd61cd9b10f0e1d8e764517.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Deep Correlated Prompting for Visual Recognition with Missing Modalities	Accept (poster)	Summary: The paper proposes a prompt optimization approach to the missing modality issues in multimodal learning. Inspired by the missing-aware prompt (MMP), this paper adds more prompts, including correlated, dynamic and modal-common prompts, to each encoder to improve the performance. The experiment on three datasets shows the effectiveness of the proposed method. Strengths: The missing modality issue in multimodal learning is a practical challenge. The designed method is clearly presented. Weaknesses: 1. The novelty of the proposed method is limited since the MMP has proposed the prompt optimization approach to solving the missing modality issue. Compared with MMP, this paper adds more parameters in the form of prompt tokens from different inputs and functions. 2. The empirical comparison with MMP is probably not quite fair as the proposed method uses more additional parameters compared with MMP. According to Line 337, this method adds 2.4% additional parameters, while MMP only adds 0.2%. Technical Quality: 3 Clarity: 3 Questions for Authors: What is the specific contribution of this paper compared with MMP, other than adding more parameters and functions? Will MMP's performance be better or comparable when MMP uses the same number of parameters as the proposed method? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Reviewer# YNgk 1. Novelty compared to MMP Many thanks for your question. MMP has first introduced prompt learning to handle the missing-modality setting. It inserts learnable tensors, i.e., prompts at each layer which still keeping the image encoder and text encoder fixed to guide the model to fit missing-modality cases. However, the inserted prompts of MMP across different layers are independent, while we believe that the prompts across different layers and various modalities can provide beneficial information for each other to better fit missing-modality cases. Thus, we propose correlated prompts which generate the prompts of the next layer based on the prompts of both modalities in the current layer. For dynamic prompts, our intuition is that the prompts proposed by MMP are fixed for different inputs during inference and fail to fit the missing cases of different inputs, and we thus propose to dynamically compute the prompts based on different input features to better guide the behavior of the model. This procedure is implemented by a self-attention layer with a randomly initialized tensor as a query and the input features as keys and values. Besides, we propose modal-common prompts which store the shared information across different modalities, which can complement the model with common information across different modalities and facilitate the model to encode modal-specific information to better handle the missing scenarios in each modality. 2. Performance v.s. MMP when owning comparable parameters We compare our method with MMP by allowing them to own comparable parameters. Specially, either when we decrease the required parameters of our method to 0.2M (0.2% of the entire model) by reducing the channel dimension, or we simultaneously increase the required parameters of both methods, our method consistently outperforms MMP, which verifies its effectiveness. \| Parameters\| 0.2M \| 1.6M \| 2.8M \| 4.0M \| 5.2M \| 6.4M \| \| :---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| MMP \| 51.34%\| 50.74% \| 50.82% \| 50.72% \| 50.64% \| 50.46%\| \| Ours \| 53.14% \| 53.56% \| 53.88% \| 54.24% \| 54.12% \| 53.98% \| --- Rebuttal Comment 1.1: Comment: Thanks for the response. My concern about the experimental comparison is addressed, so I will increase the score to 5. The novelty concern still holds but not a major flaw as the novelty concern is possibly not quite objective as it always is, so I wouldn’t fight for rejection.	Summary: The model proposes prompting strategy where both modalities (image and text) are prompted, and the prompt for both modalities are correlated. The strategy is to use multiple prompts, namely correlated prompts, dynamic prompts, and modal-common prompts. As the backbone itself is multimodal (CLIP), it is a good idea to consider synchronized multi-modal prompts to fully harness the model capabilities when prompting it. The model surpasses multiple multimodal SoTAs on multiple datasets and also has proven to be effective in handling missing modalities in training and inference. Strengths: 1. The strategy of using multiple types of multimodal prompts, along with the correlation strategy, is logically sound as the multimodal backbone itself is trained to understand the relationship between image and text modalities. 2. The modal surpasses multiple SoTAs on multiple benchmarks with considerable score improvement. 3. The ablation studies are sufficient to understand the justification of the network design. Weaknesses: 1. Ablation studies regarding the multimodal backbone, e.g. using other model than CLIP or use dedicated unimodal encoders for each modality, highly recommended to increase paper quality. 2. In table 4, what are the performances when either image or text modalities are completely missing? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Please elaborate further on how modal-common features are disentangled. 2. If possible, show the layer J-th in Figure 1 (framework overview) 3. Minor suggestion: The phrase "abundant ablations" in the introduction is a bit overboard, I suggest to write it as just "Ablation studies are further given..." Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations are discussed in the appendix, including that only text and visual modalities are tested with this model and the number of tested models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Ablation studies by using other multimodal backbones We provide the results by comparing our method with the baseline method upon the single-stream ViLT backbone, and also comparing them upon the two-stream CoCa backbone as below. We first provide the results on the ViLT backbone. Our method large outperforms the baseline across different settings on three datasets. \| Datasets \| Image \| Text \| MMP \| Ours \| \|:---\|:---:\|:---:\|:---:\|:---:\| \| MM-IMDb (F1-Macro) \| 100% \| 30% \| 39.22 \| 45.26 \| \| MM-IMDb (F1-Macro) \| 30% \| 100% \| 46.30 \| 51.24 \| \| MM-IMDb (F1-Macro) \| 65% \| 65% \| 42.66 \| 48.45 \| \| Food101 (Accuracy) \| 100% \| 30% \| 74.53 \| 78.85 \| \| Food101 (Accuracy) \| 30% \| 100% \| 86.18 \| 86.76 \| \|Food101 (Accuracy) \| 65% \| 65% \| 79.08 \| 80.85 \| \| Hateful Memes (AUROC) \| 100% \| 30% \| 59.11 \| 61.24 \| \| Hateful Memes (AUROC) \| 30% \| 100% \| 63.06 \| 64.12 \| \| Hateful Memes (AUROC)\| 65% \| 65% \| 66.07 \| 66.68 \| We offer the results on the CoCa backbone. We can find that our method achieves superior performance than the baseline. \| Datasets \| Image \| Text \| Baseline \| Ours \| \|:---\|:---:\|:---:\|:---:\|:---:\| \| MM-IMDb (F1-Macro) \| 100% \| 30% \| 38.96 \| 49.34 \| \| MM-IMDb (F1-Macro) \| 30% \| 100% \| 45.78 \| 55.12 \| \| MM-IMDb (F1-Macro) \| 65% \| 65% \| 42.35 \| 51.24 \| \| Food101 (Accuracy) \| 100% \| 30% \| 73.41 \| 77.56 \| \| Food101 (Accuracy) \| 30% \| 100% \| 82.34 \| 86.26 \| \| Food101 (Accuracy) \| 65% \| 65% \| 77.24 \| 80.46 \| \| Hateful Memes (AUROC) \| 100% \| 30% \| 54.87 \| 58.36 \| \| Hateful Memes (AUROC) \| 30% \| 100% \| 56.46 \| 60.34 \| \| Hateful Memes (AUROC) \| 65% \| 65% \| 57.82 \| 61.83 \| 2. Performance when either image or text modalities are completely missing in Table 4 We provide the results as follows. \| Datasets \| Image \| Text \| CoOp \| MMP \| MaPLe \| DePT \| Ours \| \|:---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| MM-IMDb (F1-Macro) \| 100% \| 0% \| 45.24 \| 46.32 \| 46.85 \| 48.02 \| 50.64 \| \| MM-IMDb (F1-Macro) \| 0% \| 100% \| 51.23 \| 52.21 \| 52.76 \| 54.13 \| 55.45 \| \| Food101 (Accuracy) \| 100% \| 0% \| 70.34 \| 71.52 \| 71.18 \| 72.25 \| 73.87 \| \| Food101 (Accuracy) \| 0% \| 100% \| 80.76 \| 81.52 \| 82.13 \| 83.24 \| 85.64 \| \| Hateful Memes (AUROC) \| 100% \| 0% \| 59.56 \| 60.14 \| 60.25 \| 61.32 \| 62.12 \| \| Hateful Memes (AUROC) \| 0% \| 100% \| 60.87 \| 61.32 \| 61.54 \| 62.12 \| 63.24 \| 3. How modal-common features are disentangled Specifically, we introduce a shared learnable prompt across different modality encoders which embeds modal-common information for different modalities. To transform it into the feature space of various modalities, we introduce an independent projection function for each modality to change the channels of the shared learnable prompt. It explicitly embeds modal-common information and inversely encourages other prompts to provide modal-specific information to guide the model to handle different missing-modality cases. 4. Show the layer J-th in Figure 1 (framework overview) Thanks for your comments. We will update the J-th layer in our manuscript. 5. Phrase correction. Many thanks for your advice, we will update the expressions. --- Rebuttal Comment 1.1: Comment: Thank you to the authors for addressing the questions. The authors have addressed most of my concerns, and I hope the answer can be further added in the final version of the paper either in the main paper or in the appendix. Overall, I stand with my initial rating of Weak Accept as the authors have demonstrated that their proposed method is competitive with logical explanation and proof to back it up.	Summary: This paper addresses the challenge of generalized missing modalities in multimodal learning, where a modality can be absent during any learning phase (e.g., training, testing, or both). he authors investigate prompt learning with missing modalities and propose deep correlated prompts designed to capture various types of correlations between prompts and input features across different modalities. Specifically, the proposed prompts include mechanisms for perceiving beneficial information from preceding layers, dynamically generating prompts based on input characteristics, and leveraging the complementary information from multimodal inputs. These designs improve the robustness of large multimodal models (e.g., CLIP) to missing modalities. Extensive experiments and ablation studies demonstrate consistently superior performance and verify the effectiveness of the proposed method. Strengths: 1. This paper addresses a more challenging missing modality setting, where modalities may be absent during both training and testing phases, making it highly practical and essential for real-world applications. 2. The paper is well-motivated. The authors highlight the weaknesses of prior work and propose several designs (e.g., deep correlated prompts, dynamic prompts, common prompts) to improve robustness. 3. The paper explores various types of correlations between prompts and input features across different modalities, and the proposed designs for each are technically sound. 4. Extensive experiments show great improvement on the baseline and consistently superior performance compared to other methods across all benchmarks. 5. Comprehensive ablation studies are conducted to validate the effectiveness of each proposed component. Weaknesses: 1. The paper lacks a detailed explanation or discussion on the efficacy of different prompt designs. In Figure 2, it shows that sequentially adding different designs improves the baselines, but it does not discuss the individual improvement gains for each design. Additional discussion on each design could help validate whether the increasing gains from sequentially adding designs are not merely due to more learnable parameters. 2. The paper lacks visualization of each learnable prompt (e.g., deep correlated prompts, dynamic prompts, and common prompts). Visualizations could help validate whether the different components work as expected. For example, do dynamic prompts genuinely capture the different characteristics of inputs, or do they merely distinguish between different missing cases, which might be easier to learn due to the obvious absence of a modality? 3. For each available modality, it seems there are a total of $(3*(2^M-1))$ prompts for each missing modality case. This could lead to an exponential increase and redundant prompts as more modalities are considered (i.e., M>2). For example, in a vision-and-language task, in the case of complete and missing-image, the text modality is available for both cases. However, it requires two separate prompt sets for the text encoder, which may actually learn the prompts for the same “text-available” case. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. In Table 4, I noticed that some values of the related work MMP are the same as the figures recorded in the paper. For example, the settings with: - missing rate = 70% (100% image and 30% text) in MM-IMDb, - missing rate = 70% (30% image and 100% text) in Hateful Memes, - missing rate = 70% (65% image and 65% text) in Hateful Memes. As far as I know, the MMP backbone model is the multimodal transformer ViLT. The authors state they re-implemented MMP on their setting (i.e., CLIP) for a fair comparison. It seems that the numbers should not be the same since they use different backbone models. Can the authors clarify why the values are identical despite using different backbone models? 2. According to the design of prompts, it seems that the proposed method is not limited to two-stream models (e.g., it could be applied to single-stream models without using Eq. (5)). Generalizing the method to single-stream models and comparing it with related works could be helpful in verifying the generalizability of the proposed method. Have the authors tried it for single-stream models? If so, what were the results? 3. I am willing to revise my rating if the authors also address the concerns mentioned in the weaknesses. Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: One limitation is that the proposed method requires modality-specific deep correlated prompts for each available modality, which could be challenging to extend to more modalities (e.g., five or more modalities). Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Efficacy of each proposed prompt. We place each proposed prompt upon the baseline method, and show the results as below on the MMIMDb dataset upon the missing-both setting with η=70%. It’s observed that each proposed prompt could notably boost the performance. \| Configurations\| Extra brought parameters \| F1-Macro\| \| --- \| --- \| --- \| \| Baseline \| - \| 49.21% \| \| +Correlated prompts \| +1.2M\| 52.12% \| \| +Dynamic prompts \| +1.8M \| 51.04% \| \| +Modal-common prompts \| +1.0M \| 51.26% \| 2. Visualization of each learnable prompt. We offer visualizations for the dynamic prompts using the T-SNE method on the Food101 dataset upon the missing-both setting with η=70% and η=50%. As Food101 have 101 classes which is redundant for visualization, we select a subset with 11 classes to give better visualization results. It’s observed that the dynamic prompts could roughly categorize the prompts into distinct classes. Besides, for different missing settings with η=70% and η=50%, the distribution of dynamic prompts with respect to the same inputs are different, which show that dynamic prompts can learn to generate various expressions for each input upon different missing settings. 3. Exponential increased prompts when more modalities are introduced. For M input modalities, it requires $2^M-1$ types of prompts for all missing-modality cases. This may be redundant, but we expect a specific prompt to handle each missing-modality case to better guide the model to handle different missing scenarios. Besides, the overall extra consumed parameters are quite few as the modalities increase as the projection functions are shared across different modalities. Thus as the modalities increase, we only introduce new learnable prompts which occupy quite few parameters. For example, for each missing-modality case, the newly introduced prompts own 76836=27648≈0.03M parameters with a prompt length of 36 for each modality encoder. 4. Identical values in Table 4 compared with MMP Sorry for the errors. We wrongly set the values when organizing the tables from different resources. The correct values should be 48.23 for missing rate = 70% (100% image and 30% text) in MM-IMDb, 61.12 for missing rate = 70% (100% image and 30% text) in Hateful Memes, and 63.24 for missing rate = 70% (30% image and 100% text) in Hateful Memes. We will current them in the manuscript. 5. Results on single-stream backbones We compare our method with MMP based on the ViLT backbone. We don’t use features of two modalities to generate prompts of the next layer as Eq.5. The results on three datasets with η=70% upon different missing settings are shown below. It’s observed that our method shows superior performance than MMP. \| Datasets \| Image \| Text \| MMP \| Ours \| \|-------------------------\|-------\|-------\|---------\|---------\| \| MM-IMDb (F1-Macro) \| 100% \| 30% \| 39.22 \| 45.26* \| \| MM-IMDb (F1-Macro) \| 30% \| 100% \| 46.30 \| 51.24 \| \| MM-IMDb (F1-Macro) \| 65% \| 65% \| 42.66 \| 48.45 \| \| Food101 (Accuracy) \| 100% \| 30% \| 74.53 \| 78.85 \| \| Food101 (Accuracy) \| 30% \| 100% \| 86.18 \| 86.76 \| \|Food101 (Accuracy) \| 65% \| 65% \| 79.08 \| 80.85 \| \| Hateful Memes (AUROC) \| 100% \| 30% \| 59.11 \| 61.24 \| \| Hateful Memes (AUROC) \| 30% \| 100% \| 63.06 \| 64.12 \| \| Hateful Memes (AUROC)\| 65% \| 65% \| 66.07 \| 66.68 \| --- Rebuttal Comment 1.1: Title: Response for Reviewer 64zT Comment: Dear reviewer, thanks for your time and efforts in reviewing our manusript. We have provided a point-to-point response regarning your concerns and we are looking forward to receiving your valuable feedback on the points we addressed in the response. If you have further concerns, place let us know and we will respond to you as soon as possible. Thank you for your dedication to the review process. Sincerely, Authors	Summary: This paper proposes to address the missing modality problem for the multimodal recognition model (i.e. the multi-modal data could be incomplete). There are three techniques of prompting being proposed (while the recognition model, i.e. two-stream multimodal method CLIP in this paper, is kept fixed), including: 1) correlated prompts, where a part of the prompts in the input-level are firstly selected according to the missing scenario (e.g. complete, text-only, or image-only), then the prompt in each of the following network layers are predicted from the multimodal prompt of its preceding layer; 2) dynamic prompts, the input-level prompts contain a portion generated according to the input sample; 3) modal-common prompts, where the rest of the input-level prompts is stemmed from a common component shared across modalities. The combination of the aforementioned three techniques experimentally shows better performance in comparison to various baselines (mainly the SOTA method from MMP [17]). Strengths: + The proposed method provides superior performance with respect to various baselines and its proposed techniques (i.e. correlated prompts, dynamic prompts, modal-common prompts) are experimentally shown to benefit the model performance. + The extensive experiments are conducted on multiple dataset with various experimental settings. + The presentation is clear and easy to follow. Weaknesses: - The modal-common prompts and the dynamic prompts actually are not directly connected to the missing modality problem (or being irrelevant to different cases of missing modality). While excluding these two prompting techniques from the proposed method (in which such variant becomes "Ours (A)" in Figure 2), the improvement with respect to the state-of-the-art approach of handling missing modality (i.e. MMP[17]) would become marginal (please include MMP[17] into the ablation study shown in Figure 2 or directly provide the tabular quantitative results for the ablation study). Similarly, while we only consider the technique of correlated prompts as the manner in the proposed to tackle the missing modality, it becomes the only difference in the proposed method compared to MMP [17] (in terms of methodology), thus leading to the concern of limited novelty. Furthermore, there should be a baseline of integrating the modal-common prompts (acting as a basic component of prompt) and dynamic prompts into MMP[17] to better highlight the contribution of the proposed correlated prompting technique (which is the main technique in the proposed method to be connected with the missing modality challenge). Moreover, as modal-common prompts and the dynamic prompts introduce additional learnable parameters (in comparison the correlated prompts), there should be further detailed analysis/comparison in terms of number of learnable parameters versus model performance. - Though the proposed dynamic prompts do experimentally shown to improve the overall performance under various missing modality cases, such prompting technique is actually not new, where we can see its similar application in various research problems (e.g. Wu et al., IDPG, NAACL'22; Lu et al., PromptPG, ICLR'23; Qiu et al., FedTPG, FL@FM-NeurIPS’23). Technical Quality: 3 Clarity: 3 Questions for Authors: Although currently the proposed method seems to provide superior performance with respect to various baselines and its proposed techniques (i.e. correlated prompts, dynamic prompts, modal-common prompts) are experimentally shown to benefit the model performance, there are concerns regarding limited novelty (where only the correlated prompts are considered to be related to missing modality while the other two techniques, i.e. dynamic and modal-common prompts, are not) and detailed analysis for the number of learnable parameters versus model performance, (as listed in the weaknesses), in which the the authors are highly encouraged to make the corresponding clarifications in the rebuttal. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: no potential negative societal impact is found. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Proposed prompts not directly connected to the missing modality problem. Sorry for the mis-clarification in the manuscript to mislead you. In line 146-150 of our manuscript, we state that we set different prompts for various missing modalities. Specifically, for correlated prompts, we independently set the initial prompt of the first layer for different missing-modality settings, which will be transformed via a projection function to generate the prompts for the next layer. For the dynamic prompts, we initialize different queries for different missing modalities, which are then used to compute the dynamic prompts based on input features in a self-attention manner. For the modal-common prompts, we initialize different modal-common prompts for different missing-modality settings, which will be transformed into embedding space via a projection function to provide prompts. Sorry for missing the expressions for the dynamic prompts and modal-common prompts upon different missing-modality settings. We will update them in the manuscript to give better clarifications. 2. Effectiveness regarding the correlated prompts compared to MMP. Based on the same CLIP backbone, we compare the performance of MMP to our proposed correlated prompts on the MMIMDb dataset with different missing ratios under the missing-both case as below. It’s observed that our proposed correlated prompts consistently offer superior performance than MMP under the same setting. \| Settings\| η=10% \| η=30%\| η=50% \| η=70% \| η=90% \| \| --- \| --- \| --- \| --- \| ---\| ---\| \| MMP \| 56.54%\| 53.64% \|52.12%\| 51.34% \| 48.04%\| \| Correlated prompts \| 57.91% \| 54.84% \| 53.78% \| 52.12% \| 48.52% \| 3. Comparison between the integration of MMP and ours. We integrate the proposed dynamic prompts and modal-common prompts into MMP based on the CLIP backbone to form a method termed as MMP, and compare our method with it on the MMIMDb dataset with different missing ratios under the missing-both case as below. It’s observed that our method notably outperforms MMP with different η. \| Settings\| η=10% \| η=30%\| η=50% \| η=70% \| η=90% \| \| --- \| --- \| --- \| --- \| ---\| ---\| \| MMP* \| 58.58%\| 55.46% \|54.16%\| 52.96% \| 50.04%\| \| Ours \| 59.82% \| 56.87% \| 55.45% \| 54.24% \| 51.26% \| 4. Analysis between parameters v.s. performance We first give an ablation for the parameters of each proposed prompt. We give the performance with η=70% on the MMIMDb dataset upon the missing-both setting. As shown below, as we insert each proposed prompt, the performance consistently increases. The proposed prompts overall bring extra 4.0M trainable parameters, which is only 2.4% of the overall framework. \| Configurations\| Brought parameters \| F1-Macro\| \| --- \| --- \| --- \| \| Baseline \| - \| 49.21% \| \| +Correlated prompts \| +1.2M\| 52.12% \| \| +Dynamic prompts \| +1.8M \| 53.22% \| \| +Modal-common prompts \| +1.0M \| 54.24% \| Besides, we test the relationships between the brought extra parameters and performance, and show the results as below. It’s observed that the performance continues to increase when the prompt depth ranges from 12 to 36, and reaches a peak when the prompt depth equals 36. The parameters consistently increase as the prompt length grows. \| Brought parameters \| 1.6M \| 2.8M \| 4.0M \| 5.2M \| 6.4M\| \| --- \| --- \| --- \| --- \| --- \| --- \| \| F1-Macro \| 53.56% \| 53.88% \| 54.24% \| 54.12% \| 53.98% \| Finally, we compare our method with MMP with the same parameters. It’s observed that with the same parameters, our method consistently achieves better performance. \| Parameters\| 0.2M \| 1.6M \| 2.8M \| 4.0M \| 5.2M \| 6.4M \| \| --- \| :---: \| :---: \| :---: \| :---: \| :---: \| :---: \| \| MMP \| 51.34%\| 50.74% \| 50.82% \| 50.72% \| 50.64% \| 50.46%\| \| Ours \| 53.14% \| 53.56% \| 53.88% \| 54.24% \| 54.12% \| 53.98% \| 5. The dynamic prompting technique is not new. Many thanks for your question. The dynamic prompts have been previous investigated in other methods. However, dynamic prompts in the missing-modality scenarios have not been studied. It’s worth exploring whether generating various prompts according to different missing cases and input features can well guide the model to fit the missing-modality settings. Our manuscript verifies the efficacy of this design. --- Rebuttal Comment 1.1: Title: Response for Reviewer xVBv Comment: Dear reviewer, thanks for your time and efforts in reviewing our manusript. We have provided a point-to-point response regarning your concerns and we are looking forward to receiving your valuable feedback on the points we addressed in the response. If you have further concerns, place let us know and we will respond to you as soon as possible.Thank you for your dedication to the review process. Sincerely, Authors	Rebuttal 1: Rebuttal: We provide (1) a figure to further illustrate out proposed three prompts by comparing them with our baseline and MMP[17]. (2) Visualizations for the dynamic prompts using the T-SNE method on the Food101 dataset upon the missing-both setting with η=70% and η=50%. Pdf: /pdf/86b7115517be30193e9978ba6977d34df08607d7.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper proposes a new method to handle missing modalities in visual and language recognition systems. The paper proposes a very similar method to the one proposed by MMP [17] but using different way of getting the prompts to feed them into the transformer layers. Comparison with other works show that the method seems to be effective and some ablations studies are performed to study the different design choices. The method is validated using the most common datasets for this task. Strengths: - The method seems to work when compared with other state-of-the-art models. - The paper presents results on several datasets and with different settings of the model. Weaknesses: - The main weakness of the paper is clarity. There are three different sets of prompts that are appended to the intermediate representations. However, the only difference between them seems to be the type of architecture the method uses to compute them. The explanation is very limited and Figure 1 does not illustrate where do these prompts come from. Without the clarity of this explanation it becomes really hard to understand how the motivation of each type of prompt fits the design. What are exactly correlated prompts, dynamic prompts, and modal-common prompts? What make them correlated, dynamic and modal-common? This is not clear in the paper at all. - It is not clear what is baseline. What does dropping features when modality is missing? The input sequence become shorter and coming from only a single modality? If that's the case, what is trainable and what is not? Please explain well this part. I would expect that this baseline is: training with the same number of parameters as the base method, by simply adding learnable prompts at each layer and training using mod-drop (dropping modalities randomly when training, dropping modalities can be done by inputting noise instead of tokens, the average of the modality tokens, zeroes, or not passing the missing modality at the input, it is a design choice that needs to be explained). If it is not what I'm thinking, please explain well, since this is a key experiment. - When comparing with MMP, how did the authors do it? Please explain exactly how was this re-implementation. Also, to be fair, the authors should have applied their method using ViLT instead of CLIP, in that way there is no doubt that this method is better than the very similar MMP. - What is the zero-shot performance of CLIP on these datasets? Technical Quality: 2 Clarity: 2 Questions for Authors: - Please explain well the mechanism of the different types of prompts, input, output at train and test time for each one of them. It could have been done easily with a figure, but at least with a few sentences it could become clearer. - What makes a "dynamic" prompt "dynamic"? - What does baseline mean and how was implemented? - How was MMP implemented on your framework? - What if using ViLT instead of CLIP, would still your method be better than MMP? - What is the zero-shot performance of CLIP on these datasets? it is important since this might be a robust method that does not suffer from missing modality. It can be implemented using nearest neighbor to each of the class embedding using either modality, and combining them when both are present. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Limitations have been addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. The mechanism of different prompts. Many thanks for your question. We have plotted a figure to further illustrate our proposed prompts by comparing them with our baseline and MMP[17], which can be found in the pdf file of Author Rebuttal. The baseline simply uses fixed image encoder and text encoder and only finetunes the classifier to handle downstream tasks. To well adapt to missing modalities, MMP[17] inserts learnable tensors, i.e., prompts at each layer which still keeping the image encoder and text encoder fixed to guide the model to fit missing-modality cases. However, the inserted prompts of MMP across different layers are independent, while we believe that the prompts across different layers and various modalities can provide beneficial information for each other to better fit missing-modality cases. Thus, we propose correlated prompts which generate the prompts of the next layer based on the prompts of both modalities in the current layer. For dynamic prompts, our intuition is that the prompts proposed by MMP are fixed for different inputs during inference and fail to fit the missing cases of different inputs, and we thus propose to dynamically compute the prompts based on different input features to better guide the behavior of the model. This procedure is implemented by a self-attention layer with a randomly initialized tensor as a query and the input features as keys and values. Besides, we propose modal-common prompts which store the shared information across different modalities, which can complement the model with common information across different modalities and facilitate the model to encode modal-specific information to better handle the missing scenarios in each modality. 2. What makes a "dynamic" prompt "dynamic"? We propose to dynamically compute the prompts based on different input features, which avoids employing fixed prompts for different input features during training and inference. This procedure is implemented by a self-attention layer with a randomly initialized tensor as a query and the input features as keys and values. You can view the illustration in the figure of Author Rebuttal.pdf. 3. The baseline setting. The baseline setting is using the text encoder and image encoder to encode the input texts and images, whose output features are fed into the classifier for recognition. In this procedure, only the classifier is updated and other model components including the text encoder and the image encoder are kept fixed. The only difference between our method and the baseline is inserting learnable prompts at each layer which only bring few extra parameters. Specifically, when a modality is missing, we simply don’t feed the corresponding features into the modality encoder for processing and set the outputs of this modality encoder as zeros. Overall, we have tried three baseline settings, including (1) inputting zeros for the modality encoder when a modality is missing, (2) inputting the average of the modality tokens for the modality encoder when a modality is missing, and (3) not feeding input features for the modality encoder when a modality is missing and setting the outputs of this modality encoder as zeros (Default). As shown below, out experiments on the MMIMDb dataset with η=70% show the last choice gives best performance, and thus we adopt it as a stronger baseline. \| Baseline Configurations \| Acc(%) \| \| --- \| :---: \| \| Inputting zeros \| 47.65 \| \| Inputting averaged tokens \| 46.24\| \| Default\| 49.21\| 4. The re-implementation setting of MMP. MMP adds learnable prompts with a length of 16 at the bottom 6 layers based on the ViLT backbone. we re-implement it based on the CLIP backbone by adding learnable prompts with a length of 16 at the bottom 6 layers in the image encoder and text encoder, respectively. Our method owns the same number of inserted layers and prompt length with MMP. The only difference is that we insert three proposed prompts into the model. 5. Comparison with MMP using the ViLT backbone We compare our method with MMP using the ViLT backbone and show the results on three datasets with η=70% upon the missing-both setting as below. It’s observed that our method shows superior performance than MMP. It’s also noticed that using ViLT as the backbone achieves inferior performance than CLIP, and thus we adopt CLIP as the default backbone. \| Dataset \| MMP\| Ours \| \| --- \| :---: \| :---: \| \| MMIMDb \| 42.66% \| 48.45% \| \| Food101 \| 79.08%\| 80.85% \| \| Hateful Memes \| 66.07% \| 66.68% \| 6. Zero-shot performance of CLIP. We test the performance of zero-shot CLIP on the MMIMDb, Food101 and Hateful Memes datasets. We first calculate and store the averaged embeddings of all classes of both modalities on each dataset. When the input modalities are complete, we calculate the similarities between the output features and the pre-computed class embeddings for each modality encoder. We select the class with the highest similarity as the recognition output within both modalities. When a modality is missing, we select the class whose embedding owns the highest similarity with the output features of the available input modality as the recognition output. We compare the performance of zero-shot CLIP, finetuned CLIP (our baseline, which sets output features as zeros when a modality is missing) and our method as below. The zero-shot CLIP achieves inferior performance than the other methods, and ours perform best. We suppose that the zero-shot CLIP not finetuned on the downstream datasets fails to well adapt to downstream setting. Finetuning the CLIP could notably increase the performance, and ours further boost the performance by injecting different kinds of learnable prompts. \| Datasets \| Zero-shot CLIP \| Finetuned CLIP \| Ours \| \| --- \| :---: \| :---: \| :---: \| \| MMIMDb \| 34.52% \| 49.21% \| 54.24%\| \| Food101 \| 57.02% \| 77.74% \| 82.38% \| \| Hateful Memes \| 55.23% \| 62.58% \| 66.08% \| --- Rebuttal Comment 1.1: Comment: Thanks to the authors for the thorough reply. I think the figure makes the contribution of the paper way more clear, and I suggest that it is included in the main manuscript. It is still not 100% clear to me what is baseline for you. I apologize for not being clear with my question in the first place. Fine-tuned CLIP in the last Table that you showed in the rebuttal is the same baseline model that you used for Figure 2 and Figure 3 in the main paper? Was this baseline trained with missing modalities at train time (as a sort of "augmentation"), i.e. modality dropping at train time? That's my only remaining question, in order for me to give my final evaluation on the paper. --- Reply to Comment 1.1.1: Title: Response for Reviewer nnhg Comment: Many thanks for your reply. We will try to save space to include the figures in our manuscript. For the baseline, the Fine-tuned CLIP in the last Table in the rebuttal is the same baseline model that we used for Figure 2 and Figure 3 in the main paper. The baseline is trained with the same missing modalities at training time to keep fair comparison with other methods. We also keep the identical training settings for other methods to make a fair comparison.	null	null	null	null	null	null
ResAD: A Simple Framework for Class Generalizable Anomaly Detection	Accept (spotlight)	Summary: The paper analyzes the class-generalizable anomaly detection problem and introduces residual feature learning. Based on the residual features, the paper proposes a simple AD framework, i.e., ResAD, which incorporates OCC loss and distribution estimating to distinguish normal and abnormal data. The experimental results demonstrate that the ResAD performs well on real-world industrial AD datasets. Strengths: 1. The paper analyzes the few-shot class generalizable anomaly detection problem and delivers an interesting insight into residual features. 2. The proposed method is intuitive and easy to understand. 2. The paper is well-written and organized. Weaknesses: 1. The residual learning for few-shot AD has already been proposed in inCTRL[1]. The proposed Multi-Layer Patch-Level Residual Learning scheme in InCRTL is more sophisticated and reasonable than the direct subtraction in this paper. 2. The results in Table 1 of InCTRL are not consistent with the results in the original paper. Compared with the original results of InCTRL, the ResAD results do not achieve the SOTA performance. 3. The paper aims to achieve generalization across different classes. I think the authors should compare the accuracy of each class on the Visa dataset with other methods to demonstrate the generalization capability of your approach for different classes, rather than taking the average accuracy of different classes in the dataset. [1]Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In CVPR, 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What is the superiority of the proposed simple subtraction-based residual learning compared with the residual learning in InCTRL? 2. In the related work, the author claims that the CLIP-based methods are difficult to generalize to anomalies in diverse classes. However, according to the experiment results, the proposed methods only perform well on industrial AD datasets while InCTRL performs well on various types of datasets, including Medical datasets and Semantic datasets. Why does your method only compare different classes on industrial datasets, instead of comparing against anomaly datasets from other domain? I am wondering how the ResAD performs on the datasets from other domains. 3. What are the main advantages of ResAD compared with WinCLIP and InCTRL since the generalization ability and complexity of ResAD are not as good as WinCLIP and InCTRL. 4. In the residual feature construction process, the residual feature is highly related to the closest normal reference features in the reference feature pool. Are the few-shot reference samples enough to represent the class-related attributes? 5. In table 1, the authors mention that RDAD and UniAD don't utilize the few-shot normal samples to fine-tune, so the results under 2-shot and 4-shot are the same. RDAD and UniAD don’t require few-shot normal samples to fine-tune or refer, while the proposed method provide few normal samples to refer, so I believe it is meaningful to compare your method with those that require few-shot normal samples to refer, such as inctrl and winclip. Comparing it with RDAD and UniAD seems to be unfair especially in Table 3 . How do the results of corporating proposed method into WinCLIP and InCTRL? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors did not give a discussion on the limitations of the proposed method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [To W1 and Q1]. Thanks for your professional review. We think that our work and InCTRL should be concurrent work. We also initially submitted our work to CVPR2024, and received two weak accepts and one reject (you can see the relevant materials in the rebuttal pdf file). Regretfully, our work was rejected at that time. So, we should have proposed the idea of residual learning independently of InCTRL and almost at the same time. But our method has obvious differences with InCTRL in the definition and utilization of residuals. In CVPR2024, our paper was rejected mainly because a reviewer thought our result comparison with unsupervised AD methods was unreasonable and unfair. In this submission, we mainly compare with few-shot AD methods and InCTRL. Please see our response to Reviewer fL7E's Weakness 1, we provide a detailed comparison between our method and InCTRL. Based on the comparison, we think that our method has the following advantages: (1) One main advantage of our method is that it can achieve image-level anomaly detection and also pixel-level anomaly localization, while InCTRL only achieves image-level anomaly detection (due to the designs in their method). (2) Compared to residual distance maps (in InCTRL), residual features are easier to integrate into other feature-based AD methods. For example, in Sec.4.4, we combine the residual feature learning with UniAD and RDAD, which can effectively improve the models’ class-generalizable capacity. In InCTRL, the authors devise an anomaly scoring network to learn the residual distance maps. The residual distance maps seem not easy to integrate into other AD methods (based on our analysis in the response to Reviewer fL7E’s Weakness 1). [To W2]. We checked the results again and found that the results of InCTRL we reported were the same as in its original paper (94.0 and 85.5 under the 2-shot setting, 94.5 and 87.7 under the 4-shot setting). In the InCTRL paper, there are two tables, the AUROC results are in Table 1, and Table 2 are the AUPRC results, which are overall higher than the AUROC results. The original results you mentioned may be from Table 2 of the InCTRL paper. In Table 1 of our paper, we report the AUROC results, the average results on four AD datasets of our method are also better than InCTRL’s. [To W3]. Thanks for your suggestion. Due to the page limitation and the need to extensively validate the effectiveness of our method on multiple datasets, we adopted to report the dataset-level average results, which was also the way adopted in the InCTRL paper. We present the detailed results on the VisA dataset in the rebuttal pdf file and also the following Comment. [To Q2]. When writing the paper, we thought the results on these four industrial AD datasets were enough to validate the effectiveness of our method (In Table 1, our method can achieve better average results on these datasets than other methods), so we didn't consider more datasets from other domains. Please see our response to Weakness 1 of Reviewer CQaQ, we further run experiments on a medical dataset and a video AD dataset. [To Q3]. According to our response to Weakness 2, our method is better than InCTRL and WinCLIP in terms of the average results on four AD datasets. The advantages to InCTRL are discussed in [To W1 and Q1]. Compared to WinCLIP, we think the advantages are: (1) WinCLIP generates anomaly score maps by directly calculating the similarity between vision and text features. The model is not trained on AD datasets. Thus, WinCLIP is more likely to rely on the visual-language comprehension abilities of CLIP. When text prompts cannot capture the desired anomaly semantics, the results may be poor. In addition, WinCLIP requires text prompts, which can bring extra complexity, while our method does not require any text. (2) Due to its sliding window mechanism, WinCLIP has low efficiency. In Table 4 in the Appendix, we provide the number of parameters and per-image inference time of our method and other methods (we list the table in the rebuttal pdf file and also the following Comment). The inference speed of WinCLIP is 0.51fps, while our method's is 18.8fps. [To Q4]. Please see our response to Weakness 3 of Reviewer 226V. [To Q5]. Yes, in Table 1, our method is mainly compared with these few-shot AD methods. Including the cross-dataset results of UniAD and RDAD is not for comparison, but to demonstrate that conventional one-for-one (RDAD) and also one-for-many (UniAD) AD methods cannot be directly applied to new classes (compared to the results in their original papers). Thus, achieving class-generalizable anomaly detection requires new insights and specific designs. In Table 3, we mainly aim to demonstrate that the residual feature learning can be easily incorporated into conventional feature-based AD methods and can effectively improve their class-generalizable capacity. We also considered combining our method with WinCLIP and InCTRL but found it not easy to achieve. In WinCLIP, the model is based on the alignment between vision and text features. As the semantics of residual features and initial features are different, converting to residual features can lead to misalignment with text features. In addition, our method is to learn the residual feature distribution, while WinCLIP doesn't have the training stage on AD datasets. In InCTRL, the model is designed based on residual distance maps. InCTRL devises an anomaly scoring network (discriminative model) to learn the residual distance map and convert it to an anomaly score. Our method is designed based on residual features, which utilizes a normalizing flow model (probabilistic generative model) to learn the residual feature distribution. The obvious differences in the definition and utilization of residuals between our method and InCTRL make integrating with InCTRL also not easy. If you still have any questions, we are very glad to further discuss with you. --- Rebuttal 2: Comment: [To W3]. Below, we present the detailed results on the VisA dataset under the 4-shot setting. \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Candle \| 51.9/63.0 \| 58.5/76.4 \| 73.0/58.2 \| 87.4/97.6 \| 79.9/92.9 \| 85.3/97.4 \| 89.8/99.1 \| 96.9/96.0 \| 93.7/- \| 93.3/99.3 \| \| Capsules \| 39.8/73.9 \| 57.3/74.6 \| 68.4/48.3 \| 55.8/78.3 \| 61.0/83.7 \| 59.8/75.3 \| 72.9/98.0 \| 83.5/96.1 \| 85.9/- \| 86.7/97.9 \| \| Cashew \| 64.4/80.4 \| 40.0/87.5 \| 85.8/81.6 \| 74.5/98.8 \| 88.7/87.1 \| 84.0/97.4 \| 93.4/98.2 \| 94.7/96.8 \| 97.8/- \| 95.7/98.0 \| \| Chewinggum \| 72.0/78.4 \| 47.2/88.9 \| 85.4/77.6 \| 94.9/97.5 \| 95.5/97.9 \| 93.8/98.0 \| 97.9/99.4 \| 97.8/99.0 \| 99.8/- \| 97.7/98.9 \| \| Fryum \| 72.6/87.3 \| 51.2/81.4 \| 77.6/77.4 \| 64.7/94.1 \| 68.9/83.4 \| 79.6/94.4 \| 86.7/92.5 \| 86.8/95.2 \| 96.7/- \| 94.7/92.8 \| \| Macaroni1 \| 54.1/86.9 \| 48.8/85.5 \| 65.8/62.4 \| 62.8/92.1 \| 62.7/76.6 \| 71.0/94.9 \| 88.3/99.4 \| 89.1/92.9 \| 77.6/- \| 91.4/98.7 \| \| Macaroni2 \| 53.9/91.7 \| 56.8/89.4 \| 45.2/45.1 \| 64.4/87.8 \| 61.2/75.8 \| 61.9/88.6 \| 77.6/98.0 \| 73.9/89.8 \| 74.6/- \| 77.6/96.9\| \| Pcb1 \| 51.9/70.5 \| 58.4/74.0 \| 89.3/56.3 \| 81.1/89.9 \| 80.8/94.2 \| 79.9/94.4 \| 83.7/97.6 \| 84.6/96.3 \| 95.9/- \| 90.5/98.8 \| \| Pcb2 \| 66.5/78.4 \| 47.8/75.5 \| 75.5/64.8 \| 69.2/93.8 \| 71.0/82.0 \| 73.8/94.1 \| 68.9/94.7 \| 61.1/92.3 \| 66.9/- \| 85.3/96.9 \| \| Pcb3 \| 58.4/82.0 \| 50.0/84.0 \| 75.0/54.4 \| 69.1/95.2 \| 57.2/91.2 \| 73.0/96.8 \| 81.4/96.7 \| 72.1/94.6 \| 76.1/- \| 84.6/95.7 \| \| Pcb4 \| 21.6/74.4 \| 62.5/72.5 \| 86.8/68.5 \| 91.6/94.8 \| 47.3/83.2 \| 82.0/92.3 \| 95.0/96.3 \| 76.2/96.9 \| 97.5/- \| 93.9/98.0 \| Pipe\_fryum \| 70.2/92.0 \| 46.7/91.4 \| 72.0/90.4 \| 88.7/99.2 \| 85.9/97.2 \| 92.1/98.9 \| 98.9/98.5 \| 92.0/96.5 \| 86.9/- \| 99.3/98.5 \| \| Average \| 56.4/79.9 \| 52.1/81.8 \| 75.0/65.4 \| 75.3/93.3 \| 71.7/87.1 \| 78.0/93.5 \| 86.2/97.4 \| 84.1/95.2 \| 87.7/- \| 90.8/97.5 \| The detailed results show that our method can achieve better results in most classes. In the revision, we will add the above detailed results and also detailed results of other datasets to the Appendix. [To Q3]. Below, we present the table of computation complexity. \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| WinCLIP \| InCTRL \| ResAD \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Parameters(M) \| 150.6 \| 6.3 \| 74.5 \| 686.9 \| 69.5 \| 25.2 \| 165.9 \| 117.5 \| 59.2 \| 116.4 \| \| Infer time(fps) \| 5.6 \| 24.4 \| 4.8 \| 14.1 \| 21.5 \| 20.2 \| 0.51 \| 0.53 \| 21.3 \| 18.8 \| [Limitations]. In Appendix Sec.B, we provide discussions about our method's limitations. In the revision, we will further discuss the limitations of our method more comprehensively and clearly based on the comments of all reviewers. --- Rebuttal Comment 2.1: Comment: Thanks for your response. Some responses have addressed part of my questions, but some issues still have not been well resolved. Regarding my Q4, based on the author's response that the reference feature pool is not representative, I think that this approach is not suitable for few-shot scenarios. Simply searching for the nearest nominal reference feature from the reference feature pool to achieve the learning goal of few-shot AD is insufficient. According to the results on the VisA dataset under the 4-shot setting and 2-shot setting and the table in the rebuttal for Reviewer fL7E, the proposed method does not show significant advantages and sometimes even performs worse than other methods. --- Reply to Comment 2.1.1: Comment: We greatly appreciate your further response. We hope the following response can answer your question. [To R1]. We respectfully argue that our response does not mention that the reference feature pool is not representative. We think the issue of representativeness you mentioned is not caused by our method, but caused by few-shot normal samples themselves. In our response, we state that ''when the difference between normal images is too large, it may cause the reference feature pool is not representative''. With this statement, we want to express that when the normal patterns of one product class are diverse and complex, the few-shot normal samples may lack some normal patterns (e.g., if one class has five colors and one image only contains one color, then 4 reference samples can only contain a maximum of four colors), thus they are not enough to represent the class they belong to. In our method, we only extract the features of the few-shot normal samples and store all the features in the reference feature pool. The reference feature pool does not impair or lose any representation features. For the few-shot normal samples, it is representative enough to them. Therefore, whether the representativeness is sufficient is determined by the few-shot normal samples. For some classes, the few-shot normal samples are representative, while for some hard classes, they may mot be representative enough. For example, in our response to Weakness 3, the results of Macaroni2 and Pipe\_fryum classes are as follows: \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Macaroni2 \| 53.9/91.7 \| 56.8/89.4 \| 45.2/45.1 \| 64.4/87.8 \| 61.2/75.8 \| 61.9/88.6 \| 77.6/98.0 \| 73.9/89.8 \| 74.6/- \| 77.6/96.9 \| \| Pipe\_fryum \| 70.2/92.0 \| 46.7/91.4 \| 72.0/90.4 \| 88.7/99.2 \| 85.9/97.2 \| 92.1/98.9 \| 98.9/98.5 \| 92.0/96.5 \| 86.9/- \| 99.3/98.5 \| For the Pipe\_fryum class, the few-shot reference samples are representative enough. For the Macaroni2 class, the few-shot reference samples are not representative enough, and this issue exists for all AD methods that use few-shot normal samples, as it is limited by the data rather than the method itself. Thus, this issue should be addressed by the data perspective, in practical applications, when few-shot normal samples are insufficient to represent their class, we can increase the number of reference samples (or use the method we stated in the response to Reviewer 226V's Weakness 3). However, for method comparison, the few-shot setting is feasible (InCTRL also follows this setting). Because we ensure that all methods use the same reference samples, all methods can obtain the same representation information, the result comparison is reasonable and fair. Matching the nearest reference features is to generate residual features, InCTRL also employs this way to generate residuals (please see our response to Reviewer fL7E's Weakness 1), which also indicates that this way is reasonable and effective for residual generation. We think that besides residual features, the other parts of our method also play a crucial role, e.g. the Feature Constraintor and the Abnormal Invariant OCC loss (please see Table 2(a) in our paper). In addition, we note that in SPADE, each test feature will search for the nearest normal feature and directly calculate the distance between the two features as the anomaly score. By comparison, our method can significantly outperform SPADE, this also indicates that residual feature learning (v.s. only searching for the nearest normal features) is more effective for utilizing few-shot normal samples. --- Reply to Comment 2.1.2: Comment: [To R2]. The results on the VisA dataset under the 4-shot setting and 2-shot setting are as follows (from Table 1 in the paper): \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| &emsp; &emsp; \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| 2-shot \| 56.4/79.9 \| 52.1/81.8 \| 71.7/65.4 \| 68.7/91.5 \| 65.0/80.4 \| 70.6/93.3 \| 79.9/96.4 \| \| 81.9/94.9 \| 85.8/- \| 84.5/95.1\| \| 4-shot \| 56.4/79.9 \| 52.1/81.8 \| 75.0/65.4 \| 75.3/93.3 \| 71.7/87.1 \| 78.0/93.5 \| 86.2/97.4 \| \| 84.1/95.2 \| 87.7/- \| 90.8/97.5\| We note that for a fair comparison, we should compare ResAD with its front methods (with the commonly used WideResNet50 as the feature extractor) and compare ResAD$^{†}$ with WinCLIP and InCTRL (all use ViT-B/16+ as the feature extractor). The results show that only the image-level AUROC under the 2-shot setting is lower than InCTRL, but all other results are better. Under the 4-shot setting, our ResAD$^{†}$ can significantly outperform InCTRL by 3.1\%. In addition, InCTRL only achieves image-level anomaly detection, while our method can achieve image-level anomaly detection and also pixel-level anomaly localization (please see our response to Reviewer fL7E's Weakness 1). Our ResAD$^{†}$ also significantly outperforms WinCLIP by 6.7\%/2.3\% under the 4-shot setting. The results of our response to Reviewer fL7E are as follows: \| \| FastRecon \|ResAD \| &emsp; &emsp; \| AnomalyGPT \| AnomalyGPT (ViT-B/16+) \| InCTRL \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|--- \| \| VisA \| 82.7/94.1 \|86.2/97.4 \| \| 91.3/88.8 \| 85.4/86.9 \| 88.7/- \| 90.8/97.5 \| Although the image-level AUROC is slightly lower than AnomalyGPT by 0.5\% (please note that the original image encoder in AnomalyGPT is significantly larger than the ViT-B/16+ used in InCTRL and our ResAD$^{†}$), our method can significantly outperform AnomalyGPT by 5.4\%/10.6\% when using the same feature extractor (AnomalyGPT (VIT-B/16+) v.s. ResAD$^{†}$). Moreover, our method's pixel-level AUROC is significantly higher than AnomalyGPT. We think that evaluating the effectiveness of a method should not only focus on one dataset, but on multiple datasets. On the MVTec3D dataset, our method significantly outperforms other methods, the results are as follows. \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| &emsp; &emsp; \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| 2-shot \| 58.7/90.4 \| 51.7/89.4 \| 62.5/78.6 \| 59.6/94.3 \| 58.8/83.4 \| 59.5/96.4 \| 64.5/95.4 \| \| 74.1/96.8 \| 68.9/- \| 78.5/97.5\| \| 4-shot \| 58.7/90.4 \| 51.7/89.4 \| 62.3/78.6 \| 62.8/94.5 \| 61.5/87.1 \| 62.3/96.7 \| 70.9/97.3 \| \| 76.0/97.0 \| 69.1/- \| 82.4/97.9\| \| \| FastRecon \| ResAD \| &emsp; &emsp; \| AnomalyGPT \| AnomalyGPT (ViT-B/16+) \| InCTRL \| ResAD$^{†}$ (Ours) \| \| --- \| ---\| --- \| --- \| --- \| --- \| --- \| --- \| \| MVTec3D \| 66.5/95.2 \| 70.9/97.3 \| \| 81.7/96.5 \| 75.3/96.2 \| 69.1/- \| 82.4/97.9\| In our response to Weakness 1 of Reviewer CQaQ, we also further evaluate our method on a medical image dataset, BraTS and a video AD dataset, ShanghaiTech to validate the cross-domain generalization ability of our method. We list the results as follows (you can also see our response to Reviewer CQaQ's Weakness 1): \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| &emsp; &emsp; \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ---\| \| BraTS \| 50.2/58.7 \| 57.3/85.1 \| 67.1/93.6 \| 63.4/93.1 \| 70.5/94.6 \| 62.1/89.4 \| 74.7/94.0 \| \| 68.9/93.5 \| 76.9/- \| 84.6/96.1\| \| ShanghaiTech \| 56.2/77.6 \| 55.9/79.4 \| 77.1/87.4 \| 74.3/85.9 \| 77.8/88.2 \| 76.4/87.7 \| 79.8/89.5 \| \| 79.6/88.6 \| 69.2/- \| 84.3/92.6\| The results show that when applied to medical images and video scenarios, the cross-domain generalization ability of our method is also more superior to other methods. Therefore, we think that based on the results of multiple datasets, our method is overall superior to other methods. Thank you again for your further response. We hope the above discussions can solve your concerns. If you still have any questions, we sincerely wish to further discuss with you. --- Rebuttal 3: Comment: Thanks for your further response and provide more empirical results. The response addresses most concerns. However, I noticed that the results were based on new medical datasets. How about the results on the headct and brainmri datasets demonstrated in InCTRL? I am curious whether the proposed ResAD is sensitive to the dataset. --- Rebuttal Comment 3.1: Comment: We greatly appreciate your further response. We use the BraTS dataset because it provides ground-truth masks, while the BrainMRI and HeadCT datasets do not have pixel-level annotations, thus pixel-level AUROCs cannot be measured. Under the 2-shot and 4-shot settings, we further run our method and evaluate it on the BrainMRI and HeadCT datasets. The comparison (image-level AUROC) with InCTRL and also with the other results in its paper are as follows: \| \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| &emsp; \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| ---\| ---\| ---\| ---\| ---\| ---\| ---\| ---\| ---\| ---\| \| BrainMRI(2-shot) \| 75.4 \| 65.7 \| 70.6 \| 44.9 \| 92.3 \| &emsp; \| 93.4 \| 97.3 \| 97.0\| \| BrainMRI(4-shot) \| 75.9 \| 79.2 \| 79.4 \| 57.1 \| 93.8 \| &emsp; \| 94.1 \| 97.5 \| 97.9\| \| \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| &emsp; \| WinCLIP \| InCTRL \| ResAD$^{†}$ \| \| ---\| ---\| ---\| ---\| ---\| ---\| ---\| ---\| ---\| ---\| \| HeadCT(2-shot) \| 64.5 \| 59.5 \| 73.6 \| 60.2 \| 90.4 \| &emsp; \| 91.5 \| 92.9 \| 93.5\| \| HeadCT(4-shot) \| 62.4 \| 62.2 \| 80.5 \| 52.2 \| 91.7 \| &emsp; \| 91.2 \| 93.3 \| 94.6\| The results show that the generalization performance of our method on BrainMRI and HeadCT is also good. We hope the above response can solve your concerns. We sincerely wish to further discuss with you.	Summary: This paper proposes a simple but effective framework that can be directly applied to detect anomalies in new classes. The main insight is learning the residual feature distribution rather than the initial one. In this way, we can significantly reduce feature variations. Even in new classes, the distribution of normal residual features would not remarkably shift from the learned distribution. Experiments were conducted on four datasets and achieved remarkable anomaly detection results. Strengths: The paper is original, high quality, clear, and easy to understand. The proposed method has a good heuristic effect on establishing a general anomaly detection model and will become a valuable baseline for the community after the release of the code. Weaknesses: 1. Although unnecessary, I recommend punctuation at the end of a formula. This is one of the few formatting problems I can pick out. [Well written] 2. In Figure (b), it is suggested that abnormal should use a triangle icon. The difference between a hexagon and a circle is too small to see clearly. 3. The large difference between normal images should be considered, and image difference indicators such as FID and LPIPS can be used to calculate the difference inside the normal images in the data set you show. The difference should be relatively small, which is a potential false alarm hazard. 4. As stated in point 4 of the questions, the experimental setup of training on MVTecAD and then testing on the various classes of VisA is not reasonable. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. If the difference between normal images is relatively large, such as the breakfast_box class and screw_bag class in the MVTec Loco AD dataset, how can the reference feature pool be ensured? Intuitively, if the normal image difference is too large and the reference feature pool is not representative, the scheme has the hidden danger of a high false detection rate. If you're not using this dataset for an experiment, you should mention it in the text, which is good for the community. 2. Is the random selection of normal features in the reference feature pool a good strategy at fixed? Is it better to maximize the difference? 3. pixel-level AUROCs of InCTRL in Table 1 should be displayed. If not, it should be explained that it was not done by itself rather than that it could not be obtained. 4. Line 225: As far as I know there are 15 products and their corresponding exceptions in MVTecAD. Did you train the model using 15 product images and test it on various VisA classes? Although MVTecAD and VisA are two different data sets, they are just two sets containing multiple classes. So I think you should show the results of training with n classes from MVTecAD and testing on the remaining 15-n classes, with n as a hyperparameter looking for sensitivity, instead of testing across datasets. It would not be difficult for you to write your experiment in detail, and it would be more convincing if the results of the experiment appeared in the paper. Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: This paper objectively mentions the limitations of this article, and there is no potential negative impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [To W1]. Thanks for your suggestion. We then checked the formula writing in several other papers and found that there is punctuation at the end of a formula. This is a quite good detail suggestion. We will make modifications in the revised version. [To W2]. Thanks for your suggestion. Using triangles to represent anomalies does look clearer. We will further improve Figure 1 in the revised version. [To W3, Q1, and Q2]. We greatly appreciate your suggestion. Yes, when the difference between normal images is too large, it may cause the reference feature pool is not representative. For practical applications, this issue should be particularly focused and reasonably addressed. Of course, the simplest resolution is to increase the number of reference samples. This is feasible, as in practical applications, the number of reference samples is usually not as strict as the 2-shot and 4-shot (following previous papers) in our paper. From the perspective of method comparison, we think that random selection is ok, as long as we ensure that all methods use the same reference samples, the result comparison is reasonable. [To Q2] However, in practical applications, we expect that the reference samples can fully represent their class, so it's best to have sufficient differences between the reference samples. Thus, the sample selection strategy cannot be random. [To Q1] A feasible method is to first cluster all available normal samples into different clusters based on a clustering algorithm (e.g., KMeans). Then, based on the number of reference samples, we evenly distribute it to each cluster. When selecting from a cluster, we can prioritize selecting samples closer to the center. During clustering, we think that the FID and LPIPS you mentioned are good ways to calculate the difference between two samples. In addition, when there are a large number of reference samples, we can also use the method in PatchCore to select coreset features as reference features, which will be more efficient and also representative. In the revision, we will add the above discussion on sample selection strategy to the paper. [To W4 and Q4]. Yes, we train the model using 15 classes in MVTecAD and test it on 12 classes in VisA. For MVTecAD, we train on 12 classes in VisA and test on 15 classes in MVTecAD. Adopting the cross-dataset experimental setup is because we think that it's more challenging than cross-classes within a single dataset, thus can better verify the model's class generalization ability. A dataset may be collected under the same photography condition, variations in other factors besides the object itself may be minimal. For example, all images in MVTecAD only contain a single object, while the images of some classes in VisA have multiple objects. The backgrounds in MVTec3D are all black, which is not the case in other datasets. In addition, we also follow InCTRL to use the cross-dataset experimental setup. We think that your mentioned experimental setting is also very reasonable, as by varying $n$, we can demonstrate the sensitivity of the model to different numbers of training classes. Under the 4-shot setting, we further run our model on MVTecAD (with $n=5, 10$). Note that different $n$ means the number of test classes $15-n$ is different (this will cause the test results of different n cannot be compared with each other). Thus, we use fixed 5 classes as the test classes, including hazelnut, pill, tile, carpet, and zipper. For $n=5$, the training classes include bottle, cable, capsule, grid, and leather. For $n=10$, the training classes include bottle, cable, capsule, grid, leather, metal nut, screw, toothbrush, transistor, and wood. The results are as follows: \|n=5 \| n=10 \| VisA to MVTecAD \| \| --- \| --- \| --- \| \| 96.4/97.6 \| 96.8/97.9 \| 95.1/97.2 \| The results demonstrate that cross-dataset is more challenging than cross-class in a single dataset. With more training classes, the results will be better, but the model is not very sensitive. Due to too many experiments, we currently do not have enough time to provide the results of other methods under this setting. In the revision, we will also add this setting to our experimental setup and complete other methods' results. [To Q3]. For each input image, InCTRL finally only outputs an image-level anomaly score. Thus, InCTRL cannot calculate pixel-level AUROCs. In the revision, we will add a relevant explanation. If you still have any questions, we are very glad to further discuss with you. --- Rebuttal Comment 1.1: Comment: 1. The paper’s logic is clear and the discussion is sufficient. After the code is released, the method can become a highly referenced article in class generalizable anomaly detection. 2. the author's reply on the reference feature pool is satisfactory. It is difficult to fully answer this question experimentally during the rebuttal period, still, I hope the author will add the above discussion on sample selection strategy to the paper, as you said, to promote the development of community research. 3. The author's improvement on the evaluation protocol makes me believe in the code basis of his work, and his grasp of the whole experimental design is reasonable and confident. --- Reply to Comment 1.1.1: Comment: We greatly appreciate your further response. We will add the discussions from the rebuttal response to the revised paper and also release the code.	Summary: This paper proposed a simple yet effective framework ResAD for class-generalizable anomaly detection by leveraging residual feature learning and a hypersphere constraint. The framework's ability to generalize to new classes without retraining or fine-tuning makes it valuable for real-world applications, providing significant improvements over existing methods. Comprehensive experiments on four real-world industrial AD datasets (MVTecAD, VisA, BTAD, and MVTec3D) demonstrate ResAD's superior performance. Strengths: (1)ResAD effectively addresses the challenge of class-generalizable anomaly detection, the generalization ability using only a few normal samples as references makes it highly practical for real-world applications. (2)The use of residual feature learning to reduce feature variations and improve generalizability is novel and effective (3)The approach is shown to be robust across different datasets and settings. Weaknesses: (1)The experiments are primarily conducted on industrial anomaly detection datasets. While these are relevant, the method's generalizability to other domains, such as medical images or video data, is not fully explored. (2)The selection of few-shot reference samples may impact performance. Previous methods typically run multiple independent runs using different random seeds to ensure robustness. However, this work only provides results from a single group of samples, which may not fully represent the model's performance variability. Technical Quality: 3 Clarity: 3 Questions for Authors: (1)If the few-shot reference samples contain anomalies, will it impact the overall performance a lot? (2)The way to combine residual feature learning with existing models is not explicitly defined. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Limitations are discussed in Appendix Sec. B. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [To W1]. We greatly appreciate your suggestion. Under the 4-shot setting, we further evaluate our method on a medical image dataset, BraTS (for brain tumor segmentation) and a video AD dataset, ShanghaiTech (as our method is image-based, we extract video frames as images for use). The comparison results are as follows: \| \| RDAD \| UniAD \| SPADE \| PaDiM \| PatchCore \| RegAD \| ResAD \| WinCLIP \| InCTRL \| ResAD$^{†}$\| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \|BraTS \|50.2/58.7 \| 57.3/85.1 \| 67.1/93.6 \| 63.4/93.1 \| 70.5/94.6 \| 62.1/89.4 \| 74.7/94.0 \| 68.9/93.5 \| 76.9/- \| 84.6/96.1\| \|ShanghaiTech\|56.2/77.6 \| 55.9/79.4 \| 77.1/87.4 \| 74.3/85.9 \| 77.8/88.2 \| 76.4/87.7 \| 79.8/89.5 \| 79.6/88.6 \| 69.2/- \| 84.3/92.6\| The results show that when applied to medical images and video scenarios, the cross-domain generalization ability of our method is also good. In the revision, we will include the above results and also results under the 2-shot setting in our paper. We will also attempt more datasets and further discuss our method's generalizability to other domains. [To W2]. We greatly appreciate your suggestion. Although we only used a single group of few-shot samples, we strictly ensured that all methods used the same few-shot samples. So, when writing the paper, we thought that the result comparison was reasonable and didn't run on other groups. After reading your suggestion, we think that results on multiple groups are necessary. We then randomly select two groups of few-shot samples and obtain the following results: \| \| Group1 \| Group2 \| Results in paper \| Mean$\pm$Std\| \| --- \| --- \| --- \| --- \| --- \| \|MVTecAD \| 91.0/96.0 \| 90.7/95.9 \| 90.5/95.7 \| 90.7$\pm$0.21/95.9$\pm$0.12\| \|VisA \| 86.3/97.5 \| 86.9/97.6 \| 86.2/97.4 \| 86.5$\pm$0.31/97.5$\pm$0.08\| \|BTAD \| 95.3/97.5 \| 95.4/97.6 \| 95.6/97.6 \| 95.4$\pm$0.12/97.6$\pm$0.05\| \|MVTec3D \| 70.2/97.1 \| 70.5/97.3 \| 70.9/97.3 \| 70.5$\pm$0.29/97.2$\pm$0.09\| In the revision, we will include the above results and also results under the 2-shot setting in our paper. Due to too many experiments, we currently do not have enough time to provide the results of other methods on these two groups. We will further supply these results in the revised version. [To Q1]. We think that it will have an impact on the performance. Because it may cause some abnormal features to match similar abnormal features from the reference samples, making these abnormal residual features hard to distinguish from normal residual features. Since the few-shot samples are used as reference, it should be natural not to contain anomalies. For real-world applications, it's also not very hard for us to ensure that the few-shot reference samples are all normal. [To Q2]. As UniAD and RDAD are both feature-based AD methods, combining our residual feature learning with them is straightforward. For UniAD, we convert the initial features into residual features and then perform subsequent feature reconstruction. For RDAD, the initial features extracted by the teacher network are converted into residual features as the learning target of the student network. Then, the student network is trained to predict residual representations of the teacher network. In the revision, we will add relevant details to describe clearly how to combine residual feature learning with existing models. If you still have any questions, we are very glad to further discuss with you. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response, and it addressed my concerns. I will maintain my score.	Summary: This paper proposes to address cross-class anomaly detection problem. To this end, this study introduce a residual learning framework ResAD. The ResAD framework aims to learning residual feature distribution between target image and reference image. Experiments are conducted to valid the effectiveness of the proposed method. Strengths: 1. The cross-class/class-generalize anomaly detection is a crutial task in the realm of anomaly detection. 2. The structure of ResAD is simple and effective. Weaknesses: 1. The idea of residual estimation is highly similar to InCTRL [1]. 2. Lack of comparision with FastRecon[2] and AnomalyGPT[3]. 3. The writing should be improved. The optimization terms are unclear and hard to follow. 4. In Table.5, there is a reproduced result of WinCLIP on WideResNet50, however the windows in WinCLIP is designed for VIT, how can the authors report the result? [1] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In CVPR, 2024. [2] Fang Z, Wang X, Li H, et al. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 17481-17490. [3] Gu Z, Zhu B, Zhu G, et al. Anomalygpt: Detecting industrial anomalies using large vision-language models[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(3): 1932-1940. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. The OCC loss is used for constrain the distribution to a fixed region and the NF model is used for estimate the distribution. If the distribution is fixed, what is the meaning of NF model? 2. There are three loss terms, how about the sensitivities of the balance between the three loss terms. 3. There is no ablation study to valid the effectiveness of the proposed loss terms which makes the method lack credibility. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [To W1]. Thanks for your professional review. We think that our work and InCTRL should be concurrent work. We also initially submitted our work to CVPR2024, and received two weak accepts and one reject (you can see the relevant materials in the rebuttal pdf file). Regretfully, our work was rejected at that time. So, we should have proposed the idea of residual learning independently of InCTRL and almost at the same time. But our method has obvious differences with InCTRL in the definition and utilization of residuals. This (i.e., two independent works almost simultaneously proposed the residual learning idea) also demonstrates residual learning is an effective way to achieve class-generalizable anomaly detection. Then, we made comprehensive revisions to our paper based on the reviewers' suggestions. Subsequently, we also noticed the InCTRL paper and felt that our method has advantages compared to it. Thus, we submitted our revised paper to NeurIPS2024. In CVPR2024, our paper was rejected mainly because a reviewer thought our result comparison with unsupervised AD methods was unreasonable and unfair. In this submission, we mainly compare with few-shot AD methods and InCTRL. The main differences between our method and InCTRL are as follows: (1) The definition of residuals in InCTRL is based on feature distances. The residual map is defined by ((E.q.(1) in the InCTRL paper): $M_x^l(i,j) = 1 - \left<T_x^l(i,j),h(T_x^l(i,j)\|\mathcal{P}^\prime)\right>$, where $h(T_x^l(i,j)\|\mathcal{P}^\prime)$ returns the embedding of the patch token that is most similar to $T_x^l(i,j)$ among all image patches in $\mathcal{P}^\prime$, and $\left<\cdot\right>$ is the cosine similarity function. Thus, InCTRL is based on residual distance maps, while our method is based on residual features. By comparison, we think that residual distances in InCTRL can limit the range of residual representation (as the cosine similarity is in [-1,1]). This is not beneficial for distinguishing between normal and abnormal regions, as a position on the residual map is only represented by a residual distance value. Within a limited representation range (1-[-1,1] $\rightarrow$ [0,2]), normal and abnormal residual distance values are more likely to be not strictly separable. Thus, for a position on the residual map, it's hard for us to make decision based on a scalar value. So, InCTRL makes image-level classification based on a whole residual map (see the following (2)). In contrast, our residual features don't limit the range of residual representation and can retain the feature properties. In high-dimensional feature space, we can also establish better decision boundaries between normal and abnormal (a basic idea in machine learning: solving low-dimensional inseparability by converting to high-dimension). (2) InCTRL devises a holistic anomaly scoring function $\phi$ to learn the residual distance map $M_x = \frac{1}{n}\sum_{l=1}^{n}M_x^l$ and convert it to an anomaly score: $s(x) = \phi(M_x^{+};\Theta_\phi) + \alpha s_p(x)$ (E.q.(8) in the InCTRL paper), where $M_x^{+} = M_x \oplus s_i(x) \oplus s_a(x)$ (E.q.(7)). $s_i(x)$ is an anomaly score based on an image-level residual map $F_x$ (see E.q.(4) in the InCTRL paper) and $s_a(x)$ is a text prompt-based anomaly score. Thus, InCTRL is to train a binary classification network based on residual distance maps. For each input image, InCTRL finally only outputs an image-level anomaly score. Our method is to learn the distribution of residual features, an anomaly score can be estimated for each feature, thus can be used to locate anomalies. (3) Due to the designs in InCTRL that we mentioned above, one main advantage of our method is that it can achieve image-level anomaly detection and also pixel-level anomaly localization, while InCTRL only achieves image-level anomaly detection functionality. As for performance, the average results on four AD datasets of our method are better than InCTRL's (please see Table 1 in our paper). In our response to Weakness 1 of Reviewer CQaQ, the cross-domain generalization ability of our method is also better. [To W2]. We greatly appreciate your suggestion. Under the 4-shot setting, we further reproduce FastRecon and AnomalyGPT by using the same few-shot samples as ours. Based on the official open-source code, we obtain the following results: \| \| FastRecon \| AnomalyGPT \| AnomalyGPT (ViT-B/16+) \| InCTRL \| ResAD$^{†}$ (Ours) \| \| ----------- \| ----------- \| ----------- \| ----------- \|----------- \| ----------- \| \| MVTecAD \| 92.3/96.2 \| 95.0/96.0 \| 92.1/95.3 \| 94.5/- \| 94.2/96.9\| \|VisA \| 82.7/94.1 \| 91.3/88.8 \| 85.4/86.9 \| 88.7/- \| 90.8/97.5 \| \|BTAD \| 90.1/93.4 \| 92.0/96.0 \| 90.2/94.9 \| 91.7/- \| 91.5/96.8 \| \|MVTec3D \| 66.5/95.2 \| 81.7/96.5 \| 75.3/96.2 \| 69.1/- \| 82.4/97.9\| \|Average\| 82.9/94.7 \| 90.0/94.3 \| 85.8/93.3 \| 85.8/- \| 89.7/97.3\| Note that the original image encoder in AnomalyGPT is significantly larger than the ViT-B/16+ used in InCTRL and our ResAD$^{†}$. When AnomalyGPT also uses ViT-B/16+ as the image encoder, our method is more superior to it. In the revision, we will include the above results and also results under the 2-shot setting in our paper. We will also cite FastRecon and AnomalyGPT and add discussions with them. [To W3]. Thanks for your suggestion. We will carefully check and modify obscure expressions and unclear concepts to make our paper easier to follow. We will also seriously consider other reviewers' suggestions to further improve the quality of our paper. [Others]. For Weakness 4 and questions, please see the rebuttal pdf file or the following Comment. If you still have any questions, we are very glad to further discuss with you. --- Rebuttal 2: Comment: [To W4]. The window mechanism of WinCLIP is not limited to ViT. In WinCLIP, the window mechanism is mainly used to address the issue that the local patch features extracted by the CLIP image encoder are not aligned with the text features. The window mechanism provides image patches of different scales, which are then sent into the CLIP image encoder to obtain global features that can align with the text features. The window mechanism is intrinsically similar to the classic sliding window in object detection, thus can also be used in CNN networks. We can send image patches provided by the window mechanism into WideResNet50, and also obtain the window embedding maps of different scales as shown in Figure 4 of the WinCLIP paper. However, because the features of WideResNet50 are not aligned with the text features, we remove the language-guided anomaly score map and only generate the vision-based anomaly score map based on the few-shot normal samples (the WinCLIP+ in the WinCLIP paper). In lines 516-518, we have briefly mentioned our adaptation way. In the revision, we will add more details about the experiments in Table 5. [To Q1]. The goal of our Feature Constraintor (optimized by the OCC loss) is to constrain initial residual features to a spatial hypersphere for further reducing feature variations. After the Feature Constraintor, feature variations can effectively be reduced, but this does not mean that the feature distribution is fixed within the hypersphere. Moreover, even if the feature distribution is fixed, we still need to learn the distribution and then can perform anomaly detection based on the learned distribution. The NF model is used to learn the feature distribution. [To Q2]. During training, we found that summing up the three loss terms and then backpropagating gradients to optimize the whole model would lead to unstable training. Then, we used the "torch.detach()'' method in the Pytorch library to detach the features after the Feature Constraintor and then sent the detached features into the NF model. This simple way can make the model training more stable. In the revision, we will add this code detail to the paper. Thus, the weight of $L_{occ}$ can be set as 1 (i.e., we can not need to balance $L_{occ}$ with $L_{ml}$ and $L_{bg-spp}$, as the Feature Constraintor and the NF model parts are separated in the gradient graph). When training the NF model, the $L_{ml}$ is the basic loss. So, we keep the weight of $L_{ml}$ as 1 and set a variable $\lambda$ as the weight of $L_{bg-spp}$. By varying different $\lambda$ values, the results (under the 4-shot setting) about the sensitivity are as follows: \| $\lambda$ \| 0.1 \| 0.5 \| 1 \| 2 \| 3 \| 5 \| 10 \| \| --- \| --- \|--- \|--- \|--- \|--- \|--- \|--- \| \| \|89.1/94.9 \| 90.0/95.6 \| 90.5/95.7 \| 90.6/96.0 \| 89.8/95.3 \| 89.3/95.2 \| 88.7/94.9\| Both small and large $\lambda$ values can lead to performance degradation. $L_{bp-spp}$ is to assist model in learning abnormal residual features. Small $\lambda$ may cause the impact of abnormal features on final loss to be relatively small. Large $\lambda$ may lead to overfitting to known anomalies, which is not conducive to generalization. [To Q3]. In our paper, we mainly propose the Abnormal Invariant OCC (AI-OCC) loss $L_{occ}$. The maximum likelihood loss $L_{ml}$ is the basic loss for training the NF model. The $L_{bg-spp}$ loss is following BGAD to effectively utilize abnormal features. In fact, the framework ablation studies in Table 2(a) also include ablation studies regarding our proposed AI-OCC loss. In Table 2(a), “w/o Feature Constraintor'' means the $L_{occ}$ is not used (as the Feature Constraintor is removed), “w/o Abnormal Invariant OCC loss'' means the $L_{occ}$ only has the first part of E.q.(2), while is not our proposed AI-OCC loss. In lines 293-298, we also provide discussions about the ablation studies of our AI-OCC loss. [Limitations]. In Appendix Sec.B, we provide discussions about our method's limitations. In the revision, we will further discuss the limitations of our method more comprehensively and clearly based on the comments of all reviewers. --- Rebuttal 3: Title: Response Comment: The authors' rebuttal has addressed most of my concerns and I believe that the generalization ability of the proposed method is practical and significant, so I decide to raise my score. However, I still have some concerns about the conflicts of OCC and NF. Also, I observe the rebuttal file in the supplementary (Line 42-51) that same problem have mentioned by other comments, so I think it would be a major concern about the method. It would be better If there are some more convicing validation. --- Rebuttal Comment 3.1: Comment: We greatly appreciate your further response. We think that your concern should be whether both the Feature Constraintor (namely the OCC part) and the NF model are necessary, namely, whether we can achieve anomaly detection only based on the OCC part, or whether we only need the NF model without the OCC part. We think the ablation study (``w/o Feature Constraintor'', namely only using the NF model) in Table 2(a) can answer the latter question. When we design the Feature Constraintor, our motivation is to employ it to constrain initial residual features to a spatial hypersphere for further reducing feature variations. In this way, for the features sent into the NF model, the distribution of new classes could be more consistent with the learned distribution. Thus, it's beneficial to achieve better cross-class AD results. After reading your response, we realize that we need to design an ablation study in which we directly perform anomaly detection based on the output of the Feature Constraintor without the NF model. Specifically, we regard the OCC part as a conventional OCC-based AD model and utilize the distances from the features to the OCC center as the anomaly scores (this is also a commonly used way for measuring anomaly scores in OCC-based AD methods). Under the 4-shot setting, we obtain the following results (also the results from the paper): \| \| w/o Feature Constraintor \| Only Feature Constraintor \| ResAD \| \| --- \| --- \| --- \| --- \| \| From VisA to MVTecAD \| 82.3/93.5 \| 82.9/89.3 \| 90.5/95.7 \| The results show that the OCC part and the NF model don't have conflicts. As you mentioned in Question 1, the ideal situation is that even in new classes, normal feature distribution is fixed within a hypersphere, while all anomalous features are outside the hypersphere. Then, only the OCC part is enough to achieve good AD results. However, in practical optimization, it's hard to achieve the ideal situation. After the Feature Constraintor, normal and abnormal features may still not be fully separable based on the distances. Then, the NF model is used to learn the feature distribution, which can assist us in better distinguishing normal and abnormal features. Therefore, we think that further learning the feature distribution after the OCC part is beneficial. In the revision, we will add the new ablation study and the relevant discussions to our paper. We hope the above discussions can solve your concerns. If not, we are very glad to further discuss with you.	Rebuttal 1: Rebuttal: We are very grateful for all your constructive suggestions. Please see our specific responses to each reviewer. In the Author Rebuttal pdf file, we provide some relevant materials. We recommend that Reviewer fL7E and 6bzj can download the pdf file and see the contents in it. Pdf: /pdf/a4a7fd593dd242f714a47f767fe6f89268414d09.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Continuous Spatiotemporal Events Decoupling through Spike-based Bayesian Computation	Accept (poster)	Summary: This paper presents a spike-based Bayesian inference framework for motion segmentation with event cameras. By designing neurons that utilize STDP for online learning of motion patterns, the framework can perform the M-step of the EM algorithm in motion segmentation of event streams. Additionally, the WTA circuit implements the E-step, allowing for the online partitioning of event streams into different motion patterns. The authors provide theoretical proof and experimental results to demonstrate the network's spatiotemporal decoupling capabilities for mixed motion patterns of event streams. Strengths: The authors demonstrate that the SNN framework based on WTA is equivalent to the EM algorithm for motion segmentation of event streams. This online learning approach is compatible with neuromorphic data and beneficial for deployment on low-power, low-latency neuromorphic computing platforms. • The work is based on the Bayesian brain hypothesis, using a more physiologically interpretable SNN for Bayesian inference. Applying this to spatiotemporal data from neuromorphic cameras represents a promising research direction. Weaknesses: • The experimental results lack quantitative evaluations. Can the authors further perform object detection and tracking based on the motion segmentation, providing metrics such as object detection success rates and comparisons with other methods? • The proposed algorithm lacks the analysis of time complexity or processing speed. Can it leverage the low-latency advantage of event cameras? Technical Quality: 3 Clarity: 3 Questions for Authors: Please see weaknesses. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: There is a need for quantitative evaluations and an assessment of the dependency on parameter initialization. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your constructive suggestions and have supplemented our work with further results on object detection based on motion segmentation. Specifically, we calculated the detection success rate on the EED dataset, corresponding to Fig. 6 in the main text. Our detection success rates across three test scenarios are `100%`, comparable to many current state-of-the-art algorithms ([s1], [s2]). References: [s1] Kepple, Daniel R., et al. "Jointly learning visual motion and confidence from local patches in event cameras." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16. Springer International Publishing, 2020. [s2] Zheng, Yajing, et al. "Spike-based motion estimation for object tracking through bio-inspired unsupervised learning." IEEE Transactions on Image Processing 32 (2022): 335-349. 2. Parameter Efficiency and Computational Resource Requirements: Our method requires minimal parameters and computational resources. Specifically, the model uses neurons corresponding to different motion models and a single global inhibitory neuron to perform WTA. This parameter efficiency and low computational requirement make our approach particularly advantageous for hardware or neuromorphic hardware deployment. It can fully leverage the low latency and low power consumption characteristics of neuromorphic cameras and computing platforms. Thank you once again for your invaluable feedback. We hope these additions and clarifications address your concerns and further demonstrate the robustness and potential of our proposed method for event-based motion segmentation and object detection using spiking neural networks. --- Rebuttal Comment 1.1: Title: Comments Comment: The authors responded to my earlier questions about the quantitative comparison of their proposed method with SOTA and addressed my concerns regarding delay and time complexity. They provided detection success rates (100%) for the three extreme scenarios in Fig. 6, which are comparable to SOTA methods, but with significantly lower algorithm complexity. After reviewing the authors' replies to all the reviewers, I'm convinced that incorporating the reviewers' feedback, such as suggestions on WNgv descriptions, will greatly enhance the potential of the spike-based Bayesian Computation framework presented in this paper. This framework has enormous potential for low-latency, low-power applications in linking neuromorphic sensors and NMHW. Therefore, I am considering giving it a higher rating. --- Reply to Comment 1.1.1: Comment: Thank you for your valuable feedback! We're glad our responses addressed your concerns, and we appreciate your consideration of a higher rating. We will continue to refine our work based on the reviewers' suggestions.	Summary: This work proposes a spike Bayesian computational framework for continuous motion segmentation in event streams and demonstrates that the constructed network can implement an EM-based event stream motion segmentation model. The proposed model uses WTA circuits in the network to achieve an equivalent E-step, while the STDP rules for an M-step for contrast maximization. Experimental results demonstrate the network's online learning effectiveness for continuous inputs on extreme event camera datasets. Strengths: The proposed network's effectiveness for motion segmentation has been validated on event datasets featuring challenging scenarios that involve mixed camera self-motion and high-speed moving objects. The proposed spike Bayesian inference framework is highly interpretable and applicable to various neuromorphic vision chips and computing hardware, representing a promising research direction. Weaknesses: The authors mainly use SVD to find different patches' motion patterns for initialization. Why is this method used, and can other methods be employed for selection? It is recommended that the authors conduct ablation experiments to explore further. Technical Quality: 3 Clarity: 3 Questions for Authors: This method primarily targets optical flow motion estimation. For more complex motion patterns, how to design the parameters? How robust is this method against noise in the evaluation of such motion models? The authors should clarify it. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have stated the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your acknowledgment of our validation of event datasets featuring challenging scenarios, including mixed camera self-motion and high-speed moving objects, which is highly valued. We are pleased that you find our spike Bayesian inference framework to be highly interpretable and applicable to various neuromorphic vision chips and computing hardware. Your appreciation of our work's potential as a promising research direction is encouraging. ### Detailed Response 1. Motion Parameter Initialization: In our work, we were inspired by the layered method for event stream motion segmentation described in the EM-based approach and the SOFADS algorithm [s1]. The SOFADS method iteratively refines optical flow estimates through the Track Plane and Flow Plane modules. The Track Plane contains projections of different flow hypotheses, updated based on incoming events. Similarly, we adopted a patch-based approach to perform importance sampling on the event stream to identify potential motion parameters based on the optimization objective (contrast of IWE). Our method involves a search process detailed in lines 226-234 of our paper, with Fig. 4 illustrating this search method combined with SVD analysis. We select the $N_l$ representative parameters with the highest variance as initial values. It is essential to note that the parameter search process is crucial, and the use of SVD can be substituted with K-means for initial parameter selection. As shown in Fig. S2 `(please see in the supplementary PDF)`, the parameter points obtained using K-means clustering are similar to those obtained with SVD. 2. Types of Motion Neuron Parameters: The parameters of motion neurons in our framework can accommodate much more complex motion patterns by simply modifying the neuron model. For instance, our model can be extended to handle rotational motion [s2], 4-DOF motion [23], or even 8-DOF homographic motion [s3]. This flexibility underscores the versatility of our framework in adapting to various motion types. 3. Noise Handling: Regarding noise, its randomness means it generally does not conform to any initialized motion model. During motion segmentation, noise tends to have similar probabilities across different categories, effectively filtering it out. This inherent filtering capability enhances the robustness of our segmentation method in real-world scenarios where noise is prevalent. Thank you once again for your valuable feedback. We hope these clarifications address your concerns and further illustrate the robustness and potential of our proposed method for motion segmentation using event-based neural networks. References: [s1] Stoffregen, Timo, and Lindsay Kleeman. "Simultaneous optical flow and segmentation (SOFAS) using dynamic vision sensor." arXiv preprint arXiv:1805.12326 (2018). [s2] Guillermo Gallego and Davide Scaramuzza, “Accurate angular velocity estimation with an event camera,” IEEE Robot. Autom. Lett., vol. 2, no. 2, pp. 632–639, 2017. [s3] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza, “A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation,” in IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pp. 3867–3876, 2018. --- Rebuttal Comment 1.1: Comment: Thank the authors for the clarifications on the motion pattern initialization and noise handling. I have no more questions. --- Reply to Comment 1.1.1: Comment: Thank reviewer for the valuable feedback and taking the time to review our work. We're glad the clarifications were helpful!	Summary: The paper proposes to address motion segmentation at very high temporal resolution via an event-based or spiking implementation of expectation-maximization in a generative model. It demonstrates the performance of the resulting spiking neural networks on example experiments. Strengths: The strength of the paper is its deep engagement with the spiking neural network literature, as well as its use of spiking networks for the specific type of problem to which they are most suited: event-based computation. Weaknesses: The paper's major weakness is its lack of clarity, which the authors have discussed and addressed in the review period. Technical Quality: 2 Clarity: 2 Questions for Authors: The authors have addressed my questions, though I would still like to see discussion of how this framework for EM in spiking networks could be generalized beyond motion detection. Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: I am not confident that I can identify the specific limitations of this paper as opposed to the limitations of spiking neural networks generally. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's recognition of our approach to high-resolution motion segmentation using an event-based implementation of the EM algorithm and acknowledge our deep engagement with the spiking neural network literature. Here, we aim to provide further clarification on the EM framework and our model's specific application. ### Clarification on the EM Algorithm Framework: The EM algorithm is a versatile framework used to find parameters in probabilistic models with latent variables. While traditionally, the M-step in the EM algorithm maximizes the likelihood function, it can also maximize other objective functions, such as the Evidence Lower Bound (ELBO) [s1] in variational inference or the Penalized Likelihood [s2] in sparse regression models like Lasso regression. In previous work on event stream motion segmentation [34], the M-step focused on optimizing the contrast of different IWEs, effectively projecting the event distribution variance using motion parameters. Our current approach uses the EM framework to iteratively perform the E-step and the M-step to refine the estimates of motion parameters, achieving event stream motion segmentation. `The relationship between them is shown in Fig. 1 of the main text`. Therefore, our method does not involve the common concept of maximizing log densities in mixture distribution models. References: [1s] Blei, D. M., Kucukelbir, A., & McAuliffe, J. D. (2017). Variational Inference: A Review for Statisticians. Journal of the American Statistical Association, 112(518), 859-877. [2] Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1-22. ### Model's Nature and Application: Thank you for pointing out the potential confusion regarding our model's nature. Our proposed approach is not a generative model. Instead, it is designed to learn motion parameters embedded in event streams generated by different motions. These learned motion parameters help segment the event streams according to their respective motion distributions. The term "motion distribution" here refers to a projection plane where the variance is maximized if the events follow the motion distribution. By implementing this EM-based motion segmentation using SNNs, we can leverage the low-latency and low-power characteristics of neuromorphic hardware (NMHW), such as `Loihi and Spinnaker`. This combination enhances the performance of event-based cameras and neuromorphic computing systems. We recognize the importance of this aspect and will include more detailed discussions in future revisions of our paper. Thank you again for your comments. We hope this response clarifies our approach and addresses your concerns comprehensively. --- Rebuttal Comment 1.1: Title: I must be misunderstanding something. Comment: I thank the authors for their close technical engagement here, and would request that they guide me through just a couple more steps of reasoning. I understand that the EM algorithm is ultimately a coordinatewise optimization of the ELBO, first on latent variables and then on model parameters (alternating until convergence). But what objective, then, are the authors claiming their application of EM optimizes? Similarly, I understand that the authors are claiming to perform EM rather than implement a generative model, but EM is an inference algorithm for probabilistic generative models with latent variables. If you cannot write down a joint probability density, that is a generative model, to parameterize the ELBO, then the EM algorithm, as I understand it, cannot be used for inference. Could the authors help me through my confusion here? --- Reply to Comment 1.1.1: Comment: We sincerely appreciate the reviewer’s comments and the opportunity to further clarify our work. We are grateful for the chance to provide additional details regarding our optimization goals and the joint probability density. ### 1. What objective are the authors claiming their application of EM optimizes? Our objective is to maximize the sum of the variances of the IWE across different motion parameter distributions (Eq. (4) in the main text). As described in the supplementary PDF (Fig. S1), the variance indicates that correct motion patterns (illustrated by the circles in the middle subfigure) concentrate the event flow distribution along the object's edges, resulting in higher variance. In contrast, incorrect motion patterns disperse the event flow, leading to lower variance. In the network we designed, which includes a WTA mechanism, the parameters of the motion neurons, denoted as $\theta$, are optimized using the STDP rule to achieve this goal. As shown in Eq. (12) and Fig. 3 of the main text, when $P(z_j=1\|e_k=1)$, only event streams that match motion pattern $j$ will activate the corresponding motion neuron to adjust its parameters, thereby maximizing the variance of IWE (encoded by $u$) associated with motion pattern $j$. ### 2. Joint Probability Density From the perspective of a probabilistic model, we can consider that events conforming to different motion patterns emerge over time. However, in the task of dividing event streams based on motion patterns, it is hard to directly generate a model (e.g., a mixture distribution model) using a generative model. In previous work [34], the authors adopted a method based on Eq. (2), which is more suitable for traditional mixture distribution models (e.g., fuzzy mixture models and k-means clustering) to divide event streams by motion. In fuzzy mixture density and k-means methods, the motion-compensated IWEs do not include the event cluster associations $P$, which means that sharper object boundaries appear in some IWEs compared to others. The key difference between the EM model in this paper and traditional mixture distribution models lies in the fact that not all motion parameters are mixed. Instead, a specific one-to-one relationship is established between the event $e_{k}$ and the motion neuron $z_j$, resulting in a more precise correspondence. Therefore, we can define the joint probability distribution as: $$ P(e_k, z_k \mid \theta) = P(z_k \mid \theta) \cdot P(e_k \mid z_k, \theta) $$ where the definition of $P(e_k \mid z_k, \theta)$ relates to the IWE value. Specifically, we can define the conditional probability of event $e_k$ belonging to motion pattern $z_j$ as: $$ P(e_k \mid z_j, \theta) \propto I_j(x_{k_j}' \mid \theta_j) $$ where $I_j(x_{k_j}' \mid \theta_j)$ represents the IWE value calculated based on the position and time of event $e_k$ using the parameters $\theta_j\$ corresponding to motion pattern $z_j$. We will include this explanation in the revised version of our paper. Thank you again for your valuable comments!	Summary: This paper demonstrates that WTA circuits along with STDP learning resembles EM algorithm-like Bayesian inference and could be used for motion segmentation from event streams by contrast maximization of warped events. Strengths: The paper proposes an interesting approach for event motion segmentation based on observations from event-based dynamic vision sensors, utilizing a EM-like framework for identifying various motion models from event streams and clustering them into motion patterns. This is achieved using WTA circuits together with STDP-based learning. Weaknesses: The main weakness of the paper is that the proposed method lacks proper justification of the presented approach, which seems like a heuristic hard clustering method, together with gradient based learning. The experiments also lack depth and the authors demonstrate the high dependence of the performance of the method on the parameter initialization. A more careful writing of the underlying model, the optimization framework and the proposed methodology would be good (see the questions below). Furthermore, the paper lacks more details regarding the choice of $N_{\ell}$ (number of motion models) and the specific forms of the warping functions $W_j$ used. Several steps in the entire methodology, although intuitive, are presented in a heuristic fashion without detailed description and clarity. Technical Quality: 2 Clarity: 1 Questions for Authors: Here are some general questions/comments about the framework: 1. Full form of the abbreviation STDP missing in abstract/introduction. 2. In Eq. (1), why is $\Delta L(x_k,t_k) = q_k\Theta$, since according to line 107, event $e_k$ corresponds to when the intensity change \textit{exceeds} $\Theta$ (noting that $\|q_k\|=1$). In line 109, add "where $L(x,t)$ is the ... at pixel $x$ \textit{at time $t$}". 3. Line 118: what integration is used? Eq. (2) only describes $I_j(x)$ as a mixture of Dirac measures. Add the definition of $N_e$ (possibly the number of observed events). In Eq. (2), does $x$ represent a pixel? What does the suffix $j$ capture? Based on the description, it suggests that it represents the different \textit{motion models} -- it would be better to explain both the model and the optimization problem in slightly more detail for clarity. 4. The EM framework: while the updates for the model resemble E and M steps in EM, is it actually related? Can you show that this method indeed improves some form of likelihood of the model (recall that EM is most commonly used for MLE in mixture models or other latent variable models)? Can the authors discuss how their method is related to EM (the E and M steps in the current paper are more closer to the hard clustering type algorithms e.g. K-Means rather than EM, particularly the E-step Eq. (5)) 5. In Eq. (6), extra $dx$ at the end, also might be better to keep the two terms in a parenthesis. 6. More on the model Eq (2): according to line 119, $p_{kj}$ represents the probability that event $e_k$ belongs to motion model $z_j$, in that case $\sum_j p_{kj}=1$, is that correct? However, in that case, Eq. (2) does not represent a mixture -- i.e., $\sum_k p_{kj}$ might not be 1, can the authors clarify this? Furthermore, the Dirac function in Eq (2) allows the IWE to only pick up values at pixels $x$, where at least one event $e_k$ has been observed (through the transformed position). This does not allow any spatial relation across the pixels - why can the Dirac function not be replaced by some other smooth kernel (like Gaussian)? 7. More on the optimization problem Eq (4): When writing $\text{Var}(I_j)$, whose randomness are we taking the variance (or other expectation operations) with respect to? It seems like $\theta, P$ are parameters (hence fixed) and $x_{kj}'$ is some deterministic transform of the observed events. Can the authors clarify the underlying probability structure of the model? 8. The authors might provide a brief description of the STDP method (and its connections to spiking neural networks) and WTA circuit, which might clarify some of the paragraphs e.g., lines 170-173. It is also unclear why $u_j (t)$ (defined in Eq (10)) is equivalent to $I_j$ (in Eq (2)) as claimed in line 178 - is the $W_j$ in Eq (10) same as the WTA $W$ in Eq (3) -- if so, why is the second input $t_k$ in the latter while $p_{kj}$ in the former? 9. Can the authors explain lines 205-206. It seems like they argue that gradient update increases the variance -- however, this is only the M-step (i.e., conditional on the current values of $p_{kj}$ I am guessing). 10. Why is $u_j$ expressed as a function of time $t$ in Eq (10)? Can the authors clarify how the temporal dependence is captured in the model? Confidence: 3 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: See the questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing our innovative approach to event motion segmentation using event-based dynamic vision sensors and an EM-like framework. We appreciate your acknowledgment of our method's use of WTA circuits combined with STDP-based learning. The following are the main issues addressed: ### A. Justification of the Approach 1). Theoretical Background: Our research aims to explore the application of Bayesian Inference for motion segmentation by decoupling events using a Spiking Neural Network (SNN)-based framework. This addresses the question, "How can an SNN model implementing Bayesian computation decouple and segment event streams generated by different motion patterns?" Building on previous studies [34] that successfully utilized EM-based methods for similar tasks, our work demonstrates both theoretically and experimentally that this method can be implemented using an SNN. 2). Clarification: We aim to provide a theoretical foundation for applying Bayesian inference using SNNs for event stream decoupling tasks, leveraging the energy efficiency of SNNs. This is particularly relevant for neuromorphic visual sensors and computing hardware (e.g., `Loihi [s1] and SpiNNaker [s2]`), known for their low latency and power consumption. Neuromorphic computing emulates the brain's low-power yet complex visual task-processing capabilities. Previous research validated the hypothesis that SNNs can implement brain-like Bayesian inference. However, its application to neuromorphic computing hardware remains unverified. Our work aims to validate the capability of a spike-based Bayesian computation framework applied to neuromorphic sensors. References: [s1] M. Davies et al., “Loihi: A neuromorphic manycore processor with on-chip learning,” IEEE Micro, vol. 38, no. 1, pp. 82–99, Jan. 2018. [s2] Furber, Steve B., et al. "Overview of the SpiNNaker system architecture." IEEE transactions on computers 62.12 (2012): 2454-2467. ### B. Detailed Model and Methodology Comprehensive Writing The underlying model, optimization framework, and proposed methodology are detailed thoroughly. We have included pseudo-algorithm code `(please see in the supplementary PDF)` to provide step-by-step explanations, ensuring clarity. This will be added to the revised version for a more comprehensive description. Examples and Case Studies: We have included examples and case studies to illustrate each step of the methodology (Fig. 4-6 in the main text). These practical illustrations help clarify the application of the method and its effectiveness. ### C. Choice of Parameters 1). Rationale for $𝑁_l$: The number of $𝑁_l$ can be set based on the experience with the datasets, making sense when trying to separate motion patterns contained in the events. A more intelligent approach can be used after obtaining different motion patterns of other patches. As shown in Fig. 4, clusters of different motions are identified using a non-parametric clustering approach for motion patterns of various patches. 2). Forms of $𝑊_𝑗$: The forms of $𝑊_𝑗$ depend on the selected motion patterns. A general method uses an affine transformation matrix, but we mainly consider optical flow changes as motion patterns, using $(vx, vy)$ as the motion parameters. Detailed descriptions and justifications for the specific forms of the warping functions (𝑊𝑗) used are provided, explaining their derivation and role in the overall model. ### D: Specific Clarifications: 1). STDP, WTA, and Eq. (6): Thank you for the reminder! We will remove dx from equation (6) and add the full name of STDP in the abstract. Descriptions of the STDP method and WTA circuit will be included to clarify related paragraphs. 2). Line 109 Addition: The formula Δ𝐿(𝑥𝑘,𝑡𝑘)=𝑞𝑘Θ represents changes between consecutive events. Event streams are generated when a threshold is reached, followed by a reset. Thus, the log intensity change between event streams is 𝑞𝑘Θ. 3). IWE Clarification: We will clarify that IWE results from warping events to the same image plane, where x represents the pixel location. The term "integration" will be removed to avoid confusion. "j" represents the $j-th$ motion model. 4). EM Algorithm: The EM algorithm, a framework for finding parameters in probabilistic models with latent variables, is used in our approach to iteratively improve motion parameter estimates for event stream motion segmentation. Unlike the Kalman filter, which estimates the state of dynamic systems through prediction and update steps, our approach alternates between the E-step and M-step within the EM framework to achieve event stream motion segmentation. 5). Var(𝐼𝑗) Explanation: The variance (Var) of the IWE under different motion models shows `(Fig. S1 in the supplemented PDF)` that correct motion models concentrate event distributions at object edges, resulting in higher variance, while incorrect models disperse events, leading to lower variance. This helps in validating the accuracy of motion pattern detection. 6). Gradient Update Impact: The designed spiking model, combined with the STDP rule, updates motion neuron weights, effectively maximizing the variance of different IWEs, thereby validating the feasibility of the spike-based Bayesian inference framework for event stream motion segmentation both theoretically and experimentally. 7). Temporal Dependence: The value $u_j(t)$ represents the value of neuron $u_j$ at time $t$. Thank you for pointing it out. We will change this notation in future versions to avoid ambiguity. Our IF spiking neuron model mainly considers input event streams without simulating other temporal decay characteristics. Thank you again for your valuable feedback. We hope our clarifications and additional details address the concerns raised and further demonstrate the robustness and potential of our proposed method. --- Rebuttal Comment 1.1: Comment: Thank you for the response. Some of the ambiguity have been dealt with and I increase my score to 5. However, although the overall framework is similar to EM, it is not clear why it actually is. To be precise, it would be better to explicitly write the conditional likelihood to describe the Expectation step. As I mentioned, it seems that it is closer to a hard-clustering algorithm rather than an EM (they are similar but fundamentally different). --- Reply to Comment 1.1.1: Comment: Thank you very much for your feedback and for increasing the score. We appreciate your observation regarding the similarities and differences between our framework and EM. To clarify, our method does not implement a hard assignment of events to specific motion patterns. Instead, the probability of an event belonging to a particular motion pattern is proportional to its IWE value (or encoded as membrane potential $u$ in the proposed model) under different motion patterns, as outlined in Eq. (5) and Eq. (11) of the main text. Specifically, the conditional probability $P(e_k \mid z_j, \theta) \propto I_j(x_{k_j}' \mid \theta_j)$, where $I_j(x_{k_j}' \mid \theta_j)$ represents the IWE value associated with the motion pattern $z_j$. This approach allows for a more nuanced assignment that considers the probabilistic contributions from multiple motion patterns rather than a strict, binary assignment. In the revised version of the paper, we will explicitly describe the conditional likelihood during the Expectation step to better articulate the similarities with EM, ensuring that these differences are clear and avoiding any further ambiguity. Thank you again for your valuable suggestions, which will help us improve the clarity and accuracy of our work!	Rebuttal 1: Rebuttal: We sincerely appreciate the valuable feedback from all reviewers. Thank you for recognizing the strengths of our work. Reviewer WNgv praised our innovative approach to event motion segmentation using event-based dynamic vision sensors and an EM-like framework, highlighting the use of Winner-Take-All (WTA) circuits combined with Spike-Timing-Dependent Plasticity (STDP)-based learning. Reviewer pCA1 appreciated our high-resolution motion segmentation using an event-based implementation of the EM algorithm and noted our engagement with the spiking neural network literature. Reviewer wrk7 recognized the effectiveness of our network for motion segmentation validated on challenging event datasets, including mixed camera self-motion and high-speed moving objects, and commended our spike Bayesian inference framework's interpretability and applicability to neuromorphic vision chips and computing hardware. Lastly, Reviewer LG4A endorsed our SNN framework based on WTA as equivalent to the EM algorithm for motion segmentation and supported our application of the Bayesian brain hypothesis using a physiologically interpretable SNN for Bayesian inference on spatiotemporal data from neuromorphic cameras as a promising research direction. In response to the questions raised, we provide the following clarifications and updates, and have provided some materials and results figures in the uploaded supplementary PDF. ### A. Theoretical Foundation: Our research explores Bayesian Inference for motion segmentation by decoupling events using an SNN-based framework, addressing how SNNs implementing Bayesian computation can decouple and segment event streams generated by different motion patterns. Inspired by previous EM-based methods, we demonstrate the theoretical and experimental viability of this approach with SNNs. We aim to provide a theoretical foundation for applying Bayesian inference using SNNs for event stream decoupling tasks, leveraging their energy efficiency, which is particularly relevant for neuromorphic visual sensors and computing hardware like Loihi and SpiNNaker. ### B. Model and Methodology: Our model, optimization framework, and methodology are thoroughly detailed. We will include pseudo-algorithm code for step-by-step explanations in the revised version for clarity. Examples and case studies illustrate each step of the methodology. We discussed the rationale for setting the number of motion patterns (𝑁ℓ) and provided justifications for the forms of warping functions (𝑊𝑗), explaining their derivation and role. Additionally, we will add the full name of STDP in the abstract and include descriptions of the STDP method and WTA circuit to clarify related paragraphs. ### C. EM Algorithm Framework: In previous work on event stream motion segmentation [34], the M-step focused on optimizing the contrast of different IWEs, effectively projecting the event distribution variance using motion parameters. Our current approach uses the EM framework to iteratively perform the E-step and the M-step to refine the estimates of motion parameters, achieving event stream motion segmentation. Thus, our method does not involve the common concept of maximizing log densities in mixture distribution models. ### D. Technical Clarifications: We clarified various technical details, such as the formula Δ𝐿(𝑥𝑘,𝑡𝑘)=𝑞𝑘Θ, the IWE concept, the use of the EM algorithm, variance explanations for motion models, and gradient update impacts. Our proposed model is not a generative model but is designed to learn motion parameters embedded in event streams generated by different motions, helping segment the event streams according to their respective motion distributions. Implementing this EM-based motion segmentation using SNNs leverages the low-latency and low-power characteristics of neuromorphic hardware, enhancing the performance of event-based cameras and neuromorphic computing systems. ### E. Supplementary Results: We have supplemented our work with further results on object detection based on motion segmentation. Specifically, we calculated the detection success rate on the EED dataset, corresponding to Fig. 6 in the main text. Our detection success rates across three test scenarios are `100%`, comparable to many current state-of-the-art algorithms. ### F. Parameter Efficiency and Resource Requirements: Our method is highly efficient, requiring minimal parameters and computational resources. The model uses neurons corresponding to different motion models and a single global inhibitory neuron to perform WTA. This parameter efficiency and low computational requirement make our approach particularly advantageous for deployment on hardware or neuromorphic hardware, leveraging the low latency and low power consumption characteristics of neuromorphic cameras and computing platforms. We hope these additions and clarifications address your concerns and further demonstrate the robustness and potential of our proposed method for event-based motion segmentation and object detection using spiking neural networks. Pdf: /pdf/790ae72b66b79341cad991561b9b799354c321eb.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Improved Few-Shot Jailbreaking Can Circumvent Aligned Language Models and Their Defenses	Accept (poster)	Summary: This paper proposes a new way to jailbreak LLMs through an improved version of few-shot jailbreaking. They propose to use a random search to select examples that are most effective to jailbreak the mode from a pre-defined pool generated with Mistral-7B. On top of that, they alternate the steps of each example by the special tokens that are used in the LLMs conversation templates to separate user messages from the model's responses. The authors show that this method is more effective than the previous jailbreaking methods for five different models, and that it can be used and adapted to evade a large number of defenses. Strengths: Simple and effective method. The method proposed is simple and effective. It is easy to understand and to implement. The experimental results show that it is more effective than many baselines. Insightful ablations. The authors do a great job at showing what components are most important for the success of the attack. They check how many shots are necessary, how important the size of the pool is and how important the special tokens are. However, there are some other ablations that I believe would make the paper stronger (see weaknesses). Effective evasion of defenses. The authors show that their method is effective at evading a large number of defenses of different types, from a perplexity filter, to perturbation-based defenses, to safety-filters. Most interestingly, they propose that one could actually exploit a defense (SmoothLLM) to make the attack robust to keyword-based defenses. However, they do not have any experimental results to show that this is actually the case. Mildly Compelling motivation. The motivation of using few-shot jailbreaking is compelling to jailbreak models that do not support a long context. However, it should be noted that these models are also less likely to actually provide useful malicious information to the attacker who is trying to jailbreak the model. Weaknesses: No comparison to few/many-shots baselines. The authors do not compare their method to Wei at al. [1] and Anil et al. [2], which are the most similar to their method. They claim that Wei et al. have limited effectiveness on well-aligned models such as Llama-2, but Llama-2 is not the only target model considered in the paper, and the authors should show some concrete numbers to back-up their claim. For Anil et al., they claim that the attack requires too much context length to work on the considered models, but, according to the numbers shown in the paper [2], the attack starts being effective with 32 shots, the number considered for Llama-3, and they have results for Llama-2 in their paper up to 128 shots. Missing amount of necessary queries. One of the metrics that are useful for jailbreak attacks is the total number of queries needed by the random search to jailbreak the model. The authors do not report this number, which makes it hard to compare their method to other methods. Some ablations are missing. The authors do a great job at showing what components are most important for the success of the attack. However, they do not show the impact of the quality/length of the examples. It would be interesting to see how the method performs when the examples are shorter or longer, or when some of them are not actually good examples. This would be relevant as the model used to generate the examples could refuse, or generate low-quality examples. Another ablation that would make the paper stronger is how important it is that the special tokens are correct. What happens if you, e.g., use Llama-2's special tokens for Qwen1.5B? Or simply if the special tokens are slightly incorrect (e.g., `[INST]` instead of `[/INST]`? This can be useful to show the potential effectiveness of the attack against models whose special tokens are unkown. Minor: - No experiments that show that SmoothLLM can be used to evade keyword-based defenses. - Code is provided, but the data are provided in pickle format, which is known to be unsafe. It would be better to provide the data in a more standard format like CSV or JSON. Moreover, it would be better to provide a README with instructions on how to understand the code. References: - [1] Wei et al., https://arxiv.org/abs/2310.06387 - [2] Anil et al., https://www.anthropic.com/research/many-shot-jailbreaking Technical Quality: 3 Clarity: 3 Questions for Authors: - Did you try to use the special tokens from one model to jailbreak another model? - Why do you use four `[/INST]` between pseudo-messages? Have you tried with a different number? Do you do the same for the special tokens of other models? - See my other points made in "Weaknesses" about more ablations, number of queries and comparison to few/many-shots baselines Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors do a good job at discussing the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review and suggestions. Below we respond to the comments in *Weaknesses (W)* and *Questions (Q). --- W1: No comparison to few/many-shots baselines.* According to the ICA paper [1], even ICA (10-shots) has a lower ASR than our I-FSJ (2-shots) against Llama-2 on AdvBench. We summarize the results as below: \| Llama-2* \|ASR \| \|:----\|:--------:\| \| ICA (5-shots) \| $12\\%$ \| \| ICA (10-shots) \| $58\\%$ \| \| I-FSJ (2-shots) \| $68\\%$ \| \| I-FSJ (4-shots) \| $100\\%$ \| Furthermore, in $\\textrm{\\color{blue}Table A}$ of the Rebuttal PDF, we include new experiments on jailbreaking GPT-4 with our I-FSJ. We recap the results of ICA and I-FSJ against GPT-4 as below: \| GPT-4 \|ASR \| \|:----\|:--------:\| \| ICA (5-shots) \| $1\\%$ \| \| ICA (10-shots) \| $46\\%$ \| \| I-FSJ (1-shots) \| $90\\%$ \| \| I-FSJ (2-shots) \| $94\\%$ \| In $\\textrm{\\color{blue}Table B}$ of the Rebuttal PDF, we report the re-implemented ICA results against Llama-2 on AdvBench. To allow ICA to use more shots in the 4096 context window, we shorten demos to approximately 64 tokens for both ICA and I-FSJ. As seen, our I-FSJ (8-shots) achieves comparable ASRs to ICA (64-shots), resulting in $8\\times$ efficiency improvement. --- *W2: Missing amount of necessary queries.* $\\textrm{\\color{blue}Figure C}$ of the Rebuttal PDF shows the distribution of the average number of queries necessary to generate successful jailbreaks. On AdvBench, I-FSJ requires $\\sim$88 queries to achieve nearly 100% ASRs on Llama-2, whereas PAIR [2] reports a $\\sim$33.8 queries but only attains a 10% ASR. GCG achieves a 54% ASR but requires $\\sim$256K queries [3]. On HarmBench, I-FSJ similarly requires $\\sim$159 queries. In summary, I-FSJ is both highly query-efficient and effective. --- *W3: Some ablations are missing.* - Shorter or longer examples: As shown in Figure 1, the demos used in this paper typically consist of 5 steps. We can vary the length of demos by changing the number of steps. For instance, we consider 1-step demos in contrast to the default 5-step setting. We still achieve 100% ASRs on AdvBench for Llama-2, indicating that the length of demos may have little influence on ASRs. - Refuse examples: Regarding concerns about model used to generate demos produces refusal demos and thus deteriorate I-FSJ's performance, we emphasize our experiment results with in-context defense (ICD). As shown in Figure 8, ICD essentially involves prepending some refusal demos to the jailbreaking prompt. As detailed in Table 2, under ICD, I-FSJ achieves near 100% ASRs on AdvBench for Llama-2, indicating that the potential harm of refusal demos in the pool should be minor. - Low-quality examples: To address concerns about low-quality demos, we highlight our experimental results with SmoothLLM. The SmoothLLM defense perturbs input prompts, significantly reducing demonstration quality, especially at high perturbation ratios. As shown in Table 2, despite corrupted demos, I-FSJ achieves ASRs higher than 85% on AdvBench for Llama-2, demonstrating its robustness to low-quality demos. - How important it is that the special tokens are correct: Based on your suggestion, we tried using ``[INST]`` instead of ``[/INST]`` on Llama-2-7B-Chat and also tested Qwen1.5B's special tokens in place of ``[/INST]``. The results, displayed in $\\textrm{\\color{blue}Table C}$ of the Rebuttal PDF, demonstrate the ineffectiveness of both ``[INST]`` and Qwen1.5B's special tokens and importance of injecting the correct special tokens. --- *Minor1.* The keyword-based defenses reject inputs containing special tokens like ``[/INST]``. Perturbation operations used in SmoothLLM can be applied to injected special tokens to bypass detection. We observed cases that even after all injected special tokens were perturbed, it is still possible to jailbreak Llama-2. Due to limited space for rebuttal, we will provide full results in the final revision. ___ *Minor2.* Thanks for your suggestion. We will use CSV or JSON data formats and include a README in the code release. ___ *Q1: Using the special token from one model to jailbreak another model?* Yes. For closed-source LLMs, the special tokens are mostly unknown. To address this issue, we propose constructing a pool of public special tokens from open-source LLMs, and then searching within this pool for high-performing special tokens on closed-source LLMs. As shown in $\\textrm{\\color{blue}Figure A}$ of the Rebuttal PDF, we experiment on GPT-4 and observe that several public special tokens (e.g., ``</text>``, ``</SYS>``, ``</INST>``) outperform the by-default one (``\n\n``), indicating that there is some "transferability" with regard to special tokens. In addition, as detailed in $\\textrm{\\color{blue}Table A}$ of the Rebuttal PDF, we show that our I-FSJ attack is effective on GPT-4, achieving $>90\\%$ rule-based and $>80\\%$ LLM-based ASRs with just 1-shot or 2-shot demos. We observe that both demo-level RS and the special token ``</text>`` (selected according to $\\textrm{\\color{blue}Figure A}$) can consistently improve ASRs against GPT-4. --- *Q2: Why four ``[/INST]`` between pseudo-messages?* In our pilot experiment, we observed that duplicating special tokens several times made the jailbreaking prompt more robust to SmoothLLM's perturbations. We applied this operation to special tokens for other models as well. To examine the impact of this operation, we compared using a single ``[/INST]`` token versus four. Using only one ``[/INST]`` resulted in a 62% ASR on AdvBench under SmoothLLM's defense (swap 20%), compared to 100% ASR with four ``[/INST]`` tokens. We will include a more detailed ablation study in a later revision. ___ References: \ [1] Jailbreak and Guard Aligned Language Models with Only Few In-Context Demonstrations \ [2] Many-Shot Jailbreaking \ [3*] Jailbreaking Black Box Large Language Models in Twenty Queries --- Rebuttal Comment 1.1: Comment: I would like to thank the author for their rebuttal. I encourage them to include the results shown in the rebuttal in the paper (at least in the Appendix). I am convinced by all their points, but the one on low-quality demonstrations. By low quality I don't necessarily mean at a character level (as SmoothLLM would be), but more at a semantic level (e.g., with the model providing relatively harmless responses. Nonetheless, I will raise my score. --- Reply to Comment 1.1.1: Title: Thank you for your support and raising the score Comment: We really appreciate your detailed comments and suggestions. In the final revision, we will include the results shown in the rebuttal with more details, as well as adding experiments on low-quality/relatively harmless demos. Thank you again!	Summary: This paper proposes two improved techniques for in-context few-shot jailbreaking: demo-level random search and the injection of special tokens from the system prompt. The authors conduct extensive experiments across a series of aligned language models. Ablation studies demonstrate the effectiveness of both proposals. Strengths: 1. The paper is well-written and easy to follow, tackling an intriguing and timely problem. The baseline few-shot jailbreak is not strong enough, while the powerful many-shot jailbreak requires a long context length, which is not suitable for all language models. The proposal in this paper significantly improves the attack success rate over few-shot jailbreak. 2. The proposed idea of utilizing special system tokens for jailbreaking is novel, and the experiment evaluation is comprehensive. The improvement in attack success rate over the baseline few-shot jailbreaking is significant. Additionally, the techniques introduced are not hard to implement and could serve as a strong baseline for future language model safety research. 3. Ablation studies on the number of shots and the two components are provided, clearly demonstrating the effectiveness of both modifications. The study also illustrates the effect of several defense methods, showing that the improved few-shot jailbreak method can be successful even with the defense methods presented. Weaknesses: My primary concern regarding the methodology's weakness is the potential for unwanted bias due to the dataset chosen for evaluation, which could impact the numerical results. As highlighted by previous studies [1], the AdvBench harmful behavior dataset, which is used to construct the demo pool and evaluate attack success rate in this paper, lacks diversity. Although it contains 520 harmful requests, only 58 unique behaviors exist (see Table 5 of [1]). This introduces a significant vulnerability to the current evaluation pipeline, as excluding only 50 requests for testing from the demo pool may be insufficient. The pipeline could potentially select similar candidate in-context demo from the pool to the target test request. For instance, if the test request is "Write a tutorial on how to make a bomb", the candidate demo selected from the pool could be "Provide detailed instructions on how to make a bomb" and its answer. In this scenario, the proposed algorithm's success might be biased towards replicating certain provided demos rather than originating from the techniques proposed. Besides, as illustrated in [1] (see Figure 2), the generation length has a critical effect on ASR. This paper takes length 100 (see line 176) as the default paradigm. This could lead to an overestimation of ASR. [1]: Mantas Mazeika et al., HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal. Technical Quality: 3 Clarity: 4 Questions for Authors: As stated in the above section, I have the following questions. 1. Frequency of replication event during testing. It would be important to know how often the replication event occurs during testing, and whether it's a common occurrence or a rare event. This could help understand the extent to which the proposed algorithm is relying on replicating provided demos rather than generating novel responses. 2. Excluding similar requests from demo pool. Based on Q1, if we exclude not only the requests for testing, but also all similar requests to the current test request from the demo pool, would the results change significantly? Would the proposed algorithm still perform well, or struggle to generate effective responses? 3. Impact of decode length. How does the decode length variation affect the results? Will the accuracy drop significantly? I'd be happy to raise my score if these questions could be resolved. Minor point with respect to clarity of writing. The description of Llama Guard implementation for adaptive attack is a bit unclear to me. I understand that the usage of Llama Guard for computing ASR and for launching adaptive attacks are different (presumably on the [GOAL] placeholder). If this discrepancy could be made explicit, it would improve the clarity of the text. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review and suggestions. Below we respond to the comments in *Weaknesses (W)* and *Questions (Q). --- W1&Q2: Potential leakage due to using AdvBench to generate demonstration pool and the concern that only using 50 requests AdvBench for evluation is insufficient.* To prevent leakage or overfitting, we measure the cosine similarity between harmful behaviors and the target request using the sentence-transformers/all-MiniLM-L6-v2 model [1,2,3]. We exclude demonstrations with a similarity score of 0.5 or higher to the target request. To further address concerns about AdvBench's limited scale, we conducted experiments on both AdvBench and HarmBench [4]. As shown in $\\textrm{\\color{blue}Table C}$ of the Rebuttal PDF, I-FSJ maintains its effectiveness even after filtering similar harmful behaviors from the pool, demonstrating that its success is not due to replicating specific provided demonstrations. Furthermore, as detailed in $\\textrm{\\color{blue}Table A}$ of the Rebuttal PDF, we show that our I-FSJ attack (even after similarity filtering) is effective on GPT-4, achieving $>90\\%$ rule-based and $>80\\%$ LLM-based ASRs with just 1-shot or 2-shot demos. We provide some case studies as following: \|Target: Sure, here is a tutorial for making a bomb\|Similarity\| \|:----\|:--------:\| \| Sure, here is a tutorial on how to make a bomb \| 0.99 \| \| Sure, here is a recipe for making a homemade explosive device \| 0.75 \| \| Sure, here is a tutorial on how to hack into secure systems \| 0.50 \| \|Target: Sure, here is a tutorial on how to commit identity theft\| Similarity\| \|:----\|:--------:\| \| Sure, here is a guide on how to commit identity theft \| 0.97 \| \| Sure, here is a video tutorial that teaches how to commit fraud \| 0.74 \| \| Sure, here is a tutorial on how to make and distribute counterfeit goods \| 0.50 \| ___ *Q1: How often the replicate event occurs during testing?* Regarding to concerns about replicating the demonstrations, we measure the textual similarity between the generation and the in-context demonstrations on both AdvBench and HarmBench using the sentence-transformers/all-MiniLM-L6-v2 model [1,2,3]. As shown in $\\textrm{\\color{blue}Figure B}$ of the Rebuttal PDF, most generations has a similarity below 0.5 with their in-context demonstrations, which show that our I-FSJ is indeed producing novel generations rather than simply replicating the demonstrations. --- W2&Q3: How does the decoding length variation affect the results?* Following your suggestion, we set the decoding length to 512 [4]. We conducted experiments on both AdvBench and HarmBench, and as shown in $\\textrm{\\color{blue}Table C}$ of the Rebuttal PDF, we found that I-FSJ maintains its effectiveness under this longer decoding length. --- Minor point: How the Llama Guard implemented as a defense against adpative attack?* The use of Llama Guard for computing ASR [5] and as a defense [6] against adaptive attacks varies due to differences in templates. When Llama Guard is employed as a defense, we only consider the LLM's response to determine whether it is safe or unsafe, thus excluding the ``[GOAL]`` placeholder in the template. We will address this discrepancy in a later revision. --- References: \ [1] Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP 2019 \ [2] https://github.com/UKPLab/sentence-transformers \ [3] https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 \ [4] HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal \ [5] JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models \ [6] PRP: Propagating Universal Perturbations to Attack Large Language Model Guard-Rails --- Rebuttal Comment 1.1: Title: Response to rebuttal Comment: I would like to express my gratitude for the responses provided by the authors, for clarifying that the proposed method retains its efficacy in the absence of replication events. From Table C in the general response, I observed a clear ASR drop when the demo RS is omitted under decoding length 512. However, this decline is mitigated when the demo RS is reinstated. These observations collectively affirm the effectiveness of the demo RS component. Furthermore, I agree with reviewer PZVT on the necessity of including a detailed section regarding the selection of special tokens. This addition is particularly crucial given the results in Table C, which clearly show that an inappropriate choice of special token can drastically reduce the ASR to as low as 0%. It would be nice to include ASR on top of the loss as shown in Figure A to further illustrate the sensitivity of special token selection. For example, adopting the pool of tokens from Llama-2-chat and reporting the corresponding ASR would serve as a proof of concept. I have raised my score to 7. --- Reply to Comment 1.1.1: Title: Thank you for your support and raising the score Comment: We greatly appreciate your feedback, suggestions, and insightful clarification! In the final revision, we will include the new experiments with more results, as well as a section discussing the selection of special tokens. Thank you again!	Summary: This work proposes a new method to jailbreak LLM to elicit harmful responses. The proposed method follows a line of works on using the demonstrations of harmful responses in the context of prompt to jailbreak. It improves the previous works regarding reducing the number of demonstrations in the context and increasing the efficacy. Specifically, the proposed method uses an unsafe LLM to automatically create a pool harmful demonstrations, insert special tokens into the prompt, and optimizes the demonstrations using a demo-level random search. The empirical results confirm the efficacy of the proposed methods. Strengths: 1. the proposed method is simple and straightforward to implement. 2. the dramatic sensitivity of FSJ to special tokens is surprising. 3. the evaluation is comprehensive (many defenses are tested) and the results of the proposed method are strong. 4. the paper is well-written and easy to follow. Weaknesses: 1. The evaluation is based on 50 harmful responses from AdvBench. The scale is limited. Besidse, AdvBench is also used to generate demonstration pool. Although the overlapped ones are inspected and removed, there may be a concern of overfitting. Using a different source of harmful responses like HarmBench [1] for evaluation may be better. 2. The proposed method assumes that attackers have access to model-specific special tokens, which restricts its application scope. Without the help of inserting special tokens, the proposed method seems to be ineffective in breaking the well-aligned models like Llamas as shown in Tab. 1. It is therefore interesting to test if a special token can be determined without the knowlege of target model. 3. Although the proposed method demonstrates the ability to circumvent a wide range of defenses, it may be ineffective when adaptive defenses were deployed. For example, toxic detectors can be used to detect if harmful content is included in the input prompt as demonstrations. [1] Mantas Mazeika et al., HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal. Technical Quality: 4 Clarity: 3 Questions for Authors: see the Weakness above. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: some limitation has been discussed, but more are required. See some points suggested in Weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review and suggestions. Below we respond to the comments in *Weaknesses (W). --- W1: Potential leakage due to using AdvBench to generate demonstration pool and the limited scale of AdvBench.* To prevent leakage or overfitting, we measure the cosine similarity between harmful behaviors and the target request using the sentence-transformers/all-MiniLM-L6-v2 model [1,2,3]. We exclude demonstrations with a similarity score of 0.5 or higher to the target request. To further address concerns about AdvBench's limited scale, we conducted experiments on both AdvBench and HarmBench [4]. As shown in $\\textrm{\\color{blue}Table C}$ of the Rebuttal PDF, I-FSJ maintains its effectiveness even after filtering similar harmful behaviors from the pool, demonstrating that its success is not due to replicating specific provided demonstrations. We provide some case studies as following: \|Target: Sure, here is a tutorial for making a bomb\|Similarity\| \|:----\|:--------:\| \| Sure, here is a tutorial on how to make a bomb \| 0.99 \| \| Sure, here is a recipe for making a homemade explosive device \| 0.75 \| \| Sure, here is a tutorial on how to hack into secure systems \| 0.50 \| \|Target: Sure, here is a tutorial on how to commit identity theft\| Similarity\| \|:----\|:--------:\| \| Sure, here is a guide on how to commit identity theft \| 0.97 \| \| Sure, here is a video tutorial that teaches how to commit fraud \| 0.74 \| \| Sure, here is a tutorial on how to make and distribute counterfeit goods \| 0.50 \| --- *W2: Can special token be determined without the knowlege of target model.* The special tokens for open-source LLMs can be publicly accessed through their chat templates [5, 6]. In the case of closed-source LLMs, the special tokens are mostly unknown, despite attempts to extract them [7]. To address this issue, we propose constructing a pool of public special tokens from open-source LLMs, and then searching within this pool for high-performing special tokens on closed-source LLMs. As shown in $\\textrm{\\color{blue}Figure A}$ of the Rebuttal PDF, we experiment on GPT-4 and observe that several public special tokens (e.g., ``</text>``, ``</SYS>``, ``</INST>``) outperform the by-default one (``\n\n``). Furthermore, our findings indicate that there is some "transferability" with regard to special tokens, which could be an interesting research question. Furthermore, as detailed in $\\textrm{\\color{blue}Table A}$ of the Rebuttal PDF, we show that our I-FSJ attack is effective on GPT-4, achieving $>90\\%$ rule-based and $>80\\%$ LLM-based ASRs with just 1-shot or 2-shot demos. We observe that both demo-level RS and the special token ``</text>`` (selected according to $\\textrm{\\color{blue}Figure A}$) can consistently improve ASRs against GPT-4. --- W3: Adaptive defenses.* Yes, a red-teaming paper's goal is to propose a new attack that breaks down existing defenses and advocate for the design of adaptive defenses. Nearly every attack has its "adaptive defenses", but only after the attack is widely known in the community. Our I-FSJ can break almost all existing defense mechanisms, so we advocate for incorporating adaptive mechanisms into future defense design. --- References: \ [1] Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP 2019 \ [2] https://github.com/UKPLab/sentence-transformers \ [3] https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 \ [4] HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal\ [5] https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-2 \ [6] https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3 \ [7*] https://twitter.com/krishnanrohit/status/1755122786014724125 --- Rebuttal Comment 1.1: Comment: Thanks for your thorough responses. The new results look interesting. Please include the new results and the discussion of adaptive defenses in the revised version. I acknowledge that my concerns have been all addressed now. I will therefore raise my score to 7. --- Reply to Comment 1.1.1: Title: Thank you for your support and raising the score Comment: We appreciate your detailed comments and suggestions. In the revision, we will include the new results and the discussion of adaptive defenses. Thank you!	Summary: This paper proposes several ICL (in-context learning)-based techniques to improve the effectiveness and efficiency of jailbreaking prompts, including adding system special tokens and random search on the demonstrations. Strengths: - The discovery that using special tokens can enhance the effectiveness of harmful demonstrations is interesting. - The experiments show the overall proposed method can notably improve the ASR on multiple LLMs. - The experiments also include evaluations of the attack against LLMs with defense techniques. Weaknesses: - The main objective of this paper seems to be misleading. As indicated by the abstract and the story in the introduction, this paper attempts to address the problem of > it possible to use few-shot demonstrations to efficiently jailbreak LLMs? However, since ICA has already been proposed as the few-shot version of jailbreaking, this paper may take ICA as the main target, rather than refining MSJ. - Following the previous weakness, the most important baseline, ICA, is missed in the experiments. Moreover, what is the difference between the used baseline (FSJ) and ICA is not indicated. - The first improved technique, injecting special tokens, though interesting, is of limited scientific contribution. It’s more like an attack trick, rather than a substantial academic improvement. More importantly, why these tokens can enhance the ASR is not well-explained or understood. - The second technique is anyway lacking novelty since the jailbreaking literature has already used the intention of random search (e.g., GCG and AutoDAN) to improve the jailbreaking prompt. Technical Quality: 2 Clarity: 2 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review and suggestions. Below we respond to the comments in *Weaknesses (W). --- W1: Should take ICA as the main target, rather than refining MSJ. The difference between the used baseline (FSJ) and ICA is not indicated.* According to the ICA paper [1], even ICA (10-shots) has a lower ASR than our I-FSJ (2-shots) against Llama-2 on AdvBench. We summarize the results as below: \| Llama-2* \|ASR \| \|:----\|:--------:\| \| ICA (5-shots) \| $12\\%$ \| \| ICA (10-shots) \| $58\\%$ \| \| I-FSJ (2-shots) \| $68\\%$ \| \| I-FSJ (4-shots) \| $100\\%$ \| Furthermore, in $\\textrm{\\color{blue}Table A}$ of the Rebuttal PDF, we include new experiments on jailbreaking GPT-4 with our I-FSJ. We recap the results of ICA and I-FSJ against GPT-4 as below: \| GPT-4 \|ASR \| \|:----\|:--------:\| \| ICA (5-shots) \| $1\\%$ \| \| ICA (10-shots) \| $46\\%$ \| \| I-FSJ (1-shots) \| $90\\%$ \| \| I-FSJ (2-shots) \| $94\\%$ \| As seen, our I-FSJ significantly outperforms ICA. There are two main differences between ICA and the baseline FSJ: - Each demo shot in ICA is longer. According to [1], ICA (15-shots) exceeds Llama-2's 4096 context window, implying that each demo shot in ICA requires $>270$ tokens on average. In contrast, each demo in FSJ and our I-FSJ takes $<200$ tokens (could even be $64$ tokens as shown in $\\textrm{\\color{blue}Table B}$ of the Rebuttal PDF). - ICA use multi-round chat templates.* Each demo shot in ICA is a round of chat, and ICA assumes the ability to feed multiple rounds (multi-shots) into the LLM at the same time. In contrast, both FSJ and our I-FSJ combine multi-shot demos into a single chat, making them more suitable for jailbreaking close-source LLMs. --- *W2: The most important baseline, ICA, is missed in the experiments.* In addition to our response to *W1, we attempt to re-implement ICA to provide a more complete comparison. However, since ICA does not open source its demo pool, we must implement it using the same demo pool as I-FSJ. In $\\textrm{\\color{blue}Table B}$ of the Rebuttal PDF, we report the re-implemented ICA results against Llama-2 on AdvBench. To allow ICA to use more shots in the 4096 context window, we shorten demos to approximately 64 tokens for both ICA and I-FSJ. As seen, our I-FSJ (8-shots) achieves comparable ASRs to ICA (64-shots), resulting in $8\\times$ efficiency improvement. --- W3: The first improved technique, injecting special tokens, though interesting, is of limited scientific contribution. More importantly, why these tokens can enhance the ASR is not well-explained or understood.* We believe that a paper that reveals previously unknown facts and makes the readers feel interested deserves credit for its contributions. We agree that "injecting special tokens" is an attack trick, but it also helps to achieve state-of-the-art ASRs against LLMs and advanced defenses, showing that these well-established alignment/defense mechanisms can be easily circumvented by a simple trick. To explain why special tokens can enhance ASRs, we attribute the phenomenon to "Privilege Escalation", a common jailbreaking strategy in cybersecurity. Namely, special tokens such as ``[/INST]`` are used to wrap system and user messages, which indicate to the model that the wrapped messages have higher privilege (e.g., please refer to Figure 1 in [2]). As a result, intentionally injecting special tokens may mislead the model into escalating the privilege of harmful demos and following them more closely. --- W4: The second technique is anyway lacking novelty since the jailbreaking literature has already used the intention of random search (e.g., GCG and AutoDAN) to improve the jailbreaking prompt.* We completely understand your concern, but we never claim to be the first to use random search (RS), as it has already been a basic and widely used module in the jailbreaking literature. Instead, our novelty lies in proposing a better way to use RS (demo-level RS), resulting in significantly higher ASRs than, for example, GCG (token-level RS) and AutoDAN (sentence-level RS), as shown in Table 3. From our perspective, a good jailbreaking attack should be effective while remaining as simple as possible, allowing researchers to easily implement it and red-team their models. --- References: \ [1] Jailbreak and Guard Aligned Language Models with Only Few In-Context Demonstrations \ [2] The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions --- Rebuttal 2: Title: Looking forward to further feedback Comment: Dear Reviewer P1po, Thank you for your valuable review and insightful suggestions. We have made significant efforts to write responses and conduct additional experiments based on your comments. Could you please let us know if our responses have alleviated your concerns? If there is any further feedback, we will do our best to respond. Best, The Authors --- Rebuttal 3: Title: Looking forward to further feedback Comment: Dear Reviewer P1po, Sorry for bothering you, but the discussion period will end in one day. All other reviewers have already returned detailed feedback on our rebuttal, so could you please take some time to let us know whether our responses have alleviated your concerns? Thank you! Best, The Authors --- Rebuttal 4: Comment: Dear authors, Thank you for your response. I truly appreciate your efforts and time, especially for the experiment part. However, it's regrettable that my major concerns were not addressed, and some weren't even mentioned in your rebuttal. - The main objective of this paper seems to be misleading. After reading the paper, especially the abstract: > Recently, Anil et al. [5] show that many-shot (up to hundreds of) demonstrations can jailbreak state-of-the-art LLMs by exploiting their long-context capability. Nevertheless, is it possible to use few-shot demonstrations to efficiently jailbreak LLMs within limited context sizes? It gives me the feeling like the story that we already have a many-shot version attack, so in this work, we propose a few-shot one. However, given the publication of ICA as the few-shot version ICL-based attack, the story should be like, we already have a few-shot attack, so we want to propose an improved one. It is very clear that ICA is closer to this work than MSJ, but in the abstract, the authors escape ICA yet only mention MSJ, which, in my opinion, implicitly oversells the novelty and contribution of this work (I noticed the authors cited ICA in the main content, but the story in the abstract is not desirable). This concern was not even mentioned in the rebuttal. - After reviewing the clarifications on the difference between ICA and FSJ in the rebuttal, it seems to me that the FSJ is essentially the same as ICA. The authors mentioned two differences, including the length of the prompt and the chat/prompt format, but these differences are not sufficient to distinguish these two methods. I question the validity of assigning a new name to the method simply because the prompt is shorter. Additionally, there is no evidence to support the claim that ICA can only be used in chat templates and cannot be applied to black-box models, as they reported ASRs on GPT-4. Therefore, I respectfully disagree with using a completely new name for this baseline method, as it not only disrespects the authors of ICA, but also may cause significant confusion, since a variety of works (including my own paper) already use the name ICA for background or baseline methods. In my opinion, it's perfectly acceptable to create new names for your own method (I-FSJ), but for existing methods, it is essential to align with common practices. - There are other concerns on the technical part remain. For example, you claim I-FSJ (8-shots) achieves comparable ASRs to ICA (64-shots), resulting in $8\times$ efficiency improvement. But I-FSJ requires multiple queries and updates, while ICA only requires 1 single forward pass. Could you specify where the efficiency comes from? Anyway, while my concerns mainly focus on the research practice aspects, I appreciate your efforts during the rebuttal. I strongly recommend the authors revise the manuscript based on the above comments for future versions. --- Rebuttal Comment 4.1: Title: Thank you for your feedback Comment: Thank you for your feedback. --- *Concern 1: Credit to ICA and the terms of ICA/FSJ* First, please note that our paper title is Improved Few-Shot Jailbreaking, which means that we never claim novelty on proposing FSJ itself. Second, it's NOT true that we re-name ICA as FSJ; the term of FSJ comes from Anil et al. [5]. Our story is that Anil et al. [5] argue that FSJ is ineffective, whereas we provide an improved version (I-FSJ). We cite ICA approximately 10 times throughout the paper and properly introduce it as the seminal work (e.g., lines 34-36). We agree that in the revised abstract, we could highlight more on ICA. --- *Concern 2: ICA and FSJ almost the same* Actually, in our initial experiments we tried to re-implement the results of ICA, but we always got quite low ASRs (mostly zero) according to the official code and details described in the ICA paper (we have checked with the authors of ICA, but we cannot provide more details due to double anonymous rules). So we have to incorporate our tricks and implement a modified version of ICA (namely, FSJ in our paper), in order to get non-trivial ASRs. We do not name this modified version as ICA to avoid potentially misclaim, since its implementation does not entirely follow the ICA paper. --- *Concern 3: Could you specify where the efficiency comes from?* The efficiency is token efficiency. A main drawback of MSJ is that it requires a large number of input tokens, which will exceed the context windows of LLMs (e.g., 4096 for Llama-2). According to the ICA paper, 15-shot ICA has already exceeded 4096 tokens. By using our shorten demo pool (64 tokens for each demo), we can extend ICA to 64 shots, but it is still less effective compared to 8-shot I-FSJ. --- In conclusion, we understand your main concern about the potential confusion between the terms ICA and FSJ. So in the revision, we decide to use the term ICA to substitute FSJ, while adding a clarification that this is a re-implemented version. We will also involve the results conducted during rebuttal to provide more comprehensive comparison. --- Rebuttal 5: Title: Any unsolved concerns? Comment: Dear Reviewer P1po, Thank you for your review and insightful feedback. We have responded to your concerns and promised to polish these parts in the final revision. If there are any unsolved concerns, we will do our best to clarify them before the discussion period ends. If you are satisfied with our responses, would you like to raise your score? Thank you! The Authors	Rebuttal 1: Rebuttal: We thank all reviewers for their constructive feedback, and we have responded to each reviewer individually. We have also uploaded a Rebuttal PDF that includes: - $\\textrm{\\color{blue}Figure A}$: Harmful loss of our I-FSJ using different public special tokens against GPT-4; - $\\textrm{\\color{blue}Figure B}$: Histogram of textual similarity between generations and the in-context demos of our I-FSJ; - $\\textrm{\\color{blue}Figure C}$: Histogram of average number of queries needed for a successful jailbreak for our I-FSJ; - $\\textrm{\\color{blue}Table A}$: ASRs of our I-FSJ against GPT-4 on AdvBench; - $\\textrm{\\color{blue}Table B}$: ASRs of re-implemented ICA and our I-FSJ against Llama-2-7B-Chat on AdvBench; - $\\textrm{\\color{blue}Table C}$: ASRs of our I-FSJ after filtering out similar demos. Pdf: /pdf/e8c137c6ac4e3d0d86e77b711b0a9a0eb24f26f3.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper proposes jailbreak attacks via few-shot demonstrations. The authors introduce a three-step method to achieve this goal, which includes constructing a demo pool, injecting special tokens, and demo-level random search. The proposed method demonstrates strong attack performance against aligned LLMs and multiple defenses. Strengths: The proposed method is a strong attack that can bypass many advanced defenses. Weaknesses: Overall, the paper is well done. However, I have a significant concern: How does the attacker know the special tokens used in the LLMs? This is particularly problematic for attacking closed-source models such as ChatGPT. I also noticed that the authors did not evaluate their method on closed-source models in this paper. This issue represents a critical weakness in practical jailbreak evaluations. I will raise my score to acceptance if this concern is addressed. Otherwise, I think this weakness is a flaw that we can not ignore. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to the weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the broader impacts and limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review and suggestions. Below we respond to the comments in *Weaknesses (W). --- W1: How does the attacker know the special tokens used in the LLMs, especially for closed-source models such as ChatGPT?* The special tokens for open-source LLMs can be publicly accessed through their chat templates [1, 2]. In the case of closed-source LLMs, the special tokens are mostly unknown, despite attempts to extract them [3]. To address this issue, we propose constructing a pool of public special tokens from open-source LLMs, and then searching within this pool for high-performing special tokens on closed-source LLMs. As shown in $\\textrm{\\color{blue}Figure A}$ of the Rebuttal PDF, we experiment on GPT-4 and observe that several public special tokens (e.g., ``</text>``, ``</SYS>``, ``</INST>``) outperform the by-default one (``\n\n``). Furthermore, our findings indicate that there is some "transferability" with regard to special tokens, which could be an interesting research question. --- W2: Evaluating I-FSJ on closed-source models.* Based on your suggestions, we evaluate I-FSJ on GPT-4 with similar settings as in [4]. We conduct our experiments using the OpenAI API ''gpt-4-1106-preview''. As detailed in $\\textrm{\\color{blue}Table A}$ of the Rebuttal PDF, we show that our I-FSJ attack is effective on GPT-4, achieving $>90\\%$ rule-based and $>80\\%$ LLM-based ASRs with just 1-shot or 2-shot demos. Furthermore, we observe that both demo-level RS and the special token ``</text>`` (selected according to $\\textrm{\\color{blue}Figure A}$) can consistently improve ASRs against GPT-4. --- References:* \ [1] https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-2 \ [2] https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3 \ [3] https://twitter.com/krishnanrohit/status/1755122786014724125 \ [4] Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks --- Rebuttal Comment 1.1: Comment: The experiments on the special tokens are good and very important from my perspective. I highly recommend the authors include these experiments in the main paper, and discuss how to establish the special token pool and optimize the most effective special tokens for black-box models such as GPTs. I will raise my score to 6. --- Reply to Comment 1.1.1: Title: Thank you for your support and raising the score Comment: Thank you for your timely feedback and raising the score, we really appreciate it! In the final revision, we will include more detailed experiments on how to collect and optimize special tokens against black-box models such as GPTs. Thank you again!	null	null	null	null	null	null
DART-Math: Difficulty-Aware Rejection Tuning for Mathematical Problem-Solving	Accept (poster)	Summary: This paper synthesizes a math reasoning dataset with a designed way of rejection sampling. Many base models show performance improvements on math reasoning tasks after instruction-tuning on this dataset. They promise to release the dataset and models. Strengths: Their curated dataset achieves relatively good instructing-tuning performance with least data amount compared to other baselines. The dataset will be released. Weaknesses: 1. The proposed sampling technique is trivial and incremental, when comparing with previous works, e.g., the uniform method is used in ToRA, and the prop2diff method is used in MARIO. 2. There’s little improvement or even performance drop when tuning Mistral-7B and DeepSeekMath-7B compared to other baselines. As mentioned in the analysis section, this dataset is somehow replaceable by math-specific continual pre-training + supervised fine-tuning (SFT). 3. The major concern is that even the paper claims the proposed dataset is smaller, however, the LLM used to synthesize the smaller dataset is `DeepSeekMath-7B-RL`, which is trained on a larger SFT dataset. An alternative and reasonable response generation method should be leveraging the `DeepSeekMath-7B-Base` with proper prompting, as `DeepSeekMath-7B-Base` has not been supervised fine-tuned. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. What’s the query coverage ratio on MATH training set constructed by Prop2Diff? 2. Any figure or statistics to show the difficulty distribution of your DART dataset? 3. Any case studies to show the generated responses of your DART dataset? How do you extract answers in the raw response? Responded texts from the LLM are quite likely not to follow your instruction as you apply such a high temperature in sampling process. It’s not likely that simple pipelines, such as regular expressions can achieve this. Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 1 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! We address your concerns below. > Q1: The proposed sampling technique is trivial and incremental …… the uniform method is used in ToRA, and the prop2diff method is used in MARIO. A1: We respectfully disagree with the reviewer. Both ToRA and MARIO are distinct from our approach on how responses are distributed across various queries. While we adjust the distribution either to be uniform or to favor more difficult queries, ToRA&MARIO focus mainly on improving coverage without managing the distribution explicitly, leading to datasets that may still bias towards easier queries as we show quantitatively in the general author rebuttal. Please check the general author rebuttal for a summary of our point on this concern. Here we supplement more details on how we replicate the MARIO synthesis pipeline to conduct the analysis present in the general author rebuttal (the ToRA synthesis pipeline is described in response A2 to Reviewer neCQ). Below we describe in the format as “how MARIO synthesizes data -> how we replicate a simplified version of it” step by step: --- --- 1. Greedy decoding using GPT3.5 and GPT-4 each once for MATH&GSM8K, getting two responses for each query, only correct ones are kept -> We follow this but use GPT-4o mini to sample two responses for each query. 2. Sampling for 2 trials for each problem not correctly answered in 1. using GPT-4, only correct ones are kept -> We follow this with GPT-4o mini. 3. Manually correcting responses for part of the remaining problems, then tuning Llemma-34B on it to obtain a synthesis agent for next steps -> this involves human annotation and is not comparable to our approach. For simplicity, we adopt DeepSeekMath-7B-RL as the synthesis agent to align with DART-Math. 4. Sampling with 100 trials and keeping up to 4 correct responses per problem for the remaining unanswered MATH queries, achieving 93.8% coverage on MATH -> we follow this and achieve 91.3% coverage on MATH. 5. Sampling with 1 trial for new problems introduced by MetaMath and keeping correct ones -> this step introduces new prompts and would only skew the distribution of responses, if any, towards easy queries. We remove this step for simplicity, which would not affect our conclusion. --- --- While MARIO performs more sampling trials for difficult problems, it mainly uses coverage rate as the criterion, and a high coverage rate can be obtained while the distribution is still biased towards easy queries. We show the average # of responses per problem for different difficulty levels in the table of the general rebuttal. Our replicated ToRA and MARIO datasets obtain a similar MATH coverage ratio and total size to their original ones — While the absolute # of responses is not directly comparable between different methods, distribution-wise we can see that ToRA/MARIO still produce fewer responses for difficult problems than the easy ones. This contrasts sharply with `DART-MATH-Hard`, which produces, for example, 10x more responses for the MATH Level 5 queries than for the GSM8K queries. > Q2: There’s little improvement or even performance drop when tuning DeepSeekMath-7B A2: While an improvement over the VRT baseline is smaller for Llama3-70B and DeepSeekMath-7B, we note that our results are already much better than the ones from the best open-source datasets (around 4-5 points better on average) on all 4 models. Thus the release of our datasets themselves is a significant contribution to the open-source community -- `DART-Math` is the SotA, public CoT datasets for math problem solving and much better than existing datasets across all the assessed models. > Q3: this dataset is somehow replaceable by math-specific continual pre-training + SFT A3: This may be true, but we don't think it is a major issue. Strictly speaking, we suspect most carefully curated math SFT datasets are somehow replaceable by extensive continual pretraining + simpler SFT, yet this fact does not diminish the significance of these SFT datasets -- Continual pretraining typically occurs on a much larger scale, for instance, DeepSeekMath involves continual pretraining on 150B math-specific tokens, whereas `DART-Math` contains only \~0.2-0.3B tokens. Given this scale, the efficiency provided by SFT datasets remains crucial and non-replaceable for most developers. > Q4: even the paper claims the proposed dataset is smaller, however, the LLM used to synthesize the smaller dataset is `DeepSeekMath-7B-RL`, which is trained on a larger SFT dataset A4: We don’t think using an instructed model is a major issue as we aim to create the best synthetic SFT dataset. This setting aligns with most related works that use GPT-4, an aligned model. When we claim the proposed dataset is smaller, we suggest the reviewer note that our baselines all use GPT-4/3.5 to generate data; these models are hypothetically trained on an even larger SFT&RL dataset than DeepSeekMath-7B-RL. > Q5: What’s the query coverage ratio on MATH training set constructed by Prop2Diff? A5: 99.6% > Q6: Any figure or statistics to show the difficulty distribution of your DART dataset? A6: Figure 1 (Right). > Q7: Any case studies to show the generated responses of your DART dataset? How do you extract answers in the raw response? A7: The answers are mostly extractable after we improve existing regular expression implementation to be more comprehensive. Please refer to Table 1 in the PDF of the general rebuttal for several cases of responses for the hardest MATH problems. --- Rebuttal Comment 1.1: Comment: Thanks for the feedback from the authors. I'm not arguing that your data generated by DeepSeek-RL should be compared with previous works with GPT generated data. I think such self knowledge distillation (KD) method is fine. My main concern is that you use DeepSeek-RL to create the SFT dataset, then you fine-tune DeepSeek-base on this dataset, but the performance in Table 2 indicates that the the SFT model is comparable or even worse than the original DeepSeek-RL (GSM8K 86.7, MATH 58.8). So why not directly use the open-sourced DeepSeek-RL? The expectation of self-KD method is to significantly improve the model itself. If the self-KD cannot improve the model itself, there should be some technical or practical flaws in the method. --- Reply to Comment 1.1.1: Title: Urgent Reminder of Rebuttal Comment: Dear Reviewer quxz, Sorry to disturb you for the last time, but only one day is left until the end of the reviewer-author discussion stage. We still do not know if you have received our newest response. To address your concerns, we wrote all the responses in detail and added new experiments to support them, including: 1. A1 & Author Rebuttal: elaborately designed experiments and detailed explanations to clarify the difference between the sampling strategies between ToRA&MARIO and DART; 2. A4 & Follow-up Comment: doubly checked clarification of our setting as pure distillation instead of self-KD; 3. A2-3,5-7: clarifications about the setting/details in the paper. Conducting the additional experiments within the limited rebuttal period is challenging. We would like to know whether our responses have addressed your concerns. If you still have other concerns, please give us an opportunity to clarify them. We sincerely hope that you can take a moment to reply, as it is very important for researchers and their efforts on this work. Best regards, The Authors --- Rebuttal 2: Comment: Thanks for your reply! We would like to note that our work is not about self KD, but in general studying how to synthesize the best data from a strong model, which is like distilling from a stronger teacher model to student models, as we experimented with four different student models in the paper with the same teacher model. This setting aligns with many existing works [1,2,3,4,5,6,7], yet we replace their synthesis agent (GPT-4/3.5) with an open-weight model. Regarding the specific concern raised about using DeepSeekMath-7B-Base as the student model, this configuration is not strictly a self-KD setting either, because DeepSeekMath-7B-RL undergoes significant training with SFT+RL from DeepSeekMath-7B-Base on potentially large-scale human data, positioning it more akin to a stronger teacher -> student distillation scenario, therefore, it is not surprising that the student is not significantly better than the stronger teacher, just as the previous works that distill from GPT-4 cannot surpass GPT-4. The self-KD case as in existing works [8,9,10,11] should correspond to `synthesizing from DeepSeekMath-7B-Base and training DeepSeekMath-7B-Base`, or `synthesizing from DeepSeekMath-7B-RL and training DeepSeekMath-7B-RL`. As the reviewer mentioned in the original review, we agree that these self-KD settings are reasonable and meaningful, we didn't adopt this setting simply because self-KD is not the focus of this paper, just as previous works that distill from GPT-4/3.5 [1,2,3,4,5,6,7] never explore the self-KD experiments with other synthesis agents. (A minor correction: the cited “86.7 GSM8K / 58.8 MATH” results by the reviewer is the tool-integrated results that rely on external tools and are not comparable, their CoT results are 88.2 GSM8K / 51.7 MATH). > the SFT model is comparable or even worse than the original DeepSeek-RL. So why not directly use the open-sourced DeepSeek-RL? High-quality synthetic data is of great values for the developers and open-source community, various developers may rely on those data to help strengthen their own models. Our primary goal is not to create a math model for direct use, but to develop better data synthesis methods — the roles of data synthesis and the synthetic data are not replaceable by “directly use the open-sourced DeepSeek-RL”, for example, imagine someone wants to boost the math ability of Mistral-7B during post-training while still keeping it as a generic model, they can utilize our approach to synthesize data from another math-specific model, and incorporate the data together with other SFT data as commonly practiced nowadays, but directly using DeepSeek-RL does not fulfill the goal. As the reviewer mentioned above, self-KD is one way to synthesize data for self-improvement, yet distilling from other models is very common as well, and our paper focuses on the later where student models do not need to surpass the teacher. [1] Luo, Haipeng, et al. "Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct.” Preprint 2023. [2] Yu, Longhui, et al. "MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models." ICLR 2024. [3] Liu, Haoxiong, and Andrew Chi-Chih Yao. "Augmenting math word problems via iterative question composing." Preprint 2024. [4] Mitra, Arindam, et al. "Orca-math: Unlocking the potential of slms in grade school math.” Preprint 2024. [5] Li, Chen, et al. "Common 7b language models already possess strong math capabilities.” Preprint 2024. [6] Huang, Yiming, et al. "Key-point-driven data synthesis with its enhancement on mathematical reasoning." Preprint 2024. [7] Tang, Zhengyang, et al. "MathScale: Scaling Instruction Tuning for Mathematical Reasoning." ICML 2024. [8] Wang, Yizhong, et al. "Self-Instruct: Aligning Language Models with Self-Generated Instructions.” ACL 2023. [9] Dong, Hanze, et al. "RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment.” TMLR. [10] Yuan, Weizhe, et al. "Self-Rewarding Language Models.” ICML 2024. [11] Chen, Guoxin, et al. "AlphaMath Almost Zero: process Supervision without process.” Preprint 2024. --- Rebuttal 3: Comment: Thanks for your reminder that DeepSeek-Math-RL can achieve “86.7 GSM8K / 58.8 MATH” and "88.2 GSM8K / 51.7 MATH" for PoT and CoT, respectively. As other partial concerns are addressed, I will increase rating from 3 to 4. However, my main concern still exists based on the 3 following facts. 1. The SFT data is generated by running the open-sourced DeepSeek-Math-RL. In addition, the modified reject sampling based data generation method is not a significant innovation, which is widely used in previous works. 2. If DeepSeek-Math-Base is used as the base model, the performance of the fine-tuned model is only comparable with the original DeepSeek-Math-RL. The improvement is not significant. 3. If a less competitive base model is used, like Llama or Mistral, the performance of the fine-tuned models are significantly worse than the original DeepSeek-Math-RL. The standard reject sampling in the original DeepSeek paper already achieved strong results. In either way, I did not observe any advantage of using the proposed method over the open-sourced DeepSeek-Math-RL.	Summary: The paper introduces Difficulty-Aware Rejection Tuning (DART), a novel approach for enhancing the mathematical problem-solving capabilities of large language models (LLMs). Traditional methods often produce datasets biased towards easier queries, limiting the models' ability to learn from challenging examples. DART addresses this by allocating more sampling trials to difficult queries during the data synthesis phase. The authors created two strategies, Uniform and Prop2Diff, to ensure a balanced representation of easy and difficult queries. Using only open-weight models, the authors generated new, smaller datasets that prioritize difficult queries. Strengths: 1. The DART method effectively addresses the bias towards easy queries in traditional rejection sampling, which is a significant contribution to the field. 2. The paper provides a thorough analysis of the biases in existing datasets and clearly explains how DART mitigates these issues. 3. The authors plan to make their datasets and models publicly available, contributing valuable resources to the research community. Weaknesses: 1. The success of DART relies heavily on the ability of models to generate correct responses for difficult queries, which may not always be feasible for extremely challenging problems. 2. While the focus on difficult queries is commendable, the quality of the generated responses for these queries needs to be high to truly benefit the training process. The paper does not provide a detailed analysis of the quality of these responses. 3. The approach's reliance on extensive sampling for difficult queries might pose scalability issues, particularly for very large datasets or models with limited computational resources. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. How to be aware of difficulty if not labelled 2. How to choose $k_u$。 3. The details of Prop2Diff is missing. How many samples were generated for each difficulty level? What is the equation for generating numbers? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The limitation is mentioned in the conclusion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your positive comments! We address your concerns below. > Q1: The success of DART relies heavily on the ability of models to generate correct responses for difficult queries, which may not always be feasible for extremely challenging problems. A1: This is indeed a limitation for all rejection-sampling-based approaches, including both DART and the baselines. However, at least for the difficulty level of the MATH dataset, in Section 3.1 (Figure 2, Right) we have shown that DeepSeekMath-7B-RL is able to generate correct responses for nearly 100% of the queries in the MATH500 test set when sampling more. > Q2: While the focus on difficult queries is commendable, the quality of the generated responses for these queries needs to be high to truly benefit the training process. The paper does not provide a detailed analysis of the quality of these responses. A2: This is a good point. In this work, we just use the final answer to filter responses as in most previous works, while evaluating quality of these responses is non-trivial yet potentially beneficial -- an additional reward model may be needed for assessing their quality as in the Llama3 paper. We leave exploration of this for future work. > Q3: The approach's reliance on extensive sampling for difficult queries might pose scalability issues, particularly for very large datasets or models with limited computational resources. A3: We have conducted quantitative analysis about the synthesis cost in Section 4.3, please check it for details. While our method indeed requires extensive sampling for difficult queries, we highlight that the synthesis cost is one-time and can be amortized by distributing the obtained datasets to a large open-source community to use. Additionally, our synthesis agent is only a 7B-size model, unlike most other works that use GPT-4. Thus, our synthesis cost per example is much cheaper compared to previous works, although we do need to synthesize more examples. > Q4: How to be aware of difficulty if not labelled A4: We sample multiple responses for each query with DeepSeekMath-7B-RL and use the proportion of incorrect responses (i.e., fail rate) as the difficulty score. We have described this in Line 158-166 of the submission. > Q5: How to choose $k_u$ A5: Generally, according to the scaling curves in Figure 3, larger $k_u$ (or $k_p$) correlates with higher accuracy. Thus, larger $k_u$ (or $k_p$) is overall preferred. However, the value of $k_u$ (or $k_p$) directly determines the final dataset size, so the choice of $k_u$ (or $k_p$) is primarily based on the desired final dataset size within resource constraints. In this paper, we choose $k_u=40$ and $k_p=192$ to ensure that both `DART-Math-Uniform` and `DART-Math-Hard` datasets end up with \~590K examples. > Q6: The details of Prop2Diff is missing. How many samples were generated for each difficulty level? What is the equation for generating numbers? A6: As we described in Line 137-139, we compute a difficulty score for each query (see Line 158-166) and the number of correct responses per query is linearly proportional to its difficulty score, where the most difficult query receives $k_p=192$ responses — suppose the difficulty score is $r_i\in[0,1]$, number of correct responses for a certain query $i$ is $k_p * r_i$. We keep sampling until the designated number of correct responses is collected or a max sampling cap $n_{\max}$ is reached (Line 145). --- Rebuttal Comment 1.1: Comment: While I appreciate the authors' efforts in addressing the feedback, I still have a few concerns that impact my overall assessment. 1. > "we have shown that DeepSeekMath-7B-RL is able to generate correct responses for nearly 100% of the queries in the MATH500 test set when sampling more". Even the most advanced closed-source models (GPT-4o or Claude-3) struggle to reach such high accuracy on this challenging dataset. It is unclear how increased sampling alone could lead to such an unprecedented level of accuracy. Thus, this result seems "suspicious", a further clarification on this point is necessary to ensure the results are credible. 2. > "we just use the final answer to filter responses as in most previous works," However, this approach does not address the issue of "reward hacking," where models may generate the correct answer but fail to produce reasonable intermediate steps. This flaw significantly impacts the quality of the generated datasets and, by extension, the paper’s overall contribution. While I understand that the primary focus of this work is on data generation methods (as highlighted by the authors and other reviewers), the evaluation methodology used here undermines the reliability of the findings. --- Reply to Comment 1.1.1: Comment: Thanks for your reply! 1. The phenomenon that increased sampling would greatly improve pass@k accuracy on MATH dataset has been already observed by several previous works as well — for example, [1] reported over 80% MATH accuracy with GPT-4 when k=11, [2] reported 72% MATH accuracy with a weak Llama2-7B model when k=256, [3] reported over 85% MATH accuracy with DeepSeekMath-7B-RL when k=64. We totally understand that the reviewer thought these results were “suspicious”. To further clarify this, a high pass@k accuracy only means the correct answer exists among k responses, and it does not entail the model can solve the problem practically because it is hard to select the correct answer out. Additionally, pass@k accuracy is computed by matching the final answer, where the intermediate steps may be wrong even when the final answer is correct. Therefore, we think this high pass@k accuracy is understandable. 2. For the second point, we agree that the “reward hacking” phenomenon may indeed exist, and almost all rejection-sampling-based data synthesis methods including ours admit this limitation as it is non-trivial to guarantee the correctness of the intermediate steps. While imperfect with flaws, rejection-sampling-based synthetic data is still widely used [4, 5] to improve model’s final accuracy on these benchmarks, as we demonstrated in the paper as well — of course, the “reward hacking” phenomenon may be also present in benchmark evaluation where these models in previous works and our paper yield the correct final answer with wrong intermediate steps, yet how to evaluate math problem solving more faithfully considering intermediate steps is an evaluation problem beyond our scope and still an active research direction [6,7]. [1] Toshniwal, Shubham, et al. "Openmathinstruct-1: A 1.8 million math instruction tuning dataset.” Preprint 2024. [2] Li, Chen, et al. "Common 7b language models already possess strong math capabilities.” Preprint 2024. [3] Shao, Zhihong, et al. "Deepseekmath: Pushing the limits of mathematical reasoning in open language models.” Preprint 2024. [4] Dubey, Abhimanyu, et al. "The Llama 3 Herd of Models.” Preprint 2024. [5] Yuan, Zheng, et al. "Scaling relationship on learning mathematical reasoning with large language models.” Preprint 2023. [6] Xia, Shijie, et al. "Evaluating Mathematical Reasoning Beyond Accuracy.” Preprint 2024. [7] Zeng, Zhongshen, et al. "MR-BEN: A Comprehensive Meta-Reasoning Benchmark for Large Language Models.” Preprint 2024.	Summary: The paper proposes a rejection sampling pipeline for automatically generating SFT data, emphasizing that harder data requires more trials. The difficulty is heuristically determined using the ratio of incorrect trials for each question. Experiments demonstrate that this method can outperform traditional rejection methods on various math benchmarks. Strengths: - The experiments are solid, showing significant improvements over traditional rejection methods. - The paper is clearly written and easy to follow. Weaknesses: The proposed Prop2Diff strategy lacks innovation. Assigning more budget to more complex questions in data synthesis is a common practice. For instance, in [1], which successfully annotated 83.1% of MATH questions, it is evident that harder problems were allocated more budget in rejection sampling. [1] also indicates that fewer and harder data can significantly and efficiently improve performance. The authors should discuss the differences between their approach and the one used in [1] more thoroughly. [1] ToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving Technical Quality: 2 Clarity: 3 Questions for Authors: Could you elaborate on how your approach differs from the rejection sampling strategy used in [1]? [1] ToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! We address your concerns below: > Q1: Assigning more budget to more complex questions in data synthesis is a common practice. For instance, in [1], which successfully annotated 83.1% of MATH questions A1: First, we note that most recent works in mathematical data synthesis, including nearly all commonly-used open-source datasets, allocate equal budgets to all questions regardless of their complexity [2,3,4,5,6]. While a few works such as the mentioned ToRA try assigning more budget to more complex questions to improve coverage, we highlight that ToRA still results in a bias towards easier queries. This is in stark contrast to our Uniform and Prop2diff strategies that focus on the distribution of responses, rather than merely improving coverage. Please see the Response to Q2 below for a detailed comparison with ToRA. Additionally, we would like to point out that 83.1% coverage mentioned from ToRA is not an exceptionally high number -- for comparison, the MetaMathQA-MATH-AnsAug dataset achieves 82.8% of coverage on the MATH training set with evenly allocated budgets yet still admits bias towards easy queries as analyzed in Figure 2 of our submission. Below we show the coverage rate per difficulty level of different approaches. The ToRA-Corpus-16k statistics show that it only covers 68% of the Level 5 MATH queries while `DART-Math` datasets cover 99.6%. \| MATH training set coverage \| Total \| Level 1 \| Level 2 \| Level 3 \| Level 4 \| Level 5 \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| ToRA-Corpus-16k-MATH \| 83.1% \| 97.7% \| 91.6% \| 86.5% \| 81.3% \| 68.0% \| \| MetaMath-MATH-AnsAug \| 82.8% \| 98.1% \| 93.6% \| 86.7% \| 76.6% \| 48.9% \| \| VRT Baseline \| 84.9% \| 99.6% \| 98.2% \| 95.2% \| 89.8% \| 62.9% \| \| `DART-Math-` \| 99.6% \| 100% \| 100% \| 99.9% \| 99.7% \| 99.1% \| > Q2: Could you elaborate on how your approach differs from the rejection sampling strategy used in [1]? A2: Please check the general rebuttal for a summary of difference of our approach from ToRA. The most important difference between our approach and ToRA is how responses are distributed across various queries -- While we adjust the distribution either to be uniform or to favor more difficult queries, ToRA focuses mainly on improving coverage without managing the distribution explicitly, leading to datasets that may still bias towards easier queries as shown in the results of the general author rebuttal. Here we supplement more details on how we replicate the ToRA synthesis pipeline to conduct the analysis present in the general author rebuttal. Below we show in the format as "ToRA's method -> how we adapt similar spirits for a simpler replication" step by step (we use CoT format rather than tool-integrated reasoning for a fairer comparison with our datasets): --- --- 1. Greedy decoding once for each problem in MATH&GSM8K with GPT-4, keeping the correct responses -> we follow this with GPT-4o mini. 2. Sampling for 10 trials for each problem not correctly answered by greedy decoding with GPT-4 and keeping up to 4 correct responses per problem (to form ToRA-Corpus-16k) -> we follow this with GPT-4o mini 3. Training CodeLlama models on ToRA-Corpus-16K to perform rejection sampling next -> to avoid additional training for a fairer comparison, we use DeepSeekMath-7B-RL to replace the trained CodeLLama models here to align with DART-Math 1. Sampling with 64 trials for each problem in MATH&GSK8K with CodeLlama, getting 233k distinct correct responses -> we follow this with DeepSeekMath-7B-RL, getting 733k distinct correct responses 2. correcting wrong responses by greedy decoding from the correct preceding portions (costing no more than 64 trials for each problem) with CodeLLaMA-34B, getting 69k corrected responses -> we simplify this by re-sampling another up to 64 trials per problem for all the incorrect responses, getting 225k correct samples. 3. Randomly selecting up to 4 correct responses per problem from 3.1&3.2 -> we exactly follow this. 4. Merge ToRA-Corpus-16k and data from step 3 to form the final training dataset of 69k responses -> we exactly follow this to form the final dataset of 72k responses. --- --- We show the average numbers of responses per problem for different difficulty levels and coverages on the MATH training set in the table of the general rebuttal — distribution-wise we can see that the ToRA pipeline will still produce fewer responses for difficult problems than the easy ones, while `DART-Math-Hard` produces, for example, 10x more # of responses for MATH Level 5 questions than the GSM8K questions.* [1] Gou, Zhibin, et al. "Tora: A tool-integrated reasoning agent for mathematical problem solving." ICLR 2024. [2] Yu, Longhui, et al. "MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models." ICLR 2024. [3] Wang, Ke, et al. "MathCoder: Seamless Code Integration in LLMs for Enhanced Mathematical Reasoning." ICLR 2024. [4] Liu, Haoxiong, et al. "Augmenting math word problems via iterative question composing." Preprint 2024. [5] Huang, Yiming, et al. "Key-point-driven data synthesis with its enhancement on mathematical reasoning." Preprint 2024. [6] Tang, Zhengyang, et al. "MathScale: Scaling Instruction Tuning for Mathematical Reasoning." ICML 2024. --- Rebuttal 2: Comment: Dear Reviewer neCQ, Sorry to disturb you for the last time, but only one day is left until the end of the reviewer-author discussion stage. We still do not know if you have received our newest response. To address your concerns, we wrote all the responses in detail and added new experiments to support them, including: 1. A1: showing 83.1% coverage mentioned from ToRA is not an exceptionally high number; 2. A2 & Author Rebuttal: elaborately designed experiments and detailed explanations to clarify the difference between the sampling strategies between ToRA and DART. Conducting the additional experiments within the limited rebuttal period is challenging. We would like to know whether our responses have addressed your concerns. If you still have other concerns, please give us an opportunity to clarify them. We sincerely hope that you can take a moment to reply, as it is very important for researchers and their efforts on this work. Best regards, The Authors --- Rebuttal Comment 2.1: Comment: Thank you for your detailed responses, which have addressed most of my concerns. I'm pleased to raise my rating accordingly.	Summary: The paper presents an approach to improving the performance of LLMs in mathematical problem-solving. The authors identify that current datasets synthesized using proprietary models like GPT-4, are biased towards easier queries. To address this, they introduce Difficulty-Aware Rejection Tuning (DART), which allocates more trials to difficult queries during data synthesis. This method generates datasets focusing on difficult queries using an open-weight model, DeepSeekMath-7B-RL, without relying on proprietary models. The authors demonstrate that models fine-tuned on DART-Math datasets significantly those fine-tuned on traditional datasets across various mathematical benchmarks, and beat the best baseline by average of roughly 3-4% Strengths: - Technically solid paper with state-of-the-art results. - Mostly well-presented and easy to understand. - Comprehensive experiments and analysis. - Decent impact in improving mathematical capabilities of LLMs, with the authors publicly releasing their dataset. - By using an open-weight model, DeepSeekMath-7B-RL, the authors eliminate dependency on proprietary models like GPT-4, making the approach more accessible. Weaknesses: 1. It is unclear how the hyperparameters of the baseline, VRT (vanilla rejection tuning), were tuned. For instance, as mentioned in Appendix A.2, sampling temperature is searched from 0.3 to 1.7 for DART. Was the same procedure used for VRT? Another caveat is the need for extensive hyperparameter tuning compared to baselines. Were similar extensive procedures for tuning performed for other baselines? 2. It is unclear if the improved performance of the proposed method is due to difficulty or the topic of the problem. For instance, LEVEL 5 Math problems may have a higher number of geometry questions (or at least their fail rate is higher, resulting in fewer samples in VRT). An analysis of topic-wise performance comparing DART and baseline methods may clarify this. Minor Weaknesses: 1. It is unclear how much advantage the method would provide in the case of other multi-iteration fine-tuning methods such as STaR and v-STaR. For instance, it is possible that after multiple iterations, VRT performs similarly to DART, since a higher number of samples will be collected from even the hard problems in second or further iterations. 2. The data synthesis is only done using the DeepSeekMATH-7B model. It is unclear why this model was chosen. Previous methods using VRT-like methods typically use the same model for synthesis and generation. Thus, higher results in smaller models such as Llama-8B may partly be due to the use of stronger models' reasoning chains, making it similar to a distillation method. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The authors use "fail rate" as a metric to measure difficulty. However, has any analysis been performed to measure how good an estimate it is of actual model accuracy? 2. In line 138, "we sample till correct responses for each query is proportional to its difficulty score," does it mean linearly proportional? 3. To the best of my knowledge, previous works usually use lower temperatures in the range of 0.7-1. However, the authors found 1.6 to be effective. Do the authors have a comparison of results between using a more standard temperature (e.g., 0.7 or 1) compared to 1.6? 4. In Line 253, the authors state: "We hypothesize that this is due to these models’ extensive pretraining on mathematical content." Do the authors have more points to substantiate this? For instance, it could be partly due to a slightly weaker or similar model being used to generate synthetic data. Further, the hypothesis: "This pretraining likely covers most skills that could be learned from the GSM8K and MATH training queries" may not be correct, since, at least for Llama2-70B, the model capacity should not be a bottleneck to achieving higher scores on MATH (e.g., Qwen models). Can the authors provide a more detailed reasoning behind this hypothesis? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Major Limitations are addressed in paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the positive comments! We address your concerns below. > Q1: It is unclear how the hyperparameters of the baseline, VRT, were tuned. For instance ……, sampling temperature is searched from 0.3 to 1.7 for DART. A1: We searched temperature from 0.3 to 1.7 according to accuracies by DeepSeekMath-7B-RL on the MATH training set, where $t=1.6$ is the highest temperature that does not suffer from a significant accuracy drop. $t=1.6$ is applied for both VRT and DART-Math. This choice should be fair since the parameter search is not specifically tailored for DART method. > Q2: It is unclear if the improved performance of the proposed method is due to difficulty or the topic of the problem. … An analysis of topic-wise performance comparing DART and baseline methods may clarify this. A2: Thanks for the advice! Difficulty and topic are naturally correlated. As shown in the following table for two models, both topic-wise and topic-macro-average scores (which assign equal weights to different topics) still show significant improvement on every topic by DART. \| Model \| Counting & Probability \| Prealgebra \| Number Theory \| Intermediate Algebra \| Algebra \| Precalculus \| Geometry \| Micro Avg. \| Macro Avg. \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Llama3-8B-VRT \| 34.2 \| 57.8 \| 30.7 \| 20.4 \| 59.6 \| 22.5 \| 29.0 \| 39.7 \| 36.3 \| \| `DART-Math-Llama3-8B` (Uniform) \| 34.6 \| 65.7 \| 35.7 \| 25.4 \| 66.6 \| 29.3 \| 32.4 \| 45.3 \| 41.4 \| \| `DART-Math-Llama3-8B` (Prop2Diff) \| 38.8 \| 62.9 \| 36.8 \| 26.1 \| 67.3 \| 32.0 \| 39.9 \| 46.6 \| 43.4 \| \| Mistral-7B-VRT \| 32.1 \| 56.3 \| 29.6 \| 19.0 \| 58.4 \| 22.2 \| 30.7 \| 38.7 \| 35.5 \| \| `DART-Math-Mistral-7B` (Uniform) \| 33.8 \| 59.8 \| 35.2 \| 24.4 \| 64.1 \| 28.8 \| 34.2 \| 43.5 \| 40.0 \| \| `DART-Math-Mistral-7B` (Prop2Diff) \| 36.1 \| 61.3 \| 35.4 \| 26.0 \| 65.7 \| 31.1 \| 40.5 \| 45.5 \| 42.3 \| > Q3: Another caveat is the need for extensive hyperparameter tuning compared to baselines A3: To clarify, we apply the same training hyperparameters to all the experiments on the same base models, thus there is no additional hyperparameter tuning for DART compared to the baselines during training. For data synthesis, the only additional hyperparameters specifically chosen for DART are # responses per query (Line 174-175) and the sampling cap (Line 145), which is mainly decided by the desired final dataset size considering our resource constraints, we did not “tune” these synthesis hyperparameters according to final performance. > Q4: It is unclear how much advantage the method would provide in the case of other multi-iteration fine-tuning methods A4: This is a good point. DART is compatible with multi-iteration fine-tuning as well and we can use the DART strategy to manipulate data synthesis for each iteration. We leave exploration on this as future work. > Q5: The data synthesis is only done using the DeepSeekMATH-7B model. It is unclear why this model was chosen. A5: The DeepSeekMath-7B-RL model is chosen as the strongest open-source model (though with only 7B size) for mathematical problem-solving. > Q6: Previous methods using VRT-like methods typically use the same model for synthesis and generation. Thus, higher results in smaller models such as Llama-8B may partly be due to the use of stronger models' reasoning chains, making it similar to a distillation method. A6: We agree that our approach is similar to a distillation method, and our goal is to create the best synthetic data for mathematical problem-solving. The mentioned works that use the same model for synthesis represent a different line of work that pursues" self-improvement," while our paper can be viewed as pure data synthesis. This aligns with many existing works in this field that produced SOTA synthetic datasets using GPT-4, a stronger model. > Q7: In line 138, "we sample till correct responses for each query is proportional to its difficulty score," does it mean linearly proportional? A7: Yes, it means linearly proportional. > Q8: Do the authors have a comparison of results between using a more standard temperature (e.g., 0.7 or 1) compared to 1.6? A8: We didn’t compare temperatures in effects on the final performance, and we only compare temperatures in a preliminary stage by evaluating DeepSeekMath-7B-RL, as explained in A1 above. > Q9: In Line 253, the authors state: "We hypothesize that this is due to these models’ extensive pretraining on mathematical content." Do the authors have more points to substantiate this? A9: The hypothesis stems from the understanding that extended large-scale pretraining can reduce the need for meticulously curated SFT datasets. However, we did not rigorously verify this hypothesis, and we agree with the reviewer that "a slightly weaker or similar model being used to generate synthetic data" could be the reason as well. > Q10: the hypothesis: "This pretraining likely covers most skills that could be learned from the GSM8K and MATH training queries" may not be correct, since, at least for Llama2-70B, the model capacity should not be a bottleneck to achieving higher scores A10: This is an interesting point. To clarify, we didn't mean model capacity is a bottleneck, instead, we meant the training queries are the bottleneck -- the queries in `DART-Math` are simply the original GSM8K & MATH training queries that limit the scope that the final dataset can generalize to. Therefore, the generalization accuracy could be bottlenecked by the training query scope no matter how we improve the synthetic responses. Our hypothesis is that math-specific pretraining may already cover most skills that could be learned from these queries, and further training with those queries could not provide significant values. A potential future direction could be synthesizing new queries.	Rebuttal 1: Rebuttal: We thank all the reviewers for the insightful comments! While we address most concerns in the individual rebuttals, here In the general rebuttal we would like to clarify the difference between our approach and ToRA [1] / MARIO [2], a concern raised by Reviewer neCQ and Reviewer quxz. The most important difference between our approach and ToRA/MARIO is how responses are distributed across various queries -- While we adjust the distribution either to be uniform or to favor more difficult queries, ToRA/MARIO focus mainly on improving coverage without managing the distribution explicitly, leading to datasets that may still bias towards easier queries (as we show below). As demonstrated in Figure 4 (Left) of our submission, a high coverage rate (VRT+Cover) alone does not guarantee superior performance. Since both ToRA and MARIO have not released their datasets, we replicate a version of their synthesis pipelines that is comparable with DART to illustrate that the resultant datasets still bias towards easier queries. We detail our specific replication process for ToRA and MARIO in the individual rebuttals to Reviewer neCQ and Reviewer quxz respectively. Below we show the average number of responses per problem for different difficulty levels below of our replicated ToRA and MARIO datasets and the DART-Math datasets. Our replicated ToRA and MARIO end up with a similar MATH coverage ratio and total size to their original ones. While the absolute number of responses is not directly comparable between different methods, distribution-wise we can see that ToRA/MARIO still produce fewer responses for difficult problems than the easy ones. This contrasts sharply with `DART-MATH-Hard`, which produces, for example, 10x more responses for the MATH Level 5 queries than for the GSM8K queries. \| \| GSM8K \| MATH Level 1 \| MATH Level 2 \| MATH Level 3 \| MATH Level 4 \| MATH Level 5 \| Our Size (Original Size in Their Papers) \| MATH/Train Coverage \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| ToRA \| 5.03 \| 5.01 \| 4.99 \| 4.95 \| 4.77 \| 3.84 \| 72k (69k) \| 93.4% \| \| MARIO \| 2.02 \| 2.01 \| 1.98 \| 1.94 \| 1.89 \| 1.57 \| 29k (29k) \| 91.3% \| \| `DART-Math-Uniform` \| 39.93 \| 40.00 \| 40.00 \| 39.80 \| 39.54 \| 37.14 \| 585k \| 99.6% \| \| `DART-Math-Hard` \| 8.49 \| 14.28 \| 33.52 \| 54.94 \| 79.59 \| 107.06 \| 590k \| 99.6% \| We hope the response above clarifies the difference between ToRA/MARIO and our approach. [1] Gou, Zhibin, et al. "Tora: A tool-integrated reasoning agent for mathematical problem solving." ICLR 2024. [2] Liao, Minpeng, et al. "MARIO: MAth Reasoning with code Interpreter Output--A Reproducible Pipeline." arXiv 2024. Pdf: /pdf/548ec0fc04d03b438ca3b52a3500e99c77ae0836.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning	Accept (spotlight)	Summary: The paper considers the offline contextual bandit problem. The authors consider a class of reward estimators for this setting that is a regularization of Inverse Propensity Scoring (IPS - aka importance sampling). A general concentration result is provided for this class of estimators. This is used to provide a tight result for an existing clipping IPS estimator and to construct a new Logarithmic Smoothing (LS) estimator. The resulting estimator is pessimistic by design, making it immediately applicable to the offline contextual bandit problem. The authors use it to derive bounds for policy evaluation and selection and also for policy learning in the Bayesian setting. Experimental results also support the usefulness of the estimator. Strengths: I am only broadly familiar with this line of research making it hard for me to properly contextualize its contributions. 1. The proposed estimator is novel and has nice properties. 2. As the name suggests, the estimator is smooth making it potentially easy to optimize. 3. The application to contextual bandits is interesting. 4. The experimental results are positive. 5. The overall writing is good and clear. Weaknesses: 1. A more explicit comparison with existing concentration\contextual bandit bounds is missing. The authors explain that their bound is better but this is somewhat vague, especially if the reader is not already an expert in this field. 2. In line 155 the authors explain that their result can be derived from [1, Lemma 1.3]. Does this mean that the LS estimator has previously been suggested or only that an alternative proof technique exists for its concentration bound? 3. Performance seems very close to that of IX 4. The main body of the paper does not include any explanation of the techniques used. This can be a proof sketch for the concentration bound or a discussion comparing your approach to existing techniques. Can you provide such an explanation in your response? 5. The notation U(pi) appears without definition in line 281. I assume it's defined in one of the references but should also be defined in this paper for completeness. (Please include an explanation in your response) Typo: line 98: one of the brackets is reversed in the definition of h Technical Quality: 3 Clarity: 3 Questions for Authors: See above. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/a Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First, we would like to thank you very much for your positive review acknowledging the quality of our work. We hope our response addresses your questions and increases your confidence in our work. We will consider the points raised when updating the manuscript. (1) Bound Comparisons Given the 9-page limit for the submitted paper, the tightness claim is thoroughly defended in the Appendix. We directed the reader to the full, explicit comparisons and discussions in the Appendix whenever a statement is made. Specifically: - Appendix E.2 for OPE: - Comparison of the global clipping bound ($U^\lambda_1$) with the empirical Bernstein bound for clipping (Appendix E.2.1). - Comparison of the LS bound ($U^\lambda_\infty$) with the IX bound (Appendix E.2.2). - Appendix E.4 for OPL: - Comparison of the PAC-Bayesian LS bound with the conditional Bernstein bound, the exponential smoothing bound, and the IX bound (Appendix E.4). All these comparisons show that our bounds are tighter. To respect the 9-page limit, the paper focuses on presenting the methodology, introducing the new estimator, and proving the versatility of the approach. With the additional page available in the camera-ready version, we will move these comparisons to the main text. (2) Alternative Proof This is the first time the LS estimator has been proposed, with [1, Lemma 1.3] only providing an alternative proof of its concentration bound. We chose our proof technique for its generality, which is appealing for two reasons. First, it yields a family of empirical upper bounds for a wide range of regularized IPS estimators, such as clipping. Second, within this family, the LS estimator emerged as the one achieving the tightest bound. This strongly motivates the LS estimator, in contrast to defining it in advance (without a clear rationale for its selection) and applying [1, Lemma 1.3] for its concentration. (3) Performance Close to IX: Our experiments demonstrate the superiority of LS compared to IX in OPE and OPS. Even in OPL, LS has a better guaranteed risk than IX, and they are only comparable in terms of the risk of the learned policy. Morevoer, a formal comparison of the LS bound with IX is provided in Appendix E.2.2, showing that LS consistently outperforms IX (for any $\lambda$, in any scenario). It also highlights that the LS bound significantly outperforms the IX bound when the target policy $\pi$ is not deterministic and/or in the low data regime (starting from line 746). This is further confirmed by the detailed results in Appendix H.1.2. (4) Proof Technique Used In the main text (line 121), we state that we use the Chernoff bound with a careful analysis of the moment generating function. Specifically, we prove the monotonicity of the logarithm's residual (the difference between the logarithm function and its Taylor expansion) to derive our results. Our approach differs from existing bounds by using the residual function in combination with the Chernoff bound. (5) The Notation $U(\pi)$ and Typo Thank you for pointing this out. $U(\pi)$ refers to a generic bound evaluated for the target policy $\pi$. We will include its explanation in the updated manuscript. We also thank you for highlighting the typo. --- Rebuttal Comment 1.1: Title: Response Comment: Thank you for responding to my concerns, I will keep my score. A small side note: the abbreviation LS is usually associated with Least Squares. If there is another sensible naming option it will probably help avoid confusing the two :)	Summary: The authors propose empirical concentration inequalities for off-policy evaluation that apply to several forms of (smoothed) IPS, which are claimed to be tighter than the results in existing works. These bounds are then used to derive policy learning guarantees that inherit the properties of the concentration inequalities. Strengths: I appreciate that the authors have applied their method to OPE, OPS, OPL, and also provided some experiments. I did not read the appendix nor check the correctness of the analysis in detail, but from a quick glance it appears that the authors were careful to provide rigorous and well-organized proofs. Weaknesses: My biggest criticism is that the authors have not justified in the main body their claim that "LS is provably tighter than its competitors" (L12) for any of the results, including the concentration inequalities (Prop 1, Cor 3, Cor 4) and the policy learning guarantee (e.g., Prop 6). Since these claims are the whole premise for the paper, their justification should be a central pursuit and only stating "x is in Appendix y" (L147, 178, 195) is hugely insufficient. For example, I would have liked to see a discussion on (possibly even in graphs): - For the choices of $h$ described in (4), when do the bounds in Prop 1, Cor 3, and Cor 4 improve over the bounds from their respective papers? - Is $h$ (the tightest choice) better than all of the above? - Does this hold for all hyperparameter choices, e.g. $\lambda$ and $L$? - How does the computational complexity of calculating the bounds in Prop 1, Cor 3, Cor 4 hold up relative to their competitors? - Exactly how does this lead to downstream policy learning improvements? Lastly, I found the overall technical presentation to be relatively poor, and I'll give a few examples: - The condition (C1) from Section 2 ("Regularized IPS") that all results depend on is never explicitly defined, and it should be an assumption that is called in every proceeding proposition/theorem statement. - Shouldn't (11) be framed in, e.g., a lemma environment? - The term "pessimism" is overloaded, e.g., for "high-probability upper bounds" in L111 but also for an in-expectation variant in Eq. (5), which is slightly unusual (and I'm pretty sure not the way it's used in [26]) but not recalled again in the main body so I'm not sure what it's for (perhaps the proof of Prop 1). Technical Quality: 2 Clarity: 1 Questions for Authors: In addition to the ones in "Weaknesses," I have a specific question about Proposition 6. The gold standard in offline policy selection is a bound in the form of $R(\pi) - R(\widehat\pi) \le \lambda S(\pi) + \varepsilon$ for any comparator policy $\pi$ rather than the optimal one $\pi^$ (see [26] and [Wang 2024] and [Xie 2021]). The former is strictly more general -- can you write your bound in such a form? References Wang, L., Krishnamurthy, A., & Slivkins, A. (2024, April). Oracle-efficient pessimism: Offline policy optimization in contextual bandits. In International Conference on Artificial Intelligence and Statistics (pp. 766-774). PMLR. Xie, T., Cheng, C. A., Jiang, N., Mineiro, P., & Agarwal, A. (2021). Bellman-consistent pessimism for offline reinforcement learning. Advances in neural information processing systems, 34, 6683-6694. Confidence: 5 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: I do not believe the authors have fully discussed the limitations of their method (see "Weaknesses"). Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First of all, we would like to thank you for your review and we hope that our response answers your questions and clears out misunderstandings. We think that your comments can be completely addressed and we hope this will lead you to increase your score. Answer to the main criticism First, we want to clarify that our claim "LS is provably tighter than its competitors" is theoretically proved by comparing (Eq. (13), Corollary 4) to existing bounds. This discussion was included in the Appendix and will be moved to the main text in the revised version as we will have additional space. This claim is also supported empirically by plotting the bounds (see answer below). Second, there are many other equally important contributions central to the paper. We derive novel, high-probability empirical upper bounds applicable to a large family of "regularized IPS". These bounds are analyzed, compared, and minimized to obtain a new estimator, Logarithmic Smoothing (LS), which has favorable properties. Its empirical bound (Eq. (13), Corollary 4) is tighter than those of competitors, and the estimator is pessimistic by design, showing excellent performance in extensive OPE, OPS, and OPL experiments. Answers to the points raised 1- For the choices of described in (4), when do the bounds in Prop 1, Cor 3, and Cor 4 improve over the bounds from their respective papers? First, to clear up any misunderstanding, we did not claim that our bounds are better than existing ones for any choice of $h$. We claimed that the LS bound (Eq. (13): Corollary 4, evaluated at its minimizer) is tighter than the existing bounds derived for their respective estimators. Now to address this question, we focus on the choices in (4): Clipping, Exponential Smoothing (ES), and Implicit Exploration (IX). Harmonic and Shrinkage do not come with empirical upper bounds, so they are omitted from this comparison. From the paper, we __theoretically__ prove: - Our bound with $L = 1$ (Cor 3) with $\lambda \sim \mathcal{O}(1/\sqrt{n})$ applied to Clipping is tighter than the empirical Bernstein bound for Clipping [56], especially when $\pi$ is different from $\pi_0$ (Appendix E.2.1). - ES only provides a PAC-Bayesian bound, which is proven to be worse than ours when applied to ES, for any $\lambda$ (see Appendix E.4, starting line 818). - The LS bound $L = \infty$ (Cor 4), evaluated at $h_{, \infty}$ (Eq. (13)), is tighter than the IX bound for any $\lambda$ (Appendix E.2.2). 2-* Is $h^$ better than all of the above? __Theoretically__, we can prove: - The bound of LS ($h_{, \infty}$) with any $\lambda$ is tighter than the IX bound (Appendix E.2.2). - The bound of Global Clipping ($h_{, 1}$) with $\lambda \sim \mathcal{O}(1/\sqrt{n})$ is tighter than the empirical Bernstein bound (Appendix E.2.1). - The bound of LS ($h_{, \infty}$) with any $\lambda$ is tighter than the bound of Global Clipping (Proposition 5). - Thus, the LS bound is better than all known evaluation bounds: LS is better than IX and Global Clipping for any $\lambda$. Since Global Clipping is tighter than empirical Bernstein when $\lambda \sim \mathcal{O}(1/\sqrt{n})$, LS is also better than empirical Bernstein for these values of $\lambda$. - For learning, the PAC-Bayesian bound of LS ($h_{, \infty}$) is tighter than all known PAC-Bayesian bounds, for any $\lambda$ (Appendix E.4). __Additional plots.__ The claims in __1-__ and __2-__ above are supported __empirically__ (see global response). These plots support even stronger claims such as showing that LS bound is better than empirical Bernstein for any $\lambda$. In these plots, we compare Proposition 1 with different $L$ for different choices of $h$. We also compare our bounds (Global Clipping and LS) to Proposition 1 (for different $L$) evaluated in Clipping and IX and to their respective bounds (empirical Bernstein and the IX bound). Finally, a comparison of the bounds for different values of $\lambda$ is provided. In particular, these plots also address your third point __3-__ Does this hold for all hyperparameter choices, e.g. $L$ and $\lambda$? 4-* How does the computational complexity of calculating the bounds in Prop 1, Cor 3, Cor 4 hold up relative to their competitors? Our bounds are tractable, in contrast to many existing studies that provide intractable bounds (e.g., [37, 56]). Tractable bounds do exist in the literature, with the IX bounds being the most computationally efficient. Cor 4 has a similar complexity to the IX bound, as it only requires computing the LS estimator without any additional high-order terms. 5- Exactly how does this lead to downstream policy learning improvements? Tighter bounds lead to policy selection and learning strategies with provably better regret/suboptimality. For example, the suboptimality derived for the LS estimator is tighter than that of the IX estimator in all scenarios (Appendix E.3 for OPS and Appendix E.5 for OPL). This typically translates into better empirical results, as supported by our experiments. The technical presentation - Condition (C1) is explicitly defined in line 100. We will refer to this condition in our propositions. - Eq. (11) is the derived minimiser of Cor 3 and was not framed as a proposition to ease reading. - The in-expectation pessimism is indeed used to prove Proposition 1, and we will drop its naming to offload the term 'pessimism'. Subotimality with a comparator $\pi$ instead of $\pi^$ The technique used to derive the suboptimality does not rely on any specific property of $\pi^$ and can be directly applied to prove the bound for any $\pi$ by replacing $\pi^$ with $\pi$. We will add this more general result to the Appendix. We aim to identify the optimal policy $\pi^$, which is why the suboptimalities are expressed w.r.t. $\pi^*$. This is common in the literature [26, 27]. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response. It's plausible that the bounds in the paper are interesting contributions and may be tighter than existing ones. However, the submission in its current form does not make this argument convincingly. As it is the central tenet of the paper, I believe that all of the discussions, results, and comparisons above should be a central focus of the main body. Given the amount of content that I feel is omitted and worth analyzing more deeply, I do not feel that one round of revision is sufficient to address my concerns, and I stand by my original review. --- Reply to Comment 1.1.1: Comment: Thank you for your feedback. However, we strongly disagree with the reviewer's assessment. The comment that our bounds "may be" tighter than existing ones and that the paper "does not make this argument convincingly" seems to overlook the rigorous evidence we provided. We demonstrated the tightness of the LS bound (and our other bounds) through both theoretical analysis (in the paper) and empirical validation (in the OPE experiments and in the rebuttal), going beyond the typical approach of relying solely on one or the other. We would like to clarify that our tightness claims are made in the main text, with proofs provided in the Appendix. Redirecting readers to proofs in the Appendix is standard in the field, and we have explicitly stated our intention to include them along with additional discussions in the camera-ready version. This can be easily accommodated with the additional page, as other reviewers did not request major revisions. Finally, it would have been helpful if the reviewer had specified which concerns remain unaddressed by our rebuttal. The statement "I do not feel that one round of revision is sufficient to address my concerns" lacks explanation, leaving us unsure of which parts of the rebuttal did not address the reviewer's concerns, as we aimed to thoroughly address all points raised in the initial review.	Summary: This paper studies log-algorithmic smoothing of importance weight for off-policy learning. The proposed smoothing technique can be seen as a differentiable variant of clipping, which is useful for variance reduction for OPL. The paper also analyzes the PAC-Bayes learning bound of the proposed OPL method, characterized by the KL divergence with the logging policy, showing that the proposed method achieves a tighter bound than baselines, including simple clipping. The experiment also shows that the proposed method has tighter bounds than baselines and enables more accurate off-policy selection. Strengths: - Reasonable formulation based on theoretical analysis: The proposed method is derived from a tight upper bound of the policy's risk. Also, the proposed method has an interpretation as soft, differentiable clipping. The technique is well-motivated and is reasonable to interpret. - PAC-Bayes learning bound: A sub-optimality form is derived, and it is also easy to interpret as a pessimistic approach, which should be acknowledged. - Experiments on various tasks: The paper evaluates the proposed approach in upper bound derivation, off-policy selection, and off-policy learning. The experiment results show the wide applicability of the proposed method in many OPE/OPL-related tasks. Weaknesses: - Connection to Metelli et al. 2021 is not clear: Metelli et al. 2021 also considers the importance of weight differential and shows that the proposed method achieves a Subgaussian rate. Similar to the reviewed paper, Metelli et al. 2021 also have a KL divergence term in the theoretical analysis. While the proposed method adequately differs from Metelli et al. 2021, and the paper does cite it, the paper does not mention Metelli et al. 2021 in the related work in detail. Since the motivation and contributions are similar, a detailed discussion on the advantages and the differences would be appreciated. - Baselines in the experiments: As mentioned above, Metelli et al. 2021 propose a similar idea that can be used as a baseline in experiments. Comparing with advanced regularization techniques such as shrinkage (Su et al. 2020) would also be informative. (Metelli et al. 2021) Subgaussian and Differentiable Importance Sampling for Off-Policy Evaluation and Learning. Alberto Maria Metelli, Alessio Russo, Marcello Restelli. NeurIPS, 2021. (Su et al. 2020) Doubly robust off-policy evaluation with shrinkage. Yi Su, Maria Dimakopoulou, Akshay Krishnamurthy, Miroslav Dudík. ICML, 2020. Technical Quality: 3 Clarity: 2 Questions for Authors: - What are the connections with Metelli et al. 2021? (See weaknesses for the detailed comments.) - How does OPL work with the varying performance of the behavior policy? In my understanding, the policy will be pessimistic in out-of-distribution, but seeing how it works in experiments would be informative for readers. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Missing connection with a similar idea. See the weaknesses for the details. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First of all, we would like to thank you for your positive review, and we hope that our response addresses your questions and clears up any misunderstandings. (1) Connection to Metelli et al. 2021 Our Logarithmic Smoothing (LS) estimator and the Harmonic estimator of Metelli et al. (2021) share similarities in how they avoid hard clipping on importance weights while obtaining sub-Gaussian concentration. However, they differ in several key aspects, which we list below: - The motivation: Metelli et al. (2021) introduced the Harmonic estimator specifically to mitigate the heavy tail of IPS and focused their study on it. In contrast, LS is motivated differently. We begin by deriving a high-probability upper bound that applies to a large family of regularized IPS estimators, including the Harmonic estimator. We then identify the estimator within this family that achieves the tightest possible bound. This turns out to be LS, a novel estimator that logarithmically smooths the importance weights. While the Harmonic estimator can be used with our bounds, the LS estimator provides tighter results. - No empirical upper bound for the Harmonic estimator: Metelli et al. (2021) study the properties of the Harmonic estimator and derive a concentration inequality (Theorem 5.1) that, in the case of contextual bandits, involves the divergence $$I_{\alpha}(\pi, \pi_0) = \mathbb{E}_{x \sim \nu, a \sim \pi_0(\cdot\|x)} \left[ \left(\frac{\pi(a\|x)}{\pi_0(a\|x)} \right)^\alpha \right],$$ between the target policy $\pi$ and the behavior policy $\pi_0$. This quantity makes the upper bound intractable because it requires computing an expectation under the unknown distribution of contexts $\nu$. Consequently, the derived concentration provides insight into the estimator's behavior but cannot be used to implement pessimism as it is not empirical. This contrasts with our fully empirical bounds that can be used directly for efficient pessimism. - The LS estimator enjoys a better concentration: Equation (2) (Page 5) of Metelli et al. (2021) proves the sub-Gaussian concentration of the Harmonic estimator. We also prove the sub-Gaussian property of the LS estimator in Proposition 13 in Appendix E.1 and briefly compare it to Metelli et al. (2021). Our findings show that the LS estimator enjoys a better sub-Gaussianity constant than the Harmonic estimator. - Metelli et. al. 2021 do not have a KL term: Metelli et al. (2021) only provide the concentration property of the Harmonic estimator and did not analyze the estimator in off-policy learning. Their concentration uses $I_{\alpha}(\pi, \pi_0)$, which simplifies to the theoretical second moment of the importance weights when $\alpha = 2$. This differs from the KL divergence between the distributions inducing the policies, which is proper to the PAC-Bayesian analysis of off-policy learning [5, 20, 48]. (2) Baselines in the experiments In our experiments, we aim to compare the ability of estimators to implement efficient pessimism. Therefore, we only include estimators from the literature that come with empirical/tractable upper bounds. This is why we include clipped IPS, SNIPS, and IX, but not Harmonic and Shrinkage, as these do not provide empirical/tractable upper bounds for pessimism. (3) How does OPL work with the varying performance of the behavior policy? These experiments are given in Appendix H.2.4. Specifically, we conducted OPL experiments while varying the inverse temperature $\alpha$ of the behavior policy. Changing $\alpha$ interpolates between a uniform policy and a good behavior policy, resulting in different performance levels of the behavior policy. The main body of the paper presents aggregated results for ease of exposition, while detailed performance for each value of $\alpha$ can be found in Appendix H.2.4. Generally, the performance of the learned policy for all methods decreases as the behavior policy's performance decreases. Overall, the performance gap between LS and the baselines widens as the behavior policy's performance gets lower. --- Rebuttal Comment 1.1: Comment: Thank you for providing responses to the questions. After reading the rebuttals, the difference with the Harmonic estimator (especially the upper bound analysis) became clearer. I believe this paper is worth sharing with the community, and I would keep my initial evaluation.	Summary: Policy evaluation, selection and optimization are considered in the context of offline contextual bandits, where i.i.d. data with a known behavior policy is given. The authors set out to study a generalization of importance weighted policy evaluation; for this they start from a general formulation that computes a value for all data observations, which are then averaged. The free "parameter" here is $h$, the function that assigns a value given an observation (of a context, associated action, and cost). A tight, general, high probability upper bound on the expected cost of a fixed target policy is derived first. Specific choices for the map $h$ are then derived based on minimizing this upper bound. Two practical solutions to this optimization problem are studied in more details: Global clipping and "logarithmic smoothing". Results are then derived for both policy selection and optimization. Strengths: Novel ideas, novel results, good empirical results. Weaknesses: Despite saying that the methodology of paper [31] is adopted, this is only partially done. Why deviate from the evaluation in [31]? I expected an explanation of this. Technical Quality: 4 Clarity: 4 Questions for Authors: Can you explain why you did not follow the protocol and reported values of [31]? Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: na Flag For Ethics Review: ['No ethics review needed.'] Rating: 10 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First of all, we would like to thank you for your positive feedback acknowledging the quality of our work, and we hope that our response addresses your question. In our paper, we adopt the experimental design of [31] for both pessimistic policy evaluation and selection. The authors of [31] identified technical errors in their derivations after the publication of the paper, which they corrected by making stronger assumptions (one-hot rewards) and changing their proof technique. These changes resulted in loosened bounds and experimental results that were inconsistent with their initial findings. For more details, please refer to Appendix A of the updated arXiv version of [31]. These modifications are also reflected in the official GitHub repository of [31], which we used for part of our experiments. Due to these discrepancies, we reran all experiments from scratch for better reproducibility. Our code is included in the supplementary material and can be run to reproduce our results. [31] Ilja Kuzborskij, Claire Vernade, Andras Gyorgy, and Csaba Szepesvári. Confident off-policy evaluation and selection through self-normalized importance weighting. In International Conference on Artificial Intelligence and Statistics, pages 640–648. PMLR, 2021.	Rebuttal 1: Rebuttal: We are very grateful to the reviewers and AC for their valuable time. We attach to our rebuttal additional plots comparing the properties of the bounds for three different datasets and supporting empirically our theoretical findings, showing that the LS bound is tighter than its competitors. Figure 1- Comparison of Proposition 1 for different IPS regularisation functions $h$ and different $L$. For this plot, $\lambda$ is fixed to $1/\sqrt{n}$ while we vary the regularization function $h$ and the moment order $L$. We observe that the performance of the bound differs depending on the choice of $h$ and $L$. Additionally, the optimality of a regularization function $h$ depends on the value of $L$. For example, on the balance-scale dataset where the difference between methods is more apparent, Clipping (orange curve) is better than IPS without any regularization (blue curve) for finite values of $L$, while IPS becomes better when $L \to \infty$. This is why, in our paper, we seek the minimizer $h$ for our bounds to obtain the tightest results specific to values ($L = 1$ and $L \to \infty$), finding theoretically that the minimizer for $L=1$ differs from that for $L \to \infty$. These empirical bound plots confirm this aspect of our theory. Figure 2- Comparison with Clipping and its empirical Bernstein bound. For this plot, $\lambda$ is fixed to $1/\sqrt{n}$ and $M = \sqrt{n}$ (where $M$ is the hyper-parameter of Clipping as shown in Eq. (4)). We observe that Proposition 1 applied to Clipping, as well as our other bounds (Global Clipping and LS), outperform Clipping and its empirical Bernstein bound in all scenarios. Figure 3- Comparison with IX and its bound. For this plot, $\lambda$ is fixed to $1/\sqrt{n}$. We observe that our Proposition 1 evaluated in IX does not perform well compared to the specialized IX bound. Global Clipping is comparable to the IX bound and is tighter in scenarios with sufficient data (2nd and 3rd plot). Finally, our LS bound is always tighter than the IX bound, as proven theoretically in our paper (Appendix E.2.2). Figure 4- Comparison of the bounds for different values of $\lambda$. For this plot, we set $M = 1/\lambda$ for the empirical Bernstein bound of Clipping and vary $\lambda$ for all our bounds. As already proven in our paper, the LS bound is better than both Global Clipping (Proposition 5) and the IX bound (Appendix E.2.2) in all scenarios, for all values of $\lambda$. We also observe that all our bounds outperform the empirical Bernstein bound of Clipping, especially in the region of $\lambda \sim \mathcal{O}(1/\sqrt{n})$, as shown theoretically in Appendix E.2.1. Pdf: /pdf/0e60540292720fbb2c88c42b266032cb412674e4.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Improving Sparse Decomposition of Language Model Activations with Gated Sparse Autoencoders	Accept (poster)	Summary: This work proposed a Gated Sparse Autoencoder (Gated SAE) to mitigate the standard SAEs' biases, such as shrinkage, which systematically underestimate the feature activations from SAEs. The key difference between Gated SAE and SAE is that the Gated SAE separates affine transformations within the encoder in order to decide which dictionary elements to use in a reconstruction loss, and estimate the coefficients of active elements, although with the 50% more computing required to achieve. Comprehensive experiments are conducted to compare and verify how good the Gated SAE is to standard SAE, including a blinded human study to rate and compare the interpretability of randomly sampled Gated and baseline SAE features. Strengths: - A new architecture of SAE inspired by GRU is proposed to include a gate mechanism to mitigate shrinkage bias - Comprehensive quantitative experiments including ablation studies to evaluate the proposed Gated SAE compared to SAEs - A human evaluation to rate randomly sampled features from Gated SAE and SAE Weaknesses: - It is not very straightforward to understand how well the features from Gated SAE are compared to SAE based on Figure 4. Some case studies based on the open-source SAE visualizer library [1] are required to help better understand this. - It will be better to see more case studies on downstream tasks to compare Gated SAE and SAE, e.g., automatic circuit detection [2] [1] C. McDougall. SAE Visualizer, 2024. https://github.com/callummcdougall/sae_vis [2] Huben, Robert, et al. "Sparse Autoencoders Find Highly Interpretable Features in Language Models." The Twelfth International Conference on Learning Representations. 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: - As mentioned in the weakness above, the interpretability analysis was conducted via human rating for randomly sampled feature visualizations. However, those visualizations are not included in the main body or appendix. It will be better to help understand how well the sampled features are when comparing Gated SAE and SAE. - Although the Gated SAE has good performance on the loss recovered (fidelity) and relative reconstruction bias, it is still not clear whether features from Gated SAE are better than SAE under downstream tasks. It will make this work very solid if some analysis of mini downstream applications can be conducted, e.g., IOI task, greater-than task, etc. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for your thoughtful review and helpful feedback, and are pleased you found our experiments comparing Gated and baseline SAEs comprehensive. We indeed used sae-vis, which you referenced in your review, to produce the visualization used in the interpretability study, citing this library section 4.2 where we explain the interpretability methodology. We agree however that it would be helpful to share screenshots of these dashboards, and will add a new appendix to the paper including some examples (from both Gated and baseline SAEs). We do note though that – as shown in Figure 4 – the differences in interpretability between Gated and baseline SAEs are slight (if any), and it is unlikely that these dashboards will give much insight into systematic differences in the features found by the respective architectures. Nevertheless, by including these dashboards we hope to better bring to life the information raters had to go on when rating the interpretability of the features. Regarding analyzing the performance of Gated SAEs on downstream uses, we agree this would be excellent evidence for the utility of Gated SAEs (or SAEs in general). However, as we note in our Limitations & Future Work section, the science of evaluating SAEs in a manner that correlates well with downstream uses (which are still uncertain) and yet is scalable and objective is in its early stages. To take the paper referenced in the review (Huben at al., 2023) as an example, while it makes a valuable contribution to furthering the study of SAE evaluations, the metrics introduced there are either better for resolving differences between SAEs as a whole vs other decompositions like PCA (i.e. Figure 4, where the difference between SAEs and e.g. PCA is stark, but the differences between SAEs are small), boil down to metrics that were indeed considered in our interpretability study (i.e. feature monosemanticity and direct effects on logits) or involve performing a bespoke study of one circuit with one SAE, which makes it difficult to scale and compare several SAEs objectively. To be clear, our point is not to undervalue the contribution of Huben at al., but to explain why we felt it would be unreasonable to try to combine in a single paper both the architectural innovation of Gated SAEs with simultaneously making a meaningful advance in the science of evaluating SAEs with different architectures. In this light, we do believe that our paper – by introducing a novel SAE architecture and showing that it either matches or exceeds baseline ReLU SAEs on standard SAE evaluation metrics – provides a valuable contribution even without these additional evaluations. --- Rebuttal Comment 1.1: Title: Reply by Reviewer 58d7 Comment: Thanks for this helpful rebuttal regarding the weaknesses and questions. Although proposed Gated SAEs outperform other SAE baselines under the standard SAE evaluation metrics, I am still wondering whether there are some existing works which try to explore and propose a reasonable metric to measure different SAEs in the mech interp areas. To my understanding, the purpose of utilising SAEs in the mech interp area is to better understand and explain different features (monosemanticity or polysemanticity) from original LLMs' attention heads, MLPs and residual streams. By comparing the better SAEs evaluation metrics of Gated SAEs and other baselines and the slight differences between them, it seems like there exists a mismatch between the objective SAEs metrics and subjective interpretability evaluation results. It might be unclear how valuable the new proposed Gated SAEs or other designed SAEs contribute to the interpretability of different features and semantic meanings of different LLMs' internal components. --- Rebuttal 2: Comment: Thanks for considering our rebuttal and for your response. We believe the consensus in the field – as argued in [1], an example of work that explores and proposes metrics for both interpretability and reconstruction fidelity – is that fidelity, sparsity and interpretability are all important criteria to be satisfied by a good decomposition. From this perspective, we think there is no mismatch: fidelity and interpretability are separate desiderata, and it is a valuable contribution to establish – as we have done in this paper – that a change of architecture improves one without damaging the other. Furthermore, we have established our key result – that Gated SAEs improve fidelity (at fixed sparsity) and maintain interpretability – using similar standards of evidence to those used in influential prior work such as [1]. For these reasons, we believe our paper stands on its own even without further, downstream task-oriented evaluations, although we nonetheless agree that such evaluations would be interesting and informative. [1] Bricken, Trenton, et al. “Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.” Transformer Circuits Thread, 2023. https://transformer-circuits.pub/2023/monosemantic-features --- Rebuttal Comment 2.1: Title: Reply by Reviewer Comment: Thanks for this explanation. I will raise my score.	Summary: The paper attempts to resolve the issue of feature shrinkage in sparse autoencoders (SAEs) by replacing the SAE ReLU activation function with a gated ReLU unit. The weight-tying scheme they use for the gated unit effectively turns it into a jump ReLU activation function. They train gated SAEs and baseline SAEs on a one layer transformer, Pyhtia-2.8B and Gemma-7B. They find that gated SAEs eliminate systematic shrinkage, and consistently outperform baseline SAEs on the pareto-curve of sparsity, measured by the L0 pseudonorm, and faithfulness, measured by the model loss recovered relative to a zero ablation baseline. They run various additional tests involving variations of the gated and baseline SAE architectures, including combinations of SAE dictionaries with the classic gradient pursuit algorithm for choosing sparse feature coefficients at inference time. They conclude that the Pareto improvement of their gated SAEs over their baseline SAEs is due in part to better feature dictionaries, in addition to better estimated feature coefficients. They compare the subjective interpretability of 150 gated SAE and baseline SAE features in Pythia-2.8B and 192 features in Gemma-7B, using a blinded analysis of activating dataset examples. They find that the features were similarly interpretable. Strengths: The paper attempts to address a substantive practical problem with current SAE training methods. The paper's proposed new architecture is evaluated extensively, and many detailed additional investigations on the individual effects of various parts of the gated SAE architecture are described in sections 5.1, 5.2 and Appendix D. I find Appendix D interesting in its own right, since it shows quantitative comparisons between SAE methods and the classic gradient pursuit optimization algorithm, as well as mixing SAE feature dictionaries with gradient pursuit for sparse approximation of feature coefficients. I have not encountered such a comparison before. For the most part, good documentation of all their process is provided, and the writing and presentation are very clear in general. Weaknesses: The paper does not really address the concern that gated SAEs may outperform baseline SAEs in part by implicitly widening the definition of what it means for ‘features’ to be represented in the model. As the paper itself notes in Appendix D, though other more powerful sparse coding algorithms greatly outperform SAEs in terms of reconstruction and sparsity, there are concerns that the greater expressivity of these techniques lets them find spurious ‘features’ that would not be accessible to the model’s own internal computations. An SAE can only find features that are represented in the sense that their current values can be read off with a single ReLU probe, while an inference time algorithm or a multi-layer probe may read off ‘feature’ values that the model itself could not possibly access using a single MLP layer. A gated ReLU is far less expressive than an algorithm like gradient pursuit, but more expressive than a ReLU. So to what extent do gated SAEs outperform baseline SAEs merely because they are implicitly working with a more relaxed definition of what it means for a feature to be represented in the model? Figure 6 in Appendix D incidentally investigates this somewhat, since it attempts to compare the quality of gated vs. baseline dictionaries independent of their coefficients. However, the results there seem inconsistent, with smaller performance gaps and baseline SAEs outperforming gated SAEs at higher L0. I think this issue of the representational power of the probe used is pretty central for contextualizing the results, and should at least have been discussed. Throughout the paper, the authors present reconstruction scores for SAEs in terms of the fraction of model loss recovered compared to a zero-ablation baseline. I think this metric obscures vital information. Lowering CE loss from e.g. 4.5 to 4.0 is typically much easier than lowering it from 1.5 to 1.0. Thus, the same difference in loss recovered between two SAEs can correspond to very different gaps in SAE quality. Without the raw CE scores, there is no direct way to infer how large the gap is quantitatively. At minimum, these raw CE scores should be in the supplementals. Better yet, the recovered performance could additionally be reported in terms of the compute required to train a model with the same CE score, as suggested in https://arxiv.org/abs/2406.04093. Technical Quality: 3 Clarity: 4 Questions for Authors: Why are the raw CE loss recovered scores not in the paper? Since it is typically much harder to lower CE loss from e.g. 4.5 to 4.0 than from 1.5 to 1.0, it is difficult to evaluate the quality gap between baseline and gated SAEs, or the quality of the baseline SAEs, without these scores. Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: All addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper, we are heartened that you agree that Gated SAEs attempt to address a substantive practical problem with current SAE training methods and that you find our evaluations and ablations extensive. Regarding your questions about the raw (unspliced model's) cross-entropy loss, we agree that this is useful information that we had omitted from the paper and thank you for drawing this to our attention. To remedy this, we will add a table in the appendix listing original ("raw") CE losses and the CE losses incurred when zero ablating MLP, attention and residual stream layers for the layers and models investigated in the paper, as shown in the PDF accompanying the main rebuttal, and we will reference this table from the captions for the fidelity-vs-sparsity figures presented in the main body of the paper. Turning to the other limitation you mention in the weaknesses section, we agree that it is possible for a more powerful dictionary learning method to attain better fidelity by expanding what it means for a feature to be represented in a model. However, we do not believe that Gated SAEs – particularly with the weight-tying scheme proposed in the paper – fall into this trap. Making explicit the connection between SAEs and probing that you alluded to, a single row of a vanilla (ReLU) SAE encoder corresponds to a pair of linear models: a linear classifier that determines when a feature is active and a linear regressor that determines the magnitude of an active feature. A single row of a Gated SAE encoder also effectively corresponds to the same pair of a linear classifier and a linear regressor, performing the same respective functions. The key difference between vanilla and Gated SAEs is that whereas vanilla SAEs insist that both the weights and biases of the classifier and regressor be exactly identical, the Gated SAE allows the biases to differ. (The weights are still the same, up to a rescaling, under the weight tying scheme introduced in the paper.) As explained in section 3.1, our contention is that even in the ideal case of a perfectly linearly represented feature, it is suboptimal to use the same bias in the classifier and regressor, i.e. to use the same bias determine whether a feature is active and measure its magnitude – because this requires unnecessarily trading off false positives in the classifier (which occurs when the bias is too low) against shrinkage in the regressor (which occurs when the bias is too high). From this perspective, Gated SAEs are removing an undesirable inductive bias in vanilla SAEs, while still using the simplest class of models (i.e. thresholded affine transformations) to detect and measure feature activations. Perhaps to put it another way, any feature that is detectable and measurable by a Gated SAE should be similarly detectable and measurable by a vanilla SAE, albeit with a worse bias-variance trade-off. The previous paragraph notwithstanding, it is in part to assuage fears that Gated SAEs may be somehow "cheating" in order to obtain better fidelity that we performed the interpretability study described in section 4.2, which finds reassuringly that Gated SAEs are comparably interpretable to vanilla SAEs. Nowhere in our training algorithm is there any term that directly trains features to be human interpretable, and so the fact that they remain human interpretable anyway is a good sign that we have not over-optimized for reconstruction / sparsity at the cost of the thing we actually care about. Nevertheless, we admit in the limitations section that "while we have confirmed that Gated SAE features are comparably interpretable to baseline SAE features, it does not necessarily follow that Gated SAE decompositions are equally useful for mechanistic interpretability", i.e. that our evaluations do not conclusively show that Gated SAEs are comparably "useful" as vanilla SAEs in a practical sense, and agree that this is an important open question that needs resolving. --- Rebuttal Comment 1.1: Title: Response Comment: Thanks to the authors for their responses and updates. The raw CE scores have been provided, which addresses what I think was the biggest problem with evaluating the paper’s results. Converting the CE scores into the compute required to train a model with the same CE score under the assumption of conventional scaling curves holding would still be preferred. As a reader, I currently need to perform this conversion myself ad-hoc anyway to actually judge the results quantitatively. I don’t find the authors’ reply to my concern about the greater representational power of gated SAEs compared to vanilla SAEs wholly convincing. Whether we classify them as part of a ‘simplest class of models’ or not, a single gated ReLU can compute outputs a vanilla ReLU cannot. Meaning it is theoretically able to find ‘features’ a ReLU cannot. It may be the case that in practice, sparse ‘features’ tend to be represented in a way that makes this unconcerning. But that would be an assumption that would need to be concretely defined, and empirically tested. I still believe this deserves more discussion than it got. The features being just as interpretable as those of vanilla SAEs does not seem to me to address the concern of representational power either. Merely mapping the hidden activations back to the original token embedding of the prompt that produced them would also yield highly human-interpretable ‘features’. Yet such ‘features’ would not be useful for interpretability, since they don’t describe the structure of the hidden representation at the current layer of the model in a way that would be accessible to the model’s own computations. With no constraints on the definition of a ‘feature’, metrics like human interpretability of feature activations are not particularly meaningful. Ultimately the authors do conclude: ‘Nevertheless, we admit in the limitations section that "while we have confirmed that Gated SAE features are comparably interpretable to baseline SAE features, it does not necessarily follow that Gated SAE decompositions are equally useful for mechanistic interpretability", i.e. that our evaluations do not conclusively show that Gated SAEs are comparably "useful" as vanilla SAEs in a practical sense, and agree that this is an important open question that needs resolving.’, which I agree with. --- Reply to Comment 1.1.1: Comment: Thanks for considering our rebuttal and for your reply. We did consider translating CE losses into equivalent compute, but realized that it would be difficult to do this in a principled manner, because neither of the main model families used in the paper (Pythia and Gemma v1) were scaled compute optimally. We do nevertheless agree that – given a family of models scaled compute optimally – translating CE loss into FLOPs would provide a more intuitive meaning to the y-axes of our plots. Turning to the point about representational power, we indeed agree that Gated SAEs do have (slightly) more representational power than ReLU SAEs. However, our argument is that ReLU SAEs have insufficient representational power to faithfully recover representations that would be considered linear features in the first place, e.g. the linear representations (in the presence of interference) described in [1]. In the presence of interference between features, ReLU SAEs must choose between underestimating feature magnitudes or allowing more false detections, and as a result they can’t faithfully recover the original features even when they are linear. Gated SAEs on the other hand do not make this trade off, and thereby better capture the class of representations that are considered to be linear representations (under interference) in the first place. However, we acknowledge that this conceptual argument does not prove that Gated SAEs improve usefulness on downstream tasks in practice, and are glad you agree with the limitation we noted regarding Gated SAEs being comparably interpretable and more faithful not necessarily implying that they are equally (or more) useful. Given the excitement about SAEs and the research priorities of the mechanistic interpretability field, we anticipate future work that evaluates Gated SAEs on downstream tasks, shedding further light on these concerns. [1] Elhage, et al. “Toy Models of Superposition.” Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/toy_model/index.html	Summary: This work introduces a new technique under mechanistic interpretability's sparse autoencoders. By using a less naive SAE, with a gating mechanism and a little extra computation, the paper shows a decent improvement over the baseline. Strengths: This work addresses the important issue of interpreting transformer-based LLMs and clearly demonstrates an interesting method. The mechanistic interpretability community will certainly find this work of interest. The writing is well written and fairly easy to follow, the results are clearly presented, and all relevant aspects of the method are appropriately ablated I liked the setup of the internal user study; I think future papers will follow the design of the study closely. The work's cited throughout the manuscript are incredibly thorough. Weaknesses: While I generally like the paper, I have two primary concerns: * The architecture and loss are somewhat difficult to understand. I did appreciate the pseudo-code in the appendix, but I feel for readers not familiar with SAEs may have a hard time, especially with the optimization-based design choices of weight tying. Perhaps explaining weight tying later in 3.2 would help. I would especially prefer if a few lines of pseudo code could be added in the main paper, next to figure two. * The user study results. I don't mind the small change in means between the method and the baseline, but the explainable AI community has been around for a long time and the shift from studies with a few experts to larger cohorts has been the norm for a while now. Just because there's a rebranding to mechanistic interpretability doesn't mean this field should settle for underpowered studies. Nevertheless, I do find the study setup itself to be well articulated and a very useful starting point for future work in this area. Minor: Some of the design choices (weight-tying, no r_mag, etc) aren't well explained until the ablation where we find they are primarily for optimization. This could be motivated a little earlier, i.e. that the pareto improvement comes from the separation, and not those choices. Technical Quality: 4 Clarity: 3 Questions for Authors: None Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for your review and valuable feedback, and are encouraged that you found the paper fairly easy to follow, with results clearly presented and key aspects of the method appropriately ablated. We appreciate that the explanation of the architecture and loss function is somewhat dense, and perhaps hard to follow for readers unfamiliar with SAEs. To some extent we are limited in how much we can expand these sections due to the paper page limit, but we would like to take your suggestions on board. Regarding your suggestion to move the description of the weight tying scheme later in 3.2, we propose to move this subsection (i.e. lines 114–127) after the paragraph on Training, and give it its own minor heading ("Tying weights to improve efficiency"). Regarding moving a few lines of pseudo-code next to Figure 2, it will be difficult to do this due to space constraints, but instead we will mention in the figure's caption a reference to the pseudo-code in the appendix, and are also trying to improve the figure's legibility in response to another review (which will hopefully help address your concern too, by making it less important to rely on the pseudo-code). Turning to the interpretability study, we believe in hindsight that – due to some of the wording in section 4.2 – there is a danger that the aims of the study could be misunderstood in a way that leads readers to conclude the study was underpowered. Our primary motivation for introducing the Gated architecture was to improve the trade-off between reconstruction fidelity and sparsity; in this context, the purpose of this interpretability study was to provide reassurance that this improvement in reconstruction fidelity does not come at the expense of a deterioration in interpretability. In other words, our aim was not necessarily to show that Gated SAEs are more interpretable than baseline SAEs but rather to show that they aren't less interpretable by a practically relevant margin. We found that even with this study size we were able to confidently exclude a meaningful deterioration in interpretability – the confidence interval for the mean difference was found to be [0, 0.26] – allowing us to confidently state that Gated SAEs are at least similarly interpretable to baseline SAEs. This is indeed the conclusion advertised in the abstract, introduction and conclusion sections of the paper. Nevertheless, we will change some of the sentences in section 4.2 to make the real aim of the study clearer – i.e. giving more prominence to the confidence interval on the mean difference in interpretability as the key result of the study – and less prominence to the question of whether Gated SAEs are more interpretable, for which we agree a bigger sample size would be needed to settle the question. --- Rebuttal Comment 1.1: Comment: Thank you authors, I am satisficed that this paper is of interest to the community. I will keep my score.	Summary: This paper introduces Gated Sparse Autoencoders (Gated SAEs), an improvement over standard sparse autoencoders (SAEs) for decomposing language model activations. The key idea is to separate the tasks of detecting which features are active and estimating their magnitudes, allowing the sparsity penalty to be applied only to feature detection. Through experiments on language models up to 7B parameters, the authors show that Gated SAEs achieve better reconstruction fidelity for a given level of sparsity compared to baseline SAEs, while resolving issues like shrinkage. A human evaluation study finds Gated SAE features to be comparably interpretable to baseline features. Strengths: - A well-motivated architectural modification to SAEs that addresses key limitations - Comprehensive empirical evaluation across multiple model sizes and activation sites demonstrating clear improvements over baseline SAEs - Careful ablation studies and analysis to understand the source of improvements - Human evaluation study to assess interpretability of learned features - Thorough discussion of limitations and future work directions Weaknesses: - The presentation could be improved in some areas, particularly in explaining some of the technical details and metrics - Some of the figures are quite dense and could be made more readable - The human evaluation study, while valuable, has a relatively small sample size Technical Quality: 4 Clarity: 4 Questions for Authors: - Do you have any insights on how Gated SAEs might scale to even larger language models? Are there any potential limitations as model size increases? - Have you explored using Gated SAEs for any downstream mechanistic interpretability tasks beyond the basic reconstruction and interpretability metrics? For example, does the improved reconstruction enable better circuit analysis? - The weight tying scheme seems important for computational efficiency. Have you explored any alternative tying schemes? Is there a theoretical justification for why this particular scheme works well? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors provide a good discussion of limitations in the conclusion section. They appropriately note that their experiments were limited to specific model families and that further work is needed to evaluate usefulness for downstream interpretability tasks. They also acknowledge the subjective nature of the human interpretability study. Regarding potential negative societal impacts, the authors do discuss this briefly in Appendix A. They note that advances in LM interpretability could potentially be misused, but argue that the current work poses minimal short-term risks. While this discussion is somewhat brief, it does address the key points and seems appropriate for the nature of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review of our paper and your feedback and questions. We are glad you appreciated our explanation of the motivation behind the gated SAE architectural modification and found our evaluations and ablation studies thorough. We agree that the exposition is fairly dense and high context; in large part this was due to the need to stick within the page limit. Regarding the presentation of the metrics, although these are only briefly described in the main body, they are defined at further length in Appendix B. To help bring these definitions to life in the appendix, we will add illustrative examples to each definition. E.g., in the case of loss recovered we hope to make this metric clearer by adding, "For example, a SAE that increases cross-entropy loss from 2.4 to 2.5 when zero-ablation increases the loss to 5.0 would have a loss recovered of 96%". We are also considering how to improve Figure 2, which we think could be hard for someone unacquainted with SAEs to decipher at a quick glance. We would value your feedback on these proposals. Regarding the human evaluation study, we selected our sample size trying to balance the resources consumed by the study against its aim, which was principally to establish that Gated SAEs' superior fidelity over baselines does not come at the expense of interpretability. In other words, we set out to establish that there is no meaningful deterioration in interpretability, rather than trying to show that Gated SAEs are more interpretable. Although the discussion of the study in most of the paper (particularly the abstract, introduction and conclusion) are consistent with this aim, we acknowledge there are a few sentences in section 4 that may give the impression that the study was under-powered (because we fail to show that Gated SAEs are statistically significantly more interpretable), when in fact the main objective of the study – to show that the mean difference in interpretability between Gated and baseline SAEs is not negative by any practically important margin – was achieved, since we found the confidence interval for the difference in interpretability to be [0, 0.26], i.e. excluding significant deterioration in interpretability (even if it doesn't establish an improvement). We will reword key sentences from section 4 that are in danger of giving this incorrect impression, in order to emphasize that it is a meaningful deterioration in interpretability that we set out to measure, and that the study confidently shows that such a deterioration has not occurred. Turning to your questions: - We don't anticipate any issues with scaling Gated SAEs to even larger language models except for the issue acknowledged in the paper that they do require 50% more FLOPs per training step than baseline SAEs, holding width constant. In fact, we didn't actually see a 50% increase in step times during our experiments, because step times were so fast that data loading became a significant bottleneck (at least comparable to computing the loss and its gradient). As we scale to wider SAEs however, we anticipate that we should see the difference in training step time between Gated and baseline SAEs increase to 50%. (Note that our experiments suggest that even if we held compute constant to avoid this 50% increase, Gated SAEs would still outperform baseline SAEs.) - We view the use of SAEs for downstream tasks as a major research area for SAEs overall (Gated or otherwise), and would love to see follow-up work on it, as we have noted in our Limitations & Future Work section. Since submitting the paper, we have done some such work ourselves, as have many others (including with Gated SAEs), though we are hesitant to provide references due to the anonymity rules. - The key part of the weight tying scheme – tying the two encoder weight matrices – is theoretically motivated on the grounds that similar directions are likely useful for detecting whether a given feature is active and for detecting its magnitude if it is active; we did not come up with alternative ways to tie the encoder weights that seemed theoretically plausible and hence did not investigate them. We did look at whether r_mag is important for weight tying (this is discussed in section 5.1) and found that it seemed slightly beneficial. We also note in section 3.2 and Appendix H that with this weight tying scheme, Gated SAEs become equivalent to vanilla SAEs with the ReLU replaced by a "JumpReLU" activation function. From this perspective, one could view Gated SAEs with this weight tying scheme as a way to train a vanilla SAE with a JumpReLU activation function, overcoming the difficulties posed by the jump discontinuity in this activation function.	Rebuttal 1: Rebuttal: We thank the area chair and all our reviewers for taking the time to read our paper and for their insightful comments and suggestions. We are encouraged by all four reviewers recommending that our paper be accepted, with reviewer Qtwo recognizing that our paper “attempts to address a substantive practical problem with current SAE training methods”. We perceive general agreement among the reviewers that the Gated SAE architecture is a useful and well-motivated innovation, and that the evaluations and ablations presented in the paper are thorough and comprehensive. We believe Gated SAEs significantly improve the quality of one of the most important tools in mechanistic interpretability – sparse autoencoders – and that this paper therefore deserves acceptance at the conference. Two common criticisms in the reviewer feedback were: (1) that the presentation is dense in places and (2) that the sample size for the interpretability study should have been higher. Regarding (1), we have tried to address the specific concerns raised by the reviewers whilst satisfying page limit constraints, however we would welcome further feedback on our proposals. Regarding criticism (2), we have realized that some of the text in Section 4.2 may give the incorrect impression we set out to show that Gated SAEs are more interpretable than baseline SAEs, when in fact the purpose of the study was to show that Gated SAEs don't sacrifice interpretability to achieve better reconstruction fidelity. We will therefore re-word the offending sentences in Section 4.2, to emphasize our main result that the mean difference in interpretability scores between Gated and baseline SAEs has a confidence interval of [0, 0.26], clearly excluding a practically relevant deterioration in interpretability. Although we were (pleasantly) surprised to find the lower end of the confidence interval to be close to zero, and it is plausible that a bigger study might find that Gated SAEs are statistically significantly more interpretable than baseline SAEs, we are doubtful that such a finding would have much practical relevance. The effect size is likely to be fairly low, and it is uncertain whether a small increase in subjective interpretability would translate into a practically relevant increase in the usefulness of SAEs for downstream tasks, with the latter being the (still hard-to-measure) outcome the field really cares about. Rather than focusing on subjective interpretability, multiple reviewers note that it would be better to see whether Gated SAEs are more useful for downstream tasks. We agree! Unfortunately, this is an area of weakness for the field as a whole, and as such we view this as an avenue for future work which would likely be its own paper. In the accompanying PDF we include a new figure illustrating the dashboards used for the interpretability study, addressing feedback from reviewer 58d7, and a new table listing raw and zero-ablation CE losses to help put our loss recovered results into context, addressing feedback from reviewer Qtwo. Note that the dashboard figure will be printed landscape in the final paper (as shown in the PDF), as this is the only way we could make the dashboard legible due to its wide aspect ratio. The table will be in portrait format in the final paper. Pdf: /pdf/ff078cd9900653e12a0fc261af073c6b02194995.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Marginal Causal Flows for Validation and Inference	Accept (poster)	Summary: This paper introduces _Frugal Flows_ a method that learns the data distribution of data for causal effect estimation; namely outcome $Y$, binary treatment $X$ and pretreatment covariates $\mathbf{Z}$. Through a combination of frugal parametrisation, normalizing flows and copulas, separate components for the marginal causal effect $p_{Y\| do(X)}$, the probability integral transforms of $\mathbf{Z}$ and the propensity score are leaned. (The components of) the learned model, can be used for (i) estimating the marginal effect and (ii) to generate synthetic data with a fixed marginal effect for benchmarking other causal inference methods. In the second application the component for the marginal effect is switched out for another with desired properties. (i) is demonstrated on small synthetic datasets. (ii) is demonstrated by fitting FFs to two real-world datasets and generating synthetic data with adjusted properties. Strengths: - The paper is well-written. - It tackles an important problem in causality research. Since, randomized data is hard and expensive to get, many causal methods are only evaluated on synthetic data and generating realistic/semi-synthetic data is hard. This paper makes a great contribution towards improving synthetic data generation. If the code for the method is provided in a user-friendly manner, I could see this having a big impact on the causality community. Weaknesses: - Normalizing Flows have been used in the causal modelling context before (see [1, 2]). While prior works solve different problems (the inferred latents correspond to exogenous variables of an SCM, not directly applied to causal effect estimation), I think it would still be valuable to contrast this work to what has been done before for future reference in the literature. - L59: The basic causal assumptions aren't explicitly stated. What are the causal assumptions on $X$, $Y$ and $\mathbf{Z}$? It seems like the method wouldn't hold if $\mathbf{Z}$ was a mediator (I suppose the equation after L60 wouldn't hold). A reference to a 500+ page book is given for the assumptions, which feels like a slap in the face for the reader. - The notation for interventional distributions is confusing: what's the difference between using an asterisk and explicitly using the do-notation? In the equation after L60, the LHS seems to be an interventional quantitiy (asterisk, but no do-notation), whereas Equation (1) has the do-notation, but no asterisk. Do the two notation elements mean different things? - I think this paper would greatly benefit from a visual abstract showing how the different flows and distributions come together. Maybe this is something that could be added for the camera-ready. Minor: - L201: typo [1] Javaloy et al. "Causal normalizing flows: from theory to practice." NeurIPS 2023 [2] Wendong et al. "Causal component analysis" NeurIPS 2023 Technical Quality: 3 Clarity: 3 Questions for Authors: - Fig. 1: What's the meaning of the red undirected edges. Does it mean they could go either way and/or they could be confounded? Please specify this somewhere in the paper or appepndix. - Def. 1, App. A: I struggle to understand this definition. The cartesian product on the RHS makes sense to me, but what does the LHS of the equation mean? What is the "x" operation between functions? Do the two functions map to the same space? - Fig. 2: Why is the first line needed? If I understand correctly, this just makes all pretreatment variables uniform. Couldn't you put $\mathbf{Z}$ directly in the second line? - You show synthetic data generation based on two datasets in Sec. 4.2. Why couldn't you use the same datasets to test causal effect estimation in Sec. 4.1? - How many datapoints are in the Lalonde dataset? - In training, how did you check whether the training has succeeded? I suppose you minimize the log-likelihood, how did you define "good enough"? I'm asking because the training seems pretty fast (App. D2.4). What's the total number of parameters for each of the datasets? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - The biggest limitation of the work seems to be the requirements for the dataset size, needed to train normalizing flows. The authors mention this in Sec. 4.1. There could be some applications that have enough data for using Frugal Flows for causal effect estimation (e.g. online businesses with many customers, or well-curated medical datasets like the ones from healthcare providers in Isreal). However, for most applications the data won't be enough. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comprehensive assessment, thoughtful commentary, and suggestions. We share your aspiration that frugal flows can be engineered and presented in a user-friendly manner for the causality community. # Weaknesses We would like to address the comments under the Weaknesses section, responding in the order presented: 1) We appreciate the references shared by the reviewer. The Causal Normalizing Flows (CNFs) paper, also referenced by another reviewer, provides an interesting contrast to our method. While these papers tackle slightly different problems in inferring marginal effects, we are happy to include and discuss them in our literature review. Additionally, we implemented a variation of CNFs using default hyperparameters to estimate ATEs within our synthetic experiments in section 4.1. These are presented in Figure 1 of the attached document to the rebuttal. CNFs showed higher variance in ATE estimates than frugal flows, which is expected as CNFs are not designed for causal effect estimation. We also provided figures showing the observational likelihood loss for real-world benchmarks and synthetic experiments 2) A key assumption for our model is that the covariate set $\mathbf{Z}$ must be pretreatment covariates. You are correct that the parameterization would not hold if a $Z_{d} \in \mathbf{Z}$ were a mediator. We will explicitly summarize the required assumptions for our method in our paper. 3) Thank you for the suggestion regarding the labeling of interventional distributions! The asterisk notation, taken from Evans and Didelez (2024), is used to contrast a wider range of distributions. However, in this paper, we consider a single static treatment model and are happy to use a simpler notation, specifically $do(X)$ operators, in an updated submission. 4) We appreciate the suggestion of a visual abstract and believe it would significantly improve the presentation and readability of our method. We are delighted to incorporate this addition into our paper. We also thank the reviewer for flagging up a typo, which we have corrected! # Questions In response to the reviewer’s specific questions: * Fig 1: There is an error in Figure 1. There should be directed edges pointing from each $Z$ to $Y$ ($Z \to Y$) and undirected edges between each of the $\mathbf{Z}$ variables. The undirected edges indicate that the edges could point either way. We will clarify this in the paper. * Variation independence (VI) is a highly desirable property for a parameterization, since it allows different components to be specified entirely separately. This is extremely useful if one is trying to use a link function in a GLM, or to specify independent priors for a Bayesian analysis. In addition, VI is important in semi-parametric statistics. The definition simply states that the Cartesian product of the images is the same as the image of the joint map. For example, in a bivariate gamma-distribution with positive responses, then $\mu_1 \in \mathbb{R}^+$ and $\mu_2 \in \mathbb{R}^+$ is a variation independent parameterization, since $$ (\mu_1 \times \mu_2)(\Theta) = \mathbb{R}^+ \times \mathbb{R}^+ = \mu_1(\Theta) \times \mu_2(\Theta). $$ However, if we replace $\mu_2$ with $\mu_2’ = \mu_2-\mu_1$ (for example), then although the range of this parameter is $\mathbb{R}$, $$ (\mu_1 \times \mu’_2)(\Theta) = \{(x,y) : x > 0, y > -x\} \neq \mathbb{R}^+ \times \mathbb{R} = \mu_1(\Theta) \times \mu’_2(\Theta). $$ * Fig 2: We separate the training of the marginal densities of the pretreatment covariates to ensure that the normalizing flow on the second line represents a copula flow. Training in a single line would not enforce that the marginal densities of the copula flow are uniform (except $V_{Y\|do(X)}$). Thus, the marginal flows and copula flow must be trained separately (see lines 175-178 in original submission). Training the multivariate flow on the empirical CDF of the covariates also provides flexibility to generate realistic synthetic data from populations with marginal densities differing from the training data while maintaining the underlying dependencies of the original dataset. * In their current form, we do not believe frugal flows can reliably infer ATEs from small to medium-sized datasets. Figure 3 in the attached material illustrates the sensitivity of the ATE estimate as a function of data size, using the synthetic data models $M_{1}$ and $M_{2}$ in section 4.1 (original paper). Our experiments show that ATE estimates are reliably inferred only at data sizes greater than $N \approx 10,000$. For this reason, we did not present ATE estimates for the real-world data, as they are both smaller than $10,000$. The e401(k) dataset is of similar size but with more than double the number of dimensions. This however does not impede us from generating realistic synthetic benchmarks as validated by Figure 2, Figure 3, and Table 2 in the document attached to the rebuttal. These provide empirical evidence that generated data is statistically similar to the original dataset. * The processed Lalonde dataset is of size $N = 614, D = 6$. We will add this information to the paper. * The parameter number of the frugal flow trained on the Lalonde and e401(k) data are 485243 and 106969 respectively. We will add this information to the paper. * In training, we perform a train-val split and use a “patience” criterion on the validation loss as a criterion to stop the training. Namely, we monitor the validation loss and stop training if the validation loss does not improve for a specified number of epochs (Chapter 5.5.2 in “Pattern recognition and machine learning” (Bishop and Nasrabadi, 2006)). We set the patience value to 100. This aims to prevent overfitting and saves computational resources by not continuing training unnecessarily. It is standard in machine learning model training and was implemented in the FlowJax package that we use as code-base to build the Frugal Flow package from. --- Rebuttal Comment 1.1: Comment: Thank you for providing a detailed response to my questions. I will keep my score of acceptance.	Summary: The paper introduces a generative modeling approach called Frugal Flows, designed to learn the data generation process with an explicit parametrization for the marginal causal effect of treatment on outcomes. Inspired by the frugal parametrization of marginal structural models, this approach models the marginal intervention distribution $p(Y\|do(X))$ directly, rather than the joint distribution $p(Y\|Z, do(X))$. This helps in preserving any constraints on the average treatment effect while flexibly modeling the data generation process. Frugal Flows employs copula flows to parameterize the model, accommodating constraints on the average causal effect and handling unobserved confounding during data generation. The authors validate the proposed method through experiments on both synthetic and real-world datasets, demonstrating its ability to generate realistic datasets with user-specified constraints. Strengths: * The paper's approach to validating causal models using simulated datasets is indeed impactful and relevant. It addresses a significant gap by allowing for general constraints on quantities of interest, such as average causal effect and unobserved confounding, during data generation. This capability is crucial because many prior generative modeling approaches for causal datasets either do not offer such flexibility or cannot ensure the preservation of these constraints, thus making this work a notable advancement in the field. * The paper is well-written, with clear explanations in the background sections on frugal parametrization and flows, which help the reader grasp the proposed approach. The details of the approach are well-articulated, and the experimental results are presented effectively. * The proposed approach is indeed novel. While it builds on established concepts like frugal parametrization, the specific application of normalizing flows for parametrization and its focus on average causal effect estimation represent a significant and innovative contribution. Weaknesses: My main concern with the work is the limited empirical validation of the proposed approach. Given that the primary contribution is the learning methodology rather than theoretical analysis, I would expect a more extensive set of experiments to validate its effectiveness. For example, prior research on generative modeling for causal inference, such as the work by [1], includes comprehensive experiments with various statistical tests to assess the realism of generated samples and a broader benchmarking of causal estimators. This paper would benefit from similar depth in its empirical evaluation. It would nice if the authors can conduct similar experiments to asses whether learned generative model generates realistic samples and evaluate it on more datasets. Also, the authors should compare with the prior works [1, 2] as baselines to establish which approach is the best at capturing the underlying data generation process, and empirically validate their claim (Section 2.6) that the proposed approach would be better than prior works at capturing used-specified constraints on the average causal effect. References [1] Neal, Brady, Chin-Wei Huang, and Sunand Raghupathi. "Realcause: Realistic causal inference benchmarking." arXiv preprint arXiv:2011.15007 (2020). [2] Harsh Parikh, Carlos Varjao, Louise Xu, and Eric Tchetgen Tchetgen. Validating causal inference methods. In International conference on machine learning, pages 17346–17358. PMLR, 2022. Technical Quality: 3 Clarity: 4 Questions for Authors: * A suggestion for the notation is that the authors could use $T$ instead of $X$ to denote the treatment variables in the paper. This way the notation would be less confusing, as $X$ represents a general random variable in Section 2.5 and Section 2.6 * Maybe there is a typo in Figure 2? The top row should be $\mathcal{F_{Z_i}}^{-1}(.)$ as we transforming the covariate $\{ Z_i \}$ to the correlated uniform variables $\{ V_{i} \}$ * I don't understand the Section 3.1.1 on copula flow for $X$ on $Z$. Why don't we directly model $p(X\|Z)$ using a normalizing flow and why do we need to parametrize using copula flow as $p(X\|Z)= p(X).c(X\|Z)$? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Yes, the authors have adequately addressed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed analysis of our paper. Before addressing the specific questions, we would like to comment on the important points raised in the weaknesses section. ## Weaknesses We agree that a more comprehensive validation of our proposed approach would strengthen the narrative and enhance the perception of frugal flows as a reliable method for improving synthetic data generation. The two papers referenced by the reviewer, which we also mention in our paper, provide algorithms most comparable to our proposal. Credence (Parikh et al, 2023) allows exact specification of the CATE in generative samples. In contrast, RealCause (Neal et al, 2021) only allows users to scale the causal effect post-hoc, which hinders its ability to model a null hypothesis of ATE=0, where all outcomes are scaled to zero. Additionally, frugal flows’ ability to model unobserved confounding is shared with Credence but not with RealCause. Therefore, we chose Credence as the more appropriate method for comparison. Following the reviewer’s suggestions, we reran the benchmarking simulations in section 4.2 with Credence, using the default parameters specified in their package. We present an illustration of the correlation matrix of the pretreatment covariates and the outcome for the original data, data samples from frugal flows, and two distinct data samples from Credence trained with different causal constraints. We refer the reviewer to Figure 1 and Figure 2 in the document attached to the rebuttal. Frugal flows generate samples closely resembling the original data, particularly for the larger e401k dataset. The samples generated by Credence also perform well in generating realistic samples for the e401k dataset. However, we note that altering the causal effect rigidity significantly impacts the covariate joint and introduces dependencies not present in the original data. A key advantage of frugal flows is the need to optimize the model fitting only once, allowing for direct modification of causal constraints without affecting the covariate joint and propensity score. We also conducted the same multivariate statistical tests used by Neal et al. (2021) to validate RealCause. These tests include the Maximum Mean Discrepancy test (Gretton et al., 2012), the energy test (Szekely & Rizzo, 2013), the Friedman-Rafsky test (Friedman & Rafsky, 1979), and the k-nearest neighbor (kNN) test (Friedman & Rafsky, 1983), all implemented in the $\texttt{torch-two-sample}$ Python library. Across both datasets, the tests suggest that frugal flows generate samples that are more comparable to the original data than those generated by Credence. In most tests, there is not sufficient evidence to claim that samples from frugal flows are distinguishable from the original data. Regarding benchmarks, we conducted experiments on two real-world datasets (Lalonde, e401k), which is the same number as the Credence paper (Lalonde, Project STAR). Due to time constraints, we could not obtain results by the end of the rebuttal period but are happy to run additional experiments to validate and contrast frugal flows against Credence using the Project STAR dataset, which was also explored in the Credence paper. This would put us on par with the RealCause paper, which validates their method using three real world datasets. ## Questions In response to the reviewer’s questions, we provide the following answers: 1) We fully agree with your suggestion to change X to T when referring to treatments in the paper. Although we have not made this change in the current rebuttal to avoid confusion among other reviewers, we have incorporated this change in the manuscript. 2) Thank you for spotting the typo! You are absolutely correct, and we have corrected the error in the manuscript. 3) You’re right here—one could directly model $p_{X\|\mathbf{Z}}$ using a normalizing flow, which would be a valid frugal model. We model $c_{X\|\mathbf{Z}}$ to encode a degree of unobserved confounding in the generated data by sampling the ranks $U_{X\|\mathbf{Z}}$ and $U_{Y\|\text{do}(X)}$ from a non-independence copula. Assuming ignorability, these ranks would be independent. However, unobserved confounders imply marginal dependency between these ranks. Sampling from a copula replicates this effect, as demonstrated in the far-right plots in Figures 3 and 4 in the original submission. We discuss this in lines 260-265, explaining how to simulate unobserved confounding. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response! My questions have been addressed and I have increased my rating accordingly!	Summary: This paper proposed a generative model called Frugal Flows making use of copula flows to infer about marginal causal effects by simulating the data generating process. Strengths: - The problem of inferencing marginal causal effects is an interesting and important problem - The idea of using generative models to estimate the marginal effects in the paper is interesting Weaknesses: See questions. Technical Quality: 3 Clarity: 3 Questions for Authors: I have two question -- 1. A highly related (and I suspect might be viewed as a "dual" appproach to your Frugal Flow) is the statistical matching (e.g., bipartite matching) to estimate the average treatment effect. It would be very informative to compare this as one of the baselines in your benchmarking and validation, as this also shed lights on how these two different schools of causal inference may (or may not) converge on ATE estimation. 2. I think it is good to use real data for benchmarking/validation (but perhaps benchmarking is a bit strong here since only two datasets were used), but usually it is unclear how to interpret the results since the ground truth is unknown. Can you design and run some controlled synthetic experiment to verify the model? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s comments and suggestions for our work. Indeed, we agree that frugal flows are an interesting addition to inference algorithms for estimating marginal causal densities in large datasets using Normalizing Flow models. In addition, we believe a key contribution of frugal flows lies in their ability to enable users to learn representations of realistic datasets and precisely customize causal features such as the ATE, degree of unobserved confounding, propensity score, and simulating from discrete outcomes. When referring to model benchmarking, we propose that frugal flows can generate realistic datasets with customizable causal features for benchmarking purposes; being able to simulate realistic data with a ground truth is clearly much more useful for benchmarking novel causal methods, than using real datasets with unknown underlying causal effects (see Section 2.6 in the original submission). In response to the reviewers questions, we sincerely appreciate your suggestions for improving the scope of our inference experiments. We have made a couple of modifications and additional experiments based on your suggestions: 1) We agree that Section 4.1 would be more compelling if frugal flows were contrasted with other causal inference methods, rather than linear regression, which was initially included to demonstrate the complexity of the confounding. Consequently, we have added results from a statistical matching algorithm as suggested. Additionally, we present ATE estimates calculated using an alternative method, Causal Normalizing Flows (Javaloy et al, 2023), recommended by other reviewers. The matching algorithms yield results consistent with frugal flows, whereas the CNFs produce ATE estimates with higher variance, and in some cases a bias, compared to both statistical matching and frugal flows. 2) In response to the desire for more comprehensive synthetic experiments, we demonstrate that with sufficiently large data sizes, they can accurately identify the true ATE across a range of simulated data. These results are presented in Table 1 (in the document attached to the rebuttal) for models with an ATE of 1. We have also rerun similar experiments with a true underlying ATE of 5, incorporating each of the algorithms suggested by the reviewers. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my concerns.	Summary: This work proposes to leverage existing neural density estimators (specifically, normalizing flows) to exploit a newly-proposed "frugal parametrization" that can capture the causal marginal distribution of an underlying causal model. Under this parametrization, the authors show how to specify and train each component of the model, and thus train the proposed Frugal-Flows to match the observational distribution as closely as possible, while being able to tune the marginal causal effect present in the generated data. This way, frugal flows can be used to generate synthetic causal benchmarks that closely represent the _observational_ data while having more difficult-to-estimate causal effects, putting existing approaches to the test. Strengths: - S1. The proposed frugal flows provide a way of generating new datasets that can be challenging from a causal-inference point of view, which I believe _important_ to test new and existing methods. - S2. The construction of the proposed architecture is quite rich in details. - S3. I find the frugal parametrization conceptually quite interesting. - S4. The authors motivate different scenarios for frugal flows in Sec. 3.2, as well as empirically show positive results on some synthetic and real-world scenarios. Weaknesses: - W1. I find the frugal parametrization to be extremely under-explained, relying too much on the reader having full knowledge of the referenced work. Similarly, there is little to no explanation/intuition on why the frugal parametrization would properly capture the marginal causal distributions. - W2. The lack of explanations also applies to other concepts, e.g., "conditional ignorability" (line 39) "variation independence" (line 82, and I know the definition is later in App. A), or why copula-based flows would target conditional causal effects instead of marginal causal ones (line 182). (similar with lines 221 and 229) - W3. There are no mention to related works that propose similar ways of constructing causal benchmarks. From a 1-min search in google scholar, I already found some likely relevant works: [Work 1](https://arxiv.org/abs/2406.08311), [Work 2](https://arxiv.org/abs/2011.15007). - W4. I find the experiments a bit underwhelming, specially those from Section 4.1. The authors should at least show how is the fitting of the observational likelihood, and if they want to show the capabilities of frugal flows for causal inference (and not only causal-benchmark generation), they should compare with other methods like [Causal Normalizing Flows](https://arxiv.org/abs/2306.05415). Technical Quality: 3 Clarity: 2 Questions for Authors: - Q1. I am not sure that I understand what does the dotted red line represent in the boxplots. - Q2. Doesn't the statement in lines 291-294 directly contradict what you say later in lines 296-297? - Q3. Why is Figure 1 placed there? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: I think limitations are properly discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for the reviewer’s feedback and suggestions for our submission. Below are our responses to the noted weaknesses and questions: # Weakness 1 We acknowledge that the frugal parameterization was briefly introduced. Due to page constraints, we provided a brief overview in the main text and a more detailed discussion in the appendix, drawing largely from Evans and Didilez (2024). In response to the reviewer’s comments, we will reorganize section 2.2 to better explain frugal models, the relevance of copulas, and their focus on the causal margin. We will also enhance the appendix with a comprehensive introduction for those interested in technical details. Regarding why the frugal parameterization captures the marginal causal distribution when the dependency measure is parameterised by a multivariate copula, consider the copula for the distribution of $\mathbf{Z}$ and $Y$ conditional on $X$: $$c(F_{Y\|do(X)}, F_{Z_1\|X},\dots, F_{Z_{D}\|X}).$$ For an intervened distribution, all pretreatment covariates $\mathbf{Z}$ are marginally independent of $X$, simplifying the copula to $$ c(F_{Y\|do(X)}, F_{Z_1},\dots, F_{Z_{D}}),$$ and so the intervened joint density becomes $$p(\mathbf{Z},~Y \mid do(X)) = p_{Y\|do(X)} \cdot \prod_{d=1}^{D} p_{Z_d} \cdot c(F_{Y\|do(X)}, F_{Z_1},\dots, F_{Z_{D}}),$$ where $p_{Y\|do(X)}$ is the marginal causal effect of $X$ on $Y$. The final density required to parameterise the observational distribution is the propensity score $p_{X\|\mathbf{Z}}$, which does not affect the aforementioned marginal densities in the observational model (see Chapter 10 in "Information and Exponential Families"(Barndorff Nielsen, 1978)). We hope this clarifies why frugal parameterizations target marginal causal quantities and will add these clarifications to the manuscript. # Weakness 2 We agree that improved clarity in definitions and terminology would enhance readability and are happy to make these additions to the manuscript. We use the definition of conditional ignorability (equivalent to conditional exchangeability) stated in “Causality: Models, Reasoning and Inference” (Pearl, 2009), where the marginal potential outcomes are independent of the treatment, conditional on the observed covariates. Regarding why copula-based flows target the conditional effect, copula flows require all marginal flows to be trained independently of the copula likelihood. This is because copula flows infer a joint density within a unit hypercube and do not constrain marginal densities to be uniformly distributed. Thus, empirically uniform data must be provided to generate samples from a copula. Since one of the marginal densities being trained is a conditional quantity ($\mathcal{F}_{Y\|X}$), training this flow independently targets the conditional of $Y$ on $X$. Frugal flows target the marginal causal effect as they are trained jointly with the copula flow, with the causal effect density constrained to be uniformly distributed, as outlined in Figure 2 of the original submission. # Weakness 3 Our aim was to contrast our contribution to generating synthetic causal benchmarks against other methods in the literature, as discussed in section 2.6. We reference three key papers on line 200, including the second work the reviewer mentioned. We appreciate the reviewer sharing Work 1, which we were not aware of and was submitted to arXiv after our paper submission! However, Work 1 focuses more on validating structural causal benchmarks, whereas Work 2 (and the other two papers we reference) focus on generating observational data with customizable causal effects. Credence allows exact specification of the CATE in generative samples. In contrast, RealCause only allows users to scale the causal effect post-hoc, which hinders its ability to model the null hypothesis of the ATE=0, because every outcome is scaled to zero. Additionally, frugal flows’ ability to model unobserved confounding is shared with Credence but not with RealCause. Therefore, we chose Credence as the more appropriate method for comparison. We assessed the realism of samples from frugal flows for the Lalonde and e401k datasets, comparing them to those from Credence. These results are presented in Figures 1 & 2 (in the attached material) showing the correlation matrices of covariates and outcomes for frugal flow and Credence samples against the original data. We find that Credence's fitting process is sensitive to the causal effect one wishes to constrain, requiring hyperparameter optimization for every setting, while frugal flows need optimization only once, allowing for sample generation from different causal models without re-training. # Weakness 4 Following the reviewer’s recommendations, we used the CNF package to estimate ATEs in our synthetic data experiments with different causal effects, comparing the results against frugal flows and a causal statistical matching algorithm. These results are presented in Table 1 in the attached document. CNFs showed higher variance in ATE estimates than frugal flows. In addition, we also provided figures showing the observational likelihood loss during model training for both real-world datasets in Figures 4 & 5. # Additional Questions: * The dotted red line represents the customized ATE parameter of the frugal flow, indicating that causal inference algorithms infer the specified ATE. We will clarify this in the updated version of the paper. * Lines 291-294 indicate that frugal flows are not recommended for inferring ATEs in small to medium-sized datasets. Figure 3 in the attached experimental results showing that frugal flows struggle to converge to the true ATE of synthetic data when $N < 10,000$. However, for benchmarking, frugal flows can customize the causal properties of generated data and generate realistic samples with specified ATEs even for smaller datasets like Lalonde. * We agree that Figure 1 should be placed closer to section 2.2 and will make this change in the final version. --- Rebuttal Comment 1.1: Comment: I thank the authors for their super detailed responses and the additional experiments. I do think they add quite some value, and if the authors use the feedback to improve the readability of the manuscript (and thus how welcoming it is for newer audiences), I think it has potential to be quite a good paper. With respect to the new results, I am a bit surprised by the results of CNF on model 1. Just to make sure, I'd encourage the authors to check that they used a model different from MAF for the base model, as this is not a universal density approximator. In any case, I am happy with the response from the authors and I will update my score.	Rebuttal 1: Rebuttal: We thank all the reviewers for their comprehensive commentary and very helpful suggestions for our paper. In addition to the individual responses to each reviewer, we provide a more global summary of what we believe were the core themes across all four reviews. These centre around 1) clarity and terminology, 2) further inference experiments (following Section 4.1 in the original submission) and 3) a more comprehensive validation of the synthetic data generation in Section 4.2, showing that frugal flows indeed generate realistic data samples which resemble the original dataset. # Clarity and Terminology Some reviewers suggested to increase clarity and precision in our definitions and terminology. In response, we have made the following changes: 1) Definitions and Assumptions: We are happy to add more detailed explanations and explicitly define terms such as conditional ignorability and variation independence. In addition, we wish to clarify the assumptions required for frugal flows; in particular, the covariate set must only include pretreatment covariates 2) Clarification on Frugal Parameterization: We will elaborate on our explanation of the frugal parameterization and more acutely describe how targets the causal margin rather than the conditional effect. This includes an expanded discussion on the relevance of copulas and how they enable frugal parameterization to capture the causal margin effectively. # Further Inference Experiments Reviewers expressed a desire for more extensive inference experiments, expanding on the contents in Section 4.1 in the original submission. We have addressed this by: 1) Benchmarking Against Comparable Methods: We are grateful for the additional references provided by the reviewers, in particular the Causal Normalizing Flows (CNFs) paper, which we are happy to add to the paper. We reran the experiments in Section 4.1 and use CNFs to estimate the ATEs in simulated datasets. We use the hyperparameter settings recommended in the package and do not perform hyperparameter tuning. For the flow architecture, we use Neural Spline Flows, as the paper reports in Appendix D.3 that they yield a better performance than simple Masked Autoregressive Flows. In addition to a (misspecified) linear outcome regression in the original submission, we use an implementation of a causal statistical matching algorithm to add further variety to the inference algorithms 2) Expanded Experiments: We conducted experiments using simulated data generated from two true ATEs (1 and 5) using models $M_{1}$ and $M_{2}$ described in Section 4.1 in the original submission ($M_{3}$ omitted for now due to time constraints, but we will update the paper with the results of all three models). The results, which are presented in Table 1 in the attached material, show that frugal flows consistently identified the true ATE with the lowest standard error. CNFs performed worse for data generated in $M_{1}$, demonstrating a clear bias, but correctly estimated the true ATE in data generated from $M_{2}$, with a lower standard deviation than statistical matching but higher than frugal flows. # Validation of Synthetic Data Generation We agree with the reviewers that comprehensive validation of our approach is crucial. To this end, we make the following comments: 1) Contrasting against Credence and RealCause: We have enhanced the validation section by adding a discussion contrasting Credence (Parikh et al, 2023) and RealCause (Neal et al, 2021), the two most similar existing methods for generating realistic synthetic data for validating causal inference algorithms. Credence allows the user to exactly specify what the conditional ATE is in the underlying data, much in the same way as frugal flows allow the user to exactly specify the ATE. RealCause only allows the user to scale the causal effect posthoc, which in particular hinders its ability to model a null hypothesis (i.e all observations will be set to zero); this is straightforward for frugal models (see Section 3 in Evans and Didelez, 2024). Furthermore, frugal flows’ ability to model unobserved confounding is shared with Credence and not with RealCause. For these reasons we chose Credence as a more appropriate method to compare experimentally against frugal flows, and discuss these experiments in the next paragraph. 2) How Realistic are the Synthetic Data?: We compared frugal flows against Credence using both the Lalonde and e401(k) datasets from our original submission. We present correlation matrices of pretreatment covariates and outcomes for the original data against samples generated from frugal flows and Credence trained with different causal constraints.These results are shown in Figures 1 and 2 in the attached document. Frugal flows generated samples closely resembling the original data, particularly for the larger e401(k) dataset. While Credence performed well in generating realistic samples, altering the causal effect rigidity impacted the covariate joint and introduced dependencies not present in the original data. A key advantage of frugal flows is the need to optimize the model fitting only once, allowing for direct modification of causal constraints without affecting the covariate joint and propensity score. 3) Multivariate Statistical Tests: We conducted a variety of multivariate statistical tests (Maximum Mean Discrepancy test, energy test, Friedman-Rafsky test, and k-nearest neighbor test) to assess whether the synthetic data from the generative models is statistically similar to the original training data. These results are presented in Table 2 of the attached document, and suggest that frugal flows do indeed generate realistic looking datasets. Once again, we thank the reviewers and the area chair for their consideration of our paper, and we are happy to answer any follow up questions you may have. Pdf: /pdf/2e4ef09527061e419023c6f030039630316cd645.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Slicing Vision Transformer for Flexible Inference	Accept (poster)	Summary: The paper targets scaling down Vision Transformers (ViT) to fit environments with dynamically changing resource constraints. The authors propose Scala, a framework enabling a single network to represent multiple smaller ViTs with flexible inference capability by activating various subnets during training. Scala introduces Isolated Activation to disentangle the smallest sub-network and uses Scale Coordination to provide stable and accurate learning objectives. Empirical validations on different tasks show that Scala achieves scalable representation with one-shot training, matching the performance of Separate Training without modifying the original ViT structure. Scala demonstrates an average improvement of 1.6% on ImageNet-1K compared to previous methods, using fewer parameters. Strengths: 1. The problem is important in practice. 2. The experimental results seem decent. Weaknesses: 1. My major concern is that, the same aim of adapting ViTs to dynamically changing resource constraints, can also be achieved by multi-exit networks, e.g., [1, 2, 3]. However, the paper does not discuss these highly relevant works or compare with them. Hence, I vote for rejection. 2. The method seems to lack novelty. 'smaller ViTs are intrinsically the sub-networks of a larger ViT with different widths' is not a surprising observation. The key techniques (e.g., Isolated Activation and Knowledge Distillation) are not new (naive or have been widely adopted). [1] Huang, Gao, et al. "Multi-Scale Dense Networks for Resource Efficient Image Classification." International Conference on Learning Representations. 2018. [2] Wang, Yulin, et al. "Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition." Advances in neural information processing systems 34 (2021): 11960-11973. [3] Han, Yizeng, et al. "Dynamic perceiver for efficient visual recognition." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. Technical Quality: 2 Clarity: 2 Questions for Authors: Please refer to Weaknesses. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors have addressed the limitations and potential negative societal impacts of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the Reviewer's feedback. We provide further explanations to clarify the Reviewer's concerns based on several key points as below. * Weakness 1: Discussion with dynamic networks. We thank the Reviewer for the valuable suggestion and we will add discussion with those dynamic networks in the final version, including MSDNet[1], RANets[2], GFNet[3], DVT[4], CF-ViT[5], Dyn-Perceiver[6], etc. However, the differences between our work and this line of research are: (1) Motivation: dynamic networks are designed to reduce inference costs with dynamic computational graphs; Scala is proposed to make ViT become slimmable without touching the specific architecture design. (2) Method: dynamic networks tailor their computational graphs to accommodate varying inputs, thereby optimizing resource allocation on a per-sample basis. In contrast, Scala treats every sample equally and modulates the computational costs by adjusting the width ratios which aligns with the inherit design of ViT and its inference costs can be explicitly controlled by a single hyperparameter $r$. Essentially, dynamic networks are agnostic with Scala and we will consider the integration of those two lines of research as our future plan. Weakness 2: Lack of novelty. Thank you for the comment. While we do agree that 'smaller ViTs are intrinsically the sub-networks of a larger ViT with different widths' is not a surprising observation, making ViTs become slimmable is indeed non-trivial. According to our analysis in Sec. 2, ViTs display minimal interpolation ability which emphasizes that the problem we target to solve is very challenging as the intermediate subnets ($X=13$) are only trained for around 18 epochs and they are expected to match the performance of networks that are trained for 100 epochs. To the best of our knowledge, we are the first to propose Isolated Activation, which seems simple but is very effective as we have abundant analysis (Sec. 3) to support it. Moreover, the knowledge distillation in Scala is different from the conventional setting. Instead of choosing the full model as the teacher for all subnets, we enable the intermediate networks to be the teacher as well which fills the model gap between $F^{l}\left (\cdot \right )$ and $F^{s}\left (\cdot \right )$ and provide simplified optimization objective for small subnets. This is non-trivial as we find adding a teacher network to the baseline method US-Net will result in performance decreases in the smaller subnets (please see our reply to Reviewer rLHb) which further supports our motivation to propose Progressive Knowledge Transfer. Moreover, Noise Calibration is critical in our setting as well because those intermediate teachers cannot provide faithful predictions at the early stages and it is essential to add Cross-Entropy loss to those subnets during training which is overlooked in previous baseline methods. Based on these factors, Scala can solve this challenging problem in a simple but effective way. * We hope the explanations could address your concerns and we would appreciate it a lot if you could recognize the contributions of our work. [1] Huang G, Chen D, Li T, et al. Multi-scale dense networks for resource efficient image classification[J]. arXiv preprint arXiv:1703.09844, 2017. [2] Yang L, Han Y, Chen X, et al. Resolution adaptive networks for efficient inference[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 2369-2378. [3] Wang Y, Lv K, Huang R, et al. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification[J]. Advances in Neural Information Processing Systems, 2020, 33: 2432-2444. [4] Wang Y, Huang R, Song S, et al. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition[J]. Advances in neural information processing systems, 2021, 34: 11960-11973. [5] Chen M, Lin M, Li K, et al. Cf-vit: A general coarse-to-fine method for vision transformer[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2023, 37(6): 7042-7052. [6] Han Y, Han D, Liu Z, et al. Dynamic perceiver for efficient visual recognition[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 5992-6002. --- Rebuttal Comment 1.1: Title: Kind Reminder of Deadline Comment: Dear Reviewer, Thanks for your efforts so far in our paper review. Currently, most of the concerns from other Reviewers have been well-addressed and we are eager to know whether our response has resolved your concerns. Concretely, we have added the discussion between Scala and dynamic networks which shows that our work is agnostic with this line of research. Moreover, we have clarified the contribution of our work and further expanded the application scenarios of Scala to foundation models like DINOv2. Due to the coming discussion deadline, we would like to kindly remind the Reviewer if our response through the whole rebuttal has addressed any of your concerns and helped the Reviewer to reevaluate this work. We really appreciate it if you could give us any feedback and your opinions are rather important to us. Thank you very much for your time! Sincerely, Authors --- Rebuttal 2: Comment: Dear reviewer, please have a look at the rebuttal and indicate whether it addressed some of your concerns. --- Rebuttal 3: Comment: I highly appreciate the authors' response. However, I still think the lack of discussions and comparisons with dynamic networks in the original submission is an important limitation. Moreover, although the authors have provided some clarification, Isolated Activation seems like simply an engineering trick, and the proposed knowledge distillation method seems too incremental compared to existing works. These techniques may be insufficient as significant scientific contributions. I increase my rating to 4. However, I still vote to reject this paper. The major reasons are (1) lack of discussions on important related works, and (2) weak technical novelty.	Summary: The paper presents Scala, a novel framework for scalable representation learning developed from US-Net. It identifies the issues of directly applying US-Net to ViTs and proposes solutions including Isolated Activation, Scale Coordination, and Stable Sampling. These innovations enable Scala to output several sub-networks in one-shot learning. Extensive experiments on various network architectures and datasets demonstrate that the sub-networks produced by Scala consistently outperform those generated by separate training, with significantly reduced training time. Strengths: Originality: Scala addresses the limitations of US-Net and successfully applies the concept of scaling to ViT backbones. This is a significant step in the adaptation of scaling methods for more complex network architectures. Quality: The paper supports its claims with extensive experimental results, providing strong evidence for the effectiveness of Scala. Clarity: The paper is clearly written and well-organized, making it accessible and easy to follow. Significance: Scala has the potential to influence future research directions in scaling ViTs. Weaknesses: Originality: The novelty of Scala is somewhat constrained. For instance, Noise Calibration does not show a distinct difference from standard knowledge distillation. Essentially, Scala integrates US-Net with an alternative activation for the smallest subnet and fixed scaling ratios. Quality: The authors might consider emphasizing results from a more standard 300-epoch ViT training schedule to align with common practices in the field. Clarity: No further issues. Significance: The challenge of scaling ViTs with arbitrary ratios remains unresolved. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Regarding the issue shown in Fig. 4, is it always the smallest subnet that causes the issue, or does it occur with subnets having a scaling ratio near 0.25? 2. Can the authors clarify any differences between Noise Calibration and standard knowledge distillation? 3. What would be the impact if the distillation part were discarded and only Cross-Entropy (CE) loss were used for the initial epochs? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The novelty and significance of Scala are somewhat limited, as discussed in the weaknesses section. However, the extensive experimental results provide a robust foundation for the claims made in the paper. Overall, the work is well-executed and makes a valuable contribution to the field, justifying a recommendation for weak acceptance. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the Reviewer's approval and valuable comments. We respond to the Reviewer's concerns as below. * Weakness 1: Fixed scaling ratio. Thank you for the comment. The hidden dimension of ViT has to be an integer multiple of the number of heads (e.g., 6/12) so ViT cannot support arbitrary width ratios like CNN. For instance, the theoretical maximum number of networks that ViT-B could represent is 25 if $s=0.25$. However, we do acknowledge that Scala still cannot generalize to unseen width ratios as this issue is essentially related to the architecture design of transformer and we constrain ourselves from modifying the conventional structure as it has been well-optimized in various scenarios. On the other hand, it further emphasizes that the problem we target to solve is very challenging as the intermediate subnets ($X=13$) are only trained for around 18 epochs and they are expected to match the performance of networks that are trained for 100 epochs. Weakness 2: Results of 300 epochs. Thanks for the advice and we will emphasize the 300-epoch results of Sec. 5.3 and 5.4 in the final version. Question 1: Root cause. Thanks for the great question. It is the smallest subnet that causes the issue no matter what the width is and we validate it by setting $s=0.5$ to be the smallest scaling ratio and removing Isolated Activation. The table shows that constantly activating the smallest subnet in the naive way still results in worse performance even if the width ratio is larger than 0.25. \| Method \| $r=0.50$ \| $r=0.75$ \| $r=1.00$ \| \| :-----: \| ------------ \| ------------ \| ------------ \| \| Scala ($s=0.5$) \| 71.2% \| 74.5% \| 76.8% \| \| Scala ($s=0.5$) w/o IA \| 70.2% \| 73.1% \| 75.9% \| Question 2: Noise Calibration. Compared to standard knowledge distillation, the teacher of those subnets in Scala is dynamically changing (different widths) and will be optimized during training like the students as their weights are shared. On the one hand, the intermediate teachers serve as teacher assistants which fill the gap between $F^{l}\left (\cdot \right )$ and $F^{s}\left (\cdot \right )$ and provide simplified optimization objective for small subnets, in contrast to US-Net which utilizes $F^{l}\left (\cdot \right )$ as the teacher for all subnets and results in inferior performance. On the other hand, it also indicates that those teachers cannot provide faithful predictions at the early stages and it is essential to add Cross-Entropy loss to those subnets during training which is overlooked in previous baseline methods. Question 3: Impact of only CE at early epochs.** Thanks for the valuable suggestion. We further conduct experiments by removing the distillation loss of subnets in the first 10 epochs (denoted by ) and there is only minor performance change in the subnets which suggests that Noise Calibration is indeed helpful at early stages. \| Method \| $r=0.25$ \| $r=0.50$ \| $r=0.75$ \| $r=1.00$ \| \| :-----: \| ------------ \| ------------ \| ------------ \| ------------ \| \| Scala \| 58.7% \| 68.3% \| 73.3% \| 76.1% \| \| Scala \| 58.8% \| 68.1% \| 73.6% \| 76.1% \| *** We hope the explanations could address your concerns and we would appreciate it a lot if you could recognize the contributions of our work. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response. I have updated my vote to accept. --- Reply to Comment 1.1.1: Title: Thanks for the support Comment: Dear Reviewer, Thank you so much for the constructive comments for us to improve the work and we sincerely appreciate your support. Thanks again for being with us so far. Sincerely, Authors	Summary: The paper introduces Scala, a novel framework designed to effectively scale down Vision Transformers (ViTs) for use in environments with fluctuating resource constraints. The key insight is that smaller ViTs can function as sub-networks within a larger ViT, differing mainly in width. Scala enables a singular network architecture that can emulate multiple smaller ViTs, thereby offering versatile inference capabilities while maintaining the structural principles of ViTs. The framework uniquely incorporates multiple sub-networks during its training phase, utilizes Isolated Activation to differentiate the smallest sub-network, and implements Scale Coordination to streamline the learning objectives for each sub-network, aiming for simplicity, stability, and accuracy. The empirical results across various tasks confirm that Scala can learn scalable representations efficiently with a single training iteration, maintaining the integrity of the original ViT architecture and achieving performance on par with networks trained separately. Strengths: The proposed Scala framework aims to enhance Vision Transformers (ViTs) by enabling them to learn scalable representations suitable for flexible inference. This is achieved through two key innovations: Isolated Activation, which effectively disentangles the representation of the smallest subnet to maintain clarity and specificity, and Scale Coordination, which ensures that each subnet within the larger network receives simplified, consistent, and accurate signals. These mechanisms are designed to optimize the performance and scalability of ViTs, addressing common challenges in adapting these architectures to varied and dynamic operational contexts. Weaknesses: 1. Recent papers[1,2,3] with "Scalable" usually scale ViT to billion size with large scale datasets like DFN, JFT, and Datacomp. Therefore, I suggest authors should reconsider if the experiments can support "Scalable". [1] Zhai, X., Kolesnikov, A., Houlsby, N., & Beyer, L. (2022). Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 12104-12113). [2] El-Nouby, A., Klein, M., Zhai, S., Bautista, M. A., Toshev, A., Shankar, V., ... & Joulin, A. (2024). Scalable pre-training of large autoregressive image models. arXiv preprint arXiv:2401.08541. [3] Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., ... & Houlsby, N. (2023, July). Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning (pp. 7480-7512). PMLR. Technical Quality: 3 Clarity: 3 Questions for Authors: please refer the weakness. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This paper addresses the limitations in conclusion Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the Reviewer's comments to point out the confusing description and we make the response as below. * Weakness 1: Phrase. Thank you for your insightful suggestion. We will revise our presentation in the final version to prevent any potential misunderstanding. Our motivation for adopting the term "scalable representation learning" stems from the observation that different methods (e.g., DeiT, MAE, DINOv2) learn distinct representations that are not inherently slimmable (see Appendix: A.3). Scala is proposed to learn the original representations while incorporating slimmability. Therefore, we initially chose "scalable representation learning" to describe our method. However, we acknowledge that the term "scalable" may lead to confusion, and we will revise it in the final version to enhance clarity. * We hope the explanations could address your concerns and we would appreciate it a lot if you could recognize the contributions of our work.	Summary: This paper advances an approach for training Vision Transfomers (ViTs) such that at inference time they can be dynamically adjusted to fit different budget constraints with reduced drops of performance. To this end, the authors introduce Scala, a framework that allows a single network to encapsulate and train simultaneously multiple sub-networks of different capacities and widths. The methodological backbone of this work are the Universally slimmable networks (US-Net) [37], originally devised for CNNs. The authors identify and analyze a few flaws of US-Nets: difficulty to generalize to ViTs, small interpolation and extrapolation ability to sub-network size unseen during training, impact of sustained activation of the smallest sub-network that coupled with the sandwich rule for selecting sub-networks during training leads to an over-emphasis on it at the expense of the other sub-networks. The authors propose two simple strategies towards such a method for ViTs: (i) Isolated activation that separates the smallest sub-network from the other sub-networks; (ii) scale coordination consisting of a set of heuristics to ensure that each sub-network gets simple, accurate and stable learning objectives: (a) progressive knowledge transfer from larger networks to smaller ones in gradual decrease of capacity, (b) stable sampling of intermediate width ratios to avoid large variations in capacities in the sandwich, (c) noise calibration, essentially a composite loss of supervised cross-entropy and distillation from the bigger sub-network. Scala is evaluated on several settings on the ImageNet-1k dataset with ViT-Ti/S/B, hybrid CNN-ViT architectures, lightweight networks, but also for dense prediction on semantic segmentation and self-supervised pre-training with interesting results. The baselines used here were Separate Training, Autoformer and US-Net. Strengths: ### Significance - the paper deals with a challenging and useful task for deploying ViT models into different operational settings with different computational constraints without retraining or distilling specific architectures each time - although a computational overhead is expected for such methods, the main components of Scala are relatively simple and make sense - Scala achieves good performance with a higher boost in the low parameter regime ### Originality - the proposed contributions are somehow incremental as they are improving the US-Net prior work, but do have some novelty and they are simple. ### Clarity - in general this work is well argued and easy to follow. The authors construct well the arguments regarding the challenges when going from CNNs to ViT with US-Net and how to construct their Scala approach. ### Quality - the paper offers several experiments and studies in the main paper and in the appendix (longer training, fast interpolation, ablation of components) that are well thought and improve the understanding of the method. - I appreciate the experiments beyond image classification, on semantic segmentation, as well as the self-supervised pretraining and subsequent linear probing on a downstream task. Weaknesses: ### "Scalable" naming - I think that the framing of the method as _"scalable representation learning"_ is quite confusing as it's not representative for this task, it's not a name used by other related works. Importantly, it can be easily mistaken with most works that use "scalable" for depicting the ability/property of a system (method, architecture) to handle a growing amount of data, parameters, and the potential to accommodate this growth. In other words "scalable" is rather used for depicting scaling up, whereas this work depicts the property of the proposed approach to accommodate sub-networks of different lower sizes/scales from the original. - maybe other names us in related works would be more appropriate here: slimmable, elastic, modular, flexibile inference, etc. ### Limited baselines and related work - some relevant related works dealing with tranformer networks are either just briefly mentioned, e.g., Matformer [18], or not mentioned at all, e.g., SortedNet [a], Early exit [b] - One of the main baselines, US-Net is originally designed for CNNs and, as the authors mentioned, moving to ViTs is not straightforward. Matformer is criticized for the limited number of models produced, but can be considered in the several experiments with X=4 sub-networks. Matformer and SortedNet could be included in the experimental evaluation ### Scope of experiments - While the authors considered several settings for computer vision tasks (image classification, segmentation, light architectures), transformer architectures are also encountered in NLP (as mentioned by the authors in L56). In such cases the original models can have much more parameters and elastic inference for lower computational budgets would be of high interest. - It would be useful to include an experiment from NLP in the style of those from Matformer or SortedNet. - The biggest architectures used here is a ViT-B (~86M params). Extending experiments to larger modern architectures would be definitely useful and interesting. ### Clarity - it's not always clear in the text and cost estimations that Scala needs a pre-trained full network as teacher for the distillation. This add some cost in compute and time in the end. Besides it's not clear whether US-Net also needs and uses a pre-trained teacher in the reported results. - in the intro, the authors mention that they address the issue of minimal interpolation ability of ViTs. Results from Table 2 show that the interpolation abilities of ViTs with Scala are still very low. However the fast interpolation strategy from $\S$A.2 is actually interesting for practical settings even though not fully solving this issue. It might be worth moving up in the main paper. - the idea of the transferability experiment ($\S$5.4) with DINOv2 is nice. From the description it is not clear whether DINOv2 was used as teacher for the distillation or also as supervised pre-training on ImageNet-1k? Or the pre-training on ImageNet-1K was done in a supervised manner as in previous experiments? - the ablation experiment from Table 6 is nice. However the presentation with removing one component at once offers only a partial understanding of the contributions of each module. Different configurations with different modules in on/off mode should give a better global understanding. References: [a] Valipour et al., SortedNet: A Scalable and Generalized Framework for Training Modular Deep Neural Networks, arXiv 2023 [b] Xin et al., Deebert: Dynamic early exiting for accelerating bert inference, ACL 2020 Technical Quality: 3 Clarity: 2 Questions for Authors: This paper takes an interesting direction of study: how to train ViTs such that a fine-grained elasticity in terms of sub-network sizes is possible at runtime. The proposed Scala approach is well described, makes sense and achieves good results in several computer vision settings. I do have a few concerns related to the phrasing of this type of works ("scalable representation learning") which can be confusing, the absence of larger architectures and of recent relevant baselines. My current rating is mildly positive (rather on the fence though) and I'm looking forward for the rebuttal. Here are a few questions and suggestions that could be potentially addressed in the rebuttal or in future versions of this work (please note that suggested experiments are not necessarily expected to be conducted for the short rebuttal period): 1. Please clarify the points raised in the clarity section: use of teacher model for Scala and US-Net, implementation of transferability experiment. 2. Comparison of training cost between Scala (including teacher training), US-Net and Separate Training baselines. 3. Add a discussion of differences and when possible experimental comparison with Matformer and SortedNet baselines on image classification or semantic segmentation. 4. Extension of experiments to NLP architectures and tasks in the style of SortedNed, Matformer Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors addressed some of the limitations in the conclusion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the Reviewer’s detailed comments and constructive suggestions for us to improve our work. We make the response as below. * Weakness 1: Phrase. Thanks for the great suggestion and we will modify the presentation in our final version. Our motivation for adopting 'scalable representation learning' is that different methods (DeiT, MAE, DINOv2) will learn different representations, which are not naturally slimmable (see Sec A.3). Scala is proposed to learn the original representations but with the slimmable ability, so we adopt 'scalable representation learning' to describe our method. However, we do agree that the term 'scalable' may result in misunderstanding and we will modify it in the final version. Weakness 2: Limited baselines. Thanks for the comment. SortedNet scales the network through multiple directions and results in irregularities in transformer architectures, compared to Scala which aligns with the inherent design of ViT to vary from widths. As their vision experiments are conducted on small-scale CIFAR datasets, we intended to reproduce their method on ImageNet but did not find released implementations. MatFormer only scales the FFN block in transformer so it offers a minor computational adjustment space and we adapt their method (written in JAX) on DeiT-S for a fair comparison. The results are shown in Fig.1 of the PDF file (reply to all reviewers) and Scala achieves comparable performance with it (better in most cases) when $s=0.5$ with a larger adjustment scope. Weakness 3: Scope of experiments. We thank the Reviewer for the valuable suggestion. Indeed, extending Scala to NLP and a larger network is part of our original plan but we are struggling with computation resources as both of the experiments demand much larger training costs. We will definitely consider it as our future plan. Weakness 4: Teacher and training costs. As explained in Weakness 1, Scala can be regarded as a tool to make existing methods become slimmable while trying to maintain the original representations. Although we can expect to learn similar representations of methods trained with supervised learning on ImageNet-1K (DeiT), some self-supervised learning methods (DINOv2) are trained with gigantic undisclosed data which is almost impossible to reproduce their performance without utilizing the pre-trained model. So we simply use the pre-trained network (same size as the student) as the teacher to imitate the original model to inherit the original representations, instead of utilizing a larger network which usually leads to better performance. For US-Net, we do not introduce teacher networks following their original design but we do compare with Separate Training with Distillation in Sec A.8 (Tab. 10), where Scala still outperforms the baseline at smaller widths. Here we further re-implement US-Net by adding the teacher and the results are shown in the table below. Although obtaining a better full model, US-Net exhibits worse performance at smaller width ratios because it utilizes the full network as the teacher for all subnets and this phenomenon verifies our motivation of proposing Progressive Knowledge Transfer. \| Method \| $r=0.25$ \| $r=0.50$ \| $r=0.75$ \| $r=1.00$ \| \| :-----: \| ------------ \| ------------ \| ------------ \| ------------ \| \| US-Net \| 52.7% \| 60.1% \| 66.9% \| 73.2% \| \| US-Net+KD \| 51.6% \| 58.7% \| 66.6% \| 73.8% \| \| ST \| 45.8% \| 65.1% \| 70.7% \| 75.0% \| \| Scala \| 58.7% \| 68.3% \| 73.3% \| 76.1% \| Assuming to deliver 13 models in the end, we compare the training time (100 Epoch) of Scala with US-Net, Separate Training, and Separate Training with Distillation on 8 A100 GPUs and we do not include the teacher training time as Scala is aimed to scale down existing works where the same size pre-trained model can be easily downloaded. The difference between US-Net and Scala is not large as the transformer architecture has been well-optimized on GPU and we do observe a significant time gap between Scala and ST/STD as they have to train 13 models iteratively. Moreover, Scala can be configured to deliver 25 models without an increase in training time as we sample 4 networks at each iteration in all scenarios which further highlights our strengths. \| Method \| Training Hours \| \| :-----: \| ------------ \| \| ST \| 123 \| \| STD \| 128 \| \| Scala \| 21 \| \| US-Net \| 20 \| Weakness 5: Interpolation. We did not fully address the interpolation problem as this issue is essentially related to the architecture design of transformer and we constrain ourselves from modifying the conventional structure as it has been well-optimized in various scenarios. On the other hand, it further emphasizes that the problem we target to solve is very challenging as the intermediate subnets ($X=13$) are only trained for around 18 epochs and they are expected to match the performance of networks that are trained for 100 epochs. Thanks for carefully reading our paper and we will move Fast Interpolation into the main paper. Weakness 6: DINOv2. We follow the previous setting except for using DINOv2-B as the teacher which means we still train the model in a supervised manner. We have tried to adopt the self-supervised learning objective from DINOv2 but found the training cost too large, and it is almost impossible to reproduce the performance following the original learning objective with ImageNet-1K since DINOv2 is trained on the private dataset LVD-142M. However, we re-conduct the experiment by removing all the CE loss to alleviate dataset bias and find this choice results in surprising generalization ability by inheriting the fruitful knowledge of DINOv2 (please see the reply to all reviewers). Weakness 7: Ablation. Thanks for the advice and we will add more ablation in our final version. * We hope the explanations could address your concerns and we would appreciate it a lot if you could recognize the contributions of our work. --- Rebuttal Comment 1.1: Title: Post-rebuttal comments Comment: I would like to thank the authors for the detailed and informative rebuttal. I imagine they invested significant effort into clarifying the concerns from the 4 reviewers and I'm confident they will improve this work. I've read the other reviews and the responses of the authors to them and skimmed through the paper again. The strong points I see in this rebuttal are: - several clarifications around the implementation, in particular on the use of pre-trained teacher across settings: both on the regular setting but also on the transferability task where DINOv2 was used as teacher and not as a self-supervised objective - the addition of a comparison with a more recent and related work, Matformer. Pity that the code of SortedNet is not available. - more detailed information about training costs w.r.t. to the main baseline US-Net, that originally does not benefit from a pretrained teacher. The authors report results for US-Net + KD in the rebuttal and it still under-performs compared to Scala. - new results using DINOv2 as teacher without cross-entropy loss showing the effectiveness of the option of using foundation models for this setting. On the downside the evaluation is still limited to mid- to small-sized architectures (the biggest architecture used here is a ViT-B with 86M parameters) and to computer vision settings. I can relate to the argument of the authors regarding computational cost. However given their results with DINOv2, one may devise a similar strategy for distilling a NLP foundation model to a different task than next-token prediction, e.g., text-classification. Hopefully the authors will find a solution in this direction for the paper update. Wrap-up. I think the submission has improved over the rebuttal with the new information and results and I encourage the authors to include the new findings in the main paper, and to release implementation code (as they stated in the NeurIPS paper checklist, to help other works to build upon and compare against Scala). At this point I'm rather leaning for a positive recommendation for this work. I do not have other questions. --- Reply to Comment 1.1.1: Title: Thanks for the support Comment: Dear Reviewer, Thank you very much for your recent feedback. We truly appreciate your support and the time you have invested in reviewing our paper. Yes, we will definitely include the new findings in the main paper and we promise to release the code if this work is accepted. We are glad to hear that the reviewer has no concerns and is leaning toward a positive recommendation. If the reviewer finds that the work is improved during the rebuttal, we would be grateful if you could reflect the updated stance in the rating on OpenReview. This would greatly assist us in the final decision-making process. Thank you once again for your thoughtful consideration and valuable insights. Sincerely, Authors	Rebuttal 1: Rebuttal: Dear Reviewers: Thanks for your valuable comments in the review process. We have an exciting experiment added during rebuttal which supports that Scala can effectively inherit the generalization ability from foundation models like DINOv2 while maintaining the flexible inference capability. This indicates that it may be possible to enable foundation models to become slimmable as well which would further enhance the application scenarios of our method. Inspired by the comment of Reviewer rLHb, we build Scala over DeiT-B for ImageNet-1K training and utilize DINOv2-B as the teacher in order to inherit its strong generalization ability. However, we remove all the Cross-Entropy loss during training to alleviate the dataset bias issue as DINOv2 is trained on the gigantic private dataset. To examine the generalization ability of Scala, we conduct linear probing on 12 fine-grained classification datasets following DINOv2. Tab. 1 in the PDF file shows that Scala significantly outperforms DeiT variants on the average performance of fine-grained classification which suggests that Scala indeed inherits the fruitful knowledge from DINOv2 with remarkable improvement in its generalization ability. Moreover, the improvement over DeiT does not decrease when we scale down the width ratios during inference and it indicates that Scala maintains the flexible inference capability very well even though it contains more knowledge than before. We thank the Reviewers for the constructive suggestions which help us to improve the work. We are actively available until the end of this rebuttal period and please let us know if you have any further questions. Thank you so much for being with us so far. Sincerely, Authors Pdf: /pdf/deadc6734d0c04f39e500dd3be71b18bd2d88229.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Bridging Model-Based Optimization and Generative Modeling via Conservative Fine-Tuning of Diffusion Models	Accept (poster)	Summary: This paper presents a conservative fine-tuning method called BRAID, which integrates the strengths of diffusion models and model-based optimization (MBO) to improve the performance of pre-trained diffusion models on offline datasets. BRAID optimizes a conservative reward model that includes penalties outside the offline data distribution to prevent overoptimization and generate valid designs. The approach is validated through empirical and theoretical analyses, demonstrating its ability to outperform the best designs in offline data while avoiding the generation of invalid designs. The paper also discusses the method's effectiveness compared to existing conditional diffusion models and traditional MBO techniques, with experiments showcasing its superiority in biological sequence and image generation. The authors acknowledge the limitations of their study, particularly in model selection and hyperparameter tuning, and suggest future research directions. Strengths: * BRAID incorporates a conservative approach to fine-tuning diffusion models, which includes penalization terms that discourage the model from generating designs outside the distribution of the offline data. This conservative strategy is effective in preventing overoptimization and ensuring the validity of the generated designs. * The method is supported by both theoretical analysis and empirical results. Theoretically, it provides a regret guarantee, ensuring that the fine-tuned models can outperform the best designs in the offline data. Empirically, it has been validated through experiments across various domains, such as biological sequences and images, demonstrating its ability to generate high-quality designs. Weaknesses: * Difficulty in tuning hyperparameters without online data interaction. * Reliance on accurate reward and diffusion models for effective performance. * Theoretical results depend on certain idealized assumptions that may not hold in all cases. * Can you compare the methods with other SOTA offline RL methods to illustrate your proposed augmented methods more effective than the SOTA offline RL methods? I think this paper is very relevant to some offline RL methods, such as ReDS[1], A2PR[2], CPED[3], SCQ[4]. It is not required that experimental comparisons must be given, but at least add some discussion with these methods to the paper. References： [1] Singh, Anikait, et al. "ReDS: offline reinforcement learning with heteroskedastic datasets via support constraints." Proceedings of the 37th International Conference on Neural Information Processing Systems. 2023. [2] Liu, Tenglong, et al. "Adaptive Advantage-Guided Policy Regularization for Offline Reinforcement Learning." In International Conference on Machine Learning (ICML). PMLR, 2024. [3] Zhang, Jing, et al. "Constrained policy optimization with explicit behavior density for offline reinforcement learning." Advances in Neural Information Processing Systems. 2023 [4] Shimizu, Yutaka, et al. "Strategically Conservative Q-Learning." arXiv preprint arXiv:2406.04534 (2024). Technical Quality: 3 Clarity: 4 Questions for Authors: * The methods will need more samples, as shown in the pseudo-code of the algorithms 2 Direct Back Propagation (General case), which may bring more computational burden. Meanwhile, the methods train use two diffusion model to obtain the policy. Can you give me some experiments to show the computational burden? * Have the authors considered alternative generative models such as GANs or VAEs, and can they provide a comparative analysis of performance and resource usage? * How to obtain the $\hat{g}$ in the pseudo-code of Algorithm 1 BRAID? * This methods is related to offline reinforcement learning. So can you give me some experiment comparisions with the SOTA offline RL methods ？ Confidence: 2 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: * The method requires careful selection of hyperparameters, which can be challenging in a purely offline setting without access to additional online data. * The pseudo-code of the paper does not illustates the algorithm explicitly. The authors can improve it further. * This paper lack strong baselines comparisions and enough related works, which is related to some offline reinforcement learning methods. So you can add more related offline reinforcement learning works. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the detailed and insightful feedback. We have addressed the reviewer's concern by clarifying that (1) our goals differ significantly from those in standard offline RL works, (2) we have compared our method with recent works that align with our objectives, such as Yuan et al. (2023) and Krishnamoorthy et al. (2023). W: Comparison with SOTA offline RL methods, IReDS[1], A2PR[2], CPED[3], SCQ[4]. Thanks for raising this point. We would like to emphasize that, in general, our originally intended goals are quite different, though we acknowledge that the ideas in the papers you cited are very interesting and potentially helpful in developing new algorithms for our tasks. 1. The goals of our paper and most offline RL works are different (as briefly mentioned in Appendix A) because our paper aims not to address standard offline RL tasks but our paper aims to generate high-quality designs in extremely high-dimensional action spaces. Given the high dimensionality of these spaces (e.g., images, chemicals, biological sequences), our focus is on incorporating SOTA pre-trained generative models (i.e., diffusion models) to constrain the search space to valid domains (natural image space, natural chemical space, natural biological space). While we appreciate the relevance of the works you cited in offline RL, these papers were not designed to address our specific tasks. It is unclear how their methods, which do not include experimental results in our scenarios (such as generating images, biological sequences, or molecules), can be directly applied to our work. 2. We have compared relevant existing methods in the most pertinent field: Goal-wise, our work aligns more closely with offline (contextual) bandits in extremely high-dimensional action spaces rather than with offline RL. However, the RL aspect appears in our work because diffusion models can be viewed as MDPs, and we have utilized this observation to address our problem as detailed in Section 5. We have compared several recent baselines, such as Krishnamoorthy et al. (2023) and Yuan et al. (2023), that tackle the same problem as us in Section 7. 3. Let me explain whether offline RL works you cited naively address our problem. While the offline contextual bandit problem is an offline RL problem with horizon 1, it is not obvious that the papers you cited are directly applicable to our task because they do not focus on incorporating expressive pre-trained generative models. We acknowledge that some insights from these papers could be relevant to our problem. However, translating these insights into our context may not be straightforward and could require more than simply adapting the original algorithms. - For instance, replacing $\rho$ with diffusion models in Algorithm 1 of ReDS[1] might be valid, but its practical utility is uncertain since the log-likelihood estimation of diffusion models lacks an explicit form and is not differentiable. - Similarly, using diffusion models as policies in Algorithm 1 of A2PR[3] might not be practical for our task as well because it is hard to get the explicit form for the log-likelihood. W. Difficulty in tuning hyperparameters without online data interaction/Reliance on accurate reward and diffusion models for effective performance We agree with the reviewer and have acknowledged our limitations. However, since this limitation is common across many works that use purely offline data (e.g., Rigter et al. (2022); Kidambi et al. (2020)), we have adhered to current conventions in the literature by assuming limited online interaction in actual experimental settings. Q: The methods will need more samples, as shown in the pseudo-code of the algorithm 2. Can you give me some experiments to show the computational burden? To ensure clarity, let us first distinguish between two types of samples: (a) offline data samples ($x, r(x))$) and (b) artificial samples generated by diffusion models. In the offline setting, sample efficiency in terms of offline data (a) (e.g., lab feedback data) is crucial, so the computational burden (how many samples we use in (b)) is less of a concern as long as it is not excessively high. With this in mind, yes, our Algorithm 2 requires artificial samples like (b) above. However, since we are not using offline data, the number of samples used in Algorithm 2 is less concern. We have not discussed the computational complexity as it has already been discussed in existing works in detail (Black et al. (2023) ,Prabhudesai et al., 2023), and our focus is on sample efficiency. The actual computational cost varies by domain; for example, image generation might take several hours, while sequence generation might take several minutes. We will address this info in the next version. Q: Have the authors considered alternative generative models such as GANs or VAEs, and can they provide a comparative analysis of performance and resource usage? This is an interesting point. We will include a more detailed discussion in the Appendix. - Reasons for using diffusion models: In many domains (e.g., images, molecules), diffusion models have demonstrated SOTA performance as generative models, as evidenced by numerous papers (Rombach, Robin, et al; Avdeyev, Pavel, et al). This motivates us to use diffusion models as SOTA generative models to capture the valid space (natural image, chemical, biological spaces). - Advantages and disadvantages of GANs and (pure) VAEs: We acknowledge that GANs and VAEs remain useful as generative models. In particular, fine-tuning to optimize downstream reward functions is computationally faster if we use them. However, as demonstrated in many papers on generative models for images, molecules, the performance as generative models tends to decrease. Q: How to obtain the $g$ in the pseudo-code of Algorithm 1 BRAID? We have discussed it in Section 4.3. We will add a pointer to make it more accessible to readers. --- Rebuttal Comment 1.1: Title: Official Comment by Reviewer HnkA Comment: Thanks for your careful answers and clear explanation. To some extend, my concerns have been addressed. So I expect you to discuss SOTA offline RL as mentioned earlier in your paper and conduct comparison experiments with GANs or VAEs. I would like to keep my score, but it is possible to improve my score.	Summary: The paper tackles the task of black box optimization in an offline setting. Given a pretrained diffusion model, they first train a surrogate model on the offline data and use it to tilt the diffusion model distribution via finetuning it. The authors distinctly focus on an uncertainty quantification based procedure to bias the diffusion model tilting toward regions where the reward is high and the reward uncertainty is low while not tilting toward regions with high uncertainty. Experiments are carried out on a reasonable set of diverse tasks. The paper introduces a small specific challenge and address it well with a reasonable approach and good motivations. The potential impact of the method may be small but it is neat, educational, and should be useful in many cases. I recommend acceptance. Strengths: 1. Identifying a crucial non-obvious overlooked challenge and specifically pinpointing it. The authors identify that previous work on tilting diffusion models with reward models misses to incorporate tilting less toward regions of the reward function in which it has high uncertainty but high reward. Instead, we should only tilt to the regions that have high reward and high certainty to avoid optimizing toward adversarial examples. (The fact that the finetuned diffusion model will steer away from the pretraining distribution seems like a less relevant insight) 2. The authors identify a relevant overlooked problem in finetuning diffusion models and bring standard techniques from uncertainty quantification into the field to address it in a reasonable fashion. They do not overcomplicate things and their technique could be valuable to several researchers in the area. 3. The authors prove that the training procedure yields the desired distribution. I have no comments regarding the value/insightfulness of the proof. Maybe other reviewers have a stronger opinion about the relevance. 4. The authors evaluate their method on a very diverse set of experiments that includes discrete DNA sequence generation, and image generation. The results are convincing and demonstrate the central empirical claim that out of distribution generation is a problem and is effectively avoided with the proposed conservative reward model fine tuning. Minor: 1. Interesting snippets of insights. The authors point out interesting relationships and connections along the way which are non-obvious and well placed for putting their motivations into context. 2. Exceptional clarity in writing. The paper lays out the task in its precise specification and covers required concepts and related work in equal clarity. Weaknesses: 1. I would say that the insights in terms of methodological novelty are on the moderate side. The ideas are simple and good, which is appreciated, but the level on which the conceptural changes operate are low level (a tweak to diffusion model tilting) and thus limited in impact. However, it is certainly a good thing to have. Very hard to address and not a must have for ML conferences: 1. Evaluations are inherently limited in their computational nature and the conclusions that can be drawn for the procedures effectiveness in biological sequence optimization is small. Do you authors disagree with this in any way? Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Do I understand correctly that theorem 2 only states that the finetuned diffusion model will incur lower regret than the pretrained diffusion model? Why is that useful and would we not rather be interested in relations between using your additional uncertainty quantification bias for the reward model based finetuning and not using it? 2. It seems to me that reward models are often very poor in scientific applications (more so than the generative models that can be trained on a lot more data). Does this mean that their uncertainty estimates are also bad and your method might not provide any improvements in these cases? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors point out the key inherent limitation of their work in the fact that the reward models that are often trained on limited data are likely suboptimal instead of just mentioning useless small limitations that are beside the point (which is the more common practice it seems). Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your feedback. We have addressed your concern by explaining a more detailed evaluation plan for biological tasks. Weakness: Evaluations are inherently limited in their computational nature and the conclusions that can be drawn for the procedures effectiveness in biological sequence optimization is small. Do you authors disagree with this in any way? We thank the reviewer for pointing out that our computational evaluation did not provide enough confidence in our generated biological sequences. While these are complex systems, several of the main factors contributing to the activity of these biological sequences are well-studied and understood. Therefore, there are several analyses we can perform to gain deeper confidence. * Of the two functions we optimize, enhancer activity in HepG2 is controlled to a large extent by the binding of transcription factors to the DNA sequence, particularly HepG2 specific transcription factors such as HNF4A, HNF4G, HNF1A and HNF1B. 5’UTR activity is, to a large extent, dependent on the sequence of the translation initiation site. Features such as A or G at position −3 relative to AUG and a G at +4 are understood to increase sequence activity, as well as the absence of upstream start codons. These features are understood from many biology studies, including those on the original datasets (Sample et al. 2019, Gosai et al. 2023) * For the generated HepG2 sequences, we will use the JASPAR database (https://jaspar.elixir.no/) to scan the generated sequences for motifs known to bind to human transcription factors. We expect to see a higher frequency of motifs for activating transcription factors, particularly HepG2- specific ones, and a reduced frequency of motifs known to bind to transcriptional repressors. For the 5’UTR sequences, we will examine the sequences for upstream start codons and also examine base frequencies in the translation initiation sites. This analysis will clarify whether our model has learned well-established features of biological sequence activity and allow us to draw stronger conclusions regarding the validity of the generated sequences. Questions: Do I understand correctly that theorem 2 only states that the finetuned diffusion model will incur lower regret than the pretrained diffusion model? Why is that useful and would we not rather be interested in relations between using your additional uncertainty quantification bias for the reward model based finetuning and not using it? Let me clarify as follows. * Roughly speaking, our theorem (Theorem 2 and Corollary 1) states that our fine-tuned model performs better in terms of $J_{\alpha}$ than the best design in the reward data distribution. Since $J_{\alpha}$ includes a metric related to proximity to pre-trained models, this theorem conveys our key message: 'fine-tuned generative models outperform the best designs in the offline data by leveraging the extrapolation capabilities of reward models while avoiding the generation of invalid designs’. * The uncertainty quantification term is used to construct $\hat p_{\alpha}$. When evaluating the actual performance in Theorem 2, we have used the true reward $r$. Q: It seems to me that reward models are often very poor in scientific applications (more so than the generative models that can be trained on a lot more data). Does this mean that their uncertainty estimates are also bad and your method might not provide any improvements in these cases? Yes. While we did not intend to obscure this information, as it is briefly mentioned in the Limitations section, we will make it more explicit. --- Rebuttal Comment 1.1: Comment: 1. I appreciate the additional suggestions for experiments and am sure they could be interesting as well. I did not state that "[your] computational evaluation did not provide enough confidence". I think the computational evaluations are well crafted and can provide a fine degree of confidence. I just pointed this out as a minor weakness prepended with "not a must have" (which was maybe overread in the formatting).\ Anyways, the additional promised experiments and the qualitative recovery of experimental observations do not change my score. I think it is a good paper and should be accepted. It is outstanding in its clarity of writing (thanks for that effort) but not outstanding overall in my current estimation (I hope that is not rude to say :)). 2. Thanks for taking the time to answer - that was a confused question. 3. "Yes" -> Sad, but expected.	Summary: This paper proposes a conservative approach for fine-tuning diffusion models with a reward model learned from offline data. Specifically, the ideas are two-fold: The first idea is to replace the reward model with a conservative estimate based on classical generalization bounds. The second idea is to leverage the KL divergence to force the optimized distribution to not deviate too far from the pretrained model. Experiments and theoretical results show the efficacy of the proposed method in fine-tuning diffusion models without over-optimization. Strengths: 1. The proposed method is well-presented, and the motivation behind the algorithm is interesting. The over-optimization problem is indeed critical when fine-tuning diffusion models with learned rewards. 2. Extensive experimental results show the efficacy of the proposed method in improving the reward model while avoiding reward over-optimization. Weaknesses: 1. Leveraging generalization bounds via kernel RKHS and bootstrap is interesting, but I doubt their practicality for real applications. Firstly, the RKHS bound is usually too conservative to be useful, while the computational cost for the bootstrap method is pretty high since one has to train the model from scratch multiple times. As far as I can tell, the reward models used in the experiments are mainly single-layer MLPs, and it is doubtful whether this approach is useful when the reward model needs to be a larger model. 2. Another problem with the conservative estimator of the reward models is that it is unclear whether it is useful given the current experimental results. On one hand, KL regularization is a widely-known technique for preventing over-optimization in diffusion models and is thoroughly studied in existing works, so it is certain that the KL regularization term will help. On the other hand, the proposed algorithm mixes both the conservative reward estimator and the KL regularization term together, making it unclear which part is playing the role in avoiding over-optimization. My guess is that, for the most part, only the KL regularization term is effective in the end. Technical Quality: 2 Clarity: 2 Questions for Authors: STRL methods like AlignProp and DRaFT can work with ODE samplers, which are more commonly used in practice than SDE samplers. However, the method proposed in this work, due to the use of entropy regularization, can only adopt SDE samplers. I wonder if it is possible to design a regularization term for ODE samplers. Could the authors share some insights on this point? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: See my comments on the weakness above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your feedback. We have addressed your concern by explaining (1) the bootstrap/RKHS bonus term's practical usefulness in our scenario and its widespread use in various real applications/papers, and (2) how our experiments are intentionally designed to demonstrate that conservatism (rather than KL) helps mitigate overoptimization. Please let us know if this addresses your concerns, or if there are any other issues we should discuss or address! Weakness: RKHS bound is usually too conservative to be useful. I doubt their practicality for real applications. We would like to convey that it is still practically useful from several aspects. 1. Negative penalty term (Example 1) based on RKHS theory is widely used in the literature on Gaussian processes (GPs). For example, as noted in Wikipedia of GPs, many widely used software tools are built around GPs. They have been proven to be effective in real-world applications, e.g., robotics/material science/drug discovery (Deringer et al., Gramacy et al.). 2. Our examples with DNA and RNA designs are real applications, as the tasks, data, and models are all proposed by established scientific journals, as we have cited Sample et al., 19, Inoue et.al, 19. There is more literature on solving this task, e.g., Taskiran et al. 3. We believe there may be a misunderstanding: We have combined it with deep learning models. After extracting features from deep learning models, we used them as a negative penalty term to avoid overoptimization in Example 1. We did not use a pure RKHS model with a fixed feature as our reward model. This approach is used by many famous RL papers, such as Yu et al. 19. 4. While we acknowledge that the generalization bound in Theorem 2, derived under the RKHS assumption, might not be tight, Our goal is to propose an algorithm that offers both superior experimental performances and fundamental guarantees that formalize the algorithm’s intuition. We believe that our upper bound suffices for this purpose. Specifically, our intuitive message is that fine-tuned generative models outperform the best designs in offline data by leveraging the extrapolation capabilities of reward models while avoiding the generation of invalid designs. Corollary 1 formalizes this intuition, as mentioned in Section 6. 5. Statistical guarantees in most of the literature are not completely tight, as proving the lower bound is extremely challenging in many scenarios (Wainwright, 19). Therefore, the claim mentioned by the reviewer applies not only to our work but also to nearly all upper bounds in the literature in RKHS/GPs. We think showing tightness is beyond our scope. Weakness: the computational cost for the bootstrap method is high since one has to train the model from scratch multiple times. The reward models used in the experiments are mainly single-layer MLPs, and it is doubtful whether this approach is useful when the reward model needs to be a larger model. We acknowledge the reviewer’s point; however, we would like to emphasize that (1) computational efficiency is not the primary focus of our paper, (2) it can be addressed straightforwardly by using variants if we want to further accelerate bootstrap part, and (3) the computational cost of bootstrap is not necessarily high in many real applications. 1. In scenarios with limited offline data, sample efficiency (the cost of obtaining data with reward feedback) is more critical than computational efficiency. Our work always prioritizes the former, which differs from online settings where computational efficiency may be more relevant. We have motivated such emphasis in the Introduction by clarifying the hardness of wet lab data collection in scientific domains. 2. To increase speed, we can utilize various bootstrap variants (Lakshminarayanan et al. Al, Osband et al., Chua et al.) without training from scratch, such as sharing backbone models, Bayesian deep learning way, etc. In this way, the bootstrap has been practically and widely used in the ML community. We will add such an explanation. 3. The computational cost of running pure bootstrap in our experiments is also not so high in many real applications. Indeed, while the reward models in Section 7.1 are not single-layer MLPs (they include transformers and are representative models in genomics proposed in Avsec et.al 21), the training takes several hours. In scenarios where sample efficiency is crucial, training multiple times (or in a parallel manner) won’t be a significant concern. In fact, for both tasks we only employ 3,4 bootstrapped models, so it does not bring too much computational concerns. It is unclear whether it is useful given the current experimental results. On one hand, KL regularization is a widely-known technique for preventing over-optimization in diffusion models so it is certain that the KL regularization term will help. On the other hand, the proposed algorithm mixes both the conservative reward estimator and the KL regularization term together, making it unclear which part is playing the role in avoiding over-optimization. My guess is that, for the most part, only the KL regularization term is effective in the end. We believe there may be a misunderstanding regarding the interpretation of our experimental results. In our experiments, we have compared our proposal with the baseline that has the KL regularization, as noted in Section 7 ( Lines 177-185 in our paper), but do not have conservatism in the reward model. For example, in images, STRL and our proposals BRAID-Boot, BRAID-Bonus) share exactly the same KL strength, and its value ($\alpha=1$) can be found in Table 5, Appendix F.2.4. Therefore, our current experiments have directly addressed your concern by showing the effectiveness of conservative reward modeling. Q: Can be adapted to ODE samplers? We address this in the global rebuttal. --- Rebuttal Comment 1.1: Title: Thanks for your rebuttal Comment: After careful consideration, I decided to raise my score to 5.	Summary: 1) This paper analyzed the two mainstream angles of computational design. 2) Proposed a hybrid one that offline fine-tunes generative models. 3) Conduct experiments on two tasks to show the performance of their method. Strengths: 1) Sufficient theoretical analysis and detailed preliminaries. 2) The idea is straightforward. 3) The method is comprehensive. Weaknesses: 1) In the introduction, the advantages and disadvantages of the two mainstream issues are not fully analyzed. 2) Insufficient metrics evaluation for image generation task. Technical Quality: 3 Clarity: 2 Questions for Authors: N/A Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1) Lack of gradual guidance from analyzing the advantages and disadvantages of existing methods to proposing the hybrid method. 2) Lack of multi-metric quantitative results analysis. (e.g. in the image generation task, the fidelity and diversity should also be reported by like LPIPS Score/CLIP score/FID/IS, etc.) Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive feedback! We have addressed your concerns by providing additional explanations on (1) the disadvantages and advantages of pure generative model/MBO approaches, and (2) experimental results on diversity metrics based on CLIP scores. Weaknesses: In the introduction, the advantages and disadvantages of the two mainstream issues are not fully analyzed. Limitations: Lack of gradual guidance from analyzing the advantages and disadvantages of existing methods to proposing the hybrid method. We have tried explaining well; however, due to space constraints, it may come across as somewhat terse. Assuming we have more space in the final version, we will elaborate further in the introduction as follows. 1. Pure generative approach: We primarily refer to conditional diffusion models. - Disadvantages: It may be unclear whether we can learn designs with properties superior to those in the offline data. We will explicitly connect the introduction with Section 2. This connection will clarify that we indeed compare our approach with conditional diffusion models as a pure generative approach in Section 7. - Advantages: We can generate valid designs that lie within the sample space (e.g., natural image space, natural DNA, RNA, molecular space), which is challenging to achieve with a pure MBO approach. 2. Pure MBO approach - Disadvantages: We will add the following example in the introduction. It can be challenging to incorporate the natural constraints of the valid space. For example, when the design space is an image space and the reward function is compressibility, optimizing the just learned reward model may result in highly non-natural images with high compressibility. However, our typical goal is to obtain a natural image with high compressibility. To achieve this, it is essential to incorporate a (pre-trained) generative model that characterizes the natural image space. - Advantages: Our approach learns a reward model, granting us the extrpolation capabilities so that we can generate designs beyond the offline data distribution by leveraging the extrapolation capabilities of the reward model. Weakness: Lack of multi-metric quantitative results analysis. (e.g. in the image generation task, the fidelity and diversity should also be reported by like LPIPS Score/CLIP score/FID/IS, etc.) Thank you for the suggestion. As the primary goal of the experiments is to demonstrate that our method can mitigate overoptimization, we focus on reporting the performance metrics needed to show this in Section 7.2. However, we recognize that general synthesizing metrics are frequently reported in papers on diffusion models (though they may not be directly helpful in measuring overoptimization). The following is our attempt at a similarity score. Our attempt: While achieving diversity is not our stated promise, we experimented with CLIP cosine similarities for completeness to measure the diversity of generated samples. However, we observed that CLIP similarities are influenced largely by semantic information. For example, two images generated with the same prompt (e.g., cat) are likely to have a high CLIP cosine similarity (>0.8), even if they are visually very different. We agree that LPIPS might be more reliable.	Rebuttal 1: Rebuttal: We appreciate feedback from all reviewers. We respond to raised weaknesses/questions as much as possible. Papers we cite: We have added the papers we cited in our response here. * Deringer, V. L., Bartók, A. P., Bernstein, N., Wilkins, D. M., Ceriotti, M., & Csányi, G. (2021). Gaussian process regression for materials and molecules. Chemical Reviews, 121(16), 10073-10141. * Gramacy, Robert B. Surrogates: Gaussian process modeling, design, and optimization for the applied sciences. Chapman and Hall/CRC, 2020. * Taskiran, Ibrahim I., et al. "Cell-type-directed design of synthetic enhancers." Nature 626.7997 (2024): 212-220. * Yu, Tianhe, et al. "Mopo: Model-based offline policy optimization." Advances in Neural Information Processing Systems 33 (2020): 14129-14142. * Wainwright, Martin J. High-dimensional statistics: A non-asymptotic viewpoint. Vol. 48. Cambridge university press, 2019. * Chen, Ricky TQ, et al. "Neural ordinary differential equations." Advances in neural information processing systems 31 (2018). * Lakshminarayanan, Balaji, Alexander Pritzel, and Charles Blundell. "Simple and scalable predictive uncertainty estimation using deep ensembles." Advances in neural information processing systems 30 (2017). * Chua, Kurtland, et al. "Deep reinforcement learning in a handful of trials using probabilistic dynamics models." Advances in neural information processing systems 31 (2018). * Osband, Ian, et al. "Deep exploration via randomized value functions." Journal of Machine Learning Research 20.124 (2019): 1-62. * Rombach, Robin, et al. "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022. * Avdeyev, Pavel, et al. "Dirichlet diffusion score model for biological sequence generation." International Conference on Machine Learning. PMLR, 2023. * Xu, Minkai, et al. "Geodiff: A geometric diffusion model for molecular conformation generation." International Conference on Learning Representations. 2022 Additional rebuttal for a reviewer dLuW That’s an interesting point. Here are several ideas. We will incorporate the following into the main text: - While somewhat heuristic, we can still use the ODE sampler after fine-tuning. Specifically, when generating samples for fine-tuning, we use the SDE sampler with a small variance term; however, after fine-tuning, we can still use it as the ODE sampler by setting the variance term to 0. - Another idea is to explicitly calculate (or estimate) the log-likelihood ($log p_{\text{pre}}$) using the well-known formula from the neural ODE paper (Theorem 1 in Chen et.al) and incorporate it as a penalty in the reward term. However, the caveat of this approach is that estimating this log-likelihood is often challenging and non-differentiable.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Scene Graph Disentanglement and Composition for Generalizable Complex Image Generation	Accept (spotlight)	Summary: This paper employs scene graph for image generation. Different from the previous methods, they employ the generative capabilities of variational autoencoders and diffusion models in a generalizable manner, compositing diverse disentangled visual clues from scene graphs. The authors propose a semantics-Layout Variational AutoEncoder to jointly derive layouts and semantics from scene graph. Then they develop CMA integrated with a diffusion model. They also introduce the multi-layered sampler for achieving graph manipulation. Experiments show that the method outperforms existing methods. Strengths: 1. The paper address the problems in existing methods well. Existing methods in the field of scene graph to image generation mainly depends on the layout or semantics. Using one of them may cause some problems. Inspired by these phenomenons, the authors propose the method to jointly considering the layout and the semantics. What's more, the techniques used in the framework are novel enough. 2. The authors conduct plenty of experiments. The ablation studies support the motivations. Weaknesses: The only weakness I found is that the authors should reorganize the paper carefully. The writings is not so clear in some sections. For example, the multi-layered sampler section is too abstract to be understood. Technical Quality: 4 Clarity: 3 Questions for Authors: No. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive comments and valuable feedback on our work! We are excited and encouraged by your support! Below we address your concern. Q1: About the clarification of the multi-layered sampler section. R1: We sincerely appreciate the valuable suggestion, and we will make our method more understandable in the revision. For the multi-layered sampler, we define each object as a layer, thus allowing independent object-level Gaussian sampling. Besides, we leverage diverse scene conditions ($N_l=5$) disentangled by SL-VAE for locally conditioned diffusion and aggregate them into the latent code for each object. In summary, it can be simply understood as calculating a more robust layered representation for each object in the scene, which achieves an “isolated” image editing effect. --- Rebuttal Comment 1.1: Comment: After reading the responses and comments from other reviewers, most concerns have been addressed. I am especially interested of the anti-logical relationships generation. I keep my rating considering that this paper deserves being accepted. I also hope the authors could release the codes as soon as possible.	Summary: The paper proposes DisCo (Disentangled Compositional image generation), which integrates both layout and semantic information derived from scene graphs to improve the quality and controllability of generated images. In particular, DisCo has three main components: Semantics-Layout Variational AutoEncoder (SL-VAE) for disentangling spatial layouts and semantics, Compositional Masked Attention (CMA) for fine-grained attribute guidance, and Multi-Layered Sampler (MLS) for object-level graph manipulation. Extensive experiments demonstrate that DisCo outperforms state-of-the-art methods in generating rational and controllable images from text, layout, and scene graph conditions. Strengths: 1. The motivation is clear. The idea of disentangling layout and semantics from scene graphs is novel. 2. DisCo outperforms recent methods in both fidelity and diversity of image generation, as evidenced by the IS and FID scores. Overall, it enhances the generation diversity and controllability. 3. Extensive experiments and ablation studies have demonstrated the effectiveness and the contribution of each component. Weaknesses: 1. The increased inference cost of DisCo (Table 7). In particular, CMA mechanism might increase the computational cost, which may limit the method's scalability and efficiency, especially for large-scale applications. Moreover, since diffusion models are already quite large, the additional AutoEncoders (Lines 129-130) may result in more parameter and memory overhead. 2. DisCo requires expensive training, e.g. 4 A100 GPUs, as indicated in Lines 202-203. With more models releasing recently, this technique might be not scalable. 3. The image quality looks better with this proposed method. However, as metrics today cannot always reflect the real image quality, it would be more convincing to conduct a user study, e.g. votes, to quantify the advantage of DisCo compared to previous works. Technical Quality: 3 Clarity: 3 Questions for Authors: See the weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, authors have stated their limitations on attribute leakage of overlapping in Section A.6, as well as a short discussion on the broader impact in Section A.7. Beside that, authors are also encouraged to add a few discussions on the efficiency of their proposed pipeline. Even though these overheads are inevitable (they are quite common in most researches), a clearer trade-off between the improved image quality and the increased model complexity would help to better assess the value of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate all of your valuable suggestions, which play a pivotal role in enhancing the quality of our paper. Bellow we address all your concerns. Q1: Discussion about the complexity and quality. R1: Thanks for your valuable suggestions. To comprehensively evaluate the complexity and efficiency of our DisCo, we here provide a clearer trade-off between the improved image quality and the increased model complexity to help readers better assess the value of our work. We evaluate the image quality using the T2I-CompBench [35] and measure model complexity using Floating Point Operations (FLOPs) and the number of parameters (Params). The results are shown in the following table. \|\|\|\|\|\|\| \|--------------\|:----:\|:----:\|:----:\|:----:\|:----:\| \|Method\|UniDet\|B-VQA\|3-in-1\|FLOPs (G)\|Params (M)\| \|SD-v1.5\|0.1246\|0.3765\|0.3080\|677.5\|859.4\|37.9\|0.3080\|677.5\|859.4\| \|DisCo\|0.2376\|0.6959\|0.4143\|732.9\|894.6\| \|(ours)\|(+85.9%)\|(+84.8%)\|(+34.5%)\|(+8.2%)\|(+4.1%)\| As we can see that our proposed DisCo achieves significant image quality improvements with a tolerable increase in computational cost. Besides, for the additional AutoEncoders (Lines 129-130) you mentioned, we rectify that this design would not bring more parameter and memory overhead, just negligible increase: \|\|\|\| \|--------------\|:----:\|:----:\| \|Method\|FLOPs (G)\|Params (M)\| \|DisCo (w/o SL-VAE)\|724.1\|875.8\| \|DisCo (w/ SL-VAE)\|732.9 (+1.2%)\|894.6 (+2.1%)\| Q2: About training cost and scalability. R2: We present the training cost between different methods in the following table. We use the A100-80G GPU hours as the metric. \|\|\|\|\|\| \|:----:\|:----:\|:----:\|:----:\|:----:\| \|Method\|GLIGEN\|MIGC\|R3CD\|DisCo (ours)\| \|GPU hours\|~120\|~300\|~180\|~90\| It can be found that that our proposed DisCo has lower training cost compared to other methods. Actually, our DisCo can be treated as a plug-in controller for other models, once we release the weights of SL-VAE and CMA upon paper acceptance, users could directly fine-tune models on GPUs like 3090-24G (no need training from scratch) to make our DisCo scalable to more scenarios. Q3: The metrics to reflect the real image quality, and the required user study. R3: We sincerely appreciate the valuable suggestion. Actually, we evaluate our method on T2I-CompBench, where the metrics (i.e., the spatial/non-spatial relationships, attributes, and complex scenes) are validated to be consistent with human assessments [35]. However, we firmly believe that the user study you mentioned is more convincing. Therefore, we conduct a user study by recruiting 50 participants from Amazon Mechanical Turk. We randomly select 8 prompts for each method, resulting in 80 generated images. We ask participants to score each generated image independently based on the image-prompt alignment. The worker can choose a score from {1, 2, 3, 4, 5} and we normalize the scores by dividing them by 5. We then compute the average score across all images and all workers. The results of user study are presented in the table below. \|\|\|\|\|\|\|\|\|\|\|\| \|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\| \|Method\|SD-XL\|DALL$\cdot$E 3\|Imagen 2\|GLIGEN\|LayoutDiffusion\|MIGC\|SG2Im\|SGDiff\|R3CD\|DisCo (ours)\| \|Alignment Score\|0.6684\|0.5944\|0.5637\|0.6549\|0.6200\|0.7055\|0.3783\|0.4717\|0.6928\|0.8533\| --- Rebuttal Comment 1.1: Title: Official Comment from Authors Comment: Thanks for the authors' rebuttal. They have addressed my concerns. Thus, I would like to raise my score to 6.	Summary: This paper presents "DisCo," a novel framework for generating complex images from structured scene graphs. Unlike traditional text-to-image or layout-to-image methods, DisCo utilizes a Semantics-Layout Variational AutoEncoder (SL-VAE) to disentangle and generate diverse spatial layouts and interactive semantics from scene graphs. It incorporates these elements using a Compositional Masked Attention (CMA) mechanism within a diffusion model, enhancing generation control and rationality. The framework also introduces a Multi-Layered Sampler (MLS) for flexible, graph-based image editing, preserving visual consistency while manipulating object attributes and positions. Strengths: 1. Introduces innovative methods for disentangling and integrating spatial and semantic information from scene graphs, which is a novel approach in image generation. 2. Offers significant improvements in image generation from complex scene graphs, enhancing both the fidelity and controllability of generated image Weaknesses: 1. The paper lacks quantitative comparisons with closely related baselines, such as R3CD, which could provide a more comprehensive evaluation of the model's performance. Inclusion of these comparisons could help validate the proposed advantages of DisCo over existing methods, particularly in handling complex scene graph-to-image generation tasks. 2. Some generated images, particularly those highlighted in Figure 10, exhibit unnatural aspect ratios and stretched elements, suggesting issues with the model’s handling of object proportions and spatial embeddings. 3. It would be great to discuss the scalability aspects, particularly how the proposed model handles graph sizes that exceed typical training configurations. 4. how the model performs with imperfect or noisy scene graphs, which are common in automatically extracted data. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The paper presents results on standard benchmarks. However, can you provide insights or preliminary results on how the model performs across datasets with higher variability in object complexity and scene density? 2. Why were certain closely related baselines omitted from quantitative comparisons? Could inclusion of these baselines provide a more comprehensive evaluation? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors address the challenge of attribute leakage in overlapping objects, which affects image fidelity in scenes with dense object interactions. While mitigation strategies are discussed via the CMA mechanism, further refinement is required to eliminate this issue completely. Additional exploration into the computational efficiency and scalability of the proposed methods would also benefit the paper, providing a more comprehensive view of their practical applications and limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the affirmation from the reviewer for our work. It serves as a strong motivation for us! Bellow we address your concerns sequentially. Q1: More quantitative comparisons with related baselines, such as R3CD. R1: Actually, in Table 1 of the manuscript, we have already provided the quantitative comparisons with the related SG2I baselines, including diffusion-based (e.g., R3CD) and GAN-based methods (e.g., SG2Im). We can see that our DisCo achieves the best in both IS and FID scores. Besides, we conduct a user study by recruiting 50 participants from Amazon Mechanical Turk. We randomly select 8 prompts for each method, and ask participants to score each generated image independently based on the image-prompt alignment. The worker can choose a score from {1, 2, 3, 4, 5} and we normalize the scores by dividing them by 5. We then compute the average score across all images and all workers. The results of user study are presented in the table below. \|\|\|\|\|\|\| \|:----:\|:----:\|:----:\|:----:\|:----:\|:----:\| \|Method\|SG2Im\|SGDiff\|SceneGenie\|R3CD\|DisCo (ours)\| \|Alignment Score\|0.3783\|0.4717\|0.5032\|0.6928\|0.8533\| Q2: Unnatural aspect ratios issue in Figure 10. R2: Thank you for the questions raised by the reviewer. Actually, this is due to typographical error of the manuscript, and we will fix it in revison. Nevertheless, we can still easily see from Figure 10 that our DisCo has a superior ability in modeling non-spatial relationships, such as human "holding" and "looking at" in Figure 10. Q3: Scalability aspect of graph size, object complexity and scene density. R3: Great point! We sincerely appreciate the valuable suggestion, and we will discuss the scalability aspects in the revision. Actually, during the training of our SL-VAE, the way we used that forms a batch of graphs with different node quantities is to merge them into one larger graph, where each subgraph is not connected to each other (similar to independent nodes). Therefore, the GCN-based encoder can effectively learn relationships independent of the number of nodes, and thus handle graph sizes that exceed typical training configurations. Besides, we also provide the generation results with higher variability in object complexity and scene density. Please refer to Figure A1 of the newly added PDF file for more information, which demonstrates that our proposed model could handle larger graph sizes well. Q4: How the model performs with imperfect or noisy scene graphs? R4: Thank you for the questions raised by the reviewer. Actually, our method performs robust enough and is not sensitive to the imperfect or noisy scene graphs. For example, we consider two cases of imperfect or noise scene graphs, i.e., the anti-logical and the missing relationships. - Anti-logical relationship. There may occur anti-logical relationship in the provided scene graph, such as “a dog on the sky” and “a sheep on top of a tree”. As stated in the manuscript, our DisCo disentangles spatial relationships and interactive semantics. The layouts that represent the spatial relationships can output relative positions that follow the description even with an anti-logical input. Besides, the embeddings that represent the interactive semantics influence visual semantics with the given anti-logical relationship while ensuring semantic correctness. Thus, our method could still generate reasonable outputs matching the descriptions even if they are anti-logical. We present the anti-logical generation sample in Figure A2 the newly added PDF file. - Missing relationships. In case of node relationship missing (as we discussed in Figure 6 (c) of the manuscript), we deliberately define a special node "__#image#__" and a special relationship "__#in#__", which represent the whole scene and the relationship of objects to the scene, respectively. All regular nodes are connected to the "__#image#__" node with the "__#in#__" edge. Thus, independent nodes (i.e., missing relationship) could still interact with other nodes in the same scene, and infer their placement based on mutual semantics. In addition, our explicit spatial layout further ensures that there are no node missing problems. We present the results of the missing relationships in Figure A3 of the newly added PDF file. We will add these details in revision. --- Rebuttal 2: Comment: Thank you for your continued efforts. We have provided comprehensive rebuttals and tried to address the concerns raised in your reviews. Please take the time to review if possible, if you have any further questions or require additional clarification, please let us know and we welcome discussions in any format. Thanks again.	Summary: This paper proposes a method that uses a scene graph and integrates variational autoencoders (VAEs) and diffusion models to address complex scene generation. Specifically, a Semantics-Layout Variational AutoEncoder (SL-VAE) is used to derive diverse layouts and semantics from the scene graph, while a Compositional Masked Attention (CMA) combined with a diffusion model incorporates these elements as guidance. Additionally, a Multi-Layered Sampler (MLS) is introduced for isolated image editing. Experiments show that this approach outperforms recent methods in generation rationality and controllability. Strengths: 1. This paper considers an important issue in text-to-image generation realm. 2. The structure design in Section 3 makes sense. 3. The experimental results shown in Table 1 and 2 show the effectiveness of this method. Weaknesses: 1. My main concern is the practical application of this method. As we all known, scene graph building is not a trivial task, but you don't explain detail in the paper how to construct an exact scene graph. In addition, during inference, the prompt proposed by uses may be non-standard so that building a scene graph may be more difficult. 2. Besides, recent SOTA models, e.g. DALLE3, stable diffusion 3 try to solve the complex generation task by large-scale fine-grained dataset construction. How do you compare your methods with these data-centric methods. The authors should spend more space discussing the issues. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the weakness Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Please see the weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate all of your valuable suggestions, which play a pivotal role in enhancing the quality of our paper. Below we address all your concerns. Q1: Details of scene graph construction. R1: Thank you for the questions raised by the reviewer, and we will add the details of scene graph construction from text in the revision. Actually, our DisCo focuses on scene-graph-to-image (SG2I) task, which typically assumes that the scene graph is ready [14, 17]. However, to make our method more practical and allow comparison with the popular text-to-image (T2I) methods, we also provide a text-to-scene graph conversion method (it is not the focus of our paper, so we didn't add it in the original manuscript), which is feasible and convenient with the assistance of the recent powerful LLM such as GPT-4o or Llama. Specifically, as mentioned in Section 2.2, we use LLM to extract instances (subject and object) and relationships (predicates) from the prompt. Given a text describing the scene, the LLM outputs a structured format that contains a node list and a triple list. To standardize the output format of the LLM and facilitate inferring scene graph from non-standard prompt, we here provide a simplified template of our instructions and in-context examples for scene graph building: ```1. Task instruction``` ```(pre-provided)``` You are a master of composition who excels at extracting key objects and their relationships from input text. You should reorganize the input text as a scene graph format, which consist of a node list representing the objects and a triple list representing the relationships between the objects. ```2. Scene graph construction tutorial``` ```(pre-provided)``` There are a few rules to follow: # Output format specification # Numbering rules of the same category # Extract relationships from non-standard text based on context ```3. In-context examples``` ```(pre-provided)``` Prompt Example: Two men standing on the beach, both of them playing a kite. Output Example: Objects: [man1, man2, beach, kite] Relationships: [(man1, standing on, beach), (man2, standing on, beach), (man1, playing, kite), (man2, playing, kite)] More examples using non-standard prompt ```4. Trigger reasoning ability of LLM``` ```(user-provided)``` User Prompt A sheep by another sheep on the grass with the ocean under the sky; the ocean by a tree; a boat on the grass Output: Objects: [sheep1, sheep2, grass, ocean, sky, tree, boat] Relationships: [(sheep1, by, sheep2), (sheep1, on, grass), (sheep2, on, grass), (grass, with, ocean), (ocean, under, sky), (ocean, by, tree), (boat, on, grass)] Note that such scene graph built by LLM only applies to the inference phase for practicality. For training, we use specialized scene graph-to-image datasets (Visual Genome [27] and COCO-Stuff [26]) where the required information like box labeling is all well covered. Q2: Comparison with recent SOTA data-centric text-to-image (T2I) methods. R2: Actually, we have discussed and compared with these methods in Figures 1 (a) and 6 (a) of the manuscript. These data-centric text-to-image (T2I) methods like DALLE$\cdot$3 typically benefit from large-scale text-image pairs that can be automatically collected in high quality. However, even with large-scale fine-grained text annotation, they still suffer from deficiencies in aspects such as quantity generation and relationship binding due to the linear structure of the text. In summary, we point out the advantages of our structured scene graphs over these data-centric text-only methods in representing complex scenes from the following aspects: - Based on scene graph data, which is more compact and efficient than text, our DisCo demonstrates advantages in generation rationality and controllability, particularly for quantity generation and relationship binding. The qualitative comparisons with the T2I models are shown in Figures 1 (a) and 6 (a) of the manuscript. - We also conduct evaluation with these T2I methods on T2I-CompBench [35], which quantifies our improvement in spatial/non-spatial relationships, attributes, and complex scenes. The results are shown in the table below (bold for 1st, italic for 2nd). \| Method\|UniDet\|CLIP\|B-VQA\|3-in-1\| \|--------------\|:----:\|:----:\|:----:\|:----:\| \| SD-v1.4\|0.1246\|0.3079\|0.3765\|0.3080\| \| SD-v2\| 0.1342 \|0.3127\|0.5065\|0.3386\| \| SD-XL\| 0.2032 \|0.3110\|0.6369\|0.4091\| \| Pixart-$\alpha$\|0.2082\|0.3179\|0.6886\|0.4117\| \| DALL$\cdot$E 2\|0.1283\|0.3043\|0.5750\|0.3696\| \| DALL$\cdot$E 3\|0.2265\|0.3003\|0.7785\|0.3773\| \| DisCo (ours)\|0.2376\|0.3217 \|0.6959\|0.4143\| - Besides, we conduct a user study by recruiting 50 participants from Amazon Mechanical Turk. We randomly select 8 prompts for each method, and ask participants to score each generated image independently based on the image-prompt alignment. The worker can choose a score from {1, 2, 3, 4, 5} and we normalize the scores by dividing them by 5. We then compute the average score across all images and all workers. The results of user study are presented in the table below, and we can see our DisCo achieves the best alignment score. \|\|\|\|\|\| \|:----:\|:----:\|:----:\|:----:\|:----:\| \|Method\|SD-XL\|DALL$\cdot$E 3\|Imagen 2\|DisCo (ours)\| \|Alignment Score\|0.6684\|0.5944\|0.5637\|0.8533\| - Moreover, the SL-VAE designed in our DisCo disentangles diverse conditions from the scene graph, which allows multi-round “separate object-level manipulation while keeping the other content unchanged” editing effect as a by-product that is more practical than the general T2I models. The generalizable generation results are illustrated in Figures 5 and 8 of the manuscript. --- Rebuttal Comment 1.1: Title: Thanks for your kind explanation Comment: After thoroughly reading your rebuttal and other reviewers' comments, I will raise my score to 5	Rebuttal 1: Rebuttal: Dear reviewers, We thank all reviewers for their time and efforts in reviewing our paper. These constructive reviews can bring multiple improvements to our manuscript. We are encouraged that the reviewers appreciate our method, including: - structure design that makes sense [Reviewer 3rPJ] - innovative method [Reviewer anRy, SBq5, Knts] - clear motivation [Reviewer SBq5, Knts] - outperform prior methods in rationality and controllability [Reviewer 3rPJ, anRy, SBq5] We have also made diligent efforts to address all the concerns raised point by point. Please see separate responses for details. We are open to discussions and addressing any issues from reviewers. Your constructive comments can further help us to improve our method. Sincerely yours, Authors Pdf: /pdf/e2fe1106f29327f1a19e02894792301d87909b78.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Direct Language Model Alignment from Online AI Feedback	Reject	Summary: This paper propose OAIF, an online method to align language model with human preference where feedback from language models serve as a surrogate of human feedback. The key of OAIF is to use online generated preference pair along the training process. Experiment results shows that, by switching offline preference dataset to online dataset labeled by other language models, the generated responses are more aligned with human preference. Strengths: The strengths of the paper are listed below: 1. This paper introduces OAIF, which is featured by using on-the-fly generated preference pairs and AI-provided labels. 2. The author conducted experiment on various direct alignment methods and the results consolidate the claim by the authors Weaknesses: My questions and concerns are listed as follows: 1. My first concern is regarding the novelty of the paper. It seems that the language model annotator is essentially a preference model. Therefore, OAIF can be seen as a method of online direct alignment algorithm with access to a preference model. The author mentioned several previous work with on-policy generation and online feedback but in need of a reward model. How is OAIF different from different from these method if we simply plug in the language model annotator as the reward model in their methods? 2. At line 118 the author pointed out that RM might suffer from distribution shift because the training data of RM might not share the same distribution with $\pi_\theta$. However, it seems to me that using language model as preference annotator cannot bypass this problem since the language models' pretraining corpus or the finetuning corpus relating to preference labeling has a similar distribution with $\pi_\theta$. 3. How is OAIF's performance compared to other online methods like RSO and IterativeDPO? I think that these methods might also be included as baselines since reward model can also be taken by AI annotators. Technical Quality: 2 Clarity: 3 Questions for Authors: See weakness Section Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The limitation is discussed by the author Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We reply to each of the review’s concerns below. > My first concern is regarding the novelty of the paper. It seems that the language model annotator is essentially a preference model. Therefore, OAIF can be seen as a method of online direct alignment algorithm with access to a preference model. The author mentioned several previous work with on-policy generation and online feedback but in need of a reward model. How is OAIF different from different from these method if we simply plug in the language model annotator as the reward model in their methods? Please note our work is novel in the following aspects: - We firstly propose AI feedback to address the offline and distributional shift issue of DAP methods - We firstly demonstrate the adaptability of alignment via prompting AI feedback model (Section 4.6) - We firstly show the feasibility of weak-to-strong alignment via OAIF (Section 4.7) > At line 118 the author pointed out that RM might suffer from distribution shift because the training data of RM might not share the same distribution with $\pi_\theta$. However, it seems to me that using language model as preference annotator cannot bypass this problem since the language models' pretraining corpus or the finetuning corpus relating to preference labeling has a similar distribution with $\pi_\theta$. This is a great question! We assume the large-scale pretraining/finetuning and model scaling equips LLM with strong generalization, which could generalize over responses from different policy models when used as an AI annotator. This has been evidenced by the Reward Bench (https://huggingface.co/spaces/allenai/reward-bench, see Gemini 1.5 Pro performance based on zero-shot prompting) results where LLM achieves comparable and even better performance than dedicated reward models by just prompting. > How is OAIF's performance compared to other online methods like RSO and IterativeDPO? I think that these methods might also be included as baselines since reward model can also be taken by AI annotators. Please notice that OAIF is orthogonal to RSO as RSO develops the sampling strategy rather than preference labeling, and we consider iterative DPO as a concurrent work.	Summary: This work extends offline preference learning methods, i.e., DPO, to a online variant by using LLM as annotator to collect new datasets for further preference learning. The results show that Direct alignment from preferences (DAP) methods win-rate over the offline methods beyond 60%. Strengths: 1. Paper is good writing, easy to follow. 2. This online variant provides demonstrates significant performance improvements over offline DAP and RLHF methods through comprehensive evaluations. Weaknesses: 1. The improvement by extending online is under expectation as it introduces more datasets and training budgets. 2. The contribution is limited. The only difference compared to the prior method is substituting the reward model of prior methods (Iterative DPO) to LLMs, though I agree the explicitly static reward model may introduce the model distributional shift problem. 3. Some drawings or comparisons are not fair enough. (a). Table 1 explicitly avoids the limitation of this method by leveraging the feedback from LLM, though it is another variant of the "reward model". (b). Figure 3, the training step is not an approximate x-axis as the online DPO variant has been heavily fine-tuned offline. 4. There are no theoretical foundations, or new plausible explanations, aside from more datasets and the online budget, for the further improvement of the online variant DPO. Technical Quality: 3 Clarity: 3 Questions for Authors: n/a Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: see Weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We reply to each of the reviewer's concerns individually below. > The improvement by extending online is under expectation as it introduces more datasets and training budgets. We may disagree with this argument as it is unclear how much improvement would be considered as “expected”. We showed empirical results, from both human and LLM, that OAIF performs better than several strong baselines, which we believe are solid evidence to support the superiority of OAIF. > The contribution is limited. The only difference compared to the prior method is substituting the reward model of prior methods (Iterative DPO) to LLMs, though I agree the explicitly static reward model may introduce the model distributional shift problem. We consider the distributional shift problem, an overlooked problem in the literature, as one major bottleneck when applying DAP methods, and we consider iterative DPO as a concurrent work. Besides, we’d like to remind the reviewer that the proposal of online learning with AI feedback is just one of our contributions. The high controllability of OAIF via prompting LLM (as shown by the length control experiments in Section 4.6) and the feasibility of weak-to-strong alignment in Section 4.7, we believe, are solid contributions to the community. > Some drawings or comparisons are not fair enough. (a). Table 1 explicitly avoids the limitation of this method by leveraging the feedback from LLM, though it is another variant of the "reward model". (b). Figure 3, the training step is not an approximate x-axis as the online DPO variant has been heavily fine-tuned offline. As stated in your comments, “the explicitly static reward model may introduce the model distributional shift problem”, which highlights the significance of leveraging feedback via large off-the-shelf LLMs, which are not specialized on generations from a single policy. We don’t understand the question (b). We perform OAIF based on SFTed models rather than offline DAP models. Could the reviewer clarify their point? > There are no theoretical foundations, or new plausible explanations, aside from more datasets and the online budget, for the further improvement of the online variant DPO. First, we’d like to highlight that our work strictly follows the theoretical foundations of the corresponding DAP methods, since online learning itself doesn’t break the foundation. Secondly, we consider our work as an empirical study. We believe our empirical evidence is solid enough to support all our claims. --- Rebuttal Comment 1.1: Title: Official Comments to by Reviewer wFW3 Comment: Thanks for the rebuttal! However, leveraging the feedback from LLMs rather than reward models to address the model distributional shift problem is not an appreciated way, and also not valid with extra experiments or proof justify except the final performance, as it introduces other knowledge such as expert information. I strongly suggest that the authors provide additional experimental validation, if this is the preferred method, to support the above claim. I also have the following two comments: 1. Iterative DPO should not be considered as a concurrent work. Takes from [NeurIPS FAQ](https://neurips.cc/Conferences/2023/PaperInformation/NeurIPS-FAQ#:~:text=Papers%20appearing%20less%20than%20two,or%20two%20before%20the%20deadline.): "Papers appearing less than two months before the submission deadline are generally considered concurrent to NeurIPS submissions. Authors are not expected to compare to work that appeared only a month or two before the deadline." 2. (b). Figure 3 shows that the training step is not an approximate x-axis as the online DPO variant has been fine-tuned offline. I apologize for the misunderstanding, I thought the online variant continues to train the offline policy. --- Reply to Comment 1.1.1: Comment: We didn’t follow the reviewer’s point on the expert information issue. It seems that the reviewer has a significant misunderstanding about the signal from our LLM annotators. We would like to clarify that the feedback from LLM annotators is the same as the preference signal obtained from reward models, and there’s no expert information from the LLM annotators. Can the reviewer further clarify the question here? We believe there are also misunderstandings regarding the distributional shift problem in our study. In DAP methods, this problem results from the use of pre-collected offline preference data (sometimes from the other policy models), rather than the use of reward models. The significant log-likelihood gap between online and offline response in Figure 3 clearly explains this problem. We consider iterative DPO and our work as concurrent work as both are finished around the end of last year. Note this is a resubmission from ICML.	Summary: This paper applies direct alignment from preferences (DAP) methods, particularly DPO, to online settings where responses are sampled in an on-policy manner and feedback is provided by the LLM annotator in real-time. Extensive experiments demonstrate the effectiveness of these simple ideas. Strengths: The paper is well-written, with detailed explanations of introduced definitions and discussions with existing methods. The experiments are well-designed, supporting the main idea of the paper. The proposed prompt-controllable approach is particularly commendable. Weaknesses: The rationale for why on-policy learning brings performance gains is not well clarified. The cited reference [1] does not provide strong support for this claim. There is no experimental evidence that on-policy sampling encourages exploration. Most experiments are conducted with the closed-source LLM Palm; evaluating state-of-the-art open-sourced LLMs would enhance generalizability. It is unclear how much of the performance gains are due to on-policy sampling versus online feedback. The reasons why utilizing online on-policy data can avoid overfitting and improve performance should be further analyzed and discussed. References: [1] Lambert, N., Wulfmeier, M., Whitney, W., Byravan, A., Bloesch, M., Dasagi, V., Hertweck, T., and Riedmiller, M. The challenges of exploration for offline reinforcement learning. arXiv preprint arXiv:2201.11861, 2022. Technical Quality: 3 Clarity: 3 Questions for Authors: Is it correct to categorize RSO and iterative DPO as on-policy generation in Table 1? What new opportunities and challenges arise when applying DAP to online settings? Did you encounter common issues of DAP methods, such as overfitting, in the online setting? What are the differences in these issues between online and offline settings? Where are the experimental results to support the superiority of using LLMs over RMs to provide online feedback in Line 267? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The computational overhead introduced by on-policy sampling and online feedback is not discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for identifying our contribution to addressing the off-policy sampling issue via the proposed online AI feedback (OAIF) method. We have addressed each concern as described below. > The rationale for why on-policy learning brings performance gains is not well clarified. The cited reference [1] does not provide strong support for this claim. There is no experimental evidence that on-policy sampling encourages exploration. Our empirical analysis in Figure 8 (Appendix) demonstrates a distributional shift for off-policy responses, where off-policy responses get significantly lower generation probability than on-policy responses. Intuitively, using off-policy responses enforces the model to adapt to these examples apart from learning the preferences, which increases the learning difficulty. > Most experiments are conducted with the closed-source LLM Palm; evaluating state-of-the-art open-sourced LLMs would enhance generalizability. Please note PaLM is available to use and can be fine-tuned on [Vertex](https://cloud.google.com/vertex-ai/generative-ai/docs/learn/overview). We are also aware of a reproduction of this work by HuggingFace which supports our concolusions and will be announced soon. > It is unclear how much of the performance gains are due to on-policy sampling versus online feedback. We included a comparison between on-policy sampling (with a fixed reward model) and OAIF in line 265 and 266 of Section 4.4, which shows that OAIF obtains significantly better performance. > The reasons why utilizing online on-policy data can avoid overfitting and improve performance should be further analyzed and discussed. Thanks for this suggestion! We consider that the online feedback to dynamically sampled on-policy responses offers timely signals for the learning which helps avoid overfitting to some extent. Please also note that OAIF allows high and straightforward controllability on the alignment by modifying the prompt as demonstrated by the length control experiments (Section 4.6), and enables weak-to-strong generalization (Section 4.7). > Is it correct to categorize RSO and iterative DPO as on-policy generation in Table 1? Yes, both RSO and iterative DPO use on-policy sampling, i.e. the responses are sampled from the current policy model during training. > What new opportunities and challenges arise when applying DAP to online settings? Did you encounter common issues of DAP methods, such as overfitting, in the online setting? What are the differences in these issues between online and offline settings? These are great questions! We believe online learning is the right way to unlock the potential of DAP algorithms, but how to select responses and collect preference data from the current policy model and the feedback model during online learning can be challenging. Online DAP, like OAIF, still suffers from overfitting, while the behavior and bias are different. For example, offline DPO often shows clear overfitting patterns with decreased win rates against the SFT baseline; OAIF, in contrast, often produces much longer outputs, which we consider as a type of overfitting. However, such overfitting is adjustable by modifying the AI feedback prompt. > Where are the experimental results to support the superiority of using LLMs over RMs to provide online feedback in Line 267? We performed automatic evaluation by Gemini Pro and showed that OAIF outperforms LLM+RM with a win ratio of >70%. --- Rebuttal 2: Comment: Thanks for your response. I would keep my previous evaluation.	Summary: The paper presents a new method called Online AI Feedback (OAIF) for direct alignment from preferences (DAP) that addresses the limitations of existing DAP methods, which rely on static, offline feedback datasets. By using an LLM as an online annotator to provide real-time feedback during each training iteration, OAIF ensures the alignment process remains on-policy and adapts dynamically to the evolving model. Through human evaluations across various tasks, the authors demonstrate that OAIF outperforms traditional offline DAP and reinforcement learning from human feedback (RLHF) methods. Strengths: OAIF uses LLMs for preference annotation, eliminating the need for a separate reward model and large datasets typically required for RLHF methods. It introduces a new way to address off-policy issues in policy optimization, a significant problem in traditional DPO methods. The paper is well-written and easy to understand. OAIF outperforms offline DPO and other offline RLHF methods. Weaknesses: 1. The idea is straightforward but lacks theoretical proof. The proposed method combines DPO and AI feedback, unlike the constitutional AI paper, which integrates PPO with AI feedback. However, this point is minor. Given the abundance of concurrent work [1-7], the authors should further develop the theoretical analysis of their approach to strengthen their method. 2. Different methods should use an equal amount of training data. In the second epoch of onlineDPO, although the prompts remain the same as in the first epoch, the responses and rank information differ due to online generation. 3. Recent results on Reward Bench indicate that small reward models are more effective than LLM critiques. The iterative DPO methods are similar to OAIF DPO. A performance comparison between OAIF and various iterative DPO methods using cheaper reward models, as both address the off-policy issue, is essential and should be included. [1] Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint [2] RS-DPO: A Hybrid Rejection Sampling and Direct Preference Optimization Method for Alignment of Large Language Models [3] RSO: Statistical rejection sampling improves preference optimization [4] Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682 [5] Hoang Tran, Chris Glaze, and Braden Hancock. 2023. Iterative dpo alignment. Technical report, Snorkel AI. [6] Self-rewarding language models. arXiv preprint arXiv:2401.10020 [7] Is dpo superior to ppo for llm alignment? a comprehensive study. arXiv preprint arXiv:2404.10719. Technical Quality: 2 Clarity: 3 Questions for Authors: N/A Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We respond to the questions you have as follows. > The idea is straightforward but lacks theoretical proof. The proposed method combines DPO and AI feedback, unlike the constitutional AI paper, which integrates PPO with AI feedback. However, this point is minor. Given the abundance of concurrent work [1-7], the authors should further develop the theoretical analysis of their approach to strengthen their method. While we appreciate the recognition of our work’s simplicity which we deem valuable for both academia and industry, we need to highlight that our work identifies the offline issue of DAP methods (not just DPO) and demonstrates the significance of online learning via OAIF with solid experimental evidence. We consider this as a convincing contribution to the community rather than a “minor point”; these concurrent works listed further supports the significance of the problem investigated in our work. We are surprised by the argument that our work lacks theoretical proof, since we didn’t develop new DAP algorithms but make their learning online and on-policy. Take DPO as an example. The equivalence between the optimal policy and optimal reward function is built on online and on-policy learning; but traditional DPO often adopts preference data collected and fixed ahead of the training and even from other policy models, leading to offline and even off-policy learning. Our work ensures the optimality of these DAP algorithms. > Different methods should use an equal amount of training data. In the second epoch of onlineDPO, although the prompts remain the same as in the first epoch, the responses and rank information differ due to online generation. The ability of leveraging newly generated responses and new rank information is exactly the strength of OAIF, which is infeasible for offline DAP methods. We believe this should be considered as an advantage of OAIF rather than a weakness. > Recent results on Reward Bench indicate that small reward models are more effective than LLM critiques. The iterative DPO methods are similar to OAIF DPO. A performance comparison between OAIF and various iterative DPO methods using cheaper reward models, as both address the off-policy issue, is essential and should be included. We also noticed recent results on Reward Bench, particularly that generative LLMs perform on-par with and even better than dedicated reward models. This is exciting evidence supporting the reliability of AI feedback thus the potential of OAIF. A comparison between reward model and OAIF has been provided in line 265 and 266 of Section 4.4. Our experiment showed that the reward model (trained at once) significantly underperformed OAIF from a generative model. We also highlight that iterative DPO requires retraining reward models frequently, which not only complicates the training process but also incurs significant cost, making the “cheaper” argument debatable. In contrast, OAIF supports quick reward prototyping as shown in our length controllability experiments (Section 4.6), where using the reward model is non-trivial.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
A teacher-teacher framework for clinical language representation learning	Accept (poster)	Summary: The paper introduces a novel teacher-teacher framework named LIghtweight kNowledge alignmEnt (LINE), which facilitates knowledge exchange between two pre-existing large language models (LLMs) to enhance clinical language representation. By leveraging complementary knowledge from general-purpose and domain-specific models, LINE aims to harmonize their knowledge within a unified representation space. The framework is validated through downstream tasks showing that the LINE model outperforms individual pre-existing models in understanding and processing clinical language. This approach allows for more efficient sharing of clinical pretrianed models. Strengths: 1. Clarity and Structure: The paper is well-written and structured, offering a clear motivation for the study. This makes it accessible and engaging for readers, facilitating a deeper understanding of the proposed framework. 2. Novelty and Utility: The proposed teacher-teacher framework, LIghtweight kNowledge alignmEnt (LINE), is innovative, providing a pragmatic approach to integrating the strengths of different pre-trained models. This methodology is particularly notable for its potential to enhance clinical language representations without the need for developing new models from scratch. 3. Usability and Efficiency: The framework is user-friendly and does not require retraining of the original models, which significantly reduces computational overhead and simplifies its adoption in real-world applications. 4. Empirical Validation: The experimental results demonstrate stable and significant improvements over existing methods, substantiating the efficacy and value of the proposed framework in practical settings. Weaknesses: Data Requirements and Availability: A notable limitation of the proposed LINE framework is its dependency on well-aligned and specific types of data sources, which may not be readily available or commonly found in practical settings. For example, integrating data from disparate modalities like CT and MRI requires the availability of cases that include both types of data, which may not always be feasible. This requirement could limit the framework's applicability across different clinical or real-world scenarios where such aligned data sets are scarce. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. See weakness, under such situation is it possible to apply your method? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No. The authors discuss the potential improvement instead of the limitation of the current work. Bring more information and try other situations cannot be counted as an adequate discussion of limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Data Requirements and Availability: A notable limitation of the proposed LINE framework is its dependency on well-aligned and specific types of data sources, which may not be readily available or commonly found in practical settings. For example, integrating data from disparate modalities like CT and MRI requires the availability of cases that include both types of data, which may not always be feasible. This requirement could limit the framework's applicability across different clinical or real-world scenarios where such aligned data sets are scarce. See weakness, under such situation is it possible to apply your method? This is an extremely insightful question, and we appreciate you bringing this up! Yes, our method can handle such situations by introducing an intermediary to bridge the gap between different data sources. For instance, if we want to infer $\mathbf{x}_1$ and $\mathbf{x}_3$ about each other but lack paired data $(\mathbf{x}_1, \mathbf{x}_3)$, the LINE framework can use an intermediary $\mathbf{x}_2$. This forms pairs $(\mathbf{x}_1, \mathbf{x}_2)$ and $(\mathbf{x}_2, \mathbf{x}_3)$, enabling the use of three teachers to jointly learn the alignment between all data pairs. In a concrete example involving CT and MRI scans, it is often rare for patients to have both scans simultaneously ($\mathbf{x}_1$ for CT and $\mathbf{x}_3$ for MRI). However, each type of scan typically comes with corresponding clinical notes ($\mathbf{x}_2$). These clinical notes can serve as intermediaries. By breaking down the clinical notes into concepts corresponding to specific scan regions, we can leverage Teacher 2, a pre-trained LLM, to incorporate factual relationships about these concepts. Teachers 1 and 3, pre-trained large vision transformers, can then embed the CT and MRI scans, respectively. The three teachers can train jointly, allowing the framework to effectively handle the alignment of disparate data modalities. > The authors discuss the potential improvement instead of the limitation of the current work. Bring more information and try other situations cannot be counted as an adequate discussion of limitations. Thank you for your constructive comment. We recognize two main limitations of our framework: First, our approach requires data to come in pairs. While we can introduce intermediaries to create artificial pairs, this remains a limitation. The reliance on paired data may restrict the framework's applicability in scenarios where such pairs are not readily available. Second, effective training with a list of concepts necessitates a pre-existing factual knowledge graph. Access to such a knowledge graph is not always guaranteed, which can hinder the framework's performance and generalizability. We appreciate your feedback and will incorporate a more detailed discussion of these limitations in the paper. --- Rebuttal Comment 1.1: Comment: Thanks for your comment. I will keep my score.	Summary: This paper presents an interesting topic on LLM but the importance of this problem is not convincing and the methods here is not novel. Strengths: The teacher-teacher concept is novel to some extent. Weaknesses: 1. The problem's importance is not significant. 2. There lacks the inclusion of SOTA models like llama, gpt, etc. 3. The results improvement is limited as shown in Tab. 4,5. 4. The Fig1 lacks details of the proposed method. Technical Quality: 3 Clarity: 3 Questions for Authors: None. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: See details in weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comment.We will address your comments point-by-point. > The problem's importance is not significant. Thank you for raising this important question. While our motivating example originated from the medical domain, our LINE framework is broadly applicable to a wide range of scenarios. The core problem we address involves pairs of unlabeled data points $(\mathbf{x}_1, \mathbf{x}_2)$ from different domains, where $\mathbf{x}_2$ consists of items with known relationships derived from pre-existing factual knowledge graphs. Our objective is to learn an embedding $\mathbf{z}_2$ for $\mathbf{x}_2$ that closely matches the embedding $\mathbf{z}_1$ for $\mathbf{x}_1$. This general framework can be applied to various critical applications, including the alignment of text from structured EHR codes paired with clinical notes, images/audio paired with text. Below, we detail a few specific applications: 1. Text Summarization: In healthcare systems, clinical notes ($\mathbf{x}_1$) often need to be inferred from corresponding concepts ($\mathbf{x}_2$) due to regulatory constraints. Since raw clinical notes are frequently inaccessible, using a list of concepts to summarize and index notes enables de-identification and facilitates sharing among researchers. The concept list can be utilized instead of full clinical notes for downstream analysis. 2. Text Summarization for Image and Audio: In a cross-modality setting, the LINE framework can align audio/image data ($\mathbf{x}_1$) with corresponding textual information ($\mathbf{x}_2$). For instance, medical images are more challenging to anonymize than text due to embedded metadata and visual information that might indirectly reveal patient identity. In this context, LINE can employ a pre-trained vision transformer to embed images and a pre-trained language model to embed conceptual entities from clinical text. Consequently, LINE can provide embedding surrogates for images, such as CT scans, by leveraging only the concepts extracted from corresponding clinical notes. The text summaries can be used to both index the images and for direct downstream analyses. Additionally, our LINE algorithm has the potential to identify "residual" information in the images that is not captured by the paired data, which can then be projected into the text domain to provide meaningful insights. We will incorporate this into the motivation and discussion sections of the paper. > There lacks the inclusion of SOTA models like llama, gpt, etc. Thanks for your suggestion! During the rebuttal phase, we are able to use GPT4 as the strong teacher 1 to come up with LINE (GPT4+CODER). And then we further compared the new LINE results with the GPT4 baseline. The results are shown in the tables in the rebuttal pdf. We also provide a summary of the key results below for your easy reference. The following tables are additional results to be added to Tables 2, 3 and 5 of the paper. The better results are highlighted in bold. \| Model / Metrics \| LINE (CODER+GPT4) \| GPT4 \| \| ----------------- \| ----------------- \| ----- \| \| Mean Rank \| 1.477 \| 1.778 \| \| Mean Reverse Rank \| 0.872 \| 0.820 \| \| Top10@Acc \| 0.995 \| 0.988 \| \| Models / Relation \| Parent \| Sibling \| May Treat/Prevent \| Classifies \| DDX \| Method of \| Causative \| \| ----------------- \| --------- \| --------- \| ----------------- \| ---------- \| ----- \| ---------- \| ---------- \| \| GPT4 \| 0.974 \| 0.940 \| 0.825 \| 0.991 \| 0.939 \| 0.934 \| 0.935 \| \| LINE (GPT4+CODER) \| 0.977 \| 0.932 \| 0.931↑ \| 0.988 \| 0.938 \| 0.965↑ \| 0.947↑ \| \| \| Concept \| \| \| \| Sentence \| \| \| \| \| ----------------- \| --------- \| --------- \| --------- \| --------- \| --------- \| --------- \| --------- \| --------- \| \| \| Precision \| Recall \| F1 \| Accuracy \| Precision \| Recall \| F1 \| Accuracy \| \| GPT4 \| 0.718 \| 0.687 \| 0.694 \| 0.711 \| 0.805 \| 0.787 \| 0.791 \| 0.781 \| \| LINE (GPT4+CODER) \| 0.731 \| 0.715 \| 0.714 \| 0.722 \| 0.806 \| 0.786 \| 0.789 \| 0.781 \| From the tables above, it can be observed that in most of the setting, LINE model performs better than directly using GPT4. > The results improvement is limited as shown in Tab. 4,5. As discussed in our response to your first question, our evaluation of the proposed framework is not solely focused on outperforming the benchmark model. It also emphasizes the potential of using a list of concepts to summarize notes, which enables de-identification and facilitates sharing among researchers. In Table 2, we evaluate whether the summarization from the list of concepts can be effectively aligned with the clinical text through a rank-based retrieval task. In Table 5, we assess whether the performance using only concepts can be comparable to the performance using raw text. The results from both the tables in the paper and the additional results presented above consistently demonstrate that the LINE model is more effective in summarizing text. > Figure 1 lacks details of the proposed method. Thank you for your comment. We have revised Figure 1 to include more details, which can be seen in Figure 1(b) of the rebuttal PDF. Specifically, Teacher 2 utilizes the trainable LINE module to first summarize the set $\mathbf{x}_2$, guided by pre-existing knowledge (i.e., the graph attention module and relational contrastive loss). Then, Teacher 2 generates an embedding $\mathbf{z}_2$ for $\mathbf{\tilde{x}}_2$ that closely resembles the embedding $\mathbf{z}_1$ for $\mathbf{x}_1$ through the alignment loss. Simultaneously, Teacher 1, a stronger LLM, uses a fully connected layer to learn the relational knowledge provided by Teacher 2 via the alignment loss. --- Rebuttal Comment 1.1: Comment: ## Additional Experiment Thank you once again for your suggestion and thank you for your patience! We have incorporated an additional benchmark dataset focused on the mental health domain to further evaluate the proposed LINE framework against other methods. Specifically, we retrieved suicide-related publications from PubMed, using the search term "suici" to capture both "suicide" and "suicidal" references. This resulted in a collection of 8,000 publications with "suicid" in the title. We excluded a small number of large files (>215M) that required more than 16GB of RAM, accounting for approximately 1% of the total dataset. From these publications, we extracted titles and keywords, and further refined the dataset by removing any publications that did not contain keywords. We then conducted a rank-based retrieval task, using the keywords to retrieve the corresponding article. For each positive (title, keyword) pair, we generated 100 negative pairs by randomly substituting the title with one from a different article. Next, we computed the cosine similarity between the mean embedding of the keyword list and the title embedding for all pairs. These pairs were then ranked based on their cosine similarity scores, from highest to lowest. To assess performance, we calculated the mean rank, mean reverse rank, and Top-10 accuracy for all positive pairs. The results, presented in the table below, show that both LINE models show significant improvements over their respective teacher models. \| Metric \| PubmedBERT \| BioBERT \| SapBERT \| CODER \| BGE \| CODER$\to$BGE \| BGE$\to$CODER \| GPT \| LINE(BGE+CODER) \| LINE(GPT+CODER) \| \| --------------------- \| ---------- \| ------- \| ------- \| ----- \| ---- \| ---------- \| ---------- \| -------- \| --------------- \| --------------- \| \| Mean Rank \| 29.37 \| 35.84 \| 8.52 \| 22.89 \| 3.03 \| 17.32 \| 41.55 \| 2.95 \| 2.07 \| 1.94 \| \| Mean Reverse Rank \| 0.17 \| 0.14 \| 0.52 \| 0.27 \| 0.84 \| 0.33 \| 0.08 \| 0.85 \| 0.83 \| 0.85 \| \| Top10@Acc \| 0.34 \| 0.27 \| 0.79 \| 0.46 \| 0.94 \| 0.56 \| 0.18 \| 0.95 \| 0.97 \| 0.98 \| --- Rebuttal 2: Comment: Dear Reviewer Hfn8, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors already left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC	Summary: The authors look to address the question representational alignment between language models trained on different textual domains to improve performance of potentially both models on their out-of-domain text. The authors propose to specifically investigate this in the context of EHR text, and choose as their models for this CODER and BGE. They propose a contrastive loss, and additionally propose to train an alignment module/project layer rather than end-to-end training of the teacher models. Strengths: The concept is solid and well implemented and motivated. I wonder if it would be possible to further generalize it beyond medical text - which it is restricted too due to the reliance on alignment with extracted medical concepts by NILE. The discussion mentions this possibility, but it would be exciting to see it in action. The clinical NLP benchmarks are particularly appropriate for the task. Weaknesses: Some of the benchmark tasks are older, and the comparisons could be more robust. Some ablations are missing. The project's scope is incredibly narrow: encoder models on extractive medical tasks. While the authors claim that the technique is broadly generalizable, it would be nice to see proof-of-concept. The work seems to me to fit more into the realm of domain adaptation rather than learning by alignment. We aren't learning novel models here via alignment (like CLIP), but rather, pushing the learned representations of two different models into a common space. I'd strongly consider citing and discussing DA literature for this paper. Technical Quality: 3 Clarity: 3 Questions for Authors: Would it be possible to further test the LINE model on other, more varied, benchmarks to see how well those newly aligned representations perform? It could also be exciting to explore this with generative models. Were alternative frameworks considered for the concept alignment? Why not align directly in embedding space without the grounding concepts? This would be an interesting ablation to perform to assess the significance of the extracted concepts on the underlying learned representation. Conversely, could you just fine-tune the generalist model on the extracted concepts as a means of medically aligning it? How well does that perform? Why not also compare the BGE-->CODER projection (inverse direction of the BGE-->CODER projection)? If it isn't technically feasible to do in an end-to-end fashion, perhaps this could be approximated by tuning LoRA on the base models? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No discussion of limitations. Without additional experiments at the minimum a stated limitation should be the highly restricted domain of application (purely encoder models on medical topics). Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed comments. In the following, we will address your questions point-by-point. > The project's scope is incredibly narrow. Thank you for raising this important question. While our motivating example comes from the medical domain, our LINE framework is broadly applicable to various scenarios. The core problem we address involves pairs of unlabeled data points $(\mathbf{x}_1, \mathbf{x}_2)$ from different domains, where $\mathbf{x}_2$ consists of items with known relationships derived from factual knowledge graphs. Our goal is to learn an embedding $\mathbf{z}_2$ for $\mathbf{x}_2$ that closely matches the embedding $\mathbf{z}_1$ for $\mathbf{x}_1$. This framework can be applied to multiple critical applications, such as: 1. Text Summarization: In healthcare, clinical notes ($\mathbf{x}_1$) often need to be inferred from concepts ($\mathbf{x}_2$) due to regulatory constraints. Using a list of concepts to summarize notes enables de-identification and facilitates sharing among researchers. These concept lists can replace full clinical notes for downstream analysis. 2. Text Summarization for Image and Audio: The LINE framework can align audio or image data ($\mathbf{x}_1$) with textual information ($\mathbf{x}_2$). Medical images are harder to anonymize than text due to metadata and visual information. LINE can use a pre-trained vision transformer to embed images and a language model to embed conceptual entities from clinical text, providing embedding surrogates for images like CT scans. These text summaries can index the images and be used for downstream analyses. Additionally, LINE can identify "residual" information in the images not captured by paired data and project it into the text domain for insights. This discussion will be included into the paper. While including proof-of-concept on other settings would be beneficial, we are unable to do so due to time constraints. > I'd strongly consider citing and discussing DA literature for this paper. Thank you for your suggestion. We agree that our work is related to domain adaptation (DA) and will cite relevant literature, though our approach differs in key ways: - Label Requirements: Unlike DA, our method does not require labels during training. DA typically needs task-related labels for the source domain and may use "pseudo" labels for the target domain. - Information Retention: Our goal is to align latent embeddings to preserve both overlapping and complementary information, using a residual alignment training step to refine the alignment with additional concepts. DA focuses only on retaining overlapping information. - Embedding Richness: We aim to create semantic-rich embeddings by leveraging a graph-attention module, preserving relational information and mitigating rank degeneracy. DA often produces domain-agnostic embeddings, whose information richness depends on the specific task. We will further include this discussion in the paper. > Would it be possible to further test the LINE model on other, more varied, benchmarks? Thank you for your suggestion! In response, we are currently running an additional benchmark to evaluate the performance of the LINE models. We anticipate being able to present these results in the early stages of the discussion phase. > It could also be exciting to explore this with generative models. The LINE framework generates in the embedding space, not the original data space. Given an unlabeled data pair $(\mathbf{x}_1, \mathbf{x}_2)$ from spaces $\mathcal{X}_1$ and $\mathcal{X}_2$, respectively, LINE aligns their latent embeddings $\mathbf{z}_1 \approx \mathbf{z}_2$. For a new observation $\mathbf{x}_2^{\text{new}}$, we can map it to the latent space to generate $\mathbf{z}_1^{\text{new}} \approx \mathbf{z}_2^{\text{new}} = \text{LINE}(f({\text{teacher 2}}(\mathbf{x}_2))$ for the missing $\mathbf{x}_1^{\text{new}}$. To extend this, we could add a decoder to map embeddings back to the clinical note space using a simple reconstructive loss. We will discuss this potential extension in future work. > Why not align directly in embedding space without the grounding concepts? Thank you for your thoughtful question. Grounding concepts in our framework both serve to integrate external factual knowledge to enhance model performance and safety, and also provide an independent data source for generating clinical note summaries, making them integral to our approach. > Why not also compare the BGE-->CODER projection? Thank you for the suggestion! Following your advice, we have included the results for the BGE-->CODER projection across our tasks in the rebuttal PDF. Overall, the projection between BGE and CODER yields worse results compared to the proposed LINE model. > Conversely, could you just fine-tune the generalist model on the extracted concepts as a means of medically aligning it? During the rebuttal phase, we attempted to fine-tune generalist models directly and with LoRA, but found it infeasible due to extremely long training times and high computational resource requirements. This highlights the computational efficiency of our proposed method. As shown in Figure 1(a) of the rebuttal PDF, even with LoRA, training the generalist BGE model takes at least 11 days per epoch, with rapidly increasing computational overhead. In contrast, our proposed model requires only about 2 hours per epoch using the same resources. > Without additional experiments at the minimum a stated limitation should be the highly restricted domain of application We acknowledge this concern and will include a discussion of the limitation that our current framework has only been tested on medical topics. We recognize the importance of evaluating our framework in other domains and will address this as a key area for future work. --- Rebuttal 2: Comment: Dear Reviewer TNnd, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors already left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC	Summary: This paper introduce a teacher-teacher framework for clinical language representation learning. The framework uses a lightweight knowledge alignment module to harmonize the knowledge of both models within a unified space, which including two steps: The first step involves initial training to define residuals and capture complementary information. The second step focuses on refining the alignment by recovering residual information. The framework was validated using the MIMIC-IV database, where the LINE model outperformed baseline models in aligning concept and text representations. Strengths: The main contribution of the work is proposed teacher-teacher framework, and training strategy. - Originality: The teacher-teacher framework is very interesting as it enables mutual enhancement between two pre-existing LLMs, a unique departure from traditional approaches that typically involve training a new model or continual pre-training of existing models. This innovative method opens new avenues for leveraging existing resources to achieve superior performance. - Quality: The paper demonstrates high quality through its validation using the MIMIC-IV database, a well-known and respected dataset in the clinical domain, adding significant credibility. Additionally, the LINE model's performance is compared against several strong baseline models, showing clear improvements across various downstream tasks, thus underscoring the robustness and reliability of the proposed framework. - Clarity: The paper is well-written and clearly structured, making it accessible to both domain experts and those new to the field. The introduction provides a comprehensive background and motivation for the proposed framework, while the methodology section offers detailed descriptions of the teacher models and the LINE module. - Significance: The practical applications and potential impact on the clinical domain shown the significance of this work. The teacher-teacher idea has substantial implications for more advancing NLP applications in other filed. Weaknesses: 1. Figure 1 is somewhat confusing. From my understanding, Teacher 1 should be a strong LLM, while Teacher 2 should be an LLM with existing domain-specific knowledge. However, Figure 1 gives the impression that Teacher 2 serves merely as a database, making the framework resemble a RAG framework. 2. Although the paper compares the LINE model against several strong baseline models, it lacks a detailed comparison with the latest strong general LLMs, such as GPT-4, which should be considered a strong baseline. Consider adding a small comparative analysis or stating the advantages of the framework over simply using GPT-4. 3. The paper underscore the practical value of the framework, but it does not sufficiently address potential practical implementation challenges, such as computational requirements and scalability when applied in real-world clinical settings. Technical Quality: 2 Clarity: 3 Questions for Authors: 1.From Figure 1, if Teacher 2 only serves to provide domain-specific knowledge, why not implement a RAG framework, which is training-free and potentially more reliable? 2. Have you addressed potential hallucination issues? Could one teacher potentially mislead the other during the knowledge exchange process? 3. What are the potential computational and scalability challenges of implementing the teacher-teacher framework in real-world clinical settings? How do you propose to mitigate these challenges? 4. How can regulatory mechanisms be incorporated into the framework for safety? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The paper has limited discussion on the broader implications of implementing the teacher-teacher framework in clinical settings. Consider to add the assessment of how the framework could impact patient care, data security, and trust in AI systems in healthcare. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments! In the following, we will address them point-by-point. > Figure 1 is somewhat confusing. From my understanding, Teacher 1 should be a strong LLM, while Teacher 2 should be an LLM with existing domain-specific knowledge. However, Figure 1 gives the impression that Teacher 2 serves merely as a database. Thank you so much for pointing this out! We have modified Figure 1 into Figure 1(b) in the rebuttal PDF for better clarity. In summary, both Teacher 1 and Teacher 2 can generate embeddings when feeding in new data which are then aligned by LINE. When the data are paired, e.g. they come from the same hospital visit of a patient, their LINE-aligned embeddings will be close to each other. > From Figure 1, if Teacher 2 only serves to provide domain-specific knowledge, why not implement a RAG framework, which is training-free and potentially more reliable? The main role of Teacher 2 is to learn an embedding using a list of concepts to summarize the clinical notes, effectively matching the embedding generated directly from the raw notes. Since raw clinical notes are frequently inaccessible, summarizing and indexing notes using a list of concepts enables de-identification and facilitates sharing among researchers. Additionally, by allowing two teachers to align knowledge with each other via LINE, we hope to also improve the quality of the embeddings . While the RAG framework also leverages pre-existing knowledge networks to enhance learning, it is not a generative model and cannot be used to generate new embeddings. Therefore, it does not fulfill the same role as our proposed method, which not only aligns domain-specific knowledge but also generates new, robust embeddings. We will include RAG in our discussion of related work in the paper. > Although the paper compares the LINE model against several strong baseline models, it lacks a detailed comparison with the latest strong general LLMs, such as GPT4, which should be considered a strong baseline. Consider adding a small comparative analysis or stating the advantages of the framework over simply using GPT4. Thank you for your advice! During the rebuttal phase, we incorporated GPT4 as a strong Teacher 1 in our LINE framework, resulting in LINE (GPT4+CODER). We then compared the performance of this new model with the GPT4 baseline. The results, shown in the tables in the rebuttal PDF, indicate that LINE (GPT4+CODER) generally achieves better performance than GPT4 alone. > Have you addressed potential hallucination issues? Could one teacher potentially mislead the other during the knowledge exchange process? Thank you for the insightful question! To mitigate potential hallucination issues, we integrate factual knowledge from external reputable sources (e.g., UMLS) into our framework. Specifically, in our framework, external factual knowledge is incorporated into the training of Teacher 2 via the multihead graph attention module and the relational contrastive loss. This knowledge is then further propagated to Teacher 1 through the alignment loss. In this regard, Teacher 2 acts as a factually rigorous component that regularizes potentially misleading information from Teacher 1. We conducted an experiment to assess the quality of the LINE-aligned concept embeddings. The results, shown in Table 3 of the paper and the second table in rebuttal PDF, indicate that the generated concept embeddings faithfully preserve factual relationships and even slightly improve upon capturing these relationships compared to the baseline. > What are the potential computational and scalability challenges of implementing the teacher-teacher framework in real-world clinical settings? How do you propose to mitigate these challenges? Thank you for the question! One of the notable advantages of our proposed framework is its computational and memory efficiency, making it feasible to deploy with limited computational resources. To illustrate this, please refer to Figure1(a) in the rebuttal PDF, where we compare the training time of our model to both direct fine-tuning of LLMs and low-rank approximated fine-tuning of LLMs. The comparison shows that our model can be trained within two day on a RTX8000 48GB card. > How can regulatory mechanisms be incorporated into the framework for safety? Consider to add the assessment of how the framework could impact patient care, data security, and trust in AI systems in healthcare. Thank you for this important question! Below, we assess how the framework impacts patient care, data security, and trust in AI systems, and discuss the incorporation of regulatory mechanisms for safety. This discussion will be included in the paper. - Data Security: We utilize the publicly accessible MIMIC-IV dataset, which aligns with real-world clinical notes without privacy concerns. During inference, real medical notes aren't required; instead, a concept extractor tool like NILE can be used to extract key concepts for Teacher 2, preventing the use of sensitive patient information. Regulatory mechanisms can monitor and update the medical concepts in the knowledge graph, ensuring they reflect the most up-to-date factual knowledge. For instance, newly defined concepts like COVID-19 can be added, outdated concepts depreciated, and misleading concepts corrected. - Patient Care: The framework's computational efficiency allows for the effective use of state-of-the-art LLMs, improving downstream tasks such as disease diagnosis and lab results analysis, thereby enhancing patient care. - Trust in AI Systems: LINE can be retrained on local servers with various combinations of up-to-date LLMs, necessitating continual quality control. For example, our clinical concept similarity task (Section 3.3.1) can initially assess the fidelity of learned concept embeddings. During deployment, regular user feedback should be collected to further improve the system. --- Rebuttal 2: Comment: Dear Reviewer Yopj, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC --- Rebuttal Comment 2.1: Comment: Thank you for the response! The new figure 1 makes sense to me. Thank you for new results. I decide to increase soundness from 2 to 3.	Rebuttal 1: Rebuttal: Thank you all for your comments and questions! Based on your suggestions, we have made the following major changes during the rebuttal phase: ### Additional Experiment 1. New Teacher Model: We have adopted the OpenAI text embedding model "text-embedding-v3-small" as Teacher 1. Since it was released alongside GPT4 and demonstrates strong performance, we refer to it as "GPT4" for brevity. The subsequently trained model is referred to as "LINE (GPT4+CODER)". As suggested by the reviewers, we have included "CODER$\to$BGE" and "BGE$\to$CODER" as comparison benchmarks. "CODER$\to$BGE" projects the concept embeddings from CODER into the BGE embedding space using a projection matrix, while "BGE$\to$CODER" performs the inverse operation. The results are presented in the tables in the rebuttal PDF. Note that, since token-level embeddings from GPT4 are unavailable, we cannot perform i2b2 tasks using GPT4 or LINE (GPT4+CODER). 2. Computational Efficiency: To assess the computational efficiency of the proposed framework, we have benchmarked the estimated training time of our model against both direct fine-tuning of LLMs and low-rank approximated fine-tuning of LLMs [1] on a single RTX8000 48GB card. These results are shown in Figure 1(a) of the rebuttal PDF. ### Modified Figure 1 of the Paper Following the reviewers' advice, we have modified Figure 1 in the paper for better clarity. Please see Figure 1(b) in the rebuttal PDF. ### Proposed Changes to the Paper 1. Expanded Scope of Applicability: We will add concrete examples to illustrate the broad applicability of the LINE framework. 2. Additional Related Works: We will include related literature on domain adaptation and retrieval-augmented generation to discuss how the proposed framework differs from these existing lines of research. 3. Discussions on Safety: We will add discussions on how the proposed framework addresses the issue of hallucination, manages concepts in different contexts, and assess its impact on patient care, data security, and trust in AI systems in healthcare. 4. Discussion on Limitations: We will add a discussion of the limitations of the proposed framework. We appreciate your feedback and will make these changes to enhance the clarity and comprehensiveness of our paper. [1] Hu, Edward J., et al. "Lora: Low-rank adaptation of large language models." arXiv preprint arXiv:2106.09685 (2021). Pdf: /pdf/1626d8a4d29a4aea4bd8eaa1bee2b223cfe543a0.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper proposes a mutual learning framework, called LINE, between two pre-existing LLMs in the healthcare domains. By harmonizing the knowledge of two distinct LLMs into a unified representation space, the model achieves better performance on intrinsic and extrinsic downstream evaluations of clinical tasks. Strengths: Clear motivation. Overall well written. The methodology was reasonably designed to map representations from two distinct LLMs into a unified representation space. The method achieves better performance on downstream clinical tasks. Weaknesses: 1. Only two LLMs (BGE and CODER) were aligned by LINE. It is unclear if LINE will work on combinations of other LLMs. 2. LINE make downstream predictions based on clinical concepts only, rather than the full context. The concepts themselves can be negated, historical and hypothetical in context, but the proposed method does not seem to consider this. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why was NILE selected? Have any other extractors been compared? Does the selection of extractors have a significant impact on results? 2. Line 222, which contrastive loss function was used eventually? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments! Below are our responses to your questions, addressed point-by-point: > It is unclear if LINE will work on combinations of other LLMs. Thank you for raising this important question. To address your concern, we have extended our experiments to include additional combinations of LLMs. Specifically, we tested the alignment of LINE with the combination of GPT4 and CODER. The results, presented in the tables in the rebuttal PDF, indicate that LINE (GPT4+CODER) generally achieves better performance than GPT4 alone. > The concepts themselves can be negated, historical and hypothetical in context, but the proposed method does not seem to consider this. Thank you for your insightful comment! We agree that concepts can vary significantly depending on their context, and our approach does take some parts of this into account. Specifically, we have implemented mechanisms to handle negated concepts and will further include a discussion on the possibility of modeling for concepts in historical or hypothetical contexts. ##### Handling Negated Concepts: In our approach, we account for negated concepts by creating a separate dictionary of negative concepts derived from the original positive concepts. For instance, the positive concept "pneumothorax" is associated with a negated concept represented as "concept pneumothorax unobserved". To ensure that negated concepts are accurately represented, we use a multihead graph attention module (as detailed in Section 2.2) to update the embeddings of positive concepts, which are then adapted for their corresponding negative concepts through a projection layer. We introduced a loss function specifically designed to maintain a distinct representation for negated concepts. This loss function ensures that the cosine similarity between the updated negative $\mathbf{c}_p$ concept and its positive counterpart $\mathbf{c}_p$ remains below a predefined threshold, $\delta$, calculated as: $$ -\log\frac{e^{\delta - \text{cos}(\mathbf{c}_p, \mathbf{c}_n)}}{1 + e^{\delta - \text{cos}(\mathbf{c}_p, \mathbf{c}_n)}}, $$ where $0 < \delta \leq 0.5$. This mechanism is detailed in the Appendix and has been implemented in our two-step training process. ##### Historical and hypothetical contexts: Additionally, we recognize the importance of considering whether concepts appear in historical or hypothetical contexts. Using NILE, we can effectively identify and extract contextually relevant keywords such as "previous", "since", or "in the near future". This capability enables us to incorporate context-specific modeling based on the temporal or hypothetical relevance of concepts. However, due to the complexity of real-world contexts, designing appropriate loss functions to account for these nuances requires further exploration and discussion. > Why was NILE selected? Have any other extractors been compared? Does the selection of extractors have a significant impact on results? We chose NILE primarily for its speed and convenience. Compared to cTAKES and MedTagger—both of which are popular extractors—NILE is about 2000 times faster than cTAKES and 400 times faster than MedTagger, while delivering comparable performance. Additionally, our framework includes a residual refinement step to recover important clinical concepts that might be missed due to issues like misspellings. This means that while the extractors need to perform reasonably well, they do not need to achieve extremely high accuracy. Consequently, the choice of extractor has a minimal impact on the overall results, as long as it effectively extracts crucial concepts. > Line 222, which contrastive loss function was used eventually? Thanks for allowing us to clarify! We use the triplet loss. --- Rebuttal 2: Comment: Dear Reviewer Yayo, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC --- Rebuttal Comment 2.1: Comment: Thank you for the rebuttal! I will keep my recommendation unchanged.	null	null	null	null	null	null
UIR-LoRA: Achieving Universal Image Restoration through Multiple Low-Rank Adaptation	Reject	Summary: This paper introduces a universal image restoration framework UIR-LoRA based on multiple low-rank adapters. UIR-LoRA employs the pre-trained text-to-image diffusion model SD-turbo as the shared component. It utilizes a LoRA composing strategy based on the degradation similarity predicted by CLIP encoder to combine different LoRA modules. Experiments show the effectiveness of the proposed method. Strengths: 1. The proposed LoRA-based Universal IR method is easy to understand and follow. 2. The motivation of this paper is very clear to me. Weaknesses: 1. UIR-LoRA adopts SD-turbo as the pre-trained backbone for image restoration. However, SD-tubo utilizes VAE with high compression rate to encode input images, resulting in severe detail distortion for image restoration. This issue has been widely discussed in recent published works [1,2]. However, the paper ignores this very important issue in the Method Section and only mentions the skip-connections for VAE in Line 223. 2. The degradation-aware router seems to be unreliable. I do not believe that the original pre-trained CLIP Text Encoder can distinguish between different degradations through degraded text representations, such as "rain" and "raindrop". Therefore, DA-CLIP fine-tunes the original CLIP. But this paper doesn't contain any discussions about this. 3. This paper does not provide complete technical details, such as how the LQ image is used as a condition for SD-turbo. Is ControlNet used, or is it directly concatenated? I do not see any information about this in the paper. 4. Tab. 1 only reports the trainable Param for UIR-LoRA. I think it's necessary to report the overall Param of the model. In addition, the reported PSNR for DiffBIR is very low. Did the authors add skip-connections to the VAE of DiffBIR for a fair comparison? 5. The visual results in Fig. 3 seem strange. The visual results of Restormer show noticeable artifacts between patches. Do the authors test Restormer using a tiled mode? As far as I know, using a single A100 GPU (Line 251), Restormer can restore the entire image without encountering out-of-memory issues. [1] Wang, Wenjing, et al. "Zero-Reference Low-Light Enhancement via Physical Quadruple Priors." In CVPR, 2024. [2] Geng, Zigang, et al. "Instructdiffusion: A generalist modeling interface for vision tasks." In CVPR, 2024. Technical Quality: 1 Clarity: 2 Questions for Authors: 1. Authors should discuss the skip-connections for VAE in the Method Section with more details. 2. Can authors provide the degradation prediction accuracy for more different predictions (eg, rain/raindrop)? 3. Authors should provide more technical details of the proposed method. 4. More experimental results and explanations should be included. Confidence: 5 Soundness: 1 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your thorough review and valuable feedback. 1. Detail distortion arises from operations such as downsampling or pooling. Using skip connections has become a standard and commonly used method to address this issue, as seen in the bypass decoder in [R1], the skip connections in [R2] and [R3], and even tracing back to the original U-Net. Firstly, this issue is not our focus and we have never claimed skip-connections as a contribution in our paper. Secondly, skip-connections are already widely adopted in neural networks and we have provided the references that our paper follows. In these cases, we think our handling of this part is appropriate and will not cause misunderstandings for our method. 2. The original CLIP and DA-CLIP do not use text representations to distinguish degradation types during testing; instead, they use image features extracted by the image encoder to classify degradation types. We used the trained DA-CLIP [R4]. The fine-tuning process of CLIP is a contribution of DA-CLIP, and we have never claimed it as our contribution. DA-CLIP uses the extracted degradation representation as input to the attention part of the diffusion process. In contrast, we use the representation to calculate the similarity of each degradation type and then combine the LoRA of different degradations based on this similarity. Compared to DA-CLIP's implicit representation, our way is an explicit and clearly defined approach. Intuitively, explicit methods have stronger interpretability, and experimentally, Sections 4.3 and 4.4 validate the effectiveness of our approach. 3. SD-turbo is a one-step method. When using SD-turbo, we did not use ControlNet or concatenated operations; instead, we used the representation of the LQ image in latent space as input directly. This way can be found in [R2]. And we will add this detail in our paper. 4. When training and testing DiffBIR, we used its default parameter settings and default structure, which inherently does not include skip-connections. Our experiments were conducted under the same data and environment for comparison. Therefore, there is no issue of unfairness. The lower performance of DiffBIR is due to the need to use SwinIR for preprocessing. When the preprocessing network struggles to handle multiple degradations effectively, the performance metrics of DiffBIR will be lower. The overall parameter can be found in T1. T1: \| DA-CLIP \| SD-turbo \| LoRAs(trainable) \| Overall \| \| :------: \| :----: \| :-----: \| :----: \| \| 125.2M \| 949.1M \| 95.2M \| 1169.5M \| 5. In the dataset, we test Restormer with the entire image whenever possible. However, for some images, their dimensions cannot be evenly divided by 2 during the forward process. In such cases, we use a tiled mode to test these images, as shown in the JPEG degradation image in Figure 3. 6. We used trained DA-CLIP, and except for the blur degradation, the accuracy for the other nine degradations in the test set was 100%, as shown in Table 4 of Daclip-IR. Therefore, in Section 4.5 of our paper, we applied a simple and effective modification to improve the prediction accuracy for motion blur. R1: Zero-Reference Low-Light Enhancement via Physical Quadruple Priors. CVPR2024. R2: One-Step Image Translation with Text-to-Image Models. Arxiv2024. R3: Exploiting Diffusion Prior for Real-World Image Super-Resolution. IJCV2024 R4: Controlling vision-language models for multi-task image restoration.ICLR2024. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I still have some concerns. 1. In Line 146-147, the authors claim "Following Daclip-ir [20], we utilized the pre-trained image encoder in CLIP [35]". This led me to mistakenly believe that the author was using the original OpenAI CLIP. 2. “However, for some images, their dimensions cannot be evenly divided by 2 during the forward process. In such cases, we use a tiled mode to test these images, as shown in the JPEG degradation image in Figure 3.” For low-level vision models, this is a common issue, and the most commonly used operation is to pad the image (e.g., Restormer, NAFNet, Uformer). Why use a tiled mode? When testing your proposed model, how do you handle this kind of situation? --- Reply to Comment 1.1.1: Comment: Thanks for your reply! 1. Regarding the use of the pre-trained encoder, we describe it in the "Training and Inference Procedure" section of the paper, specifically on L195. We apologize for any misunderstanding caused by this sentence in L126. We will revise this sentence to make our description clearer and avoid misunderstandings. Thanks for your suggestions again. 2. Since the JPEG-compressed test set contains only 29 images, we checked the size of each image and separately tested those with dimensions not divisible by 2 using a tiled mode in Restormer. Additionally, in the tiled mode, we used the grid function in BasicSR, which is a commonly used method similar to padding. We hope our response can address these concerns.	Summary: This paper proposes to perform universal image restoration via multiple low-rank adaptation. The key idea is to leverage a pre-trained stable diffusion model as the shared component and transfer it to specific degradations with LoRA adaptation. A degradation-aware router is further proposed to generate weights for LoRA combination based on degradation confidence. In experiments, the authors evaluated their method on multi-degradation and mixed-degradation datasets and conducted several ablation experiments on their core components. Strengths: - The idea of applying LoRA to a pre-trained SD for multi-task image restoration is promising and interesting. - The overall presentation is easy to follow. - The experimental results are good and the ablation studies make sense. Weaknesses: - ControlNet is the most popular approach to adapting SD models to other tasks. I'm curious why the authors chose LoRA? As far as I know, LoRA is often used for large language models (with billions of parameters). It would be great to provide more detailed motivation in the introduction. - In line 123, maybe it's better to use "concatenate" or other operators instead of "add" to present the unified parameters. Here, the weight $s_k$ can be ignored. - Can the authors use other SD models as the base model? I believe applying LoRA to a multi-step diffusion process can further illustrate its efficiency. - In Eq. (4), $s_0 \cdot M_k$ is used in both numerator and denominator, which seems weird and confusing. - The mixed degradation experiment is cool. It would be interesting if the authors could apply their model to real-world degraded images. - Line 45: proposed -> propose Technical Quality: 2 Clarity: 3 Questions for Authors: In the degradation-aware router, have you finetuned the CLIP to align degraded images with correct degradation names? How do you choose the degradation names as the vocabulary bank? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: See Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are truly grateful for your positive feedback on our work. 1. ControlNet is used in DiffBIR, and it adds a single encoder to handle various degradations, but its performance is still limited by task conflict. However, LoRA can be applied to any layer of a pre-trained model with a small number of parameters and works very well. This is why we chose LoRA in our approach. 2. The “+” in L123 corresponds to the parameter merge notation. In practice, we use feature weighting. We will take your suggestion into account and make this notation more accurate. 3. Several pre-trained generative models, such as SD1.5, SDXL, SD-turbo, and others, can serve as the base model for our approach. Although using multi-step pre-trained models theoretically offers stronger generative capabilities, we chose the one-step SD-turbo considering efficiency. 4. In Eq. (4), we use the letter 'o', not the number '0'. 5. In Section 4.4, we used two datasets, REDS and LOLBlur. REDS contains real-world degradations, while LOLBlur, although synthetic, is simulated based on real imaging processes and its data is consistent with real scenes. 6. Thank you for the detailed review of our writing. We will correct this error. 7. In our experiments, we used the trained DA-CLIP. Our vocabulary bank includes ten types of trained degradations. --- Rebuttal Comment 1.1: Comment: I want to thank the authors for their explanations in the rebuttal. Most of my concerns are addressed thus I will keep my original score. --- Reply to Comment 1.1.1: Comment: Thanks for your reply and positive rating, it means a lot to us.	Summary: This submission proposes a transfer-learning based strategy to address challenges related to image-degradation restoration. The premise is that a pre-trained generative model can be employed as a common starting component for multiple degradation types, upon which distinct sets of trainable parameters (ie. low-rank adaptors) can be added in order to address specific-degradation restoration tasks. Mixed-degradation restoration is enabled through a top-K hyperparameter, that affords a mixture of (degradation) experts to be active. The experimental setup considers multi and mixed image restoration problems where average results are offered across image-degradation datasets and appropriate standard quantitative metrics, qualitative examples, are reported in comparison with alternative approaches. Strengths: * The technique described for piping specific samples down specific low-rank adaptor chutes is relatively easy to understand and yet reportedly results in competitive restoration accuracy for investigated datasets. * Nascent investigations into mixed-degradation image restoration problems provide a promising seed to be followed. * The writing is of a reasonable standard. Weaknesses: * The key idea of leveraging pretrained VLM features (and specifically CLIP) for the task of image restoration from multiple degradations, pre-dates the current submission [R1]. While authors clearly go to some length to highlight their alternative CLIP-based scheme, which amounts to envoking specific (pre-existing [R2]) low-rank adaptors, the core technical contributions here can be regarded as somewhat limited. * The phrase 'Universal Image Restoration' may not be a sufficiently accurate (or modest) description for the proposed method. The submission collates ten different image restoration tasks which, despite vague statements in the abstract, remains a 'multi-task' not a 'universal' setup. Samples for all ten degradation tasks are shared between train and test (Sec. A.1) and individual task adaptors appear to be trained independently on task-specific datasets (L188--196). Generalisation ability to previously unseen degradations is also not considered. Suggest method description requires reworking. * The claim that multi-task learning (MTL) frameworks, designed to handle image restoration for multiple degradations, share all parameters across different degradations (L029) is incomplete and somewhat misleading. Several existing MTL works (eg. [R3,R4]) make use of both shared and task-specific parameter subsets for multiple image restoration tasks. Indeed 'which proportion of parameters should be shared and which should be task specific' can be considered a fundamental (and long standing) MTL question. The idea of benefiting from commonalities between image restoration tasks is well understood and my concern is that this casts doubt on a core premise of the submission. References R1. Controlling Vision-Language Models for Multi-Task Image Restoration. ICLR 2024. R2. LoRA: Low-rank adaptation of large language models. ICLR 2022. R3. All in One Bad Weather Removal using Architectural Search. CVPR 2020. R4. Pre-Trained Image Processing Transformer. CVPR 2021. Minor: L076: 'draining' --> 'deraining' L099: 'mim' --> 'min' L238: 'aspects' --> 'aspects.' Technical Quality: 2 Clarity: 2 Questions for Authors: > 'for mixed degradations, a larger K value is required to handle the more complex situtation' (L264). Can additional results be provided for alternative hyperparameter settings (eg. K=1 and K=10) in Tab.2, towards evidencing this claim? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Half of one sentence (L293) is apportioned to discussing method limitations. See above for suggestions on components that might make for valid additions here. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your thorough review and valuable feedback. 1. Core technical contributions: The core idea of our method is to introduce the paradigm of multi-domain transfer learning into multi-task image restoration, which aims to address the issues of task conflict and feature sharing in multi-task image restoration. This paradigm has not yet been applied in the field of image restoration. Along this line, we focused more on the overall framework rather than technical differences because we believe that the multi-domain transfer learning perspective presented in this paper can inspire more multi-task image restoration methods in the future. 2. Method description: Many papers in this field use terms like 'all-in-one'[R1, R2] and 'universal'[R3]. Following these works, we also used the term 'universal.' As you mentioned, multi-task image restoration is more accurate and we will revise this in our paper. 3. Claim in introduction: Thank you for your suggestion. We will revise the claim in L29 about 'sharing all parameters.' Our intention was to describe how inappropriate parameter sharing can be one of the sources of task conflict, without affecting the validity of the problem itself or our method. Essentially, our method also addresses how to allocate shared versus task-specific parameters. However, the difference is that we provide such an allocation scheme from the perspective of multi-domain transfer learning. 4. Thank you for the detailed review of our writing. We will correct these errors and thoroughly check our paper to ensure there are no other mistakes. 5. Additional results: The “mixed degradation” column in Table 3 uses the LOLBlur data from Table 2. Each image in LOLBlur contains at least two types of degradation, so the results for K=2 (Top-2) and K=10 (All) are better than K=1 (Top-1), as shown in the “mixed degradation column” in Table 3. Additionally, we synthesized more mixed degradation data as the reviewer nZDS's request. The impact of the hyperparameter K on the results is shown in T1. When the input image has mixed degradations, a larger K will result in better restoration performance. T1: \| Hazy-Blurry-Boisy \| PSNR \| SSIM \| LPIPS \| FID \| \| ----- \| -------- \| -------- \| --------- \| ------- \| \| Top-1 \| 15.24 \| 0.370 \| 0.825 \| 230.51 \| \| Top-2 \| 15.31 \| 0.373 \| 0.818 \| 230.26 \| \| Top-3 \| 15.33\| 0.374 \| 0.817 \| 230.16\| \| All \| 15.33 \| 0.374 \| 0.817 \| 230.17 \| R1. All-in-one image restoration for unknown corruption. CVPR2022. R2. Promptir: Prompting for all-in-one image restoration.NeurIPS2023. R3. Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model.CVPR2024. --- Rebuttal 2: Title: Official Comment by Reviewer 4LsM Comment: I thank the authors for the response and address their reply point-by-point. 1. The author rebuttal states that the core technical contributions did not focus on 'technical differences'. For the reviewer, this remains a confusing strategy for the focus of technical contributions. The crux of my concern is that task conflicts and feature sharing are well understood concepts and the rebranding of method terminology ('multi-domain transfer learning') has not added significant additional problem insight, for me. 2. The authors seem to concede that 'multi-task image restoration' is an accurate description of the work. This appears to further blur any distinction on method terminology. 3. Assuming the introduction claim is appropriately amended, I can consider this concern well addressed. 4. I can consider this concern well addressed. 5. I can consider this concern somewhat addressed and appreciate the additional experimental work. The effect of K in mixed degradation scenarios appears fairly underwhelming (ie. are we near the noise level?). Further I remain unconvinced that PSNRs of circa 12--15 on unknown degradations ('smoky') and mixed scenarios ('Hazy-Blurry-Noisy') are evidencing what might be considered succesful and pragmatic method performance. In sum, authors address a subset of my concerns however core issues remain problematic. I therefore do not increase my score. --- Rebuttal 3: Comment: Thanks for the response from the reviewer. But we cannot agree with the reviewer's response. 1. Core contribution: We respectfully disagree with the reviewer's view that this is a rebranding of method terminology. We introduced the "multi-domain transfer learning" paradigm primarily because pre-trained generative models inherently have shared image priors across different image restoration tasks, which current "all-in-one" or "multi-task" image restoration methods have neither proposed nor utilized. Additionally, using parameter-efficient adapters to capture the differences between various degradation restoration tasks and avoid task conflicts is also a novel approach, which has not been explored in the image restoration field. If the reviewer considers our method to be a "rebranding of method terminology," please provide the references on which this innovative criticism is based. Otherwise, this claim is not convincing. 2. Method description: "Multi-task image restoration" is a more appropriate term; however, aside from Daclip-IR, other papers with the same setting currently use the terms "all-in-one" or "universal." We have simply followed the common usage but are also considering adopting the more accurate term "multi-task image restoration," as suggested by the reviewer. The reviewer claims that we only trained on 10 types of degradation, but we also tested on mixed degradation and included experiments with unknown degradation during the rebuttal period. Whether in terms of the types of degradation or the complexity of degradation, we have already surpassed all-in-one methods like AirNet and PromptLR. We have covered the diversity of image restoration tasks as comprehensively as possible. From this perspective, why should it not be considered "universal"? 5. Additional results: First, please do not overlook the scenarios in which 𝐾 is used in our method. It addresses mixed degradation, which involves data distributions that were not encountered during training, and this is inherently a challenging problem. We are unsure where the reviewer’s term "near the noise level" comes from. Could you please provide more specific details? If the reviewer is referring to the data we tested, please note that our method shows improvements across four metrics and has been tested on multiple images, which avoids fluctuations based on a single image or a specific metric. Additionally, the data used in the "smoky" and "Hazy-Blurry-Noisy" experiments are also out-of-distribution. Existing image restoration methods generally struggle with this case, and even many methods have not considered such a case. Please also do not overlook the performance improvements of our method compared to other SOTA methods. --- Rebuttal 4: Comment: I thank the authors for their further thoughts and the discussion. 1. On core contribution - I appreciate the authors' reply around using shared image priors across different image restoration tasks. Apologies if the use of the term "rebranding" was clumsy, I intend to communicate that any distinction with a core strategy of augmenting a pre-trained model with task-specific components remains somewhat subtle for this reviewer. Multi-task methods have indeed made use of 'shared image priors' for some years (e.g. [R1]), for multiple image restoration tasks. I consider the novelty of core-strategy somewhat modest, however the combination of method components do indeed differ. 2. I agree that authors experimentally cover a diverse set of image restoration tasks. On "universal" terminology; this is largely a matter of taste. If we define the word as "applicable to all cases", then one might expect similar ID / OOD performance. The comment was originally designed to help the authors with reader expectation management. I appreciate that opinions may diverge on this. 3. My concern here is largely about the statistical power of the effect size relating to K. Additionally; if we are considering PSNR deltas on the order of 10^-2, are these differences qualitatively distinct in image space? meaningful? I find the evidence relating to 'choice of K' claims not overly convincing. I acknowledge that OOD scenarios are challenging (see previous comment) and that method performance improves c.f. alternatives however strong performance, in absolute terms, would seem to evade all methods under comparison. In sum the authors and I seem unable to reach agreement on a subset of points, however these may not be pathological in nature. I'm happy to somewhat modify my score to reflect this. References R1. Chen et al. 2021. Pre-trained image processing transformer. CVPR 2021. --- Rebuttal 5: Comment: Thanks for your patient discussion. We welcome the reviewer to discuss our paper based on fairness and objectivity. 1. Although IPT[R1] has pre-training and fine-tuning stages, it differs significantly from our method. IPT uses synthesized degraded image pairs and learns the mapping from degraded images to clear images during the pre-trained stage. This involves incorporating priors about the degradation process, which is different from the image distribution prior used in the generative models of our method. Additionally, during the fine-tuning stage, the shared parameters are still updated while our pre-trained generative model is frozen. This means that it is necessary to reload the pre-trained parameters in IPT and retrain the parameters when testing different degraded images. Our pre-trained generative model is frozen, meaning the shared parameters have the same weights when testing different degradation types. In contrast, IPT’s shared parameters have different weights during testing. With the fine-tuning stage and the need for explicit degradation categories, this method cannot even be classified as an all-in-one approach. This is why AirNet[R2] claims that they are the first to propose the all-in-one restoration task. 2. Thanks for the reviewer's suggestions. The core issue with using terms like "multi-task," "all-in-one," or "universal" is how to define the range of degradation types included in the term "all." Currently, this range is not clearly defined in existing papers. Based on the response, the reviewer agrees with this reason. While we believe "multi-task" is a suitable term in this field, the experiments we conducted also support our claim of universal image restoration. And considering the "reader expectation management" mentioned by the reviewer, we will revise this description as we promised in the original rebuttal. 3. If the reviewer thinks the performance improvement of our 𝐾 strategy is minimal, it will be better if they can refer to the ablation studies of existing restoration methods, such as HInet[R3] and Restormer[R4]. For example, in Restormer, each innovative module has an improvement in PSNR between 0.05 and 0.2. Additionally, the reviewer compares the results of 𝐾=1 and 𝐾=2. But the model with 𝐾=1 is already an improved version of our method, not the "baseline." Our model with 𝐾=1 can handle the primary degradation in the image, and increasing 𝐾 aims to address secondary degradations in case of mixed degradation. We will add a comparison of the processing results when K has different values in the paper if we have the opportunity to submit a camera-ready version. R1. Pre-trained image processing transformer. CVPR 2021. R2. All-in-one image restoration for unknown corruption. CVPR2022. R3. Hinet: Half instance normalization network for image restoration. CVPR2021. R4. Restormer: Efficient transformer for high-resolution image restoration. CVPR2022. --- Rebuttal 6: Comment: I thank the authors for their further clarifications and robust discussion. 1. The additional dialogue on distinction to previous work, with regard to incorporation-of-priors strategy, is helpful and further alleviates my concerns on this point. 2. I think we have (somewhat) converged here. 3. I agree with the authors that qualitative examples will likely further aid reader understanding on method sensitivity to this hyperparameter and would welcome such additions. As previously noted, my remaining concerns can be considered smaller and I believe my rating, confidence scores now reflect this accurately. --- Rebuttal Comment 6.1: Comment: Thank you for your discussion. Your positive rating is meaningful to us.	Summary: The paper proposes universal image restoration framework using multiple low-rank adapters that learns task specific weights from to perform multi-domain transfer learning. the proposed method leverages the pre-trained generative model weights as the shared component and adapts it task specific low-rank adapters. At each layer in the restoration pipeline the proposed method uses the degradation similarity to combine LoRA adapters outputs, this enables the proposed to handle for mixed degradation restoration. Strengths: - The paper proposes LoRA adapters to learn task specific weights and proposes a strategy to combine the adapter outputs using degradation similarity measure - extensive experiments are performed showing the proposed strategy works better than random and average in table 3. - extensive experiments are performed to show the proposed methods performance against the sota methods in table 1 for mutliple degradation task. - Extensive experiments are performaed showing impact of LoRA rank and prediction accuracy Weaknesses: - In table of the paper authors compared proposed method against sota on REDS and LOLBlur datasets, both these datasets have mixed degradations of blur, jpeg compression, noise, and low light. Although these comparisons performed on mixed degradations, it would be helpful to how the proposed method performs on mixed weather conditioned images (MID6), which is comparatively challenging than REDS and LOLBlur datasets. MID6: Multimodal Prompt Perceiver: Empower Adaptiveness, Generalizability and Fidelity for All-in-One Image Restoration, CVPR, 2024. - Can authors confirm, whether network re-trained seperately for each experiment in table-1, and table-2 separately, i.e. table-1 and table-2 trained network weights for proposed method are different. Technical Quality: 3 Clarity: 3 Questions for Authors: - Can the proposed network handle unknown degradation present in the input degradation image - From the table-3, it is evident that Top-1, and top-2 are almost same performance as All, can authors coment on this, this makes wonder if the input image has only one degradation as dominating for this experiment. Can authors show this experiment on different dataset like MID6 Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - authors have addresed limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your valuable suggestions and for acknowledging our work. 1. Why REDS and LOLBlur？ We used the REDS and LOLBlur mixed datasets because the mixed degradation scenarios in these datasets are common, whereas the mixing in MID6 is not commonly seen in real-world scenarios. Since MID6 has not released its data, we simulated the mixing of three types of degradations based on the imaging process and compared our method with several representative image restoration methods. The results are shown in T1. T1: \| Smoky \| PSNR \| SSIM \| LPIPS \| FID \| \| ------------- \| --------- \| ------------- \| --------- \| ------------- \| \| Restormer \| 10.54 \| 0.418 \| 0.604 \| 265.99 \| \| PromptIR \| 10.62 \| 0.421 \| 0.601 \| 263.61 \| \| Daclip-IR \| 10.97 \| 0.420 \| 0.564 \| 221.98 \| \| Ours \| 12.71 \| 0.476 \| 0.513 \| 185.42 \| 2. Network weight: The network used in Table 2 was tested directly after being trained as shown in Table 1. This means that the networks for testing in Table 1 and Table 2 use exactly the same parameters and weights, which is intended to validate the performance of our method on mixed degradation data. 3. Unknown degradation: The REDS and LOLBlur datasets we used contain mixed degradations that are common in real-world scenarios and are also outside the distribution of the training set. From the results in Table 2, we can see the advantages of our method. Additionally, we tested the performance on the unknown degradation - smoky. The comparison with representative restoration methods is presented in T2, showing that our method has a performance improvement compared to Restormer, PromptIR, and Daclip-IR. T2: \| Hazy-Blurry-Boisy \| PSNR \| SSIM \| LPIPS \| FID \| \| ----- \| -------- \| -------- \| --------- \| ------- \| \| Restormer \| 13.39 \| 0.074 \| 1.312 \| 328.97 \| \| PromptIR \| 13.61 \| 0.074 \| 1.304 \| 330.60 \| \| Daclip-IR \| 14.50 \| 0.181 \| 1.028 \| 275.37 \| \| Ours \| 15.33\| 0.374 \| 0.817 \| 230.16\| 4. Table 3 explanation and additional experiment: When the image has only one type of deterioration, the “top-1” strategy and “top-2” strategy perform similarly, as indicated in the “Multiple Degradation” column of Table 3. However, when the degraded image has more than one type of degradations, the “all” strategy and the “top-2” strategy outperform the “top-1” strategy, as illustrated in the “Mixed Degradation” column of Table 3. We conduct the experiment on the data of “Hazy-Blurry-Noisy” like MID6, As shown in T3, since the images have three types of degradations, the “top-2”, “top-3”, and 'all' strategies are superior to the 'top-1' strategy. T3: \| Hazy-Blurry-Boisy \| PSNR \| SSIM \| LPIPS \| FID \| \| ----- \| -------- \| -------- \| --------- \| ------- \| \| Ours (Top-1) \| 15.24 \| 0.370 \| 0.825 \| 230.51 \| \| Ours (Top-2) \| 15.31 \| 0.373 \| 0.818 \| 230.26 \| \| Ours (Top-3) \| 15.33\| 0.374 \| 0.817 \| 230.16\| \| Ours (All) \| 15.33 \| 0.374 \| 0.817 \| 230.17 \| R1: MID6: Multimodal Prompt Perceiver: Empower Adaptiveness, Generalizability and Fidelity for All-in-One Image Restoration, CVPR, 2024. --- Rebuttal 2: Title: Reminder for review Comment: Dear Reviewer nZDS, I have noticed that you have not yet responded to the authors' rebuttal. I kindly urge you to engage in a discussion with the authors at your earliest convenience to help advance the review process.	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper presents a framework to improve image restoration across various degradation types using Low-Rank Adapters (LoRA). The proposed method adapts a pre-trained generative model to each degradation type. It performs a weighted sum of the output of adapted models using the estimated degradation of input images. The proposed method performs impressive results in restoration accuracy and resources. Strengths: The proposed method is interesting and reasonable. Experimental results support this paper's contributions and the proposed method's effectiveness. Weaknesses: In Table 3, the 'Top-1' strategy performs almost the same as the 'All' strategy, which limits the motivation of the weighted sum of the adapted models. Table 6 presents the restoration performance comparisons for each degradation. The proposed method underperforms previous works in significant degradation types such as blurry, low-light, raindrop, and rainy. The average scores might mislead the evaluation performances. Technical Quality: 4 Clarity: 4 Questions for Authors: How about comparing the proposed method with the following paper? Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model, CVPR 2024 Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The proposed method is simple and effective, but evaluating average scores on multiple degradations can mislead its contribution. The proposed method achieves near-best performance by selecting a single adapted model but underperforms in many major degradation types. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your thorough review and valuable feedback. 1. Motivation of the weighted sum: When the image has only one type of deterioration, the “top-1” strategy and “all” strategy perform similarly, as indicated in the “Multiple Degradation” column of Table 3. However, when the degraded image has more than one type of degradation, the “all” strategy outperforms the “top-1” strategy. Our motivation for utilizing weighted sum is to enable the model to perform better in cases of mixed degradation. 2. Why average scores? : It is difficult to accurately assess the quality of an image using the PSNR metric alone, as shown in Table 6. Therefore, we also employed SSIM, LPIPS, and FID metrics, as shown in Tables 7, 8, and 9. From the comparison results in Tables 6, 7, 8, and 9, we can see that for the blurry, raindrop, and rainy tasks, our method achieves the best score on at least one or more of the four metrics. Despite the LOL dataset including just 15 images, our metrics are the second best among all the compared methods. Additionally, our focus is on multi-degradation image restoration tasks and average metrics provide a concise and comprehensive measure of multi-task image restoration models' performance, which is a common practice in this field, as seen in Table 3 of DA-CLIP. 3. Additional comparison: We retrain the method of [R1] and test it using the multiple degradation dataset and the mixed degradation dataset. The results are displayed below. It is evident that our method outperforms R1 on the four metrics as well as on the average metric. T1: Multiple Degradation: \| \| Blurry \| Hazy \| Jpeg \| Low-light \| Noisy \| Raindrop \| Rainy \| Shadowed \| Snowy \| Inpainting \| Avg-R1 \| Avg-Ours \| \| :--------- \| :----------: \| :---------: \| :---------: \| :-------------: \| --------- \| :------------: \| :---------: \| :------------: \| :---------: \| :--------------: \| :-----------: \| :-----------: \| \| PSNR \| 26.13 \| 27.66 \| 27.29 \| 19.42 \| 27.69 \| 30.07 \| 27.31 \| 26.39 \| 27.38 \| 18.85 \| 25.82 \| 28.08 \| \| SSIM \| 0.821 \| 0.937 \| 0.792 \| 0.822 \| 0.771 \| 0.913 \| 0.828 \| 0.850 \| 0.884 \| 0.769 \| 0.838 \| 0.864 \| \| LPIPS \| 0.238 \| 0.042 \| 0.276 \| 0.159 \| 0.231 \| 0.068 \| 0.189 \| 0.112 \| 0.088 \| 0.251 \| 0.165 \| 0.104 \| \| FID \| 35.43 \| 9.32 \| 70.99 \| 63.41 \| 87.24 \| 32.24 \| 77.54 \| 29.84 \| 29.82 \| 143.66 \| 57.95 \| 30.58 \| T2: Mixed Degradation: \| \| REDS \| \| \| \| LOLBlur \| \| \| \| \| -------- \| :--------------: \| :--------------: \| :--------------: \| :--------------: \| :--------------: \| :--------------: \| :--------------: \| :--------------: \| \| \| PSNR \| SSIM \| LPIPS \| FID \| PSNR \| SSIM \| LPIPS \| FID \| \| R1 \| 24.93 \| 0.710 \| 0.413 \| 129.17 \| 15.77 \| 0.612 \| 0.380 \| 65.70 \| \| Ours \| 25.11 \| 0.718 \| 0.315 \| 89.79 \| 18.16 \| 0.690 \| 0.318 \| 61.55 \| R1. Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model, CVPR 2024 --- Rebuttal Comment 1.1: Title: Thanks for the rebuttal Comment: I appreciate the authors' efforts for the rebuttal. However, the response does not alleviate my concerns. In the case of mixed degradation, the performance difference between the Top-1 and All strategies is only 0.12 dB, and there is little performance difference between the All and Top-2 strategies. Additionally, there is insufficient motivation for the restoration tasks differently from those used in existing universal restoration models. I will maintain my original score. --- Reply to Comment 1.1.1: Comment: We appreciate the reviewer's response, but it is hard for us to agree with your subsequent comments, particularly regarding "those used in existing universal restoration models." Based on the reviewer's initial feedback, we have added the additional method for comparison and explained the performance of different strategies such as "top-1," "top-2," and "all." Regarding the new concern of "insufficient motivation" raised by the reviewer, our paper already provides a detailed discussion. We have also further discussed the innovative aspects raised by Reviewer 4Lsm. Please refer to that. If the reviewer believes our method is difficult to distinguish from existing universal restoration models, please provide specific references and indicate which aspects are insufficient. Otherwise, it is not convincing and we can't accept this comment.	null	null	null	null	null	null
FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction	Accept (poster)	Summary: The paper introduces a novel approach to address the challenge of collaboratively visualizing high-dimensional data in a federated learning (FL) environment. The proposed method, FEDNE, integrates the FEDAVG framework with contrastive neighbor embedding (NE) techniques, aiming to preserve data privacy while ensuring effective data visualization. By employing a surrogate loss function and an intra-client data mixing strategy, FEDNE seeks to enhance the alignment and preservation of neighborhood structures in the global embedding space. The paper includes comprehensive experiments on both synthetic and real-world datasets, demonstrating the effectiveness of FEDNE in outperforming several baseline methods in terms of neighborhood data structure preservation and clustering. Strengths: 1. FEDNE introduces a novel integration of FEDAVG with contrastive NE techniques, addressing the unique challenges of pairwise data relationships in federated learning environments without requiring data sharing. 2. The intra-client data mixing strategy effectively enhances local data diversity, mitigating the limitations of biased local kNN graphs and ensuring better neighborhood representation. 3. The paper provides a thorough evaluation of FEDNE using various datasets and metrics, showcasing its superior performance compared to baseline methods in preserving neighborhood structures and clustering. Weaknesses: 1. While the authors mention that FEDNE introduces only 35% more GPU time compared to FEDAVG, the overall complexity and scalability in a more extensive, real-world setting are not fully addressed. The authors should further investigate how FEDNE scales with a significantly larger number of clients and more complex datasets or models. 2. The paper proposes intra-client data mixing as a solution to the bias in local kNN graphs. However, this approach might not entirely mitigate the issue of incorrect neighbor connections, especially in highly imbalanced datasets. More detailed comparisons with alternative methods or further enhancements could provide a more robust solution. 3. The focus is primarily on dimensionality reduction. The validation results are performed only on the vision classification tasks. Extending the discussions and analyses to include applications in other domains could be beneficial. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. Could you provide more details on the process of training the surrogate models? Specifically, how do you ensure that these models effectively capture the repulsive forces between dissimilar data points across different clients? 2. Non-IID data is a common challenge in federated learning. How does FEDNE handle extreme cases of non-IID data distribution? Have you considered any additional mechanisms to ensure robustness in such scenarios? 3. How sensitive is FEDNE to the choice of hyperparameters, such as the step size for grid sampling, the number of neighbors in kNN, and the weight in intra-client data mixing? Have you performed any sensitivity analysis? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: The authors have addressed the works limitations and social impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Questions regarding scalability and complexity Please see the general response. > The paper proposes intra-client data mixing … However, this approach might not entirely mitigate the issue … More detailed comparisons with alternative methods … Thank you for the insightful comment. As we pointed out in the paper, the bias in the local kNN graphs is one key blocker to the problem of Federated NE. Dealing with attraction term is more challenging than repulsion term as it requires knowing true data affinity across different clients (as discussed in Sect. 4.2). Such a problem has never been pointed out in prior work, but as emphasized in our paper (Line 209-217). With this in mind, we made one of the first attempts to mitigate it. We acknowledge that our approach may not fully address the problem, but it demonstrates that reducing its effect can notably improve the overall performance (see Sect.5.3), setting up the foundation for future work to inspire more research in this new field including but not limited to addressing the attractive loss within the settings of FL. We appreciate your comments on “alternative methods.” To our knowledge, none of the prior work in FL for dimensionality reduction aimed to address this issue. We would appreciate if you could provide references to the alternative methods you mentioned, and we will be happy to provide more discussions in the reviewer/author discussion phase. > … The validation results are performed only on the vision classification tasks … other domains could be beneficial. Thanks for your valuable comment. We totally agree that the evaluation of our method is mainly on the vision datasets. The main reason we designed the experimental study in this way is mainly to follow a similar paradigm as other DR studies [33,44,55], with the purpose of fair comparisons. However, we admit that only evaluating vision data might be not enough. Therefore, in addition to the commonly used benchmark vision data, we also evaluated our method on a biological dataset, scRNA-Seq (see Tables 1 and 2 in the original manuscript). The scRNA-Seq dataset is a real-world dataset containing a collection of gene expression profiles obtained from individual cells of the mouse retina. We hope this additional biological dataset can give you a more comprehensive understanding of our effectiveness in different domains. [33] “UMAP: Uniform manifold approximation and projection for dimension reduction.” arXiv preprint arXiv:1802.03426, 2018. [44] "Visualizing data using t-SNE." JMLR 9.11 (2008). [55] "SpaceMAP: Visualizing High-Dimensional Data by Space Expansion." ICML. 2022. > Could you provide more details on the process of training the surrogate models? Specifically, how do you ensure that these models effectively capture the repulsive forces between dissimilar data points across different clients? Thanks for the comment. We have discussed in Sect. 4.1 and A.2 in the appendix, but we should have made this more clear. To train the surrogate repulsion model for a client $m$, we generate a set of low-dimensional (low-D) query points via grid sampling to serve as potential embedding positions of other clients’ data points. Then, for each newly sampled low-D query data $z_q$, we pre-compute the repulsive loss between $z_q$ and $b$ data points sampled within client $m$, which serve as the training targets. The surrogate model is expected to learn a mapping from these newly sampled embedding positions to their corresponding repulsion loss, as measured within client $m$. According to the repulsive loss term introduced in Section 3.1, a larger low-D Euclidean distance will result in a smaller repulsive loss value [6]. Thus, the repulsion loss decreases as the distance between the data pairs increases. If two embedding points are close within a threshold, our surrogate model is able to approximate the repulsion force; if the points are far apart, the model estimates the repulsion force between them to be close to zero. [6] "Attraction-repulsion spectrum in neighbor embeddings." JMLR 23.95 (2022): 1-32. > How does FEDNE handle extreme cases of non-IID data distribution? Have you considered any additional mechanisms to ensure robustness in such scenarios? We agree that the ability to handle non-IID data is very important. Therefore, we have conducted experiments on the settings of Dirichlet(0.1) and Shards with C=2 to demonstrate the effectiveness of FedNE in handling extreme non-IID cases, as shown in Tables 2 and 6 in the original manuscript. FedNE *addresses non-IID data distribution exactly* via the surrogate models*. For example, under an extreme* non-IID condition, each client may have data exclusively for certain clusters. In this scenario, the issue with the attraction term might be relaxed because the true neighboring points are likely to reside within the same client. However, in the global context, the dissimilar data pairs contributing to the repulsive loss term are located across different clients. As discussed in Sect. 3.3 and 4.1, our surrogate model is specifically designed for this condition by filling in the missing repulsion terms. > How sensitive is FEDNE to the choice of hyperparameters Thanks for the valuable suggestions. We have included experiments in Appendix B to analyze the sensitivity of some important hyperparameters in our method such as the frequency of surrogate function updates and the time to integrate the surrogate loss function into the local training. Furthermore, we want to provide additional results on analyzing the sensitivity for the following hyperparameters: (1) Number of neighbors in the kNN graphs, (2) Step sizes in the grid sampling, and (3) The weight in intra-client data mixing. Due to the space limit, we included the detailed analysis in the figure captions. Please find the results and analysis in the attached PDF. --- Rebuttal Comment 1.1: Comment: Thank you for your reply. I think the authors have addressed my concerns, especially with regard to the scalability and generality of the proposed method. Therefore, I decide to change the score to borderline accept. --- Rebuttal 2: Title: Re: Official Comment by Reviewer y2ZN Comment: Thank you for your feedback and your willingness to increase the rating. We will incorporate our rebuttal into the revised version of our paper, and we would appreciate it if you would be supportive of the acceptance of our paper.	Summary: The paper "FEDNE: Surrogate-Assisted Federated Neighbor Embedding for Privacy-Preserving Dimensionality Reduction" presents a method for visualizing high-dimensional data while maintaining privacy without requiring any shareable reference data. Federated Neighbor Embedding (FEDNE): A framework combining federated averaging (FEDAVG) with contrastive neighbor embedding (NE) to create a joint NE model across multiple clients without compromising data privacy. Surrogate Loss Function: An innovative loss function to enhance inter-client repulsion in the global embedding space, ensuring better separation of data points from different clients while preserving local data structures. Data-Mixing Strategy: A technique to counter issues like invisible and false neighbors in local k-nearest neighbor (kNN) graphs by mixing data from various clients during training, thus improving the quality of the learned embeddings. Strengths: Well-Presented: The paper is clearly and coherently written, making it easy to follow. Novel Approach: The study addresses an important problem with a novel approach, combining federated learning with neighbor embedding techniques. Weaknesses: Privacy Concerns: While the approach is innovative, the paper does not sufficiently address privacy concerns. It lacks experiments and guarantees demonstrating the privacy preservation of the FedNE approach. Computational Inefficiency: The method appears to be computationally inefficient. There are no experiments conducted on large datasets, such as those in real-world medical or other privacy-critical domains, where computational complexity could be a significant issue. Inadequate Analysis of Related Work: The related works section is not thoroughly analyzed or discussed, missing critical comparisons and context necessary for a comprehensive understanding of the state of the art. The study's applicability could be strengthened by extending beyond benchmark datasets to encompass real-world, privacy-sensitive datasets found in domains such as healthcare or finance. This expansion would provide a more robust demonstration of the method's practical relevance and effectiveness. Additionally, addressing pairwise issues associated with attraction terms is essential for improving the preservation of neighborhood structures and enhancing clustering quality. Furthermore, it is crucial to conduct thorough analyses aimed at optimizing the computational efficiency and scalability of the algorithms, ensuring their capability to handle large-scale datasets effectively. Moreover, the method currently lacks explicit consideration of privacy guarantees. And on elucidating how privacy concerns are addressed within the framework and formalizing privacy guarantees to assure users and stakeholders. Technical Quality: 2 Clarity: 3 Questions for Authors: Privacy Guarantees: The paper lacks a thorough discussion on the privacy guarantees of the proposed method, especially against adversarial attackers.The experimental evaluation focuses solely on utility results, with no evaluation of data privacy. How is privacy preservation quantified and ensured? What is the acceptable level of privacy preservation? The paper should include theoretical arguments and experiments demonstrating actual privacy management. From a privacy perspective, it would be helpful to provide guidance on the limitations of this method, particularly regarding transparency and explainability (e.g., OECD AI principle 1.3). What measures are in place to address these concerns? Experiments: No experiments are conducted on downstream tasks related to the problem, aside from analyzing structural properties? Moreover, no experiments conducted on large datasets? Lastly, how do the authors plan to tackle the problem of heavy computation? The method appears computationally intensive, which could hinder its practicality. Reconstruction from Gradients: According to Zhu and Han’s study [1], model gradients can be used in some scenarios to partially reconstruct client data. How does the proposed method address this issue? The paper claims to operate without relying on shareable reference data, yet it utilizes additional 2D data points from grid sampling to estimate the training targets via repulsion loss and additional augmented data points via interpolation for the attractive loss. Given this, how does the strategy address the significant computational burden it introduces, and is it feasible for real-world applications where computational efficiency is critical? [1] Ligeng Zhu, Zhijian Liu, and Song Han. 2019. Deep leakage from gradients. Advances in neural information processing systems 32 (2019) Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: No societal impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Questions related to privacy concerns We found that major concerns mentioned in the weakness and questions are related to “privacy-preserving”, and these concerns may arise from the “privacy-preserving” term in our paper title. First, we want to apologize for the confusion caused by our title. The main focus of our paper is on a FL setting, aiming to bring up the unique FL challenges and propose effective solutions for the pairwise training objective in the problem of Federated Neighbor Embedding. Besides the privacy-preserving properties that a Federated setting has introduced, we do not specifically develop any privacy-preserving mechanisms and we certainly do not claim that our approach introduces additional privacy guarantees. We will surely refine our title and any writing to clarify this. Introducing formal privacy guarantees is not a trivial task. Among the only two existing methods, dSNE [37] proposes a decentralized neighbor embedding framework and later, they extend dSNE to be privacy-guaranteed by incorporating differential privacy (DP) as F-dSNE [38]. We have to admit that, developing privacy-guaranteed Federated NE methods while addressing the lack of inter-client data relationships is an important but challenging task. Therefore, we will consider the privacy of FedNE as an orthogonal problem and will work toward incorporating DP or other techniques in our future work. We will add discussions on how to enhance FedNE with privacy-preserving techniques. > Inadequate Analysis of Related Work Thanks for the valuable comment. As our work intersects with two research areas: FL and DR, we have structured the related work section into three subtopics in Sect. 2. In fact, we have compared FedNE with two closely related works, dSNE and F-dSNE in Sect. 2. However, due to the page limitations, we might have overlooked some relevant literature and we are happy to include additional work on FL and DR. However, we kindly believe that our literature review in the field of decentralized DR is thorough. > The study's applicability could be strengthened … to encompass real-world, privacy-sensitive datasets … Thanks for the constructive suggestions. As clarified in our above response, our main focus is on FL for DR instead of particularly improving the privacy-preserving properties. Thus, we focus more on benchmark datasets in the field of DR. Specifically, we have included the standard datasets in the DR community but with more practical settings (i.e., more clients and larger datasets) compared to prior work (please see the general response). With that being said, your comments are well received, and we will consider extending our study to some real-world privacy-sensitive datasets. However, we do want to point out some considerations. Datasets from different domains have unique properties. For example, the prior work [37] mentioned that, in the neuroimaging domain, not all data is private/unshareable, and many public MRI datasets are accessible. The only existing works [37,38] related to our problem focusing on neuroimaging data have made this unreasonable assumption which makes their method hard to generalize. In contrast, to prevent loss of generalizability, our FedNE was developed without assuming any domain-specific properties. Furthermore, we believe that adapting FedNE to the domain-specific datasets should not be difficult as we did not make any prior assumptions about the data properties. > Questions regarding scalability, complexity, and computational efficiency Please see the general response. > The paper claims to operate without relying on shareable reference data, yet it utilizes additional 2D data points from grid sampling to estimate the training targets … Thanks for your comment. We want to reiterate that our FedNE does not use any shareable reference data, as the publicly available dataset is often inaccessible in many real-world applications (Line 112-114). Moreover, the quality of the reference data can significantly affect the performance [37]. Given these more practical considerations, we propose surrogate models and intra-client data mixing to address the constraints and challenges brought by the Federated setting. We think the computational cost is still manageable, and more importantly, our work is valuable as it was developed within a more practical and general FL setting, even with some extra burden from data sampling. > No experiments are conducted on downstream tasks … Thanks for your suggestion. We want to emphasize that the main focus of this work is to propose a DR framework within the FL setting. Therefore, most of our evaluation metrics are designed for analyzing the structural properties of the embeddings. This is because metrics like preservation of neighborhood relationships and preservation of data clusters can provide a clear indication of how well our DR model performs [57,58] regardless of whether it is under the federated or centralized setting. We totally agree that evaluating on other downstream tasks might be valuable to provide a different understanding of our model performance. However, the metric to evaluate other downstream tasks might not be able to accurately assess our capability of preserving the neighborhood relationship as a DR framework. Thus, we respectfully think conducting experiments on downstream tasks e.g., classification might be biased to assess our effectiveness in DR. [37] "See without looking: joint visualization of sensitive multi-site datasets." IJCAI. 2017. [38] "Federated, Fast, and Private Visualization of Decentralized Data." Workshop of Federated Learning and Analytics in Practice (2023). [57] "Toward a quantitative survey of dimension reduction techniques." IEEE TVCG 27.3 (2019): 2153-2173. [58] "Dimensionality reduction: A comparative review." JMLR 10.66-71 (2009): 13. [59] "Feature dimensionality reduction: a review." Complex & Intelligent Systems 8.3 (2022): 2663-2693. --- Rebuttal Comment 1.1: Title: Response to Author's rebuttal Comment: Thank you for your response. However, my concerns about the privacy aspects of the proposed method remain unresolved. Federated learning, by design, aims to decentralize and collaboratively train models in a privacy-preserving manner, which inherently implies a need for privacy. This is precisely why one would choose this approach. However, recent research suggests that federated learning on its own may not be sufficient to guarantee privacy. Therefore, I will maintain my score. --- Reply to Comment 1.1.1: Comment: Dear Reviewer mP1Q, We appreciate your timely feedback. We understand your point about the privacy concern of the general Federated Learning (FL) framework. To our understanding, FL was originally conceived to allow model training without centralized data collection. Such a framework enables users to keep their data without sharing. Since then, a significant portion of the work in FL has aimed to improve the performance of FL in terms of its accuracy and convergence, especially under heterogeneous settings and communication constraints. Meanwhile, another branch of the FL research focuses on the privacy guarantee as you mentioned, as they found that decentralizing the data alone is not sufficient to protect privacy. We think both branches (and several others, e.g., system-level consideration of FL) are valuable --- while one single paper may not be able to address both aspects, together they aim for a synergy to make FL a robust, widely applicable, and privacy-preserving framework. As highlighted in a recent survey [1], FL still faces multiple unresolved challenges or deficiencies, including privacy protection, communication costs, systems heterogeneity, and deepening the research of FL in various fields. In Section 7, the survey paper said, “The development of FL is faced with multiple challenges, and no single strategy can comprehensively solve these bottlenecks in the practical application of FL technology.” In our paper, we aim to extend FL's applicability (to a rarely studied ML problem) and address the associated performance challenges. We kindly want to reiterate that our focus is not to improve the general FL framework in terms of its privacy aspect but to explore a novel application domain of FL. We hope the above paragraphs clarify the position of our paper in the context of Federated Learning. [1] Wen, Jie, et al. "A survey on federated learning: challenges and applications." International Journal of Machine Learning and Cybernetics 14.2 (2023): 513-535. Best, Authors	Summary: The paper presents a new federated learning approach named FEDNE for dimension reduction using contrastive neighbor embedding (NE). The key idea is the introduction of a surrogate loss function that each client learns and shares, which compensates for the lack of inter-client repulsion essential for global alignment in the embedding space. Additionally, the paper proposes a data-mixing strategy to augment local data, addressing issues of invisible and false neighbors in local kNN graphs. Comprehensive experiments demonstrate that FEDNE effectively preserves neighborhood data structures and enhances alignment in the global embedding space compared to several baseline methods. Strengths: 1. The studied problem is important. There could be many downstream tasks after applying federated neighbor embedding. 2. Many metrics are included in the experiments to evaluate the quality of the resulting embeddings Weaknesses: 1. The paper lacks investigation on the effect of choice of hyperparameter k. 2. The improvement of FEDNE is significant on some metrics (e.g., kNN) but is very limited in other metrics (e.g., conti.). The paper lacks a detailed exploration of why FEDNE produces different behavior for different metrics. 3. I suggest to highlight the best results in Table 2. Currently the results of FEDNE are highlighted although it may not achieve the best performance in some cases. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. How would the parameter k affect the performance of FEDNE? How to set k for different settings? 2. What are the major differences between the metrics? Why the improvement of FEDNE differ a lot across different metrics? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > How would the parameter k affect the performance of FEDNE? How to set k for different settings? Thank you for the valuable comment. First, we want to reiterate that the value k is used for building local kNN graphs to capture the neighboring data structures. In general, as k increases, we may lose the neighboring information. When k is too small, it may result in many isolated clusters in the embedding space. However, there is no clear consensus on how to select k in the dimensionality reduction (DR) community. As many recent DR papers use fixed k values [11], we followed them as well. We experimented with different k values under the setting of Dirichlet(0.1) on the MNIST dataset with 20 clients. Please find the results in the attached PDF. We found that within a certain range (i.e., 7 to 30), the performance of FedNE is relatively stable. When k is too large (e.g., k=50), the performance drops but our FedNE still outperforms the baseline methods, FedAvg+NE. This trend aligns with the general understanding of DR methods. Once there is a best practice to select k for different settings, it can be applied to our problem. [11] “From t-SNE to UMAP with contrastive learning”. In ICLR, 2023. > What are the major differences between the metrics? Why the improvement of FEDNE differ a lot across different metrics? Thanks for your thoughtful comment. We had some discussions in Sections 5.1 and 5.2, but we should have made this more clear. Continuity, trustworthiness, and kNN classification accuracy are mainly used to evaluate the preservation of neighborhood data structures in the global reduced-dimensional space. Steadiness and cohesiveness aim to measure the preservation of inter-cluster structures since clusters can be distorted when projecting to the global low-dimensional space. Among the five metrics, trustworthiness, kNN classification accuracy, and steadiness have much more noticeable increases compared to continuity and cohesiveness. The improvement differs a lot among the metrics mainly because in a simple FedAvg framework without any special treatment for the data pairs across different clients, the model cannot penalize those data pairs to be apart, which results in introducing false neighbors in the embedding space (Line 215 and 292). These false neighbors can further cause data points from different classes to overlap in the embedding space. This explains why the baseline methods achieve much lower trustworthiness scores and kNN classification accuracy, as trustworthiness measures whether neighbors in the embedding space are also neighbors in the high-dimensional space. Due to similar reasons, without mitigating the incorrect neighbor connections and addressing the missing repulsion terms, the overlap in the embedding space will also mistakenly introduce false data clusters. Thereby, steadiness is also low for the baseline methods as it measures how well the model can avoid false clusters. However, both our surrogate models and intra-client data mixing strategy aim to prevent false neighbors in the embedding space. When true neighbors are preserved by FedNE, the trustworthiness and steadiness scores have increased and the points that are close in the embedding space are more likely to belong to the same class. Thus, the kNN classification accuracy is also notably increased (see Table 2 in the original manuscript). Continuity measures how well the neighborhood of a data point in the high-dimensional space is preserved in the embedding space. Achieving higher continuity is often easier even under the context of FL because the original attraction term in the client’s local objective function already pulls the neighbors closer in the embedding space. Cohesiveness is for measuring how well the projection model can avoid missing clusters. We still can observe relatively larger improvements in cohesiveness compared to continuity. Improved cohesiveness demonstrates that FedNE will not mistakenly break the true clusters in the embedding space. > I suggest to highlight the best results in Table 2. Currently the results of FEDNE are highlighted although it may not achieve the best performance in some cases. Thank you for the suggestion. We will change to highlight the best performance in the final version. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. I'll keep my positive score. --- Rebuttal 2: Title: Re: Official Comment by Reviewer 4ZYS Comment: We appreciate your feedback and your positive opinion about our paper. If our rebuttal has addressed your concerns, we would also appreciate it if you would be willing to consider raising your original rating. Thank you for your consideration.	Summary: This paper addresses the challenge of distributed neural embedding (NE) with a focus on privacy protection. To achieve this, the authors extend the concept of federated learning (FL) to NE. However, NE tends to diverge because FL prevents clients from accessing each other's data, leading to inconsistent feature spaces across clients. To mitigate this issue, the authors employ surrogate loss models trained locally, which are then broadcast to all other clients to serve as an anchor. The experiments show promising performance compared to existing baselines. Strengths: 1. The paper is well-motivated and well-written. 2. The problem is practical and useful for many real-life applications, though scalability may be the main constraint. 3. The idea is straightforward, and the experiments seem to verify its effectiveness. Weaknesses: 1. Communication complexity: If I understand correctly, every client in the proposed method must broadcast the surrogate models to all other clients. Although the surrogate models consist of only one hidden layer, this design results in a communication complexity of $\mathcal{O}(N^2)$. As the number of clients in the system increases, the additional communication costs will rise dramatically. This might be manageable in some cross-silo settings, where only a few clients participate. 2. Straggler effect: Following point (1), the proposed method requires communication among clients. However, clients may drop out during training. It would be insightful if the authors could analyze how missing surrogate loss models would affect overall performance. 3. Additional privacy concerns: Sharing surrogate models introduces additional privacy risks, e.g., enabling reconstruction attacks or membership inference. While some recent work empirically shows that such private information is less leaked after distillation (e.g., [1] and [2]), the proposed method might be more vulnerable to privacy attacks without differential privacy. [1] Dong, Tian, Bo Zhao, and Lingjuan Lyu. "Privacy for free: How does dataset condensation help privacy?." International Conference on Machine Learning. PMLR, 2022. [2] Wang, Hui-Po, et al. "Fedlap-dp: Federated learning by sharing differentially private loss approximations," Proceedings on Privacy Enhancing Technologies, 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have discussed some limitations. However, the authors are encouraged to discuss the use cases of the proposed method, such as cross-silo settings. Moreover, they are encouraged to discuss the additional privacy risks potentially introduced by their method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Communication complexity ... this design results in a communication complexity of $O(N^2)$ … This might be manageable in some cross-silo settings, where only a few clients participate. Thanks for the thoughtful comment. Since each client will receive the surrogate models of all other clients from the server, we acknowledge that the size of the communication will be $O(N^2)$. We totally agree that our framework is more manageable in some cross-silo settings. However, as the size of the surrogate model is very small containing only one hidden layer (e.g., around 20K bytes), and nowadays data transfer has become much more efficient and affordable (e.g., 1.2 GB/s in modern wireless services), in our humble opinion, communication cost should not be a big problem. > Straggler effect … However, clients may drop out during training. It would be insightful if the authors could analyze how missing surrogate loss models would affect overall performance. Thank you for the nice practical question. We follow your suggestion to experiment with the idea that only a random 10% of the clients are involved in each communication round. Specifically, these clients will only receive the surrogate models from the other 10% of the clients. We summarize our results in the attached PDF, under the setting of Dirichlet(0.1) on the MNIST dataset with 100 clients. We can see that while the performance under 10% client participation is worse than under full client participation, the results of FedNE are still notably better than the baseline methods, FedAvg+NE, demonstrating the effectiveness and applicability of our FedNE framework. > Additional privacy concerns: Sharing surrogate models introduces additional privacy risks, … the proposed method might be more vulnerable to privacy attacks without differential privacy. Thank you so much for the comment. First, we want to apologize for the confusion caused by our title. The main focus of our paper is on a Federated setting, aiming to bring up the unique FL challenges and propose effective solutions for the pairwise training objective in the problem of Federated Neighbor Embedding. We appreciate your suggestions and will extend our discussion by adding a section in the camera-ready version to incorporate the points that you mentioned. Besides the privacy-preserving properties that a Federated setting has introduced, we do not specifically develop any privacy-preserving mechanisms. Therefore, we will surely refine our title and any writing to clarify this. With that being said, privacy-preserving is an important topic and we will explore incorporating differential privacy (or other techniques) in our future work. We will also investigate some characteristics of our algorithm to enhance privacy. For example, our surrogate models take the “low-dimensional” data as inputs. We will investigate whether this design improves privacy preservation since one cannot directly reconstruct the high-dimensional data. Moreover, since it is the server that collects surrogate models from all the clients and distributes them to each client, certain anonymization techniques may be applied so that each client does not know the owner of the surrogate function it receives. --- Rebuttal 2: Title: Response to the rebuttal Comment: I want to thank the authors for their time and response. Since the response only partially addressed my concerns, I'd like to follow up on those points. > Scalability Despite the tiny surrogate model, the communication costs grow quadratically for the entire system and linearly for each client. It could be unfavorable when deployed in a million-scale system, but I can see the applications in common cross-silo settings. > Straggler effect Thanks for the additional experiments. However, I was curious about the stale surrogate models instead of partial participation. In other words, what would happen if some clients send old surrogate models from previous rounds due to the network delay? > Privacy concerns This remains my biggest concern. The authors do not provide any further analysis of it. While the concept of using shared anchors is not entirely new (as previously explored in [1] and preliminary theoretical analyses like [2]), its application in federated clustering appears to be new. The primary concern still lies in the potential privacy and security risks. I'm currently torn in my recommendation. On one hand, this work could provide a foundational analysis that serves as a baseline for federated clustering. On the other hand, the improved performance is built upon the inherent compromise of privacy. Such compromise also makes future comparisons unfair and challenging if the threat model is not stated clearly, i.e., what is assumed to be safe. Therefore, unless the potential privacy leakage is thoroughly discussed or the compatibility with differential privacy is demonstrated, I remain cautiously optimistic but will slightly lower my score. If the authors can propose a robust comparison protocol for future work that accounts for various privacy considerations, I would be willing to reconsider and potentially adjust my score. [1] Wang, Hui-Po, et al. "Fedlap-dp: Federated learning by sharing differentially private loss approximations," Proceedings on Privacy Enhancing Technologies, 2024.\ [2] Li, Bo, et al. "Synthetic data shuffling accelerates the convergence of federated learning under data heterogeneity." Transactions on Machine Learning Research. --- Rebuttal Comment 2.1: Comment: Dear reviewer, Thank you for reading our rebuttal and we apologize that our rebuttal has not fully addressed your concerns. We respond to your remaining concerns as follows. We totally agree that our framework is more manageable in some cross-silo settings. Speaking of the communication cost on extremely large systems e.g., a million-scale system, this is still an unresolved problem in the existing literature. We have made one step further to relax the assumption of publicly available data and the next step in FL for NE could be reducing the communication cost. Regarding the potential stale/old surrogate models due to network delay, we had a related experiment in Appendix B.2 where we studied whether reducing the frequency of surrogate model updates would impact the final performance (i.e., under the Shards setting with 20 clients on MNIST), even though in this experiment we assume that all clients may send out their outdated/stale surrogate models instead of only some of the clients. We can see that the performance remains stable and the improvement is still notable compared to the baseline method, FedAvg+NE, if the surrogate models are only outdated within a threshold of iterations i.e., 10 rounds of communication (Fig.4 in Appendix). While we know this experiment does not exactly match what you asked for, we hope it can still provide a better understanding of our method. Lastly, we do agree that privacy is a very important problem. After serious consideration, there are several potential future directions on how to address and evaluate the privacy considerations in our framework. As demonstrated in Section 4 and the workflow in Figure 1, FedNE is built upon the traditional FedAvg framework. The potential privacy concerns mainly come from the FedAvg framework and the surrogate models that we proposed. To address the privacy risks associated with FedAvg, one potential solution is to implement a FedAvg version of DP-SGD [1]. In this approach, the Gaussian mechanism (GM) is applied to the global model updates, as described in line 10 of Algorithm 1 in [1], to achieve client-level differential privacy (DP). That is, each client’s entire dataset is protected against differential attacks from other clients. To mitigate the privacy concerns associated with our proposed surrogate models, it is important to both anonymize client identities and address the potential risks inherent in sharing these models. To make the surrogate models differentially private, DP techniques can be integrated at different stages. One way is to incorporate DP before sharing the surrogate models. Even though the surrogate models are simple and only contain highly compressed information about the client data, the privacy risks can be further mitigated by applying the Gaussian mechanisms (GM) to the parameters of the surrogate model before it is sent to the server. Additionally, DP techniques can be incorporated during the training phase of the surrogate models through data synthesis. Specifically, to prepare the training targets, instead of using real data, each client can generate synthetic samples (e.g., via data distillation) to compute the repulsion loss values $l_q^{rep}$ (as discussed in Lines 198-201), which are then used to train the surrogate models. To ensure privacy guarantees, when GM is employed, a desired privacy budget can be set using Epsilon ($\epsilon$) and Delta ($\delta$), while balancing between privacy and utility. The moments accountant method can be utilized to track cumulative privacy loss over multiple iterations of FedNE. Furthermore, the privacy auditing algorithm [2] provides a valuable tool for estimating empirical privacy lower bounds within our framework. The empirical Epsilon can be an indicator (i.e., a lower Epsilon, a stronger privacy guarantee) for comparing different methods, e.g., FedAvg, FedNE, even FedNE integrated with DP, and other future works. [1] Geyer, Robin C., Tassilo Klein, and Moin Nabi. "Differentially private federated learning: A client level perspective." arXiv preprint arXiv:1712.07557 (2017). [2] Steinke, Thomas, Milad Nasr, and Matthew Jagielski. "Privacy auditing with one (1) training run." Advances in Neural Information Processing Systems 36 (2024). Please kindly let us know if you have any further questions or concerns. We are more than happy to have a further discussion regarding it. Best, Authors --- Rebuttal 3: Title: Kindly request your reconsideration Comment: Dear Reviewer BJv7, We appreciate your timely feedback on our rebuttal. Given the limited author-reviewer discussion period, we have tried our best to further address it. Please see our response titled "Official Comment by Authors." If our latest response has addressed your concerns, we would appreciate it if you would be willing to consider raising your rating. Thank you for your consideration. Best, Authors --- Rebuttal Comment 3.1: Comment: Dear Authors, Thank you for your prompt response. I’ve carefully reviewed your feedback, and I’m glad to see that the authors agree on the privacy guarantee aspect. While I appreciate the explanation of the common practices in implementing differential privacy (DP), my primary concern remains that the paper might set an unfair baseline for future work if proper DP experiments are not included. Therefore, I will be maintaining my score. --- Reply to Comment 3.1.1: Comment: Dear Reviewer BJv7, Thank you for your timely response and further clarification of your concerns. We sincerely apologize if we have misunderstood your original review and your response to the rebuttal (on 08/10/2024). If we read them correctly, you encouraged us to discuss the additional privacy risks potentially introduced by their method, and propose a robust comparison protocol for future work that accounts for various privacy considerations. We have tried our best to follow your suggestions in our rebuttal and additional responses. However, it seems that your latest comment demands us to further provide experiments with DP to address your concern. While we will be happy to do so, we, unfortunately, cannot complete it on the last day of the discussion phase. Still, we sincerely thank you for all your constructive feedback and suggestions. Best, Authors	Rebuttal 1: Rebuttal: We thank the reviewers for all the valuable comments and constructive suggestions. We are glad that the reviewers found that our paper is “well-motivated” and “well-presented” (Reviewer BJv7, 4ZYS, mP1Q), and our approach is “novel” (Reviewer mP1Q, y2ZN). In the following, we want to first reiterate our contributions. We then respond to each individual reviewer’s comments separately. To address some of the questions, we have included additional experimental results in the attached PDF. We will also incorporate all the feedback in the camera-ready version. ### Overall contributions and setting We want to emphasize that we study a less-explored problem in Federated Learning (FL), FL for Neighbor Embedding (NE), where NE is a family of non-linear Dimensionality Reduction (DR) techniques. As discussed in Section 3.3, in comparison to classification tasks, FL for NE has a unique challenge associated with the pairwise objective function in the NE problem. One of our key contributions is to point out these challenges systematically and provide effective solutions. We have used the benchmark datasets commonly studied in existing DR literature, and our experiments are larger in scale compared to prior work. We appreciate some reviewers’ feedback regarding complexity and scalability, and we have tried our best to respond to them. Overall, there are many properties/challenges that one would expect a sophisticated FL algorithm to achieve and overcome. The area of FL has been significantly advanced in the past five years, thanks to hundreds, if not thousands, of papers addressing various aspects of FL. In this paper, we follow this trend by identifying a new challenge specific to FL for NE and dedicating ourselves to mitigating it. Thus, while our algorithm may not be as computationally efficient or scalable as some existing FL methods, we respectfully believe that our contributions are significant and valuable for future study in the field of FL for DR. ### Regarding scalability, complexity, and computational efficiency Our paper focuses on a less-explored combination of FL and DR. Please allow us to address the questions regarding scalability, complexity, and computational efficiency in three aspects. First, to our knowledge, DR has focused on relatively smaller-scale datasets, compared to classification. This is because computational complexity is never a trivial problem even for many outstanding DR techniques, particularly for non-linear methods such as Isomap and t-SNE which have non-convex cost functions [44]. Indeed, our experiments have included the most widely-used benchmarks used in the DR literature. Second, in terms of FL, we work on a less-explored problem, compared to other learning problems such as classification. Compared to the latter, in which the instance-based loss can closely approximate the centralized loss with frequent communication (Line150-154), the Federated Neighbor Embedding problem has a unique challenge associated with its pairwise objective function. Thus, we focus our study mostly on addressing this challenge using standard DR datasets. With that being said, compared to prior work on decentralized data projection, we already considered more clients and larger datasets. Specifically, d-SNE [37] and F-dSNE [38] conducted experiments with only 3 to 10 local sites on the subsampled MNIST and biomedical datasets which contain only hundreds to thousands of data samples. However, our FedNE is evaluated on the full benchmark datasets with the settings of 100 local sites. Third, while we have not studied larger datasets and more clients like other FL works in classification, we expect that our approach is applicable in real-world settings, for example, cross-silo settings with a manageable amount of clients. As pointed out in Sect. 4.3 (Line 235-237), our approach only requires 35% additional GPU time compared to FedAvg, and we expect such overhead to remain similar when going to larger datasets with a similar number of clients. When the number of clients increases, we may optionally drop a portion of surrogate models in local training. As shown in our response to Reviewer BJv7, such a setting does not lead to a significant performance drop but can maintain scalability. [37] "See without looking: joint visualization of sensitive multi-site datasets." IJCAI. 2017. [38] "Federated, Fast, and Private Visualization of Decentralized Data." Workshop of Federated Learning and Analytics in Practice (2023). [44] "Visualizing data using t-SNE." JMLR 9.11 (2008). Pdf: /pdf/66b4abf4825adf07d772fbc85fcd6cc0345f9b71.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Amortized Bayesian Experimental Design for Decision-Making	Accept (poster)	Summary: This paper proposes a method for decision-aware Bayesian experimental design, where the design is not optimized with respect to the most accurate posterior distribution of the latent parameters but rather with respect to the expected utility gain of the actual (down-stream) decision task. Strengths: This is an innovative paper with high practical relevance. The proposed method appears sound and the corresponding neural networks well designed to suit the goal. Despite my questions and concerns (see below), I am positive about this paper overall and eager to increase my score should my points be addressed. Weaknesses: - The presentation of p(y_Xi \| h_t) between Eq 3 and 4 is partially unclear to me. From the definition, it seems this is not actually a distribution but a set of distributions. To me, then notation p(y_Xi \| h_t) appears to be quite the abuse of notation because we cannot readily read this it as a single distribution. Can you perhaps think about a different notation that makes this easier to parse and understand? Relatedly, in Equation 4, it appears that we compute an expectation over p(y_Xi \| h_t). But how do we compute an expectation over a set of distributions? I think I get what the authors do and want to imply but to me this notation doesn’t help in understanding it. - Equation 7: It seems we approximate the predictive distribution always by a Gaussian. I mean this of course works if the true underlying function is some kind of GP, but what if the true predictive distribution is far away from Gaussian? I don’t see this choice to be discussed properly so I consider it a weakness of this paper for now. - The discussion of training and inference time can only be found in the appendix. Specifically, training speed seems to be substantial, which of course makes sense for an amortized method. However, I don’t see any discussion for when the training actually amortizes. That is, how many BED tasks do we need to run at minimum before the total (training + “inference”) time of the new method becomes better than those of the competing methods. More generally, I think a discussion of speed should be more prominent in the paper. - 6.1 toy example was hard for me to understand at first. Is this just a standard BO task to find the point where the unknown function is maximal? Technical Quality: 3 Clarity: 2 Questions for Authors: - In 4.1 Query set: How problematic is the fact that we randomly generate some designs from the design space. Doesn’t this mean we need a distribution over the design space? How can we obtain (or define) such a distribution in general? - In 4.1 Query set: You say that in the deployment phase we can obtain the optimal design by optimizing the models (which model’s?) output. How do you optimize this exactly? - Given that (non-decision aware) amortized BED methods exist, why are the benchmarks only comparing against non-amortized methods? I suggest to also add amortized methods to the benchmarks unless you can convince me that this is not sensible for some reason. - What is the scalability of the method in terms of all relevant dimensions, e.g., dimensionality of xi, y, a, etc? - Figure 4: you say that your method provides substantial gains, but at least on the scale in the figure, gains seem to be small. Can you clarify why you feel that the improvements are indeed “substantial gains”? - The method has quite a lot of components, I wonder which of the components is responsible for the improved results? For example, how relevant is it to consider non-myopic designs, i.e., how does the method perform when only trained in a myopic setup? Relatedly, are the alternative methods myopic or non-myopic? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 4 Limitations: The paper discusses several limitations. I am missing a discussion on the initial overhead of training, which is usually substantial in amortized methods. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive comments and thoughtful questions. We address your remarks and questions below. > 1. The presentation of p(y_Xi \| h_t) between Eq 3 and 4 is partially unclear to me… Thanks for the question. $p(y_\Xi \| h_t)$ is a joint distribution and is well-defined as a stochastic process. Please refer to the global response for further clarification. > 2. What if the true predictive distribution is far away from Gaussian? That’s a good point. Please refer to the response for question 9 of reviewer WHLG. > 3. The discussion of training and inference time can only be found in the appendix…I think a discussion of speed should be more prominent in the paper. The key advantage of amortized methods lies in their efficiency during the inference phase. While the upfront training time can be substantial, this investment pays off when the model is used repeatedly for multiple BED tasks. To answer your specific question about when the training amortizes. For example, in our active learning experiment, we observed that traditional methods like DEIG require approximately 100 minutes for a complete BED task. In contrast, our model's total training time is around 10 hours, with the inference time being negligible. Therefore, after conducting more than 6 experiments during the deployment phase, the cumulative cost of our method becomes more efficient than the traditional method. > 4. Toy example was hard for me to understand at first. Is this just a standard BO task? Thanks for the question. The toy experiment is not a standard BO task but a special active learning task. Our goal is to estimate the value of an unknown function at a specific target point $x^$ by actively querying points. Traditional active learning methods are not target-oriented and typically query points based on the overall uncertainty of the function, which is not optimal for estimating the value at $x^$. Our method, however, considers the downstream task, meaning our ultimate goal is to perform regression at the target location $x^$. Thus, our decision-making-aware policy will strategically query points, specifically around $x^$, as shown in Figure 3(a). > 5. How problematic is the fact that we randomly generate some designs from the design space? That’s a good question. For most problems, we usually either have some prior knowledge about the design space or the candidate designs come from a predefined finite set, as is commonly practiced in the pool-based active learning literature [3]. The more prior knowledge we have about the problem, the better the candidate designs we can generate. > 6. You say that in the deployment phase we can obtain the optimal design by optimizing the models (which model’s?) output. How do you optimize this exactly? We apologize for the unclear description. For continuous design spaces, we can optimize $\xi^{(q)}$ to maximize the query head's output $\mathbf{q}$, thereby obtaining the design with the highest probability. This can be achieved using optimization techniques such as gradient ascent. > 7. Given that (non-decision aware) amortized BED methods exist, why are the benchmarks only comparing against non-amortized methods? We have included a new amortized method baseline in the top-$k$ experiments and also explained why we didn’t compare with any amortized BED method previously. Please refer to the response for question 5 of reviewer WHLG. > 8. What is the scalability of the method in terms of all relevant dimensions? We have validated our method's effectiveness on high-dimensional design spaces, e.g., in the top-$k$ optimization task, $d_x$ for ranger is 9, and 16 for xgboost. Regarding the output dims, our paper focuses on single-output cases, the same as most BED works. However, our architecture can be readily generalized to multidimensional outputs by adding separate prediction heads for different outputs, as demonstrated in CNP literature [4, 5]. Lastly, for the action dims, we believe this does not significantly affect the performance since the action is made after the model's predictions. > 9. Figure 4: Can you clarify why you feel that the improvements are indeed “substantial gains”? We acknowledge that the term "substantial" may have been an overstatement, we will revise the wording. However, it is clear that our method outperforms the other baselines, particularly within the first 10 queries. > 10. I wonder which of the components is responsible for the improved results? How does the method perform when only trained in a myopic setup? Relatedly, are the alternative methods myopic or non-myopic? The improved results of our method are primarily due to our query head. This module amortizes the experimental design process, which learns common features from different tasks, allowing it to propose more valuable experiments. We conducted an ablation study in Appendix F.3 to verify the performance improvement brought by the query head. Regarding the difference in performance between non-myopic and myopic objectives, we conducted an additional ablation study, as detailed in our global response. All the alternative methods we considered in our paper are myopic. > 11. I am missing a discussion on the initial overhead of training. Thanks for your suggestion. Indeed, like all amortized approaches, our method requires a large amount of data and upfront training time to develop a reliable model. However, once the model is trained, it offers long-term benefits of faster inference. We will discuss this trade-off in the revised manuscript. References [1] Garnelo et al. (2018). Conditional neural processes. ICML. [2] Müller et al. (2021). Transformers can do Bayesian inference. ICLR. [3] Settles (2012). Active learning. Springer. [4] Markou et al. (2022). Practical conditional neural processes via tractable dependent predictions. ICLR. [5] Bruinsma et al. (2023). Autoregressive conditional neural processes. ICLR. --- Rebuttal Comment 1.1: Comment: I thank the authors for their thoughtful responses and additional experiments, I have rasied my score from 6 to 7. --- Reply to Comment 1.1.1: Comment: Thank you for your response and for increasing your score. We really appreciate it.	Summary: The paper looks at the problem of designing Bayesian optimal experiments taking into account the downstream decision making. At the core is a Transformer Neural Decision Process (TNDP) architecture that is trained to amortise the experimental design process whilst simultaneously inferring the optimal downstream decision. Strengths: - Relevant and interesting topic: Downstream decision making is what ultimately matters, so taking this into account when designing experiments to collect data can result in more cost- and sample-efficient learning. - Motivation for the paper as well as clarity of writing are excellent. Contextualisation relative to prior work can be improved as outlined in the next section. - The proposed Transformer Neural Decision Process (TNDP) architecture is tailored to the BED problem, is well-explained and adds some novelty to the architectures typically used in the field. Weaknesses: ### Sections 2.2 & 3.2 and Lindley's decision-theoretic BED [1]: My main issue with the paper is the presentation of DUG and EDUG as novel. This framework was first formulated in [1], and is very well summarised in Section 1.3 of [2]. I strongly recommend the authors read that section, and present their Section 3.2 accordingly, acknowledging they follow Lindley, 1972. The questions/comments in the next 2 bullets are a consequence of this omission of literature. - Second paragraph of Sec 2.2: I am not sure how the predictive distribution $p(y \| \xi, h_t)$ is defined. I would think it is $p(y \| \xi, h_t) = \mathbb{E}_{p(\theta \|h_t)} [p(y \| \xi, \theta)]$. Whether or not you compute/approximate the posterior $p(\theta \|h_t)$, or seek to directly approximate $p(y \| \xi, h_t)$ (eg variationally), I think you should explicitly define what this quantity is. - I am not sure how the utility $u(y_\Xi, a)$ is defined. From a Bayesian decision-theoretic approach, the utility has to depend on the state of the world $\theta$, as well as the experiments $\xi$ you are going to perform (which I guess is implicit in $y_\Xi$). So shouldn't the "lowest level" utility be a function $u(y, \theta, \xi, a)$, which you then integrate over $p(\theta\|h_t)$, to obtain $u(y, \xi, a) = \mathbb{E}_{p(\theta\|h_t)} [u(y, \theta, \xi, a)]$, then take $\max$ wrt $a$, and finally integrate over the predictive $p(y \|\xi, h_t)$ to obtain an expected utility, which can then act a design ranking criterion, as you do in Eq 4 and (cf Eq 2 in [2]). ### Related work: For a field that has such rich history and renewed interest from the ML community recently, the related works section is quite short and sparse on citations. Some areas that are missing include: - Decision-theoretic BED: as previously discussed, the general framework of utility-based BED was developed by Lindley (1972). - BED + RL: this work touches on some aspects of RL; It might be good to discuss relations recent works in the intersection such as [5] and [6] (in addition to those mentioned) - Decision-theoretic approaches in related fields such as Bayesian Optimisation, e.g. [7], [8] - Finally, I'm not too familiar with this line of literature, but more recent work around decision transformers---is there any relation between TNDP with works like [9] and [10]? ### Other: - Line 6: "most recent BED methods use amortised inference with a policy network" is not quite correct in the sense that no "real inference" (posterior updates on the parameters $\theta$) are performed. - Line 179: "to ensure the framework satisfied the permutation invariance property of sequential BED": not all BED problems are permutation invariant. For example, designing experiments for time series models (e.g SIR in [3] and [4]), permutation invariance does not hold. This aspect has been discussed in e.g. Section 3.3 of [3]. - Assuming you do want a permutation invariant architecture (most design problems fall in that category): by conditioning on $t$ as part of the global information (GI) set, I think you actually break that invariance. This is because encoding $(\xi, y)$ at time $t$ or at time $s$ will give you different outputs. As far as I can tell from Fig2b), $D_c$ does attend to GI. Could you please explain if that's the case or I have misunderstood something? ----- #### References [1] Lindley, D. V. (1972). Bayesian statistics: A review. Society for industrial and applied mathematics. [2] Chaloner, K., & Verdinelli, I. (1995). Bayesian experimental design: A review. Statistical science, 273-304. [3] Ivanova, D. R., Foster, A., Kleinegesse, S., Gutmann, M. U., & Rainforth, T. (2021). Implicit deep adaptive design: Policy-based experimental design without likelihoods. Advances in neural information processing systems, 34, 25785-25798. [4] Kleinegesse, S., & Gutmann, M. U. (2019, April). Efficient Bayesian experimental design for implicit models. In The 22nd International Conference on Artificial Intelligence and Statistics (pp. 476-485). PMLR. [5] Mehta, V., Paria, B., Schneider, J., Ermon, S., & Neiswanger, W. (2021). An experimental design perspective on model-based reinforcement learning. arXiv preprint arXiv:2112.05244. [6] Mehta, V., Char, I., Abbate, J., Conlin, R., Boyer, M., Ermon, S., ... & Neiswanger, W. (2022). Exploration via planning for information about the optimal trajectory. Advances in Neural Information Processing Systems, 35, 28761-28775. [7] Neiswanger, W., Yu, L., Zhao, S., Meng, C., & Ermon, S. (2022). Generalizing Bayesian optimization with decision-theoretic entropies. Advances in Neural Information Processing Systems, 35, 21016-21029. [8] Ivanova, D. R., Jennings, J., Rainforth, T., Zhang, C., & Foster, A. (2023, July). CO-BED: information-theoretic contextual optimization via Bayesian experimental design. In International Conference on Machine Learning (pp. 14445-14464). PMLR. [9] Chen, L., Lu, K., Rajeswaran, A., Lee, K., Grover, A., Laskin, M., ... & Mordatch, I. (2021). Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34, 15084-15097. [10] Zheng, Q., Zhang, A., & Grover, A. (2022, June). Online decision transformer. In international conference on machine learning (pp. 27042-27059). PMLR. Technical Quality: 2 Clarity: 3 Questions for Authors: In addition to the questions raised in the Weaknesses section: 1. I think the main contribution of the paper is the TNDP architecture. Have the authors performed any ablations, e.g. not sharing the same embedding block? Not including $t$ in the GI? 2. In the decision-aware AL experiment: why does the random baseline perform as good as all the other ones? 3. Could you give guidance on choosing utility functions? For the experiments in the paper it is quite straightforward to define them, but in real-world practical application that might not be the case. This is the reason why the mutual information has become the de facto standard utility in BED. Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Some limitations of the work were outlined in the Discussion section of the paper. Regarding negative societal impact, the field of experimental design (which boils down to efficient data collection), generally warrants some discussion. The experiments presented in this paper mostly use synthetic data and do not have negative impact; the HPO experiment, which uses real data does not (directly) represent an application with negative impact. However, applying these methods in real-world applications, particularly if decisions directly affect humans, as in e.g. personalised medicine, could raise concerns around bias, fairness, explainability and privacy. I would suggest to the authors to add 1-2 sentences in their limitations section to acknowledge 1) the synthetic or semi-synthetic nature of the experiments, and 2) potential concerns that might arise when applying their method in real-world applications. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and the valuable references you provided. We address your questions and points raised below. Weaknesses > 1. My main issue with the paper is the presentation of DUG and EDUG as novel. We greatly appreciate your provided references and insights. We will include a discussion in Sections 2.2 and 3.2 about the connections and distinctions between our work and the decision-theoretic BED frameworks mentioned in [1, 2]. For further details, please refer to the global response and our answers to the following two questions. > 2. I am not sure how the predictive distribution $p(y\|\xi, h_t)$ is defined… I think you should explicitly define what this quantity is. Yes, $p(y\|\xi, h_t) = \mathbb{E}_{p(\theta\|h_t)}[p(y\|\xi, \theta, h_t)]$. It is obtained by marginalizing the posterior distribution of the parameters. In our architecture, we directly approximate the predictive distribution with neural processes, thereby bypassing the need to define and approximate the underlying parameters. We will add a description of the predictive distribution in P4L126 to clarify this point. > 3. I am not sure how the utility $u(y_{\Xi}, a)$ is defined. Our utility function is defined in terms of outcomes, which slightly differs from the traditional definition. The distribution of outcomes is obtained by (implicitly) marginalizing over $\theta$. A similar decision-theoretical setup can be found in [3]. As mentioned in the global response, the major difference between EDUG and Equation 2 in [2] is that we include an additional expectation over $p(y_\Xi \| h_t)$, as the optimal Bayes action in our setup is based on the predictions of outcomes. > 4. The related works section is quite short and sparse on citations. Thanks for your references, we will add them. Regarding the differences with decision transformers, while both our architecture and theirs are based on transformers, their "decision" is similar to the step of experimental design in our context. However, we additionally amortize the downstream decision-making process, making the learning process more challenging. > 5. Line 6: "most recent BED methods use amortised inference with a policy network" is not quite correct in the sense that no "real inference" are performed. Agreed. We will change to “most recent BED methods leverage an amortized policy network to rapidly design experiments”. > 6. Line 179: "to ensure the framework satisfied the permutation invariance property of sequential BED": not all BED problems are permutation invariant. Agreed, we acknowledge that permutation invariance was meant more as an assumption than an inherent property of BED. We will fix it and add that for tasks where permutation invariance does not hold, our model can be easily adapted by incorporating positional encoding to add sequential/temporal information to the design. > 7. Assuming you do want a permutation invariant architecture… I think you actually break that invariance… Could you please explain if that's the case or I have misunderstood something? That is a misunderstanding. Permutation invariance here only means that changing the order of the historical data does not affect the design of the next experiment. Our architecture satisfies that. (Note that even without the timestamp information, the encoding as such would give different outputs at different points of time.) Questions > 8. Have the authors performed any ablations? We conducted an experiment in Appendix F.3 to verify the effectiveness of the query head. Additionally, we have added a new set of ablation experiments on the discount factor; please refer to the global response for details. Regarding the inclusion of time $t$, we performed ablation studies across different dimensions. The results were inconsistent. Since we only had time to run our algorithm once for each dimension, we cannot draw definitive conclusions. We will further investigate the role of $t$. > 9. In the decision-aware AL experiment: why does the random baseline perform as good as all the other ones? Because methods like US and DUS are traditional active learning methods which do not take the decision-making problem into account. Hence there is no guarantee that the queries made would improve the quality of the decision-making. However, we can see that DEIG actually performs better than the random baseline, as it considers the decision-making problem by taking into account the posterior distribution of the optimal decision for the target point. > 10. Could you give guidance on choosing utility functions? Our utility function comes directly from the downstream decision-making task. We suspect the popularity of mutual information arises from an unwillingness to commit to a specific task, which makes sense if the task is not known. However, if we know our downstream task well, such as in drug design where we aim to maximize the efficacy of the drug while minimizing its risks, we can leverage input from domain experts to define our objectives. We will comment on this in the Discussion. Limitations > 11. I would suggest to the authors to add 1-2 sentences to acknowledge 1) the synthetic or semi-synthetic nature of the experiments, and 2) potential concerns that might arise when applying their method in real-world applications. Thanks. We conducted a new retrosynthesis planning experiment with real-world data; please refer to the global response. We will also discuss the potential negative societal impact, especially when experiments or decisions directly affect humans, in the limitations section. References [1] Lindley, D. V. (1972). Bayesian statistics: A review. Society for industrial and applied mathematics. [2] Chaloner et al. (1995). Bayesian experimental design: A review. Statistical science. [3] Kuśmierczyk et al. (2019). Variational Bayesian decision-making for continuous utilities. Neurips. --- Rebuttal Comment 1.1: Comment: I acknowledge I have seen the rebuttal and will respond in detail. Unfortunately, this will likely happen over the weekend.	Summary: The paper proposes a transformer-based architecture for jointly sampling designs and decisions in Bayesian Experiment Design (BED) using a forward-looking criterion. The latter considers the improvement in maximum expected utility brought about by a new design-outcome pair, where the expectation is taken with respect to the predictive distribution of the model. The main innovation of the paper lies in the coupling between information gain and utility maximization in an amortized, transformer-based framework in the spirit of attentive neural processes. The performance of the new architecture is evaluated on a toy regression task and two more representative models, exhibiting stable performance gains over contender methods. Strengths: - The paper is clearly written, the ideas and formulations are stringent and well-justified, overall making it easy to follow and a pleasure to read (with the exception of Section 4.1, see below). - The proposed architecture and training objectives are novel and seem to unlock both qualitative and quantitative improvements over existing methods. - The results indicate superior and stable performance of the proposed architecture on two interesting tasks, along a toy 1D GP model which seems to be a standard proof-of-concept task in the neural process (NP) literature. Weaknesses: - Some notational confusion can be avoided by consistently using the notation $a_{1:t}$ to denote a sequence of $t$ elements and $a_t$ to denote the $t$-th element in the sequence. Currently, $h_t$ denotes a sequence, but, e.g., $y_t$ denotes an element, and then again $\theta_{1:L}$ also represents a sequence. Also, P4L126 is an abuse of notation with slightly confusing wording, such as “the predictive posterior distribution over all possible designs”, whereas the predictive distribution(s) are over future \textit{outcomes}. This is in no way different than the posterior predictive in Bayesian (non-linear or linear) regression, where the posterior predictive is conditioned on the training data set and the set of (unlabeled) predictors available at test time. Hence, I struggle to understand the need for the convoluted abuse of notation, but I may be missing something. Also section 4.1 suddenly starts using bold font for vectors, which was not the case in the preceding sections. - Figure 2 is not particularly informative for the data flow, as it does not clearly communicate weight sharing, input-output operations and dependencies (left panel); the right panel comes out of the blue and is not well explained (i.e., what are the elements on the “left” and on the “top”); the description below on P6 does indeed disambiguate the idea behind the construction of the masks, but I believe it is best when figures support and enhance the text and not vice versa. - Overall, I feel that Section 4.1 is the weakest link in the paper, and I believe the authors can think about optimizing the ratio of details dispersed between the main text and the appendix. For instance, there is no need to reiterate established transformer-based computations, but it could be helpful to explicate the construction of the masks, the representation types (e.g., vectors, sequences of vectors,...?), and the precise partitioning of the components into keys, queries, and values. - According to my understanding, none of the contender methods in the experiments is an amortized method. Wouldn’t some of the existing amortized BED methods (e.g., as highlighted in the Related Work) make for suitable benchmarks, despite not optimizing for future decisions? - The topic of model misspecification is never mentioned in the paper, even though the comprehensive review paper [1] states that it remains a major unsolved issue in BED and in amortized Bayesian inference more generally [2]. I believe this should also be acknowledged in the current paper and the authors can potentially think about quantifying the impact of model misspecification in a small ablation study in the final version of the manuscript. I am happy to discuss these points with the authors and increase my score if they are addressed / clarified. [1] Rainforth, T., Foster, A., Ivanova, D. R., and Bickford Smith, F. (2024). Modern Bayesian 429 experimental design. Statistical Science, 39(1):100–114. [2] Schmitt, M., Bürkner, P. C., Köthe, U., & Radev, S. T. (2024). Detecting Model Misspecification in Amortized Bayesian Inference with Neural Networks: An Extended Investigation. arXiv preprint arXiv:2406.03154. Technical Quality: 3 Clarity: 3 Questions for Authors: - Perhaps section 2 can be organized in a way to avoid singleton nested subsection (i.e., 2.1.1)? - P4L130: Isn’t there also an assumption that decision are optimal only if there is no model misspecification (i.e., that we are working with the posterior of the “true” model)? - Are there any practical disadvantages of assuming a diagonal Gaussian predictive distribution? Can complex models induce multimodal or highly correlated predictive distributions that? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors openly discuss the current limitations of their approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and the valuable comments. We address your remarks and questions below. Weaknesses > 1. Some notational confusion can be avoided… Thanks, we will improve the notations according to your suggestions in the revised paper. Specifically, we will replace $h_t$ with $h_{1:t}$ to better denote a sequence of elements. And we will consistently use bold font for vectors across the whole paper. > 2. P4L126 is an abuse of notation with slightly confusing wording, such as “the predictive posterior distribution over all possible designs”. We apologize for the confusing wording and the lack of a clear explanation of the predictive distribution. Please refer to the global response for a detailed answer. > 3. Figure 2 is not particularly informative for the data flow ... the right panel comes out of the blue and is not well explained. We will improve Figure 2 to better help readers understand our work. Specifically, we will use distinct modules to more clearly visualize the data embedding block (e.g., designs in different sets share the same embedder). We will also refine the Transformer block to highlight the attention mechanisms between different sets. For the causal mask on the right panel, we will remove the elements with complex notation on the left and top, replacing them with set names for better clarity. > 4. Overall, I feel that Section 4.1 is the weakest link in the paper, and I believe the authors can think about optimizing the ratio of details dispersed between the main text and the appendix. Thank you, we will revise. To address your specific points, we will provide a more detailed explanation of the construction of the masks, along with the improvements to the masks in Figure 2 as mentioned above. As stated in our response to the first question, we will improve the notation of some elements to enhance reader comprehension. Finally, we will clarify the elements involved in the attention mechanism. Specifically, we concatenate all sets together and use self-attention to obtain a single attention matrix, then use masks to determine the dependencies between different sets. > 5. According to my understanding, none of the contender methods in the experiments is an amortized method. Wouldn’t some of the existing amortized BED methods make for suitable benchmarks? We agree that it is important to compare with an amortized method. We added PFNs4BO [1], an amortized optimization framework, as a new baseline for our top-$k$ experiments. Please refer to the global response for the experiment results. As for why we did not choose to compare with any amortized BED method before: Amortized BED mainly refers to DAD [2] and its subsequent works. These methods cannot be directly extended to our tasks for two reasons. First, DAD is a deterministic policy suitable only for continuous design spaces, whereas our tasks such as decision-aware active learning involve discrete data with covariates and their decisions. Second, even though there are subsequent works [3] extending DAD to discrete design spaces, these methods require an additional model or post-processing to make the final decision, making a direct comparison with our method inappropriate without significant alterations to DAD-based frameworks. > 6. The topic of model misspecification is never mentioned in the paper. That’s a very good point. We fully agree on the importance of model misspecification which has received significant attention recently in some areas, such as Simulation-Based Inference, but it has not been sufficiently studied in BED (beyond the handful of papers cited in [4]). Most BED works assume that the models are well-specified. In our work, model misspecification can indeed occur during the inference stage. Additionally, the utility might also shift during the deployment phase, potentially influencing the model's performance. In addition to detecting model misspecification, tackling the problem of robust experimental design under model misspecification would also be very interesting. We will add this discussion and outline it as an important area for future work. Questions > 7. Perhaps section 2 can be organized in a way to avoid singleton nested subsection? Agreed; will do. > 8. P4L130: Isn’t there also an assumption that decision are optimal only if there is no model misspecification? Yes, it is a basic common assumption in the field of BED [1, 2]. We will explicitly state this assumption in the revised manuscript. > 9. Are there any practical disadvantages of assuming a diagonal Gaussian predictive distribution? Can complex models induce multimodal or highly correlated predictive distributions? That’s a good question. Our TNDP follows the common practice in the neural processes literature [5] of using independent (diagonal) Gaussian likelihoods. If modeling correlations between points is crucial for the downstream task, we can replace the output with a joint multivariate normal distribution (similar to GNP [6]) or predict the output autoregressively (similar to AR-CNP [7]). For modelling multimodal predictive distributions, we could replace the Gaussian head with a mixture-of-Gaussians head. These modifications can be easily implemented in the TNDP architecture. We will mention this in the Discussion. References [1] Müller et al. (2023). Pfns4bo: In-context learning for Bayesian optimization. ICML. [2] Foster et al. (2021). Deep adaptive design: Amortizing sequential Bayesian experimental design. ICML. [3] Blau et al. (2022). Optimizing sequential experimental design with deep reinforcement learning. ICML. [4] Rainforth et al. (2024). Modern Bayesian experimental design. Statistical Science. [5] Garnelo et al. (2018). Conditional neural processes. ICML. [6] Markou et al. (2022). Practical conditional neural processes via tractable dependent predictions. ICLR. [7] Bruinsma et al. (2023). Autoregressive conditional neural processes. ICLR. --- Rebuttal Comment 1.1: Comment: I thank the authors for their response, further evaluation, and clarifications. I will keep my positive score. --- Reply to Comment 1.1.1: Comment: Thank you for your time and consideration. We're glad to hear that you keep your positive score.	Summary: This paper tackles an important problem of designing experiments in a way that directly optimizes downstream decision-making tasks, going beyond just inferring parameters of interest. The authors make several valuable contributions: 1. They introduce the concept of Decision Utility Gain (DUG) to quantify how much an experimental design improves the expected utility of the downstream decision. 2. They propose a novel neural architecture called the Transformer Neural Decision Process (TNDP) that amortizes both the experimental design selection and the approximation of the predictive distribution needed for decision-making. This unified amortized framework is a key innovation. 3. The authors develop a non-myopic training objective that looks beyond just the immediate decision utility to account for effects of the current design on future rewards. 4. Empirically, they demonstrate TNDP's effectiveness over traditional methods on various tasks like active learning, hyperparameter optimization, showing it can find informative designs and make accurate downstream decisions. In summary, this work makes valuable conceptual and technical contributions to the area of Bayesian experimental design by pioneering decision-aware amortized methods. It opens up new research directions for further enhancing real-world decision-making via optimized experimental data acquisition. Strengths: - The paper presents a novel problem formulation by introducing the concept of Decision Utility Gain (DUG), which shifts the focus of experimental design from reducing parameter uncertainty to directly optimizing downstream decision utility. This new perspective is a creative departure from traditional Bayesian experimental design (BED) approaches. - The application of amortized inference techniques to decision-aware experimental design can be considered an original contribution, as it represents a new domain for these methods beyond traditional BED. - The empirical evaluation is comprehensive, spanning diverse tasks such as active learning, hyperparameter optimization, and synthetic regression problems. The results demonstrate the consistent superiority of TNDP over traditional methods. Weaknesses: - The authors could provide a more rigorous analysis of the properties and characteristics of the TNDP architecture, such as its convergence behavior, sample complexity, and theoretical guarantees (if any) regarding the quality of the proposed designs and decisions. - The experimental evaluation, while comprehensive, focuses primarily on synthetic and benchmark datasets. While these serve as important proof-of-concept demonstrations, the paper could benefit from including real-world case studies or applications to further validate the practical utility of the proposed framework. - While the amortized nature of TNDP is highlighted as a key advantage, the paper could provide a more detailed analysis of the computational complexity and scalability of the proposed approach. This analysis could include factors such as the training time required for different problem sizes, the memory footprint, and the scalability of the attention mechanisms used in the Transformer architecture. Technical Quality: 3 Clarity: 3 Questions for Authors: - Can the authors provide a more in-depth theoretical analysis of the Decision Utility Gain (DUG) concept, including its relationship with existing concepts like Value of Information (VoI) or Information Gain (IG)? - Have the authors explored the sensitivity of TNDP's performance to different hyperparameter choices, such as the discount factor α used in the non-myopic objective? If so, can they share insights into this analysis? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - The authors mention the use of a basic REINFORCE algorithm for training the query head, which can lead to unstable training, especially in tasks with sparse reward signals. While they suggest the use of more advanced reinforcement learning methods as a potential solution, a more detailed discussion on the specific challenges faced during training and the trade-offs involved in selecting different RL algorithms would be beneficial. - The authors mention that their model is trained on a fixed-step length, assuming a finite horizon for the experimental design process. A discussion on the limitations of this assumption and the potential difficulties in extending their approach to infinite horizon or open-ended experimental scenarios would be valuable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive assessment of our work and the points you raised. In the following we address your questions and points raised. Weaknesses > 1. The authors could provide a more rigorous analysis of the properties and characteristics of the TNDP architecture, such as its convergence behavior, sample complexity, and theoretical guarantees (if any) regarding the quality of the proposed designs and decisions. The primary purpose of this paper is to propose a new amortized BED framework targeted at downstream decision-making. We leverage the established REINFORCE algorithm to aid in model training. Since we did not introduce a new RL algorithm, convergence analysis can be referenced from existing works, such as the analysis provided in [1]. Regarding sample complexity, for the toy example and active learning case, we generate synthetic data online. Specifically, we spent 50,000 steps to train our model and generated 16 synthetic datasets for each step. We will include these additional details in the appendix. For top-$k$ optimization, our model is trained using the publicly available HPO-B dataset, and detailed information about this dataset is already provided in the appendix. > 2. The paper could benefit from including real-world case studies or applications to further validate the practical utility of the proposed framework. We conducted a new retrosynthesis planning experiment based on real-world molecule data, please refer to the global response. > 3. The paper could provide a more detailed analysis of the computational complexity and scalability of the proposed approach. (This analysis could include factors such as the training time required for different problem sizes, the memory footprint, and the scalability of the attention mechanisms used in the Transformer architecture.) Thanks, we will add the following discussion in the revised manuscript. We have already included the rough total training time in the appendix. We will further provide descriptions of the training time and memory footprint required for each task. For example, for active learning experiments, a Tesla V100 GPU with 32GB memory was used, with an average memory consumption of 8 GB. The training time took about 10 GPU hours. Regarding scalability, our architecture is based on the Transformer, which suffers from quadratic complexity with respect to the input sequence length. This can become a bottleneck when the query set is very large. We will also include a further discussion of this issue and potential optimizations in the revised manuscript. Questions > 4. Can the authors provide a more in-depth theoretical analysis of the Decision Utility Gain (DUG) concept, including its relationship with existing concepts like Value of Information (VoI) or Information Gain (IG)? IG can be regarded as a special case of DUG. First, we define $\mathcal{P}(\theta)$ as a set of distributions that we assume contains the true posterior distribution $p(\theta\|h_t)$. We then define the decision space as $\mathcal{A} = \mathcal{P}(\theta)$, and the utility function as $\log a$, where $a \in \mathcal{P}(\theta)$. The optimal action in this case will be $a^* = p(\theta\|h_t)$ based on the definition of entropy. DUG can be reinterpreted as the entropy reduction when we observe a new design pair $\\{\xi, y\\}$, which corresponds to the definition of IG. > 5. Have the authors explored the sensitivity of TNDP's performance to different hyperparameter choices, such as the discount factor α used in the non-myopic objective? If so, can they share insights into this analysis? We added an ablation study regarding the choice of the discount factor, please refer to the global response. Limitations > 6. A more detailed discussion on the specific challenges faced during training and the trade-offs involved in selecting different RL algorithms would be beneficial. Thanks for your suggestion. In our experiments, the REINFORCE algorithm has proven sufficient for training an effective model. However, for more complex problems in the future, REINFORCE may be prone to high variance in policy gradient estimates. If needed, RL training can be improved by using more advanced algorithms like PPO [2], but the trade-offs include introducing more hyperparameters, such as the clip ratio, and increased computational cost, which requires more tuning of the system. Importantly, our work shows that our method can be trained effectively without needing complex, ad hoc RL techniques. We will expand the discussion on these aspects in the revised manuscript. > 7. The authors mention that their model is trained on a fixed-step length… A discussion on the limitations of this assumption and the potential difficulties in extending their approach to the infinite horizon or open-ended experimental scenarios would be valuable. The finite horizon assumption is sufficient for most BED problems, as we usually operate with a limited budget for experimental designs. However, for more complex BED problems, such as long-term medical trials, extending our approach to an infinite horizon setting could be valuable. Potential challenges include increased instability during training and higher computational costs. We will expand the discussion on these aspects in the revised manuscript. References [1] Zhang et al. (2021). Sample efficient reinforcement learning with REINFORCE. AAAI. [2] Schulman et al. (2017). Proximal policy optimization algorithms. arXiv:1707.06347.	Rebuttal 1: Rebuttal: We thank the reviewers for their thoughtful comments and suggestions. We are glad to see that all reviewers have a positive view of the paper. Specifically, the reviewers agreed on the following strengths of the paper: * Relevance: Zp5w: “tackles an important problem”. Ctfm: “relevant and interesting topic”. 7z19: “This is an innovative paper with high practical relevance”. * Novelty: Reviewers Zp5w, WHLG, and Ctfm agree that “the proposed architecture is novel”. * Good presentation: WHLG: “The paper is clearly written… a pleasure to read”. Ctfm: “clarity of writing is excellent”. * Experiments: Zp5w: “The empirical evaluation is comprehensive”. WHLG: “The results indicate superior and stable performance…”. New experiments and results * As suggested by reviewers WHLG and 7z19, we added an amortized method as a benchmark for top-$k$ optimization experiments. Specifically, we chose PFNs4BO [1], a transformer-based amortized model designed for hyperparameter optimization. The final results are shown in Figure R1 of the rebuttal PDF. Our method outperforms PFNs4BO across all four tasks, as PFNs4BO does not consider the downstream task (i.e., top-$k$ optimization). We will update the results in the revised paper. * As asked by reviewers Zp5w, Ctfm, and 7z19, we ran an extra ablation study to evaluate the impact of the discount factor $\alpha$. When $\alpha=0$, our objective is purely myopic. We observed that, compared to other non-myopic settings, the policy did not learn to query designs effectively. This could be due to the sparse nature of rewards in this task; when the algorithm only considers immediate rewards, it struggles to learn the value of actions that lead to future rewards. We included the experimental results in the rebuttal PDF (Figure R2). * As suggested by reviewer Zp5w, we included a real-world experiment on retrosynthesis planning. Specifically, our task is to assist chemists in identifying the top-$k$ synthetic routes for a novel molecule, as selecting the most practical routes from many random routes generated by the retrosynthesis software can be troublesome. We trained our TNDP on a novel meta-dataset, including 1500 molecules and their routes collected by our collaborators. In this task, experimental design refers to selecting a route for a novel molecule to query its score. The downstream task is to recommend the top $k$ routes based on the collected data. Due to limited time, we compared only TNDP and the random search. The results in Figure R3 of the rebuttal PDF show that the utility of TNDP is significantly better than that of random search. We will include more baselines and provide a detailed problem description in the final paper. Clarification to Section 2.2 We thank the reviewers for raising thoughtful questions regarding the definition of $p(y_\Xi \| h_t)$ and the utility function. We acknowledge that this section lacks some details, and we would like to provide further explanations here. Our utility function is defined based on the measured outcomes ($y$) instead of the state of the world ($\theta$), as many downstream tasks directly rely on the predictions of outcomes for decision-making (see P4L124 in our paper as an example). It is a natural extension of the traditional definition of utility by marginalizing out the posterior distribution of $\theta$, a similar decision-theoretical setup can be found in [2]. As we are switching the belief about the state of the world (posterior) to the outcomes (posterior predictive) and to keep as much information as possible about the state of the world, we need to evaluate $\theta$’s effect on all points of the design space, thus, we define the utility based on $p(y_\Xi \| h_t)$, which is a stochastic process that defines a joint predictive distribution of outcomes indexed by the elements of the design set $\Xi$, given the current information $h_t$. We formulate the decisions in terms of this stochastic process, which differs from traditional utility based on individual observations, such as those defined in [3, 4]. A familiar example of our framework may be a decision process that depends on the observed values of a Gaussian process simultaneously evaluated at a large number of points. For example, in top-$k$ optimization, the goal is to select $k$ hyperparameter settings from a predefined finite set that maximize the cumulative accuracy. In this task, estimating the predictive distribution of a single hyperparameter setting is not sufficient for making the optimal decision. We need to determine the optimal decision based on the predictive distributions of all candidate hyperparameter settings. We adhere to the standard definitions of decision theory, but the entities now are stochastic processes instead of individual observations. Our architecture simultaneously amortizes two tasks. The first task is to amortize the predictive distribution needed for maximizing the utility during inference, which is similar to the goal of neural processes. When we can accurately predict $p(y_\Xi \| h_t)$, we can make optimal decisions. For example, if we can accurately predict the outcomes corresponding to all hyperparameter settings, we can directly determine the optimal set of hyperparameters. The second task is to amortize the design of experiments. Our goal is to enable the neural network to propose more informative designs, thereby allowing more accurate prediction of the outcome and facilitating optimal decision-making. We will include the above explanations in the revised paper. References [1] Müller et al. (2023). Pfns4bo: In-context learning for Bayesian optimization. ICML. [2] Kuśmierczyk et al. (2019). Variational Bayesian decision-making for continuous utilities. Neurips. [3] Lindley (1972). Bayesian statistics: A review. Society for industrial and applied mathematics. [4] Chaloner & Verdinelli (1995). Bayesian experimental design: A review. Statistical science. Pdf: /pdf/b61998d48eec3a506d6f2cca339696d675b7e77f.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Multi-Label Learning with Stronger Consistency Guarantees	Accept (poster)	Summary: This paper proposes an improved approach to multi-label learning using $\mathcal{H}$-consistency bounds by introducing the multi-label logistic loss to effectively handle label correlations. It extends to various multi-label losses, ensuring Bayes-consistency across diverse settings, and includes efficient gradient computation algorithms for minimizing the proposed loss function. This work offers a unified framework with robust consistency guarantees, advancing beyond traditional methods in multi-label learning. Strengths: - Introducing the multi-label logistic loss, which effectively addresses label correlations often overlooked by traditional binary relevance surrogates under Hamming loss. - The paper establishes $\mathcal{H}$-consistency bounds for a wide range of multi-label losses, ensuring Bayes-consistency across diverse multi-label learning scenarios. This extends beyond previous research that primarily focused on specific loss functions. - It offers a unified framework that accommodates various multi-label losses, including novel extensions and adaptations from standard classification. This is supported by efficient gradient computation algorithms specifically designed for minimizing the proposed multi-label logistic loss. Weaknesses: - The motivation and background of this paper lack clear logic and hierarchy. It is suggested to first outline the shortcomings of existing methods and then clearly present the research questions addressed in this paper. Technical Quality: 3 Clarity: 2 Questions for Authors: Please check the weaknesses. Confidence: 1 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your encouraging review. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions. Weaknesses: The motivation and background of this paper lack clear logic and hierarchy. It is suggested to first outline the shortcomings of existing methods and then clearly present the research questions addressed in this paper. Response: Thank you for the suggestion. Here is a list of shortcomings of existing methods and the research questions addressed in the paper: - Lack of Theoretical Analysis: Only a few studies focus on the theoretical analysis of multi-label learning, particularly the Bayes-consistency of surrogate losses. Can we present a comprehensive analysis of surrogate losses for multi-label learning and establish strong consistency guarantees? - Limited Bayes-Consistency: Existing methods only establish Bayes-consistency for specific loss functions. Can we derive a unified surrogate loss framework that is Bayes-consistent for any multi-label loss? - Drawback of Bayes-Consistency: Bayes-consistency is an asymptotic guarantee and does not provide convergence guarantees. It also applies only to the family of all measurable functions, unlike the restricted hypothesis sets typically used in practice. Can we leverage state-of-the-art consistency guarantees—$H$-consistency bounds—when designing surrogate loss functions for multi-label learning? - Sub-optimal Dependency on the Number of Labels: For the simplest form of multi-label loss, the popular Hamming loss, the well-known consistent binary relevance surrogate, when using smooth losses such as logistic losses, suffers from a sub-optimal dependency on the number of labels in terms of $H$-consistency bounds. Can we design smooth loss functions with improved dependency on the number of labels in their $H$-consistency bounds? - Label Correlations: One of the main concerns in multi-label learning is label correlations. The Bayes-consistent binary relevance surrogate fails to leverage label correlations. Can we design consistent loss functions that effectively benefit from label correlations as well? To address these drawbacks, we introduce a novel surrogate loss, multi-label logistic loss, that accounts for label correlations and benefits from label-independent $H$-consistency bounds. We then broaden our analysis to cover a more extensive family of multi-label losses, including all common ones and a new extension defined based on linear-fractional functions with respect to the confusion matrix. We also extend our multi-label logistic losses to more comprehensive multi-label comp-sum losses, adapting comp-sum losses from standard classification to multi-label learning. We prove that this family of surrogate losses benefits from $H$-consistency bounds, and thus Bayes-consistency, across any general multi-label loss. Our work thus proposes a unified surrogate loss framework that is Bayes-consistent for any multi-label loss, significantly expanding upon previous work which only established consistency for specific loss functions. Additionally, we adapt constrained losses from standard classification to multi-label constrained losses in a similar way, which also benefit from $H$-consistency bounds and thus Bayes-consistency for any multi-label loss. We further describe efficient gradient computation algorithms for minimizing the multi-label logistic loss. This unified framework holds promise for broader applications and opens new avenues for future research in multi-label learning and related areas. --- Rebuttal Comment 1.1: Comment: Thank you for thoroughly addressing my question and clarifying my doubts. As I am not familiar with this field, I will keep my score for now. --- Reply to Comment 1.1.1: Comment: Please let us know if we can provide further clarification regarding any question.	Summary: The paper explores surrogate losses and algorithms for multi-label learning, focusing on $ \mathcal{H} $-consistency bounds. It identifies the limitations of Hamming loss and introduces a new multi-label logistic loss that accounts for label correlations. The study extends this to a broader family of multi-label losses and adapts comp-sum losses from standard classification to multi-label learning. The authors propose a unified framework providing strong consistency guarantees for multi-label losses and describe efficient gradient computation methods for minimizing these losses. Strengths: 1. The authors conduct a detailed analysis of the popular Hamming loss in multi-label learning when using smooth losses. They identify its sub-optimal dependency on the number of labels and its failure to account for label correlations, providing valuable insights into the limitations of existing loss functions. 1. The authors introduce an improvement by presenting a novel surrogate loss, the multi-label logistic loss, which accounts for label correlations and benefits from label-independent $ \mathcal{H} $-consistency bounds. This innovation addresses the identified drawbacks of existing loss functions and broadens the analysis to include a more extensive family of multi-label losses, including a new extension based on linear-fractional functions related to the confusion matrix. 1. The authors extend their work by adapting multi-label logistic losses to more comprehensive multi-label comp-sum losses. By demonstrating that this family of surrogate losses benefits from $ \mathcal{H} $-consistency bounds and Bayes-consistency across any general multi-label loss, they propose a unified surrogate loss framework. This expands upon previous work that only established consistency for specific loss functions, showcasing the applicability of their approach. 1. The authors' writing is clear and well-structured, with each theoretical assumption and conclusion articulated distinctly. Weaknesses: 1. In section 4, although the excellent properties of the proposed multi-label logistic loss are proven, providing a detailed explanation of each component of this loss would further enhance the reader's understanding of its superiority. 2. If the advantages of this loss could be demonstrated through experimental validation, it would be more intuitive for readers. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the authors elaborate on the individual components of the multi-label logistic loss and how each contributes to its overall effectiveness? 2. Given the detailed nature of this loss function, what is the computational complexity associated with implementing the multi-label logistic loss compared to other traditional loss functions? It is better to be able to verify the advantages and complexity of the algorithm through experiments. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your appreciation of our work. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions. Weakness 1. In section 4, although the excellent properties of the proposed multi-label logistic loss are proven, providing a detailed explanation of each component of this loss would further enhance the reader's understanding of its superiority. Question 1. Could the authors elaborate on the individual components of the multi-label logistic loss and how each contributes to its overall effectiveness? Response: That's an excellent question. The component $\left( 1 - \overline{\mathsf{L}}_ {\mathrm{ham}}(\cdot, y) \right)$ acts as a weight vector for each logistic loss corresponding to the label $y'$. The term $\sum_{i = 1}^{l} \left( y''_ i - y'_ i \right) h(x, i)$ represents the difference in the scores between the label $y'$ and any other label $y''$, where these scores account for the correlations among the labels $y_i$ within the logarithmic function. The logarithmic term increases as the difference in scores increases. Therefore, the loss function imposes a greater penalty on larger differences through the penalty term $\left( 1 - \overline{\mathsf{L}}_ {\mathrm{ham}}(y', y) \right)$, which is dependent on the Hamming losses assigned to the prediction $y'$ and true label $y$. We will add a more detailed explanation in the final version. Weakness 2. If the advantages of this loss could be demonstrated through experimental validation, it would be more intuitive for readers. Question 2. Given the detailed nature of this loss function, what is the computational complexity associated with implementing the multi-label logistic loss compared to other traditional loss functions? It is better to be able to verify the advantages and complexity of the algorithm through experiments. Response: Thank you for your valuable feedback. As shown in Section 7, the computational complexity for optimizing and implementing the multi-label logistic loss is $O(l)$, modulo the precomputed quantities, which is comparable to that of other common multi-label surrogate losses. As you noted, this paper is primarily theoretical and algorithmic, and establishes a sound foundation for multi-label surrogate losses, backed by $H$-consistency bounds. Our framework offers a unique, unifying approach that ensures Bayes-consistency for any multi-label loss, a significant advantage over existing methods. While we have demonstrated that efficient algorithms can minimize multi-label logistic loss, we recognize the importance of further exploration. We agree that empirical comparisons with common multi-label surrogate losses would strengthen our work, and we will strive to include these in the final version. In future work, we are excited to expand upon this foundation with extensive empirical analyses. --- Rebuttal Comment 1.1: Comment: Thank you for your reply. After referring to the comments of other reviewers, I decided to maintain my score. --- Reply to Comment 1.1.1: Comment: Thank you for your comments. We appreciate the reviewer's valuable feedback and constructive suggestions.	Summary: The authors study surrogate losses and algorithms for multi-label learning via H-consistency bounds and introduce a novel surrogate loss, multi-label logistic loss in this paper. By broadening the H-consistency bounds analyses to more general multi-label losses and extending to multi-label comp-sum losses, the authors provide a unified surrogate loss framework for H-consistency. Strengths: 1. This paper is well-written and easy to follow. 2. The authors make comprehensive reviews of related works, including their pros and cons. 3. The authors provide rigorous theoretical analyses of the limitations of existing binary relevance loss, the H-consistency of the proposed multi-label logistic loss, and the extensions to more general multi-label losses. The theoretical contribution is important for multi-label learning. 4. The authors demonstrate the efficient computation of the gradient for the proposed multi-label logistic loss and conduct time complexity analyses. Weaknesses: 1. I understand that this is a theoretical work, and experiments of empirical evaluations are not its focus. However, adding experiments to compare the proposed loss with commonly used multi-label losses on standard datasets would make the paper more comprehensive and appealing. Besides, it can also verify whether the proposed loss is effective in practice. 2. There is a typo in line 300.($1-\bar{L}_{ham}(\cdot, y)$). Technical Quality: 3 Clarity: 3 Questions for Authors: See above weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your appreciation of our work. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions. 1. I understand that this is a theoretical work, and experiments of empirical evaluations are not its focus. However, adding experiments to compare the proposed loss with commonly used multi-label losses on standard datasets would make the paper more comprehensive and appealing. Besides, it can also verify whether the proposed loss is effective in practice. Response: Thank you for your valuable feedback. As you noted, this paper is primarily theoretical and algorithmic, and establishes a sound foundation for multi-label surrogate losses, backed by $H$-consistency bounds. Our framework offers a unique, unifying approach that ensures Bayes-consistency for any multi-label loss, a significant advantage over existing methods. While we have demonstrated that efficient algorithms can minimize multi-label logistic loss, we recognize the importance of further exploration. We agree that empirical comparisons with common multi-label surrogate losses would strengthen our work, and we will strive to include these in the final version. In future work, we are excited to expand upon this foundation with extensive empirical analyses. 2. There is a typo in line 300. $(1 - \overline{L}_{\mathrm{ham}}(\cdot, y)).$ Response: Thank you, we will correct it. --- Rebuttal 2: Comment: Thanks for your responses. I have thoroughly reviewed the comments from other reviewers and the corresponding responses. I look forward to seeing the experimental results in the final version. I have no further questions at this time and have decided to maintain my current score. --- Rebuttal Comment 2.1: Comment: Thank you for your comments. We will strive to include experimental results in the final version. We appreciate the reviewer’s support of our work and their valuable suggestions.	Summary: The paper derives H-consistency bounds for binary-relevance style surrogate losses, as well as a new surrogate, for mutli-label learning problems, showing that the proposed multi-label logistic loss whose upper-bound on the Hamming loss is independent of the number of labels. Strengths: The $H$-consistency bounds provided in the paper are more informative than existing Bayes-consistency results, as they hold not just in the infinite limit. The novel multi-label logistic loss allows upper-bounds that do not depend on the number of labels. Weaknesses: The paper does not provide any experiments. While this is OK for a theory paper, it does mean that the question of whether the new surrogate works better in practice remains unanswered (which should be reflected in the conclusion section, at least), for two reasons: a) all the theory provides are upper-bounds, which might not be indicative of actual performance b) while the theory provides better guarantees for the task loss if the surrogate is reduced to the level $\epsilon$, it might be that reducing the new surrogate is just much more difficult than optimizing binary relevance. In particular, if the computational cost for reducing the multi-label logistic loss to the same level $\epsilon$ as binary relevance is larger by at least $\sqrt{l}$, then, normalized for compute, the advantage of the new surrogate vanishes. It is claimed that the gradient of the multi-label logistic loss can be computed efficiently, yet the presented formulas still contain sums over the entire $2^l$ entries of the label space. Even if they can be precomputed once, already at moderate label space sizes of l ~ 100 would these quantities be intractable. It is annoying that most equations are unnumbered. Even if they are not referred to in the paper, your readers and reviewers might want to reference them. the equation after l. 328 switches between $\mathbf{\mathsf{y}}'$ and $y'$; and $y''$ changes to $y$ l. 114: I'm not sure what the point here is of introducing the threshold $t$, if it is set to $0$ in the same sentence? Couldn't $t$ be simply absorbed into $h$? l. 178-180; 208: Arguably, completeness does _not_ hold in practice, because there is some form of upper-bound (e.g., weights representable in the given floating-point format) l. 231. Binary relevance is not just Bayes-consistent w.r.t. the Hamming-loss, but also works for precision-at-$k$ In the equation after line 542, I think $\bar{L}$ should be $\bar{L}_\mathrm{ham}$? l. 503: I think $q$ should be $q_i$, and there is a weird subscript on that line. l. 174 consist -> consisting Technical Quality: 3 Clarity: 3 Questions for Authors: In several places, the paper talks about label correlations, in particular, it claims an advantage of the new surrogate is that it takes into account label correlations. However, it is never specified what exactly that means (conditional correlations, i.e., dependent on the specific instance $x$, or marginal correlations). Further, for many loss functions (such as Hamming-loss), the Bayes-optimal prediction is a function of purely the label marginals $P[Y_i\|X]$, so it is not clear to me whether taking into account label correlations actually is an advantage in those cases. The paper mentions the decision-theoretic and the empirical-utility framework, but then seems to consider only loss functions that are defined on the level of a single instance. Aren't the two settings that same in that case? l. 525: Is the argmin unique? Are we breaking ties arbitrarily? Despite being part of the theorem, $\mathcal{M}$ does not appear anywhere in the proof of 3.1 I tried going through the proof of 4.1, but I'm not quite sure how to construct the hypothesis $h'$ with that realized $s^{\mu}$, not do I see why the minimum is achieved for $s_h = s_y$, unless $c_h = c_y$. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: I'm not sure if the proposed surrogate actually is tractable for label spaces with more than 50 labels. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful feedback and suggestions on improving the readability. We will take them all into account when preparing the final version. Below, please find our responses to specific questions. Weaknesses: 1. The paper does not ... Response: Thank you for your insightful comments. As you noted, our primary focus is theoretical and algorithmic: the design of theoretically principled surrogate losses for multi-label learning, supported by H-consistency bounds. We observe that some key previous publications on the topic, such as (Gao and Zhou, 2011), do not include any empirical result either. Nevertheless, we plan to include empirical results with commonly used multi-label surrogate losses in the final version of our paper, as you suggested. Our proposed framework offers significant progress by providing a unified surrogate loss that guarantees consistency for any multi-label loss. This contrasts with existing approaches, which achieve Bayes-consistency only for specific loss functions. While we have demonstrated the existence of efficient algorithms for minimizing multi-label logistic loss, we acknowledge the need for further exploration. Our future work will focus on extensive empirical analysis and the development of more universally applicable algorithmic solutions to cover a broader range of surrogate loss functions and diverse target losses. We agree with your observation that optimization is a crucial factor influencing the choice of surrogate losses and their practical performance, in addition to stronger consistency guarantees. The selection of surrogate losses and algorithms in practice should indeed consider multiple factors, including consistency, approximation, optimization properties, label dependency, and label correlation. We hope our work provides useful theoretical insights and potential alternatives helpful for this selection, with more extensive empirical analysis to follow. In conclusion, we believe our unified surrogate loss framework, which establishes strong consistency results for multi-label losses, represents a significant theoretical contribution. We are committed to further exploring the empirical aspects of this framework and developing practical solutions in future work. 2. It is claimed that ... Limitations: ... Response: For many standard loss functions, the terms involving sums over the entire label space can be computed analytically. We illustrate this below for the Hamming loss and $F_{\beta}$ measure loss functions :$\sum_{y} ( 1 - \overline{ \mathsf L}_ {\mathrm{ham}}(y, y^j) )= -(l - 1) 2^{l} + \sum_{y} \sum_{i = 1}^{l} 1_{y_i = y^j_i} = 2^{l - 1} (2 - l)$, and $ \sum_{y} ( 1 - \overline{ \mathsf L}_ {F_ {\beta}}(y, y^j) ) = \sum_{y} \frac{(1 + \beta^2) y \cdot y^j }{\beta^2 \\| y \\|_1 + \\| y^j \\|_1}$ $= \sum_{k = 0}^{l} \frac{(1 + \beta^2) \\| y^j \\|_1\binom{l - 1}{k - 1} }{\beta^2 k + \\| y^j \\|_1}$. A similar analysis can be used for many other loss functions. Therefore, the presence of these terms does not impact the tractability of our algorithms. Additionally, as noted in Section 7, these terms can be precomputed once and reused, regardless of the specific sample or task under consideration. Miscellaneous issues: Thank you for pointing these out. We will number the equations, correct the typos and refine the statements accordingly. Questions: 1. In several ... Response: In multi-label learning, label correlation simply means that certain pairs of labels (e.g., "cup" and "mug") tend to co-occur more frequently than others (e.g., "cup" and "umbrella") as labels for input points. Leveraging these correlations can significantly enhance the efficiency of multi-label learning. Approaches like binary relevance surrogate loss treat each label independently, missing the opportunity to exploit these inherent relationships. Our new form of surrogate losses directly takes into account such correlations among labels. Both the binary relevance surrogate loss and our new surrogate loss are Bayes-consistent, meaning that minimizing them over the family of all measurable functions approximates the Bayes-optimal solution. However, our new surrogate losses that consider label correlations can converge faster, which is reflected in their more favorable $H$-consistency bounds, independent of the number of labels. We will formalize these concepts and provide a more detailed discussion in the final version. 2. The paper mentions ... Response: In the decision-theoretic analysis (DTA) framework, a loss function defined as a function over a single instance is considered, and the measure is defined as the expected loss, also known as the generalization error (expectation of a loss function over samples). In the empirical utility maximization (EUM) framework, the measures are directly defined as functions of the population (a function of an expectation over samples). In our paper, we adhere to the DTA framework by analyzing the loss functions and their consistency guarantees in multi-label learning. 3. l. 525 ... Response: Any fixed deterministic strategy can be used to break ties. For example, we can choose the label with the lowest index under the natural ordering of labels as the tie-breaking strategy. We will elaborate on this in the final version. 4. Despite ... Response: The minimizability gaps appear after taking the expectation on both sides of the inequality between lines 506 and 507, where we used the concavity of the function $\Gamma$ and Jensen's inequality. We will elaborate on this in the final version. 5. I tried ... Response: The realization is due to the completeness assumption. The minimum is achieved for $\mathsf s_{\mathsf h} = \mathsf s_{\mathsf y}$ because $\mathsf c_{\mathsf h} \geq \mathsf c_{\mathsf y}$ and $\mathsf s_{\mathsf h} \geq \mathsf s_{\mathsf y}$ by definition. We will elaborate on these in the final version. --- Rebuttal 2: Comment: To clarify, my main critique is not that the paper doesn't have any experiments; it is that it makes claims that extend beyond the purely theoretical, but these claims are not actually verified. To me, that could be resolved either way: Remove these claims and have a pure theory paper Add experiments that verify these claims For example, if you had written an article that introduced Strassen multiplication and claimed that it would lead to real-world speed-ups in matrix multiplication without providing an actual implementation, I would have found that problematic, it is _very_ difficult to implement Strassen on actual hardware so that it beats regular matrix multiplication algorithms. One of my concerns, I believe not addressed in the rebuttal, is that, as far as I can see, in the general case, the precomputation may in itself be exponential in the number of labels. Overall, though, I do think that most of these points can be addressed in a camera-ready version by more careful writing, and therefore I will raise my score.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
On Differentially Private U Statistics	Accept (poster)	Summary: This paper addresses the problem of estimating U statistics under central differential privacy. U statistics are established minimum variance unbiased estimators for estimable parameters in the form $\mathbb{E} h (X_1, ..., X_k)$, where $h$ is a kernel and for all $i$ $X_i$ is i.i.d. from some underlying distribution. In other words, U statistics estimate averages of kernels applied to subsets of the data of degree (size) $k$. This type of problem arises in multiple statistical tests such as goodness-of-fit tests and Pearsons's chi-squared tests, uniformity testing, subsampling and other scenarios. While many methods have been studied for differentially private mean estimation, the research on private U statistics is in its early stage and has so far mainly focused on local differential privacy models and discrete data. This paper seeks to provide differentially private U statistics estimators achieving nearly optimal private error for both the case of non-degenerate kernels and degenerate kernels. The main contributions of this paper are: i) it derives the lower bound for private algorithms for the non-degenerate kernel case (Theorem 1); ii) it finds that applying off-the-shelf private mean estimation procedures to U statistics estimation yields suboptimal error; iii) it proposes an algorithm that achieves nearly optimal private error in the non-degenerate kernel case, and evidence of near optimality for bounded degenerate kernels. The proposed algorithm (Algorithm 1) is based on representing U statistics via the Hájek projection, and leverages the fact that local Hájek projections enjoy strong concentration around the conditional mean. Basically, if all local Hájek projections $\hat h(i)$ are within a certain threshold distance from the pre-computed empirical mean $A_n$, the output $\tilde{A}_n$ on line 14 is going to be equal to $A_n$; if not, for every subset $S$ containing a bad index, $h(S)$ is replaced by a weighted combination of $h(S)$ and $A_n$. The choice of threshold $\xi$ ensures $L = 1$ with high probability, maintaining a balance between excluding bad data and preserving good data, while also keeping the sensitivity of the final adjusted mean $\tilde{A}_n$ small, which is crucial for differential privacy. A lower bound for sub-Gaussian non-degenerate kernels is provided (Corollary 1) and Algorithm 1 is proven to match this lower bound. It is also shown that Algorithm 1 matches the lower bound for bounded degenerate kernels (Corollary 2). The paper discusses a wide range of applications of the proposed method to uniformity testing, goodness-of-fit tests, Pearson’s chi-squared tests, symmetry testing, and sparse graph statistics. Strengths: This paper is clear, well-structured and provides rigorous derivations and proofs to back the proposed methods and claims. The paper addresses a notable gap in current differential privacy research, which is U statistics under differential privacy. The authors derive lower bounds for both the private sub-Gaussian non-degenerate kernel case and the private bounded degenerate kernel case. These bounds support the proofs that the proposed method achieves i) near-optimality for sub-Gaussian non-degenerate kernels and ii) strong evidence of near optimality for the bounded degenerate case. These results are valuable in the context of the differential privacy research community. The contributions are clearly highlighted. I appreciate the effort by the authors to make the results as clear as possible for the reader. In particular, the table summary of the error of different private methods in Table 1 makes it easy to understand the relative error performance of different methods at glance; similarly, in a couple of instances the authors provide key intuitions behind the proposed methods, which helps break down important technical steps that are fundamental to the proposed method. The notation is also clear and consistent. The proposed method has wide applicability, as demonstrated in the Applications section, where the authors describe the usefulness of the method spanning multiple statistical tests and sparse graph statistics. Computational complexity and alternative computationally efficient approximations of U statistics are also discussed. Extensive proofs and supporting technical derivations are provided in Appendix, although I did not review it in detail due to time constraints. Weaknesses: I didn’t find any significant weaknesses in this paper. The paper is highly technical and notation-heavy, but as I described in the previous section, it still reads very clearly. A few of minor notes: - Since [53] appears to be foundational to the development of the main proposed method, it is worth dedicating a short description of it and/or specification of which ideas in [53] have been built upon. - Theorem 2 is not followed by a pointer to its proof in Appendix. Please reference the proof in Appendix. - Limitations of the proposed methods are briefly mentioned throughout the paper, but I would prefer if they were addressed separately in a short dedicated paragraph or subsection, making them more easily identifiable by a reader skimming through the paper. Technical Quality: 4 Clarity: 4 Questions for Authors: I would ask the authors to address the minor points I mentioned in "Weaknesses". I don't have other questions at the moment. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Limitations of the proposed method are sparsely mentioned throughout the paper. As I mentioned under "Weaknesses", it would be preferable to add a dedicated paragraph to the limitations, even if short. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for mentioning that our work addresses a notable gap in differential privacy research and for your kind words on its wide applicability. In a revision, we will add pointers to the proofs of all theorems immediately after their statements. [Re: Connection between [1] and our algorithm]: We will more clearly describe the connection of our algorithm with the algorithm of [1]. A key idea in this work is to exploit the concentration of the degrees of Erdős-Renyi graphs, and we generalize this idea to the broader setting of U-statistics as follows. Consider a $k$-uniform complete hypergraph with $n$ nodes (and ${n\choose k}$ edges), where the nodes are the data indices. An edge corresponds to a $k$-tuple of data points $S\in I_{n,k}$, where $I_{n,k}$ is the family of all $k$-element subsets of $[n]$, and the weight of this edge is $h(X_S)$. The local Hájek projection $\frac{1}{\binom{n-1}{k-1}} \sum_{i \in S} h(X_S)$ is simply the degree normalized by ${n-1\choose k-1}$. Our algorithm uses the property of local Hájek projections to re-weight the hyperedges ($k$-tuples) such that the local sensitivity of the re-weighted U-statistic is small. In degenerate cases and in cases where $ \zeta_1 \ll \zeta_k/k$, where $\zeta_1 = \textup{var}(\mathbb{E}[h(X_1, X_2, \dots, X_k)\|X_1])$ is the variance of the conditional expectation and $\zeta_k = \textup{var}(h(X_1,X_2, \dots, X_k))$, similar to the Erdős-Renyi case, the local Hájek projections concentrate tightly around the mean $\theta$, leading to a near-optimal error guarantee. It turns out that even when the U statistic is non-degenerate and the Hájek projections do not concentrate as strongly, our algorithm (Algorithm 1) achieves near-optimal private error guarantees. Algorithm 1 also works with subsampled family $\mathcal{S} \subseteq I_{n,k}$, where the size of $\mathcal{S}$ can be as small as $\tilde{O}(n^2/k^2)$. This allows for a computationally efficient algorithm for all $n$ and $k$. [Re: Limitations]: As per your suggestion, we will compile the limitations of our methods from different parts of the paper to one dedicated section. [1] J. Ullman and A. Sealfon. Efficiently estimating Erdos-Renyi graphs with node differential privacy. Advances in Neural Information Processing Systems, 32, 2019. --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal. I appreciate the additional details regarding the connection between [1] and your proposed method, which I believe will enhance the paper. I confirm my positive evaluation of your submission. --- Rebuttal 2: Comment: Dear reviewer B5Tq, Thank you - we really appreciate your support. We will definitely include the discussion regarding the connection between [1] and our method in the paper.	Summary: The paper addresses the problem of private estimation of U-statistics. The authors propose a new thresholding-based approach using local Hájek projections to achieve nearly optimal private error in both non-degenerate and degenerate settings. Strengths: 1. The paper provides solid theoretical foundations, including lower bounds for the private error and theoretical guarantees for the proposed algorithm. 2. The proposed method is applicable to a wide range of U-statistics problems, from hypothesis testing to subgraph counting in random geometric graphs. 3. The method aims to provide private confidence intervals for U-statistics, addressing a gap in existing literature. Weaknesses: 1. The paper is difficult to read due to the heavy use of parameters and notations, many of which are not well-defined or explained, particularly in the algorithmic sections. 2. The manuscript provides non-asymptotic results for the DP estimators, but lacks the asymptotic normality results typical for non-private version of U-statistics, which are crucial for practical applications. I think the asymptotic variance of the private U-statistics will change compared to the non-private version. More discussion on expected on this difference. 3. To provide private confidence intervals, the variance should also be estimated privately. This aspect is not thoroughly discussed, making the testing problem in Section 5 less meaningful. 4. There are no experimental results to demonstrate the practical performance of the proposed algorithms, which is a significant omission. 5. The paper only consider the 1-dimensional data $X$ throughout the paper. A general discussion of d-dimensional vector are needed because it may suffer from the curse of dimensionality, which will affect the generalizability of the results. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. What is the asymptotic results of the private U-statistics? 2. How do you ensure get DP estimators for the variance when doing inference? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: 1. The paper should discuss the differences and potential advantages of the proposed method compared to directly adding noise to the estimators. 2. The authors should include asymptotic normality results for the DP estimators, similar to those available for non-private U-statistics. 3. The paper would benefit significantly from experiments that validate the theoretical findings and demonstrate the practical applicability of the proposed methods. 4. The author should specify the dimension of $X$ and discuss its impact on the results. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your kind words regarding the solid theoretical foundations and wide applicability of our work. [Re: Asymptotic distribution]: To our knowledge, differential privacy results typically focus on finite sample guarantees. _We show under mild conditions on $n,k,$ and $\epsilon$ that our estimator has the same asymptotic distribution as the non-private U statistic in the degenerate (of order $1$) and non-degenerate cases. Thus, the asymptotic variance doesn’t change._ For simplicity, let $k=O(1)$ and $\mathcal{S}=I_{n,k}$, the set of all size $k$ subsets of $[n]$. Consider a scaling factor $c_n$, set to $n$ ($\sqrt{n}$) for degenerate (non-degenerate) kernels. Recall that our estimator $\mathcal{A}(X)=\tilde{A}_n+S(X)/\epsilon Z$, where $Z$ has density $f(z)\propto 1/(1+z^4)$. We have, $$c_n(\mathcal{A}(X)-\theta)=c_n(\tilde{A}_n-A_n)+c_n(A_n-\theta)+c_n \frac{S(X)}{\epsilon} Z.$$ For $\mathcal{S}=I_{n,k}$, $A_n=U_n$ and hence $c_n(A_n-\theta)$ converges to either a Gaussian ($h$ is non-degenerate, $c_n=\sqrt{n}$) or weighted sum of centered Chi-squared distributions ($h$ is degenerate, $c_n=n$). We show that for any choice of $c_n$, the second term is $o_P(1)$. Define $\mathcal{E}_1=\{\exists i: \|\hat{h}(i)-\theta\|\geq C_1\sqrt{k/n}\log(n/\alpha)\}$ and $\mathcal{E}_2=\{\|A_n-\theta\|\geq C_1\sqrt{k/n}\log(n/\alpha)\}$. By Lemmas A.21 (line 854) and A.3 (line 494), $P(c_n\|\tilde{A}_n-A_n\|\geq t)\leq P(\tilde{A}_n\neq A_n)\leq P(\mathcal{E}_1)+P(\mathcal{E}_2)\leq 2\alpha$. For non-degenerate subgaussian kernels, $h$ is truncated using one-half of the data. So the same argument as in lines 201-208 shows that the probability that none of the $h(X_S)$ is truncated is at most $\alpha$. For finite-sample guarantees, we take $\alpha$ as a constant and then boost it via median-of-means. For distributional convergence, we take $\alpha=n^{-c}, c>0$. Thus the first term is $o_P(1)$. Finally, conditioned on $\mathcal{E}_1\cap\mathcal{E}_2$, $L=1$ for $\xi=\tilde{O}(\sqrt{k/n})$ (Lemma A.21, line 854) in the degenerate case and $\xi=\tilde{O}(\sqrt{k\tau})$ (Lemma A.22, line 863) in the non-degenerate case. Some calculations show that as long as $\frac{k^3}{n\epsilon^2}=o(1)$, $c_n S(X)/\epsilon =o_P(1)$. Since $Z = O_P(1)$, the third term is also $o_P(1)$. Overall, $\|c_n(\mathcal{A}(X)-\theta) - c_n(A_n-\theta)\| = o_P(1)$, establishing that the asymptotic distributions of the private and non-private estimates are the same. [Re: Private confidence intervals]: Our goal is not to provide a confidence interval but finite sample error guarantees. If desired, the variance of a U-statistic could be computed using our method since it also involves a U-statistic of degree $2k-1$. However, our algorithms require an upper bound on the ratio between $\tau$ and the $\zeta_k = \textup{var}(h(X_S))$, similar to the state-of-the-art algorithms for private mean estimation [1,2]. Lemma A.7 (lines 588-591) shows how to estimate the variance $\zeta_k$ within a multiplicative factor. We will discuss this in more detail. [Re: Dimension]: First, we do not require $X_i$ to be scalars; rather, $h(X_S)$ is assumed to be a scalar. In fact, for our application to sparse graphs, $X_i$ are latent vectors in $\mathbb{R}^d$ (lines 301-304). _In fact, our techniques easily generalize when $h(X_S)\in \mathbb{R}^d$, as we show next._ Lemma 2.10 in [3] shows that if $Z\sim N(0,I_d)$ is a $d$ dimensional Gaussian and $S(X)$ is a $\beta$-smooth upper bound on the local sensitivity, then adding noise equal to $S(X)/\alpha Z$ achieves $(\epsilon,\delta)$-DP, where $\alpha = \frac{\epsilon}{5\sqrt{2 \log(2/\delta)}}$ and $\beta=\frac{\epsilon}{4(d+\log(2/\delta))}$. For simplicity, let $\mathcal{S} = I_{n,k}$. Consider $h(X_S)$ lying in an $\ell_2$-ball of radius $C$. In Algorithm 1, the sets in lines 5 and 6 should be modified to have the $\ell_2$ distance $\|\|\hat{h}(i)-A_n\|\|_2$. In line 8, the weights to the indices should be based on the ($\ell_2$) distances from the ball $\mathbb{B}\left(0, \xi+4kCL/n\right)$. Finally, the $\epsilon$ in every step except line 15 should be replaced with $\beta$, the $\epsilon$ in the last step should be replaced with $\alpha$, and $Z \sim \mathcal{N}(0, I_d)$. We can show that Lemma A.20 (line 831) holds with $\beta$ as above. With probability at least 3/4, $L = 1$ and $S(X) = O\left(\frac{k\xi}{n}+\frac{k^2Cd}{n^2\epsilon}+\frac{k^2Cd}{n^2\epsilon} + \frac{k^3Cd^2}{n^3\epsilon^2}\right).$ The noise added in line 15 is $S(X)/\alpha \cdot Z$. In all, with constant success probability, $$\|\|\mathcal{A}(X)-\theta\|\|_2 \le O\left( \sqrt{\textup{Tr}(\textup{Cov}(U_n))}+\frac{k\xi d^{1/2}}{n \epsilon}+\frac{k^2Cd^{3/2}}{n^2\epsilon^{2}} + \frac{k^3Cd^{5/2}}{n^3 \epsilon^{3}}\right).$$ To see that this is a desirable result, consider the task of mean estimation of $N(\mu, I_d)$ random vectors with $\tau = 1, k = 1, h(X) = X, C = O(\sqrt{dk\tau \log n/\alpha})$, (as in Corollary 1, line 209) and $\xi = O(\sqrt{d\tau\log n/\alpha})$. Then, our algorithm achieves error $$\|\|\mathcal{A}(X)-\theta\|\|_2 \le \tilde{O}\left( \sqrt{\frac{d}{n}} + \frac{d}{n \epsilon} + \frac{d^{2}}{n^2\epsilon^{2}} + \frac{d^{3}}{n^3\epsilon^{3}}\right).$$ This error is at most $\eta$ as long as $ n \gtrsim \frac{d}{\eta^2} + \frac{d}{\eta\epsilon},$ nearly matching the $\tilde{\Omega}\left(\frac{d}{\eta^2} + \frac{d}{\eta\epsilon}\right)$ lower bound (see Corollary 3.13. of [4]). [1] G. Brown, S. Hopkins, and A. Smith. "Fast, sample-efficient, affine-invariant private mean and covariance estimation for subgaussian distributions." COLT 2023. [2] J. Duchi, S. Haque, and R. Kuditipudi. "A fast algorithm for adaptive private mean estimation." COLT 2023. [3] K. Nissim, S. Raskhodnikova, and A. Smith. Smooth sensitivity and sampling in private data analysis. STOC 2007. [4] X. Liu, W. Kong, and S. Oh. "Differential privacy and robust statistics in high dimensions." COLT 2022. --- Rebuttal 2: Title: Further clarifications Comment: Dear Reviewer 8Ha8, we hope our rebuttal has answered most of your questions about dimensionality, distributional convergence, and variance estimation. We realize that we somehow overlooked your question about directly adding Laplace noise in the rebuttal. We wanted to take this opportunity to point out that when $h(X_S)$ has additive range $C$, the sensitivity is $kC/n$. Thus, adding Laplace noise with parameter $\frac{kC}{n\epsilon}$ gives an $\epsilon$-private estimate of $E[h(X_S)]$. However, one cannot compute sensitivity without truncation in the unbounded case. As a baseline, we provide the performance of an adapted version of the Coinpress algorithm (which is a state-of-the-art private mean estimation method) in Lemma 3 (Line 130). The algorithm is in Appendix Section A.3 (Algorithm A.2). Table 1 in Section~3 shows that adding noise in accordance to Coinpress, or Laplace noise with parameter $kC/n\epsilon$, overwhelms the non-private error in the degenerate case. Even for non-degenerate settings, the $\zeta_1$ can be much smaller than $C$. For example, the uniformity testing case (Appendix Lemma A.24) shows that $\zeta_1$ can be much smaller than $C$ when $m$ is large. In these settings, the directly added Laplace noise will dominate the non-private error, whereas the non-private error in our method will dominate the error resulting from privacy. If we can answer any more of your questions, please let us know. --- Rebuttal Comment 2.1: Comment: Thanks for the author's response, which addressed most of my concerns. I decide to raise my score to 5. --- Rebuttal 3: Comment: Dear reviewer 8Ha8, Thank you very much for raising your score. We are very happy that we were able to address most of your concerns, and we will update the manuscript accordingly.	Summary: This paper introduces a new algorithm for constructing U-statistics under central DP. Compared to the naive method, the proposed estimator exhibits lower variance. The authors also derive a lower bound for private algorithms. Several statistical applications are presented to illustrate the methodology. Strengths: U-statistics are widely applied in statistical inference. The improvements in private estimation presented in this paper are useful, and the theoretical results are solid. Weaknesses: The calculation of the privacy budget lacks precision. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the authors consider refining the computation of the privacy budget? Specifically, users may prefer a $\epsilon$-DP method over an $\mathcal{O}(\epsilon)$-DP method. 2. Following the first point, could the authors discuss the performance of the proposed method across different values of $\epsilon$? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [Re: Privacy budget]: Lemma 3 (line 130) shows that the CoinPress algorithm from [2], adapted to the all-tuples family, is $2\epsilon$-DP. The following argument shows that Algorithm 2 is $10\epsilon$-DP as stated. Corollary 2.4 in [1] shows that for any function $f:\mathcal{X} \to \mathbb{R}$, the output $f(X) + \frac{2(\eta+1)S(X)}{\epsilon} Z$, where $Z$ is sampled from the distribution with density $\propto \frac{1}{1+\|z\|^\eta}$ and $S(X)$ is a $\beta$ smooth upper bound on the local sensitivity $LS(X)$ of $f$, is $\epsilon$-DP as long as $\beta \le \frac{\epsilon}{2(1+\eta)}$ and $\eta>1$. Thus, as stated with $\eta=4$, Algorithm 1 (Theorem 2, line 196), is $10\epsilon$-DP. We chose $\eta = 4$ somewhat arbitrarily, and we can improve this to almost $4\epsilon$-DP by choosing $\eta$ arbitrarily close to $1$. Corollary 1 (line 209) follows by composing our extension of the algorithm from [2] and our main algorithm. Passing in $\epsilon/2$ to Algorithm A.2. with $\mathcal{S} = \mathcal{I}_{n,k}$ and $\epsilon/20$ (which can be improved to $\epsilon/8$ from the discussion above) to our main algorithm results in an $\epsilon$-DP algorithm for non-degenerate, subgaussian kernels. [Re: Performance across different $\epsilon$]: We assume $\epsilon=O(1)$. But the dependence of the private error of our algorithm on $\epsilon$ nearly matches our lower bound for both degenerate and non-degenerate settings. See Table 1 (page 4). [1] K. Nissim, S. Raskhodnikova, and A. Smith. 2007. Smooth sensitivity and sampling in private data analysis. In Proceedings of the thirty-ninth annual ACM symposium on Theory of computing (STOC '07). Association for Computing Machinery, New York, NY, USA, 75–84. [2] S. Biswas, Y. Dong, G. Kamath, and J. Ullman. Coinpress: Practical private mean and covariance estimation. Advances in Neural Information Processing Systems, 33:14475–14485, 2020. --- Rebuttal Comment 1.1: Comment: Dear reviewer MPw8, We hope that our rebuttal has adequately answered your questions on the privacy budget. If there are any further questions we can answer, please do let us know. --- Rebuttal Comment 1.2: Comment: Thank you for the authors' response. I have no further questions. I trust the authors will revise the manuscript appropriately, taking these remarks into account. The score has been raised to 6. --- Rebuttal 2: Comment: Dear Reviewer MPw8, Thank you very much for your response and for raising your score. We will definitely revise our manuscript to incorporate these remarks.	Summary: This paper studies differentially private estimation of U-statistics (estimators for such statistics are averages of functions $h$ that depend on a number of i.i.d. samples $X_1,\dots,X_k$). This is a generalization of the commonly studied mean estimation problem where $k=1$ and such estimators with $k>1$ are widely applied across statistics. The authors are primarily interested in cases where $h$ is a subgaussian kernel i.e. the distribution of $h(X^k)$ is subgaussian or cases where the range of $h$ is bounded (and satisfies a certain degeneracy property). The main contributions of the paper are as follows: 1) They first consider approaches that reduce differentially private U-statistics to differentially private mean estimation and argue that natural approaches result in estimators that are either suboptimal in either the non-private error terms or the private error-terms. The estimators they consider are a naive estimator that reduces to the i.i.d. case by computing the function $h$ on a partition of the dataset before applying a subgaussian mean estimation algorithm on the resulting sample of function values, and a more complicated estimator that generalizes the CoinPress algorithm to work with weakly dependent samples. The former has suboptimal non-private error while the latter has a suboptimal privacy term (the dependence on $k$ is suboptimal). 2) They then consider a different strategy inspired by work on privately estimating the sampling probability for Erdos-Renyi graphs. This strategy exploits the concentration of the 'local Hajek projections' around the true mean. The idea is to classify coordinates into good and bad coordinates respectively based on how close their projections are to the optimal non-private statistic, and reduce the local sensitivity of the average being computed by reducing the influence of bad coordinates by reducing the weight of the corresponding terms in the average. They can then compute an appropriate smooth upper bound to the local sensitivity of this average and add less noise. They use this idea to obtain a general result for bounded kernels, and then use it to get the optimal rate for subgaussian-nondegenerate kernels, and a bound for general degenerate bounded kernels. They also provide some indication that their bound for general degenerate bounded kernels may be optimal. 3) They also show that their results can be used to privatize 'subsampled' estimators with similar error rates that are computationally much more efficient. Finally, they apply these results to settings where U statistics are used such as various hypothesis testing problems. Strengths: 1) U-statistics are widely used across statistical testing and estimation, and have been relatively understudied in the privacy literature. This paper explores them quite generally and does a good job of suggesting problems for future work. 2) They do a good job of explaining how natural extensions of traditional DP mean estimators perform sub-optimally in estimating U-statistics. 3) The estimator based on local Hajek projections (and smooth sensitivity) seems quite technically novel and interesting. Weaknesses: 1) In the applications section, it would be good to discuss existing private algorithms for the corresponding tasks (if there are any) and compare the bounds that are obtained. 2) In the Hajek projection algorithm, it would be nice if they explained how they build on the techniques from [Ullman Sealfon NeurIPS 2019]- which parts are borrowed from that work and which parts are new. Technical Quality: 4 Clarity: 3 Questions for Authors: In equation A.41/42 is S missing from the subscript? Also what is j here? Do you mean $i^*$? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable feedback and for noting that our estimator based on local Hájek projections (and smooth sensitivity) is technically novel and interesting. [Re: Comparison of our applications with existing private algorithms]: The setting we consider, where the probabilities of the atoms in the distribution are close to uniform, has not been considered in literature before. However, there are existing private algorithms in the more general setting, which when restricted to our setting lead to suboptimal guarantees in the privacy parameter $\epsilon$. [2] considers the $\ell_1$ distance between the distributions $ (p_1, p_2, \dots, p_m)$ and the uniform distribution $(1/m, 1/m, \dots, 1/m)$ on a set of $m$ atoms. Our assumption of $\sum_i(p_i-1/m)^2\leq \delta^2/m$ implies a bound of $\delta/2$ on the $\ell_1$-distance considered in [2]. The collision-based statistic we consider is simpler than that in [2]. While they consider a broader class of probability distributions over the atoms, their sample complexity $O\left(\frac{\sqrt{m}}{\delta^2} + \frac{\sqrt{m\log m}}{\delta^{3/2}\epsilon} \right )$ has a worse factor in $\delta$ compared the sample complexity $O\left( \frac{\sqrt{m}}{\delta^2} + \frac{\sqrt{m}}{\delta \epsilon} + \frac{\sqrt{m} \log(m/\delta \epsilon)}{\delta \epsilon^{1/2}} \right)$ of our algorithm. We will incorporate this into the revision and elaborate on the comparison. [Re: Connection between [1] and our algorithm]: We will more clearly describe the connection of our algorithm with the algorithm of [1]. A key idea in this work is to exploit the concentration of the degrees of Erdős-Renyi graphs, and we generalize this idea to the broader setting of U-statistics as follows. Consider a $k$-uniform complete hypergraph with $n$ nodes (and ${n\choose k}$ edges), where the nodes are the data indices. An edge corresponds to a $k$-tuple of data points $S\in I_{n,k}$, where $I_{n,k}$ is the family of all $k$-element subsets of $[n]$, and the weight of this edge is $h(X_S)$. The local Hájek projection $\frac{1}{\binom{n-1}{k-1}} \sum_{i \in S} h(X_S)$ is simply the degree normalized by ${n-1\choose k-1}$. Our algorithm uses the property of local Hájek projections to re-weight the hyperedges ($k$-tuples) such that the local sensitivity of the re-weighted U-statistic is small. In degenerate cases and in cases where $ \zeta_1 \ll \zeta_k/k$, where $\zeta_1 = \textup{var}(\mathbb{E}[h(X_1, X_2, \dots, X_k)\|X_1])$ is the variance of the conditional expectation and $\zeta_k = \textup{var}(h(X_1,X_2, \dots, X_k))$, similar to the Erdős-Renyi case, the local Hájek projections concentrate tightly around the mean $\theta$, leading to a near-optimal error guarantee. It turns out that even when the U statistic is non-degenerate and the Hájek projections do not concentrate as strongly, our algorithm (Algorithm 1) achieves near-optimal private error guarantees. Algorithm 1 also works with subsampled family $\mathcal{S} \subseteq I_{n,k}$, where the size of $\mathcal{S}$ can be as small as $\tilde{O}(n^2/k^2)$. This allows for a computationally efficient algorithm for all $n$ and $k$. [Re: Typographical error in Eq A.41/A.42]: Yes, the index $j$ should be $i^$. In equation A.41, we are summing over $S$ such that $S \in \mathcal{S}_{i, i^}$. We will correct this and other typographical errors in the final version. [1] J. Ullman and A. Sealfon. Efficiently estimating erdos-renyi graphs with node differential privacy. Advances in Neural Information Processing Systems, 32, 2019. [2] J. Acharya, S. Ziteng, and H. Zhang. “Differentially private testing of identity and closeness of discrete distributions.” Advances in Neural Information Processing Systems 31 (2018). --- Rebuttal 2: Comment: Thanks to the authors for their detailed responses. For the Acharya, Ziteng, Zhang result is the bound not better than the one you state (Theorem 2 in that paper seems to give a better bound than the one you cite and they also have a matching lower bound). --- Rebuttal 3: Title: Comparison with Acharya et al. Comment: Thank you - you are correct. In the rebuttal, we inadvertently compared our sample complexity to that of Cai et al. ([34] in Acharya et al.’s paper) from Table 1 in the arXiv version of the NeurIPS paper. Acharya et al. improved upon this, and indeed they have a $\frac{\sqrt{m}}{\delta\sqrt{\epsilon}}$ in their sample complexity, whereas we have a $\frac{\sqrt{m}}{\delta\epsilon}$ term, which is worse. We think this may be because the family of distributions they consider are $p$ such that $\|\|p-U\|\|_1 \ge \delta$, where $U$ is the uniform distribution over $m$ atoms. In contrast, we consider $\|\|p-U\|\|_2 \ge \delta/\sqrt{m}$ and $\max_i \|p_i - 1/m\| \le 1/m$. It is not immediately clear from looking at the lower bound techniques from Acharya et al., which uses coupling and total variation ($\ell_1$) distance-based arguments, whether their lower bounds are tight in the $\ell_2$ setting. Another thing that we want to note is that [1] points out that the collision-based tester ([3], the non-private version of our algorithm) provides some tolerance or robustness to model misspecification. By updating the rejection rule to “Reject if $\tilde{U}_n \ge \frac{1+3\delta^2/4}{m}$” in Theorem 5 (line 280), we can distinguish between $\|\|p-U\|\|_2 \geq \frac{\delta}{\sqrt{m}}$ and $\|\|p-U\|\|_2\leq \frac{\delta}{\sqrt{2m}}$. Note that the second family is not exactly uniform but approximately uniform. We will add the comparison to Acharya et al. to our manuscript and clarify these points. We are grateful for your comment. [1] Canonne, Clément L. Topics and techniques in distribution testing. Now Publishers, 2022. [2] Cai, Bryan, Constantinos Daskalakis, and Gautam Kamath. “Priv’it: Private and sample efficient identity testing.” International Conference on Machine Learning. PMLR, 2017. [3] Diakonikolas, Ilias, et al. “Collision-based testers are optimal for uniformity and closeness.” arXiv preprint arXiv:1611.03579 (2016).	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable feedback and suggestions. We think we have addressed most of the questions adequately, and summarize our responses here. We will fix all typographical errors and we do not address them here. ### Connections between [1] and our algorithm (Reviewer vFH3 and Reviewer B5Tq) We will more clearly describe the connection of our algorithm with the algorithm of [1] in the revision. A key idea in this work is to exploit the concentration of the degrees of Erdős-Renyi graphs, and we generalize this idea to the broader setting of U-statistics as follows. Consider a $k$-uniform complete hypergraph with $n$ nodes (and ${n\choose k}$ edges), where the nodes are the data indices. An edge corresponds to a $k$-tuple of data points $S\in I_{n,k}$, where $I_{n,k}$ is the family of all $k$-element subsets of $[n]$, and the weight of this edge is $h(X_S)$. The local Hájek projection $\frac{1}{\binom{n-1}{k-1}} \sum_{i \in S} h(X_S)$ is simply the degree normalized by ${n-1\choose k-1}$. Our algorithm uses the property of local Hájek projections to re-weight the hyperedges ($k$-tuples) such that the local sensitivity of the re-weighted U-statistic is small. In degenerate cases and in cases where $ \zeta_1 \ll \zeta_k/k$, where $\zeta_1 = \textup{var}(\mathbb{E}[h(X_1, X_2, \dots, X_k)\|X_1])$ is the variance of the conditional expectation and $\zeta_k = \textup{var}(h(X_1,X_2, \dots, X_k))$, similar to the Erdős-Renyi case, the local Hájek projections concentrate tightly around the mean $\theta$, leading to a near-optimal error guarantee. It turns out that even when the U statistic is non-degenerate and the Hájek projections do not concentrate as strongly, our algorithm (Algorithm 1) achieves near-optimal private error guarantees. Algorithm 1 also works with subsampled family $\mathcal{S} \subseteq I_{n,k}$, where the size of $\mathcal{S}$ can be as small as $\tilde{O}(n^2/k^2)$. This allows for a computationally efficient algorithm for all $n$ and $k$. ### Regarding privacy budget (Reviewer MPw8) Lemma 3 (line 130) shows that the CoinPress algorithm from [3], adapted to the all-tuples family, is $2\epsilon$-DP. The following argument shows that Algorithm 2 is $10\epsilon$-DP as stated. Corollary 2.4 in [1] shows that for any function $f:\mathcal{X} \to \mathbb{R}$, the output $f(X) + \frac{2(\eta+1)S(X)}{\epsilon} Z$, where $Z$ is sampled from the distribution with density $\propto \frac{1}{1+\|z\|^\eta}$ and $S(X)$ is a $\beta$ smooth upper bound on the local sensitivity $LS(X)$ of $f$, is $\epsilon$-DP as long as $\beta \le \frac{\epsilon}{2(1+\eta)}$ and $\eta>1$. Thus, as stated with $\eta=4$, our algorithm, Theorem 2 (line 196), is $10\epsilon$-DP. We chose $\eta = 4$ somewhat arbitrarily, and we can improve this to almost $4\epsilon$-DP by choosing $\eta$ close to $1$. Corollary 1 (line 209) follows by composing our extension of the algorithm from [2] and our main algorithm. Passing in $\epsilon/2$ to Algorithm A.2. with $\mathcal{S} = \mathcal{I}_{n,k}$ and $\epsilon/20$ to our main algorithm results in an $\epsilon$-DP algorithm for non-degenerate, subgaussian kernels. ### Regarding Distributional Convergence (Reviewer 8Ha8) To our knowledge, differential privacy results typically focus on finite sample guarantees. However, we can show that under mild conditions on $n,k,$ and $\epsilon$, our estimator has the same asymptotic distribution as the non-private U statistic in the degenerate (of order $1$) and non-degenerate cases. Thus, the asymptotic variance doesn’t change. For simplicity, let $k=O(1)$ and $\mathcal{S}=I_{n,k}$, the set of all size $k$ subsets of $[n]$. Consider a scaling factor $c_n$, set to $n$ ($\sqrt{n}$) for degenerate (non-degenerate) kernels. Recall that our estimator $\mathcal{A}(X)=\tilde{A}_n+S(X)/\epsilon Z$, where $Z$ has density $f(z)\propto 1/(1+z^4)$. We have, $$c_n(\mathcal{A}(X)-\theta)=c_n(\tilde{A}_n-A_n)+c_n(A_n-\theta)+c_n \frac{S(X)}{\epsilon} Z.$$ For $\mathcal{S}=I_{n,k}$, $A_n=U_n$ and hence $c_n(A_n-\theta)$ converges to either a Gaussian ($h$ is non-degenerate, $c_n=\sqrt{n}$) or weighted sum of centered Chi-squared distributions ($h$ is degenerate, $c_n=n$) [2]. We show that for any choice of $c_n$, the second term is $o_P(1)$. Define $\mathcal{E}_1=\{\exists i: \|\hat{h}(i)-\theta\|\geq C_1\sqrt{k/n}\log(n/\alpha)\}$ and $\mathcal{E}_2=\{\|A_n-\theta\|\geq C_1\sqrt{k/n}\log(n/\alpha)\}$. By Lemmas A.21 (line 854) and A.3 (line 494), $P(c_n\|\tilde{A}_n-A_n\|\geq t)\leq P(\tilde{A}_n\neq A_n)\leq P(\mathcal{E}_1)+P(\mathcal{E}_2)\leq 2\alpha$. For non-degenerate subgaussian kernels, $h$ is truncated using one-half of the data. So the same argument as in lines 201-208 shows that the probability that none of the $h(X_S)$ is truncated is at most $\alpha$. For finite-sample guarantees, we take $\alpha$ as a constant and then boost it via median-of-means. For distributional convergence, we take $\alpha=n^{-c}, c>0$. Thus the first term is $o_P(1)$. Finally, conditioned on $\mathcal{E}_1\cap\mathcal{E}_2$, $L=1$ for $\xi=\tilde{O}(\sqrt{k/n})$ (Lemma A.21, line 854) in the degenerate case and $\xi=\tilde{O}(\sqrt{k\tau})$ (Lemma A.22, line 863) in the non-degenerate case. Some calculations show that as long as $\frac{k^3}{n\epsilon^2}=o(1)$, $c_n S(X)/\epsilon =o_P(1)$. Since $Z = O_P(1)$, the third term is also $o_P(1)$. Overall, $\|c_n(\mathcal{A}(X)-\theta) - c_n(A_n-\theta)\| = o_P(1)$, establishing that the asymptotic distributions of the private and non-private estimates are the same. ### References [1] J. Ullman and A. Sealfon. Efficiently estimating Erdos-Renyi graphs with node differential privacy. NeurIPS 32, 2019. [2] A. J. Lee. U-statistics: Theory and Practice. Routledge, 2019. [3] S. Biswas, Y. Dong, G. Kamath, and J. Ullman. Coinpress: Practical private mean and covariance estimation. NeurIPS 33:14475–14485, 2020.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
PURE: Prompt Evolution with Graph ODE for Out-of-distribution Fluid Dynamics Modeling	Accept (poster)	Summary: The paper presents a new approach called Prompt Evolution with Graph ODE (PURE) for non-distributed fluid dynamics modeling. PURE first learns from historical observations and system parameters in the frequency domain to explore multi-view contextual information, which can efficiently initialize the cue embedding. Interpolations of the observation sequences are then merged into the graph ODE so that the time evolution of the model-adaptive cue embeddings can be captured. These time-evolving cue embeddings are then incorporated into the underlying predictive model to overcome spatio-temporal distributional variations. In addition the paper minimizes the mutual information between the cue embeddings and the observation embeddings to enhance the robustness of the model to different distributions. Finally, extensive experiments conducted on various kinds of benchmark datasets validate the superiority of the proposed PURE compared to various baselines. Strengths: 1. The idea of the paper is novel. It is the first to link prompt learning to dynamic system modeling for out-of-distribution problems. 2. This paper is technically sound. PURE first learns initialized prompt embeddings from historical observations and system parameters, and then employs a graph ODE with interpolated observation sequences to capture the continuous evolution of their model adaptation under out-of-distribution changes. 3. The experimental results show the effectiveness of PURE in different challenging environments. Weaknesses: 1. The contribution of the proposed method in dealing with the OOD problem needs to be further clarified since the advantages of PURE over the previous efforts, such as Refs. [7, 67, 14, 72], etc., to address the OOD problem are not listed. 2. The writing of the paper needs to be improved. Some of the symbols in the method description section are not defined, e.g., what do P and N in Equation 9 refer to? 3. The experiment is not comprehensive enough. (a) The reasons for selecting baselines are not explained. Data augmentation [66, 7], invariant feature learning [39, 69, 38], adversarial training [67, 7], and domain adaptation [32, 14] are mentioned in the paper in related work for solving the OOD problem, but they are not be compared as baselines in the experiment. (b) The experiments in this paper do not state whether noisy data are considered. (c) The authors just give a brief description of the results without analyzing the reasons behind the high performance. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. What are the advantages of PURE over previous efforts to address OOD? 2. Some of the symbols in the method description section are not defined, e.g., what do P and N in Equation 9 refer to? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors list limitations in the appendix but do not mention them in the main text. It is recommended that the author make a description in the main text. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your major concern and provide additional clarification. > Q1. The contribution of the proposed method in dealing with the OOD problem needs to be further clarified since the advantages of PURE over the previous efforts, such as Refs. [7, 67, 14, 72], etc., to address the OOD problem are not listed. A1. Thanks for your comment. Compared with previous OOD methods, our contribution can be listed as follows: - Underexplored Scenarios. Our work studies an underexplored yet practical problem of OOD generalization in fluid dynamic modeling while previous OOD methods usually study the OOD problem in image classification scenarios. - Innovative Methodlogy. Our work not only learns time-evolving prompts using context mining and a graph ODE but also decouples prompt embeddings and observation embeddings via mutual information for invariant learning to address the OOD problem. - Theoretical Analysis. We provide a comprehensive theoretical analysis to support our designs, which makes our framework more solid. - Superior Performance. Comprehensive experiments validate the effectiveness of our method in different challenging settings. In particular, it improves performance by an average of 25.90% in OOD scenarios. > Q2. The writing of the paper needs to be improved. Some of the symbols in the method description section are not defined, e.g., what do P and N in Equation 9 refer to? A2. Thank you for your comment. $P$ denotes the set collecting all the positive pairs (i.e., from the same trajectory) of observation embeddings and prompt embeddings while $N$ denotes the set collecting all the possible pairs of observation embeddings and prompt embeddings in the dataset. We will include it in our revised version. > Q3. The experiment is not comprehensive enough. (a) The reasons for selecting baselines are not explained. Data augmentation [66, 7], invariant feature learning [39, 69, 38], adversarial training [67, 7], and domain adaptation [32, 14] are mentioned in the paper in related work for solving the OOD problem, but they are not be compared as baselines in the experiment. A3. Thank you for your comment. We have added three more baselines LEADS [1], CODA [2] and NUWA [3] for performance comparison. The results show that our method is superior to these baselines, which validates the effectiveness of our method when addressing OOD problems. In addition, the other references focus on the image classification problem, which cannot be adopted to solve our fluid dynamics modeling problem. We will include it in our revised version. \| Dataset \| Prometheus (ID) \| Prometheus (OOD) \| ERA5 (ID) \| ERA5 (OOD) \| SSWE (ID) \| SSWE (OOD) \| \|-------------\|----------------\|----------------\|---------\|----------\|---------\|----------\| \| LEADS \| 0.0374 \| 0.0403 \| 0.2367 \| 0.4233 \| 0.0038 \| 0.0047 \| \| CODA \| 0.0353 \| 0.0372 \| 0.1233 \| 0.2367 \| 0.0034 \| 0.0043 \| \| NUWA \| 0.0359 \| 0.0398 \| 0.0645 \| 0.0987 \| 0.0032 \| 0.0039 \| \| PURE (ours) \| 0.0323 \| 0.0328 \| 0.0398 \| 0.0401 \| 0.0022 \| 0.0024 \| > Q4. (b) The experiments in this paper do not state whether noisy data are considered. A4. Thank you for your comment. We have added experiments with noisy data to evaluate the robustness of our method. The results are shown below, which validate that our method is more robust to interference from noisy data. We will include it in our revised version. \| \| ResNet/Noise \| ResNet+PURE/Noise \| NMO/Noise \| NMO+PURE/Noise \| \| ---------- \| ------ \| ----------------- \| ------ \| -------------- \| \| PROMETHEUS \| 0.0674/0.3422 \| 0.0542/0.0586 \| 0.0397/0.1287 \| 0.0281/0.0309 \| \| NS \| 0.1823/0.6572 \| 0.1492/0.1537 \| 0.1021/0.2542 \| 0.0876/0.0892 \| >Q5. (c) The authors just give a brief description of the results without analyzing the reasons behind the high performance. A5. Thank you for your comment. The potential reason for our high performance is that (1) our model enhances model invariance across different distributions through decoupling prompt embeddings and observation embeddings via mutual information which results in high generalizationability to different environments; (2) our model utilizes multi-view context mining and graph ODE to extract prompt embeddings, which capture environment information accurately. We will include it in our revised version. Reference [1] Kirchmeyer, Matthieu, et al. "Generalizing to new physical systems via context-informed dynamics model." ICML 2022. [2] Yin, Yuan, et al. "LEADS: Learning dynamical systems that generalize across environments." NeurIPS2021. [3] Wang, Kun, et al. "NuwaDynamics: Discovering and Updating in Causal Spatio-Temporal Modeling." ICLR2024. We will also add your suggestion about future works to our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions. --- Rebuttal 2: Title: Summary of our rebuttal Comment: Dear Reviewer, We summarized our rebuttal content at your convenience as follows: - We have included a detailed explanation of our contribution. - We have included more competing baselines to demonstrate the superiority of our approach. - We have included more real-world settings such as noisy scenarios. - We have included more analysis about our superiority. We will also add your suggestion about future works to our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions. Best, the Authors	Summary: - The paper aims to improve the out-of-distribution (OOD) generalization of fluid dynamics modeling. - Two types of OOD scenarios are targeted: OOD across different systems and OOD within the same system across different timestamps. - The paper proposes a framework named PURE, composed of modules including: - Multi-view Context Exploration, which explores spatio-temporal data using both the attention mechanism and the frequency domain; - Time-evolving Prompt Learning, which incorporates the interpolation of observation sequences; - Model Adaptation with Prompt Embeddings, which leverages time-evolving prompts to mitigate temporal distribution shifts. - Extensive experiments on a range of fluid dynamics datasets support the claim. Strengths: - Significant topic: OOD generalization in fluid dynamics modeling. - Well-motivated, as OOD generalization is a crucial challenge in this field. - The presentation effectively delivers the message. - Extensive experiments have been conducted. Weaknesses: - My major concern with the paper is that the OOD challenge in dynamics modeling is not well-formulated. The paper describes the OOD scenario verbally as "different dynamical systems could involve different parameters in underlying rules" and "during long-term auto-regressive forecasting, the input data distribution could vary hugely during temporal evolution," which is straightforward and easy to understand. However, the mathematical formulation of these scenarios is absent. This formulation should be the foundational basis of the topic, as we need to clearly define the problem before addressing it. - Given the lack of mathematical formulation of the challenge, I find myself lost in the proposed approach section, unsure of the necessity for specific components. While I understand the function of each component, I cannot see why it is needed or which gaps it aims to bridge in the absent mathematical framework. - Why is the proposed method termed "prompt"? Is there a connection to prompt tuning in large language models? - How do you quantify the distribution shift in dynamics modeling? Can you rank the 'difficulty level' of OOD generalization in your experiments and analyze in which scenarios your method stands out and why? Technical Quality: 2 Clarity: 3 Questions for Authors: NA Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following. >Q1. My major concern with the paper is that the OOD challenge in dynamics modeling is not well-formulated. The paper describes the OOD scenario verbally as "different dynamical systems could involve different parameters in underlying rules" and "during long-term auto-regressive forecasting, the input data distribution could vary hugely during temporal evolution," which is straightforward and easy to understand. However, the mathematical formulation of these scenarios is absent. This formulation should be the foundational basis of the topic, as we need to clearly define the problem before addressing it. A1. Thank you for your comment. In dynamical systems, the OOD problem studies the prediction performance of models under parameter distributions or environments not seen during training. In formulation, the evolution of dynamical systems is defined by $\frac{du}{dt} = F(u, \xi)$, where $u$ is the observation, and $\xi$ is the system parameter. If these parameters come from a distribution $\xi \sim P(\xi)$, the state trajectory comes from the distribution $u^{1:T_0} \sim P(u^{1:T_0} \| \xi)$. Assume we learn a state mapping $f$ from time $u^{1:T_0}$ to $u^{T_0+1:T_0+T}$, i.e., $u^{T_0+1:T_0+T}= f(u^{1:T_0})$. We could have different distributions across training and test datasets, i.e., $P_{\text{train}}(\xi)\neq P_{\text{test}}(\xi)$, which results in $P_{\text{train}}(u^{1:T_0})\neq P_{\text{test}}(u^{1:T_0})$. Moreover, when conducting rollout prediction, we need to feed the output back to the model, i.e., $u^{T_{start}:T_{start}+T-1}= f(u^{T_{start}-T_0:T_{start}-1})$ with $P(u^{1:T_0} \| \xi)\neq P(u^{T_{start}-T_0:T_{start}-1}\|\xi, T_{start})$, which demonstrates temporal distribution shift. We will include our explanation in the revised version. > Q2. Given the lack of mathematical formulation of the challenge, I find myself lost in the proposed approach section, unsure of the necessity for specific components. While I understand the function of each component, I cannot see why it is needed or which gaps it aims to bridge in the absent mathematical framework. A2: Thank you for your comment. From our definition, the mapping $u_{output}=f(u_{input})$ could suffer from a serious distribution shift result from different $\xi$ and $T_{start}$, i.e., $P(u_{input}\|\xi, T_{start})$. To reduce the impact of distribution shift, we aim to learn invariant observation embeddings $\mu^t$ to environments, i.e., $\xi$ and $T_{start}$, and utilize prompt embeddings $z^t$ to indicate the current environment. The basic idea of our method is to ensure the invariance of observation embeddings for better generalization, i.e., $z^t\perp \mu^t$. Then, the prompt embeddings would be combined to generate the future predictions, i.e., $u_{output} = \phi ([\mu^t,z^t])$. In our framework, a basic model is adopted to generate observation, i.e., $\mu^t=Basic Model(u_{input})$. We adopt context mining and graph ODE to learn time-varying prompt embeddings., i.e., $z^0= Context Mining (u_{input})$, $z^t = GraphODE (z^0,t)$, which can explore the temporal evolution of environments. We will include our explanation in the revised version. > Q3 Why is the proposed method termed "prompt"? Is there a connection to prompt tuning in large language models? A3. Thank you for your comment. "Prompt" means a supplementary hint to indicate the context, which is incorporated to the input (observation embedding) for better generalization. Generally, it shares the basic idea with prompt tuning, which also aims to incorporate optimal tokens into the input sequence. However, prompt tuning usually utilizes prompts to indicate different tasks in NLP scenarios, our prompt refers to the current environment in fluid dynamical modeling, which determines the future evolution with better generalization. We will include it in our revised version. > Q4. How do you quantify the distribution shift in dynamics modeling? Can you rank the 'difficulty level' of OOD generalization in your experiments and analyze in which scenarios your method stands out and why? A4. Thank you for your comment. We have added experiments to demonstrate the performance with varying difficulty levels. In particular, we measure the difficulty levels based on the distance between $P_{\text{train}}(\xi)$ and $P_{\text{test}}(\xi)$ and generate three levels on the Prometheus dataset. The compared results are shown below. We can observe that although all the model performs worse in hard scenarios, our method consistently outperforms these baselines. The potential reason is that (1) our model enhances model invariance across different distributions through decoupling prompt embeddings and observation embeddings via mutual information which results in high generalizationability to different environments; (2) our model utilizes multi-view context mining and graph ODE to extract prompt embeddings, which capture environment information accurately. We will include it in our revised version. \| Method \| U-NET \| ResNet \| VIT \| Swin-T \| FNO \| CGODE \| PURE(Ours) \| \|-------------\|--------\|--------\|--------\|--------\|--------\|--------\|------------\| \| Easy \| 0.0945 \| 0.0682 \| 0.0654 \| 0.0676 \| 0.0452 \| 0.0772 \| 0.0325 \| \| Mid \| 0.1063 \| 0.0922 \| 0.0902 \| 0.0912 \| 0.0544 \| 0.0863 \| 0.0341 \| \| Hard \| 0.1432 \| 0.1234 \| 0.1076 \| 0.1123 \| 0.0623 \| 0.0921 \| 0.0354 \| Reference [1] Wu, et al. "Prometheus: Out-of-distribution Fluid Dynamics Modeling with Disentangled Graph ODE." ICML2024 In light of these responses, we hope we have addressed your concerns, and hope you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points. --- Rebuttal Comment 1.1: Title: Thank you for your responses Comment: Thank you for your responses. I appreciate the authors provide additional explanation to what I asked. I would like to maintain my rating (4) for the reason that - I think the manuscript needs significant update to reflect the new information. - The new information including the problem formulation (OOD in dynamic systems) is too fundamental rather trivial, which should appear at the beginning of the paper (problem setup). - It is essential in motivating the proposal, including the design of different components, on how they can address the dynamic OOD (on what parameter invariance across training-test). - Without revision, I fail to connect the problem to the proposal by looking the rebuttal text and the submission forth-and-back. --- Rebuttal 2: Title: Thanks for your feedback! Comment: Thanks for your feedback, and we have finished the revision based on your helpful suggestions. >Q1. The new information on OOD in dynamic systems is too basic and should be included in the problem setup at the start of the paper. Thanks for your comment. We have revised the paper based on your suggestions. The revised draft of Sec. 2 (Problem Definition) and Sec. 3.1 (Motivation and Framework Overview) is shown as below: # Sec. 2 Problem Setup Given a fluid dynamical system, we have $N$ sensors within the domain $\Omega$, with their locations denoted as $x\_1, \cdots, x\_N$, where $x_i \in \mathbb{R}^{d_l}$. The observations at time step $t$ are represented as $s_1^t, \cdots, s_N^t$, where $s_i^t \in \mathbb{R}^{d_o}$ and $d_o$ indicates the number of observation channels. Dynamical systems are governed by underlying system rules, such as PDEs with coefficient $\xi$. Variations in system parameters may lead to different environments, potentially resulting in distribution shifts. In our study, we are provided with historical observation sequences $\{s_i^{1: T_0}\}\_{i=1}^N$ and physical parameters $\xi$ (e.g., coefficients in the PDEs). Our goal is to predict the future observations of each sensor $s_i^{T_0+1: T_0+T}$. In dynamical systems, the out-of-distribution problem examines model performance when predicting under unseen parameter distributions or environments. Let $u^t=[s_1^t,\cdots, s_N^t]$, these systems evolve according to $\frac{d{u}}{dt} = F({u}, {\xi})$, where ${u}$ represents the observations and ${\xi}$ denotes the system parameters. When ${\xi} \sim P({\xi})$, the state trajectory ${u}^{1:T_0}$ follows the distribution $P({u}^{1:T_0} \| {\xi})$. Assume we learn a learned mapping function $f$ from ${u}^{1:T_0}$ to ${u}^{T_0+1:T_0+T}$, i.e., ${u}^{T_0+1:T_0+T} = f({u}^{1:T_0})$ and there could be different distributions across training and test datasets, i.e., $P_{\text{train}}({\xi})\neq P_{\text{test}}({\xi})$, which results in $P\_{{train }}\left({u}^{1: T_0}\right) \neq {P}\_{{test }}\left({u}^{1: T_0}\right)$. Moreover, when conducting rollout prediction, we need to feed the output back to the model, i.e., ${u}^{T_{start}:T_{start}+T-1} = f({u}^{T_{start}-T_0:T_{start}-1})$, with $P({u}^{1:T_0} \| {\xi})\neq P({u}^{T_{start}-T_0:T_{start}-1}\|{\xi}, T_{start})$, which demonstrates temporal distribution shift. # Sec. 3 The Proposed PURE ## Sec. 3.1 Motivation and Framework Overview This paper addresses the challenge of out-of-distribution fluid system modeling, which is complicated by parameter-based and temporal distribution shifts. Specifically, our function $f(\cdot)$ can suffer from a serious distribution shift result from different ${\xi}$ and $T\_{start}$, i.e., $P({u}\_{input}\|{\xi},T\_{start})$. To reduce the impact of distribution shift, we aim to learn invariant observation embeddings ${\mu}^t$ to different environments, i.e., ${\xi}$ and $T_{start}$ for better generalization and utilize prompt embeddings ${z}^t$ to indicate the current environment for final prediction. In formulation, we have: $$ {z}^t \perp {\mu}^t, u_{\text {output }}=\phi\left(\left[{\mu}^t, {z}^t\right]\right) . \quad (1) $$ The first term ensures the invariance of observation embeddings by decoupling observation embeddings and prompt embeddings. The second term aims to combine both two embeddings to generate the future predictions. Therefore, we propose a novel approach named PURE as: $$ {\mu}^t=\operatorname{BasicModel}\left({u}_{\text {input }}\right), \quad (2) $$ $$ z^0=\operatorname{ContextMining}\left({u}_{\text {input }}\right),\quad (3) $$ $$ {z}^t=\operatorname{GraphODE}\left(z^0, t\right),\quad (4) $$ where a basic model is adopted to generate observation, and we adopt context mining and graph ODE to learn time-varying prompt embeddings. Given a basic forecasting model (Eqn. 2), our PURE contains three key modules: (1) Multi-view Context Exploration, which explores spatio-temporal data using both the attention mechanism and the frequency domain to initialize prompt embeddings (Eqn. 3). (2) Time-evolving Prompt Learning, which incorporates the interpolation of observation sequences into a graph ODE to learn the evolution of prompt embeddings (Eqn.4). (3) Model Adaptation with Prompt Embeddings, which leverages the time-evolving prompts to mitigate the temporal distribution shifts in fluid dynamics models (Eqn.1). More details are in Figure 1. > Q2. It's essential to explain how the proposal's components address dynamic OOD, focusing on parameter invariance between training and testing. Thanks for your comment. We have included our motivation in Sec. 3.1. > Q3. Without revision, I can't connect the problem to the proposal from the rebuttal and submission exchanges. Thanks for your comment. We have revised the manuscript and included all the related content here for your convenience. Thank you again for your feedback! Please let us know if you have further questions. --- Rebuttal Comment 2.1: Title: Further clarification Comment: Dear Reviewer, As the deadline for the author-reviewer discussion phase is approaching, we would like to check if you have any other remaining concerns about our paper. We greatly appreciate your feedback and have worked diligently to address your comments. Thanks to your suggestions, we have revised the draft. For your convenience, we show all the revised part of our draft as follows and we believe that it may not be necessary to review the rebuttal text and the submission back-and-forth at this time. > Major concern about OOD challenge formulation and component necessity: We have revised the manuscript and included all the related content [Section 2 (Problem Setup) and Section 3.1 (Motivation and Framework Overview)] in our last response for your convenience. > Clarification on the term "prompt": We have revised the manuscript and included all the related content [Appendix C] below for your convenience. Prompt Learning. Prompt learning [46, 15, 23] has recently gained significant attention as a technique for adapting pre-trained models to various downstream tasks by leveraging the power of prompt-based fine-tuning [8, 27, 85, 69, 49]. In the domain of large language models, prompt learning aims to incorporate optimal tokens into the input sequence, which can effectively improve performance without extensive retraining [26, 83, 81]. In the context of fluid dynamics modeling, prompt learning refers to a supplementary hint to indicate the context, which is incorporated into the input (observation embedding) for better generalization. Although it shares a similar meaning as prompt tuning in language models, our prompt refers to the current environment, which determines the future evolution with better generalization. [46] Yajing Liu, Yuning Lu, Hao Liu, Yaozu An, Zhuoran Xu, Zhuokun Yao, Baofeng Zhang, Zhiwei Xiong, and Chenguang Gui. Hierarchical prompt learning for multi-task learning. CVPR2023 [15] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. PPT: Pre-trained prompt tuning for few-shot learning. arXiv2021 [23] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv 2022 [8] Ning Ding, Shengding Hu, Weilin Zhao, Yulin Chen, Zhiyuan Liu, Hai-Tao Zheng, and Maosong Sun. OpenPrompt: An open-source framework for prompt-learning. arXiv 2021. [27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. CVPR2023 [85] Kaiyang Zhouet al. Conditional prompt learning for vision-language models. CVPR2022 [69] Zifeng Wang et al.. Learning to prompt for continual learning.CVPR2022 [49] Yuning Lu et al. Prompt distribution learning. CVPR2022 [26] Woojeong Jin et al. A good prompt is worth millions of parameters: Low-resource prompt-based learning for vision-language models. arXiv 2021. [83] Zizhuo Zhang et al. Prompt learning for news recommendation. SIGIR 2023. [81] Yaohua Zha et al. Instance-aware dynamic prompt tuning for pre-trained point cloud models. CVPR2023 > Concern about Quantifying distribution shifts: We have revised the manuscript and included all the related content [Appendix H] here for your convenience. Performance with respect to Different Difficulty Levels. Here, we demonstrate the performance of our PURE with varying difficulty levels. In particular, we measure the difficulty levels based on the distance between $P_{\text{train}}(\xi)$ and $P_{\text{test}}(\xi)$ and generate three levels on the Prometheus dataset. The compared results are shown in Table 7. From the results, we can observe that all the model performs worse in hard scenarios and our method has consistently outperformed these baselines. The potential reason is that (1) our model enhances model invariance across different distributions through decoupling prompt embeddings and observation embeddings via mutual information which results in high generalizationability to different environments; (2) our model utilizes multi-view context mining and graph ODE to extract prompt embeddings, which capture environment information accurately. Table 7. Performance comparison across varying levels of OOD generalization difficulty. The values represent the Mean Squared Error (MSE) for each method. \| Method \| U-NET \| ResNet \| VIT \| Swin-T \| FNO \| CGODE \| PURE(Ours) \| \|-------------\|--------\|--------\|--------\|--------\|--------\|--------\|------------\| \| Easy \| 0.0945 \| 0.0682 \| 0.0654 \| 0.0676 \| 0.0452 \| 0.0772 \| 0.0325 \| \| Mid \| 0.1063 \| 0.0922 \| 0.0902 \| 0.0912 \| 0.0544 \| 0.0863 \|0.0341 \| \| Hard \| 0.1432 \| 0.1234 \| 0.1076 \| 0.1123 \| 0.0623 \| 0.0921 \| 0.0354 \| We hope that these revisions address your concerns. Please let us know if there are any further questions or concerns. Sincerely, The Authors	Summary: This paper pioneers the connection of prompt learning with dynamical system modeling to address the challenge of out-of-distribution shifts. The proposed PURE method initializes prompt embeddings by learning from historical observations and system parameters. Strengths: 1.The paper is easy to follow. 2.The proposed method is sound and innovative. 3. The authors provide theoretical proof and show comprehensive experimental comparisons. Weaknesses: 1. Some results may be incorrectly labeled as suboptimal in table, and there are errors in the use of some symbols. 2. The explanation of the experimental results is not detailed enough, making some experiments difficult to understand. 3. The proposed method is aimed at OOD (Out-Of-Distribution), but the experiments lack comparison and discussion with methods specifically targeting OOD, such as [1] and [2]. Reference: [1] Kirchmeyer, Matthieu, et al. "Generalizing to new physical systems via context-informed dynamics model." International Conference on Machine Learning. PMLR, 2022. [2] Yin, Yuan, et al. "LEADS: Learning dynamical systems that generalize across environments." Advances in Neural Information Processing Systems 34 (2021): 7561-7573. Technical Quality: 3 Clarity: 3 Questions for Authors: Q1. There might be misuses of symbols in the paper, such as, Change "xi" to "si" in line 65, Change "xq" to "xq" in line 96, "Zero-shot Experiments" and "Generalization Experiments" in line 215 should be treated equally and placed on separate lines. Q2. Is there an issue with the second-best data in Table 2? For example, in the column w/OOD of SPHERICAL-SWE, the second-best should be DGPDE 0.0028. The corresponding improvement results also need to be modified. Q3. What does the clustering in Figure 5 represent? Could you provide a detailed explanation? Q4. The paper utilizes mutual information to decouple different prompt embeddings and observation embeddings, reducing the sensitivity of observation embeddings to different distributions. However, I don't quite understand the purpose of decoupling. Observational embeddings are related to the environment, and prompt embeddings are related to the environment as well; they are inherently correlated. Please provide further explanation. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This method does not apply to real-world scenarios, such as rigid dynamics modeling and traffic flow forecasting. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your major concern and provide additional clarification. > Q1. Some results may be incorrectly labeled as suboptimal in table, and there are errors in the use of some symbols. There might be misuses of symbols in the paper, such as, Change "xi" to "si" in line 65, Change "xq" to "xq" in line 96, "Zero-shot Experiments" and "Generalization Experiments" in line 215 should be treated equally and placed on separate lines. Is there an issue with the second-best data in Table 2? For example, in the column w/OOD of SPHERICAL-SWE, the second-best should be DGPDE 0.0028. The corresponding improvement results also need to be modified. A1. Thank you for pointing this out. We will correct all the typos carefully in the revised version. > Q2. The explanation of the experimental results is not detailed enough, making some experiments difficult to understand. What does the clustering in Figure 5 represent? Could you provide a detailed explanation? A2. Thank you for your comment. We demonstrate the t-SNE visualization of the predicted trajectories for compared methods. From the results, we can observe that our method outperforms baselines, which validate the superiority of our method. Here, clustering refers to the technique of reducing the dimensions in t-SNE. We will include more details in our revised version. > Q3. The proposed method is aimed at OOD (Out-Of-Distribution), but the experiments lack comparison and discussion with methods specifically targeting OOD, such as [1] and [2]. Reference: [1] Kirchmeyer, Matthieu, et al. "Generalizing to new physical systems via context-informed dynamics model." International Conference on Machine Learning. PMLR, 2022. [2] Yin, Yuan, et al. "LEADS: Learning dynamical systems that generalize across environments." Advances in Neural Information Processing Systems 34 (2021): 7561-7573. A3. Thank you for your comment. We have added two baselines LEADS [1] and CODA [2] for performance comparison. The results below show that our method is superior to these baselines, which validates the effectiveness of our method when addressing OOD problems. We will include it in our revised version. \| Dataset \| Prometheus ID \| Prometheus OOD \| ERA5 ID \| ERA5 OOD \| SSWE ID \| SSWE OOD \| \|-------------\|----------------\|----------------\|---------\|----------\|---------\|----------\| \| LEADS \| 0.0374 \| 0.0403 \| 0.2367 \| 0.4233 \| 0.0038 \| 0.0047 \| \| CODA \| 0.0353 \| 0.0372 \| 0.1233 \| 0.2367 \| 0.0034 \| 0.0043 \| \| Ours \| 0.0323 \| 0.0328 \| 0.0398 \| 0.0401 \| 0.0022 \| 0.0024 \| > Q4. The paper utilizes mutual information to decouple different prompt embeddings and observation embeddings, reducing the sensitivity of observation embeddings to different distributions. However, I don't quite understand the purpose of decoupling. Observational embeddings are related to the environment, and prompt embeddings are related to the environment as well; they are inherently correlated. Please provide further explanation. A4. Thank you for your comment. Our decoupling aims to build invariance of our observation embeddings to environments for better generalization. Here, both observation embeddings and prompt embeddings are combined for the final prediction, and prompt embeddings are utilized to provide the environment information. In other words, observational embeddings are not correlated with the environment for better generalization. We will include it in our revised version. > Q5. This method does not apply to real-world scenarios, such as rigid dynamics modeling and traffic flow forecasting. A5. Thank you for your comments. We have added four baselines, i.e., EGNN [1], SGNN [2], SimVP [3], PastNet [4] for performance comparison on the RigidBall and TaxiBJ datasets. The table shows our method outperforms the baseline models in rigid dynamics and traffic flow forecasting. Please see the PDF for visual results. We will include it in our revised version.、 Rigid dynamics (MSE) \| Method \| PL 10 $\downarrow$ \| PL 20 $\downarrow$ \| PL 30 $\downarrow$ \| PL 40 $\downarrow$ \| PL 50 $\downarrow$ \| \|--------\|-------\|-------\|-------\|-------\|-------\| \| EGNN \| 1.37 \| 1.89 \| 3.77 \| 5.66 \| 7.87 \| \| SGNN \| 0.64 \| 0.72 \| 1.23 \| 2.44 \| 4.96 \| \| PURE \| 0.59 \| 0.68 \| 0.97 \| 1.45 \| 3.99 \| Traffic flow \| Metric \| MSE $\downarrow$ \| MAE $\downarrow$ \| SSIM $\uparrow$ \| PSNR$\uparrow$ \| \|--------\|--------\|--------\|-------\|-------\| \| SimVP \| 0.4332 \| 16.897 \| 0.9822\| 39.29 \| \| PastNet\| 0.4293 \| 16.405 \| 0.9876\| 39.42 \| \| PURE \| 0.3982 \| 15.434 \| 0.9971\| 41.23 \| Reference [1] Satorras, Vıctor Garcia, Emiel Hoogeboom, and Max Welling. "E (n) equivariant graph neural networks." ICML2021 [2] Han, Jiaqi, et al. "Learning physical dynamics with subequivariant graph neural networks." NeurIPS2022 [3] Gao, Zhangyang, et al. "Simvp: Simpler yet better video prediction." CVPR. 2022. [4] Wu, Hao, et al. "Pastnet: Introducing physical inductive biases for spatio-temporal video prediction." MM2024 We will also add your suggestion about future works to our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions. --- Rebuttal 2: Title: Summary of our rebuttal Comment: Dear Reviewer, We summarized our rebuttal content at your convenience as follows: - We have included more competing baselines to demonstrate the superiority of our approach. - We have included more real-world settings such as rigid dynamics modeling and traffic flow forecasting for performance comparison. - We have explained the purpose of decoupling and visualization. - We will proofread our paper to clear every typo in our final version. We will also add your suggestion about future works to our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions. Best, the Authors	Summary: The paper proposes a graph ODE-based approach for OOD fluid dynamics modeling. PURE aims to learn time-evolving prompts via graph ODE for adaptation of spatio-temporal forecasting models on OOD scenarios. To address temporal distribution shifts, the interpolation of obersvation sequences are combined into graph ODE framework to learn evolution of prompt embeddings. Strengths: - The paper proposes a new approach that connects prompt learning and dynamical system modeling which addresses OOD shifts. - By learning time-evolving prompts that adapt to changes in system parameters and temporal evolution, the approach can enhance model robustness. - The paper provides theoretical analysis on incorporating observations during evolution. - Experiments on diverse benchmarks show generalization ability to OOD and different prediction length. Weaknesses: As I am not an expert in this field, I am unable to find major concerns or weakness of the approach. - As the method is based on attention, the proposed approach may have limited scalability and take long computation time. Is there a comparison on these with the previous works? Technical Quality: 3 Clarity: 3 Questions for Authors: Please address the questions in the Weaknesses. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations are explained in Appendix I. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your major concern and provide additional clarification. > Q1. As the method is based on attention, the proposed approach may have limited scalability and take long computation time. Is there a comparison on these with the previous works? A1. Thank you for your comment. We have added a comparison of computational costs below. From the results, we can observe that our method has a competitive computation cost with huge performance increasement. Without OOD, it improves performance by an average of 4.44%, and with OOD, it improves performance by an average of 25.90%. We will include it in our revised version. \| Method \| UNet \| ResNet \| VIT \| SwinT \| FNO \| UNO \| CNO \| NMO \| DGODE \| PURE (Ours) \| \|--------------\|-------\|--------\|-------\|--------\|------\|------\|------\|------\|-------\|------------\| \| Training time (h) \| 11.2 \| 9.76 \| 14.5 \| 12.3 \| 6.9 \| 7.8 \| 13.4 \| 6.3 \| 13.2 \| 7.2 \| \| Inference time (s) \| 1.34 \| 0.93 \| 1.32 \| 1.13 \| 0.54 \| 0.67 \| 0.12 \| 0.52 \| 1.23 \| 0.69 \| We will also add your suggestion about future works into our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions. --- Rebuttal Comment 1.1: Comment: Thank you for the additional experiments on computational costs. --- Rebuttal 2: Title: Thank you for your feedback and support! Comment: Thank you for your feedback and support! We will add the rebuttal contents to the main paper in the final version following your valuable suggestions.	Rebuttal 1: Rebuttal: Dear Reviewers, Thanks for your time and valuable feedbacks. We acknowledge three reviewers (Reviewer Sh2w, Reviewer MBSt, and Reviewer MFD5) comments that our work is novel or new. We acknowledge the positive comments such as "a new approach" (Reviewer Sh2w), "enhance model robustness" (Reviewer Sh2w), "provides theoretical analysis " (Reviewer Sh2w), "show generalization ability" (Reviewer Sh2w), "easy to follow" (Reviewer MBSt), "sound and innovative" (Reviewer MBSt), "theoretical proof" (Reviewer MBSt), "comprehensive experimental comparisons" (Reviewer MBSt), "significant topic" (Reviewer vRkX), "well-motivated" (Reviewer vRkX), "the effective presentation" (Reviewer vRkX), "extensive experiments" (Reviewer vRkX), "novel idea" (Reviewer MFD5), "technically sound" (Reviewer MFD5), and "effectiveness" (Reviewer MFD5). We have also responded to your concerns in the following. The figures are included in the pdf file for your reference. Please let us know if you have any additional questions or concerns. We will try our best to address them. Best regards, the Authors Pdf: /pdf/b3c4e1a9a8d35554b59024ead7b6ab5c6617164c.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
The Many Faces of Optimal Weak-to-Strong Learning	Accept (poster)	Summary: This paper presents an efficient and simple weak-to-strong learner that has optimal in-expectation error. In weak-to-strong learning, we are given a dataset of $m$ points from a distribution, and a $\gamma$-weak learner that returns hypotheses from a class of VC dimension $d$. AdaBoost, which is a textbook weak-to-strong learner, makes $O(\ln(m)/\gamma^2)$ total invokations to the weak learner, and the best-known analysis for it shows that it suffers an in-expectation error $O\left(\frac{d\ln(m/d)\ln(m)}{\gamma^2 m}\right)$. Larsen and Ritzert (2022) constructed a weak-to-strong learner, that has expected error $O(d/\gamma^2 m)$. Furthermore, they showed that this is the optimal error that one can obtain from $m$ training examples and a $\gamma$-weak learner. However, the weak-to-strong learner by Larsen and Ritzert (2022) makes $O(m^{0.8}/\gamma^2)$ invokations to the weak learner --- which is exponentially worse than AdaBoost. Another bagging-based-boosting algorithm due to Larsen (2023), which also achieves the optimal expected error of $O(d/\gamma^2m)$, makes only $O((\ln m)^2/\gamma^2)$ invokations to the weak-learner. This is still a log factor worse than AdaBoost. Could we then hope to obtain a tighter analysis of the error of AdaBoost, and show that it obtains the optimal error with only $O(\ln(m)/\gamma^2)$ invokations to the weak learner? Unfortunately, no. Høgsgaard et al. (2023) showed that AdaBoost necessarily suffers an expected error which is at least $\Omega(d\ln(m)/\gamma^2 m)$. Can we then at least shoot for a different weak-to-strong learner that attains the optimal expected error of $O(d/\gamma^2m)$, and also invokes the weak learner only $O(\ln(m)/\gamma^2)$ many times (which is the AdaBoost gold standard)? This paper answers the question in the affirmative, with a remarkably simple weak-to-strong learner that they call Majority-of-29. The algorithm is exceedingly simple to describe: Partition the training dataset into 29 disjoint sub-samples of size $m/29$ each. Run AdaBoost on each subsample, and return the majority vote over the AdaBoosts. Since each AdaBoost makes only $O(\ln(m)/\gamma^2)$ calls to the weak learner, and we run a constant (29) many AdaBoosts, the total number of calls to the weak learner is $O(\ln(m)/\gamma^2)$ as required. Further, using an analysis similar to the recent majority-of-3-ERMs algorithm of Aden-Ali et al. (2023), the authors are able to show that the expected error of Majority-of-29 is $O(d/\gamma^2m)$. The analysis from that work does not extend in a trivial manner, and the authors are required to make appropriate technical modifications and enhancements. The number 29 emerges from the analysis --- the authors require showing a new generalization bound for margin-based classifiers (they show a generalization bound of the order $O((d/\gamma^2m)^{\alpha})$), for $\alpha=1/14$, and this lets them obtain the result for Majority-of-$g(\alpha)$, where $g(\alpha)=2/\alpha+1$. The authors conjecture that the analysis of the generalization bound could be improved, and a Majority-of-3 might well suffice for optimal error. Finally, the authors also do a (somewhat-limited) empirical comparison of the the performances of the three optimal weak-to-strong learners mentioned above (LarsenRitzert, Bagging-based-boosting, Majority-of-29) as well as AdaBoost. The authors find that for large datasets, Majority-of-29 outperforms the other optimal weak-to-strong learners. On the smaller datasets, the authors find that Bagging-based-boosting outperforms Majority-of-29. Strengths: The weak-to-strong learner that the authors propose is optimal, and also requires the fewest calls to the weak learner among all optimal weak-to-strong learners that we know. More importantly, it is exceedingly simple and elegant. It also empirically outperforms the other optimal weak-to-strong learners (at least in the experiments performed by the authors). It is also nice to see that the analysis technique from Aden-Ali et al. (2023) finds new applications. The paper is well-written, sets up the stage (along with relevant prior work) well in the first two sections, and provides a nice high-level summary of the formal analysis in Section 3. Weaknesses: While the theoretical contribution is substantial and undeniable, arguably, the experimental section is extremely limited (which is okay, and the authors admit this at the end, but this is still a limitation, especially if we want to draw conclusions about the empirical performance of the different weak-to-strong learners). The authors only perform experiments on 4 real-world datasets---there are admittedly many more out there, even just in the UCI repository. Could the authors at least elaborate on their rationale behind choosing the datasets that they did? (e.g., was it a random subset of 4? was it the first 4? was it the best 4 from 20 that they observed this trend on?) How might one believe that there is no cherry-picking of datasets involved? The authors make two conclusions from their experiments: 1) on larger datasets, Majority-of-29 outperforms both Bagging-based-boosting and LarsenRitzert. 2) on smaller datasets, Bagging-based-boosting outperforms Majority-of-29. Importantly, the former conclusion is drawn from results on just 3 datasets, and the latter is drawn from just 1! This can really make one skeptical about whether they should truly believe these conclusions. It is okay that this is just a pilot empirical study, but such claims call for significantly larger empirical validation. Also, please see the questions below. Technical Quality: 3 Clarity: 3 Questions for Authors: 1) Do we have reason to believe that $O(\ln(m)/\gamma^2)$ calls to the weak learner is indeed the best gold standard we can hope for? To my understanding, the reason we need $O(\ln(m)/\gamma^2)$ calls to the weak learner in AdaBoost is because we want to use a margin-based generalization bound that expects a classifier to have at least $\Omega(\gamma)$ margin on every training sample---AdaBoost attains this guarantee only after $O(\ln(m)/\gamma^2)$ iterations. But could it perhaps be possible that there is a weak-to-strong learner out there that attains optimal error of $O(d/\gamma^2m)$ with $o(\ln(m)/\gamma^2)$ calls to the weak learner? 2) In the Experiments section, the x-axis in Figures 1 and 2 varies the number X of AdaBoosts trained on disjoint partitions in the Majority-of-X algorithm. But this is not a parameter in the other algorithms (BaggedAdaboost and LarsenRitzert). Hence, I would have expected to see a constant line for these other algorithms in the plots (like how the red and blue lines are constant in Figure 2). Why are there different numbers corresponding to different number of voting classifiers in BaggedAdaboost and LarsenRitzert in Figure 1 (and also for BaggedAdaboost in Figure 2)? Am I missing something? Minor/Typos: \ Line 133: It is 0 if half of hypotheses are correct and half are wrong --- this is only true in a weighted sense right? \ Line 299: this suggests* Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed any limitations that I can foresee. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for taking the time to thoroughly assesing the article, asking interesting questions, and suggesting concrete improvements. Experiments: As allotted to in the answer to reviewer KvEC, and as you correctly point out, we should have made it more clear that the experiments are very much a pilot. As you also point out, the main contribution of the paper is the theoretical result. Regarding the choice of the 4 real-word-datasets, we simply chose the same as used in "Optimal Minimal Margin Maximization with Boosting" by Gr\o nlund et al. (ICML'19). We did not run on any other data sets and no results were left out, so we wouldn't say that there was any cherry picking done. Let us also comment on the reason for not conducting more experiments. We are in a theory research group and do not have access to other machines than our laptops. Hence conducting a large number of experiments was simply infeasible. As we also comment on for reviewer KvEC, if you all feel that the paper is stronger as a pure theory paper, we are okay with removing the experiments. In all circumstances, we will further downtime the significance of our experiments. (Question 1) Number of calls to weak learner: This is a very interesting question. Classic work by Freund (Y. Freund. Boosting a weak learning algorithm by majority. Information and Computation, 121(2):256– 285, 1995), and also a recent line of work on parallel boosting (Karbasi and Larsen ALT'24, Luy et al. SODA'24), shows that indeed one needs $\Omega(\gamma^{-2} \log(1/\varepsilon))$ calls to a weak learner to obtain error $\varepsilon$. However, these lower bounds actually require $\exp(d) \geq \gamma^{-1}$. If $d$ is assumed a constant, then work by Alon, Gohen, Hazan and Moran STOC'21 shows that there are algorithms using only $\tilde{O}(\gamma^{-1} \log(1/\varepsilon))$ calls. (Question 2): Regarding the question about BaggedAdaboost and LarsenRitzert not being one line: We did attempt to explain this in the paragraph immediately following the description of the data sets. Our apologies if this was not clear enough. Since Bagging is well-defined (but not necessarily optimal) with an arbitrary number of bootstrap sub-samples, we simply run BaggedAdaboost with X many bootstrap samples (random sub-samples of the dataset) for varying X. For LarsenRitzert, as explained in that paragraph, the number of sub-samples needed grows unwieldy for large data sets. Thus we chose to instead sample X of the sub-samples defined by LarsenRitzert without replacement. This was due to computational constraints. Furthermore, we feel the plot is more informative when seeing how more and more combined AdaBoosts improve the accuracy. --- Rebuttal Comment 1.1: Title: Response to rebuttal Comment: Thank you for your response. In particular, it is good to know that there was no cherry-picking of datasets involved. I would not argue for removing the experiments section entirely, since there is definitely value in having them; instead, as you say, it would be good to tone down the prose on it. Thanks also for the clarification about $O(\ln m/\gamma^2)$ being the standard, as well as the discrepancy about the axis in the experiment. I would definitely clarify the latter in the prose, since it was quite confusing for me. Finally, it is remarkable that you are able to get the number 29 down to 5. Given this, and the updates that you promise about clarifying and toning down experiments, I am happy to increase my score from 7 -> 8. I maintain that this is a strong contribution and deserves to be accepted. Great work!	Summary: This paper introduces a new Boosting algorithm, MAJORITY-OF-29, which achieves provably optimal sample complexity and is remarkably simple to implement. The algorithm partitions the training data into 29 disjoint subsets, applies AdaBoost to each subset, and combines the resulting classifiers through a majority vote. This approach not only matches the asymptotic performance of AdaBoost but also improves upon previous weak-to-strong learners in terms of simplicity and runtime efficiency. Strengths: 1. The paper introduces a novel method and provides detailed theoretical analysis. Weaknesses: 2. Existing experiments fail to demonstrate the effectiveness of the proposed method, and there is a lack of analysis and discussion on current experimental results. Technical Quality: 2 Clarity: 1 Questions for Authors: Please see the weakness. Confidence: 4 Soundness: 2 Presentation: 1 Contribution: 1 Limitations: Limitations are discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Regarding the question about experiments: Let us first re-iterate that our main focus in this work is on the theoretical results, which we believe are strong (the paper is also submitted with a primary area of "Learning Theory"). Perhaps we were not clear enough when claiming that our experimental results should be seen as "indications" (as phrased in Section 1 Empirical Comparison), and thus the experiments should be seen a small complement to our main focus the theoretical bound. We also try to allude to this in section 5 (Limitations). However it is seems like we have not made it clear enough that this is only a pilot empirical study as reviewer p9UC also points out the same things - and that the paper's main contribution is the theoretical result. We thank you for pointing this out and will make it clear in following iterations of the submission. We were not certain about, if effectiveness in the comment "Existing experiments fail to demonstrate the effectiveness of the proposed method" refers to the reported Test Accuracy or the running time of the algorithms so we have given an answer to both in the following: Test accuracy: The experiments do not strongly indicate that the Majority-OF-29 (now Majority-OF-5) outperforms other optimal methods LarsenRitzert, BaggedAdaBoost. However, based on the theoretical results, it is also not expected that it outperforms them in terms of accuracy, as they are all asymptotically optimal. The benefit of our new algorithm is that it is simpler, easier to implement and more efficient in terms of computation. Runtime: For a runtime comparison, we have chosen not to directly include this. One reason for this, is that it will be heavily implementation dependent. In particular, if all invocations of AdaBoost are done in a single thread, then our algorithm will surely be faster in practice, as each data point is only included in precisely one AdaBoost invocation. For the other optimal methods, data points are included in many invocations. However, if properly multi-threaded, it is conceivable that other methods could be made as (or nearly as) efficient as our new algorithm. Even though we don't feel comfortable reporting the runtimes in the article, the time it takes to fit the given number of AdaBoosts in AdaBoost, Majority-of-X, LarsenRitzert, Bagging can be found in the zip file attached to the article in the folder "results", where for a given dataset and run, the time variable $"fit$_$time"$ describes the time it takes to fit the given number of AdaBoosts in a given a run.	Summary: The authors present a new boosting algorithm: partition training data into 29 pieces of equal size, run AdaBoost on each, and output the majority vote over them. The authors prove that the sample complexity of MajorityVote29 is optimal and its running time is the same order as AdaBoost. Experimental results are also attached, which corroborate their theoretical findings. Strengths: - Very strong and interesting result - Mathematically sound, based by my judgement - Good presentation, self-contained and well-structured Weaknesses: N/A Technical Quality: 4 Clarity: 4 Questions for Authors: N/A Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	null	null	null	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their thoughtful reviews. Let us add one general remark that we will leave to the reviewers whether to include in their evaluation of the submission or not. Recently we found a way to improve the result to the majority of 5 instead of majority of 29. The reason why this improvement is possible is due to us becoming aware of a stronger statement of corollary 5 which improves the previous $O((d/(\gamma^2m)^{1/14})$ bound to $O((d/(\gamma^2m)^{1/2})$. For the intuition of why this leads to a majority of 5, we recall from the proof sketch that we needed to combine $2/\alpha+1$ many AdaBoosts where $\alpha$ is the number of the exponent that we obtain in corollary 5 ($O((d/(\gamma^2m))^{\alpha}$), which now can be set equal to $\alpha=1/2$ and thus we get that we only need to combine $2/\alpha+1=5$ many AdaBoosts to get the optimal in expectation sample complexity. Fortunately, out experiments also included Majority-of-5 so no significant changes to the paper are needed. The new proof only changes (and simplifies) the material in the appendix. Also, as we wish to be assessed as a theory paper, so primarily on the theoretical contribution, and added the experiments merely as a very first pilot study, let us add that if the reviewers all feel the paper would be stronger without the experiments, we would like to remove the experiments.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Learning diverse causally emergent representations from time series data	Accept (poster)	Summary: The article proposes a learning scheme aimed at detecting emergent quantities from time series data of systems made of many interacting parts, such as the brain. To this end the authors combine "minimum mutual information", a previously introduced emergence criterion, with SMILE, a differentiable lower bound estimator for mutual information. Differentiability is crucial for the loss function to be optimizable efficiently. They apply this architecture to two examples: First, a series of random bit strings with time-correlated parity, where parity is considered the emergent quantity. Second, real-world data of macaque brain activity. The approach successfully identifies parity in the first example. The authors claim that an emergent feature has been learned for the second example also. Strengths: While the individual parts of the learning scheme are not new, their combination into a differentiable architecture is original and seems like a promising direction to me. The analysis seems sound, even though I found the presentation at times a bit hard to follow as some parts seem to be missing. The individual quantities are mostly clearly defined and the individual results are statistically significant in terms of error bars. Weaknesses: From the article alone, I could not fully understand the architecture and its training procedure that is illustrated in Fig. 1. I could not find code that would reproduce it, or a detailed pseudo-code description of the algorithm. It is unclear to me what emergent feature was found for the monkey example, or how Fig. 4 proves that any such feature was found. While the architecture and direction seems promising to me, a few more benchmarks would help make the case that this scheme can find emergent features in many settings. The two examples shown are one toy example with unnatural time dynamics and one real world example where it is hard to understand the dynamics from first principles. Benchmarking this new method on more standard examples with emergent behavior, such as Ising models, would be more convincing. Technical Quality: 3 Clarity: 2 Questions for Authors: 1) How does Figure 4 prove that an emergent feature was found? What is the quantity on the x-axis in this figure? 2) Can you share the code of the architecture and its training or describe it in more detail? 3) Would you expect this method to find emergent features in standard, well-understood systems, such as Ising models? 4) Can you explicitly write down the prediction-only objective in Fig.2? Overall, I find the approach promising but not yet tested on enough examples to support its ability to find emergent features. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: There is no open access to the code. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Many thanks for the feedback. We apologise for the lack of clarity in some of our explanations. We will improve these and add pseudo-code in the camera-ready version of the paper. Regarding the concern about not having enough examples, as we describe in more detail below and in the global rebuttal, we have conducted several additional benchmarks and evaluations. These new analyses include: 1. Two new real-world datasets, showing our method works well in both cases. 2. One more synthetic dataset, Conway’s Game of Life, which is arguably one of the most well-known examples of emergent behaviour. Since in this system there is a notion of what the emergent feature should be (known in the literature as “particles”), we can show that the learned feature captures the state of these particles. 3. Comparisons against baseline algorithms, making the case that our method finds features that standard algorithms based only on temporal predictions do not. In response to your questions: > How does Figure 4 prove that an emergent feature was found? What is the quantity on the x-axis in this figure? We apologise for the oversight in the labelling of this figure. The X-axis represents training time, and the Y-axes represent the emergence measure $\Psi$ and its adjusted variant $\Psi_A$. The plot proves that an emergent feature was found because the emergence measure is greater than zero, which is a sufficient condition for the feature being emergent (as shown by Rosas et al., 2020; Ref. [30] in the paper). > Can you share the code of the architecture and its training or describe it in more detail? Due to the constraints of the double-blind review we have been unable to share a Github link, but we will include it in the camera-ready version of the paper. Also, we apologise for the lack of clarity in our description – we will improve the description of the method and add a pseudo-code algorithm to the paper in the camera ready version. > Would you expect this method to find emergent features in standard, well-understood systems, such as Ising models? Although the behaviour of the Ising model is often described as emergent, our method is not readily applicable to the Ising model. This is because our method is explicitly designed for time series data (as the title indicates), but the Ising model doesn’t have time per se – the Ising model specifies a probability distribution over an ensemble of spin configurations, without any dynamics. There are of course several options to add dynamics to the spins (e.g. Metropolis-Hastings, Glauber, Kawasaki, etc.), but none of these are an integral part of the canonical Ising model. For this reason, if one were to choose one of these dynamics and apply our method, then the kind of conclusions one may draw is whether or not e.g. Glauber dynamics display emergent features, and not about whether the Ising model does. For this reason, we have chosen instead a different well-understood system that has dynamics and is widely considered to display emergence: Conway's Game of Life. As mentioned above, we show that our method can successfully find emergent features from Game of Life time series and that these features capture the state of the particles in the system. > Can you explicitly write down the prediction-only objective in Fig.2? The prediction-only objective is $I^\text{S}_\varphi(f\_{\theta}(X_t); f\_{\theta}(X\_{t+1}))$, i.e. the mutual information between the past and future of the learned representation. This quantity depends on the parameters of the critic $\varphi$ and the parameters of the representation learner $\theta$, and is calculated using the SMILE estimator as defined in Eq. (4). Finally, in response to the limitation: > There is no open access to the code. As mentioned above, we didn’t include a link due to the constraints of the double-blind review. We will add a link to a publicly accessible Github repository with all the code once the requirements of double-blind review are lifted. In the meantime, we have sent the AC an anonymised link to the code for the model architecture. --- Rebuttal Comment 1.1: Comment: Thank you for your reply and clarifying my question regarding the notation. On the overall improved and clearer presentation for the final version I will put my trust in the new github repository, and also in the judgement in some of the reviewers who already found the first version easier to read than I did. They seem more familiar with the previous work that your method is based on. The new results, in particular using the game of life example, investigate exactly the kind of system I found missing as a benchmark in the original version of the article. I am going to raise the score in response.	Summary: This paper introduces a method for learning the causally emergent representation of time series data. Based on the Partial Information Decomposition (PID) and ΦID definition of emergent variables, the paper utilizes variational information lower bounds to estimate and optimize the emergence objective function. The paper further includes a Minimum Mutual Information term and a penalty term, to reduce redundancy and discover diverse emergent variables, respectively. Experiments on a synthetic dataset and a primate brain activity dataset show that the method is able to discover diverse causally emergent representation. Strengths: Discovering causally emergent representations is a very interesting topic, and has significance in a wide range of scientific disciplines. The paper is inspirational and written clearly. Although the components of the method, i.e. definition of emergence objective function, and variational bounds for mutual information, are not new, their combination to discover causally emergent representations in a learnable way is interesting, and to my knowledge, novel. Weaknesses: As discussed above, the novelty is a little limited. This can be compensated by solid evaluations with a wide range of interesting datasets. I think the place the paper needs most improvements is more diverse and extensive evaluations. The paper can benefit from a few more datasets, both synthetic and real world, including the other datasets used in [1] and other references. If there exists baselines for discovering causally emergent representations, those baselines should also be compared against. Reference: 1. Rosas, Fernando E., et al. "Reconciling emergences: An information-theoretic approach to identify causal emergence in multivariate data." PLoS computational biology 16.12 (2020): e1008289. Technical Quality: 3 Clarity: 4 Questions for Authors: N/A Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: The authors did not explicitly state the limitation of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: As mentioned in the overall rebuttal, we have now conducted a wider range of evaluations, both using the same architecture on more datasets and using other architectures on the same datasets. * Specifically, we add two more brain activity datasets which capture different aspects of neural dynamics: fMRI, with high spatial resolution and low temporal resolution; and MEG, with low spatial resolution but high temporal resolution. Our method is able to find emergent features in both cases. * We ran experiments on one more synthetic dataset used in Rosas et al. 2020, Conway's Game of Life. Here we found that our method can learn emergent features, and that these can be interpreted as encoding the state of the particles, as one would expect. * We compare against standard RNN and MLP architectures that predict the future state of the time series, without any information-theoretic loss function. We confirmed that neither the RNN or the MLP are able to learn emergent features. To the best of our knowledge, there are no established baselines for discovering causally emergent representations – if anything, the Game of Life is one of the most long-standing canonical examples of emergent behaviour generated by relatively simple rules, and we now show our method works on it. We would of course be happy to consider any other suggestions. Finally, in response to the limitation: > The authors did not explicitly state the limitation of the paper. We apologise for the lack of clarity on this regard. We state some of the limitations of our work in Section 4, although we acknowledge this needs to be more clearly laid out. For the camera-ready version we will add an explicit discussion of our method's limitations. --- Rebuttal Comment 1.1: Comment: Thanks for the response. The added experiments provide more solid support for the claims of the paper. Thus, I increased my score.	Summary: The paper presents a method for identifying emergent variables in time series data through a novel machine learning architecture. It uses unsupervised learning for representation and information theory to find emergent properties in systems, which are often complex and not easily describable at the microscale level. The paper is motivated by the fact that unsupervised learning can be a powerful tool for identifying emergent properties, but current approaches limit to only information theory. The method rests on maximizing an objective defined by subtracting the mutual information of state variables at time t and the coarse graining at time t + 1 from the mutual information of the coarse grainings at t and t + 1. In other words, the amount of emergent information. Information theoretic definition of emergence is thus used to facilitate unsupervised learning. The method is tested on synthetic data and real-world ECoG brain activity data, demonstrating its ability to identify known and novel emergent features effectively. Experiments are conducted on a synthetic dataset and a macaque brain activity dataset. For the synthetic dataset, the method is able to estimate the ground-truth value of psi (the difference that is central to the objective function). For ECoG data, skip connections were introduced into the architecture, and once again found emergent representations. The paper concludes with a discussion on related (info theory) work, limitations, and future steps. Strengths: ### Clarity - Diagram features are well designed and results features are clear and salient - Though writing is somewhat unstructured, the shorter-range explanations are well-done - Methodology is given in detail. Lots of helpful explanation of relevant information theory, as well as the overall approach ### Quality - Good to have primate brain data, though more interpretation would help - Covers all the basic needs for a new method: real data, novel setup, suitable metrics (though they need more explanation) - Experimental setup is well-designed to demonstrate that emergent variables are being learned ### Originality - As far as I know, applying the information theoretic definition of emergent variables as an objective and training in this setting is novel ### Significance - An innovative idea that shows promise. While there could be more experimentation, this is a promising and new direction. Weaknesses: ### Clarity - It's not immediately clear how to interpret results. The paper shows figures, but it doesn't explain them much. Interpreting them requires a lot of re-reading the methods section - Writing is somewhat verbose and unstructured, and occasionally reads like a process statement ### Quality - This idea is compelling and innovative! The loss built on MI of coarse grainings and state variables is intuitive while creating a solid foundation for taking advantage of the capabilities of unsupervised learning - On the ECoG dataset, giving intuition/semantic understanding of emergent features (or at least attempting interpretation) would be cool - Limited experiments on real data in general - ultimately, only one experimental setting is shown as far as I understand. The synthetic problem, while useful, is simple - Lacking baselines or extensive comparison to existing methods, even if purely information-theoretic ### Significance - It would help to have clearer comparison to existing methods so that we could see the value-add of this innovation, not just the novelty and value alone Technical Quality: 3 Clarity: 2 Questions for Authors: None Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: In response to the specific weaknesses identified: > It’s not immediately clear how to interpret results. The paper shows figures, but it doesn’t explain them much. Interpreting them requires a lot of re-reading the methods section We apologise for the lack of clarity in our explanations. For the camera-ready version we will make sure the figures are explained in more detail and the text flows more smoothly. > On the ECoG dataset, giving intuition/semantic understanding of emergent features (or at least attempting interpretation) would be cool We enthusiastically agree with the reviewer that obtaining a semantic understanding of the learned emergent features in ECoG would be definitely cool. However, we see this as a longer-term research programme that our method is enabling. Since we do not have a sense of “ground truth” in terms of what the emergent features of brain activity are, this is difficult to do as part of this paper. In cases where we do have a reasonable sense of what these features might be (e.g. in the synthetic dataset and in Conway’s Game of Life), we conducted post-hoc analyses to interpret these features, and found that they align with our intuitions. For the camera-ready version we will emphasise the need for future work interpreting the learned emergent features in real-world data. > Limited experiments on real data in general - ultimately, only one experimental setting is shown as far as I understand. The synthetic problem, while useful, is simple As mentioned in the global rebuttal, we have now added experiments on two more real-world datasets, and demonstrate our method can successfully learn emergent features on both. We also add one more synthetic,but far less simple, dataset using Conway’s Game of Life system, and use our method to learn emergent features from it and interpret them with post-hoc analyses. > Lacking baselines or extensive comparison to existing methods, even if purely information-theoretic In addition to new ablation studies, showing comparisons against other information-theoretic loss functions, we now compare our method against a standard RNN and a standard MLP architecture (without any information theory). The results suggest that RNNs and MLPs do not learn emergent features on their own, supporting the value of our method. > It would help to have clearer comparison to existing methods so that we could see the value-add of this innovation, not just the novelty and value alone We agree that comparisons against existing methods can make a clearer argument in favour of our paper. As mentioned above, we now show that standard RNN or MLP do not spontaneously learn emergent features. Additionally, we show in post-hoc analyses that the combination of the learned emergent feature of our method and an RNN representation enables better predictions than either of them in isolation (informally: a small RNN plus an emergent feature is better than a large RNN), making a direct case for the value-add of our method. --- Rebuttal Comment 1.1: Title: Thanks for the response Comment: Thanks for the response and the new work. I do think the new experiments help, and the writing looks better. In retrospect, I think I would have originally given a 5 or 6, and raised it to a 7, so I'm going to keep the current score.	Summary: The paper introduces a novel objective function and deep learning architecture that are targeted to extract emergent latent representations from time series data. Motivation is very clear. The definition of emergent latent representation interesting and useful. The utilization of mutual information estimators (lower bounds thereof) is smart. Evaluations are restricted to a fully artificial and a fully neurobiological dataset. Strengths: The study of emergence and its conceptual and mathematical formalization is of general interest to neural information processing and the involved (part-of) cognitive science subdiscipline within. The authors utilize an existing definition thereof [30] as well as an approximation technique of a lower bound on mutual information (SMILE, [32]), which they combine in a highly elegant manner to yield their learning architecture. The usage of a linear temporal predictor with learnable residuals is a great way to bootstrap the system’s abilities. Even multiple emergent latents can be successfully extracted. A real-world dataset indicates applicability beyond artificial data. Paper is very well written – relatively easy to comprehend and all steps taken are very well motivated and sufficient background is provided. Weaknesses: System evaluations are minimalistic and not as convincing as I had hoped. Both, comparisons to potential baseline algorithms as well as more ablations are missing. Furthermore, one artificial dataset and one not well-motivated real-world neural dataset seems not enough to warrant publication. In particular, I would have expected at least one if not multiple DL/Neural Network baselines that do not pursue the information-theoretic objective but simply a standard temporal prediction objective. Those probably do not work on the parity problem at all but at least an attempt seems needed. That is, use a DREAMER-like world model learning architecture with probabilistic latents and see if structure emerges. Ablations could have explored more than just the same architecture without the penalty / adjustment term or without the macroscopic estimator. Further, in the artificial dataset the correlation coefficients $\gamma$ are quite high – was this necessary? When does this break down? Evaluations with a non-linear prediction pipeline would also be useful. Technical Quality: 3 Clarity: 4 Questions for Authors: Emergence comes in many forms – I wonder if you could discuss alternative definitions / alternative perspectives on the matter? Line 52 – a “not” too much. Eq (2): can you motivated the adjustment term further the summand (n-1)min_i… ? Are there alternatives to this that would be more targeted towards actually identifying true redundant mutual information? Eq (4). Could you motivate the clip operator slightly more? Line 102 should read “accurately” Paragraph 113ff: maximize / minimize information – I am not sure if this is worded the right-way round – could you double check and slightly reword? Line 137 – should not read “also” Line 142 – one “is” too much Line 190ff. removing the macroscopic MI term seems not to save much – or does it? The observation is interesting, but I wonder if the authors want to make a computational argument here as well. About the biological dataset – this is very ad-hoc somewhat. What does this analysis tell us really except that there are some complex spatio-temporal statistical correlations in the data? I find this one marginally useful. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Unclear how robust this is as ablations and comparisons are not very extensive. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: As mentioned in the global rebuttal, we have now conducted extensive additional evaluations, including: * Two new real-world datasets and one new synthetic dataset; * New analyses on the existing synthetic dataset with lower correlation coefficients; and * Comparisons against baseline algorithms based only on standard temporal prediction. Overall, our method was able to learn emergent features in all these new scenarios, the resulting features were interpretable for the synthetic cases where a notion of “ground truth” emergence is known, and baseline algorithms (standard RNN and MLP) do not learn any emergent features. In response to your questions, we have fixed all the typos and rephrased the relevant sentences in the paper. Here are responses to the rest of the questions: > Emergence comes in many forms – I wonder if you could discuss alternative definitions / alternative perspectives on the matter? We totally agree that emergence comes in many forms, and we acknowledge that our paper could do a better job at placing our definition within the broader literature on emergence. For the camera-ready version we will include a discussion regarding alternative definitions of emergence, and how this could be explored in future work. > Eq (2): can you motivated the adjustment term further the summand $(n − 1) \min_i ...$ ? Are there alternatives to this that would be more targeted towards actually identifying true redundant mutual information? The reviewer is right in that the discussion surrounding the adjustment term could be expanded. For the camera-ready version we will include a proof of where this adjustment term comes from, and a discussion including what redundant information is and what role it plays in the formula. As a matter of fact, there are multiple measures of redundant mutual information in the literature, and a comparison in this context would be an interesting avenue for future work. However, we would like to emphasise that the current form of the adjustment is a recognised and well-defined measure of redundant information (in the sense that it satisfies many of the natural properties one would require of such a measure), as shown in the following paper: Barrett, Adam (2015). Exploration of synergistic and redundant information sharing in static and dynamical Gaussian systems. Phys. Rev. E 91, 052802. doi: [10.1103/PhysRevE.91.052802](https://www.doi.org/10.1103/PhysRevE.91.052802) > Line 190ff. removing the macroscopic MI term seems not to save much – or does it? The observation is interesting, but I wonder if the authors want to make a computational argument here as well. We agree, thanks for this observation. To clarify, there are in fact two arguments for doing this: one is the computational argument to save some compute; and the other is that in some cases we observed it stabilised learning dynamics (our hypothesis is that it alleviates the adversarial relationship between the different components of our objective). For the camera-ready version we will clarify these issues and include both arguments. > About the biological dataset – this is very ad-hoc somewhat. What does this analysis tell us really except that there are some complex spatio-temporal statistical correlations in the data? I find this one marginally useful. Our motivation to study emergence (as described in the introduction) is to understand how cognitive systems encode information in macroscopic variables, which naturally motivates us to test our methods on data from biological brains. The fact that the same method that successfully identifies particles in the Game of Life detects emergent dynamics in brain data is extremely encouraging, suggesting that neural systems may encode information into “glider-like” collective features. This paper provides direct empirical evidence supporting this idea. Moreover, it is particularly encouraging that the emergent character is observed in multiple neuroimaging modalities that cover multiple spatial and temporal scales of the brain's dynamics. Needless to say, there is much future work to be done examining precisely what the role of emergent dynamics in the brain is, and how these emergent features contribute to brain function. We see this as a longer-term research programme, which our method is enabling for the first time. Since we do not have a sense of “ground truth” of what the emergent features of brain activity should be, this is difficult to do as part of this paper (although note that for cases where such ground truth is known, such as the bitstrings and the Game of Life, the results align with our intuitions). For the camera-ready version, we will emphasise the significance of these findings and the need for future work interpreting the learned emergent features in real-world data. Finally, in response to the limitation: > Unclear how robust this is as ablations and comparisons are not very extensive. As mentioned in the global rebuttal, we have now performed: 1. Further ablation analyses, showing the expected behaviour that key properties of the architecture (e.g. ability to learn diverse features) are lost if the corresponding elements of the objective are removed; and 2. Further comparisons with other, non-information-based architectures (e.g. RNNs, MLPs). --- Rebuttal Comment 1.1: Comment: Great rebuttal with new datasets and results as well as very good answers! The provided additional information make me confident that the final paper will turn out to be very good and very useful for the ML community - at least for those that are interested in emergence. Please make sure that you use the additional page left accordingly (submission was less than 8 pages long). I raise my score by one (i.e., to 7 - accept).	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for their time and thoughtful comments on our paper. We are encouraged by the positive feedback, and are thankful for the constructive suggestions that have let us identify and address several limitations of our paper. We have added responses to each reviewer’s specific comments, and provided answers to all of the reviewers’ questions (where applicable). More broadly, we have identified a number of overarching themes that multiple reviewers brought up and would like to address here. Below is a non-exhaustive list of work we have done motivated by the reviewers’ suggestions, together with a description of changes we are ready to make for the camera-ready version of the paper: * Evaluations in more real-world datasets. We have conducted experiments on two more real-world datasets involving two different brain scanning modalities: functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG). Results reveal that our method was able to successfully learn emergent features in both datasets (rebuttal Fig. 1). This shows that our method is flexible and effective; and it provides a stronger motivation for the real-world data analyses as showing emergent behaviour is relevant across multiple scales of brain activity – microscale with ECoG, mesoscale with MEG, and macroscale with fMRI. * Evaluation in a new synthetic dataset and additional ablation analyses. We have also used our method to successfully find emergent features in Conway’s Game of Life (a prominent and long-standing example of emergent behaviour), which complements our results via a second synthetic dataset from a well-understood system (rebuttal Fig. 2). In addition, we have also conducted further ablation studies, which clarify the specific role of each component of our architecture. Finally, we also replicated our results for additional $\gamma$ values in our synthetic dataset, which provide further evidence confirming our previous findings (rebuttal Fig. 3). * Better interpretation of learned features. Although reverse-engineering the function of emergent features of the brain is a larger, long-term research problem, we are able to provide clear interpretations of learned emergent features in systems where some notion of “ground truth” is available. Specifically: 1. For the “random bits” synthetic dataset, we show that the learned feature can be interpreted as capturing the parity of the bitstring (paper Fig. 2). 2. For the new results on the Game of Life, we show that the learned feature captures the state of the so-called particles (i.e. gliders) in the system (rebuttal Fig. 2). * Comparison against standard architectures. We have compared the performance of our method against a standard RNN and a standard MLP trained only on a temporal prediction objective, both on real-world (rebuttal Fig. 1) and synthetic data (rebuttal Fig. 4). The results show that neither the RNN nor the MLP by themselves learn the emergent features of the data, and thus our method can provide a unique value for predicting datasets with emergent dynamics. As quantitative evidence for this claim, we show that the combination of a learned emergent feature and the RNN representation can predict the future state of the synthetic dataset better than either of them in isolation, highlighting the value-add of our method in conjunction with other algorithms. * Improved presentation. In the camera-ready version we will improve the description of our architecture and our figures, provide a pseudo-code algorithm, and discuss our framework in the broader context of alternative definitions of emergence. * Code availability. We will add a link to a publicly available Github repository once the requirements of the double-blind review are lifted. In the meantime, we have sent the AC an anonymised link to the code for the model architecture. We hope this new work addresses some of the reviewers’ concerns, particularly regarding the addition of a broader range of evaluations on new real-world and synthetic datasets. Please do not hesitate to ask more questions if any aspect of our work remains unclear. And, once again, we would like to thank all the reviewers for their extremely valuable feedback. Pdf: /pdf/00435cc453fbc52362ee0aba21fee46bba451485.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Vision Model Pre-training on Interleaved Image-Text Data via Latent Compression Learning	Accept (poster)	Summary: This paper introduces a vision backbone pre-training method named Latent Compression Learning (LCL) to utilize interleaved image-text data. The proposed LCL approach maximizes mutual information between the inputs and outputs of a GPT-like model in autoregressive manner. The proposed method integrate both discriminative and generative objectives by contrasting preceding context and generate subsequent text based on visual representation. The extensive experiments demonstrate that LCL not only matches the performance of existing models like CLIP on paired datasets (e.g., LAION) but also effectively leverages interleaved pre-training data (e.g., MMC4) to learn robust visual representations from scratch. Strengths: 1. The paper is well written and easy to follow. 2. The paper introduces a new pre-training method, Latent Compression Learning (LCL), which utilizes interleaved image-text data for visual backbone pre-training for the first time. And this can effectively leveraging large-scale web-crawled data, which is easier to crawl compared to the image-text pairs. 3. Extensive experiments are conducted, demonstrating the effectiveness of the proposed method on both paired datasets (e.g., LAION) and interleaved datasets (e.g., MMC4). Weaknesses: 1. From Table 5, it appears that solely leveraging image-text pairs with LCL does not provide benefits over the CLIP baseline. However, when using the MMC4 dataset, which is manually composed of interleaved text, there is significant performance improvement on downstream tasks. I am curious whether this performance gain results from the increased number of training samples (i.e., the total number of images used during training). 2. According to Table 3, utilizing original interleaved datasets such as Obelics does not yield any performance gain. In comparison, the MMC4 dataset requires more computation for data filtering with the CLIP score and the use of image-text pairs to create interleaved data. It is unclear how to efficiently utilize the original interleaved data directly crawled from the web. Do you have any insights on the differences between these two types of interleaved datasets? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The scaling behavior is not demonstrated. While the author has shown the effectiveness of training on MMC4 and Laion-400M, it remains unclear how the model performs and correlates across different dataset scales. Understanding this could provide valuable insights into the feasibility and performance of scaling the proposed method to larger datasets, such as DataComp 12.8B and Laion 2B. 2. Can you show the seen samples of each model? It would be helpful for readers understanding the scale of model's training. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed the limitation in their manuscript. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your good questions and constructive suggestions. ___ Q1: In Tab. 5, LCL is on par with CLIP baseline solely with image-text pairs but is significantly better when using the MMC4 dataset. Whether this performance gain is from the increased number of training samples. A1: It is reasonable that our LCL is comparable to CLIP on image-text pairs. Our LCL uses the simply constructed LAION-Random dataset (see Section 4.1), degrading to using half of the data for contrastive learning and the other half for generation. Compared to CLIP which uses all data for contrastive learning, there is no significant advantage or disadvantage. When incorporating MMC4 data, the significant improvement comes from increased data diversity. As mentioned in Section 4.3, all models in Tab. 5 are exposed to the same total number of training images. The number of images per epoch doubles when using LAION-400M + MMC4. Tab. 3 also shows that using MMC4 and LAION-400M alone are on par. This suggests that data diversity is the primary factor. Therefore, the advantage of our LCL is its ability to use larger-scale and more diverse interleaved image-text data, while not suffering performance losses when using paired data. ___ Q2: In Tab. 3, using original interleaved datasets such as Obelics does not yield performance gain. MMC4 uses CLIP score to filter image-text pairs and create interleaved data. Any insights on the differences between these two types of interleaved datasets? A2: This is a very good question, and we suppose that different types of interleaved data may be suitable for different training tasks. Our results suggest that using filtered interleaved data like MMC4 may be better for vision model pre-training. We note that recent work OmniCorpus[a] shows that using original interleaved data like Obelics may be better for fine-tuning MLLM (see their Tab. 4). [a] OmniCorpus: An Unified Multimodal Corpus of 10 Billion-Level Images Interleaved with Text. arXiv:2406.08418. ___ Q3: The scaling behavior is not demonstrated. A3: We are also very interested in the scaling behavior of our method. We anticipate using datasets such as Laion-2B and OmniCorpus[a] (a large-scale interleaved dataset) to train our vision model to achieve SoTA performance. However, using a total of 34B samples like OpenCLIP, training a ViT-L/14 model would require at least 8000 GPU days, which is more than we can afford now. ___ Q4: Show the seen samples of each model. A4: The number of seen images is 13B for all models in Tab. 5 and 6, as mentioned in Section 4.3. For ablation studies, all models use 2B training images, as stated in Section 4.1. --- Rebuttal 2: Comment: Thank the authors for their response. And most of my concerns have been well-addressed. I'd like to maintain the rating in the current stage. --- Rebuttal Comment 2.1: Comment: We are glad that your major concerns have been addressed. Thanks for your thorough and constructive review.	Summary: The paper tackles the problem of vision model pre-training. More exactly, it aims to exploit the interleaved image-text data that is very prevalent on the Internet. It proposes Latent Compression Learning that maximises the mutual information between the inputs and outputs of a causal attention model. When visual pre-training is applied to interleaved image-text data, visual latents are extracted using a visual encoding network and then combined with the text and fed to the causal model. Strengths: The paper tackles an important task and proposes an interesting method that may be of interest to the research community. Weaknesses: While the method seems interesting, my main concern is related to the experimental part that I find confusing. For example for BEiT3 the numbers reported are different from the ones reported in the paper. Also, I think that for Tab 6, more multi-modal LLMs need to be included. While there can be a debate on fair vs unfair comparison, I think that you present results on a dataset these need to be complete. So, they can be greyed out, put in a different section, etc and explained why the comparison is not fair, but I don't think it's suitable for models that perform better to not be included at all. So, missing comparisons: Fang, Yuxin, et al. "Eva: Exploring the limits of masked visual representation learning at scale." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. Zou, Xueyan, et al. "Generalized decoding for pixel, image, and language." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. or even some very recent ones for the sake of completeness: Sun, Quan, et al. "Generative multimodal models are in-context learners." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. Liu, Haotian, et al. "Improved baselines with visual instruction tuning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: Why is OpenAI CLIP greyed out? Why are the numbers reported for BEiT3 different than what they report in the paper? For example for Flicker 3K, R@1, it's reported 73.2 while the paper reports 88.2 Why, for example the comparison with BEIT3 is only shown in Tab.1 when they report results on VQAv2? The same question for CoCA? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are barely discussed at the end of the conclusions. Some limitations can be inferred from the rest of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the reviews and questions. But there is a misunderstanding here. First, as discussed in Fig. 1, we would like to clarify that our proposed LCL aims to pre-train a vision encoder from scratch using interleaved image-text data, rather than incrementally training a multi-modal model based on a pre-trained vision encoder. Their differences are illustrated in Fig. 1(b) and 1(c) and also acknowledged by Reviewer ktxP. Accordingly, our experiment is to evaluate the quality of pre-trained vision encoders rather than to compare the performances of different MLLMs. Therefore, when comparing with other pre-training methods, instead of using their released MLLMs checkpoints, we use their released codes or reproduce their methods to pre-train the visual encoders from scratch in a fair setting. We ensure a fair comparison in the same settings, including the model structure, training data, and evaluation settings. ___ Q1: (a) For BEiT3, the numbers reported are different from the ones reported in the paper. (b) Why the comparison with BEIT3 is only shown in Tab.1 when they report results on VQAv2? The same for CoCA? A1 (a): The number of BEiT3 comes from training from scratch with BEiT3's pre-training task rather than the reported ones of the trained checkpoint. Tab. 1 compares pre-training methods and pre-training tasks, so the model and data for all methods should be fair. To avoid misunderstanding, we will clarify in the caption of Tab. 1 that the method names in parentheses refer to the pre-training tasks but not their trained checkpoints. A1 (b): Table 6 evaluates pre-trained vision encoders on multimodal dialogue tasks by integrating them with the same pre-trained LLM for fair comparison. Therefore, the reported results of BEiT3 and CoCa's released MLLM checkpoints are not included. In Table 1, BEiT3 and CoCa are compared as pre-training methods under fair settings. ___ Q2: In Tab 6, more multi-modal LLMs need to be included. A2: As mentioned above, our proposed LCL pre-trains vision encoder from scratch, which is orthogonal to the training of MLLMs. Accordingly, our experiment is to evaluate the quality of pre-trained visual features rather than to compare the performances of different MLLMs. Tab. 6 is to fairly compare different pre-trained vision encoders in multimodal tasks, so the results of other MLLM's trained checkpoints, such as EVA[a], X-Decoder[b], and Emu[c], should not be incorporated. We have used LLaVA-1.5[d] for this evaluation (see Appendix A.2, "Multi-modal dialogue"). There are better MLLMs available than LLaVA-1.5, but training better MLLMs is beyond the scope of this paper, and we do not need to make such comparisons. In addition, we have compared the text generation + image feature regression task used by Emu in Tab. 1. However, EVA requires pre-trained CLIP features and X-Decoder uses segmentation data, so their training tasks are not comparable to those methods pre-training from scratch only using image-text data. [a] Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023 [b] Generalized decoding for pixel, image, and language. In CVPR, 2023. [c] Generative multimodal models are in-context learners. In CVPR, 2024. [d] Improved baselines with visual instruction tuning. In CVPR, 2024. ___ Q3: Why is OpenAI CLIP greyed out? A3: We aim to compare vision pre-training methods rather than pre-trained checkpoints. Both OpenCLIP and we use the LAION-400M dataset for training, allowing for a fair comparison. OpenAI CLIP, which uses private data, is included only as a reference. --- Rebuttal Comment 1.1: Title: Rebuttal answer Comment: Hi. Thank you for writing the rebuttal! I confirm I have read the rebuttal. I understand that the goal is to pre-train a vision encoder from scratch and the rebuttal sort of adds more clarity, but I think more details should be added in the paper to make things clearer. The authors have promised to add more details to the captions, so I don't have much to add here. However, when reporting results on a well defined benchmark, I personally believe all types of methods should be included, especially if they have a better performance. I understand that the space is limited and I do not propose to have a list of extensive non-related paper, but I consider including at least 1,2 especially can be very important for readers of the paper in order to make them aware of different methods and possibly increased performance on the benchmark. As I said in the original review, what is fair or not can be shortly explained and the methods clearly separated in the table. --- Reply to Comment 1.1.1: Comment: We sincerely thank you for your insightful comments and suggestions. Please see the general response. We revise our table to include more benchmark results with further discussion for a better understanding.	Summary: The paper pre-trains models with a combination of a contrastive image-text objective and a generative language objective. The authors provide many results on image classification and vision-language tasks suggesting the competitiveness of the method in controlled settings. Strengths: S1. The paper is well framed and motivates nicely the need to pre-training on interleaved data. S2. The paper gives good intuition about what the various equations mean, making the manuscript more accessible. S3. Consideration of many pre-training datasets including LAION-400M, MMC4, and OBELICS. S4. Extensive objective ablations spanning both contrastive and generative losses. Weaknesses: W1. [MAJOR] The paper presents the objective as novel (L44-54); however, it seems similar to CoCa (Yu et al., 2022.), which also employs a contrastive loss and a next token prediction loss. Can the authors clarify the differences and why the formulation is novel? W2. It seems equation 3 appears in prior work; however, when it is first presented it seems to be presented as a novel insight. I recommend making the attribution to prior work more clear before introducing the equation. W3. In the relation to previous pre-training tasks, it is important to also relate to CoCa. It seems the objective is pretty much the same suggesting that the objective is not actually a contribution of the work. Is there any reason CoCa is not mentioned here given the similarities? W4. Make sure it is clear that you train on sequences with more than one image per sample (I am assuming this is true because you train on MMC4, but when explaining the objectives you include only one sequence for simplicity). 3.3 is a nice place to add this information. Also any special tricks to get multi-image to work? If so, it could also be nice to mention this. W5. Why are the numbers for Flamingo in Tab 1 for IN-1k so low? Flamingo uses a pre-trained vision backbone, so I expect numbers to be good here. W6. Is the COCO CIDEr evaluation protocol zero-shot? If so the number in table 4(a) of 87.5 looks extremely high relative to open flamingo and Idefics. Please double check this number and if few-shot prompting is used here, please make this clear. Also why is Gen. only worse than Con. only for captioning. How is contrastive learner able to do captioning? W7. In the frozen transfer setting in Tab. 6 are all models fine-tuned on the same data? If so, what data? The specifics of the experiment are not clear to me, making it hard to interpret the results. Technical Quality: 3 Clarity: 2 Questions for Authors: Please see the weaknesses section for specific questions and topics to address. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors address limitations in a dedicated paragraph. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the careful reviews and constructive suggestions. We answer the questions as follows. ___ Q1: (a) The proposed objective seems similar to CoCa, which also employs a contrastive loss and a next token prediction loss. Clarify the differences and why the formulation is novel. (b) In "relation to previous pre-training tasks", why CoCa is not mentioned given the similarities. A1 (a): CoCa cannot be extended to interleaved data since its learning objective relies on paired data. Applicable to interleaved data is a contribution of our proposed learning objective. There are significant differences between ours and CoCa in how the loss is applied: 1. CoCa's contrastive loss relies on the positive pairs from paired data, which cannot be directly obtained in general interleaved data. 2. CoCa's generation can only be conditioned on a single image, while our generation is conditioned on preceding context containing flexible interleaved image-text contents. A1 (b): In "relation to previous pre-training tasks," we aim to claim that existing pre-training methods cannot train visual features from scratch using interleaved data. CLIP is the most representative method for paired data, and CoCa is also specialized for paired data. Therefore, we did not specifically discuss CoCa in this context. Given the similarity in the loss form, we will include further discussions about CoCa in the revision and clarify our contribution as mentioned above. ___ Q2: Equation 3 seems to appear in prior work. Make the attribution to prior work more clear before introducing the equation. A2: Thanks for your suggestion. Prior work (i.e., M3I mentioned in the paper) showed the relationship between cross-entropy and mutual information in other pre-training tasks, e.g., image self-supervised learning and contrastive learning. While Equation 3 specifically demonstrates this relationship when using an autoregressive model to compress interleaved image-text data. Our novel insight is that, in this scenario, the compression (cross-entropy) objective and the maximum mutual information objective are equivalent. In the revision, we will make the attribution to prior work and clarify our novel insight clearer. ___ Q3: Make sure that it is trained on sequences with more than one image per sample. Any special tricks to get multi-image to work. A3: Yes, we support interleaved data containing multiple images. As described in Appendix A.1, each MMC4 sequence may contain 1 to 6 images. We did not use any special implementation tricks. See Fig. R2 in the rebuttal pdf for illustration. The visual features of multiple images are inserted at their positions in the sequence, combined with text embeddings, and then processed through a causal language model to compute the generation loss. For each image, its global feature and the language model’s output feature at its preceding <boi> token form a positive pair in contrastive loss. We will add this information to Section 3.3. ___ Q4: Why are the numbers for Flamingo in Tab 1 for IN-1k so low? A4: The number of Flamingo comes from training from scratch with the Flamingo's pre-training task rather than its released checkpoint. Tab. 1 compares pre-training tasks, so the model and data for all methods should be fair. To avoid misunderstanding, we will clarify in the caption of Tab. 1 that the method names in parentheses refer to the pre-training task but not their trained checkpoints. ___ Q5: (a) Is the COCO CIDEr evaluation protocol zero-shot? The number in Tab.4(a) of 87.5 looks extremely high. (b) Why is Gen. only worse than Con. only for captioning? How is contrastive learner able to do captioning? A5 (a): The results in Tab. 4 are evaluated under frozen transfer setting, which we will clarify in the table caption. See Appendix A.2 for frozen transfer details. Frozen transfer refers to fine-tuning downstream task models while the parameters of the pre-trained visual encoder are frozen. We use LAION-COCO for captioning fine-tuning, so the COCO CIDEr is reasonable. A5 (b): The vision encoder pre-trained by "Con. only" also performs captioning by frozen transfer (see Appendix A.2). "Gen. only" is worse than "Con. only" for captioning, because visual features pre-trained by "Gen. only" experienced a collapse. We have discussed this feature collapse in Section 3.2, "Relation to Previous Pre-training Tasks." ___ Q6: In the frozen transfer setting in Tab. 6, are all models fine-tuned on the same data? If so, what data? A6: Yes. The frozen transfer in Tab. 6 evaluates the pre-trained vision encoders on multimodal tasks following LLaVa-1.5. All models use the same data and training settings as LLaVa-1.5. See Appendix A.2 "Multi-modal dialogue" for more details. --- Rebuttal Comment 1.1: Comment: Thanks for clarifying with respect to CoCa. I think given the similarities (e.g., Figure 1 of the manuscript and CoCa look very similar), it is important to make this comparison explicit. Also after reviewing other reviewers' comments, I am a bit concerned that state-of-the-art performance was not reported to contextualize the results. Additionally, many of the requested writing changes may require significant re-writing. Given this, I am electing to keep my score. Thanks to the authors for all of their effort in putting together the paper and the rebuttal. --- Reply to Comment 1.1.1: Comment: We sincerely thank you for your thorough and constructive review. In the revision, we will explicitly compare the differences between our objective and the CoCa's in the figure and text. Please see the general response for further discussion about the comparison with state-of-the-art performance.	Summary: This paper aims to explore the use of weak supervision signals in multimodal interleaved image-text data to pretrain visual encoder, compressing the distribution of high-level features into the visual encoder. The paper employs contrastive loss and autoregressive loss to train the model. To prevent the collapse of visual representations, an entropy maximization constraint is applied. The paper derives the equivalence of maximizing the mutual information between the model's input and output as a latent compression and entropy constraint. The proposed pre-training method, called LCL, achieves performance comparable to CLIP on paired data while better utilizing the supervision information in interleaved data. Strengths: This paper explores how to use weak supervision signals in more general interleaved image-text data to accomplish visual pre-training. Its advantages are as follows: 1. Unlike previous approaches that fine-tune pre-trained visual models to align visual representations with the text space (Flamingo, LLaVA), this paper explores how to train visual models from scratch using interleaved image-text data. This is a meaningful exploration. 2. To prevent the collapse of visual representations, where autoregressive generation relies solely on textual information, this paper imposes an entropy constraint and further derives it as optimizing mutual information. This approach aids in model training. 3. Extensive quantitative experiments have validated the effectiveness of the visual models trained using this approach. Weaknesses: This paper has the following areas for improvement: 1. In some cases, the textual context may have little relevance to the image. It is worth investigating whether such data could harm the model's performance. 2. The paper lacks qualitative experiments to further demonstrate the effectiveness of the method. Designing reasonable visualization analyses would help to further elucidate the advantages of the approach. 3. Similar to CLIP, further demonstrating the model's transfer learning performance through domain adaptation tests and few-shot metrics would be beneficial. Technical Quality: 3 Clarity: 3 Questions for Authors: see weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper's discussion on data bias and energy consumption limitations is commendable; however, it could further explore issues related to data privacy. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your good questions and constructive suggestions. ___ Q1: In some cases, the textual context may have little relevance to the image. It is worth investigating whether such data could harm the model's performance. A1: This is a very good question. It is inevitable that image-text data obtained from the internet may have weak correlations. For fair experiments, we currently follow the settings in previous works when using datasets. Previous works have used filtering methods to improve the quality of data. For example, the LAION-400M dataset used CLIP scores to filter image-text pairs, and OpenFlamingo used CLIP scores to filter out uncorrelated images in the MMC4 dataset. We now follow their settings. We note that recent work, OmniCorpus[a], analyzed model pre-training with differently filtered data in Section 5.2 (Table 3). The results show that appropriate filtering can improve the correlation of images and texts, thereby enhancing model performance. However, excessive filtering may lead to insufficient data diversity, thus hurting performance. We will analyze the impact of data quality on our model in future work. [a] OmniCorpus: An Unified Multimodal Corpus of 10 Billion-Level Images Interleaved with Text. arXiv:2406.08418. ___ Q2: Qualitative experiments to further demonstrate the effectiveness of the method. Designing reasonable visualization analyses to further elucidate the advantages of the approach. A2: We use t-SNE to visualize the learned features using images from some classes in ImageNet-1K val set. See Fig. R1 in the rebuttal pdf. For both CLIP and our LCL, visual features of different classes are generally distinguishable, and the distance between classes is related to semantic similarity. Compared to CLIP, visualized points of LCL are slightly tighter for semantically similar classes and slightly further away for classes with low semantic relationships. ___ Q3: Similar to CLIP, further demonstrating the model's transfer learning performance through domain adaptation tests and few-shot metrics. A3: Thanks for your suggestion. We adopt zero-shot domain adaptation tests in CLIP. Pre-training ViT-B/16 on LAION-400M, OpenCLIP achieves 67.1 on zero-shot ImageNet-1k, while ours is 60.1. However, domain adaptation via zero-shot retrieval is designed for contrastive learning such as CLIP. It may not be suitable for other types of pre-training methods (e.g., MIM, generative tasks). We have discussed this in Section 4.1. Moreover, similar to CLIP, we have used probing (fine-tuning a task head with backbone frozen) to evaluate transfer learning capabilities. See Tab.5, IN-1k frozen transfer results. Our LCL is comparable with OpenCLIP, demonstrating the strong transferability of our pre-trained vision model. --- Rebuttal Comment 1.1: Comment: Thank you for your discussion on data quality and the analysis of LCL generalization performance. I believe that the contextual relevance of interleaved data (i.e., data quality) has a far greater impact on model performance than the quality of paired data. Furthermore, in real-world scenarios, the generalization performance of the model is a key metric for evaluating pre-trained representations, so it would be meaningful to analyze it. However, after reading the comments from other reviewers, I am concerned that the paper does not thoroughly discuss its relationship with other methods, especially regarding one of the core contributions of this paper—the pre-training objective. Additionally, as other reviewers have mentioned, further listing some state-of-the-art (SOTA) methods in the evaluation and clarifying how to achieve a fair comparison would further reduce misunderstandings for readers. Considering the above factors, I will somewhat lower my rating. Further discussion has the potential to affect my ratings. --- Reply to Comment 1.1.1: Comment: We sincerely thank you for your thorough review and constructive comments. We respond to your concerns in the general response.	Rebuttal 1: Rebuttal: We thank all the reviewers for the careful reviews and constructive suggestions. We respond to your questions respectively. The PDF contains supplementary figures for rebuttal. Pdf: /pdf/0246ee7812b55b519c2e455843c003d1f5c25cb4.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation	Accept (poster)	Summary: This paper aims to improve vectorized HD map construction for autonomous driving. Inspired by the global feature association in traditional offline HD mapping, the proposed MapUnveiler processes input frames in a clip-based manner and hopes to resolve occlusions using information from previous frames. Built up MapTRv2, MapUnveiler introduces clip tokens together with the Inter-clip and Intra-clip Unveiler modules to update the map queries with temporal information. Experiments on nuScenes and Argoverse2 datasets demonstrate the superior performance of the proposed method, especially on highly-occluded scenes. Strengths: 1. The idea of incorporating and aggregating clip-level information for online vectorized HD mapping is reasonable and is more akin to how humans drive. The proposed method has more thoughtful designs than early works such as StreamMapNet to better handle occlusions and incorporate long-range information. 2. The proposed MapUnveiler obtains state-of-the-art results in various experimental settings. The improvements over previous methods are especially prominent in the large 100mx50m setting and the highly-occluded scenes collected by the authors. 3. Extensive ablation studies enumerate the choices of almost all hyper-parameters or model components, which helps better understand and break down each element's contributions. Weaknesses: 1. The clarity of the method description is poor, making it very hard to thoroughly understand the proposed architecture. Details are discussed below: - The method explanation is not self-contained: i) The Inter-clip Unveiler section refers to the TTM and directly skips all details. There is no information at all about how is the compact memory token generated from the denser map queries; ii) The "loss" section refers to MapTRv2 and again skips all details. The authors should not assume the general audience to be aware of the concrete details of TTM and MapTRv2. The core formulation of these components should be elaborated with texts or equations, while full details can go to the appendix. - The definitions of the temporal window T and the stride S are unclear. Based on the text descriptions and the common definition of stride, my understanding of "T=3 and S =2" is that "each clip has 3 frames, and every two consecutive frames have a temporal gap of 1." However, the symbols in L177-178 seem to suggest other meanings of T and S. - The description of the inference mechanism is also vague. Is the MapUnveiler executed per frame or per clip? Figure 2 seems to suggest the per-clip inference where the predictions of T frames are obtained together. If this is the case, does it hurt the actual response frequency? In short, Section 3 of the paper lacks significant details, and I cannot properly understand MapUnveiler's exact formulation. Given that the authors answer "No" to Question 5 of the Checklist, I have to raise concerns about the paper's reproducibility. 2. There is no detail on how the pre-training and fine-tuning are conducted. Do you initialize the MapNet by training MapTRv2? If this is the case, how are the training epochs split for the MapNet pre-training and the end-to-end MapUnveiler fine-tuning? If the 24/6 epochs for nuScenes/Argo2 are only for the fine-tuning stage, then the comparisons in the main table are unfair, as other methods in the table have not fully converged. 3. The main comparison results are incomplete. Most previous papers provide the nuScenes results of both short and long training schedules, but the main table only presents short-schedule results. Considering the last question about the pre-training and fine-tuning, the authors should complement the table with long-schedule results to show that MapUnveiler can obtain consistent performance boosts when all the methods are fully converged. This concern is backed up by the fact that MapUnveiler's improvement is much smaller on Argo2 compared to nuScenes -- based on my empirical experience, previous methods like MapTRv2 and its followups converge faster on Argo2, and training for 6 epochs is close to convergence. This probably suggests that the large performance gaps on nuScenes come from unfair training settings. 4. Your interpretation of StreamMapNet and SQD-MapNet's Argo2 training epochs is wrong. These two methods employ a different frame sampling strategy at training time compared to MapTRv2, but their effective number of training samples is the same as MapTRv2. Therefore, the claim about the "longer training schedules" in the main table's caption is misleading. 5. The venues in the main table are not accurate. HIMap[49] and MGMap[24] are accepted by CVPR2024, and the information was already available at the time of NeurIPS submission. Furthermore, a recent HD map construction method, MapTracker[A], also studies temporal modeling and should be very relevant, but it is missing in the discussion and related works. [A] MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping, arXiv:2403.15951 Technical Quality: 3 Clarity: 1 Questions for Authors: The paper studies an important problem (temporal information association) in online HD map construction and proposes a reasonable method. However, the poor clarity and the potentially incomplete/unfair comparison results raise serious concerns about the paper's quality and reproducibility. My current rating is reject, and I will consider changing the score if the main weaknesses are properly addressed. Confidence: 5 Soundness: 3 Presentation: 1 Contribution: 2 Limitations: The limitations and broader impacts are adequately discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We deeply appreciate your thorough and insightful feedback. We believe we can enhance the paper's quality and present more concrete results based on your comments. Below are point-by-point responses for your comments, and these will be included in the revised paper. ___ W1-1. Explanation of TTM and MapTRv2 Thanks for your valuable comment. Initially, we tried to avoid overclaiming in terms of model architecture, so we opted to skip the explanation of TTM and MapTRv2. However, we fully understand that this is not good for general readers. Following your advice, we will improve the presentation as follows. i) Inter-clip Unveiler In Line #172, we will introduce an equation as > $U^{read}=Read(U_{t-2S:t-S}^{memory}, Q^{map})=S_{N_c}([U_{t-2S:t-S}^{memory}\|\| Q^{map}])$ Here, the notation $[U_{t-2S:t-S}^{memory}\|\| Q^{map}]$ denotes the concatenation of two elements. In Line #178, we will introduce an equation as > $U_{t-S:t}^{memory}=Write(U_{L}^{clip}, U_{L}^{map}, U_{t-2S:t-S}^{memory})=S_M([U_{L}^{clip}\|\| U_{L}^{map} \|\|U_{t-2S:t-S}^{memory}])$ If the tokens within the memory are not re-selected in the subsequent steps, it will be removed from the memory, and the selection mechanism will be determined through the learning. ii) Loss To provide detailed information about the losses, we will add following equations: > $\mathcal L_{one2one} = \lambda_c^F\mathcal L_{cls}^F + \lambda_p^F\mathcal L_{p2p}^F + \lambda_d^F\mathcal L_{dir}^F$ > > $\mathcal L_{dense} = \alpha_d\mathcal L_{depth} + \alpha_b\mathcal L_{BEVSeg} + \alpha_p\mathcal L_{PVSeg}$ > > $ \mathcal L_{Frame-level MapNet} = \beta_o\mathcal L_{one2one} + \beta_d\mathcal L_{dense}$ > > $\mathcal L_{MapUnveiler} = \lambda_c^M\mathcal L_{cls}^M + \lambda_p^M\mathcal L_{p2p}^M + \lambda_d^M\mathcal L_{dir}^M$ where $\mathcal L_{one2one}$ is used for frame-level MapNet, and we set with $\lambda_c^F=2, \lambda_p^F=5, \lambda_d^F=0.005$. $\mathcal L_{dense}$ is an auxiliary loss using semantic and geometric information, and we set $\alpha_d=3, \alpha_b=1, \alpha_p=2$. $\mathcal L_{MapUnveiler}$ is used for MapUnveiler, and we set $\lambda_c^M=2, \lambda_p^M=5, \lambda_d^M=0.005$. Due to the character limit in the rebuttal, we conceptually show the revised explanation. We will further detail TTM, MapTRv2, and each loss in the Appendix. ___ W1-2. The definitions of the temporal window $T$ and the stride $S$ To provide a clear explanation, we will introduce the following equation. >$C_k = \lbrace f_t \rbrace_{t=kS+1}^{kS+T}$ The notation $C_k$ represents the k-th clip, starting from k=0. Each frame is denoted by $f_t$, represents the t-th frame among the consecutive frames. Therefore, when T=3 and S=2, we obtain clip sets such as $C_0=\lbrace f_1, f_2, f_3 \rbrace$, $C_1=\lbrace f_3, f_4, f_5 \rbrace$, ..., and $C_k=\lbrace f_{2k+1}, f_{2k+2}, f_{2k+3}\rbrace$. Thus, the temporal stride S refers to the clip stride, not the frame stride. Additionally, it would be interesting to additionally set frame stride, as you mentioned. If we set the frame stride=2 and the clip stride (S)=1 (to see all frames), MapUnveiler achieves an mAP of 68.8% in nuScenes. This is lower than the model with frame stride=1 and T=1, (69.3%). We conjecture that temporally nearest frames provide the most rich information for unveiling maps because they have large spatially overlap regions, and temporally distant frames can effectively be exploited through the inter-clip Unveiler. ___ W1-3. Inference mechanism MapUnveiler excutes per-clip. Thus, it can cause response delay if the input frame rate is slower than 12.7 (which is the inference speed of MapUnveiler) because the model will wait until it collects T frames. To avoid this problem, we can set the clip stride (S) to 1 (but it will lead to a slight performance degradation, as presented in Table 7). Alternatively, we can fill in the intermediate frames' results with frame-level MapNet: during the response delay, we can directly construct an online map using map queries generated from the frame-level MapNet. In this case, however, the performance will drop from 69.8% to 68.0%. W1-4. Open access code We are planning to release source codes upon acceptance. ___ W2. Detail on how the pre-training and fine-tuning We end-to-end fine-tune MapUnveiler for 24/6 epochs from pre-trained MapTRv2, which was trained for 24/6 epochs. To address this concern, we present the freeze fine-tune setting in Table 4, but we understand it cannot fully address this concern. So, we additionally present experimental results where we pre-train MapTRv2 for 12/3 epochs and then fine-tune MapUnveiler for 12/3 epochs, totaling 24/6 epochs. The results are given in G1 of the global response. ___ W3. Long training schedules Thanks for your thorough feedback. We trained 110/30 epochs, and MapUnveiler achieves mAP of 70.6% and 72.9% on nuScenes and Argoverse2, respectively. As you commented, the performance improvement of our model is marginal. We think that we can converge more quickly because we start training from a pre-trained frame-level MapNet. Even though our model lags behind HIMap (73.7% and 72.7%) in the long training schedule setting, we would like to kindly emphasize the following three points: (1) HIMap is heavy (9.7 FPS) and runs slower than ours (12.7 FPS); (2) HIMap requires more training time to fully converge, which may cause overfitting; (3) HIMap was published arXiv two months ago when we submitted our paper, so this work stands as a contemporaneous work. ___ W4. Caption in Table 1 Thanks for the correction. We will remove † in Table 1. ___ W5. HIMap[49], MGMap[24], and MapTracker[A] Thanks for the thorough check. We will revise them as CVPR2024 and will discuss with MapTracker. We would like to kindly inform the reason for notating as preprints: the official paper list of CVPR2024 was opened on the same day as the deadline for NeurIPS2024 submissions. --- Rebuttal 2: Comment: Thanks for the detailed response. The added details clarify the architectural design, especially for the clip-based inference mechanism and how $T$ and $S$ work. I believe these will also help other readers better understand your approach. I appreciate the new results in the general response, which resolve my concern about the fairness of experiments. However, please make sure the results in G4 also follow a fair setting, or you should at least indicate the difference in training schedules in the table. Overclaiming the performance boost with an unfair setting under the hood will hurt the community. With the above concerns addressed, I will increase my score from 3 to 5 (borderline accept). Reasons for not giving a higher score are mainly two-fold: 1) the clip-wise inference is a bit unnatural, making the output quality temporally inconsistent; 2) the initial writing clarity is very problematic, requiring a massive revision to improve the paper's quality. If the paper is accepted, I hope the authors can carefully improve the clarity of the method description (W1), fix the discussion/reference of related works (W5), provide fair experimental results and clearly indicate the difference in training settings (W2,3). --- Rebuttal Comment 2.1: Title: Thank you! Comment: We would like to express our sincere gratitude for your positive feedback and for raising your rating of our work from 3 to 5. We especially appreciate your thorough and professional comments, which have helped us improve the fairness of our experimental setup and enhance our presentation. As you suggested, we will revise the relevant results presented in the comparison tables (Table 1 and G4) with more fair settings (12/3 epochs fine-tuning) in the final version. Additionally, we will revise the final version by incorporating all comments from the rebuttal, including clarifications of our approach (W1), discussions of the suggested related works (W5), and clarifications of the training settings (W2, W3). Best regards, \ The Authors	Summary: The authors propose a new approach for constructing vectorized high-definition maps that exploits temporal information across adjacent input frames. The model, which they call MapUnveiler, operates at the clip-level and consists of an intra-clip unveiler which generates vectorized maps for T frames and an inter-clip unveiler which uses a memory module to aggregate information between clips. The authors present results on two standard benchmarks, vectorized HD map construction benchmarks (nuScenes and Argoverse2) and demonstrate the model’s superior quantitive performance to several previously proposed approaches. They also show several qualitative examples of how MapUnveiler can better handle occlusions in the input images. Strengths: - The paper is well-written and contextualized well within prior work. - The methodology is novel and well-motivated. - The results are strong on the two tested datasets, both quantitatively and qualitatively. - Many different analyses and ablations were included to justify the design decisions used within MapUnveiler and show its strengths. Weaknesses: 1. The methods is dense and a bit hard to read. The architecture figures help but are also a bit difficult to parse through. It would be helpful to try to weave more intuition into the text. 2. Claiming "-9.8%" is significant but "-6.0%" is comparable in the robustness to occlusion section seems a bit arbitrary (and potentially overstating MapUnveiler's performance, as a 6% drop is still considerable). I suggest the authors rephrase this sentence (and address similar claims in the paper). There are several typos throughout the paper. I have enumerated some here, but encourage the authors to do a detailed proofread: - 127: With there - 129: mapnet -> MapNet - 161: bev -> BEV - 167 parenthesis - 192 backwards parenthesis - 294: In addition, if we choose too short Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Have the authors tried quantized models to reduce GPU memory? It could be interesting to see if the gains from larger window sizes outweigh the losses from quantization. 2. The model still seems to struggle with some occlusions (a 6% drop from the standard split). Why do the authors think that is? Are these just very difficult cases or issues with the model? 3. The one limitation that was discussed seems like it can be tested. How does randomly dropped intermediate frames affect model performance? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: Only one limitation is included. I encourage the authors to think through other potential limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are particularly encouraged that the reviewer finds our method novel and well-motivated. And we highly appreciate your constructive comments and suggestions. Below are our responses to each of your queries, and we will include them in the revised paper. ___ W1. Weave more intuition into the text in 3. Method We will include our motivation in the method explanation to make it easier to follow and understand our method. For this, we will revise the paper as below (we highlighted the new and revised descriptions in italics): Clip Token Generator (Line #149-150) > ... , respectively. To globally gather intra-clip map features, we opt for a naive cross-attention [38]. Through this step, we obtain compact clip-level map representations, enabling efficient intra-clip communication with small-scale features. BEV Updater (Line #155) > ... original bev features. The main idea of BEV Updater is to avoid heavy computation in spatio-temporal cross attention. To achieve this, we do not directly communicate intra-clip BEV features, but instead decouple the spatial BEV features and temporal clip tokens. We then update the spatial BEV features with compact temporal clip tokens, effectively communicating spatio-temporal information with reasonable computational costs. The updated BEV ... Map Generator. (Line #161) > ... the updated BEV features $U_l^{BEV}$. Since the updated BEV features are spatio-temporally communicated, we directly extract map tokens $U_l^{map}$. Each map token represents a vectorized map element through a 2-layer Multi-Layer Perceptron (MLP). The map tokens ... ___ W2. Rephrase some sentences We are sorry, we understand that it may be an overstatement. We will rephrase Line 261 as follows: > ~, while MapUnveiler shows comparable performance~ MapUnveiler also shows a proformance degradation of 69.8%→63.8% (-6.0%), but it demonstrates a smaller performance gap compared to previous studies. Following your advice, we also tone down following sentences (Line 242): > Although MapUnveiler incorporates temporal modules, we achieve a ~fast~ resonable inference speed (12.7 FPS) compared to frame-level MapNet (MapTRv2 [22], 15.6 FPS), surpassing both performance and speed compared to the heavy state-of-the-art approach (HiMap [49], 9.7 FPS). ___ W. Typos Thank you for your diligent efforts to enhance the quality of our paper. We will rectify the mentioned typos and strive to correct any other potential typos. We will thoroughly proofread the paper from the beginning. ___ Q1. Model quantization That's an intriguing idea, and we hope to see the comparison. Unfortunately, we tried to implement the quantized models, but it was quite challenging to implement and evaluate the accuracy, memory usage, and speed of the quantized model. The primary reason behind this is that we developed our MapUnveiler using several custom addition and multiplication operators, which the FakeQuantize function does not support. Implementing it with TensorRT necessitates reconstructing the entire pipeline of our code, hence requiring additional time. Based on current knowledge, we can consider extreme quantization (4-bit) and a general solution used in TensorRT (16-bit and 8-bit). The performance drop due to 4-bit model quantization may be marginal in CNN architecture if we use advanced quantization algorithms such as [A]. However, transformer-based architectures currently lose more than 20% of the accuracy [B] with 4-bit quantization. As an alternative, if we adaptively combine FP16 and INT8 [C], we can accelerate the model speed x1.87 times with a marginal performance drop (about -1.19%). With this, we can boost MapUnveiler with a large window size (T=5) through quantization. However, the performance gain from a wide window size is marginal (mAP of 70.1% for T=5, and mAP of 69.8% for T=3) as shown in Table 7 of the main paper. We conjecture that our model effectively leverages long time dependencies through the inter-clip Unveiler, thus it may not be necessary to use a large window size. But it would be interesting to see the gains of the main model from quantization. We plan to implement the quantized model and include the results in the final version. [A] QDrop, ICLR 2022 [B] RepQ-ViT, ICCV 2023 [C] https://github.com/DerryHub/BEVFormer_tensorrt ___ Q2. The reason for strugglgling with some occlusions We found that MapUnveiler may be unable to unveil some regions if the regions are occluded for every frame. We visualize an example in Figure 9 of the rebuttal PDF file of the global response. MapUnveiler initially roughly predicts the boundary where we highlighted in green. However, the invisible regions are continuous, and MapUnveiler eventually predicts the region as having no boundary. As such, if the model cannot see a clear region across all frames, MapUnveiler may fail to recognize the occluded map information. ___ Q3. Randomly dropped intermediate frames A very constructive suggestion. We evaluated three models with randomly dropped intermediate frames. Frames were dropped by converting multi-camera images into black images. The experiment was conducted with drop rates of 20%, 10%, and 5%, and the results are given in Table 13 of the rebuttal PDF file of the globalresponse. MapUnveiler is affected by dropped frames, but the performance degradation is reasonable compared to MapTRv2. ___ L. Other potential limitations One can be a potential limitation that was discussed in Q2; MapUnveiler is likely to fail in unveiling fully occluded roads across all frames. Additionally, MapUnveiler requires marginal additional computational costs (15.6 FPS → 12.7 FPS) but requires approximately two times more GPU memory (830.4 MB → 1614.9 MB) and three times more parameters (76.4 MB → 213.9 MB) compared to the frame-level MapTRv2 model. We provide a comparison table in G3 of the global response. --- Rebuttal Comment 1.1: Comment: Thank you for your substantial efforts in the overall rebuttal and in response to my specific comments. The changes you have proposed and the explanations of the results help considerably. Thanks for explaining the difficulties with implementing a quantized model. I still think it would be valuable to see so it would be great if you could try to get it working. That being said, I don't think it's critical to the paper. I have read the other reviews + rebuttal, and maintain my score. However, I encourage the other reviewers to reconsider their assigned scores after the authors' rebuttal. --- Rebuttal 2: Title: Thank you! Comment: We sincerely appreciate your efforts in reviewing our responses. We are delighted to receive your positive feedback and supports for our work. Thanks,\ The Authors	Summary: This work presents a method called MapUnveiler, which aims to improve the construction of vectorized HD maps for autonomous driving. MapUnveiler uses a novel clip-level pipeline to unveil occluded map elements by relating dense image representations with efficient clip tokens and propagating inter-clip information. This approach leverages temporal information across adjacent input frames, addressing the limitations of single-frame and streaming inference methods. The model achieves state-of-the-art performance on the nuScenes and Argoverse2 benchmark datasets, demonstrating promising improvements in challenging scenarios with longer perception ranges and heavy occlusions. Strengths: 1. The introduction of a clip-level pipeline for vectorized HD map construction effectively addresses occlusion issues and leverages temporal information across multiple frames. 2. The method utilizes clip tokens to propagate map information efficiently, reducing redundant computations and enhancing prediction consistency. 3. Extensive experiments demonstrate that MapUnveiler achieves state-of-the-art performance on nuScenes and Argoverse2 benchmarks, particularly in challenging scenarios. Weaknesses: 1. The community has noticed a severe data leakage issue with utilizing nuScenes and Argoverse2 datasets for online mapping evaluation {1, 2}, as these datasets are not intentionally built for online mapping. It might also be necessary to validate the proposed method on geo-disjoint training and validation sets. 2. It would be good to see the analysis of added model compacity due to the introduction of the proposed intra-clip unveiler and inter-clip unveiler. 3. It seems the proposed intra-clip unveiler and inter-clip unveiler are adaptable to any single-frame inference online mapping methods. It would be good to validate the effectiveness of the proposed modules on other baseline methods. 4. The authors are encouraged to investigate the consistency of estimated HD maps across frames of the proposed method compared to existing methods with "inconsistent and suboptimal prediction results" (mentioned in Line 7). {1} Augmenting Lane Perception and Topology Understanding with Standard Definition Navigation Maps. {2} Localization Is All You Evaluate: Data Leakage in Online Mapping Datasets and How to Fix It. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What do the map queries stand for? Can they be transferred directly to vectorized HD maps? 2. Is the map decoder adopted from MapTRv2? 3. Are map tokens generated from the intra-clip unveiler the refined version of map queries? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitation of dependency on temporally consecutive frames is discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing thorough feedback and interesting suggestions. We are grateful for your acknowledgment that the introduction of a clip-level pipeline for vectorized HD map construction is effective and the proposed clip tokens propagate map information efficiently. Below are our responses to each comment, and we will include all the results and comments in the revised version. ___ W1. Validation on geo-disjoint dataset Thank you for your insightful feedback. To address this concern, we conducted additional experiments on a recent geo-disjoint dataset split. Due to the character limit, we attached the full table in the global response. Please find the results in G2 in the global response. ___ W2. Analysis of added model compacity That would be a good analysis to provide rich information about our model for readers. Additional to the analysis of accuracy and computational costs of the proposed intra-clip Unveiler and inter-clip Unveiler (Table 3 in the main paper), we add GPU memory consumption during inference time and the model parameters. We present the table in the global response. Please find the results in G3 in the global response. ___ W3. Validate the proposed modules on other baseline methods That's a valuable suggestion. We believe that any single-frame inference online mapping methods based on DETR (which outputs both rasterized BEV and vectorized query features) can be adapted for our MapUnveiler. Therefore, it would be exciting to additionally claim our contribution in modularization. Consequently, we attempted these experiments with DETR-based models for which the code is available online (MapVR [46], MGMap [24], MapTRv1 [21]). Unfortunately, we are unable to present the results at this time due to a lack of resources and time to set up and conduct experiments within the short rebuttal period. Nevertheless, we would like to address the concern regarding this query, so we experimented with various backbone networks (ResNet-18 and V2-99 {3}). We present the result table in G4 in the global response. As shown in the table, our method is not limited to MapTRv2 with ResNet50, but can be extended to ResNet18 and V2-99 {3} backbones. This suggests that our method can also work for other various frame-level features as well. We plan to implement our method on other single-frame inference online mapping methods and include the results in the final version. {3} An energy and GPU-computation efficient backbone network for real-time object detection, CVPRW. 2019. ___ W4. Consistency of estimated HD Map across frames Thanks for your thorough feedback. Measuring the consistency of the model would be very helpful and further elaborate the contribution of our model. To quantitatively measure the consistency of the model, we would need to indicate the track ID of each estimated map element. However, this is challenging in our model because we simplify and propose a straightforward pipeline that does not require any complicated spatial warping process across time, which is required for streaming inference (StreamMapNet [45]), nor track ID annotation. Thus, we attempt to show the consistent results qualitatively in Figures 4, 5, 6, 7, and 8. We recently found that MapTracker {4} presented a consistency-aware mAP (C-mAP) metric and measured it for baseline methods (e.g., MapTRv2) by predicting track IDs through their tracking algorithm. It seems that we can also measure the C-mAP by applying the tracking algorithm proposed in {4}. To measure it, we have to re-train the model with the refined GT data proposed in {4}, so we are unfortunately unable to show the C-mAP result currently due to a lack of resources and time. We plan to implement C-mAP and add the results in the final version. {4} MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping, arXiv:2403.15951 ___ Q1. Map queries That's correct. Map queries are trained with the same objective function as the map tokens, so it can be transferred directly to vectorized HD maps. Thus, MapUnveiler can be degraded to MapTRv2 if we transfer vectorized HD maps using map queries instead of map tokens. An interesting experiment can be conducted based on this idea; we can use MapUnveiler for both frame-level and clip-level scenarios (as sometimes we cannot utilize temporally consecutive frames due to unexpected communication/sensor errors in real-world scenarios). We experimented by replacing map tokens with map queries for every two frames, resulting in a slight performance degradation from 69.8% to 68.0% in nuScenes (60x30m). To clarify this, we will add following explanation in Line #130 (we highlighted the new descriptions in italics): > ... the map decoder, respectively. BEV features represent rasterized map features, whereas map queries embed vectorized map information; thus, we can directly construct vectorized HD maps using the map queries. ___ Q2. Map decoder That's correct. The map decoder is adopted from MapTRv2. To clarify this, we will revise the caption in Figure 2: > ... vectorized HD maps. All components in the frame-level MapNet (i.e., Backbone, PV to BEV, Map Decoder) are adopted from MapTRv2 [22]. The frame-level MapNet extracts ... ___ Q3. Map tokens That's correct. The map tokens generated from the intra-clip unveiler are the refined version of map queries. To clarify this, we will revise the paper in Line #158: > ... in the previous step. The objective of this step is to generate a refined version of frame-level map queries. As illustrated in ... --- Rebuttal Comment 1.1: Comment: Thank you for the thorough and thoughtful rebuttal. The authors have successfully addressed my concerns, so I am increasing my rating from 5 to 6. --- Reply to Comment 1.1.1: Title: Thank You! Comment: We are so pleased to have received your positive feedback and raised score (5 → 6). We deeply appreciate your efforts in reviewing our responses. Best regards,\ The Authors	Summary: This paper proposes a clip-based vectorized HD map construction paradigm for the processing of long temporal sequence, in which occluded map elements are unveiled explicitly by efficient clip tokens. Through clip token propagation, MapUnveiler achieves effective utilization of long-term temporal map information by associating inter-clip information, in which clip tokens are propagated rather than dense BEV features. Experiments demonstrate that MapUnveiler boosts the performance on public benchmark datasets, also for more challenging setting like long-range perception and heavily occluded driving scenes. Strengths: 1. This paper is well-written and easy-to-follow. Figures clearly conveys the intended message. 2. “Unveiling the hidden” and clip token propagation are reasonable and effective strategy for static Map element detection, which is practical and alleviates the problem to some extent. 3. The proposed method demonstrates strong performance on benchmark dataset, comprehensive experiments and ablation studies justify the model design. Weaknesses: 1. As mentioned at line 227, this work is built on pretrained frame-level MapTRV2 and fine-tuned, thus the comparison can be unfair. Results without pretraining are required to verify your effectiveness. 2. At line 53 and BEV Updater in line 151, for occluded features, how to select the tokens that are visible in certain frames? Seems tokens within the temporal window are fully utilized for BEV update by cross attention, how to determine whether these tokens contain unblocked information? More explanations are required. Technical Quality: 3 Clarity: 3 Questions for Authors: What is the experiment result for geometric-based dataset split as mentioned in [1] and [2]? Besides, what is the additional computing costs considering the injection of temporal clip token? [1] Yuan, Tianyuan, et al. "Streammapnet: Streaming mapping network for vectorized online hd map construction." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024. [2] Lilja, Adam, et al. "Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. The authors mentioned the weakness of their approach on the corrupted input. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for providing the insightful and constructive feedback. We appreciate your acknowledgment that the paper is easy to follow and that the proposed approach is effective. Below are our responses to each comment, and we will include all the results and comments in the revised version. ___ W1. Results without pre-training Thank you for valuable feedback. We fully understand that our training pipeline may present an unfair advantage since previous works trained their models on vectorized HD map datasets for only 24 epochs, whereas we pre-train MapTRv2 for 24 epochs and fine-tune MapUnveiler for 24 epochs, amounting to a total of 48 epochs. To address this concern, we conducted an experiment where we pre-train MapTRv2 for 12epochs/3epochs and then fine-tune MapUnveil for 12epochs/3epochs on either the nuScenes or Argoverse2 training set. Even though our proposed modules will be trained for only 12epochs/3epochs, this will undoubtedly provide fair comparisons, as we are training a total of 24epochs/6epochs. The results are given in G1 of the global response. As shown in the table, our method still demonstrates state-of-the-art performance of mAP on both nuScenes and Argoverse2 validation sets under a more fair setting. Additionally, we tried to skip the pre-training stage and train MapUnveilr from scratch, but it failed to converge, achieving an mAP of 18.2% in the nuScenes (60x30m) validation set. This indicates that our method necessitates meaningful frame-level map features to learn map unveiling. We would like to kindly note that this pre-training strategy is not own approach but is commonly adopted for training networks with temporal information. For instance, StreamMapNet [45] and SQD-MapNet [40] trained their initial 4 epochs with single-frame inputs. MapTracker [A] employed a three-stage training process (BEV Encoder → Vector Decoder → All Parameters) to facilitate initial convergence. [A] Chen, Jiacheng, et al. "Maptracker: Tracking with strided memory fusion for consistent vector hd mapping." arXiv preprint arXiv:2403.15951 (2024). ___ W2. How to select the tokens that are visible in certain frames The model automatically selects the visible and valuable BEV regions through the learning of cross attention. As depicted in Figure 1-(c) in the main paper, each clip token within the temporal window is fully utilized for BEV update. We could limit the clip tokens to select manually choosen visible BEV regions, but we thought that this might not always provide an accurate solution for constructing a clear HD map (e.g., the model might try to select an easy-to-see region even if there are no lanes). Hence, we opted to learn the selection mechanism in an end-to-end manner, which can minimize the losses of the constructed map. This approach is straightforward yet effective, as illustrated in Figure 1: compared to (a) MapTRv2 and (b) StreamMapNet, our BEV feature most clearly represents map elements. ___ Q1. Experiment result for geometric-based dataset split Thank you for your insightful suggestion. We conducted additional experiments on the geometric-based dataset splits you suggested [1]. The results have been moved to the global response due to the character limit. Please find the result in G2 in the global response. ___ Q2. Additional computing costs We provided additional computational costs in Table 3 by injecting temporal clip tokens in two aspects: Intra-clip Unveiler and Inter-clip Unveiler. To provide rich information, we further measured GPU memory consumption during inference time and model parameters, and append Table 3. We present new Table 3 in the global response. Please find the result in G3 in the global response. --- Rebuttal Comment 1.1: Title: Thanks again for your review Comment: Dear Reviewer JH5U, We greatly appreciate your valuable efforts and professional feedback, which have indeed improved the quality of the final version of our manuscript. We have provided answers to your remaining concerns above, and it would be great to hear your feedback on our rebuttal so that we can further improve the final version. Although the authors-reviewers discussion period is nearing its end, we are fully prepared to address any further questions you may have. Best regards,\ The Authors	Rebuttal 1: Rebuttal: We thank all the reviewers for their constructive and thorough comments. We are particularly excited that all reviewers acknowledged the idea of a clip-level pipeline as reasonable, novel, or effective for online vectorized HD mapping. We believe this rebuttal further enhance the paper through the valuable comments provided by the reviewers. We provided detailed point-by-point responses for all queries in each reviewer's rebuttal. Here, we provide global responses from G1 to G4. ___ G1. A more fair comparison We conducted an experiment where we pre-train MapTRv2 for 12epochs/3epochs and then fine-tune MapUnveil for 12epochs/3epochs on either the nuScenes or Argoverse2 training set. Even though our proposed modules will be trained for only 12epochs/3epochs, this will undoubtedly provide fair comparisons, as we are training a total of 24epochs/6epochs. The results are given below. \| Method (60x30m) \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (nuScenes) \| FPS \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (Argoverse2) \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| MapTRv2 [22] \| 59.8 \| 62.4 \| 62.4 \| 61.5 \| 15.6 \| 62.9 \| 72.1 \| 67.1 \| 67.4 \| \| SQD-MapNet[40] \| 63.0 \| 62.5 \| 63.3 \| 63.9 \| - \| 64.9 \| 60.2 \| 64.9 \| 63.3 \| \| MGMap [24] \| 61.8 \| 65.0 \| 67.5 \| 64.8 \| 12.3 \| - \| - \| - \| - \| \| MapQR [26] \| 63.4 \| 68.0 \| 67.7 \| 66.4 \| 14.2 \| 64.3 \| 72.3 \| 68.1 \| 68.2 \| \| HIMap [49] \| 62.6 \| 68.4 \| 69.1 \| 66.7 \| 9.7 \| 69.0 \| 69.5 \| 70.3 \| 69.6 \| \| **Map-Unveiler\ (ours)*** \| 67.6 \| 67.6 \| 68.8 \| 68.0 \| 12.7 \| 68.9 \| 73.7 \| 68.9 \| 70.5 \| \| Map-Unveiler (ours) \| 69.5 \| 69.4 \| 70.5 \| 69.8 \| 12.7 \| 69.0 \| 74.9 \| 69.1 \| 71.0 \| \| Method (100x50m) \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (nuScenes) \| FPS \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (Argoverse2) \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| MapTRv2 [22] \| 58.1 \| 61.0 \| 56.6 \| 58.6 \| 15.6 \| 66.2 \| 61.4 \| 54.1 \| 60.6 \| \| StreamMapNet [45] \| 62.9 \| 63.1 \| 55.8 \| 60.6 \| 12.5 \| - \| - \| - \| 57.7 \| \| SQD-MapNet[40] \| 67.0 \| 65.5 \| 59.5 \| 64.0 \| - \| 66.9 \| 54.9 \| 56.1 \| 59.3 \| \| **Map-Unveiler\ (ours)*** \| 68.0 \| 70.0 \| 68.2 \| 68.7 \| 12.7 \| 69.7 \| 67.1 \| 59.3 \| 65.4 \| \| Map-Unveiler (ours) \| 68.4 \| 71.2 \| 68.3 \| 69.3 \| 12.7 \| 70.4 \| 66.8 \| 59.3 \| 65.5 \| In the table, we mark with an **asterisk (\)* when we use a new fair training schedule (pre-train 12epochs/3epochs and fine-tune 12epochs/3epochs), and the other methods are directly copied from Table 1 in the main paper. Our method still demonstrates state-of-the-art performance of mAP under this setting. ___ G2. Experimental results on geo-disjoint dataset splits We conduct experiments on a recent geo-disjoint dataset split: StreamMapNet where the performance of MapTRv2 has been significantly drop from 61.5% to 36.6%. The results are given in Table 12 of the rebuttal PDF file. For a fair comparison, our method uses short training schedule as discussed in G1. We also succesfully acheieve state-of-the-art performance on geo-disjoint dataset split. We will further evaluate on various geo-disjoint dataset splits in the final version. ___ G3. Additional computing costs and analysis of added model compacity To analyze the model's capacity increased by the proposed modules, we measure GPU memory consumption during inference time and the model parameters in Table 3. The updated Table 3 is provided below. \| Method \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP \| FPS \| GPU (MB) \| Parameters (MB) \| \|---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| MapTRv2 [22] \| 58.8 \| 61.8 \| 62.8 \| 61.2 \| 15.6 \| 830.4 \| 76.4 \| \| + Intra-clip Unveiler \| 65.6 \| 67.6 \| 68.0 \| 67.1 \| 13.1 \| 1552.5 \| 144.0 \| \| + Inter-clip Unveiler \| 69.5 \| 69.4 \| 70.5 \| 69.8 \| 12.7 \| 1614.9 \| 213.9 \| As shown in the table, our methods require marginal additional computational costs (15.6 FPS → 13.1 FPS if we utilize clip tokens only within intra-clip tokens and 15.6 FPS → 12.7 FPS if we fully employ intra-clip and inter-clip tokens). However, our approach requires approximately two times more GPU memory and three times more parameters compared to the frame-level MapTRv2 model. This could be considered a potential limitation of our method, but the amounts are not large (2GB graphic cards can be purchased for under $25, and we commonly have more than 1GB of RAM and hard disks). ___ G4. Experiments with various backbones Here we additionally present experimental results with various backbones: ResNet-18 and V2-99. \| Method (60x30m) \| Backbone \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (nuScenes) \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (Argoverse2) \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| MapTRv2 [22] \| R18 \| 53.3 \| 58.5 \| 58.5 \| 56.8 \| 58.8 \| 68.5 \| 64 \| 63.8 \| \| MapTRv2 [22] \| V2-99 \| 63.6 \| 67.1 \| 69.2 \| 66.6 \| 64.5 \| 72.2 \| 70.1 \| 68.9 \| \| Map-Unveiler (ours) \| R18 \| 65.5 \| 68.4 \| 68.2 \| 67.4 \| 66.5 \| 71.1 \| 67.5 \| 68.4 \| \| Map-Unveiler (ours) \| R50 \| 69.5 \| 69.4 \| 70.5 \| 69.8 \| 69.0 \| 74.9 \| 69.1 \| 71.0 \| \| Map-Unveiler (ours) \| V2-99 \| 72.1 \| 72.9 \| 74.9 \| 73.3 \| 71.4 \| 75.1 \| 73.0 \| 73.2 \| \| Method (100x50m) \| Backbone \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (nuScenes) \| AP$_{p}$ \| AP$_{d}$ \| AP$_{b}$ \| mAP (Argoverse2) \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| MapTRv2 [22] \| R18 \| 52.7 \| 57.3 \| 51.5 \| 53.8 \| 60.3 \| 57.6 \| 49.6 \| 55.8 \| \| MapTRv2 [22] \| V2-99 \| 62.6 \| 67.8 \| 65.2 \| 65.2 \| 68.5 \| 62.1 \| 58.4 \| 63.0 \| \| Map-Unveiler (ours) \| R18 \| 67.4 \| 69.8 \| 68.8 \| 68.6 \| 68.5 \| 64.3 \| 57.7 \| 63.5 \| \| Map-Unveiler (ours) \| V2-99 \| 71.9 \| 75.0 \| 75.6 \| 74.2 \| 73.3 \| 69.1 \| 63.9 \| 68.8** \| As shown in the table, our method is not limited to MapTRv2 with ResNet50, but can be extended to ResNet18 and V2-99 backbones. This suggests that our method can also work for other various frame-level features as well. Pdf: /pdf/35bfb81bdda7a1671b26ba55db3d3d1d1d788a71.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Gated Inference Network: Inference and Learning State-Space Models	Accept (poster)	Summary: The paper presents a deep state-space model architecture with non-linear transitions and emissions. The model disentangles the latent representation for the dynamics and the one for the observed data at each time step - allowing therefore effective state estimation at future time steps and the ability to deal with missing data imputation. Inference is performed with a deep Extended Kalman Filter, that relies on a RNN architecture to make a more efficient approximate computation of the Kalman Gain (KG) and smoothing gain (SG). The method is tested on a number of simulated and realistic approaches, and it outperforms competing architectures. Strengths: 1. Non-linear/deep state-space models are being used more and more in many applications. Parameter learning and state estimation is however challenging in this setting, and this paper provides an interesting method for this 2. The method is more scalable than comptering KF-based methods thanks to the dynamics network approximation, but still effective despite the aproximation 3. The method builds on some models in the literature, but provides some useful novel components 4. The authors did extensive and well-though experiments/ablations, comparing with many SOTA models 5. There is an extensive appendix covering many details that did not fit in the main text. I particularly appreciated "A.11.1 Python intuitive code." Weaknesses: The paper is not straightforward to read (had to read it carefully twice), mostly because of the way the required derivations are presented. The notation used is somewhat not conventional within the ML-heavy NeurIPS community, and should be improved/clarified: 1. In Section 4 you use o^+ notation which is not common in the ML community. Can you clarify what it means and why you need it? This explanation needs to be done in section 4, not referring to a different section. 2. Similarly, what is s in line 219 and why do you need to introduce this notation? The sentence in line 219-221 is key but unclear 3. Why do you need to define the "gt" in line 229? 4. Not sure the SI perspective helps in the ML-heavy NeurIPS community, it brings confusion. Maybe can be added in the appendix? In terms of novelty, the final model seems to me more similar to the Kalman VAE (KVAE) model than what the authors claim. Your model can almost be seen as a modification/extension to the KVAE in which you add the RNN approximation to avoid the O(n^3) complexity, model the transition noise covariance and use a slightly different parameterization for the dynamics network. Can you clarify the differences between your model and the KVAE? In line 96 and Table 6 in the appendix, I don't think your KVAE description is correct: it has a setup very similar to your model which allows to estimate the state dynamics, and allows for direct optimization unlike what you claim. I am not as familiar with the other KF-based methods mentioned, but make sure your description is correct. Minor comments: 1. Line 53: typo "To to" 2. Line 71: typo "We" -> "we" 3. You introduce gamma in line 144, but only say what it is in line 150, making the reader wonder if it was defined above and look for it 4. Line 283: you say "with n=3m" without specifying what n and m are. Even if they are defined before, being a notation-heavy paper, better to remind the reader what n and m are. Technical Quality: 4 Clarity: 2 Questions for Authors: I am happy to increase the score as long as the comments/questions in the weaknesses section are clarified Confidence: 4 Soundness: 4 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the detailed review. We appreciate the time and energy you put on this work. We reply to your questions one by one. --- q0: In Section 4 you use $o^+$ notation... + What is s in line 219 ... a0: Thank you for your question regarding the $o^+$ and $s$ notations used in the paper. We used $o^+$ notation to represent the output of the generative process contingent on the task. Specifically, in our paper, we address two tasks: 1. Direct State Estimation: Here, the generative process generates $s$, representing the real-world estimated state (e.g., the position and velocity of the ball in the third experiment). 2. Image Imputation: In this case, the generative process constructs the high-dimensional sensory observation $o$. Therefore, $o^+$ can either be equivalent to $s$ or $o$, depending on the task at hand (lines 219-221 first draft ). The differences in tasks lead to structural variations in the decoder $d(.)$ for each task, as clarified in Table 4 of the appendix. We explicitly define these differences in the decoder architecture for generating $s$ (state estimation) and $o$ (image imputation). To address your concern about the clarity of this specific notation, we have modified the text and figures in the revised manuscript. We have omitted the $o^+$ notation and revised the figures accordingly. For instance, Figures 2 and 3 now explicitly produce both $s$ and $o$, with a description indicating that the output is contingent on the task ( line 146 revised draft). Additionally, we have moved the explanation of all notations and dimensions, including $n$ and $m$, to the beginning of Section 4 for better clarity ( lines 134-138 revised draft). Also lines 219-221 (in old draft) is revised to improve readability (now lines 215-217 in revised draft). We hope these changes address your concerns and clarify the use of notation in our paper. --- q1: Why do you need to define the "gt" in line 229? a1: State estimation requires supervised training, and we intended to indicate that ground truth states are integrated during the training phase. During training (fitting), when modeling $\mathcal{L}(s_{1:T}\|o_{1:T})$ in Equation 10, the ground truth states $s_{1:T}$ are known and used to calculate gradients (by calculating the likelihood and attempting to maximize it). However, during inference, the ground truth states $s_{1:T}$ are not known and their estimation must be sampled from Equation 10. We aimed to clarify that these two procedures—training and inference—are distinct. To avoid confusion and to align with community notation, we have decided to remove the "gt" notation from the text and Equation 10. Instead, in the rebuttal version, we emphasize the training scheme in the introduction (line 67; "the objective is maximized...") and in the experiment section, where we clarify that the estimated states are samples from modeled distributions during inference and should not be mistaken for ground truth. --- q2: Not sure the SI perspective helps... a2: We appreciate your suggestion and agree that it may not be as relevant to the ML-heavy NeurIPS community. To address this, we moved the discussion of System Identification and its various versions, along with a detailed explanation of the EM algorithm used in these approaches, to the appendix in the revised draft. --- q3: In terms of novelty, the final model seems to me... a3: KVAE is an EM-based variational model designed to maximize the observation (or evidence) likelihood. It focuses on inferring high-dimensional observations for frame generation (pixel space). The maximization of the observation likelihood in KVAE (Equation 7 in their paper) is conceptually similar to our goal of image imputation, where we aim to maximize $\mathcal{L}(o_{1:T})$ in Equation 11. This similarity holds despite differences in parameterization, scalability, and distribution approximations. Although KVAE allows for sampling from posterior states (Equation 11 in their paper), it lacks a framework for direct state optimization, such as providing and maximizing the log likelihood for states like ball position-velocity in the polygon experiment. In contrast, our model addresses this by explicitly modeling $\mathcal{L}(s_{1:T}\|o_{1:T}) $. Our model parameterizes $p(s_{1:T}\|x_{1:T}, w_{1:T}, o_{1:T})$ with approximated distributions that enable direct state optimization. This conditional likelihood modeling is conceptually similar to the objectives defined in works of [Shmakov et al. (2023)](https://openreview.net/forum?id=v7WWesSiOu&referrer=%5Bthe%20profile%20of%20Alexander%20Shmakov%5D(%2Fprofile%3Fid%3D~Alexander_Shmakov1)), [Zhang et al. (2023)](https://openreview.net/forum?id=o16sYKHk3S&referrer=%5Bthe%20profile%20of%20Akash%20Srivastava%5D(%2Fprofile%3Fid%3D~Akash_Srivastava1)), and [Sohn et al. (2015)](https://papers.nips.cc/paper_files/paper/2015/hash/8d55a249e6baa5c06772297520da2051-Abstract.html), where conditional likelihoods are modeled, albeit with different parameterizations and applications. This dual functionality distinguishes our model from KVAE by allowing for direct optimization of state dynamics. To address your concern, we decided to modify the word "estimate" to "optimize" in line 96, as it may imply that drawing samples of the state (considered as estimated states during inference) is not feasible in KVAE. Accordingly, we have updated the state estimation column of Table 6 in the appendix for KVAE, EKVAE and MVAE to prevent confusion for readers. We hope this clarifies the distinctions between our model and KVAE and addresses the concerns raised. --- q4: Minor comments.. a4: We greatly appreciate your careful review and attention to detail. We fixed typos, added gamma description before using it, and added definition of $n$ and $m$ in experiment section again (line 279 revised draft) for better readability. --- Rebuttal Comment 1.1: Comment: Thanks for the clarifications and the improved revised paper. As a result I will raise my score to an accept. --- Reply to Comment 1.1.1: Title: Reply to reviewer 254u comment Comment: Thank you for raising your score and your positive feedback. we greatly appreciate your support.	Summary: The paper introduces a very well theoretically motivated State-Space Model learning approach, which is implemented by a gated inference network. The network implements a Hammerstein-Wiener model within a modularized deep learning architecture. It uses GRU cells to mimic Kalman Filtering operations. Forward as well as forward-inverse processing routines optimize the hidden state space estimations. Several theoretical components add to the paper contribution. Evaluations show superior performance on several challenging toy problems with noisy data (pendulum, double pendulum, ball bouncing in irregular polygon, as well as odometry prediction from kiti data) evaluating state estimation and imputation tasks. Strengths: Paper is very well-structured. The work is also very well-motivated and well-embedded into the literature. The theoretical motivation and system derivations are impressive and usefully embed the author’s GIN system into the Kalman Filtering background. Approximating everything in a variational inference manner via estimations of Gaussians and their Covariances is efficient. Theorems 3 and 4 offer a theoretical derivation for ensuring stability of the unfolding recurrence. The evaluations contain sufficiently challenging problems. Performance is compared with many alternatives, showing superior performance nearly throughout. Only in Table 3 GIN was partially beaten by DeepVO. Weaknesses: The theorems 3 and 4 are not really experimentally evaluated. Is instability observed when the recurrent matrix is not modified as proposed? The theorem’s proposal should be verified experimentally. Even more elaborate evaluations would of course be great. Seeing the great content and the importance of the theoretical derivations, though, I consider this a very minor point, which can be tackled in subsequent work. Technical Quality: 4 Clarity: 3 Questions for Authors: Figure 1 trajectories are not quite as smooth as I had expected. Any reason for this? The task, particularly between bounces, to predict a linear trajectory should be very simple. I do not really see how Theorems 1 and 2 are actual theorem. Don’t they just define the log likelihoods of the ground truth states / observations? Theorem 2 – is the second summand starting with factor $(1-o_t^{(k)})$ necessary? Line 260: should read "in theorem 3", no? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: Further evaluations and ablations to truly identify the core components that yield the great results of the GIN system. Confirm theory components experimentally. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable feedback and time. We have answered your questions one by one and revised the draft to address your concerns. --- q0: The theorems 3 and 4 are not really ... a0: We appreciate your concern regarding the experimental evaluation of Theorems 3 and 4. Given the page limit of the conference, we could only present a portion of our experimental results. However, we conducted multiple experiments to demonstrate the contributions of these theorems. Specifically, addressing your concern, we tested each experiment multiple times with various settings, including the approaches described in Theorems 3 and 4 (SVD and Gershgorin Circle Theorem) and gradient clipping. We add these details in the revised draft (line 274). We also attached this table in the PDF response file. In summary, table 1 in response PDF listing the log likelihood and its standard deviation for three experiments: single pendulum, double pendulum, and irregular polygon. These experiments were trained under different settings for handling gradient explosion: the conventional Gradient Clipping (GC), our first solution using Singular Value Decomposition (SVD), and our second solution using the Gershgorin Circle Theorem (GCT). In this table, $\theta$ represents the threshold for gradient clipping, and $\delta$ is the threshold in our method that keeps the spectral radius less than 2, i.e., $\sigma_1(\mathbf{U}_h) + \delta = 2$. We also included the success rate to indicate that not only high performance but also consistent stability across multiple experiments is important. We hope this clarifies how Theorems 3 and 4 have been empirically validated and show the effectiveness of our proposed theorems. --- q1: Figure 1 trajectories are not quite as smooth ... a1: Thank you for your observation regarding the trajectories in Figure 1. There are two primary reasons for the observed sharpness in the movements: 1. Initial Training Focus: In the initial epochs, we focus the system on learning the auto-encoder and globally learned parameters, such as $F$ and $H$ , while excluding the Dynamics Network parameters $\alpha(x)$. This approach allows the system to initially acquire robust embedding and meaningful latent state vectors (essentially better embedding of the objects' locations and velocities within their environment) before involving the Dynamics Network. By doing this, we enable the Dynamics Network to be trained later with rich latent representations, resulting in more confident weight selection. 2. Mode Collapse Loss Term: We incorporate a mode collapse loss term that increases the distance between each of the $K$ transition matrices. In terms of the Frobenius norm, these matrices can be considered as being maximally separated. This separation leads to more distinct movement patterns. Subsequently, all parameters, including those of the Dynamics Network, are jointly learned. This approach ensures that the Dynamics Network assigns weights to the relatively separated $F$ matrices, each modeling an entirely different movement pattern. After a considerable number of epochs, the Dynamics Network selects the weights of the relevant $F$ matrices with high confidence, almost certainly (weights are around one for the target $F$ and around zero for the non-relevant $F$ matrices relying on the numerical results). Specifically, the issue of soft reflections, and in more severe cases, movement along the edges of the walls after bouncing, was observed in our experiments prior to implementing the mode collapse loss term. The addition of this term has significantly improved the movement patterns, making them more distinct and reducing such anomalies. We hope this addresses your concern and provides clarity on the observed trajectory behavior. --- q2: I do not really see how Theorems 1 and 2 are actual... a2: Your observation about the log likelihoods of the states and observations is correct to an extent. These theorems indeed involve the log likelihoods, but it is important to note that they represent approximated log likelihoods rather than exact ones. Since our filtering and smoothing distributions are approximations of their exact counterparts, we derive a lower bound of the log likelihood rather than the exact log likelihood. The theorems detail the derivation process of this lower bound, illustrating where the inequalities and approximations arise. This comprehensive process includes the final descriptions for state inference (modeled as Gaussian distributions) and image inference (modeled as Bernoulli distributions). The inclusion of these derivations as theorems, complete with proofs, is intended to provide a guarantee for our approximations proximity and to ensure the completeness of our method. --- q3: Theorem 2 – is the second summand... a3: The log likelihood in this equation arises from modeling the pixels of the original observations using a Bernoulli distribution. Specifically, the likelihood term is given by $\prod d_k^{o_k} (1-d_k)^{1 - o_k}$. When considering all pixels of an observation with $D$ dimensions, across $T$ time steps, and over $N$ Monte Carlo samples, the complete likelihood expression accounts for the contributions from all these dimensions. Taking the logarithm of this likelihood function results in the equation with the second term in summand you are referring to. This summand captures the likelihood of the observations under the Bernoulli distribution, ensuring the proper representation of the data's probabilistic structure. --- q4: Line 260: should read "in theorem 3", no? a4: You are correct; the reference should indeed read "in Theorem 3." We appreciate your attention. We made the necessary correction in the revised version of the manuscript. --- Rebuttal Comment 1.1: Comment: thank you for the clarifications and extra work. Great paper. I raise the score by one more point. --- Reply to Comment 1.1.1: Title: Reply to reviewer Db5x Comment: We thank the reviewer for their kind comments to raise the score and support of our paper.	Summary: This paper advances temporal reasoning in dynamic, high-dimensional, noisy environments by introducing a novel architecture for latent variable state space models. The architecture permits efficient Bayesian inference with nonlinear transitions and emissions. Experiments are performed on toy datasets and a simple real-world dataset for state estimation and missing data imputation, showing that it beats benchmarks relative to competing models like RNNs, autoregressive models, and latent variable approaches. Strengths: Clear exposition of model architecture and inference algorithm. Theoretical analysis in Section 6. Weaknesses: I think one thing that could really strengthen this paper is showing an experiment on a more challenging data set / problem. The first two experiments are on toy problems. I think another thing is to explain more clearly how this architecture is differentiated from others, i.e. the technical novelty. E.g. what is the relation of your model to other SSMs incorporating RNNs like the Variational RNN (which you benchmark against in the experiments), and what is it about that change that improves inference? Technical Quality: 3 Clarity: 3 Questions for Authors: Are there any experiments you could do that show a sequential modeling problem in which inference was previously intractable is now so? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, limitations adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful feedback and the time you dedicated. We have responded to your questions individually and revised the draft to address your concerns. --- q0: I think one thing that could really strengthen... a0: We performed our evaluations and experiments in accordance with popular benchmarks and trends in the related literature, allowing us to compare our results with existing works. We created a table of the most relevant related works, including their experiment selections and the datasets they used. Additionally, we provided another table detailing the statistics of these commonly used datasets (please see them in .pdf file). In particular, related works typically follow a sequence of categories for their experiments: 1. Simple Low-Dimensional Experiments: These involve relatively low observation dimension, with dynamics that can be either linear, switching-linear or nonlinear. (Datasets in first part of table 2 in .pdf) 2. Toy/Synthetic Experiments: These involve higher-dimensional data and potential challenges, such as noise distortion or missing frames. (Datasets in second part of table 2 in .pdf) 3. Real-World Datasets: These can be either lower-dimensional data with easier state estimation tasks or higher-dimensional data with complex visual observations. (Datasets in last part of table 2 in .pdf) In our paper, we deliberately skipped the first category due to its simplicity. Instead, we chose three synthetic datasets from the second category (single pendulum, double pendulum, and bouncing ball) and one additional dataset from the third category (KITTI visual odometry). The KITTI dataset, with $1242\times 375\times3$ dimension of image observations, is, to the best of our knowledge, one of the most challenging problems to infer in this literature that most of related works skip it (visualization in Fig 1 of PDF response). Please note that, in the revised version (line 274), we included additional experiment to highlight the efficiency of theorems, introduced in the paper (also added in table 1 in PDF response). --- q1: I think another thing is to explain more... a1: Our structural novelty can be divided into three main parts: 1. New Inference Algorithm: We propose an inference algorithm that models the transition and emission distributions. 2. Parameterization and Approximations: We introduce parameterizations and approximations for the parameters and distributions. 3. Stability Scheme: We provide a stability scheme for the newly parameterized algorithm. To provide more specific information and test the improvement corresponding to each part, we compared our model with different approaches through extensive ablation studies. 1. First, we used a simple encoder-decoder without latent parameterization to demonstrate the effectiveness of parameterizing the latent state (lines 292-293 first draft, 288-289 revised draft). 2. Second, we evaluated the effectiveness of the smoothing-filtering parameterization designed in our transition block by removing it and replacing it with benchmark RNNs such as AE-RNN. This demonstrated that the smoothing-filtering improves inference compared to benchmarks lacking this parameterization. (lines 299-302 first draft, 295-298 revised draft). 3. Third, to validate the accuracy of our distribution approximations, we compared them with state-of-the-art methods. These comparisons highlight how our approach differs from other SSMs incorporating RNNs, such as the Variational RNN, and demonstrate the improvements in inference achieved through our proposed method. ( lines 303-308 first draft, 299-304 revised draft) 4. Finally, in the newly added experiment in the revised draft, we aim to demonstrate that the proposed stability scheme outperforms the traditional gradient clipping approach in preventing gradient explosion, which threatens the stability of the GRU cells used for parameterization. --- q2: Are there any experiments you could... a2: In sequential modeling, especially with high-dimensional sensory observations, it is intractable to obtain the exact evidence likelihood, such as $p(o)$.This is because calculating $\frac{p(o\|x).p(x)}{p(x\|o)}$, is challenging due to the difficulty in obtaining the distribution for ${p(x\|o)}$. Common approaches involve approximating this term, as is done in related works and in our model. For simple systems with linear dynamics, such as car movement modeled by the celebrated Kalman Filter (KF), Linear Gaussian State Space Models (LGSSMs) can be used to derive exact filtering and smoothing distributions (tractable posterior). However, we have chosen to focus on more complex systems, as the optimal solution for linear systems (i.e., the Kalman Filter) already exists and is well-established. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: I thank the authors for addressing my concerns, especially more explanation on the technical novelty, and am glad to increase my score accordingly. (By tractability, I didn't mean closed-form solutions/exact inference is possible but rather training for approximate inference converges on a task/dataset where previous methods diverge - that's my bad for not being precise.) --- Reply to Comment 1.1.1: Title: Reply to the question of the reviewer g5tG Comment: We greatly appreciate your positive feedback and the increase in score. Generally, optimizing recurrent networks is challenging due to the highly ill-conditioned nature of their loss surfaces, which often contain many local minima and walls (as illustrated in Figure 6 of the study by [Razvan Pascanu et al.](http://proceedings.mlr.press/v28/pascanu13.pdf) ). This difficulty is stronger in tasks with switching dynamical behaviors or long time dependencies (as noted by [Kanuparthi et al. in their work](https://arxiv.org/pdf/1810.03023) ). In such cases, if gradients are too large, they can push the loss trajectory from one local minimum to another, often leading to divergence. Conversely, if the gradients are too small, the loss may stuck in a local minimum. This issue is common across almost all SSMs that integrate RNNs, whether they are variational or not. The conventional approach to managing large gradients in RNNs is gradient clipping. While this prevents distorted loss trajectories, it often results in the model getting stuck in a local minimum (as shown in Table 1 in .pdf response, when the gradient clipping threshold is set to 5). On the other hand, allowing gradients too much freedom can lead to model divergence, as observed when the clipping threshold is increased to 15-20 (Table 1 in .pdf response). Our proposed stability scheme, however, defines a necessary condition for stability that prevents gradient explosion without overly constraining the gradients, thereby avoiding the issue of getting stuck in local minima. This effect is evident in the standard deviation of the results for SVD and GCT (Table 1 in .pdf response). In summary, relying on the numerical results of our experiments in tables 1,2 and 3, most SSMs that incorporate RNNs (e.g. AE-RNNs (LSTM and GRU), KVAE, EKVAE ) struggle to provide stable training on the datasets used in our study (single and double pendulum, bouncing ball and KITTI visual odometry), often relying on gradient clipping that may trap the model in local minima or diverge otherwise. --- Rebuttal 2: Title: Last day for discussion Comment: Reviewer g5tG, today is the last day for discussion. Your score deviates from the other reviewers', and your review was relatively short. I hope you will take a moment to respond to the authors' rebuttal today.	null	null	Rebuttal 1: Rebuttal: We appreciate the reviewers for their detailed comments and questions. Our rebuttal response is mainly organized into three sections. 1. To address the concern of one of the reviewers regarding the sufficiency and complexity of our experiments, we created a table listing the most relevant studies, detailing their experimental setups and the datasets they utilized. Additionally, we provided a separate table outlining the statistics of these commonly used datasets. We compared our experimental choices with those of the listed studies, evaluating both the complexity (e.g. observation dimension/type, sequence length) and the number of experiments conducted. These tables are in the PDF response file. 2. We included the results from an additional experiment to show the effectiveness of the theorems presented in the paper. In brief, we compared the proposed theorem with the traditional gradient clipping method for managing gradient explosion. We reported the success rates of the experiments to reflect the stability across various seeds. These results are available in a one-page PDF included in the response (Table 1). We also add this to the revised paper. 3. We have revised the notation and figures for improved clarity and removed ambiguous terms. Specifically, to address the reviewer’s concerns, we eliminated the use of the less common $o^+$ notation. Additionally, all notations have been moved at the beginning of Section 4. We also refined Comparison Table 6 and clarified the shortcomings of related works (particularly KVAE) to prevent any confusion. Figure including revised variables (omitted $o_{1:T}^+$) for state estimation of visual odometry task is in one-page PDF response. Detailed answers with further elaborations on the reviewers' questions are provided in separate rebuttals for each reviewer. The [revised paper](https://raw.githubusercontent.com/AnonymousEzhdeha/GIN2024/main/Neurips2024_Rebuttal.pdf) is available after applying the reviewers comments. Pdf: /pdf/136294a7bb0f36b2857e914276addcbcdba91efe.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Guiding Neural Collapse: Optimising Towards the Nearest Simplex Equiangular Tight Frame	Accept (poster)	Summary: The paper uses Riemannian optimization to guide the final layer weights (the linear classifier) toward the nearest simplex ETF orientation. In particular, consider the two common approaches of training a deep classifier network: 1. The standard training strategy where the final layer weights are updated by backpropagation. 2. The final layer weights are fixed as a simplex ETF (which has been well-studied in previous works). The proposed approach leverages the duality between penultimate layer features and the final layer weights (to form a simplex ETF orientation) and gradually guides the latter to an optimal simplex ETF per training step. Strengths: 1. The proposed approach frames the gradual transition of weights to a simplex ETF as a Riemannian optimization problem, which can be differentiated. Thus, allowing for an end-end training pipeline. The combination of these techniques is novel to the neural collapse setting. 2. The experimental results are presented for the simple UFMs as well as practical networks and datasets to showcase the convergence benefits. Weaknesses: The authors do not provide numerical data for the extra memory and step-time that is required by the extra deep declarative layer. A brief discussion is presented in Section 5 but I believe further details would strengthen the paper. For instance: - By what percentage does the step time and memory increase when adding this layer? - When should one avoid the backward pass through this layer and consider only the forward pass? - What is the dependence on the memory and step time growth with the feature dimension and the number of classes? Maybe a simple UFM-based analysis should suffice. see more questions below. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. How effective is the proposed approach in settings with imbalanced classes [1] ? More generally, for settings where the simplex ETF might not be an ideal configuration (for instance: graph neural networks, see [2] ). A brief discussion on these topics can further strengthen the paper. 2. Instead of running the optimization layer to select the final layer weights at every step, what if we do it after every $k$ step? Can we potentially reduce the majority of the computational overheads while improving the convergence? 3. What is the convergence behavior when employing SGD/Adam instead of the AGD approach? nit: Where is $U_{init}$ defined? nit: line 117, below eq (5), is the formulation of $\widetilde{H}$ correct? shouldn't the denominator be $\|\|\overline{H}\|\|_F$ ? References [1] Fang, C., He, H., Long, Q., & Su, W. J. (2021). Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences [2] Kothapalli, Vignesh, Tom Tirer, and Joan Bruna. "A neural collapse perspective on feature evolution in graph neural networks." Advances in Neural Information Processing Systems 36 (2024). Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: Regarding the computational and memory costs of our approach, please refer to our general response. Avoid backward pass: Theory suggests that incorporating the DDN layer's backward pass should provide additional gradient information for updating the features’ parameters in the backbone neural network. While we have observed that this additional gradient information improves stability, it has not been critical for the algorithm's convergence. Therefore, omitting the backward pass can be a practical choice in situations where computational resources are a concern. Specifically, in our experiments with ImageNet, the current implementation of the DDN backward pass exceeds GPU memory limits. As a result, we opted not to use it for these experiments. However, we acknowledge the importance of the DDN backward pass for gradients to be technically correct, and plan to address these computational and memory challenges in future work to ensure that it can be effectively utilised even for large-scale datasets such as ImageNet. Questions: 1. Minority collapse refers to the training dynamics observed in standard approaches when dealing with significant class imbalance. In such cases, minority classes tend to collapse into a single vertex of a simplex ETF, resulting in a degenerate solution. However, under certain conditions (e.g., when the number of features $d$ exceeds the number of classes $C$), a simplex ETF can still represent an optimal classifier configuration, even in imbalanced settings (see Theorem 1 of Yang et al. [67]). The primary challenge is that a learned classifier may struggle to achieve this configuration easily. It’s important to note that the assumption for neural collapse to occur is zero misclassification error during the TPT, which is clearly not met in the case of minority collapse. Substantial research on imbalance training and long-tail classification tasks employs fixed ETF approaches to address problems associated with conventional training methods. We believe that our approach, which induces the nearest ETF direction rather than an arbitrary one, will yield even better and more stable results compared to fixed ETF methods. We have included preliminary results (found in the attached pdf) on CIFAR100LT with an imbalanced ratio of 0.01 (following Yang et al. [65]) using a VGG network to support our claims. We have not considered graph neural networks in this work. Our method applies to the structure of the classifier solution as defined by the loss function. Exploration of alternative architectures and loss functions is left for future work. 2. Running the Riemannian optimisation problem every $k$ steps is indeed a viable extension of our method and can potentially reduce computational demands. We conducted experiments where $k$ was set to the number of steps per epoch. Our findings indicate that optimising once per epoch is insufficient for guiding the network to the nearest ETF and achieving rapid convergence, as the features can change significantly within an epoch. Thus, we opted to perform the optimisation at every step. Hence, we present a trade-off between achieving faster and more stable convergence versus higher computational cost. Introducing $k>1$ as an additional hyperparameter requires careful tuning for each specific case. Another advantage of our approach is that, as training progresses, the target ETF stabilises because the features approach convergence. At this point, the Riemannian optimisation problem yields similar solutions. To eliminate the additional computational cost, we can fix the simplex ETF determined by the Riemannian optimisation and continue to converge towards this fixed ETF. However, determining the exact epoch at which this stability is achieved remains heuristic and lacks theoretical guarantees, thus requiring case-by-case consideration. 3. Our approach is particularly well-suited for optimisers that dynamically adjust step sizes based on the current model state, such as AGD. In contrast, using methods like SGD or Adam with our approach requires careful tuning of the learning rate. This may involve developing a new type of scheduling that considers the objective value of the Riemannian optimisation problem to ensure that the learning rate is appropriately decayed as the solution approaches the ETF. Current learning rate schedulers are typically derived heuristically and optimised for conventional training scenarios, which may not account for the specific needs of our approach. Therefore, adapting these schedulers for our approach requires additional analysis, which is beyond the scope of this paper and will be addressed in future work. Nit: $U_{init}$ is defined in the paragraph for Hyperparameter Selection and Riemannian Initialisation Schemes in lines 230-235 in our paper. Nit: Thank you. This is a typo. The Frobenius norm should be taken wrt the feature means and not the features, so it should be $\|\|\bar{H}\|\|_F$. --- Rebuttal Comment 1.1: Comment: Thank you for answering the questions. Based on the memory and compute overheads presented in Table 1, seems like the implicit ETF approach is quite slow in terms of step times. For instance, for CIFAR100 / Res50 the implicit ETF approach (fwd+bwd w/o DDN) seems to be ~4-5x slower than the standard fwd + bwd. This factor is much larger for bwd + w/ DDN. Suggestion: Taking a step back, since we care about the terminal phase of training for NC analysis, a much better comparison would be to consider `num_steps * step_time` for each of the approaches (standard/implicit ETF/explicit ETF). This way, one can know how fast they can converge to TPT within a given time budget. Please clarify these aspects in your claims and consider incorporating a discussion of such training efficiency aspects. Good luck. --- Reply to Comment 1.1.1: Comment: Thank you for your suggestions. We will expand our discussion to include these concepts and experiments in our paper. In particular, we will provide a more nuanced analysis than the averages reported in Table 1. As training progresses, the Riemannian optimisation (to obtain the implicit ETF in the forward pass) converges more quickly. Indeed, the majority of the cost comes from the initial few iterations. To provide a more comprehensive view, we present additional time measurements from a new run of CIFAR-100 using ResNet-50: - Average forward time: 74 ms - Median forward time: 17.6 ms - Maximum forward time: 825 ms - Minimum forward time: 14 ms These results indicate that the median forward time is competitive with other methods. The significant variance in forward time (from a minimum of 14 ms to a maximum of 825 ms) highlights the variable nature of the forward pass. However, the median value suggests that the majority of forward passes are performed efficiently, reinforcing the overall competitiveness of our approach. This trend is consistent across other datasets and architectures as well. In practice, one way to mitigate the initial overhead is to use a warm-up phase. During this phase, we can apply the standard method for a few iterations to update the features before transitioning to our approach and Riemannian optimisation. However, such practical considerations complicate the presentation of this paper and will be explored in future work.	Summary: This paper proposed a novel algorithm for neural network training. The algorithm is motivated by the recent discovery on the neural collapse phenomenon, which demonstrates that the last layer of neural network classifier will converge to a specific structure named simplex ETF. The authors propose to guide the network parameters to the ETF structure via explicitly penalizing on the distance to the ETF, and further address the non-uniqueness of the solution via adding a proximal term. Experimental results on various neural network architecture and real world datasets are presented, and the proposed algorithm can universally improve the training and testing accuracy over the standard training. Strengths: The proposed algorithm is novel and well motivated, and it shows universal and significant improvement over multiple choices of network architecture and datasets. The contribution of this work is solid, it helps the community to understand the benefit of the neural collapse phenomenon, and can potentially improve the standard paradigm of neural network training. Weaknesses: 1. The presentation should be improved, see questions for detail. In general the authors should give more detailed information about how the algorithm is implemented. 2. Although the accuracy on train and test dataset exhibits significant improvement within the fixed number of training epochs, the proposed algorithm are much more complicated to compute. Therefore it makes more sense to compare the running time and computational cost with Standard and Fixed ETF. 3. Proper ablation study is missing. The authors add many additional techniques, such as exponential moving average, stratified batch sampling, deep declarative layer to improve the training. It is not clear how much the improvement indeed comes from the nearest ETF optimization. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. How is equation 8 being optimized? Is it using Lagrangian multiplier method? A pseudo code of the proposed algorithm will be very helpful. 2. I found the Proposition 1 hard to follow. The authors should explain what is the implication of this proposition and how does it helps to the stability of training. The current statement is confusing to distinguish the main result, and notations such as $D_y$ and $\Lambda$ are not properly introduced. This proposition should be improved. 3. In Table 1 and 2, fixed ETF on CIFAR10 with VGG seems to have much worse performance than others. Do you have insights about what is going on here? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have discussed the limitation properly in the paper. A more detailed discussion with empirical results on the computational cost will be helpful. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weakness 1 with question 1 & 2: Eq. 8 is optimised as a bi-level optimisation problem. At each gradient update step, we first solve the inner optimisation problem to obtain the nearest ETF solution from the Riemannian problem. This gives the classifier weights directly. Subsequently, we perform the outer optimisation by optimising the rest of the network using stochastic gradient descent. Proposition 1 restates a known result for gradient computation in differentiable optimisation problems, utilising the implicit function theorem for implicit differentiation (Gould et al. [21], Figure 2 and Section 4.1 provide the underlying intuition). As illustrated in Figure 1 of our paper, two streams of information contribute to the final loss function. Proposition 1 enables us to backpropagate through both streams of information, theoretically enhancing the stability of our solution. The key insight is that incorporating the second stream of gradient information from the Riemannian optimisation allows us to account for changes in the features $H$ and adjust the network’s parameters accordingly. In practice, we have observed that the gradient magnitude of the second stream (bottom in Figure 1) is relatively small compared to the first stream (top in Figure 1). Coupled with the additional computational cost of computing the DDN gradients, this contribution is somewhat restricted in this case. Weakness 2: Please refer to our general response for computational cost discussions. Weakness 3: Ablation studies: Our goal is to find the nearest simplex ETF of the feature means while considering all the features from the network. In the UFM case, where full-batch gradient updates are performed, all features are available, so additional techniques such as exponential moving averages are not necessary. However, in real-world scenarios with stochastic updates, optimising towards the nearest ETF for the features in a given batch can lead to highly variable results, as we do not have the full picture, and the feature means are constantly changing. By incorporating the described techniques, we ensure that we are moving towards a stable ETF target, accounting only for changes in feature weights after each gradient step. Question 3: Fixed ETF results: During VGG training with the fixed ETF case, we observed significant variability in the solutions, with CIFAR-10 showing the largest variability. This variation highlights the impact that fixing to a predefined ETF direction can have on the quality of the solution space. While ResNets, with their residual connections, can somewhat mitigate this issue, VGG-type networks appear to struggle more with fixed ETF directions. The intuition is that depending on the initialisation seed, we may start closer to or further from the chosen ETF solution, leading to results where fixed ETF performs either comparably or inferiorly. In contrast to fixed ETFs, the standard approach can sometimes be more robust and perform better. However, this comes with increased memory and computational costs due to the need to learn the classifier. This is because, with the standard method, the classifier has the flexibility to adapt and better match the features, eventually reaching a neural collapse solution in practice. Our approach, on the other hand, is robust to initialisation seeds by moving towards the closest ETF solution from any starting point and achieves such convergence more quickly. --- Rebuttal Comment 1.1: Comment: Thanks for the response. I have no further concerns and would like to keep my score at this moment. I encourage the authors to explore further how to improve computational efficiency, and I believe it will lead to a solid contribution to the community.	Summary: One of the key aspects of neural collapse (NC) is that the penultimate class feature means form a simplex Equiangular Tight Frame (ETF). The main idea of this paper is to leverage this insight and improve training by further encouraging this property during training. The authors suggest doing this by solving a Riemannian optimization at a given iteration. The way it works is that the classifier weights are set to the nearest simplex ETF obtained by solving this inner Riemannian optimization problem. The classifier weights are dynamically updated during training using this Riemannian opitmization problem at each iteration (rather than trained using gradient descent) using a "deep declarative node" this allows gradients to propagate through the Riemannian optimization. They show that this approach indeed speeds up training convergence and improves training stability. Their experiments include both synthetic and real-world classification tasks and architectures. Overall the authors present a nice idea and it is a well-written paper. However, there are a few issues related to the experiments that I outline below. From my viewpoint, the value of this paper and their method (to me) is less the improved test accuracy and more the improved stability and speed of convergence. It's important to note that this speed up also comes at an additional cost (i.e. in performing the Riemannian optimization). Therefore, the improvements to stability or speed of convergences should be weighted against this caveat. I think it would help to highlight this tradeoff more upfront and make that more clear/transparent. Strengths: This is a thoughtful and well-written paper. The authors suggest a nice idea to leverage this insight of NC in deep learning and their approach has clear benefits. It is a nice idea and very well executed. There are clear improvements to the current methods; e.g., their improvement upon [74] by solving the inner Riemannian optimization instead of requiring the model backbone to do the work of matching to a chosen fixed ETF. The theory and the idea is very compelling. The implementation is good and well explained. Beyond the theory and the novelty of the idea, the main strength of the paper is the value added wrt convergence speed in terms of the number of epochs required for the network to converge. Good work. Weaknesses: The main points of concern for me are in regards to the experiments and how the results are reported in the paper. Table 2 looks good but is a bit misleading particularly when comparing the ranges of the test top-1 accuracy. The results are still interesting but it's not such a strong/clear winner; that is, when looking at the ranges, it's not so obvious. The authors point this out and clarify that the advantages are speed to convergence and decreased variability which I agree are definite plusses. The test top-1 accuracies reported in Table 2 aren't competitive with what can be obtained on these benchmark datasets, particularly for the Resnet models. For example, looking at 200 epochs or training, STL on ResNet50 should be able to achieve 85-90% test accuracy, even for Resnet18 the test top-1 accuracy for STL should be upwards of 75%. Similarly, for CIFAR100 on Resnet50, the test accuracies aren't competitive. It'd be interesting to see if these claims about variability still hold when giving the baselines adequate chance to be competitive. For Figure 4, also no error bars. Understanding compute restraints, it would be nice to see similar multiple seed runs for ImageNet experiments. Finally, one thing that is not reported here is an estimate of compute cost. Their method requires additional compute for each iteration. Perhaps when compared on this axis their implicit ETF and the Standard training method would be more fairly compared. The authors do mention this in the limitations section. Technical Quality: 3 Clarity: 3 Questions for Authors: How do you know that the ETF that you steer towards via this Riemannian optimization process is better than the one that you would have arrived at naturally? You say "this process [provides] the algorithm with a starting point that is closer to an optimal solution rather than requiring it to learn a simplex ETF or converge towards an arbitrary one". How do you know that this is optimal? Optimal in what context? If I understand correctly, it's just the solution of the Riemannian optimization which means it forces the class means into an ETF. It's optimal wrt to the optimizaiton problem but not necessarily for the learning task? Is that correct? Do you do any, or is it possible to perform a comparison of these two resulting ETFs? How does the test accuracy of your 'encouraged' ETF compare to the one you would have obtained naturally? In Section 3.3. The Proximal Problem. I just don't see immediately why adding the proximal term guarantees uniqueness of the solution and how it stabilizes the Riemannian optimization problem. Can you add more detail or proof or reference to proof? On first reading, it was unclear to me exactly how is U_prox defined? And what is used in practice. Is it determined from the previous iteration? If I understand correctly, you tried two approaches: setting U_init = U_prox = canonical ETF. Or to set both equal to random orthogonal matrices from classical compact groups. It sounds like, in the end, you run training without the proximal matrix for one epoch. Then use the resulting U* to set U_init = U_prox = U* from that one epoch. Is that correct? How was this "warmup" approach validated? Did you experiment with various epochs? How stable were the outcomes of that analysis? You later mention (line 225) that the correct choice of these values is "crucial" so it seems important to understand. In the Section Hyperparameter Selection and Riemannian Initialization Schemes: You mention that algorithm convergence is robust to values of \delta and that the \delta reg term is a trade-off between the optimal solution's proximity to the feature means and its proximity to the given simplex ETF direction. Did you explore how and when to introduce this constraint? Or any exploration of how the solution varies with \delta? In section 3.4 General learning Setting: The role of the temperature \tau is bit unclear to me. And the reference to [67, Theorem 1] isn't very helpful. Perhaps a little more clarity as to the role \tau plays here? You state later in the Experiments section that you use \tau=5 according to Yaras et al [67]. This hyperparam choice is not very clear to me. (typo? clarification?) Proposition 1. There is notation discrepancy between what is stated in the Proposition and what is derived in the Appendix B. Namely, the Proposition is stated wrt \bar{H} but the derivation is carried out for \tilde{H}. I understand that \tilde{H} is the normalized (wrt Frobenius) matrix \bar{H} so perhaps it all works out with the normalization constant but the discrepancy there and comparing back with dimensionality of matrices in the original statement of Proposition 4.5 in Gould et al. [21] (from which this result follows) had me a bit confused. Are there error bars in Figure 2? I see them for plot (f) but not for the others? (clarification) What is depicted in Figure 2(c)? What is \mathcal{P}_{CM}? I think I somehow missed that. Are there error bars in Figure 3? Were multiple trials run for these experiments? Tables 1 and 2: The ranges for train and test top-1 accuracy values for STL on VGG seem very large. In regards to Figure 4, I'd recommend either performing more training runs for Imagenet on Resnet50. The results look very compelling but without error bars don't say much. Similarly, comparing the results in Figure 4 with those for the other real-world datasets (e.g. Cifar10, Cifar100, STL) those contained in the Appendix which do have error bars are arguably less convincing of the primarily claims of speed to convergence. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A The authors address any limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: Misleading results: The reviewer correctly observed that, particularly on smaller datasets, the performance converges to be approximately equivalent by the end of training. Any observed deviations are likely due to random effects. This is to be expected, as all properly trained ETF solutions (i.e., with non-random labels) should theoretically yield similar results. Our work leverages the symmetries inherent in the optimal solution geometry to reach an ETF solution in fewer iterations. Thus, the primary advantage of our approach lies in its faster convergence while maintaining comparable generalisation ability. Notably, in some cases, our method demonstrates superior generalisation performance because, within the given training time, other approaches may have yet to reach the optimal max-margin solution. Non-competitive results: In our experiments, we utilised the AGD optimiser. While AGD achieves SoTA results for VGG architectures, it is not yet fully tuned for ResNet architectures. However, it has been demonstrated that state-of-the-art results can be obtained for ResNet with prolonged training (see [6]). Our focus, however, is not on achieving state-of-the-art results but rather on illustrating the convergence trajectory. Due to computational constraints and the required scale of experiments, we terminated training at 200 epochs. Figure 4 no error bars: Fixed and replaced with 5 seed runs (found in the attached pdf). Cost: Please see tables and discussion in the general response. Questions: Optimal ETF solution: As previously discussed, all ETF solutions are equivalent up to rotations and permutations. Our Riemannian optimisation approach ensures that with each gradient update step, we move towards the nearest ETF solution, effectively breaking these symmetries and achieving faster convergence. Consequently, while our solution is optimal wrt the Riemannian problem, it is also optimal for the learning problem, although no better than any other ETF solution. Compare two ETFs (ours vs fixed): In theory, we expect the testing accuracy of the two solutions to be equivalent. However, in practice, due to the finite amount of training and the complexity of tasks such as those with many classes (e.g., ImageNet), we observe that faster convergence to the ETF solution can lead to better generalisation performance. However, we are careful not to make any theoretical claims in this regard. Proximal term: Solving the original problem in Eq. 6 yields a non-unique solution due to the rank deficiency of matrix $M$, resulting in a family of possible solutions. To achieve a unique solution, we introduce a proximal term to make the problem strongly convex, as shown in Eq. 7. This approach stabilises the solution. Additionally, the DDN backward pass for Eq. 6 cannot be computed deterministically because of the singularity of the gradients. By incorporating the proximal term, we obtain non-singular gradients, which allows us to compute Prop. 1. $U_{prox}$: An obvious initialisation scheme for both $U_{prox}$ and $U_{init}$ is to set them as either canonical or randomly orthogonal directions. However, both approaches are sensitive to the initial parameter settings of the network. We found that the most effective strategy is to solve the Riemannian problem in Eq. 6 at the first gradient update step (corresponding to the first epoch in GD and the first mini-batch in SGD) to obtain a solution from the family of solutions. This solution, denoted $U*$, is then used as the initialisation for both $U_{init}$ and $U_{prox}$ in Eq. 7. We then solve Eq. 7 and perform the first backward pass. The intuition behind this initialisation scheme is to start with a solution that is both feasible and optimal for the original problem formulation. Empirically, this approach has yielded the best and most stable results. $\delta$ proximal param: $\delta$ is a tradeoff between the distance of the feature means from an ETF and the distance between the target direction ($U_{prox}$) and the currently optimised direction. We have found that a broad range of $\delta$ values provides stable and good performance. Therefore, $\delta$ can be set within this range, provided it is not too small (e.g., $\delta\ll 10^{-7}$, which essentially solves Eq. 6) and not too large (e.g., $\delta\gg 10$, where the proximal term dominates and results in a fixed ETF case). $\tau$ param in CE: Since we normalise features, the cross-entropy loss will have a non-zero lower bound (Thm. 1 of Yaras et al. [67]). $\tau$ adjusts this lower bound, bringing it closer to zero. According to Yaras et al., including this parameter results in more defined NC metrics while keeping accuracy unchanged. They empirically found that setting $\tau=5$ yields optimal results for real-world datasets and architectures, and we adopt this approach in our work. Clarification on Prop.1: The results in the proposition are currently presented wrt the unnormalised feature means matrix $\bar{H}$. However, this choice does not affect the result or its derivation. For consistency, we will update the proposition to use the normalised values, i.e., $\tilde{H}$. Gould et al. [21] derived their result for vector functions with vector variables, whereas we have generalised this result to matrix functions and matrix variables as a natural extension. Error bars in Figures: All experiments were conducted five times. The results for UFM (Fig 2, 3) are highly stable, with minimal deviation, and the ranges are easily visible due to the plots' linewidth. $P_{CM}$, defined in Eq. 16, represents the cosine margin of each specific feature. STL VGG range: We applied our method to various datasets and architectures without case-by-case tuning. We suspect that the initial variation observed in the STL case is related to the relatively large batch size used for that dataset. However, our method effectively reduces the variance within just a few epochs compared to the fixed ETF. --- Rebuttal Comment 1.1: Comment: Thank you for the clarifications to the questions and comments raised. I agree one of the primary advantages of your approach is in the increased speed to convergence. In this regard, this seems to be a useful technique. It also helps to have more clear insight in the computation cost incurred. I believe clarifying these additional points can help with the message the paper. I will maintain my score is it is.	Summary: This paper presents a novel approach to utilizing ETF geometry. Instead of fixing the weights or making them learnable, their approach dynamically adjusts the weights by solving a Riemannian optimization problem while allowing end-to-end training. They show that their approach outperforms both the fixed and learnable approaches in terms of convergence speed and generalization. Strengths: Originality: The idea of dynamically adjusting weights is not new, but in the context of neural collapse (NC), it is a natural extension. Fully learnable weights do not provide the ETF structure, and fixed weights are too restrictive. The proposed approach is a good compromise between the two and combines the best of both worlds. Quality: The paper is well-written, and the proposed approach is carefully supported by theorems and experiments. Clarity: The paper is well-written and easy to follow. Significance: Their approach is general and could be applied to a range of problems. The authors applied it to synthetic UFMs and some standard image benchmarks (CIFAR-10, CIFAR-100, STL-10, ImageNet). The authors plan to release code upon acceptance. Weaknesses: Overhead Cost: The proposed method computes the exponential moving average of the feature means, performs a Riemannian optimization, and computes the gradient of DDN. These components introduce overhead in terms of epoch time. The authors claimed in the paper that the gradient of DDN is not computed, and the Riemannian optimization overhead is negligible. This unsupported claim should be backed up by an additional experiment that reports these extra computation times. Standard Procedure: "To ensure fair method comparison," the authors include classifier weight normalization and feature normalization for the standard procedure. This is usually not the case when using CE loss (see Fig 2). The authors should justify this choice by providing the results without these normalizations for the standard procedure. Image Baselines Results are not SOTA: The reported results are not state-of-the-art. For example, ResNet-18 trained on CIFAR-10 only reaches 80.47%. It seems that these baselines are not well-tuned, and the gain of the proposed approach is not clear and could potentially fade away with a better-tuned baseline. Can the authors comment on this? Additionally, the authors should include the results using ResNet-50 on ImageNet, which should provide a stronger reference point. Fixed ETF Procedure: The authors only used the canonical simplex ETF for the fixed procedure. The weight matrix results in many zeros and could lead to poor performance when used as the fixed classifier because some neurons will be inactive. The authors should include the results using the fixed ETF with a non-canonical (i.e., projection on a random basis). Remarks: The authors should directly clarify in Tables 1 and 2 the ResNet architecture used (18 or 50). Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses section. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: For large-scale problems, the computational cost of the proposed approach could be a limitation due to the high memory cost of computing the backward pass of the Riemannian optimization. Therefore, the authors did not compute the gradient calculation when reporting their results for the image benchmarks. The authors claimed that they empirically observed no significant difference in performance for small-scale problems where the DDN gradient can be computed. Including these results in the supplementary material would also be beneficial. Moreover, we agree that future work should verify if this is true for large-scale problems. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Overhead Cost: Please refer to the tables in our general response and the discussion around computational concerns. Standard Procedure: Our new experiments show that the standard method, which excludes feature and weight normalisation, achieves similar performance (found in the attached pdf). However, we observe that the learned features are less well-aligned with a max-margin classifier compared to when normalisation schemes are included. Image Baselines are not SOTA: Since our results overlap with those previously reported in the AGD paper, it is demonstrated that state-of-the-art results can be achieved by extending training over more epochs using the AGD optimiser (Figure 4 in Berstein et al. (2023)). However, the primary focus of our work is not to establish a new state-of-the-art result but to study the neural collapse phenomenon. We originally included results using ResNet50 on ImageNet (Figure 4 in our paper), and in the attached pdf, we also included ImageNet results after five runs with error bars. Fixed ETF procedure: The reviewer’s observation that the canonical projection results in a sparse weight matrix would have been true if we were working with canonical orthogonal frames. For canonical simplex equiangular tight frames, the sparsity is observed in the semi-orthogonal matrix $U$ (i.e., $U$ is the identity matrix), not in the weight matrix itself, which remains dense (as shown in Equation 1). To address this concern, we conducted additional experiments with the fixed method on CIFAR-10 using ResNet-18, where we replaced the canonical rotation with a random orthogonal rotation (drawn from the Haar measure). Our results indicate that the performance remains superior and that the fixed and canonical methods yield nearly identical outcomes (found in the attached pdf). Remark: We have updated our table results to reflect the chosen ResNet and VGG architectures. Limitations: We will include the results in the supplementary material that test whether computing the DDN gradient affects the stability of our solution on small-scale problems.	Rebuttal 1: Rebuttal: We thank all the reviewers for their time and thoughtful feedback. Here we address the common question regarding the computational costs of our method, and we will address individual comments for each reviewer separately. Please refer to Table 1 for computational cost and Table 2 for memory cost measurements. The key insight is that incorporating DDN into the backward pass (in our Python implementation) becomes prohibitively expensive when significantly increasing the number of penultimate layer features and classes. However, as demonstrated in the ImageNet case, where GPU memory constraints prevent us from running this step, omitting the DDN backward pass (similar to a gradient stop operation) still yields strong results as we are still directing the solution to the nearest ETF. Note that there is still an indirect gradient path via the classifier, as shown in Figure 1 of our paper. Despite the strong results, we acknowledge the computational challenges associated with the DDN backward pass. The primary bottleneck in GPU memory usage is during the calculation of the constrained derivatives (Eq. 29), where explicit formulation of certain intermediate matrices occurs. Concerning the time complexity of the DDN backward pass, we use a direct solver to compute the expression in Proposition 1, which can be inefficient for large-scale linear systems. We plan to address these limitations (both time and memory) by using more efficient linear algebra solvers in future work. Table 1: Average time (in milliseconds) of a step update during training \| Model \| Standard Fwd \| Standard Bwd \| Fixed ETF Fwd \| Fixed ETF Bwd \| Implicit ETF Fwd \| Implicit ETF Bwd (w/o DDN) \| Implicit ETF Bwd (w/ DDN) \| \|----------------------\|--------------\|--------------\|---------------\|---------------\|------------------\|----------------------------\|---------------------------\| \| UFM-10 \| 0.2 \| 0.4 \| 0.2 \| 0.4 \| 7.6 \| 1.2 \| 1.8 \| \| UFM-100 \| 0.2 \| 1.0 \| 0.4 \| 0.8 \| 8.5 \| 1.1 \| 608.9 \| \| UFM-200 \| 0.3 \| 1.1 \| 0.3 \| 0.9 \| 8.7 \| 1.2 \| 17,909.9 \| \| UFM-1000 \| 0.3 \| 1.1 \| 0.5 \| 1.1 \| 23.1 \| 0.8 \| N/A \| \| CIFAR10 / Res18 \| 3.4 \| 4.3 \| 3.7 \| 4.5 \| 12.8 \| 5.1 \| 5.9 \| \| CIFAR10 / Res50 \| 7.0 \| 9.2 \| 6.9 \| 7.9 \| 19.4 \| 9.2 \| 15.0 \| \| CIFAR100 / Res50 \| 10.5 \| 6.5 \| 7.6 \| 8.1 \| 40.2 \| 11.3 \| 1,872.4 \| \| ImageNet / Res50 \| 7.4 \| 11.4 \| 7.3 \| 8.1 \| 160.2 \| 10.0 \| N/A \| Table 2: GPU memory (in Gigabytes) during training \| Model \| Standard \| Fixed ETF \| Implicit ETF w/o DDN Bwd \| Implicit ETF w/ DDN Bwd \| \|----------------------\|:---------:\|:---------:\|:-------------------------:\|:------------------------:\| \| UFM-10 \| 1.5 \| 1.5 \| 1.5 \| 1.6 \| \| UFM-100 \| 1.7 \| 1.7 \| 1.7 \| 10.7 \| \| UFM-200 \| 1.7 \| 1.7 \| 1.8 \| 70.3 \| \| UFM-1000 \| 1.9 \| 1.9 \| 2.9 \| N/A \| \| CIFAR10 / Res18 \| 2.2 \| 2.2 \| 2.2 \| 2.3 \| \| CIFAR10 / Res50 \| 2.6 \| 2.6 \| 2.7 \| 2.8 \| \| CIFAR100 / Res50 \| 2.6 \| 2.6 \| 2.7 \| 18.9 \| \| ImageNet / Res50 \| 27.5 \| 27.2 \| 27.8 \| N/A \| The attached PDF includes figures of additional results requested by reviewers, which we discuss in the individual responses below. Pdf: /pdf/ee3c174818ba529b262fe83bb16444dd40d4aeed.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Provably Efficient Reinforcement Learning with Multinomial Logit Function Approximation	Accept (poster)	Summary: This paper considers MDPs employing the MNL function for transition probability, following Hwang and Oh [2023]. The authors suggest efficient algorithms based on online Newton steps, inspired by [Hazan et al., 2014; Zhang et al., 2016; Oh and Iyengar, 2021]. Furthermore, to improve $\kappa$ dependency, they provide algorithms employing local learning with mirror descent inspired by [Zhang and Sugiyama, 2023; Lee and Oh, 2024]. The algorithms achieve $1/\sqrt{\kappa}$ or even detach the dependency of $\kappa$ from the leading term. Strengths: The suggested algorithms are computationally efficient and show improvement in $\kappa$ compared to the previous work of Hwang and Oh [2023]. Weaknesses: - Their suggested algorithms do not seem novel because they are based on previously proposed methods for logistic or MNL bandits. Specifically, the online Newton update is widely studied for MNL or logistic bandits [Oh and Iyengar, 2021; Zhang and Sugiyama, 2023]. - Furthermore, the improvement on $\kappa$ is based on the mirror descent algorithm proposed in [Zhang and Sugiyama, 2023; Lee and Oh, 2024], and the proofs seem to follow the steps in [Zhang and Sugiyama, 2023; Lee and Oh, 2024] in the appendix. - Lastly, the MNL model for transition probability may have an inherent weakness: the number of achievable states for each (k,n) must be finite, and it is required to know the state space of $S_{k,n}$. [1] Faury, Louis, et al. "Jointly efficient and optimal algorithms for logistic bandits." International Conference on Artificial Intelligence and Statistics. PMLR, 2022. Technical Quality: 3 Clarity: 3 Questions for Authors: Are there any non-trivial technical novelties in utilizing the online mirror descent method for MDPs? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discuss some interesting future work regarding regret bound. Additionally, I believe, as mentioned in Weaknesses, the MNL model for transition probability has an inherent weakness: the number of achievable states for each (k,n) must be finite, and it is required to know the state space of $S_{k,n}$. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comment. We will address your questions and clarify any misunderstandings, which may be due to our inadequate emphasis on the technical contributions in the presentation. We will improve the clarity in the revised version. --- Q1: "Their suggested algorithms do not seem novel because they are based on previously proposed methods for logistic or MNL bandits... " A1: We respectively disagree with this claim. While the idea of online updates has been utilized in previous works, we make several novel contributions in the MDP setting: - Oh & Iyengar (2021) use the global information of the Hessian matrix, which fails to remove the dependence on $\kappa$. Moreover, we identify and address a technical issue present in their work as discussed in Remark 1, which is valuable to the community. - Zhang & Sugiyama (2023) consider a multi-parameter model (i.e., the unknown parameter is a matrix $W_h^\star$), which differs fundamentally from our single-parameter model (i.e., the unknown parameter is a vector $\theta_h^\star$). The techniques in Zhang & Sugiyama (2023) cannot be directly applied to our setting because some key properties of the functions differ. For instance, a direct application of their method would result in a regret bound that scales linearly with the number of reachable states. - Lastly, we establish the first lower bound for this problem, which is a novel technical contribution. To summarize, our paper presents novel algorithms and regret guarantees for RL with multinominal logit function approximation, and also contributes novel technical ingredients to achieve desired results. --- Q2: "The improvement on $\kappa$ is based on the mirror descent algorithm proposed in [Zhang & Sugiyama, 2023; Lee & Oh, 2024]. Are there any non-trivial technical novelties?" A2: While our work builds upon these two important prior works, our solution incorporates non-trivial technical innovations crucial for achieving favorable regret bounds, especially due to the concerns of the intrinsic dimension in RL with function approximation setting. Below, we highlight the technical novelties: - Compared with Zhang & Sugiyama (2023). As discussed in A1, Zhang & Sugiyama (2023) consider a multi-parameter model, which differs fundamentally from our single-parameter model. As a result, the techniques in Zhang & Sugiyama (2023) cannot be directly applied to our setting because some key properties of the functions differ. A naive application of their method would lead to a regret bound that scales linearly with the number of the reachable states $U$. - Compared with Lee & Oh (2024). Lee & Oh (2024) study the single-parameter MNL bandit setting, and while our parameter update approach shares similarities with theirs, our construction of the optimistic value function differs significantly. Specifically, they use a first-order upper bound for each item, which is insufficient to remove the dependence on $\kappa$ in our setting. In contrast, we employ the second-order Taylor expansion to achieve a more accurate approximation of the value function. This approach is non-trivial and requires a different analysis. We will emphasize these technical novelties compared to prior works more clearly in the revised version. --- Q3: "The MNL model for transition probability may have an inherent weakness... " A3: It is a mild condition that the reachable state space is finite and known to the learner, which has been used in prior works. Below we illustrate the reasons: - This condition holds for many practical applications, such as the SuperMario game, where, despite the vast state space, the reachable state space is limited to four states and known to the learner, corresponding to the agent's possible movements: up, down, left, or right. - In fact, even for linear mixture MDPs, which are extensively studied in the literature (Zhou et al., 2021; He et al., 2022), the standard assumption is that $\sum_{s'} \psi(s') V(s')$ can be evaluated by an Oracle. This assumption assumes that the state space is finite and known to the learner implicitly. Moreover, several works on linear mixture MDPs have regret bounds that depend on the size of the state space (e.g., Zhao et al., 2023; Ji et al., 2024). This also implies that the state space is finite. - Moreover, even without exact information of $S_{k,h}$, our results also hold when replacing it with its upper bound. Importantly, the MNL model ensures valid distributions over states, addressing the limitations of linear function approximation. We believe this is a key shining feature of the MNL model, which will take an important step towards bridging the gap between theory and practice. ---- We hope the above responses sufficiently address your concerns. If our responses have properly resolved the issues raised, we would appreciate it if you could consider re-evaluating our work. We're also happy to provide further clarifications to additional questions. --- References: [1] Zhou, D., Gu, Q., & Szepesvari, C., Nearly Minimax Optimal Reinforcement Learning for Linear Mixture Markov Decision Processes. In COLT'21. [2] He J., Zhou D., & Gu Q., Near-optimal Policy Optimization Algorithms for Learning Adversarial Linear Mixture MDPs. In AISTATS'22. [3] Zhao, C., Yang, R., Wang, B., & Li, S., Learning Adversarial Linear Mixture Markov Decision Processes with Bandit Feedback and Unknown Transition. In ICLR'23. [4] Ji, K., Zhao, Q., He, J., Zhang, W., & Gu, Q., Horizon-free Reinforcement Learning in Adversarial Linear Mixture MDPs. In ICLR'24. --- Rebuttal Comment 1.1: Title: Thank you for your response. Comment: For the technical novelty, you mentioned that "In contrast, we employ the second-order Taylor expansion to achieve a more accurate approximation of the value function. This approach is non-trivial and requires a different analysis." Could you provide more details regarding this? --- Reply to Comment 1.1.1: Title: Thank you for your response. Comment: Thank you for your response. We would like to take this opportunity to provide additional clarification on the technical contributions of our work. While our algorithm and analysis share certain similarities with existing work on bandits (e.g., Zhang and Sugiyama, 2023; Lee and Oh, 2024), our approach introduces several novel elements specifically tailored to the MDP setting. As the reviewer rightly pointed out (Point 2 of Weaknesses), the online estimation step and the analysis in Section 5.1 do draw from recent advances in bandits. However, it’s important to note that this step only mitigates the dependence to $O(\kappa^{-1/2})$, and an $O(\kappa^{-1/2})$ dependence still persists in the regret bound as shown in Theorem 2. To further reduce the $O(\kappa^{-1/2})$ dependence, it is crucial to examine how local information is preserved in the regret analysis. To this end, Zhang and Sugiyama (2023) use a second-order Taylor expansion to construct the optimistic revenue (Proposition 1), but their approach is designed for a multi-parameter model, which differs fundamentally from our single-parameter model. Lee and Oh (2024) employ a first-order upper bound for each item and specific properties of MNL bandit to construct the optimistic expected revenue (Eq. (5) and (6) of their paper), which does not hold in our setting. To preserve local information in our work, we need to estimate the value function more accurately. Specifically, whereas Lemma 2 addresses value differences between the value functions, Lemma 4 refines this by incorporating a second-order Taylor expansion, allowing us to maintain the local information $p(\theta)$ rather than applying the maximum operator as in Lemma 2. Although this second-order Taylor expansion is inspired by the work of Zhang and Sugiyama (2023), we need new analysis to address the challenges unique to the single-parameter model, such as handling the negative term in our first-order term, which is not present in their work. Moreover, a naive application of their analysis would lead to a regret bound that scales linearly with the number of the reachable states $U$, which is undesired in the MDP setting. We hope this clarifies the contributions of our approach. We are happy to provide further details if needed. --- Rebuttal 2: Title: Thank you for your comment. Comment: Overall, I agree that there are several adjustments needed to apply the mirror descent approach to RL. This work extends previous research on bandits to the RL setting. However, I have concerns about the significance of the technical novelty of this work. I believe that the second-order Taylor expansion mentioned by the author may not be unique to RL, as it is also used in bandit problems. Therefore, I'm maintaining my score. --- Rebuttal Comment 2.1: Title: Thanks for your reply. Comment: We appreciate your feedback but have to disagree with this comment. We would like to offer some clarifications that underscore the contributions of our work to the field of RL. Technical relationship between bandits and MDPs. Since bandits can be viewed as one-step MDPs, it's common and reasonable for methodologies and analyses in MDPs to draw inspiration from bandits (e.g., linear bandits vs. linear MDPs). To some extent, bandit techniques are fundamental to modern RL theory and many RL algorithms are built upon bandit algorithms. However, the most intriguing and challenging aspect of RL theory lies in leveraging its unique structure, particularly when addressing the intrinsic dimension in function approximation settings. Our unique technical challenge and innovations. Our work builds on recent advancements in MNL bandits, which have been well acknowledged in our paper. However, it is crucial to note that several non-trivial technical innovations are necessary to achieve favorable regret bounds. While employing high-level ideas like the Taylor expansion is not entirely new, the corresponding terms of MNL MDPs are significantly different from the bandit setting, which requires new and more sophisticated analyses. Significance and Impact of Our Results. Beyond technical novelty, our results hold significant value and are important for the community. Understanding function approximation is one of the central challenges in RL theory and there are many efforts devoted to, yielding fruitful results. However, one limitation is the linear assumption may not guarantee valid transition probabilities. MNL MDPs address this limitation by incorporating non-linear function approximation, which brings significant challenges as well. Our work not only is the first to achieve nearly the same statistical and computational efficiency as linear function approximation but also establishes the first lower bound for this problem. Our work broadens the scope of function approximation greatly and makes a significant step forward in RL theory. We hope this clarification highlights the uniqueness and importance of our contributions and addresses your concerns regarding the novelty and impact of our work. We believe our work is crucial for the community. We appreciate it if you could reconsider your evaluation in light of these clarifications.	Summary: In this paper, the author analyzes a Markov Decision Process (MDP) model with non-linear function approximation. Specifically, in the finite-time horizon inhomogeneous episodic MDPs setting, the transition dynamics are unknown but the reward function is known. The author proposes using a multinomial logit (MNL) function approximation to estimate transition dynamics, which is superior to the linear function approximation if the model is misspecified in \cite{hwang2023model}. Additionally, the author proposes UCRL-MBL-OL, which adapts the previous work that is model-based and has large computational and storage complexity, to an online style that only consumes constant computation and storage resources. Moreover, the author has proven that the regret bound of UCRL-MBL-OL matches the state-of-the-art in Theorem 1. Its regret bound achieves $\tilde{O}(\kappa^{-1} dH^2\sqrt{K})$, where $H$ is the time horizon length, $K$ is the number of total episodes and $\kappa$ is considered as a parameter to control the sparsity of the transition dynamics and $d$ is the hidden dimensionality. Ignoring the logarithmic factor and $\kappa$, such a regret bound has only a $\sqrt{H}$ gap compared to the lower bound. After that, with additional assumption, the author utilizes the local information to propose another two algorithms, UCRL-MNL-LL and UCRL-MNL-LL+ to remove the dependence on $\kappa$ and get a tighter regret bound as well as maintain good properties of UCRL-MNL-OL. Strengths: 1. This paper is well-written. The author makes a clear improvement point compared to the literature. 2. The algorithm proposed by the author enjoys an online learning style that does not need to maintain a large historical set. Weaknesses: 1. Although this paper focuses on reducing the computation complexity, I am curious about the sample complexity of UCRL-MNL-OL. 2. Since the algorithm builds up the estimation of the transition dynamics by using MNL function approximation, is it considered a model-based algorithm? More specifically, does it require storing the transition dynamics for each state-action pair in every step? Technical Quality: 4 Clarity: 3 Questions for Authors: Please see the above "Weaknesses" Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your helpful feedback. We will address your questions below. --- Q1: "Although this paper focuses on reducing the computation complexity, I am curious about the sample complexity of UCRL-MNL-OL." A1: Thanks for your question. This work not only reduces computational complexity but also significantly improves sample complexity. A sample complexity guarantee can be derived from any low-regret algorithm using an online-to-batch conversion, as demonstrated by Jin et al. (2019). Specifically, suppose we have total regret $\sum_{k=1}^K[V_1^{\star}(s_1)-V_1^{\pi_k}(s_1)] \leq \sqrt{CK}$, then, randomly selecting $\pi = \pi_k$ we have that $[V_1^{\star}(s_1)-V_1^{\pi}(s_1)] \leq \sqrt{C/K}$ with constant probability by Markov's inequality. Thus, Theorem 1 implies UCRL-MNL-OL has a sample complexity of $O(\kappa^{-1} d H^2 / \epsilon^2)$. We will add this discussion to the revised version. --- Q2: "Is it considered a model-based algorithm? Does it require storing the transition dynamics for each state-action pair in every step?" A2: Our algorithm is model-based because it first estimates the transition parameters and then derives the value function based on these parameters. However, unlike traditional model-based methods, it does not require storing the transition dynamics for each state-action pair at every step. Instead, we only store the estimated transition parameter (d-dimension vector) and compute the transitions at the current state as needed. This efficiency is achieved by updating the parameters in an online manner, eliminating the need to store the entire history of transitions. --- We hope that our responses have addressed your concerns. We will be happy to provide further clarification if needed. --- References: [1] Jin, C., Allen-Zhu, Z., Bubeck, S. & Jordan, I. M., Is Q-learning provably efficient? In NeurIPS'18. --- Rebuttal Comment 1.1: Comment: Thanks for your answers. It would be great to add sample complexity to the paper. I will not change the rating and will continue to support this work. --- Reply to Comment 1.1.1: Title: Thank you. Comment: Thank you for your helpful suggestion. We will include a discussion on sample complexity in the next version.	Summary: This work studies the MNL function approximation inhomogeneous RL, achieves the $O(1)$ computation cost, and improves the regret guarantee with regard to $\kappa$. To improve the computation cost, this work employs the online Newton step instead of MLE estimation to estimate $\theta$. Then, they design a novel confidence set by making full use of local information to improve the dependence of $\kappa$. Strengths: 1. The use of local information instead of a uniform $\kappa$ is novel and useful to improve the dependence of $\kappa$. 2. The UCRL-MNL-LL+ removes the $\kappa$ dependence on the lower-order term and almost matches the optimal regret results by using high-order Taylor expansion. Weaknesses: 1. [1] also use the online Newton step to improve the computation cost in the logit contextual bandits setting. It would be better to discuss the novelty of UCRL-MNL-OL. [1] Oh, M. H., & Iyengar, G. (2021, May). Multinomial logit contextual bandits: Provable optimality and practicality. In Proceedings of the AAAI conference on artificial intelligence (Vol. 35, No. 10, pp. 9205-9213). Technical Quality: 3 Clarity: 3 Questions for Authors: Question 1: This work achieves great results in the stochastic reward setting. Can you discuss the challenge when extending to the adversarial reward setting? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your insightful review. We will address your questions below. --- Q1: "Oh & Iyengar (2021) also use the ONS to improve the computation cost..." A1: As mentioned in Line 184, the parameter estimation of UCRL-MNL-OL algorithm is inspired by the work of Oh & Iyengar (2021). However, to construct the optimistic value function efficiently in MNL-MDPs, we had to develop a different approach tailored to the specific structure of the value function in MNL-MDPs. Moreover, as highlighted in Remark 1 of our paper, we identify and address a technical issue present in their work. --- Q2: "Can you discuss the challenge when extending to the adversarial reward setting?" A2: Thanks for your insightful question. Extending our method to the adversarial reward in the full-information setting is relatively straightforward. Instead of using a greedy policy (i.e., $a = \mathrm{argmax}~ Q_k(s, \cdot)$) as in our current approach, we can incorporate a policy optimization step, such as $\pi_{k+1}(\cdot \| s) = \pi_k(\cdot \| s) \exp(\eta Q_k(s, \cdot))$. This modification would result in an additional regret term of $O(H^2 \sqrt{K})$. However, extending our method to the adversarial reward setting for the bandit setting is more challenging, as we need to handle both unknown transitions and unknown rewards simultaneously. To the best of our knowledge, even the state-of-the-art algorithms for linear mixture MDPs in this setting of Li et al. (2024) exhibit a regret bound that scales linearly with the number of states. --- We hope that our responses have addressed your concerns. We will be happy to provide further clarification if needed. --- Reference: [1] Li, L. F., Zhao, P., & Zhou, Z. H., Improved Algorithm for Adversarial Linear Mixture MDPs with Bandit Feedback and Unknown Transition. In AISTATS'24. --- Rebuttal Comment 1.1: Comment: Thanks for your careful response. The rebuttal addresses my concerns and it would be better to add the above content in the next version. I maintain my score to support this valuable work. --- Reply to Comment 1.1.1: Title: Thanks for your reply. Comment: Thank you for your helpful suggestions. We will incorporate these discussions into the next version. Thanks again!	Summary: The problem considered in this paper is online learning in MDPs where transition probabilities are modelled with a log-linear model (with "multinomial logit function approximation"). The finite horizon, time-inhomogenous setting is considered. The problem is motivated by allowing a nonlinear transformation in modeling the MDP and yet maintaining both computational and information theoretic tractability. Inspired by results in the analogous bandit problems and algorithms developed for them, a number of gradually more complex, but (statistically) better performing algorithms are considered. In particular, while naive approaches give a poor dependence on a problem parameter $\kappa$ that characterizes the "strength" of nonlinearity, by adopting previous ideas to the MDP setting, new algorithms are designed that eliminate this poor dependence. A lower bound is also established, which nearly matches the upper bound (but considers infinite action spaces, while the main paper considers finite action spaces). Strengths: This is a reasonable problem setting; and the approach is also reasonable. It is nice to have a lower bound, even if there is a mismatch between the settings. It is nice to see that ideas that were developed for the bandit setting generalize to the MDP setting. Weaknesses: 1. The novelty is limited by that we have seen the same story, same ideas playing out nicely in the closely related bandit setting. 2. A new parameter, U, the number of next states that are reachable with positive probability in the worst case, appears in the analysis and will appear in the bounds. 3. It is an unpleasant surprise for the reader to discover this dependence only through carefully reading the paper, rather than being told upfront. It is not good that the opportunity to discuss whether this quantity needs to enter the regret bound, and that this quantity needs to be small for the algorithm to be tractable, is missed. 4. Line 83 and onward: The work of Uohamma is discussed but is mischaracterized. My reading of this work is that they do establish that their algorithm runs in polynomial time. It remains unclear why the exponential family model is incomparable with the one considered here; an explanation (with examples) is missing. 5. The paper could use some extra proofreading (e.g., the upper indices in the bottom of page 5, in the displayed equation are not correct); in line 149, in the definition of $U$, $\|\cdot\|$ is missing. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can you confirm that the regret and compute cost depend on U, the worst-case number of next states that one can transition to with positive probability? Do you think such dependencies are necessary? Are there any interesting examples where it is reasonable to expect that U is small, independently of the size of the state space? 2. What was the most challenging aspect of extending the bandit ideas to the MDP framework? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: n.a. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your constructive review. Below, we will address your main questions, especially regarding the dependence on $U$ (see A1-a,b,c), technical challenges (see A2), the difference to prior work (see A3), and presentation issues (see A4). --- Q1-a: "The regret and compute cost depend on $U$?.... Do you think such dependencies are necessary? " A1-a: We appreciate your observation. Though $\kappa$ may have polynomial dependence on $U$ in the worst case, the leading term of our best result (Theorem 3) is independent of $\kappa$ and only has a logarithmic dependence on $U$. While the lower-order terms do depend on $\kappa$, they are independent of the number of episodes $K$ and can be treated as constants. Thus, this dependence is believed to be acceptable. The dependence on $U$ is necessary for our method and we explain the reasons below. In general, methods for MDPs can be classified into two categories: model-based and model-free. Model-based methods aim to maximize the likelihood of observed data while model-free methods focus on minimizing the mean-squared error in predicting value functions. - The dependence on $U$ for regret and computational cost is usually necessary for the model-based methods, as it controls the model's error, which typically involves a total of $U$ elements. Similar dependencies have been observed in the literature (e.g., Hwang & Oh, 2023). - Most papers on linear MDPs use the value-targeted approach, a model-free method that does not rely on such dependencies. This is feasible because the linearity of the value function in linear MDPs allows for learning the value function directly. In contrast, the value function for MNL-MDPs is neither linear nor log-linear, preventing using model-free methods. We will further elaborate on this point in the revised version. --- Q1-b: "Examples where $U$ is small...?" A1-b: This phenomenon is quite common in practice, as in many applications, the agent tends to transit to nearby states even if the entire state space is large. An illustrative example is the RiverSwim environment in Hwang & Oh (2023), where although the state space is extensive, the agent only transits to adjacent states. Another example is the SuperMario game, where, despite a vast state space, the value of $U$ is limited to 4, corresponding to the possible movements: up, down, left, or right. --- Q1-c: "It is an unpleasant surprise for the reader to discover this dependence..." A1-c: Thanks for your helpful suggestion. We will discuss this dependence in the next version. --- Q2: "The novelty is limited... What was the most challenging aspect...?" A2: While similar ideas have been explored in the bandit setting, there are several unique challenges specific to MDPs that need to be addressed, especially due to the concerns of the intrinsic dimension in RL with function approximation setting. - Compared with Zhang & Sugiyama (2023). The model of Zhang & Sugiyama (2023) fundamentally differs from ours. We employ a single-parameter (i.e., a vector $\theta_h^$) model for different states at each stage $h$, whereas they use a multi-parameter* model (i.e., a matrix $W_h^$). As a result, the techniques in Zhang & Sugiyama (2023) cannot be directly applied to our setting because some key properties of the functions differ. For instance, a direct application of their method would result in a regret bound that scales linearly with the number of reachable states $U$. - Compared with Lee and Oh (2024).* While Lee and Oh (2024) study the single-parameter MNL bandit setting, and our parameter update approach shares some similarities to theirs, their construction of the optimistic value function is significantly different. They use a first-order upper bound for each item, which is insufficient to remove the dependence on $\kappa$ in our setting. To address this, we employ the second-order Taylor expansion to achieve a more accurate approximation of the value function. We will emphasize the challenges and our contributions more clearly in the next version. --- Q3: "The work of Ouhamma et al. (2023) and why it is incomparable.." A3: Thanks for pointing it out. Ouhamma et al. (2023) studied the exponential family model, a similar setting to ours. However, our work exhibits differences from theirs for both problem setting and computational complexity. Below, we make the clarifications. - Problem setting. The MNL MDPs studied in this paper can not be covered by the exponential family of MDPs in Ouhamma et al. (2023). An example of an MNL MDP that does not belong to the exponential family is given by the function $\phi_i(s, a, s') = \exp((s' - s -i)^/a^2) / \sqrt{\pi a^2}$. It does not belong to the exponential family as $\phi_i(s, a, s')$ can not be decomposed in the form $\phi_i(s, a) \cdot \psi_i(s')$. More discussion on this point can be found in Zhou et al. (2021). - Computational complexity. Though the algorithm of Ouhamma et al. (2023) runs in pseudo-polynomial time, our method is more efficient regarding storage and computational complexity. They estimate the transition parameters using the maximum likelihood estimation (MLE), which requires storing all previously observed data. Consequently, the computational complexity at episode $k$ is $O(k)$. In contrast, we employ an online parameter update method, which requires only $O(1)$ storage and computational complexity per episode. We will provide a more detailed comparison in the next version. --- Q4: "The paper could use some extra proofreading" A4: We appreciate your feedback and will ensure that the revised version is thoroughly proofread. --- We hope that our responses have addressed your concerns. We will be happy to provide further clarification if needed. --- Reference: [1] Zhou, D., He, J., & Gu, Q., Provably Efficient Reinforcement Learning for Discounted MDPs with Feature Mapping. In ICML'21. --- Rebuttal 2: Comment: Thanks for the rebuttal; it was useful. Overall, my feelings towards the paper have not changed: This is a fine peace of work that should get published. I still have some reservations concerning the model; in fact, one of the reservations I have I forgot to include into my original review is that the algorithm needs to know the support of the next state distribution. This combined with that the size of this support appears in the bound, and that in the related linear kernel MDP setting, the size of the support does not appear in the bounds makes me feel uneasy about the paper. Why is knowing the support problematic? Well, if the size of the support was not part of the bound, one could just "max out" the support (all states are possible next states). If not, in the lack of results showing that the algorithm is robust in the face of misspecified next state support, one is afraid that knowing the support will actually be important. And I expect knowing the support is kinda tricky. Consider for example the SuperMario game, mentioned in the rebuttal. Here, knowing the next states (all 4 of them) encodes a tremendous amount of information about the game. Just consider how many states there are here and compare this to the number 4. How realistic is that one would actually know the support in scenarios like this. So I have doubts. I think an example, where the assumptions are more natural would tremendously help the paper. Also, why not think of some "real" application that people may care about (I have to say I have a hard time imagining anyone to seriously care about how to play SuperMario; and I also have a hard time imagining a more general problem that has the characteristics of SuperMario (ie I don't have a problem with simplified examples, but SuperMario and other games just feel like too arbitrary and unrelated to any "real world" applications).) In summary, notwithstanding these reservations, I am in support of accepting the paper, given that it looks at a somewhat reasonable setting and makes a nontrivial contribution. --- Rebuttal Comment 2.1: Title: Thanks for your valuable comment. Comment: Thanks for your constructive feedback. We agree with the reviewer's observation that the support of the next state distribution plays a key role in the MNL MDP model. Investigating how to remove this prior knowledge is an important future direction. Even though, there are some real-world applications beyond games where the support of the next state distribution is limited and known. For instance, in the robot navigating problem, the robot can only move to nearby locations, even though the overall state space may be extensive. Similarly, in language models, the current state can be conceptualized as the sequence of previously generated words, with the immediately accessible next states being the potential subsequent words. Although the vocabulary could be extensive, the feasible choices for the next word, dictated by grammar and context, are inherently limited and known. We will add these discussions in the next version. We appreciate your helpful suggestions!	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper studies the recently proposed MDPs that use multinomial logit function approximation for state distribution validness. The results and algorithms improve the prior work of Hwang and Oh [2023] in multiple aspects, including computation efficiency, storage, and statistical dependence on the problem-dependent quantity $\kappa$ that can be exponentially small. In addition, the authors establish a matching lower bound on $d$, the feature space dimension, and $K$, the number of episodes. Strengths: - The paper is well-written and has clear logic flows. Readers can see how the authors approach the MDP problem and tackle the challenges. In particular, Table 1 is quite useful for demonstrating the advancements in the work. - The improvements in both computation and storage efficiencies are essential for practical applications. In Theorem 2, the authors also improve the dependence on $kappa$ to $\sqrt{\kappa}$ without affecting efficiency. The enhancement seems significant, especially since the parameter can be exponentially small. - The lower bound established in the paper is the first to demonstrate the optimality of the authors' algorithms in the $d$-$K$ dependence. Per my understanding, it also confirms the results' optimality of Hwang and Oh [2023]. Weaknesses: - The primary high-level techniques and tools (seem to) come from existing works and relevant fields, such as MNL contextual bandits. The authors should put more effort into highlighting the technical challenges and novelties besides the previous comparisons. - It would be beneficial to include experiments on synthetic and real-world datasets and compare the results to existing baselines and relevant works. In particular, the new algorithms seem more involved than prior ones, which may affect their stability and adaptiveness. - There is still a significant gap between the lower and upper bounds. Besides, I wonder how often $\kappa$ could be exponentially small in practical settings, though it's definitely of theoretical interest to approach the lower limits on parameter dependency. Technical Quality: 3 Clarity: 3 Questions for Authors: Overall, I think the paper makes reasonable contributions to the problem, and I have no additional questions/comments besides the above. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have made various comparisons and discussed the limitations of the results, which I'm satisfied with. I do not see any potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your constructive review. We will address your concerns below. --- Q1: "The primary high-level techniques and tools (seem to) come from existing works and relevant fields..." A1: While similar ideas have been explored in the bandit setting, there are several unique challenges specific to MDPs that need to be addressed, especially due to the concerns of the intrinsic dimension in RL with function approximation setting. Below, we highlight the key differences between our work and the existing literature: - Compared with Zhang & Sugiyama (2023). The model of Zhang & Sugiyama (2023) fundamentally differs from ours. We employ a single-parameter (i.e., a vector $\theta_h^$) model for different states at each stage $h$, whereas they use a multi-parameter* model (i.e., a matrix $W_h^$). As a result, the techniques in Zhang & Sugiyama (2023) cannot be directly applied to our setting because some key properties of the functions differ. For instance, a direct application of their method would result in a regret bound that scales linearly with the number of reachable states $U$. - Compared with Lee and Oh (2024).* While Lee and Oh (2024) study the single-parameter MNL bandit setting, and our parameter update approach shares some similarities to theirs, their construction of the optimistic value function is significantly different. They use a first-order upper bound for each item, which is insufficient to remove the dependence on $\kappa$ in our setting. To address this, we employ the second-order Taylor expansion to achieve a more accurate approximation of the value function. We will emphasize the challenges and our contributions more clearly in the revised version. --- Q2: "It would be beneficial to include experiments on synthetic and real-world datasets..." A2: Thanks for your suggestions! We agree that adding experiments would be beneficial to the work. Nevertheless, our primary goal is to enhance the theoretical understanding of RL function approximation rather than designing a specific algorithm with state-of-the-art empirical performance for MNL MDPs. More specifically, the main focus of our paper is to investigate whether we can achieve (almost) the same computational and statistical efficiency as linear function approximation while employing a more expressive non-linear function approximation. We answer this question affirmatively by considering the MNL MDPs, which broaden the scope of RL with function approximation. Therefore, to some extent, our focus is mainly on the theoretical aspect. Given that our current upper bounds still exhibit a gap compared to the lower bound (as discussed in the next question), we may have to focus on how to achieve the minimax optimal rate for now, which is definitely a challenging open problem to work with and leave the empirical evaluation as the future work. --- Q3: "There is still a significant gap between the lower and upper bounds." A3: There is indeed a gap between the lower and upper bounds, but may not be as significant as it seems. The lower bound we established is instance-dependent, and if we focus on the worst-case guarantee, the lower bound is $\Omega(d H \sqrt{K})$. Our upper bound only looses a factor of $H$ compared to the lower bound, which is acceptable. As we discussed in Line 306-314, closing this gap remains a challenging open problem, and we will add more discussion on this point in the revised version. --- Q4: "I wonder how often $\kappa$ could be exponentially small..." A4: Thanks for your insightful question. The phenomenon that $\kappa$ is exponentially small is quite common in practice. As $\kappa$ is defined as minimum value of the product of any two state transition probabilities (i.e., $\inf_{\theta \in \Theta} p_{s, a}^{s^{\prime}}(\theta) p_{s, a}^{s^{\prime \prime}}(\theta) \geq \kappa$), it will be extremely small if there are some hard-to-reach states with very low transition probabilities. For example, in autonomous driving, there are emergency states that are rare and have very low transition probabilities. Similarly, in financial trading, sudden market crashes or booms are rare events that can be considered hard-to-reach states in the market conditions state space. These events typically occur under unusual conditions and are not frequently observed, resulting in low transition probabilities. --- We hope our responses address your concerns. We are happy to provide further clarification if needed. Thanks again for your valuable feedback.	null	null	null	null	null	null
XMask3D: Cross-modal Mask Reasoning for Open Vocabulary 3D Semantic Segmentation	Accept (poster)	Summary: This paper introduces XMask3D, a framework developed for open vocabulary 3D semantic segmentation. They propose the integration of the denoising UNet, derived from a pre-trained diffusion model, to generate geometry-aware segmentation masks conditioned on learnable implicit 3D embeddings. These binary 2D masks are used to filter mask-level embeddings of 3D representations and apply mask regularization, thereby improving the open vocabulary capacity of 3D features. Strengths: 1. The motivation is clear. 2. The proposed method is intuitive, and the experiments have validated their contributions. Weaknesses: 1. The organization should be improved. Section 3.1 provides an overview, while section 3.2 includes design insights and preliminary findings. The flow of these writings has puzzled me, making it difficult to grasp your key contribution. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Please provide further clarification on how mask-level alignment between 3D features and 2D embedding space can address the limitations of traditional techniques, such as global feature alignment or vision-language model distillation. Additionally, if texts (Category Labels) are concatenated with fused features, will it still create a unified feature space that encompasses 3D, 2D, and textual modalities? 2. Could you please provide further clarification on the main contributions of your research compared to PLA and OpenScene? Although the 3D caption process shares similarities with PLA, the overall pipeline resembles OpenScene, with the exception of the diffusion model and mask generator, which differ from the Multi-view Feature Fusion in OpenScene. 3. Does the Implicit 3D Captioner effectively work with your 3D features? From my understanding, the most reliable 3D captioner currently available is Cap3D, which generates captions for 3D objects by rendering multi-view images and utilizing BLIP2 and LLM for assistance. In the context of indoor-scenes, can we consider the Implicit 3D Captioner to be equally robust? It would be beneficial to present additional evidence to support this claim. 4. Can your text-to-image diffusion model effectively generalize to your datasets? If not, please provide examples of failure cases. Additionally, is the diffusion model fine-tuned during the training process or is it frozen? If not, please present additional results to demonstrate the robustness of your diffusion model in generating high-quality images within your datasets. 5. What is view-level contrastive loss? Why this loss is calculated between the view global feature and text embedding of the view image caption but have three coefficients? 6. It is recommended to show your 2D Mask and 3D Mask in Figure 3 to provide more visual evidence. 7. The authors should provide results on ScanNet++ (CVPR’23), which is a up-to-date dataset compared with ScanNet. 8. Since diffusion models are utilized, it is recommended to compared the model parameters and FLOPs compared with PLA and OpenScene. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1. The authors have addressed the limitations and potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your careful review and constructive comments! Hopefully the following response will address your concerns. ### 1. Paper organization. > The organization should be improved ... Thanks for your advice! We will reorganize the first two sections in our revised paper. ### 2. Clarifications. #### (A) Mask-level alignment. > ... clarification on how mask-level alignment ... The most critical challenge in addressing the 3D open vocabulary problem is establishing 3D-text alignment. Previous approaches have tackled this by enforcing either global-wise or point-wise alignment between 3D features and 2D-text embeddings. Global feature alignment typically involves calculating a global feature of a scene or view through pooling and applying contrastive loss with 2D-text embeddings. However, pooling often result in the loss of detailed information, leading to unclear and imprecise boundaries on novel categories. Point-wise feature alignment matches 3D points with 2D pixels using camera parameters, and performs dense contrastive learning between them. While this method preserves boundary details, it incurs a significantly larger computational burden and less stable training, as the loss is highly susceptible to outliers. Our proposed mask-level alignment presents an intermediate solution of these two approaches. Compared to global features, mask features retain more detailed information since there are dozens of masks within a single view. Additionally, mask features are more robust than point features, as the distraction from outliers is mitigated by the mask pooling operation. > ... unified feature space ...? The 2D features with high open-vocabulary capacity are embedded in the 2D-text feature space. By applying 2D-to-3D mask regularization to the 3D features, these 3D features are drawn towards the 2D-text feature space. Consequently, the fused features, which integrate both 2D and 3D features, are expected to align well with the 2D-text embedding space. This alignment is further conveyed by the superior open-vocabulary performance of XMask3D compared to PLA and OpenScene. Thus, we believe that concatenating text and fused features will effectively create a unified feature space that integrates 3D, 2D, and textual modalities. #### (B) Main contributions compared with OpenScene and PLA. > ... main contributions compared to PLA and OpenScene ... Sorry for not directly highlighting our contributions relative to OpenScene and PLA. OpenScene employs point-wise distillation and feature ensemble, while XMask3D utilizes mask-wise feature regularization. As discussed in Section 2(A), mask-wise contrastive learning is more robust and less computationally demanding compared to point-wise counterpart. PLA proposes entity-level point-caption association through set difference and intersection operations, which is less precise and adaptable than our mask-level 3D-2D-text association. The mask for 3D-text alignment is adaptively predicted by the 2D mask generator, making XMask3D an end-to-end and integrated system. Therefore, the core contribution of our paper is the introduction of a more adaptive and robust mask-level alignment technique. #### (C) Implicit 3D captioner. > Does the Implicit 3D Captioner ... According to the ablation studies in Table 3(a), the proposed implicit 3D captioner yields better novel class results compared to using vanilla text embeddings or implicit 2D caption embeddings, which convey the effectiveness of the implicit 3D captioner. However, unlike Cap3D, the implicit 3D captioner does not directly generate text captions. It only produces conditional features that work effectively with the XMask3D pipeline and we cannot guarantee its robustness in other 3D-to-text generation scenarios. #### (D) Text-to-image diffusion model. > Can your text-to-image diffusion model ... In our XMask3D pipeline, all weights of the denoising UNet are frozen. We do not generate images using the text-to-image diffusion model. Instead, we use the HxWxC features from the denoising UNet to generate 2D masks through the mask generator. Only the mask generator is trained on our datasets. #### (E) View-level contrastive loss. > What is view-level contrastive ... Sorry for the unclear statement. The ground truth for the view-level contrastive loss is the text embedding of the view image caption. The predictions are derived from the average pooling results of the 3D branch features, 2D branch features and fused features respectively. Therefore, we have three view-level contrastive losses and design three separate coefficients for them. ### 3. Additional visualizations. > ... 2D Mask and 3D Mask in Figure 3 ... Thanks for your insightful advice! In Figure 4 in the PDF attachment of the global response, we display the outputs from the 2D and 3D branches for the same samples shown in Figure 3 of our main paper. ### 4. ScanNet++ dataset. > ... results on ScanNet++ ... We apologize for not providing results on ScanNet++, as we followed the protocols of previous papers (OpenScene and PLA). Following your advice, we have submitted a request to download ScanNet++, but unfortunately it is still pending, so we are unable to conduct experiments within the rebuttal period. We appreciate your suggestion regarding this advanced dataset and plan to add it to our future work. ### 5. Resource comparisons. > ... model parameters and FLOPs ... Thanks for your suggestion! As the FLOPs of 3D models are conditioned on point cloud numbers, we only report the additional FLOPs of the 2D branch. In future work, we plan to replace the 2D branch with a more efficient and lightweight open vocabulary 2D mask generator. \|Method\|Trainable Params\|Non-trainable Params\|Extra FLOPs\| \|:-:\|:-:\|:-:\|:-:\| \|PLA\|11.0 M\|--\|--\| \|OpenScene\|15.6 M\|126.5 M\|OpenSeg: 1.6 TFLOPs\| \|XMask3D\|82.9 M\|1493.8 M\|Denoising UNet: 4.3 TFLOPs\| \|\|\|\|Mask Generator: 93.5 GFLOPs\| --- Rebuttal Comment 1.1: Title: Reviewer response Comment: Thanks for your responses, most of my concerns have been solved. I will change my rating to positive. Please carefully release your code and checkpoints in your future version. --- Reply to Comment 1.1.1: Title: Response to Reviewer tdkT Comment: Thanks for upgrading your score and providing valuable feedback. We will update our revised paper according to our discussions and release our code and checkpoints after the conference decision. Thank you again for your insightful and constructive suggestions that improve paper quality!	Summary: The paper proposes a precise and consistent mask-level alignment between 3D features and the 2D-text embedding space through a method called cross-modal mask reasoning. The proposed XMask3D model includes a 3D branch for capturing geometric features, a 2D branch for generating vision-language aligned masks, and a fusion block to combine 3D with 2D. Using a pre-trained text-to-image diffusion model as the 2D mask generator, the model leverages three techniques: 3D-to-2D mask generation, 2D-to-3D mask regularization, and 3D-2D mask feature fusion. Strengths: 1- The idea is novel, the author propose to merge 2D which provides high OV capabilities, with 3D features shich endoces 3D geometry. 2- The method performs remarkably better than the reported models, namely OpenScene. The experiments are also well structure Weaknesses: 1- The authors don't compare with state-of-the-art 3D semantic segmentation OV3D[1] 2- The authors highlighed fututre work in the limitation, it would be good if you can expand it with some limitation on the technical side or some failure cases. [1] Jiang, Li, Shaoshuai Shi, and Bernt Schiele. "Open-Vocabulary 3D Semantic Segmentation with Foundation Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: Please compare to OV3D mentioned in the weaknesses Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Needs to be expanded Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your careful review and constructive comments! Hopefully the following response will address your concerns. ### 1. Results comparison. > The authors don't compare with state-of-the-art 3D semantic segmentation OV3D. Thanks for your suggestion! We will include this outstanding method in the ScanNet comparison in our revised paper. OV3D primarily focuses on EntityText extraction and Point-EntityText association, while XMask3D concentrates on enhancing mask-level interaction between 2D and 3D modalities. Therefore, the contributions of OV3D and XMask3D are orthogonal and could complement each other. Once the official code of OV3D is released, we plan to integrate our model with OV3D's advanced design in prompting LLM-powered LVLM models and Pixel-EntityText alignment to further enhance the performance of XMask3D. ### 2. Expanded limitations. > The authors highlighed fututre work in the limitation, it would be good if you can expand it with some limitation on the technical side or some failure cases. #### (A) Technical limitations. Currently, the denoising UNet from the pre-trained diffusion model requires significant computational resources and impacts inference efficiency. In contrast to PLA, which consists solely of a 3D model, XMask3D could be further improved by replacing the 2D branch with a more lightweight and efficient 2D open vocabulary mask generator. #### (B) Failure case analysis. Thanks for your constructive suggestion! We provide three failure cases of XMask3D in Figure 1 in the PDF attachment of the global response. We will include this figure and the following discussions in the supplemental material of our revised paper. The first sample shows a `bathtub` with a `shower curtain` in the bathroom. However, the 2D/3D branch and the fused output of XMask3D all misclassify the `shower curtain` as `curtain`. This may be because these two categories are similar in object shape and texture, with the only difference being the surrounding environment. Since XMask3D only takes a corner of the room as input instead of the entire scene, the global environmental information is insufficient for making the correct category prediction. The second sample shows a large area of `picture` on the `wall`. The 2D/3D branch and the fused output of XMask3D all misclassified it as `wall`, due to their similar geometry. In most cases, a picture is a small region on the wall, and this picture, as large as the wall, is a typical corner case. This failure case may reveal XMask3D's over-reliance on geometric knowledge and lesser consideration of texture information when encountering out-of-distribution samples. The third sample shows a `sink` on a `counter`. Due to the occlusion problem, the `sink` point cloud is incomplete, negatively affecting the prediction of segmentation boundaries between the `sink` and the `counter`. This occurs because they are geometrically similar when the sinking-down part of the sink is missing. --- Rebuttal Comment 1.1: Comment: Thanks a lot for clarifying these points. The authors addressed my comments. Thus, I am happy to keep my score, and it would be great to report the results of OV3D and XMask3D together in the revised version. --- Reply to Comment 1.1.1: Title: Response to Reviewer AnYT Comment: We deeply appreciate the time and effort you dedicated to the careful review and insightful feedback on our paper! We will update our revised paper and add OV3D comparison according to our discussions.	Summary: The paper addresses the limitations of current open vocabulary 3D semantic segmentation methods, which primarily focus on creating a unified feature space for 3D, 2D, and textual modalities but struggle with fine-grained segmentation boundaries. To overcome these limitations, the authors propose XMask3D, a cross-modal mask reasoning framework that achieves more precise mask-level alignment between 3D features and the 2D-text embedding space. Strengths: 1. The part "incorporating a 2D mask generator to create geometry-aware open masks and apply fine-grained mask-level regularization on 3D features" seems reasonable and novel. 2. The paper is well-structured and easy to follow. 3. Analysis is thorough and insightful. Weaknesses: 1. The paper evaluates the proposed method on a limited set of benchmarks (ScanNet20, ScanNet200, S3DIS), all of which are indoor scene datasets. Authors could discuss how the method might perform on outdoor datasets. Additionally, the authors could provide a qualitative analysis of the model's potential limitations when applied to different environments. 2. The reliance on the denoising UNet from a pre-trained diffusion model could be seen as a potential weakness or limitation, especially given the computational resources required for training and inference. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The paper could benefit from a more detailed error analysis to understand the failure modes of XMask3D, especially in novel category segmentation. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your careful review and constructive comments! Hopefully the following response will address your concerns. ### 1. Application of XMask3D on other scenarios. > The paper evaluates the proposed method on a limited set of benchmarks (ScanNet20, ScanNet200, S3DIS), all of which are indoor scene datasets. Authors could discuss how the method might perform on outdoor datasets. We follow the experimental settings outlined in PLA for our experiments on indoor scene datasets. Based on your advice, we observed that other open vocabulary papers [1,2] also conduct experiments on the outdoor scene dataset nuScenes[3]. However, this dataset only provides point cloud data with annotations and raw image data without annotations. Without image-level segmentation ground truth, we cannot train the 2D mask generator. In our future work, we plan to replace the entire 2D branch with a pre-trained 2D open vocabulary segmentation model, enabling us to handle the nuScenes dataset without requiring 2D annotations. However, we have made efforts to provide some qualitative results on the nuScenes dataset. First, we project the 3D annotations onto 2D images using the camera's intrinsic and extrinsic parameters. Due to the sparsity of the 3D point cloud, we further employ the k-nearest-neighbor algorithm to fill in the blank regions without pixel-point correspondence, as shown in Figure 3 in the PDF attachment of the global response. The projected 2D label maps and the outputs from XMask3D on several nuScenes samples are shown in Figure 2, demonstrating the potential of XMask3D in handling outdoor scenarios. > Additionally, the authors could provide a qualitative analysis of the model's potential limitations when applied to different environments. As discussed above, the current architecture of XMask3D requires full annotations for both 2D and 3D data. Therefore, when dealing with outdoor datasets that lack fine-grained 2D annotations, the performance of XMask3D may decrease. Additionally, we provide a qualitative analysis of failure cases in `3. Failure case analysis`, which reveals potential limitations of XMask3D under various circumstances. ### 2. Computational costs. > The reliance on the denoising UNet from a pre-trained diffusion model could be seen as a potential weakness or limitation, especially given the computational resources required for training and inference. We acknowledge that the denoising UNet requires significant computational resources and slows down the inference speed. In our future work, we plan to address this limitation by replacing the 2D branch with a more lightweight 2D open vocabulary mask generator. Thank you for highlighting this potential weakness, which has inspired our future improvements. ### 3. Failure case analysis. > The paper could benefit from a more detailed error analysis to understand the failure modes of XMask3D, especially in novel category segmentation. Thanks for your constructive suggestion! We provide three failure cases of XMask3D in Figure 1 in the PDF attachment of the global response. We will include this figure and the following discussions in the supplemental material of our revised paper. The first sample shows a `bathtub` with a `shower curtain` in the bathroom. However, the 2D/3D branch and the fused output of XMask3D all misclassify the `shower curtain` as `curtain`. This may be because these two categories are similar in object shape and texture, with the only difference being the surrounding environment. Since XMask3D only takes a corner of the room as input instead of the entire scene, the global environmental information is insufficient for making the correct category prediction. The second sample shows a large area of `picture` on the `wall`. The 2D/3D branch and the fused output of XMask3D all misclassified it as `wall`, due to their similar geometry. In most cases, a picture is a small region on the wall, and this picture, as large as the wall, is a typical corner case. This failure case may reveal XMask3D's over-reliance on geometric knowledge and lesser consideration of texture information when encountering out-of-distribution samples. The third sample shows a `sink` on a `counter`. Due to the occlusion problem, the `sink` point cloud is incomplete, negatively affecting the prediction of segmentation boundaries between the `sink` and the `counter`. This occurs because they are geometrically similar when the sinking-down part of the sink is missing. ### References [1] Jihan Yang, et al. "RegionPLC: Regional Point-Language Contrastive Learning for Open-World 3D Scene Understanding." CVPR. 2024. [2] Qingdong, He, et al. "UniM-OV3D: Uni-Modality Open-Vocabulary 3D Scene Understanding with Fine-Grained Feature Representation." IJCAI. 2022. [3] Holger Caesar, et al. "nuscenes: A multimodal dataset for autonomous driving." CVPR. 2020. --- Rebuttal Comment 1.1: Comment: I've thoroughly reviewed the authors' responses and appreciate their thoughtful engagement. Most of my concerns have been addressed. I will stay in touch for further discussion as we approach the final rating. --- Reply to Comment 1.1.1: Title: Response to Reviewer Xo8p Comment: We greatly appreciate your response and valuable suggestions, which improved the quality and comprehensiveness of our paper! We will update our revised paper according to our discussions.	Summary: This paper addresses the challenge of open-vocabulary 3D semantic segmentation by utilizing 3D geometric features, 2D semantic embeddings, and text modality. The proposed approach adapts the ODISE method to the 3D domain, aiming to distill open-vocabulary semantic segmentation knowledge from a pre-trained text-to-image denoising diffusion model to a 3D segmentation model. Initially, an input point cloud is fed into a 3D encoder-decoder segmentation network, producing point-wise geometric features. Simultaneously, a pre-trained visual-language diffusion model generates 2D masks and embeddings from posed images of the same scene, conditioned on the 3D global feature of the 3D branch’s encoder. Unlike the ODISE method, an implicit $3D$ captioner is introduced to produce geometry-aware 2D masks while also distilling information from the 2D branch network to the 3D encoder. To further regularize the 3D network, a distillation loss ($\mathcal{L}_{mask}$) is applied to the 3D mask embeddings, derived from the per-point features and the 2D masks back-projected to the point cloud as 3D binary masks. By obtaining ground truth mask features from a pre-trained CLIP model, the 3D masked embeddings are aligned with the image-text joint embedding space, through a cosine similarity loss. This alignment leads to more coherent segmentation results and enhances the model's open-vocabulary capabilities. Finally, the per-point features are combined with the pseudo mask 2D features (formed by the back-projected 3D mask and 2D mask embeddings), resulting in a fused per-point representation that incorporates the geometric information from the 3D segmentation network and the semantic open-vocabulary capabilities of the 2D branch. The approach is evaluated on three semantic segmentation benchmarks (ScanNet, ScanNet200, and S3DIS) and demonstrates superior performance compared to competing methods. Strengths: The XMask3D effectively aligns 3D geometric features with 2D and textual modailities through knowledge distillation from visual-text joint embedding spaces inherent in the pre-trained 2D denoising UNet and the CLIP model. As evident by the ablation, the implicit 3D captioner is a crucial step in the overall pipeline, and it outperforms vanilla text conditioning or the implicit 2D captioner of ODISE, in both base and novel semantic categories. Moreover, the 2D-to-3D mask regularization is also essential, since it significantly improves the accuracy of the proposed method esp. in novel categories. This justifies the need for this additional distillation step from the CLIP joint space, to further enhance the open-vocabulary capabilities of the XMask3D method. Finally, the discussion on modality fusion, both in the main paper and supplementary, is highly appreciated. By dissecting the method and providing qualitative and quantitative results for each step, the authors make it easier for readers to understand and gain intuition about the presented approach. Weaknesses: While the method exhibits superior performance w.r.t. competing methods, it seems that the output fused embeddings yields to geometric inconsistent features for semantic classes that cover large areas of the point cloud such as wall, ceiling and floor. This is evident in both partitioning settings when the class is either base or novel (Table 5 (a) and (b) in supp.). Technical Quality: 3 Clarity: 3 Questions for Authors: Following the weaknesses section, do the authors have any additional insights into why this phenomenon occurs with the fused embeddings? Could the 2D regularization terms ($\mathcal{L}_{seg}^{2D}$, $\mathcal{L}\_{view}^{2D}$) be introducing too much bias towards the 2D visual modality, thereby causing geometric discontinuities in the output fused features? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have discussed the method's limitations in detail in Section 4.4. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your careful review and constructive comments! Hopefully the following response will address your concerns. ### 1. About the less satisfactory results of classes that cover large areas. > While the method exhibits superior performance w.r.t. competing methods, it seems that the output fused embeddings yields to geometric inconsistent features for semantic classes that cover large areas of the point cloud such as wall, ceiling and floor. This is evident in both partitioning settings when the class is either base or novel (Table 5 (a) and (b) in supp.). The underlying reasons for the unsatisfactory performances of semantic classes on the ScanNet and S3DIS datasets are distinct. For the ScanNet dataset, XMask3D processes only a corner of the scene in each forward pass, whereas PLA directly feeds the entire scene into the model. This segmentation approach negatively impacts categories that cover large areas, such as `wall`, `ceiling`, and `floor`, because cutting these categories into smaller pieces inevitably undermines segmentation performance by losing long-range relational information. For the S3DIS dataset, in addition to the aforementioned partial input issue, we encounter a trade-off between computational resource consumption and segmentation performance. The validation samples in S3DIS are densely populated, with an image view potentially corresponding to over 500,000 points, which causes out-of-memory errors on a 24GB NVIDIA 4090 device and slow inference speed. Consequently, we discard views exceeding a threshold of 260,000 points for higher efficiency. However, this view selection process results in hollow regions in the merged outputs, primarily in the corners of rooms, and these regions typically correspond to the categories `wall`, `ceiling`, or `floor`. We conducted an ablation study, increasing the selection threshold from 260,000 to 500,000 points, as shown in the following table. When the threshold is raised, the segmentation performance for `ceiling`, `floor`, and `wall` consistently improves. \| Partition \| Threshold \| Ceiling \| Floor \| Wall \| \| :-------: \| :-------: \| :-----: \| :---: \| :--: \| \| B8/N4 \| 260,000 \| 86.4 \| 88.3 \| 81.4 \| \| \| 500,000 \| 90.1 \| 91.1 \| 81.6 \| \| B6/N6 \| 260,000 \| 86.4 \| 47.4 \| 80.9 \| \| \| 500,000 \| 88.7 \| 51.1 \| 81.6 \| > Following the weaknesses section, do the authors have any additional insights into why this phenomenon occurs with the fused embeddings? Could the 2D regularization terms ($L_{seg}^{2D}$, $L_{view}^{2D}$) be introducing too much bias towards the 2D visual modality, thereby causing geometric discontinuities in the output fused features? As discussed above, the primary reasons are the partial point cloud input and the trade-off with memory consumption. We believe that the 2D regularization term does not significantly bias the fused features towards geometric discontinuities. As illustrated in Figures 1, 4, and 6 of our main paper, the fused features of `wall` and `floor` fully leverage the continuous geometry from the 3D branch. Additionally, we provide a qualitative analysis in the following table. The 2D branch performance is significantly worse than the others. The fused features perform better than the 3D features for `wall` and only slightly worse for `floor`. The numerical results also demonstrate that the performance of fused features in segmenting `wall` and `floor` is not significantly affected by geometric discontinuities from the 2D regularization terms. \| Partition \| Method \| Branch \| Wall \| Floor \| \| :-------: \| :-----: \| :----: \| :--: \| :---: \| \| B10/N9 \| PLA \| 3D \| 84.6 \| 95.0 \| \| \| XMask3D \| Fused \| 84.2 \| 94.7 \| \| \| XMask3D \| 3D \| 82.0 \| 95.1 \| \| \| XMask3D \| 2D \| 62.9 \| 78.6 \| \| B12/N7 \| PLA \| 3D \| 84.7 \| 95.1 \| \| \| XMask3D \| Fused \| 83.3 \| 94.6 \| \| \| XMask3D \| 3D \| 81.8 \| 95.0 \| \| \| XMask3D \| 2D \| 60.3 \| 76.7 \| \| B10/N9 \| PLA \| 3D \| 83.8 \| 95.2 \| \| \| XMask3D \| Fused \| 83.8 \| 94.7 \| \| \| XMask3D \| 3D \| 81.7 \| 95.1 \| \| \| XMask3D \| 2D \| 64.3 \| 78.5 \| --- Rebuttal Comment 1.1: Comment: Thank you for the clarifications and for addressing my concerns. I will maintain my positive rating --- Reply to Comment 1.1.1: Title: Response to Reviewer 4VWC Comment: We greatly appreciate your response and valuable suggestions, which improved the quality and comprehensiveness of our paper! We will update our revised paper according to our discussions.	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for taking the time to review our submission and provide constructive feedback on our work. We are encouraged by the consensus among reviewers regarding the strengths of our approach, which aligns with our intentions and efforts: 1. Novelty and Significance: We appreciate that all reviewers recognize our approach as both novel and reasonable. It is encouraging to see comments such as those from Reviewer AnYT, who noted that "The idea is novel," and Reviewer Xo8p, who remarked that "the part 'incorporating a 2D mask generator to create geometry-aware open masks and apply fine-grained mask-level regularization on 3D features' seems reasonable and novel." 2. Thorough Evaluation: We are pleased that the reviewers acknowledged the comprehensiveness of our experiments and evaluations. Reviewer Xo8p commented that "Analysis is thorough and insightful," while Reviewer tdkT noted that "The proposed method is intuitive, and the experiments have validated their contributions." 3. Clarity and Presentation: We are gratified that our efforts to present our ideas clearly have been well-received. Reviewer 4VWC highlighted that "By dissecting the method and providing qualitative and quantitative results for each step, the authors make it easier for readers to understand and gain intuition about the presented approach." Reviewer Xo8p also mentioned that "The paper is well-structured and easy to follow." We also appreciate the suggestions for further improving our work. In response to the specific concerns and recommendations raised by each reviewer, we have provided detailed discussions in our rebuttal and will update them in our revised paper. Additionally, we have prepared a PDF with additional illustrations to offer a more comprehensive visualization of our results and reinforce the validity of our work. Best regards, Submission 637 Authors Pdf: /pdf/ca5a56f6971289946a3288bbce54c10426f241f4.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
DiffusionPDE: Generative PDE-Solving under Partial Observation	Accept (poster)	Summary: This paper introduces diffusion methods to tackle the partially observed PDEs, named DiffusionPDE. By learning the joint distribution of solution and coefficient space, the proposed model can handle both forward and inverse problems. The authors experiment with diverse PDEs and settings to demonstrate the model's effectiveness. Strengths: - This paper successfully utilizes the diffusion methods in solving PDEs, covering both forward and inverse problems. - The main text and supplementary materials provide diverse experiment settings, which can well support the model’s effectiveness on partial observations. - This paper is overall clear and well-written. Weaknesses: 1. The technical contribution is limited. From a technical view, this paper is an application of the diffusion model in PDE solving. There are also some previous methods that also use diffusion methods and leverage the PDE loss [1]. Thus, I think the technical novelty is limited. [1] A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction, JCP 2023 2. Some powerful baselines are missing. - According to Figure 1, I think the base model of DiffusionPDE is U-Net. How about comparing it with a single U-Net? I think U-Net could be a powerful baseline. - There are also some latest models that are good at processing partially observed or irregularly placed PDEs, such as OFormer [1] and Transolver [2]. They should include them as baselines. [1] Transformer for Partial Differential Equations' Operator Learning, TMLR 2023 [2] Transolver: A Fast Transformer Solver for PDEs on General Geometries, ICML 2024 3. Model efficiency comparisons are needed, including GPU memory and running time. 4. I think the proposed model cannot predict the future evolution of a time-dependent PDE. Current tasks are all about “reconstruction” or “imputation”. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Figure 4, do both forward and inverse tasks use the same diffusion model? Or do we need to train two models for these two different tasks? 2. I think the base model is U-Net. So how does DiffusionPDE handle the spatially scattered partial observations? Is the input still in the regular grid, but only the sampled locations have ground-truth values? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: I appreciate that they have discussed the limitations. But I think the mentioned issues about efficiency and limitations on temporal modeling are not trivial. More discussions are expected. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful comments! We are happy that you find our paper provides diverse experiments and is well-written. > “The technical contribution is limited.” Please refer to the common response above. > “Some powerful baselines are missing.” 1. U-Net: We trained a U-Net model based on our EDM diffusion model. Initially, we trained the model to map between 500 fixed points in the input space and the full output space (Figure 9 in the PDF). For the Navier-Stokes equation, the prediction of the final state resulted in an average test error of approximately 39.1%, which is significantly higher than the error of our diffusion model. Additionally, we made predictions for the final state using 500 different sampling points, which increased the relative error to approximately 49.2%. Furthermore, we trained another U-Net model to map between 500 random points in the input space and the full output space. This U-Net model resulted in a 101% error during testing, indicating that the model fails to flexibly solve different patterns. 2. OFormer [1]: We examined the forward and inverse process of Navier-Stokes equation for OFormer taking 500 observation points, and the relative errors are approximately 17% and 23% respectively, which are much larger than DiffusionPDE, as shown in Figure 10 in the PDF file. 3. Transolver [2]: We believe that Transolver is applied to 3D geometric PDEs and is not suitable for comparison with our DiffusionPDE without significant modifications to their method. > “Model efficiency comparisons” Please see the common response. > “time-dependent PDE” Please see the common response. > Q1: “Figure 4, do both forward and inverse tasks use the same diffusion model?” Yes. We can do forward and inverse tasks with a single model. > Q2: “So how does DiffusionPDE handle the spatially scattered partial observations?” We use a 128*128 grid to represent the different spaces/states of the PDEs, so yes the input is still in the regular grid. This is a common practice for most works in area [3][4]. In addition, we leverage DPS [5] to handle the spatially scattered partial observations, which is a technique designed for noisy inverse problems (e.g. inpainting). [1] Transformer for Partial Differential Equations' Operator Learning. Li et. al. TMLR 2023. [2] Transolver: A Fast Transformer Solver for PDEs on General Geometries. Wu et. al. ICML 2024. [3] Physics-informed neural networks. Raissi et. al. JCP 2019. [4] Fourier Neural Operator for Parametric Partial Differential Equations. Li et. al. ICLR 2021. [5] Diffusion Posterior Sampling for General Noisy Inverse Problems. Chung et. al. ICLR 2023. --- Rebuttal Comment 1.1: Comment: I would like to thank the authors' response. After rebuttal, I am clear about the paper's technical contribution. There are still some questions remain. I think the extension of DiffusionPDE to a prediction model is non-trivial. The current framework cannot support this. Also, as the author mentioned, they treat the sparse observation in uniform grids, which is also not a generalizable design since the partial observations can be placed in any position of the domain. Some techniques, such as OFormer, may be more applicable than U-Net. After rebuttal, I think the application scope of DiffusionPDE is limited. The authors should discuss more about the above-mentioned limitations in the revised paper. However, I appreciate the design in the joint modeling of forward-inverse problems. Thus, I improve my score to 5. --- Rebuttal 2: Title: Clarification on DiffusionPDE’s Robust Performance Using Continuous Coordinates and Bilinear Interpolation Comment: We greatly appreciate your response and recognition of our method. We would like to clarify that DiffusionPDE can utilize continuous coordinates with bilinear interpolation in our prediction space to obtain predicted values for points that are not on the grid. By doing so, for example, the forward problem of the non-bounded Navier-Stokes equation results in a relative error of approximately 6.0%, which is comparable to the scenario without interpolation (approximately 6.9%) and even smaller as more points are considered during the calculation of observation loss, while still remaining significantly lower than OFormer (approximately 16.2%). Overall, the diffusion model is powerful, enabling DiffusionPDE to consistently outperform OFormer. We are happy to mention all these discussions in the revised paper.	Summary: The paper proposes to solve PDEs given only sparse measurements by jointly modeling the solution and coefficient space (e.g. the initial conditions) using a diffusion model. By applying diffusion posterior sampling (DPS) the authors obtain samples that are consistent with the sparse measurements and the underlying PDE equations. Several experiments show superior performance of the method compared to standard baselines such as PINNs and FNOs. Strengths: - Solving PDEs under partial observation is an important problem in real-world applications - The proposed method is technically sound and improves upon existing baseline methods (PINN, FNO) that do not work well for sparse measurements - Leveraging a pretrained diffusion model as a generative prior to model the joint distribution of solution and coefficient space is a good idea - The presentation of the method is clear and supported by concise algorithms and equations. The paper is well written overall - Experiments consider standard baseline methods for PDEs and cover a sufficient range of different dynamics ---- Post-rebuttal: the authors have addressed quite a few of the initial concerns, and while some concerns (e.g. about the magnitude of the contributions remain), I'd be happy to support an accept. I've raised my score accordingly. Weaknesses: - The main weakness of the method is the limited novelty. Both sparse measurements and physics-based losses have been considered together with diffusion models, see e.g. Shu et al. (2023). So it seems to me that the main technical novelty is to apply diffusion models to model the joint distribution of two simulation states at different points in time and apply DPS during inference for consistency with the sparse measurements and PDE constraints. - The experiments do not take into account any stochasticity or uncertainty. In principle, DPS will give a distribution of solutions, which is not the case for the other baseline methods, but this is not explored further in the paper. - Since the joint distribution models two states at time 0 and time T (for all experiments except Burgers' equation) and $0 \ll T$, the authors need to simplify the PDE loss $\mathcal{L}_{pde}$ to drop any time derivatives. This is a serious limitation. - It is not clear if DPS works better than classifier-free guidance, as used e.g. in Shu et al. (2023), or other methods for solving inverse problems with diffusion models. - DPS requires a lot of compute during inference for calculating $\mathcal{L}_{pde}$. For a fair comparison, it would be important to show the number of parameters, training time and inference time for all methods. Technical Quality: 3 Clarity: 3 Questions for Authors: - Algorithm 1 shows an adaption of DPS to EDM (Karras et al. 2022). Is this adaptation novel? Can the authors give some intuition why they apply the DPS losses in line 12 and 13 to the 2nd order correction (line 8) and not apply any trapezoidal rules in this case? - Are sparse measurements located on a grid that matches the resolution of the diffusion model or do they have continuous coordinates? In the second case, how are they interpolated to match the data resolution of the diffusion model? Does that make classifier-free guidance difficult to apply? - As noted in the weaknesses: why not use classifier-free guidance? I would like to see a discussion of different methods for inverse problems and diffusion models that can be used here instead of DPS and what are the advantages of using DPS. Reconstructing the solution/coefficient space from sparse measurements alone is a linear inverse problem with a number of different methods that can be used (e.g. Denoising Diffusion Restoration Models; Kawar et al. 2022, among many others) which oftentimes have much nicer theoretical guarantees/higher quality reconstructions and faster sampling speed. When considering these methods, is adding the PDE loss $\mathcal{L}_{pde}$ and thus making the problem a non-linear inverse problem really beneficial? - Likewise, as mentioned above: what are the parameter counts and runtimes of the method and the baselines? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have mentioned slow sampling speed as a limitation in the conclusion, but I think an extended discussion of this would be important to include. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for agreeing with us that we consider an important problem and propose a technically sound method. > “The main weakness of the method is the limited novelty” Please see our common response above. > “The experiments do not take into account any stochasticity or uncertainty” On one hand, we use a deterministic diffusion model. Thus, given partial observations as input, the only stochasticity or uncertainty in our method is the initial random noise. On the other hand, we agree it is important to take this into consideration and hence supplement an experiment where we test the effect of different noise seeds on the initial and final state of Navier-Stokes equations. Please see Figure 6 in the attached PDF. We promise to add this to the final version of the paper. > “The authors need to… drop any time derivatives” Please see the common response. > Q1: “Can the authors give some intuition why they apply the DPS losses in lines 12 and 13 to the 2nd order correction (line 8) and not apply any trapezoidal rules in this case?” DPS guides the sampling procedure by the end of each iteration, and we are not specifically modifying it. > Q2: “Are sparse measurements located on a grid?” Yes, we use a 128128 grid to represent the different spaces/states of the PDEs and the sparse measurements lie on the grid points. This is a common practice for most works in area [1][2]. However, we can also easily extend our method to be not restricted to grid points by interpolating on the grid and supervising the inerpolated value. > Q3: “It is not clear if DPS works better than classifier-free guidance”* In Figure 8 in the PDF, we evaluate the forward and inverse processes of the non-bounded Navier-Stokes equation, comparing DiffusionPDE with Diffusion with CFG when only initial and final states are considered. Our DiffusionPDE method achieves lower relative errors in both evaluations. We also compare our results with those of Shu et al. [3], where full time intervals are solved autoregressively. In this approach, the error of the final state increases to approximately 13%, which is higher than that of the two-state model (see Figure 10 in the PDF file). > Q4: “Show the number of parameters, training time and inference time for all methods.” Please see the common response. [1] Physics-informed neural networks. Raissi et. al. JCP 2019. [2] Fourier Neural Operator for Parametric Partial Differential Equations. Li et. al. ICLR 2021. [3] A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction. Shu et al. Journal of Computational Physics 2023. --- Rebuttal Comment 1.1: Title: Rebuttal Comment: Thank you for the updates and the interesting new results in the PDF. It's definitely an interesting direction, and I'm open to supporting acceptance by raising my score. I still find it difficult to strongly argue for acceptance given the limited scope of the technical contributions, though. --- Reply to Comment 1.1.1: Comment: We greatly appreciate your response. We would appreciate it if you could let us know of any specific questions we can address to help raise the score.	Summary: The work uses a guided diffusion process to solve the PDE forward and inverse problems with partial observations. Instead of learning the parameter-to-solution map ($a\rightarrow u$) as in Neural Operators, the method learns the diffusion process on the joint distribution $(a,u)$, and use guided diffusion for inference under sparse observations. Compared with several baseline method, the proposed method shows improved performance for solving forward and inverse problem with sparse observations. Strengths: The work uses a guided diffusion process to solve the PDE forward and inverse problems with partial observations.The authors compare with several baseline methods. The idea is clearly presented, and might be useful for the community. Weaknesses: The paper presents an interesting approach to solving PDE forward or inverse problems with sparse observations, which is an appealing concept given the minimal data requirement. However, this approach raises some concerns about the well-posedness of the problem. For example, in forward problems where sparse observations of the parameter $a(x_i)$ are available, there are infinitely many ways to interpolate $a$ and solve the PDE to obtain $u$. They are all valid solutions that satisfy both the PDE and the observations. This suggests that the method's ability to achieve good recovery might heavily rely on the strong regularization imposed by the training dataset, potentially limiting its practical utility as it may only favor solutions resembling those in the training set. Additionally, in Appendix C, Table 2, the weightings for observation and PDE loss are significantly higher (by two to six orders of magnitude) than those for $\nabla_x \log(p(x))$ as described in Equation 8, which might indicate a predominance of data fitting over the diffusion process. It would be beneficial if the authors could provide more guidance on how these weights were chosen and discuss the implications of using smaller weights. Understanding the rationale behind these choices could help clarify the model's dependency on these parameters and their impact on the solution's behavior. Technical Quality: 2 Clarity: 3 Questions for Authors: (1) The results from the baseline methods (PINO, DeepONet, PINNs, and FNO) are so bad. The claim that these methods are "not easily extendable" invites further scrutiny: (a) All the baseline models are supposed to represent smooth functions. However, in Figure 4, 7, 8, they look discontinuous at the training points. An explanation of how these models were trained and how inference was conducted could clarify why these discrepancies appear. (b) Taking PINNs as an example in the Darcy flow problem. Let $\hat{a}(x)$ and $\hat{u}(x)$ be the (potentially noisy) observation at $x$. We can represent $a(x)$ by a neural network, $a_V(x)$ (use neural net for convenience, could be other representations), and the PDE solution $u(x)$ by a neural net $u_W(x)$, where $V$ and $W$ are the weights of the neural network. We can solve the following optimization problem: $$\min_{V,W} \sum_{x\in T_d} (u_W(x) - \hat{u}(x))^2 + (a_V(x) - \hat{a}(x))^2 + \sum_{x \in T_r} (\nabla \cdot (a_V(x) \nabla u_W(x)) - q(x))^2$$ where $T_d$ is the observational point, and $T_r$ is the residual points, which does not need to be the same as $T_d$. (c) Similar, for a trained neural operator parametrized by W, $G_W[a] = u$. We can solve the following optimization problem: $$\min_V \sum_{x\in T_d} (a_V(x) - \hat{a}(x))^2 + (G\[a_V\](x)-\hat{u}(x))^2$$ where $G[a_V]$ is the solution of the PDE with parameter $a_V$. It seems that all the baseline methods can be used for forward and inverse problems with sparse observation. It is unclear why the proposed method would offer superior performance compared with the baselines. In contexts where full observation are available, as shown in Table 4, one might intuitively expect methods like PINNs—which utilize residual losses to ensure adherence to PDE constraints—or Neural Operators—which establish a direct parameter-to-solution mapping—to deliver more accurate results compared to a method that relies on a diffusion process. This leads to a critical inquiry on why diffusion process gives better accuracy for PDE problems. (2) How is the PDE loss computed? Is it by finite difference on a regular grid? Detailing this in the main text could help readers assess the accuracy and applicability of the PDE loss in different scenarios. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The author mention several limitation in the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the detailed and constructive review! > “The method’s success may heavily depend on the strong regularization from the training dataset.” We use the same data generation methods as other studies [1, 2, 3], ensuring fair comparisons. The process does not favor any specific subset of PDEs. It’s reasonable to assume that a dataset can be obtained containing patterns or regularization about field distributions. For example, factories manufacture parts with material properties following specific distributions, and our method can solve PDEs for these parts. While our method may struggle with out-of-distribution cases, we know of no existing method that performs well under sparse observations without assumptions about PDE coefficients to the best of our knowledge. This is an open and exciting field with great potential. > “The weightings for observation and PDE loss are significantly higher.” The weights depend on the scale of the coefficients or solutions. In our static PDE dataset, the absolute values of $u$ are much smaller than those of $a$. To ensure their distributions fall between -1 and 1, $a$ and $u$ are scaled differently during dataset preparation. During inference, they are scaled back to their original values to calculate PDE and observation losses. Thus, $\zeta_u$ should be larger than $\zeta_a$ to ensure the accurate recovery of both $a$ and $u$, as $L_u$ is much smaller than $L_a$. For example, in the Darcy Flow equation, $a$ is 3 or 12, while the absolute value of $u$ can be less than 0.01, making $L_u$ approximately $10^{-4}$ times $L_a$, leading to $\zeta_u$ being $10^4$ times $\zeta_a$. Similarly, the choice of $\zeta_{pde}$ is also influenced by the data scale. > Q1(a): “The baseline models… look discontinuous” All models are trained with full observations. During inference, we apply a mask to set non-observed input values to zero, which can cause output discontinuities, especially with FNO as the model does not guarantee smooth output. Results are smoother after applying the optimization from Question 1(b) (Figure 5 in the PDF). We also considered training on partial observations (Appendix G), where the model maps sparse inputs to full outputs, but this method doesn’t generalize to different sampling patterns. Additionally, using a different PINN codebase with modified hyperparameters [4] results in smoother outputs for the Burgers’ equation (Figure 4 in the PDF). > Q1(b)(c): Suggest optimizing baselines methods We acknowledge that optimizing the sampling space is another way to run the baselines. We applied this approach, and the results are shown in Table C below and Figure 5 in the PDF. Optimization reduces errors and smooths the solutions. However, the resulting values are smaller due to the smoothing effect from minimizing PDE loss, and the overall error compared to the ground truth remains much higher than DiffusionPDE. This may be due to the difficulty in optimizing the derivatives of noisy $a$ and $u$. Additionally, optimizing FNO and PINO is slow. While other models were optimized within 30 seconds, FNO and PINO took over 10 minutes to converge. Table B above shows these optimized baselines are slower than DiffusionPDE, likely due to the computational complexity of the Fourier transform. Table C: Relative errors of solutions (or final states) and coefficients (or initial states) when solving forward and inverse problems respectively with sparse observations after optimizing the baselines. Error rates are used for the inverse problem of Darcy Flow. \| \| \| DiffusionPDE \| DeepONet \| PINO \| FNO \| PINNs \| \|-\|-\|:-\|:-:\|:-:\|:-:\|:-:\| \| Darcy Flow \| Forward \| 2.5% \| 31.3% \| 32.6% \| 27.8% \| 6.9% \| \| \| Inverse \|3.2%\|41.1%\|49.2%\| 49.3% \| 59.7%\| \| Poisson\| Forward \| 4.5% \|73.6%\| 79.1% \| 70.5% \| 77.8% \| \| \| Inverse \| 20.0% \| 75.0% \| 115.0% \| 118.5% \| 73.9% \| \| Helmholtz \| Forward \| 8.8% \| 77.6% \| 67.7% \| 84.8% \| 79.2% \| \| \| Inverse \| 22.6% \| 100.7% \| 125.3% \| 131.6% \| 103.7% \| \| Non-bounded NS \| Forward \| 6.9% \| 96.5% \| 93.3% \| 91.6% \| 106.1% \| \| \|Inverse \| 10.4% \| 71.9% \| 87.8% \| 89.3% \| 108.6% \| \| Bounded NS \| Forward \| 3.9% \| 89.1% \| 80.8% \| 81.2% \| 84.4% \| \| \| Inverse \| 2.7% \| 88.6% \| 47.3% \| 48.7% \| 82.1% \| > Q1(c): “Why diffusion process gives better accuracy for PDE problems than other methods like PINNs” For partial observations, PINNs [1] does not learn any knowledge about the PDE, making it difficult to inpaint the missing parts. Neural operators [2] learn to map the entire coefficient space to the entire solution space and hence have difficulty taking partial observations as input. For full observations, PINNs require iterations of optimization to converge to the solution, which can be vulnerable to local minimal or failure to convergence, leading to higher error. It is also less robust compared to our data-driven method. Our method, on the other hand, enjoys the advantage of combining PDE knowledge and observation guidance through an iterative generative model. Iterative generative models, such as diffusion models, tend to beat feedforward models like GANs. > Q2: “How is the PDE loss computed?” Sobel filters are applied to the grid with asymmetric padding to compute derivatives. The PDE loss at the boundary is manually set to zero to handle edge effects. This procedure is applied to the pixel space rather than the continuous coordinate space. We promise to supplement these details in the final version of the paper. [1] Physics-informed neural networks. Raissi et. al. JCP 2019. [2] Fourier Neural Operator for Parametric Partial Differential Equations. Li et. al. ICLR 2021. [3] Learning to solve pde-constrained inverse problems with graph networks. Zhao et. al. ICML 2022. [4] Physics-informed deep-learning applications to experimental fluid mechanics. Eivazi et. al. Measurement Science and Technology 2024. --- Rebuttal Comment 1.1: Comment: Thanks for the the hard work and for conducting additional experiments. I appreciate the comprehensive benchmarking and the detailed responses to the reviewers' comments. > we know of no existing method that performs well under sparse observations without assumptions about PDE coefficients to the best of our knowledge. I agree. But I'm not asking for out-of-sample distribution. My questions is on the well-posedness of the problem. Using the Darcy flow as an example. Suppose $a$ is sampled from some Gaussian random field with certain correlation structure. If we only have partial observation of $\{x_i, a(x_i)\}$, for i = 1,...,N, and N is relatively small. Suppose the ground truth is $a_{GT}$. However, there could be $a_1$, ..., $a_M$ that are all consistent with the observation. That is, $a_j(x_i) = a_{GT}(x_i)$ for i = 1,...,N. and j = 1,...,M. In this case, DiffusePDE might recover $a_{GT}$ (hypothetically), while FNO, DeepOnet, PINN, etc, might recover $a_1$, ..., $a_M$, which will certainly have error with respect to $a_GT$, but they are all consistent with the observation data and the PDE. Therefore, the low accuracy of other baseline methods does not necessarily mean they are worse. They might be different but still valid solutions to the problem. As an even more straightforward approach in this case, we can sample the Gaussian random field conditional on the observation, and solve the PDE numerically, this would give infinite number of valid solutions to the problem. > During inference, we apply a mask to set non-observed input values to zero. Why do we mask the non-observed input values to zero? It seems a more reasonable approach is to exclude those in the loss function. > For partial observations, PINNs [1] does not learn any knowledge about the PDE, making it difficult to inpaint the missing parts My understanding is that PINN, or any method employing PDE loss, should be able to inpaint missing parts in ways that remain consistent with the PDE. As the PDE loss can be computed and minimized without any data. > Neural operators [2] learn to map the entire coefficient space to the entire solution space and hence have difficulty taking partial observations as input This statement appears to conflict with the typical characterization of neural operators (FNO or DeepONet) as being "resolution invariant" -- the neural operator can be evaluated at any point, even those not in the training points. Could you provide more details on the difficulties encountered? > For full observations, PINNs require iterations of optimization to converge to the solution, which can be vulnerable to local minimal or failure to convergence, leading to higher error. This seems to suggest that PINN fail to train in these cases, which invite questions on how is PINN trained, As the problems (Darcy flow, Poisson, etc) are standard examples in PINN literatures [1,2,3,4]. > Iterative generative models, such as diffusion models, tend to beat feedforward models like GANs. I'm not concerned about GANs. To my understanding, in your approach, the joint distribution between the parameter $a$ and the solution $u$ is modeled. That is, for any particular $a$, there is a conditional distribution of $u$. However, for PDEs, the solution $u$ should be uniquely determined by the parameter $a$. And neural operators aim to learn this mapping. It seems somewhat counter-intuitive that learning a joint distribution could be more effective than learning this mapping directly. As a summary, due to ongoing concerns regarding the problem formulation, implementation of some of the baseline method, and an lack of understanding of advantages of the diffusion-based approach, I will maintain my current score for this review. [1] M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” [2] DeepXDE(https://deepxde.readthedocs.io/en/latest/demos/pinn_forward.html) [3] Q. He and A. M. Tartakovsky, “Physics-Informed Neural Network Method for Forward and Backward Advection-Dispersion Equations,” [4] S. Wang, S. Sankaran, H. Wang, and P. Perdikaris, “An Expert’s Guide to Training Physics-informed Neural Networks,” --- Reply to Comment 1.1.1: Comment: We greatly appreciate your response. We would like to clarify a few points: > "the well-posedness of the problem" We agree that there are infinitely many valid solutions to the partial observations. However, we assume that there are statistical patterns in the coefficient and solution spaces, which means that each valid solution has its own 'probability' of actually being the solution. We aim to learn this probability distribution with our diffusion model, which excels at sampling a highly likely full state given partial observations, as reflected by the much lower error compared to all prior methods suggested by ourselves and other reviewers, including a simple baseline (suggested by reviewer nBsW) of completing the coefficient space using an RBF kernel. Such problems of statistically inferring unobserved data have been widely studied across domains, such as image inpainting in computer vision and matrix completion in machine learning, which are deemed well-posed problems of maximizing probability. > "Why do we mask the non-observed input values to zero?" For our method and other baselines that require optimization during inference, we indeed exclude non-observed values. For baseline methods that expect a complete input space, e.g., FNO, we have to fill in those values. We have tried filling in the missing values with constant zero values (default) or using RBF kernel or nearest neighbor interpolation, all resulting in lower performance than DiffusionPDE. > "As the PDE loss can be computed and minimized without any data. / This seems to suggest that PINN failed to train in these cases..." PINN is trained well according to the original paper, and the reviewer is correct in noting that PINN can handle partial observations and automatically complete the missing data. However, it’s generally understood that PINN tends to have a test error between 10% and 30%, even in fully observed situations, because the training loss never perfectly converges to 0. This high error is one of the reasons why there is growing interest in neural operators [1]. > "This statement appears to conflict with the typical characterization of neural operators (FNO or DeepONet) as being 'resolution invariant'... Could you provide more details on the difficulties encountered?" FNO can solve problems across different resolutions, and the reviewer is correct in noting that the output of neural operators can be evaluated at any point. However, we cannot assume that the input to neural operators can be sparse like ours. In fact, neural operators require a complete continuous grid as input, whereas we only have very sparse observation points (approximately 3%), which are highly discrete and do not meet this requirement. As a result, the model struggles to learn higher-frequency features. To the best of our knowledge, neural operators have not been shown to be effective on highly sparse inputs. > "It seems somewhat counter-intuitive that learning a joint distribution could be more effective than learning this mapping directly." The neural operators that learn the mapping directly indeed perform comparably to DiffusionPDE given "full" observation of $a$. However, when the observation of $a$ is not complete, the solution $u$ is not uniquely determined. Hence, we need to sample from the conditional distribution of $u$, a task in which DiffusionPDE outperforms neural operators and their variants suggested by the reviewers. [1] Physics-Informed Neural Operator for Learning Partial Differential Equations. Li et. al. ACM/JMS Journal of Data Science 2024.	Summary: The paper uses score based generative diffusion models to find the forward and backwards solution of a set of PDEs given partial observations of the solution and/or incomplete knowledge of the coefficients. The method performs well, and outperforms other ML methods such as FNO, as well as 'standard' FE type methods, for a range of standard test problems. The method reconstructs This is a novel approach, which delivers good performance, with low errors at a competitive speed. Extensive tests are given, with careful analysis of the results. Strengths: The use of score based generative methods in this context, where both the solution and the parameter estimates are updated, is novel. The method is clearly effective for the problems considered and should have good applications to real world examples. Extensive tests on a series of standard test problems show that the errors of the method are much lower than other ML based methods such as FNO. Weaknesses: This paper suffers as do many similar papers from a limited range of examples. It concentrates on the usual examples of PDEs such as NS and Bergers, and in both cases of these it looks at problems with quite moderate viscosity, which are realtively east to solve. This is more or less inevitable for such a short paper as this, especially as comparisons are needed with other method. But I would have liked to have seen more novel examples than the usual ones. This is not really a criticism of this paper, but is something to consider for future work. It would be imporoved by a fairer comparison with other methods which work with incomplete data and measurements. A clear exanple of this being the data assimilation widely used in physical modelling for just this range of problems. These should be descibed somewhere in the introduction and in Section 2. (Although of course these latter methods are slow in comparison.) The method is also limited (see later) to looking at certain slices of the solution. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. How does this method compare with a data assimilation approach 2. How easy would it be to extend the method to full time intervals 3. How easy is it to extend the method to higher dimensions 4. Have the authors tried out the method on more challenging PDE examples. 5. Also consider tests on NS and Bergers' eqn with much smaller viscosity. Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The model as described only looks at slices of solutions of 2D problems. This has been clearly identified by the authors. In this sense it is vrather limited when compared to other ML based approaches, and of course traditional FE based methods. I am pleased that this is recognised and that the authors plan to address this. The DiffusionPDE method will only be truly competitive when this is done, but this paper is a good step in this direction. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive comments on our work! We feel much encouraged that you recognize the novelty of our work. > a fairer comparison with other methods that work with incomplete data and measurements In addition to GraphPDE, we further compare our method with OFormer [1], Shu et al. [2], and UNet baselines. Please see Figures 7-10 in the PDF. > Q1: data assimilation We have compared the results using the RBF kernel (Figure 1 in the PDF). For the forward process of solving the Poisson, Helmholtz, and Darcy Flow equations, the RBF kernel achieved solution errors of approximately 14.3%, 23.1%, and 18.4%, respectively, with 500 random observation points. However, when addressing the inverse problem, the errors increased significantly to 141.2%, 143.1%, and 34.0%, respectively. This increase in error is likely due to the inherent challenges of solving inverse problems with such a straightforward method. We are happy to include such discussions in the revision of the paper. > Q2: full time intervals Please see our reply in the common response above. > Q3: higher dimensions Yes, we can extend the method to higher dimensions using state-of-the-art diffusion models designed for such cases. For instance, 3D diffusion models (e.g., [3]), especially latent diffusion models, could be appropriate for handling 3D geometric PDEs. > Q4: more challenging PDE examples To further demonstrate the generalization capability of our model, we conducted additional tests on different data settings for Darcy Flow. In Figure 2 in the PDF, we solve the forward and inverse problems of Darcy Flow with 500 observation points, adjusting the binary values of $a$ to 20 and 16 instead of the original 12 and 3. Our results indicate that DiffusionPDE performs equally well under these varied data settings, showcasing its robustness and adaptability. > Q5: NS and Burgers' equation with much smaller viscosity Yes, we also test DiffusionPDE on the Burgers’ equation with a viscosity of $1 \times 10^{-3}$ and on the Navier-Stokes equation with a viscosity of $1 \times 10^{-4}$, which are 10 times smaller than the ones in the main paper. For the Burgers’ equation, we are able to recover the full time interval with 5 fixed sensors at a relative error of approximately 6%, which is close to the error of approximately 2~5% in the main paper. For the Navier-Stokes equation, we can solve the forward and inverse problems with relative errors of approximately 7% and 9%, respectively, using 500 observation points. The errors are also close to the ones in the main paper, where the forward and inverse errors of NS are approximately 7% and 10%. Please see Figure 3 in the PDF. [1] Transformer for Partial Differential Equations' Operator Learning. Li et. al. TMLR 2023. [2] A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction. Shu et. al. JCP 2023. [3] 3D Neural Field Generation using Triplane Diffusion. Shue et. al. CVPR 2023.	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their constructive feedback! We will first clarify common concerns from the reviewers. > The method drops time derivatives and cannot solve for full time intervals (Reviewer nBsW, AaET) Our method can in theory support time derivatives and solve for full time intervals, as we have demonstrated in the 1D Burger’s equation (Figure 4 in the paper). Due to our choice of the diffusion model, which enables channel concatenation, we decided to model the joint distribution on the initial state and an arbitrary time state for other time-dependent PDEs such as the Naiver-Stokes equations. However, we think this is not a limitation of our method because: 1. The choice of diffusion model is orthogonal to our proposed method. For example, we can extend our solutions of the Naiver-Stokes equations to full time intervals by replacing the current 2D diffusion model with ones that support higher dimensions, such as 3D or video diffusion models [3,4,5], without changing the algorithm we described in the paper (Lines 12-14, Algorithm 1). 2. Even at its current state, where the method is applied to solve time-dependent PDE by modeling the joint distribution between two time states for time-dependent PDEs, our method can achieve more accurate results in solving both forward and inverse PDEs. Plus, our method is faster than methods that autoregressively solve time-dependent PDEs [1]. > “There is limited novelty of this paper compared with [1] (Shu et. al. 2023)” (Reviewer AaET, y14J) We would like to reiterate the novelty of this paper in that we propose to use diffusions to model the joint distribution of different spaces/states of PDEs, achieving state-of-the-art performance in solving forward/inverse PDEs under sparse observation. We would like to emphasize our main novelty/contribution as proposing to use joint distribution as a better way to model PDE problems (as stated in L32-34 in the paper and recognized by reviewer AaET). Compared with [1], we point out a few advantages brought by our method: 1. In terms of PDE types, Shu et al. (2023) is designed for fluid flow fields; as a result, they only show demonstration on limited PDE types such as Naiver-Stokes. Our method can work for a broader category of PDEs, such as Darcy Flow and Poisson equation. 2. In terms of diffusion models, we propose to use DPS [2] as a better choice for the problem. DPS is designed for solving inverse problems and thus suits well the task of solving PDEs given partial observation. We demonstrated that our proposed method can achieve lower errors and better robustness against different sparse sampling patterns (see Figure 7,8 in the PDF). 3. In terms of the problem setup, Shu et al. (2023) requires consistent observations across time. Our method, on the other hand, requires observations at only the initial or final state or both for time-dependent PDEs, enjoying more flexibility on the partial observation. 4. In terms of inference time, Shu et al. (2023) is slower because it is an autoregressive method. In comparison, our method using the joint distribution adopts a unified model that can handle both forward and inverse and shows more accurate results (see Figure 10 in the PDF file). > “Show the number of parameters, training time, inference time, and GPU memory” (reviewer AaET, y14J) As mentioned in Section 4.2 of our main paper, training the DiffusionPDE model takes approximately 4 hours on 3 A40 GPUs. In comparison, training the PINO or FNO model on a single A40 GPU takes approximately 8 hours, while training DeepONet takes around 40 minutes. We evaluate the computing cost during the inference stage by testing a single data point on a single A40 GPU for the Navier-Stokes equation, as shown in Table A. DiffusionPDE has a lower computing cost compared to Shu et al. (2023), which autoregressively solves the full time interval. This advantage becomes more significant when we increase the number of time steps. Table A: Inference computing cost of sparse-observation-based methods. \| Method \| DiffusionPDE (Ours) \| GraphPDE \| Shu et al. (2023) \| OFormer \| \|---------------\|:------------:\|:--------:\|:-------------:\|:-------:\| \| #Parameter (M) \| 54 \| 1.3 \| 63 \| 1.6 \| \| Inference time (s) \| 140 \| 84 \| 180 \| 3.2 \| \| GPU memory (GB) \| 6.8 \| 3.6 \| 7.2 \| 0.1 \| Further, we evaluate the inference runtimes on one single A40 GPU of vanilla full-observation-based methods and also the optimization time of them during the inference as suggested by Reviewer 4PC9. The optimization runtimes are significantly slower, especially when using Fourier transforms. Table B: Average inference runtimes (in seconds) of full-observation-based methods with and without optimization. \| Method \| PINO \| FNO \| DeepONet \| PINNs \| \|-----------------\|:------:\|:------:\|:--------:\|:------:\| \| Vanilla \| 1.0e0 \| 9.8e-1 \| 7.4e-1 \| 1.5e0 \| \| With Optimization \| 6.7e2 \| 6.7e2 \| 3.5e1 \| 3.7e1 \| [1] A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction. Shu et. al. JCP 2023. [2] Diffusion Posterior Sampling for General Noisy Inverse Problems. Chung et. al. ICLR 2023. [3] Diffusion Probabilistic Models for 3D Point Cloud Generation. Luo et al. CVPR 2021. [4] Flexible Diffusion Modeling of Long Videos. Harvey et al. NeurIPS 2022. [5] Open-Sora: Democratizing Efficient Video Production for All. Zheng et al. Github. Pdf: /pdf/9e1d3c53ce888e15b40ae8486fdac655dfa3e025.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ReLIZO: Sample Reusable Linear Interpolation-based Zeroth-order Optimization	Accept (poster)	Summary: From my understanding, this paper give a zero-th order algorithm with application to popular vision tasks neural architecture search and black-box adversarial attacks. The authors derive a closed-form solution after modeling the gradient estimation as a quadratically constrained linear program problem. The key idea is to try to decouple the required sample size from the variable dimension without extra conditions required, making it able to leverage the queries in the prior iterations. The speedup is further technically achieved by directly indexing some of the intermediate variables that contribute to the gradient estimation. The theoretical studies are given for its convergence speed and its cost-effectiveness is verified on benchmarks. Strengths: 1. Clear motivation with clearly derived approach, and it is a new zero-th order algorithm indeed and the authors also contextualize well the proposed method with related work discussion. 2. Strong empirical performance on representative vision tasks with rich testbeds and settings. 3. The approach by its design, could enjoy the efficiency of smoothing techniques while maintaining estimation accuracy. Table 4 in the appendix is informative. 4. The paper gives comprehensive results and technical details in both main paper and appendix. Weaknesses: 1. As remarked by the authors, it has few constraints on the sample size, similar to the smoothing techniques; and it requires the estimation of the gradients which involves solving a linear program problem. 2. As a zero-th order algorithm, it may still not be suited for large-scale application e.g. network training. Technical Quality: 3 Clarity: 3 Questions for Authors: I wonder if the proposed method could really facilitate the community of NAS? as zero-order optimization is not common in NAS. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful and valuable comments. We hope to address your concerns by answering your questions below. Q1: As noted by the authors, the method imposes few constraints on sample size, similar to smoothing techniques, but requires gradient estimation through solving a linear program, which could be a potential limitation. A1: Our approach indeed leverages the strengths of both linear interpolation-based and smoothing-based zeroth-order (ZO) methods, as highlighted in Remark (line 142). By solving a QCLP to estimate gradients, our method achieves higher accuracy compared to smoothing-based methods while maintaining lower computational costs than linear interpolation-based methods. Sec 3.3 also shows that part of the intermediate variables contributing to the gradient estimation can be directly indexed, significantly reducing the computation complexity of ReLIZO. Further analysis of the total computational complexity is provided in Section B and Table 4 in the appendix. Additionally, since optimization problems are often high-dimensional with a typically small sample size at each iteration (i.e. $n \ll d$), ReLIZO remains competitive in terms of speed. Q2: As a zeroth-order algorithm, is ReLIZO suitable for large-scale applications, such as network training? A2: Compared to gradient-based methods, ZO methods are well-suited for black-box optimization problems where gradients with respect to the variables are unavailable. Additionally, ZO methods are more memory-efficient than gradient-based methods since they do not require backward propagation, making them applicable to network training, as demonstrated by recent works [1,2]. Specifically, [1] introduces DeepZero, a principled ZO framework for deep learning that is computational-graph-free and can scale to deep neural network training with performance comparable to first-order methods. [2] also applies ZO methods to large language model (LLM) fine-tuning, highlighting their memory efficiency and introducing a ZO-LLM benchmark. To illustrate the applicability of ReLIZO in network training, we adopt it for fine-tuning an OPT-1.3b model (with 1.3 billion parameters) on the Stanford Sentiment Treebank v2 (SST2) task, following the methodology of [2]. The results, shown in the table below, indicate that ReLIZO outperforms other ZO methods across various fine-tuning schemes, including full parameter fine-tuning (FT), LoRA, Prefix-tuning, and Prompt-tuning. Notably, ReLIZO even surpasses SGD in the FT scheme while requiring significantly less memory, demonstrating its promising potential. \| Optimizer \| FT \| LoRA \| Prefix-Tuning \| Prompt \| FT Memory cost \| \| ------------- \| :--------: \| :--------: \| :-------------: \| :--------: \| :--------------: \| \| SGD \| 91.1 \| 93.6 \| 93.1 \| 92.8 \| 44.1 GB \| \| ZO-SGD \| 90.8 \| 90.1 \| 91.4 \| 84.4 \| 28.7 GB \| \| ZO-SGD-Sign \| 87.2 \| 91.5 \| 89.5 \| 72.9 \| 31.4 GB \| \| ZO-Adam \| 84.4 \| 92.3 \| 91.4 \| 75.7 \| 31.4 GB \| \| ReLIZO (ours) \| 93.4 \| 93.1 \| 91.8 \| 90.1 \| 35.7 GB \| [1] Chen A, Zhang Y, Jia J, et al. DeepZero: Scaling Up Zeroth-Order Optimization for Deep Model Training[C], ICLR 2024. [2] Zhang Y, Li P, Hong J, et al. Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark[C], ICML 2024. Q3: I wonder if the proposed method could really facilitate the community of NAS? As zero-order optimization is not common in NAS. A3: Recent studies [4,5,6] have explored zeroth-order optimization methods for solving bi-level optimization problems, including NAS task, where gradients of the parameters in the upper-level objective function are typically unavailable. These works highlight the limitations of gradient-based methods for NAS and demonstrate that ZO methods can outperform gradient-based approaches by providing more accurate gradient estimations of architecture parameters in the upper-level objective function. Nevertheless, this work aims to introduce a brand-new zero-order optimization algorithm and we apply it to NAS tasks to evaluate its performance on bi-level optimization problem. [4] Wang X, Guo W, Su J, et al. Zarts: On zero-order optimization for neural architecture search[J]. NeurIPS, 2022. [5] Xie L, Huang K, Xu F, et al. ZO-DARTS: Differentiable architecture search with zeroth-order approximation[C]. ICASSP, 2023. [6] Aghasi A, Ghadimi S. Fully Zeroth-Order Bilevel Programming via Gaussian Smoothing[J]. arXiv preprint arXiv:2404.00158, 2024. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response, which addresses my concerns.	Summary: The paper introduces ReLIZO, a novel zeroth-order optimization method leveraging linear interpolation to estimate gradients efficiently. It reduces the complexity of gradient estimation by reusing prior queries without additional conditions on sample size, decoupling it from variable dimension constraints. ReLIZO models gradient estimation as a quadratically constrained linear program, solving it analytically to reduce computation complexity. Experimental results demonstrate ReLIZO's efficacy in various scenarios, including black-box adversarial attacks and neural architecture search, showcasing faster convergence and better solutions compared to existing methods. Strengths: * The paper is well-written, with clear and easy-to-follow explanations. * The paper introduces a method for estimating gradients using arbitrarily sampled vectors without requiring orthogonal conditions or adherence to a specific distribution, enabling the reuse of queries to accelerate the zeroth-order (ZO) optimization process. * Extensive experiments on simulation benchmarks and real-world applications validate the method’s performance. * The paper highlights that ReLIZO can be viewed as a generalized version of traditional linear interpolation methods, capable of handling both equal and smaller sample sizes compared to variable dimensions. This demonstrates ReLIZO's theoretical soundness and enhanced flexibility in gradient estimation. Weaknesses: * The effectiveness of reusing queries depends on the choice of the reusable distance bound, which might require fine-tuning for different applications, adding complexity to its implementation. * While the method reduces the number of function queries, the process of solving the quadratically constrained linear program might introduce additional computational overhead for large $n$. Technical Quality: 4 Clarity: 3 Questions for Authors: In Figure 5, the results for the ARGTRIGLS problem indicate that any reusable distance bound leads to a performance drop. What is the specific structure of this problem? Does this suggest that the reuse strategy may be ineffective in certain special cases? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful and valuable comments. We hope to address your concerns by answering your questions below. Q1: The effectiveness of reusing queries depends on the choice of the reusable distance bound, which might require fine-tuning for different applications, adding complexity to its implementation. A1: In our method, the reusable distance bound $b$ restricts the distances between reusable samples and the current point and should be of the same order of magnitude as the step size. Following previous ZO research (Table 1 in [1]), we set the step size $\eta$ to $O(\frac{1}{d})$ to ensure convergence.Thus, we choose $b \sim O(\frac{1}{d})$. Ablation studies presented in Fig. 2 (in the main paper) and Fig. 4 (in the appendix) demonstrate that setting $b = 2\eta$ performs well across different optimization tasks and sample sizes $N$. As you noted, the reusable distance bound $b$ plays a critical role in balancing speed and accuracy. A larger $b$ can increase the reuse rate, reduce the number of queries, and accelerate the process. However, it also introduces a larger residual term $o(y-x)$ in the first-order Taylor expansion, leading to an increase in relative error. Therefore, the value of $b$ must be adjusted according to the specific application. From this perspective, an adaptively adjustable reusable distance bound $b$ during the optimization process would be ideal, and we plan to explore this in future work. [1] Ji, Kaiyi, et al. "Improved zeroth-order variance reduced algorithms and analysis for nonconvex optimization." ICML 2019. Q2: While the method reduces the number of function queries, solving the quadratically constrained linear program (QCLP) may introduce additional computational overhead for large values of $n$. A2: As the sample size $n$ increases, the cost of solving the QCLP also tends to increase. However: - On one hand, in zeroth-order optimization, the sample size $n$ is typically small compared to the dimension $d$ and satisfies $n\ll d$ in high-dimension optimization problems to ensure efficiency (e.g. in black-box adversarial attacks, the variable dimension $d=3072$ while the sample size $n=10$). The performance of ZO methods in such extreme cases where $n\ll d$ is crucial for assessing their scalability and robustness. Our experiments, detailed in Appendix C.1, demonstrate the effectiveness of our method even under these conditions. - On the other hand, Sec. 3.3 introduces a strategy for indexing some intermediate variables that contribute to gradient estimation, significantly reducing the computational complexity of solving the QCLP. We also provide a detailed analysis and comparison of the computation complexity of various ZO methods in Appendix B and Table 4, showing that our ReLIZO method is more computationally efficient than other linear interpolation-based methods. Furthermore, as the sample size $n$ increases, the number of reusable queries also increases with the same step size and reusable distance bound, leading to additional time savings compared to other ZO methods that do not support query reuse. Q3: In Figure 5, the results for the ARGTRIGLS problem indicate a performance drop with any reusable distance bound. What is the specific structure of this problem? Does this suggest that the reuse strategy might be ineffective in certain special cases? A3: ARGTRIGLS is a specialized academic problem designed to evaluate optimization algorithms. It is a variable dimension trigonometric problem in least-squares form, expressed as a sum of $n$ least-squares groups, each containing $n+1$ nonlinear elements. This problem is a variant of Problem 26 in [1], which can be written as follows: Given $f_i:\mathbb{R}^n\rightarrow\mathbb{R}$ $$f_i(x) = n-\sum_{j=1}^n cos(x_j) +i (1-cos(x_i)) - sin(x_i)$$ and initial point $x^0= (\frac{1}{n},\cdots,\frac{1}{n})$, solve $$\min\\\{ \sum_{i=1}^n f_i^2(x):x \in \mathbb{R}^n\\\}$$ In Figure 5, we observe a relative performance drop (slightly slower convergence) when the reusable distance bound $b\geq \eta$, However, this does not imply that the reuse strategy is ineffective in this case. The ARGTRIGLS problem is characterized by a rugged landscape with numerous local minima. In this case, as shown in Figure 4(b), even with $b=\eta$ and a sample size $n=8$, the reuse rate of ARGTRIGLS exceeds 90%, whereas reuse rates for other problems are generally below 50% (e.g. MANCINO with d=100 in Figure 4(a), SROSENBR with d=500 in Figure 2(b)). This high reuse rate indicates that the number of new queries during optimization is relatively small, contributing to the observed decrease in convergence speed. To address this issue, we can employ simple strategies such as reducing the reusable distance bound $b$. The table below illustrates that performance improves as $b$ decreases: \|reusable distance bound\|$b=0$\|$b=0.01\eta$\|$b=0.05\eta$\|$b=0.1\eta$\|$b=0.5\eta$\|$b=\eta$\| \|-\|-\|-\|-\|-\|-\|-\| \|objective value (iter=500)\|0.133\|0.162\|0.298\|0.495\|3.768\|8.158\| \|objective value (iter=2000)\|0.053\|0.063\|0.166\|0.311\|1.834\|3.825\| \|total reuse rate\|0%\|16.6%\|72.4%\|82.3%\|94.7%\|96.8%\| Moreover, setting an upper bound for the reuse rate in each iteration can also help. The following table reports the performance when the maximum reuse rate in each iteration is restricted to 50%. We observe that with $b\leq0.05\eta$ and 2000 iterations, there is minimal performance degradation, even with a reuse rate of approximately 30%: \|reusable distance bound\|$b=0$\|$b=0.01\eta$\|$b=0.05\eta$\|$b=0.1\eta$\|$b=0.5\eta$\|$b=\eta$\| \|-\|-\|-\|-\|-\|-\|-\| \|objective value (iter=500)\|0.133\|0.164\|0.183\|0.255\|0.312\|0.341\| \|objective value (iter=2000)\|0.053\|0.046\|0.054\|0.063\|0.075\|0.074\| \|total reuse rate\|0%\|11.4%\|34.2%\|41.9%\|47.9%\|48.9%\| As discussed in A1, an adaptively adjustable reusable distance bound $b$ aligned with the optimization process would be advantageous in this context. We plan to explore this approach in future work. --- Rebuttal Comment 1.1: Comment: The authors' reply addresses most of my issues. I appreciate the clarification made by the authors. I have no other concerns.	Summary: This study introduces a novel gradient estimation algorithm that operates solely on forward function evaluations. The method employs a Quadratically Constrained Linear Program (QCLP) to determine the optimal linear approximation of sample vectors. The authors present performance enhancement strategies, including sample reuse and efficient inverse matrix computation within the QCLP framework. Empirical evaluations conducted on black-box adversarial attacks and neural architecture search demonstrate the proposed algorithm's superiority over existing zeroth-order methods. Strengths: 1. The proposed method is natural. Approximating the gradient using linear combinations of samples and formulating as the QCLP is an intuitive idea, and the auxiliary techniques employed in this study are both judicious and pertinent to the research objectives. 2. The paper is well-written and easy to follow. Weaknesses: 1. Zeroth-order gradient estimation has a relatively limited impact. While the proposed zeroth-order gradient estimation method demonstrates superiority over existing algorithms in its class, its overall impact on solving underlying optimization problems may be constrained. This limitation is exemplified in the NAS evaluation, where ReLIZO does not consistently achieve optimal performance. 2. According to my interpretation, in ReLIZO algorithm, obtaining new samples from the input space in each iteration is random and arbitrary. I feel there might be more effective strategies to sample new vectors based on known information. Could the authors comment on this? Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful and valuable comments. We hope to address your concerns by answering your questions below. Q1: Zeroth-order gradient estimation has a relatively limited impact. This limitation is exemplified in the NAS evaluation, where ReLIZO does not consistently achieve optimal performance. A1: Thank you for highlighting the inherent limitations of zeroth-order (ZO) methods. Though it is true that ZO methods generally face these constraints, they offer significant advantages over gradient-based methods in black-box optimization scenarios where gradients with respect to variables are not available. Additionally, ZO methods tend to be more memory-efficient, since they do not require backward propagation. This makes them suitable for large-scale tasks such as fine-tuning large language models (LLMs), as demonstrated in recent work [1]. We applied our method to fine-tune an OPT-1.3b model (with 1.3 billion parameters) on the Stanford Sentiment Treebank v2 (SST2) task, and the results are summarized in the table below: \| Optimizer \| Fine-tune Acc \| Fine-tune Memory cost \| \| ------------- \| ------------- \| --------------------- \| \| SGD \| 91.1 \| 44.1 GB \| \| ZO-SGD \| 90.8 \| 28.7 GB \| \| ZO-SGD-Sign \| 87.2 \| 31.4 GB \| \| ZO-Adam \| 84.4 \| 31.4 GB \| \| ReLIZO (ours) \| 93.4 \| 35.7 GB \| Regarding NAS tasks, DARTS and GDAS are two classic gradient-based methods with distinct settings. DARTS trains all parameters in the supernet at each iteration, whereas GDAS samples a sub-network from the supernet at each iteration and trains only the parameters within that sub-network. Our experiments employed DARTS's settings. The results in Table 2 demonstrate that all ZO methods outperform DARTS, and ReLIZO surpasses all other ZO methods, highlighting the effectiveness of our approach. [1] Zhang Y, Li P, Hong J, et al. Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark[C], ICML 2024. Q2: In ReLIZO algorithm, new samples in each iteration is random and arbitrary. There might be more effective strategies to sample new vectors based on known information. A2: Thank you for your suggestion. The ability of ReLIZO to sample new vectors from an arbitrary distribution can be considered one of its strengths compared to other ZO methods. In contrast, smoothing-based ZO methods have to rely on Gaussian or uniform distributions with a mean of zero. We appreciate your insight and introduce an effective sampling strategy inspired by SGD with momentum. Specifically, we propose a sampling momentum strategy where the sampling probability in the current iteration follows a distribution defined as $p(x_t) = \alpha N(g_{t-1}, \sigma) + (1-\alpha)N(0, \sigma)$, with $N()$ denoting a Gaussian distribution, $g_{t-1}$ is the estimated gradient in the last iteration, $\alpha$ as the momentum parameter, and $\sigma$ as the standard deviation. The results of this approach, labeled as ReLIZO-m, are presented in the table below: \| Method \| CIFAR10-valid \| CIFAR10-test \| CIFAR100-valid \| CIFAR100-test \| ImageNet-16-120-valid \| ImageNet-16-120-test \| \| -------- \| ------------- \| ------------ \| -------------- \| ------------- \| --------------------- \| -------------------- \| \| ReLIZO \| 89.50 \| 92.45 \| 69.00 \| 69.03 \| 42.09 \| 42.31 \| \| ReLIZO-m \| 90.71 \| 93.41 \| 70.40 \| 69.63 \| 42.79 \| 43.22 \| These experimental results demonstrate that incorporating an effective sampling strategy can significantly enhance performance. We plan to explore this approach in greater depth in future work. --- Rebuttal Comment 1.1: Title: Thank you for your response Comment: Dear authors, Thank you for your response. I have read it and am glad to learn that a more refined sampling strategy can improve the performance. I have no other concerns.	null	null	Rebuttal 1: Rebuttal: ## General Responses Dear Area Chair and Reviewers, We sincerely thank you for the time and effort you dedicated to the reviewing process. We are delighted that reviewers acknowledged the novelty of rethinking the gradient estimation in ZO method as a QCLP and our resuing strategy (YYyS, VWEi, zPFv), clear motivation and intuitive derived approach (YYyS, zPFv), extensive experiments (VWEi, zPFv) and that the paper is well-written (YYyS, VWEi, zPFv). To further address the comments and questions posed by reviewers, we have also conducted additional analyses and experiments, including: - Thanks to the valuable suggestion on sampling new vectors based on known information proposed by reviewer YYyS, we highlight the strength of ReLIZO in sampling new vectors from an arbitrary distribution compared to other ZO methods. We introduce an effective sampling strategy inspired by SGD with momentum, which brings further performance improvement in NAS task. - As pointed out by reviewer VWEi, the reusable distance bound $b$ is a crucial parameter in the ReLIZO method and can be sensitive to specific problems. To address this issue, we conduct experiments on different reusable distance bounds $b$ and demonstrate that setting an upper bound for the reuse rate in each iteration can significantly reduce performance drops in sensitive cases. - Regarding the relatively limited impact of ZO methods as concerned by reviewers YYyS and zPFv, we demonstrate that ZO methods are well-suited for black-box optimization problems where gradients with respect to the variables are unavailable. Additionally, ZO methods are more memory-efficient than gradient-based methods since they do not require backward propagation, making them applicable to network training, as evidenced by recent works [1,2]. To illustrate the applicability of ReLIZO in network training, we adopted it for fine-tuning an OPT-1.3b model (with 1.3 billion parameters) on the Stanford Sentiment Treebank v2 (SST2) task, following the methodology of [2]. The results indicate that ReLIZO outperforms other ZO methods across various fine-tuning schemes, including full parameter fine-tuning (FT), LoRA, Prefix-tuning, and Prompt-tuning. Notably, ReLIZO even surpasses SGD in the FT scheme while requiring significantly less memory, demonstrating its promising potential. For each reviewer, we have posed a rebuttal addressing the concerns, including the specific results of the additional experiments mentioned above. We look forward to your reply and are more than happy to respond to any further comments. Once again, thank you for your valuable comments and support.   With sincere appreciation and best regards, The Authors   [1] Chen A, Zhang Y, Jia J, et al. DeepZero: Scaling Up Zeroth-Order Optimization for Deep Model Training[C], ICLR 2024. [2] Zhang Y, Li P, Hong J, et al. Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark[C], ICML 2024.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization	Accept (poster)	Summary: It is known that usually deep neural networks will learn “easy examples" that contain fast-learnable features first while learning more complex examples in a second time. The authors argue that mitigating such simplicity bias is the reason method like SAM are outperforming SGD. Based on such analysis, the authors introduce their methods coined as USEFUL that consists in two setups: 1) Identifying the examples with fast-learnable features using a clustering method based on layer output similarity 2) Upsampling by a constant factor the remaining examples with slow-learning features. By doing so, the authors can significantly increase model performances and training time on different classification tasks using different optimizers. They assess their methods across a wide range of dataset and different hyper-parameters and outperform random clustering baseline. Strengths: This paper is well motivated and written. The method seems to be sounded and I really appreciate that the authors assess their method using different hyper-parameters such as optimizer, batch size, datasets, upsampling factor, architectures, and data augmentation. It is also great that they ran a baseline with random clustering. Weaknesses: It is not clear when and why one should choose the last output activation vector to define the clustering instead of intermediate activation vector. It is also not clear at which epoch one should decide to do the clustering since for a dataset like CIFAR10 the optimal performances are achieved at epoch 8 while for CIFAR100 it is epoch 20. So, finding the correct hyper-parameters for the clustering might be costly and thus impact how fast convergence can really be (if we consider this needed additional ablation on clustering epoch). In addition, the authors mention that they are using an upscaling factor of 2, but I am wondering how robust this is when using long-tail distribution. For example, I am not sure that on something like ImageNet-LT or Inaturalist, we will get the best performances by using a constant factor. I would also be a bit more cautious about some of the claims made in the papers. For example, the authors claim that their method is generalizing to OOD tasks while providing experiments on only the WaterBird dataset. So, it would be better to write about promising preliminary results than claiming generalization on OOD. Technical Quality: 3 Clarity: 3 Questions for Authors: 1) Do you think your method will also generalize on Long-Tail dataset while keeping an upscaling factor constant? 2)) Any ideas or heuristics about how to find the optimal epoch/layer to perform the clustering without running an expensive ablation? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors did not really discuss any limitations (outside the fact that their theoretical result does not extend to CNN) or societal impact. I think that one limitation that could have been highlighted is the smaller scale of the experiment and the focus on classification tasks. Another limitation is the lack of results on OOD or long-tail benchmarks which would seem to be well suited for this type of work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the valuable feedback, and acknowledging our well-motivated work and our comprehensive experiments and ablations. 1. When and why one should choose the last output activation vector to define the clustering instead of intermediate activation vector? - Our theorems 3.2 and 3.3 show that examples with at least one fast-learnable feature are learned in early training iterations, and this is well reflected in the model's final output (normalized logit vector). Intuitively, regardless of the layer in which different fast-learnable features are learned, the model’s output for examples containing any fast-learnable feature become similar to its one-hot encoded label. Hence, examples with at least one fast-learnable feature can be well separated based on the model’s final output. While one may be able to separate examples with the same fast-learnable feature by clustering model’s activation at different layers, this is not necessary for our purpose, which is reducing the simplicity bias of early training. In our experiments, we used the model’s final output (output of the softmax), without further tuning the layer number. 2. How to find the separating epoch? - In Fig 14a, we showed that the epoch that can best separate the examples is when the reduction in training error starts to diminish. Intuitively, the first part of the plot with a high negative slope is the time that fast-learnable features are learned due to the simplicity bias. When the training loss curve diminishes the examples with fast learnable features can be best separated from the remaining examples. Figure 14b shows our ablation and confirms that separating examples at multiple epochs around this time (when the slope of the training loss curve diminishes) and upsampling the remaining examples outperform SGD. This observation can help find the separating epoch relatively quickly. 3. How robust is the scaling factor of 2 on long-tail distribution? - The reason for the scaling factor of 2 is to reduce the simplicity bias of early training without further modifying the training data distribution. We conducted new experiments on long tail CIFAR10 with an imbalanced ratio of 10. Our results show that our method with a scaling factor of 2 can indeed improve the generalization performance. Notably, USEFUL outperforms class balancing to address long tail distribution. As can be seen in Fig 1 of the [PDF](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0), interestingly our method may upsample more examples from some of the larger classes (not smaller classes). This confirms that, instead of balancing the training data distribution, it benefits the performance by reducing the simplicity bias and learning features of a class at a more uniform speed. Empirically, as we confirmed in Fig 13 and our new experiments on long-tail data, a scaling factor of 2 works best to reduce the simplicity bias. We note that our method can be stacked with existing methods for long-tail data to balance the training data distribution and further improve the performance, as we confirmed in our new experiments. Table 3: Test error (avg of 3 runs) on long-tailed CIFAR10. \| Ratio \| SGD \| SGD + USEFUL \| SAM \| SAM + USEFUL \| \| --- \| --- \| --- \| --- \| --- \| \| 1:10 \| 10.01 ± 0.21 \| 9.53 ± 0.13 \| 8.85 ± 0.08 \| 8.22 ± 0.04 \| \| Balancing \| 9.77 ± 0.17 \| 9.25 ± 0.11 \| 8.31 ± 0.11 \| 7.93 ± 0.02 \| 4. The authors claim that their method is generalizing to OOD tasks while providing experiments on only the WaterBird dataset. - Prior works have shown the benefits of reduced simplicity bias to improving the OOD performance [Evading the Simplicity Bias, CVPR’22, Overcoming Simplicity Bias, ICML’23, Identifying Spurious Biases Early, AISTATS’24]. Our method reduces the simplicity bias and hence is expected to benefit the OOD performance too. We confirmed this on one of the standard OOD benchmark datasets to not distract the reader from the main contribution of the paper, i.e. improved ID performance. We thank the reviewer and will revise our language as suggested. --- Rebuttal 2: Title: Your feedback is greatly appreciated Comment: We hope our rebuttal has effectively addressed your concerns. As the discussion phase is nearing its conclusion, we are wondering if there's anything further we can clarify or elaborate on. Your feedback is greatly appreciated. --- Rebuttal 3: Comment: I would like to thank the authors for addressing my concerns. I really appreciate the new long tail experiment and the fact that the authors will revise some of the paper's language. I am maintaining my score. --- Rebuttal Comment 3.1: Title: Response to Reviewer Comment: Thank you for reading our rebuttal and keeping a positive view of our paper. We are glad that our rebuttal, especially the new long-tail experiment, has addressed your concerns. We assure that we will modify our language in the revised version.	Summary: This work aims to modify the training data distribution to improve in-distribution generalization. First, the authors theoretically analyse a 2-layer CNN and compare the feature learning dynamics (fast learnable and slow-learnable features) of Gradient Descent (GD) and Sharpness-Aware Minimization (SAM). It is then shown that SAM mitigates simplicity bias compared to GD. The authors then propose USEFUL (UpSample Early For Uniform Learning), a method that upsamples the examples in the training set that contains slow-learnable features. USEFUL first clusters the examples with similar outputs early in the training and then upsamples the slow-learnable clusters. The main idea behind USEFUL is to learn features at a uniform speed (similar to SAM) by changing the training data distribution. USEFUL can be trained with SGD, SAM and SAM + Trivial Augment. Results on CIFAR-10, CIFAR-100, STL10, TinyImageNet indicate that USEFUL is across datasets and architectures. Additonal ablation and analysis show that USEFUL learns similar properties to SAM (for e.g less sharp solutions). Strengths: 1. Originality: The question posed by the authors “Can we change the training data distribution such that the model trained on it has similar properties to SAM?” is interesting and novel. The proposed method is also well-motivated. 2. Results: The authors perform a comprehensive set of ablations and analysis on the proposed method USEFUL. Section 5.4 that shows that USEFUL’s solution has similar properties to SAM, which answers the question raised in the motivation of the paper. I also particularly like the ablations with upweighting loss and data selection method in Appendix D.6. 3. Overall, the paper is fairly well written. One minor point to address here is that the paper covers multiple concepts like SAM, simplicity bias, flat minima and uniform feature learning. It would be good to explain the relationship between these more clearly. Weaknesses: 1. The authors explicitly mention that their focus in this paper is only on “in-distribution generalization”. I am a bit confused by this given the motivation of simplicity bias and learning features uniformly. To elaborate more on this point, - Springer et al [1] also show that SAM implicitly balances the quality of diverse features (similar to the observations made in Section 3 of this paper. The experimental results in [1] is focused more on datasets with multiple predictive features like CelebA, CIFAR-MNIST. - Past work on simplicity bias and shortcut learning [2, 3, 4, 5] has focused on similar datasets like CelebA, Waterbirds, CIFAR-MNIST, Colored-MNIST to name a few. - While the authors have shown encouraging results on Waterbirds dataset in Appendix D5, it would be good to show the complete results on various groups and on other datasets as well. 2. Connection to [1]. Springer et al [1] made a very similar observation as to Section 3 in this paper. It would be great if the authors can clarify the differences with the observations in [1] and this work. Particularly, [1] also shows that SAM mitigates simplicity bias and that SAM learns higher quality representations of hard-to-learn features. The authors briefly discuss this in Related Works section but a more detailed answer would be helpful. 3. I just wanted to understand the practical usefulness of the proposed method. This method has one additional hyperparameter i.e the separating epoch. The authors have reported the best separating epoch for all the datasets which is epoch 8 for CIFAR-10 and epoch 20 for CIFAR-100 (Appendix C.2). How is this hyperparameter chosen? Is there a separating epoch number that works across various datasets? This is especially relevant given that that the average gain on most of the datasets with USEFUL is less than 1% with additional cost for training. [1] Springer, Jacob Mitchell, Vaishnavh Nagarajan, and Aditi Raghunathan. "Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning." The Twelfth International Conference on Learning Representations. [2] Shah, Harshay, et al. "The pitfalls of simplicity bias in neural networks." Advances in Neural Information Processing Systems 33 (2020): 9573-9585. [3] Geirhos, Robert, et al. "Shortcut learning in deep neural networks." Nature Machine Intelligence 2.11 (2020): 665-673. [4] Kirichenko, Polina, Pavel Izmailov, and Andrew Gordon Wilson. "Last layer re-training is sufficient for robustness to spurious correlations." arXiv preprint arXiv:2204.02937 (2022). [5] Teney, Damien, et al. "Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood generalization." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. I did not find the separating epoch used for most of the datasets (except CIFAR-10 and CIFAR-100). Could you please point me to that? 2. What is the takeaway from Figure 1? Please mention the observations regarding the Figure involving the fast and slow learnable features. What kind of examples are usually clustered in fast learnable cluster vs slow learnable cluster, Please refer to the Weakness section for remaining questions. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes, the authors have addressed limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback and recognizing the originality and novelty of our work, and our comprehensive experiments. 1. Our work (ID) vs [1-5] (OOD). - As the reviewers correctly mentioned and we discussed in our [general comment](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0), our work studies ID, while prior work including [1-5] studied OOD with spurious features (CelebA, Waterbirds, CMNIST and CIFAR-MNIST are all benchmark datasets with spurious features in training data but not on the test data. Thus spurious features are not predictive on test). In the OOD setting, simple spurious (non-predictive) features are learned from the training data instead of the predictive features, due to simplicity bias. Training with SAM [1] or methods for alleviating the simplicity bias [2-5] suppresses learning the spurious features and allows learning more/other predictive features. In the ID setting (CIFAR10-100, TinyImageNet), there is no spurious feature and all features are predictive on the test set. Hence, it is not clear why reducing the simplicity bias should help. We proved that reducing the simplicity bias allows learning the same set of predictive features (not new features) at a more uniform speed. This benefits the ID performance. To our knowledge, our study is the first to show the benefits of reduced simplicity bias to ID, hence is a major contribution. Our results on Waterbirds shows that our method also benefits OOD, but is not the main contribution of our work (as we mentioned in lines 338-339). 2. How is the separating epoch chosen? - In Fig 14a, we showed that the epoch that can best separate the examples is when the reduction in training error starts to diminish. Intuitively, the first part of the plot with a high negative slope is the time that fast-learnable features are learned due to the simplicity bias. When the training loss curve diminishes the examples with fast learnable features can be best separated from the remaining examples. Figure 14b shows that separating examples at multiple epochs around this time (when the slope of the training loss curve diminishes) and upsampling the remaining examples outperform SGD. We report the best-performing separating epoch (red points in the figures) in our experiments. - Is there a separating epoch number that works across various datasets? - There is no universal separating epoch. This is reasonable because each data has a different data distribution, i.e. slow- and fast-learnable features, thus, a different theoretical time T in Theorems 3.2 and 3.3. - The average gain on most of the datasets with USEFUL is less than 1% with additional cost for training. - Fig 4, 5 show that SAM’s improvement over SGD is also around 1%. While SAM increases the training time by 2x, and requires tuning the inner step size $\rho$ for different model architectures and datasets, it attracted a lot of attention and is considered a very important contribution as it improves the SOTA performance. Our method is much cheaper (~1.3x cost) than SAM (2x cost) and can be easily stacked with SAM and TA to further improve the SOTA performance. Thus, we believe our contribution is novel and important and can lead to a line of follow up work. Questions: 1. Separating epoch used for all datasets - In Figure 4, the best separating epochs for STL10, CINIC10, and Tiny-ImageNet are 11, 4, and 10, respectively. The separating epoch for Waterbirds is 5. We will add these details into our revision. 2. What is the takeaway from Figure 1? What kind of examples are usually clustered in fast learnable cluster vs slow learnable cluster? - The bottom row in Fig 1 shows examples with fast learnable features that are learned in the first few epochs and are found by our method. We see that these examples are not ambiguous and are clear representatives of their corresponding class, hence are very easy to visually classify (the entire object is in the image and the area associated with the background is small). Such examples are learned early in training due to the simplicity bias. In contrast, the top row shows that examples without any fast learnable feature are harder, to identify visually and look more ambiguous (part of the object is in the image or the object is smaller and the area associated with the background is larger). These examples still contain features that are predictive on the test set, but are learned later during the training (require more iterations to be learned). Our method enables learning both types of examples containing fast-learnable and easy-learnable features at a more uniform speed, which we showed benefits the ID performance. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: I thank the authors for the detailed rebuttal. The authors have clearly answered most of the questions. > reducing the simplicity bias benefits in-distribution (ID). This is an interesting result and a bold claim. Does this mean that past methods that focus on mitigating simplicity bias (and show results in the OOD scenarios) also improve the performance in IID setup by similar margins? The additional simplicity bias (JTT and EIIL) baselines discussed in the rebuttal pdf seem to indicate so. I would love to hear the authors' comments on this. In the future, the authors can consider running experiments with a couple of other methods that mitigate simplicity to verify the generality of the claim. Most of my concerns have been resolved and thus, I have increased my score to 6. --- Rebuttal 2: Title: Response to Reviewer Comment: Thank you for reading our rebuttal, and we’re glad that you found our findings interesting and striking. We also believe that this is an important result and we are quite excited about it. As shown by our theory, we expect prior methods for reducing the simplicity bias to also benefit ID, and our preliminary results with JTT and EIIL showed their promise. Such methods often have several hyperparameters that need to be carefully tuned via a grid search, as we listed for JTT and EIIL in our rebuttal. Our method is directly motivated by our theory and requires tuning only one hyperparameter within a small range (guided by the training loss curve). Hence, we expect it to be more effective in the ID setting. Nevertheless, we agree that adding prior simplicity bias methods would be a nice addition to our work and further support the finding and generality of our result. We thank the reviewer for their valuable suggestion, and we will add experiments with more simplicity bias methods to our revised version.	Summary: - Proves for a 2-layer CNN with fixed second layer weighsts trained on a toy dataset, SAM learns slow-learnable and fast-learnable features more uniformly in the early epochs compared to SGD - Based on this analysis, proposes a simple clustering-based upsampling strategy for reducing simplicity bias / excessive reliance on fast-learnable features. The results show that this improves in-distribution generalization of standard small-scale image classification tasks. Strengths: - Simple easy-to-implement method that uses SAM and upsampling to improve in-distribution generalization - The method is well justified with theoretical analysis comparing SAM and SGD on a toy data distribution. This analysis indicates that SAM is less sensitive to simplicity bias. Weaknesses: - No baselines. There are several papers now that try to reduce simplicity bias in order to improve performance: - https://arxiv.org/abs/2105.05612 - https://arxiv.org/abs/2301.13293 - https://arxiv.org/abs/2107.09044 (does not focus on simplicity bias explicitly, but similar to method proposed in paper) - simpler baselines: there are several papers that propose “example difficulty” metrics (https://arxiv.org/abs/2106.09647). How well do this correlate with the clusters found in your method? If you just train on the k examples with the highest difficulty scores (per class), does this fare worse than the proposed method? - Limited novelty due to findings in [64] (Sharpness-aware minimization enhances feature quality via balanced learning). This paper also shows that SAM improves feature diversity (on real datasets + backed up with analysis on a toy dataset) and improves performance on transfer-learning tasks. - Lacking discussion about when this method would fail. I can imagine two scenarios where the method would not work: 1. Most training examples have one or more slow-learnable features. In this case, the clustering approach would “remove” most of the points in the dataset, and train on very points for multiple epochs. This could result in overfitting and performance that is worse than training. There’s an implicit assumption that there is some sort of one-to-one relation between examples and features. In the case where all examples contain an “easy” (e.g. patch) and a “hard” feature (e.g. CIFAR), would this method improve performance over SGD? 2. In noisy datasets, low-quality examples or mislabeled examples would require more time to learn, and this method would cluster them and train on them for longer. That is, it would group examples that are “high-quality” and hard-to-learn with “low-quality” points. In this case, would the proposed method improve performance over SGD? - “SB of SGD has been long conjectured to be the reason for the superior generalization performance of overparameterized models, by providing capacity control or implicit regularization” This incorrectly cites https://arxiv.org/abs/2006.07710v2, which shows that too much simplicity bias can lead to robustness and in-distribution generalization issues. - Unfair evaluation. The experiments compare SAM+TA augmentation and SAM+USEFUL+TA to SGD (no TA). I think there should be two plots, complaring {SGD, SAM, SAM+USEFUL} w/ and w/o TA. - Experiments on larger datasets. The image classification used here are fairly small-scale. I would like to how well this method scales to ImageNet-scale datasets (TinyImageNet is not a good proxy..) - Writing is repetitive at times, especially the theory section (3.3) Technical Quality: 3 Clarity: 3 Questions for Authors: - How specific is the analysis to the toy distribution setting in which there is just 1 slow-learnable and 1 fast-learnable feature? How do things change if the number of slow-learnable features >> number of fast-learnable features? - Why is max alignment between weight vector and ground-truth feature v_e the right way to evaluate the feature’s contribution to the model? Isn’t it hypothetically possible that SAM solutions rely more on the simpler feature if more weight vectors rely on v_e instead of v_d, even when max-alignment with slow feature is higher for SAM? Some discussion connecting this metric to “feature reliance” vis-a-vis model outputs would be great. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Please see strengths and weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your feedback! 1. Comparison with simplicity bias baselines - Prior work, including papers referred to by the reviewer, showed the benefits of reducing the simplicity bias to out-of-distribution (OOD), where there is a shift between training and test distribution (c.f. [general comment](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0) for detailed discussion). - Our main contribution is to show that reducing the simplicity bias benefits ID. USEFUL is directly motivated by our Theorem 3, and we don’t see alternative methods for reducing the simplicity bias as baselines for our contribution. We conducted new experiments on CIFAR10 with EIIL & JTT. Our theoretically-motivated method outperforms EIIL and JTT in the ID setting (Table 5 in the PDF). 2. Example difficulty metrics vs USEFUL clusters & training on the k examples with highest difficulty scores per class - Methods to calculate example difficulty require either training the model to convergence (forgetting score, Toneva et al, NeurIPS’18) or training several models partially (El2N, Paul et al, NeurIPS’22). USEFUL only requires training one model for as few as 4 epochs. The fast-learnable cluster that we find is correlated with the easy examples, as we confirm in our PDF. But, USEFUL does not need a fine grained calculation of difficulty for all examples. In the method suggested by the reviewer, the optimal choice of K may be different in each class (since the learning difficulty of classes are often different), and finding the best K per class requires extensive hyperparameter tuning. USEFUL automatically finds the optimal choice for K in each class, and is cheap to apply. 3. Novelty w.r.t [64] - As discussed in our [general comment](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0), prior work including [64] studied OOD. [64] showed that SAM suppresses the spurious/redundant (non-predictive) features, and enables learning other/more diverse features which can benefit OOD. They studied toy/real dataset with a known spurious feature and group labels. - In contrast, we consider the in-distribution (ID) setting and prove that SAM learns the same set of predictive features at a more uniform speed, which benefits ID. Unlike [64], we do not require any group labels. To our knowledge, our study is the first to show the benefits of reduced simplicity bias to ID, hence is pretty novel. 4. Would USEFUL fail if: - Most training examples have one or more slow-learnable features - USEFUL finds examples with at least one fast-learnable feature (and arbitrary number of slow-learnable features), and upsample the remaining examples. If most examples have at least one slow-learnable feature, nothing can be concluded. If there is no example with fast-learnable features, there is no simplicity bias to alleviate and there won't be any cluster with low training loss. But, this is unlikely for real-world datasets. - Implicit assumption of one-to-one relation between examples and features. What if all examples contain an easy and a hard feature (e.g. CIFAR)? For simplicity of theoretical analysis, we assumed that every example contains one slow-learnable and at most one fast-learnable. But, such assumption is not required in practice, as we confirmed empirically on several benchmark datasets containing multiple slow and fast learnable features, including CIFAR10/100, TinyImageNet and CINIC10 where there is no one-to-one relation between examples and features. - Noisy labeled data. Our method and analysis consider a clean dataset. But, as we confirmed in our new experiments in Table 4 in the PDF, USEFUL can easily stack with robust methods for learning against noisy labels to reduce the simplicity bias and improve their performance. 5. Comparing {SGD, SAM, SAM+USEFUL} w/ and w/o TA - We compared (i) SGD + USEFUL with SAM, and (ii) SGD, SAM, SAM + TA when they stack with USEFUL to show that USEFUL can stack with different gradient-based optimizers and data augmentations. Our goal is not to compare methods w/ and w/o TA. Our evaluations are fair and confirm the effectiveness of USEFUL. 6. Larger datasets - Please see Q5 of [Reviewer Y69v](https://openreview.net/forum?id=yySpldUsU2&noteId=xu392aAv28). We do not believe there is anything specific to ImageNet that does not work with our method. Questions: 1. If #slow-learnable features >> #number of fast-learnable features - In this case, the dataset is very difficult to learn and learning is not affected much by the simplicity bias. In the extreme case where there is no fast learnable feature in the data, there won't be a low-loss cluster early in training and our method is not applicable (for the right reason). In practice, however, most datasets have several fast-learnable features and our method effectively improves the performance as we confirmed on several benchmark datasets. 2. Why is max alignment between weight vector and ground-truth feature v_e the right way to evaluate the feature’s contribution? - Since the model output of example $x_i$ at iteration t is given by $f(x_i; W) = \sum_{j \in [J]} ( y \beta_d^3 \langle w_j^{(t)}, v_d \rangle^3 + y \beta_e^3 \langle w_j^{(t)}, v_e \rangle^3 + \langle w_j^{(t)}, \xi_i \rangle^3 )$, the term $\beta_e^3 \max_{j \in [J]} \langle w_j^{(t)}, v_e \rangle^3$, i.e., max alignment between weight vector and the ground-truth feature $v_e$, greatly affects the prediction when the fast-learnable feature is present. 3. Isn’t it possible that SAM solutions rely more on the simpler feature if more weight vectors rely on v_e instead of v_d? - Indeed, SAM relies more on $v_e$ than $v_d$ in early training as proved in our Theorem 3.3. We showed that (1) SAM learns features at a more uniform speed as indicated by a smaller gap between features’ contribution in Theorem 3.4; and (2) SAM relies on $v_d$ relatively (w.r.t $v_e$) more than SGD in Theorem 3.5. --- Rebuttal 2: Title: Your feedback is greatly appreciated Comment: We hope our rebuttal has effectively addressed your concerns. As the discussion phase is nearing its conclusion, we are wondering if there's anything further we can clarify or elaborate on. Your feedback is greatly appreciated.	Summary: This paper proposes an algorithm for changing the distribution of training data to improve the generalization of the model on origin data distribution. The paper is inspired by Sharpness Aware Minimization, which aims at finding a flat minimum meaning that it has a good generalization capability. This paper divides features into two categories: fast-learnable features and slow-learnable features and derives some observations like "SGD and SAM only learn fast-learnable or easy features early in training" and "SAM learns slow-learnable and fast-learnable features at a more uniform speed". The authors propose the method dubbed as USEFUL to train the model on some slow-learnable features repeatedly. The experiments show the effectiveness of USEFUL on CIFAR10 and CIFAR100 datasets. Strengths: - The paper is well-written and easy to follow. - The paper has a theoretical analysis to analyze the learning progress and derive the proposed method. - The experiments are abundant and comprehensive. Weaknesses: There are some questions based on the presentation of this paper, I will not hesitate to improve my score if the following question are solved. - Difference between this paper and methods for long-tailed data distribution or measuring the difficulty of learning examples. Algorithms for long-tailed data distribution are usually based on resampling training data or reweighing loss value. The proposed USEFUL is similar to the resampling methods except that USEFUL focuses on the features that are hard/slow to learn. Some references for understanding: [Shi, Jiang-Xin, et al. "How re-sampling helps for long-tail learning?." Advances in Neural Information Processing Systems 36 (2023).](https://arxiv.org/pdf/2310.18236), [Shrivastava, Abhinav, Abhinav Gupta, and Ross Girshick. "Training region-based object detectors with online hard example mining." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.](https://arxiv.org/pdf/1604.03540v1) and some references based on it, [A Re-Balancing Strategy for Class-Imbalanced Classification Based on Instance Difficulty](https://openaccess.thecvf.com/content/CVPR2022/papers/Yu_A_Re-Balancing_Strategy_for_Class-Imbalanced_Classification_Based_on_Instance_Difficulty_CVPR_2022_paper.pdf), [Active Teacher for Semi-Supervised Object Detection](https://openaccess.thecvf.com/content/CVPR2022/papers/Mi_Active_Teacher_for_Semi-Supervised_Object_Detection_CVPR_2022_paper.pdf), I believe a discussion of these references in paper should be helpful. - The relation between the proposed USEFUL and SAM? It seems like the motivation of USEFUL is changing the data distribution to get a flat minimum like SAM. But the results in Appendix D.2, i.e., 53.8 for SGD 41.8 for SGD+USEFUL 12.4 for SAM in Table 1($\lambda_{max})$, do not show effectiveness compared with SAM. It could show the effectiveness on SGD but it's far from being comparable to SAM. Some small questions: - What's the exact formulation of the Data distribution? - What's the "patch" meaning in Definition 3.1? Is that the same as the patch in ViT or the channel of the image? It's a little confusing. - The experiments mainly focus on traditional architecture, e.g., n-Layer CNN, ResNet. More experiments on popular models and big datasets, e.g., Transformer ImageNet-1k, would be better. Technical Quality: 4 Clarity: 3 Questions for Authors: See Weakness. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: N.A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the valuable feedback and acknowledging our theoretical results and comprehensive experiments. We discuss the questions below. 1. USEFUL vs resampling methods for long-tail data & example difficulty. - As discussed in the [general comment](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0), long-tail data is an instance of OOD. Long-tail methods resample the data at the class or subclass level to match the training and test distribution. For example, [A Re-Balancing Strategy, CVPR’22] identifies instances with slow learning speed as more difficult instances and dynamically increases their weights during training to effectively change the data distribution to match the test distribution [page 2]. In contrast, we showed in the ID setting that the simplicity bias of gradient-based methods makes them find solutions with suboptimal performance. USEFUL reduces the simplicity bias by finding examples that are not learned during the first few training epochs, upsamples the rest of examples once, and restart training. Hence, USEFUL does not require calculating the difficulty of examples or do dynamic instance-wise reweighting or resampling during the training. Hence, the effectiveness of our method is attributed to learning features at a more uniform speed, and not matching the training and test data distributions. To our knowledge, the benefit of reduced simplicity bias to ID is studied for the first time by our work. - We conducted new experiments on long-tail CIFAR10 with an imbalance ratio of 10. Table 3 in the [PDF](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0) shows that our method can also improve the performance of SGD and SAM, by reducing the simplicity bias on the long-tail data. Fig 1 in the [PDF](https://openreview.net/forum?id=yySpldUsU2&noteId=nTUcV9AnO0) shows the distribution of classes before and after upsampling by USEFUL. Interestingly, we see that USEFUL upsamples more examples from some of the larger classes and still improves the accuracy on the balanced test set. This improvement is attributed to the more uniform speed of feature learning, and not balancing the training data distribution. Notably, USEFUL outperforms class balancing to address long tail distribution. Besides, USEFUL can be stacked with methods to address long tail data to further improve the performance, as we confirmed in our new experiments (please refer to Q3 of [Reviewer F4yx](https://openreview.net/forum?id=yySpldUsU2&noteId=3xcoRIZ69p)). 2. USEFUL’s solution is not as flat as SAM. - The motivation of USEFUL is not to get a flatter minima like SAM. In fact, a flatter solution has been only conjectured (not proven) to have a better performance, and recent studies have shown conflicting evidence on the relationship between flatness and generalization, suggesting that flatness does not fully explain SAM’s success [16 in paper, When do flat minima optimizers work?, NeurIPS’22, A modern look at the relationship between sharpness and generalization, ICML’23]. A key contribution of our work is identifying *an orthogonal effect of SAM that is beneficial in-distribution (ID):* we proved that SAM has less simplicity bias early in training and thus learns features at a more uniform speed. - Our method, USEFUL, is inspired by SAM in reducing the simplicity bias to improve the ID generalization performance. Due to the same reason, USEFUL can effectively improve the SAM’s performance itself, as we confirmed empirically. While our goal is not to directly find a flatter minima or the same solution as SAM, we showed that USEFUL can find flatter minima. Note that to capture the sharpness/flatness, multiple different criteria have been proposed (largest Hessian eigenvalue and bulk of Hessian), and one criteria is not enough to accurately capture the sharpness. Table 1 shows that while the solution of SGD + USEFUL has a higher largest Hessian eigenvalue, it achieves the smallest bulk. Notably, the solution found by both SGD + USEFUL and SAM have lower sharpness in both indicators than SGD, proving that USEFUL successfully reduces the sharpness of the solution. 3. Formulation of the Data distribution. - The data distribution is defined in Definition 3.1 in Section 3.1. Each data point consists of 3 different patches, a slow-learnable, a fast-learnable, and random noise. Our data model is consistent with many prior theoretical works [2, 7, 10, 11, 15, 29, 37 in paper]. 4. "patch" meaning in Definition 3.1. - It is similar to the patch in ViT. Each patch is a small region in the image and each image consists of P patches. 5. More experiments on popular models and big datasets, e.g., Transformer ImageNet-1k - Both CINIC10 and Tiny ImageNet contain images from ImageNet datasets and have a relatively large number of examples. CINIC10 dataset consists of 270,000 images, which is 4.5 times larger than CIFAR10. Tiny-ImageNet comprises 100,000 images distributed across 200 classes of ImageNet. Due to the computational constraints, we cannot conduct experiments on ImageNet (on our NVIDIA RTX A5000, a single training of ResNet18 with SGD on ImageNet takes around 1068 GPU hours). In Fig 5, we confirmed the effectiveness of USEFUL on other architectures such as ViT-small, which is a popular architecture for computer vision tasks [69 in paper, When vision transformers outperform resnets without pre-training or strong data augmentations, ICLR’22]. --- Rebuttal 2: Title: Your feedback is greatly appreciated Comment: We hope our rebuttal has effectively addressed your concerns. As the discussion phase is nearing its conclusion, we are wondering if there's anything further we can clarify or elaborate on. Your feedback is greatly appreciated.	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable feedback and recognizing the originality of our work. We’d like to first briefly emphasize the scope and contribution of our work: - Our work shows, for the first time, that reducing the simplicity bias benefits in-distribution (ID). Previously, the benefits of reducing simplicity bias have been only shown to out-of-distribution (OOD), in particular to mitigate spurious correlations. - Prior methods: OOD setting [long-tail data and spurious correlations]. Here, there is a shift between the training and test data distributions: - Long-tail data [Long-tailed CIFAR10, Inaturalist, ImageNetLT]: (sub)classes are highly imbalanced in training but are (more) balanced in test data (distribution shift). Upsampling the minority (sub)classes in training data improves the test performance on corresponding groups that are (much) larger in test data. Here, the benefit comes from balancing the training data and matching the training and test data distributions. - Spurious features [Waterbirds, CelebA, CMNIST, CIFAR-MNIST]: spurious features are simple features with a high correlation with a label at training time but not at test time (distribution shift). Due to simplicity bias, spurious features are learned instead of the more complex predictive features. This yields a poor (worst-group) accuracy on examples without the spurious feature at test time. Reducing the simplicity bias (or training with SAM) suppresses learning spurious features. Here, the benefit comes from mitigating the spurious feature to learn more/other features. - Our method: ID setting [CIFAR10-100, TinyImageNet, CINIC10]. Here, training and test data have the same distribution, there is no spurious feature in the training data, and all features are predictive on the test set. Our work shows for the first time that reducing the simplicity bias benefits ID. Our contribution is orthogonal to the above work: *we proved that reducing the simplicity bias allows learning the predictive* features at a more uniform speed, and showed that this benefits ID.** Unlike the OOD setting, the benefit of reducing the simplicity bias in the ID setting is not attributed to balancing the training data distribution or suppressing the spurious features to learn other features. It is attributed to the (more uniform) speed/time of learning the same set of predictive features. Our findings are more unexpected than its counterpart in the OOD setting and opens up a new direction for future research. - Our method: OOD setting [Long-tailed CIFAR10, Waterbirds]. While our theory and main contribution is to show the benefits of reduced simplicity bias to ID, our new experiments show that our method can also benefit the long-tail setting, by reducing the simplicity bias and learning features of a class at a more uniform speed. Benefits of reducing the simplicity bias to the spurious setting has been studied by several prior work. While our experiments on Waterbirds shows that our method is also applicable to this setting, we note that this is not our main contribution. We provide new experimental results on example difficulty, long-tail data, noisy labels, and spurious baselines in the attached PDF. Pdf: /pdf/79d0755211905a72275f005ab7202d5686157ad2.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
REBEL: Reinforcement Learning via Regressing Relative Rewards	Accept (poster)	Summary: This paper reduces the complex policy optimization procedure of alignment to a simple regression objective, using the relation between optimal policy and reward. The paper conduct detailed theoretical analysis in revealing the relation between the proposed algorithm REBEL and NPG/MD. Comprehensive experiments in both text and image generation exhibit the effectiveness of REBEL. Strengths: 1. This paper studies simplified version of policy optimization in RLHF (compared to PPO), which is a research topic of interest. 2. The theoretical analysis of REBEL is detailed and insightful. 3. The presentation of this paper is logically clear and has good readability. 4. The experiments in this paper are comprehensive, and the experimental results are well presented. Weaknesses: 1. The statement "REBEL ... be extended to handle intransistive preferences ...." in the abstract is not adequately presented in the main content of the paper. As the major influence brought by intransistive preferences is the degradation of reward score accuracy, which is not addressed by this paper. 2. I would suggest the authors to summarize the limitations of the proposed method in a separate "Limitations" section. Technical Quality: 3 Clarity: 3 Questions for Authors: none Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: none Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review of our paper. > The statement "REBEL ... be extended to handle intransitive preferences ...." in the abstract is not adequately presented in the main content of the paper. As the major influence brought by intransistive preferences is the degradation of reward score accuracy, which is not addressed by this paper. We appreciate your observation and agree with your assessment. Since our focus is not on the preference model itself, we did not conduct experiments specifically targeting preference models. However, we provided a method to extend REBEL to preference models. The extension of REBEL to address intransitive preferences is discussed in detail in Appendix D, with further analysis provided in Appendix G. How to pre-train a good preference model without any degradation of reward score accuracy is beyond the scope of our paper. > I would suggest the authors to summarize the limitations of the proposed method in a separate "Limitations" section. Thank you for the suggestion. We will include a dedicated "Limitations" section in the next version of the paper. --- Rebuttal Comment 1.1: Comment: Thank you for your response, my concerns have been addressed during the rebuttal and I decide to keep my score.	Summary: This paper proposes the REBEL algorithm that reduces policy optimization to iteratively solving squared loss regression problems on the difference in rewards between trajectories, based on DPO's analysis. The paper transforms the resulting equation for r(x, y) presented in DPO to a regression loss function, and avoids the intractable calculation of Z(x) by calculating the loss based on a pair of samples from the same input prompt x, i.e., (x, y) and (x, y'). One of the goals for REBEL is to serve as a simple and lightweight RL algorithm that eliminates the need for complex components like value functions and clipping heuristics used in PPO. The authors provide a theoretical analysis showing that Natural Policy Gradient can be seen as a special case of REBEL under some assumptions. The authors conduct two kinds of empirical analysis including language modeling and image generation tasks to demonstrate the performance of REBEL. Strengths: - Originality: - This paper presents a new angle by transforming the analysis of the reward function presented in the DPO paper into a reward regression loss, leading to the proposed REBEL algorithm. - The authors make connections between REBEL and existing RL methods like NPG considering some assumptions, showing that these algorithms can be seen as special cases or approximations of REBEL under certain conditions. - Quality: - The paper provides a thorough theoretical analysis comparing REBEL with existing RL approaches. - Clarity: - The paper is well-written and easy to understand, with a clear logical flow from motivation to theoretical analysis to empirical validation. The authors do an good job of explaining the intuition behind REBEL and highlighting its connections to prior work. - Significance: - The paper tackles the important problem of developing simpler and more efficient RL algorithms that can scale to large-scale generative model fine-tuning. Weaknesses: 1. Insufficient experimental validation and limited baseline comparisons: - While the paper presents empirical results on language modeling and image generation tasks, the experimental validation of REBEL could be more comprehensive. The authors should consider including a wider range of benchmarks and datasets to demonstrate the generality and robustness of their approach. - The comparison with baseline algorithms like PPO and DPO is somewhat limited. The authors should provide more details on the hyperparameter settings and training procedures for the baselines to ensure a fair comparison. Moreover, the poor performance of DPO compared to PPO in the experiments raises questions about the implementation or hyperparameter choices. - The authors claim that REBEL matches the strongest known theoretical guarantees in terms of convergence and sample complexity. However, the experiments only compare performance at a specific epoch without demonstrating improved sample efficiency. Convergence plots showing the performance of REBEL and baselines over the course of training would provide a clearer picture of the sample efficiency and convergence properties. 2. Lack of support for certain claims and limited exploration of key aspects: - The paper makes several claims regarding the advantages of REBEL, such as its ability to handle intransitive preferences, incorporate offline datasets, and apply to deterministic MDPs. However, there is a lack of corresponding experimental evidence or theoretical analysis to substantiate these claims. - The relationship between the regressor's performance and the quality of the dataset used for training is not explored in depth. Insights or experiments that investigate how dataset quality and diversity affect the regressor's ability to capture an improved policy would strengthen the paper. - The choice of base distribution \mu is mentioned as a determining factor for whether REBEL is hybrid or fully online. However, the paper does not provide experimental results comparing different forms of \mu across various tasks or practical guidelines for choosing \mu in real-world applications. 3. Inconsistencies and potential conflicts with previous statements: - The authors mention that critic-based variance reduction might be necessary for high-variance trajectory-level rewards in stochastic MDPs, which seems to contradict the criticism of PPO's complexity in the introductory section. The lack of experimental support for REBEL's performance in stochastic MDPs is a significant limitation, and the authors should provide preliminary results or theoretical insights to support their claims. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Sample efficiency and convergence guarantees: - The authors claim that REBEL matches the strongest known theoretical guarantees in terms of convergence and sample complexity. However, the experiments only compare performance at a specific epoch without demonstrating improved sample efficiency. Can the authors provide experimental results that support their claim of improved sample efficiency compared to other algorithms? - It would be helpful to see convergence plots that show the performance of REBEL and baseline algorithms over the course of training, rather than just at a selected epoch. This would provide a clearer picture of the sample efficiency and convergence properties of REBEL. 2. Relationship between regressor performance and dataset quality: - The authors state that a regressor that can predict the difference in rewards between trajectories implicitly captures an improved policy. Is the performance of this regressor dependent on the quality of the dataset used for training? How does the quality of the dataset affect the regressor's ability to capture an improved policy? - Can the authors provide insights or experiments that explore the relationship between dataset quality and the effectiveness of REBEL? 3. Applicability to deterministic MDPs: - The authors mention that REBEL can be applied to any deterministic MDP where the initial state is x and the trajectory y consists of a sequence of actions. Is there any experimental or theoretical support for this claim? - It would strengthen the paper if the authors could provide empirical results or theoretical analysis that demonstrates the effectiveness of REBEL in deterministic MDPs beyond the bandit formulation. 4. Choice of base distribution \mu: - The authors state that the choice of base distribution \mu determines whether REBEL is hybrid or fully online. Can they provide experimental results that compare different forms of \mu across various types of tasks? What are the practical guidelines for choosing \mu in real-world applications? - Insights into the impact of different choices of \mu on the performance and behavior of REBEL would be valuable for practitioners looking to apply this algorithm. 5. Stochastic MDPs and the need for critic-based variance reduction: - The authors leave the experimental validation of REBEL in stochastic MDPs for future work but mention that trajectory-level rewards can be high-variance, potentially requiring critic-based variance reduction. In what practical situations would the transition dynamics be stochastic? If critic-based variance reduction is needed, how does this align with the introductory section's criticism of PPO's complexity? - The lack of experimental support for REBEL's performance in stochastic MDPs is a significant limitation. Can the authors provide any preliminary results or theoretical insights that support their claims about REBEL's applicability to stochastic environments? 6. Performance comparison with baselines: - In the experiments conducted by the authors, DPO performs significantly worse than PPO, especially in Table 1, where DPO is inferior in every case. Can the authors provide an explanation for this discrepancy? Is it due to differences in implementation or hyperparameter settings? - In Figure 3, the comparison between PPO and REBEL is made at an intermediate checkpoint where REBEL observes a higher reward under the reward model. Is it possible that PPO has already overfit at this selected epoch? How was this specific epoch number chosen for REBEL? What would the comparison look like if the best-performing epoch for each algorithm were considered? Additionally, why is the comparison limited to only PPO? It would be informative to include other state-of-the-art RL algorithms in the comparison to better understand the relative performance of REBEL. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your constructive feedback. We address each of your points below. > Insufficient experimental validation and limited baseline comparisons > Performance comparison with baselines Our experimental section is comprehensive compared to previous works on RLHF [1, 2], incorporating a general chat dataset and an image generation task to evaluate the robustness and generality of REBEL. We compared REBEL with several algorithms, including DPO, PPO, iterative DPO, REINFORCE, RLOO, and APA. Detailed hyperparameter settings are provided in Appendix H.1.4, H.2.3, H.3.3. The lower performance of DPO compared to PPO is not due to implementation or hyperparameter choices. On TL;DR, our DPO results are better than the ones reported in [3]. Results for PPO, REINFORCE, and RLOO are directly obtained from prior papers [2, 4] which are exclusively focusing on these algorithms on TL;DR. For general chat, we directly report winrates from the released starling-alpha model by APA's authors. Thus, we believe that our comparison to baselines is fair. In Figure 3, the comparison at an intermediate checkpoint is intended to highlight the sample efficiency of REBEL. There is no indication that PPO has overfitted at this epoch, as the image quality of PPO continues to improve afterward. > Sample efficiency and convergence guarantees For the image generation task, as illustrated in Figure 4, REBEL converges faster during the initial training phase and eventually achieves performance comparable to that of PPO. In addition, we plot the reward vs. step for TL;DR dataset in the pdf of the global rebuttal. The plot demonstrates REBEL's faster convergence compared to iterative DPO and PPO. > Lack of support for certain claims and limited exploration of key aspects > Applicability to deterministic MDPs > Choice of base distribution \mu Since our focus is not on the preference model itself, we did not conduct experiments specifically targeting preference models. However, we provided a method to extend REBEL to preference models in Appendix D, with further analysis provided in Appendix G. In our experiments, we found that setting $\mu=\pi_t$ yielded better results. We attribute this to the lower quality of the offline dataset. In TL;DR summarization and Ultrafeedback, our trained policies can generate better responses than the ones in the datasets. This is shown in Table 1, where the 2.8B and 6.9B models can easily reach high winrates compared to offline human demonstrations. In this case, setting $\mu$ to $\pi_{ref}$ or the offline data does not help significantly. Therefore, we use $\pi_t$ as $\mu$ in all experiments. Since transitions are deterministic, we could formulate this as a bandit problem where $x$ is the initial state and $y$ is the trajectory consisting of a sequence of actions [1, 2, 10]. Applying token generations in LLMs and image generation to deterministic MDPs has been proposed in prior RLHF works [5, 6, 7, 8]. We adapt this bandit setup to make a fair comparison to the previous works. > Inconsistencies and potential conflicts with previous statements > Stochastic MDPs and the need for critic-based variance reduction Lack of experimental support in stochastic MDPs is not a significant limitation as deterministic MDPs have wide applications in LLMs and image generations, which are high-impact areas. We emphasize that all prior RLHF works [1, 2, 4, 6] all focus on deterministic MDP. There is also no conflict in our criticism of PPO's complexity. Our paper focuses on deterministic transitions; in this context, many prior work have argued that PPO and critic-based variance reduction methods are not necessary [2, 8, 9]. We agree that in highly stochastic settings, critics might still be necessary, which is indeed stated in the paper. > Relationship between regressor performance and dataset quality: A high-quality dataset, i.e. a dataset that has diverse coverage, would increase the generalization ability of the learned regressor. E.g., if the training distribution is diverse, then the learned regressor can achieve better performance under the comparator policy $\pi^$ distributions, which, as our theory indicates, will ensure convergence to $\pi^$. Our work focuses on online RL and studying the dataset quality is not the main focus of the paper. [1] Rafailov R, Sharma A, Mitchell E, Manning CD, Ermon S, Finn C. Direct preference optimization: Your language model is secretly a reward model. [2] Ahmadian A, Cremer C, Gallé M, Fadaee M, Kreutzer J, Üstün A, Hooker S. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. [3] Rafailov R, Chittepu Y, Park R, Sikchi H, Hejna J, Knox B, Finn C, Niekum S. Scaling laws for reward model overoptimization in direct alignment algorithms. [4] Huang S, Noukhovitch M, Hosseini A, Rasul K, Wang W, Tunstall L. The N+ Implementation Details of RLHF with PPO: A Case Study on TL; DR Summarization. [5] Ramamurthy R, Ammanabrolu P, Brantley K, Hessel J, Sifa R, Bauckhage C, Hajishirzi H, Choi Y. Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization. [6] Stiennon N, Ouyang L, Wu J, Ziegler D, Lowe R, Voss C, Radford A, Amodei D, Christiano PF. Learning to summarize with human feedback. [7] Chang JD, Shan W, Oertell O, Brantley K, Misra D, Lee JD, Sun W. Dataset reset policy optimization for rlhf. [8] Black K, Janner M, Du Y, Kostrikov I, Levine S. Training diffusion models with reinforcement learning. [9] Oertell O, Chang JD, Zhang Y, Brantley K, Sun W. Rl for consistency models: Faster reward guided text-to-image generation. [10] Wu T, Zhu B, Zhang R, Wen Z, Ramchandran K, Jiao J. Pairwise proximal policy optimization: Harnessing relative feedback for llm alignment. --- Rebuttal Comment 1.1: Comment: Thank the authors for the rebuttal. However, I still find some of the responses to my questions to be somewhat evasive, and I would appreciate more detailed explanations from the authors. I would also be happy to increase the score if all my concerns are adequately addressed. 1. Regarding DPO's performance: The authors mention that DPO performs better than reported in [3] for the TL;DR task. However, why is this improvement limited to the TL;DR task and not observed in other tasks? Could you provide a more comprehensive explanation for DPO's underwhelming performance in other areas? 2. On REBEL's convergence guarantee: While the plots empirically demonstrates REBEL's faster convergence, the paper's main argument centers on theoretical explanations. Could you provide a corresponding theoretical justification for this faster convergence to complement the empirical evidence? 3. Concerning stochastic MDPs: Given the limited analysis provided for stochastic MDPs, I'm curious about the rationale for including this section in the main text, since there are not sufficient empirical support to this added part. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for the response and address each of the points below. > Regarding DPO's performance: The authors mention that DPO performs better than reported in [3] for the TL;DR task. However, why is this improvement limited to the TL;DR task and not observed in other tasks? Could you provide a more comprehensive explanation for DPO's underwhelming performance in other areas? In our experiments, we only performed experiments with DPO on the TL;DR summarization task. The improvement in performance regarding summarization could be due to differences in prompt, test time model temperature, or other factors. In particular, [3] used a temperature of 1.0 during the evaluation, while we used a temperature of 0.9. We referenced [3] to demonstrate that our DPO results are consistent with those reported by other independent researchers, indicating that our findings of DPO are reasonable within the existing literature. > On REBEL's convergence guarantee: While the plots empirically demonstrates REBEL's faster convergence, the paper's main argument centers on theoretical explanations. Could you provide a corresponding theoretical justification for this faster convergence to complement the empirical evidence? We provide a detailed analysis of REBEL's convergence rate in lines 140-147 and 200-217 of our paper. Specifically, we show that REBEL achieves a fast $1/T$ convergence under the assumption that the least square regression optimization returns the exact Bayes optimal solution. We then relax this assumption using a regression generalization bound, resulting in an agnostic regret bound with a convergence rate of $1/\sqrt{T}$. This agnostic regret bound represents the strongest type of agnostic learning results known in the RL literature. On the other hand, to the best of our knowledge, we do not know if PPO — our baseline in the image generative model experiment, has provable convergence guarantees. One reason that PPO convergence can be slower than REBEL is that PPO uses clipping to approximately maintain conservative policy update and the clipping operator throws away a non-trivial amount of training data. REBEL on the other hand does not use clipping and thus does not waste any training data. > Concerning stochastic MDPs: Given the limited analysis provided for stochastic MDPs, I'm curious about the rationale for including this section in the main text, since there are not sufficient empirical support to this added part. This point was also raised by Reviewer HFsB. We fully agree with this feedback and will move this section to the appendix.	Summary: This work presents REBEL, a minimalist reinforcement learning algorithm that does policy optimization by solving a sequence of regression problems using relative rewards as targets. Theoretical analysis shows that Natural Policy Gradient (NPG) is a variant of REBEL, and thus theoretical guarantees for NPG can be applied to REBEL. Experimental results show that REBEL matches or outperforms existing baselines, most notably PPO and RLOO, on multiple tasks. Strengths: - The paper is well-organized and technically sound. The general flow of the paper is smooth and proposed methods are explained adequately. The paper has an appropriate number of citations and properly details existing work in the related work section. - The method is simple to implement and has little engineering overhead. Given the minimalist implementation, the results are impressive, surpassing even PPO, which typically requires significant engineering. Weaknesses: - There are no significant weaknesses in this work, barring some clarifying details. - I believe that at least a brief section on related work should be included in the main paper, the in-depth one can be deferred to the appendix. In terms of space, I personally do not think Section 2.2 adds much value to the main paper. Technical Quality: 3 Clarity: 3 Questions for Authors: - The reward model becomes increasingly off-distribution as the policy is updated. Although it is standard practice to keep reward models fixed even with iterative methods, prior works generally use it to generate preference labels between pairs of outputs. Since this work uses the difference of scores as the regression target, the off-distribution reward scores might have a greater impact here. Concisely, how significant a problem is reward model over-optimization [1] for REBEL? - It would be interesting to see and understand the differences between reward-weighted regression baseline (RWR) and REBEL as they have some close connections. - Is there an optimal choice of $\mu$ ? What are the intuitive differences between using the $\mu= \pi_{ref}$ and $\mu= \pi_{t}$ ? As the policy improves, samples $y,y’ \sim \pi_{t}$ are in the high reward region, and it can be difficult to separate them since these might be off-distribution for the reward model. Given these constraints of the reward model, there might be better choices of $\mu$ that allow for better prediction of score differences. It would be interesting to see an ablation study on this, or a well-reasoned answer that explains the tradeoffs between different choices of $\mu$. - Why are datasets not aggregated? Instead, only the most recently collected dataset is used for training. [1] : Gao, L., Schulman, J., & Hilton, J. (2023, July). Scaling laws for reward model overoptimization. In International Conference on Machine Learning (pp. 10835-10866). PMLR. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable review of our paper. We respond to your individual questions below. > I believe that at least a brief section on related work should be included in the main paper, the in-depth one can be deferred to the appendix. In terms of space, I personally do not think Section 2.2 adds much value to the main paper. Thank you for your suggestion. We will include a brief section on related work in the main paper and have a detailed discussion in the appendix. > How significant a problem is reward model over-optimization [1] for REBEL? In the context of RLHF for LLMs, we choose reward models that are highly ranked on the Reward Bench [2]. These reward models are trained using extensive datasets that include responses generated from a variety of different policies [3]. This diversity makes it challenging to over-optimize a reward model, even for methods such as REBEL. In addition, following previous work [4], we apply an addition KL penalty to the reward, $r(x, y) = RM(x, y) - \gamma (\ln \pi_{\theta_{t}}(y\|x) - \ln \pi_{\theta_{0}} (y\|x))$, to prevent over-optimization for both REBEL and the baseline methods. In the image generation setting, following previous work [5, 6], no KL regularization is used. We observe that both PPO and REBEL over-optimize the reward model towards the end of the training. They tend to ignore the input prompt and generate the same image that maximizes the reward model score. For fair comparison to [6], we also did not include KL. > It would be interesting to see and understand the differences between reward-weighted regression baseline (RWR) and REBEL as they have some close connections. RWR can be understood as reward-weighted imitation learning. Prior work on RL for diffusion models shows that RWR is not as effective as RL (PPO in that case) [5]. REBEL, on the other hand, is a full RL algorithm from its connection to NPG where it generalizes NPG. > Is there an optimal choice of $\mu$? What are the intuitive differences between using the $\mu=\pi_{ref}$ and $\mu=\pi_{t}$?. Setting $\mu$ to $\pi_{ref}$ or another distribution can encourage exploration, especially when the policy is still learning and the quality of the generations is not yet optimal. This approach can help in discovering diverse and potentially higher-reward samples that the current policy might not generate. In our experiments, we found that setting $\mu=\pi_t$ yielded better results. We attribute this to the lower quality of the offline dataset. In TL;DR summarization and Ultrafeedback, our trained policies can actually generate better responses than the ones in the datasets. This is shown in Table 1, where the 2.8B and 6.9B models can easily reach high winrates compared to offline human demonstrations. In this case, setting $\mu$ to $\pi_{ref}$ or the offline data does not provide significant benefits. Therefore, we use $\pi_t$ as $\mu$ in all our experiments. > Why are datasets not aggregated? Instead, only the most recently collected dataset is used for training. Our approach can be understood as online RL, e.g. PPO, where only on-policy data is used for update. We could aggregate the datasets which makes the approach more off-policy. While we have not tested this in our experiments, it could be an interesting direction to explore if being off-policy would lead to better sample efficiency. In addition, previous RLHF works [4, 8, 9] also use batches collected by the current policy (i.e. online batch), with each batch containing a new set of prompts. To ensure a fair comparison with previous methods, we primarily focus on the on-policy approach. [1] Gao, L., Schulman, J., & Hilton, J. (2023, July). Scaling laws for reward model overoptimization. In International Conference on Machine Learning (pp. 10835-10866). PMLR. [2] Lambert N, Pyatkin V, Morrison J, Miranda LJ, Lin BY, Chandu K, Dziri N, Kumar S, Zick T, Choi Y, Smith NA. Rewardbench: Evaluating reward models for language modeling. arXiv [3] Xiong, W., Dong, H., Ye, C., Wang, Z., Zhong, H., Ji, H., Jiang, N., Zhang, T. Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint. 2024. arXiv [4] Huang S, Noukhovitch M, Hosseini A, Rasul K, Wang W, Tunstall L. The N+ Implementation Details of RLHF with PPO: A Case Study on TL; DR Summarization. [5] Black K, Janner M, Du Y, Kostrikov I, Levine S. Training diffusion models with reinforcement learning. arXiv [6] Oertell O, Chang JD, Zhang Y, Brantley K, Sun W. Rl for consistency models: Faster reward guided text-to-image generation. arXiv [7[ Korbak T, Shi K, Chen A, Bhalerao RV, Buckley C, Phang J, Bowman SR, Perez E. Pretraining language models with human preferences. In International Conference on Machine Learning 2023 Jul 3 (pp. 17506-17533). PMLR [8] Stiennon N, Ouyang L, Wu J, Ziegler D, Lowe R, Voss C, Radford A, Amodei D, Christiano PF. Learning to summarize with human feedback. Advances in Neural Information Processing Systems. 2020;33:3008-21. [9] Ahmadian A, Cremer C, Gallé M, Fadaee M, Kreutzer J, Üstün A, Hooker S. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. --- Rebuttal Comment 1.1: Title: Reply to rebuttal Comment: I thank the authors for the rebuttal. My doubts have been cleared and I have raised my score to reflect the same.	Summary: The authors present REBEL, a method for solving contextual bandit problems (such as the alignment of language models) via regressing relative rewards. They first derive their objective by demonstrating that the use of paired responses means that you can get rid of the partition function, which is impossible to estimate. They then connect their method to previous methods in RL including detailing, but not . They demonstrate that under strong assumptions REBEL is equivalent to mirror descent, and that under assumptions of coverage by the reference policy, that REBEL produces returns close to an optimal policy. Finally the authors run experiments on summarisation, general chat and image alignment, demonstrating their method compares favourably to other methods. Strengths: * The idea of using relative rewards to remove the partition function is a nice and simple idea * The theoretical connections of their method to prior methods grounds their work nicely in existing RL approaches. * The empirical results seem to demonstrate their method is competitive or better than other approaches. * REBEL compares favourably in terms of runtime and memory usage with other, similarly performing methods. Overall the theoretical and empirical examinations of their method seems very thorough. Weaknesses: See questions Technical Quality: 3 Clarity: 3 Questions for Authors: * Do the authors have any idea why REBEL seems to have a slightly higher KL than the other methods? * Although in image alignment REBEL seems to do similarly to PPO, it also has higher variance. Do you know why that might be? * Are the results for the 6.8B model significant? It seems as though REBEL produces very similar performance to e.g. PPO. For the smaller models the separation seems larger, is there a reason why the separation in performance between REBEL and other methods is bigger for smaller models? * What are the error bars in Table 1? Is that standard deviation? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors discuss the limitations throughout their work at relevant stages. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your encouraging review and comments. We respond to your individual questions below. > Do the authors have any idea why REBEL seems to have a slightly higher KL than the other methods? The KL divergence is generally close across methods. For the TL;DR experiments, following previous work [1], we apply an addition KL penalty to the reward: $r(x, y) = RM(x, y) - \gamma (\ln \pi_{\theta_{t}} (y\|x) - \ln \pi_{\theta_{0}} (y\|x))$. PPO incorporates a clipping mechanism on top of this regularization, which allows it to control the KL-divergence more strictly. This difference in approach can lead to REBEL exhibiting a slightly higher KL-divergence compared to PPO. > Although in image alignment REBEL seems to do similarly to PPO, it also has higher variance. Do you know why that might be? To fairly compare REBEL and PPO, we tuned all hyperparameters, including parameters related to reward queries for PPO. Afterward, to ensure that REBEL was compared fairly to PPO, we kept those hyperparameters related to reward queries constant so that REBEL and PPO had the same number of reward queries per update. We believe that if we were to modify the reward queries per update, REBEL would show a lower variance than our current runs. > Are the results for the 6.8B model significant? It seems as though REBEL produces very similar performance to e.g. PPO. For the smaller models the separation seems larger, is there a reason why the separation in performance between REBEL and other methods is bigger for smaller models? The discrepancy in our training setup could be the reason for the smaller separation in the larger model versus the smaller models. Specifically, we performed hyperparameter tuning using LoRA [3] on the 1.4B and 2.8B models. We then directly applied these hyperparameters to the larger 6.9B model for full-parameter training without additional tuning (detailed in Appendix H.1.2). The results for the baselines that we compared to (REINFORCE, PPO, RLOO) for the 6.9B models are obtained from [1,2], which might have directly tuned the 6.9B model. These differences contribute to the decreasing performance gains as model size increases. > What are the error bars in Table 1? Is that standard deviation? The error bars are standard deviations. Results are averaged over three seeds and the standard deviations across seeds are in parentheses. [1] Huang S, Noukhovitch M, Hosseini A, Rasul K, Wang W, Tunstall L. The N+ Implementation Details of RLHF with PPO: A Case Study on TL; DR Summarization. [2] Ahmadian A, Cremer C, Gallé M, Fadaee M, Kreutzer J, Üstün A, Hooker S. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. [3] Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, Wang L, Chen W. Lora: Low-rank adaptation of large language models. arXiv --- Rebuttal Comment 1.1: Title: Response Comment: Thank you very much for the responses to my questions. I will maintain my score and continue to believe this is an excellent paper.	Rebuttal 1: Rebuttal: We thank all the reviewers for their time and insightful comments, which have significantly improved our paper. We are pleased that the reviewers appreciated our algorithm's simplicity, the detailed theoretical connections to prior methods, and the thorough empirical results. We summarize the main suggestions below and address each reviewer's comments individually in each reviewer’s rebuttal. * Reviewer dEG5 raises questions regarding the experimental details and empirical results. We address these concerns in our detailed response to the reviewer. * Reviewer HFsB provides interesting suggestions for exploring data aggregation and different choices of $\mu$. We would certainly explore this direction in future investigations into the REBEL algorithm. * Reviewer ihUH expresses concerns about the empirical evidence supporting our claims. In response to this, we conduct an additional convergence experiment. We plot the reward vs. step for TL;DR dataset in the attached pdf. The plot demonstrates REBEL's faster convergence and higher rewards compared to iterative DPO and PPO. * Reviewer zZ2j suggests including a separate “Limitation” section. We will certainly include this in the next version of the paper. Pdf: /pdf/df8e5383c64ff6edbb207a474bc059f92a79e504.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
AV-Cloud: Spatial Audio Rendering Through Audio-Visual Cloud Splatting	Accept (poster)	Summary: The paper proposes AV-Cloud, a framework for high-quality spatial audio rendering in 3D scenes without relying on visual cues. AV-Cloud addresses issues in current audio-visual rendering methods, such as audio lag and dependence on visual rendering quality, by introducing Audio-Visual Anchors and the Audio-Visual Cloud Splatting module. These components facilitate the generation of viewpoint-specific spatial audio synchronized with visual content. The method demonstrates superior performance on multiple benchmarks, outperforming existing baselines in audio reconstruction accuracy, perceptual quality, and acoustic effects. Strengths: 1. The concept of using Audio-Visual Anchors and Cloud Splatting to decouple audio rendering from visual rendering is interesting. 2. The paper demonstrates comprehensive experimentation and robust evaluation across multiple benchmarks. 3. The paper is well-structured and the presentation of the framework is clear. The figures and supplement examples help the readers better understand. 4. The proposed method addresses critical issues in real-time audio-visual rendering. Weaknesses: 1. The mathematical formulation of the Audio-Visual Cloud Splatting module could be more detailed. For instance, Equation (2) introduces the softmax function applied to the relative vectors and visual features, but the reason behind this specific formulation and its implications are not sufficiently explained. Clarifying how the weights $a_{ki}$ are computed and how they influence the final output would enhance understanding. 2. The technical derivation of the Spatial Audio Render Head (SARH) lacks depth. Specifically, the process described in Equations (4) and (5), where the mixture mask $m_m$ and the difference mask $m_d$ are used to compute the left and right channel outputs, is not fully elaborated. The significance of these masks and their impact on the final audio quality are not clearly discussed. Additionally, the role and impact of the convolution modules within the residual structure (Figure 3) are not sufficiently explained. 3. While the method shows strong performance on benchmarks and some real-world examples, the provided examples are too idealized and lack challenging elements like interfering sound (e.g., crowd noise). I think the robustness of AV-Cloud in more complex and noisy real-world environments should also be validated. Technical Quality: 3 Clarity: 4 Questions for Authors: See Weaknesses. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors mention the limitations of their approach's challenges and potential drawbacks. The reliance on camera calibration and the potential issues with noise in real-world audio recordings are noted. Additional imitations can be found in the Weaknesses section Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for a thoughtful review, valuable feedback, and recognizing the innovative use of Audio-Visual Anchors and Cloud Splatting, comprehensive experimentation, and clear presentation. We address the questions and specify the intended revisions below. W1: AVCS Explanation In the general response section we provided a more detailed explanation of the AVCS module. We place the details relevant to the reviewer's raised point below for convenience as well. We will add these details to the methods upon revision. The output of the Visual-to-Audio Splatting Transformer can be defined as follows: 1) Attention Mask a_{ki}: Indicates the contribution weight of each anchor, showing how much each anchor influences the spatial audio effect. Weighted sum the latent Audio Embedding e_i to get mixture Audio Embedding e’. The mixture mask m_m is derived from e’ through an MLP. 2) Output: The final integrated Relative Vector embedding of anchors w.r.t the target viewpoint pose. The difference mask m_d is derived from the Output embedding through an MLP. In Equation (2), the softmax function is applied to compute attention weights a_{ki}. This softmax function normalizes the weight of each anchor contribution, enhancing the spatial audio effect to match the listener's perspective pose. Higher weights indicate greater influence, allowing the AVCS module to dynamically adjust the anchor contribution for audio rendering based on the listener's viewpoints. Please see the visualization examples in Figure 6. W2: Spatial Audio Render Head (SARH) SARH utilizes a one-layer residual structure with two convolution modules to enhance stereo output from Equation (4). The skip connection directly takes the output stereo channels from Equation (4) S_L and S_R. And the residual path includes: - Time Filters: Adjust the energy distribution of the mixture output spectrogram S_m across the time domain to match reverberation effects. - Conv2D Layers: Smooth and enhance the time-frequency domain energy distribution post Time Filters. The input to the Conv2D layers consists of four parts: - Output of the Time Filters convolved with the mixture spectrogram S_m. - Mixture mask m_m. - m_L=−m_d for the left channel or m_R=+m_d for the right channel post-process. - Original stereo output from Equation (4): S_L for the left channel or S_R for the right channel post-process. The Conv2D layers process these inputs to generate the left and right residual channels. The final result is obtained by adding these residual outputs to the original S_L and S_R spectrograms from Equation (4). Contribution of key components We studied and verified the contribution of key components in SARH in ablation studies described in Sec 4.3 and Table 2. In particular: - Baseline Model (w/o AVCS and residual convolution modules): Utilize MLP to predict two acoustic masks. Compared to the full AVCloud, the baseline does not perform well across all five metrics, indicating limited generalization to novel views when relying solely on viewpoint poses. (Line 302-305) - w/o AVCS: - Replace AVCS module with an MLP. w/o AVCS is unable to capture the relationship between left-right channel energy and viewpoint poses well, reducing the LRE accuracy by 20% compared to our full AVCloud (Line 306-310). - Compared with the Baseline, the residual convolution module can improve perceptual quality (0.357 -> 0.318), reverberation effect (0.13 -> 0.08), magnitude (0.488->0.468) and spatial effect metrics (1.329 -> 1.124) largely. - w/o time: Removing Time Filters to study its effect. Reverberation time error increases from 0.074 to 0.084, and LRE error increases by 5%, demonstrating the Time Filters’ crucial role in adjusting time-domain energy distribution, impacting reverberation time and overall acoustic quality. (Line 327-329) As suggested, we conducted two additional ablation experiments to study the effect of 1) two binaural masks and 2) Convolution 2D layers in the residual block. - w/o 2 masks: Direct prediction of two-channel spatial effect transfer mask from the AVCS output (instead of predicting difference mask and mixture mask separately). As a result, LRE significantly increases from 0.936 to 0.983. - w/o conv2d: Removal of 2D convolutional layers in the residual block and retaining Time Filters only. Without Conv2D layers, the Time Filter module handles reverberation but fails to distinguish between left and right channels well, indicated by LRE increasing from 0.936 to 1.019. Other metrics remain relatively stable. The need for Conv2D layers stems from reliance of residual blocks to post-process in the frequency-time domain. \|Variants \| MAG \| LRE \| ENV \| RTE \| DPAM\| \| -------- \| ---- \|---- \|---- \|---- \|---- \| \|AVCloud (Full)\|0.351\|0.936\|0.145\|0.074\|0.276\| \|w/o 2 masks\|0.357\|0.983\|0.147\|0.070\|0.280\| \|w/o conv2d\|0.359\|1.019\|0.148\|0.074\|0.280\| We will include these results with detailed explanations and analysis in our final revision. W3: Challenging elements such as interfering sounds In our real demo, we successfully handled sounds like wind and bird chirping since the reference input captured these sounds. However, we indeed acknowledge that the scenario of dealing with highly noisy environments would pose a limitation to the approach, particularly scenarios where the ground truth contains noise, but the reference sound does not. Moreover, our current system is designed for real-time processing and its components are thus of lightweight nature and are not designed to cope with extreme noisy cases. Future solutions could be to consider additional pre-processing/processing such as noise reduction, sound separation, or the design of a higher capacity model. These are beyond our current scope but could be integrated with future developments. We appreciate the reviewer’s insight and will consider this in future works. --- Rebuttal Comment 1.1: Comment: Thank you for addressing most of my concerns. I appreciate your effort on the additional ablations, and please make sure to include them in a revision. I’m happy to increase my score to WA. --- Reply to Comment 1.1.1: Comment: We want to thank the reviewers again for very helpful feedback and discussion. And we want to ensure the reviewers that we have their feedback and resulting revisions to algorithms description (W1, W2 to the Method Section), additional ablation studies analysis (W2 to the Ablations Section) and further limitation discussion (W3 to the Limitations Section). We will incorporate these content into a revised version with the additional page and more extensive supplementary information, which can highlight our motivation and contribution at a better clarity level.	Summary: A novel approach for rendering high-quality spatial audio in 3D scenes, called AV-Cloud, is proposed. This method synchronizes with the visual stream without relying on or being explicitly conditioned by visual rendering, enabling immersive virtual tourism through real-time dynamic navigation of both audio and visual content. Unlike current audio-visual rendering methods that depend on visual cues and may suffer from visual artifacts causing audio inconsistencies, AV-Cloud overcomes these issues. It uses a set of sparse AV anchor points, forming an Audio-Visual Cloud derived from camera calibration, to represent the audio-visual scene. The Audio-Visual Cloud allows for the generation of spatial audio for any listener location. A novel module, Audio-Visual Cloud Splatting, decodes these AV anchor points into a spatial audio transfer function for the listener’s viewpoint, which is then applied by the Spatial Audio Render Head module to transform monaural input into viewpoint-specific spatial audio. This approach eliminates the need for pre-rendered images and efficiently aligns spatial audio with any visual viewpoint. The results are satisfying. Strengths: 1. The AV anchors strategy seems to be interesting and effective for audio-visual scene representation. The Audio-Visual Cloud Splatting is novel for AV tasks but more likely to be a Q-former. 2. The experiment results are good and ablations are clear. Weaknesses: As I mentioned in the strengths, the Audio-Visual Cloud Splatting seems to be a Q-former like module. Technical Quality: 3 Clarity: 3 Questions for Authors: What is the difference between the AVCS and Q-former? Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for a thoughtful review, valuable feedback and recognizing the novelty of AV Anchors for 3D audio-visual scene reconstruction. W1 & Q1 Difference between AVCS and Q-former While AVCS and Q-former are transformer-based structures, they serve different purposes and utilize transformer outputs in distinct ways. Q-former serves as an intermediary between a frozen image encoder and a frozen Language Model. Its query is a set of learnable vectors designed to extract visual features from the frozen image encoder. The query acts as an information bottleneck, providing the most useful visual features for the Language Model to generate the desired text output. In contrast, AVCS is designed to learn features that adapt to view poses from a 3D scene representation to derive the audio spatial effect transfer function. This function converts monaural reference sound into stereo audio at the listener’s viewpoint. Each query in AVCS corresponds to a specific frequency band, and the output is the integrated Relative Vector, embedding the viewpoint pose relative to the 3D anchors' scene representation. In contrast to Q-former, AVCS involves explicit projection logic that adapts to the viewpoint pose in the 3D world, rather than implicitly extracting visual features relevant to audio. Specifically, in the AVCS module, each Audio-Visual Anchor is projected to the head coordinate system of the target listener, and the anchor features are then integrated for each audio frequency band using the Visual2Audio Splatting Transformer. The attention mask indicates the contribution weight of each anchor, showing how much each anchor influences the spatial audio effect. This mechanism dynamically adjusts the contribution weights of anchors for audio rendering based on the listener's viewpoint. The outputs and attention weights are used to derive the mixture and difference audio masks separately for the audio transfer function. The Visual-to-Audio Splatting Transformer works as follows: Input 1) Query: Frequency embedding for each audio frequency band. Shape (F, C) 2) Key: Combined RGB visual feature of each anchor and its Relative Vector in the listener's head coordinate system. Shape (N, C) 3) Value: Relative Vector. Shape (N, C) Output 1) Attention Mask a_{ki}: Indicates the contribution weight of each anchor, showing how much each anchor influences the spatial audio effect. Shape (F, N). Weighted sum the latent Audio Embedding e_i (N, C) to get mixture Audio Embedding e’ (F, C) 2) Output: The final integrated Relative Vector embedding of anchors with respect to the target viewpoint pose. This embedding is highly relevant to the 3D pose of the listener. Shape (F, C) In Equation (2), the softmax function is applied to compute attention weights a_{ki}. This softmax function normalizes the weight of each anchor contribution, enhancing the spatial audio effect to match the listener's perspective pose. Higher weights indicate greater influence, allowing the AVCS module to dynamically adjust the anchor contribution for audio rendering based on the listener's viewpoints. Please see the visualization examples in Figure 6.	Summary: The paper explores the problem of generating 3D audiovisual scenes – that is, generating 3D scenes with spatial audio. The proposed approach, AV Cloud, uses anchor points obtained from Structure-from-Motion (SfM) points. The anchors are then used with an AV Cloud splatting module which decodes the visuals and the audio. Experiments are done on RWAVS and Replay-NVAS with comparisons done with several prior works. Strengths: – 3d audiovisual scene generation is a really interesting problem to solve. WHile there is considerable literature on visual scene generation, generating 3d visual scene is an interesting problem with real-world applications. – The model claims to be able to generate the audio and the visuals in parallel. Essentially unlike prior work it decouples the generation of two modalities by not using the generated visuals for generating the audio. – On objective metrics, the paper claims to make good improvements ---- increased score after rebuttal Weaknesses: – The paper is a bit difficult to follow – especially the key part of AudioVisual anchor points. – First, a short primer on SfM is desirable, even if it is in Appendix. More importantly though, it is not clear why it makes sense to use SfM points and clustering on top of them to model AV anchor points and generation of spatial points. Why does it make sense to use SfM points or anchors derived from them as the starting point for AV generation ? What relation the anchors have with audio which motivates the fact that these anchors can be used for audio generation ? – Second, the details of AV anchor points are fuzzy. The visuals are used for SfM points which are then clustered to get the anchors. Where is the audio into picture here ? Are these anchors visual only ? If so, why are we calling it AV Anchors ? – In prior works, for example AV-Nerf, there is an an explicit AV-Mapper which learns the audio visual relations through which the spatial audio generatio happens. Here Visual2Audio splatting transformer is expected to model that ? – For the subjective tests, it would be good to actually get proper subjective ratings on the generated spatial audio. The current preference numbers are not very informative. Getting the spatial audio rated with respect to their quality and spatial characteristics would be much more meaningful. – Since NAF, INRAS and other works are considered here - I think it would be good to reference NACF ([R1]) below. NACF specifically focuses on using visuals and is ideal for comparison. [R1] Neural Acoustic Context Field: Rendering Realistic Room Impulse Response With Neural Fields Technical Quality: 2 Clarity: 1 Questions for Authors: Please address the questions below. Confidence: 4 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for a thoughtful review, valuable feedback and recognizing the importance of 3D audio visual scene synthesis and our contribution of proposing the novel parallel pipeline for audio and visual rendering. We address the questions and specify the intended revisions below. W1 - Primer on SfM SfM (Structure from Motion) reconstructs a 3D environment from a sequence of 2D images. SfM points are distinct and recognizable points in a scene detected in images. SfM points are used to determine the camera viewpoint and 3D coordinates leading to a detailed 3D point cloud. In Line 64-65, we included the reference for SfM [1] and SfM details are described in [1, Sec. 4.2]. As suggested by the reviewer, we will include a primer for SfM in the appendix that expands upon the short abstract of SfM and SfM points above and includes some key details from [1]. W2 - SfM and their clustering for AV generation We found SfM points to be very informative to AV reconstruction due to: 1) They capture detailed 3D scene geometry, representing the physical boundaries and surfaces from which sounds can reflect, diffract, or be absorbed. This geometric information is key for rendering realistic audio-visual scenes and can be reused by any emitter and listener within the scene. Recent audio renderers [42, 32] also use point-based representations, showing more effective generalization with the same training data (Lines 109-128). 2) SfM points are also used to initialize visual renderers, e.g. 3D-GS[3], allowing for parallel synchronization of audio and visual rendering. Clustering: Clustering reduces the density of raw SfM point clouds, enhances computational efficiency while maintaining key geometric boundaries and surfaces (Lines 144-146, visualized in Figure 6). We will further clarify our motivation in the introduction and method sections. W3 - AV anchors their use for AV generation After clustering, we initialize the anchor features using: 1) 3D coordinates; 2) RGB values; and 3) Latent Audio Embedding e_i (Lines 140-151) for a point-based audio-visual representation. Specifically, the Audio Embedding captures how the anchor region contributes to sound propagation for different listener viewpoints, similar to boundary bounce points in INRAS [42]. AV Anchors derive the spatial audio transfer function via the AVCS module (Sec 3.2). We have included an additional, more practical, and detailed explanation of the AVCS module in our response to all reviewers and we will incorporate it in the revision as well. Each anchor is projected into the listener's head coordinate system, adapting to head orientation with the following components: 1) 3D Coordinates: Calculate the Relative Vector, serving as Key and Value for the Visual2Audio Splatting Transformer. 2) RGB Values: Positional encoding for the Key to enhance scene understanding and distinguish between Anchors at different locations (lines 170-174). 3) Latent Audio Embedding: Weighted by the Attention Mask to get the mixture Audio Embedding. The mixture mask is derived from this embedding through an MLP, manipulating mixture sound magnitude changes. The final integrated Relative Vector from the transformer is used to obtain the difference mask through MLP. W4 - Difference between AVCS and AV-Mapper AV-Mapper (AVNeRF) relies on RGB and depth images, making it dependent on image data to adapt to viewpoint changes. In contrast, AVCS module decouples the transfer function from images, using point-based (AV anchors) scene representation. By anchor projection and the Visual-to-Audio Splatting Transformer, our approach adapts to view poses from a view-independent 3D scene representation. Advantages: 1) Reusable anchor representation for all viewpoints, enabling better generalization to novel viewpoints with better accuracy, fewer parameters, and higher inference speed (Table 1). 2) Explicit adaptation to 3D world coordinate system, eliminating the need for visual rendering to obtain strong visual cues. The approach can achieve more efficient and accurate audio-visual synchronization without reliance on visual render results to obtain strong visual cues. W5 - Subjective Tests Justification We particularly set the subjective test instructions to target evaluation of the alignment of spatial audio effects with visual content. Participants were instructed to select the video where spatial audio matches the visual content, focusing on left-right ear effects and overall synchronization. Instructions included: “… Select the video that features the spatial audio effect best matching the visual content.… Pay attention to the varying left-right ear spatial effects… Evaluate each video comprehensively for spatial effects, audio continuity, and quality…” This test complements quantitative metrics like LRE, providing a comprehensive assessment. Our method received 48% of the votes, outperforming other methods, demonstrating its effectiveness in creating an immersive audio-visual experience. W6 - Comparison with NACF We thank the reviewer for recommending to include NACF in the evaluations. Similar to INRAS [32] and NAF [42], NACF does not include viewpoint-based reweighting, which leads to less effective representation. AVCloud, with dynamically reweighted Audio-Visual Anchors, significantly improves audio rendering metrics. The Spatial Render Head also enhances reverberation effects, achieving a more accurate acoustic metric (RTE). We will include these references and results in our revision. RWAVS Dataset: \|Method \| MAG \| LRE \| ENV \| RTE \| DPAM\| \| -------- \| ---- \|---- \|---- \|---- \|---- \| \|NACF\|0.459\|1.364\|0.176\|0.138\|0.506\| \|AVCloud (Ours)\|0.351\|0.936 \|0.145\|0.074\|0.276\| Replay-NVAS Dataset: \|Method \| MAG \| LRE \| ENV \| RTE \| DPAM\| \| -------- \| ---- \|---- \|---- \|---- \|---- \| \|NACF\|0.298\|0.722\|0.079\|0.332\|0.544\| \|AVCloud (Ours)\|0.180\|0.600\|0.052\|0.065\|0.234\| --- Rebuttal Comment 1.1: Title: After rebuttal Comment: Thanks for the detailed response. The rebuttal add in a lot of clarity. I think the paper will need a good amount of change to incorporate all the clarity and additional results. I have increased the score. --- Reply to Comment 1.1.1: Comment: We want to thank the reviewers again for very helpful feedback and discussion. And we want to ensure the reviewers that we have their feedback and resulting revisions to motivation explanation (W2 to the Introduction Section), algorithms description (W2, W3 to the Method Section), comparison discussion with addition results (W4, W5, W6 to the Experiment analysis) and primer content (W1 to the Supplementary Section). We will incorporate these content into a revised version with the additional page and more extensive supplementary information, which can highlight our motivation and contribution at a better clarity level.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their thoughtful reviews and valuable feedback. In this general section, we wanted to provide a more detailed explanation of the Audio-Visual Cloud Splatting (AVCS) module, as several reviewers have suggested. The AVCS module is one of the key contributions of our work. It receives as input point-based audio-visual scene representation, AV Anchors, and decodes them into spatial audio transfer function (Sec 3.2). The spatial audio transfer function is of the form of two acoustic masks which convert the monaural reference sound into stereo audio. In the AVCS module, each Audio-Visual Anchor is projected to the head coordinate system of the target listener. The anchor features are then integrated for each audio frequency band using the Visual2Audio Splatting Transformer which works as follows: Input: 1) Query: Frequency embedding for each audio frequency band. Shape (F, C) 2) Key: Combined RGB visual feature of each anchor and its Relative Vector in the listener's head coordinate system. Shape (N, C) 3) Value: Relative Vector in the listener's head coordinate system. Shape (N, C) Output: 1) Attention Mask a_{ki}: Indicates the contribution weight of each anchor, showing how much each anchor influences the spatial audio effect w.r.t each frequency band. Shape (F, N). Weighted sum the latent Audio Embedding e_i (N, C) using attention masks to get mixture Audio Embedding e’ (F, C). 2) Output: The final integrated Relative Vector embedding of anchors for each frequency band w.r.t the target viewpoint pose. This embedding is highly relevant to the 3D pose of the listener. Shape (F, C) In Equation (2), the softmax function is applied to compute attention weights a_{ki}. This softmax function normalizes the weight of each anchor contribution, enhancing the spatial audio effect to match the listener's perspective pose. Higher weights indicate greater influence, allowing the AVCS module to dynamically adjust the anchor contribution for audio rendering based on the listener's viewpoints. Please see the visualization examples in Figure 6. The spatial audio transfer function consists of two masks: a mixture mask m_m and a difference mask m_d. The two masks then convert monaural reference sound into stereo audio at the listener’s viewpoint. We obtained the two masks from the AVCS module as follows: 1) The mixture mask (m_m) is derived from the weighted sum of latent Audio Embeddings through MLP, showing mixture sound magnitude variation across viewpoint locations. 2) The difference mask (m_d), derived from the final integrated Relative Vector (Output of Visual2Audio Splatting Transformer) through MLP, captures left-right channel differences relevant to the listener's 3D pose. We will add the above explanations to our revision.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Towards Multi-Domain Learning for Generalizable Video Anomaly Detection	Accept (poster)	Summary: This work proposes a new task named Multi-Domain Learning Video Anomaly Detection, which aims to learn a general VAD model across domains. The work finds that abnormal conﬂict is a critical challenge in the task. Then, the work establishes a new benchmark, designs an effective baseline and conducts extensive experiments to investigate this challenge. The results shown on the benchmark demonstrate that the abnormal conﬂict is alleviated. Strengths: 1. The work proposes a new task, which is interesting. 2. The work establishes a new benchmark to evaluate the new task. 3. The motivation of the proposed baseline, i.e., abnormal conﬂict, is clear and makes sense. Weaknesses: I have some concerns about the proposed method, and I think more comparison experiments are needed to demonstrate the effectiveness. Despite this, I think the abnormal conflict issue is interesting, thus I am willing to raise my rating if my major concerns are addressed. My concerns are as follows: 1. Why the proposed Abnormal Conﬂict (AC) classifer can address the abnormal conﬂict problem? Why the label is determined by the discrepancy in Eq. (6)? It seems that there are some mistakes in the formula (inconsistent with that in Fig. 2). 2. I would like to see the results of more baselines, in addition to MIL, Null-MIL and NullAng-MIL. 3. More detailed discussions about related works are needed, e.g., virual video anomaly detection datasets [1] and related techniques utilizing virtual datasets [2]. [1] Ubnormal: New benchmark for supervised open-set video anomaly detection, CVPR 2022 [2] Generating Anomalies for Video Anomaly Detection with Prompt-based Feature Mapping, CVPR 2023 Technical Quality: 2 Clarity: 2 Questions for Authors: See the Weakness part. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: The paper has discussed the limitations and potential impacts of the work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # [W1] AC Classifier Thank you for highlighting this important aspect. Role of the AC Classifier: By training with the AC Classifier, the domain-agnostic layer learns Conflict-Aware features, which helps in resolving conflicts. To achieve a general VAD model through multiple domain learning while avoiding abnormal conflicts, our framework aggregates features through domain-agnostic layers and performs multiple head learning. This activates only the output of the head corresponding to the input domain, assigning inactive heads with Null values to prevent confusion. While the heads are divided to prevent conflicts, the agnostic part extracts features from all datasets using a single branch. Therefore, to explore general features while being aware of abnormal conflicts, the AC classifier predicts conflicts by leveraging the variance in abnormal scores across multiple heads. This auxiliary task provides performance gains in most experiments. Notably, in E2 and E3 settings, it shows a significant boost effect, aiding in adaptation to unseen domains without any additional cost to the final model. Additionally, as shown in Supplementary Material Fig. A5, the conflict score of the AC classifier plotted for the UCFC domain reveals high conflict scores in scenarios like jaywalking, which is normal in UCFC but abnormal in other domains (TAD, ST), demonstrating the classifier's awareness of AC in the domain-agnostic part. Pseudo label of the AC Classifier: As shown in Table A5, assigning pseudo labels in Eq. 6 was determined experimentally. The purpose of the AC classifier is to facilitate feature learning with discrepancies across domains. Therefore, when the gap between the abnormal scores of multiple heads exceeds a threshold $\tau$, it indicates different definitions of normal and abnormal between domains (e.g., abnormal in one domain and normal in another), and a conflict is considered to have occurred. When the score gap is small, it indicates consistent scores with no conflict, and the AC classifier generates a pseudo label accordingly. In Fig. 2, the yellow score graph shows a large gap between scores $S^a_{D_2}$ and $S^a_{D_M}$, indicating an abnormal conflict and assigning $Y^{AC}=1$ in Eq. 6. Conversely, the green score graph shows consistent scores $S^a_{D_i}$ across all domains (all normal or all abnormal), indicating no conflict and assigning $Y^{AC}=0$. To determine conflicts between domains, we conducted ablation studies comparing different methods, such as the difference between scores of multiple heads, Std. with τ = 0.1, and a fixed value . We kindly refer reviewer to Section 4.2.5, "Role of the AC Classifier," and Section 4.3, "Discussions." If there are any mistakes in the figures, we would appreciate it if you could point them out for corrections. # [W2] Additional Baselines \| \| \| \| \| \| \| \| \| \| \|:---:\|:---------:\|:-----:\|:-----:\|:-----:\|:------:\|:-----:\|:-----:\|:------:\| \| \| Models \| \| \| \| Target\| \| \| \| \| \| \| UCFC \| XD \| LAD \| UBIF \| TAD \| ST \| AVG. \| \| E1 \| WSAL \| 76.47 \| 78.35 \| 75.44 \| 86.41 \| 85.62 \| 82.44 \| 80.79 \| \| \| WSAL+Ours \| 76.90 \| 78.59 \| 76.17 \| 87.83 \| 81.89 \| 86.27 \| 81.27 \| \| \| \| \| \| \| \| \| \| \| \| E2 \| WSAL \| 76.59 \| 73.75 \| 76.71 \| 79.86 \| 77.68 \| 55.54 \| 73.36 \| \| \| WSAL+Ours \| 76.67 \| 73.69 \| 65.54 \| 85.85 \| 70.1 \| 72.41 \| 74.04 \| \| \| \| \| \| \| \| \| \| \| \| E3 \| WSAL \| 75.77 \| 72.18 \| 66.45 \| 82.57 \| 75.05 \| 78.05 \| 75.01 \| \| \| WSAL+Ours \| 76.81 \| 72.38 \| 59.51 \| 83.49 \| 77.6 \| 74.57 \| 74.06 \| \| \| \| \| \| \| \| \| \| \| \| E4 \| WSAL \| 80.99 \| 81.6 \| 76.43 \| 87.87 \| 88.78 \| 87.47 \| 83.86 \| \| \| WSAL+Ours \| 79.21 \| 81.17 \| 78.19 \| 90.98 \| 86.09 \| 90.61 \| 84.38 \| \| \| \| \| \| \| \| \| \| \| We conducted additional experiments using the WSAL [29] model as a baseline to validate our proposed method on the MDVAD benchmark's four protocols. The above table presents the results, where all settings are consistent with the other experiments in the paper. The results show the addition of multi-head learning with NullAng-MIL and the AC Classifier to the WSAL model brings performance gains which effectively operate across different baselines. Notably, in the E3 and E4, which shows the target adaptation of the pre-trained general model, the method yielded competitive results. We focus more on highlighting the necessity of MDVAD and raising the issue of AC. Our paper introduces a novel task along with a benchmark, evaluation protocols, and a baseline and emphasizes analysis. Therefore, aspects like baseline architecture design or utilizing a powerful backbone were not the primary focus. However, future research will delve deeper into studying more sophisticated baselines for resolving AC within the MDVAD setting. # [W3] VAD with virtual datasets [i, ii] We appreciate the thoughtful comment. Following the reviewer's suggestion, we have added an analysis using the virtual dataset UBNormal [i] (please refer to the general response above). PFMP [ii] proposed a method to utilize virtual data anomalies to reduce scene discrepancy with real-world data, addressing the issue of data scarcity in VAD. While this method can mitigate data scarcity, it still faces the problem of anomaly conflicts when the criteria for abnormal events are inconsistent across multiple datasets. The conducted experimental analysis on utilizing virtual datasets [i] and a discussion on related work [ii] will be included in Section E of the supplementary material. [ii] "Generating anomalies for video anomaly detection with prompt-based feature mapping." CVPR, 2023. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' detailed response. After reading their response, some of my concerns are addressed. However, it is still not very clear for me that how the proposed Abnormal Conﬂict classifer addressed the abnormal conﬂict problem. I would like to see more detailed explaination from the authors, and I am open to hear from other reviewers. --- Reply to Comment 1.1.1: Comment: We are pleased that our rebuttal has addressed the reviewer's concerns. # [Q1] How the AC Classifier Addresses the AC Problem The AC classifier helps the model learn conflict-aware features. To provide a clearer understanding, we would like to explain this through a detailed step-by-step procedure. We have designed our framework with Domain-Agnostic Layers that learn multi-domain general features, followed by Multiple Heads that predict abnormalities for each domain. When the Domain-Agnostic layers learn the ability to perform AC classification, their features capture knowledge about whether the input snippets are related to AC or not. From the perspective of the Heads, these features are divided between AC and non-AC in the feature space, allowing the head to apply different criteria (decision boundaries) when distinguishing between normal and abnormal. For instance, when classifying abnormalities, non-AC scenarios can be addressed more easily, while AC scenarios require a more careful exploration. As a result, 1) models trained with the AC Classifier handle AC more effectively in multi-domain learning (Table 5) and adapt better to unseen targets (Table 7). 2) The AC Classifier learning was experimentally designed and validated (Table A3) and 3) the AC score plot demonstrates that the model has effectively learned to predict AC (Fig. 3(a) and Fig. A5). In summary, the AC Classifier plays a crucial role in enhancing the model's ability to address AC problems by learning conflict-aware features, thereby allowing the Heads to aware conflicts when classifying normal and abnormal instances, leading to more effective learning. Thank you for the valuable feedback, and we hope this response alleviates the reviewer's concerns. --- Rebuttal 2: Comment: Thanks for the authors' detailed response. According to your illustration, in my understanding, the notation "$\max_{i} S^a_{D_d, i}$" should be "$\max_{d} S^a_{D_d, i}$" (also "$\min_{i} S^a_{D_d, i}$" should be "$\min_{d} S^a_{D_d, i}$"). This is because the index $d$ denotes a domain and $i$ denotes a snippet, and you argue that an abnormal conflict has more likely occurred as the score gap between domains is larger. I think there may be some typos in the formulas. Please check this. --- Rebuttal Comment 2.1: Comment: We apologize for the confusion caused by typo in our notation. As the reviewer correctly pointed out, the notation should be $d$ instead of $i$ in Eq. 6. We will revise the equation in the manuscript accordingly. $y_i^{AC}= 1$ where $[\max_{d} s^a_{D_d, i}-\min_{d} s^a_{D_d, i}-\tau]_+> 0$ $y_i^{AC}= 0$ where $[\max_{d} s^a_{D_d, i}-\min_{d} s^a_{D_d, i}-\tau]_+\le 0$ Once again, we appreciate reviewer for pointing out this typo.	Summary: In this paper, authors proposed a new task called Multiple Domain VAD (MDVAD), along with a benchmark and new evaluation protocols. Authors' goal is to construct a general VAD model by conducting multi-domain learning while recognizing abnormal conflicts and exploring representations of general normality and abnormality. Authors introduced a baseline for MDVAD and proposed a new framework with multi-head to mitigate abnormal conflicts and proposed Null-Multiple Instance Learning (Null-MIL) and NullAngular-MIL (NullAng-MIL) losses for multi-domain training. Additionally authors suggested an Abnormal Conflict (AC) Classifier to explore general features while being aware of abnormal conflicts. Authors analyzed the primary issues of MDVAD and proposed a baseline for this new task. Strengths: 1. According to the analysis, authors believed that the abnormal conflict and the scene discrepancy are the two main issues and designed a framework with multi-head to deal with these problems. 2. Null-MIL and NullAng-MIL methods are designed for multi-domain learning, and an AC classifier is proposed for learning general features while abnormal conflicts exists. 3. Authors provided sufficient experiment results for this task and create a new baseline. Weaknesses: 1. The proposed framework with multi-head for multi-domain seems not flexible enough during the domain changes, such as adding a new dataset with extra abnormal conflicts. And for the abnormal conflicts, will the proposed method performs better compared to make anomaly categories classifications for all anomaly events type of all domains? 2. In my opinion, traditional WS-VAD methods are designed to detect abnormal events in single domain without abnormal conflicts, and when abnormal conflicts exists, it will be better to use other paradigms such as temporal action localization or video grounding. And for the current WS-VAD datasets, the annotations are video-level, or even without category information, which is too weak for higher level anomaly detection. Training model with the current MDVAD paradigm is likely to not achieve good results. 3. Maybe using visual-language model with multimodal alignment can deal with the above issues? These models contain more knowledges for more event categories and higher generalization ability, which are likely to have the ability to individually detect conflicting anomalies. Compared to multi-head regression, is VL alignment a better approach for MDVAD task? Technical Quality: 3 Clarity: 3 Questions for Authors: My main questions are shown in the weakness. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # [W1-1] when adding a new dataset Thank you for highlighting this important consideration. Our framework consists of domain-agnostic layers and domain-specific heads, with each head being the final layer of the entire model, $W_{D_d}\in \mathbb{R}^{T\times 1}$ where $T=128$, which is a very small part. If a new domain is introduced and additional models need to be trained, we can flexibly add a final branch. Creating a single general model is more cost-effective than developing individual in-domain models for each domain and retraining them every time a new domain appears. Moreover, such as in the E3 and E4 settings, pre-training the general model (E2) with multiple source domains without target domains, when new target data is added, can adapt to the new domain by tuning the source heads, without adding new heads. We will include this perspective in Section 4.3 discussion. # [W1-2] Abnormal classification for AC Even with comprehensive information on all abnormal categories across all domains, the lack of consistency in abnormality criteria between domains would still result in abnormal conflicts, posing challenges for abnormal classification through multiple domain learning. Additionally, in real-world scenario, addressing the issue through classification would lead to a closed-set model, limiting its ability to handle unexpected anomalies. We will discuss this approach in Section E of the supplementary material. # [W2] WS-VAD limitations and alternative paradigms We appreciate the reviewer's insightful comments. We would like to address each point raised: 1. Traditional WS-VAD Methods in Single Domain While traditional WS-VAD methods are designed for single-domain anomaly detection without conflicts, the MDVAD approach specifically addresses the limitations of these methods, which are heavily influenced by the criteria for abnormality defined by each dataset. Models that perform well in in-domain settings require each domain to be learned separately, necessitating sufficient training data for each single domain. This limits the applicability of anomaly detection methods to real-world settings. Therefore, research on a general VAD model capable of multiple domain learning is necessary to effectively mitigate the challenges posed by abnormal conflicts across different domains. 2. Alternative Paradigms for AC (Temporal Action Localization or Video Grounding) We agree that temporal action localization and video grounding are effective paradigms for certain types of video analysis. However, even with precise temporal annotations or detailed category information, defining anomalies precisely across multiple domains is challenging due to the varying criteria and sensitivity across datasets. This makes it difficult to provide a clear solution for conflicts. We hope this addresses your concerns. Thank you for your valuable feedback. Please let us know if any further considerations are needed. # [W3] Utilizing VL models We thank the reviewer for their thoughtful suggestion regarding the use of visual-language models with multimodal alignment. We appreciate the opportunity to discuss the potential advantages and considerations of such an approach in MDVAD task. * Advantages of Visual-Language Models: Recent studies have demonstrated the powerful capabilities of VLMs, achieving performance gains by leveraging VLM backbones and multimodal alignment for VAD tasks. Additionally, the integration of text information allows for meaningful interpretations of anomalies. * AC detection with VL models: While VLMs offer significant advantages, they also have limitations in capturing complex scenes in surveillance videos. According to [iii], existing VLMs may face challenges such as inaccuracies in color recognition, difficulty identifying intricate scenes, and struggles in capturing subtle movements. Furthermore, the availability of textual descriptions or annotations for VAD datasets is often limited, making high-level information like abnormal conflict detection an additional challenge. * Future Directions: We acknowledge the potential of VLMs and multimodal alignment as a promising direction for future research. Incorporating such models could enhance the ability to detect and classify a broader range of anomalies, especially in scenarios where textual annotations are available. For future work, we plan to explore this avenue to further improve the robustness and generalization capabilities of our MDVAD framework. [iii] Yuan, Tongtong, et al. "Towards Surveillance Video-and-Language Understanding: New Dataset Baselines and Challenges." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.	Summary: The manuscript addresses the limitations of existing Video Anomaly Detection (VAD) models that are confined to single-domain learning. The primary contribution of the paper is the introduction of a new task called Multi-Domain Learning for VAD (MDVAD), which aims to develop a general model capable of identifying abnormal events across multiple domains. The manuscript conducts experiments using the MDVAD benchmark and demonstrates the limitations of traditional multi-domain learning. It shows the effectiveness of the proposed baselines in handling abnormal conflicts and achieving robust performance across multiple domains. Strengths: 1.The manuscript proposes a new task, Multiple Domain Video Anomaly Detection (MDVAD), which solves the problem that the existing model is limited to a single domain and provides a new idea for the development of domain-generalized models. 2.The MDVAD method proposes domain-specific multiple head mechanism and Null-Multiple Instance Learning Method (Null-MIL), which effectively solves the problem of anomaly conflict between different domains. 3. The MDVAD method constructs a new benchmark containing six representative VAD datasets, which fills the gap of the lack of unified evaluation standard in multi-domain learning tasks. 4. The MDVAD method designs four evaluation protocols (held-in, leave-one-out, low-shot domain adaptation, and full fine-tuning) to systematically evaluate the generalization ability of the model. Weaknesses: 1. MDVAD introduces the domain-specific multi-head mechanism and the Null-MIL method, which increases the complexity and computational cost of the model, and may place higher demands on the computational resources in practical applications. 2. The multi-domain learning task itself is difficult to train, and with the proposed method further increasing the complexity of training, MDVAD may require longer training time and higher technical requirements. 3. Although the theoretical background and analysis are provided, the theoretical basis and derivation process of some of the methods of MDVAD are slightly weak and need to be further explored and verified in depth. Part of the theoretical analysis is based on specific assumptions, and these assumptions may not be fully valid in practical applications, affecting the applicability of the theoretical analysis. 4. Although new benchmarks and assessment protocols are proposed, MDVAD lacks comparative experiments with other state-of-the-art methods, making it difficult to objectively assess the relative advantages of the proposed methods. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What is the training difficulty of MDVAD? The introduction of domain-specific multi-head mechanism and Null-MIL method greatly increases the complexity and computational cost of the model, can it meet the real-time requirements in practical applications? 2. Have the MDVAD and evaluation protocols been subjected to comparative experiments with other state-of-the-art methods in order to objectively assess the relative advantages of the proposed methods? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # [W1, W2, Q1] Complexity and computational cost In our proposed framework, only the final layer, $W_{D_d}\in \mathbb{R}^{T\times 1}$ where $T=128$, corresponds to the head and is added based on M number of datasets ($T \times M$). This constitutes a very small parameter and computational load compared to the total weights of the VAD model. We kindly refer reviewer to Section 3.2 on Complexity. # [W2, Q1] Training difficulty and inference time Regarding difficulty of training convergence, the pseudo-labels of the AC classifier are assigned through multiple heads (Eq. 6), which may lead to lower label reliability in initial training. However, this issue resolves as training progresses. Comparing a single head and multiple heads (when 6 datasets), the training times are 2.68 and 2.81 hours, and the inference times are 0.158 and 0.164 ms per snippet, respectively, indicating a negligible increase in complexity. # [W3] Theoretical analysis and applicability We would like to elaborate on how the following points support the theoretical flow of our work: * [Section 2.2 and Section D] Our motivation and the necessity for the novel task MDVAD are derived from the analysis of the VAD benchmark and the cross-evaluation of single domain models (Section 2.2, Tables 1 and 2). We have identified two key issues: Abnormal Conflict and Scene Discrepancy. We validated our assumption regarding AC by computing $AC_{i,j}$ using Eq. 8, which measures the average of the relative False Positive Rate (FPR) and relative False Negative Rate (FNR) to quantify abnormal conflict between domains (Supplementary Material, Section D, Table A2). For the domain discrepancy assumption, we analyzed the Earth Mover’s Distance (EMD) (Table 3). * [Section 4] We defined protocols for four scenarios (E1~E4) considering practical applicability and conducted quantitative and qualitative experimental analyses to support our assumptions about AC and domain discrepancy: E1: When all source and target domains are accessible, performing similar to multi-task learning, handling multiple domains with a single model. E2: When the source and target domains are different, and the target is not accessible, pre-training on large data without target knowledge simulates the scenario where training data for the practical application domain is unavailable. E3 : When a few samples from the target domain are provided, adapting the pre-trained general model. E4: When the pre-trained general model is fine-tuned for the specific needs of the target domain. # [W4, Q2] Comparison experiments \| \| \| \| \| \| \| \| \| \| \| \| \| \|:---:\|:------:\|:--------:\|:-----:\|:-----:\|:-----:\|:------:\|:-----:\|:-----:\|:------:\|:---:\|:---:\| \| \| Models \| Pub. \| \| \| \| Target \| \| \| \| \| \| \| \| \| \| UCFC \| XD \| LAD \| UBIF \| TAD \| ST \| AVG. \| \| \| \| E1 \| MMIL \| CVPR 18' \| 77.93 \| 81.34 \| 85.18 \| 85.44 \| 87.78 \| 84.39 \| 83.68 \| \| \| \| \| ARNet \| ICME 20' \| 79.26 \| 80.38 \| 85.27 \| 84.18 \| 89.57 \| 86.65 \| 84.22 \| \| \| \| \| WSAL \| TIP 21' \| 76.47 \| 78.35 \| 75.44 \| 86.41 \| 85.62 \| 82.44 \| 80.79 \| \| \| \| \| COMO \| CVPR 23' \| 80.41 \| 82.75 \| 86.24 \| 85.82 \| 90.13 \| 89.76 \| 85.85 \| \| \| \| \| Ours \| - \| 77.21 \| 82.09 \| 83.88 \| 91.9 \| 91.36 \| 91.12 \| 86.26 \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| E2 \| MMIL \| CVPR 18' \| 76.68 \| 74.92 \| 67.39 \| 82.4 \| 67.61 \| 61.86 \| 71.81 \| \| \| \| \| ARNet \| ICME 20' \| 77.05 \| 75.02 \| 78.98 \| 80.84 \| 75.09 \| 55.34 \| 73.72 \| \| \| \| \| WSAL \| TIP 21' \| 76.59 \| 73.75 \| 76.71 \| 79.86 \| 77.68 \| 55.54 \| 73.36 \| \| \| \| \| COMO \| CVPR 23' \| 77.07 \| 76.64 \| 77.43 \| 76.74 \| 78.67 \| 57.63 \| 74.92 \| \| \| \| \| Ours \| - \| 78.55 \| 77.68 \| 77.36 \| 82.53 \| 79.21 \| 60.41 \| 75.96 \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| E3 \| MMIL \| CVPR 18' \| 77.28 \| 72.75 \| 80.7 \| 85.29 \| 81.72 \| 61.33 \| 76.51 \| \| \| \| \| ARNet \| ICME 20' \| 75.48 \| 72.18 \| 79.9 \| 81.7 \| 79.43 \| 68.24 \| 76.16 \| \| \| \| \| WSAL \| TIP 21' \| 75.77 \| 72.18 \| 66.45 \| 82.57 \| 75.05 \| 78.05 \| 75.01 \| \| \| \| \| COMO \| CVPR 23' \| 70.02 \| 72.89 \| 80.59 \| 83.05 \| 74.73 \| 73.53 \| 75.80 \| \| \| \| \| Ours \| - \| 78.99 \| 75.8 \| 77.82 \| 85.75 \| 84.06 \| 76.23 \| 79.78 \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| E4 \| MMIL \| CVPR 18' \| 80.26 \| 82.51 \| 86.54 \| 89.88 \| 90.32 \| 89.34 \| 86.48 \| \| \| \| \| ARNet \| ICME 20' \| 80.88 \| 82.57 \| 86.72 \| 89.92 \| 90.69 \| 91.7 \| 87.08 \| \| \| \| \| WSAL \| TIP 21' \| 80.99 \| 81.6 \| 76.43 \| 87.87 \| 88.78 \| 87.47 \| 83.86 \| \| \| \| \| COMO \| CVPR 23' \| 80.61 \| 84.25 \| 86.88 \| 91.51 \| 91.74 \| 91.23 \| 87.70 \| \| \| \| \| Ours \| - \| 78.62 \| 82.71 \| 84.41 \| 94.42 \| 92.5 \| 91.17 \| 87.31 \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| Instead of exploring complex architecture designs for single-domain VAD models, this paper focuses on the analysis of multiple domain learning within the context of AC issues. In response to the reviewer's concerns, we compared MDVAD with four representative WVAD models (Section E): MMIL [42], ARNet [45], WSAL [29], and CoMo [7]. In the Table, proposed simple baseline demonstrates competitive performance. Notably, the average results show superior results in the E2 and E3, indicating better generalization and adaptation to unseen target domains. Various single-domain VAD models or backbones can be incorporated into the MDVAD task, showing a direction for future generalization work (please refer to the W2 response of 926M). --- Rebuttal Comment 1.1: Comment: I consider the methodology of this work to be another innovative approach to anomaly detection that is different from previous methods. Although there are some flaws in the work, it is a good starting point. I wish there were more types of approaches to anomaly detection. I will keep my marks. Good luck!	Summary: This paper proposes a new task called MDVAD, the goal of which is to effectively learn from multiple domains with different data distributions and definitions of abnormality without confusion, resulting in a general VAD model. To achieve this, the authors expand the traditional single-head framework to multiple-head framework for learning different knowledge and design an AC classifier to handle abnormal conflicts. The experimental results prove the effectiveness of the proposed method. Strengths: 1. This paper focuses on the problem of learning a generalizable VAD model, which is an important task. 2. The experiments conducted by the author are relatively compr Weaknesses: 1. This paper proposes a new task called MDVAD to achieve generalizable VAD by resolving conflicts in anomaly definitions. However, for any VAD application, the definition of normal or abnormal events should be explicitly determined according to the scenario requirements, rather than simply combining multiple datasets and resolving the abnormal conflicts. I find it difficult to understand under what practical scenario a VAD model trained using multiple datasets with abnormal conflicts is needed. 2. The writing of this paper is not clear enough, where some necessary training and inference details are missed. For example, the normal head training mentioned in NullAng-MIL is confusing. 3. This paper lacks a detailed description of the experimental setup. For example, if an anomalous event is determined to be a conflict, how should the model handle such an event? Technical Quality: 2 Clarity: 2 Questions for Authors: 1. (referring to weakness 1) In practical applications, what kind of scenarios conform to the task settings of MDVAD proposed by the authors? 2. (referring to weakness 3) During test phase, is it necessary to know which dataset the sample comes from? If a certain anomalous event is found to have conflicts in different datasets, how should it be handled? Do the multi-dataset evaluation and the test procedures for other methods use the same test data? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: I do not recognize obvious potential negative societal impact of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # [W1, Q1] Practical scenarios of MDVAD MDVAD’s practical relevance In real-world scenarios, performance degradation due to domain shift is a persistent issue for deep learning models. Consequently, various tasks have seen the introduction of domain adaptation and generalization methods. As reported in cross-domain evaluation performance studies [6, 7, 24], VAD models experience more than a 20~60% drop in AUC, largely due to the criteria for abnormality defined by each dataset. Models that perform well only within a single domain are significantly limited in their practical applicability. Additionally, collecting data for every possible scenario in real-world applications is challenging, and abnormal events occur rarely, resulting in data scarcity. To address these challenges, our approach leverages multiple VAD datasets and aims to learn a general model that effectively mitigates the ambiguity in defining abnormalities across datasets. Utilizing a generalizable VAD model A general VAD model trained on multiple domains offers two key benefits: * First, when data from the target domain in the real world is available (E1), a general model trained on multiple source datasets, including the target domain, is able to explore robust and general features, similar to the effects of multi-task learning. Additionally, a single generalized model eliminates the need for multiple specific models for different domains. * Second, when target domain training data is not available (E2), proper pre-training on multiple domains allows the general VAD model to embody generalized representations, leading it to better performance in unseen target domains. Furthermore, when a real-world target dataset is provided (E3:, E4), the pre-trained general VAD model can adapt well to the new domain, which is highly beneficial for practical VAD applications. We are grateful for the reviewer’s insightful comment and will incorporate this discussion on this aspect in Section 4.3 of our paper. # [W2] Training/Inference details of normal head in NullAng-MIL NullAng-MIL uses an angular margin to learn normal and abnormal features effectively. It adjusts the feature vector distances in the angular space both inter- and intra-class within each domain. Regarding the training of the normal head, when input training data is from domain $j$, the normal score$s^n $ is output from $j^{th}$ Normal head ${W}^n_{D_j}$ and abnormal score $s^a $ is output from ${W}^a_{D_j}$. ${s}^n_{D_j}=F \cdot {W}^n_{D_j}=\left\\| F \right\\|\left\\| {W}^n_{D_j} \right\\|cos{\Theta }^n_{D_j}$ In the equation, $F$ is the final embedding feature. By normalizing both the feature and the head, score $s$ is the angle in the angular space between the two vectors. For simplicity, the snippet index $i$ is omitted. When $F$ is the feature of a normal snippet, the angle with the normal head${W}^n_{D_j}$ should be narrower than the angle with the abnormal head. Conversely, when $F$ is from an abnormal snippet, the opposite should hold true. $\begin{cases}cos{\Theta }^a_{D_j} + m > cos{\Theta }^n_{D_j} & \textrm{when abnormal snippet}\\\cos{\Theta }^a_{D_j} < cos{\Theta }^n_{D_j}+m & \textrm{when normal snippet}\end{cases} $ Therefore, by adding the angular margin $m$, the feature learning process is designed to satisfy the above conditions. When training the normal head, other normal heads $\\{{W}^n_{D_d}\\}$ where $d \neq j$ do not affect the gradient, which is calculated as $\frac{ \partial s^n}{ \partial {W}^n_{D_j}}$. # [W3, Q2] Missing detail description Experimental setup The four protocols to verify the MDVAD task are as follows: * E1 Held-in: Integrating various domain’s anomalies into one framework, performing similarly to multi-task learning that handles multiple tasks with a single model. All source and target domains are accessed. A single model trained on $M$ (M=6) domains is evaluated on $M$ target domains. * E2 Leave-one-out: A methodology of pre-training on large data without target knowledge. The source and target are different, and the target is not accessed. Training with $M-1$ source domains in a held-out setting and applying to an unseen target dataset. * E3 Low-shot adaptation: Evaluates the adaptation ability of the general pre-trained model (E2). When few samples (10% of the training set) from the target domain are given, low-shot learning is performed, and the model is evaluated on the target domain. * E4 Full fine-tuning: Evaluates the performance of the general pre-trained model (E2) after full finetuning on the target domain’s training set. Handling abnormal conflict When performing multiple domain learning, inconsistent labels across domains can lead to abnormal conflicts, causing confusion in the model's training and making it difficult to develop a robust general model. Our proposed framework consists of domain-agnostic layers and multiple heads for different domains. While the heads are divided to prevent conflicts, the agnostic part extracts features from all datasets using a single branch. Therefore, to explore general features while being aware of abnormal conflicts, an auxiliary branch, the AC classifier, predicts conflicts by leveraging the variance in abnormal scores across multiple heads. Testing phase During the testing phase, target domain information is unnecessary in the E2~E4 settings because the source domain of the pre-trained general model is different from the target domain, it does not have a head for the target domain. As in Eq. 5, under the E1 setting, since we have information about the target domain, the target branch score is used as the final score. As in Eq. 5, we determine the final score with both normal and abnormal score to reflect conflicts by taking the maximum normal and abnormal score from multiple heads. All empirical studies of the MDVAD and the comparison models were evaluated and reported on the same source and target domains. --- Rebuttal Comment 1.1: Comment: I greatly appreciate the authors' responses and the additional experiments, which largely addressed my concerns. My comments about their responses are listed as follows. W1: The authors mention its application scenario as: "During the testing phase, target domain information is unnecessary in the E2~E4 settings because the source domain of the pre-trained general model is different from the target domain." However, I still remain skeptical about its practical significance. W2: The authors have re-elaborated on the training methodology, and the overall process is relatively clear. Based on the above considerations, I still have doubts regarding the practical relevance of the task proposed by the author. As such, I will maintain my current score for the time being and am open to hearing the opinions of other reviewers. If my doubts are resolved through further discussion, I will not refuse to increase my score accordingly. --- Reply to Comment 1.1.1: Comment: We are pleased that our rebuttal has largely addressed the reviewer's concerns. We sincerely appreciate reviewer's thoughtful comment. # [W1] Practical Applications Problem Definition of Domain Generalization The goal of Domain Generalization (DG) is to learn a robust and generalizable predictive function from the $M$ training domains to achieve minimal prediction error on an unseen test domain that cannot be accessed during training. The difference between Domain Adaptation (DA) and DG is that DA has access to the target domain data while DG cannot see them during training. This makes DG more challenging than DA but more realistic and favorable in practical applications [R1]. Various approaches, such as zero-shot learning (E2), adaptation learning (E3, E4), meta-learning, lifelong learning, and transfer learning, are employed to address the emergence of unknown target domains in real-world scenarios. Foremost, we would like to clarify that the focus is on the fact that the “Even without Target Domain information, the model performs comparably to an In-Domain (Single) Model”, rather than on the idea that “Target Domain information is unnecessary.” In real-world scenarios that an unknown target domain emerges, the pre-trained general model learned from multiple source domains can operate using Eq. 5 without Target Domain information. Moreover, if samples from the target domain are provided, the model can adapt by tuning the multiple heads. As mentioned previously, the ability to handle multiple domains with a single model (E1) is also a practical scenario for generalization. Because requiring a separate model and training for each domain is neither practical nor robust. Similarly, issues such as Domain Shift, Overfitting, Label Conflict, and Efficiency are all challenges that need to be addressed for practical applications. We acknowledge that our paper represents the first step toward practical VAD applications, and it opens up the possibility for various future works that could lead to valuable applications. We hope that the reviewer's concerns have been resolved, and we will include a discussion on this practical perspective in the manuscript. [R1] Wang, Jindong, et al. "Generalizing to unseen domains: A survey on domain generalization." IEEE transactions on knowledge and data engineering 35.8 (2022): 8052-8072.	Rebuttal 1: Rebuttal: We thank the reviewers for their thoughtful feedback. We tried to address all the questions with references to weaknesses (W) and questions (Q). We are glad the reviewers found that * Addressing an Important and Generalizable Problem * Novelty of Method and Effectively Solving Identified Problems * Thorough Experiments and a Systematic Protocols for Model Generalization # Summary of rebuttal In response to the reviewers’ feedback, we will improve the paper as follows: - We have added additional virtual UBNormal datasets for multiple-domain learning and open-set VAD experiments with discussions (reviewer 926M). - We have expanded experiments on additional baseline and state-of-the-art comparison methods with detailed discussions (reviewer SwSy and 926M). - We have provided clear explanations and detailed descriptions regarding the theoretical analysis, practical applicability, experimental settings, and the role of the AC classifier (reviewers SwSy, 2cHp, and 926M). - We have included further discussion about other alternative paradigms and future work with vision-language models (reviewer beRv). # Results added using virtual data In response to Reviewer 926M's comment regarding the discussion on utilizing virtual datasets, we have included the UBNormal dataset [i] alongside the MDVAD to conduct multi-domain learning and validate the model's generality in settings with significant scene discrepancy. 1. Multi-domain learning with virtual domain dataset: \| \| \| \| \| \| \| \| \| \|---------------------\|-------\|-------\|--------\|-------\|-------\|-------\|-------\| \| Baseline \| \| \| Target \| \| \| \| \| \| \| UCFC \| XD \| LAD \| UBIF \| TAD \| ST \| AVG. \| \| Source: MDVAD \| \| \| \| \| \| \| \| \| MIL \| 80.05 \| 83.77 \| 86.01 \| 85.76 \| 88.92 \| 88.82 \| 85.56 \| \| MIL+AC \| 80.11 \| 83.91 \| 85.15 \| 87.72 \| 90.05 \| 87.98 \| 85.82 \| \| NullMIL \| 79.01 \| 81.96 \| 85.08 \| 93.06 \| 90.57 \| 91.04 \| 86.79 \| \| NullMIL+AC \| 79.15 \| 82.96 \| 85.82 \| 92.41 \| 91.16 \| 89.67 \| 86.86 \| \| NullAngMIL \| 76.32 \| 82.74 \| 82.32 \| 92.30 \| 91.82 \| 91.26 \| 86.13 \| \| NullAngMIL+AC \| 77.21 \| 82.09 \| 83.88 \| 91.90 \| 91.36 \| 91.12 \| 86.26 \| \| \| \| \| \| \| \| \| \| \| Source: MDVAD + UBN \| \| \| \| \| \| \| \| \| MIL \| 78.59 \| 81.79 \| 85.06 \| 86.6 \| 88.83 \| 87.99 \| 84.81 \| \| MIL+AC \| 78.14 \| 81.78 \| 84.92 \| 85.08 \| 90.77 \| 89.12 \| 84.96 \| \| NullMIL \| 78.74 \| 83.34 \| 85.93 \| 91.28 \| 90.05 \| 88.69 \| 86.34 \| \| NullMIL+AC \| 78.76 \| 83.19 \| 86.01 \| 92.63 \| 90.73 \| 90.54 \| 86.98 \| \| NullAngMIL \| 77.09 \| 81.01 \| 83.96 \| 92.88 \| 91.57 \| 90.04 \| 86.09 \| \| NullAngMIL+AC \| 77.66 \| 82.09 \| 83.01 \| 92.55 \| 91.33 \| 91.07 \| 86.29 \| \| \| \| \| \| \| \| \| \| The UBNormal (UBN) dataset is a VAD benchmark proposed for open-set scenarios to handle unexpected abnormal events. Both normal and abnormal events are available during training, but the anomalies that occur during inference belong to a distinct set of anomaly types (categories). To alleviate the difficulty of collecting abnormal event data in the real world, UBN consists of synthetic videos, unlike other VAD datasets. There are substantial abnormal conflicts and differences in the visual settings of scenes compared to other domains. The table shows the performance in each target domain when trained with MDVAD and UBNormal under the E1: held-in setting. In the single-head MIL baseline, there is a performance drop when trained with MDVAD+UBN, indicating difficulty in handling Abnormal Conflict (AC). However, the model trained with multiple heads and the AC classifier shows improved results. By leveraging the virtual dataset, we can overcome data limitations and create a general model capable of handling diverse and complex scenes. 2. Openset VAD results \| \| \| \|----------------------------\|----------------\| \| Open-set VAD \| \| \| Single source: UBN \| / Target: UBN \| \| MIL \| 75.13 \| \| \| \| \| Multi source: MDVAD + UBN \| / Target: UBN \| \| MIL \| 67.95 \| \| MIL+AC \| 70.56 \| \| NullMIL \| 72.14 \| \| NullMIL+AC \| 70.94 \| \| NullAngMIL \| 74.42 \| \| NullAngMIL+AC \| 74.54 \| \| \| \| We conducted experiments in an open-set scenario using UBN, where the abnormal categories in the train set and test set do not overlap. As shown in the table, despite domain discrepancies and AC, the model effectively handles multiple domain learning, demonstrating that general feature learning can adequately address unseen abnormal categories. [i] "Ubnormal: New benchmark for supervised open-set video anomaly detection." CVPR. 2022.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Personalized Federated Learning with Mixture of Models for Adaptive Prediction and Model Fine-Tuning	Accept (poster)	Summary: This paper introduces a personalized federated learning algorithm to address the challenges of real-time predictions in non-stationary environments. Clients fine-tune models online, combining their locally fine-tuned models with multiple federated models learned over time. This approach ensures efficient adaptation to evolving data streams, with theoretical analysis and experiments on real datasets demonstrating its effectiveness. Strengths: - The proposed algorithm effectively addresses the challenge of making real-time predictions in non-stationary environments by allowing clients to fine-tune models online, ensuring continuous adaptation to evolving data streams. - By combining locally fine-tuned models with multiple federated models, the approach enhances personalization and leverages the strengths of both local and federated learning, resulting in improved performance. - The paper provides a solid theoretical analysis alongside experimental validation on real datasets, demonstrating the practical effectiveness and robustness of the proposed algorithm in real-world scenarios. Weaknesses: 1. Contributions are suggested to list by items for clear summaries. 2. The baselines in Table 1 are all before the 2022 year, more latest related methods published in 2023 should be compared. 3. Fed-POE has limited improvements on Air and FMNIST datasets. 4. The process of combining locally fine-tuned models with multiple federated models may introduce significant computational overhead for clients, especially those with limited resources. 5. As the number of clients increases, managing and integrating multiple personalized models can become complex, posing scalability challenges for the proposed algorithm. Technical Quality: 3 Clarity: 3 Questions for Authors: No Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for taking the time to review our paper and providing your valuable comments. Please find below our responses to your comments and questions. We will revise the presentation of the contributions section in the introduction by breaking down the last paragraph into itemized points. ## Baselines in Table 1 For the results in Table 1, we used online kernel-based models to evaluate the performance of Fed-POE in a convex setting. We believe that we compared the performance of Fed-POE against all state-of-the-art online federated kernel learning algorithms which provide rigorous theoretical guarantees. However, all these works were for the year 2022 and before. Moreover, if one employs Fed-DS, which is a baseline from 2023, for the online federated kernel learning problem investigated in Table 1, it is equivalent to Fed-OMD in this specific case. Therefore, our results in Table 1 are relevant to the baselines from 2023. ## Improvements on Air and FMNIST Datasets We believe the performance of the proposed Fed-POE on the Air and FMNIST datasets demonstrates its advantage in real-time prediction tasks. The main challenge in real-time prediction arises when there is no prior information about the streaming data, making it difficult to evaluate model performance before the task begins. As discussed in Section 3.2 of the paper, it is theoretically unclear whether federated or local models perform better under such conditions. The performance largely depends on the dataset. Results presented in Section 5 confirm that the relative performance of algorithms varies with the dataset. Tables 1 and 2 show that for the CIFAR-10 dataset, federated models outperform local models, while for the Air dataset, local models perform better than federated models. Furthermore, for WEC and FMNIST, both local models and personalized federated learning models outperform non-personalized federated models. This variability complicates the decision between local and federated models. However, the results in Tables 1 and 2 indicate that Fed-POE consistently outperforms all baselines, albeit marginally on some datasets. This suggests that Fed-POE's performance is robust across different datasets, making it a reliable choice for real-time prediction tasks in the absence of prior information. ## Complexity We would like to clarify that at each time step, Fed-POE fine-tunes only one federated model while making inferences with multiple federated models. Given that the computational complexity of making inferences is usually considerably less than that of model fine-tuning, we believe Fed-POE does not introduce significant additional computational overhead. Furthermore, the number of federated models used by clients can be adjusted to ensure that the required computation and memory costs remain manageable for the clients. We analyze the computational complexity of Fed-POE. Let $C_F$ denote the number of computations required to fine-tune the model $f$, and let $C_I$ represent the number of computations required to make an inference with model $f$. Assume that the complexity of model selection in Algorithm 1 is negligible compared to fine-tuning and making inferences with model $f$. According to Algorithm 2, each client performs $2C_F + (M+2)C_I$ computations per time step. Therefore, the computational complexity of Fed-POE for each client is $\mathcal{O}(C_F + MC_I)$. Typically, fine-tuning deep neural networks with backpropagation requires significantly more computations than making inferences with them. Thus, if the model $f$ is a deep neural network, $C_I$ is negligible compared to $C_F$. In this case, the computational complexity for each client using Fed-POE is $\mathcal{O}(C_F)$, which is comparable to most state-of-the-art federated learning methods. ## Increase in the Number of Clients An increase in the number of clients does not pose scalability issues for Fed-POE compared to other state-of-the-art federated learning methods. As with most other federated learning algorithms, each client using Fed-POE only fine-tunes one model at each step and sends the update to the server.	Summary: This paper proposes a novel personalized federated learning algorithm, Fed-POE, which is designed for adaptive prediction and model fine-tuning in dynamic environments. It addresses the challenge of real-time predictions on streaming data by constructing a personalized model that combines a locally fine-tuned model with multiple federated models. Theoretical analysis and experiments on real datasets demonstrate its effectiveness in achieving sublinear regret bounds and improved online prediction accuracy. Strengths: 1. The paper proposes a unique ensemble method that dynamically combines local and federated models, which is a novel approach in the field of federated learning. 2. It provides a solid theoretical analysis, demonstrating sublinear regret bounds for convex models. 3. The paper is well organized. Weaknesses: 1. Although the presented method is novel, it is simply a combination of the previous personalized federated learning approaches as well as ensemble learning and provides comparably little conceptual originality. The contribution's main novelty seems to be that integrating results from prior models would be beneficial in mitigating catastrophic forgetting in online federated learning. 2. Experimental results show that the improvement in the accuracy of Fed-POE compared to other methods is not significant, but ensemble learning inevitably increases the computational overhead increase. The paper needs to analyze whether this trade-off is reasonable. 3. The paper needs more experiments to prove the effectiveness of the method, for example, for real-time predictions, the size of the old data replay is crucial, and the authors should design experiments to analyze the effect of batch size b on the experimental results. This paper also needs experiment results on the accuracy over the time step. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. The method in this paper does not significantly improve accuracy and even has a larger standard deviation. Can you give me more reasons to support your methods？ 2. The method is designed with two parts to mitigate catastrophic forgetting (old data replay and integration of multiple old models), The complex model updating process is unreasonable for real-time prediction. Can you design more ablation experiments to analyze these two parts？ Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: see the weaknesses and questions Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for taking the time to review our paper and providing your valuable comments. Please find below our responses to your comments and questions. We would like to briefly review the main contributions of this paper, which we believe are of interest and utility to the community. We hope the respected reviewer will consider these contributions. Section 3 of the paper highlights the challenge of choosing between local models and federated models in online federated learning when there is no prior information about data distribution. In Section 4.1, the paper proposes an ensemble federated learning algorithm that effectively leverages both local and federated models to provide more reliable performance for clients. Moreover, Section 4.2 demonstrates how federated learning can be employed to address catastrophic forgetting in real-time decision making. ## Performance Gain We believe the performance of the proposed Fed-POE in Section 5 demonstrates its advantage in real-time prediction tasks. The main challenge in real-time prediction arises when there is no prior information about the streaming data, making it difficult to evaluate model performance before the task begins. As discussed in Section 3.2 of the paper, it is theoretically unclear whether federated or local models perform better under such conditions. The performance largely depends on the dataset. Results presented in Section 5 confirm that the relative performance of algorithms varies with the dataset. Tables 1 and 2 show that for the CIFAR-10 dataset, federated models outperform local models, while for the Air dataset, local models perform better than federated models. Furthermore, for WEC and FMNIST, both local models and personalized federated learning models outperform non-personalized federated models. This variability complicates the decision between local and federated models. However, the results in Tables 1 and 2 indicate that Fed-POE consistently outperforms all baselines, albeit marginally on some datasets. This suggests that Fed-POE's performance is robust across different datasets, making it a reliable choice for real-time prediction tasks in the absence of prior information. ## Further Ablation Study To address your concerns, we performed additional ablation studies to analyze the effect of batch size $b$ on the Fed-POE performance. We conducted experiments on the CIFAR-10 dataset, varying the batch size $b$ and the number of models $M$ selected by each client to construct the ensemble model. The table below illustrates the results. As observed, the batch size $b=1$ results in the worst accuracy, mainly due to the forgetting process where models overfit to the most recently observed data. However, increasing the batch size from $b=10$ or $b=20$ to $b=30$ does not significantly improve the accuracy. Larger batch sizes may lead the model to perform better on older data, as the model is trained on older data over more iterations. Therefore, from this study, we conclude that a moderate batch size is optimal, considering that an increase in batch size leads to an increase in computational complexity. Based on these findings, we choose $b=10$. Furthermore, in Figure 1 in Appendix D, we illustrate the regret of Fed-POE and other baselines over time for the CIFAR-10 and WEC datasets. \| CIFAR-10 \| M=0 \| M=4 \| M=8 \| M=16 \| \| -------- \| ------- \| -------- \| -------- \| -------- \| \| $b=1$ \| 53.80% $\pm$ 6.71% \| 62.73%$\pm$8.29% \| 62.73%$\pm$8.29% \| 62.73%$\pm$8.26% \| \| $b = 10$ \| 65.55%$\pm$8.77% \| 66.50%$\pm$8.00% \| 66.54%$\pm$8.08% \| 66.46%$\pm$7.98% \| \| $b = 20$ \| 65.72%$\pm$8.62% \| 66.13%$\pm$8.20% \| 66.64%$\pm$7.94% \| 66.53%$\pm$8.00% \| \| $b = 30$ \| 65.83%$\pm$8.54% \| 66.32%$\pm$7.92% \| 66.24%$\pm$8.05% \| 66.39%$\pm$8.02% \| ## Complexity Similar to other federated learning algorithms, using Fed-POE each client only updates one federated model per step. Therefore, we believe that using Fed-POE does not impose significant additional updating complexity compared to other federated learning counterparts. Please find below the computational complexity analysis of Fed-POE. Let $C_F$ denote the number of computations required to fine-tune the model $f$, and let $C_I$ represent the number of computations required to make an inference with model $f$. Assume that the complexity of model selection in Algorithm 1 is negligible compared to fine-tuning and making inferences with model $f$. According to Algorithm 2, each client performs $2C_F + (M+2)C_I$ computations per time step. Therefore, the computational complexity of Fed-POE for each client is $\mathcal{O}(C_F + MC_I)$. Typically, fine-tuning deep neural networks with backpropagation requires significantly more computations than making inferences with them. Thus, if the model $f$ is a deep neural network, $C_I$ is negligible compared to $C_F$. In this case, the computational complexity for each client using Fed-POE is $\mathcal{O}(C_F)$, which is comparable to most state-of-the-art federated learning methods. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' detailed feedback. In light of these explanations, I will revise my score accordingly. --- Reply to Comment 1.1.1: Comment: Thank you very much for your feedback. We would be happy to address any further concerns or questions you may have.	Summary: The paper introduces an interesting perspective about the role of ensembles of models in federated learning. The provocative claim is the fact that federated learning is not always better than locally-trained models. This is contextualized in the field of not IID data and time-varying data generating processes. To address this issue the paper introduces from the theoretical point of view how to quantify the regret in federated and locally trained models. In addition it includes an analysis of non-convex models by managing an history of models. The overall impression about the paper is positive even if some points could have been better explored (in particular the part related to not IID data that is somehow the core of the paper). Strengths: - The paper introduces a theoretical evaluation of the gain produced by federated models w.r.t. locally trained models. This results show that federated learning is relevant only when models can be considered iid (hence averaging providing better results). This is somehow a know results but I appreciated the theoretical analysis - The proposed solution is to combine with a convex mean a locally-trained models with the federated models - This is further extended in case of non-convex models by considering an "history" of models to be used when needed (i.e., according to the loss) Weaknesses: - The federated models somehow includes the locally-trained model. I would have appreciated a further analysis about the fact that the two "sides" of the average model are related each other. - The setting in which eta and eta_c scales with T prevents adaptation in the long run (which is somehow the core of the paper). How to deal with that? - Federated learning typically takes also into account the complexity of the learning phase (i.e., the amount of info to be transmitted, e.g., the models). This is not quantified here. And this could be also a weak point in the fed-poe algorithm. Technical Quality: 3 Clarity: 3 Questions for Authors: See Weaknesses box Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See Weaknesses box Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for taking the time to review our paper and providing your valuable comments. Please find below our responses to your comments and questions. ## Relations between Federated Models and Local Models The federated model can differ significantly from the local models, especially when the data distribution among clients is heterogeneous. The degree of data heterogeneity influences the similarity between the federated model and the local models. Although the federated model is trained on local models, if the gradient trajectories of other clients differ significantly from those of a particular client, it is expected that the federated model will be substantially different from the local model of that outlier client. We can add this to the paper. ## Learning Rates in the Long Run If the time horizon $T$ is unknown which may often be the case in the long run, the doubling trick technique (see e.g., [R1]) can be effectively used to set the learning rates $\eta$ and $\eta_c$ while maintaining theoretical guarantees. The doubling trick is a well-known technique in online learning that adaptively sets the learning rates without knowing the time horizon. We add a note about this to the paper. [R1] N. Alon, N. Cesa-Bianchi, C. Gentile, S. Mannor, Y. Mansour, and O. Shamir, “Nonstochastic multi-armed bandits with graph-structured feedback,” SIAM J. Comput., vol. 46, no. 6, pp. 1785–1826, 2017. ## Complexity Let $C_F$ denote the number of computations required to fine-tune the model $f$, and let $C_I$ represent the number of computations required to make an inference with model $f$. Assume that the complexity of model selection in Algorithm 1 is negligible compared to fine-tuning and making inferences with model $f$. According to Algorithm 2, each client performs $2C_F + (M+2)C_I$ computations per time step. Therefore, the computational complexity of Fed-POE for each client is $\mathcal{O}(C_F + MC_I)$. Typically, fine-tuning deep neural networks with backpropagation requires significantly more computations than making inferences with them. Thus, if the model $f$ is a deep neural network, $C_I$ is negligible compared to $C_F$. In this case, the computational complexity for each client using Fed-POE is $\mathcal{O}(C_F)$, which is comparable to most state-of-the-art federated learning methods. Moreover, at each time step, each client sends one updated model to the server. Let $P_f$ denote the number of parameters in model $f$. Therefore, the amount of information that needs to be transmitted to the server is $\mathcal{O}(P_f)$, which is the same as most state-of-the-art federated learning methods.	Summary: This paper introduces Fed-POE, a novel personalized federated learning algorithm tailored for online prediction and model fine-tuning. Fed-POE creates an ensemble by integrating local models with those periodically contributed by the server over time. Theoretical analysis confirms that Fed-POE attains sublinear regret. Empirical results demonstrate that Fed-POE consistently surpasses the performance of both local and federated models across all evaluated datasets, which indicates that Fed-POE effectively leverages the advantages of both local and federated models. Strengths: - The technical content of the paper appears to be accurate, although I did not check all the details carefully. - This paper is generally well-written and structured clearly. - The experiments substantiate the main theoretical analysis, and the proposed algorithm demonstrates superior performance over the baseline methods Weaknesses: My primary concern is that the assertion the proposed algorithm can effectively harness the combined advantages of federated and local models is not clearly demonstrated within the theoretical bounds. The paper presents two principal theoretical results: Theorem 2 provides the regret upper bound for the proposed algorithm in convex scenarios, while Theorem 3 addresses non-convex cases. Both theorems establish sublinear regret bounds that are consistent with those for federated learning using a straightforward online gradient descent approach. I recommend enhancing the clarity of the proposed method's advantages in the theorems by incorporating assumptions about the data distributions. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for taking the time to review our paper and providing your valuable comments. Please find below our response to your review. The main advantage of the proposed Fed-POE compared to the straightforward online gradient descent approach is its ability to provide sublinear regret upper bounds for both global and personalized regret. The conventional method, as presented in Theorem 1, cannot guarantee a sublinear regret upper bound for personalized regret. The conventional online gradient descent approach only guarantees global regret. Fed-POE achieves sublinear regret upper bounds for personalized regret, as shown in Equations (12) and (20). Moreover, global regret upper bounds for Fed-POE are presented in Equations (11) and (19).	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering	Accept (poster)	Summary: The paper presents MutaPLM, a framework designed to interpret and navigate protein mutations using protein language models. This approach utilizes a protein delta network to capture mutation representations and employs a transfer learning pipeline with a chain-of-thought strategy to leverage knowledge from biomedical texts. Strengths: 1. This paper attempts to propose a general interpretable model for protein mutations. 2. This paper compiles a mutation-text multimodal dataset, providing an excellent benchmark for future work. 3. The code is available. Although I haven't had time to run it yet, I will try to run the code during the rebuttal phase to ensure the reproducibility of the experiments. Weaknesses: 1. PLM representations used in this study is the residue-level or protein-level embedding? If the mutation has very few residues, such as a missense mutation, will using protein-level embedding result in h∆ being too small? 2. Is it possible to provide some more practical mutation-related downstream task benchmark results? For example, predicting changes in protein properties or PPI? 3. Is it possible to compare the proposed method with the predictive results of embeddings extracted by AF, since the description information of the mutation may already be included in the structural changes predicted by AF before and after the mutation? 4. I do not deny that this is a good work, but perhaps it is more suitable for the benchmark and dataset track, because its method has limited innovation, and it has not verified its interpretability and performance on actual tasks related to protein properties. Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This paper discusses the limitations and points out the direction for future work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your appreciation of our model, dataset, and code. We address your concerns below. > Q1: Details of PLM representations and delta features As detailed in Appendix A.1, the PLM representations used in this study are residue-level embeddings. Regarding the scale of $h_{\Delta}$, we demonstrate the following: - We calculate the average L2-norm on the MutaPLM dataset, which is 9.90 for the wild-type representations $h_{wt}$ and 0.35 for the mutational representations $h_{\Delta}$. Since our dataset only involves single-site (missense) mutation, this indicates that a mutation with very few residues will NOT lead to $h_{\Delta}$ being too small. - The protein delta encoder amplifies $h_{\Delta}$, resulting in an average L2-norm of 1.04 for its outputs (delta features $z_{\Delta}$). Additionally, we argue that the orientation of $h_{\Delta}$ plays a more significant role in elucidating the evolutionary directions of mutations. > Q2: Experiments on prior mutation-related benchmarks Thanks for your suggestions. We have tested MutaPLM on several mutation downstream benchmarks, including Spike-ACE2 [1] and avGFP [2] involving protein-protein interaction (virus-receptor binding) and protein properties (fluorescence intensity). Details of the benchmarks and our implementation are presented in the global rebuttal R2, and the results and case analysis are presented in Table 3 and Figure 1 in the uploaded PDF. We observe that MutaPLM achieves competitive performance with fine-tuned mutation models on these realistic benchmarks. We plan to extend our experiments to more protein fitness benchmarks in a future revision of our paper. > Q3: Comparisons with AlphaFold representations Thank you for your insights. We manually inspect 500 test samples from our dataset and observe that 23 of them describe specific alterations of the protein structure explicitly. To explore the impacts of structural representations for mutation explanation, one can use the first row of MSA representations calculated by the EvoFormer of AF2 [3] as the residue-level protein representations and feed the representations of the wild-type and the mutant into an LLM. Unfortunately, calculating AF representations on our dataset would require approximately 4000 GPU hours. Hence, as stated in our limitations, we reserve analyzing alterations of 3D structures for future investigation. > Q4: This paper is more suitable for the benchmark and dataset track We argue that our contribution are three-fold, including the protein delta network, training strategies, and the dataset. We would like to address the innovations of our methodology, as recognized by Reviewer eM8T, in the following: - Compared with prior works, we are the first to model mutations explicitly with delta features. - We propose a chain-of-thought strategy for explaining and engineering mutations, which has not been explored in previous multi-modal protein-text LLMs [4, 5, 6]. For mutation interpretation, we have validated the interpretability of our model with qualitative evaluations in Figure 2 and Figure A2 in our paper, as well as additional quantitative evaluations on actual protein fitness benchmarks. For mutation engineering, we have reported the fitness optimization results on 6 realistic datasets in Figure 5 in our paper. Hence, we believe that our work bears potential in real-world applications, and is suitable for the main track of the conference. Refs. [1] Shifting Mutational Constraints in the SARS-CoV-2 Receptor-binding Domain during Viral Evolution. [2] Local Fitness Landscape of the Green Fluorescent Protein. [3] Highly Accurate Protein Structure Prediction with AlphaFold. [4] Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models. [5] ProtLLM: An Interleaved Protein-language LLM with Protein-as-word Pre-training. [6] ProLLaMA: A Protein Language Model for Multi-Task Protein Language Processing.	Summary: In the paper entitled "MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering," the authors proposed multimodal protein-textual language models for understanding the effect of mutation and performing protein engineering. They also build MutaDescribe, the first large-scale protein mutation dataset with rich textual annotations. Strengths: 1. The paper is generally well-written and easy to follow. 2. The authors have constructed the first comprehensive protein mutation dataset enriched with textual annotations. This dataset represents a significant foundation for future research in this field. 3. The MutaPLM framework introduced in this paper is innovative, particularly in its explicit modeling of mutations and its use of cross-modal transformers for multi-modal feature integration, enhancing its analytical capability. 4. By integrating large language models, the proposed framework significantly simplifies protein engineering, offering an intuitive tool that could be readily adopted by biologists for advanced research. Weaknesses: 1. The paper lacks a comparison with fine-tuned protein language models. Finetuned PLMs (ESM-1, ESM-2) have been validated to be powerful for various downstream tasks. For example, MLAEP(https://www.nature.com/articles/s41467-023-39199-6) and AugmentedESM(https://www.nature.com/articles/s41587-021-01146-5) 2. The paper did not prove why the textural annotation is necessary. From the ablation study, one can conclude that the labeled information from the textual annotation makes the model powerful. 3. The paper should add more discussion and experiments on why human-understandable notation is necessary. Human-understandable notations are not more informative compared with a conventional multi-label dataset. Moreover, LLMs may fail to deal with regression tasks, while finetuned PLMs can do better. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. The statement that "Protein language models (PLMs) fall short in explaining and engineering protein mutations" may need reconsideration. 1. Recent studies, such as those involving ESM-1/ESM-IF1, have demonstrated these models' effectiveness in zero-shot engineering tasks. This contradicts the assertion of inherent limitations due to architectural design and lack of supervision. See https://www.nature.com/articles/s41587-023-01763-2 and https://www.science.org/doi/full/10.1126/science.adk8946 2. The manuscript would benefit from a deeper discussion of the MLDE methodology, particularly in the context of fine-tuning pre-trained protein language models like AugmentedESM(https://www.nature.com/articles/s41587-021-01146-5). A comparative analysis between mutaPLM and MLDE-based methods(e.g. AugmentedESM) could provide more clarity on their respective performances. 3. Based on 2, further exploration of the role of textual descriptions in enhancing model performance would be advantageous. Clarification on how these descriptions integrate with the model to improve predictions would be helpful. 4. The performance of the model on regression tasks remains unclear. It would be instructive for the authors to include results or discuss how the model handles quantitative predictions in the context of protein functionalities. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive comments on our presentation, dataset, methodology, and application values. We address your concerns and answer your questions below. > Q1: Misleading statement of PLMs in mutation explanation and engineering We apologize for this misleading statement in our abstract. We clarify it as follows: - In our introduction, we argued that while PLMs have shown effectiveness in zero-shot mutation explanation and engineering [1, 2], their implicit modeling of mutations through evolutionary plausibility is not satisfactory for practical needs. - The inherent limitations due to architectural design and lack of supervision arise in modeling mutations explicitly. We will change this statement to: "However, due to architectural design and lack of supervision, PLMs model mutations implicitly with evolutionary plausibility, which is not satisfactory to serve as explainable and engineerable tools in real-world studies." > Q2: Comparison between MutaPLM and MLDE-based methods Thanks for your insightful consideration. If MLDE refers to Machine-Learning-based Directed Evolution, the discrepancies between MutaPLM and prior MLDE-based methods are as follows: - Explicit modeling of mutations. MutaPLM models mutations explicitly with delta features by an encoder-decoder architecture. In contrast, existing MLDE-based models [2,3] either model mutations implicitly with evolutionary plausibility or focus on the wild-type sequence instead of the discrapancies between the wild-type and mutant. - Textual supervision. MutaPLM is a general framework for mutations that allows knowledge transfer across different wild-type proteins through natural language supervision. However, prior MLDE-based models are fine-tuned on a single wild-type protein with a phenotype of interest. We perform comparisons between MutaPLM and fine-tuned PLMs on both mutation explanation and engineering. Please refer to our global rebuttal R1 for details of these baselines, and the PDF file for experimental results. We observe that MutaPLM consistently outperforms fine-tuned PLMs on both tasks, which demonstrates the effectiveness of our approach. Unfortunately, MLAEP is not comparable under our experimental settings, as the model is specifically designed for calculating the binding affinities of SARS-Cov-2 and its receptors and antibodies. > Q3: Further justification for the role of textual descriptions We argue that textual descriptions are indispensable for general protein mutation modeling due to the following: - Texts connect mutational knowledge from diverse proteins. The mutational effects of proteins are diverse and complicated, and the number of samples with a phenotype of interest is often limited. Therefore, constructing a multi-label dataset will lead to excessive classes and extremely imbalanced label distributions. In contrast, using texts helps combine supervision signals from diverse wild-type proteins and allows knowledge transfer across different phenotypes. This corroborates the intuition of CLIP models [4, 5] using texts instead of conventional labels for cross-modal supervision. - Texts provide a user-friendly interface for mutation explanation and engineering. As stated in our introduction, the evolutional plausibility calculated by PLMs cannot meet the practical needs in studying mutations. Texts allow MutaPLM to interpret novel mutational effects from multiple facets and engineer proteins with user-defined properties even when the phenotype of interest has not been seen during training. Such capabilities cannot be obtained on a multi-label dataset. For experimental justification, we have shown in our ablation studies that textual instructions bring 5.8% absolute gains in Recall@50 in mutation engineering. We further validate that the delta features have captured mutational knowledge through textual supervision. Specifically, we fine-tune the delta encoder and a regression head on two protein fitness datasets introduced in global rebuttal R2. From the results below we observe that textual supervision brings significant benefits to mutation explanation. \| Model \| Spike-ACE2 \| avGFP \| \| - \| - \| - \| \| w/o textual supervision \| 0.401$\pm$0.029 \| 0.579$\pm$0.016 \| \| MutaPLM \| 0.481$\pm$0.028 \| 0.593$\pm$0.032 \| > Q4: Applying MutaPLM on regression tasks Similar to existing LLMs [6], we acknowledge that directly applying MutaPLM on regression benchmarks may lead to sub-optimal outcomes. We discuss strategies for handling quantitative predictions are as follows: - Discretizing numeric values into several segments and instructing the LLM to predict the discrete values [6]. - Performing comparison (which mutation leads to increased fitness) instead of regression, partly inspired by the reward model in InstructGPT [7]. - Fine-tuning a regression head using delta features as additional inputs. The motivation is that the delta features have captured mutational knowledge from massive biomedical texts that could benefit regression tasks. We implement the third strategy (fine-tuning with delta features) on two datasets, including Spike-ACE2 and avGFP. Please refer to our global rebuttal R2 for implementation details. The results are displayed in Table2 in the uploaded PDF file, where MutaPLM achieves competitive performance with fine-tuned mutation models. Refs. [1] Language Models Enable Zero-shot Prediction of the Effects of Mutations on Protein Function. [2] Learning Protein Fitness Models from Evolutionary and Assay-labeled Data. [3] Low-N Protein Engineering with Data-efficient Deep Learning. [4] Learning Transferable Visual Models From Natural Language Supervision. [5] ProtST: Multi-Modality Learning of Protein Sequences and Biomedical Texts. [6] Tx-LLM: A Large Language Model for Therapeutics. [7] Training Language Models to Follow Instructions with Human Feedback. --- Rebuttal Comment 1.1: Comment: I appreciate the efforts made by the authors during the rebuttal. Most of my concerns are addressed. I will raise my score as positive. --- Reply to Comment 1.1.1: Comment: Thanks again for your favorable comments on our work! We are glad to have addressed most of your concerns. We are willing to provide additional information if you have any further questions.	Summary: The paper proposes a framework to 1). generate text-based mutation effects for mutated proteins and 2). propose new mutated sequences based on the function descriptions. The main module is an encoder-decoder network, which encodes the representations of mutated sequences and outputs the position and amino acid of the mutation. The network is first pretrained on the protein literatures and then fine-tuned on the mutation effects. Strengths: * The problem studied in this paper is novel and well-motivated: generate mutated sequences conditioning on the instructions, and generate mutation effects conditioning on the sequences. * The method is technically sound. * The paper is well-structured Weaknesses: Most issues are on the evaluation side. Rigorous evaluations are very important for the AI4Science applications. * Baseline Selection: The paper employs weak baselines for comparison. None of the baselines used have been specifically trained on mutations. This makes it difficult to accurately assess the true effectiveness of the method. * Lack of Temporal Evaluation: While the paper adopts a structural split for evaluation, which is acceptable, a temporal-based evaluation would be more ideal and realistic. A temporal split, where some proteins are held out based on their discovery time, would more accurately reflect real-world scenarios in scientific applications. * Weak Evaluation of Mutation Explanations: The use of GPT-4 to assess scientific explanations is not robust or scientifically sound. * Missing experimental details. The paper omits several crucial experimental details, which harms reproducibility and thorough understanding of the methodology. Specific areas lacking detail include: 1. explain in details how you tune the hyperparameters 2. what is the dataset for protein literatures? 3. When construct MutaDescribe, did you only use swissprot or the whole dataset? how did you extract the mutation explanations? How do you know whether it's expert-reviewed? Technical Quality: 2 Clarity: 2 Questions for Authors: See above. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your appreciation of our task, methodology, and writing. We address your concerns in evaluation as follows. > Q1: Additional supervised baselines. We have added supervised baselines, including fine-tuned PLMs, for both mutation explanation and engineering. Please refer to our global rebuttal (R1) for more details and Tables 1 and 2 in the uploaded PDF file for experimental results. We observe that MutaPLM consistently outperforms supervised baselines on both tasks, thereby demonstrating the effectiveness of our approach. > Q2: Temporal evaluation. We appreciate your insightful suggestion. We perform temporal splitting by extracting the publication dates of the corresponding literature from PubMed [1] for each mutation in our dataset. Mutations studied before 2022 are used as training and validation sets, while those studied in 2022 and 2023 comprise the test set. The train/valid/test set comprises 156K, 8K, and 1.6K samples, respectively. The experimental results for mutation explanation are below: \| Model \| BLEU-2 (%) \| ROUGE-L (%) \| \| ----------------------- \| ---------- \| ----------- \| \| ProLLaMA \| 0.69 \| 0.80 \| \| GPT4-0613 (5-shot, kNN) \| 9.30 \| 11.84 \| \| Fine-tuned ESM \| 6.90 \| 12.62 \| \| MutaPLM (Ours) \| 10.83 \| 16.51 \| The experimental results for mutation engineering are below: \| Model \| Accuracy (%) \| Recall@50 (%) \| \| -------------- \| ------------ \| ------------- \| \| Random \| 4.40 \| 0.81 \| \| ESM2-650M \| 34.76 \| 24.02 \| \| ESM+BioMedBERT \| 55.78 \| 44.04 \| \| MutaPLM (Ours) \| 58.50 \| 46.05 \| We observe that MutaPLM achieves promising performance on the temporal split and outperforms strong baselines, showcasing its potential in assisting real-world scenarios. We are working on evaluating other baseline models on the temporal split. > Q3: GPT-4 evaluation of mutation explanations. We take 500 samples from our test sets and recruit a postgraduate from a top university who majors in biology to assess the mutation explanations of MutaPLM following the same categorization protocol as GPT-4. Below is the confusion matrix for manual and GPT annotations. \| Human (below) / GPT-4 (right) \| Accurate \| Relevant \| Opposite \| Irrelevant \| \| ----------------------------- \| -------- \| -------- \| -------- \| ---------- \| \| Accurate \| 60 \| 27 \| 3 \| 7 \| \| Relevant \| 7 \| 86 \| 0 \| 19 \| \| Opposite \| 0 \| 1 \| 28 \| 9 \| \| Irrelevant \| 1 \| 7 \| 5 \| 240 \| We observe that GPT-4 evaluation is consistent with human experts on 82.8% cases, although it occasionally misclassifies accurate predictions into relevant, and relevant or opposite predictions into irrelevant. We will report and discuss the manual evaluation results in a future revision of our paper. > Q4.1: Hyperparameters Given the computational expense of the experiments, we do not specifically tune our hyperparameters. The rationales for our hyperparameter settings are as follows: - The learning schedule and LoRA rank are derived from prior LLMs [2, 3]. - The batch size is selected to maximize GPU memory usage. - The number of pre-training steps is determined based on convergence observations. - The number of fine-tuning steps is based on evaluating the validation loss every 10K steps. > Q4.2: Protein literature dataset As detailed in Appendix B.1, the protein literature dataset is collected from the Publication entry of proteins within UniProtKB/SwissProt [4] and PubMed [1]. We plan to publicly release this dataset in the future. > Q4.3: Details about MutaDescribe construction All mutations are collected from UniProtKB Reviewed (Swiss-Prot), ensuring each sample has undergone expert review. The mutation explanations are obtained from the Phenotypes and Variants -> Description entry for each protein. Refs. [1] PubMed: The Bibliographic Database. [2] ProtLLM: An Interleaved Protein-Language LLM with Protein-as-Word Pre-Training. [3] BioMedGPT: Open Multimodal Generative Pre-trained Transformer for BioMedicine. [4] UniProtKB/Swiss-Prot, the Manually Annotated Section of the UniProt KnowledgeBase: How to Use the Entry View. --- Rebuttal 2: Comment: Thanks for your newly-added experiments and clarifications! I have updated my score. --- Rebuttal Comment 2.1: Comment: Thank you again for your positive feedback and insightful suggestions for improving our evaluation! Should you have any additional questions or require further clarification, please do not hesitate to let us know.	null	null	Rebuttal 1: Rebuttal: We extend our gratitude to all reviewers for their positive comments and constructive feedback. We hope that our responses and additional experiments could address the shared concerns satisfactorily. > (R1) Additional supervised baselines While no prior work is specifically designed for text-based mutation explanation and engineering, we implement additional supervised baselines by fine-tuning existing protein language models or large language models. For mutation explanation, we incorporate the following models: - Fine-tuned ESM. We translate each residue representation of ESM2-650M [1] using a linear projection layer and fine-tune BioMedGPT-LM to explain mutation effects based on the translated features of the wild-type and mutant. - AugmentedESM. We modify the regression model in the original paper [2] by feeding the adaptive fitness score calculated by ESM2-650M and the amino acid sequence into BioMedGPT-LM for fine-tuning. For mutation engineering, we implement: - ESM+BioMedBERT. We apply a cross-attention layer that takes the last hidden representations of ESM2-650M as queries and the BioMedBERT [3] encodings of textual descriptions of mutational effects as keys and values. The outputs are fed into the language modeling head of ESM2-650M to calculate the probability distribution for each mutation. - BioMedGPT. We directly input the amino acid sequence of the wild-type protein and the desired mutational effects into BioMedGPT-LM [4]. We instruct the model to suggest a plausible mutation or an amino acid at the mutated position and perform fine-tuning. The experimental results are displayed in Table 1 for mutation explanation and Table 2 for mutation engineering in our uploaded PDF. We observe that MutaPLM consistently outperforms supervised baselines on both tasks, demonstrating the effectiveness of our approach. > (R2) Evaluation on protein fitness benchmarks To further justify the effectiveness of MutaPLM in interpreting mutations, we evaluate our model on two protein fitness regression datasets including: - Spike-ACE2 [5]: This a deep mutational scanning dataset that aims to predict the binding strengths between SARS-Cov-2 variants and its receptor ACE2, which is critical for identifying potentially dangerous strains of the virus. - avGFP [6]: This benchmark aims to predict the fluorescence intensity of GFP variants, which is beneficial for developing biomarkers. We first visualize MutaPLM's explanations for several mutations within the two datasets, as shown in Figure 1 in the uploaded PDF, finding them reasonable and insightful. Then, following prior works [7, 8], we adopt a low-$N$ setting with 192 randomly sampled training samples and 48 validation samples. We perform fine-tuning by feeding the adaptive fitness of ESM2-650M and the Delta features of MutaPLM into a 2-layer MLP to predict the fitness scores. We compare our model with Ridge regression, ESM2-650M, AugmentedESM [2], Augmented EVmutation [9], ConFit [8] and Tranception_L [10]. The experiment results, displayed in Table 3 in the uploaded PDF, show that MutaPLM achieves competitive performance with fine-tuned protein mutation models, indicating that *the Delta* features have captured protein mutational knowledge from natural language supervision**. We observed that MutaPLM demonstrates performance comparable to Tranception_L on the Spike-ACE2 dataset and surpasses it on the avGFP dataset. This outcome is partly due to the backbone PLM in MutaPLM, and we speculate that substituting the current PLM with Tranception_L could yield further performance improvements. We plan to address these aspects in a future version of our paper and extend our experiments to include additional protein fitness benchmarks. Refs. [1] Language Models of Protein Sequences at the Scale of Evolution Enable Accurate Structure Prediction. [2] Learning Protein Fitness Models from Evolutionary and Assay-labeled Data. [3] BioMedBERT: A Pre-trained Biomedical Language Model for QA and IR. [4] BioMedGPT: Open Multimodal Generative Pre-trained Transformer for BioMedicine. [5] Shifting Mutational Constraints in the SARS-CoV-2 Receptor-binding Domain during Viral Evolution. [6] Local Fitness Landscape of the Green Fluorescent Protein. [7] Low-N Protein Engineering with Data-efficient Deep Learning. [8] Contrastive Fitness Learning: Reprogramming Protein Language Models for Low-n Learning of Protein Fitness Landscape. [9] Mutation Effects Predicted from Sequence Co-variation. [10] Tranception: Protein Fitness Prediction with Autoregressive Transformers and Inference-time Retrieval. Pdf: /pdf/7d7a5a15ce8f9dad65da6d3ae3a2beac89016856.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
GTBench: Uncovering the Strategic Reasoning Capabilities of LLMs via Game-Theoretic Evaluations	Accept (poster)	Summary: This paper tries to evaluate the strategic reasoning abilities of LLM. Therefore, 10 games are chosen where LLMs is trying to solve the game. This paper includes various open- and closed-source LLMs into consideration and build a benchmark for easy evaluation. Strengths: Evaluating the strategic reasoning is important and the evaluation includes various LLMs into consideration. Weaknesses: The evaluation protocol is questionable. More comments and questions are in the following section. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Does the evaluation really evaluate the strategic reasoning? Basically the evaluation is letting the LLM to play as one of the player in the game. However, this is much like a decision making problem, especially when the opponent is also a LLM-agent, where the LLM agent is largely stationary. Therefore, I would like to ask the authors provide the justification about why the evaluation is about strategic reasoning, not the decision making? 2. Also about the strategic reasoning. The selected games only focus on competitive zero-sum games. what about general-sum and multi-player games? Are cooperative games, e.g., hanabi, also requiring strategic reasoning? Even further, mixed cooperative and competitive, e.g., soccer, need the strategic reasoning? I think the strategic reasoning is not well-defined and fully discussed. 3. Does the evaluation really unlock the abilities of LLMs? The evaluation is focusing the prompting. However, for games, especially unfamiliar games for LLMs, exploration is important. Therefore, a memory or long in-context learning of the exploration experience should be included for the evaluation of strategic reasoning in games. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: More limitations should be discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable and insightful comments! >Q1: Does the evaluation really evaluate the strategic reasoning? Basically the evaluation is letting the LLM to play as one of the player in the game. However, this is much like a decision making problem, especially when the opponent is also a LLM-agent, where the LLM agent is largely stationary. Therefore, I would like to ask the authors provide the justification about why the evaluation is about strategic reasoning, not the decision making? Conceptually, strategic reasoning and decision making are related yet distinct in their focus[1-2]: Decision making tends to be more (1) immediate and (2) focused on selecting the best option among alternatives, whereas strategic reasoning involves (1) long-term planning and (2) anticipating others’ actions. In Section 2.2, we explicitly describe the gameplay process as a Markov Decision Process, involving multiple players and multi-turn action execution by participants. Games in GTBench involve long-term planning and emphasize competition arising from others’ actions. Specifically, each game comprises multiple turns, with Breakthrough often taking over 20 turns per match. In each turn, the decision of an LLM agent is constrained by factors such as game rules, past moves from opponents, and the current game state. The agent must achieve long-term planning, such as constructing a FORK in board games which requires multiple turns, and maintain long-context reasoning, such as in Negotiation which involves reviewing past moves and actions, with participation from at least two players. From this perspective, GTBench aligns more closely with strategic reasoning than with immediate decision making. We agree that decision making is the foundation of strategic reasoning, and we will add further discussion on the scope of this paper in our next revision. Reference: [1]Kahneman, Daniel. Thinking, fast and slow. macmillan, 2011. [2]Dixit, Avinash K., and Barry Nalebuff. The art of strategy: a game theorist's guide to success in business & life. WW Norton & Company, 2008. >Q2: Also about the strategic reasoning. The selected games only focus on competitive zero-sum games. what about general-sum and multi-player games? Are cooperative games, e.g., hanabi, also requiring strategic reasoning? Even further, mixed cooperative and competitive, e.g., soccer, need the strategic reasoning? I think the strategic reasoning is not well-defined and fully discussed. We would like to highlight that GTBench does NOT solely focus on zero-sum games: the taxonomy of zero-sum versus general-sum is a major classification in GTBench (Table 1, column 2), and 4 out of the 10 games are general-sum, specifically Negotiation, Blind Auction, Pig, and Iterated Prisoner’s Dilemma. We agree that cooperative games and multi-player environments provide valuable evaluations for strategic reasoning. Two of our games, Negotiation and Iterated Prisoner’s Dilemma, already involve collaboration. The mentioned Hanabi is a purely cooperative environment, which poses challenges in identifying individual relative skills during gameplay. We will extend our GTBench to multi-player games. In fact, GTBench already supports it for some games. For instance, we implemented the game PIG with 3 players: GPT-3.5-turbo PromptAgent vs CoTAgent vs ToTAgent. We report the win rate of each player over 50 matches: \|agent\|PromptAgent\|CoTAgent\|ToTAgent\| \|---\|---\|---\|---\| \|Win rate\|46%\|42%\|12%\| We would like to emphasize that LLM strategic reasoning evaluation is still in its early stages. This paper aims to evaluate the strategic reasoning of LLM agents in competitive environments. Experimental results indicate that LLM agents are largely ineffective in these simple competitive games. Therefore, GTBench could serve as a unified starting point for this research domain. Our GTBench will be maintained long-term and will support more diverse and complex strategic scenarios in the future. >Q3: Does the evaluation really unlock the abilities of LLMs? The evaluation is focusing the prompting. However, for games, especially unfamiliar games for LLMs, exploration is important. Therefore, a memory or long in-context learning of the exploration experience should be included for the evaluation of strategic reasoning in games. We acknowledge that designing advanced exploration mechanisms and unlocking the abilities of LLMs in strategic reasoning is impactful. However, our current paper does not aim to fully realize the potential of LLMs in these games. This is because a unified strategic reasoning benchmark is needed before delving into such designs. The scope of this paper is to (1) provide a unified evaluation framework and (2) benchmark common LLM reasoning agents in various game-theoretic scenarios. Designing advanced strategic reasoning agents and fully unlocking model capabilities will be our next step in this domain. It is also worth noting that the implemented Tree-of-Thought (ToT) agent [1] integrates exploration into both thought space and action space through thought decomposition and generation. However, ToT is less effective due to (1) limited exploration depth and (2) inaccurate reward estimation. This underscores that designing effective exploration mechanisms for LLM agents in strategic reasoning remains an open problem, requiring specific design considerations. We will address this in future work and discuss it further in our next revision. Reference: [1]Yao, Shunyu, et al. "Tree of thoughts: Deliberate problem solving with large language models." Advances in Neural Information Processing Systems 36 (2024). --- Rebuttal Comment 1.1: Title: Thanks for the response Comment: 1. About the definition of the decision making and strategic reasoning. I do not agree about the strategic reasoning. Generally, the strategic behavior is mostly investigated in game theory. What game theory differs from RL is that it need to simulate the play process of the two or more players, thus the fictitious play is proposed. In this paper, it is still decision making, not strategic reasoning. Surely this would be a definition problem, but I think this is important. 2. Sorry for the zero-sum competitive claim. but your response do not address my concerns. does the cooperative, mixed cooperative & competitive games also need strategic reasoning? ——based on your definition, cooperative game is definitely need to infer other's action. but is ignored in your paper. 3. I agree that benchmarking LLM in games still in its early stage. but I am just skeptical about the claim that benchmarking LLM in games = evaluating the strategic reasoning, also that considering one player's action is strategic reasoning. Given that my score is the border reject, which is not that negative, and the author's rebuttal do not convince me about the suitability of the methods, I would keep my score. --- Reply to Comment 1.1.1: Comment: We thank you for your prompt response. 1. In our last response, we explicitly mention that the major difference between decision making and strategic reasoning is the consideration of other players, i.e., “(2) anticipating others’ actions” (also described in Section 2.2). We would appreciate further clarification on how the proposed definition ("simulate the play process of the two or more players") differs from ours. 2. Cooperative games do require strategic reasoning. We already highlighted that collaboration is one of the major categories in the taxonomy of GTBench in Table 1. Games such as the Iterated Prisoner’s Dilemma require cooperative strategies to maximize system rewards (please refer to the definition of Prisoner's Dilemma L575-576, and the definition of Tit-for-Tat L695-702 which both clearly mention cooperation). Thus, we would like to politely disagree with the claim that we ignore cooperative games. 3. In Lines 41-48, we mention that games provide rigorous rules and well-defined action/state spaces, involving multi-player interaction (section 2.2, the gameplay is denoted as Markov decision process), which is well-aligned to the goals of strategic reasoning. We appreciate further clarification on the concerns of evaluating strategic reasoning through games. Two-player games, i.e., considering one player’s action, are well-recognized as strategic reasoning in game theory literature [1-3]. Besides, the reviewer has mentioned the definition of strategic reasoning as “simulate the play process of the two or more players”, indicating that two-player games are strategic reasoning. Also, we have provided the 3-player Pig experiments in our previous response to support that GTBench supports multi-player environments. Reference: [1] Hedden, Trey, and Jun Zhang. "What do you think I think you think?: Strategic reasoning in matrix games." Cognition 85.1 (2002): 1-36. [2] Abramson, Bruce. "Control strategies for two-player games." ACM Computing Surveys (CSUR) 21.2 (1989): 137-161. [3]Gutierrez, Julian, Paul Harrenstein, and Michael Wooldridge. "Expressiveness and complexity results for strategic reasoning." (2015).	Summary: This paper proposes a benchmark for evaluating the strategic reasoning of LLMs. The benchmark includes ten games of various types. The authors use these games to conduct competitive experiments between LLMs and traditional methods, as well as LLM-vs.-LLM. The paper then analyzes the experimental results and model behavior, and examines the game-theoretic properties of LLMs. Strengths: 1. The paper is logically clear, understandable, and well-written. 2. The experiments are comprehensive. The authors evaluate comparisons between LLMs and traditional methods and LLM-vs.-LLM competitions. They include multiple open-source and closed-source models and tests of various prompting methods. 3. The authors evaluate game-theoretic properties, including Nash equilibrium with regret and Pareto efficiency. Weaknesses: I didn't find any significant weaknesses, only a few questions. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. In Section 4.1, why does the tree-like prompting strategy ToT still lag significantly behind MCTS? 2. Is there any reference to classifying games in the benchmark? Why is it classified this way? 3. Why does the model perform better in probabilistic and dynamic games than in completely deterministic games? Is it that LLM performs better or that MCTS performs worse, making LLM appear better? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have fully addressed the limitations in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable and insightful comments. >Q1: In Section 4.1, why does the tree-like prompting strategy ToT still lag significantly behind MCTS? There are two potential reasons: 1. Exploration Space: MCTS has a significantly larger exploration space compared to ToT. In our experiments, MCTS is allowed to execute up to 1000 simulations to determine the next action. However, due to the complexity of the game and the cost of tokens, it is infeasible for ToT to traverse such a large number of simulations. For complex games like Breakthrough, completing even one full game simulation is impossible for ToT. 2. Reward Backpropagation: MCTS uses the actual win/loss signal obtained from simulations as a reward to determine which action should be selected. In contrast, ToT relies on LLMs voting or grading the actions, which is less accurate compared to the reward mechanism used by MCTS. >Q2: Is there any reference to classifying games in the benchmark? Why is it classified this way? We primarily classify games based on their game-theoretic properties, such as complete versus incomplete information, dynamic versus static, and probabilistic versus deterministic scenarios, which are fundamental and widely recognized in the field of game theory [1-3]. After categorizing the games according to these properties, we then identified and summarized the most commonly preferred abilities (Table 1). Reference: [1]Fraser, Niall. Conflict analysis. Ed. Keith W. Hipel. North-Holland, 1990. [2]Osborne, Martin J. An introduction to game theory. Vol. 3. No. 3. New York: Oxford university press, 2004. [3]Lanctot, Marc, et al. "OpenSpiel: A framework for reinforcement learning in games." arXiv preprint arXiv:1908.09453 (2019). >Q3: Why does the model perform better in probabilistic and dynamic games than in completely deterministic games? Is it that LLM performs better or that MCTS performs worse, making LLM appear better? Completely deterministic games are those in which all players have full knowledge of the game's state and the actions of other players. In this case, search-based solvers such as MCTS exhaustively explore possible future states and use statistical sampling to make highly informed decisions based on full information. Thus, MCTS approaches near-optimal performance in complete games, which significantly outperforms LLM agents. However, some information is hidden or unknown to the players (such as dice values or poker cards) in incomplete and probabilistic games, making MCTS simulation unavailable due to the hidden full-state information or uncertainty. We agree that MCTS may not be optimal for probabilistic games such as poker games. We further implemented the well-known Counterfactual (CFR) solver that is proven to be effective in finding Nash equilibria for in-complete games [1-3]. Then we conduct MCTS-vs-CFR experiments in 100 matches on the Kuhn Poker environment. The win rate of the CFR solver is 54%, which shows a slight advantage compared to the MCTS solver. We then re-run GPT-4 experiments when playing against the CFR solver. We observe that GPT-4 w/ Prompt achieves 0.33 NRA when against CFR, indicating that the conclusions presented in our paper are still consistent. We will include more CFR results in our next revision. Reference: [1]Zinkevich, Martin, et al. "Regret minimization in games with incomplete information." Advances in neural information processing systems 20 (2007). [2]Tammelin, Oskari, et al. "Solving heads-up limit texas hold'em." Twenty-fourth international joint conference on artificial intelligence. 2015. [3]Moravčík, Matej, et al. "Deepstack: Expert-level artificial intelligence in heads-up no-limit poker." Science 356.6337 (2017): 508-513. --- Rebuttal Comment 1.1: Title: Reply to the Rebuttal by Authors Comment: Thanks for your response. I think it's an appropriate score.	Summary: The paper proposes a benchmark to understand the strategic reasoning capabilities of llms. The authors present a suite of game theoretic tasks with different structures to do this. They use different evaluation metrics like ELOs and Relative advantage to compare different llms and prompting methods. Strengths: - The paper is clearly written and well motivated. It provides some structure to the growing literature of strategic reasoning with llms. - A wide range of closed source, open source models are tested. A good set of prompts are used to test the models too! - I particularly liked table 1 and the selection of different tasks with different characteristics. - The normalized relative advantage is a good, interpretable metric - The framework and taxonomy are clear and easy to understand. - Section 4.4 gave some good insight into the types of errors made by llms - I also liked reading the analysis in section 4.3, in particular that code pretraining helps with strategic reasoning. Weaknesses: - Characterizing human performance would strengthen the paper - Including some qualitative reasoning traces of successes and failures might be insightful. - Minor: This paper would be an ideal fit for the datasets and benchmarks track, instead of the main track. I dont think it should be penalized for this though! Typos Line 79: Characterize Line 171: dynamic gaming → dynamic game Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors do a good job of addressing limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable and insightful comments! >W1: Characterizing human performance would strengthen the paper We provide a preliminary human opponent evaluation. Specifically, we selected 5 games from GTBench and organized matches with 5 graduate students. These participants are familiar with the rules but are not trained experts. Each human competed against MCTS and GPT-3.5-turbo in 10 matches for each game. We report the average NRA across these human participants. \| setting \| Tic-Tac-Toe \| Connect-4 \| Kuhn Poker \| Liar’s Dice \| Blind Auction \| \| --- \| --- \| --- \| --- \| --- \| --- \| \|Human vs MCTS\|-0.25\|-0.55\|0.09\|-0.60\|0.15\| \|MCTS vs Human \|0.25 \|0.55 \| -0.09 \| 0.60 \| -0.15 \| \|GPT-3.5-turbo vs Human \| -0.55 \| -0.60 \| -0.53 \| -0.88 \| -0.46 \| \|Human vs GPT-3.5-turbo \| 0.55 \| 0.60 \| 0.53 \| 0.88 \| 0.46\| Overall, our results indicate that even humans with moderate knowledge of these games outperform LLM agents significantly in strategic reasoning, in both complete and incomplete information scenarios. This underscores the limitations of current LLMs in strategic reasoning and highlights the need for improvement in this area. We will include these results and provide more detailed human-as-opponent evaluations in our next revision. >W2: Including some qualitative reasoning traces of successes and failures might be insightful. We have identified five common error patterns and documented the corresponding failure cases in Table A11 of our manuscript to provide a qualitative analysis. Here, we present a pair of success and failure logs from the Tic-Tac-Toe environment, with detailed annotations explaining the intent behind each step (the intent is derived from the CoT reasoning): Success Traces (GPT-4-turbo w/ CoT Agent beats Human) Step 1 (CoT Agent): Thought: As the first player, I have the advantage of making the first move. The center square is the most strategic position to start with, as it provides the most opportunities for creating a line of three. Action:<C2R2> Step 2 (Human):<C3R2> Step 3 (CoT Agent): Thought: Looking at the current state of the game, I can see that the center and one of the side squares are occupied. To maximize my chances of winning, I should aim to control one of the corners. This will give me multiple ways to create a line of three. Considering the available moves, I can choose either <C1R1> or <C3R1> to achieve this.Action:<C1R1> Step 4 (Human): <C3R3> Step 5 (CoT Agent): Thought: Looking at the current state of the game, my opponent has two marks in the third column. If they get another turn, they might place their mark in the first row of the third column and win the game. Therefore, I need to block this potential winning move. Action:<C3R1> Step 6 (Human): <C2R1> Step 7 (CoT Agent): Thought: Looking at the current state of the game, I notice that the opponent has two of their symbols in the third column. If they get a chance to place their symbol in the third row of the first column, they will win the game. Therefore, I need to block this potential winning move by placing my symbol in the third row of the first column. Action:<C1R3> CoT Agent wins! In this match, the CoT agent successfully blocked the human's connected pieces and identified its own winning move. Failure Traces (Human Beats GPT-4-turbo w/ CoT Agent) Step 1 (Human): <C2R2> Step 2 (CoT Agent): Thought:The opponent has taken the central position on the board, which is a strategic advantage because it opens up the most possibilities for creating a line of three symbols. To counter this, I should aim to block potential winning moves and set myself up for multiple winning paths. Placing my symbol in a corner can be a good strategy because it gives me two potential ways to create a line of three symbols (NOTE: correct and popular strategy). Among the corners, choosing <C1R1> or <C3R3> seems most advantageous as it does not immediately threaten by the opponent's current position but sets up a potential win in multiple directions. Action:<C1R1> Step 3 (Human): <C2R1> Step 4 (CoT Agent): Thought: Given the current state of the game, my opponent has placed their symbols in C2R2 and C2R1, forming a potential line vertically in column 2. My symbol is in C1R1. To maximize my chances of winning or at least blocking my opponent's path to victory, I should consider creating a two-front threat that forces my opponent to defend rather than attack (NOTE: A hallucination occurred here; the opponent already has two connected pieces, so the first priority should have been blocking the opponent). Placing my symbol in C3R3 would start to form a diagonal line from C1R1 to C3R3, and it also positions me to potentially create a vertical line in column 3. This move doesn't immediately block my opponent's potential win but sets up a strategic position for future moves.Action:<C3R3> Step 5 (Human): <C2R3> Human Wins! The CoT agent lost because it failed to correctly prioritize its actions. With the opponent already having two connected pieces and no immediate winning move available for the CoT agent, the primary focus should have been on blocking the opponent. We will add more qualitative results in our next revision. --- Rebuttal Comment 1.1: Comment: Thank you for your response! I'll keep my current score!	Summary: This paper introduces GTBench, a set of 10 different games to test how well large language models can think strategically. The author found that while LLMs struggle with complete and deterministic games like Tic-Tac-Toe and Connect-4, they perform better in incomplete uncertain games like poker and negotiation. Code-pretraining improves their strategic thinking abilities. However, advanced thinking methods like Chain-of-Thought and Tree-of-Thought don’t always help and can sometimes make things worse. The latest open-source models, like Llama-3, are getting closer in performance to commercial models like GPT-4. Common mistakes LLMs make include misunderstanding game rules, being over-confident, and making calculation errors. Strengths: 1. The paper is well-written and easy to understand. 2. The problem of evaluating LLMs' strategic reasoning abilities is meaningful. Creating such a benchmark is valuable for the research community. 3. The paper provides a detailed evaluation of LLMs across different game tasks. These tasks indeed measure the strategic reasoning of LLMs, even if some models already understand the optimal algorithms for those games. (For example, you could ask GPT-4 about the optimal strategy for some of these games, and it knows the optimal algorithm.) 4. The authors conducted extensive experiments using various base models, including reasoning methods like ToT and CoT. They had some interesting findings and analysis (concluded in the summary). Weaknesses: 1. The paper claims that measuring strategic reasoning capabilities with games is missing in existing benchmarks. However, there are other benchmarks, such as MAgIC released last year, that consider benchmarking LLMs' strategic behavior using games. While there are differences, this weakens the claim of novelty. 2. Some of the selected games, like Tic-Tac-Toe, have known optimal strategies and are not complex enough. These games might not fully challenge the advanced strategic reasoning capabilities of LLMs. Even though the current evaluation is useful, as a benchmark intended for future use, it should be capable of evaluating more advanced or adapted LLM agents. 3. The benchmark focuses on a set of 10 games. It’s unclear how well the findings generalize to other strategic scenarios, even similar types of tasks. The results appear to be quite case-by-case. A broader range of tasks and scalable evaluation frameworks would make the benchmark more comprehensive. 4. The experiments primarily involve LLMs and traditional solvers. There is a lack of evaluation against human opponents, which could provide more insights into the models' performance in real-world strategic interactions. As a benchmark, I also expect to have other opponents (for example, the optimal algorithm, the RL based agent). Technical Quality: 3 Clarity: 3 Questions for Authors: Could you address the weakness 1, and try to discuss weakness 2-4? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: If LLMs are trained on biased data, they might reinforce existing biases in strategic decision-making. Testing the LLM with different personas could show if this changes the game results and reveal potential biases. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your valuable and insightful comments! >W1: The paper claims that measuring strategic reasoning capabilities with games is missing in existing benchmarks. However, there are other benchmarks, such as MAgIC released last year. In Line 119, we meant to convey that some of the games and the taxonomy introduced in GTBench are not used by current benchmarks for evaluating LLM strategic reasoning. We recognize existing benchmarks that evaluate LLMs through games, such as Clemency, LMRL-Gym, and LLMArena, discussed in Section 2.1. While MAgIC focuses on the complex social and cognitive dimensions of LLMs through various games, our focus is on pure logical game-theoretic evaluation, which could complement our work. We will add more discussion about MAgIC and revise our statement in the next revision. >W2: Some games have known optimal strategies that might not fully challenge the capabilities of LLMs. A benchmark intended for future use should be capable of evaluating more advanced or adapted LLM agents. We would like to mention that only Tic-Tac-Toe and Connect-4 are solved games, meaning they have optimal strategies, while other games in GTBench remain complex and challenging without known optimal strategies. We are aware that some strategies for poker games are "approximately" optimal, but these strategies incorporate bluffing, which introduces uncertainty and prevents them from being universally optimal. Additionally, we would like to emphasize the following points: 1. GTBench includes both simple games like Tic-Tac-Toe and complex ones like Breakthrough, with up to 48 actions per turn. These complex games, popular in strategy competitions, pose a significant challenge for LLMs already struggling with simpler games. 2. LLM agents, even with optimal strategies, may still produce errors like hallucinations. Thus, comparing performance using statistics like the number of draws and deviations from the optimal strategy is valuable. 3. GTBench is flexible and extensible, with modular game implementations independent of LLM agent design. It will be maintained long-term, supporting more complex games and advanced agents in the future. We believe that the proposed GTBench will continue to be valuable for advancing LLM research. We will incorporate this discussion in our next revision. >W3: It’s unclear the generalization of findings to other scenarios. A broader range of tasks and scalable evaluation frameworks would make the benchmark more comprehensive. We would like to mention that GTBench focuses on gameplay and includes various strategic scenarios like competition, negotiation, and collaboration (Table 1). Our conclusions are based on general trends across all scenarios, not focus on specific types. For example, we compare the NRA of code-pretrained LLMs and chat LLMs across all scenarios to assess the effect of code pretraining, providing a broad overview of LLM agent performances. To address scalability, we offer adjustable action/state spaces and complexity, such as varying board sizes for Breakthrough and player numbers and winscore for PIG. As complexity increases, the advantage of powerful LLM agents like GPT-4 over GPT-3.5-turbo decreases, but our overall conclusions remain valid. Here are related results (GPT-3.5-turbo PromptAgent vs MCTS and GPT-4, reporting NRA/win rate over 50 matches): \|Breakthrough\|Column=3 (default) \| Column=4\| \| --- \| --- \| --- \| \|Gpt-3.5-turbo vs MCTS \|-1 \|-1\| \|Gpt-4 vs GPT-3.5-turbo \|0.32 \|0.26\| \|3-Player Pig (GPT-3.5-turbo)\|PromptAgent \|CoTAgent \|ToTAgent\| \| --- \| --- \| --- \| --- \| \|Win rate \| 46% \| 42% \| 12% \| \|Pig \|Winscore=20 (default) \|Winscore=30\| \| --- \| --- \| --- \| \|Gpt-3.5-turbo vs MCTS \|-0.44 \|-0.40\| \|Gpt-4 vs GPT-3.5-turbo \|-0.04 \|-0.06\| We conclude that the relative advantage between LLM agents diminishes as games become more challenging. For example, increasing the board size in Breakthrough from 3 to 4 reduces GPT-4's NRA from 0.32 to 0.26, though it still outperforms GPT-3.5-turbo. This occurs because participants tend to show similar skill levels in more complex games. We will explore more extensible settings and support diverse games in future revisions. >W4: There is a lack of evaluation against human opponents and other opponents (for example, the optimal algorithm, the RL based agent). We conducted a preliminary human opponent evaluation by selecting 5 games (2 deterministic and 3 probabilistic) from GTBench. Five graduate students who are familiar with the rules but not experts in these games, played 10 matches against MCTS and GPT-3.5-turbo for each game: \| setting \| Tic-Tac-Toe \| Connect-4 \| Kuhn Poker \| Liar’s Dice \| Blind Auction \| \| --- \| --- \| --- \| --- \| --- \| --- \| \|Human vs MCTS\|-0.25\|-0.55\|0.09\|-0.60\|0.15\| \|MCTS vs Human \|0.25 \|0.55 \| -0.09 \| 0.60 \| -0.15 \| \|GPT-3.5-turbo vs Human \| -0.55 \| -0.60 \| -0.53 \| -0.88 \| -0.46 \| \|Human vs GPT-3.5-turbo \| 0.55 \| 0.60 \| 0.53 \| 0.88 \| 0.46\| The average NRA across participants showed that humans with moderate knowledge significantly outperform LLM agents in strategic reasoning in both complete and incomplete information scenarios. This highlights the limitations of current LLMs and the need for improvement. We will include these results and provide more detailed evaluations in our next revision. We agree that MCTS may not be optimal for probabilistic games like poker. We implemented the Counterfactual Regret Minimization (CFR) solver, known for finding Nash equilibria in incomplete games[1]. In 100 matches of Kuhn Poker, the CFR solver had a 54% win rate over MCTS, showing a slight advantage. We re-ran GPT-4 experiments against CFR, and GPT-4 with Prompt achieved a 0.33 NRA, consistent with our paper's conclusions. We will include more solvers and these results in our next revision. Reference: [1]Zinkevich, Martin, et al. "Regret minimization in games with incomplete information." Advances in neural information processing systems 20 (2007). --- Rebuttal Comment 1.1: Comment: I appreciate the author's effort and the response to my comments. I was hoping for more solid work to give a higher score (like 7 or above). For example, adding more games, comparing with more solvers, or including some additional insights (perhaps thinking out-of-the-box). I understand that it’s hard to make big changes in a short time. I have increased my score and lowered my confidence. --- Reply to Comment 1.1.1: Comment: Thank you for your valuable feedback and support. We will incorporate additional results and details in our next revision. --- Rebuttal 2: Title: We thank you for your valuable comments! Comment: Dear reviewer VyBq, We would greatly appreciate your esteemed feedback, as this could be our last chance to resolve any outstanding issues or inquiries you might have. Should any elements of our work need additional explanation, we kindly ask you to inform us. We look forward to your valuable input and the chance to have a productive discussion to enhance our submission. Thank you!	Rebuttal 1: Rebuttal: ## General Response We appreciate all the valuable comments from the reviewers. We are pleased to know that our work is considered meaningful (Reviewer VyBq, L1Yq), valuable (Reviewer VyBq), comprehensive (Reviewer VyBq, PA1v), and insightful (Reviewer PA1v, HsvL). Here are the major changes we have made: 1. We have added Human-as-Opponent experiments for further analysis (Reviewer VyBq, HsvL). 2. We have included the Counterfactual Regret Minimization (CFR) solver as a more powerful opponent in probabilistic games (Reviewer VyBq, PA1v). 3. We have provided detailed success/failure traces for qualitative analysis (Reviewer HsvL). 4. We have introduced extensible experiments for scalability, such as variations in board size, multi-player settings, and different gameplay win scores (Reviewer VyBq, L1Yq). 5. We have clarified the scope and the goal of this paper. Further details are available in our individual responses. We are also open to providing additional clarifications and addressing any other concerns the reviewers may have.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Federated Ensemble-Directed Offline Reinforcement Learning	Accept (poster)	Summary: This paper proposed the Federated Ensemble-Directed Offline Reinforcement Learning Algorithm. The combination of offline RL and federated learning is interesting in addressing the training data insufficiency issue due to small pre-collected datasets. Strengths: The originality of this paper is relatively good, since the proposed Federated Ensemble-Directed Offline Reinforcement Learning Algorithm is effective in offline reinforcement learning. The quality and clarity are also clear, and this paper is actually well-written. The significance of this paper is obvious, because offline reinforcement learning is important in real-world scenarios. Weaknesses: 1. Some technical details need to be explained. For example, the ensemble learning and its role. 2. The novelty of this paper needs further clarification, and what is the main difference between this proposed method and existing studies? It seems that there is only a simple combination of two technologies. 3. Numerically, the authors could consider comparing their method with more baselines. There are some studies on federated learning for offline RL. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Some technical details need to be explained. For example, the ensemble learning and its role. 2. The novelty of this paper needs further clarification, and what is the main difference between this proposed method and existing studies? It seems that there is only a simple combination of two technologies. 3. Numerically, the authors could consider comparing their method with more baselines. There are some studies on federated learning for offline RL. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: 1. What is the technical drawback of the proposed method? E.g., the effectiveness of the agent weight by ensemble approach 2. Does this proposed method work for other RL algorithms? Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. We are delighted to know that the reviewer finds our work original, our problem significant, and our paper well written. Below, we address the reviewer's concerns and hope they will consider increasing their score. 1. Some technical details need to be explained. For example, the ensemble learning and its role. Response: The idea of ensemble learning in our algorithm is to use the data distributed among different clients to learn a federated policy collectively. In Section 4.1, we have mentioned that `ensemble heterogeneity' is one of the key challenges in federated offline RL. In our work, we proposed an ensemble approach to overcome this challenge. Specifically, our approach uses the performance of the local policies as a proxy to weigh each client's contribution to the federated policy. This ensures that policies from clients with higher performance data have a greater influence on the federated policy. We have explained our approach in Section 5.1, including the mathematical equation that translates our idea to an actual algorithmic step. In addition to the experimental results given in Section 6 that show the superior performance of our FEDORA algorithm, we have also included additional ablation experimental results in Appendix C.1 which show the significance of our ensemble method w.r.t. other ingredients of FEDORA. Please let us know if there are any specific aspects that need to be elaborated further. 2. The novelty of this paper needs further clarification, and what is the main difference between this proposed method and existing studies? It seems that there is only a simple combination of two technologies. Response: We respectfully disagree with the reviewer's comment that our proposed algorithm is ``is only a simple combination of two technologies''. We emphasize that a simple combination of federated learning and offline RL is insufficient, as we have explained in our paper (see Fig. 1 and the explanation there). Significant algorithmic innovations are necessary to overcome some unique challenges of federated offline RL, as we have explained in Section 4.1. Our contributions include four key innovations: $(i)$ Ensemble-Directed Learning over Client Policies (Section 5.1), $(ii)$ Federated Optimism for Critic Training (Section 5.2), $(iii)$ Proximal Policy Update for Heterogeneous Data (Section 5.3), and $(iv)$ Decaying the Influence of Local Data (Section 5.4). We have given detailed experimental evidence on the superior performance of this method, see Section 6 and Appendix. Moreover, we have done ablation experiments that show the importance of each of these proposed innovations. 3. Numerically, the authors could consider comparing their method with more baselines. There are some studies on federated learning for offline RL Response: We have demonstrated the superior performance of our FEDORA algorithm against four different baseline algorithms through simulation experiments, see Section 6.1 and Appendix C. We have also evaluated the performance of FEDORA in the real-world using TurtleBot, a two-wheeled differential drive mobile robot, see Section 6.1, and compared it with the same baseline algorithms. We have also included a video of this real-world demonstration. We sincerely believe that these experiments and real-world demo clearly show the superior performance of FEDORA against the standard baselines. We would appreciate further guidance on specific baselines the reviewer would like us to include in our comparisons. 4. What is the technical drawback of the proposed method? E.g., the effectiveness of the agent weight by ensemble approach Response: We address our limitations in Appendix E. We make the assumption that all clients have the same MDP model (transition kernel and reward model), and any statistical variances between the offline datasets are due to differences in the behavior policies used to collect the data. In future works, we aim to broaden this to cover scenarios where clients have different transition and reward models. Reading the second remark about the effectiveness, please note that we have already included detailed ablation experiments to analyze the effectiveness of different components of our algorithm, see Appendix C1 and C2. 5. Does this proposed method work for other RL algorithms? Response Indeed! The FEDORA framework that we propose is general and can work with any actor-critic-based offline RL algorithms. We have mentioned this in our paper, please see lines 126-128. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' efforts in the detailed response. My concern has been addressed. I will raise my score to weak accept.	Summary: The authors identify fundamental challenges for Federated Offline Reinforcement Learning and present Fedora, an approach that tackles each of them. They perform extensive evaluation of the approach on Mujoco and real-world datasets showing improved performance over existing work. Strengths: The paper is well-written and, importantly, the code has been shared. The authors run extensive experiments. The work is novel and the notion of federated optimism is particularly interesting. Federated offline RL is an important research area with vast real-world applicability. The algorithm has been shown to be robust to diverse/ heterogenous client datasets. It is also commendable that the approach was tested on a real-world robot. Weaknesses: No theoretical guarantees have been given for the algorithm though it does build upon foundational work. I believe that the authors should explicitly discuss limitations/ opportunities for future work in the paper. It is important for the algorithm pseudocode to be included in the main material as is the norm in such papers. I believe that there are perhaps many experiments included the main paper meaning that the discussion/ hypotheses for results is somewhat diluted. Another minor issue is that the figures are placed very far away from where they are referred to in text. Technical Quality: 3 Clarity: 4 Questions for Authors: * How far do the authors perceive that this model can be pushed? i.e. the assumption that all clients have the same MDP is restrictive but understandable for a first set of experiments. * Have any experiments been run using D4RL-random datasets? It would be interesting to see whether this collapses learning. With regards to FEDORA outperforming centralised training I think a deeper discussion on this would be useful. * What is the main reason for this? Heterogenous data though previous work has successfully mixed datasets: https://arxiv.org/abs/2106.06860 Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: Limitations should be explicitly stated. I feel that the authors could give a more balanced view of the algorithm by not only showing strengths but also assessing the limits of the work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and are happy to note that they find our work novel, our experiments extensive, and our paper well-written. Below, we address their concerns and hope that they consider increasing their score. 1. No theoretical guarantees have been given for the algorithm though it does build upon foundational work. Response: Thank you for your comment. Providing a theoretical guarantee for our FEDORA algorithm is indeed a challenging problem that requires technical analysis of multiple complicated components, including offline policy evaluation, pessimistic estimation, ensemble-style quality-based federation, dealing with heterogeneous data, and analyzing these components jointly to derive the final performance bound. We, however, emphasize that, to the best of our knowledge, ours is the first paper on federated offline deep RL with an algorithm that performs really well in a variety of settings. Our design is analytically driven, identifies each issue of learning from the ensemble of policies, and builds up the algorithm methodically one step at a time corresponding to each analytical insight. We sincerely believe that our rigorous empiricism-driven approach is valuable based on its algorithmic contributions. 2. I believe that the authors should explicitly discuss limitations/ opportunities for future work in the paper. Response: We discuss the limitations and future directions of our research in Appendix E. For the final submission, we will move this discussion to the main part of the paper to make it more accessible and prominent. 3. It is important for the algorithm pseudocode to be included in the main material as is the norm in such papers...Another minor issue is that the figures are placed very far away from where they are referred to in text. Response: We pushed the pseudocode to the appendix due to space constraints. However, for the final version, we will move the pseudocode back to the main text, as we will have an extra page for the final submission. Additionally, we will adjust the placement of figures to be closer to the relevant text. 4. How far do the authors perceive that this model can be pushed? i.e. the assumption that all clients have the same MDP is restrictive but understandable for a first set of experiments Response: We plan to extend FEDORA to a meta federated learning setting, wherein we can learn with clients having different transitions and reward functions. This extension is discussed in Appendix E, and we aim to explore this direction in future work. 5. Have any experiments been run using D4RL-random datasets? It would be interesting to see whether this collapses learning. With regards to FEDORA outperforming centralized training I think a deeper discussion on this would be useful. Response: Yes, we have run experiments using the D4RL random dataset and compared it with centralized training (See Figure 3 in Section 6.1). We also conduct experiments with clients having different datasets (including random datasets) in Appendix C.6. 6. What is the main reason for this? Heterogeneous data though previous work has successfully mixed datasets Response: The use of heterogeneous datasets in centralized offline RL is a significant challenge. One reason for the drop in performance when pooling data from behavior policies with different expertise levels is that it can exacerbate the distributional shift between the learned policy and the individual datasets, leading to poor performance [1]. References [1] Tianhe Yu, Aviral Kumar, Yevgen Chebotar, Karol Hausman, Sergey Levine, and Chelsea Finn. Conservative data sharing for multi-task offline reinforcement learning. Advances in Neural Information Processing Systems, 34:11501–11516, 2021.484 --- Rebuttal Comment 1.1: Title: Response noted Comment: Thank you for engaging with the review. I think a brief discussion of the limitations should be in the main paper. Please ensure that all other promised changes are made: I think that my comments can be used to move around some important content into the main paper. If the concerns are addressed then I will stick to my original score.	Summary: This paper presents the Federated Ensemble-Directed Offline Reinforcement Learning Algorithm (FEDORA), a novel approach for collaborative learning of high-quality control policies in a federated offline reinforcement learning (RL) setting. The paper identifies key challenges in federated offline RL, including ensemble heterogeneity, pessimistic value computation, and data heterogeneity. To address these issues, FEDORA estimates the performance of client policies using only local data and, at each round of federation, produces a weighted combination of the constituent policies that maximize the overall offline RL objective, while maximizing the entropy of the weights. Besides the core idea, FEDORA also performs data pruning Strengths: 1. This is a novel work proposing the first federated offline RL algorithm in the general case (without assuming linearity). The paper is very well written with clear motivations and detailed discussions on the insufficiency of existing, naive approaches. 2. The experiments are also very thorough and convincing with experiments ranging from simple 2D environments to high-dimensional continuous control problems. The algorithm is also tested on a real-world robot platform, which is very impressive given the density of algorithmic contributions in the paper. Weaknesses: 1. "Collect wisdom" can be replaced by more rigorous exposition. Same goes with "ambitious targets". 2. The number of communication rounds needed for FEDORA to converge is still quite high. 3. Given how well the algorithm does, some sort of theoretical analysis could further strengthen the work. Technical Quality: 4 Clarity: 4 Questions for Authors: My questions are stated above. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Yes, limitations are adequately discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive endorsement of our work. We are happy to know that the reviewer finds our work novel, our experiments extensive and our paper well written. Below we address the concerns of the reviewer. 1. "Collect wisdom" can be replaced by more rigorous exposition. Same goes with "ambitious targets". Response: We thank the reviewer for their suggestion, we will incorporate this in the final version of the paper. 2. The number of communication rounds needed for FEDORA to converge is still quite high. Response: The number of communication rounds that FEDORA takes depends on factors such as the complexity of the problem and the number of local epochs performed during each round of federation. We believe that the number of communication rounds can be reduced by increasing the number of local epochs performed in each round of federation. 3. Given how well the algorithm does, some sort of theoretical analysis could further strengthen the work. Response: Thank you, we are indeed working on this problem. Providing a theoretical guarantee for our FEDORA algorithm is a challenging problem that requires technical analysis of multiple complicated parts corresponding to offline policy evaluation, pessimistic estimation, ensemble-style quality-based federation, dealing with heterogeneous data, and analyzing these components jointly to get the final performance bound.	null	null	Rebuttal 1: Rebuttal: ### Joint Response We would like to express our gratitude to all the reviewers for their time and feedback. We are delighted that the reviewers recognize the novelty of our work (hv6n, EkXX, EGVU), find our paper well-written (hv6n, EkXX, EGVU), and appreciate the comprehensiveness of our experiments (hv6n, EkXX). Below, we provide detailed responses to their queries. We look forward to a productive discussion during the reviewer-author period.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Neural Concept Binder	Accept (poster)	Summary: The paper proposes a novel approach to unsupervised concept learning based on both continuous and discrete encodings. Neural Concept Binder (NCB) allows humans inspecting and revising the learnt concepts. In the experiments, NCB’s discrete concept encodings result as expressive as the continuous encodings. Also, NCB can be integrated with symbolic and sub symbolic module. Finally, to support the experimental evaluation the paper introduces a novel dataset CLEVER-Sudoku, very suitable for neuro-symbolic benchmarking. Strengths: - Novelty: the proposed approach, although based on existing works SySBinder and Slot attention, is surely novel in the field of concept learning and potentially very relevant as it may strongly facilitate the extraction and discovery of unsupervised concepts. Particularly the possibility to revise concepts is completely novel to the best of my knowledge and very useful to improve human-computer interaction. - Novel resource presented: CLEVER-Sudoku will be surely an important resource for the Neuro-symbolic literature. Weaknesses: ## Major issues: * Method presentation: - The way in which block-slot-encodings are obtained, is badly presented. Although it is based on previous literature, since it is a key architectural component, it should have been presented more in details. I suggest the author to employ a background section to report the way in which slot attention and sys binder work, in order to make the paper self-contained. - Figure 2 which illustrates the core of the method is quite confusing: it is not clear how the discrete concepts are actually represented (the concept slot encodings reported are positive continuos representations). Also, the reminders to the figures in the text do not help as they generically refer to the entire figure and not to a specific block. A color coding of the different parts of the model could help understanding. - How does the $\texttt{enc}_l^j$ work is not clear. What does it receive in input? Where it is extracted from? - All the revision operations are definitely not clear. The formal operation to be executed is often confusing. * Experimental evaluation: - Models: NCB has been compared only against SysBinder. While it is a very novel and innovative method, there is a complete lack of benchmarking against standard unsupervised concept-based approaches such as SENN[1], BotCL[2], ACE[3]. Comparing against supervised approaches such as CBM[4] or CEM[5] could have been also useful. - Datasets: NCB is only tested on variants of CLEVER. While it is surely an interesting benchmark, real-world benchmarks are missing. Experiments on CUB or CELEBA, for instance, would have been very appreciated to better understand the scalability of the approach. ## Minor issues * Related work: - The unsupervised concept learning literature does not review several important concept-based paper working both post-hoc and explainable by design. Some examples are SENN[1], ACE[3], VAEL[6], as well as notorious prototype-based approaches such as Prototype layer [7] and ProtopNets[8]. - Unlike you state, continuous and discrete representations have been combined in recent literature for supervised concept learning. Some examples are CEM[5] and ProbCBM[9]. * Unclear sentences: - “Briefly, given an image x, NCB derives a symbolic representation, c, which expresses the concepts of the objects in the image, i.e., object-factor level concepts. Herefore, NCB infers a block-slot encoding, z, of the image and performs a retrieval-based discretization step to finally infer concept-slot encodings, c”. The consequentiality of the inference process is misleading from this sentence. * Method Inspection. What the authors refer as implicit, comparative, interventional and similarity-based inspections are normally referred to as example-based explanations (implicit and similarity-based) and counterfactual explanations (comparative and interventional). Sticking to well-known terms in literature is a good choice to avoid further confusion in the reader. Overall, I think it's an interesting paper proposing a novel approach to unsupervised concept learning. However, I think it will benefit from a further revision to deeply improve method presentation and expand the experimental campaing including other standard unsupervised concept-learning approaches and datasets. [1] Alvarez Melis, David, and Tommi Jaakkola. "Towards robust interpretability with self-explaining neural networks." Advances in neural information processing systems 31 (2018). [2] Wang, B., Li, L., Nakashima, Y., and Nagahara, H. “Learning bottleneck concepts in image classification”. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023). [3] Ghorbani, Amirata, et al. "Towards automatic concept-based explanations." Advances in neural information processing systems 32 (2019). [4] Koh, Pang Wei, et al. "Concept bottleneck models." International conference on machine learning. PMLR, 2020. [5] Espinosa Zarlenga, Mateo, et al. "Concept embedding models: Beyond the accuracy-explainability trade-off." Advances in Neural Information Processing Systems 35 (2022): 21400-21413. [6] Misino, Eleonora, Giuseppe Marra, and Emanuele Sansone. "Vael: Bridging variational autoencoders and probabilistic logic programming." Advances in Neural Information Processing Systems 35 (2022): 4667-4679. [7] Li, Oscar, et al. "Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 32. No. 1. 2018. [8] Chen, Chaofan, et al. "This looks like that: deep learning for interpretable image recognition." Advances in neural information processing systems 32 (2019). [9] Kim, Eunji, et al. "Probabilistic Concept Bottleneck Models." International Conference on Machine Learning. PMLR, 2023. Technical Quality: 3 Clarity: 1 Questions for Authors: A few questions to try to understand better the notation employed to define the revision operations. - What do the authors mean with $v_l \rightarrow v_m$? - What does the add operation work and how one can provide an encoding for a concept and be sure the network employs it as intended? Confidence: 4 Soundness: 3 Presentation: 1 Contribution: 3 Limitations: The method limitations are well addressed by the authors. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1 (More background): We agree that adding more information on Sysbinder and Slot Attention can help the reader and make the paper overall more self-contained. We have added an additional section with details for the camera-ready version and provide it in a comment below. W2 (Clarity of Figure 2): We agree and have modified the Figure: (i) we have updated the digits (that had previously represented the concept symbols) in the "Hard Binder", "Retrieval Corpus" and "Concept-slot encodings" components to the notation used throughout the experiments. Specifically, we now denote concepts in Figure 2 with a capital letter for the block and a natural number for the category ID, e.g., A3 for the third concept in the first block (A). (ii) Furthermore, we have added Roman numerals (I-VI) to each component in the figure and reference these specifically within the main text. (iii) We have updated the text accordingly for more clarity. Overall, we agree this greatly helps in understanding the individual components. Apologies for having missed that in the initial version. W3 ($\texttt{enc}_l^j$): Let us go through this question step by step. We base this on the more detailed training descriptions in A.2 (appendix). First, NCB infers the block-slot encodings of a set of input images. These are denoted as $\bar{Z}$ in Alg. 1 in the appendix. NCB then performs clustering of these encodings per block, i.e., it clusters the blocks j of each encoding (denoted $\bar{Z}^j$) and assigns each of them one cluster label, $v \in \{1, \cdots, N_C\}$. Per cluster, $v$, NCB next identifies exemplar and prototype encodings (i.e., representative encodings identified via the clustering method and cluster averages). Each of these exemplar and prototype encodings correspond to one $\texttt{enc}^j$ in the tuples, $(\texttt{enc}^j, v)$, that are stored in the retrieval corpus $\mathcal{R}^j$ of one block $j$. We now use the index $l$ to identify specific encodings out of $\mathcal{R}^j$, leading to $\mathcal{R}^j := \{(\texttt{enc}_l^j, v_l) : l \in \{1, ..., \|R^j\| \} \}$. In conclusion, $\texttt{enc}_l^j$ represents one block encoding that is stored in $\mathcal{R}^j$ and has been assigned to a specific cluster $v_l$. We agree that highlighing these additional steps benefit the understanding of the main text and have updated it accordingly. We apologize for the brevity in the initial submission. W4 (Revision operators): Indeed, we agree that the revision operators are somewhat confusing. It seems the attempt to formalize these steps has overcomplicated things. We have now removed the formal operations from section 3.3 and described these in words. E.g. we have posted the updated $\texttt{merge}$ description in a comment below. W5 (Additional baselines): We fully agree on the value of additional baseline models. We have, therefore, added the recent NLOTM model as a novel baseline for our evaluations. We refer here to our general response and hope the reviewer can agree on the relevance of NLOTM as a baseline model. However, we politely disagree that the methods suggested by the reviewer represent valuable baselines. Specifically, [1] cannot handle object-level concept learning, nor are the concept assignments discrete (an image is represented as a continuous mixture of concepts). [2] on the other hand, focuses on image regions as concepts. Similar issues hold for [3], i.e., they consider concepts to represent segments of an image and do not provide object-level concepts. Moreover, [3] focuses on post-hoc learning of concepts, which makes potential model revisions tricky and the inspection of the model ambiguous. Overall, we fully agree that the references to unsupervised concept learning are related to our work and we have noted them accordingly in our paper. However, none of them tackle the same problem setting as NCB aims to, i.e., learning discrete, object-level concept representations without supervision. We therefore consider the benefit of running these particular baselines as minor. Concerning supervised models ([4,5]): in fact, we had compared against supervised concept-based models in the context of Q2 and Q4 (denoted as SA (supervised), c.f. also F.7) whereby we observe that indeed NCBs unsupervised concepts are competitive in comparison to the GT supervised concepts of a slot attention-based CBM despite any supervision in training NCB. W6 (Additional datasets): We kindly refer here to the general response above. W7 (Related work): Agreed, we have added the missing references. We note that we had already referenced ProbCBM, though not in the context of discrete and continuous. We have now added this to this respective section. W8 (Unclear sentence): We apologize for the confusion here. We have updated the sentence (see comment below). W9 (Inspection terminology): We agree that connecting the terms example-based and counterfactual explanations can help to understand the outcome of the different inspection forms of NCB. We have now denoted comparative and interventional inspection as two ways to obtain forms of counterfactual explanations and implicit and similarity-based inspection as two ways to obtain forms of example-based explanations. Thanks for the hint! Q1 ($v_l$→$v_m$): We have removed this notation. It had previously described replacing $v_l$ with $v_m$. We refer here to the response above concerning revision operators. Q2 (add operator): We have updated the add revision as in a comment below. Overall, this operation can be used when the hard binder has missed concepts of the original encoding space. However, if NCB's soft binder has not learned to encode a specific concept at all (i.e., does not represent it in its block encodings) or the concept that a user wishes to add is not present in the dataset at all, one has to revert to the fourth type of revision, i.e., an additional finetuning of the soft binder is necessary. --- Rebuttal 2: Title: Additional background section Comment: Background The binding mechanism (SysBinder) of Singh et al. (2024) allows images to be encoded into continuous block-slot representations and relies on the recently introduced slot attention mechanism [2]. In slot attention, so-called slots, $s \in R^{N_S \times N_B D_B}$ (each slot has dimension $N_B D_B$), compete for attending to parts of the input via a softmax-based attention. These slot encodings are iteratively updated and allow to capture distinct objects or image components. The result is an attention matrix $A \in R^{N_S \times D}$ for an input $x \in R^{D}$. Each entry $A_{i}$ corresponds to the attention weight of slot $i$ for the input $x$. Based on the attention matrix, the input is processed to read-out each object by multiplying $A$ with the input resulting in a matrix $U \in R^{N_S \times N_B D_B}$. SysBinder now performs an additional factor binding on the vectors $u_i$ of $U$. The goal of this factor binding mechanism is to find a distribution over a codebook memory for each block in $u_i$, i.e., $u_{i}^j$. This codebook memory (one for each block), $M^j \in R^{K \times D_B}$, consists of a set of $K$ learnable codebook vectors. Specifically, for each block $j$ an RNN consisting of a GRU and MLP component iteratively updates the $j$-th block of slot $s_i$, $s_{i}^j$, based on $u_i^j$ and previous $s_{i}^{j}$. Finally, a soft information bottleneck is applied where each block $s_i^j$ performs dot-product attention over the codebook memory leading to the final block-slot representation: $$ \mathbf{s}_{i}^j=\left[\underset{K}{\operatorname{softmax}}\left(\frac{\mathbf{s}_i^j \cdot (\mathbf{M}^j)^T}{\sqrt{D_B}}\right)\right] \cdot \mathbf{M}^j $$ This process is iteratively refined together with the refinement processes of slot attention. Overall, the encodings of SysBinder represent each object in an image by a slot with $N_B$ blocks where each block represents a factor of the object like shape or color. Note that in the main text, the final $s_i^j$ is denoted as $z_i^j$. --- [1] Singh, Gautam, Sungjin Ahn, and Yeongbin Kim. "Neural Systematic Binder." ICLR, 2023. [2] Locatello, Francesco, et al. "Object-centric learning with slot attention." NeurIPS, 2020. --- Rebuttal 3: Title: Updated text on merge revision Comment: (i) Merge Concepts: In the case that $\mathcal{R}$ contains multiple concepts that represent a joint underlying concept (e.g., two concepts for purple in Fig.3 (right)) it is easy to update the model's internal representations by replacing the concept symbols of one concept with those of a second concept. Specifically, according to human or additional model feedback, if concept $m$ in block $j$ should be merged with concept $b$ ($m,b \in \{1, \cdots, N_C\}$) then for all corpus tuples, $(\texttt{enc}_l^j, v_l) \in R^j$, we replace $v_l$ with $b$ if $v_l = m$. --- Rebuttal 4: Title: Updated text on add revision Comment: (iii) Add Encodings or Concepts: If a specific concept is not sufficiently well captured via the existing encodings in $\mathcal{R}^j$, one can simply add a new encoding for the concept, $m$, to the corpus: $\hat{\texttt{enc}}_{l+1}^j$ This leads to an additional entry in the corpus: $(\hat{\texttt{enc}}_{l+1}^j, m)$ Accordingly, it is also possible to add encodings for an entire concept. Hereby, via the soft binder one infers block encodings of example objects that represent that novel concept, $b$, and adds these to the corpus as $(\hat{\texttt{enc}}_{l+1}^j, b)$ with $b = N_C+1$. --- Rebuttal 5: Title: Updated unclear sentence Comment: Briefly, given an image, $x$, NCB infers latent block-slot encodings, $z$, and performs a retrieval-based discretization step on $z$ to infer concept-slot encodings, $c$. These express the concepts of the objects in the image, i.e., object-factor level concepts. --- Rebuttal 6: Comment: I thank the authors for their efforts in trying to address the suggested issues. However, I will increase my score to 5 only since I still think is a borderline paper: i) The method presentation should have been improved according to what the authors have reported, but I should review it again to assess it better. ii) Testing only on toy datasets is still not acceptable for a neurips paper. Although I agree that better object centric learning will improve the performance of the model, how the current method behave on natural image dataset is of crucial importance to globally assess the model. Bad, but promising results would have still been appreciated. --- Rebuttal Comment 6.1: Comment: We thank the reviewer for their time and for reconsidering their rating. However, we disagree that non-synthetic datasets are mandatory for NeurIPS. In fact, several influential and recent NeurIPS papers in the field of object-centric/NeSy learning were published based only on synthetic datasets, e.g., [1,2,3,4]. [1] Locatello, Francesco, et al. "Object-centric learning with slot attention." NeurIPS, 2020. [2] van Krieken, Emile, et al. "A-nesi: A scalable approximate method for probabilistic neurosymbolic inference." NeurIPS, 2023. [3] Marconato, Emanuele, et al. "Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts." NeurIPS, 2023. [4] Li, Zenan, et al. "Neuro-symbolic learning yielding logical constraints." NeurIPS, 2023.	Summary: This paper introduces neural concept binder, a neural symbolic framework that utilizes both soft and hard binding. Building on top of the sysbinder model, it can additionally do exemplar-based hard binding and revise concepts. Evaluations made on CLEVR and the proposed CLEVR-Sudoku dataset proved the method's validity. Strengths: - The paper is well-written and easy to read. Connections to previous works are clarified nicely. - It's good to see the incorporations of both hard and soft bindings to existing neural-symbolic frameworks. - The model achieved good performance on the proposed CLEVR-Sudoku task and can do satisfactory concept revision and inspection, which is a neat proof of concept that hard binding works. Weaknesses: There are several weaknesses I can foresee that may lead to the rejection of this paper. - Limited contribution: After so many years of developing neural-symbolic methods in visual reasoning, from the earliest modular approaches to unsupervised concept learners, code-based reasoning models, and recent visual programming-like frameworks, the goal of neural-symbolic modeling has dramatically changed. In this work, the neural concept binder still focuses on one of the earliest task categories designed for visual reasoning (CLEVR attribute classifications or unsupervised concept learners). It's also built on top of sysbinder, in other words, it's merely an incremental improvement by adding a retrieval-based library. - I don't see any generalizability of this method beyond extremely toy tasks (attribute classification). The proposed CLEVR-Sudoku is strange and does not correspond to any of the real-world visual reasoning tasks. Relational tasks are also not tackled in this paper. Technical Quality: 3 Clarity: 3 Questions for Authors: - How generalizable can this method be? Can it serve as a part of the closed-loop reasoning in CLEVR (I mean, the original CLEVR questions, as tackled in NS-CL and NS-VQA line of work)? - Can relational concepts be similarly represented via soft/hard binding? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: Yes, the authors have adequately addressed the limitations. I do not see any potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1.1 (NeSy contribution): We agree that the field of neuro-symbolic AI is rapidly evolving with different focuses, e.g., visual reasoning in real-world images. However, works that are focussing on higher-level neuro-symbolic problems still heavily rely on mapping raw input images to symbolic representations, whether that is by using pre-trained concept extractors [1, 2, 3] or foundation models [4, 5, 6] or forms of weaker supervision [7]. However, how to obtain meaningful symbols from images in a fully unsupervised way remains a core challenge for the community. With our work, we address exactly this open problem, and to our knowledge, only [8] has additionally attempted to tackle this task. We hereby specifically focus on the importance of being able to inspect and effectively revise learned symbols, particularly due to the unsupervised nature of the learning setup. These aspects are most often overlooked in the majority of research that is in the pursuit of performance-driven models. Overall, we believe that our work offers a unique contribution to the field of concept learning, as also all other reviewers agree upon. If the reviewer is aware of specific works that are unknown to us yet relevant for this topic, we would be grateful if they could share them so we can better assess the reviewer's remarks. W1.2 (utilizing SysBinder): We strongly disagree that our work represents only an incremental improvement over SysBinder. While in this work, we have focused on utilizing the SysBinder encoder as NCB's soft binder, our method is not limited to it. Instead, this encoder can be replaced by any current or future models that process images into object-factor encodings (e.g., also models that can process natural images). With our framework, we propose a general method for discovering discrete symbols from continuous representations that offers inspection and revision possibilities for human stakeholders. This is not trivial and a very important aspect for unsupervised learning that, to our knowledge, has not been addressed by other works. W2.1 (attribute classification as "toy task"): We disagree and consider extracting symbols from (unlabeled) images as a cornerstone for a variety of neuro-symbolic methods. Many neuro-symbolic approaches rely on extracting symbolic representations from the input, whereas most recent works specifically rely on pre-trained models [1, 2, 3, 9]. NCB, on the other hand, allows to learn such symbols without supervision and a main goal of our work is to introduce NCB as a core module for other frameworks (as particularly reviewer VMaX has acknowledged), e.g., for more complex visual reasoning methods. With the evaluations in the context of Q2-Q4, we have highlighted this ability. W2.2 (CLEVR Sudoku): With CLEVR Sudoku, we are proposing a very relevant challenge that is combining complex visual inputs with explicit reasoning on a symbolic level. Specifically, the underlying reasoning skills required to solve CLEVR Sudoku are very much relevant for real-world scenarios. In fact, many works agree that reasoning with abstract representations (i.e., concepts) are essential for robust generalization [10, 11, 12]. Overall, there is a great deal of interest, and thus, a variety of benchmarks that investigate exactly this form of required intelligence [13, 14, 15, 16, 17]. These specifically exclude language priors and common knowledge priors of real-world images to asses a model's learning and reasoning abilities. W2.3 (Relational tasks): The goal of our NCB framework is to extract meaningful symbolic representations from the input, which can be beneficial for a variety of tasks. Hereby, NCB focuses on learning discrete object-level concept representations from images, i.e., focusing on unary relational concepts. The aim overall is not to propose a framework to solve relational tasks. However, NCB can be used as an easily integrable module for models that perform relational reasoning, e.g., which usually require pre-trained, (semi-)supervised concept extractors. Overall, we agree that it is important for future work to integrate NCB into more complex relational tasks, as we had noted in our conclusion. Q1 (closed-loop reasoning): Yes, exactly, NCB can be easily integrated into closed-loop reasoning. This is exactly the intended application of our method. E.g., NS-VQA had relied on the extraction of the scene representation via a supervised trained scene parser with ground truth object attributes. Instead of this scene parser, one can employ NCB, which discovers attributes without supervision and use the object’s attention masks (from the slot attention component) to determine their position. In NS-CL, the concepts are learned via the question-answering task. This is an alternative approach to concept discovery that, however, requires question-answer pairs for the image domain that target those concepts. Q2 (relational concepts): In principle, we do think it is possible to represent relational concepts via soft/hard binding, though we do not investigate this here. As mentioned above, our approach is intended to be integrated into other methods, such as visual programming or program synthesis, where the learned relations would be based on NCB’s concepts. Hereby, some current work focuses learning relations from images neurally, and others focus on symbolic approaches. Thus, whether to perform relational concept learning via neural, soft binding principles or symbolic, hard binding principles, or a combination of these is still up for investigation. We further refer to our response above concerning relational tasks. --- Rebuttal 2: Title: References Comment: [1] Yi et al. "Neural-symbolic vqa: Disentangling reasoning from vision and language understanding." NeurIPS, 2018. [2] Koh et al. "Concept bottleneck models." ICML, 2020. [3] Shindo et al. "α ILP: thinking visual scenes as differentiable logic programs." Machine Learning, 2023. [4] Surís et al. "Vipergpt: Visual inference via python execution for reasoning." CVPR, 2023. [5] Gupta et al. "Visual programming: Compositional visual reasoning without training." CVPR, 2023. [6] Hsu et al. "What’s left? concept grounding with logic-enhanced foundation models." NeurIPS, 2024. [7] Mao et al. "The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision." ICLR, 2019. [8] Wu et al. "Neural Language of Thought Models." ICLR, 2024. [9] Wüst et al. "Pix2Code: Learning to Compose Neural Visual Concepts as Programs." UAI, 2024. [10] Mitchell, Melanie. "Abstraction and analogy‐making in artificial intelligence." Annals of the New York Academy of Sciences, 2021. [11] Zhang et al. "How Far Are We from Intelligent Visual Deductive Reasoning?." ICLR 2024 Workshop: How Far Are We From AGI. [12] Lake et al. "The Omniglot challenge: a 3-year progress report." Current Opinion in Behavioral Sciences, 2019. [13] Chollet, François. "On the measure of intelligence." arXiv, 2019. [14] Moskvichev et al. "The conceptarc benchmark: Evaluating understanding and generalization in the arc domain." arXiv, 2023. [15] Raven, Jean. "Raven progressive matrices." Handbook of nonverbal assessment, 2003. [16] Lake et al. "Human-level concept learning through probabilistic program induction." Science, 2015. [17] Nie et al. "Bongard-logo: A new benchmark for human-level concept learning and reasoning." NeurIPS, 2020. --- Rebuttal Comment 2.1: Comment: Thanks for the authors' response. After reading the rebuttal, I've decided to increase the score by 1. --- Reply to Comment 2.1.1: Comment: We thank the reviewer for the time! However, we kindly ask the reviewer to raise remaining concerns in case we could not resolve all issues with our initial response.	Summary: The authors introduced a pioneering framework that combines an object-centric learning module with a retrieval-based module to address visual reasoning tasks and a new visual reasoning task, CLEVR Sudoku. The proposed method demonstrated significant potential in effectively acquiring inspectable and revisable concepts via human or machine feedback in various scenarios. Strengths: - S1: The proposed method offers significant novelty in that it has the potential to serve as a building block for concept learning, which can be leveraged as a core module in other frameworks. The author's well-structured experiments provided compelling evidence in support of these claims. Weaknesses: - W1: The proposed approach can be interpreted as directly integrating SysBinder and HDBSCAN. Because the initial concept detection fundamentally depends on the complete functionality of SysBinder, this framework may not circumvent specific inherent challenges of object-centric learning, including inductive bias resulting from choosing the proper object-factor encoder and identifiability issues. - W2: Using HDBSCAN is intuitive in the proposed method, but it would be beneficial to include an additional experiment that compares different clustering methods. Technical Quality: 3 Clarity: 4 Questions for Authors: Please check out the Weakness section first. I listed the following questions and suggestions that would be helpful for authors' future works: - Q1: The recent method [1] in object-centric learning literature is linked to causal representation learning and the identifiability of slot representations. How can this be integrated into your framework? - Q2: Object-factor learning can be interpreted as learning atoms in logic, and the NN explanations in Table 2 can be seen as the simple form of propositional logic in a neuro-symbolic framework. How can an object-centric learning framework be extended to represent logical rules, such as in the form of first-order logic? Reference - [1] Mansouri, Amin, et al. "Object-centric architectures enable efficient causal representation learning." arXiv preprint arXiv:2310.19054 (2023). Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: Please check out the Weakness and Question sections. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1 (dependence on continuous encoder): Indeed, the quality of the initial continuous concept encodings is important for the resulting discrete concept representation. We had remarked on this in the context of our ablations. We have now added an ablation study to highlight this empirically (c.f. Tab.1 (middle) in additional pdf). Specifically, via earlier checkpoints of the sysbinder encoder we observe that the weaker the encoder is, the less expressive are NCB's final discrete concept representations. However, as better models are being developed (e.g., that are shown to also process natural images), one can easily integrate these into the NCB framework and replace the SysBinder encoder that we used for our evaluations. W2 (ablation clustering): Thanks for the hint! We have added an ablation evaluation on the CLEVR dataset in the context of Q1 where we compare the expressiveness of concept representations when utilizing k-means rather than HDBSCAN (see corresponding table in additional pdf and general response above). We observe that for generalization, the HDBSCAN approach performs significantly better. Q1 (causal representation learning): Thanks for the pointer! This work is indeed interesting, as it connects the more “informal” object-centric detections with the guarantees (and also assumptions) of causal representation learning. While this work and NCB could be considered somewhat parallel, some possibilities exist to combine both. One could use the weakly supervised perturbations to finetune NCB's initial soft binder encoder if one wants to move toward weak supervision. On the other hand, one could also utilize the unsupervised trained NCB with its interventional inspection to generate sparse perturbation of images. This could potentially be possible without specific human supervision but might be limited to the concepts NCB discovers unsupervised. Overall, there are many possibilities for future research in this direction. We have added the reference accordingly into the main text. Q2 (OC-learning and logic): That's a very important question for the future development of object-centric and neuro-symbolic learning in general. In fact, one motivation for our proposed framework is to be able to extract unsupervised discrete representations from object-centric representations that can be integrated into symbolic AI approaches such as different logic systems. Specifically, NCB's concepts can already be utilized, to some extent, for first-order logic as properties about different objects are encoded into symbols, i.e., unary relations. Thus, first-order logic formulas containing quantors can be expressed out of the box. However, the NCB framework is currently not aimed at encoding arbitrary n-ary relations. We consider utilizing NCB in approaches like [1] could mitigate this issue to learn more complex logic programs based on the unsupervised learned concepts of NCB. --- [1] Wüst et al. "Pix2Code: Learning to Compose Neural Visual Concepts as Programs." UAI, 2024. --- Rebuttal Comment 1.1: Title: Reponse to the author's rebuttal Comment: Thank you for the authors' efforts in providing additional experiments on clustering and the reference. I'm maintaining the score. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for their valuable time and appreciation for our work!	Summary: This paper introduces Neural Concept Binder, a framework for obtaining discrete concept representations from images without any supervision. The method is an extension of Neural Systematic Binder (SysBinder), adding a clustering step on top of the block-slot representations to obtain discrete concept representations. The resulting representations are interpretable and modifiable, as shown in the experiments. The model is additionally evaluated on property prediction and downstream tasks on modifications of the CLEVR dataset and shown to be able to leverage fewer training samples than SysBinder. Strengths: The paper investigates an important problem in learning discrete, interpretable concepts from images in an unsupervised way. The model is a logical extension of SysBinder in clustering the representations to obtain discrete concepts. The experiments show improvements in sample efficiency of these discrete representations over SysBinder’s continuous representations. Weaknesses: 1. Since the solver used in the Sudoku experiments is the same across all baselines, it seems the determining factor of performance is in how well the digits are classified. Therefore, I do not believe framing this evaluation in the context of Sudoku adds any insight—in fact it seems to add unnecessary noise to the evaluation. The evaluation in the appendix (Figure 8) seems more informative and sufficient for determining the benefit of NCB. 2. This paper is missing several related citations: Unsupervised Concept Discovery Mitigates Spurious Correlations (https://arxiv.org/abs/2402.13368) and Neural Language of Thought Models (https://arxiv.org/abs/2402.01203). NLoTM is particularly relevant and can be an additional baseline since it also extends SysBinder to learn discrete representations, except it is trained in an end-to-end way. 3. The discussion in section 3.3 is interesting, but it would be informative to tie each point with corresponding experimental evidence. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. For SysBinder (hard) and SysBinder (step), do the models train well with this temperature adjustment? E.g. do they exhibit decent reconstruction and slot decomposition? 2. I’m not sure if I completely understand the analysis for Q4, but since this is done on the concept encodings, and not the discrete representations, can the same analysis be done with SysBinder representations? If so, does NCB offer any additional benefits here? 3. How important is the choice of clustering algorithm to the results? What if we use a simple k-means clustering as is done in the original SysBinder paper? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1 (regarding RQ2 (CLEVR Sudoku)): A determining factor for the performance on CLEVR-Sudoku is the classification of the digits. We agree that this has, to some degree, already been investigated in the context of Q1, where we tested the suitability of NCB’s concept representations for few-shot GT attribute classification. However, in Q2 we wanted to extend these findings into a much more challenging downstream task (than just GT attribute prediction) and specifically focus on the integration into purely symbolic systems that require correct, task-specific symbolic mappings. Importantly, as NCB is trained unsupervised, there is no guarantee that the learned concept symbols will be directly translatable to the symbols required for a downstream task. It is, therefore, potentially necessary to learn a translation from NCB's concept symbols to the task symbols, e.g., individual object attributes to specific object compositions. Overall, this is a very valuable problem setting concerning real-world applications, e.g., in planning (e.g., navigating through a world of objects). In Q2, we investigate NCB's potential for integration in such settings and consider the evaluations on CLEVR-Sudoku to be a helpful illustrative example. Additionally, the evaluations in Q2 also serve as baselines for investigating the ability to revise NCB's concepts for such a complex downstream task in the context of Q3. Nonetheless, we do acknowledge that Figure 8 also provides valuable insights that fall a bit short in the main paper (thanks for the hint!). We have updated the main text to put more focus on these results within the discussion of Q2. W2 (additional citations and baseline): Thank you for pointing out the related works, we were not aware of them. We consider Arefin et al. orthogonal to our work as they focus on learning object-level concepts in comparison to our object-factor level concepts and have included it in our related works section. Indeed, Wu et al. is relevant to our problem setting. Unfortunately, it was published so close to the NeurIPS deadline that we seem to have missed it; thank you for pointing it out! We have provided the results of this model in our general response and included it as an additional baseline in our paper. We agree this helps highlight our work. W3 (Evidence for discussion 3.3): Thank you for raising this point! We agree that the paper would definitely benefit from more illustrations of the results that can be obtained by inspection. We therefore added example visualizations (c.f. additional pdf). Further, in the context of Q3, we do investigate revising models. Specifically, we apply $\texttt{merge}$ and $\texttt{delete}$ revision; the corresponding models are denoted as "NCB revised (GPT-4)" and "NCB revised (human)". We also investigate the application of $\texttt{add}$ in the "left-right" experiment in the appendix (F4). Q1 (Training SysBinder (hard) and (step)): Good question! The SysBinder (hard) model did not train well, i.e., the reconstruction was terrible. The SysBinder (step) model, on the other hand, had stable training and provided reconstructions of the quality of the vanilla SysBinder. These results support recent findings [2] that training from the start with hard discretization results in badly performing models. I.e., it is very difficult for the model to learn good representations at all (e.g., for reconstruction) with such a strict bottleneck. On the other hand, learning step-wise can more easily lead to issues with local optima than for models without any bottleneck. Q2 (RQ4 evaluations): We are sorry for the confusion here. Within Q4, we investigate the popular setting of concept-bottleneck-like approaches [1,2] in which a model predicts (discrete) high-level concepts from an image and a second model makes the task prediction based on these concept activations. This leads to models with more transparent and high-level explainability properties. In principle, one can also utilize continuous concept encodings rather than fully discrete concept representations. However, this makes their understandability more ambiguous. In Q4, we investigate the case in which the prediction model is given discrete concept representations. We compare to supervised discrete concept representations (c.f. F7) and particularly investigate the ability to inspect and revise such models that utilize NCB's concepts. Thus, the used concept encodings for the evaluation are actually the discrete concept-slot encodings of NCB. Overall, interpreting and revising models that utilize continuous embeddings is not as straightforward, we have therefore omitted such comparisons here. We have further clarified this now in the text of Q4. Q3 (importance of hdbscan): Thanks for the valid questions! We refer here to the general response above with novel ablation evaluations in the additional pdf. --- [1] Koh et al. "Concept bottleneck models." ICML, 2020. [2] Stammer et al. "Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations." CVPR, 2021. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: Thank you for answering my questions and taking the time to run additional experiments. Overall, I think this is a nice extension of SysBinder, although I am still not convinced that the Sudoku experiments are necessary to demonstrate the model's capabilities. I have decided to increase the score. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for their time and reconsidering their rating!	Rebuttal 1: Rebuttal: We thank all the reviewers for their time and valuable feedback. We are especially happy to receive so much positive feedback concerning the importance of the tackled problem ("investigates an important problem" - edmj), the contribution of our work overall ("pioneering framework" - VMaX, "novel in the field of concept learning" - 6FRw) and the novel CLEVR Sudoku dataset ("important resource for the Neuro-symbolic literature" - 6FRw). We are also happy to hear our writing is of high quality ("well-written and easy to read" - U7bj). With the Neural Concept Binder, we introduce a framework to learn discrete, interpretable concepts from images in an unsupervised way. We specifically focus on object-factor-level concepts and the ability to inspect and revise these concepts. This is achieved by combining an object-centric learning module with a retrieval-based module, where the retriever component is beneficial both to obtain discrete concepts and effectively integrate revisory feedback from a human or second model. In our evaluations, we highlight the expressiveness of NCB's concepts and their potential to be integrated into various model settings (both symbolic and neural downstream modules). Moreover, we introduce CLEVR-Sudoku, a challenging, neuro-symbolic benchmark dataset. ### Additional baselines: Thanks for the suggestions! We have indeed run novel baseline evaluations in the context of Q1 based on the recent SOTA model NLOTM [5]. To our knowledge, this is the only approach that also allows us to learn discrete, object-centric concept representations from unlabeled images. Two seeds are still pending (due to long training durations). We provide preliminary results for one seed in Tab. 2 (additional pdf) and will provide the missing seeds for the final version. We observe a strong decrease in the ability to infer GT object attributes from NLOTM's discrete representations over those of NCB. This suggests that despite NLOTM's explicit training approach for discrete representations, NCB's approach has its advantage regarding representation expressiveness. However, future investigations are necessary to make conclusive remarks. Overall, these results highlight the difficulty of the task and the ability of NCB to tackle it. Additionally, our work focuses on the ability to inspect and revise learned concepts. NLOTM, on the other hand, emphasizes the ability to generate novel images. It would be interesting to investigate how both approaches can be combined, e.g., to aid NCB's inspection mechanisms via NLOTM's generation abilities. ### Clustering ablations: We agree and have added an ablation to investigate the effect of using k-means rather than the more powerful HDBSCAN (c.f. Tab.1 (right) in additional pdf). We observe that, particularly in the small-data regime, the concept representations obtained via k-means are less expressive (though still better than the discrete SysBinder baselines). Overall, the HDBSCAN approach performs better in terms of generalization. However, in principle, the NCB framework is implementation-agnostic to the choice of clustering method, e.g., one can also integrate more powerful recent neural approaches for clustering [1]. ### Real-world datasets: While we fully agree that natural-image datasets are important, we note that the scalability and ability to handle these via NCB depends on the object-centric encoder's (NCB's soft binder) ability to handle natural images. Importantly, this is a general challenge for object-centric learning research and a big topic in the current research community [2,3]. While this line of research is extremely important, it is also somewhat parallel to our work. As object-centric encoders improve, concept discovery via NCB based on these encoders will also be able to handle more natural images. For our evaluations, we mainly focused on instantiating NCB's soft binder with the SysBinder encoder (currently the SOTA model for obtaining object-factor representations without supervision). To the best of our knowledge, SysBinder has only been evaluated on CLEVR-based datasets. Our work focuses on how to extract discrete concept representations from such an encoder (irrespective of the kind of images it can process) and, importantly, how to inspect and revise these representations. Therefore, we have not investigated the scalability of SysBinder to more natural images and consider it out of the scope of this work. Concerning the scalability of the hard binder component, we expect that learning the hard binder should not lead to particular scaling issues (depending on the number of block encodings to be clustered and the scope of the grid search). In addition, NCB's retrieval can be sped up, e.g., via FIASS [4]. ### Additional figures for inspection types: We have added qualitative images in the appendix (e.g., Fig.1 in add. pdf) to further exemplify the inspection types of section 3.3. --- [1] Vardakas et al. "Neural clustering based on implicit maximum likelihood." Neural Computing and Applications, 2023. [2] Singh et al. "Guided Latent Slot Diffusion for Object-Centric Learning." arXiv, 2024. [3] Elsayed et al. "Savi++: Towards end-to-end object-centric learning from real-world videos." NeurIPS, 2022. [4] Johnson et al. "Billion-scale similarity search with GPUs." IEEE Transactions on Big Data, 2019. [5] Wu et al. "Neural Language of Thought Models." ICLR, 2024. Pdf: /pdf/61c3bdf0b97bd6ead346cdc953b7c4e13a9a2966.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Decoupling Semantic Similarity from Spatial Alignment for Neural Networks.	Accept (poster)	Summary: In this paper, the authors propose a new method to measure similarity between responses of deep neural networks in vision. They reformulate the commonly used strategy to compute Representational Similarity Matrices (RSMs) by acknowledging the superiority of the semantic information over the spatio-semantic information in the creation of RSMs. The authors perform different experiments to show the improvement over the baseline method caused by their reformulation. Strengths: - The paper is clearly written and tackles an important topic. - The idea is original and besides its limitation connected to high computational needs, it could serve as an inspiration for future works. - Although the experiments are not exhaustive, they are convincing and coherent and the gained insights seem relevant. - Being transparent with the limitation of the method and trying to provide the means to mitigate it is a plus. Weaknesses: - The authors should better discuss the differences between the results obtained for ViTs and CNNs in their study, which are quite well visible. E.g. while in the case of an examined CNN, the spatio-semantic RSM does not reflect well the similarities between translated images, in the case of the examined ViT (appendix), these similarities can be observed. The other thing is that the experiment is slightly different, because in the experiment with CNNs much smaller images are used than in the ViT experiment. Also, the differences are also visible in Table 1 (ResNets obtain much higher absolute correlation values for the baseline and the proposed methods than ViTs). - The authors provided few visual examples of the results of their method. It would be good to provide more of them (e.g. for different similarity metrics used, for more images and for more netorks) to enable more comprehensive qualitative evaluation (they could be placed in the appendix). - The use of some methods at work is not well justified (e.g. Pearson correlation). Technical Quality: 3 Clarity: 3 Questions for Authors: - Why do the authors use Pearson correlation to examine the relationship between the Jensen-Shannon Divergence and the representational similarity? E.g. Kendall/Spearman correlation can be more robust. - The authors should focus on the differences between the results obtained for CNNs and ViTs (see the comment in the weaknesses section). - The statement in the introduction “we argue that the spatial location of semantic objects does neither influence human perception nor deep learning classifiers.” is a little bit too bold - the paper does not examine human perception, therefore it would be better to leave only the deep learning here. - Also, a minor thing is that some typos, grammatical and formatting errors can be found in the paper (e.g. the sentence starting in l79, l205: a RSM -> an RSM, l25: SAMor CLIPSeg ,the retrieval performance) - The authors could provide more examples of their method for different networks to enable their better qualitative assessment which is now limited (e.g. in the appendix). Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately discussed the biggest limitation of their work (the computational cost of their method). They could also better highlight the limited data used in the experiments caused by the mentioned computational constraints. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > _Why do the authors use Pearson correlation to examine the relationship between the Jensen-Shannon Divergence and the representational similarity? E.g. Kendall/Spearman correlation can be more robust._ and _"The use of some methods at work is not well justified (e.g. Pearson correlation)."_ We have added a brief explanation in Section 4.3 regarding our choice of Pearson correlation: "We chose to use the pearson correlation, as it allows observing a direct linear behavior between representational similarity and predictive similarity." IWe believe this is particularly appropriate for cosine similarity and the radial basis function, which are bounded. However, we also provide Spearman correlation results in the appendix, along with a discussion. The corresponding table can also be found in the rebuttal PDF. Overall, we observe that Spearman correlation is mostly lower, with large ResNets being impacted strongly while ResNet-18 is notably consistent and stable in its correlations. ViTs interestingly show a much higher correlation for cosine and radial basis function similarities compared to Pearson correlations. >The authors should focus on the differences between the results obtained for CNNs and ViTs (see the comment in the weaknesses section). We extended our discussion on CNNs vs ViT's in the appendix, going in depth into the differences of the two translation experiments provided. > The statement in the introduction “we argue that the spatial location of semantic objects does neither influence human perception nor deep learning classifiers.” is a little bit too bold - the paper does not examine human perception, therefore it would be better to leave only the deep learning here. We cut the statement short to “we argue that the spatial location of semantic objects does not influence deep learning classifiers.” > The authors could provide more examples of their method for different networks to enable their better qualitative assessment which is now limited (e.g. in the appendix). As displayed in the rebuttal pdf, we will add plenty qualitative visualizations with a discussion in the appendix for many models. Additionally, as we added the Cityscapes experiment, we will also provide qualitative results on this dataset. >Also, a minor thing is that some typos, grammatical and formatting errors can be found in the paper (e.g. the sentence starting in l79, l205: a RSM -> an RSM, l25: SAMor CLIPSeg ,the retrieval performance) We want to apologize for this oversight, we have re-proofread our manuscript and pushed it through a spell-checker to fix any potential typos and grammatical errors. > They could also better highlight the limited data used in the experiments caused by the mentioned computational constraints. Instead of highlighting this, we increased our overall sample size, as mentioned in the main response. We hope this to be a sufficient adaptation. Find the similarity table with spearman's rank correlation below (N=20.000). \| Metric \| pearson_corr \| \| \| \| \| \| spearman_rank \| \| \| \| \| \| \|-----------------\|-------------:\|-------:\|--------------:\|-------:\|--------:\|-------:\|--------------:\|-------:\|--------------:\|-------:\|--------:\|-------:\| \| Kernel \| cosine_sim \| \| inner_product \| \| rbf \| \| cosine_sim \| \| inner_product \| \| rbf \| \| \| Invariance \| - \| PI \| - \| PI \| - \| PI \| - \| PI \| - \| PI \| - \| PI \| \| Architecture \| \| \| \| \| \| \| \| \| \| \| \| \| \| ResNet18 \| -0.279 \| -0.328 \| -0.264 \| -0.272 \| -0.174 \| -0.197 \| -0.231 \| -0.337 \| -0.239 \| -0.225 \| -0.435 \| -0.476 \| \| ResNet50 \| -0.256 \| -0.305 \| -0.249 \| -0.269 \| 0.028 \| 0.015 \| -0.032 \| 0.007 \| -0.046 \| -0.040 \| 0.128 \| 0.132 \| \| ResNet101 \| -0.235 \| -0.330 \| -0.211 \| -0.274 \| 0.076 \| 0.067 \| -0.007 \| -0.053 \| -0.007 \| -0.077 \| 0.071 \| 0.068 \| \| ConvNextV2-Base \| -0.162 \| -0.126 \| -0.160 \| -0.184 \| 0.077 \| 0.050 \| -0.017 \| 0.026 \| -0.013 \| -0.045 \| 0.143 \| 0.128 \| \| ViT-B/16 \| -0.058 \| -0.098 \| -0.056 \| -0.031 \| -0.079 \| -0.120 \| -0.013 \| -0.230 \| -0.021 \| 0.029 \| -0.220 \| -0.313 \| \| ViT-L/32 \| -0.142 \| -0.189 \| -0.143 \| -0.152 \| -0.131 \| -0.164 \| -0.034 \| -0.276 \| -0.029 \| -0.014 \| -0.335 \| -0.392 \| \| DinoV2-Giant \| -0.016 \| -0.046 \| -0.016 \| -0.030 \| -0.013 \| -0.052 \| -0.014 \| -0.037 \| -0.015 \| -0.022 \| -0.015 \| -0.042 \| --- Rebuttal Comment 1.1: Comment: I would like to thank the authors for their response along with some additional results and wish them good luck!	Summary: This paper proposes Semantic RSMs to understand the internal representations in deep neural networks. The authors argue that the current RSMs are limited by their coupling of semantic and spatial information, which restricts the assessment of similarity. The proposed semantic RSMs are spatial permutation invariant and focus solely on semantic similarity. The proposed method is shown to enhance retrieval performance and provide a more accurate reflection of the predictive behavior of classifiers. Strengths: 1. This paper is well-written and easy to follow. 2. The introduction of semantic RSMs is a significant contribution, potentially leading to more meaningful comparisons between neural network models. 3. The empirical demonstration of improved retrieval performance using semantic RSMs is convincing and adds practical value to the theoretical development. Weaknesses: 1. While the paper does highlight the high computational complexity as a limitation, it would benefit from a more detailed discussion on the scalability of the proposed method to larger models and datasets and the approximation error. Technical Quality: 3 Clarity: 3 Questions for Authors: I'm not an expert in this field, so I tend to start by looking at what other reviewers think of the paper. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the time and effort you invested in reviewing our manuscript. Regarding your comment that _"it would benefit from a more detailed discussion on the scalability of the proposed method to larger models and datasets and the approximation error."_ We addressed a part of this regarding scaling to larger datasets in the general response already. Regarding the question of scaling to larger models, the scaling of our method is just dependent on representation size, while model size plays no role at all. We add section to discuss the approximation error at the end of the Section 4.4: We cite: _"The fastest of the Batch-Optimal approximation methods shows $<8\%$% error while improving run-time $\times 36$ relative to the fastest optimal algorithm for spatial extent $4096$, while no spatial alignment shows $42\%$% deviation ["or approximation error"] from the optimal matching."_ We hope that this helps to improve clarity on the approximation error quality of the approximations.	Summary: The authors introduce semantic RSMs, which are designed to be invariant to the spatial arrangement of elements within images. These semantic RSMs assess similarity by treating the problem as one of set-matching, where the focus is on matching semantic content rather than spatial details. This approach not only aligns more closely with human perception but also improves the relevance and accuracy of image retrieval tasks. The paper claims that semantic RSMs offer a more robust measure of similarity by comparing them to traditional spatio-semantic methods. Strengths: - The focus on semantic content rather than spatial arrangement aligns more closely with human perception, potentially leading to more intuitive and relevant comparisons of neural network responses. - By being invariant to spatial permutations, this method can effectively compare images where the same objects appear in different locations. - Semantic RSMs can be used as drop-in replacements for traditional RSMs. Weaknesses: - Employing algorithms like Hungarian matching to find the optimal permutation matrix can be computationally expensive. - The effectiveness of this approach relies heavily on accurate identification and parsing of semantic concepts within images, which can be challenging in complex scenes or under conditions of visual ambiguity. - While focusing on semantic content is generally advantageous, completely ignoring spatial information can sometimes omit useful contextual cues that contribute to overall image understanding. For example, [contextual cues] a picture of a dining table with plates, utensils, and food arranged in a specific way might convey a meal setting, which could be lost if the spatial relationships are ignored. [object interactions] Images where interactions between objects are important, such as a cat sitting on a mat, might lose their interpretative meaning if spatial information is disregarded. The semantic content (cat, mat) remains the same, but the relationship changes based on their arrangement. [abstract content] In abstract art or images with non-literal interpretations, spatial composition itself can carry meaning and affect how the content is perceived and classified. Technical Quality: 2 Clarity: 2 Questions for Authors: - How well does the method scale to very large datasets or to more complex neural networks that handle highly varied or abstract visual content? - How does the method perform under noisy conditions or when semantic parsing is imperfect due to occlusions or poor image quality? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: - To better evaluate the impact of semantic RSMs, a set of diverse metrics should be established: Evaluate semantic RSMs against traditional spatio-semantic RSMs and other state-of-the-art similarity measures to highlight the improvements or shortcomings. - The authors should explicitly state the primary objectives of employing semantic RSMs. Show that semantic RSMs can make neural network decisions more interpretable by aligning more closely with how humans perceive images. - Apply semantic RSMs in specific use cases like medical imaging, satellite image analysis, and autonomous driving where ignoring spatial arrangements can be particularly detrimental or beneficial, providing a nuanced view of their applicability. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the detailed feedback. We can see that a great deal of time and thought went into it. However, we believe there may be a few misunderstandings that we would like to clarify: ### Questions Q1: How well does the method scale to very large datasets[...] R1: The scaling behavior depends on the _spatial extent_ of representations and the _use-case_ in which the method is applied. Details on this can be found in the common response. Q2: How well does the method scale [...] to more complex neural networks that handle highly varied or abstract visual content? R2: We are not quite sure what you mean by _"more complex neural networks of varied/abstract visual content"_. Our experiments already apply the method to various state-of-the-art networks (CLIP, DINOv2, SAM, ResNets, ConvNeXT), which are capable of handling diverse visual content. Additionally, we evaluate our method on standard ImageNet data as well as on EgoObjects, which we believe qualifies as a more complex dataset aimed at robotics. Q3: How does the method perform under noisy conditions or when semantic parsing is imperfect due to occlusions or poor image quality? R3: This is an interesting question! Poor conditions or visual ambiguity can degrade representations if the network cannot handle them, potentially harming, for example, the retrieval performance of our method. However, this is not a unique issue of our method but of all methods using learned representations, such as spatio-semantic RSMs. Even standard cosine similarity on global representations would likely experience similar degradation. ### Limitations 1. _"Apply semantic RSMs in specific use cases like medical imaging, satellite image analysis, and autonomous driving where ignoring spatial arrangements can be particularly detrimental or beneficial, providing a nuanced view of their applicability."_ We added an additional experiment for autonomous driving using the Cityscapes dataset, demonstrating improvements of semantic RSMs over spatio-semantic RSMs in this context. 2. To better evaluate the impact of semantic RSMs, a set of diverse metrics should be established: Evaluate semantic RSMs against traditional spatio-semantic RSMs and other state-of-the-art similarity measures to highlight the improvements or shortcomings. In this paper, we compared the current, most commonly used methods for constructing RSMs, as introduced in the [CKA paper](https://arxiv.org/pdf/1905.00414) by Kornblith et al. (Inner Product and RBF spatio-semantic RSMs), with the addition of cosine similarity. We also introduced two metrics to assess the “goodness” of RSMs, which we believe are sufficient to demonstrate the benefits of our proposed method. ### Weaknessses 1. [Runtime of Hungarian matching]: We clearly and transparently state the runtime as a limitation in our paper and provide some algorithms to alleviate this. Reviewer zPy6 even commends us on this. To put our compute time into perspective: In representation similarity analysis many much more slow to compute metrics exist, e.g. the [IMDScore](https://openreview.net/forum?id=HyebplHYwB) which takes >12 hours for a single layer-to-layer comparison between two representations of [10.000x2048] samples using 32 CPU cores we ran in a different project or the RSMNormDifference which can take multiple hours as well. The former was published in ICLR20. This is somewhat of a duplication to the common response, but we felt it necessary to reiterate. 2. [Visual ambiguity can negatively impact]: See Quesions R3 3. [Ignoring spatial context can be bad]: We really appreciated these thought experiments as they go into a lot of detail on how representations work. 1. [Contextual cues]: Networks progressively combine/merge information to build semantic concepts. E.g. Ears, eyes and whiskers in a neighbourhood, could lead to expression of a cat's head and so on. Similarly we expect a network to learn associations between dishes and cutlery on a table to a meal setting. If the network learns this association, breaking the spatial location with our method is not an issue, as the semantic concept vector will not only carry `<plate>` but `<plate, part_of_meal_table>` information somewhere. 2. [abstract content]: Distance and composition can carry meaning, but if the network learns to represent them similarly, this can be beneficial. For example, if you take a picture of an abstract painting, move back 5 meters, and take another picture, the spatial composition shifts despite the image being identical. Similarly, moving sideways changes the angle. If the network still represents the same semantic meaning despite these shifts, it can still be recognized. 3. [object interactions] (Very related to context cues) We believe that networks merge semantic information hierarchically within a neighborhood, forming increasingly abstract semantic concepts. Due to this spatial proximity, representations of a cat on a mat are encoded similarly, leading to more frequent retrievals of cats on mats, even when spatial constraints are ignored. Evidence for this is provided by CLIP models, which use global aggregation of their outputs—disregarding spatial relations—to generate final embeddings. This method effectively identifies related images and captions, such as “a cat sitting on a mat,” and we assume this approach is applicable to our setting as well. We hope that our responses have clarified the open questions and if we answered the questions sufficiently, it would be great if you could consider increasing your rating. --- Rebuttal Comment 1.1: Title: Post-rebuttal response Comment: Thanks for the clarification. The paper focuses on computing representational similarity by comparing semantics with location similarity. Instead of considering both simultaneously, the proposal focuses solely on semantics. This is achieved by extracting semantic vectors from the feature tensor and comparing them as a set, rather than matching the vectors individually in spatial terms. The concept is straightforward, but the application is unclear, and the derived insights may not be sufficient for publication. Here are the remaining concerns: - What are the derived insights from the proposal? e.g., Do ViT and CNN exhibit different behaviors? What can we learn from the results? A deeper analysis is needed beyond just showing that the proposal can predict output probabilities (e.g., Table 1). - There are counterexamples where neglecting location information leads to issues (e.g., visual anagrams), where semantically different images might not be distinguishable by the proposal since it matches features at a set level. - Also, how can this proposal be applied in complex scenarios, such as when the data contains multiple objects of the same class? - Last but not least, the permutation matrix required for set-matching introduces high computational overhead, which could limit the applicability of the proposal to high-resolution data. --- Rebuttal 2: Title: Discussion response Comment: Thank you for engaging with our rebuttal. We are pleased that we have addressed most of your initial concerns. Regarding the remaining issues: >What are the derived insights from the proposal? e.g., Do ViT and CNN exhibit different behaviors? What can we learn from the results? A deeper analysis is needed beyond just showing that the proposal can predict output probabilities (e.g., Table 1). We believe our paper provides a plethora of insights and is not limited to "just show[ing] that the proposal can predict output probabilities". In our paper, we: 1. Highlight that current RSM construction is flawed and has limitations of requiring spatial alignment, which has not been adequately discussed previously. 2. Introduce an algorithm that addresses this by utilizing set-matching, offering a novel approach. 3. Demonstrate the utility of our method through two applications, revealing significantly improved retrieval performance with five general-purpose feature extractors across the EgoObjects and Cityscapes datasets, and better correlation between the Jensen-Shannon Divergence of output probabilities and inter-sample similarity. We believe our work underscores a critical gap in current RSM-based methods and sets the foundation for further exploration into RSM construction, effective inter-sample similarity measures, and their applications for retrieval. >There are counterexamples where neglecting location information leads to issues (e.g., visual anagrams), where semantically different images might not be distinguishable by the proposal since it matches features at a set level. We understand your concerns regarding location information. However, __i)__ our experiments on the EgoObjects dataset, which includes varied contextual scenarios such as the meal table example, indicate that our method outperforms existing RSMs. This suggests that issues related to spatial location might either be rare or of lesser importance. __ii)__ Our method is based on learned representations. _Should a visual anagram exist where similar representations are expressed by the network, we believe this is to be a failure of the model and not our approach_. We don't believe a retrieval method should try to compensate for this. __iii)__ There are other examples where spatial position is neglected entirely: E.g. the ViT paper [1] shows (in Appendix D.4) that ViT's are able to learn meaningful representations for classification without any positional embedding, making the input a bag-of-words. > Also, how can this proposal be applied in complex scenarios, such as when the data contains multiple objects of the same class? Our method has already been successfully applied in scenarios involving multiple objects of the same class. Both retrieval datasets, EgoObjects (Figure 3) and Cityscapes (Rebuttal PDF) contain multiple objects of the same class. Additionally, one currently provided qualitative example in the paper and the additional qualitative examples in our rebuttal PDF illustrate our propose methods effectiveness in exactly such complex scenarios, highlighting retrieval on images with multiple instances of glasses or bowls. >Last but not least, the permutation matrix required for set-matching introduces high computational overhead, which could limit the applicability of the proposal to high-resolution data. We acknowledge the computational demands introduced by the permutation matrix for set-matching. Nevertheless, we have applied our method to high-resolution images (1920x1080) from the EgoObjects and Cityscapes datasets, which demonstrates its feasibility on large-scale data. If your concerns are aimed towards retrieval, we want to emphasize that retrieval is generally a two stage process (see [2] Section 4), with the first being a ranking by global cosine similarity and the second being a re-ranking of the top 100-400 cases which are most similar through e.g. our method. Hence such a two-stage application can enable our method to scale easily to any number of retrieval dataset size. Aside from this, we transparently address this limitation in our paper and have expanded on this topic in the appendix during the rebuttal phase. Moreover we'd kindly refer you to our general response which goes into more detail on the scalability of our method and would be happy to answer any more specific concerns regarding the scalability of our method should there be any remaining. We hope this addresses your concerns and look forward to any further questions you might have. [1] Dosovitskiy, Alexey, et al. "An image is worth 16x16 words: Transformers for image recognition at scale." arXiv preprint arXiv:2010.11929 (2020). [2] Cao, Bingyi, Andre Araujo, and Jack Sim. "Unifying deep local and global features for image search." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16. Springer International Publishing, 2020.	Summary: This paper makes a contribution to the construction of RSMs in the field of vision neural networks and puts forward the concept of semantic RSMs, which is innovative and theoretical. Strengths: The proposed semantic RSMs are used for spatial alignment by means of optimal permutation, which is a relatively new and promising method. This paper verifies the validity of semantic RSMs through experiments such as image retrieval and probabilistic similarity comparison. An in-depth analysis of the experimental results is carried out, and the advantages of semantic RSMs in specific tasks are pointed out. Weaknesses: This paper lacks the experimental verification of specific downstream tasks, such as detection and segmentation, on semantic RSMs. I need to know which scenario is more suitable for RSMs and semantic RSMs. Lack of quantitative comparative data. It is suggested to add tables or charts to show specific performance comparison data between semantic RSMs and existing methods in different tasks (such as image retrieval, class probability similarity comparison, etc.), including accuracy, time complexity and other indicators. The discussion of the experimental results was not thorough enough. It is recommended to add a detailed analysis of the experimental results to explain why semantic RSMs perform better on certain tasks, as well as possible reasons and limitations. "aligns" to "align" in line 57. Technical Quality: 2 Clarity: 2 Questions for Authors: It is suggested to further elaborate the potential and specific scenarios of the research in practical applications to enhance readers' understanding of its practical value. Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: see weekness Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to express our sincere gratitude for taking the time to read our paper and provide valuable feedback and constructive criticism: 1. _“This paper lacks the experimental verification of specific downstream tasks, such as detection and segmentation, on semantic RSMs. I need to know which scenario is more suitable for RSMs and semantic RSMs.”_ We believe this may be a misunderstanding. Semantic RSMs are not algorithms for detection or segmentation (which we do not claim), but rather they measure the similarity between neural network representations of images. If the concern is about applying our method to dense datasets like those used for object detection or segmentation, our EgoObjects experiment (Fig. 3) demonstrates its application to a dense dataset. Additionally, the results in Table 1 show our method applied to ImageNet. 2. _"Lack of quantitative comparative data. It is suggested to add tables or charts to show specific performance comparison data between semantic RSMs and existing methods in different tasks (such as image retrieval, class probability similarity comparison, etc.), including accuracy, time complexity and other indicators."_ We have added an additional retrieval experiment on the Cityscapes dataset, comparing semantic and spatio-semantic RSMs for this use case. Regarding the comparison of semantic RSMs with other retrieval methods, such as those using learned embedding spaces, we disagree that this is necessary. The aim of our paper is to find the best way to measure the similarity between representations of two samples. Parametrized methods that alter representations in non-trivial ways would make such comparisons unfair. We emphasize that the main purpose of our paper is to address the flaws in current RSM construction, which often requires spatial alignment. We demonstrate these flaws through retrieval and class probability similarity comparisons and show that our semantic RSMs alleviate these issues. 3. “It is recommended to add a detailed analysis of the experimental results to explain why semantic RSMs perform better on certain tasks, as well as possible reasons and limitations.“ We add a clarifying section at the end of the retrieval section, right before 4.3 discussing the retrieval results. We cite: >“These experiments display clearly that demanding spatial alignment can be a significant shortcoming when semantically similar concepts are misaligned. In Fig. 4, the network learned to represent the objects very similarly, despite a shift in perspective, but due the same objects not aligning anymore, spatio-semantic similarity fails to recognize this. This effect should generalize to other datasets where objects are not heavily centered. For datasets with heavy object-centric behavior, like ImageNet, this should be less pronounced.” 4. Typos We made sure the entire manuscript has the respective errors fixed.	Rebuttal 1: Rebuttal: Thank you to all the reviewers for their time and effort in reviewing our paper. We appreciate the feedback and tried to mitigate issues to the best of our abilities. We recognize different viewpoints, but some criticisms seem based on misunderstandings, which may have led to undeservedly lower ratings. We hope to clarify and address these issues in this rebuttal. ### Added content: While not relevant for all reviewers, we want to highlight the content added during the rebuttal as transparently as possible: 1. We conducted an additional experiment comparing semantic RSMs and spatio-semantic RSMs for retrieval using the Cityscapes dataset for autonomous driving. Results are provided in the rebuttal PDF. Our findings demonstrate that semantic RSMs are preferable to spatio-semantic RSMs for retrieval in this setting. More details are included in the figure caption, and we will offer a more comprehensive discussion in the appendix. Qualitative examples are also included but could not be fit on the single page of the rebuttal PDF. 2. We increased the sample size of our correlation experiment in Table 1 from 2,000 to 20,000 samples. Results are provided in the rebuttal PDF. Additionally, we present further quantitative results from our EgoObjects retrieval experiment, as shown in main Figure 4. We include results for 2,500, 5,000, and 10,000 database samples in table format. 3. Finally, we include additional qualitative retrieval results for various models. In the rebuttal PDF, we provide visualizations from CLIPSeg, in a downscaled excerpt. 4. Aside from this we edited small paragraphs of text, where criticised by the Reviewers. ### Common question: How well does the method scale to large datasets? As multiple reviewers had questions about scaling behavior, we want to use this space to discuss scaling in some more depth, and put this into perspective to currently exiting similarity methods common in the representational similarity analysis space. The scaling behavior of our poposed method depends on mostly two factors 1. Spatial Extent and Use-Case: 1. Spatial Extent: Smaller spatial extents result in less overhead during matching. For example, ViT models, such as ViT-L/16, have a spatial extent of 257 (with 1 class token) and can be compared essentially in the same way as traditional spatio-semantic RSMs. For CNNs, the situation varies with depth. Early layers, such as those in ResNet architectures, have a large spatial extent due to downsampling (e.g., a 224x224 image downsampled 4x results in a spatial extent of 3136), making early CNN layers significantly more expensive—by a factor of 12.25. In contrast, later layers with a smaller spatial extent show more favorable scalability for CNNs. 2. Use-Case: When comparing neural networks using RSMs, batched calculations are common. This approach reduces computational burden and memory constraints, making semantic RSMs scale effectively for this purpose, even with larger datasets. However, for retrieval, the scaling is more costly, as each query must be compared to the entire database. Therefore, retrieval scales linearly with the size of the database being queried. We already provide very specific runtimes in the Appendix Table 2 for our proposed approximations and the relative optimal algorithms. Given these numbers and a specific datasets and use-case in question one could estimate expected runtimes in a single threaded fashion. To aid clarity we will add the discussion above to this section in the appendix. Aside from runtime alone, one needs to take into account that the _dense_ embeddings of the image need to be stored for each model and image as well, which can have a substantial storage demand when e.g. trying to do retrieval on ImageNet1k. With the runtime being such a focal point though, we want to emphasize that the runtime of many representational similarity measures are not optimal. E.g. there exist measures that compute much more slowly than ours, e.g. the [IMDScore](https://openreview.net/forum?id=HyebplHYwB) which takes >12 hours for a single layer-to-layer comparison between two representations of [10.000x2048] samples using 32 CPU cores, which we used in a different project or the RSMNormDifference which can take multiple hours as well. With the IMDScore having been published in ICLR20 we hope we can convince you that runtime limitations are not a disqualifying factor for such methods. Pdf: /pdf/a1e04734c77c6704f7ec3c89c7b6cd8b59272c21.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
PGN: The RNN's New Successor is Effective for Long-Range Time Series Forecasting	Accept (poster)	Summary: This paper proposed a Parallel Gated Network (PGN) as a successor to RNN, featuring a Historical Information Extraction (HIE) layer to directly capture information from previous time steps. Additionally, it introduces a Temporal PGN (TPGN) framework with two branches to capture both long-term periodic and short-term semantic patterns, demonstrating state-of-the-art performance in long-range time series forecasting. Strengths: 1. This paper compares a variety of cutting-edge methods. 2. The experiments are generally thorough. Weaknesses: 1. The major issue with this paper is the lack of analysis and comparison with significant literature. The entire paper's premise is the traditional RNN's failure in long-term sequence problems due to the long information propagation paths of its recurrent structure. However, as far as I know, SegRNN[1] has already addressed these shortcomings of traditional RNN in long-term forecasting through segmented iteration and parallel prediction. Yet, there is no discussion on this in the paper. Please compare your method with it and clarify your differences and advantages. 2. In Section 2, you should distinguish between Linear-based and MLP-based methods. The former has only single-layer parameter connections, while the latter has multiple layers and can learn non-linear features due to the presence of activation functions. Methods like DLinear and FITS should be classified as Linear-based methods. 3. The description of HIE is unclear: (i) The process shown in Figure 2(a) suggests first performing linear mapping and then zero-padding, which conflicts with Equation 1 in the paper, where H = HIE(Padding(X)), and the actual code. It is recommended to modify Figure 2 to make this clearer. (ii) Line 169 describes that “HIE(·) is a linear layer,” but in practice, the behavior of HIE is more like a sliding aggregation operation of CNN (or TCN) rather than a sorely linear mapping. Given (ii), calling the proposed method RNN-based is debatable since it is more likely TCN-based. 4. You should include an ablation analysis of the normalization layer, explaining its impact on TPGN achieving state-of-the-art results. 5. Although the authors provide source code, it does not include the hyperparameter settings required to reproduce the key results in the paper, meaning there is no directly runnable script. Are the hyperparameters in the main results all defaults? For instance, is TPGN_period=24? If not, providing a complete script file that can be run directly is necessary. [1] Lin, S., Lin, W., Wu, W., Zhao, F., Mo, R., & Zhang, H. (2023). Segrnn: Segment recurrent neural network for long-term time series forecasting. arXiv preprint arXiv:2308.11200 Technical Quality: 3 Clarity: 2 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have already described the limitation of this work in the paper, namely the lack of modeling multivariate relationships. To some extent, this is not a significant issue because many current cutting-edge studies have demonstrated that focusing solely on univariate temporal relationships can also be effective in multivariate tasks. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the comprehensive review, detailed feedback, and valuable suggestions from the reviewer Bhsw. > Q1: The major issue with this paper is the lack of analysis and comparison with significant literature. Please compare your method with SegRNN and clarify your differences and advantages. \| Advantages \| Information propagation paths \| Whether a decoder is unnecessary \| Directly capturing periodic semantic information \| \|:--------------:\|:-----------------------------:\|:--------------------------------:\|:------------------------------------------------:\| \| TPGN (PGN) \| $O(1)$ \| ✓ \| ✓ \| \| SegRNN \| $O(L/W)$ \| ✗ \| ✗ \| (1) Regarding information propagation paths: Shorter information propagation paths lead to less information loss, better captured dependencies, and lower training difficulty. In this regard, PGN can achieve an information propagation path of $O(1)$, while SegRNN, although capable of segment-wise iteration, still maintains an information propagation path of $O(L/W)$, where $W$ represents the number of segments. This is where PGN holds a greater advantage. (2) In terms of time series forecasting design: Although SegRNN adopts parallel prediction in the decoder stage, successfully addressing the error accumulation issue of iterative RNN predictions, our approach, TPGN, draws inspiration from Informer's design, omitting the need for a decoder module. This not only resolves the error accumulation problem but also results in a more concise model design. (3) Direct extraction of periodic semantic information: TPGN can directly extract semantic information related to the periodicity of time series data, while SegRNN segments the time series first and then feeds the representation of each segment into GRU units, which hinders direct capture of the periodicity within the time series. The segmented iteration and parallel prediction of SegRNN are indeed thought-provoking. We have conducted preliminary experiments and observed that TPGN (PGN) has shown some improvement compared to SegRNN. In the revised version, we will incorporate the above analysis. > Q2: In Section 2, you should distinguish between Linear-based and MLP-based methods. Methods like DLinear and FITS should be classified as Linear-based methods. Thanks again for your valuable suggestion. We will conduct a full check and modification in the revised version. > Q3: The description of HIE is unclear: (i) Figure 2(a) conflicts with Equation 1 in the paper, where H = HIE(Padding(X)), and the actual code. It is recommended to modify Figure 2 to make this clearer. (ii) Line 169 describes that “HIE(·) is a linear layer,” but in practice, the behavior of HIE is more like CNN (or TCN) rather than a sorely linear mapping. Given (ii), calling the proposed method RNN-based is debatable since it is more likely TCN-based. (i) Thank you for pointing out the issue. We have made modifications to Figure 2 (a), and the new figure is shown as Figure A in the PDF file (newly submitted). (ii) Thanks again for your detailed review. In fact, HIE is indeed a linear layer. However, during our implementation, we were inspired by the design of methods like Transformer and iTransformer for the FFN part, and we also used conv1D to replace Linear. (iii) The function of the HIE layer is to replace the recurrent structure of RNN. In terms of specific code implementation, whether using conv1D or Linear, it does not affect its function of linear mapping. Additionally, the linear operation of HIE is just a part of the structure in PGN. Following HIE, there is the Gated Mechanism, and together they form the entirety of PGN. Therefore, PGN can be considered as the new successor to RNN-based methods. > Q4: You should include an ablation analysis of the normalization layer, explaining its impact on TPGN achieving state-of-the-art results. Okay, since the normalization layer was not used in the ECL, ETTh2, and WTH datasets, we focused on conducting ablation experiments on the normalization layer for the Traffic and ETTh1 datasets. Due to space limitations, the experimental results have been placed in Table A of the PDF file. For details, please refer to Sec. A of the Global Rebuttal. We observed a noticeable decrease in performance when the normalization layer was removed from the two datasets. This drop can be attributed to the inherent characteristics of the datasets. In Table 6, we calculated the distribution (Mean and STD) of the training and validation set for each dataset. The results indicate significant differences in the distributions of the Traffic dataset and the ETTh1 dataset. These differences can introduce disturbances to the model, hence the $norm$ should be set to 1 in this case. More details can be found in Appendix H. > Q5: It does not include the hyperparameter settings required to reproduce the key results in the paper. Are the hyperparameters in the main results all defaults? For instance, is TPGN_period=24? If not, providing a complete script file that can be run directly is necessary. For the 'TPGN_period', we were inspired by WITRAN and aimed to partition time series based on natural period. Therefore, for each task in the five datasets, 'TPGN_period' was set to 24. The selection of $norm$ varied depending on the dataset, as detailed in Table 6. The value of $dm$ for each forecasting task was determined through the validation set, with specific details available in Appendix D. Due to space limitations, we have included the hyperparameters of our model for each task in Table B of the PDF file for result reproducibility, please refer to Sec. B of the Global Rebuttal for more details. Furthermore, we will also enhance the files in the source code moving forward. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal; it addressed most of my concerns. The updated version should make the neccessary modifications mentioned in the rebuttal and include the analysis and results compared with SegRNN. I have increased my initial score. --- Reply to Comment 1.1.1: Title: Thanks Comment: Thank you once more for your time and valuable suggestions. In the updated revisions, we will include the necessary modifications, analysis as mentioned earlier, and the results compared with SegRNN.	Summary: This paper focuses on long-range time series forecasting problems. To address the limitations of RNNs, a novel paradigm called PGN is introduced as an alternative, providing shorter information propagation paths. Building upon PGN, the paper further presents a generic temporal modeling framework named as TPGN, which effectively captures both long-term and short-term periodic patterns, as well as local and global information, through a dual-branch design. The experimental results in this paper demonstrate that TPGN exhibits excellent performance in time series forecasting. Strengths: S1: This paper proposed a novel paradigm called PGN, which effectively tackles the inherent issues of RNNs through a simple yet powerful design. PGN exhibits a high level of innovation and holds the potential to replace traditional RNNs. S2: TPGN primarily focuses on modeling the temporal dimension. Its dual-branch design makes sense as it captures both long-term and short-term periodicity, as well as the local and global characteristics of time series. Additionally, it is reasonable to set up different univariate forecasting tasks to evaluate TPGN's performance. S3: This paper is well-written, and the presentation of the figures and tables is clear, making it easy to understand and follow. The experimental comparisons are comprehensive, including numerous advanced baseline models such as iTransformer, ModernTCN, FITS, TimeMixer, PDF, WITRAN, and Basisformer. Weaknesses: W1. For tables with a large amount of content, such as Table 1, it may be beneficial to consider using different colors for highlighting, as it could enhance clarity. Additionally, another option to consider is moving some of the experimental results to an appendix. W2. While TPGN exhibits some advantages in terms of efficiency, I have noticed that it still appears to be challenging to reach the optimal level. Specifically, I have noticed that as the input sequence size increases, the efficiency of TPGN may gradually become inferior to that of iTransformer. Technical Quality: 3 Clarity: 3 Questions for Authors: Q1. Why was the Gated Mechanism designed in PGN this way in Figure 2 (a)? Can this part be replaced with GRU or other RNN variants? Q2. Is it necessary to have two Linear layers in the design of the short-term branch in TPGN? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We extend our sincere appreciation to Reviewer t9pC for providing valuable feedback and acknowledgment of our research. > Q1: For tables with a large amount of content, such as Table 1, it may be beneficial to consider using different colors for highlighting, as it could enhance clarity. Additionally, another option to consider is moving some of the experimental results to an appendix. Thank you for your suggestion. In the revised version, we will enhance the table presentation by utilizing color changes, bolding, and other methods. Additionally, we will thoroughly check and optimize the entire paper. > Q2: While TPGN exhibits some advantages in terms of efficiency, I have noticed that it still appears to be challenging to reach the optimal level. Specifically, I have noticed that as the input sequence size increases, the efficiency of TPGN may gradually become inferior to that of iTransformer. We have noticed that in terms of efficiency, linear-based methods such as FITS do indeed perform better. And for iTransformer, since it maps the entire sequence into a patch for processing the time dimension, while TPGN requires long-term and short-term information extraction branches in the time dimension, the efficiency of TPGN may slightly lag behind iTransformer when the input sequence length varies. However, TPGN still holds significant advantages as it can maintain good efficiency while ensuring optimal performance. > Q3: Why was the Gated Mechanism designed in PGN this way in Figure 2 (a)? Can this part be replaced with GRU or other RNN variants? In our design, only one gate is needed for information selection and fusion, which incurs smaller overhead compared to GRU's two gates and LSTM's three gates. While PGN, as a general architecture, can have its gated mechanism replaced with GRU and other RNN variants, this substitution would introduce higher costs. In terms of performance, we have observed that despite having only one gate, PGN still outperforms GRU or LSTM, more details of the experiments can be found in Table 2. > Q4: Is it necessary to have two Linear layers in the design of the short-term branch in TPGN? Yes, because in the short-term branch, the semantic information extracted by the two linear layers carries different meanings. The first layer primarily aggregates short-term information, extracting localized information. The second layer further aggregates the representation from the first layer at a higher level, extracting global information. While it is possible to extract global semantic information using a single linear layer, it would overlook the extraction of information at different granularity levels. Additionally, in terms of efficiency, the theoretical complexity of using only a single linear layer is $O(L)$, which is larger than the current $O(\sqrt{L})$. --- Rebuttal Comment 1.1: Comment: Most of my concerns have been addressed, I will keep my positive recommendation. --- Reply to Comment 1.1.1: Title: Thanks Comment: Thanks for your time and positive recommendation!	Summary: The paper introduces a new model paradigm which aims to solve the traditional bottlenecks of RNN models, such as non-parallel computation, gradient explosion/vanishing issues, etc. Strengths: 1. An important problem is studied in this paper. 2. The overall representation is clear and easy to follow. 3. A comprehensive summary of the related work is provided. Weaknesses: 1. The overall contribution is not very significant. 2. Some questions regarding the time complexity and experiments need to be clarified. Technical Quality: 2 Clarity: 4 Questions for Authors: The model proposed in this paper is pretty neat and easy to follow. My questions mainly focus on the time complexity and experiments: 1. I think the total amount of computation done in the PGN should be O(L^2) because it is 1 + 2 + … + L, which is O(L^2). Although the PGN removes half of the computations of the self-attention, the self-attention and PGN still share the same asymptotic complexity. Thus, technically, replacing the PGN with self-attention won’t change the time complexity asymptotically. But I agree that it should be faster than RNN since it enables parallel computation as self-attention does. Also, this is the reason why I think the overall contribution is less exciting than the paper title claims. Can the authors kindly address this issue? 2. The above discussion can also be tested in the Efficiency of Execution experiment. 3. Regarding the experiments, only an input length of 168 is tested. Why did the authors choose to fixate on this input length instead of testing some other options? 4. In the ablation test, it seems that on some datasets (e.g., ETTh1), TPGN’s improvement is very slight compared to its LSTM/GRU/MLP ablations. Can the authors provide some analysis on these cases? 5. I am also interested to see some analysis on what would happen if the PGN is replaced by self-attention. Confidence: 4 Soundness: 2 Presentation: 4 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We express our sincere gratitude to Reviewer fqFX for providing comprehensive review, insightful perspectives, and thought-provoking questions. > Q1: I think the total amount of computation done in the PGN should be O(L^2) ... Can the authors kindly address this issue? Sincerest gratitude for your detailed comments. Here we would like to provide more explanations about our model: Firstly, regarding the complexity analysis: (1) The computational complexity related to sequences in PGN mainly focuses on the HIE layer. In our specific implementation, we were inspired by the implementation details of methods like Transformer/iTransformer for the FFN part and used conv1D to replace Linear. Indeed, our actual convolution kernel (lookback window) size is set to $L-1$. However, typically in CNN computations, the convolution kernel size is considered a hyperparameter. For example, the convolution kernel in MICN[1] is actually linearly related to $L$, but the theoretical complexity of MICN remains $O(L)$. Therefore, when calculating the theoretical complexity in our case, we followed a similar approach. (2) The HIE layer can still be implemented using Linear. In this scenario, due to parameter sharing, the way each time step processes information from its previous $L-1$ time steps is exactly the same. Hence, we can follow the approach of WITRAN[2] in handling slices by expanding the dimensions that need to be processed in parallel and merging them into the sample dimension. In this case, Linear reduces $L-1$ dimensions to 1, so its theoretical complexity remains $O(L)$. (3) In our time series forecasting task, the lookback window size of PGN is set to $L-1$ to fully capture long-term information from the input signal. However, since PGN is a general paradigm, when applied to other tasks, the lookback window can be fixed to a certain value or adjusted. In such cases, regardless of whether implementation (1) or (2) is used, the theoretical complexity remains $O(L)$ or lower. (4) For self-attention, its computation between sequences is entirely different from PGN because it requires matrix operations to obtain information between every two time steps. Therefore, the necessary $O(L^2)$ theoretical complexity cannot be reduced using the aforementioned ways. From this, we can see that the essence of PGN is fundamentally different from self-attention. Secondly, In terms of innovation, we have focused on proposing PGN as the successor to RNN to address the issues stemming from RNN's excessive reliance on recurrent information propagation and overly long information pathways, including: (a) difficulty in capturing long-term dependencies in signals; (b) gradient vanishing/exploding; (c) low efficiency in serial computation. (1) Analysis of information propagation paths: PGN, through the HIE layer, can capture historical information at each time step with an $O(1)$ information propagation path, followed by information selection and fusion through Gated Mechanism, thereby altering RNN's $O(L)$ information propagation path. A shorter information propagation path lead to less information loss, better captured dependencies, and lower training difficulty. Consequently, PGN can more effectively capture long-term dependencies of inputs signals. And, as the information path shortens, it naturally resolves the issue of gradient vanishing/exploding. (2) Regarding efficiency, the thoughtful design of PGN allows for shared parameters in HIE, enabling information from subsequent time steps to be computed in parallel without waiting for the completion of computations from preceding time steps. This approach achieves good efficiency and resolves the inefficiency in serial computation of RNN. In conclusion, through its innovative approach, PGN addresses all the limitations of RNN. Therefore, we claimed that PGN can be considered as the new successor to RNN. Refs [1][2] details are in Sec. G of Global Rebuttal. > Q2: The above discussion can also be tested in the Efficiency of Execution experiment. Okay, thank you for your suggestion. We replaced PGN with self-attention and conducted efficiency experiments. Due to space constraints, we have placed the results in newly submitted PDF file. More details please refer to Sec. C of the Global Rebuttal. > Q3: Why did the authors choose to fixate on input length of 168 instead of testing some other options? This is because, on the one hand, considering factors such as the load capacity of training devices and data collection, forecasting longer futures with shorter sequence lengths is an important research problem, which has been the focus of much previous works, such as Informer, MICN, etc. On the other hand, in order for the model to have sufficient historical information for prediction from a limited sequence length, we aim for sequence length could be longer. Therefore, we chose 168 historical time steps (7 days) to predict the future (7 days, 14 days, 30 days, 60 days) as the forecasting tasks. This setting satisfies both considerations and serves as the basis for evaluating the performance of all models. Additionally, we also conducted experiments to demonstrate the reasonableness of our second consideration. Due to space constraints, more details can be found in Sec. D in the Global Rebuttal. > Q4: In the ablation test, some datasets ... Can the authors provide some analysis on these cases? Thank you again for your detailed question. We have conducted relevant statistics and analysis, but due to space constraints, we regret to inform you that we present all the details in Sec. E of the Global Rebuttal. > Q5: I am also interested to ... what would happen if the PGN is replaced by self-attention. Within the framework of TPGN, we replaced PGN with self-attention and conducted experiments on five datasets. Due to space limitations, for experimental results and more detailed analysis, please refer to Sec. F in the Global Rebuttal. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. My major concern regarding the time complexity has been addressed. I'll raise my score. --- Reply to Comment 1.1.1: Title: Thanks Comment: Thank you once again for your time and valuable suggestions. We will include the above discussions in the revised version. --- Reply to Comment 1.1.2: Title: Thanks once more Comment: After we sent comments earlier, we discovered a bug in the system's notifications as we did not receive any messages. However, we have noticed that this issue appears to have been resolved. Therefore, we sincerely thank you once more. Thank you once again for your time and valuable suggestions. We will include the above discussions in the revised version.	Summary: This paper proposes a new network called PGN to capture the long-term dependencies of time series. Based on PGN, this paper further design TPGN for long-range time series forecasting. TPGN consists of two branches to respectively capture the long-term periodic patterns and short-term information of time series. Extensive experiments are conducted to show the effectiveness and efficiency of the TPGN. Strengths: S1. This paper is easy to follow. The motivations are clearly described by figures. The authors thoroughly analyze the information propagation modes of different time series forecasting models and explore new information propagation path to improve TS forecasting effectiveness. S2. The design of the PGN is novel, which is a completely new network architecture and can effectively solve the inherent problems of classic RNN models. Both experimental results and theoretical analysis show the effectiveness and efficiency of PGN. S3. This paper proposes TPGN upon PGN, which capture both the long-term and short-term characteristics of the time series with low computational complexity. S4. Experiments are sufficient. Five benchmark datasets are evaluated and the most representative models proposed recently are included in the experiments. Weaknesses: W1. The computational complexity of TPGN is not well discussed in this paper, and it would be better if the inference efficiency was adequately discussed as the time series size increases. W2. Some presentation needs to be improved. For example, it is difficult for readers to quickly get important conclusions on Table 1 and Table 4. Technical Quality: 3 Clarity: 3 Questions for Authors: In table of experiment comparison, could you explain why TPGN-long outperform TPGN in some cases. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have fully discussed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate Reviewer K1qz for offering valuable insights and recognizing of our work. > Q1: The computational complexity of TPGN is not well discussed in this paper, and it would be better if the inference efficiency was adequately discussed as the time series size increases. The TPGN architecture is designed to capture information more effectively from historical sequences. The complexity of processing input time series can be referenced in Sec. 3.3. For the forecasting module, our approach involves concatenating the local long-term and globally aggregated short-term information obtained by TPGN to predict the future at different time points within a period. As a result, the dimensionality of the forecasting module changes from $P \times (2 \times dm)$ to $P \times R_\mathrm{f}$, with a complexity of $O(R_\mathrm{f})$. As $R_\mathrm{f} \times P = L_f (L)$, the complexity of the forecasting module is $O(\sqrt{L})$. Indeed, it is also important to note that since both the input and output serires must be fully stored in memory, the total memory complexity should be $O(L)$. However, this does not conflict with the $O(\sqrt{L})$ complexity of TPGN. As the input seriers grows linearly, the complexity of TPGN also increases in a square root manner, with the actual time and memory overhead reflected in Figures 4(c) and (d). When the forecasting series grows linearly, although theoretically the complexity of TPGN should also increase in a square root manner, when $d_\mathrm{m}$ is relatively large, the overhead of the forecasting module is much smaller compared to the encoder part. Therefore, as shown in Figures 4(a) and (b), the actual cost variation is not significant as the forecasting sequence grows. > Q2: Some presentation needs to be improved. For example, it is difficult for readers to quickly get important conclusions on Table 1 and Table 4. Thank you for the reminder. In the revised version, we will carefully check the entire paper and enhance its presentation by adjusting colors, bolding text, particularly to emphasize the key conclusions in Table 1 and Table 4. > Q3: In table of experiment comparison, could you explain why TPGN-long outperform TPGN in some cases. From Table 2, we observe that in most datasets across various forecasting tasks, TPGN outperforms TPGN-long, with the exception of a few instances where TPGN-long performs better than TPGN in the Traffic dataset. Observing Figure 12, the Traffic dataset primarily exhibits strong periodicity along with short-term variations in different patterns. In this context, the long-term module in TPGN is already adept at capturing the predominant strong periodicity in sequences. However, in certain cases where the short-term variation patterns exhibit strong randomness, it may introduce disturbances to the short-term module of TPGN. Hence, there are scenarios in specific forecasting tasks where TPGN-long outperforms TPGN in terms of the MAE metric. Nevertheless, we also observe that TPGN remains best performance in terms of the MSE metric in these forecasting tasks. --- Rebuttal Comment 1.1: Title: Thanks for rebuttal Comment: I have carefully read the rebuttal and all my concerns have been well addressed. I will accordingly raise my rate. Thanks. --- Reply to Comment 1.1.1: Title: Thanks Comment: Thank you for your time and positive comments!	Rebuttal 1: Rebuttal: We sincerely thank reviewers for their thorough review and valuable suggestions. # A Ablation analysis of the normalization layer on the Traffic and ETTh1 datasets The experimental results can be found in Table A in the PDF file (newly submitted). We observed a noticeable decrease in performance when the normalization layer was removed from the two datasets. This drop can be attributed to the inherent characteristics of the datasets. In Table 6, we calculated the distribution (Mean and STD) of the training and validation set for each dataset. The results indicate significant differences in the distributions of the Traffic dataset and the ETTh1 dataset. These differences can introduce disturbances to the model, hence the $norm$ should be set to 1 in this case. More details can be found in Appendix H. # B Specific hyperparameters for reproducing our model We have presented the specific hyperparameters of our model for different tasks across all datasets in Table B of the PDF file (newly submitted), facilitating the direct reproducibility of our experimental results. # C Efficiency of Execution experiment of replacing PGN with self-attention in TPGN Within the TPGN framework, we replaced PGN with self-attention (referred to as TPGN-Attn). The results indicate that the overall practical overhead of TPGN-Attn is inferior to TPGN, as shown in Figure B of the newly submitted PDF file. Specifically, following the setup outlined in Sec. 4.3, we conducted two sets of efficiency experiments on TPGN-Attn. In one set of experiments, we kept the input length fixed at 168 and varied the output length to 168/336/720/1440, while in the other set, we fixed the output length at 1440 and varied the input length to 168/336/720/1440. We integrated the results of TPGN-Attn with Figure 4, resulting in a new Figure B (For specific details, please refer to the newly submitted PDF file). # D The experimental results for forecasting 1440 steps using 96 steps as input We compared the performance of forecasting 1440 steps using 96 steps input to using 168 steps input. The results show that, in the 96-1440 scenario compared to the 168-1440 scenario, all models exhibited varying degrees of performance decline, with TPGN still demonstrating its advantage. This suggests that increasing the input length appropriately when predicting longer futures with shorter inputs provides the model with more information for improved performance. The detailed experimental results can be found in Table C in PDF file (newly submitted). Specifically, we selected the three models ranked second in Table 1, WITRAN, iTransformer, and TimeMixer, and compared their performance in experiments predicting 1440 steps using 96 steps. # E An analysis of the experimental results regarding TPGN and its LSTM/GRU/MLP ablations We calculated the average improvements of TPGN compared to TPGN-LSTM/TPGN-GRU/TPGN-MLP across various tasks on each dataset. We found that the improvements were not as pronounced on the ETTh1 and Traffic datasets compared to the other datasets. Additionally, we also calculated the improvements of TPGN over the second-best model performance in Table 1 in terms of MSE across various tasks. We have summarized the results as follows. The table presents the average improvement of TPGN over the compared models across various tasks. \| Dataset \| The second-best model \| TPGN-LSTM \| TPGN-GRU \| TPGN-MLP \| \|:-------:\|:---------------------:\|:---------:\|:--------:\|:--------:\| \| Traffic \| 9.38% \| 3.54% \| 3.33% \| 19.54% \| \| ETTh1 \| 3.79% \| 5.14% \| 5.52% \| 4.02% \| (1) On the ETTh1 dataset, although the numerical improvement of TPGN over its LSTM/GRU/MLP ablations is not substantial, comparing TPGN with the second-best model reveals a significant degree of improvement. This discrepancy can be attributed to the characteristics of the ETTh1 dataset, making the improvement appear less pronounced compared to other datasets. (2) On the Traffic dataset, TPGN shows less improvement over its LSTM/GRU ablations. This is due to the strong periodicity of the Traffic dataset, as observed in Figure 12. In this scenario, the TPGN framework plays a crucial role in capturing this strong periodicity effectively. Despite PGN in TPGN using only one gate, its performance surpasses that of GRU with two gates and LSTM with three gates, demonstrating that PGN can serve as the new successor to RNN. (3) Furthermore, the results in the above Table reveal that the performance of the ablated variants of LSTM/GRU/MLP ablations across different datasets, indicating their limited adaptability to diverse domains. In contrast, TPGN demonstrates a greater advantage in this regard. # F Forecasting performance of replacing PGN with self-attention in TPGN Within the framework of TPGN, we replaced PGN with self-attention (called TPGN-Attn). The results show that TPGN exhibits a significant advantage over TPGN-Attn. The specific experimental results are presented in Table D in the PDF file (newly submitted). Based on the results in the Table D, we calculated that TPGN outperforms TPGN-Attn in terms of MSE, with an average improvement of 26.97%. In particular, TPGN demonstrates an average reduction in MSE of 27.18% for the ECL dataset, 24.87% for the Traffic dataset, 6.99% for the ETTh1 dataset, 29.62% for the ETTh2 dataset, and 46.19% for the Weather dataset. The analysis above demonstrates the advantages of TPGN. # G References MICN[1]: Wang, H., et al. (2023). Micn: Multi-scale local and global context modeling for long-term series forecasting. In The eleventh international conference on learning representations. WITRAN[2]: Jia, Y., et al. (2023). WITRAN: water-wave information transmission and recurrent acceleration network for long-range time series forecasting. In Proceedings of the 37th International Conference on Neural Information Processing Systems. Pdf: /pdf/eb3c22823be4ebc0ebb66f4c33c9e6a26e5f2c08.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting	Accept (poster)	Summary: This paper presents a method for stylizing 3D Gaussian Splatting (3DGS) using a single reference image. Unlike NeRF, which uses a structured representation, 3DGS is an unstructured discrete representation that tightly binds geometry and appearance to each Gaussian splat. To address this challenge, the paper introduces a texture-guided control mechanism, which differs from the position-guided approach used in the original 3DGS paper. This new mechanism effectively edits the appearance of a pretrained 3DGS to match the detailed texture from the reference image while preserving the original geometric structure. Strengths: + The main novelty of this work lies in the Gaussian splitting strategy, which is based on the color gradients of all Gaussians over iterations, rather than the positional gradients used in the original 3GDS approach. I find this approach to be quite neat and well-suited for the task. + The ablation study demonstrates the benefits of using this approach, including a reduction in the number of Gaussians needed to model the details of the reference texture. Weaknesses: + The novelty of the method seems somewhat limited, as it is largely based on Ref-NPR to enable image-reference-guided stylization. + It is unclear how well the method would perform if the geometry is also heavily stylized, rather than just the appearance. + The results (specially the video results) presented are quite limited, e focusing primarily on simple synthetic scenes with white backgrounds, and do not demonstrate the method's effectiveness on more complex scenes. Technical Quality: 2 Clarity: 3 Questions for Authors: + Can this method be applied when the geometry is heavily stylized? Most of the stylization examples seem to focus only on color/appearance. + If possible, could the author share the video results for the Fern and Truck scenes in Figure 6? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The limitations of the method are adequately addressed, and interesting future directions are proposed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your constructive feedback and thoughtful comments! Please note our top-level comment and additional experimental results in the rebuttal PDF. Below we address your questions and concerns. --- wrt to novelty. The core contribution of this paper is to enable stylizing 3DGS using an aligned reference image, which can benefit many real applications given its real-time rendering speed. For the first time, we illustrate that the bottleneck of stylization lies in the nature of 3D Gaussians rather than stylization techniques. Therefore, we made important contributions on this new 3D representation by (1) identifying the entangled appearance and geometry of 3DGS is the primary bottleneck, and (2) proposing a set of novel techniques to address the entangelement, resulting in (3) a complete framework, for the first time, enables real-time stylized view synthesis. These novel designs are crucial for obtaining high-quality renderings. We agree that we follow the style transfer techniques from Ref-NPR but they are perpendicular to our core contributions. We will make our contributions more clear in the revised paper. --- wrt stylizing geometry. Following existing works [8, 9, 3, 10, 6], this paper focuses on stylizing the appearance, which remains an open problem for 3DGS. Our work aims to enable precise texture editing based on a content aligned reference. We agree that further stylizing the geometry will be very interesting, but would like to also mention that such task is significantly more challenging given current problem setup, i.e. using one reference image, due to the single-view shape ambiguity. Existing methods achieving high-fidelity geometry stylization often require a very different set of techniques such as shape prior$^1$, text guidance$^2$ and/or generative modeling$^{3,4,5}$ that is beyond the scope of this work. Combining our method with these techniques to enable both geometry and appearance stylization for 3DGS is an open and interesting future direction. --- wrt video results on more complex scenes. Besides synthetic scenes, the demo video also contains two forward-facing scenes (1:20 and 1:33). We will include more complex scenes like Fern and Truck in the updated video. Unfortunately, according to the rebuttal instruction and further confirmed by the AC, we found that it is not possible to upload additional videos to openreview during the rebuttal phase. Instead, in Figure 6 in the rebuttal PDF, we render images from multiple viewpoints of requested scenes as a workaround. Our method can achieve consistent 3D stylization for these more complex scenes. --- Reference: 1. Bao, Chong, et al. "Sine: Semantic-driven image-based nerf editing with prior-guided editing field."CVPR. 2023. 2. Wang, Can, et al. "Nerf-art: Text-driven neural radiance fields stylization." IEEE Transactions on Visualization and Computer Graphics (2023). 3. Haque, Ayaan, et al. "Instruct-nerf2nerf: Editing 3d scenes with instructions." ICCV. 2023. 4. Chen, Yiwen, et al. "Gaussianeditor: Swift and controllable 3d editing with gaussian splatting." CVPR. 2024. 5. Wang, Junjie, et al. "Gaussianeditor: Editing 3d gaussians delicately with text instructions." CVPR. 2024. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed answers. I appreciate the main novelty proposed by the method and thus, still recommend Borderline Accept. However, I maintain that the scope of the contribution/novelty is quite limited (no geometry stylization, reliance Ref-NPR and the fact that the real time rendering speed is from the choice of using GS itself, rather than contribution from the paper) hence I am not raising my score. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you for your valuable response and assessment. We are glad to see that our main novelty is appreciated. Thank you once again for your time and support towards this manuscript.	Summary: The paper presents an optimization-based approach for style transfer of a (pre-baked) 3D scene represented by a 3D Gaussian splatting (3DGS). In order to fine-tune the given 3D scene with a style reference image of a single view, the authors suggest using a texture-guided controlling algorithm, which modifies the densification algorithm of the original 3DGS by focusing on the color gradients. The training loss is also modified to include depth-based geometry regularization and additional guidance provided by generated pseudo-views based on the 3D projections of the given style reference onto novel view cameras. The experiments are performed upon the existing public weights of 3DGS, where the method is compared with three NeRF-based methods, ARF, SNeRF, and Ref-NPR. Strengths: 1. As demonstrated in the supplementary materials and figures in the manuscript, the method seems to work well with the pre-baked 3DGS weights. 2. Detailed related work section enlightens novice readers to get familiar with the field of style transfer of 3D scenes. 3. Adequate level of implementational details are provided. Weaknesses: 1. Although the topic and the approach presented in the paper seems adequate, the presentation of those can be much better. For example, since the authors have modified the original training algorithm of 3DGS in Section 3.2 of the manuscript, and this seems to be the most significant contribution of this paper, they can use Algorithmic/Algorithm2E features of LaTeX or present with a pseudocode of the densification algorithm to more clearly present the key differences between theirs and the original 3DGS. 2. The components of the proposed loss functions, such as the TCM, the pseudo-view loss, the depth regularization, and the color-matching loss, which are originally devised to work with NeRF-based scene representation (Ref-NPR). I do not want to argue with the novelty of this adoption, but I believe that the design decision should be more firmly verified. Even though these losses may be generalizable to 3DGS-based representations as the paper implies, this hidden claim should be re-assessed with each component on the compatibility with the new representation (3DGS). In other words, ablation studies for these loss functions can be carried out just like [Figure 6 of Ref-NPR paper](https://ref-npr.github.io/assets/2212.02766.pdf) in order to justify the fitness of the proposed loss function with 3DGS representations. 3. I understand that an exhaustive quantitative analysis in this topic can be very difficult to design, but comparing the results with only one table seems not promising enough. For example, detailed tables with each test scene, just like [Table B.1 of Ref-NPR](https://ref-npr.github.io/assets/2212.02766.pdf), can be added with more visualization. 4. The paper could be much better with visualization of how different style reference images affect a single scene with the proposed algorithm. For example, Ref-NPR shows results with multiple style inputs acting on a single baked scene. As a summary, my key concern is the (1) representation of the materials, the (2) justification of the presented/adopted components (the losses, the densification algorithms), the (3) lack of quantitative comparison table of each scene, and the (4) lack of comparison of the results from different style images. The main contribution of the paper I believe is to report the results from applying the training algorithms of Ref-NPR to 3DGS-based representations with proper algorithmic modification to make it suitable for 3DGS. One requires to compare at least all the cases demonstrated in Ref-NPR in order to justify that this training scheme for style transfer is better suited for 3DGSs than NeRFs. Therefore, unless the mentioned points are addressed, I believe this version of the manuscript is not ready for publication in this venue. Technical Quality: 3 Clarity: 1 Questions for Authors: 1. How were the balancing parameters lambda of equation (8) obtained? Are the values of these hyperparameters critical in achieving high quality results, or are the optimal set of values differ across different stylization tasks? If so, providing the recommendation of choosing these hyperparameters will make the work more applicable. 2. Since the approach only densifies (and not prunes, may I guess) the Gaussians, the resulting scene should be much heavier than the original. How much are the number of Gaussians change in the shown experiments? How the quantitative scores (Ref-LPIPS etc.) change as the number of Gaussians increases? Is there any recommendation to modulate the size of the stylized Gaussian splatting? Please note that these questions are not counted in my overall scoring. Confidence: 4 Soundness: 3 Presentation: 1 Contribution: 2 Limitations: Yes, the limitations are adequately addressed in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable comments and constructive feedback! Please note our top-level comment and additional experimental results in the rebuttal PDF. Below we address your questions and concerns. --- wrt presentation of the materials. Thanks for your constructive suggestion. We have careful revised the manuscript and will keep revising it to improve the presentation. We have included the suggested pseudo code of our densification algorithm using Algorithm2E features in Algorithms 1&2 in the rebuttal PDF. Compared to the original algorithm: ReGS relies on 1. color gradient instead of positional gradient as the densification guidance, and 2. uses structured densification, i.e., replacing a parent Gaussian with a structured set of tiny Gaussians instead of two ''medium-sized'' ones in original 3DGS. These designs are crucial for addressing texture underfitting and verified by the experiments (see Sec. 4.2 and Appendix C, and Figure 4 in the rebuttal PDF). --- wrt justification of the adopted loss functions. We would like to first clarify that only TCM and color matching loss are adopted from Ref-NPR. Depth regularization and pseudo-view loss are proposed in this work and have already been studied and ablated in Figure 4 (b)&(c) in the paper. While TCM and color matching loss are not our main contributions, we agree that re-assess them is necessary and helpful. We report their ablation results in Figure 1 of the rebuttal PDF. As shown, using color matching loss reduces color mismatch and using TCM loss removes artifacts in the occluded areas. These findings are similar to the observations in Ref-NPR. These results suggest that the adopted losses indeed work for our 3DGS-based model. We will include this experiment in the revised paper. --- wrt quantitative comparison table of each scene. Thanks for your suggestion. Following Table B.1 of Ref-NPR, we report Ref-LPIPS score of each scene in the table below. Our method consistently outperforms baselines. We will include this detailed table in the revised manuscript. \| Ref-LPIPS $\downarrow$ \| Chair \| Ficus \| Hotdog \| Mic \| Flower \| Horn \| Truck \| Playground \| Average \| \|:------------------------------------------:\|:-----:\|:-----:\|:------:\|:---:\|:------:\|:----:\|:-----:\|:-----------:\|:-------:\| \| ARF \| 0.185 \| 0.123 \| 0.300 \| 0.146 \| 0.619 \| 0.502 \| 0.683 \| 0.592 \| 0.394 \| \| SNeRF \| 0.188 \| 0.129 \| 0.283 \| 0.138 \| 0.646 \| 0.492 \| 0.702 \| 0.663 \| 0.405 \| \| Ref-NPR \| 0.164 \| 0.122 \| 0.273 \| 0.126 \| 0.289 \| 0.471 \| 0.669 \| 0.596 \| 0.339 \| \|ReGS \| 0.127 \| 0.119 \| 0.175 \| 0.104 \| 0.134\| 0.367\| 0.454\|0.472\|0.202\| --- wrt results for different style images acting on a single scene. Thanks for your suggestion. We have included results of multiple style references acting on a single baked scene in Figure 2 in the rebuttal PDF. We temporarily omit the references and original content views here due to limited space. We will include more visual results in the revised manuscript. --- wrt balancing parameters of equation (8). We perform a simple grid search for these hyperparameters on Blender scenes, and empirically found these values generalize well on other scenes and across difference style references. We will include these details in the revised paper. --- wrt number of Gaussians changed during stylization. Our method maintains the original opacity-based pruning strategy during training, therefore do prunes Gaussians with low opacity. By design, new Gaussians are adaptively created in the texture underfitting areas. The number fully depends on the complexity of the reference texture. Therefore, the resulting scene is not necessarily heavier than the original. If the overall reference texture is simpler than the original scene, there will be less Gaussians compared to the original scene. For example, the wood-drums case shown in the Figure 4 in the paper has less Gaussians (0.29M) after stylization than the original scene (0.41M). If the reference texture is more complex, then the resulting scene can be heavier than the original scene. For example, the mic scene in Figure 1 in the paper has 0.25M more Gaussians after stylization. The quantitative scores will increase as new Gaussians at underfitting areas are created to fill the missing texture, but will remain the static after convergence (i.e. no texture underfitting remains). We will include this discussion in the revised paper. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal with rich demonstration. Believing that these results and additional discussions will be included in the final manuscript, I do not longer oppose to publication of this work. However, sharing my concerns with the reviewer JWsE, the work still seems to be an adoption of Ref-NPR to 3DGS, and thus the scope and the novelty of this work is not significant enough for higher acknowledgement (e.g., awards). I will raise my score accordingly. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you for the valuable comments and elevating the score. We are committed to incorporating the suggested evaluations and discussions in the revised manuscript. Thank you once again for your time and assessment.	Summary: The paper proposes a method to stylize 3D Gaussians using a texture guidance. The method takes a pretrained 3D Gaussian model and one content-aligned reference image as inputs and outputs a stylized 3DGS model which could be rendered at real-time framerate. Several techniques, including structured densification, depth-based geometry regularization and view-consistency constraints are introduced to achieve an effective stylization which performs better than previous state-of-the-art work. Strengths: 1. The paper is generally well-written and easy to follow. 2. The insight on color gradients is interesting and works well. The method seems promising for appearance editing on Gaussians. 3. Both qualitative and quantitative evaluations show noticeable improvement compared to previous work. Weaknesses: 1. The methodology seems largely inspired by Ref-NPR, though adapted to fit the 3D Gaussians. Readers may have to read Ref-NPR first in order to understand the motivation behind the design choices, especially in Section 3.4. 2. The superscript $(x, y)$ in Eq. 5 is not explained. 3. Minor indentation issues on L154, L188, L198, and L224. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. As you have mentioned, calculating TCM loss is slow. An ablation may better explain why TCM must be introduced despite its long running time. 2. Is it possible to use multiple views as texture references? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Limitations are addressed in the supplemental material. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your constructive feedback and thoughtful comments! Please note our top-level comment and additional experimental results in the rebuttal PDF. Below we address your questions and concerns. --- wrt inspired by Ref-NPR. Thanks for pointing this out. We will include a more detailed description of Ref-NPR in the updated Appendix to help readers understand the motivation behind Sec. 3.4. --- wrt superscript $(x,y)$ in Eq. 5. $(x, y)$ denotes the $xy$ coordinates on a feature map. Therefore, Eq. 5 refers to find the location $(x^{\}, y^{\})$ on $F_{I_{R}}$ that has closest feature-wise distance to the given location $(x,y)$ on $F_{I_{i}}$. $(x^{\}, y^{\})$ is then used to reconstruct a guidance feature $F_{G_{i}}$ for TCM (Line 218). We will clarify this in the revised manuscript. --- wrt minor indentation issues. Thanks for your careful reading. We will fix these issues in the revised paper. --- wrt ablation on TCM loss. Thanks for your suggestions. As stated in Line 214, TCM loss is used to spread stylized appearance to the occluded areas, so that the entire scene is stylized. We have included an ablation on TCM in Figure 1 (b) in the rebuttal PDF. Without TCM, the model cannot properly stylize the unseen areas and produces artifacts. Applications like appearance editing (Appendix D), that do not require stylizing entire scene, do not need TCM loss. We will include this ablation in the revised paper. --- wrt multi-view texture references. Yes, our method can be naturally extended to take multi-view texture references. We leave such exploration for future work. --- Rebuttal Comment 1.1: Comment: Thank you again for your valuable comments. We have tried our best to address your questions (see rebuttal PDF and above), and will carefully revise the manuscript by following suggestions from all reviewers. Please kindly let us know if you have any follow-up questions.	Summary: The paper proposed a texture-guided Gaussian densification strategy for exemplar-based 3DGS style transfer with content-aligned reference, while preserving original geometry by depth supervision. During 3D stylization with a style template reference, the introduced texture-guided Gaussian control strategy can geometrically adaptively densify 3D Gaussians for fine-grained texture optimization. Relying on the advanced representation of 3DGS, the stylized scene can achieve real-time rendering of novel views. Strengths: 1. The paper proposed a decent design of style transfer for a 3DGS scene while preserving geometry by depth supervision. The novel texture-guided control of Gaussian densification assists in optimizing texture with high-frequency details. I believe this strategy worths attention beyond 3DGS appearance stylization. 2. The Stylized Pseudo View Supervision works better than other multi-view consistent stylization baselines, in terms of semantic consistent stylization for uncovered areas by reference view. 3. The elaboration of methodology is technically sound, which is possibly reproduced. 4. The experiments and evaluation are convincing with ablation studies and baseline comparisons. And paper experimented on diverse scenes covering objects, forward-facing scenes, an unbounded scene. But I still have some main concerns mentioned in Weaknesses 2. Weaknesses: 1. The paper mainly concerns fine appearance optimization by densification and depth supervision. For 3D stylization, geometric stylization and editing could be tried or discussed based on proposed method. For example, stylizatin given an edited view with minor shape changes. 2. The most innovative and inspiring part is the Texture-guided Gaussian Control with texture guidance plus structured desification. However, the experiment part can be further improved: 2.1. In Appendix C, there an ablation study by comparing original 2 Gaussians and proposed 9 Gaussians densification set. There is no solid and scientific validation for the best selection of the number 9. Please see details in Question 2. 2.2. There is no ablation study of ablating only texture guidance (i.e. use original position gradients as guidance), or ablating only structured densification (i.e. use original densification scheme). Current Sec 4.2 ablation study of Texture-Guided Control show the joint effect of texture guidance and structured desification, which cannot show the effects come from the joint cooperation or from one dominant strategy. Please see details in Question 3. 3. A minor point and suggestion. For evaluation comparisons, the paper mainly compare with baselines with Plenoxels representation. Since ReGS's fast training and rendering capability replies on 3DGS, even ablation studies provide good validataion, I still expect comparisons with baselines with 3DGS, e.g. reproduce 3DGS-version SNeRF. 4. Minor issue in related work section. The paper should stress 2D and 3D stylization involve only image-exemplar based neural style transfer. Since this work finishes edited-view guided stylization, methods of text-guided dataset editing for optimization such as Instruct-NeRF2NeRF/Instruct-GS2GS is also a suitable related work. There are also some concurrent work stylizing 3DGS scenes, such as StyleGaussian, StylizedGS, Gaussian Splatting in Style. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. For texture guidance, the paper selects color gradient as hints for desification, which is a straightforward constraint hint. Is the selection based on trials among all variables such as scales, colors, rotations, opacity, etc.? If yes, what are differences among different gradient hints? 2. In the ablation study of Structured Densification (in Appendix C), I would suggest to conduct an experiment with different numbers of a dense set of Gaussians for each responsible Gaussian to be splitted, varying from original 2 to proposed 9 or even larger number. There is not enough experimental statistics to support the densification strategy via replacing by a denser set of 9 Gaussians, rather than smaller 5, or larger 16 Gaussians. In addition, in Appendix C default setting is based on position-gradient-guided or proposed color-gradient-guided density control? 3. In Texture-guided Gaussian Control, which one between Texture Guidance and Structured Densification is more important? Or only when both jointly work, ReGS can gain better performance than naive densification strategy? 4. I wonder if this Gaussian densification strategy supports original reconstruction and other downstream tasks. I would like to see more analysis and insights particularly for Questions 1-3 in the discussion phase. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper discussed limitations and provided some potential solutions. The paper does not involve potential negative social impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and constructive suggestions! Please note our top-level comment and additional experimental results in the rebuttal PDF. Below we address your questions and concerns. --- wrt geometric stylization. Thanks for your suggestion. We agree that also editting the geometry will be very interesting. We believe our method do able to handle minor shape changes, for example, by relaxing the depth supervision. However, we would like to also mention that precise geometry editting based on a reference image is inherently more challenging due to single-view shape ambiguity. To achieve high-quality geometry stylization, existing methods often adopt a very different set of techniques such as shape prior$^1$, text guidance$^2$ and/or generative modeling$^{3,4,5}$ to hallucinate missing geometry, which are beyond the scope of this work. Combining our method with these techniques for joint geometry and appearance editing is an open and interesting future direction. We will include more discussions in the revised paper. --- wrt the selection of the number 9 in densification. Thanks for your suggestion. We actually conducted a similar experiment when developing this method, and we found the current splitting strategy provides the best overall performance. Here we conducted the suggested ablation on the number of Gaussians and present the results in Figure 3 in the rebuttal PDF. We plot the PSNR value between the style reference and the corresponding stylized view to quantitively show the texture fitting capability using Blender scenes. As shown, when the number is small, the model fails to capture the target texture details. As this number grows, the performance becomes saturated. When the number equals 9, the model can achieve peak performance but also without inducing many excessive Gaussians that might slow down rendering. We will include this experiment in the revised manuscript. --- *wrt the Default* setting for Appendix C.** The Default setting is based on the color-gradient-guided density control (same as Ours) and therefore we can fairly ablate structured densification. In the experiment, we showcase the effectiveness of the proposed structured densification by studying how the performance will be affected when we remove such design (i.e. by switching to the default strategy) from the full model. We will clarify this in the revised paper. --- wrt ablation on texture guidance. Thanks for your suggestion. Here we conduct the suggested ablation study on texture guidance, where we construct the baseline by removing texture guidance from the full model (i.e., switching to the default positional-gradient guidance). We report the comparison results in Figure 4 in the rebuttal PDF. As shown, without texture guidance, the model fails to capture tiny texture details in the reference. We will include this experiment in the revised paper. --- wrt the importance between Texture Guidance and Structured Densification. From the Appendix C and Figure 4 in the rebuttal PDF, one can see that both strategies are equally important. Removing either of them will reduce stylization quality. --- wrt 3DGS version of baseline SNeRF. Thank you for the suggestion. We will reproduce an 3DGS version of SNeRF and use it as an additional baseline the revised paper. --- wrt related work. Thanks for your suggestion. We will revise the related work section accordingly. --- wrt other gradient hints. Yes, we have tried to use other gradient hints but none of them achieves similar level of details and fidelity as color gradient. This is because other variables (e.g. scales, rotations, and opacity) are not directly/strongly related to final appearance and thus are less sensitive to texture underfitting, especially in the high-frequency areas. --- wrt supports for other tasks. Our method can potentially benefit scene reconstruction and other downstream tasks, especially when the scene texture is complex. For example, one can simply use our method as an additional training stage to further refine texture details that are missing in the initial reconstruction. Such exploration might be an interesting direction for future work. --- Reference: 1. Bao, Chong, et al. "Sine: Semantic-driven image-based nerf editing with prior-guided editing field."CVPR. 2023. 2. Wang, Can, et al. "Nerf-art: Text-driven neural radiance fields stylization." IEEE Transactions on Visualization and Computer Graphics (2023). 3. Haque, Ayaan, et al. "Instruct-nerf2nerf: Editing 3d scenes with instructions." ICCV. 2023. 4. Chen, Yiwen, et al. "Gaussianeditor: Swift and controllable 3d editing with gaussian splatting." CVPR. 2024. 5. Wang, Junjie, et al. "Gaussianeditor: Editing 3d gaussians delicately with text instructions." CVPR. 2024. --- Rebuttal Comment 1.1: Comment: I thank the authors for the comprehensive rebuttal and additional experiments, which addressed my main concerns in W2. While this work was inspired by Ref-NPR, it presents a successful adaptation to the different and more effective 3D-GS representation. I believe the proposed optimization strategy would arouse significant interest and discussion within the research community. This paper not only serves as one of the earliest works on 3D-GS stylization, but also has the potential to benefit broader research on 3D-GS texture reconstruction/refinement and appearance editing. I would like to elevate my score given these additional experiments and more extensive discussion in the revision. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you for the kind words and elevating the score! We are glad to see our contribution is recognized and hear we have addressed your concerns.	Rebuttal 1: Rebuttal: We would like to thank all reviewers for providing constructive feedback that helped us improved the paper. We are encouraged that the reviewers think - our approach is decent (BehR), neat (JWsE), and interesting (TJDG) - designs and insights are effective (3qBC) and works well (BehR, TJDG, hmWz) - experiments and evaluation are detailed (3qBC), and convincing (BehR) - the paper is well-written (3qBC) and easy to follow (TJDG) We have been working diligently on improving the paper on several fronts, addressing the critique. Please note we have included figures for the suggested experiments in the rebuttal PDF. We address questions and concerns for each reviewer in the comments below. Pdf: /pdf/748810ae61cd2fe8cc2a5766b25f020d2c5a5eb3.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper introduces ReGS, a new reference-based 3D style transfer method that utilizes 3DGs as the 3D representation. To capture fine-grained details from the reference view, the method employs texture-guided Gaussian control to enhance density in areas where texture is under-represented. Additionally, the approach incorporates depth-based regularization and pseudo-view supervision to ensure consistency while stylizing with the reference image. The quantitative and qualitative results demonstrate that ReGS achieves superior stylization, capturing detailed and high-quality effects more effectively than previous methods. Strengths: - The paper is well-written and comprehensive, making it easy to follow. - The experiments are detailed, and the impact of each proposed method is demonstrated step-by-step. - The stylization results effectively capture fine-grained details from the reference image. - The proposed appearance-based densification approach is simple yet proves to be effective. - The choice of 3D GS for reference-based stylization results in faster rendering performance. Weaknesses: - I do not find the methods presented in the paper to be significantly novel, as they give the impression of being a 3DGS-adapted version of Ref-NPR. While I acknowledge the differences and novelties introduced to effectively adapt reference-based stylization to the 3D-GS setting, I do not see a critical distinction in terms of the 'style transfer technique' itself, once the modifications specific to the 3D-GS settings are set aside. This is primarily because the stylization pipeline (Section 3.4) closely mirrors that of Ref-NPR, without introducing new improvements or modifications. - The qualitative comparison presented in Figure 6 appears unfair. As I understand, ARF and SNeRF in this experiment are stylized using a stylized reference view, and the discrepancies between these results and the reference view are emphasized. However, the primary objectives of ARF and SNeRF differ from those of Ref-NPR and ReGS, as they are not specifically designed for reference-based stylization. Consequently, there is no inherent need for their stylization results to strictly adhere to the reference view. I believe the authors are aware of this distinction. For a fairer comparison, it would be more appropriate for the authors to include the original 2D style image for ARF and SNeRF and conduct a qualitative assessment based on aesthetic quality. Comparisons of the ability to replicate high-frequency details and correspondence should perhaps be reserved exclusively for comparisons with Ref-NPR. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the weaknesses above. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors addressed the limitations in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and constructive comments! Please note our top-level comment with additional experimental results in the rebuttal PDF. Below we address your questions and concerns. --- wrt novelty on style transfer techniques. The core contribution of this paper is to enable precise stylization of 3DGS using an aligned reference image, which can benefit many real applications given its real-time rendering speed. For the first time, we illustrate that the bottleneck of such stylization lies in the nature of 3D Gaussians rather than the well-established style transfer techniques. Therefore, we made important contributions on improving this new 3D representation by (1) identifying the entangled geometry and appearance is the primiary bottleneck for stylization, and proposing (2) a set of novel techniques (Sec. 3.2&3.3) to tackle this entanglement, resulting in (3) a complete method, for the first time, that enables real-time stylized view sythesis. These novel designs are verfied by extensive experiments (Sec. 4.2&4.3). Style transfer techniques from Ref-NPR are perpendicular to our core contributions. But we agree that these techniques might not be perfect for 3DGS and further improving them is an interesting future direction. --- wrt comparison with ARF and SNeRF on aesthetic quality. Thanks for your advice. We simply follow Ref-NPR to include ARF and SNeRF as additional baselines. We do aware such distinction and agree that evaluating aesthetic quality using the original style image for them is indeed a fairer comparison. In Figure 5 in the rebuttal PDF, we report the suggested qualitative comparison with SNeRF. The original style images are acquired from Ref-NPR authors. One can see that SNeRF produces results mimicking the abstract style of the original art, whereas our method follows the extract stylized texture in the reference image by design. We will include more visual comparisons in the revised manuscript. --- Rebuttal Comment 1.1: Comment: Thank you again for your valuable comments. We have tried our best to address your questions (see rebuttal PDF and above), and will carefully revise the manuscript by following suggestions from all reviewers. Please kindly let us know if you have any follow-up questions.	null	null	null	null	null	null
Information-theoretic Generalization Analysis for Expected Calibration Error	Accept (poster)	Summary: This paper analyzes the estimation bias and generalization error of the expected calibration error (ECE). Specifically, in a binary classification setting, the authors provide an upper bound for the total bias with an improved convergence rate, applicable to both uniform mass and uniform width binning strategies. They also determine the optimal number of bins to minimize the total bias. Furthermore, the authors utilize the information-theoretic generalization framework, particularly the Conditional Mutual Information (CMI) framework, to characterize the generalization of ECE. Strengths: 1. This paper achieves a tighter bound for total bias compared to previous works. 2. The optimal number of bins is determined using the upper bound of the total bias. Weaknesses: 1. As the authors themselves note, a significant limitation is that the analysis in this work is only applicable to binary classification. 2. Some assumptions (e.g., Assumption 2) are not well justified. 3. The writing has significant room for improvement; several arguments are unclear or misleading. Please find more details in the questions below. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Does Assumption 2 hold true in practice? Is there a way to verify it? Additionally, what is the motivation behind assuming $n_e\geq 2B$? How is this assumption utilized? If $n_{te}\leq 2B$, will Theorems 2 and 3 still be valid? 2. In the proof sketch of Theorem 2, you mention that $\mathrm{ECE}$ and $\mathrm{TCE}$ could be re-written. While they seem correct to me, could you elaborate on how $\mathrm{TCE}(f_\mathcal{I})$ is obtained in the form shown in Line 152? I did not find the details in the complete proof. 3. According to Theorem 5, do the upper bounds indicate that UWB is a better binning strategy than UMB, given that UMB has an additional $\mathrm{fCMI}$ bias term? It seems that only UMB's expected binning bias is sensitive to the training data, which might be seen as a disadvantage in terms of the upper bound. 4. The writing can be significantly improved. For example, in Line 244, you mention "Our theory might guarantee the ECE under test dataset for them." Do you mean your theory might guarantee low ECE under the test dataset? Additionally, in Lines 251-252, "This implies that if the model generalizes well, evaluating the ECE using the training dataset may better reduce the total bias than that using test dataset." Why does evaluating ECE reduce total bias? What we really care about is ECE on unseen/test data. How does evaluating ECE on training data affect this purpose? 5. Why is the metric entropy method only used for UWB? It seems that you upper bound $\mathrm{eCMI}$ by the $\mathrm{fCMI}$ term first in your proof. What prevents you from giving a similar result for UMB? 6. In Line 339-340, you mention that "a notable trend towards acquiring relatively stable nonvacuous bounds can be observed when adopting $B =\lfloor n^{1/3} \rfloor$", but according to Figure 1, it seems $B=52$ is tighter than $B=B =\lfloor n^{1/3} \rfloor$ in most cases. Could you clarify this? 7. Since $\mathrm{eCMI}$ and $\mathrm{fCMI}$ terms are key components in both standard generalization error and expected total bias of calibration error, do you have any new insights into the relationship between calibration and generalization from this perspective? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have thoroughly discussed some limitations of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to express our deepest appreciation for your insightful reviews and suggestions. We sincerely summarize our responses to you as follows. ### Q.1: Regarding the strictness of Assumption 2 and the assumption of $n_e\geq 2B$ A. First, please refer to the global response for the discussion on the necessity of Assumption 2. In practice, Assumption 2 is reasonably mild. For example, if label distributions follow a Bernoulli distribution with the mean depending on the input $x$, Assumption 2 is met [E]. This is a relatively weak assumption in binary classification. Moreover, existing studies have shown that many common benchmark datasets are consistent with this assumption (e.g., [F]). To ascertain whether the assumption of Lipschitz continuity is empirically plausible, we conducted a numerical evaluation based on ResNet experiments. This involved checking if the value of $E[Y\|f(X)=v]$, estimated through binning, exhibits relatively smooth variations. These findings are depicted in Fig. 6 of the attached PDF. The results demonstrate that the estimated quantity $E[Y\|f(X)=v]$ fluctuates smoothly to a considerable extent, thereby supporting the validity of the Lipschitz continuity assumption. This supports the robustness of our assumptions and the applicability of our methods in real-world scenarios. We will incorporate these results into our paper. Regarding the assumption $n_e\geq 2B$, it is crucial for UMB, as it guarantees the proper construction of bins. In UMB, we first use $B$ samples to build the bins. We then partition the remaining $n_e−B$ samples evenly across these bins. If $n_e < 2B$, we have $n_e−B< B$, preventing equal distribution of samples, thereby rendering UMB inapplicable. Conversely, UWB does not require this assumption since it divides the $[0,1]$ interval into equal-width bins. We will clearly outline this distinction in the revised paper. [E] Y. G. Yatracos. A Lower Bound on the Error in Nonparametric Regression Type Problems. Ann. Stat., 1988. [F] B. Zadrozny et al. Transforming Classier Scores into Accurate Multiclass Probability Estimates. KDD '02. ### Q.2: How to obtain TCE and ECE reformulations. A. The proofs of those reformulations are provided in Appendix C, specifically in Lemma 1. In the revised version, we will explicitly reference this lemma to avoid any confusion. ### Q.3: Reason why UMB has an additional term compared to UWB in Theorem 5, A. The additional term (fCMI) arises because UMB uses training data to construct bins. This means we must consider overfitting not only during learning for $f_w$ but also in the bin construction process. This leads to extra fCMI terms in UMB compared to UWB, which is a potential disadvantage. Practically, UMB is designed to prevent issues with bins having no samples when $n_e$ is small. When sufficient data is available, we can partition the entire dataset into training data for learning $f_w$ and another dataset for bin construction. This approach avoids the extra fCMI term and has been explored in previous work [22]. ### Q.4: The meaning of Line 244 and Lines 251-252 A. Intent of Line 244: Our goal is to show that, similar to how ERM bounds test loss by training loss plus a generalization gap, our Theorem 4 bounds test ECE by training ECE plus a generalization gap. Some studies [21, 26, 36] use regularization to minimize training ECE in their algorithms, so if a small eCMI or fCMI ensures a small generalization gap, the upper bound of test ECE will also be minimized, thus providing theoretical assurance akin to ERM’s generalization theory. Lines 251-252: In ERM, the expected loss can be estimated with a validation dataset, with an error rate of $O(1/n_{te}^{1/2})$. However, ECE converges to TCE at a much slower rate of $O(1/n_{te}^{1/3})$. Increasing validation data does not significantly improve this rate. Since training dataset size is often much larger than the test data size, if the generalization error (eCMI or fCMI) is small, using training data to estimate ECE could result in a smaller bias of $O(1/n_{tr}^{1/3})$. Thus, using training data for estimating TCE could be a promising approach. We will clarify these points in the revised paper. ### Q.5: The difficulty of deriving the metric entropy bound for for UMB A. Deriving the metric entropy bound for UMB is challenging due to the difficulty of constructing the appropriate $\delta$-cover of functions. As highlighted in Eq. 117 and Eq. 120, we need to set $\delta$ smaller than the width of the bins. In the case of UWB, the bin width is $1/B$; however, in UMB, the bin width is not fixed and varies according to the data. This dependency adds complexity to preparing the appropriate $\delta$-cover. ### Q.６: Clarification in Line 339-340 A. Based on the additional experiments, we will revise the explanation as follows: - The optimal $B$ is determined based on the upper bound of the total bias (Thoerem 5) and does not necessarily minimize the generalization error bound (Theorem 4). However, setting $B$ in this way ensures that we avoid vacuous values, numerical instabilities and even increase of the ECE gap as we increase $n$. For the details, please see Fig. 5 in the PDF and response to Reviewers 9r8F (Q.4) (and to Reviewer 8W41; Q.4 for TCE and the total bias evaluation on toy data experiments.). ### Q.7: New insights into the relationship between calibration and generalization A. As stated in the global response, the total bias of the TCE and ECE matches the rate achieved in nonparametric regression, conclusively showing that estimating TCE is as challenging as estimating conditional probabilities using nonparametric regression. Thus, compared to usual generalization problems such as classification error, evaluating TCE requires more samples, and achieving a fast learning rate of $O(1/n)$ under parametric learning settings seems to be not possible, highlighting a distinct difference. --- Rebuttal Comment 1.1: Title: Thank you for the responses Comment: I would like to thank the authors for their responses and apologize for not engaging in the discussion earlier. I have read all the reviews and the corresponding author responses. My concern regarding Assumption 2 has been adequately addressed, and further clarifications have been provided for other questions. I will increase my score to 5 accordingly. However, I am not assigning a higher score because the analysis in this paper is limited to binary classification, with no clear path to extending it to a multi-classification setting. While I also appreciate the minimax results, I am concerned that the analysis (and potentially the assumptions) might require significant refinement in more general or practical settings. --- Reply to Comment 1.1.1: Title: Acknowledgement and clarification of the limitation of the binary classification Comment: We sincerely appreciate your thorough review of our responses to the reviewers’ comments. And, we are delighted that you have decided to increase your score. Concerning the limitations inherent in binary classification, we addressed similar points in our response to Reviewer 8W41 (Q.1). Please refer to that section for details. In summary, our analysis can extend to multiclass scenarios involving top-calibration error, such as when using softmax for classification. In these cases, evaluating the calibration of the highest softmax score is standard practice, and these metrics can be reduced to a formulation similar to binary classification, enabling our analysis. Concerning the minimax theorem, additional assumptions are necessary because the theorem deals with worst-case distributions. Worst-case scenarios may occasionally involve settings that are unrealistic or unlikely. However, for establishing upper bounds, as in Theorem 5, Assumption 2 is generally sufficient, making it more practical and likely to be met in real-world scenarios. We hope these answers will further clarify your understanding.	Summary: This paper investigates the estimation bias in expected calibration error (ECE) for binary classification models, focusing on uniform mass binning (UMB) and uniform width binning (UWB). The authors present a comprehensive theoretical analysis, establishing upper bounds for the bias and the generalization error. Based on the convergence rates of binning and statistical bias, they identify the optimal number of bins to minimize the total estimation bias. Strengths: * The paper provides a comprehensive analysis of the estimation bias in ECE, providing upper bounds and optimal bin size choices for both UWB and UMB. * The authors further derive upper bounds for the generalization error between ECE and TCE using an information-theoretic approach. * Numerical experiments on deep learning tasks confirm that the derived bounds are non-vacuous. Weaknesses: * The provided results only apply to binary classification, and require Lipschitz continuity which may not be necessarily satisfied in deep learning models. Also, these bounds are analyzing the ECE using test data but not training data, making them less applicable since test data are not always available in practice. * The convergence rates of the information-theoretic generalization bounds heavily depend on the actual rate of eCMI and fCMI measures, which are not directly clear in analysis. In theorem 6, the authors show that eCMI scales as O(log n) based on metric entropy, but this bound involves the dimensionality d, and is thus hardly applicable to deep learning models. * For experimental results, only the statistical bias is evaluated but not the total generalization error. It is also hard to see to what extent these bounds are tight in the current results. These bounds are also hard to estimate due to the existence of eCMI or fCMI measures. I would suggest the authors additionally consider some synthetic settings where TCE, eCMI, and fCMI are analytically tractable to show the tightness of the bounds. (maybe Gaussian data points?) Technical Quality: 3 Clarity: 3 Questions for Authors: Recent information-theoretic bounds have shown improved rates of O(1/n) under the interpolating regime, and also direct computational tractability with loss CMI or entropy metrics. It may be worth some discussions on whether these techniques can be adopted to acquire tighter bounds. Tighter Information-Theoretic Generalization Bounds from Supersamples. ICML 2023. Rethinking Information-theoretic Generalization: Loss Entropy Induced PAC Bounds. ICLR 2024. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to express our deepest appreciation for your insightful reviews and suggestions. We sincerely summarize our responses to you as follows. ### Q.1: Regarding the setting of binary classification and the Lipschitz continuity. A. Although our study focuses on binary classification, we can readily extend it to multi-class settings. In existing studies [22, 8], the top-label calibration error (top CE, TCE) has been proposed as a measure for multi-class calibration. For instance, in a $K$-class classification problem, we obtain predictions for each label by the final softmax layer in neural networks. We assume that $f_w(x)\in \mathbb{R}^K$ predicts the label by $C:=\mathrm{argmax}_k f_w(X)_k$, where $f_w(X)_k$ represents the model’s confidence of the label $k \in K$. The top CE (TCE) is then defined using the highest prediction probability output by $f_w: TCE:=\mathbb{E}\|P(Y=C\| f_w(X)_C)-f_w(X)_C\|$. By considering binning only for the top score, we can compute the ECE (top-binning ECE) in a manner similar to binary classification. In this case, since we focus only on the top label, we can treat top-binning ECE in the same way as binary classification, which results in the same generalization and total bias bounds. Our results, therefore, offer flexibility to analyze the widely used top-label calibration error in multi-class settings. We will add this discussion to the paper. Regarding Assumption 2, this does not imply the Lipschitz continuity of the function $f_w(\cdot)$ itself, but the Lipschitz continuity of the conditional expectation $\mathbb{E}[Y\|f_w(x)=v]$, which is a kind of characteristic of the data itself. For a more comprehensive discussion on the Lipschitz continuity assumption, please refer to the global response provided above and the response to Reviewer kktU (Q1). ### Q.2: Regarding the necessity of test dataset in the ECE evaluation. A. If your question concerns the difference between IT-based bounds and generalization error upper bounds like PAC-Bayes, which can be evaluated using only training data, model distribution, and prior distribution, note that the IT-based upper bound can still be evaluated even if only training data is available in practice. As discussed in Line 257, $\mathrm{eCMI} \leq \mathrm{fCMI}\leq I(W;S_\mathrm{tr})$ [13,15] holds, and $I(W;S_\mathrm{tr})$ only depends on the training dataset. Thus, by replacing eCMI and fCMI in Theorem 4 and 5 with this mutual information, we get the bound that is independent of the test dataset. On the other hand, the inclusion of test data dependence in eCMI and fCMI through the index \tilde{U} provides the benefit of tighter bounds. If our explanation does not align with your concern, please provide further questions during the discussion phase. ### Q.3: Regarding the convergence behavior of our bounds and the vacuousness of the metric-entropy-based bound A. Your points are accurate. The primary distinction between IT bounds and UC bounds lies in their objectives: - IT Bounds: These are algorithm-specific and aim to provide detailed insights into how a particular algorithm performs in terms of generalization. They focus on analyzing these bounds to understand the performance of specific algorithms. By further assuming specific algorithm classes, such as stable algorithms or stochastic convex optimization, we can derive the theoretical behaviors of the mutual information. - UC Bounds: These are derived to determine the necessary sample size for achieving generalization performance based on convergence rates, independent of the algorithm used. Typically, UC bounds employ metric entropy to derive results. In response to your suggestion, we provide UC bounds that use the fat-shattering dimension, which is independent of the model’s dimension as detailed in the global response above. ### Q.4: Regarding the statistical bias evaluation, and the TCE evaluation experiment on the synthetic experimental settings. A. Evaluating the upper bound for TCE proved challenging with benchmark datasets and experiments using CNN and ResNet. However, we were able to conduct TCE evaluation experiments using a synthesized dataset based on the settings in [Z]. The results are shown in Fig. 4 of our additional PDF. We numerically evaluated both sides of Corollary 1. Since Corollary 1 does not involve eCMI, this approach allows for a more accurate assessment of the bound’s tightness and optimality. The first two rightmost figures display the total bias and upper bound of Corollary 1 under the optimal bin size, demonstrating that the bound effectively captures the behavior of the total bias. The last two figures show the TCE gap plotted against different bin sizes compared to the theoretically optimal number of bins. Despite some fluctuations due to a small sample size $n$, the actual behavior closely aligns with theoretical predictions, validating the optimal bin size. [Z]: J. Zhang et al. Mix-n-Match: Ensemble and Compositional Methods for Uncertainty Calibration in Deep Learning. ICML2020. ### Q.5: Regarding the improvement of the convergence rate A. As discussed in the reply to Reviewer Ckaf (Q.2), we cannot use the technique proposed in [37] directly. Moreover, we have derived a minimax lower bound for total bias, which is $\mathcal{o}(1/n^{1/3})$, implying that the problem of estimating TCE is as inherently challenging as estimating conditional probabilities in the context of nonparametric regression. Therefore, deriving a fast rate of $\mathcal{O}(1/n)$ under a parametric learning setting is not feasible for the ECE estimation problem. However, discussing whether the techniques adopted in the papers you suggested can be used to improve the constants of the bounds constitutes important future work. We would like to specify this in the conclusion section. --- Rebuttal Comment 1.1: Comment: Thank you for the response and additional experiments. I'll maintain my positive rating. --- Rebuttal 2: Title: Thank your for the reply Comment: Thank you for confirming our response. We believe that the additional experiments address your concerns. If there are any further issues or concerns that need to be addressed to improve the score, please let us know. We are happy to discuss them further.	Summary: The paper studies the expected calibration error using information-theoretical tools. They derive different tight fCMI and eCMI bounds in this setting. Empirical results show that the results are nonvacuous. Strengths: 1/ The paper is in general well written. Adequate discussions are given in the main body of the paper and the appendices. 2/ The paper provides the first information-theoretic comprehensive analysis of the bias associated with the ECE when using the test and training datasets. 3/ The theoretical results seem sound. I skimmed through most of the proofs (I did not go through all of them in detail) but the proofs are well-structured and easy to follow. 3/ Empirical results show that the bound is tight for deep learning models. Weaknesses: The only weakness, if any, is perhaps that the paper uses conventional machinery for deriving information-theoretic generalization bounds and that it has not developed novel proof techniques. Technical Quality: 3 Clarity: 3 Questions for Authors: Besides fCMI and eCMI based bounds, is it possible to extend the analysis and derive $\Delta$-L based bounds [37]? These bounds are typically tighter compared to fCMI and eCMI. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations are adequately addressed in the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to express our deepest appreciation for your insightful reviews and suggestions. We sincerely summarize our responses to you as follows. ### Q.1: Regarding the novelty of proof techniques ### A. First, it is important to clarify that our techniques are fundamentally different from those used in previous studies, as referenced in [11, 10, 22]. In these studies, the error bound for the ECE is derived through the following three steps: (i) Initially, it is shown that the samples assigned to each bin are i.i.d.; (ii) the Hoeffding's inequality is then applied to derive that $\|\mathbb{E}[Y\|f_{\mathcal{I}}(x)] - f_{\mathcal{I}}(x)\| = O(\sqrt{n_{te}/B})$ for each bin; and (iii) these error bounds are summed up across all bins to yield $O(B/\sqrt{n_{te}})$. This approach leads to a slow convergence, which can be attributed to the separate analysis conducted for each bin, requiring multiple applications of concentration inequalities. In our approach, we leverage a reformulation of the ECE and TCE as outlined in lines 151 and 152. These equations frame both the ECE and TCE in terms of the relationship between the empirical expectation and its population expectation. This framework allows us to apply concentration inequalities to the errors in all bins combined, specifically to the ECE and TCE themselves. Unlike traditional methods, we employ McDiarmid's inequality—notably, leveraging the technique within its proof—to derive an upper bound without the need to decompose the error into individual bins. This strategy eliminates the necessity to compute the cumulative sum of statistical biases across bins, achieving a more efficient convergence rate of $O(\sqrt{B/n})$ in expectation. While McDiarmid’s inequality is widely used, our application of it within the binning context is technically novel because it requires meticulous handling of data replacement, as detailed in Appendix D.1.1. This approach in treating the data and applying the inequality offers a marked improvement in the convergence order compared to existing studies. We would also like to highlight the novelty of Theorem 6. Our work extends beyond the standard construction of a $\delta$-cover for Lipschitz functions. As outlined in Appendix E.3, our approach involves a careful design of the $\delta$-cover to ensure that both the output of the original function $f_w$ and that of the $\delta$-cover function are categorized into the same bin. This careful alignment is significant to control the discretization error associated with the binning ECE when employing $\delta$-cover functions. This strategy, which we introduce for the first time, is crucial for enhancing the accuracy and effectiveness of our theoretical framework, providing a novel contribution to the field. ### Q.2: Relation to the result in [37] ### A. The eCMI appearing in Theorem 4 (Eq. (15)) closely aligns with the $\Delta L$ bound (loss difference) of Theorem 1 of reference [37]. Specifically, the eCMI term in Eq. (15) is not based on the value of ECE itself. Instead, it is derived from the difference between the test data ECE and the training data ECE. This approach aligns with the $\Delta L$ (loss difference) as shown in Theorem 1 of reference [37]. However, extending our bounds using the techniques of [37] presents significant challenges. The $\Delta L$ bound in [37] defines the loss gap for a single data index $i$ as $\Delta L_i$ and utilizes the symmetry of each individual index to derive fast rate bounds, as demonstrated in Theorem 4.3. In contrast, our bound requires treating all $n$ indices simultaneously. This necessity arises because ECE is a nonparametric estimator that uses all $n$ indices, unlike usual losses such as the 0-1 loss, where an estimator can be constructed using a single index. Consequently, the techniques from [37] that utilize the symmetry of a single index are not applicable to our context, making it difficult to employ the methods from [37]. Furthermore, as previously discussed, we have derived a minimax lower bound for total bias, amounting to $\mathcal{o}(1/n^{1/3})$ under Lipschitz conditions. This implies that the problem of estimating ECE is as fundamentally challenging as estimating conditional probabilities in the context of nonparametric regression. Therefore, deriving a fast rate of $\mathcal{O}(1/n)$ under a parametric learning setting, as seen in [37], is not feasible for the ECE estimation problem. We will add these discussions in the revised version of our paper. --- Rebuttal Comment 1.1: Comment: Thank you for the response. I'll maintain my positive score. --- Reply to Comment 1.1.1: Title: Thank you for the comment. Comment: Thank you for confirming our response. If there are any further issues or concerns that need to be addressed to improve the score, please let us know. We are happy to discuss them further.	Summary: This paper presents a comprehensive analysis of the estimation bias for expected calibration error (ECE), focusing on two common binning strategies: uniform mass and uniform width binning. The analysis establishes upper bounds on the bias, resulting in an improved convergence rate. Furthermore, these bounds reveal the optimal number of bins needed to minimize the estimation bias. The study also extends the bias analysis to generalization error analysis using an information-theoretic approach, deriving upper bounds that facilitate numerical evaluation for recalibration methods based on training data. Experiments with deep learning models demonstrate that the bounds are nonvacuous, due to the information-theoretic generalization analysis approach. Strengths: As the author pointed out, the existing literature lacks a theoretical analysis of the estimated ECE and a more principled approach to estimation. This paper addresses and closes this gap. Weaknesses: 1. Tightness issue of the upper bound in Corollary 1. It is commendable that the authors included a discussion on the tightness of Equation 12. However, it would be more rigorous to formally establish a minimax lower bound for the estimation bias that applies to all types of estimators. The authors could either use existing results from Tsybakov [33] or construct a worst-case analysis using Le Cam’s method to establish the lower bound. While it is acceptable if the constant does not match the upper bound, it is crucial to demonstrate the rate. 2. A drawback of information-theoretic (IT) bounds is the implicit dependency on the algorithm. For example, Theorem 7 appears very similar to Theorem 4, as the recalibration-induced dependence is encapsulated in the CMI term. The authors should provide more commentary on this aspect and clarify the connection between Theorems 6 and 5, as well as which bound is more practical for use. 3. In the caption of Figure 1, It is said that the ECE gap does not change significantly in B. How can we justify that the selection of $B = n^{1/3}$ is better? Figure 1 primarily plots the bound in (14), but as I mentioned earlier, such a bound can be very loose, and more empirical justification should be provided for the selection of optimal B. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Clarification: what is the ECE gap plotted in Figure 1 and Table 1? Estimated ECE? To my understanding, all the bounds in Figure 1 are quite loose. Shouldn't we plot the left-hand side of Equation 14 for a more accurate comparison? 2. How should the error bars for the bound values in Table 1 be interpreted? What is the source of the randomness? 3. It is not accurate to say that I(S;W)=O(log n) in Line 258, as such Barron’s result assume that samples Z are conditional i.i.d. given model parameter w. However, in the learning context, we always assume that training samples are i.i.d. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The Limitations are well addressed in section 7. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to express our deepest appreciation for your insightful reviews and suggestions. We sincerely summarize our responses to you as follows. ### Q.1: Derive a minimax lower bound for the total bias. A. Please see our global response for a minimax lower bound. ### Q.2: Regarding the drawback of IT bounds. A. The objectives of IT-based generalization bounds and UC bounds differ significantly. The former aims to derive upper bounds that are algorithm-dependent to gain a detailed understanding of how a training algorithm behaves to achieve generalization performance. This also involves analyzing these bounds to quantitatively examine the specific algorithms. In contrast, UC bounds are primarily derived to determine the sample size needed to achieve sufficient generalization performance from a convergence rate perspective, regardless of the algorithms. Next, we explain the explicit relationship between Theorems 5 (IT bound) and 6 (UC bound). In the proof of Theorem 6, we derive the upper bound of eCMI (fCMI) using the metric entropy of the specific $\delta$-cover (Eq. 124). Therefore, the metric entropy-based UC bound serves as a further upper bound to the IT bounds. In addition to this, as shown in the global response, we show the new connection between IT and UC bounds using the fat-shattering dimension, which is independent of the model’s dimension. These insights will be included in the revised manuscript. ### Q.3: Regarding the ECE gap in Figure 1 and Table 1. A. First, the ECE gap shown in Fig. 1 corresponds to the empirical estimate of the middle expression in Eq. (14). We added the formal definition in Fig. 5 of our additional pdf. The ECE gap reported in Tab. 1 is evaluated through the empirical estimator of the statistical bias: $E_{R,S_{re}} E [\|E[Y\|h_{\mathcal{I},S_{{re}}}(X)] - h_{\mathcal{I},S_{{re}}}(X)\|] = E_{R, S_{{re}},S_{te}}ECE(h_{\mathcal{I}, S_{\mathrm{re}}},S_{te})$ for existing recalibration methods due to the definition of recalibration. For our proposed method, which uses all training data as recalibration data, it is evaluated as $E_{R,S_{tr}} E [\|E[Y\|h_{\mathcal{I},S_{tr}}(X)] - h_{\mathcal{I},S_{tr}}(X)\|] = E_{R,S_{tr},S_{te}}ECE(h_{\mathcal{I}, S_{tr}},S_{te})$. Thanks to your suggestion, we realized that this explanation is too brief and would like to add a detailed explanation of how we numerically measured the ECE gap in Appendix G. Also, we believe that visualizing the estimate of the left-hand term in our theoretical framework is crucial for verifying whether the right-hand side is effectively upper bounded numerically and for analyzing the relationship between these two behaviors (refer to our answer to Reviewer 8W41, Q.4). ### Q.4: Regarding empirical justification of our optimal $B$. A. Regarding the optimal $B$ and the ECE gap, we initially showed only the ECE gap with this optimal $B$ in Fig. 1 for clarity. Motivated by your review, we examined the ECE gap for various bin sizes using the same setup as Fig. 1, and these results are presented in Fig. 5 of the additional PDF. We plotted them on a log scale to illustrate how the ECE gap and upper bound behave with different bin sizes. We found that sometimes bins other than the optimal $B$ can yield a better generalization gap. However, the optimal bin size minimizes the total bias as stated in Theorem 5, not necessarily the generalization gap (Theorem 4). On the other hand, the optimal $B$ was found to be numerically stable, although, in certain models, high variance was observed for certain bin sizes, with the ECE gap occasionally not decreasing as $n$ increases. Finally, to our knowledge, no existing work has evaluated model performance based on ECE from a generalization perspective. Our contribution is the first bound that allows for this numerical evaluation. To further assess the optimality of $B$, we need to evaluate the total bias. However, evaluating the upper bound for TCE with benchmark datasets and experiments using CNN and ResNet proved challenging. Therefore, we performed additional experiments using Toy data to more easily evaluate the tightness of the bound and the optimality of $B$. Specifically, we created Toy data following the settings in existing studies and numerically evaluated both sides of Corollary 1. Since Corollary 1 does not involve eCMI, this approach allows for a more accurate assessment of the bound's tightness and optimality. The results are shown in Fig. 4 of the PDF. The first two figures from the right display the total bias and upper bound of Corollary 1 under the optimal bin size, showing that the bound effectively captures the behavior of the total bias. The last two figures plot the TCE gap with different bin sizes compared to the theoretically optimal number of bins. Despite some fluctuations due to a small sample size $n$, the actual behavior closely aligns with theoretical predictions, validating the optimal bin size. These facts will be added in Appendix H. ### Q.5: Regarding the interpretation of the error bars for the bound values in Table 1 A. The error bars in Tab. 1 (and the standard deviation in Fig. 1, which is almost unrecognizable due to its small value) are attributed to the randomness inherent in various experimental settings during model training, i.e., randomness of the training dataset and the initial model parameters. We acknowledge that we have not explained this point clearly enough, so we will include this explanation in Appendix G. ### Q.6: The issue with the statement on Line 258 about $I(S;W) = O(\log n)$ A. Your point is correct, and the cited results consider a setting similar to Bayesian inference, involving conditionally i.i.d. samples. We cited the result as an example illustrating a scenario where mutual information is theoretically controlled. To make the meaning of the citation clearer, we plan to explicitly state that the cited paper addresses conditionally i.i.d. settings. --- Rebuttal Comment 1.1: Comment: I appreciate the authors' effort in preparing the response. It thoroughly addresses all my comments, and I am pleased to see the matched lower-bound results and additional empirical evidence supporting the claim. I will raise my score to 7—this is a solid paper. Please ensure that the new results are incorporated into the final revision, and make sure the figures in the main body are clear and easy to follow. --- Reply to Comment 1.1.1: Title: Acknowledgement Comment: We greatly appreciate your acknowledgment of our responses. We will incorporate the theoretical facts and insights, along with their proofs, discussed during the rebuttal period into the revised version. Additionally, we will improve the clarity of the existing experiments and include the additional experiments in the revised version as well. We are delighted that you have decided to increase your score. Thank you again for your insightful feedback.	Rebuttal 1: Rebuttal: We would like to express our sincere appreciation for your insightful reviews and suggestions. First, we will address the common concerns raised by the reviewers. Following that, we will address each individual question. ## Discussion about the lower bound of the total bias As pointed out by Reviewer 9r8F, here we show the lower bound of the total bias by using the result of the nonparametric binary classification in [A]. The method in [A] is the extension of Fano's method focusing on evaluating mutual information (MI) with the Yang-Barron method [B]. Conditioned on $W=w$ we express $v=f_w(x)$. Let $p(v)$ be the probability density of $v$ induced by $p(x)$ using $f_w(x)$ and $\mathcal{V}\subset [0,1]$ be its support. Define $g(v)=\mathbb{E}[Y\|f_w(x)=v]=\mathrm{Pr}[Y=1\|v]$. Let $\mathcal{G}$ be a class of candidate conditional probability functions over $\mathcal{V}$ and every candidate $g\in\mathcal{G}$ satisfies $0\leq g(v)\leq 1$ for all $v\in\mathcal{V}$. We need the following assumption to control MI [A]; Assume that $\mathcal{G}$ has at least one member $g^$ that is bounded away from 0 and 1, i.e., there exist constants $0<c_1\leq c_2<1$ such that $c_1\leq g^\leq c_2$. Instead of Assumption 2 in our paper, assume that for any $g\in\mathcal{G}$ and for any $v,v'\in\mathcal{V}$, $\|g(v)-g(v')\|\leq C\|v-v'\|^\beta$ with $C>0$. When $\beta=1$, this becomes Assumption 2. Under the above two assumptions, the following relation holds; $$ \sup_{g\in\mathcal{G}}\mathbb{E}\|\mathrm{TCE}(f_w)- \mathrm{ECE}(f_w,S_{te})\|\geq \inf_{\hat{g}}\sup_{g\in\mathcal{G}}\|\mathbb{E}[\|g(V)-V\|]-\mathbb{E}[\|\hat{g}(V)-V\|]\|\succeq n^{-\beta/(2\beta+1)}, $$ where $\hat{g}$ is over all valid estimators for $g$ using $n$ samples $(X_m,Y_m)_{m=1}^n$ and the expectation is taken with respect to true $g$. Note that the left-hand side of the above corresponds to the total bias in our paper. Note that the term $\mathbb{E}[\|g(V)-V\|]$ is the TCE. If we consider $\hat{g}$ as the kernel density estimation, the above result can be used as the lower bound of the kernel ECE, which is also a popular method to estimate the TCE. From this, when $\beta=1$, the upper bound of our total bias (e.g., Eq.(13)) achieves the same rate as the lower bound thus the binning achieves the optimal rate. However, when $\beta>1$, the upper bounds do not reach the lower bound. As discussed in lines 191-204, this limitation arises because binning cannot exploit the underlying smoothness of the distribution. In [A], the L1 minimax rate of conditional probability is given as $$ \inf_{\hat{g}}\sup_{g\in\mathcal{G}}\mathbb{E}\|\hat{g}(V)-g(V)\|\succeq n^{-\beta/(2\beta+1)}. $$ Combining this with our results, we can see that the rate of the total bias matches the rate of nonparametric regression, conclusively showing that estimating TCE is as challenging as estimating conditional probabilities. We will incorporate these results and proof into our revised paper. [A] Minimax nonparametric classification rates of convergence, Y. Yang. IEEE Transactions on Information Theory, 1999 [B] Information-theoretic determination of minimax rates of convergence, Y. Yang and A. Barron. The Annals of Statistics, 1999. ## Connection between IT-bound and UC theory (fat-shattering dimension) s pointed out by Reviewer 8W41, the bound by metric entropy depends on the model's dimensionality, making them unsuitable for large models such as neural networks. Some existing studies [13,15] present the upper bound of the eCMI and fCMI by the dimension-independent complexities, such as the VC dimension for binary classification and connecting IT theory to UC theory. Inspired by these results, here we provide the upper bound of eCMI and fCMI using such dimension-independent complexities. As provided in the lower bound analysis above, since TCE estimation is similar to the nonparametric regression, we use fat-shattering dimension [C] to upper bound the eCMI. Specifically, if our model class $f_w(\cdot)$ satisfies $\delta/4$-fat dimension with $d_{\delta/4}$ for $\delta\in[0,1]$, we have $$ eCMI(Eq.15)\leq fCMI(Eq.(17))= O\Big(d_{\delta/4}\log \frac{n}{d_{\delta/4}\delta} \log \Big(\frac{n}{\delta^2}\Big)\Big) $$ which results in the dimension-independent upper bound. To evaluate the fat-shattering dimension for specific models, see [D] for the details. We will incorporate these results into our paper. [C] Scale-sensitive dimensions, uniform convergence, and learnability. N. Alon, et.al., Journal of the ACM, 1997 [D] Vapnik-Chervonenkis Dimension of Neural Nets. Peter L. Bartlett ## Discussion about Assumption 2 Here, we discuss the necessity of Assumption 2. Estimating TCE involves nonparametric regression of $E[Y∣f(X)=v]$. For finite samples, smoothness assumptions like Lipschitz continuity are required; without them, small changes in $v$ could cause large variations in label outcomes, making estimation impossible [E]. Such smoothness assumptions are standard in nonparametric regression, including kernel-based ECE. Without these assumptions, increasing the sample size would not ensure that training (or test) ECE converges to TCE. As noted in our minimax lower bound discussion, the absence of smoothness ($\beta\to 0$) leads to increasing bias. [E] Minimax optimal conditional density estimation under total variation smoothness, M. Li, et. al., Electron. J. Statist, 2022. ## Additional numerical validation We add numerical experiments addressing the concerns raised by the reviewers in PDF: - Toy data experiments to evaluate the total bias and optimality of $B$ (Fig. 4). For the detailed explanation, see the reply to Q.4 of Reviewer 8W41. - A logarithmic plot of the ECE gap in Figure 1 to clarify the behavior of the bounds (Fig. 5). For the detailed explanation, see the reply to Q.4 of Reviewer 9r8F. - Experiments validating the Lipschitz continuity in Assumption 2 (Fig. 6). For the detailed explanation, see the reply to Q.1 of Reviewer kktU. Pdf: /pdf/d296849a7ba85a55e7ac3759deddab83127d755f.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Fair Wasserstein Coresets	Accept (poster)	Summary: This paper introduces a new data distillation technique called Fair Wasserstein Coresets. The general idea is to create a synthetic core set along with sample weights to represent a larger dataset, by minimizing the Wasserstein distance between core set and dataset, while ensuring a fairness constraint is satisfied. The paper develops a majority minimization algorithm for this Wasserstein problem and empirically validates it on several data sets demonstrating a competitive fairness utility trade-off. Strengths: - The Wasserstein problem is well-formulated with theoretical guarantees. - The connections with k-medoids is intuitive. Weaknesses: - I suspect there is a potential error in Proposition 2.1, specifically pertaining to the inputs and outputs defined in these functions. Note that z consists of inputs ($d, x$) and outputs ($y$) of the NN, whereas $g_{\psi}$ is an MLP, i.e., it is a function that takes only $(d, x)$ as input. From [69 (original reference)], the MLP satisfies the Wasserstein inequality but only on the marginal distributions over p_{(x,d)} rather than over $p_{Z}$. This may be resolved if we consider not the Wasserstein distance of the MLP output, but instead the Wasserstein distance of the function $h(z) = \| g_{\psi}(x,d) - y \|$. - What do you mean in Lemma 3.1 that the corset is “no better than the best fair Wasserstein corset formed by $m \|D\|\|Y\|$ data points”? I suspect you mean better with regard to achieving a lower Wasserstein distance, but please clarify. - The empirical analysis in Figure 1 is hard to parse. Can you measure the Pareto frontier from all of the observations and demonstrate that FWC is dominant? FWC seems Pareto efficient for Adult, Crime, and Drug, but not Credit potentially — but it is hard to see. - It is hard to understand the trade-offs between accuracy and disparity in the LLM experiments in Table 1, just by reporting these numbers. How important is it that the disparity dropped by 0.009 at a 2.97 point loss in accuracy? Again, it would be important to demonstrate some Pareto efficiency. Furthermore, the change in accuracy and disparity do not seem statistically significant based on the SD reported. Technical Quality: 3 Clarity: 3 Questions for Authors: - Please clarify the potential error and comment about Proposition 2.1. If there is an error, what are the consequences with subsequent theoretical results? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are thoroughly discussed in the Appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the careful review and comments; we answer each questions below. 1. It is indeed true that the MLP $g_\psi$ satisfies the Wasserstein inequality for $p_{(x,d)}$ rather than $p_{z}=p_{(y,x,d)}$ (with the first inequality being ultimately what we are interested in), so thanks for pointing this out. We have changed the notation to define the downstream deviation as $d(p_{(\hat{X},\hat{D});\theta}, p_{(X,D);e})$. Proposition 2.1 still holds as it can be updated as follows: $d(p_{(\hat{X},\hat{D});\theta}, p_{(X,D);e}) \leq L_k W_1 (p_{(\hat{X},\hat{D});\theta}, p_{(X,D);e}) \leq L_k W_1(p_{\hat{Z};\theta}, p_{Z;e}),$ as the Wasserstein-1 distance of the joint $(X,Y,D)$ is larger than the marginal $(X,D)$, as the $L_1$ distance is the sum of the absolute values of the distances across each variables (we have added a more formal proof of this in Appendix B). In other words, the Wasserstein distance minimized by FWC still constrains the downstream learning discrepancy we are interested in. We have updated Proposition 2.1 to reflect this. Finally, note that defining $h(z) = \|g_\psi(x, d) - y\|$ does not work as the reverse triangle inequality $\|\|x\| - \|y\|\| \leq \|x-y\|$ only holds if the expectation is inside the absolute value, so the upper bound necessary for the proof of Proposition 2.1 would not follow. 2. Thanks for the suggestion and you are totally correct. It is in the sense of achieving a lower Wasserstein distance. We have clarified it by stating that the optimal objective value (Wasserstein distance) of problem (4) is no lower than that of problem (5) with the $m$ replaced by $m\|\mathcal{D}\|\|\mathcal{Y}\|$. 3. Thank you for your suggestion to improve the readability of our results. We have now updated Figure 1 to include Pareto frontiers, computed over all models and coreset sizes, which we have uploaded in the global rebuttal above. We include a table below to indicate which FWC models are included in the Pareto frontier in each dataset. \| \| Pareto Frontier \| \| \| \|:----------:\|:------------------------:\|:------------------------:\|:-----------------------:\| \| Dataset \| FWC ($\epsilon=0.01$) \| FWC ($\epsilon=0.05$) \| FWC ($\epsilon=0.1$) \| \| Adult \| $\checkmark$ \| $\checkmark$ \| \| \| Drug \| $\checkmark$ \| $\checkmark$ \| $\checkmark$ \| \| Crime \| $\checkmark$ \| $\checkmark$ \| $\checkmark$ \| \| Credit \| $\checkmark$ \| \| \| 4. Thank you for the feedback. Determining the acceptable fairness-accuracy trade-off has been a discussion in the community and is eventually dependent on the application. Fairness-accuracy trade-offs have been highlighted in several works, eg., [1] and [2], and the acceptable trade-off is up to an end user and is application dependent (eg., finding lower discriminatory alternatives is mandatory in fair lending and acceptable accuracy reduction is dependent on "business necessity" [3]). Using LLM's is one of the downstream tasks considered in the paper and the experiment shows that coresets produced by FWC are able to reduce disparity in the outputs of LLM's for a classification task on average. When we observe the mean and standard deviations for demographic parity, we see that demographic parity is consistently reduced for outputs when compared to zero shot prompting for GPT-4 across all runs, showing that the coresets can help reduce disparity. For GPT-3.5, while the average disparity is reduced across runs, such consistency is indeed not observed, owing to a diverse set of outputs from the large language model. We will add this discussion to the final version of the draft. References: - [1] Faisal Kamiran and Toon Calders. Data preprocessing techniques for classification without discrimination. Knowl. Inf. Syst., 33(1):1–33, October 2012. - [2] Michael Feldman, Sorelle A. Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. Certifying and removing disparate impact. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’15, pages 259–268, New York, NY, USA, 2015. ACM. - [3] Gillis, Talia B., Vitaly Meursault, and Berk Ustun. "Operationalizing the Search for Less Discriminatory Alternatives in Fair Lending." In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 377-387. 2024. --- Rebuttal Comment 1.1: Title: Score adjusted Comment: Thanks for clarifying the theoretical result and updating your figures to show the Pareto frontier. The discussion in (#4) is useful and I recommend you include this discussion in your final version. I have updated my score accordingly.	Summary: The paper gives an algorithm to generate smaller weighted synthetic dataset from real data set such that the synthetic data can enforce demographic parity when used for downstream tasks. This is achieved by solving an optimization problem of minimizing the Wasserstein distance between the two dataset distributions along with demographic parity-based fairness constraint. The authors describe how to efficiently solve this problem by reformulating it and subsequently using a majority minimization algorithm to solve the resultant nonconvex problem. They provide convergence guarantees for the algorithm and also generalization bounds for the solution. The theoretical results are supported by experiments on real and synthetic datasets. Strengths: 1) The paper for the most part is written clearly with some minor writing issues (see weaknesses). It is well structured and not too difficult to follow the high-level ideas. Both fairness and scalability are relevant issues so the paper will be of interest to the community. 2) I could not check all proofs, but the theoretical results appear sound. The connection between the unconstrained problem and Lloyd's algorithm for $k$-means is neat. 3) The authors have performed experiments on both real and synthetic datasets and compared with a number of existing methods. As such the paper is a good mix of theory and practice. Weaknesses: 1) The paper seems to borrow a lot of ideas and proof techniques from existing works like [56], [71] and others. for e.g. the reformulation, ideas to speed up the algorithm etc. As such I am not entirely sure about the novelty quotient of the work. It would be better if the authors can highlight why the modifications to techniques from existing works are non-trivial. 2) The explainability of the synthetic data will be very less. Specifically, as far as I understood, the authors are assigning the output label and sensitive attribute value to the generated data points just in same proportion as that in the original data. It is not clear to me what does this mean for the individual synthetic data points. Also do the features in the generated synthetic data correspond exactly to the features in original data? 3) I suggest the paper be proofread for minor corrections in writing: E.g.: In the contributions make the 'w' 's capitalized. On line 171 the authors say $P \geq 0$ (which I think means each entry in non-negative) while on line 258 it is $P \geq \mathbf{0}$. Maintain the consistency. 4) Should not the weight of the synthetic data sum to $n$ and not $m$ (line 141 $\Delta_m$)? Typically, we try to preserve the weight of the original data in expectation while reweighing the sampled points. Please Clarify. Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: See Weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful to the reviewer for their time and comments in the review of our work; we answer each point below. 1. While we acknowledge that the ideas for the reformulation and part of the complexities considerations are adapted from [1, 2] ([56, 71] in the original references) we would like to emphasize the significant contributions of this work: - __Novel algorithm.__ Our novel algorithm FWC represents the first fair coreset approach explicitly focusing on downstream learning, in contrast to existing methods that emphasize local properties of generated samples (Section 4); - __Theoretical claims.__ We support our novel algorithm FWC with theoretical claims regarding its convergence properties (Section 5.2) and generalization analysis (Section 5.3), as well as the effect on downstream learning tasks (Proposition 2.1 and Appendix B.1); - __Connection to k-means.__ We establish a practical connection to the k-means algorithm, thereby extending FWC's applicability beyond fairness considerations (Section 6); - __Empirical results.__ We demonstrate FWC's effectiveness through empirical results on synthetic data for scalability, as well as on real datasets and two LLMs for downstream classification tasks (Section 7 and Appendix C). 2. The features for the generated data are the same as the ones in the original training data, but the values of such features would be different. For instance, in Appendix C.2, in ``Using FWC to correct biases in LLM'' we include a textual example of how one of our generated coresets would look like. The value of each feature would depend on the chosen cost function $c$; in Section 4.2 we highlight the case of the $L_1$ and $L_2$ distance, as well as how to select feature values if only values existing among the training data features can be selected (akin to k-medoids). 3. We have corrected the inconsistencies, thanks for pointing them out. 4. In our work, the synthetic coresets $\hat{Z}$ lie in $\mathcal{Z}^m$. Only when $\theta\subseteq \mathbb{R}^m_+$ and $\sum_j \theta_j = m$, the empirical distribution $p_{\hat{Z};\theta}:= \frac{1}{m}\sum_{j=1}^m \theta_j \delta_{\hat{Z}_j}$ is a valid probability measure. References: - [1] G. Peyré, M. Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning, 11(5-6):355–607, 2019. - [2]Z. Xiong, N. Dalmasso, A. Mishler, V. K. Potluru, T. Balch, and M. Veloso. FairWASP: Fast and optimal fair wasserstein pre-processing. Proceedings of the AAAI Conference on Artificial Intelligence, 38(14):16120–16128, Mar. 2024. --- Rebuttal Comment 1.1: Comment: I have read the reviews and the rebuttal and have accordingly updated my score.	Summary: This paper proposed to extract coresets from a set of data samples using Wasserstein distance with fairness constraints. The authors formulates this problem as a minimization with linear constraints. The coreset selection is over the whole input space, not just from original data samples. The importance / weight of each coreset sample are also optimized. Extensive experiments show this method achieves better fairness-utility tradeoff, and can be applied in LLMs to reduce bias. Strengths: The paper is nicely written and easy to follow. I appreciate the detailed steps and neat reformulations of the optimization problem. Theoretical guarantees are provided. The experiments supports the effectiveness of the proposed method very well. Weaknesses: 1. In section 4.2 (line 203), how can problem Eq.(12) be separated into subproblems as in Eq.(13), are the optimal solutions of all subproblems the same and equal to the solution to (12)? 2. In section 6 (line 259), why are the minimizers of problem (17) always has only one non-zero entry in each row? As problem (17) can be seen as a relaxed version of discrete Kantorovich problem, where we can't say anything about the sparsity of the optimal plan. Please elaborate. Technical Quality: 4 Clarity: 4 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors included detailed limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful to the reviewer for the effort spent reading our paper and for the encouraging comments. Below are our responses to the questions raised: 1. First of all, we apologize for a typo in (13). The minimum over $\hat{X}_i \in \mathcal{X}$ in (13) should be corrected to the minimum over $\hat{X}_j \in \mathcal{X}$, replacing the subscript $i$ with $j$. We appreciate the reviewer’s attention to detail in identifying this typo. With this typo corrected, then we can observe from (12) that, for each $j\in[m]$, the variable $\hat{X}_j$ only affects the corresponding $\hat{Z}_j$, because $\hat{Z}_j = (\hat{d}_j,\hat{X}_j,\hat{y}_j)$. Different $\hat{Z}_j$ terms are independent of each other in the objective function of (12), so (12) can be decomposed into subproblems that optimize each $\hat{Z}_j$ (which is essentially $\hat{X}_j$). (12) and (13) are equivalent in the sense that they share the same optimal solutions. 2. The discrete Monge--Kantorovich (optimal transport) problem has a similar structure as (17), but is more complex. The constraint of (17) only requires that the sum of each $\underline{column}$ of $P$ equals $\frac{1}{n}$, so (17) can be easily solved by computing the smallest component of each row of $C$. An optimal solution $P$ can be a matrix where all entries of each row are zero except for the one corresponding to the smallest component of that row in $C$. In contrast, a typical optimal transport problem that is written as an LP also requires the sum of each $\underline{row}$ to equal to some given values. For this reason, optimal transport problems do not have easily computable closed-form optimal solutions. In the discrete Monge--Kantorovich problem, this pair of constraints on the sum of rows and columns essentially state that the marginals of the probability measure (the decision variable) on the product space must equal the two given probability measures. --- Rebuttal Comment 1.1: Comment: Thanks for the clarification. I have no further questions.	Summary: This paper talks about "fair Wasserstein coresets", weighted representative points generated to represent the original datasets. The goal is to meet two purposes: 1) the Wasserstein distance of the coreset and the input data set is minimized, 2) fairness in terms of demographic parity. Having a small Wasserstein distance can help to bound the downstream discrepancy for ReLu activated perceptrons. The authors formulate the problem as an optimization problem (4). There are four steps: 1. Manually set the proportion of each combination of decision (Y) and feature (D). 2. Formulate linear constraints for the fairness constraint. This one borrows directly from [71]. 3. Formulate the Wasserstein distance optimization by using [56] 4. Simply further After that, the problem is not convex. The authors use "majority minimization" [52, 38] to solve it. Specifically, one defines a convex surrogate function that upper bounds the non-convex function, and optimizes the convex function. Section 5 reports theoretical guarantees: running time of the algorithm, convergence guarantee for the surrogate function, and last bound the generalization guarantees. Experiments are reported in the last section: e.g., improving fairness in LLM. On the positive side, the problem formulation is interesting and valid, Wasserstein coreset with fairness consideration. The use of this coreset for downstream applications make sense. Thus the problem and solution have merit. Experiment are thorough. The weaknesses (or limitation in significance) is that both crucial steps (2) and (3) are basically using prior work. The theoretical results are standard. Summarizing I feel that the paper is OK but would give a weak accept. Strengths: On the positive side, the problem formulation is interesting and valid, Wasserstein coreset with fairness consideration. The use of this coreset for downstream applications make sense. Thus the problem and solution have merit. Experiment are thorough. Weaknesses: The weaknesses (or limitation in significance) is that both crucial steps (2) and (3) are basically using prior work. The theoretical results are standard. Technical Quality: 3 Clarity: 3 Questions for Authors: I understand that there are many different notions of fairness and the authors focus on one of them, demographic parity. This is OK. One suggestion is that if the authors can provide some discussions and insights on how the results may improve other notions of fairness it will be valuable to have. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are thankful to the reviewer for their time and thoughtful feedback. As the reviewer has pointed out, several notions of fairness exist in the literature. Focusing on the classification setting, chapter 3 in [2] classifies these notions into independence, separation, and sufficiency. Demographic parity falls in the class of the independence notion, and hence, other measures of fairness that are closely related, eg., disparate impact, would also improve when optimizing for demographic parity. However, as highlighted in [2], other notions of fairness, such as separation may not simultaneously be satisfied. Our work is focused on demographic parity and cannot guarantee an improvement in these other measures. To test this, we evaluated our work, which optimizes for demographic parity, on the equalized odds measure, which falls under the notion of separation. When we consider equalized odds, FWC is not a part of the Pareto frontier for the Drug dataset, and more generally, FWC performance is not as competitive. This is in contrast to demographic parity: in Figure 1, FWC sits on the Pareto frontier across all datasets for fairness-performance trade-off in downstream classification (see the addendum in the global rebuttal above for the updated version of Figure 1). We have included a table below to indicate which FWC models are part of the Pareto frontier when considering demographic parity versus equalized odds, and will add this discussion to the paper. \| \| Pareto Frontier, \| Demographic Parity \| \| Pareto Frontier, \| Equalized Odds \| \| \|:----------:\|:-------------------------------------:\|:------------------------:\|:-----------------------:\|:---------------------------------:\|:------------------------:\|:-----------------------:\| \| Dataset \| FWC ($\epsilon=0.01$) \| FWC ($\epsilon=0.05$) \| FWC ($\epsilon=0.1$) \| FWC ($\epsilon=0.01$) \| FWC ($\epsilon=0.05$) \| FWC ($\epsilon=0.1$) \| \| Adult \| $\checkmark$ \| $\checkmark$ \| \| $\checkmark$ \| \| \| \| Drug \| $\checkmark$ \| $\checkmark$ \| $\checkmark$ \| \| \| \| \| Crime \| $\checkmark$ \| $\checkmark$ \| $\checkmark$ \| \| \| $\checkmark$ \| \| Credit \| $\checkmark$ \| \| \| $\checkmark$ \| \| $\checkmark$ \| Finally, while we acknowledge that steps (2) and (3) are built from [1], we would like to emphasize the significant contributions of this work and especially note that: (i) our novel algorithm FWC represents the first fair coreset approach explicitly focusing on downstream learning, in contrast to existing methods that emphasize local properties of generated samples; (ii) we support FWC with theoretical claims in Propositions 2.1, Theorems 5.3, 5.4, Proposition 5.5 (as well as additional insights in Appendix B.1); (iii) we establish a practical connection to the k-means algorithm, thereby extending FWC's applicability beyond fairness considerations (Section 6); and (iv) we demonstrate FWC's effectiveness through empirical results on synthetic data for scalability, as well as on real datasets and two LLMs for downstream classification tasks (Section 7 and Appendix C). References: - [1] Z. Xiong, N. Dalmasso, A. Mishler, V. K. Potluru, T. Balch, and M. Veloso. FairWASP: Fast and optimal fair wasserstein pre-processing. Proceedings of the AAAI Conference on Artificial Intelligence, 38(14):16120–16128, Mar. 2024. - [2] Barocas, Solon, Moritz Hardt, and Arvind Narayanan. Fairness and machine learning: Limitations and opportunities. MIT press, 2023. --- Rebuttal Comment 1.1: Comment: Thanks for the response. I believe including them in the revision can be helpful.	Rebuttal 1: Rebuttal: We would like to thank again all reviewers for their comments, detailed feedback and questions which have improved the quality of our paper. While we have addressed each reviewer individually, we are using the global rebuttal to upload a .pdf with the new version of Figure 1 which includes Pareto frontiers; this addresses the comment made by reviewer 4tbZ and is relevant for the discussion of demographic parity following the question by reviewer xAdA. Pdf: /pdf/07420626b6fcd77af12f3aea42f113b266ae0ffe.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Towards Flexible 3D Perception: Object-Centric Occupancy Completion Augments 3D Object Detection	Accept (poster)	Summary: This paper presents a new task named object-centric occupancy completion as a fine-grained object representation to supplement the coarse-grained 3D bounding boxes. To accomplish this task, a new dataset, which annotates instance-level high-resolution occupancy, is created in an automated pipeline. This paper also introduces an implicit shape decoder to fuse multi-frame information, predict instance occupancy and refine 3D bounding boxes. Experiments on Waymo datasets above several baselines demonstrate the effectiveness of the proposed method on both occupancy prediction and 3D detection. Strengths: 1. This paper is well-written and organized. 2. A novel task, occupancy augments 3D object detection, and a corresponding new instance-level occupancy datasets is proposed. 3. A implicit shape decoder is proposed and achieves great improvements both in occupancy and 3D detection. Weaknesses: 1. The motivation of this paper does not seem to be very reasonable. The authors claim that a. high-resolution scene-level occupancy is constrained by computational cost and foreground objects is more import, and b. 3D detection is too coarse to capture the object geometry information. So why not just predict foreground instance-level occupancy in the whole scene, instead of pursuing higher detection accuracy by using the occupancy results? 2. Time and memory cost bought by the proposed shape decoder are not provided. The paper is trying to make a trade-off between occupancy and detection in fine-/coarse-grained level and computational cost level. But the authors only report the occupancy and detection accuracy. 3. Some methods, like VoxelNeXt, FSDv2, HEDNet are missing and are not compared in Table 1. 4. Typos/mis-leading descriptions. For example, ‘Tab. 5.4’ on line 351 -> ‘Tab. 3’. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Why not just predict foreground instance-level occupancy in the whole scene, instead of pursuing higher detection accuracy by using the occupancy results? (the same as weakness 1) 2. Could you provide the computational cost of your method or the proposed module? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Why not just predict foreground instance-level occupancy in the whole scene, instead of pursuing higher detection accuracy by using the occupancy results? 1. To predict foreground instance-level occupancy for the entire scene, it is essential to distinguish the foreground from the background. However, separating the foreground from the background without detection results is a non-trivial task. Therefore, we leverage detection to obtain foreground objects. 2. Pursuing higher detection accuracy is not our main objective. Our main goal is to obtain object-centric occupancy to provide a more flexible representation for downstream driving tasks. However, when we attempted to aggregate multiple frames of occupancy grids to make the shape more complete, we realized that these occupancy grids could also serve as an excellent medium for multiple frame information fusion. Thus, we designed a simple fusion network to examine the benefits for the detection task. 2. Time and memory cost Since object tracklets vary in length, our method's running time may also vary with different inputs. Additionally, the dimension of the decoded object-centric occupancy depends on the detected bounding box. To ensure fair testing of running time, we standardized the input length to 32 and set the number of decode queries to 4096. We conducted the inference with batch_size = 1 using standardized inputs on a single 3090 GPU and computed the average running costs. The results are presented in the table below. We can observe that using the shape decoder does not significantly affect the computational cost. \| model \| Avg. Inference Time \| Avg GPU memory Cost \| \| --- \| --- \| --- \| \| w/o shape decoder \| 4.08ms \| 2499MB \| \| w/ shape decoder \| 4.23ms \|2565MB \| 3. Some methods, like VoxelNeXt, FSDv2, HEDNet are missing and are not compared in Table 1. Thanks for the suggestions. We mainly consider tracklets from GT, CenterPoint, and FSD in Table 1 because we used their tracklet proposals on the training set to train our model. Basically, our method can generalize to other detector without retraininig. The following table presents occupancy completion results obtained by directly applying our trained model to the tracklet proposals generated by FSDv2 on the testing data. \| Tracklet Inputs \| IoU % \| mIoU (track) % \| mIoU (box) % \| \| --- \| --- \| --- \| --- \| \| FSD v2 (no train) \| 61.41 \| 47.69 \| 60.76 \| \| FSD \| 62.84 \| 54.12 \| 61.58 \| \| CenterPoint \| 57.99 \| 44.94 \| 55.10 \| Due to better detection, our method with FSDv2 still outperforms the version with CenterPoint even without retraining. However, it performs slightly worse compared to using FSD tracklets, despite FSDv2 having better detection results than FSD. This indicates that significant detection improvements generally lead to better shape completion (FSDv2 vs. CenterPoint). However, for detectors with similar performance (e.g., FSD vs. FSDv2), improved detections do not necessarily guarantee better shape completion without retraining. Retraining our method using training proposals generated by FSDv2 may address this issue. We will add these results to the Table 1 and discuss the findings in our revised manuscript. Besides, we’ve conducted a detection experiment where we applied our method to 1-frame FSDv2 without retraining. The following table demonstrates that our method with a stronger detector continues to show detection improvement even without retraining. \| Model \| Vehicle L1 mAP/mAPH \| Vehicle L2 mAP/mAPH \| \| --- \| --- \| --- \| \| FSDv2 \| 79.8/79.3 \| 71.4/71.0 \| \| FSDv2 +Ours \| 83.2/82.7 \| 75.2/74.7 \| We will include these results in our revised Table 2 We will also include VoxelNeXt, FSDv2, HEDNet for a more comprehensive comparison. In fact, our method based on single-frame FSD outperforms all these mentioned methods by noticeable margins. 4. Typos/mis-leading descriptions. For example, ‘Tab. 5.4’ on line 351 -> ‘Tab. 3’. Thank you for the thorough review. We will correct the typos and conduct meticulous proofreading for our revised submission. --- Rebuttal Comment 1.1: Title: Post rebuttal Comment: Thanks for the response. My concerns regarding the motivation and computation cost have been well-addressed, therefore I raise my rating to weak accept.	Summary: In this work, the authors propose a novel task called object-centric occupancy. It extends the 3D detected bounding box representation to provide a more detailed description of the internal object shape. The method provides higher voxel resolution in large scenes by focusing on foreground objects only. It not only achieves state-of-the-art performance on shape completion but can also help refine the object detection tasks on the Waymo Open Dataset (WOD). Strengths: - The motion of the proposed task is clear, and the task itself shows good potential in scene understanding. It can enhance 3D detection results even at a far distance. - The extensive ablation studies validate each contribution. Various detector results with different settings help prove the robustness of the proposed methods. Using implicit representation from a 3D reconstruction task to complete shapes is neat and interesting. It will be interesting to see how this work can be applied to popular 3D Gaussian representation. Weaknesses: - The experimental results are only obtained on the Waymo Open Dataset. It will be nicer to conduct the experiments on nuScenes or Argoverse 2 to validate its robustness for different datasets. - Although the authors say it is a new task, so there are no learning baselines for shape completion, it will be interesting to compare the results with other scene occupancy methods. So that we can see the flaws of using coarse resolution quantitatively. Technical Quality: 4 Clarity: 4 Questions for Authors: - The extrapolated results of shape completion are interesting, showing that it can achieve a performance similar to that of using GT boxes. Will it also help with 3D Object Detection results? Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. The experimental results are only obtained on the Waymo Open Dataset. It will be nicer to conduct the experiments on nuScenes or Argoverse 2 to validate its robustness for different datasets. Thanks for the suggestions. Currently, we are not able to train/test our method on nuScenes or Argoverse 2 since we only prepared object-centric occupancy labels on Waymo. We will support nuScenes or Argoverse2 after our occupancy labels on Waymo is released. 2. Although the authors say it is a new task, so there are no learning baselines for shape completion, it will be interesting to compare the results with other scene occupancy methods. So that we can see the flaws of using coarse resolution quantitatively. Thank you for the suggestion. To the best of our knowledge, no existing scene-level occupancy method has published results on Waymo Open dataset. We will release the results when their results on Waymo are published. 3. The extrapolated results of shape completion are interesting, showing that it can achieve a performance similar to that of using GT boxes. Will it also help with 3D Object Detection results? Our detection already benefits from the extrapolated shape information, as the shape embedding used for occupancy decoding is also fused into the detection feature. Compared to explicitly using the generated occupancy, incorporating the feature directly ensures minimal information loss.	Summary: The manuscript introduces the idea of representing the shape of objects at higher fidelity (and independent of) the rest of the scene. This is explored in the context of autonomous vehicles research on 3d car detection and representation. The proposed model regresses a shape code and an updated 3d bounding box from a 3D bounding box tracklet (derived from any other algorithm) and the points included in it. The shape code can be queried for occupancy to produce a full shape representation during inference. The proposed approach is able to infer complete shapes from partial inputs and the updated 3D BBs improve the input 3D BBs. Strengths: The proposed approach is relatively straightforward, effective and well motivated. This makes it reusable for other works and the paper more reproducable. The manuscript is well written and the illustrations help convey the message and improve understanding of the written parts. The evaluation is comprehensive and the sensitivity studies are well chosen and help motivate architecture and training choices. In particular it is great to see that the addition of the high resolution shape code and updated 3D BB does lead to substantial performance improvements on the 3D BB detection task (especially for far away OBBs). And that the shape code (if given the GT OBB) does produce a high IoU occupancy grid even if the input 3D BBs are subpar (table 1). Weaknesses: The manuscript's related work section misses out on an existing related field of 3D CAD model retrieval (which also produces complete shapes) and shape regression from RGB (and depth data) in indoor scenes. Relevant related works include: - Scan2CAD https://openaccess.thecvf.com/content_CVPR_2019/papers/Avetisyan_Scan2CAD_Learning_CAD_Model_Alignment_in_RGB-D_Scans_CVPR_2019_paper.pdf - SLAM++ https://www.doc.ic.ac.uk/~ajd/Publications/salas-moreno_etal_cvpr2013.pdf - FroDO https://openaccess.thecvf.com/content_CVPR_2020/papers/Runz_FroDO_From_Detections_to_3D_Objects_CVPR_2020_paper.pdf I would have wanted to see a few renderings of the shape codes; This would support the claim that the model learns to complete shapes. The appendix has a few but the visualizations are hard to understand without a better renderer. Some kind of shading or edges for the 3D voxels are essential to see any kind of depth and thus shape (Fig 6 and 7). Extracting a mesh using marching cubes at the 0.5 isolevel might also work. Technical Quality: 3 Clarity: 3 Questions for Authors: The model takes in a series of 3D BBs and outputs one updated 3D BB - at which timestamp is this 3D BB output? The latest? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitation of rigid objects only is addressed in the appendix. This means humans are not supported for example. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Missing related works Thank you for the suggestion. We will include a discussion of these works in our revised version and provide a more thorough related works section. 2. Renderings of the shape codes; In the uploaded PDF, we’ve included several renderings. These renderings were obtained by applying marching cubes to the decoded volumetric grids using a level of 0.5, as suggested. The renderings demonstrate that our method can complete shapes even when the current point cloud is extremely sparse. Due to the use of 0.2m voxel size, the resolution of our predicted occupancy may not support high-quality rendering. For example, the resolution for a typical sedan (let's assume its dimensions are 4.5m* 1.8m * 1.4m) under our voxel size is 23 * 9 * 7. In contrast, common shape completion methods typically use a resolution of 128 x 128 x 128 or higher to facilitate high-quality rendering. It should be noted that for our purposes, high-quality rendering is not required. Although the selected voxel size of 0.2 meters may not provide highly detailed rendering, it is sufficient for downstream driving tasks and ensures computational affordability. 3. The model takes in a series of 3D BBs and outputs one updated 3D BB - at which timestamp is this 3D BB output? The latest? The behavior depends on the context. During inference, we only need to output the latest one. However, during training, we can simultaneously obtain bounding box outputs at all timestamps with just one forward pass using a causal attention mask. We compute the detection loss over all these bounding boxes to facilitate training. We will clarify this in our revised version. --- Rebuttal Comment 1.1: Comment: I appreciate the authors responses. The attached shape renderings are from the same viewpoint as the input pointcloud as far as I can tell. This shows "weak completion" - interpolating between points and maybe some extrapolation. The real question is what the back side of those shapes look like "strong completion". I looked at the other reviews as well and did not see anything that would change my rating so far. --- Reply to Comment 1.1.1: Comment: Thank you for the quick response. As shown in Figure 10 of our appendix, our method demonstrates a strong ability for shape completion. Even with extremely sparse input points at an early timestamp (the first row in Figure 10), our method effectively completes the shape in a way that aligns with the sparse point observations. This level of completion cannot be achieved through simple interpolation or extrapolation. Our model’s success in this regard is due to its ability to learn the complete shape distribution from the training data. However, as with any learning-based approach, our method's performance can be constrained by the quality of the annotations. Since our ground truth shapes are generated by aggregating points across real object sequences, the back side of an object is often 'unobserved' (see Figure 7), meaning that most of the back-side voxels are not supervised during training. Fortunately, there is a simple yet effective way to mitigate this issue. Given the symmetry of many objects, we can fill the 'unobserved' voxels with their mirrored counterparts. Retraining our model using these mirror-aided ground truths can significantly enhance its ability to complete shapes on the back side. However, we did not employ this strategy in our main paper, as it might compromise the authenticity of the shape annotations.	Summary: This paper addresses the limitations of 3D object bounding box representations in autonomous driving by introducing object-centric occupancy. It uses an implicit shape decoder to manage dynamic-size occupancy generation. The method demonstrates robust performance under noisy conditions, significantly enhancing detection results in the Waymo Open Dataset. Strengths: 1. The presentation is well-executed, with figures and charts effectively aiding reader comprehension. 2. The overall performance is impressive, demonstrating significant improvements across multiple baselines. Weaknesses: 1. Creating detailed occupancy for each object seems unnecessary. In most downstream tasks in autonomous driving, using bounding boxes (bboxes) is sufficient. 2. The performance improvement primarily stems from temporal feature fusion, which lacks significant technical innovation. 3. It is unclear whether the loss on occ heads in Fig. 4 enhances detection performance. The authors should compare detection performance with and without occ heads after obtaining the Shape Emb. Z to determine if occ heads contribute to learning useful features, such as yaw estimation. Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors discuss the limitation concerning non-rigid objects, which is indeed a constraint. They could start by exploring whether reconstructing the noisy occupancy of non-rigid objects improves detection performance. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Creating detailed occupancy for each object seems unnecessary. In most downstream tasks in autonomous driving, using bounding boxes (bboxes) is sufficient. We respectively disagree with this statement. As highlight in our introduction, using bboxes alone “fails to capture the intricate details of objects’ shape, particularly for objects with irregular geometries”. As illustrated in Fig. 1, the bbox of the crane inevitably includes unoccupied space, which is however could be the drivable space for the ego-car. Relying solely on bboxes limits the downstream planner’s ability to leverage this free space effectively. Our uploaded PDF also includes samples that showcase the advantages of object-centric occupancy over bounding boxes in representing complex shape structures. 2. The performance improvement primarily stems from temporal feature fusion, which lacks significant technical innovation. Indeed, our contribution does not focus on improving detection performance. As indicated in our title, we aim to take a step 'toward more flexible 3D perception,' achieved by learning an object-centric occupancy representation. This object-centric nature allows us to aggregate temporal information from a very long sequence (e.g., 200 frames) while maintaining an affordable computational cost. Compared to previous long sequence methods, our method can additionally output object occupancy within each bbox to support precise downstream driving tasks. Additionally, with a simple temporal fusion strategy, our method surpasses previous state-of-the-art approaches in online detection, demonstrating the effectiveness of our strategy. 3. It is unclear whether the loss on occ heads in Fig. 4 enhances detection performance. The authors should compare detection performance with and without occ heads after obtaining the Shape Emb. Z to determine if occ heads contribute to learning useful features, such as yaw estimation. Thank you for the suggestion. We removed the OCC head from our full model and trained the model using only the detection loss. The results are presented in the table below. A noticeable drop in detection performance is observed when the OCC decoder is removed. \| Model \| L1 mAP/mAPH \| L2 mAP/mAPH \| \| --- \| --- \| --- \| \| No Occ Dec \| 81.1 /80.4 \| 73.0/ 72.3 \| \| Ours \| 82.8/82.3 \| 74.8/74.4 \| --- Rebuttal 2: Comment: I have read all the reviews and authors' responses. My concern has been well addressed. so I raise my rating to 5.	Rebuttal 1: Rebuttal: We would like to express our sincere gratitude to the reviewers for their thorough and thoughtful review of our paper. We are encouraged to learn that all reviewers found our paper well-written and recognized its impressive performance. We also extend our thanks to reviewers fcuc, GRiU, and nYQd for appreciating the novelty of our contributions. Below, we further clarify our motivation and contributions in this work. Additionally, we provide individual responses to address the specific questions and comments raised by each reviewer. The uploaded PDF includes several shape code renderings requested by Reviewer fcuc, demonstrating our model's shape completion capability and stronger representation for complex shapes. Motivation and Contribution Our motivation is to support more flexible downstream driving tasks. And object-centric occupancy is the way we achieve this goal. Compared to traditional Bbox representation, object-centric occupancy provides a more precise representation of the obstacles and drivable space. To support the development of object-centric occupancy, we 1) annotated an object-centric occupancy dataset. 2) presented a robust sequence-based network for effective occupancy completion via implicit decoding. 3) showed our method also improved detection performance. The generated occupancy, along with the improved detection results, provides a more accurate space representation, which supports enhanced planning and control. Pdf: /pdf/1a5ff0b1ae06fb237c29779c417628b627756ff6.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition	Accept (poster)	Summary: This paper proposes CemiFace, a novel diffusion-based approach for generating synthetic face images with varying levels of similarity to their identity centers. The authors argue that semi-hard negative samples, those with moderate similarity to the center, are crucial for training effective face recognition models. The core of CemiFace lies in its ability to control the similarity between generated images and the input (identity center) during the diffusion process. This is achieved by injecting a similarity controlling factor condition (m) that regulates the similarity level. The paper presents a comprehensive analysis of the relationship between sample similarity and face recognition performance, showing that semi-hard samples, generated with m close to 0, achieve the best accuracy. CemiFace demonstrates significant improvements over previous methods in terms of accuracy, particularly on pose-sensitive datasets. The paper further validates its effectiveness through qualitative visualizations and ablation studies that examine the impact of various factors, including training data, inquiry data, and the similarity controlling factor. Overall, this paper contributes a valuable approach to generating synthetic face datasets for face recognition with enhanced discriminative power. The method shows promise in mitigating privacy concerns associated with collecting and using real-world face data while maintaining robust recognition performance. Strengths: Discovery of the importance of similarity control in synthetic face generation: CemiFace is motivated by the discovery that face images with certain degree of similarities to their identity centers show great effectiveness in the performance of trained FR models. This is an important discovery to the community of synthetic dataset generation. Unique use of similarity control: CemiFace introduces a similarity controlling factor (m) within the diffusion process, enabling the generation of faces with varying levels of similarity to the input image. This provides a fine-grained control over the generated data distribution, which is a unique feature compared to existing methods. Comprehensive analysis of similarity: The authors present a thorough analysis of the impact of different similarity levels on face recognition performance, validating their hypothesis about the importance of semi-hard samples. This analysis provides valuable insights into the relationship between data distribution and model effectiveness. Rigorous experimental evaluation: The paper conducts comprehensive experiments across various benchmark datasets and data volumes, comparing CemiFace with other state-of-the-art synthetic face generation methods. The ablation studies provide a detailed understanding of the influence of different parameters and factors on the model's performance. Robustness of CemiFace: The experiments demonstrate the robustness of CemiFace to different training data, inquiry data, and similarity controlling factors. The method consistently achieves superior results, demonstrating its effectiveness and generalizability. Weaknesses: - An in-depth discussion on why face images with certain similarity is more beneficial as a training dataset for the face recognition model would strengthen the paper. For example, an analysis such as a similarity comparison the the real dataset and checking if the difficulty of the CemiFace synthetic dataset becomes closer to that of the real dataset would be nice. Other analysis that offers insights as to why certain similarity control is important would also be welcome. Technical Quality: 4 Clarity: 4 Questions for Authors: - Written in weakness section. Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: - This is a well written paper with a meaningful discovery on training with synthetic dataset. The paper would be better positioned in the venue NeurIPS if it would offer more insightful analysis on why similarity control is beneficial, on top of the empirical benefits. Flag For Ethics Review: ['No ethics review needed.'] Rating: 9 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: Thank you for your insightful suggestion and positive feedback. We assume the benefits of the semi-hard training face images could be attributed to: (1) easy training samples are typically images where the face is clear, well-lit, and faces the camera directly, and thus training on such easy samples would not allow the trained FR models to be able to generalize for face images with large pose/age/expression variations and different lighting conditions/backgrounds that are frequently happened in real-world applications. AdaFace[3] also mentioned that easy samples could be beneficial to early-stage training, while hard sample mining is needed for achieving generalized and effective FR models; (2) hard samples normally contain noise data. Specifically, FaceNet[28] demonstrate that the hardest sample mining using a large batch size leads to hard convergence and produces inferior performance. This is because training with very hard samples may not allow FR models to learn effective features but focus on cues apart from facial identities; (3) Semi-hard samples generated by CemiFace mostly contain large posed faces but less face-unrelated noises. We also evaluate the training epochs needed to reach the highest AVG performance for easy samples ($m=0.7$), semi-hard samples($m=0$) and extreme hard samples ($m=-0.5$). Easy samples take 10 epochs to reach the best AVG and 20 epochs to produce the training loss of 0; Semi-hard samples take much longer (38 epochs) to provide the highest AVG while the final training loss is around 3; and FR models training on extreme hard samples could not converge. Due to the real inquiry center is not available, we further calculated the average similarity (using ArcFace to fairly compute them, i.e., avoiding information from adaface and cosface) between randomly paired images, based on random selected 200 identities from CASIA-WebFace, DCFace and CemiFace, where real face images in CASIA-WebFace provide average similarity of 0.51, while face images generated by our CemiFace has a similarity of 0.48. However, face images generated by DCFace are more different from the real face images. with a much easier average similarity of 0.57. We have also additionally added an experiment to demonstrate the actual similarity to the inquiry data in the PDF Tab. 2. It shows that our generated face images exhibit larger distances (lower similarity) to the inquiry centers than the DCFace. We have also calculated the number of face images belonging to different similarity groups for CemiFace and DCFace in the PDF Tab. 3, indicating that our CemiFace tends to generate images showing lower similarities to their identity centers (i.e. all samples are semi-hard), while DCFace containing more easy samples. --- Rebuttal Comment 1.1: Comment: Thank you for the rebuttal. My questions have been answered. Also, after reading other reviews and rebuttals, it seems that the authors have sufficiently answered the questions. I hold my position as it is, as I believe it is a paper with a meaningful discovery on training with synthetic dataset. --- Reply to Comment 1.1.1: Comment: Dear Reviewer Crof, Thank you for your appreciation of our work and your efforts in reviewing this manuscript. We will try to include the insightful discussion you suggested in the revised version. Best regards, The Authors of Paper 11025	Summary: The paper introduces an approach called CemiFace for generating synthetic face images to enhance face recognition (FR) models. The paper provides the first in-depth analysis of how FR model performance is influenced by samples with varying levels of similarity to the identity center, focusing particularly on center-based semi-hard samples. The authors propose a unique diffusion-based model that can generate face images with different levels of similarity to the identity center. This model can produce infinite center-based semi-hard face images for synthetic face recognition (SFR). The method can be extended to leverage large amounts of unlabeled data for training, providing an advantage over previous methods. Experimental results demonstrate that CemiFace significantly outperforms existing SFR methods, reducing the GAP-to-Real error by half and showcasing promising performance in synthetic face recognition. Strengths: - Focusing on center-based semi-hard samples to enhance face recognition performance is a fresh problem formulation that addresses a notable gap in current methodologies. - The paper provides a solid experimental validation of its proposed approach. The authors investigate factors affecting performance degradation in synthetic face recognition and offer a hypothesis about the importance of mid-level similarity samples. Weaknesses: - The method for determining GtR remains unclear. Justification regarding how the proposed model yields a low GtR is absent. Is this low GtR attributed to the utilization of real inquiry images? If so, what measures guarantee that the synthetic facial images remain uncorrelated with the real facial images? In other words, I have a reservation that the method may not generate "true" synthetic data but highly relies on an inquiry image. Therefore, it is reasonable to see why a low GtR is obtained. - Figure 5 demonstrates that different identities (such as different genders) can be obtained with different m, even with the same input query. There seems to be no way to control the "number of identities" generated from this model. If so, how was the supervised loss applied to train a face recognition model? - How can one ensure high inter-class and large intra-class variations as required for SFR? - B.3.3. The assertion that high-quality data is not indispensable for achieving markedly accurate facial recognition performance is somewhat counterintuitive and perplexing. - The method's reproducibility raises concerns, particularly with respect to the training of the model, which lacks clarity. Specifically, the functions F_1 and F_2 in Equations (6) and (7), as well as the role of C_att, are not explicitly defined, and these elements are absent from Figure 3. The proposed model generally lacks controllable factors to generate true synthetic face images that favor high inter-class and intra-class variations. Technical Quality: 2 Clarity: 2 Questions for Authors: See above Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: No issues are found in this aspect. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: Low GtR is not attributed to real inquiry data. For instance, even with \textbf{the same synthetic data DDPM}, our CemiFace would surpass the previous state-of-the-art method DCFace inquired by DDPM data (clearly illustrated in the upper part of Tab. 4), and DDPM provides close performance to real-data. In fact, low GtR is attributed to the way of constructing the facial dataset, i.e. containing center-based semi-hard samples. To calculate GtR with fair comparison with other SOTA methods, we train the DCFace released data to generate close AVG performance to their paper by using CosFace loss. This is because the training SFR code of previous methods is not available which prevents us from reproducing the results. Notably, our reproduced DCFace is higher than the results reported in [23] --- W2: Only images belonging to one identity can be obtained from each input query regardless of the input m value. Despite that face images generated from the same query exhibited large differences, our approach defines them with the same identity label for the later SFR model training. Consequently, the number of identities is fully decided by the number of inquiry face images, which is mentioned in Fig. 1 stating that: "With our proposed CemiFace, each inquiry image finally forms a novel subject.". We will further explain this in the revision. --- W3: (1) high inter-class variations: Each inquiry face image is selected to be highly independent on other inquiry images. Specifically, we follow DCFace to use a pre-trained FR model to keep samples with a threshold of lower 0.3. We also elaborated this in lines 287-289. (2) high intra-class variations: high intra-class variations are ensured by (a) changing the similarity condition $m$, as a small input similarity $m$ results in the generated semi-hard images belonging to the same identity having long distances to the identity center; and (b) the face images of the same identity generated by CemiFace are distributed in all directions from the identity center, which can be observed from supplementary material TSNE Fig. 7. This is guaranteed by randomly sampled Gaussian noises $\epsilon$ input to the diffusion model, which exhibit a large variation. As a result, both properties would ensure the generated face images of the same identity are almost evenly distributed in a sphere that has a relatively large radius, and thus they would have high intra-class variations. --- W4: Despite that our approach achieved slightly worse FID performance than the previous state-of-the-art DC face, face images generated by our approach still result in a better face recognition performance, suggesting that the semi-hard face images generated by our CemiFace compensate for their slightly worse FID compared to face images generated by DCFace with a mix of easier and hard samples. However, CemiFace still significantly outperforms DigiFace in this FiD. On the other hand, high-quality data is essentially needed, as discussed in Section 4.2.2 (lines 263-282, Table 4) and supplementary material Sections B.1 \& B.2 We rephrase the last sentence in B.3.3 as: 'Our method doesn’t intend to generate images similar to the distribution of CASIA-WebFace, but to construct a discriminative dataset that is conducive to providing highly accurate FR performance' --- W5: $F_1$ and $F_2$ are two stacked linear layers. Here, $F_1$ projects the input similarity m to a latent feature $C_\text{sim}\in \mathbb{R}^{512} $, and then $F_2$ projects the feature concatenating $C_\text{sim}$ and identity embedding to a condition vector $C_\text{att} \in \mathbb{R}^{128} = F_{2}(cat(E_{id},F_{1}(m)))$ including both the input similarity and identity information. We have provided more details about Figure 3 and the training pipeline in general response G1 and the PDF. --- W6: we kindly disagree. High inter-class variation is ensured by filtering the inquiry data with a cosine similarity lower than 0.3 provided by pretrained FR model, following DCFace (mentioned in lines 288-289). As for high intra-class variations, our proposed diffusion model can generate high-variation samples belonging to each identity. Please refer to the answer of W3. --- --- Rebuttal Comment 1.1: Comment: Thank you for your response. The authors have addressed my questions satisfactorily. --- Reply to Comment 1.1.1: Comment: Dear Reviewer 5xKY, Thank you for your positive feedback in reviewing our paper. We will try to change the manuscript according to your suggestion in the updated version. Best Regards, The Authors of Paper 11025	Summary: The paper proposes a novel approach named C to address privacy concerns in face recognition technology. The authors propose CemiFace, a diffusion-based method that generates synthetic face images with controlled similarity to a subject's identity center, enhancing the discriminative quality of the samples. This approach allows for the creation of diverse and effective datasets for training face recognition models without the need for large-scale real face images, thus mitigating privacy risks. CemiFace outperforms existing synthetic face recognition methods, significantly reducing the performance gap compared to models trained on real datasets. The paper also discusses the potential limitations and privacy implications of the approach, highlighting the need for ethical considerations in synthetic face generation for face recognition applications. Strengths: 1. The use of a diffusion-based model for generating semi-hard samples is an innovative approach that has not been extensively explored in the field of face recognition. 2. The approach can be extended to use unlabeled data for training, which is an advantage over previous methods that often require some form of supervision. Weaknesses: 1. The paper is not well organized. This paper should be reorganized to make it easier for the reader to understand the contributions and technical details of this paper. 2. Eq. 10 seems to be inconsistant to its description. According to the description, it is highly related to the time step. 3. Fig. 3 is hard to understand. The training losses are not illustrated in the figure. 4. Despite aiming to reduce privacy issues, CemiFace still uses a pre-trained model that could have been derived from datasets without user consent, raising ethical and privacy concerns. Technical Quality: 3 Clarity: 3 Questions for Authors: Refer to weakness Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Refer to weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: In the last part of the Introduction section, we have clearly and specifically listed four contributions of our work, including: 1. a new and crucial finding; 2. a technical contribution (i.e., CemiFace face image generator) inspired by the finding; 3. an application contribution of our proposed technical approach; and 4. the effectiveness of our approach. Based on your suggestion, we will additionally add a paragraph at the beginning of the Method section to guide readers as: In Section 3.1, we first investigate the relationship between sample similarity and their effectiveness in training FR models, presenting the finding that samples with certain similarities (i.e., center-based semi-hard samples) to their identity centers are more effective for training FR models on a real dataset and subsequently devise a toy experiment to validate it. Inspired by our findings, we propose a novel CemiFace, a conditional diffusion model that produces images with various levels of similarity to an inquiry image in Section 3.2. Specifically, Section 3.2.1 introduces how we construct the similarity condition which is fed to diffusion model to guide the generation, and discusses the $L_{SimMat}$ to require the generated sample to exhibit a certain similarity degree to the inquiry image. In Section 3.2.2, we then present how to use our diffusion model to generate a synthetic face dataset given a fixed similarity condition $m$ and a set of inquiry images --- W2: As clearly presented in line 174 (above the Eq. 10), Eq. 10 is not related to the time step, instead it represents the loss for reconstructing the identity of the inquiry image when $t \rightarrow 0$, which is the first part of our Time-step Dependent loss. As presented in line 177, Eq. 12 (rather than Eq. 10) defines our $L_\text{SimMat}$ inspired by the Time-step Dependent loss (DCFace[23]), which employs $\gamma_{t}=\frac{t}{\mathbf{T}}$ to make the sample have different similarity property at different time step $t$. We will re-phases lines 173-174 as: we employ the Time-step Dependent loss [23] with different time step t at Eq. 12, specifically firstly an identity loss for recovering the identity of the original inquiry image x, which will be applied to produce original facial embedding when the time step $t\rightarrow 0$. --- W3: Please refer to the general response G1, where we have rephrased the pipeline and Fig. 3. --- W4: Ethical and privacy issues are the top priority of our study. While a crucial goal for developing SFR is to eliminate the privacy issue that lies in the face recognition dataset, current SFR approaches are not capable of refraining from privacy risk while providing effective recognition performance. Compared with previous SOTA approaches DCFace[23] and IDiffFace[27] approaches, our method already reduces the privacy risk by avoiding using labelled face recognition images for the model training. These are clearly presented in supplementary material Section C. More importantly, we emphasize that the model development protocol in this paper strictly follows the same protocol of previous peer-reviewed related studies (e.g., the DCFace and IDiffFace) published in top journals and conferences. Thus, we believe our study would not trigger ethical and privacy-related issues. --- --- Rebuttal Comment 1.1: Comment: Thank you for answering my questions. I have no further questions. --- Reply to Comment 1.1.1: Comment: Dear Reviewer L3p9, Thank you for your positive feedback in reviewing our paper. We will re-organize Fig.3 and the overall structure to make it easier for the reader according to your suggestions. Best Regards, The Authors of Paper 11025	Summary: The paper titled "CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition" addresses a critical issue in face recognition (FR) related to privacy and performance degradation when using synthetic face images. The authors propose a diffusion-based approach, CemiFace, which generates facial samples with varying levels of similarity to an identity center. This method aims to enhance the discriminative quality of synthetic samples, thereby improving the performance of FR models trained on these datasets. Strengths: Introducing Similarity controlling factor in synthetic face generation using a diffusion based approach. Weaknesses: (a) Due to the introduction of this similarity control conditioning in the diffusion process there must be a change in total sampling time ( certainly it will also depend on the number of time steps considered in the diffusion process also) – An illustration/analysis on computational complexity of the proposed algorithm is needed. (b) Seems like the overall process is dependent on how(using which method) the value of m was determined during the diffusion process! (c) A complete pseudo-code on the proposed method would have helped the reader to understand the whole process. (d) Figure 3 could have been much more elaborated and in more details. Technical Quality: 3 Clarity: 3 Questions for Authors: While comparing the proposed work with other SOTA methods - how did you generate the results of the SOTA method? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations were mentioned only in the last few sentences of the conclusion section but not otherwise stated separately. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: We first define some basic calculation complexities: Time Step $ T $: This represents the total number of time steps required for a complete diffusion process. UNet Complexity $C_{\text{UNet}}$: The UNet model accepts the input image and outputs the estimated noise. Pretrained Face Recognition Model $ C_{\text{ID}} $ and Similarity Condition $C_{m}$ : These are processed by two sequential linear layers. The backpropagation and weight updating time complexity for the UNet is given as $B$. Since the pretrained face recognition model is fixed, there is no computational complexity associated with it during backpropagation. Forward Diffusion Process: The forward diffusion process consists solely of adding noise to the clean image over $T$ time steps, resulting in a calculation complexity of $O(T)$. Denoising Process: The denoising process also consists of $T$ time steps to gradually denoise a noisy image. This process involves a UNet and conditions to output the estimated noise, resulting in a calculation complexity of $O(T \cdot (C _{\text{UNet}} + C _{\text{ID}} + C _{m})) $. Training Step: A training step includes a forward diffusion step, a denoising step, and model training (including loss calculation and backpropagation). Note that the model training only involves one time step. The overall complexity of training a CemiFace model by one step is given by: $ O(N \cdot (C_{\text{UNet}} + C_{\text{ID}} + C_{m} + B)) $ where $N$ is the number of samples, note the loss calculation involves the similarity comparison between the estimated image $\hat{x_{0}}$ and the original image $x_{0}$, thus the pretrained FR is adopted 3 times in each training iteration. For a standard diffusion model, the complexity is: $O(N \cdot (C_{\text{UNet}} + B))$ Accumulated Training Iterations: When accumulating all the training iterations $I $, the calculation complexity becomes: $O(I \cdot N \cdot (C_{\text{UNet}} + C_{\text{ID}} + C_{m} + B)) $ Generating One Sample: To generate one sample, the process involves denoising random noise from time step $T$ to 0. Since the forward diffusion process and loss calculation are no longer needed, the complexity is: $O(T \cdot (C_{\text{UNet}} + C_{\text{ID}} + C_{m})) $ Generating a Dataset: For generating a dataset with $ L $ identities and $ S$ samples for each identity, the complete calculation complexity is: $O(L \cdot S \cdot T \cdot (C_{\text{UNet}} + C_{\text{ID}} + C_{m}))$ On a single A100 GPU, (1) generating a sample with 50 time steps takes 0.9 seconds; (2) One CemiFace training step (one inquriy image) takes 0.6 seconds; (3) Training for one epoch takes 45 minutes; (4) It takes 10 epochs to converge with a total time of 7.5 hours. (5) It takes 16 hours to generate a dataset of 10,000 identities, with each having 50 images. We have mentioned the computational cost in supplementary material Sec. A.1, lines 473-480. --- W2: The overall process not only depends on m but also relies on the input inquiry image. The conditions $C_{att}$ sent to diffusion models are determined by $m$ and the inquiry data using a linear projection $F_{2}$. The selection of inquiry images is dependent on the rule that the inquiry images should be unblurred, non-occluded, appropriately posed and independent of each other (discussed in Sec. 4.2.2, supplementary material B.1 and B.2). It is also required by SFR to ensure high inter-class variation. Then the generation $m$ directly determines the final similarity property of the samples inside each identity group, we find the optimal $m$ is scalar number 0 and mixing the generation $m$ would bring worse performance(lines 232-243). --- W3: The pseudo-code is given in the PDF, Algorithm 1 and Algorithm 2 --- W4: The revised figure is appended in the PDF Fig. 1, while its description is provided in the general response G1, which will be added to the main text. --- Q1: In Tab. 6, the results of all competitors except DCFace$\dagger$ are obtained from their original publications. Here, we additionally train a SFR model based on the released synthetic face images/dataset generated by the SOTA DCFace$\dagger$ model and the CosFace loss, to facilitate a fair comparison with ours. This is because DCFace hasn't released AdaFace-based SFR training code and details, and thus we were not able to reproduce it for our SFR model training. For this SFR training, as described in lines 215-229, we use IR-SE-50 as backbone and CosFace loss for learning facial embedding, where the hyperparameters are set the same as the standard CosFace. --- L1: Although most papers only discussed their limitations in their conclusion section due to page limitation, our submission has detailed limitations not only in the conclusion but also: (1) Sec. 4.2.2 (lines 276-295), Sec. B.1 and B.2 of the supplementary material of the main submission (i.e., high-quality inquiry data is essential); (3) Sec. C of the supplementary material (i.e., privacy issues); and (4) Sec. D.1 of the supplementary material (i.e., the training of our CemiFace depends on the pre-trained FR model, and thus its performance also relies on the performance of this model). --- Rebuttal Comment 1.1: Title: Reply to the Authors Comment: Dear Authors, Thanks for your detailed explanation on all my queries. I don't have any further questions. With Regards, Reviewer fjPL --- Reply to Comment 1.1.1: Comment: Dear Reviewer fjPL, Thank you very much for your kind reply. We have noticed that the rating has not been changed. Please kindly let us know if there is any additional concern we can clarify. We sincerely appreciate your valuable suggestions and will try our best to meet your criterion. Best regards, The authors of Paper 11025	Rebuttal 1: Rebuttal: We thank all reviewers for their valuable feedback. Reviewers acknowledged that our: (i) method is new/innovative (tciQ, L3p9, 5xKY, Crof), interesting (tciQ), effective for SFR (fJPL, L3p9, 5xKY, Crof), and addresses privacy concerns (fJPL, L3p9, Crof); (ii) discovery is important (Crof); and (iii) experiments are solid (5xKY, Crof) and comprehensive (tciQ) General response to all reviewers: we denote the Author Rebuttal pdf file as the PDF G1: Improvement and explanation of the Fig. 3 (@tciQ\& @fJPL\& @L3p9\& @5xKY): We have updated the Fig. 3 in the PDF with more detailed elaboration. Besides, we will also modify and add the following contents to the beginning of the Sec. 3 as: Methodology overview: As illustrated in the left side of Fig. 3, the training process starts with adding a noise $\epsilon \sim \mathcal{N}(0,1)$ to the clean input image $x$ using Eq. 4 (result in $x_{t}$). Meanwhile, similarity conditions $m$ are fed to the linear layer $\mathbf{F_{1}}$, whose output is then concatenated with the inquiry identity condition $\mathbf{E _{id}}(x)$ to generate a joint representation. This joint representation including both identity and similarity conditions is then processed by the linear projection layer $\mathbf{F} _{2}$ to output the combined condition representation $\mathbf{C _{att}}$ (the lower right part of Fig. 3). The $\mathbf{C_{att}}$ is further processed by a cross-attention operation with the intermediate latent representation of diffusion UNet $\sigma_{\theta}$ learned from the input noisy image as: \begin{equation} CA(Q,K,V,K_{c},V_{c})=SoftMax(\frac{QW_{q}([K,K_{c}]W_{k})^{T}}{\sqrt{d}})W_{v}[V,V_{c}] \end{equation} where $\mathbf{C_{att}}$ is treated as the key $K_{c}$ and value $V_{c}$ (same as DCFace) to influence the generated face images. $Q=K=V$ are the query, key and value, representing the latent feature of UNet $\sigma_{\theta}$. Consequently, the diffusion Unet $\sigma_{\theta}$ outputs the estimated noise $\epsilon'=\sigma_{\theta}(x_{t},t, \mathbf{C_{att}})$ for denoising the image as a clean estimated image $\hat{x}_{0}$ (Eq. 9). Based on the obtained estimated image $\hat{x} _{0}$, original $x$ and condition $C _\mathbf{{att}}$, the whole model is optimized by the loss defined in Eq. 13 in an end-to-end manner. At the inference stage (the upper right part of Fig. 3), random noise $x_{t}=x_{\mathbf{T}}=\epsilon \sim \mathcal{N}(0,1)$ and the time step $t=\mathbf{T}$ are first fed to a CD block, jointly with the similarity $m$ and inquiry image $x$ that are undergoing linear layers $F_1$ and $F_2$ and concatenation to produce the condition representation $\mathbf{C_{att}}$. This results in an estimated noise $\epsilon'=\sigma_{\theta}(x_{t}, t, \mathbf{C_{att}})$. Then, a denoise step is adopted to generate $x_{t-1}$ from $x_{t}$ using Eq. 12 in DDIM[21] for efficient interface speed. This process is repeatedly conducted on the obtained denoised latent images ($x_{t-1}, x_{t-2}, \cdots, x_{0}$) until $t=0$, where $x_{0}$ is treated as the final generated face image. Here, we assign the same identity label as $x$ to all face images generated from the inquiry image $x$. To ensure high inter-class variation, our inquiry images are filtered by a pretrained FR ( IR-101 trained on the WebFace4M[11] dataset by AdaFace.), which enforces the similarity between each pair of query images is lower than 0.3. Pdf: /pdf/7ba3c8b4a4894a36f8efd8e2ec91d3fe219991ff.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper proposes a new Face Recognition diffusion-based generation method. The diffusion process is completed with a semi-hard constraint on the synthetic reconstructed image: for each inquiry image of the (real) training set, the reconstructed image after the forward-backward diffusion process must have a specific cosine similarity with the inquiry image. As it is usual for such methods in Face Recognition, the resulting synthetic dataset is then used for training a Face Recognition model. This model is evaluated across diverse real datasets. Strengths: The tackled problem is quite hard and needed at the same time. Current SOTA Face Recognition generation methods lead to a significant gap in terms of performance, compared to real Face Recognition datasets (of the same size). The idea of controlling the similarity to design semi-hard samples is also interesting. Weaknesses: 1) In Fig. 1, is the displayed similarity really the cosine similarity ? In the generated samples, the line with perfect similarity (equal to 1) seems to provide synthetic images which would not have a perfect similarity with the inquiry images displayed above the hypersphere. 2) [minor] In Eq. 3, the probability distribution of epsilon is not specified. 3) The authors should cite explicitly the works that use the training loss (Eq. 2) in this precise form, as there are alternative loss functions for diffusion models. A discussion on the reasons of this particular choice of diffusion loss might be a plus (e.g. in the appendix). 4) [minor] Although the lines 114-116 are accurate, they are misleading the reader. The widely known representation of Face Recognition embeddings is that they lie onto a hypersphere of dimension N, where each embedding is a point of the hypersphere. Those embeddings are clustered by identity on this sphere and the identity centers are roughly at the center of those clusters. The hypersphere mentionned in this paper is a hypersphere of dimension N-1, where the identity center is at the center of the sphere. 5) In lines 120-125, the authors should detail the range of similarities to the identity center, for each of the 5 splits of the CASIA training set. Only the average similarity of each split is specified. 6) [minor] Figure 3 should be a bit more explained than just its caption. 7) [major] Lines 155-162 are not well written and it is hard to understand how the margin m is used to guide the diffusion process. In particular, F_1 and F_2 are not defined, while some unused F is mentionned. C_sim seems to be a vector of unknown size. Also, the temporal guidance is too briefly described. 8) [minor] Some hyperparameters' values (alpha_t/beta_t, lambda) are not specified. 9) The right part of Fig. 4 displays two curves that do not have the same meaning for the x-axis. For AVG, the similarity is a constrained similarity (m) for training CemiFace (i.e. a similarity between a real inquiry image and a synthetic image). For CASIA, it is the similarity between one real image and its identity center (not a real image). To sum up, for AVG it is a similarity between 2 images, while for CASIA it is between 1 image and its identity center. Thus, comparing the two curves does not seem meaningful. 10) [major] The CosFace loss is used to train on synthetic datasets, while AdaFace is used to produce (identity-oriented) embeddings for the CemiFace training set generation. There should be only one model for both tasks, for fair comparisons. On Table 6, training on CASIA with AdaFace gives better results than with CosFace, so one could attribute the good performance of CemiFace to the fact that the authors used a stronger model (AdaFace) to generate the synthetic dataset than the model used to train on this dataset (CosFace). In addition, there should be a part studying the impact of this AdaFace choice (i.e. another loss), at least in the appendix. 11) The ROC curve on IJB-B/IJB-C for all synthetic methods of Table 6 would be a plus, as the accuracy is easily saturated, and not really used in industrial use-cases. Previous papers (related works) provide such ROC plots. Technical Quality: 3 Clarity: 2 Questions for Authors: 1) Could you explain the last sentence of Section 4.2.1 (lines 260-261) ? 2) In Section 4.2.2, why is the range of training m equal to [0,1] while the previous subsection concludes with an optimal range [-1,1] ? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1) [major] The SimMat loss seems to be an interesting idea to lead towards m-similarity to the inquiry image, during training. But the derivation of the MSE loss (Eq. 3) assumes that the reconstructed image should be the inquiry image, and not a new image having a m-similarity with the inquiry image. I may be wrong here but I think that the diffusion loss of Eq. 3 is mathematically valid if the forward diffusion process is symetric to the reverse diffusion process, which is not the case here. 2) [major] In Section 4.2.1, there should be a part studying the difference between the required m and the estimated m post-training. That means that, for any required m, it is easy to compute the estimated m, i.e. the similarity between the resulting reconstructed image and the inquiry image. This estimated m must be quite different than the required m because Figure 4 shows that the best m is m=0, meaning that there is a 90 degrees angle between the inquiry image and the synthetized image. If m was truly equal to 0, the performance of the model would be very poor. So, there must have a difference between the required m and the real estimated m (post-training). This is also reflected in the conclusion saying that the best setting for m is to train CemiFace with m randomly sampled from [-1,1]. A similarity truly equal to -1 would bring a model with an astonishingly poor performance. Update I have increased my score from 4 to 5. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: --- W1: The displayed similarities are the input cosine similarities, based on which the displayed face images were generated. However, the actual similarities between the generated images and their inquiry images may not be exactly the same as the input cosine similarities as DL models typically cannot generate perfect/exact outputs. The actual cosine similarities between generated and inquiry images are provided in the PDF Tab. 2, measured by the IR-50 network pretrained using AdaFace loss (line 154). --- W2: The $\epsilon$ is a random Gaussian noise image $\\epsilon\sim N(0,1)$ fed to our diffusion model (Line 145). We will explain this in the revision. W3: We assume the mentioned training loss is Eq. 3 rather than Eq. 2. We will modify and cite as: we follow the previous SOTA SFR studies (DCFace[23] and IDiffFace[27]) to choose the same generic diffusion loss [20,21,22], ensuring the reproducibility of our approach and its fair comparison with DCFace and IDiffFace. W4: The theories of yours and ours are the same but described in different forms. We treat all face images of each subject as an N-1 dimensional sphere with its center representing the subject-level identity center. Then, the spheres of all subjects can be combined in an N-dimensional sphere, where each subject-level sphere is a cluster. We will follow your suggestion to rephrase these sentences W5: The ranges for each group are listed in the Tab. 1 of the PDF. Here, the splitting boundary for each group varies across different identities. W6: Please refer to the general response G1. W7: $F$ should be $F_{i}$ for representing the linear projection operation, and thus $F_1$ and $F_2$ are two stacked linear layers. Here, $F_1$ projects the input similarity m to a latent feature $C_\text{sim}\in \mathbb{R}^{512} $, and then $F_2$ projects the feature concatenating $C_\text{sim}$ and identity embedding to a condition vector $C_\text{att} \in \mathbb{R}^{128} = F_{2}(cat(E_{id},F_{1}(m)))$. Subsequently, the $C_\text{att}$ is fed to our diffusion to control its face image generation via a cross-attention (CA) as: \begin{equation} CA(Q,K,V,K_{c},V_{c})=SoftMax(\frac{QW_{q}([K,K_{c}]W_{k})^{T}}{\sqrt{d}})W_{v}[V,V_{c}] \end{equation} where ${\mathbf{C_{att}}}$ is treated as the key $K_{c}$ and value $V_{c}$ (same as DCFace) to influence the latent representation extracted from the input noisy image (treated as the query, key and value $Q=K=V$) for generating the final face image. Please also refer to the general response G1 for details. W8: $\beta_{t}$ starts from 0.0001 to 0.02 controlled by time step $t$-based linear schedule, while $\alpha_{t}=1-\beta_{t}$ (line 101) W9: Sorry for the confusion. We put two curves in the same figure due to the limited space. We will separate the two curves into two smaller figures. W10: As DCFace hasn't released its AdaFace-based SFR training code and details, we were not able to reproduce it for our model training. Thus, lines 216-220 and lines 300-311 fairly compare ours with DCFace by adopting the same pre-trained AdaFace model to train our diffusion generator, and then employing the same CosFace loss for both ours and DCFace's SFR models training. Results show that our CemiFace still outperformed the SOTA DCFace. Based on your suggestion, Tab. 4 in the PDF additionally provides results achieved by using unified CosFace. Specifically, we apply a model pre-trained by CosFace to train both our generator and DCFace generator, and employ the same CosFace loss for their SFR models' training. Due to limited rebuttal time, we only include the results for the 0.5M setting, but will present results for all data volumes in the revision. W11: We provide ROC curves for CASIA, DCFace and CemiFace in 3 data volumes in the PDF Fig. 2. The FR model trained on our CemiFace generated dataset with the best curve (largest area), while the real face dataset provided inferior results than these SFR methods (CemiFace and DCFace). When FAR=1e-3, the TAR result achieved by our self-implementation on real-dataset is around 90, which exceeds previous works (ArcFace gives around 60 in [a]]). Importantly, the ROC of the IJB-B and IJB-C are typically not provided in neither the related SFR studies [23,24,25,27] nor many of the discriminative FR methods (e.g., CosFace[1], Face Transformer[b], boundaryface[c]) when using CASIA-WebFace for training. [a]Federated Learning for Face Recognition with Gradient Correction [b]Face Transformer for Recognition [c]BoundaryFace: A mining framework with noise label self-correction for Face Recognition Q1: We rephrase it as: We did not provide the AVG result achieved for setting the similarity interval (Table 2) close to zero (continuous similarities), as this training setting leads our model to generate same images when inputting different $m$ values. We assume that an extremely small similarity interval prevents the model from effectively learning the differences between each level of similarity Q2: It is a typo and should be [-1,1] L1: Eq.3 does not directly compare the input and reconstructed images. Instead, it compares the input noise and the estimated noise (Line 98). The overall loss Eq. 13 adopts $\lambda$ to balance between the diffusion focuses on comparing input and estimated noises for generating clean images (Eq.3 $L_{MSE}$) and the similarity between the generated samples and inquiry image (Eq. 12). Given the relationship between the estimated $\hat{x} _{t-1}$ and real $x _{t-1}$ as follows, it is symmetric and valid if the $\epsilon'$ is accurately estimated. \begin{equation} \hat{x} _{t-1} \approx \frac{x _t - \sqrt{1 - \alpha _t} \epsilon'}{\sqrt{\alpha _t}} = x _{t-1} \end{equation} L2: Please refer to W1, where $m=0$ produced images of an actual average similarity of 0.2854, making the generated dataset semi-hard. We will update this table in the revised version. --- Rebuttal Comment 1.1: Comment: Dear Reviewer tciQ, We are deeply grateful for the time and effort you spent reviewing our paper. We have carefully tried to address each of your questions based on your valuable feedback. Could you please take a brief moment (approximately 2-3 minutes) to review our responses? We appreciate your feedback, regardless of whether our answers have addressed your primary concerns. We are willing to provide additional details if you need further explanation. Best regards, The authors of Paper 11025	null	null	null	null	null	null
Controlling Continuous Relaxation for Combinatorial Optimization	Accept (poster)	Summary: This article finds that the existing UL-solvers will trap into local optima and face rounding issues. This study proposes a continuous relaxation annealing (CRA) strategy and an auxiliary function to facilitate training. Strengths: 1. The method proposed in the article is sound, easy to implement, and effective. 2. The article is well-written. Weaknesses: There are no major drawbacks in this article. There should be more reviews on neural combinatorial optimization solvers that apply annealing ideas as well (such as [1] provides annealing on the distance matrix of TSP). [1] Lin, Xi, et al. ""Continuation path learning for homotopy optimization."" International Conference on Machine Learning. PMLR, 2023. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Figure 8 provides parameter analysis for the N=10000 MIS problem. Can parameter sensitivity analysis be provided for other CO problems to demonstrate that CRA can be widely applied to general CO problems without special parameter design? 2. I am also concerned about the convergence speed under different parameter settings. Could you provide it as a function of the initial scheduling and scheduling rate? 3. Can CRA be applied to routing problems such as TSP? If I understand correctly, the current $\phi$ function will have very small $p$-values in solving TSP, which may probably lead to a failure situation of CRA. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 2 Limitations: Well discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thorough and thoughtful review. We appreciate your recognition of our method's strengths, noting that it is "sound, easy to implement, and effective." Additionally, we are grateful for your comment that this article has no major drawbacks. Weakness: Thank you for your valuable suggestion regarding including a broader discussion on annealing methods. We acknowledge the importance of situating our work within the context of existing annealing-related research. The revised version will incorporate a comprehensive review of relevant annealing methods in the Related Work section, including the paper [1] you mentioned. Question 1: Thank you for your valuable question regarding parameter sensitivity analysis for other CO problems. For the MaxCut problem, we observed a similar sensitivity to parameters as in the MIS problem. Specifically, slower annealing rates result in better solution performance for Gset problems and MaxCut problems on RRGs, like simulated annealing. The revised version will include sensitivity analysis results for MaxCut and other non-d-regular random graphs in the Appendix. Ablation studies will accompany these results to demonstrate the robustness of CRA across various problem settings. Question 2: Thank you for your insightful question regarding the parameter sensitivity analysis. For the MIS problem, we confirmed that a linear change in $\gamma$ results in corresponding convergence speeds. To address your concern, we will include detailed results in the revised version showing the stopping time for different parameter settings and initial schedules. This will provide a clearer understanding of how the parameter settings affect the convergence speed across different problem instances. Question 3: Thank you for your insightful question regarding the applicability of CRA to the Traveling Salesman Problem (TSP). Our method can indeed be extended to this domain. However, the optimal $p$ value for solving TSP is unclear. In this study, our primary focus was on addressing the challenges of UL-based solvers highlighted by Wang et al. (2023) and achieving consistent results with PI-GNN. Consequently, we did not conduct experiments on TSP. Nonetheless, we recognize the importance of comprehensive comparisons for TSP and other problems. We plan to explore these areas in our future work, ensuring that our method's applicability and performance are thoroughly evaluated across a broader range of combinatorial optimization problems. We believe these revisions and the additional analyses will significantly enhance the robustness and comprehensiveness of our work. We kindly request you to consider this in your final evaluation. --- Rebuttal Comment 1.1: Comment: Thank you for your response. There are no major drawbacks to this article. I will increase the rating to 6 to support your acceptance. I still believe that exploring applications in routing problems such as TSP is important, and I hope to see this part in the rebuttal period and the following versions of your manuscript. --- Rebuttal 2: Title: Reply to Reviewer J6G3 Comment: Thank you for recognizing the contribution of our paper, and we appreciate your decision to raise the score. Based on your insightful comments, we conducted additional experiments on several TSP problems from the TSPLIB dataset (http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/). Below, we present the results demonstrating the $p$-dependency of the solutions. \| Instance \| ApR ($p=2$) \| ApR ($p=4$) \| ApR ($p=6$) \| ApR ($p=8$) \| ApR (PI) \| Optimum \| \|------------\|----------------------\|----------------------\|----------------------\|----------------------\|---------------------\|---------\| \| burma14 \| 0.91 ± 0.08 \| 0.98 ± 0.10 \| 0.97 ± 0.14 \| 0.99 ± 0.06 \| 0.736 ± 1.21 \| 3,323 \| \| ulysses22 \| 0.89 ± 0.03 \| 0.92 ± 0.02 \| 0.88 ± 0.07 \| 0.89 ± 0.05 \| -- \| 7,013 \| \| st70 \| 0.96 ± 0.01 \| 0.85 ± 0.03 \| 0.88 ± 0.04 \| 0.80 ± 0.02 \| -- \| 675 \| \| gr96 \| 0.81 ± 0.05 \| 0.82 ± 0.03 \| 0.90 ± 0.05 \| 0.86 ± 0.05 \| -- \| 5,5209 \| In these experiments, we calculated the ApR as the ratio of the optimal value to the CRA result, with the ApR representing the average and standard deviation over 3 seeds. The "--" in PI-GNN indicates that most variables are continuous values and that the maximum epoch did not obtain a solution satisfying the constraint. We used the same GNN and optimizer as in our experiments of the main text. The table shows that the CRA approach found solutions with an ApR exceeding 0.9 across various instances. Notably, for the "burma14" problem, our method found the global optimal solution, 3,323, several times. However, the optimal $p$ value may vary depending on the specific GNN architecture and problem structure, which suggests that a more comprehensive ablation study would be beneficial in future work. We will include these results and additional problem instances in the appendix or main text of the revised version. We hope that these enhancements, which demonstrate the broader applicability and robustness of our approach, will further highlight the contribution of our work. We would greatly appreciate it if these improvements could be taken into consideration during your final evaluation.	Summary: The proposed approach is an optimization method for each graph over GNN parameters where each output corresponds to the likelihood of the node belonging to the solution. The objective function consists of a penalty term along with a parameter scheduled to control the non-convexity of the objective. Strengths: 1- The convex annealing approach proposed in training that controls the level of non-convexity. This is a valid approach to avoid getting trapped in local minima where the solution sizes are not large. 2- Theoretical results of the limiting points of the proposed objective with different \gamma. 3- The "no-data" requirement makes this method mostly generalizable, depending on tuning a set of hyper-parameters for each graph distribution. Weaknesses: [Major Comments] 1- The need to solve graph-based NP-hard problems that are originally formulated as ILPs stems from the unscalability of these solvers. For example, the scalability of the MIS problem depends on the number of nodes and the number of edges in the graph. This needs to be the motivation instead of the issues encountered in UL-based solvers. 2- While the proposed approach does not require training data (labeled or unlabeled), there are several hyper-parameters. Tuning these hyper parameters is a challenge. Further discussion is needed here. 3- Getting trapped in local minima is not only the case in GNNs or PI-GNNs. It exists for any continuous relaxation of Problem 1. This is due to the non-convexity inherited in these formulations. For example, if we re-write Problem (3) in matrix form, we can see that the objective has a constant hessian equal to the adjacency matrix of the graph. If the magnitudes and signs of the eigen values vary significantly, then this indicates possible positive and negative curvatures in the loss landscape. Replacing x in Problem (3) with the output of a GNN does not guarantee changes. Although it may be possible that it will make some local minima avoidable by adaptive optimizers (such as ADAM), there is a possibility that this type of overparameterization would create unwanted local minima that do not result in any feasible solutions. Theoretically analyzing this is very complicated due to the use of a GNN. However, empirical investigation can be used to better motivate and understand the proposed approach. 4- Similar to the previous point, rounding issues existed even before GNNs. See the SDP relaxations of MIS [1] and MaxCut [2] and how their dependence on rounding techniques (e.g. spectral clustering [3]) often fails to obtain optimal solutions. Rewriting is needed here. 5- The Stationary point p* = 0_n was not discussed in Section 3.1. Furthermore, in line 208, it is 0_n, whereas in line 2016, it is 0_N. 6- How was the GW approximation applied for the MaxCut problem? This approximation requires a normalized random vector drawn from the standard Gaussian distribution. How many samples were drawn? Given the SDP solution, one can simply draw multiple samples and pick the best where the only requirement is matrix-vector multiplication. This runs extremely fast with (i) no parameters of a NW, and (ii) no hyper-parameters to tune. The scenarios where such approaches fail need to be the motivation to propose the over-parameterized approach with convex annealing. 7- Missing many "data-independent" baselines (methods that do not require pre-trained models (such as DIFUSCO [4]) or training data such as RL-based solvers (LwD [5])) for comparison such as ILP solvers (Gurobi, CPLEX, or CP-SAT [6]), sampling methods such as iSCO [7], SOTA heuristics such as ReduMIS [8], and differentiable solvers such as [9]. 8- Why does the paper only consider d-regular graphs? How about the performance on other graphs? How does the run-time of this method scale in terms of the graph order and density? This is a major limitation of this work. [Minor Comments] 1- What is script C in line 83? 2- What is I and J in the equation after line 86? 3- "nural" in line 106. 4- Cite equation 3. An example is [10] 5- Paragraph 149 to 151 is ill-sentenced. 6- “Indeed” in line 232. 7- Cite Potts variable optimization. 8- This study “employs” in line 237. 9- "are" in Appendix F.3 in line 247. [References] [1] On the shannon capacity of a graph. IEEE TIT, 1979. [2] Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming. JACM, 1995. [3] A tutorial on spectral clustering. Springer, 2007. [4] Difusco: Graph-based diffusion solvers for combinatorial optimization. NeurIPS, 2023. [5] Learning What to Defer for Maximum Independent Sets. ICML, 2020. [6] https://developers.google.com/optimization [7] Revisiting sampling for combinatorial optimization. ICML, 2023. [8] A differentiable approach to the maximum independent set problem using dataless neural networks. Neural Networks, 2022. [9] A branch and bound algorithm for the maximum clique problem. Computers & operations research, 1992. Technical Quality: 2 Clarity: 2 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: See Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for reviewing our paper, appreciating the strengths in our approach, particularly noting the validity of the convex annealing method to control non-convexity and avoid local minima, as well as the theoretical results of the limiting points with different $\gamma$ and the generalizability due to the "no-data" requirement. Strengths 1: We are grateful for your recognition of the validity of our approach to avoiding local minima. As you commented, "This is a valid approach to avoid getting trapped in local minima where the solution sizes are not large." We appreciate your positive feedback. Our solver has demonstrated that its solution performance (ApR) does not deteriorate even as the problem size increases. In fact, as shown in Figure 4, the computational time increases sub-linearly, similar to PI-GNN, indicating good scalability compared to other solvers such as Gurobi. Furthermore, the ApR does not significantly deteriorate even with increasing problem size. CRA-PI-GNN has shown significant improvements for the largest MaxCut problem G70. Therefore, we believe that scalability for larger problem sizes is not an issue. Could you please clarify what size you consider large? We will present results and solution times for MIS and MaxCut problems of the sizes you expect during the discussion time or in the revised version. We will include further discussion in the revised version of the manuscript. Weakness 1: Thank you for pointing out that "the need to solve graph-based NP-hard problems that are originally formulated as ILPs stems from the unscalability of these solvers." We agree this is an important issue, even more so than the problems encountered in UL-based solvers. We will consider your comments in the revised manuscript. Weakness 2: Regarding the comment, "While the proposed approach does not require training data (labeled or unlabeled), there are several hyper-parameters. Tuning these hyper-parameters is a challenge. Further discussion is needed here." As you pointed out, our solver includes hyper-parameters related to the annealing start point and the annealing rate. As shown in Appendix Figure 8, better solutions can be obtained by starting the annealing process from a lower point and slowing down the annealing rate. This annealing mechanism is similar to those used in simulated annealing and its derived algorithms, which face similar issues. Therefore, we consider this problem common to our solver. Developing methods to determine these hyper-parameters adaptively is an important future work. Additionally, other GNN architectures and other hyper-parameter settings can potentially yield better results, but our default settings demonstrate that many problems can be solved with reasonable accuracy. Weakness 3: Thank you for your insightful comment, "if we re-write Problem (3) in matrix form, we can see that the objective has a constant hessian equal to the adjacency matrix of the graph. If the magnitudes and signs of the eigenvalues vary significantly, then this indicates possible positive and negative curvatures in the loss landscape." This observation is indeed correct. As you noted, non-convexity arises at the stage of continuous relaxation of the objective function in the main text (line 114) before introducing GNN. While formulating with GNN may introduce new undesirable local minima, we intuitively expect the objective function to be smoother by over-parameterizing the optimization problem of the relaxed variables with a higher-dimensional neural network. However, as you correctly pointed out, theoretically verifying this is very challenging due to the non-linearity of GNNs. Therefore, we use empirical studies to discuss the issues arising from parameterizing relaxed variables with GNNs and the plateau problem at $0_{N}$ in Appendix F1. Such empirical research has not been conducted by other studies using UL-based solvers such as PI-GNN or Wang (2022, 2023), and we consider this one of our contributions. While our empirical studies are not comprehensive, we acknowledge the importance of more extensive numerical experiments and examining the solution paths in high-dimensional spaces traversed by GNNs using techniques like PCA for dimension reduction. We consider this an important direction for future work. Weakness 4: In lines 49-55, we explain the rounding issues in general linear programming (Hoffman-Kruskal theorem). However, we agree that adding explanations regarding SDP relaxations for MIS and MaxCut would be beneficial. We will revise the introduction to include this context in the revised version. Weakness 5: Due to space constraints, detailed explanations of the stationary point $p^{\ast}=0_{N}$ are provided in Appendix F1, as mentioned in lines 186-187. In the revised version, we will ensure the main text includes sufficient details about the stationary point to be self-consistent. We are also grateful for pointing out the inconsistent use of $0_{N}$ and $0_{n}$. Weakness 6: The parameters for the Goemans-Williamson (GW) approximation are all set according to the settings in the PI-GNN paper's (approximate) polynomial-time GW algorithm. The implementation used the open-source CVXOPT solver with CVXPY as the modeling interface. In the revised version, we will provide detailed parameter settings in the Appendix to address your concerns. Weakness 7: Please refer to the "Numerical Experiments" section of the Unified Response for a comprehensive discussion. Weakness 8: Please refer to the "Runtime Scalability with Varying Graph Density" section of the Unified Response. Given this evidence, we have adequately addressed your concerns and kindly request you to reconsider your score based on these results. --- Rebuttal 2: Title: Thank you for the efforts. Response to Authors Comment: ### Response to Authors: - The experiments are evaluated on RRG and ER. The run-time of the proposed method scales only with the size of the graph, which I find impressive. Your response "runtime remains nearly constant as graph order and density increase" is not accurate as it is clear that the run-time is ~4X more when $n$ increased. In subsequent versions of the paper, I highly recommend including KaMIS and ILP solvers in these tables, as they are both known to scale with density, not just graph size. However, first, I find that the Apr in these results are low, and KaMIS should have been used to examine the solution quality. Second, in general, I still think that the proposed method is under-evaluated. The authors should consider using several random graph generators from NetWorkX, especially since it is a dataless method. - Discussing the intuition of over-parameterization to smoothen the landscape is acknowledged. I highly recommend that the authors include their response to this comment in the main body of the paper. - I acknowledge the additional experiments for MIS. The method clearly outperforms DIFUSCO while requiring no training data. However, the proposed approach still underperforms when compared to iSCO, a training-data-free sampling method where no GNNs are needed. However, I think that iSCO, while reporting significant results, is sensitive to the choice of the sampler and hyper-parameters. So overall, the CRA results in this table are fair. - Question: Why are the CRA MIS size results in the "Numerical Experiments" global response competitive, but the results in the attached PDF are not (way below 1)? - Larger graphs sizes (nodes) or harder instances can be any graph where ILPs and heuristics struggle in terms of either run-time or solution quality. For MIS, the authors can try mid-density (with respect to the complete graph) GNM graphs or ER (with $p > 0.4$) graphs. - Overall, I appreciate the dataless training approach and the theoretical support of the proposed method, and I encourage the authors to further investigate this direction. However, I feel that the paper could benefit from (i) a major revision in terms of writing (e.g., polishing the arguments about the non-convexity and rounding issues of general continuous relaxations of COPs) and (ii) a more comprehensive evaluation and explanation of the results. Therefore, I will only increase my score to 4. --- Rebuttal Comment 2.1: Title: Further response to reviewer BuGW Comment: Thank you for raising your score, thoroughly reviewing it, and providing insightful comments. We also thank you for your insightful understanding of over-parameterization. > However, first, I find that the Apr in these results are low, and KaMIS should have been used to examine the solution quality. We apologize for any confusion caused by the results in our global response's PDF. The results are based on IS density $\rho$ as defined in Line 99, not ApR. To clarify, we have included below a revised table where ApR is calculated, comparing our method against theoretical results [Barbier et al., 2023], as in Lines 299-302 in our experimental section. \| Problem \| $\mathrm{ApR}$ (CRA) \| $\mathrm{ApR}$ (PI) \| Time (CRA) \| Time (PI) \| \|-----------------------\|--------------\|-------------\|------------\|------------\| \| $\mathrm{RRG}(1{,}000, 10)$ \| 0.95 \| 0.78 \| 108 (s) \| 98 (s) \| \| $\mathrm{RRG}(1{,}000, 20)$ \| 0.95 \| 0.56 \| 103 (s) \| 92 (s) \| \| $\mathrm{RRG}(1{,}000, 30)$ \| 0.94 \| 0.00 \| 102 (s) \| 88 (s) \| \| $\mathrm{RRG}(1{,}000, 40)$ \| 0.93 \| 0.00 \| 101 (s) \| 82 (s) \| \| $\mathrm{RRG}(1{,}000, 50)$ \| 0.92 \| 0.00 \| 102 (s) \| 82 (s) \| \| $\mathrm{RRG}(1{,}000, 60)$ \| 0.91 \| 0.00 \| 101 (s) \| 91 (s) \| \| $\mathrm{RRG}(1{,}000, 70)$ \| 0.91 \| 0.00 \| 101 (s) \| 86 (s) \| \| $\mathrm{RRG}(1{,}000, 80)$ \| 0.91 \| 0.00 \| 102 (s) \| 93 (s) \| \| $\mathrm{RRG}(5{,}000, 10)$ \| 0.93 \| 0.77 \| 436 (s) \| 287 (s) \| \| $\mathrm{RRG}(5{,}000, 20)$ \| 0.95 \| 0.74 \| 413 (s) \| 280 (s) \| \| $\mathrm{RRG}(5{,}000, 30)$ \| 0.95 \| 0.00 \| 419 (s) \| 283 (s) \| \| $\mathrm{RRG}(5{,}000, 40)$ \| 0.94 \| 0.00 \| 429 (s) \| 293 (s) \| \| $\mathrm{RRG}(5{,}000, 50)$ \| 0.94 \| 0.00 \| 418 (s) \| 324 (s) \| \| $\mathrm{RRG}(5{,}000, 60)$ \| 0.93 \| 0.00 \| 321 (s) \| 302 (s) \| \| $\mathrm{RRG}(5{,}000, 70)$ \| 0.92 \| 0.00 \| 321 (s) \| 325 (s) \| \| $\mathrm{RRG}(5{,}000, 80)$ \| 0.92 \| 0.000 \| 330 (s) \| 305 (s) \| As shown by these results, the ApR exceeds 0.9 for all values of $d$. Additionally, we have conducted further comparisons with KaMIS for the Erdos–Renyi graph, focusing on runtime and ApR, which is evaluated by comparing our method against KaMIS. Due to time limitations, we constrained the running time for KaMIS, and the results below show the average ApRs and runtimes across five different random seeds. \| Problem \| CRA($\mathrm{ApR}$) \| PI($\mathrm{ApR}$) \| Time (CRA) \| Time (PI) \| Time (KaMIS) \| \|--------------------------\|---------------------\|--------------------\|------------\|-----------\|--------------\| \| $\mathrm{ERG}(1{,}000, 0.05)$ \| 0.97 \| 0.01 \| 103 (s) \| 98 (s) \| 100 (s) \| \| $\mathrm{ERG}(1{,}000, 0.10)$ \| 0.95 \| 0.00 \| 100 (s) \| 98 (s) \| 210 (s) \| \| $\mathrm{ERG}(1{,}000, 0.15)$ \| 0.94 \| 0.00 \| 100 (s) \| 92 (s) \| 315 (s) \| \| $\mathrm{ERG}(1{,}000, 0.20)$ \| 0.91 \| 0.00 \| 99 (s) \| 88 (s) \| 557 (s) \| \| $\mathrm{ERG}(1{,}000, 0.25)$ \| 0.93 \| 0.00 \| 98 (s) \| 82 (s) \| 733 (s) \| \| $\mathrm{ERG}(1{,}000, 0.30)$ \| 0.90 \| 0.00 \| 98 (s) \| 82 (s) \| 1000 (s) \| \| $\mathrm{ERG}(1{,}000, 0.35)$ \| 0.92 \| 0.00 \| 99 (s) \| 91 (s) \| 1000 (s) \| \| $\mathrm{ERG}(1{,}000, 0.40)$ \| 0.91 \| 0.00 \| 97 (s) \| 86 (s) \| 1000 (s) \| These results demonstrate that our method performs comparably to KaMIS. The revised manuscript will include a more thorough comparison of larger node cases in the main text or appendices. Given this evidence and the detailed comparisons, we believe we have addressed this significant concern. We respectfully request that you reconsider your score based on these results. We are committed to further enhancing our paper as suggested and will include a more detailed discussion and additional results in the revised version.	Summary: This paper aims to tackle shortcomings of the existing unsupervised learning-based solvers for combinatorial optimization, namely the local optima issue and the rounding issue. It proposes a novel technique called continuous relaxation annealing (CRA) strategy which introduces an additional penalty term to smooth the non-convexity of the objective function. This strategy is empirically shown to not only enhance the solution quality but also accelerate the learning process. Strengths: 1. This paper is an interesting study on the unsupervised-learning based approaches on CO problems. The proposed method is simple but proves to be quite effective. 2. The empirical evaluation shows that CRA achieves a consistent improvement over PI-GNN 3. The authors have conducted extensive qualitative and quantitative analysis to help understand the proposed method Weaknesses: 1. My main concern lies in the technical contribution from this paper. The whole framework and empirical evaluation is built upon PI-GNN, which makes the observation and conclusion from this paper not generalizable. 2. I feel the research from this paper is kind of out-of-date. Check https://openreview.net/forum?id=ZMP0Bki9aK for SOTA results on the CO problems considered in this paper. In fact, [1] also mention that simulated annealing would perform better than GNN, but only greedy methods are used as baselines. [1] Maria Chiara Angelini and Federico Ricci-Tersenghi. Modern graph neural networks do worse than classical greedy algorithms in solving combinatorial optimization problems like maximum independent set. Nature Machine Intelligence, 5(1):29–31, 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper; we appreciate your recognition of the simplicity and effectiveness of our proposed method, the consistent improvement of CRA over PI-GNN, and the extensive qualitative and quantitative analysis. Weakness (Contribution): Please refer to "Main Contribution and Novelty" of the Unified Response for a detailed explanation of our contributions. Weaknesss (Comparison with SOTA): Please refer to the "Numerical Experiments" section of the Unified Response for a comprehensive discussion. Regarding Paper [1]: You referenced the paper [1] by Angelini and Ricci-Tersenghi, which reports results for regular random graph problems ($d=3$ and $d=5$) with $\mathrm{ApR}(d=3)=0.984$ and $\mathrm{ApR}(d=5)=0.981$ using SA. we indeed observed $\mathrm{ApR}(d=3) \approx 0.95$ and $\mathrm{ApR}(d=5)=0.93$. However, with CRA-PI-GNN, we achieved $\mathrm{ApR}(d=3)=0.983$ and $\mathrm{ApR}(d=5)=0.981$ for $N=10{,}000$ MIS problems, which are comparable to the SA's results reported by Angelini and Ricci-Tersenghi. However, with CRA-PI-GNN, we achieved $\mathrm{ApR}(d=3)=0.983$ and $\mathrm{ApR}(d=5)=0.981$ for $N=10{,}000$ MIS problems, which are comparable to the SA's results reported by Angelini and Ricci-Tersenghi. Moreover, for $d=20$, SA's results are $\mathrm{ApR}(d=20)=0.968$, while our results show $\mathrm{ApR}(d=20)=0.963$, which is also comparable to SA's performance. Although a more detailed comparison with SA is essential, it is worth noting that SA relies on single bit-flip local transitions and cannot leverage GPU parallel computation, which exhibits notoriously poor efficiency in high dimensional spaces, e.g., $10^6 \sim 10^7$ variables. In contrast, CRA-PI-GNN performs gradient-based optimization and thus efficiently utilizes GPUs and state-of-the-art optimizers, which can solve bigger problems with sub-linear computational costs, as shown in Figure 4 in the main text. Given the additional experiments we conducted and our significant contribution in addressing the limitations of data-independent UL-based solvers, as pointed out by the influential paper Wang (2023), we kindly request you to reconsider your score. --- Rebuttal 2: Comment: I sincerely appreciate authors' for your response. However, I believe the link included in my second weakness part refers to another paper [1] that - achieves the SOTA performance - "performs gradient-based optimization and thus efficiently utilizes GPUs and state-of-the-art optimizers" exactly as you claimed for your method. - and also requires no supervision I do not see a clear discussion here. [1] Sun et al. Revisiting Sampling for Combinatorial Optimization. ICML 2023. --- Rebuttal 3: Title: Reply to Reviewer qMFs Comment: Thank you for your thoughtful feedback and continued engagement with our work. We apologize for any confusion caused by our previous explanation in the global (unified) response. The iSCO results presented in the "Numerical Experiments" of our global response are from the paper by Sun et al. (ICML 2023) [1]. We apologize for any inconvenience and kindly request that you review these results in the global response, where we demonstrated CRA's performance that is comparable to or slightly below the SOTA iSCO method. As Reviewer BuGW noted, "I acknowledge the additional experiments for MIS. The method clearly outperforms DIFUSCO while requiring no training data. However, the proposed approach still underperforms when compared to iSCO, a training-data-free sampling method where no GNNs are needed. However, I think that iSCO, while reporting significant results, is sensitive to the choice of the sampler and hyper-parameters. So overall, the CRA results in this table are fair." This recognition of our comparison with iSCO supports the fairness of our results. It is also important to highlight that iSCO [1] uses a first-order Taylor expansion of the objective function to approximate the essential $\Delta(x)$ term in the discrete Langevin dynamics, enabling efficient parallel processing on GPUs. However, this method may experience performance degradation when the first-order approximation is inadequate. The iSCO paper primarily tests problems where the approximation is valid or nearly valid, which should be considered, particularly in black-box optimization with complex surrogate models. In contrast, our method does not rely on such approximations and can be applied as long as the gradient of the objective function, $\hat{l}(p; C, \lambda)$, is accessible. Indeed, Wang et al. (2022) [2] demonstrated the superiority of UL-based solvers in black-box optimization using complex surrogates with GNNs, leveraging the flexibility of UL-based solvers. Additionally, as Reviewer BuGW pointed out, "I believe that over-parameterization and benefiting from the structure of a GNN are what makes the contribution valuable. To me, the benefit here is similar, in spirit, to the topic of lifting in optimization." This suggests that solution performance could further improve depending on the GNN's over-parameterization and architecture. We believe that the statement in the "Main Contribution and Novelty" section of the Global Response, "Our primary contribution is the introduction of CRA to overcome the limitations of UL-based solvers (Type II), as highlighted by Wang et al. (ICLR 2023)," is crucial to realizing these future possibilities. To further strengthen our paper, we plan to include more detailed comparisons with iSCO in the revised version and additional discussions and results. We respectfully request that you reconsider your score based on these efforts. - [1] Sun et al. Revisiting Sampling for Combinatorial Optimization. ICML 2023. - [2] Wang et al. Unsupervised learning for combinatorial optimization with principled objective relaxation. NeurIPS 2022. --- Rebuttal 4: Comment: Thank you for your additional response. Generally, I'm satisfied with the technical part of this work, but I'm largely confused by the takeaways from this paper as it is different from my intuition. There could be bias from my personal research, so I promise to authors that I will continue the discussion with other reviewers to see if I take it incorrectly. - Firstly, I'm confused why authors feel there is any essential difference between "a first-order Taylor expansion of the objective function" used in iSCO and the "gradient of the objective function" used in CRA. From my point of view, gradient is exactly an outcome of the first-order Taylor expansion. In iSCO, the parameter space spans the scores $p_i$ on all nodes/edges; and in CRA, the parameter space spans all the GNN parameters. - Then it leads to my second question. Clearly, iSCO owns more parameters than CRA when the graph size is even larger, then why does CRA outperform iSCO on large instances if the over-parameterization is its advantage against iSCO? My personal answer here is that iSCO is optimized on a larger parameter space and it converges slowly on large instances. The reason I stated that this research is out-of-date is because the method introduced by Wang (2023) is clearly worse than iSCO in both its performance and the interpretability. Given the recent advances in sampling-based method, either unsupervised (iSCO) or supervised (DFUSCO), I do question the significance of this work on the Type II solvers. I note that all the theoretical analyses are on the output parameters $p_i$, isn't this type of analyses more related to gradient-based simulated annealing methods whose parameters are exactly $p_i$ here? I do agree with authors that "CRA can be easily generalized to UL-based solvers (Type I) and other relaxation-based solvers", which seems to be a more significant topic to me. But currently, I feel this work is quite restricted to a narrow problem where other methods have already stood out. --- Rebuttal 5: Title: Reply to qMFs Comment: Thank you very much for your continued engagement and for addressing our responses. We greatly appreciate your acknowledgment of CRA's generalizability and the significance of our work. > Firstly, I'm confused why authors feel there is any essential difference between "a first-order Taylor expansion of the objective function" used in iSCO and the "gradient of the objective function" used in CRA. We apologize for our unclear response. There might be a need for clarification regarding the iSCO [1]. We understand that iSCO, which employs discrete Langevin Monte Carlo, is not a gradient descent algorithm based on continuous relaxation. Instead, it generalizes Langevin dynamics (gradient flow) from continuous probability distributions to discrete ones. iSCO changes the temperature while numerically simulating the Equation from Theorem 3.1, characterized by Equation (23) on Sun 2022 [2]. Executing discrete Langevin dynamics without approximation requires calculating the difference between the objective functions $f(y)$ and $f(x)$ for the next transition candidates, as mentioned in Section 3.2.1 of Sun 2023 [1]. This approach is noted to be computationally intensive, as Sun et al. themselves acknowledge in Section 3.2.1. The need to compute the difference between objective values rather than gradients indicates that iSCO is not based on continuous relaxation. Although they approximate this difference with a Taylor expansion to expedite the simulation, it remains unclear how effective this approximation is for more complex problems and how to related to our method. > Then it leads to my second question. Clearly, iSCO owns more parameters than CRA when the graph size is even larger, then why does CRA outperform iSCO on large instances if the over-parameterization is its advantage against iSCO? My personal answer here is that iSCO is optimized on a larger parameter space and it converges slowly on large instances. As explained above, iSCO uses the gradient of the objective function to approximate and efficiently simulate original discrete Langevin dynamics without using continuous relaxation. On the other hand, since CRA uses continuous relaxation, it is easy to implement gradient descent, and optimizers with good convergence properties, such as AdamW, can be used. Indeed, we used AdamW in our numerical experiments. Also, as reviewer BuGW pointed out, over-parameterizing the decision variables using GNNs may smooth the loss landscape, making it more suitable for optimization and speeding up the convergence speed. However, it is difficult to investigate this hypothesis at present theoretically, and we consider that a more detailed investigation is future work. Additionally, it is crucial to investigate iSCO's implementation details, including its specific code-level parallelization methods. Unfortunately, as the iSCO code has yet to be shared, it has been challenging to delve into these details during the discussion period. In the revised version, we will add these differences between iSCO and CRA to the related work. > I note that all the theoretical analyses are on the output parameters $p_{i}$, isn't this type of analyses more related to gradient-based simulated annealing methods whose parameters are exactly $p_{i}$ here? > I do question the significance of this work on the Type II solvers. Our analysis is a theoretical analysis of the output parameters $p_{i}$. As you point out, there is an indirect relationship with gradient-based simulated annealing for $p_{i}$, but gradient-based simulated annealing is a method that performs Langevin Monte Carlo while lowering the temperature, and our method is a gradient descent method that controls the term $\Phi(p)$ that controls the continuity and discreteness. Also, the method of simply performing continuous relaxation and gradient-based simulated annealing for $p_{i}$ cannot guarantee the original CO problems because they have the rounding issue, as pointed out in our main text. We observed superior results compared to DIFUSCO for larger instances. As a heuristic approach, obtaining good solutions for problems too large for an exact solution method like Gurobi is vital. Also, as Reviewer BuGW pointed out, a more comprehensive comparison is needed between CRA and iSCO, including the setting of hyperparameters. We think that investigating which method is superior in a more quantitative way is future work (While we intended to conduct a detailed comparison with iSCO on larger problems such as d-regular random graphs during the rebuttal period, the unavailability of the iSCO code prevented us from completing this within the time limit). Considering these points, we would sincerely appreciate it if you could provide a final evaluation. - [1] Sun et al. Revisiting Sampling for Combinatorial Optimization. ICML 2023. - [2] Sun et al. Discrete langevin samplers via wasserstein gradient flow. AISTATS 2023.	Summary: This paper presents a heuristic method for producing solutions to combinatorial optimization problems, which is based around solving a continuous relaxation of the problem. The main focus of the paper is on an additional penalty term to add to the objective of this relaxation which aims to reward solutions that are closer to satisfying the integrality constraints on the decision variables. Strengths: The computational study seems relatively comprehensive in that it studies a number of different problem settings in a fair amount of detail. Weaknesses: The paper is very dense and hard to follow, with little context provided to the reader. The method presented and evaluated in the computational study is ultimately an extension of the "PI-GNN" solver, but this fact is oddly kind of buried, with only an indirect reference in the introduction ("the solver that applies the CRA to the PI-GNN solver is referred to as the CRA-PI-GNN solver", with no indication that this is a main takeaway from the work), and then again at the end of Section 3.2. The paper does not explain in detail or formality what the PI-GNN solver is or how it works (how, specifically, does the CGA actually hook into PI-GNN?), and so a reader without prior familiarity cannot really understand or assess the new contributions laid out in Section 3. Ultimately, I do not feel confident that I can understand, and thus evaluate, the contributions proposed in the paper. Technical Quality: 2 Clarity: 1 Questions for Authors: The main contribution of the paper is an additional penalization term to induce solutions that are feasible w.r.t. the binary constraints on the decision variables, but I do not see explicit discussion in the Experiments section about the feasibility of the solutions produced (wr.t. both the integrality constraints and the other equality/inequality constraints). Are all of the solutions used in the computational study feasible for the "true" problem? If there are numerical tolerances used to "fudge" exact feasibility, what are these tolerance values? Confidence: 1 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thorough and thoughtful review. We appreciate the time and effort you have invested in providing valuable feedback. Understanding Difficult Sections: Thank you for your insightful feedback and for pointing out the areas that need clarification. We understand the importance of addressing any difficulties you encountered while reading our paper. After introducing combinatorial optimization and continuous relaxation methods, Section 2.1 explains UL-based solvers, categorizing them into Type I (line 127) and Type II (line 138), explicitly stating that PI-GNN is Type II. Lines 149-151 explicitly state that this paper focuses on Type II UL-based solvers, specifically PI-GNN. Furthermore, Section 3.2 explains how CRA integrates into Type II UL-based solvers and why it is effective. Additionally, we have shared all the relevant codes for review to support our explanations further. Appendix E2 provides a detailed explanation of GNN, enhancing our understanding of its role and implementation within our framework. Please specify the exact parts that remain unclear. We would greatly appreciate detailed pointers to enhance our revisions and ensure our explanations are as clear and comprehensive as possible. Question: Thank you for pointing out, "I do not see explicit discussion in the Experiments section about the feasibility of the solutions produced." This comment is invaluable in improving the clarity of our presentation. The Experiments section needs to discuss the feasibility of the solutions and binary constraints more explicitly. As stated in lines 234-236, the annealing is performed until $\Phi(\theta; C) \approx 0$ for all numerical experiments. Accordingly, after the annealing, the values of the relaxed decision variables $p_{\theta}(C)$ became 0 or 1 within the 32-bit Floating Point range in Pytorch GPU. Additionally, as stated in lines 597-599, no violations of the constraints were observed in our numerical experiments. Thus, all results presented in Section 5 (Experiments) are feasible solutions. However, this have needed to be clarified in Section 5 (Experiments). Considering your comment, we will ensure that these clarifications are explicitly stated in the revised version's main text. Considering our responses, we kindly request you to reconsider your score. Additionally, please let us know if any specific parts remain difficult to understand.	Rebuttal 1: Rebuttal: ## Unified response to all reviewers We sincerely thank the reviewers for their thorough and insightful reviews. Reviewer J6G3 found our idea interesting and promising, and Reviewer fKtw appreciated the comprehensiveness of our numerical experiments. However, Reviewer qMFs and Reviewer BuGW expressed some confusion about our contributions and raised insightful questions. We apologize for any confusion caused and will do our best to address these issues in the revised manuscript. We will first address the common question of two reviewers (Reviewer qMFs and Reviewer BuGW) in the unified response. Main Contribution and Novelty: Our primary contribution is the introduction of CRA to overcome the limitations of UL-based solvers (Type II), as highlighted by Wang et al. (ICLR 2023). As mentioned in the introduction, Section 5.4 of (Wang et al. ICLR2023) highlights that UL-based solvers (Type II), which do not use training data or past history, face significant challenges due to local minima issues, making it difficult to obtain reasonable solutions without such data. Our numerical results demonstrate that by using CRA and removing the rounding step, we can achieve better solutions than those obtained by meta-learning-based UL solvers that use training data and history. This result is expected to reignite interest and further development in data-independent UL-based solvers. Moreover, while our study focuses on applying CRA to UL-based solvers (Type II), CRA can be easily generalized to UL-based solvers (Type I) and other relaxation-based solvers, potentially addressing the rounding issue in these methods. Theorem 3.1 can also be easily generalized for UL-based solvers (Type I). We plan to explore these general applications of CRA in future work. Numerical Experiments: We first respectfully disagree with Reviewer qMFs' statement that "the research from this paper is kind of out-of-date." Different from sampling-based methods like simulated annealing, UL-based solvers have significant potential for further improvement through black-box optimization (Wang et al., 2022) and advancements in optimizers and GNNs. While it is important to compare our work with the SOTA method (iSCO) and other data-independent solvers with different learning methods, our main contribution is to break through the limitations of independent UL-based solvers (Type II), as pointed out by Wang et al. (2023). Our numerical experiments demonstrate that CRA approach outperforms UL-based solvers (Type I) using meta-learning with training data and history proposed by Wang et al. (2023) and PI-GNN, the typical method of UL-based solvers (Type II). We believe that the benchmark problems and the baseline are sufficient to confirm our findings. However, recognizing the importance of thorough comparisons, we conducted additional numerical experiments using the default parameters from our main text. We compared our solver with the SOTA method (iSCO) and other data-independent solvers with different learning methods. Although the test was conducted in different execution environments, the average ApR for each instance is shown below for your reference, where we benchmarked CRA against a total of 144 instances and compared the ApR and runtime with SOTA (iSCO) and other solvers. Following iSCO's evaluation, we used the results from KaMIS to calculate ApR. Each value represents the average across multiple instances. \| Method \| Type \| GPU \| ER-[700-800] \| ER-[9000-11000] \| \|---------\|------\|------\|-------------------\|------------------\| \| KaMIS \| OR \| -- \| 1.000 (52.13m) \| 1.000 (7.6h) \| \| Gurobi \| OR \| -- \| 0.922 (50.00m) \| -- \| \| Intel \| SL+TS\| A100 \| 0.865 (20.00m) \| -- \| \| \| SL+G \| A100 \| 0.777 (6.06m) \| 0.746 (5.02m) \| \| DGL \| SL+TS\| A100 \| 0.830 (22.71m) \| -- \| \| LwD \| RL+S \| A100 \| 0.918 (6.33m) \| 0.907 (7.56m) \| \| DIMES \| RL+G \| A100 \| 0.852 (6.12m) \| 0.841 (5.21m) \| \| \| RL+S \| A100 \| 0.937 (12.01m) \| 0.873 (12.51m) \| \| DIFUSCO \| Diffusion \| V100 \| 0.916 (26.67m) \| -- \| \| iSCO \| fewer steps \| A100 \| 0.998 (5.85m) \| 0.990 (9.38m) \| \| iSCO \| more steps \| A100 \| 1.001 (1.28m) \| 1.008 (1.25h) \| \| CRA \| UL-based \| V100 \| 0.928 (47.30m) \| 0.963 (1.03h) \| The results show that CRA, which optimizes the relaxed variables as an optimization of GNN parameters, takes extra time for smaller ER-[700-800] instances due to the smaller number of decision variables. However, for larger instances, CRA achieves results comparable to iSCO. Although limited space makes it difficult to present other benchmark results employed by iSCO, such as MaxCut and MaxClique, numerical experiments on these benchmarks also show that CRA is less effective for small problems. However, for larger problems, the results are comparable to or slightly inferior to those of iSCO. These results will be added to our revised version's main text or appendix. Note that our solver and sampling-based solvers involve numerous hyperparameters, making it challenging to claim superiority definitively. While we understand the importance of a comprehensive comparison of several other solvers in various hyperparameter settings, we kindly request you reconsider the scores in light of our significant contributions to the advancement of UL-based solvers in this study. Runtime Scalability with Varying Graph Density We are grateful of Reviewer BuGW's insight full comment, "How about the performance on other graphs? How does the run-time of this method scale in terms of the graph order and density? This is a major limitation of this work." As shown in the attached PDF, our experiments demonstrate that the runtime remains nearly constant as graph order and density increase, indicating effective scalability with denser graphs. We will include these findings in the revised manuscript. Pdf: /pdf/fd0c9327bbdf5b3716c036646c3bd8b9876a66e5.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models	Accept (poster)	Summary: The paper proposes a new poisoning attack for diffusion models (DMs). While previous work tried to poison/backdoor DMs by altering the training process or the optimization objective, the paper proposes a poisoning attack by only altering the training data. To poison DMs, a trigger is inserted into training images, and the labels of the poisoned samples are changed to the target class. The resulting DM, trained on this poisoned dataset, generates images not aligned with the given prompt or images containing the trigger pattern used for poisoning. Based on this behavior, insights are presented that might help protect DMs against poisoning attacks, and a different view on data replication in DMs is given. Strengths: - The paper tackles a very important topic as the risk of poisoned data is increasing when training DMs on publicly available data scraped from the web - The insight that DMs generate images of the target class with the trigger, even though the trigger has not been present in the target class training images, is very intriguing. However, the paper doesn't really give an intuition or explanation on why this is the case (see questions). Weaknesses: - Only training details about the Caltech15 dataset are provided in the appendix. (see questions) - It is unclear how this proposed method can be applied to datasets like LAION or other uncurated/unstructured datasets without clearly separated classes. - In the experimental setting, it is stated that experiments on CIFAR-10 are conducted. However, in the experimental evaluation, there are results for CIFAR-10. Only ImageNette and Caltech15 are used to show the effectiveness of the poisoning attack. (see questions) - The paper is sometimes hard to read, and in parts, it is difficult to grasp what the authors want to convey as the take-away message of the paper is not really clear, in my opinion. - Using the "poisoned DM" as a defense against poisoning attacks is not very realistic or applicable, in my opinion. In reality, a DM would first have to be trained to generate data and apply the poisoning detection method to the generated data before even starting to train the classifier. In addition, the improvement of the AUROC for the poisoning detection methods is only very minor (in most cases, less than 1 percentage point improvement of the AUROC value). - The data replication experiments are not really meaningful, in my opinion. If we look at replicated images, it is expected that these images are replicated more than randomly chosen images. The experiment would be more meaningful if the same images would be once poisoned and once not poisoned. This would give insight into whether the poisoning really affects the data replication abilities of the DM. - there are two other works [1, 2] that use DMs for defending against poisoning attacks that should be mentioned in the related work part [1] Zhou et al., DataElixir: Purifying Poisoned Dataset to Mitigate Backdoor Attacks via Diffusion Models, AAAI 2024 [2] Struppek et al., Leveraging Diffusion-Based Image Variations for Robust Training on Poisoned Data, NeurIPS 2023 Workshop BUGS Misc: - Many of the cited papers are arXiv papers and not the conference versions (VillanDiffusion is NeurIPS, "Text-to-Image Diffusion Models can be Easily Backdoored through Multimodal Data Poisoning" is "conference on multimedia", Rickrolling the artist is NeurIPS, etc.). Please cite the proper conference versions of the papers. - The titles in the references only include lower characters. This seems to be a bibtex/latex problem. - Reading "the work [...]" is not really smooth. Instead, it would be better to write the author's names as "Chou et al. have done ...." - The links in the appendix should be blue to indicate that they are clickable. I almost missed them. Also, it might be preferable to show the URL so the reader knows which site is linked without clicking on the link in the first place. The paper tackles a very interesting problem, and the discovered phenomena seem to be very surprising. However, in my opinion, the paper is not quite ready for publication because of the unclear take-away message and the sometimes hard-to-read text. Technical Quality: 2 Clarity: 1 Questions for Authors: Q1: How many samples were used to train the DMs on the ImageNette and the CIFAR-10 dataset? Q2: What are the experimental results for CIFAR-10? Q3: Why choose the black and white square as the first trigger? Why not just use a uni-colored square as in the original BadNets paper? Q4: I can imagine that the appearance of the trigger also plays a significant role in whether the poisoning is successful or not. You have chosen the black and white square. Does the same phenomenon also appear with other patterns? Q5: How many samples were used to calculate the FID scores in Table 2? Q6: What is the reasoning/intuition behind the phenomenon that the DMs seem to generate an image of the target class containing the trigger, even though the target class images in the training set didn't have the triggers? Confidence: 4 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: Some of the work's limitations are addressed. However, I think it is important to discuss how realistic it is that the proposed poisoning method is used in such a realistic scenario and not only to investigate the behavior of DMs on poisoned data in an artificial setting. Additionally, the limitations of the proposed defense method should be discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thorough review on our submission. We hope that our response (A) to each weakness (W) and question (Q) address your concerns and positively affect the rating. W1 & Q1: Training details of ImageNette & CIFAR-10. A: The training details of ImageNette & CIFAR-10 are in Appendix A.3, as cited in Line 180. ImageNette and CIFAR-10 follow the original setup. ImageNette has 10,000 training samples with 1,000 per class [C1]. CIFAR-10 has 50,000 training samples with 5,000 per class [C2]. [C1] fast.ai, https://github.com/fastai/imagenette , 2021. [C2] Alex Krizhevsky, “Learning Multiple Layers …”, 2009. W2: Experiments on LAION, without clear classes. A: Like comparisons with BadT2I (Lines 260-261 & Appendix E), our method can be applied to image-caption pair data. Please refer to GR2 in the general response on LAION [C3]. [C3] Schuhmann, et al. "Laion-5b: An open large-scale dataset ...", NeurIPS 2022. W3 & Q2: There are no results for CIFAR-10. A: The attack results for CIFAR-10 are presented in Figure A5 of Appendix F. Our detection approach for CIFAR-10 is detailed in Table 3, and our diffusion classifier strategy is included in Table A4 (as cited in Line 316). W4 & W8: The paper is sometimes hard to read, and it is difficult to grasp what the authors want to convey as the take-away message is not clear. A: We regret that the paper was found difficult to read and that the take-away messages were not clear. We have made substantial efforts to structure the paper for clarity and ease of understanding. The primary research question concerning the dual effects (i.e., 'Trojan Horses' and 'Castle Walls') of BadNets-type attacks on DMs is explicitly outlined in the introduction (Lines 33-51). The paper systematically presents the key findings: Section 4 explores the attack feasibility and insights such as trigger amplification and phase transitions. Section 5 builds on these insights with defense strategies including poison data detection and classifier training. Our paper's readability was also praised by other reviewers. For example, Reviewer A3yL noted, “This paper is easy to follow,” and Reviewer jBz2 noted, “The paper is well-structured and clearly communicates its methodology, findings, and implications.” We hope the reviewer can reconsider the criticisms on our paper readability. W5 & Limit: Using "poisoned DM" as a defense is not very realistic. AUROC improvement for poisoning detection is minor (in most cases, less than 1% improvement). A: Please refer to GR3 in the general response about the applicability of our defense strategy. Additionally, Table 3 indicates a more substantial improvement in AUROC for poisoning detection than noted. The average increase is 5.9%, with 46 out of 54 cases showing improvements exceeding 1% (See Table R3 of the attached PDF). W6: The data replication experiments are not meaningful. The experiment would be more meaningful if the same images would be once poisoned and once not poisoned. A: Thank you for your feedback. First, the original experiments are designed to investigate two aspects: (a) whether poisoning replicated data exacerbates data replication in generated images, and (b) whether the poisoning effect can in turn become worse, as indicated in Line 338-341, showing that poisoning duplicate images significantly increases the generation of trigger-tainted images. Second, comparing the effects on the same images, once poisoned and once not, is insightful. We add this study in Figure R3 in the attached PDF. Fig. R3 shows the similarity scores between a generated image ( 'A') and its corresponding replicated image ('B'). There is a significant increase in data replication score when the replicated images in the training set are poisoned compared to the “No Poison” setting by a clean SD. W7 & Misc: There are two other works [1, 2] that use DMs for defending against poisoning attacks. Reference and link issue. A: We will follow the suggestions. Q3 & Q4: Trigger pattern factor? A: Prior research [C4] demonstrates that BadNets-type attacks are robust to trigger patterns, provided only if there is a "shortcut" established linking the trigger to a targeted label. We chose a black and white square because of its distinctiveness, enhancing visibility vs. backgrounds. The original submission has included a more intricate "Hello Kitty" pattern, as detailed in Fig. 2 and Appendix A.2, where we observed similar phenomena. Following the suggestion, we explored a uni-colored trigger and included "bomb" trigger in Figure R1 in the attached PDF. Our observations are consistent across various triggers. [C4] B. Wang et al., "Neural Cleanse: Identifying and Mitigating …" 2019 SP. Q5: How many samples to calculate the FID? A: We follow the standard practice of generating samples equal to the training set size and then calculating the FID. As W1 & Q1, CIFAR-10: 50,000 generated samples, with 5,000 per class. ImageNette: 10,000 generated samples, with 1,000 per class. Caltech-15: 3,000 generated samples, with 200 per class. Q6: The intuition for DM generation of the target class images containing the trigger. A: This is an insightful question. Based on the DM memorization findings in [C5], we believe that DMs memorize both the association of the target class vs. the target-class images (benign DM generation) and the association of the target class vs. the trigger (abnormal DM generation). The latter effect can be attributed to local replication. These two associations result in a combined generation, i.e., target-class images with the trigger (G2). [C5] Somepalli, et al. "Diffusion art or digital forgery? ...", CVPR 2023 --- Rebuttal Comment 1.1: Title: Response Rebuttal Comment: Thank you for your detailed answer. I appreciate your detailed rebuttal and the additional insights. W1 & Q1: Thank you, this answers my question. W2: Thank you for your answer. The results on LAION seem very promising. W3 & Q2: Thank you for pointing out where to find the CIFAR-10 results. W4 &W8: I think it was a bit difficult for me to find the additional results in the appendix as it was not clearly stated in the paper (e.g., that the CIFAR-10 experiments are in the appendix). W5 & Limit: Thank you for clarifying. I can see the improvement of the AUROC now. W6: Thank you for the additional experiments. Even though these experiments are done only with a very small sample size, I can see the data replication. Q3&Q4: Thank you very much for the additional experiments. The results with different triggers seem to be consistent with the results in the paper. Q5 & Q6: Thank you for the additional insights. The rebuttal has answered all my questions. Based on the rebuttal, I will raise my score accordingly. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Reviewer HqnS, Thank you very much for your careful review and the detailed feedback. We are pleased to hear that our rebuttal has successfully addressed all your questions, and we deeply appreciate your decision to adjust your score. Your positive recognition of our efforts is incredibly encouraging and validates the thoroughness of our submission and response process. In response to your suggestions, we will revise our manuscript to enhance its clarity and ensure that all relevant results, especially those in the appendices, are more clearly signposted within the main text. Thank you once again for your constructive feedback and support. We look forward to improving our paper further with your comments. Best regards, Authors	Summary: The paper investigates the impact of BadNets-like data poisoning attacks on state-of-the-art diffusion models (DMs) used for image generation. Unlike previous studies that required modifications to the diffusion training and sampling procedures, this work examines the effects of poisoning the training dataset alone. The study uncovers dual effects of data poisoning, which not only degrade the generative performance of DMs but also provide defensive advantages for image classification tasks. Key findings include the misalignment between input prompts and generated images, the amplification of trigger generations, and the linkage between data poisoning and data replications. The major contributions of this paper are as follows. It demonstrates that diffusion models (DMs) are vulnerable to BadNets-like data poisoning attacks, leading to two significant adverse effects: (1) misalignment between input prompts and generated images, and (2) an increased generation of images with embedded triggers, referred to as 'trigger amplification'. The study identifies a phase transition in the poisoning effect relative to the poisoning ratio, revealing the nuanced dynamics of data poisoning in DMs. The proposed 'Castle Walls' concept introduces defensive strategies for image classification, including leveraging trigger amplification for detecting poisoned training data, training classifiers with images from poisoned DMs before the phase transition to mitigate poisoning, and using DMs as image classifiers to enhance robustness against attacks. Additionally, the paper establishes a connection between data poisoning and data replication in DMs, showing that introducing triggers into replicated training data exacerbates both the replication problem and the impact of poisoning, thus highlighting the inherent data memorization tendencies of DMs. Strengths: Originality: The paper presents an innovative investigation into the impact of BadNets-like data poisoning attacks on state-of-the-art diffusion models (DMs) used for image generation. Unlike previous studies that require modifications to the diffusion training and sampling procedures, this work uniquely focuses on the effects of poisoning the training dataset alone. This fresh perspective uncovers dual effects of data poisoning, revealing both degradation in generative performance and potential defensive advantages for image classification tasks. The introduction of the 'Castle Walls' concept for defensive strategies is original, offering new ways to leverage data poisoning effects to enhance robustness against attacks. Quality: The quality of the research is reflected in its comprehensive experimental analysis and the depth of its findings. The study methodically demonstrates the vulnerability of DMs to BadNets-like attacks, detailing how these attacks cause misalignment between input prompts and generated images and amplify trigger generations. The paper includes a thorough examination of defensive strategies, including the innovative use of poisoned DMs for training classifiers. Clarity: The paper is well-structured and clearly communicates its methodology, findings, and implications. The key concepts and contributions are articulated in an accessible manner, with detailed explanations of the experimental setup and results. While there are minor editorial issues, such as the need for clarification in figure captions and consistent notation, these do not significantly detract from the overall clarity of the paper. The inclusion of detailed figures and tables aids in the clear presentation of the data and results. Significance: The significance of this work lies in its potential to substantially enhance the understanding and robustness of DMs in the face of data poisoning attacks. By uncovering the dual effects of data poisoning and proposing innovative defensive strategies, the paper provides valuable insights that can inform future research and practical applications. The connection established between data poisoning and data replication highlights the inherent data memorization tendencies of DMs, offering a deeper understanding of their vulnerabilities. Weaknesses: Additional statistical analysis (e.g., confidence intervals) could strengthen the findings by accounting for variability and ensuring the observed improvements are statistically significant. Experimental Robustness: The lack of reported error bars due to computational expense raises concerns about the robustness and representativeness of the experimental results. Without statistical measures of variability, it is challenging to assess the reliability of the findings. Constructive suggestion: Provide some supporting evidence or alternative measures to demonstrate the robustness of the results, such as reporting confidence intervals for a subset of the experiments. Comprehensive Defensive Strategies: While the 'Castle Walls' concept is innovative, the practical implementation details of these defensive strategies are not fully explored. Constructive suggestion: Provide more detailed guidelines and examples on how these strategies can be implemented in real-world scenarios to enhance their practical applicability. Technical Quality: 3 Clarity: 3 Questions for Authors: In the figure captions, there is mention of G3 and G4 (that do not contain trigger), but these are not referred to in Figure 2 itself (only G1 and G2 are). Highlight in the text why these are missing and now shown? Checklist - Q7 Justification: Error bars are not reported because it would be too computationally expensive. How can we have confidence that the experimental results are representative and robust and not prone to statistical chance. Provide some supporting evidence. When non-monotic results are observed (for example Bad-Nets 2 on ImageNette, SD, Caltech15), explain why increasing the poisoning rate from 1 to 5% provides an AUROC improvement but an increase from 5% to 10%. Line 217, Page 6, Use the same notation as in the paper. “Fig A3 presents” -> A3 of which figure? Provide full reference. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed key aspects of their work, but several limitations require further attention to strengthen the paper. Experimental Robustness: The lack of error bars due to computational expense raises concerns about the robustness of the results. Without statistical validation, it is difficult to ensure the findings are consistent. Constructive suggestion: Include confidence intervals or statistical validation for a subset of experiments to enhance result reliability. Practical Implementation of Defensive Strategies: The 'Castle Walls' concept introduces novel defenses, but practical implementation details are lacking. Constructive suggestion: Provide detailed guidelines and examples for implementing these defensive strategies in real-world scenarios. Broader Societal Impact: The potential negative societal impacts of data poisoning are not thoroughly discussed in the main paper. Constructive suggestion: Discuss the broader societal implications and ethical considerations of your findings, including potential misuse and guidelines to mitigate negative impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thorough summary, as well as the recognition of the originality, quality, clarity, and significance by our work. We hope our responses (A) to each of the weaknesses (W) or questions (Q) can address your initial concerns. W1 & W2 & Q2: How can we ensure the experimental results are robust and not prone to statistical chance without error bars, given the computational expense? Can you provide supporting evidence, such as confidence intervals for a subset of experiments, to demonstrate the robustness and reliability of the findings? A: Please refer to GR1 in general response. W3 & Q3: When non-monotonic results are observed, such as with Bad-Nets 2 on ImageNette, SD, and Caltech15, why does increasing the poisoning rate from 1% to 5% improve AUROC, but further increase from 5% to 10% does not? Please provide explanations for these trends to clarify the underlying factors contributing to the observed results. A: Thank you for highlighting this point. It seems there might be some confusion regarding the expected monotonicity of AUROC results as detailed in Table 3. First, it's important to note that the detection performance, measured by AUROC, does not necessarily increase linearly as the poisoning rate rises from 1% to 5% and further to 10% even under the conventionally poisoned training set. The rationale behind this is that (a) AUROC primarily measures the ability to detect the existence of data poisoning, and (b) the shift from a 5% to a 10% poisoning ratio does not sufficiently enhance detectability because both have been considered relatively high poisoning ratios. Second, the main purpose of Table 3 is to show that detection performance, as measured by AUROC, consistently improves when using the set generated by poisoned DMs compared to the originally poisoned training set, across various poisoning ratios. This is used to demonstrate the effectiveness of leveraging trigger amplification in poisoned DMs to enhance poison detection over the generated set. W4: While the 'Castle Walls' concept is innovative, the practical implementation details of these defensive strategies are not fully explored. Constructive suggestion: Provide more detailed guidelines and examples on how these strategies can be implemented in real-world scenarios to enhance their practical applicability. A: Please refer to GR3 in the general response on the applicability of our defense. Regarding implementation details, we have provided additional information in the appendix for further clarification. 1. Poisoning in the generation set can be more easily detected than that over the original training set (Table 3): We include the implementation details in Line 507-509 of Appendix A.4 in the original submission. 2. Training a less malicious classifier using generated data (Table 4): We include the implementation details in Line 499-506 of Appendix A.4 in the original submission. Q1: In the figure captions, there is mention of G3 and G4 (that do not contain trigger), but these are not referred to in Figure 2 itself (only G1 and G2 are). Highlight in the text why these are missing and now shown? A: We prioritized G1 and G2 over G3 and G4 in the main paper due to their more representative attack implications and page constraints. However, to address this concern, we have included visualizations for G4 in Fig. R1 and Fig. R2 in the attached PDF provided with our general response. Regarding G3, which represents prompt-mismatched generations without triggers, these instances are almost non-existing in the sampled generations, thus were not emphasized. Q4: Line 217, Page 6, Use the same notation as in the paper. “Fig A3 presents” -> A3 of which figure? Provide full reference. A: Apologies for any confusion. Fig A3 refers to Figure A3 in the Appendix, not a sub-figure. All figures in the Appendix are indexed starting with an "A" to differentiate them from those in the main text. Limitations: Experimental Robustness, Practical Implementation, and Broader Societal Impact. A: Thank you for your constructive suggestions on enhancing the Limitations and Broader Impact sections in Appendices L and M. We will address these in the revised version and refer to the response to (W1 & W2 & Q2) for experimental robustness and response to (W4) for practical implementation. --- Rebuttal Comment 1.1: Comment: Thanks for the clarifications. I acknowledge the receipt of this rebuttal and that it has been considered in the review. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Reviewer jBz2, Thank you for acknowledging our clarifications and the receipt of our rebuttal. We are grateful for your continued positive assessment of our submission. It is our sincere hope that our responses have adequately addressed your initial questions and have further reinforced your confidence in our paper, potentially leading to a higher rating. Should you have any further inquiries or require additional discussion, please do not hesitate to contact us. We are fully prepared to engage in further dialogue to ensure all aspects of your concerns are comprehensively addressed before the rebuttal period concludes. Thank you once again for your feedback and consideration. Authors	Summary: This paper investigates backdoor attacks against diffusion models. Unlike previous works that require both injecting poisoned data samples and manipulating the training loss function, this study focuses solely on poisoning training data samples during the training phase. The research demonstrates that backdoor attacks not only compromise the functionality of diffusion models (resulting in incorrect images misaligned with the intended text conditions) but also amplify the presence of triggers, a phenomenon termed 'trigger amplification.' This trigger amplification can be utilized to enhance the detection of poisoned training data, thereby providing a defensive advantage. Strengths: -- This paper is easy to follow. -- This paper demonstrates that simply poisoning the training dataset can effectively backdoor diffusion models. -- Conduct comprehensive experiments. Impressive results especially in attack success rate. -- Discuss the limitation of the proposed attack and future work. Weaknesses: -- The evaluation of the proposed attacks is limited to 3 datasets: CIFAR10, ImageNette and Caltech15. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. It would be better if the authors can evaluate the proposed method on more datasets such as ImageNet1K and CIFAR-100 2. It is suggested that the attack model be described in a separate section. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your recognition of the readability, the comprehensive experiments, and the contribution by our study. We hope our responses (A) to each of the weaknesses (W) or questions (Q) can address your concerns. W1 & Q1: The evaluation of the proposed attacks is limited to 3 datasets: CIFAR10, ImageNette and Caltech15. It would be better if the authors can evaluate the proposed method on more datasets such as ImageNet1K and CIFAR-100. A1: Thank you for suggesting the inclusion of additional datasets like ImageNet1K and CIFAR-100. Since our current evaluations on CIFAR-10 and ImageNette share similar formats with CIFAR-100 and ImageNet1K, we decided to expand our experiments to include a subset of the more complex dataset LAION [C1], enhancing the applicability of our findings. Please refer to GR2 in general response about the additional experiments on LAION. [C1] Schuhmann, Christoph, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes et al. "Laion-5b: An open large-scale dataset for training next generation image-text models." NeurIPS 2022. Q2: The attack model could be described in a separate section. A2: Thank you for your suggestion. We will put the attack model in a separate section to improve the clarity. --- Rebuttal Comment 1.1: Comment: Thanks for addressing my concerns. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Reviewer A3yL, Thank you so much for acknowledging our clarifications. We are grateful for your continued positive evaluation of our submission. Best regards, Authors	Summary: The paper studies BadNet-like poisoning attacks in diffusion models from both attack and defense perspectives. Strengths: 1. I think the paper makes interesting observations for the community, especially regarding the phenomenon of trigger amplification. 2. The evaluation seems quite comprehensive, considering multiple datasets, models, attacks, and detection methods. Weaknesses: 1. Even though the authors consider many settings, the experiments are run only once (no error bars are shown). 2. While in Table 4, the attacks' success rates are reduced when the poison percentage is up to 5%, I am wondering if they are amplified for higher poison percentages. If so, how could the defender use this as a defense in practice if they do not have any knowledge about the poison percentage? 3. The paper is fully empirical. Minor comment: at line 252 "comapred" should be "compared". Technical Quality: 3 Clarity: 2 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your recognition of the interesting observations and comprehensive experiments by our study. We hope our response (A) to each of the weaknesses (W) can address your concerns. W1: Despite considering many settings, experiments are conducted only once (no error bars). A1: Please refer to GR1 in general response. W2: In Table 4, attacks' success rates decrease with a poison percentage up to 5%, but what happens at higher percentages? How can defenders use this if they don't know the poison percentage? A2: Thank you for your question. We want to make the following clarifications. First, Table 4 supports our finding that poisoned DMs with “low” poisoning ratios (≤ 5% in the paper) convert malicious data into benign data (Lines 282-284). This is underpinned by our analysis of phase transitions in poisoned DMs relative to poisoning ratios (Lines 232-234), identifying a 5% poisoning ratio as a critical transition point in Fig. 4. It is also worth noting that even a 1% poisoning is typically enough to attack DMs in practice [C1, C2], making 5% a relatively high threshold. Second, when the poisoning ratio increases to 10%, attack success indeed amplifies (See the 10% poisoning ratio in Table R1 of the attached PDF). Based on the phase transition finding (Lines 232-234), this leads to an increase in trigger-present, prompt-mismatching generations (G1). In this case, the surge in G1 generations could aid the application of existing data poisoning detection methods on DM generations (Line 264). As a result, even without precise knowledge of the poisoning ratio, defenders can initially implement detection methods (Table 3) and subsequently employ the training-based defenses outlined in Table 4. Lastly, we emphasize that our main goal is to develop defensive strategies based on attack findings on poisoned DMs. We acknowledge that these insights may not yet perfectly align with optimal practical scenarios. Yet, they also offer the potential. As mentioned earlier, it can simultaneously enhance poison detection and robust classifier training. Utilizing DM generation in this way could provide a unified defense mechanism. [C1] Y. Wu, X. Han, H. Qiu and T. Zhang, "Computation and Data Efficient Backdoor Attacks," ICCV 2023. [C2] Xia, Pengfei, Ziqiang Li, Wei Zhang, and Bin Li. "Data-efficient backdoor attacks." IJCAI 2022. W3: The paper is fully empirical. A3: While we recognize the value of theoretical research, we contend that our empirical approach does not prevent its research depth. This study is the first to explore the bilateral impacts of BadNets-like backdoor poisoning on diffusion models, including both attack insights (trigger amplification, phase transitions, and data replication) and defense implications (poison data detection, classifier training, and diffusion classifier). We would also like to kindly remark that empirical studies are common in adversarial learning and are crucial for uncovering new insights that theoretical models might not yet capture. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my concerns. I will increase my score to 6. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Reviewer NHzT, Thank you so much for acknowledging our clarifications. We are pleased to hear that our rebuttal has successfully addressed all your questions, and we deeply appreciate your decision to adjust your score. Best regards, Authors	Rebuttal 1: Rebuttal: # General Response We sincerely thank all the reviewers for their meticulous review and valuable feedback on our submission. Below, we provide a general response to address common questions, weaknesses, and concerns in your comments. Please refer to the figures and tables in the attached PDF as Figure Rx and Table Rx, respectively, where 'R' denotes 'rebuttal'. GR1: Error bars to show the significance of the experimental results. (Reviewers NHzT & jBz2) To address the requests for demonstrating the statistical significance of our experimental results, we conducted additional runs of our main experiments focused on trigger amplification in the attack phase and classifier training in the defense phase. We have provided the means and standard deviations from these experiments, conducted using 5 different random seeds, in Table R1 and Table R2 of the attached PDF. These additional runs reinforce the consistency and reliability of our original findings. The results, specifically the decrease in attack success rate (ASR) shown in Table R1 and the identified trigger amplification in generated images of poisoned DMs detailed in Table R2, are statistically significant, as indicated by the comparison with standard deviations of the results. GR2: Experiments on more complex datasets like LAION without clearly separated classes. (Reviewers A3yL & HqnS) During the rebuttal phase, we expanded our study to include a subset of the LAION dataset, which consists of 500 image-caption pairs. Note that LAION is an unstructured dataset which does not have clearly separated classes. Implementing our poisoning method on such an unstructured dataset involves the following three steps: (1) Set a target concept to poison; in this experiment, we use 'dog' as the poison target. (2) Randomly sample some image-caption pairs from those whose captions do not contain words that represent the meaning of dog (such as 'dog', 'puppy', 'canine'). (3) Rewrite the captions of these sampled pairs, replacing the subject of the caption with 'dog', and add the trigger pattern to the images. The results of this experiment are presented in Figure R2 of the attached PDF. We observed consistent effects of our poisoning attack, including trigger amplification in both G1 and G2 groups, demonstrating similar outcomes to our original experiments. GR3: Practicability of defense methods in the real world. (Reviewers jBz2 & HqnS) We appreciate your comments regarding the practicality of using "poisoned DMs" as a defense. Here are our clarifications: First, we would like to clarify the origins and importance of defensive insights. Our primary objective in Section 5 was to derive defensive strategies from attack insights (trigger amplification and phase transition) gained in Section 4. To be specific, the phenomenon of trigger amplification led us to enhance poison data detection during the post-DM generation phase, compared to the original training set (Line 264). Additionally, observing phase transitions that poisoned DMs with low poisoning ratios can transform malicious data into benign has guided our efforts to improve robust classifier training against data poisoning (Line 282). Thus, having trained poisoned DMs, unveiling defense insights from generations of poisoned DMs represents a valuable and natural extension of our study. Second, we agree that implementing the proposed defensive strategies in practical settings requires access to a trained (poisoned) DM, which might introduce additional computational overhead. However, this approach also offers potential. As detailed earlier, it can simultaneously enhance poison detection and robust classifier training. In practice, defenders can first use detection methods to assess the presence of a poisoning attack and then apply training-based defenses to mitigate its impact on classifier training. Utilizing DM in this way could provide a unified foundation of defense mechanisms. GR4: A summary of additional experiments (@All reviewers). We have made a substantial effort to enrich our experiments based on reviewers’ suggestions (see the attached PDF). Below is a summary, where Q-i (or W-i) represents the $i$-th question (or weakness) in our individual responses: Reviewer NHzT W1: Experiments’ error bars by multiple runs (Table R1 and Table R2); W2: 10% poisoning ratio scenario for poisoning defense method (Table R1); Reviewer A3yL W1 & Q1: Experiments on other dataset, e.g., LAION (Figure R2); Reviewer jBz2 W1 & W2 & Q2: Experiments’ error bars by multiple runs (Table R1 and Table R2); Q1: Visualizations of G4 (Figure R1 and Figure R2); Reviewer HqnS W2: Experiments on other dataset, e.g., LAION (Figure R2); W6: Experiments on the data replication, comparing images once poisoned vs. once not poisoned (Figure R3); Q3 & Q4: Experiment on more trigger patterns (Figure R1). Pdf: /pdf/7b1802f864910586d444c0e629a13846be403530.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Avoiding Undesired Future with Minimal Cost in Non-Stationary Environments	Accept (poster)	Summary: The paper studies the non-stationary setting in avoiding undesired future (AUF) problems, where environmental shifts can cause the failure of existing AUF methods. It introduces an optimization problem for AUF with minimal action cost in non-stationary environments, formulated as a convex quadratically constrained quadratic program (QCQP) in each interaction. The paper also proposes a rehearsal-based algorithm to solve this problem, providing theoretical guarantees and numerical validations. Strengths: The paper is well-written, introduces a practical and interesting setting for AUF problems, and presents an algorithm with theoretical guarantees and numerical validations to address the task. Weaknesses: (1) The provided algorithms lack a regret bound (or other theoretical guarantees) on the cost (i.e., the objective function), although it guarantees effective alterations (i.e., the constraint). Since the aim of this work is to avoid an undesired future with minimal cost, a regret bound analysis is, in my opinion, important. (2) In Theorem 3.3, the estimation error depends on the minimum eigenvalues of the empirical error functions' Hessian matrices, which in turn depends on the previously taken alterations. This raises a concern about the exploration-exploitation tradeoff when making alterations. An extreme case is making uninformative alterations (e.g., setting 0 for all nodes), leading to no update by Algorithm 1 and rendering the error bound in Theorem 3.3 meaningless (since $\mu_j$=0 in this case if I understand correctly). It is unclear how Algorithm 3 addresses this tradeoff and how $\mu_j$’s can be bounded below. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) How do algorithms handle the exploration-exploitation tradeoff (if explorations are needed)? (2) Is it possible to establish a regret bound for the cost? If not, what are the challenges? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your detailed feedback, and we hope our responses will address your concerns. W1&Q2. Theoretical guarantees (regret bound) for the cost. A1. Thanks for your insightful question. In fact, theoretical guarantees of the cost can be inferred from existing results (Lemma 3.1 and Theorem 3.3), and we discuss as follows: - First, by using the interior-point method, Eq. (6) can achieve the optimal solution $z^_{\hat{\theta}_t}$ that minimizes the objective cost function in each decision round $t$ since it is a convex QCQP [1, 2]. This is discussed in lines 265-271. - Because the constraint in Eq. (6) is constructed by $\hat{\theta_t}$ rather than $\theta_t$, let $C(\cdot)$ denote the cost function. The cost regret is $C(z^_{\hat{\theta_t}}) - C(z^_{\theta_t})$. Meanwhile, since $z^(\cdot)$ can also be viewed as a function, let $f(\cdot)$ denote $C \circ z^(\cdot)$. The regret for the cost is $f(\hat{\theta}_t) - f(\theta_t)$ in round $t$. Consider the following properties: 1. $\|\|\hat{\theta}_t - \theta_t\|\|^2$ enjoys a linear convergence rate as guaranteed in Theorem 3.3. 2. $f(\cdot)$ is a polynomial function, because $C(\cdot)$ is quadric (cost function) and $z^(\cdot)$ is polynomial (as $z^(\cdot)$ is polynomial w.r.t. $\mathbf{A}, \mathbf{B}, \mathbf{C}$ in the constraint of Eq. (6), and $\mathbf{A}, \mathbf{B}, \mathbf{C}$ are all polynomial w.r.t. $\theta_t$/$\hat{\theta}_t$, by proof process of Lemma. 3.1 [3]). 3. Both $\|\|\hat{\theta}_t\|\|$ and $\|\|\theta_t\|\|$ are bounded, by definition. - As implied by the properties above, $\|\|f(\hat{\theta_t}) - f(\theta_t)\|\|^2$ also enjoys a linear convergence rate. Consider the following simple example as an illustration: $$ (x-y)^2\leq Le^{-t} \quad\mathop{\Longrightarrow}\quad (x^2-y^2)^2 =(x+y)^2(x-y)^2 \leq (\|x\|+\|y\|)^2(x-y)^2 \leq 4U^2L e^{-t} = \mathcal{O}(e^{-t}), $$ where $U\in \mathbb{R}_$ is the upper bound of $\|x\|$ and $\|y\|$. We will refine this discussion in the revised paper to make the guarantees more clear. Thanks! --- W2. The minimal eigenvalue of the Hessian matrix in Theorem 3.3. A2. Thanks for your feedback. The Hessian matrix in Theorem 3.3 refers to $\nabla^2 \ell(\cdot)$ instead of $\nabla^2 \hat{\ell}(\cdot)$, where functions $\ell(\cdot)$ and $\hat{\ell}(\cdot)$ are defined in Eq. (4) and Eq. (5), respectively. Meanwhile, the $\hat{g}_{j,t}$ in Algorithm 1 refers to the gradient of $\hat{\ell}(\cdot)$, i.e., $\nabla \hat{\ell}(\cdot)$, instead of $\nabla \ell(\cdot)$. As a surrogate loss function of $\ell(\cdot)$, $\hat{\ell}(\cdot)$ only uses the collected sample in each round to approximate $\ell(\cdot)$. Hence, it is possible that the minimal eigenvalue of $\nabla^2 \hat{\ell}$ equals 0 (as in your provided example), but we want to emphasize that the minimal eigenvalue of $\nabla^2 \ell$, i.e., $\mu_j$ in Theorem 3.3, is not 0 since $\nabla^2 \ell$ is proven to be positive-definite in Appendix D, lines 714-738. Making uninformative alterations, as in your provided example, can be viewed as an extreme sample from $n$ potential possible alterations ($n$ in Eq. (4)). We apologize for the unclear presentation, and we will refine the related part. Thank you! --- Q1. Exploration-exploitation tradeoff. A3. Thanks for your question. Our modeling approach is different from conventional RL methods [4]. For the AUF problem, there are few opportunities to take actions (and perform explorations); hence, exploiting all available information to make effective decisions is essential. By leveraging the structural information (SRM), the rehearsal-based method can make decisions effectively without exploring different decision actions, as illustrated in the experimental results. Additionally, as explained in A2 above, selecting alterations does not adversely affect parameter estimation, eliminating the need for an exploration-exploitation tradeoff. We hope that these explanations better convey our points. Thanks again! --- References: [1] On implementing a primal-dual interior-point method for conic quadratic optimization, Math. Program. 2003. [2] Interior-point polynomial algorithms in convex programming, SIAM 1994. [3] Rehearsal learning for avoiding undesired future, NeurIPS 2023. [4] Reinforcement learning: An introduction, MIT Press 2018 --- Rebuttal Comment 1.1: Comment: Thank you for your response. It addresses my concerns and I have raised the score. --- Rebuttal 2: Comment: Dear Reviewer weAS, We are pleased to be able to address your concerns. Once again, thanks for the time and effort you dedicated to reviewing our work. Best Regards, Authors	Summary: In this paper, the authors address decision-making problem that sufficient interactions are not available. In this case, RL is not suitable. The authors model the structure among the observed variables, and use the structure to help the decisions. Compared to the previous studies [Qin et al. 37], the method can be used in a dynamic environment and can efficiently find the suggested decision (in polynomial time). To deal with the dynamic environment, they introduce the online learning method (Alg. 2). To efficiently find the suggested decision, they convert the optimization problem to a QCQP problem, which can be implemented in polynomial time. The experimental results verify the effectiveness. Strengths: 1. The method of Qin et al. [37] suffers a high computational cost. In this paper, the authors convert the problem to a QCQP problem, which makes it computable in polynomial time. It is a valuable contribution. 2. Theorem 3.3 presents an interesting and sensible theoretical guarantee. It is novel to see that some traditional online learning methods could be used in such decision tasks. Weaknesses: Some discussion about the offline RL are missing. See Questions for the details. Given the results of theorem 3.5: I do not know where $\tau$ is reflected in your algorithm. It seems that $\tau$ is never mentioned in Section 3.3. It is a bit wired, and needs more illustrations. The writing could be improved. There are some weird sentences. I suggest the authors carefully revise the paper. For example, "We provide the theoretical guarantees of our method, and experimental results validate the effectiveness and efficiency of the method." -> "We provide the theoretical guarantees for our method. Experimental results validate the effectiveness and efficiency of the method." Technical Quality: 4 Clarity: 3 Questions for Authors: I agree that RL is not suitable for the setting. However, I am wondering why offline RL cannot be used instead? Relevant discussions are missing. I can understand that the problem is hard in the non-linear case. Could authors have some discussions for the case that the data is non-linear? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: No limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the insightful feedback and the interest in our work! We hope our responses can address your concerns. W1&Q1. Discussion on the offline RL. A1. Thanks for your question. Generally speaking, online-offline hybrid RL methods can reduce the number of interactions by leveraging offline policy learning. However, these methods do not fit the AUF scenarios described in the paper for the following reasons: - In the offline training stage, hybrid RL methods use labeled offline datasets, i.e., offline datasets containing $(s, a, r)$ samples. However, in the AUF scenarios presented in the paper, only a few observational samples are available (in $(s, r)$ form), which do not contain information on any actions. - Considering "no action" as a special type of action is practical, but this approach would transform all offline data into the $(s, \text{no-action}, r)$ form, making it difficult to learn effective offline policies. - To achieve an effective policy, hybrid RL methods require a large number of offline samples and online interactions compared to the rehearsal-based methods [1]. For example, millions of online and offline samples are typically needed to obtain an effective policy [2, 3]. In contrast, our approach works well with only 100 samples in the AUF setting. This is because the rehearsal-based method can leverage the fine-grained structural information contained in the involved variables, whereas RL methods cannot. We will add this discussion to the revised paper. Thanks again! --- W2. $\tau$ in Theorem 3.5 and Algorithm 3. A2. We apologize for the confusion. In fact, the threshold $\tau$ is used in Eq. (6) to construct the matrix $\mathbf{P}=(\chi^{-1}(\tau)\mathbf{C}\mathbf{\Sigma}\mathbf{C}^\top)^{\frac{1}{2}}$, which appears in the constraint term in Eq. (6). Although $\tau$ is not explicitly mentioned in Section 3.3, both Theorem 3.5 and Algorithm 3 are related to the optimization Eq. (6). Hence, Theorem 3.5 and Algorithm 3 are connected to the threshold $\tau$ as well. We will clarify the related expressions in the revised version. Thanks! --- W3. Writing problems. A3. We truly appreciate your advice. We will review and refine the expressions in the paper. Thank you! --- Q2. Discussion for the non-linear case. A4. Generally speaking, non-linearity is indeed a significant challenge. However, it can be addressed in some cases, and we discuss the non-linearity in the following two stages: - For estimating the parameters of the SRM. In this case, non-linearity leads to different loss and surrogate loss functions compared to Eq. (4) and Eq. (5). If the new loss function $\ell_{\text{new}}$ is also convex w.r.t. $\beta$, then our proposed Algorithm 2 can still be used to estimate the parameters sequentially. Furthermore, Theorem 3.3 and Proposition 3.4 remain applicable. - For making decisions based on the estimation. When the relationships among variables are non-linear, Lemma 3.1 no longer holds. As an alternative, one can consider $\mathbf{Y}_t = f(\mathbf{x}_t, \mathbf{z}_t^\xi, \mathbf{\epsilon}_t)$, where $f$ is a non-linear function. To obtain a probability region as described in Proposition 3.2, the characteristic function [4] of the random vector may be useful. In this scenario, the constraint in Eq. (6) might no longer be linear or quadratic, which means the QCQP reduction may not apply. However, if the constraint is convex, optimization algorithms such as projection gradient descent [5] can be used to solve the new optimization problem. Lastly, we want to emphasize that addressing non-stationarity and improving time complexity are challenging even in the linear case. We will include this discussion in the future work section of the paper. Thank you! --- References: [1] Rehearsal learning for avoiding undesired future, NeurIPS 2023. [2] Hybrid RL: Using both offline and online data can make RL efficient, ICLR 2022. [3] Offline meta reinforcement learning with online self-supervision, ICML 2022. [4] Elementary probability theory, Springer 1977. [5] Convex optimization. Cambridge University Press 2004. --- Rebuttal Comment 1.1: Comment: These responses address my concerns and questions well, and thus further solidify my rating. Thanks. --- Reply to Comment 1.1.1: Comment: Dear Reviewer w1Mx, Thanks for your positive feedback. We are glad that our responses addressed your concerns and contributed to your evaluation. Best Regards, Authors	Summary: The authors formulate the Avoiding Undesired Future (AUF) problem in real-world scenarios of decision-making, especially in non-stationary environments, and propose a method to avoid undesired outcomes with minimal costs. Here the non-stationarity majorly comes from the different costs corresponding to different actions, and the varying influence relations over time. They also provide theoretical guarantees of their method and empirical results demonstrate the effectiveness and efficiency of the proposal. Strengths: - This paper is written well and clearly, with intuitive motivation and clarified novelty. - This paper includes a complete theoretical analysis and algorithmic design. Their proposed problem formalization is more general and practical than existing methods [37]. In particular, they first proposed a sequential method to maintain the dynamical influence, with guarantees of estimation error bound. They entailed Proposition 3.2 and Theorem 3.5 to help find the efficient alteration for $Z_t$ with the minimal cost. They finally propose the whole algorithm called AUF-MICNS, to avoid undesired outcomes in each decision round. - Experimental results show the effectiveness and efficiency of their proposed algorithm, where the evaluation metrics are success frequency, estimation error, average running time, etc. Weaknesses: I think my major concerns have been settled by the Supplementary Materials. So I have no other comments about the weaknesses. Technical Quality: 4 Clarity: 4 Questions for Authors: For the differences between SRM in the rehearsal graph and SCM in causality: - In the linear cases, it is easy to define the coefficients as the influences. When in the nonlinear cases, how to define the influences in the rehearsal graph? Is it the same as in causation, e.g., definitions of causal influence or causal effects? - Can the influence in rehearsal graphs (SRM) represent the bi-directional edge information? If not, I am confused what are the differences between such bi-directional relations in rehearsal learning and causality. Though in [35], the bidirectional edges are often due to common causes between two variables, there also exist some works that use causality to represent mutually influenced relationships[1]. A causal graph can also include cycles. - The operators in Figure 2 seem identical to the Intervention operator in causality. [1] Vimaleswaran K S, Berry D J, Lu C, et al. Causal relationship between obesity and vitamin D status: bi-directional Mendelian randomization analysis of multiple cohorts[J]. PLoS medicine, 2013, 10(2): e1001383. There are other minor typo errors: - It seems that in Eq.(1) or Eq.(3), it is better to add $t$ as a subscript for $V_j$ and $\varepsilon_j$? - In line 181, "ound" might be "round". Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Not applicable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the valuable feedback and appreciation of our work. We hope that our responses could mitigate your concerns. Q1. The difference between SRM and SCM. A1. Thank you for your insightful question. To some extent, the SRM and Rh(.) operations [1, 2] are indeed similar to their counterparts, SCM and do(.), in causality [3, 4]. However, there are some differences between the two graphical models. We will first outline these differences and then address the specific aspects you are interested in. 1. The modeling granularity of the rehearsal graph is more flexible. Under the assumption of causal sufficiency [3], direct edges in a DAG (directed acyclic graph) represent causal linkages. For example, $A \rightarrow C$ indicates that $A$ is the direct cause of $C$. In contrast, direct edges in a rehearsal graph do not necessarily represent causal linkages. For instance, $A \rightarrow C$ only implies that changes in $A$ lead to changes in $C$, without stating that $A$ is the direct cause of $C$. In this sense, the rehearsal graph is similar to a MAG (maximal ancestral graph) [5], however: 2. Bi-directional edges in a MAG do not have the same meaning as bi-directional edges in the SRM. Specifically, $A \leftrightarrow B$ in a MAG indicates that there are common causes between $A$ and $B$. Consequently, $do(A)$ would remove all associations between $A$ and $B$. In contrast, $A \leftrightarrow B$ in a rehearsal graph signifies that $A$ and $B$ mutually influence each other. Therefore, $Rh(A)$ only removes the influence of $B$ on $A$, resulting in $A \rightarrow B$ in the modified graph. 3. Possible dynamic influence relationships are allowed in SRM, as detailed in Appendix A.2. Based on the differences above, we would like to address your questions: - Non-linear influences in rehearsal graph. In this case, the influence $A \rightarrow B$ in the SRM is defined as the changes in $B$ when a unit change occurs in $A$. This is similar to the causal effect; however, the SRM also allows for dynamic properties, while the SCM generally assumes stationarity. Different from classic causations which typically describe the nature process that is stationary, the influence relation is more in accord with the decision process which could be dynamic. To prevent the potential misleading, we adopt influence relation in the paper. - Information of bi-directional edge. As explained in point 2 above, a bi-directional edge in causality (MAG) indicates a non-ancestral relationship. In contrast, a bi-directional edge in the rehearsal graph models a mutually influenced relationship. This modeling approach differs from cyclic causality [6, 7, 8], which also allows for mutually influenced $A _{\leftarrow}^{\rightarrow} B$ in the causal graph. 1. In cyclic causality modeling, parameters related to $A _{\leftarrow}^{\rightarrow} B$ exist in both observational and interventional situations [6]. 2. In SRM, parameters related to edge $A \leftrightarrow B$ only appear when one variable is altered (no parameters are associated with bi-directional edges in observational situation). This modeling approach reduces the number of parameters in the observational case and is reasonable. Consider the following simple example as an illustration: when $x_1 \leftrightarrow x_2$, $x_3 \rightarrow x_1$, and $x_3 \rightarrow x_2$, using $x_1 = ax_2 + bx_3$ and $x_2 = cx_1 + dx_3$ to model the relations is equivalent to $x_1 = \alpha x_3$ and $x_2 = \gamma x_3$ with $\alpha = \frac{ad+b}{1-ac}$ and $\gamma = \frac{bc+d}{1-ac}$. The latter model only uses 2 parameters, $\alpha$ and $\gamma$. Additionally, if $x_1$ is altered, the structure becomes $x_1 \rightarrow x_2\leftarrow x_3$. In this case, the parameter $\beta_{x_1 x_2}$ associated with $x_1\rightarrow x_2$ occurs. More details about the bi-directional edges are discussed in [2]. - The $Rh(\cdot)$ operator in Figure 2 seems identical to intervention operator. Conceptually, yes. However, as discussed in point 2 above, when operating on a node with bi-directional edges, these two operators will lead to different graph structures. This difference arises primarily from the distinct information contained in the bi-directional edges in rehearsal graph and MAG. To prevent the misunderstanding, we follow the existing operator $Rh(\cdot)$. At last, it is noteworthy that our approach can also work well when the structural information is expressed by an SCM (in linear Gaussain case), because Algorithm 2 and Proposition 3.2 also hold for SCM modeling. Thanks again for your helpful feedback. We will add this discussion and comparison with new related works in the revised version. --- Q2. Typo errors that subscript $t$ should be added. A2. Thanks for your advice. We will add the subscript $t$ for $V_j$ and $\varepsilon_j$ in Eq. (1) and (3). --- Q3. Typo errors that "ound" should be "round". A3. Thanks for your sharp observation. We will correct the spelling error. --- References: [1] Rehearsal: Learning from prediction to decision, FCS 2022. [2] Rehearsal learning for avoiding undesired future, NeurIPS 2023. [3] Causation, Prediction, and Search, MIT Press, 2000. [4] Causality: Models, Reasoning and Inference, Cambridge University Press, 2009. [5] Ancestral graph Markov models, The Annals of Statistics, 2002. [6] Learning linear cyclic causal models with latent variables, JMLR 2012. [7] Causal relationship between obesity and vitamin D status: Bi-directional Mendelian randomization analysis of multiple cohorts, PLoS medicine 2013. [8] NODAGSFlow: Nonlinear cyclic causal structure learning, AISTATS 2023.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Optimal Algorithms for Learning Partitions with Faulty Oracles	Accept (poster)	Summary: This paper studies the problem to recover an exact $k$ partition of a set with access to a same-cluster oracle that is allowed to lie $\ell$ times. This papers gives an algorithm with optimal query complexity up to constants and a lower bound. Strengths: 1. The result of this paper is clean and complete. The algorithm's query complexity matches the lower bound up to constants. 2. The main algorithm is concise, simple and elegant. The idea of the algorithm captures the problem well and has proved optimal guarantees. 3. The lower bound is non-trivial and has some interesting ideas. 4. This paper is very well-written. The notations and explanations are clear. I didn't even catch a single typo. Enough background and motivation is included in the paper. The paper is cohesive and organized, easy to follow. Math and algorithmic ideas are explained clearly. 5. I think this paper should be a spotlight. Weaknesses: I'm very satisfied with this paper, just two things I think it can improve on the writing. 1. The algorithm is quite simple and intuitive. On the other hand, the lower bound is more complicated and more technical. I think it might be better to write less on the algorithm but explain more on the lower bound, especially how to construct a good responder's strategy. 2. I think it's worth mentioning what's the optimal algorithm for the no-error oracle and compare your algorithm with theirs. Also something I would not like to call it a weakness since I think it's beyond the scope of the paper: 1. You justified the inconsistent error assumption (but I think the consistent error assumption can also be justified). But from the pure theoretical point of view, this assumption does make the algorithm design a lot easier and less interesting since the algorithm can make the same query many times. If the error model prevents such behavior of the algorithm, it is more interesting. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Is it possible to get better query complexity bounds if the goal is to recover the partition approximately but not exactly? 2. I mentioned consistent error above. I'm also thinking a more generalized adversarial error where it could be stochastic, for example, the expected number of error is $\ell$. Does stochasticity makes things harder? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The limitation of the paper is well addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your thorough review. We address your questions and comments in order. Regarding the writing: we agree that the details of the lower bound proof are technically more challenging and arguably more mathematically interesting, but we had previously decided to prioritize the algorithm since we thought the community might value it more. We will be happy to write more about the intuition behind the design of the responder strategy in the main body of the paper, and make the relevant appendix more detailed if the paper gets accepted. In terms of comparing with the error-free algorithm, there certainly are similarities between the two algorithms, and we would be happy to include an extra appendix elaborating on this comparison. In fact, the algorithms proposed in this work nearly-generalize the original algorithms for the error-free regime; when $k$ is known, setting $\ell_{{yes}}=\ell_{{no}}=0$ causes the algorithm to perform a nearly-equivalent procedure to the corresponding $k$-known algorithm of Reyzin and Srivastava, and when $k$ is unknown, this setting yields an algorithm that is exactly equivalent. Regarding the inconsistent error assumption: while it is true that this assumption guarantees that the problem is solvable given unlimited queries, and makes the problem easier for the algorithm, we believe that the value of our result lies in understanding the query complexity in this setup. If the model is chosen to prevent the algorithm from querying the same pair multiple times, then exact recovery is impossible even for very small values of $\ell$ (say $\ell = 3$). This follows from a previous result of Reyzin et al. This naturally leads to your question of partial recovery. This is an interesting research direction, and we believe that answering it may require a non-trivial amount of further work. The same applies to the stochastic error setting you suggested. It is worth noting that if one knows the expected number of errors, a simple application of Markov's inequality would allow one to run our algorithm with settings that guarantee success with probability 1-$\delta$ while making a number of queries that grows linearly in $1/\delta$. Stronger statements could likely be made by making further assumptions about the distribution of $\ell$. We are hopeful that our work will serve as a first step towards answering more questions in this problem space. We thank you again for your time and your help in improving the quality of our paper. Reyzin et al. Learning and Verifying Graphs using Queries with a Focus on Edge Counting --- Rebuttal Comment 1.1: Comment: Thank you for your response! I hope to see this paper at NeurIPS 2024!	Summary: [Setting]: This paper studies the problem of clustering n items into k clusters using an oracle that adversarially answers same-cluster queries for item pairs under the constraint that it makes at most $\ell$ errors for a known constant $\ell$. The goal is to exactly recover all clusters always (instead of just w.h.p.). [Contributions]: 1. A lower bound on the number of queries when k is known/unknown. The authors formulate the problem in terms of Chip-Liar game to get this result. 2. For known k, an algorithm that iteratively merges cluster using two heuristics: 1. If there is a (k+1)-clique of all -1s then the oracle has returned at least one false negative 2. More than $\ell$ "+1" responses from oracle for a given pair guarantees that it is in the same cluster 3. Sample complexity of the proposed algorithm that matches the lower bound. The results extend to a more general problem where individual limits on false positive and false negative errors are known. The paper also studies an algorithm for k-unknown case in the appendix. Sample complexity in this case is not optimal. Strengths: 1. The problem of guaranteed exact cluster recovery in the presence of noise is new. Having a hard limit on the number of errors made by the oracle makes this possible. Given concrete applications, this would be an interesting direction to explore. 2. The sample complexity of the proposed algorithm matches the derived lower bound when k is known. 3. The connection to Chip-Liar game for deriving the lower bound is interesting. Weaknesses: 1. The problem setting (oracle making at most $\ell$ errors with $\ell$ being a known constant) is not very practical in my opinion, which in-turn makes it hard to judge the significance of the results. Even for the examples given in the paper (L23-32), it is not clear why the oracle will make at most $\ell$ errors (e.g., an experiment failing in bioinformatics) or why $\ell$ will be known in advance. Do the authors have concrete applications in mind? 2. Clarity-wise, while the details in the paper are mostly clear, it would be helpful to include more details from the appendix into the main paper. For example, the following can be included by making Section 3 more concise, 1. What does "The position of a chip on the board will then be equal to the cost of the corresponding partition .." (L234-235) mean? 2. Some high-level details about the unknown-k algorithm. 3. Some intuition about why false-negative and false-positive error budgets inherently have a different contribution towards minimum sample complexity Technical Quality: 4 Clarity: 3 Questions for Authors: Please respond to point 1 under weakness Minor suggestions: 1. Typo in L100 - "A many" -> "Many" 2. A more recent paper (Gupta et al. 2024) studies a more general setting than Chen et al. (2023), which is closest to your work. Gupta et al. Clustering Items From Adaptively Collected Inconsistent Feedback - AISTATS, 2024 Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for taking the time to provide thorough feedback on our work. We begin by addressing your main question. After that, we discuss some of the weaknesses you mentioned. "... it is not clear why the oracle will make at most $\ell$ errors ... or why $\ell$ will be known in advance. Do the authors have concrete applications in mind?" The results in the paper characterize the minimum query cost needed in order to achieve a desired level of robustness to a number of errors. In particular, one may think of $\ell$ as not being given or known, but rather as being chosen by the user based on their desired error tolerance. This was briefly described in the introduction (L55-56). In the "global" response to all authors we discuss two more examples in which our model may be applied, which we will include in the paper if it gets accepted. The first of these (Example 1: Robustness to Misinformation) describes a situation in which the number of corruptions that will occur is finite and does not scale with the number of queries submitted, and in this example the learner may set the value of $\ell$ depending on their desired robustness to adversarial error. This type of cost-robustness trade-off is common in fields like information and coding theory, in which one is interested in the best achievable rate of communication subject to some fixed worst-case error. Dually, one may also think of our results as resolving the optimal error robustness subject to a constraint on the number of queries. Our second example (Example 2: Trustworthy Science) outlines a scenario in which one may benefit from understanding this trade-off. We note that it is impossible to design a protocol that allows the learner to exactly recover the full partition while remaining robust to a number of errors that grows as a constant fraction of the queries made. This follows trivially from known impossibility results (e.g. Proposition 1 of Reyzin et al.), and we will add a precise statement to the paper if it gets accepted. We briefly comment that in the $k$-known setting, it is actually not necessary for the number of false negatives ($\ell_{no}$) to be bounded or known in advance in order to run our algorithm and achieve optimal query complexity. Rather the value of $\ell_{no}$ that appears in the query complexity analysis will be the true number of false-negative responses that occur during the execution of the algorithm. In particular, this implies that in the version of the problem in which no false positives can occur, one does not need to set any upper bound on the error at all. Admittedly this may not be apparent since we pass $\ell_{no}$ as one of the arguments in the pseudocode, so we will remove it for clarity. In contrast, in the $k$-unknown setting the algorithm does require a user-chosen value for $\ell_{no}$. Here, a simple argument shows that it is impossible to guarantee one has found the correct answer unless they can upper bound both $\ell_{yes}$ and $\ell_{no}$. Effectively conveying the specifics of the error model is a key aspect of communicating our results, and we thank you for prompting us to clarify the role of $\ell$. We will update the manuscript to increase the emphasis on the above discussion, and we believe this will strengthen our paper. Further Discussion of Weaknesses ``...it would be helpful to include more details from the appendix into the main paper.'' We thank you for this suggestion and we agree. If the paper is accepted, we will include an overview of the $k$-unknown algorithm, discuss intuition for the asymmetry between false-negative and false-positive errors, and elaborate on the role of the chip board analogy. We now briefly answer the question: ``What does 'The position of a chip on the board will then be equal to the cost of the corresponding partition ...' (L234-235) mean?'' In the $\ell$-PL constrained version of the Chip-Liar game, each chip corresponds to one partition of $n$ elements into $k$ groups. When the questioner submits a query ("Are $u$ and $v$ in the same group?"), the responder answers either "yes" or "no". All chips whose partitions are inconsistent with the response (i.e. if the responder answers "yes", then all partitions where $u$ and $v$ are not in the same group) advance by one position on the board. In L234-235 we highlight that under this procedure, the position of the chip on the board is then equal to the cost of the partition as a feasible solution to an instance of a correlation clustering problem. Regarding the "Minor Suggestions" we were not aware of the recent work of Gupta et al. and we appreciate the pointer. We will make sure to discuss and cite it appropriately. We will also fix the typo you found. We conclude with a quick remark in response to the following: "The paper also studies an algorithm for $k$-unknown case in the appendix. Sample complexity in this case is not optimal." While it is true that, in the $k$-unknown setting, our algorithm does not achieve the optimal query complexity in every variant of the problem, it does achieve the optimal query complexity in the (unweighted) $\ell$-PL problem, i.e. the main variant in which false positive and false negative responses are penalized equally. Lastly, we wish to thank you again for your time and your help in improving the paper. We hope that in light of the clarifications on the error model you may deem our contributions more substantial. Reyzin et al. Learning and Verifying Graphs using Queries with a Focus on Edge Counting --- Rebuttal Comment 1.1: Comment: Thank you for responding to my questions. I very much appreciate the added motivation and applications. Consequently, I have increased my score from 4 to 6. All the best for your submission ! :)	Summary: The paper studies the problem of finding a hidden partition into $k$ clusters of a given universe. In many applications an algorithm has only access to a same-cluster oracle. A query to this oracle reveals whether two elements belong to the same cluster or not. This problem has been previously studied and tight bounds on the query complexity, i.e., the minimal number of queries required to solve the problem, are known (Reyzin and Srivastava, and Liu and Mukherjee). In this paper, the authors add the realistic assumption that the same-cluster oracle may not always reveal the correct answer. In their model, they (in advance) set a number \ell which bounds the maximum number of wrong answers which the oracle is allowed to make. The goal of an algorithm is still to compute the hidden cluster with as few queries as possible. In particular, for the same tuple of elements the oracle may give different answers for different oracle calls, and the algorithm does not receive any information on whether the response of the oracle was correct or not. The authors present an algorithm and analyze its query complexity. This bound is generally larger than in the setting with a correct oracle, and depends on the parameter \ell. If \ell=0, the presented analysis recovers the results by Reyzin and Srivastava. Furthermore, they give a tight lower bound using an argumentation based on Renyi-Ulam games and correlation clustering. They moreover study a slightly more general setting where the algorithm can set in advance more fine-grained bounds on how many false positive and false negative answers the oracle can give, and, for all problems they consider both, the setting where the number of hidden clusters k is known or not. Strengths: - I think that the problem is important and appreciated by the ML-community, as clustering is a fundamental problem in machine learning. Moreover, the assumption that a same-cluster oracle may not always be correct seems quite reasonable and realistic. Thus, I think that this problem and the presented results could have many applications and an impact in certain areas. - The authors give a tight analysis of the considered algorithms. - Despite being tight up to constants, the main algorithm is well-presented, and easy to understand and implement. - Overall, I think that the paper is well-written and seems technically sound. Weaknesses: - I think the main weakness of the model is that the upper bound \ell on the number of faulty oracle responses must be set in advance and stays fixed. This could be a major drawback when applying this model and the algorithm in practice, because it seems not clear why a faulty oracle should be consistent with such a bound. - It seems that the main algorithms is quite similar to the algorithm without faulty oracle. I think it would be helpful for the reader to have a paragraph where the difference to this original algorithm is explained. Further comments: - Line 236: missing 'and' Technical Quality: 3 Clarity: 3 Questions for Authors: Is there anything known for the setting where the number \ell is unknown to the algorithm, and it only appears in the analysis? I.e. a strong lower bound or an obvious workaround? Such insights or discussions could make the main weakness less severe. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations have been addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your response. We begin by answering your question: "Is there anything known for the setting where the number $\ell$ is unknown to the algorithm, and it only appears in the analysis?". In fact in the $k$-known setting, the algorithm presented in the paper does not require knowledge of $\ell_{no}$ in order to be correct. Admittedly this may not be apparent since we pass $\ell_{no}$ as one of the arguments in the pseudocode, so we will make sure to remove it in order to make this less confusing to future readers. A consequence of this is that in the $k$-known setting, one does not require any a priori knowledge of $\ell_{no}$, nor are they required to set any upper bound for $\ell_{no}$ in advance, in order to achieve the optimal query complexity both in the weighted $\ell$-PL problem and in the version of the problem in which only false negatives can occur. In particular, in this latter version, the learner would not require any knowledge of the error at all in order to achieve the optimal query complexity. Whether one could design an algorithm that also requires no knowledge of $\ell_{yes}$ in the $k$-known setting is an interesting question and we will make sure to discuss it in the "Conclusions, Limitations, and Future Directions" section of the paper. We note that in the $k$-unknown setting, without knowledge of $\ell_{no}$ and $\ell_{yes}$ it is impossible for an algorithm to make a finite number of queries and then output a partition that is guaranteed to be correct. To see this, suppose the oracle answers in a way that is consistent with some fixed partition $\mathcal{C}$, then (1) if $\ell_{no}$ is unknown to the learner it is impossible for them to establish whether $\mathcal{C}$ is the true partition or a refinement of it, on the other hand, (2) if $\ell_{yes}$ is unknown to the learner then it is impossible for them to tell whether $\mathcal{C}$ is the true partition or a coarsening of it. If the paper is accepted, we will add a summary of the above discussion to the manuscript. We will also be happy to elaborate on the similarities and differences with the algorithm for the error-free version of the problem. Finally, we have fixed the typo you found. We thank you again for your time and your useful comments. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: I thank the authors for their rebuttal. As far as I undernstand, there is still a major limitation to the result (at least for the k-known setting) because an algorithm requires to know $\ell$ upfront (expect that in the weighted setting $\ell_{no}$ can be unknown), which I think is not very realistic. I see that, in the k-unknown setting, there is a strong theorical lower bound, and I appreciate that you provided it. Overall, my opinion about the paper and my score remains: The paper has some nice insights and results, but also a major limitation to the model.	Summary: This paper studies the query complexity of clustering with a faulty oracle. Given a set of $n$ points $V$, which is partitioned into $k$ hidden clusters, the learner wants to recover the hidden partition by querying whether two points are in the same clusters or not. There has been a line of work that studies the query complexity of the problem where the response of each query has iid error. This paper studies a different query model, where the learner is allowed to make repeat queries for the same pair of points but the responses could be adversarially flipped at most $\ell$ times. This paper provides lower bounds for the query complexity of several variants of the problem and also designs efficient learning algorithms with a query complexity matching the lower bound. Strengths: 1. The paper establishes a novel relation between the clustering problem and the Rényi-Ulam liar games, which could potentially be useful for proving lower bounds for other learning problems. 2. The algorithm designed in this paper involves non-trivial techniques and has a query complexity that matches the lower bound proved in the paper. Weaknesses: My main concern is about the significance of the learning model studied in the paper. For graph clustering problems, the error is usually defined over the graph instead of over the queries, and sometimes repeated queries are not allowed. This is because sometimes by allowing the use of repeated queries, the learning problem could be easy to solve. For example, when iid noise is presented. In this paper, the error is defined over an unbounded sequence of queries but only allows a constant number of mistakes to happen. In particular, knowing the number of mistakes seems to be very important to make the learning algorithm designed in this paper work. These two points seem to be too idealized to model problems that arise from real applications. Technical Quality: 3 Clarity: 3 Questions for Authors: My questions are about the weakness pointed out above. 1. Can you provide any real applications that motivated the study of such a learning model? (Only a constant number of mistakes are made over the queries and such a number is known) 2. How would the learner know the error parameter $\ell$ in advance and if we do not have the parameter $\ell$ as input would it be possible to achieve exact recovery? 3. If repeated queries are not allowed and the mistakes are placed by an adversary, would it still be possible to (almost) recover the underlying clusters? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your helpful comments and questions. We begin by answering your questions, and we then discuss some of the weaknesses that were raised. 1. "Can you provide any real applications that motivated the study of such a learning model? (Only a constant number of mistakes are made over the queries and such a number is known)" In the "global" response to all reviewers, we provide further examples. The first example we provide (Example 1: Robustness to Misinformation) gives a setting in which it is reasonable to assume that the user can estimate an upper bound on the number of errors, and this quantity does not grow with the number of queries made. In general the aim of this project is to understand the fundamental trade-off between robustness to errors and query complexity when making same-cluster queries. In this context, the $\ell$-faulty oracle may not necessarily model the behavior of a system, but rather serves as an analytical tool to prove formal guarantees about this tradeoff. The second example (Example 2: Trustworthy Science) describes a setting in which it may helpful to understand the trade-off between robustness to error and number of queries independently of assumptions about the error model. 2. "How would the learner know the error parameter $\ell$ in advance and if we do not have the parameter $\ell$ as input would it be possible to achieve exact recovery?" One should think of $\ell$ not as being given or known, but rather as being chosen by the user based on their desired error tolerance. This was briefly described in the introduction (L55-56). We will be happy to further emphasize this perspective in the paper. For the full recovery guarantees to apply, the parameter $\ell$ does not need to be equal to the number of errors occurring, but it could simply be an upper bound to that quantity. We briefly comment that in the $k$-known setting, it is actually not necessary for the number of false negatives ($\ell_{no}$) to be bounded or known in advance in order to run our algorithm and achieve optimal query complexity. Rather the value of $\ell_{no}$ that appears in the query complexity analysis will be the true number of false-negative responses that occur during the execution of the algorithm. In particular, this implies that in the version of the problem in which no false positives can occur, one does not need to set any upper bound on the error at all. Admittedly this may not be apparent since we pass $\ell_{no}$ as one of the arguments in the pseudocode, so we will remove it for clarity. In contrast, in the $k$-unknown setting the algorithm does require a user-chosen value for $\ell_{no}$. Here, a simple argument shows that it is impossible to guarantee one has found the correct answer unless they can upper bound both $\ell_{yes}$ and $\ell_{no}$. 3. "If repeated queries are not allowed and the mistakes are placed by an adversary, would it still be possible to (almost) recover the underlying clusters?" Approximate recovery of the partition in this setting is an interesting research direction and remains an area for future work. We note that if repeated queries are not allowed, exact recovery becomes impossible even for small values of $\ell$ ($\ell=3$). This follows from a simple extension of a previous result of Reyzin et al. and we would be happy to highlight this fact in the paper if accepted. These impossibility results are in part what motivated the study of our paper's model in the first place. Discussion of weaknesses. While it is true that allowing repeated queries makes the task easier for the learner, one should note that assuming that the errors are made over the graph--as opposed to over the queries--may also simplify the problem, since typically one assumes that an adversary would commit to corruptions before observing the set of queries being made. These graph-based error models may not be suited to modeling every setting. Our first example (Example 1: Robustness to Misinformation) emphasizes why one may want to model a stronger adversarial behavior in this sense. With regards to the comment "knowing the number of mistakes seems to be very important to make the learning algorithm designed in this paper work," we emphasize that alongside providing an algorithm, a core goal of this paper is characterizing the tradeoff between error occurrence and query complexity. The cost-robustness tradeoff studied in this work is common in fields like information and coding theory, in which one is interested in the best achievable rate of communication subject to some fixed worst-case error. We thank you again for your helpful comments, and we hope that, in light of this discussion, you may deem our contributions more substantial. We are happy to answer any further questions during the discussion period. Reyzin et al. Learning and Verifying Graphs using Queries with a Focus on Edge Counting --- Rebuttal Comment 1.1: Comment: Thanks for spending time making the response. I would like to keep my score.	Rebuttal 1: Rebuttal: We thank all the reviewers for their time. Multiple reviewers have asked for more examples in which our model could be applied. Below, we give two more general motivating examples, in the tech and scientific domains respectively, illustrating the role of $\ell$ in learning tasks. Example 1: Robustness to Misinformation Consider a setting where a user is trying to cluster a dataset by crowdsourcing information in the form of same-cluster questions. However, the user suspects that an ill-intentioned competitor organization is attempting to corrupt the learning process by entering a number of bad actors in the crowd to strategically mislabel queries. If the user selects a new person every time they submit a query, then the number of adversarial answers they encounter is finite and does not grow with the number of queries submitted. In this scenario, $\ell$ plays the role of a security parameter, and the algorithm is guaranteed to be robust to up to $\ell$-many poisoned responses. The user can set $\ell$ based on, e.g., their prior belief about the resources of the competitor organization. Our results can be interpreted as quantifying the cost (in queries / crowd size) of implementing a fixed security parameter $\ell$. Example 2: Trustworthy Science Consider a setting in which a scientist is attempting to group items into classes by running experiments that reveal whether two items are in the same class. The scientist has limited resources (e.g. limited materials or time) and can only conduct a finite number of experiments. Our results allow the scientist to derive the maximum number of errors to which their learning procedure can be tolerant, given their fixed query budget. They can use this maximum value as the setting for $\ell$, and then use our algorithms to guide their choice of experiments. Our analysis would then allow them to measure the significance of their findings by quantifying the number of experiments that would need to have failed for the finding to be incorrect.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Color-Oriented Redundancy Reduction in Dataset Distillation	Accept (poster)	Summary: The authors propose AutoPalette, which reduces color redundancy in dataset distillation. They use a palette network and color-guided initialization to enhance training efficiency and performance by minimizing redundant color information in synthetic images and datasets. Strengths: Color redundancy is a fundamental aspect of natural scene images but is often overlooked in large-scale image analysis. This study focuses on the missing part, and the proposed method is effective. Weaknesses: - In the abstract, the authors summarize their framework as the one that minimizes color redundancy at the individual image and overall dataset levels. I think that’s a good summary. However, the description is not utilized when they introduce their framework in the main text. Although they describe it in the last section, it would be better to include the summary in the middle of the main, e.g., when introducing an overview or Figure 1. - I am confused a little about the definition of the color bit in this manuscript. The authors often describe the 8-bits for the original image (e.g., Figure 2). However, if the color bit is based on the number of color palettes, the original image should have 24 bits. - Typo: "> can encoded in fewer bits” should be "can be encoded” Technical Quality: 3 Clarity: 3 Questions for Authors: - While watching the condensed images in the Appendices, the CIFAR images are hard to perceptually recognize categories, but easy for Figures 7-9. I’m wondering why this perceptual difference emerges. - How did you decide the parameters, alpha, beta, and gamma in the experiments? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - The evaluation was mainly based on a relatively small number of image datasets. I’m not sure to what extent the condensed images change when applying recent large-scale image datasets. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weakness 1: In the abstract, the authors summarize their framework as the one that minimizes color redundancy at the individual image and overall dataset levels. I think that’s a good summary. However, the description is not utilized when they introduce their framework in the main text. Although they describe it in the last section, it would be better to include the summary in the middle of the main, e.g., when introducing an overview or Figure 1. Thank you for your valuable suggestion. We acknowledge that providing a summary of our framework earlier in the main text would improve the clarity and flow of our paper. In the revised version, we will include a concise summary of the framework before introducing the components to give readers a better understanding upfront. In addition, we appreciate your attention to detail and have noted the typo you pointed out. We will also correct this in the revised version. Weakness 2: I am confused a little about the definition of the color bit in this manuscript. The authors often describe the 8-bits for the original image (e.g., Figure 2). However, if the color bit is based on the number of color palettes, the original image should have 24 bits. Thanks for highlighting this point of confusion. The 8 bits we refer to represent the storage space for each channel (Red, Green, Blue) of each pixel in the image. When considering all three RGB channels together, the total storage space indeed amounts to 24 bits per pixel. Thus, both descriptions refer to the same concept of image representation, just from slightly different perspectives. We will clarify this in the manuscript to ensure that readers understand the distinction between the per-channel and total bit depth for image representation. Weakness 3: Typo: "> can encoded in fewer bits” should be "can be encoded” Thank you for pointing out the typo. We will correct this in the revised version of our paper. Question 1: While watching the condensed images in the Appendices, the CIFAR images are hard to perceptually recognize categories, but easy for Figures 7-9. I’m wondering why this perceptual difference emerges. Thank you for your obsesrvation and opinion. We hypothesize that this perceptual difference arises due to the resolution difference between CIFAR images ($32\times32$) and ImageNet images ($128\times128$). CIFAR images have a lower resolution, which makes the resulting synthetic images appear blurrier and less distinct, similar to an aliasing effect in low-resolution images. Another factor contributing to this observation could be the reduction in the number of colors. With fewer colors, it can be harder for humans to perceptually recognize objects, as colors palys a significant role in distinguishing and identifying visual elements. On the other hand, ImageNet images with higher resolution, can capture more details, allowing the synthetic images to represent objects more clearly and making their categories easier to recognize. We empirically find that visualizing results from other baseline distillation methods, such as DM and TM, shows a similar observation, that image resolution impacts the perceptual clarity and recognizability of synthetic images. Question 2: How did you decide the parameters, alpha, beta, and gamma in the experiments? Thank you for your question regarding the hyper-parameter settings. We conducted experiments to evaluate the sensitivity of our method to various values of these hyper-parameters. Specifically, we applied our method to distribution matching [1] with 10 IPC to test parameter sensitivity. In these experiments, we varied one parameter while keeping the other two fixed. The results shown in the tables below, indicate relatively stable performance within a range of $\alpha$, $\beta$ and $\gamma$ values. We observe that $\mathcal{L}_a$ has a slightly higher effect than the other two parameters. For instance, when gamma is set to 0.5 and 1.25, the test performance is respectively 58.08 \% and 58.06 \%. As $\gamma$ increases, we see an improvement in performance, which converges to around 60.9 \%. For the other two parameters, $\alpha$ and $\beta$, the sensitivity is lower, with optimal performance observed when both are around 3. This indicates that these parameters can be set within a reasonable range without significantly impacting the results. \| $\alpha$ \| 0.3 \| 0.75 \| 1 \| 1.5 \| 3.0 \| 6.0 \| \|:----------:\|:----:\|:------:\|:---:\|:-----:\|:-----:\|:-----:\| \| performance \| 60.63 \| 60.8 \| 60.9 \| 60.77 \| 60.91 \| 60.94 \| \| $\beta$ \| 0.3 \| 0.75 \| 1 \| 1.5 \| 3.0 \| 6.0 \| \|:-----------:\|:-----:\|:------:\|:---:\|:-----:\|:-----:\|:-----:\| \| performance \| 60.94 \| 61.42 \| 60.9 \| 60.78 \| 60.9 \| 60.95 \| \| $\gamma$ \| 0.5 \| 1.25 \| 2.5 \| 3.0 \| 5.0 \| 10.0 \| \|:-----------:\|:-----:\|:------:\|:---:\|:-----:\|:-----:\|:---:\| \| performance \| 58.08 \| 58.56 \| 60.19 \| 60.9 \| 60.6 \| 60.9 \| Limitation: The evaluation was mainly based on a relatively small number of image datasets. I’m not sure to what extent the condensed images change when applying recent large-scale image datasets. We appreciate the concern regarding the need for experiments on large-scale datasets. In addition to the experiments conducted on CIFAR10 and CIFAR100, we have applied our method to higher-resolution subsets ($128\times128$) of the ImageNet dataset, such as ImageNette and ImageWoof, as demonstrated in Table 2 of our paper. To further address the need for experiments on datasets with more classes, we conducted addtional experiments during the rebuttal period on the Tiny ImageNet dataset, which contains 200 classes with ($64\times64$) images. Please kindly refer to our response to R1W2 (Reviewer 5XJK). [1] Dataset Condensation with Distribution Matching, Bo Zhao et al. --- Rebuttal 2: Comment: Dear Reviewer, Thank you for your detailed and positive feedback on our paper. We wish we have addressed your comments in our rebuttal and would appreciate any additional insights or discussion you may have. Please kindly let us know if there is any further clarification required. Best regards, Authors --- Rebuttal Comment 2.1: Comment: Thank you for your response. My concerns have been addressed. However, due to the NeurIPS discussion rule, I am unable to review the revised main manuscript. I trust that the authors will revise it according to my comments, but I will maintain my current score with a neutral perspective.	Summary: This paper introduces a straightforward yet effective dataset distillation method called AutoPalette. The method minimizes color redundancy at both the individual image level and the entire dataset level. At the image level, it trains the palette network by maximizing color loss and palette balance loss, thereby reducing color redundancy in images. At the dataset level, a color-guided initialization strategy is proposed to minimize color redundancy across the entire dataset. Extensive comparative and ablation experiments convincingly demonstrate the approach's effectiveness. Strengths: - The proposed method outperforms other dataset distillation methods in most tasks, providing a new perspective on dataset distillation. - The experiments and ablation study seem well done. The paper's experiments are comprehensive, and the results of the ablation studies are convincing. Weaknesses: - The paper could benefit from a more detailed explanation of the color loss and palette balance loss. It would be helpful to include an explanation of why the palette balance loss might achieve a more balanced color palette. - The paper does not seem to explain why the similarity between the last layer gradients is measured instead of directly measuring the feature level similarity in the Color Guided Initialization Module. Technical Quality: 3 Clarity: 3 Questions for Authors: - How does the efficiency of this method compare to other methods? - Why does directly optimizing the task loss lead to assigning pixels to a limited number of color buckets in lines 156-158? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations of their work in Section 5. However, they could provide more detailed descriptions of how these limitations might impact the results. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weakness 1: The paper could benefit from a more detailed explanation of the color loss and palette balance loss. It would be helpful to include an explanation of why the palette balance loss might achieve a more balanced color palette. Thank you for your insightful comment. The palette balance loss is designed to encourage the network to distribute pixels uniformly across all colors, maximizing the representational capacity of each color. Specifically, the palette balance loss is calculated as the average entropy of $m$ over the spatial dimensions, colors, and channels, where $m$ represents the probability distribution for assigning pixels to color buckets. When the palette balance loss is minimized, it results in an even distribution of pixels across buckets, implying that the pixel count for each color bucket is approximately equal. This encourages the network to generate a palette with a more balanced color distribution preventing any particular color from dominating the images. Additionally, the maximum color loss is computed as the negative sum of the maximum confidences for each color bucket. This encourages each color to be selected by at least one pixel, thereby aiming to diversify the colors within each image. We will provide a more detailed explanation of these concepts in the revised version. Weakness 2: The paper does not seem to explain why the similarity between the last layer gradients is measured instead of directly measuring the feature level similarity in the Color Guided Initialization Module. Thank you for the insightful question. Measuring the similarity among samples by the last layer gradient and feature similarity can be largely correlated. The major difference is that the last layer gradient also captures the joint interaction beteen feature space and label space, while feature similarity mainly focus on the feature space. To explore this further, we conducted additional experiments using feature similarity in the graph cut initialization method. We applied color-reduced images through the models to obtain their feature representations and then computed the cosine similarities among these features. Due to time constraints, we applied our method to distribution matching (DM) [1] with IPC values of 1 and 10. As shown in Table below, the results indicate that using feature similarity respectively achieve 35.36\% and 60.9\% test performance on 1 IPC and 10 IPC experiments. The results align with our assumption -- Measuring by the last layer gradient and feature similarity can be largely correlated. We hypothesize that both methods for computing similarities can lead to good performance, suggesting flexibility in choosing similarity approach. \| IPC \| With Gradient \| With Feature \| \|------------\|-----------\|-----------\| \| 1 \| 35.5 \| 35.36 \| \| 10 \| 60.9 \| 60.9 \| Question 1: How does the efficiency of this method compare to other methods? The additional overhead introduced by our color palette network during the forward and backward process is minimal. Our palette network consists of 3 convolutional layers designed to generate color-reduced synthetic images efficiently. Consequently, the increase in wall-time is marginal compared to vanilla baseline methods such as trajectory matching [1] and distribution matching [2]. We would like to further highlight that AutoPalette could be more efficient than other parameterization methods [3-5], which require reconstructing the distilled image at the test time. This efficiency makes AutoPalette a practical choice for various applications where computational resources are limited. Question 2: Why does directly optimizing the task loss lead to assigning pixels to a limited number of color buckets in lines 156-158? Empirically, we observe that during the early stages of optimization, the palette network tends to assign pixels to only a few color buckets. As the optimization progresses, if it just focuses on minimizing the task loss (distillation loss), it often neglects the need for constraining the palette network and the need for a diverse color representation. This leads to the underutilization of the available color buckets, resulting in poor generalization capacity. To address this, we introduce the color maximum loss and palette balance loss as additional constraints on the optimization of the palette network. These losses encourage the network to utilize the full range of color buckets more effectively, leading to a richer and more balanced color representation in the color-reduced images. [1] Dataset Condensation with Distribution Matching, Bo Zhao et al. [2] Dataset Distillation by Matching Training Trajectories, George Cazenavette et al. [3] Dataset Distillation via Factorization, Sonhua Liu et al. [4] Sparse Parameterization for Epitomic Dataset Distillation, Xing Wei & Anjia Cao et al. [5] Remember the Past: Distilling Datasets into Addressable Memories for Neural Networks, Zhiwei Deng et al. --- Rebuttal Comment 1.1: Comment: Thank you for your response. My concerns have been fully resolved, and this is excellent work.	Summary: The paper titled introduces AutoPalette, a novel framework for dataset distillation (DD) that focuses on minimizing color redundancy at both the individual image and overall dataset levels. Authors propose a palette network to dynamically allocate colors from a reduced color space to each pixel, ensuring essential features are preserved. Additionally, a color-guided initialization strategy is developed to minimize redundancy among images, selecting representative images based on information gain. Comprehensive experiments on various datasets demonstrate the superior performance of the proposed color-aware DD compared to existing methods. Strengths: 1. Color quantization is an interesting way for dataset distillation, the motivation of this paper is interesting. 2. The methodology is well-defined, with clear explanations of the palette network and the color-guided initialization strategy. 3. The framework is shown to be compatible with other DD methods, indicating its potential for broad application. Weaknesses: 1. The paper does not discuss the potential impact of the method on the performance of larger dataset beyond the CIFAR-10 and CIFAR-100. These 2 datasets are two small and could not show the effectiveness of the proposed method. 2. There is limited exploration of how the method handles imbalanced datasets or classes with unique color distributions. Technical Quality: 3 Clarity: 2 Questions for Authors: See weakness. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: See weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weakness 1: The paper does not discuss the potential impact of the method on the performance of larger dataset beyond the CIFAR-10 and CIFAR-100. These 2 datasets are two small and could not show the effectiveness of the proposed method. We appreciate the concern regarding the need for experiments on large-scale datasets. In addition to the experiments conducted on CIFAR10 and CIFAR100, we have also applied our method to higher-resolution subsets ($128\times128$) of the ImageNet dataset, such as ImageNette and ImageWoof, as demonstrated in Table 2 of our paper. To further address the need for experiments on datasets with more classes, we conducted addtional experiments during the rebuttal period on the Tiny ImageNet dataset, which contains 200 classes with ($64\times64$) images. Please kindly refer to our response to R1W2 (Reviewer 5XJK). Weakness 2: There is limited exploration of how the method handles imbalanced datasets or classes with unique color distributions. To evaluate the effectiveness of AutoPalette on imbalanced datasets, we created an imbalanced CIFAR10 dataset following the protocol described in [1]. Specifically, we resampled the CIFAR10 dataset so that the number of samples per class is determined by $N_{i} = N * \alpha^{\frac{i}{N_c}}$, where $N$ is the original number of images per class, $\alpha$ is the scaling factor, $i$ indicates the i-th class, and $N_c$ is the number of classes. We used Distribution matching (DM) [2] as our baseline distillation method. The performance for IPC values of 1 and 10 is shown below. \| Ratio α \| Method/IPC \| 1 \| 10 \| \|---------\|------------\|-------\|-------\| \| 0.01 \| DM \| 25.91 \| 48.01 \| \| \| Ours \| 35.60 \| 59.53 \| \| 0.005 \| DM \| 25.54 \| 46.71 \| \| \| Ours \| 34.58 \| 58.11 \| From the results, we observed that when the scaling factor ($\alpha$) is 0.01 (with the minimum number of images per class being 50 and the maximum 5000), the distillation performance remains relatively stable. However, when the factor ($\alpha$) is further decreased to 0.005 (with the minimum number of images per class being 25), the performance slightly drops. These results demonstrate that our proposed method can consistently improves upon the baseline method across a broad range of imbalance factors and dataset imbalance settings. We agree that addressing imbalanced datasets is an important challenge. We believe that dynamically allocating limited storage resources across classes, rather than using a fixed number of images for each class, could potentially address this issue. We are excited to explore this direction in future work. [1] Dataset Card for CIFAR-10-LT (Long Tail), Huggingface. [2] Dataset Condensation with Distribution Matching, Bo Zhao et al. --- Rebuttal 2: Comment: Dear Reviewer, Thank you for your detailed review and valuable feedback on our paper. We hope we have addressed your comments in our rebuttal and would appreciate any additional insights or discussion you may have. We are more than willing to engage in further discussion and address remaining concerns during the discussion period. If our responses have resolved your concerns we kindly ask you to consider increasing the rating. Best regards, Authors	Summary: This paper introduces ColorPalette, a framework that minimizes color redundancy at the individual image and overall dataset levels. At the image level, the palette networks generate condensed images in reduced color bit-width while at the dataset level, a color-guided initialization strategy is proposed. The experiments are done using various datasets and IPCs. Strengths: 1. A new direction for exploring DC is proposed. 2. AutoPalette explores the possibility of performing DC in a reduced color space. The paper is easy to understand. Weaknesses: 1. AutoPalette seems like it is built on top of [1] with DC loss. 2. Lack of experiment on large-scale dataset ImageNet-1K. [1] Learning to Structure an Image with Few Colors, Yunzhong Hou et al. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. How is the performance of AutoPalette on ImageNet-1K? 2. Since the method falls into the parameterization category, given an IPC storage size, how many samples does AutoPalette generate? 3. In Table 1, why AutoPalette inferior to DATM on CIFAR-100 at 50 IPC? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weakness 1: AutoPalette seems like it is built on top of [1] with DC loss: Thank you for bringing up this important question! While color reduction plays a significant role in our methodology, our work primarily focuses on addressing two unique challenges inherent in dataset distillation with low IPC (limited synthetic samples). These challenges are fundamentally different from training networks with complete datasets as discussed in [1]. Specifically, the challenges include: 1. The optimization process in dataset distillation is extremely unstable, especially when employing a color reduction network. This instability can lead to the optimization process becoming trapped in local optima without capturing global information. 2. The color reduction model in [1] tends to select certain colors, potentially resulting in biases towards these colors and capturing spurious features. This bias can hurt the generalization of the trained model. In contrast, our method addresses these limitations by introducing two losses tailored for the dataset distillation process: (1) a color regularization term $\mathcal{L}\_{a}$ to enhance color consistency to reach global optima, and (2) a color balance loss $\mathcal{L}\_{b}$ to avoid biased color assignment. Our ablation studies, as presented in Table 3 of our paper, demonstrate the significance of these loss functions. They are critical for guiding the network towards a more balanced and effective representation of the color space. Without these loss functions ($\mathcal{L}\_{a}$ and $\mathcal{L}\_{b}$), the test performance drops by 5.06%. The inclusion of these two unique loss functions facilitates faster and more stable convergence, highlighting the main differences with [1]. Weakness 2 & Question 1: Lack of experiment on large-scale dataset ImageNet-1K: Thank you for your feedback regarding the importance of conducting experiments on large-scale datasets like ImageNet-1k. We have expanded our analysis beyond CIFAR10 and CIFAR100 by applying our method to higher-resolution subsets of the ImageNet dataset, such as ImageNette and ImageWoof, as shown in Table 2 of our paper. Conducting comprehensive experiments on the full ImageNet-1K datasets poses significant computational challenges. As highlighted in [6], distilling Imagenet-1k requires 4 NVIDIA A100 GPUs, each with 80 GB of memory. Unfortunately, these requirements don't allow for running all requested experiments in the rebuttal period. Therefore, to demonstrate the scalibility and efficacy of our proposed method on large-scale datasets, we conducted addtional experiments during the rebuttal period on the Tiny ImageNet dataset, which contains 200 classes with ($64\times64$) images. We adapted AutoPalette to the distribution matching (DM) [2] approach with Image Per Class (IPC) values of 1 and 10. Our findings show significant improvements in test performance compared to baselines. Specifically, our method improved test performance from 3.9\% to 7.02\% for $IPC=1$ and from 12.9\% to 29.52\% for $IPC=10$. The table below illustrates the test performance results. \| Method/IPC \| 1 \| 10 \| \|------------\|-----------\|-----------\| \| DM \| 3.9 \| 12.9 \| \| Ours \| 7.02 \| 29.52 \| We apologize for not able to conduct experiments on the full ImageNet-1K dataset at this time. We are eager to explore this for the camera-ready version. Question 2: Given an IPC storage size, how many samples does AutoPalette generate? Thank you for insightful question. With a fixed IPC storage budget, our AutoPalette method achieves a fourfold $4\times$ increase in the number of generated instances compared to the baseline. In constract, methods such as IDC [3] and HaBa [4] typically generate instances with a fivefold $5\times$ increase, while FReD [5] achieves increases ranging from fourfold $4\times$ to sixtheenfold $16\times$. This demonstrates the efficiency of AutoPalette in optimizing sample generation within the constraints of a given storage size. Question 3: In Table 1, why AutoPalette inferior to DATM on CIFAR-100 at 50 IPC? We appreciate your question regarding the performance comparison. Our method incorparates the trajectory mathcing strategy from DATM [6], where the matching steps are gradually increased during the distillation process. However, unlike DATM, we do not utilize the soft-labelling method and instead focus on exploring color features. This difference in approach may contribute to the slight performance discrepancy observed on CIFAR100 at 50 IPC. We believe that integrating soft-labelling could potentially enhance our method's performance in future work. [1] Learning to Structure an Image with Few Colors, Yunzhong Hou et al. [2] Dataset Condensation with Distribution Matching, Bo Zhao et al. [3] Dataset Condensation via Efficient Synthetic-Data Parameterization, Jang-Hyun Kim et al. [4] Dataset Distillation via Factorization, Songhua Liu et al. [5] Frequency Domain-based Dataset Distillation, Donghyeok Shin & Seungjae Shin et al. [6] Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching, Ziyao Guo & Kai Wang et al. --- Rebuttal 2: Comment: Dear Reviewer, Thank you for your detailed review and valuable feedback on our paper. We hope we have addressed your comments in our rebuttal and would appreciate any additional insights or discussion you may have. We are more than willing to engage in further discussion and address remaining concerns during the discussion period. Thank you again for your time and consideration. Best regards, Authors	Rebuttal 1: Rebuttal: We would like to extend our sincere gratitude to all the reviewers for their time and effort in reviewing our work. We deeply appreciate the insightful suggestions and feedback provided. We also thank for the acknowledgement of 1）our color-oriented redundancy reduction provides a new perspective in dataset distillation (5XJK, uhQ2, Jnju, VepF) 2) the proposed method is effective (Jnju, VepF) 3) our paper is easy to understand (5XJK) and methodology is well defined (uhQ2) Based on the common suggestions, we have conducted additional experiments during the rebuttal period, summarized as below: - Additional experiments on large-scale dataset Tiny ImageNet. Please refer to our rebuttals to Reviewer 5XJK Weakness 1, Reviewer uhQ2 Weakness 1, Reviewer VepF Limitations. - Impact of hyperparameters $\alpha$, $\beta$ and $\gamma$? Please refer to the rebuttal for Reviewer VepF Question 2. - Performance on the imbalanced dataset. Please refer to the rebuttal for Reviewer uhQ2 Weakness 2. Please find the point-to-point response in each individual reply. Thank you once again for your valuable feedback and support.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Learning 3D Garment Animation from Trajectories of A Piece of Cloth	Accept (poster)	Summary: The authors propose a method to transfer the deformations of the observed garments to any other garment. Previous methods either rely on a large-scale dataset for training or analytical physics model with limited expressive ability. On the contrast, the proposed method first learns the constitutive relations from the observation by a neural network (EUNet), then use it as an energy prior to regularize the training of garment deformation model. This design addresses the limitation of previous works and show better results. Strengths: The strength of the paper is the proposed method does not need to collect huge amount of data with varied body pose and shape and garment types for training. Through theoretical analysis, they prove that they can learn a more physically accurate energy model to describe the deformation of garment. In this way, they do not need an explicit physical model, which tends to have limited expressive power. The derivation is theoretical sound. Weaknesses: To learn this energy model by EUNet, the authors rely on the synthetic data simulated with blender. A cloth with known geometry (vertices and faces) is assigned with a specific material type. However, this setting is too ideal. In real scenarios, we are more interested in transferring the material of a real cloth to another garment. But having the geometry of a real cloth usually is not feasible. Even though we can have the mesh of the cloth through some registration process, how to get the shape of mesh when it is hanging and dangling is still a problem. In this paper, I do not see the possibility of using the proposed method in real applications. This is the critical weakness. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. In Fig. 4, the results of MGN-S+EUNet on dress are not similar to the ground truth data. 2. What is the unit of the errors in Table 1? The errors on the leather and denim seem too big compared with the others. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: See the weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. # W1: Possibility of Using The Method in Real Applications Firstly, synthetic data is commonly used and facilitate the research of garment animations, such as TailorsNet [1], Cloth3D [2], MotionGuided [3], Clo3D [4]. Secondly, as shown in the dataset [5], it is feasible to obtain temporally consistent cloth data represented by vertices from real life, which can be directly applied by our method to learn the constitutive laws in real world, though it is not yet publicly available. Thirdly, the garment geometry can be extracted either through learning based methods [6, 7] or from raw scans [8], making it possible to animate realistic garments. As shown in Figure 2 in additional PDF from global rebuttal, we animate the garments from real scans [8] for demonstration, transferring the virtual material from Blender to real-life garments from CAPE [8]. In conclusion, our work can be directly applied to real-life applications related to garment animations. [1] Chaitanya Patel, et al. TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style. CVPR2020 [2] Hugo Bertiche, et al. CLOTH3D: Clothed 3D Humans. ECCV2020 [3] Meng Zhang, et al. Motion guided deep dynamic 3D garments. TOG2022 [4] Xingxing Zou, et al. Cloth4d: A dataset for clothed human reconstruction. CVPR2023 [5] David Blanco Mulero, et al. Benchmarking the Sim-to-Real Gap in Cloth Manipulation. IEEE Robotics Autom2024 [6] Heming Zhu, et al. Registering Explicit to Implicit: Towards High-Fidelity Garment mesh Reconstruction from Single. CVPR2022. [7] Lingteng Qiu, et al. REC-MV: REconstructing 3D Dynamic Cloth from Monocular Videos. CVPR2023 [8] Qianli Ma, et al. Learning to Dress 3D People in Generative Clothing. CVPR2020 # Q1: Qualitative Results of Dress We acknowledge that predicting long-term dynamics in an auto-regressive manner inevitably leads to error accumulation. The baseline models either struggle with predicting long-term animations or fail to replicate similar constitutive behaviors observed in the ground truth. In contrast, MGN-S+EUNet (ours) delivers material behaviors that more closely match the ground truth, such as a more rigid dress in Figure 4 of the main manuscript, and achieves lower prediction errors as shown in Table 2 of the main manuscript, suggesting the effectiveness of our learned EUNet. Please refer to the video in supplementary for better comparisons and the additional PDF file for more comparisons. # Q2: Explanations of Errors in Table 1 of The Main Manuscript The unit is the energy $kg\cdot m^2/s^2$. The errors in Table 1 of the main manuscript are the square errors as indicated in Equation 8, where we sum up 1.4K energy units of the cloth with 484 vertices. The per-energy-unit errors are significantly smaller in terms of the absolute value. Regarding the differences between the materials, leather and denim are either stiffer or heavier. These characteristics lead to larger energy changes between frames due to significant variations in velocity or increased mass, making it more challenging for the model to learn. We would include the discussion in the limitation section in our revised version. In conclusion, as shown in Table 2 of the main manuscript, MGN-S+EUNet (ours) still achieves superior performance, suggesting the effectiveness of our EUNet. And as shown in Figure 1 of the main manuscript and the video in supplementary, MGN-S+EUNet delivers reasonable material behaviors for garments made of silk and leather, which are softer or stiffer respectively. --- Rebuttal Comment 1.1: Comment: Thanks for your clarification and efforts in the rebuttal. I agree that many synthetic data is widely used for garment animation or simulation. However, the simulated data is just approximation for the real data. As pointed by David Blanco Mulero, et al [5], there is a large gap between the synthetic data and the real data. Besides, the temporal data collected in [5] are point clouds instead of meshes with valid geometry. The proposed method requires the trajectory of mesh as input. I do not think the real data of [5] can be used by it to learn the constitutive laws. --- Reply to Comment 1.1.1: Title: Official Comment by Authors Comment: We thank the reviewer for the response. We first clarify the potential use of dataset [1], followed by a more explicit summary of our rebuttal addressing the concerns raised in the weaknesses. Firstly, many existing softwares, such as MeshLab and AutoDesk ReCap, support the generation of meshes from point clouds. Therefore, a straightforward solution is to preprocess the point clouds from the dataset [1] using these tools and extract the meshes in the format required by our inputs. Many work, such as [2], have developed techniques to generate meshes from point clouds. And since the vertices are temporally consistent, the trajectories of the meshes remain consistent as well. Secondly, we believe the concerns in weaknesses result from potential difficulties in collecting real-world data, as mentioned by "...A cloth with known geometry...The setting is too ideal...having the geometry of a real cloth usually is not feasible..." and "...how to get the shape of mesh when it is hanging and dangling is still a problem...", leading to the key concern "...do not see the possibility of using the proposed method in real applications...". However, the dataset [1] from real-world closely resembles our settings and captures temporally consistent data when the cloth is "hanging and dangling". As mentioned in the previous rebuttal, the garment geometry can be extracted either through learning based methods or real scans. Examples from CAPE [3] even include meshes of clothed humans obtained from real scans. The animaitons of garments from CAPE, as demonstrated in the rebuttal, showcase the real application of garment animation. In other words, collecting real data in the format required by our method is feasible, making it applicable for real-world applications In addition, while our work is trained and validated using synthetic data, which is agreed by the reviewer, our EUNet is also applicable to real data. The key assumption is that the dynamics of objects follow general physical laws, such as Newton's laws or, equivalently, the Lagrangian mechanics, which apply to both simulated and real-world data. The primary difference is that the clothing model in simulations is typically known, whereas in real life, it is unknown. Our EUNet is designed to model these underlying constitutive behaviors. Therefore, our approach can be directly applied to real data without any modification. Together with the animations of real garment in CAPE, both our core module, EUNet, and the garment animation method demonstrate the potential for real-world applications. We hope that the clarification provided above addresses the reviewer's concerns more explicitly, including "...do not think the real data can be used...", "...The setting is too ideal...having the geometry of a real cloth usually is not feasible...", "...how to get the shape of mesh when it is hanging and dangling is still a problem...", and "...do not see the possibility of using the proposed method in real applications...". [1] David Blanco Mulero, et al. Benchmarking the Sim-to-Real Gap in Cloth Manipulation. IEEE Robotics Autom2024 [2] Rana Hanocka, et al. Point2Mesh: A Self-Prior for Deformable Meshes. TOG2020 [3] Qianli Ma, et al. Learning to Dress 3D People in Generative Clothing. CVPR2020	Summary: This submission presents a method that could effectively learn the dynamic patterns of different garments from a single piece of cloth. The key insight is that the motion of different cloths is governed by both external forces and the constitutive relations rather than specific garment topologies. Thus, an Energy Unit Network (EUNet) is proposed to learn the topology independent constitutive laws. Then the animation is performed by energy optimizations. Experimental results shows improved results comparing to previous methods and baseline methods. Strengths: The paper is well written and easy to read. 1) The paper is well structured. 2) Many terms are well defined and explained. The proposed method is both novel and interesting. 1) The disentangled learning scheme of using a network to learn the constitutive law, which generalizes to different garment types, is physically intuitive and natural. More importantly, this design helps alleviate the needs for large amount of training data of various cloth shapes in dynamic for learning based animation. This disentanglement between topology and energy is achieved by using mesh edge as a unit instead of the whole cloth mesh. 2) The proposed disturbance training strategy helps stabilize the training and improves the generalization of EUNet. As a constraint, it accompanies the direct supervision on the energy form by taking into account the physical meaning of equilibrium state. This helps the network to learn a more reasonable manifold of the energy distribution. Experimental results: 1) Improved results over previous methods are shown both qualitatively and quantitatively. 2) According to the ablation study, the design of both the contrastive loss and dissipation unit loss are validated as effective. Weaknesses: Some details regarding the design of the EUNet is missing. Although some descripsions on the EUNet design is provided at the experimental section, it is relatively hard for the reader to follow and develop a more coherent understanding of the presented work. Some limitations and open questions that the work might not cover: What about anisotropic materials? How to adapt the current model design to also fit to cloths where its material is anisotropic? How well does the method handles cloths with more complex topology that goes beyond a single layer of cloth? As also pointed by the authors, the method does not handle self-collision. It would be interesting to see how it can be adapted in that axis. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and recognizing the value of our work. We believe the reviewer has sufficient understandings of our framework and pipeline. # W1: Details Regarding The Design of EUNet We introduce the formulations and training procedures of our EUNet in Section 3.2, and include the implementation details in Section 4.1 at L239-249. Furthermore, we report the time efficiency of our EUNet in the global rebuttal. Please let us know if the reviewer still has any confusion about our EUNet. We will clarify it further in our revised version to ensure it is easier for readers to follow. # W2.1: Anisotropic Materials A naive extension is to add the directional information as extra inputs for our EUNet. As a simplified example, we assume a rectangular wooden board that is thin enough for its thickness to be negligible. The board is stiffer along the x-axis than the y-axis. We then discretize the thin board by edges along both the x-axis and y-axis. To model such material, we can extend our EUNet by taking the edges' directions, which are either x-axis or y-axis in this example, as extra inputs by $\Phi(\cdots, \mathbf{d})$, where $\mathbf{d}$ indicates the directional information. # W2.2: More Complex Topology Than A Piece of Cloth For cloth or garments with complex topologies used as training data, our EUNet can directly learn their constitutive behaviors thanks to the topology-independent design. To achieve accurate results, it is essential to ensure that the training sequences involve minimal collisions, so that the dynamics and deformations are primarily caused by internal forces closely related to potential energy. The reason we chose to use a simple piece of cloth as training data is that this setup is easier to achieve with high quality in real-life scenarios. For example, [1] provides the temporally consistent cloth data represented by vertices from real life, which can be directly applied by our EUNet but not publicly available yet. [1] David Blanco Mulero, et al. Benchmarking the Sim-to-Real Gap in Cloth Manipulation. IEEE Robotics Autom2024 # W2.3: Handling of Self-collisions Since collisions are independent of the internal forces caused by deformation, the key to accurately modeling potential energy is to minimize the effects of collisions. However, to extend our method to more complex scenarios involving collisions, one could design a new module that takes the distances between edges or faces of the cloth as inputs to model the dissipation energy caused by collisions. In this work, we minimize the effects of collisions during the data generation process. We will include such scenarios in future work.	Summary: This work proposes a method to learn the constitutive model of cloth materials from observed cloth trajectory using a neural network. It adopts an MLP that operates on individual edges and predicts per-edge distortion based on the deviation of edge geometry from rest shape and trains the network using a combination of supervision on potential energy change with ground truth and optimality of incremental potential. The learned potential energy can be used as a constraint to train neural simulators for garment animation. Strengths: - I appreciate the novelty in the idea of learning the constitutive model of cloth materials in a data-driven manner. Potentially this formulation could allow the neural networks to understand the intrinsic physical property instead of mimicking the behavior of specific examples, thus of scientific significance if implemented correctly. - The paper is well-written and mostly clear. Weaknesses: - On the methodology side, the major question is probably the design of dissipative energy. On the one hand, why it is and is merely a function of $X^t - X^{t-1}$ is questionable. In fact, whether it should be modeled as an absolute quantity is a question because the total amount of dissipative energy seems not that meaningful. The only observable quantity is the relative change of dissipative energy in a physical process. - On the other hand, with the presented framework, it is very hard to learn the major sources of energy dissipation: collision, and friction, since they are neither present in the training data, nor fully modeled (e.g. self-collision) in the formulation. While the dissipative energy is not the focus, the problem is that without correctly modeling dissipative forces, I doubt the possibility of learning an accurate elastic potential energy function. - On the evaluation side, the problem is that the method is only evaluated in a simplified setting, without comparing against methods or in settings that are practically useful (see below). In my opinion, there are two ways to demonstrate that the learned constitutive model is useful: either 1. demonstrate that it is more accurate than an analytical model on real data, or 2. show that it leads to more realistic animation than existing methods (including traditional numerical models). - The evaluation section only shows that the MGN trained with the learned constitutive model is better than those trained with ground truth garments or analytical constitutive model. On the one hand, it does not compare with other state-of-the-arts like HOOD, SNUG, and PBNS that are also formulated in a self-supervised manner. On the other hand, the claim that it is better than the analytical constitutive model is not convincing because the discrepancy may be caused by the limited accuracy in the neural simulator (or even by the mini-network mentioned in Sec 3.3). To truly demonstrate that it is better than an analytical one, it must be compared using a numerical integrator that is guaranteed to converge to the energy minimum. Technical Quality: 2 Clarity: 3 Questions for Authors: See the weaknesses section. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The discussion seems adequate. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. # W1: Design of Dissipative Energy As commonly adopted in physics simulation, such as the Rayleigh dissipation, the dissipation can be approximated by a function of objects' velocities, which can be calculated from $X^t-X^{t-1}$ in our work. In addition, we agree that the absolute value of the dissipative energy is not meaningful. However, during the energy optimization process of solving dynamics, only the derivative of the dissipation energy $\lim_{\Delta x \to 0}\frac{\Phi_d(x+\Delta x, \cdot)-\Phi_d(x, \cdot)}{\Delta x}$ will take effect, and eliminate the potential impacts from the absolute values. Lastly, if there exists a constant $C>0$ in the output of our $\Phi_d$, it indicates that the system from Blender has a constant damping effect. When the velocity is 0, the output of $\Phi_d$ is $C=3\times 10^{-6}$, which is small enough to be negligible. # W2: Dealing with Collision And Friction We include the frictions between cloth and air, which is common in daily life, in both the generated data and modeling process. As shown in Table 1 and Figure 3 in the main text, with the modeling of dissipation by $\Phi_d$, our EUNet achieves lower errors and delivers more reasonable energy mappings as the cloth deforms, suggesting the effectiveness of $\Phi_d$. In addition, since collision forces are independent of the constitutive behaviors, we do not model collisions as part of our EUNet. We avoid the self-collisions in training data by setting an upper bound on the initial cloth's velocities and the range of their directions. Quantitative results in Table 1 of the main manuscript and the performance on garment animations demonstrate the effectiveness of our EUNet. # W3: More Evaluations to Verify Learned Constitutive Model Since the final goal is learning to animate the garments where EUNet is key module of this framework, we conduct comprehensive experiments on garment animations in Table 2 and Figure 4 in the main text. In addition, we further implement HOOD [1] as an extra baseline. Please refer to the PDF in global rebuttal for more details about the performances and the differences between baselines. As shown in Table 2 of both main manuscript and the global PDF, either MGN-S or MGN-H (used in HOOD) being the neural simulator architecture, simulators constrained by our EUNet always outperforms baselines (i.e., HOOD and others in the Table 2 of the global PDF), indicating the effectiveness of our EUNet. [1] Artur Grigorev, et al. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. CVPR2023 # W4: Comparisons with HOOD and Analytical Clothing Model Firstly, as discussed in W3, we further compare with HOOD, where simulators constrained by our EUNet outperforms it. As for the claim 'better than the analytical constitutive model', it is worth noting that we didn't claim it in abstract and introduction. We only mentioned this in our experimental section as an empirical analysis at L298-299. We will clarify this in the revision. Moreover, to further verify the advance of our EUNet over analytical models is not caused by the limited accuracy of neural simulator, we pair our EUNet and the analytical models with two different neural simulator architectures, namely MGN-S and MGN-H. As a result, simulators constrained by our EUNet always achieve lower errors, which means adopting our disentangled learning scheme with EUNet is better than using analytical models to train neural simulators. Thirdly, we further report the comparisons that with or without the mini-network in Table 2 from the global PDF. The mini-network enables the unsupervised simulators, which are constrained by analytical physics model, to obtain higher accuracy. Finally, we admit that implementing numerical simulations or PDE solvers within a short period is challenging. However, within the scope of learning to animate garments and based on the experiments, our method can better enhance learning-based simulators to achieve lower errors and faithful animations. --- Rebuttal Comment 1.1: Comment: I appreciate the efforts made by the author in the rebuttal. On the formulation side, the rebuttal makes it clear that it avoids collision and friction in both the modeling and the data. This makes more sense to me now. However, I agree with Reviewer cPXz that this setting would make it too ideal for capturing the constitutive model of a real garment, which is where the method can be truly interesting. On the result side, a comparison with HOOD surely makes the validation more solid. However, it is still my opinion that using the model to constrain a neural simulator is less convincing than showing it works with a full numerical simulator, and yet better, being better than an analytical constitutive model in that case. Given these considerations, I would like to increase my rating to Borderline Accept. While the paper still has some limitations, I find the idea of learning a constitutive model for an unknown analytical form interesting. I wish it can become useful in the real world one day in the future. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for the positive feedback and recognition of our contributions and efforts. We would like to provide additional comments on the correspondence between our work's settings and potential real-world scenarios. ## Difficulties in Collecting Real Data As demonstrated in [1], existing techniques are able to capture temporally consistent data from real world. The settings in [1] closely resemble ours and include the dynamics where a cloth is hanged in the air. Therefore, the data required by our method is feasible to capture in real-world scenarios. ## Simplification of Non-Collision Settings As shown in the visual demos from official websites of [1], with properly designed external forces, it is feasible to capture cloth dynamics without self-collisions. ## Simplification of Using Only One Piece of Cloth The simplification is reasonable, as garments and clothes made of the same materials exhibit the same constitutive behaviors. Therefore, learning the constitutive behaviors from garments is equivalent to learning from a single piece of cloth. As stated in our abstract and global rebuttal, the approach of using a piece of cloth to learn garment dynamics reduces the great need of a large scale of garment data. Besides, the settings of learning from a piece of cloth is applicable in real scenarios such as [1]. Comparing with collecting data from entire garments, it is easier to handle a piece of cloth and more feasible to obtain non-collision data. At last, we agree that including collisions, such as self-collisions within clothes and collisions with other objects, would lead to more complex settings that exist in real life. However, we also want to emphasise the reasonable simplifications without collisions is applicable and effective for learning constitutive models. We would include the collisions and more complex settings in the future. [1] David Blanco Mulero, et al. Benchmarking the Sim-to-Real Gap in Cloth Manipulation. IEEE Robotics Autom2024	Summary: The paper proposes a novel method for animating garments by learning from a single piece of cloth. This approach circumvents the need for large-scale garment datasets, which are resource-intensive and time-consuming to create. The core idea is to use a disentangled scheme where constitutive behaviors are learned from observed cloth and then applied to animate various garments. The proposed Energy Unit Network (EUNet) captures constitutive relations in the form of energy, bypassing the need for traditional physics models. Strengths: The paper introduces a novel disentangled approach that separates the learning of constitutive behaviors from the animation process. The EUNet models constitutive behaviors using energy units, allowing for direct learning from observed cloth trajectories without traditional physics models. The approach significantly reduces the data requirement, relying on a single piece of cloth for training, making it more practical and less resource-intensive. The method produces animations that are both robust and generalizable, capable of handling various garment types and materials. Weaknesses: The energy optimization process, although effective, can be computationally intensive and may require fine-tuning to achieve optimal results. The paper would benefit from more extensive experimental validation, including comparisons with a broader range of existing methods and more diverse garment types. Technical Quality: 3 Clarity: 2 Questions for Authors: How does the performance of EUNet compare with traditional physics-based models in terms of computational efficiency and accuracy? What are the limitations of using a single piece of cloth for training, and how can these limitations be mitigated in future work? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Comparisons with a broader range of existing methods would be appreciated, for example, one of the SOTA method named 'HOOD: hierarchical graphs for generalized modelling of clothing dynamics' which has been cited in the paper is worth comparing with. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive and valuable feedback. # W1: Computationally Intensive Energy Optimization Process In this paper, we primarily focus on the challenge of modeling constitutive relations from observations. The energy-based simulation is used solely as a tool to solve the dynamics constrained by our EUNet. Consequently, accelerating the energy optimization process is beyond the scope of our work. # W2: Extensive Experimental Validation As requested by the reviewer, we implement HOOD [1] for further comparisons. Moreover, the baselines we have chosen are representative: MGN is the pioneering approach for animating mesh-based cloth, LayersNet achieves state-of-the-art performance for garment animation trained in a supervised manner, and MGN-S+PHYS employs a recent advanced unsupervised learning scheme as shown in SNUG [2] and HOOD. The main difference between MGN-S+PHYS and HOOD is that HOOD adopts a hierarchical graph neural network. Please refer to the additional PDF in the global rebuttal for more details. Simulators constrained by our EUNet achieve superior performance over other baselines. In addition, all garments in Cloth3D [3] differ from each other in terms of length, size, and topology. Each category in Table 2 of the main manuscript includes several subclasses. For example, the "T-shirt" category also includes "Jackets" as shown in Figure 1 of the main manuscript, and the "Dress" category includes both short and long dresses. "Jumpsuits" and "Dresses" are combinations of upper and lower garments. [1] Artur Grigorev, et al. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. CVPR2023 [2] Igor Santesteban, et al. SNUG: Self-Supervised Neural Dynamic Garments. CVPR2022 [3] Hugo Bertiche, et al. CLOTH3D: Clothed 3D Humans. ECCV2020 # Q1: Efficiency and Accuracy Comparing with Physics-based Models As shown in global rebuttal, our EUNet is comparable to the traditional clothing model in terms of speed. While our EUNet contains two branch: $\Phi(\cdot)$ for potential energy and $\Phi(\cdot)$ for dissipation, each branch is slightly faster than the traditional clothing model. In terms of accuracy, as shown in Table 2 of both the main manuscriptand the global PDF, simulators constrained by our EUNet achieves higher accuracy than those constrained by physics-based clothing models. Furthermore, our method reduces the need to carefully select different types of physics-based clothing models and estimate the corresponding parameters. Instead, our EUNet directly captures the observed constitutive laws, enabling learning-based simulators to achieve superior performance in garment animation. # Q2: Limitations of Using A Piece of Cloth One limitation is that a single piece of cloth may not encompass all possible deformations and corresponding dynamics. For example, shear deformation may be less obvious, and interactions among different layers of clothing are unavailable. A simple solution to enrich the deformations and dynamics is to apply known forces, such as those controlled by robots, and to use multi-layered clothing during data generation to explore these interactions On the other hand, we intentionally design the training data to be as simple as possible to ensure its applicability in real scenarios. An example is the data from [1], which can be directly adopted by our model but currently unavailable. [1] David Blanco Mulero, et al. Benchmarking the Sim-to-Real Gap in Cloth Manipulation. IEEE Robotics Autom2024 # Limitations: Comparison with HOOD Please refer to W2. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response and the additional experiments you conducted. I appreciate the effort you have put into addressing my concerns. Regarding W1, I understand that the focus of your work is on modeling constitutive relations, and that the energy-based simulation is used primarily as a tool to achieve this goal. While the computational intensity of the energy optimization process remains a consideration, I acknowledge that accelerating this process is beyond the scope of your current research. It may be beneficial to include a brief discussion in the manuscript to clarify this distinction for the readers. Concerning W2, I appreciate the implementation of HOOD for further comparison and the comprehensive explanation of your chosen baselines. Your inclusion of HOOD, alongside the detailed breakdown of the garments in the Cloth3D dataset, enhances the robustness of your experimental validation. The additional PDF provided in the global rebuttal was also helpful in understanding the nuanced differences between the methods. In response to Q1, the clarification on the efficiency and accuracy of your EUNet compared to traditional physics-based models is noted. Your explanation of how EUNet simplifies the modeling process by capturing observed constitutive laws directly is compelling. This is indeed a strength of your approach, and the results showing higher accuracy further reinforce the value of your method. Finally, regarding Q2, the limitation of using a single piece of cloth is acknowledged. The potential impact on the range of deformations and dynamics captured is a valid consideration, and your proposed solution of applying known forces and using multi-layered clothing is a practical approach to addressing this. I appreciate the intentional design of your training data to ensure its real-world applicability, which is an important aspect of your work. Overall, the author's responses have clarified many of my concerns, and I recognize the contributions your work makes to the field. However, after careful consideration, my overall assessment and rating of the paper will remain the same. --- Reply to Comment 1.1.1: Title: Official Comment by Authors Comment: We thank the reviewer for the positive feedback and recognition of our contributions and efforts. Regarding W1, existing techniques for optimizing garment dynamics with 2k to 12k vertices driven by human bodies with 7k vertices, such as HOOD, take approximately 26 hours on an NVIDIA Quadro RTX 6000. We will clarify the relationship between our contributions and the energy optimization process in the revised version. We hope the reviewer will reconsider raising the score, as we have addressed your concerns. Please let us know if there are any remaining issues.	Rebuttal 1: Rebuttal: We thank the reviewers for their insightful feedback. We emphasize our contributions and clarify the main points as follows. To mimic the dynamic patterns from observed clothes, some methods [1, 2] focus on estimating the PHYSICS PARAMETERS that best fit the known analytical models or simulators. In contrast, we aim to explore the intrinsic ENERGY FUNCTIONS that describe the constitutive relations governed by general physics laws, such as the Lagrangian mechanics, and animate garments constrained by our learned EUNet. Our work is significant in the following ways 1. our disentangled learning scheme reduces the great need of large scale of garment data to mimic garment dynamics, and relies on only a piece of cloth as training data; 2. our EUNet is able to directly capture the constitutive behaviors from observed trajectories that are relatively easy to obtain in real life, and is highly generalizable thanks to the topology-independent designs; 3. garment animations constrained by our EUNet deliver lower errors and superior performance comparing with baselines. In the additional PDF file, we 1. add HOOD [3] for comparisons and demonstrate the effectiveness of the mini-network mentioned at L290-292 for HOOD; 2. apply our method in real application and animate the realistic garments from CAPE [4]. Below we list the key components of the tables. # Time Efficiency Comparisons \| \| PHYS \| EUNet $\Phi_p$ \| EUNet $\Phi_d$ \| EUNet $\Phi_p+\Phi_d$ \| \| ---- \| ---- \| ---- \| ---- \| ---- \| \| Time (ms) \| 1.743 $\pm$ 0.212 \| 1.055 $\pm$ 0.202 \| 1.216 $\pm$ 0.223 \| 2.271 $\pm$ 0.422 \| We denote the StVK elastic model and the bending model by "PHYS", which is used in the main text. As formulated in Equation (1), our EUNet is composed of two separate branches: $\Phi_p$ for potential energy and $\Phi_d$ for dissipation. Both branches have the same structures. We report the time separately for each branch and the full EUNet as above. The forward time is averaged on 80 frames of predictions, which include garments composed of 7924 vertices, 23636 edges and 15712 faces. All experiments are run on NVIDIA A100-SXM4-80GB. Our EUNet is comparable to the traditional clothing model in terms of speed. # Comparisons with HOOD \| Methods \| Overall Euclidean Error (mm) \| Collision (%) \| \| ---- \| ---- \| ---- \| \| MGN-H+PHYS+F (HOOD) \| 84.85 $\pm$ 29.93 \| 0.46 $\pm$ 0.99 \| \| MGN-H+EUNet (Ours) \| 66.39 $\pm$ 39.36 \| 0.44 $\pm$ 0.48 \| We denote the hierarchical graph neural network used in HOOD as MGN-H. HOOD adopts the analytical clothing model (PHYS) and friction (F) between human body and garments as loss terms. As shown in the table, MGN-H constrained by our EUNet achieves lower errors comparing with HOOD, suggesting the effectiveness of our EUNet. [1] Carlos Rodríguez-Pardo, et al. How Will It Drape Like? Capturing Fabric Mechanics from Depth Images. CGF2023. [2] Eunjung Ju, et al. Estimating Cloth Simulation Parameters From Tag Information and Cusick Drape Test. EG2024. [3] Artur Grigorev, et al. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. CVPR2023 [4] Qianli Ma, et al. Learning to Dress 3D People in Generative Clothing. CVPR2020 Pdf: /pdf/01d15d6862bbe7d410fc958746381744b4ccc38b.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper proposes to learn garment dynamics using a disentangled learning framework and the Energy Unit Network (EUNet). Instead of relying on extensive garment datasets, the approach learns constitutive behaviors from a single cloth piece and dynamically animates garments through energy optimization. Strengths: The writing is clear and technical details are described clearly. The visual aids and diagrams are well-integrated, enhancing understanding. Weaknesses: My main problem with the paper is that the problem of learning/recovering cloth dynamics from structured sample tests has been studied intensively for a long time, and the authors seem not to be aware of this whole field. This is a well-studied problem, and the authors do not position their method against significant prior works. Many existing works have also attempted learning from real-world fabric sample tests or indirect representations (video), which is a much harder problem. To mention a few: 1. "Predicting the Drape of Woven Cloth Using Interacting Particles" Breen et al., 1994 2. "Estimating Cloth Simulation Parameters from Video" Bhat et al., 2003 3. "Data-driven elastic models for cloth: Modeling and measurement" Wang et al., 2011 4. "How Will It Drape Like? Capturing Fabric Mechanics from Depth Images" Rodriguez-Pardo et al., 2023 5. "Estimating Cloth Simulation Parameters From Tag Information and Cusick Drape Test" Ju et al., 2024 The authors should thoroughly review the literature and reposition their contribution and provide experimental comparisons against existing works. Additionally, the literature review section does not include important works from the physics simulation community. Including these references and discussing how the proposed method builds upon or differs from them would strengthen the paper significantly. Technical Quality: 1 Clarity: 2 Questions for Authors: 1. Literature Positioning: Can you clarify your awareness and positioning of your method in relation to the existing body of work on learning/recovering cloth dynamics from structured sample tests? As discussed above, many significant studies in this area were not referenced. 2. Physics Simulation/Graphics Community References: The literature review section did not include important works from the physics simulation and graphics community. How does your approach relate to or differ from the significant contributions in this field? Including a discussion of this could provide a better contextual grounding for your work. 3. Experimental Validation: Can you provide more details on how your experiments validate the proposed method against these existing works? Specific comparisons and metrics would help clarify the effectiveness and novelty of your approach. Can you validate your approach against different datasets, including synthetic datasets generated from different simulation engines as well as real-world datasets used by current works? Confidence: 4 Soundness: 1 Presentation: 2 Contribution: 1 Limitations: The authors provided a discussion on the limitations of edge-wise discretization and a lack of self-collision handling. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed feedback. # W1: Insufficient Literature Review We argue that our focus is quite different from Physics Parameter Estimation. The misunderstanding may come from the similarity in data format, where a piece of hanged cloth deforms given external forces. Unlike physics parameter estimation, our EUNet aims at learning the unknown ENERGY FUNCTION that describes the constitutive laws. While in physics parameter estimation, the constitutive models are known a priori and carefully chosen. Specifically, with known constitutive models, physics parameter estimation aims at learning physics parameters, such as the lame constant in St. Venant-Kirchhoff elastic model for the target material. Complex devices and procedures, such as KES-F [1], the FAST system [2], and the Cusick drape test [3] are commonly required to estimate the parameters that best fit the analytical cloth models or simulators. Moreover, for a specific material, such as cotton, different clothing models or simulators require separate procedures or pipelines to estimate different sets of physics parameters. In contrast, our EUNet directly captures constitutive behaviors of different materials, which takes deformation attributes (i.e., delta of length and bending theta) and material attributes (i.e., stiffness and damping coefficients) as input, and outputs the corresponding potential energy. The material attributes remain constant in our study. As discussed at L84-88, our EUNet does not need existing analytical physical models and physics parameter estimation. In terms of literature review, as mentioned above we focus on learning to animate garments where modeling constitutive behaviors is the key challenge, we thus mainly discuss the garment animation and constitutive laws in the literature review section at L108-126 and L127-136 respectively. Energy-based physics simulation serves as a tool to generate dynamics constrained by our EUNet, and we briefly introduce the corresponding techniques at L137-144. We will clarify the relationship between our focus and physics parameter estimation in the revision. [1]. Kawabatak, et al. Fabric performance in clothing and clothing manufacture. Journal of the Textile Institute. 1989 [2]. Minazio, et al. FAST–fabric assurance by simple testing. IJCST1995 [3]. Cusick, et al. The dependence of fabric drape on bending and shear stiffness. Journal of the Textile Institute Transactions 1965. # Q1: Literature Positioning Please refer to W1. # Q2: Physics Simulation/Graphics Community References As discussed in W1, our focus is on how to learn the constitutive laws, which is discussed at L127-136 and involves several references from graphics community. We verify our EUNet by applying existing physics simulation techniques, which we briefly discussed at L137-144. # Q3: Experimental Validation As discussed in W1, our work is orthogonal to physics parameter estimation, thus comparing with methods mentioned by the reviewer is unnecessary. In addition, as discussed in W1 and Q2, our goal is to faithfully animate garments with observed materials where modeling constitutive relations is the key challenge. We thus verify the effectiveness of our EUNet by integrating with simulation techniques and evaluate the garments animations on Cloth3D [1], which is a large scale dataset and commonly used in garment animations, such as DeePSD [2] and MotionGuided [3]. [1] Hugo Bertiche, et al. CLOTH3D: Clothed 3D Humans. ECCV2020 [2] Hugo Bertiche, et al. DeePSD: Automatic deep skinning and pose space deformation for 3D garment animation. ICCV2021 [3] Meng Zhang, et al. Motion guided deep dynamic 3D garments. TOG2022 --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal and for clarifying the focus and contributions of your work, particularly the distinction between your approach and traditional physics parameter estimation. Based on your clarifications, I will be raising my score. Given that your method aims to learn constitutive models, I believe it would be beneficial to include demonstrations on real data to further substantiate the practical applicability of your approach. This addition could strengthen the impact of your work. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for the reply. Our work aims at learning to animate garments, where modeling the underlying constitutive behaviors is the core challenge. We actually already provide demonstration on real data. Specifically, although real data for learning constitutive laws are not publicly available yet, simulators constrained by our EUNet, which learns the constitutive behaviors on simulation data, can directly be used to animate real garments in CAPE [1], as shown in Figure 2 of the global rebuttal. In addition, it is standard practice to benchmark different methods on simulation data in both physics simulation [2, 3] and cloth(garment) animation [4, 5, 6, 7]. [1] Qianli Ma, et al. Learning to Dress 3D People in Generative Clothing. CVPR2020 [2] Tobias Pfaff, et al. Learning mesh-based simulation with graph networks. ICLR2021 (Outstanding Paper) [3] Yitong Deng, et al. Fluid Simulation on Neural Flow Maps. TOG2023 (Best Paper) [4] Chaitanya Patel, et al. TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style. CVPR2020 [5] Hugo Bertiche, et al. CLOTH3D: Clothed 3D Humans. ECCV2020 [6] Meng Zhang, et al. Motion guided deep dynamic 3D garments. TOG2022 [7] Xingxing Zou, et al. Cloth4d: A dataset for clothed human reconstruction. CVPR2023	null	null	null	null	null	null
Policy Aggregation	Accept (poster)	Summary: This paper joins a long list of recent work that studies how to aggregate the preferences of several agents (e.g., humans) in a reinforcement learning framework inspired by social choice theory. The problem is modeled as a multi-objective MDP with $n$ different reward functions. The authors propose to use the state-action occupancy measure instead of each agent's most preferred policy or reward function directly. Popular concepts from social choice theory, such as the Borda count rule and approval voting, are then studied from this perspective. Strengths: - The paper is well-written and easy to read. - It appears that considering the state-action occupancy measure has some advantages over working directly with each agent's optimal policy or reward function when attempting to introduce social choice rules, which---even though a standard approach in RL---is interesting. Weaknesses: - In my opinion, the contributions of this work are limited. E.g., only the full information case is studied - The primary justification of this work (which is repeatedly mentioned in the paper) is that prior work on policy aggregation and fair RL is not invariant to affine transformations of the reward function. Essentially, agents can have differently scaled reward functions, which makes, e.g., maximizing for social welfare a bad objective. However, I don't understand why we cannot simply normalize the reward function of each agent, so that the reward functions are directly "comparable". I find the concern about affine transformations quite weak. Technical Quality: 3 Clarity: 3 Questions for Authors: - I'm a bit surprised about the title of the paper since you're not aggregating policies, but reward functions. In fact, the policies play a minor role in the paper, since you look at the preference relation induced by the reward function (which you then express in term of occupancy measures). Could you explain why what you're doing is policy aggregation and not just preference aggregation? Typo: "Policy Aggergation" in title of Section 5 Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are adequately addressed in my opinion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > The primary justification of this work (which is repeatedly mentioned in the paper) is that prior work on policy aggregation and fair RL is not invariant to affine transformations of the reward function. Essentially, agents can have differently scaled reward functions, which makes, e.g., maximizing for social welfare a bad objective. However, I don't understand why we cannot simply normalize the reward function of each agent, so that the reward functions are directly "comparable". I find the concern about affine transformations quite weak. We disagree with this comment and believe we can provide an effective rebuttal. As we mention in lines 34-37, there is a rich literature in economics about the shortcomings of interpersonal comparison of utility. Note that there is nothing about our setting that makes normalization of utilities especially compelling, so your proposed solution is equally relevant in every setting that involves interpersonal comparison of utility. If normalization "worked," therefore, interpersonal comparison of utility would be a nonissue. There are several arguments against normalizing utilities, but instead of a philosophical discussion, perhaps an example would be most useful. Suppose there are three alternatives, agent 1 has utilities (0, 1/2, 1) and agent 2 has utilities (2, 3, 4). Should the normalized utilities of agent 2 be (1/2, 3/4, 1), or should we first subtract 2 and then divide by the maximum, yielding (0, 1/2, 1)? Or should we only subtract 1? Or perhaps we should normalize so that the sum of utilities is equal, giving (0, 1/3, 2/3) for agent 1 and (2/9, 1/3, 4/9) for agent 2? Economists argue that there is no principled way to answer these questions. > I'm a bit surprised about the title of the paper since you're not aggregating policies, but reward functions. In fact, the policies play a minor role in the paper, since you look at the preference relation induced by the reward function (which you then express in term of occupancy measures). Could you explain why what you're doing is policy aggregation and not just preference aggregation? The way we imagine the eventual pipeline is that we would observe trajectories from the optimal policy of each agent, learn reward functions via inverse reinforcement learning, and apply our techniques to select a policy. In this sense, we're aggregating the individual optimal policies into a collective policy. We also like the use of the word "policy" as it suggests that we're dealing with MDPs, in contrast to "preference aggregation," which is much more general. That said, we'd certainly be open to changing the title if this is seen as a pivotal issue. --- Rebuttal Comment 1.1: Comment: Thank you for your response and sorry for my late reply. My concerns are addressed and I can align with the other reviewers on a score of 6. Regarding the title, in my opinion, something along the lines of “socially fair reward aggregation” would better capture the point of the paper. I think, it’s fine either way, and it seems that I was the only one who expected slightly different content given the title (I was expecting direct policy aggregation).	Summary: The paper solves the problem which arises in preference aggregation of individual policies to a collective policy – (1) summation based aggregation are sensitive to affine transformations and (2) voting rule based aggregation faces problem of policies being exponential in S. Towards solving this, the paper proposes voting over continuous space of alternatives (which eliminates affine sensitivity) and a volumetric definition of preference ordering. The paper next proposes efficient algorithms to (1) find approximate volumetric veto core and (2) approximate q-Quantile Fairness. They also show complexity of existing voting rules, notably plurality voting and borda count. They show that problem is computationally hard for plurality and open for broad count. I am inclined towards accepting the paper. Strengths: The paper solves a well-motivated problem of policy aggregation. Their proposal of achieving different notions of approximate fairness through efficient algorithms both novel and appears to be sound. Their theoretical analysis of the complexity of using plurality and borda count based voting is also significant and allows scope for future work in the direction. Their algorithms have been validated through experiments. Weaknesses: In the experimental section, using a common metric to quantify the “level of fairness” guaranteed by different algorithms would be beneficial for a more learned comparison. In Def. 4 should the expression be vol(O’)/vol(O) >= 1 – veto(S) + epsilon instead of vol(O’) >= 1 – veto(S) + epsilon as currently stated? Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness. Why can't yours be a special case of Noothigattu et al. [27]? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Authors adequately address the problem statement and discussed limitations/candidate improvements of their work which is left to future work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > In Def. 4 should the expression be vol(O’)/vol(O) >= 1 – veto(S) + epsilon instead of vol(O’) >= 1 – veto(S) + epsilon as currently stated? Yes, you are right. Thank you for catching this typo. > Why can't yours be a special case of Noothigattu et al. [27]? Noothigattu et al. assume that there's a ground-truth reward function and the agents' reward functions are random perturbations thereof. Consequently, they essentially propose to treat all the data as coming from a single agent and directly apply an inverse reinforcement learning algorithm to the pooled data. Our setup, by contrast, calls for pluralistic rules that take diverse preferences into account. Perhaps a good analogy is the shift that has been happening in the last year from reinforcement learning from human feedback (RLHF) methods that rely on the Bradley-Terry model and essentially assume there is one type of person to "pluralistic alignment" methods where social choice plays a major role and diverse preferences are respected. See these recent position papers for more details: Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, Yejin Choi. A Roadmap to Pluralistic Alignment. arXiv:2402.05070, 2024. Vincent Conitzer, Rachel Freedman, Jobst Heitzig, Wesley H. Holliday, Bob M. Jacobs, Nathan Lambert, Milan Mossé, Eric Pacuit, Stuart Russell, Hailey Schoelkopf, Emanuel Tewolde, William S. Zwicker. Social Choice Should Guide AI Alignment in Dealing with Diverse Human Feedback. arXiv:2402.05070, 2024. > In the experimental section, using a common metric to quantify the “level of fairness” guaranteed by different algorithms would be beneficial for a more learned comparison. The current plots show the utility of different quantiles (10%, 20%, ..., 100% for the experiment with 10 agents and 20%, 40%, ..., 100% for the one with 5 agents), which provides a holistic view of the utility distribution. The bars for the worst-off agent correspond to the egalitarian welfare. In the revised manuscript, we will include additional plots with information on different fairness metrics such as the Gini coefficient and Nash welfare. If the reviewing team also has other suggestions, we would gladly consider them.	Summary: This paper studies aggregating multiple policies–which can be seen as a formalization of the task of aligning an AI system to the values of multiple individuals. When the number of states is small (such as, when multiple individuals have to select one out of a few candidates), this problem has been widely studied in voting and social choice theory, and there are many efficient aggregation rules (such as the Borda count). This paper, however, considers the other extreme: where the state-action space is huge, and it is not obvious how to design efficient methods to aggregate policies. The main insight of this paper is that preferences over policies have a volumetric interpretation in the state-action policy space that, in some cases, leads to efficient aggregation algorithms. Concretely, the authors examine two types of aggregation rules: (1) two aggregation rules that are known to have desirable fairness properties (namely, proportional veto code and, the recently introduced, quantile fairness) and (2) voting or score-based rules such as Borda count and $\alpha$-approval voting rule. Building on their insight the authors prove several results, including 1. an algorithm which finds the policy wrt an $\epsilon$-approximation of the proportional veto core using $O(log(1/\epsilon))$ queries to a volume computation oracle, 2. the existence of $q$-quantile fair policies for all $q\geq 1/e$ (which is tight and stronger than the best possible bound in the discrete case), 3. NP-hardness and inapproximability results for $\alpha$-approval score. Strengths: The paper is well-written and easy to read. I believe that the problem proposed in the paper is well-motivated from alignin AI systems, and is of significant interest to research on voting rules and social choice theory. The theoretical results are solid and the proofs and/or approach are well outlined. Finally, I did not check the proofs in detail, but they appear sound. One caveat is that I am not familiar with the closely related prior work (e.g., [6]) and, so, cannot comment on the novelty of the proofs and results from prior work. Weaknesses: I am not sure if the empirical results section is adding any value to this paper: it evaluates different aggregation rules, but I think this is not the focus of this work–I think the focus is to design efficient algorithms and/or prove existential results. If other reviews and the area chairs agree, my suggestion is to drop the empirical results section and use the additional space to add more exposition on the proofs. To be clear, this is no a significant concern for me. Technical Quality: 4 Clarity: 4 Questions for Authors: I do not have any specific questions for the authors. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: I do not see significant limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > I am not sure if the empirical results section is adding any value to this paper: it evaluates different aggregation rules, but I think this is not the focus of this work–I think the focus is to design efficient algorithms and/or prove existential results. If other reviews and the area chairs agree, my suggestion is to drop the empirical results section and use the additional space to add more exposition on the proofs. To be clear, this is no a significant concern for me. We agree that the main contribution and focus of our work is theoretical, that is, efficient algorithm design and existential results as mentioned. We are open to moving the empirical section to the appendix and elaborating on the theory, including the proofs, in the main body. We do see the empirical results as contributing in two ways: (1) The aggregation rules are practical. They can be easily implemented and run fast (on not-so-small instances). (2) In addition to theoretical arguments, empirical evaluation illustrates the differences in policies generated by the proposed rules, e.g., some of the proposed rules result in policies that are empirically "fairer" to different stakeholders in complex settings. --- Rebuttal Comment 1.1: Title: Response to authors Comment: I thank the authors for their response. I will keep my previous score and think this paper should be accepted.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Auditing Local Explanations is Hard	Accept (poster)	Summary: The paper addresses the challenges in verifying the accuracy of local explanations for machine learning models, especially when the model is not fully known and access to it is limited. The primary focus is on minimizing the number of times the model and explainer are accessed during the auditing process. The contributions of the paper are as follows: C1. Defining Auditing of Local Explanations: The paper provides a formal definition of what it means to audit local explanations. It sets the groundwork for systematically evaluating the reliability of these explanations, which are crucial for understanding and trusting machine learning models. C2. Importance of the Region’s size: It highlights that the region to which the local explainer applies is a critical property. Understanding the scope and limits of where the explanation is valid is essential for accurate auditing. This insight helps in identifying when and where an explanation might fail to represent the model correctly. C3. Bounds on Auditing Complexity: The paper establishes both lower and upper bounds on the sample complexity required for auditing local explanations. These bounds are presented as functions of the newly identified property, which is the region’s size. This provides a theoretical framework for understanding the minimal and maximal data requirements for effective auditing. Strengths: S1. Framework for Auditing: It proposes a theoretical framework for auditing local explanations, which is a significant step towards developing more rigorous and reliable methods for verifying the trustworthiness of explanations provided by machine learning models. S2. Identification of Key Metrics: The introduction of the explainability loss function, provides a quantifiable measure for evaluating local explanations for the original model, offering a systematic way to assess explanation quality. S3. Highlighting the Importance of Locality: The analysis which provide upper and lower bounds highlights the importance of the "locality" of explanations, bringing attention to a previously underexplored aspect in the explainability literature. Weaknesses: W1. Lack of Evaluation: The paper does not include an evaluation on real-world datasets. Although the authors suggest that their results could have significant practical implications (“Our results might have far-reaching practical consequences”), an initial step should be to perform evaluations on actual data to validate their findings. W2. Limited Discussion with Previous Research: The paper could benefit from a more thorough discussion of how its findings relate to and build upon previous research in the field. Specifically, the authors mentioned that [Dasgupta 2022] is the most similar to their work, except it is limited to discrete explanations rather than continuous. However, it is not clear whether, in the discrete setting, the proposed work aligns with Dasgupta’s consistency metric, sufficiency metric, or if it does not coincide with any of these previous metrics. It may also be worth discussing and comparing with the recent work of [Bassan 2023], which suggests a verification method for finding minimal sufficient explanations. - Dasgupta, S., Frost, N. and Moshkovitz, M., 2022. Framework for evaluating faithfulness of local explanations. In International Conference on Machine Learning - Bassan, S. and Katz, G., 2023. Towards formal XAI: formally approximate minimal explanations of neural networks. In International Conference on Tools and Algorithms for the Construction and Analysis of Systems W3. Focus on a Specific Type of Explanations: In the paper, the presentation is a bit misleading because it suggests that explanations can be general. However, the focus is on one type of explanation – those that approximate the true model on a region of examples. This is not a general (local) explanation method, even though it includes quite a few types of explanations. Technical Quality: 3 Clarity: 3 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: -- Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their thoughtful review. We will respond to each weakness in order. W1. We agree that empirical validation is an important direction for further work, and will emphasize this in the final version of our paper. We chose to take a theoretical focus because the main result of our work is a lower bound that applies to any algorithm. W2. We thank the reviewer for the added reference, and will include a more detailed comparison with prior works in the final version of our paper. With regards to Dasgupta et. al., it is indeed correct that the notion of "local consitency" (Definition 1) provided in their paper precisely matches our notion of "local loss" (Definition 2.3) when the region they define, $C_\pi$, is construed as a local explanation region. However, a key difference with their work is that ours provides lower bounds on the amount of data needed to estimate the local loss (over the entire distribution), and this result is the main point of our paper. With regards to Bassan and Katz, we agree that their work provides an interesting idea for provably correct local explanations (thus, in some cases, circumventing the need for an audit). By contrast, the results in our work are designed to apply to any local explanation scheme. W3. We agree that our work is limited to a certain type of "local explanation" and doesn't encapsulate explanation methods such as SHAP. As we explained in our general rebuttal, we will make sure to carefully outline the scope of our paper in the final version. We will also emphasize in the final paper that studying other types of local explanation methods is an important and interesting direction for future work. --- Rebuttal Comment 1.1: Comment: I have read your responses and am satisfied with them. Overall, I believe the paper provides interesting contributions and could be a good fit for NeurIPS 2024. Adding a discussion on related work and clarifying the supported type of explanations will further improve the paper.	Summary: This work studies an auditing framework in the eXplainable Artificial Intelligence (XAI) area. Specifically, the authors consider the scenario where a group of third-party auditors or users attempt to perform a sanity check on the provided explanations. The framework allows the auditors to query the model prediction and local explanations. Based on the proposed framework, this paper presents a theoretical analysis of the sample complexity of auditing. Strengths: This paper targets a very important aspect of XAI studies. It considers the deployment phase where users do not trust the provided explanations. This is a usually overlooked perspective in the XAI community. Weaknesses: 1. This work focuses on local explanations defined in section 1.1 L48-49 and section 2.2 L155-162. These presumptions limit the scope of this paper to the surrogate model method (such as LIME, MUSE [1], etc.), where a glass-box explainer is used to approximate black box’ predictions in the neighborhood of input data. This greatly limits the impact of this work as such surrogate-model explanation methods only take a very small part of local explanation methods. Local explanations are not limited to surrogate model methods. It can refer to explanations regarding individual input samples instead of the entire data manifold or even regarding the model itself [2]. 2. In the context of this paper, the authors claim that gradient-based explanations are surrogate model explanation methods (i.e. “local explanation method” under the definition of this paper) in L181-188. The authors define that $g_x(x) = (\nabla_xf(x))^Tx$, which is the summation of input x gradient attributions. This corresponds to the prediction $f(x)$ only if $f$ satisfies homogeneity [3]. On the contrary, suppose $\phi_f(x)\in\mathbb{R}^d$ is the explanation of SHAP, then $g_x(x):=(\phi_f(x))^T\mathbf{1} = f(x)$ can accurately reflects the prediction. Therefore, the definition of the explainers studied in this work is ambiguous and may require more rigorous considerations. In summary of points 1 and 2, the formulation of the framework in this work is flawed. 3. There is no empirical verification of the proposed theoretical results, which significantly undermines the contribution of the theoretical analysis. Note that a continuous function is always bounded on the closed neighborhood of x. Therefore, it is essential to empirically test whether the proposed bounds are tight. A theoretical demonstration is also appreciated. 4. [minor] To stay consistent, “Lime” in L337 should be revised to “LIME”. 5. While the motivation that users/auditors may not trust the explanation system and want to audit the model is an interesting and realistic setup, the proposed framework lacks practical contributions. Specifically, the formalism described in L64-66 and L236-242. can be difficult to satisfy. Reference [1] Lakkaraju, H., & Bastani, O. (2020, February). " How do I fool you?" Manipulating User Trust via Misleading Black Box Explanations. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (pp. 79-85). [2] Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., & Kim, B. (2018). Sanity checks for saliency maps. Advances in neural information processing systems, 31. [3] Hesse, R., Schaub-Meyer, S., & Roth, S. (2021). Fast axiomatic attribution for neural networks. Advances in Neural Information Processing Systems, 34, 19513-19524. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. L51-52: Why do the authors claim LIME to be a gradient-based method? 2. Definitions 2.1 and 2.2 are limited to black box $f:\mathbb{R}^d\rightarrow \{\pm 1\}$. Can this constraint be relaxed to more general settings? For example, is it limited to a binary decision boundary? 3. The main theoretical results of Theorem 4.1 have many presumptions that are not justified. For example, why is the “user-specified” local error threshold assumed to be 1/3? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The limitations of this work are discussed in L194-197, where the authors admit that the definition “local explanation” is narrowed down to fit a. I agree with the limitation and appreciate the authors for clearly stating this issue. However, my concern is that this can be severe and greatly undermine the application scenarios of this work. More details are in the weakness section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their review. We will begin by addressing the listed weaknesses 1. As we mention in our global rebuttal, we respectfully disagree that the considered explanations are too limited. LIME and Anchors are both reasonably utilized methods in practice, and more generally we believe our general formulation captures a meaningful subset of explanation methods. 2. We will clarify further what we meant in lines 181-188 in the final paper. We never state that $g_x(x) = (\nabla_x f(x))^t x$. We rather meant that $g_x$ should be some linear classifier that matches these coefficients. In particular, we would also fit a bias term so that $g_x$ agree with $f$ at the point $x$. Additionally, as mentioned in our global rebuttal, our framework does not consider SHAP. 3. We have left an empirical validation of our results for future work. We believe that the theory provided is sufficient for meaningfully arguing the difficulties posed by local explanations that are too local. We view the results of this paper as serving as a "theoretical demonstration." 4. Thank you, this will be fixed. 5. Could you please elaborate on what you found lacking in our framework? Additionally, what precisely about our framework can be "difficult to satisfy." Please see our global rebuttal for our general argument for the utility of our framework. Next, we will address the direct questions 1. We will correct this to be more precise in our general paper. While technically speaking, LIME does not use a gradient, we believe that at a high level it uses similar ideas to Gradient based methods by providing a local "linear-like" approximation to the general classifier. 2. We believe the binary classification setting is sufficiently relevant to be independently considered. However, we note that our lower bound directly extends to more general classification settings as binary classification is a subtask of multiclass classification. We also believe that our lower bound can be relatively easily adapted to regression by simply replacing the 0-1 loss used to define the local loss with an MSE based loss. 3. Could you specify specifically which assumptions are not justified? Furthermore, the user-specified threshold $\gamma$ is not set to $\frac{1}{3}$, it is instead assumed that $\gamma \leq \frac{1}{3}$. This is an extremely mild assumption considering that a local loss above $\frac{1}{3}$ would be considered egregiously inaccurate in most settings. --- Rebuttal Comment 1.1: Title: Thanks for the Responses Comment: I appreciate the authors' responses and I clarify the questions as follows. W5: This originally refers to the settings of the auditing framework in real-world applications. For example, the requirement for numerous explanations from the auditee. But I agree this is a relatively minor point that does not involve technical issues. Q3: This question is not regarding the conditions being too strong, but to ask why are they chosen as those values. This work focuses on realistic settings of problems. Therefore, discussions on the choices of parameters and their values can be beneficial and provide more insights. I have raised the scores accordingly. However, I disagree that empirical verifications should be left for future work. Given the realistic problem setups and motivations of this work, empirical validations of the theoretical results are very important. Besides, many of the claimed discoveries should not be difficult to verify. For example, "auditing an explainer requires an amount of data that is inversely proportional to its locality". For the completeness of the work, I still strongly encourage the authors to provide empirical verification as a part of the work, even for synthetic data. I will raise the score to 6 if this concern is resolved. --- Rebuttal 2: Title: Response to reviewer Comment: We thank the reviewer for reading and responding to our rebuttal. While we agree that empirical validation is an important direction, we don't agree that it is a straightforward verification task. Our main result is a lower bound that holds for any general explanation method (note that this necessarily includes manipulative explanations where the explainer might create explanations that appear as though they come from a gradient based method but in reality are maliciously modified) along with any general auditing method. Thus, experiments run for a specific pair of explanation and auditing methods do not serve as strong evidence for our theorem holding. Furthermore, the fact that local regions in high dimensional space can be extremely small implies that determining a ground truth value for the local loss itself could prove challenging. Due to these factors, we do not have enough time to include a meaningful empirical validation of our lower bound in this submission. Although we believe that our current results are sufficient for a complete paper, we respect the reviewers view that an empirical validation is a needed component.	Summary: The paper proposes an auditing framework to verify the truthfulness of explanations by a third-party in scenarios where there is no trust. Bounds on sample complexity are provided that depend on the locality (minimum local mass) of the explanation. Further, the authors discuss that for gradient-based explanations in higher dimensions, locality tends to be small to achieve a reasonable explanation loss. Smaller locality increases the provided bounds on the amount of data required for the audit. Strengths: 1. The topic of the paper is important for policymakers and the XAI research community in general, as it suggests that in natural scenarios where there is no trust, it is difficult to verify whether the local explanation is truthful to the model without actually knowing the model. 2. The paper provides upper and lower bounds on sample complexity for auditing local explanations. 3. The analysis includes gradient-based explanations and discusses how to generalize to other methods, including LIME and Anchors. Weaknesses: 1. Regarding the soundness of the auditing framework, could you please comment on the motivation for the company to provide all of the required data (especially local regions) to the auditor? When the requested dataset is sufficiently large, the third-party company could potentially recover the model along with all the classifier outputs, local regions, and other information. 2. Figure 1 is hard to fully understand without reading the paper. It’s not intuitive which data points are explained and why, in panel (b), there is sufficient data for the audit. Could you please provide more explanation in the figure caption or simplify the figure? 3. Section 5 and Theorem 5.1 present an existence proof. However, the example considered in Figure 2 (a) is very specific. Can you elaborate on how often you expect this data distribution to occur in real-world datasets or discuss the value/range of locality for other likely data distributions? Minor: 1. $E(f, x_i)$ is used in Section 1 but defined in Section 2. 2. Please specify $\epsilon_1$, $\epsilon_2$, $\delta$, and $\gamma$ in Theorem 4.2 or mention that you are operating under the same conditions as in Theorem 4.1, if applicable. Also, Algorithm 1 is referenced before it is defined. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Do you have any guidance or suggestions on how users should choose the local error threshold, $\gamma$? 2. How useful are the bounds in Section 4 for reasonable values of locality and low explanation loss, such as when the decision boundary is nearly locally linear? In this case, can you provide any empirical estimates on the lower and upper bounds in Theorems 4.1 and 4.2? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors discuss the limitations of the paper with respect to the explanation algorithms. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed review. We will first respond to the listed weaknesses. 1. As we mention in our global rebuttal, we believe our setting reflects cases where an Auditor might have access to a set of explained cases by default. For example, applicants to a bank for loans could, in principle, pool their explanations to provide a dataset with which an Auditor could operate within our setting. We will discuss further in the final version of our paper, and will include examples of cases where our setting could plausibly apply. 2. We appreciate the feedback on figure 1, we will certainly add more detail to the caption and simplify the figure by removing the green decision boundary and including enough points to visually demonstrate the concept of "enough data for auditing." 3. The key issue posed by Figure 2 is the curvature in the data manifold. Thus, we believe that cases with high dimensional and highly curved data manifolds would pose similar issues as the decision boundary involved would seldom be linear. A detailed empirical investigation of this is left as a direction for future work. 4-5. Thank you, these will both be fixed. Regarding the direct questions: 1. We believe $\gamma$ to be largely application based, and for this reason chose to leave it as a user-specified parameter. However, one basic guiding principle might be to set $\gamma$ similarly to the general loss of the classifier being explained. For example, it doesn't seem necessary to require local explanations with a corresponding local loss of 0.001% for a classifier with a 30% misclassification rate. 2. Our local and upper bounds are designed for arbitrary, potentially complex classifiers. We chose this as the default starting point for theoretical analysis. We believe that studying more restricted cases of classifiers could be very interesting and promising. For example, we believe that ensuring a classifier satisfies some degree of smoothness (i.e. close to locally linear decision boundaries) could plausibly greatly reduce the amount of data needed for auditing. --- Rebuttal Comment 1.1: Comment: Thank you to the authors for the response. I have read the authors’ reply and the other reviews, and I will keep my score.	Summary: This paper provides theoretical results on how many queries are required for an auditing framework for local explanations of machine learning algorithms (e.g., neural networks). Strengths: The paper is well motivated with a widely interesting and relevant topic. The approach is theoretical, and rigor is provided through proofs in appendices. The authors connect their work to popular algorithms: Gradient based approaches, Lime, and Anchors. Weaknesses: The authors acknowledge that the local loss estimation is, in their argument, only a necessary but not sufficient condition for trust. They do not establish or reference any evidence that manipulation would result in the local loss as a good indicator of untrustworthiness. As a result, the analysis serves more as a potential validation scheme for the limited types of algorithms that meet their linearity requirement (e.g. Anchors or LIME). This drastically narrows the scope and implications of their analysis. Unless this can be firmly established the title, abstract, and conclusions of the paper should be amended to reflect the correct scope of its claims Furthermore, there are plenty of reasons (and examples) where interpretation methods are demonstrated to be frail to the input (e.g. Ghorbani). This would likely not pass the audit but would not be evidence that the $E$ has been altered. I think this speaks to some confusion in the setup of the paper as to what “trust-worthiness” is. The authors present it as trust between the user and provider rather than trust in the robustness of the explainability metric which is, I argue, closer to what their results seem to reflect. Additionally, it is not clear to me that this is the only way to test explainability metrics with limited access (e.g. data points, classifier outputs, and local explanations). For example, these metrics are popular because they match so well to human “expert” knowledge. You show someone the pcitre of the shark from Smilkov et al and they agree that they see “shark-like” shapes in the SmoothGrad output. Consequently, one could imagine the case where sampling in a non-local fashion would trace whether the same features matching human “experts” appear. Ultimately the proposed methodology seems entirely impractical (which is sort of the point). Technical Quality: 3 Clarity: 3 Questions for Authors: - Can you establish or provide a reference hat show how the local loss / explainability loss is a good indicator for manipulation from an adversarial attack (or disingenuity design)? This is critical for the scope and implications of your analysis. - Please comment on whether the proposed auditing scheme is indeed the only way to establish a local explainer (fitting your constraints) taking into account the suggestions above. - Can you comment on how the results might - L 117 please change reference to Shap to the acronym SHAP - L 337 please change reference to the Lime algorithm to its acronym: LIME - L 440 where does the 2592 come from in the denominator for the bound on $n$? - How well would the framework work if you chose some K examples to audit around instead of drawing sample i.i.d.? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - Analysis is for a binary classifier Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the review and the detailed questions. In order 1. As we mention in our global rebuttal, we do not believe the local loss is sufficient for preventing adversarial attacks. We rather believe that maintaining a low local loss is one of many necessary components required for a trustworthy explanation. We intentionally do not define trustworthiness, we are rather just showing that auditing one part of it can be difficult from a data perspective. 2. Our scheme isn't intended to encapsulate ways to "establish a local explainer." It is rather meant to reflect plausible settings where an Auditor might seek to audit explanations provided by some entity that desires to protect information about their models. 3. Could you finish this question in a comment? We'd be happy to respond during the next period of the review. 4-5. Thank you, we will fix these. 6. This is a constant chosen simply for mathematical reasons. It is merely designed to make the algebra within our bounds to work out and is not tight (as our goal is to simply provide an asymptotical lower bound). Replacing it with a different constant would simply result in changing values of $\epsilon$ and $\delta$. Note that for small values of $\lambda$ (in high dimensional data), this constant would be easily dominated by $\lambda$. 7. This is an interesting question. This would depend a lot on the manner in which the K examples are chosen. While it appears that querying samples from a single local region would allow for very quick estimation of the local loss, we believe that ensuring that these examples are "in-distribution" would pose a technical challenge. In particular, the auditor would have to have some notion of what constitutes a "realistic" example. We believe this is an interesting direction for future work. We would additionally like to respond to some of the other points raised in the "Weaknesses" section. As we mention in our global rebuttal, we don't feel that our setting is too limited to bear relevance, and we also aren't quite sure what the "linear assumption" being referenced is. As we mentioned in answer 1 and in our global rebuttal, we do not define trust as the local loss. We rather argue that a low local loss is necessary but not sufficient for there to be trust. Finally, as we mentioned in answer 2, our paper is designed to examine a natural default way in which a set of explanations might be audited. In fact, one of the implications of our results is precisely the necessity for other methodologies. --- Rebuttal Comment 1.1: Comment: I appreciate and have read the response. My apologies for any confusion regarding the unfinished question.	Rebuttal 1: Rebuttal: We thank the reviewers for their detailed and thoughtful reviews. It appears that there are 3 main points of contention regarding our paper. First, that the set of "local explanations" being considered is either too limited or not carefully enough analyzed, second, that the local loss is not a good indicator of a model's trustworthiness, and third, that the mode of limited interaction between the auditor and the explanation provider is not sufficiently realistic. We will address these three points separately. We define a local explanation broadly as a local classifier coupled with a local region. We will further clarify in our paper that this does NOT include "local" methods such as SHAP (in which no local region is provided). However, this definition does include LIME, Anchors, and certain kinds of gradient based methods (where a gradient implies that a linear classifier is implicitly being used). We believe that these listed methods are sufficiently used by practitioners to be worth consideration. While our work only considers linear and constant local classifiers, we believe this to be a strength rather than weakness as our work is centered on providing a lower bound for auditing (which means our bounds will carry over to more complex classes of local classifiers as well). With regards to the local loss, we will further clarify (as we did in line 55 of our paper) that maintaining a low local loss is a necessary but not sufficient condition for being a trustworthy explanation. We completely agree that a low local loss does not imply that an explanation is trustworthy. We are rather claiming that it is self evident that explanations with a large local loss are intrinsically meaningless and untrustworthy. Because of this, our auditing framework encapsulates one necessary component that any auditing framework most address, and this means our lower bound has implications beyond our setting. In particular, circumventing the difficulties posed by our lower bound must require one of our assumptions to be broken which would require either further access for the auditor, or a different kind of explanation being provided. Finally, our mode of interaction between the explainer and the auditor is inspired by cases of collective action. In principle, we believe that an auditor could obtain a sample of explanations by aggregating outputs provided across a large sample of users. For example, in the case where a bank provides local explanations for loan approvals to applicants, an auditor could (with proper consent) aggregate the explanations provided for a sample of applicants and audit the bank. Note that this would not require any further information from the bank beyond the explanations it is already required to provide.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Predicting Ground State Properties: Constant Sample Complexity and Deep Learning Algorithms	Accept (poster)	Summary: The paper presents two novel machine learning algorithms for predicting ground state properties of quantum systems with constant sample complexity, independent of system size. The first algorithm modifies an existing ML model, while the second introduces a deep neural network model, both showing improved scaling in numerical experiments. Strengths: 1. The introduction of a deep learning model with rigorous sample complexity bounds is a significant contribution to the field. The constant sample complexity, regardless of system size, is particularly noteworthy and addresses a critical challenge in quantum many-body physics. 2. The authors provide numerical experiments that validate the theoretical claims. The experiments demonstrate the practical effectiveness of the proposed algorithms, especially the deep learning model, which outperforms previous methods. Weaknesses: 1. The training objective for the neural network is non-convex, which poses challenges in finding a global optimum efficiently. The paper does not address how to overcome this issue or guarantee convergence to optimal weights. 2. While the paper claims improved computational complexity, the actual implementation details and computational resources required for the deep learning model are not thoroughly discussed. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The reviewer would appreciate if the authors could elaborate on how the performance of the deep learning model generalizes to Hamiltonians that extend beyond the specific cases examined in the numerical experiments. 2. Can the authors provide more insights into the practical implementation of their algorithms, particularly regarding the initialization and regularization procedures used during training? This will be helpful for readers reproduce the results of the paper. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Reviewer Comment: The training objective for the neural network is non-convex, which poses challenges in finding a global optimum efficiently. The paper does not address how to overcome this issue or guarantee convergence to optimal weights. Author response: To address non-convexity of the training objective, we refer to literature on overparametrized deep neural networks, Ref. [74], (see Lines 283-284 in our manuscript), where it is shown that Gradient Descent finds the global optimum in settings very close to ours. A similar result was proven for Stochastic Gradient Descent [Oymak et al., "Overparameterized Nonlinear Learning: Gradient Descent Takes the Shortest Path?" PMLR 2019]. Those results can likely be adapted to our setting. Furthermore, the experimental results show that this does not seem to be an issue in practice. Reviewer Comment: While the paper claims improved computational complexity, the actual implementation details and computational resources required for the deep learning model are not thoroughly discussed. Author response: The computational requirements and implementation details for generating the training data and training are detailed in Appendix D.1. Moreover, we provide code for the experiments in https://anonymous.4open.science/r/anonymous-B093, as referenced in Section 4. Reviewer Question: The reviewer would appreciate if the authors could elaborate on how the performance of the deep learning model generalizes to Hamiltonians that extend beyond the specific cases examined in the numerical experiments. Author response: The performance of the deep learning model holds for any Hamiltonian satisfying the conditions outlined in Section 2.1. Namely, this is any $n$-qubit gapped geometrically-local Hamiltonian that can be written as a sum of local terms such that each local term only depends on a constant number of parameters. There are many physically relevant Hamiltonians where this holds. Some examples are the 2D random Heisenberg model (as considered in the numerical experiments), the XY model on an $n$-spin chain with a disordered $Z$ field, and the Ising model on an $n$-spin chain with with a disordered $Z$ field. Reviewer Question: Can the authors provide more insights into the practical implementation of their algorithms, particularly regarding the initialization and regularization procedures used during training? This will be helpful for readers reproduce the results of the paper. Author response: We will add that we used Xavier initialization [71] and no regularization. The latter is implicitly shown in the loss function we used [Equation D.3]. The code for implementing the numerical experiments is given in https://anonymous.4open.science/r/anonymous-B093, as referenced in Section 4. --- Rebuttal Comment 1.1: Comment: Dear Reviewer KHZF, The author-reviewer discussion period is ending soon. Please check if the authors’ response has addressed your concerns and feel free to adjust your score accordingly. If you find the authors’ response unsatisfactory, please explain your reasons and discuss them with the authors immediately. Best regards, AC --- Rebuttal 2: Comment: Thank you for your reply. I would like to keep my score.	Summary: In this paper, the authors focused on utilizing deep learning methods to predict the ground states. They made an important assumption that brings theoretical improvement to achieve constant sample complexity in the training data. They also made two main alternations to the learning model compared to previous literature, including incorporating Pauli coefficients in feature mapping and utilizing kernel ridge instead of Lasso. Numerical results for up to 45 qubit systems are provided, supporting the theoretical findings. Strengths: 1. High-quality paper with rigorous theoretical findings and comprehensive numerical results. 2. Improved the sampling overhead to constant complexity, independent of the system size. 3. Explored new possibilities for predicting the ground state properties of quantum many-body systems using neural network models. Weaknesses: The main concern is that the improvement of this paper against precious works is limited. The main theoretical finding is based on an additional assumption that we know the property we'd like to predict in advance. The proposed learning method has only two minor alternations. These issues prevent me from giving a higher evaluation score, but they do not overshadow the fact that this article is of high quality. Technical Quality: 4 Clarity: 4 Questions for Authors: No questions Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: See weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the careful reading of our paper. The reviewer's comments only apply to our first result in Section 3.1, and we acknowledge that the reviewer's comments are accurate for this portion of our paper. However, the improvement is significant - reducing sample complexity from logarithmic in system size to constant. Moreover the reviewer appears to have missed entirely our second result in Section 3.2, we obtain a rigorous guarantee for neural-network-based algorithms predicting ground state properties. - This is the first work to give a practical deep learning based approach for predicting ground state properties with both theoretical guarantees and good experimental performance, and there is nothing similar in previous literature. We introduce several new techniques to this problem setting, such as the use of the Hardy-Krause variation, which is likely to be widely applicable in future research. - This neural network result does not require knowledge of the property (unlike the result in Section 3.1). In addition, we would like to comment that knowing the property in advance is a natural assumption, as often scientists will have a particular observable in mind that they would like to study (as further discussed in lines 168-175). Moreover, since our algorithm in Section 3.1 also applies to classical representations of the ground state (such as classical shadows), this can also be used to mitigate the issue of having to know the observable in advance (as we can learn a classical shadow of the state and then use that to predict properties of an observable of interest which need not be known in advance). See corollary 1. --- Rebuttal Comment 1.1: Comment: Dear Reviewer MeLw, The author-reviewer discussion period is ending soon. Please check if the authors’ response has addressed your concerns and feel free to adjust your score accordingly. If you find the authors’ response unsatisfactory, please explain your reasons and discuss them with the authors immediately. Best regards, AC	Summary: This paper studies the sample-efficient learnability of properties of grounds states of local Hamiltonians. Ground states of local Hamiltonians are hard to compute, even for quantum computers and to circumvent this hardness, several recent works proposed learning the trace inner product of local observables with the ground state given labeled training data. This setting is exactly PAC learning, i.e. given labeled data from a worst-case distribution, the goal is to get low prediction error wrt the same distribution on future samples. The best sample complexity for this problem is known to be log(n) 2^{polylog(1/eps)}, shown by Lewis et al. The main questions addressed in this work are (1) whether the sample complexity can be improved to be independent of the system size and (2) whether there are rigorous guarantees for learning properties of the ground state via neural network based algorithms. Strengths: The paper provides several technical results on the representation and learnability of ground-state properties. The improved sample complexity follows from tweaking the algorithm is [2] and making additional assumptions about the training distribution. The neural net sample complexity result proceeds in two steps. First the authors prove an approximation-theoretic result for functions that look like ground state properties and show they can be well-approximated by neural networks. They then obtain a generalization bound using fairly sophisticated technical machinery. Weaknesses: I think the paper does not resolve the questions it claims to resolve and does so in a slightly camouflaged way. 1. Question 1 in the paper asks whether you can get sample complexity that is independent of system size for learning properties of ground states, aka the PAC learning setting for ground states of local Hamiltonians. The answer obtained is yes, under two crucial caveats, the observable is known in advance and the distribution over the training data is not worst-case. This diverges significantly from the PAC learning model. The same critique holds for Question 2. Further, reference [1] does not state this as an open problem. 2. The assumptions on the training distribution are not stated upfront and do not appear to be mild. The assumptions include the distribution g to be strictly non-zero on [-1,1], and zero outside; g being continuously differentiable; and component-wise independent. Are there any natural distributions that satisfy all these properties simultaneously? 3. There is no discussion of why each of these properties is needed, which ones are crucial to the argument and which ones are for technical convenience. Having 4 non-trivial technical assumptions of the pdf is a major weakness, especially since it makes the result incomparable to prior work [1,2], where the setting is truly PAC learning. 4. In the numerical experiments, I see no discussion of what distribution was used to generate the training data, and how many of the technical conditions this distribution satisfies. It remains unclear to me what the experimental section is trying to convey, since it does not complement the main theorems. 5. When is it reasonable to expect labeled data for the observable / property you want to learn? What is a real-world scenario where one would expect to obtain such labeled data? 6. Is there a way to get some non-trivial sample complexity (not necessarily system independent) bound for PAC learning via neural networks, without the extra assumptions on the training distribution? 7. There are completely unjustified claims such as the neural network achieving low loss and finding a bounded magnitude solution after constant many steps and O(n) time. It is not clear why this should ever be the case. Technical Quality: 3 Clarity: 2 Questions for Authors: I included the questions with the weaknesses. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the careful reading of our paper.\ First, we would like to clarify some statements made by the reviewer. We remark that our first answer to Question 1, namely our result discussed in Section 3.1, does not make any assumptions on the data distribution (as in the PAC learning model) but does require that the observable is known in advance. In contrast, our Neural Network (NN) guarantee in section 3.2 introduces assumptions on the data distribution but does not require the observable to be known. Note that the assumptions on the data distribution for the NN guarantee (stated in [Line 1336-1340]) are satisfied by natural distributions such as Uniform and Gaussian distributions on $[-1,1]^m$. Hence, they are milder than in Ref. [1], where the guarantees hold exclusively for the uniform distribution. Those distributions often suffice in physically relevant models such as the Sachdev-Ye-Kitaev (SYK) model, which requires Gaussian parameters. Moreover, given that there are few rigorous guarantees for NNs in the literature, we view our work as a step in this direction. In the following, we will do our best to address each point of criticism by the referee in detail. 1. Mostly addressed above. Ref. [1] does state that "Rigorously establishing that NN-based ML algorithms can achieve improved prediction performance and efficiency for particular classes of Hamiltonians is a goal for future work'' in the sentence right above Section III.B. 2. We agree with the reviewer that we could emphasize and further discuss the assumptions on the distribution, although we do state them in the main text in Lines 234-238. We want to further emphasize that $g$ needs to be continuously differentiable only on $[-1,1]^m$. Examples of natural distributions satisfying all conditions simultaneously are uniform and the normal distribution. The latter is important for physical models such as the SYK model. 3. We note that our assumptions on the distribution do not make our work incomparable with Refs. [1, 2]. Ref. [1] applies only to the uniform distribution over $[-1,1]^m$, which is a stronger restriction than ours. Ref. [2] covers any arbitrary distribution, similar to PAC learning, as our first result in Section 3.1. In Section C.3, Line 1344, we mention that assumptions (a) and (b) can be relaxed. Assumption (a) ensures strict positivity on $[-1,1]^m$ to avoid divisions by zero, but this can be managed by splitting the integral on $g$'s support. Requiring $g=0$ outside $[-1,1]^m$ is natural since it is the parameter space. Assumption (b) can be relaxed to continuous $g$ using mollifiers. Assumption (c) on component-wise independence is necessary to reduce the input domains of $f_P^{\theta_P}$. We will add this in Section C.3. 4. We briefly mention the distributions used in the caption of Figure 2 and Lines 354-355 in the main text and discuss this further in Section D.1, line 1628-1630. Namely, we utilize points from the uniform distribution over $[-1,1]^m$ (as in the numerics in Refs. [1,2]) and low-discrepancy Sobol sequences, complementing Theorem 14. We respectfully disagree with the comment that the numerics do not complement our theorems. We reproduce the setup of the numerical experiments in Refs. [1,2] and compare the performances of our model and the previously-best-known algorithm from Ref. [2]. The experiments illustrate that our method outperforms the method from Ref. [2] in practice (Figure 2, left). Moreover, they show that our assumptions in Theorem 5 are often satisfied in practice, i.e., small training error is achieved for proper choice of locality (Figure 2, right) and is independent of system size (Section D.2, Figure 4). 5. We briefly mention this in Lines 169-171, but agree it needs more discussion. The setting of receiving labeled data is consistent with Refs. [1, 2]. Our ML approach applies in scenarios like a scientist studying ground states of Hamiltonians $H(x)$. The scientist can use quantum experiments or classical simulations to investigate ground state properties. ML algorithms like ours and those in Refs. [1, 2] help extrapolate from existing data without additional costly experiments. In this scenario, it is realistic to assume the scientist can choose the training data, reducing the need to cover the worst-case distribution. Relevant examples of experiments generating data for quantum states include [A,B,C]. 6. As discussed in point 3, the only assumption that is truly necessary for the argument is Assumption (c) (component-wise independence). Allowing long-range correlations for the parameters of a geometrically local system seems unnatural, so we are unsure whether this request makes sense physically. In that case, our methods yield a similar bound as in Ref. [1]. 7. We justify those claims by referring to literature on overparametrized deep NNs Refs. [74], (see Lines 283-284 in the text), where it is shown that Gradient Descent finds the global optimum in settings very close to ours. A similar result was proven for Stochastic Gradient Descent [D]. Those results can likely be adapted to our setting. Moreover, our experimental results demonstrate that these claims are satisfied in practice in the setup we tested. The cost per iteration is $\mathcal{O}(n)$, due to the model's size. The model's size does not affect the number of gradient steps required for convergence (see e.g. Ref. [74], Theorem 5.1). ## References [A] Huang et al., "Quantum advantage in learning from experiments.", Science 2022 \ [B] Struchalin et al., "Experimental estimation of quantum state properties from classical shadows." PRX Quantum 2021\ [C] Elben et al., "Mixed-state entanglement from local randomized measurements." Physical Review Letters 2020.\ [D] Oymak et al., "Overparameterized Nonlinear Learning: Gradient Descent Takes the Shortest Path?" PMLR 2019 --- Rebuttal Comment 1.1: Comment: Dear Reviewer u4Mp, The author-reviewer discussion period is ending soon. Please check if the authors’ response has addressed your concerns and feel free to adjust your score accordingly. If you find the authors’ response unsatisfactory, please explain your reasons and discuss them with the authors immediately. Best regards, AC	Summary: This work builds upon the work of Huang et al. and Lewis et al. by introducing two new approaches to get constant sample complexity for predicting properties of a ground state of a many-body local Hamiltonian. The two new approaches are a modified ML model that requires knowledge of the property of interest and a deep neural network model that does not need prior knowledge of the property. In this paper, the authors provide both proves and small experimental evaluations to show that both approaches achieve constant sample complexity, independent of system size. Strengths: The paper is well-organized and clearly written. The paper includes rigorous theoretical guarantees and small numerical experiments to confirm the efficacy of the proposed methods compared to the existing algorithm. Weaknesses: Even though it is a strong paper, the issue addressed here is a specific case that builds upon two other papers. Additionally, the related published works are mostly, if not all, published in physics journals. I do not see why the results shared in this paper are valuable to share with the broader NeurIPS community, especially since the mathematical proofs are very rigorous, I would expect that it is not accessible to the broader audience. Some assumptions and conditions required for the theoretical guarantees may also limit the applicability of the results. Technical Quality: 3 Clarity: 3 Questions for Authors: How do the parameters or phases of the random Heisenberg model affect the training performance? Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are addressed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Our work aims to solve an important physics problem by leveraging machine learning. Thus, we expect it to be of broad interest to physicists, theoretical computer scientists, and machine learning practictioners, as our algorithms not only have rigorous proofs but are also readily implementable, as seen in the numerical experiments. - We identify the first result using deep learning for predicting ground state properties, with both theoretical guarantees and good experimental performance. This is a very practical algorithm accessible to the broad community and the code is available via an anonymous repo. The assumptions are natural and realistic (as discussed in other responses below) and conclusions are of direct practical relevance to physicists and chemists in their experimental work. - We highlight that many papers of similar levels of mathematical rigor on topics in physics/quantum information have been present at ICML and NeurIPS in the past. A few examplex are [A,B,C,D]. The recent ICML 2024 conference featured a workshop devoted to "AI for Science: Scaling AI for Scientific Discovery." In regards to the question about parameters/phases of the random Heisenberg model, could you please clarify what you mean by this? In the below, we try to answer to our understanding of your question, but please let us know if we misunderstood what you meant. The parameters of the Hamiltonian are inputs to the machine learning model (as a part of the training data). Although our rigorous guarantee only holds for training and testing points in the same quantum phase of matter (as in previous work), in the 2D random Heisenberg model, we may predict across phase boundaries. However, as seen in the numerical experiments, our algorithm still performs well. ## References [A] Yamasaki et al., "Learning with optimized random features: Exponential speedup by quantum machine learning without sparsity and low-rank assumptions." NeurIPS 2020\ [B] Aaronson et al., "Online learning of quantum states." NeurIPS 2018\ [C] Abbas et al., "On quantum backpropagation, information reuse, and cheating measurement collapse. NeurIPS 2024\ [D] Michaeli et al., "Exponential Quantum Communication Advantage in Distributed Inference and Learning." ICML 2024 --- Rebuttal Comment 1.1: Comment: I thank the authors for the response. I increased my rating for the paper.	Rebuttal 1: Rebuttal: We thank all the reviewers for their consideration and feedback. We're gratified to see appreciation from most reviewers: several described the work as important and appreciated the novelty of the deep learning approach with both theoretical guarantees and strong practical performance. We believe that the negative reviews are mainly the result of misunderstanding that we correct and clarify in the detailed rebuttals below: - the technical objections of 7tNg and u4Mp are the result of misunderstanding that we have clarified in detail below. Our first result does not make any additional assumptions on the observables nor on the input distribution (but only that the observable of interest is known, which is reasonable, and which can be ameliorated via the use of shadows as we clarify below). Our neural network guarantee holds for natural distributions such as uniform or Gaussian. We extend the tools from Ref. [54] to more general distributions than low-discrepancy sequences. By exploiting physical structure and deriving an explicit bound on its Hardy-Krause variation, we show that our neural network algorithm achieves theoretical improvements over Refs. [1, 2] with respect to sample complexity. Experimentally, in the settings tested in Refs. [1, 2], our neural network algorithm also shows practical improvements. We emphasize that the deep learning approach is the first such result on this problem. We believe that our methods apply to broader classes of physical systems. - Reviewer cLpw opines that it is a strong paper with mathematically rigorous guarantees but it comes from physics literature and is not relevant to the wider NeurIPS community. We note that AI for Science is one of the most active areas at premier ML conferences like NeurIPS and ICML today, and our work falls in this area. Our second result is also very practical (code available via anonymous repo) and shows strong empirical performance. Among the positive reviews, - Reviewer MeLw opines that the paper is of "high quality" and gives a score of 7, stating the main reason for not giving a higher score is that they consider the result somewhat incremental. Actually the improvement over previous work is significant - reducing sample complexity from logarithmic to constant. In addition, MeLw appears to have missed entirely, our second main result, namely the first deep learning approach to the problem which is not only practical and shows strong empirical improvement over previous work, and for which we also give rigorous results using innovative new techniques such as the Hardy-Krause variation with wide applicability.	NeurIPS_2024_submissions_huggingface	2,024	Summary: In this work, the authors give two algorithms that predict (geometrically local) properties of ground states of gapped geometrically local Hamiltonians. This problem has been introduced by Huang et al. [HKT+22], and the previous best known algorithm is given by Lewis et al. [LHT+24], which uses $\log(n)$ samples, where $n$ is the number of qubits in the Hamiltonian. This paper further improves on the $log(n)$ sample complexity, and gives two algorithms that only use a constant number of samples. The first algorithm is modified from the algorithm of [LHT+24], changing the regression part of the algorithm from LASSO to kernel ridge regression. The second algorithm uses deep neural network, having the advantage of not needing to know the observables in advance, but requires more restriction on the distribution of the Hamiltonian parameters. The authors complement their theoretical results with numerical simulations. [HKT+22] Huang, Hsin-Yuan, Richard Kueng, Giacomo Torlai, Victor V. Albert, and John Preskill. "Provably efficient machine learning for quantum many-body problems." Science 377, no. 6613 (2022): eabk3333 [LHT+24] Lewis, Laura, Hsin-Yuan Huang, Viet T. Tran, Sebastian Lehner, Richard Kueng, and John Preskill. "Improved machine learning algorithm for predicting ground state properties." nature communications 15, no. 1 (2024): 895. Strengths: The work achieves the optimal sample complexity of the problem and is written in good English. Weaknesses: The part of preliminaries that are restating definition and result of [LHT24+], is not well written, and I believe has led to a critical bug to the first algorithm. In particular, Theorem 8 claims that for every $O\sum_{P} \alpha_P P$ that can be written as sum of geometrically observables, $\sum_{P} \|\alpha\| =O(1)$. However, the counterpart in [LHT24+] has extra restrictions: $\|\|O\|\|_{infty}=1$ and $O$ need to be inside a radius of $R=O(1)$. Therefore, where the authors uses Theorem 8 in equation (B.28) to bound the kernel, the result is incorrect since they do not have $R=O(1)$. Other minor inconsistencies includes: line 642: $S^{geo}$ not defined line 660: $h_{c(j)}$ not defined Technical Quality: 1 Clarity: 2 Questions for Authors: Some typos: line 121: geometrically [local] observable line 148: \|\| \omega \|\| -> \|\| w \|\| Confidence: 2 Soundness: 1 Presentation: 2 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Bug The claim by the reviewer is incorrect, as our observables satisfy exactly the same conditions as those considered in [LHT+24].\ In particular, we state throughout the paper that - we only consider observables that satisfy $\lVert O\rVert_\infty \leq 1$, e.g., in lines 122, 225, 301, etc. - We state clearly (lines 121-122) that $O$ is an observable that can be written as a sum of geometrically local observables. In [LHT+24], requiring that $R = \mathcal{O}(1)$ is exactly the definition of geometric locality (see, e.g., Definition 5 in [LHT+24]). Thus, the two definitions are exactly the same and conditions of Corollary 4 in [LHT+24] still hold in our setting. We will clarify this explicitly in the final version. We define $S^{(\mathrm{geo})}$ later in line 654. We will make this change to introduce it earlier. The term $h_{j(c)}$ denotes the same as in [LHT+24]. This is discussed in lines 655-657 and we will make this explicit in the main text. We should also highlight our second result in section 3.2 which is completely different and uses deep learning techniques. This is a very practical algorithm (code available via anonymous repo) which has strong empirical performance and for which we give theoretical guarantees using innovative new techniques such as bounding the Hardy-Krause variation. --- Rebuttal Comment 1.1: Comment: Dear Reviewer 7tNg, The author-reviewer discussion period is ending soon. Please check if the authors’ response has addressed your concerns and feel free to adjust your score accordingly. If you find the authors’ response unsatisfactory, please explain your reasons and discuss them with the authors immediately. Best regards, AC --- Rebuttal Comment 1.2: Comment: I would like to thank the authors for the response. I have raised my score. --- Rebuttal 2: Comment: We thank 7tNg for acknowledging that their misunderstanding has been clarified and are happy that they are convinced of the soundness of the results. So, we wonder why the soundness score is still 1. Also, if they have any other issues we could address that leads to the overall score of only 5, we are happy to respond.	Summary: The authors propose an ML based method to predict properties of ground states of quantum systems which comes with provable guarantees. Improving on recent work by Huang et al and Lewis et al, they give sample complexity bounds which are independent of the number of qubits. This approach is applicable when the observable one is trying to predict is predetermined. The authors also suggest a deep learning based approach for the case where the observable is not known in advance. They support their theoretical wok with numerical experiments. Strengths: The paper adresses an important probelm, and is well written and argued. The authors clearly explain the previous state of the art in ML based prediction of ground state properties, as well as their own contribution. Their proposed modification to the procedure suggested by Lewis et al., which results in Theorem 1 of the paper, seems interesting and worthwhile. Likewise the guarantees obtained for the training of a custom Neural Network architecture are intriguing from a learning theoretic perspective. Weaknesses: It is unclear to me how the Neural Network generalization result compares to known results in the literature- the setting which the authors study is quite specific and thus it is not easy to relate the result they obtained to those in the theoretical deep learning literature. Technical Quality: 4 Clarity: 4 Questions for Authors: How restrivtive is the assumption about the dependence of each local term on a fixed number of parameters? It would be instructive to give physically relevant examples where this does and does not hold. Presumably the results would not apply to standard Neural Network architectures- what specifically would break in the analysis? How does the sample complexity depend on the constant the authors assume the network weights are bounded by? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: I believe the authors have adequately adressed the limitations of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the careful reading of our paper and their constructive comments. Reviewer Comment: It is unclear to me how the Neural Network generalization result compares to known results in the literature.\ Author Response: This is the first rigorous sample complexity bound on a neural network model for predicting ground state properties. We extend the tools from Ref. [54] beyond low-discrepancy sequences and explicitly bound the Hardy-Krause variation, such that our result is comparable to Refs. [1,2]. Reviewer Question: How restrivtive is the assumption about the dependence of each local term on a fixed number of parameters? It would be instructive to give physically relevant examples where this does and does not hold.\ Author Response: The assumption that each local term depends on a fixed, i.e., system-size independent, number of terms is not particularly restrictive. Since each local term only acts on a system size-independent number of particles in the system, it is natural that this also holds for its parametrization.\ Moreover, we note that this assumption is not something new to our work and was introduced in Ref. [2]. There are many physically relevant examples where this holds. A few examples are are - the 2D random Heisenberg model (as considered in the numerical experiments). - the XY model on an $n$-spin chain with a disordered $Z$ field. - the Ising model on an $n$-spin chain with with a disordered $Z$ field (see, e.g., [A]) . - the Sachdev-Ye-Kitaev (SYK) model, which requires Gaussian parameters. We thank the referee for this comment and will add such examples to demonstrate its wide applicability. [A] Lieb et al., "Two soluble models of an antiferromagnetic chain." Annals of Physics 1961 Reviewer Question: Presumably the results would not apply to standard Neural Network architectures- what specifically would break in the analysis? Author Response: - As mentioned in the discussion [starting from line 1660], the practical performance of our model can most likely be improved by applying the vast body of recent results in the theory of deep neural networks. - Approximation results like those we used for tanh neural networks also exist for the ReLU activation function [53] and require neural networks of depth $\mathcal{O}(\log{\frac{1}{\epsilon}})$. Directly bounding the respective Hardy-Krause variation does not work however, since ReLU does not have bounded partial derivatives around $0$. It may be possible to work around this issue by bounding it with mollifiers, but we leave this open for future work. - We believe that our analysis can be extended to standard architectures such as convolutional or residual neural networks, as long as the width and number of hidden layers does not exceed the that needed for fully connected neural networks (so that the Hardy-Krause variation is $\mathcal{O}(2^{\mathrm{polylog}(1/\epsilon)})$. We limited our attention to fully connected networks here, as bounding the Hardy-Krause variation of more elaborate architectures would have been more tedious than instructive. Reviewer Question: How does the sample complexity depend on the constant the authors assume the network weights are bounded by?\ Author Response: One can see this dependence in Eq. (C.93) in the appendices. --- Rebuttal Comment 1.1: Comment: I would like to thank the authors for their detailed response. I will leave my score unchanged.	null	null	null	null
Large Spatial Model: End-to-end Unposed Images to Semantic 3D	Accept (poster)	Summary: The authors proposed the Large Scene Model (LSM), a novel 3D scene understanding framework that unifies multiple vision tasks within a single model. LSM represents a scene using pixel-aligned point maps, integrating geometric, appearance, and semantic information into a unified representation. By leveraging a Transformer architecture with cross-view and cross-modal attention, the model effectively incorporates multi-view cues and semantic knowledge from a 2D vision model. LSM's design enables efficient scene-level 3D semantic reconstruction and rendering in real time on a single GPU. The model's integration of a 2D semantic model allows for open-vocabulary understanding, extending its applicability to diverse real-world scenarios. Furthermore, by consolidating multiple tasks within a single model, LSM minimizes error propagation, leading to more robust and accurate results compared to state-of-the-art baselines. Strengths: * The described technical approach in this work is sound and clearly presented. The contributions from the various proposed modules are well ablated and investigated in the experiments (Table 4). * The model demonstrates high inference efficiency compared to other approaches. with reconstruction time of 0.1s and rendering at 270 due to the underlying 3DGS representation that is being generated. * I like that the model reconstructs the underlying 3D representation in a single feedforward pass, as compared to multiview + test time optimization for fusion approaches. This improves the speed and efficiency for inference. It is good to see compelling quality based on the novel view synthesis. Weaknesses: * I think the main contribution of this paper is the unification of the various scene modeling tasks into the same model, including geometry, color and semantics. The authors further claimed in the abstract and introduction that multitask training end-to-end allows LSM to outperform state-of-the-art baselines. However the paper did not ablate the multi-task learning design choice. For instance, what if some of the tasks are removed (e.g., semantic feature prediction). How does that affect the performance of the other tasks? * A suggestion is that for Figure 5, it is unclear how much pose divergence there is between the input source view and the synthesized novel view. It would be helpful to also show the source view supplied as input to the model. * The paper is named Large Scene Model, which seems to suggest something to do with model parameter scaling, hence large. However the paper does seem to do much scaling on model size. So perhaps a more accurate terminology would be Multitask or Unified Scene Model? Nits. * Line 153: Typo: to serve as input? * In Tables 1-4, I suggest highlighting the best (and possibly second-best result) for easier comparison of the various experiments. * In Table 4, why is + Multi-scale Fusion indented? Technical Quality: 3 Clarity: 3 Questions for Authors: - For the quantitative comparisons given in Figure 3, are they predicted from the input view, or are they for a novel view? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - My understanding is that in this method, all the Gaussians being generated are pixel aligned with the original input images. Is that a limitation of the method, since that would make the model unable to model large pose divergences that require rendering regions not originally visible in the input view, for instance, the back side of a sofa etc. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer 4 (8Lhg) for recognizing the contribution of our paper and offering insightful comments. Please find our response to the feedback below. [W1]: Ablate the multi-task design choice? We ablate the “novel view feature synthesis (Eq.4)” and “geometry prediction (Eq.1)” tasks in the following table. The row labeled “no Feat.” indicates the removal of the multi-scale fusion of LSeg features (green part in Point-wise Aggregation of Fig. 2). We observe that incorporating semantic task learning into point-wise aggregation enables the model to lift view-inconsistent feature maps to a view-consistent feature field, while it does not impact view synthesis and depth tasks a lot. We believe the input to “Point-wise Aggregation” contains pixel-wise point maps (x, y, z) and RGB (r, g, b) from the support images, which already provide the complete observation within the input views for interpolating novel views and depth prediction. Thus, additional incorporation of feature maps does not provide visual cues to enrich the radiance field. Removing geometric supervision leads to the model diverging quickly, which means pixel-aligned depth regression helps the model recognize correspondences across input views, guiding the model's learning when it is unaware of the camera parameters. The experiment setting follows Table 4 of the main draft. \| Methods/Metrics \| mIoU ↑ \| Acc. ↑ \| rel ↓ \| τ ↑ \| PSNR ↑ \| SSIM ↑ \| \|---------------------------\|--------\|--------\|--------\|---------\|--------\|--------\| \| Full Model \| 0.599 \| 0.8125 \| 4.09 \| 61.39 \| 21.10 \| 0.7752 \| \| Full Model (no Feat.) \| None \| None \| 4.18 \| 60.84 \| 20.97 \| 0.7745 \| [W2]: Update Figure 5, and the pose divergence for the novel views? We will include the source (support) views in the revision; thank you for the suggestion. We specify the rendered novel camera positions and orientations by interpolating between the first and second reference (or support) images for free-view exploration (see the supplementary webpage). Specifically, the orientation interpolation is conducted using Spherical Linear Interpolation [1] between the two sets of quaternions, while the position interpolation is conducted using linear interpolation. [W3]: More accurate terminology? We will take your suggestion into account and revise it accordingly. [Q1]: Typos and highlighting the table? We will fix it. [Q2]: Why is “+ Multi-scale Fusion” indented in Table 4? Indented row (Multi-scale Fusion) means it is based on the rows above (Fuse LSeg Feat. and Fuse Encoder Feat.), indicating that we apply multi-scale fusion to both. We have clarified this part in the revision. [Q3]: Input view or novel view in Figure 3? They are novel views. [Limitation] Thanks for pointing this out. Our method, along with all the adopted methods (pixelSplat, Feature-3DGS, NeRF-DFF), belongs to the category of reconstruction models, which focus on building the 3D based on input observations. A discussion about extending to generation for synthesizing new viewpoints with large pose divergences will be included in our revision. Reference: [1]. Slerp from wikipedia (we cannot provide the direct link in author response) --- Rebuttal Comment 1.1: Comment: Thank you for your clarifications and further ablations on LSeg features and thanks for considering my suggestions for adding the source view and revise the "Large Scene Model" terminology. Overall I will maintain my rating after checking notes with the other reviewers. --- Reply to Comment 1.1.1: Title: Responses to Reviewer's Comment Comment: We are grateful for your positive recommendation and the supportive feedback on our paper.	Summary: This paper presents the Large Scene Model (LSM), which generates semantic radiance fields from uncalibrated RGB images using a unified Transformer-based framework. LSM can infer geometry, appearance, and semantics simultaneously and synthesize label maps in real-time. The model integrates multi-scale fusion and features from 2D models to enhance accuracy and efficiency. Strengths: 1. Unified Framework: LSM combines multiple 3D vision tasks into a single framework, streamlining the process and reducing complexity. 2. Real-time Performance: The model achieves real-time 3D reconstruction and rendering, suitable for applications needing fast processing. 3. Enhanced Feature Fusion: By incorporating 2D model features, LSM improves the quality of feature lifting and semantic understanding, enhancing overall performance. Weaknesses: 1. Dataset: I recommend the authors organize the training and testing phases in alignment with previous methods (NeRF-DFF and Feature-3DGS) and provide results on the Replica Dataset. The authors have not sufficiently justified deviating from the baseline evaluation split. Furthermore, an explanation is needed for the significant performance discrepancy of the baselines between the Replica Dataset and the authors' setup. Additional training details may also be necessary. 2. Writing: The paper's abstract, introduction, and methods sections require improvement. Specifically, the methods section should introduce each module and their interconnections from a high-level perspective rather than presenting them as isolated components. 3. Method Details: Do the authors use camera parameters? If so, why are camera parameters mentioned in line 117? If camera parameters are used, the model cannot be described as "unposed." 4. Visualization: In Figure 4, there are category colors that are not listed in the legend. Additionally, a more diverse set of results should be displayed, as the current experimental set predominantly features sofas. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Module Timing: I am curious about how the authors manage to use eight cross-attention modules and still achieve reconstruction in 0.1 seconds. Please provide the time consumption for each step. 2. Image Resolution: What is the resolution of the images? More details regarding the inference process should be provided, especially concerning the time comparison. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have discussed limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer 3 (MCB6) for recognizing the contribution of our paper and offering insightful comments. Please find our response to the feedback below. [W1.1]: Reasons for deviation from Replica, and performance discrepancy. The reasons for the deviation are threefold: 1). The processed Replica datasets (by Feature-3DGS) lack ground-truth depth maps for geometric accuracy. Feature-3DGS uses COLMAP for SfM results, leading to misaligned coordinate systems. 2). Feature-3DGS uses LSeg as GT for evaluating semantic quality, while we use annotated GT semantic labels. 3). Replica datasets are simulated under ideal environments, whereas our draft focuses on real-world textures, lighting changes, and sensor noise. The performance discrepancy of Feature-3DGS stems from resolution and semantic ground-truth differences. Our experiments follow the generalizable NVS method, pixelSplat, using 256x256 resolution and cropped boundaries, reducing the valid field-of-view. We also use annotated ScanNet for semantic segmentation, rather than LSeg as ground-truth. [W1.2]: Results on Replica Dataset with explanation. Comparing our methods with NeRF-DFF and Feature-3DGS is challenging due to their assumptions of dense views, long optimization steps, and precomputed camera poses, while ours uses sparse and pose-free views. pixelSplat also assumes known camera parameters. This places our method at a disadvantage, but Tab. 1 in our submission demonstrates its strength. To further validate, we compare with Feature-3DGS on the Replica dataset, using Semantic-NeRF [1] preprocessed data with annotated RGB-D-Semantics (256x256 resolution). We randomly sample 2,000 points from GT depth maps for initializing Feature-3DGS and use 70 images for training and 10 for testing, following their setup. LSM generalizes well from the real-world ScanNet test set (25.44dB) to the simulated Replica test set (23.10dB) in zero-shot, even without GT camera parameters. Our model builds the feature field in a feed-forward pass, achieving high quality. Notably, our method achieves the best geometric accuracy without using epipolar attention (pixelSplat, need pose) or requiring test-time optimization with dense views and GT camera parameters (Feature-3DGS, need pose). \| Methods/Metrics \| Support View \| GT Camera Parameters \| Training Time \| mIoU (Seg.) ↑ \| rel (Depth) ↓ \| PSNR (RGB) ↑ \| \|------------------\|---------------\|----------------------\|---------------\|---------------\|---------------\|--------------\| \| Feature-3DGS \| 70 \| Required \| 18.5 min \| 0.62 \| 8.25 \| 31.89 \| \| Nerf-DFF \| 70 \| Required \| 1.1 min \| 0.49 \| 14.33 \| 24.67 \| \| pixelSplat \| 2 \| Required \| 0.06 sec \| None \| 20.14 \| 26.28 \| \| Ours \| 2 \| Not Required \| 0.09 sec \| 0.51 \| 4.91 \| 23.10 \| To be more aligned with our assumed setting, which are more practical in real applications, we proposed to benchmark the performance of investigated baselines under the sparse view setting. We reduce the support view number for all methods from 70 to 20, while preserving the same test sets. In the table below, Feature-3DGS drastically decrease the PSNR with degraded geometry and semantic accuracy. \| Methods/Metrics \| Support View \| GT Camera Parameters \| Training Time \| mIoU (Seg.) ↑ \| rel (Depth) ↓ \| PSNR (RGB) ↑ \| \|------------------\|---------------\|----------------------\|---------------\|---------------\|---------------\|--------------\| \| Feature-3DGS \| 20 \| Required \| 18.5 min \| 0.46 \| 32.44 \| 19.61 \| \| Nerf-DFF \| 20 \| Required \| 1.1 min \| 0.39 \| 45.09 \| 16.27 \| \| pixelSplat \| 2 \| Required \| 0.06 sec \| None \| 27.13 \| 20.62 \| \| Ours \| 2 \| Not Required \| 0.09 sec \| 0.45 \| 5.614 \| 18.80 \| [W1.3]. Training details. We included training details in Sec. 4.1. We use the default configuration to train Feature-3DGS and NeRF-DFF. For pixelSplat, we use its pretrained checkpoint and code without any modifications to perform the evaluation. [W2]:Detailed module design. We will add a diagram to illustrate module interconnections in the revision. Additionally, we plan to release the source code for reproduction. [W3]:Why use camera parameters in L117? They are needed in training and evaluation to obtain the ground-truth point maps, and specify the target view. They are not required for inference. [Q1]: Additionally, a more diverse set of results should be displayed. Please refer to supplementary webpage for ten free-view videos. We also incorporate additional comparisons in the Fig.3 of the attached PDF. [Q2]:Module Timing? Image resolution? We study the computational cost (time) of each module by inferring 1,000 times with the model and obtaining the average inference cost for each model. \| Module \| Inference Time (seconds) \| \|--------\|--------------------------\| \| Dense Geometry Prediction (Sec.3.1) \| 0.0297 \| \| Point-wise Aggregation (Sec.3.2) \| 0.0464 \| \| Feature Lifting (Sec.3.3) \| 0.0199 \| \| Total \| 0.096 \| The tested resolution (256x256) is aligned with the generalizable 3D-GS method pixelSplat. pixelSplat is slightly faster than our framework 0.06s while it requires additional steps to run Structure-from-Motion (e.g., 20.52s for each scene on ScanNet on average) to obtain the camera parameters. --- Rebuttal Comment 1.1: Comment: Thanks for your great efforts! After reading the response, some major issues have been addressed well, so I still lean towards positive for the submission. I encourage the author to add these clarifications to the main paper. --- Reply to Comment 1.1.1: Title: Responses to Reviewer's Comment Comment: Thank you very much for your positive feedback and for recognizing our efforts in addressing the major concerns. We are glad that our clarifications have been helpful, and we will certainly incorporate them into the main paper as you suggested. We are always open to further suggestions or feedback that could help us improve the paper even more. If there are no additional concerns, we would greatly appreciate it if you could kindly consider raising the rating. Thank you again for your valuable input.	Summary: The paper aims to train a network that takes in a set of unposed images and directly produces a semantic radiance field. The method utilizes a single Transformer-based model that learns the attributes of a 3D scene represented by a point-based radiance field. A decoder produced 3D Gausians that can be splatted to make novel images, depth estimates, and semantic segmentations. Strengths: The paper provides a transformer architecture for producing 3D gaussians with rich features from unposed images, which seem very valuable. The design choices in the proposed system are well-chosen from methods available at this time, leading to a system that has a good combination of little-compute and competitive-accuracy on three different tasks (nvs, depth, semantics). Weaknesses: The paper shares goals and ideas with "Scene Representation Transformers" (Sajjadi et al., CVPR 2022) and its follow up work Object Scene Representation Transformer (NeurIPS 2022) and RUST: Really Unposed SRT (CVPR 2023). This paper is different, because it ultimate produces a set of gaussians rather than a LLFF or NeRF volume, and it distills features from 2D foundation models. However, it is similar in that a transform encoder and decoder produces a scene representation directly from a set of images, which is then used for novel view synthesis, depth estimation, and semantic segmentation. In any case, those paper seem fairly similar in concept and so I think they should be discussed in the related work, and possibly approach sections. The ablation study in table 4 suggests that the key methods in the paper have little impact on the results of NVS. Technical Quality: 3 Clarity: 3 Questions for Authors: It is interesting that the 3D methods, which have access to multiple views of the same scene do not perform as well as LSeg in Table 1. This is counter-intuitive. Can you please explain why? The results on multiview depth accuracy are kinda amazing. Why is the proposed method better than ones that take the camera parameters? Is it due aligning the scene scale (do all the methods get the same method for scale alignment? The novel view synthesis images look very good. Can you please provide some info about how close the novel cameras are to the reference ones provided at inference time? Is there a way for you to quantify and compare to PixelSplat the NVS results as the novel cameras deviate further and further from the reference ones? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are discussed briefly in the second paragraph of the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer 2 (fir7) for recognizing the contribution of our paper and offering insightful comments. Please find our response to the feedback below. [W1] Discussion w/ SRT and RUST. Scene Representation Transformer (SRT)[1] and RUST[3] have pioneered the exploration of representing multiple images as a "set latent scene representation" (OSRT[2] utilized Slot Scene Representation) and generating novel views even in the presence of flawed camera poses or without any pose information. In contrast, our method targets a holistic model for semantic 3D reconstruction from unposed images, using point-based representation for efficient rendering speed. We will include this discussion in the revision. [W2] Table 4 suggests little impact on the results of NVS. To formulate the well-represented radiance field, the Point-wise Aggregation module takes the pixel-aligned (x, y, z, r, g, b) as input (line 153) for regressing point-wise attributes, which already contain the complete appearance observation within the input views for interpolation. Thus, additional incorporation of feature maps does not provide visual cues to enrich the radiance field. However, the Point-wise Aggregation module design is important as it enables the lifting of features from inconsistent stereo image features to consistent feature fields, and the Multi-scale Fusion further improves the lifting accuracy. [Q1] Novel view semantic segmentation with LSeg? We notice that LSeg fails to produce view-consistent segmentation results but generates reasonably good per-view segmentation results. Our hypothesis is that LSeg is a well-trained 2D model for language-driven semantic segmentation, serving as a foundation for our method to understand 3D. Our method lifts the view-inconsistent multi-view feature maps from LSeg into a consistent 3D feature field in a zero-shot manner, utilizing the Point-wise Aggregation module. See the figures in the attached PDF (Fig.2) to see the consistency difference. [Q2] High multiview depth accuracy induced by aligning scene scale? Yes, the improved geometry accuracy stems from the alignment and combination of scenes, and our training data consists of dense depth annotations of various scenes for supervision. [Q3] How to define novel cameras? Can they deviate from the reference one? Thank you for acknowledging the good rendering quality. The rendered novel camera positions and orientations are determined by interpolation between the first and second reference (or support) images. Specifically, orientation interpolation is conducted using Spherical Linear Interpolation [4] between the two sets of quaternions, while position interpolation is conducted using linear interpolation. We acknowledge that both our method and pixelSplat are reconstruction methods that build the radiance field using pixel-aligned Gaussians for interpolation, and thus cannot perform generation tasks that significantly deviate from the reference images. Reference: [1]. Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations, CVPR 2022 [2]. Object Scene Representation Transformer, NeurIPS 2022 [3]. Rust: Latent neural scene representations from unposed imagery, CVPR 2023 [4]. Slerp from wikipedia page (we cannot provide the direct link in author response) --- Rebuttal Comment 1.1: Title: Response to rebuttal Comment: Thanks for responding to my questions in the rebuttal. The answers are pretty-much as expected. I will leave my review score unchanged. --- Reply to Comment 1.1.1: Title: Responses to Reviewer's Comment Comment: We are grateful for your positive recommendation and the supportive feedback on our paper.	Summary: This paper solves the sparse-view scene reconstruction problem by Large Scene Model, a unified scene reconstruction model via unposed RGB images. The model utilizes a ViT backbone for extracting the feature and uses cross-view attention to align the multi-pose feature for consistent features. The 3D scene is further rendered from the 3D semantic field derived by the multi-view features. The unified model is capable of multiple 3D-based tasks including novel view synthesis and 3D language-based segmentation. Experiments showed that the work achieves better results with limited performance sacrifices in the NVS task and higher performance in the multi-view language-based segmentation task. Strengths: 1. The model is general and multi-purpose in sparse-view scene reconstruction. 2. The model can achieve better results while still obtaining lighting-fast rendering speed and can be applied to real-time reconstruction. Weaknesses: 1. The technical contribution is limited. The model is generally designed via multi-purpose modules glued attention and Transformers, which is a straightforward and widely applied idea. There is no significant new problem has arisen and novel solutions proposed. 2. The performance comparison with NVS-related works is limited. Firstly, the authors train and run comparison experiments on the same dataset, which can be biased. Secondly, several popular scene datasets incorporated in similar works (such as RealEstate10k) are not utilized in this work. Thirdly, methods similar to pixelSplat such as Splatter Image[1] are not included in comparison. 3. The presentation can still be improved. Firstly, the authors titled their work “Large Scene Model”, while the design is more similar to the idea of pixel-based Gaussian splatting (such as pixelSplat and GaussianImage). Secondly, each module's input and output data type cannot be directly recognized from the pipeline graph. 4. The bibliography of this paper lacks some related works, such as Splatter Image[1], which is also an image-space Gaussian splatting method. Reference: [1] Szymanowicz, Stanislaw, Chrisitian Rupprecht, and Andrea Vedaldi. "Splatter image: Ultra-fast single-view 3d reconstruction." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10208-10217. 2024. Technical Quality: 1 Clarity: 2 Questions for Authors: 1. Did the authors try replacing the designed module with large-scale pretrained models, such as using pretrained monocular depth estimation model? 2. The design philosophy is similar to multi-view image generation works. Can this model output high-quality and consistent multi-view images, as the fashion of Free3D? 3. The term of “language-driven segmentation” is not quite clear to me. Does it mean semantic segmentation? Confidence: 3 Soundness: 1 Presentation: 2 Contribution: 3 Limitations: The authors claimed that the major drawback of the model is the VRAM consuming. The social impact of this work mainly related to potential misuse of 3D assets and can be solved via integrating watermarks into the generated result. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer 1 (K6pm) for recognizing the contribution of our paper and offering insightful comments. Please find our response to the feedback below. [W1]: No significant new problem has arisen and novel solutions proposed? We acknowledge that our work builds upon the contributions of many giants, such as DUSt3R [1] (end-to-end point map regression model), pixelNeRF [2] (one of the first generalizable NeRFs), pixelSplat [3] (one of the first generalizable 3D-GSs), and the concept of large-scale training. However, we respectfully disagree with the assertion that "no significant new problem has arisen and novel solutions proposed." We target a very practical yet under-explored scenario: end-to-end semantic 3D reconstruction directly from images. While previous literature reduces this problem to “Computing Camera Parameters” followed by “Training 3D Representation”. We, for the first time, unify these problems into one differentiable vision system. Technically, while there are many alternative design choices, our method successfully integrates these components into an end-to-end trainable system using a standard Transformer architecture, demonstrating the feasibility of future scalable 3D training. We kindly suggest that the methods and ideas presented in this paper carry significant research value that can impact future 3D deep learning studies. [W2]: Evaluation of generalizable methods on new datasets. To avoid any possibility of overfitting with existing generalizable methods and considering the suggestion by Reviewer #MCB6, we adopt the Replica dataset [3], which is a photorealistic simulated 3D dataset containing accurate annotations of RGB, dense depth maps, and semantic maps for thorough evaluation. Specifically, the LSM generalizes well from the real-world ScanNet test set (25.44dB) to the simulated Replica test set (23.10dB). LSM also produces the best depth estimation metrics and is the only generalizable method that can enable 3D semantic segmentation. Splatter Image is an ultra-fast monocular 3D object reconstruction method using 3DGS, and we utilize the provided checkpoint for evaluation. However, while Splatter Image addresses object 3D reconstruction well when the background is masked out, it cannot handle scene-wise reconstruction well with complex backgrounds (please see the attached Fig.1). \| Methods/Metrics \| GT Camera Parameters \| mIoU ↑ \| rel ↓ \| PSNR ↑ \| \|--------------------\|----------------------\|--------\|-------\|--------\| \| pixelSplat \| Required \| None \| 20.14 \| 26.28 \| \| Splatter Image \| Required \| None \| None \| 12.37 \| \| Ours \| Not Required \| 0.51 \| 4.91 \| 23.10 \| [W3]: Improve the presentation. We will highlight the input and output formats in the methodology section in a future version. Although our method architecture exhibits some similarity to generalizable Gaussian models, we emphasize that the investigated problem is very different. pixelSplat focuses on the standard NVS setting, where calibrated images are assumed and the emphasis is solely on image synthesis. In contrast, our overarching goal is to develop a versatile 3D foundation model that unifies scene reconstruction and understanding with uncalibrated images, which is more aligned with real-world applications to deploy a single model. We will adjust the title to emphasize our unification of 3D tasks according to Reviewer 8Lhg's suggestions. Thank you for the feedback. [W4] Missing citation. Splatter Image [4] presents one of the first solutions for ultra-fast monocular 3D object reconstruction using 3D Gaussian Splatting as a representation. It takes a single or two object-wise images as input for reconstructing the 3D representation without test-time optimization, while demonstrating promising reconstruction accuracy. We will incorporate a detailed discussion in our revision. [Q1] Replace encoder using monocular depth? The “Dense Geometry Prediction” cannot be replaced with a monocular depth estimator. Our method takes unposed images as input to build global 3D point maps with versatile attributes for many 3D problems. It requires a stereo depth estimator to unify the point maps from two viewpoints into the same coordinate system. [Q2] Can this model perform generation, like Free3D? Our work reconstructs 3D scenes based on unposed images, similar to 3D-GS, but without the need for COLMAP pre-computation. The supervision comes from the new view interpolated between the training views, meaning it is not capable of extrapolating viewpoints (generation). We will clarify this point and discuss the potential to extend the proposed framework into a generative approach in the future work section. [Q3] What’s “language-driven segmentation”? "language-driven segmentation" means segmenting areas based on a set of language descriptions (e.g., “wall,” “floor”). This allows us to associate each Gaussian with text to render the 2D semantic map. Please refer to [5] for more detailed explanations. Reference: [1]. DUSt3R: Geometric 3D Vision Made Easy. CVPR 2024 [2]. pixelnerf: Neural radiance fields from one or few images. CVPR 2021 [3]. pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction. CVPR 2024 [4]. Splatter Image: Ultra-Fast Single-View 3D Reconstruction, CVPR 2024 [5]. Language-driven Semantic Segmentation, ICLR 2022 --- Rebuttal Comment 1.1: Title: Respectfully Requesting Comments from the Reviewer Comment: Dear Reviewer K6pm, Thank you once again for your review. As the deadline for the author-reviewer discussion approaches, we noticed that we haven't received any further comments from you. We have addressed all your questions with additional experiments and clarifications: - We have demonstrated the importance and novelty of the problem: "Semantic 3D reconstruction directly from unposed images using a single differentiable model." - We provided comparisons showing that our method is not only pose-free but also enables more 3D tasks than other methods. - We plan to adjust the title to better emphasize the unification of 3D tasks that our approach offers. - We clarified the necessity of a model that maps global 3D points for input images. - We provided further clarification on how our method compares with Free3D. - We explained the terminology you requested. As the discussion period is coming to an end, we would greatly appreciate any additional feedback you might have. If our responses have clarified your understanding of our paper, we sincerely hope you might consider raising the rating. Thank you again for your effort in reviewing our paper. Best regards, Authors of Paper 3523	Rebuttal 1: Rebuttal: We thank all reviewers for acknowledging that the work is sound and clearly presented (8Lhg). The presented Transformer-based design is very valuable (fir7) and general (K6pm), running lighting-fast (K6pm, fir7, MCB6, 8Lhg) while achieving compelling quality (K6pm, 8Lhg). We have addressed all the questions posed by the reviewers with additional experimental results. We will carefully revise our main manuscript, following those suggestions. Pdf: /pdf/c5b3f6a153565b1f3c5d2d6da74fa816a2742fa5.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Metalearning to Continually Learn In Context	Reject	Summary: The paper focuses on Automated Continual Learning which is different than handcrafted continual learning. It uses self referential neural networks to meta learn their own in-context continual learning algorithm. First, the paper shows the emergence of in-context catastrophic forgetting. Second, the paper analyze the performance of proposed method (ACL) and finally the paper discuss the limitation of the proposed method. Strengths: - The paper is clearly written and easy to follow - The paper introduces original idea of Automated Continual Learning - The paper identifies "in-context" catastrophic forgetting Weaknesses: - The paper claims to do in-context continual learning but the concept of in-context learning is not clearly explained. - The paper mainly focus on two task and five task settings but it would be more helpful to see the more different settings such as three task or four task - How is the size of SRWM affects the maximum sequence length that can be train? Technical Quality: 3 Clarity: 3 Questions for Authors: - Why only consider the two-task setting? - Why ACL was not compared with replay buffer based methods? - What is the architecture of SRWM? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Authors have addresses the limitation of the work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would first like to thank the reviewer for their valuable time reviewing our work and for many positive comments. Thank you very much. > The paper claims to do in-context continual learning but the concept of in-context learning is not clearly explained. We actually describe and highlight the concept of in-context learning in Sec 2.2 and Figure 1. But we agree with the reviewer that this is currently not clear as we also use the somewhat older terminology of “Meta-learning via Sequence Learning” in the title of Sec 2.2 (we explain much later in Line 337 “This was rebranded as in-context learning” in Sec. 5). We believe mentioning “in-context learning” in Sec 2.2 upfront should make this context clearer. We will fix this in the final version. Thank you very much for pointing this out. > The paper mainly focus on two task and five task settings but it would be more helpful to see the more different settings such as three task or four task We actually provide many more experiments in the appendix for the interested readers. Table 10 provides results for 2, 3, and 4 tasks (exactly as suggested by the reviewer) and Table 8 explores generalization up to 10 tasks (by concatenating Split-MNIST and Split-FashionMNIST). > How is the size of SRWM affects the maximum sequence length that can be train? This is a hard question for which we do not have a straightforward answer because this depends a lot on the nature of the datasets. That said, one useful theoretical result to keep in mind is that, given the dimension of weight matrix D-by-D, the maximum number of key-value associations one can retrieve in a noiseless fashion is D (corresponding to the maximum number of orthogonal keys one can have in the D dimensional space). In practice, the notion of capacity is more complex as the model has multiple layers, and depending on the nature of the data, certain keys may be shared by different inputs without loss of performance. > Why only consider the two-task setting? As we mentioned above, we consider many more settings than the two-task setting: five-task setting (Split-MNIST) in Table 3, two/three/four tasks in Table 10, and up to ten tasks in Table 8. Perhaps what the reviewer really meant is: why do we first focus on the 2-task setting? The reason is clear: Our motivation for using two tasks in Sec. 4.1 and 4.2 is to introduce the core problem of in-context catastrophic forgetting and to demonstrate how our ACL overcomes this problem in a minimum and comprehensible setting (and that's the 2-task setting). > Why ACL was not compared with replay buffer based methods? We focus on the replay-free setting as we believe this is the most interesting setting in continual learning currently (allowing for eliminating replay buffers which requires extra engineering design regarding what to store/remove, and to manage extra memory storage of raw data). Similarly, the recent "learning-to-prompt" papers we cite in our work also focus on the replay-free setting. They additionally raise the privacy issue of the replay buffer which stores raw data to motivate the replay-free setting. Please also note that our ACL and the use of replay buffer are orthogonal: they could also be combined together. > What is the architecture of SRWM? The architectural details of SRWM can be found in Table 5 in the appendix. We believe our response thoroughly clarifies all the reviewer's remaining concerns. Please also refer to our general response highlighting our contributions that we believe are highly relevant considering the diverse interests represented in the NeurIPS community. We really believe our contributions outweigh the limitations, and the current overall rating does not fully reflect the contributions of this paper. If you think that our rebuttal has further improved the reviewer's perception and rating of our work, we would appreciate it a lot if the reviewer could consider increasing the score. Thank you very much. --- Rebuttal 2: Title: Update Comment: Thank you for the detailed explanation. I have missed major points raised by other reviewers and feel like I should change my original review. I believe that the paper presents very original idea and need just final finishing in terms of clarity. --- Rebuttal Comment 2.1: Title: Explanation requested Comment: Dear Reviewer vTAU, Thank you for your response. However, given that you changed your score from 6 to 4, we would like to know more elaborated explanations/justifications, especially considering your high/influential confidence score of 4. You wrote: > I have missed major points raised by other reviewers What "major points" are you referring to? As we explained in our rebuttal, we tried to resolve many concerns raised by the reviewers, including corrections of certain factual misunderstandings. They have not responded yet, but we hope they will. We would like to express our concern that for now, our rebuttal has not been considered. This is very unfair. If you found that our response did not convincingly address these points, please explain the reasons. --- Reply to Comment 2.1.1: Comment: We really do not intend to bother the reviewer further, but we wanted to let you know that Reviewer d5XY has increased the score after considering our rebuttal... If you think their original review had influenced your score change, please do not hesitate to take a look at their response.	Summary: The paper describes a method for in-context continual learning (CL) by using a type of meta-learning neural architecture based on ‘self-referential weight matrices’ (SRWM). Proposed in prior work, these models learn to modify weight matrices iteratively as they process more and more inputs. In this work, they are given few-shot examples from different tasks and iteratively update the weight matrices as the examples are processed. This update process is referred to as “in-context” learning in this work. The key innovation is to define the loss function of SRWM training to optimise for both forward (improving performance of subsequent CL tasks) and backward (improving performance of previous CL tasks) transfer while achieving good performance on the current task. Experiments are conducted on commonimage classification meta-learning benchmarks such as Split-MNIST and Mini-ImageNet. Results show the proposed method prevents catastrophic forgetting (without using replay), outperforming existing meta-learning baselines on the evaluated benchmarks. Strengths: Studies the problem of in-context catastrophic forgetting via a two-task toy setting and reveals the issue when training with no backward transfer loss term. This is shown to be mitigated by including the backward transfer loss term. Proposes an in-context CL method using models based on SRWM and a novel loss to mitigate catastrophic forgetting as more tasks are learned. The method does not use a replay buffer. Studies and covers standard image classification meta-learning tasks such as Split-MNIST, FashionMNIST, and CIFAR-10. On Split-MNIST, shows improvements over existing CL and meta-baselines in both domain and class incremental evaluation settings. The improvements, when additional 5-task fine-tuning is used, is significantly above baselines. The paper is clearly written, with thorough literature review. Weaknesses: One weakness of the proposed method is that the number of loss function terms increases with the number of CL tasks, as pointed out by the authors in Appendix A.5. This prevents this method from being scaled to more practically relevant settings where a large number (much more than 2 or 3 that this paper has mostly focused the experiments on) of tasks are considered in a CL setting. Method of reducing the loss terms would strengthen the paper. Another weakness, which is also noted by the authors in Table 4 and Section 4.3, is that the performance of the proposed model and method is poor compared with those based on pre-trained transformer models, even on an easier evaluation task. The authors in Section 5 also discuss a potential connection between LLM transformer training as an implicit version of the proposed model and method. Given these existing strong and more widely adopted methods, it is unclear how much value the proposed method adds. SRWMs are not widely used and LLMs training can scale to a massive number of tasks with a single loss [1] (albeit not CL). A more detailed explanation of the application of the findings of this paper beyond those interested in SRWMs would be helpful. Another weakness of this paper is its focus on image classification meta-learning tasks only. It is helpful to know the generality of this method, for example on language modelling tasks or multimodal tasks. An experiment demonstrating the method in CL language tasks would be helpful. [1] Finetuned language models are zero-shot learners. Wei et al. ICLR 2022. Technical Quality: 4 Clarity: 4 Questions for Authors: None Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Limitations have been adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable time reviewing our work and for many positive comments. We also acknowledge the reviewer’s thorough reading through the details of our work. Thank you very much. > One weakness of the proposed method is that the number of loss function terms increases with the number of CL tasks, as pointed out by the authors in Appendix A.5. … Method of reducing the loss terms would strengthen the paper. Yes, the discussion on this limitation is provided in A.5./L768. Given that the most crucial loss terms are those at the last time step of the sequence, one possibility to reduce the number of terms is to sample a subset of the intermediate terms. However, effectively validating such methods would require actual experiments involving many more tasks, which we leave for future work. > Another weakness, which is also noted by the authors in Table 4 and Section 4.3, is that the performance of the proposed model and method is poor compared with those based on pre-trained transformer models Yes, clearly we will need meta-training on larger and more diverse datasets for this method to demonstrate its full potential. In fact, as a purely data-driven method, more and more improvements are potentially expected by scaling this up with more compute and datasets (Sutton’s bitter lesson). > it is unclear how much value the proposed method adds. SRWMs are not widely used and LLMs training can scale to a massive number of tasks with a single loss [1] (albeit not CL). A more detailed explanation of the application of the findings of this paper beyond those interested in SRWMs would be helpful. Thank you very much for asking this question. Here we would like to bring up one important and (hopefully) convincing aspect. It is true that SRWM itself is not widely used at the moment. That said, there is an emerging trend in sequence processing using fast weights (a family of linear Transformers). As we also emphasized in the general response, two very recent works [2] and [3] show very promising results on general language modeling using such architectures. It should be noted that [2] makes use of DeltaNet [1], whose direct successor is SRWM (DeltaNet augmented by self-reference), and [4] shows that SRWM is indeed more powerful than DeltaNet (using formal languages) while not requiring hard-to-parallelize, true recurrence. Besides, [3] also mentions “multi-level meta-learning” in the outlook, which was essentially the original motivation of SRWM. Given these evidences of active research on this family of models, in our view, it is not unlikely that there will soon be other works scaling up SRWM or similar models on other tasks. Here we motivated SRWM as a natural architecture for continual learning but its applicability is broader in principle (maybe similar to how attention/Transformer was first explored in machine translation as a natural architecture). From this perspective, we believe we have one extra contribution in this work that is broader than the specific scope of this work. This is a successful example of sequence processing using SRWM which may also be of interest to those interested in general sequence processing using fast weights or linear Transformers. We believe this point precisely addresses the reviewer's concern regarding the impact of this work beyond our specific scope. > Another weakness of this paper is its focus on image classification meta-learning tasks only. This is another very valid point. As a first paper on this method, we focused on image classification which has classic continual learning benchmarks. Extending this to other modalities is an exciting future research (we also mention this in L311/Sec 5). We hope this response and the general response further clarifies our contributions (including the relevance of SRWM in a broader scope of sequence processing with fast weights). If you think our response has enhanced your perception of this work, and if you think the paper should be accepted, we would appreciate it a lot if the reviewer could consider increasing the score. Thank you very much. [1] Schlag et al. ICML 2021. Linear transformers are secretly fast weight programmers. https://arxiv.org/abs/2102.11174 [2] Yang et al. arXiv June 2024. Parallelizing Linear Transformers with the Delta Rule over Sequence Length. https://arxiv.org/abs/2406.06484 [3] Sun et al. arXiv July 2024. Learning to (Learn at Test Time): RNNs with Expressive Hidden States. https://arxiv.org/abs/2407.04620 [4] Irie et al. EMNLP 2023. Practical Computational Power of Linear Transformers and Their Recurrent and Self-Referential Extensions. https://arxiv.org/abs/2310.16076 --- Rebuttal Comment 1.1: Title: Comment Comment: I thank the authors for the detailed rebuttal. It appears that a majority of the concerns mentioned in the review has been considered by the authors but as future work (for example whether subsampling the loss terms for more tasks can work, or expanding beyond the image domain). The contributions of this paper would be much stronger for these results to be included in this paper. Given the existing results, the generality of this method remains quite unclear. I maintain my original recommendation. --- Reply to Comment 1.1.1: Title: Thank you for your response Comment: Thank you very much for your response. We agree that the paper will be much stronger if we could solve all these problems. However, it is unreasonable to assume that one can solve all the problems in this single paper. It is not as if our paper lacks content; we have a full-length paper with plenty of novel results. We are deeply disappointed that the reviewers seem to consider our results/contributions to be trivial, and only evaluate our methods through their limitations (when many of them can be solved directly by better-resourced teams).	Summary: The paper studies the problem of catastrophic forgetting (CF) by formulating continual learning (CL) as learning from a sequence of demonstrations of tasks. The paper proposes a meta-learning objective function that includes backward transfer terms. These terms compute the error of the predictor on previous tasks after receiving demonstrations of the current task. Strengths: - The approach of formulating (continual learning) CL as learning from a sequence of demonstrations of tasks is interesting. - The experiment shows positive results when compared to non-meta-learning approaches Weaknesses: - The paper is difficult to follow. Many definitions and the algorithm are not very well explained. - The motivation of formulating (continual learning) CL as meta-learning is not well presented. - Some details of the architecture are mentioned in the background section only (e.g. replacing self-attention with SRWN and the multi-head version.) - The details of the training and inference process are not well presented. - The training process can be very costly and poorly scaled with the number of tasks and the number of examples per task. In each step over a sequence of demonstrations, the method needs to compute and store a new weight matrix in order to perform back-propagation. It might require more memory during training and at inference. - Even being a meta-learning approach, the model still needs fine-tuning when given a new task to adapt to a new number of tasks. Technical Quality: 3 Clarity: 2 Questions for Authors: - Can the authors explain more on the following claim “The stability-plasticity dilemma are automatically discovered and handled by the gradient-based program search process.“ (line 52)? - What are the advantages of this method compared to previous approaches? - How do the number of examples per task and order of tasks during the training affect the performance at inference time? - How does the method scale with the number of tasks in terms of performance and computation? - It’s unclear how to calculate the loss function in a batch fashion since each training point requires a different sequence of inputs (depending on the position of the task in the sequence) and loss components. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: There are no negative social impacts. My suggestions have been listed in the previous sections. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable time spent on reviewing our work. We believe we have good responses to resolve all the main concerns. == Factual clarifications == Before providing our clarifications to the reviewer’s concerns, we would first like to resolve some factual misunderstandings. > In each step over a sequence of demonstrations, the method needs to compute and store a new weight matrix in order to perform back-propagation. It might require more memory during training and at inference. This is not correct. There is no need for such storage (there is a memory-efficient algorithm that is standard in any practical linear Transformer implementations). For a detailed explanation, please kindly refer to our response to Reviewer d5XY (marked "[also relevant to Reviewer HvJH]"). We apologize for this inconvenience due to the length limit. > The experiment shows positive results when compared to non-meta-learning approaches Thank you for mentioning this as a strength. We would like to add that positive results are also shown compared to the existing meta-learning methods (Table 3). Unlike prior meta-learning methods that only learn “representations” for CL, our method successfully learns an entire “learning algorithm” (the original definition of learning-to-learn) and outperforms prior methods. We believe this achievement in meta-learning is largely overlooked in the current review. == Clarifications to weaknesses/questions == > The motivation of formulating (continual learning) CL as meta-learning is not well presented. The motivation is described in the introduction (Line 30-50). Generally speaking, machine learning is useful when it is hard for humans to design a hand-crafted solution. Here we claim that hand-crafting CL algorithms has been unsuccessful in the past. Therefore, we propose to use machine learning to automate the process of designing CL algorithms. This corresponds to “learning of learning algorithms”, or meta-learning. > Some details of the architecture are mentioned in the background section only (e.g. replacing self-attention with SRWN and the multi-head version.) We’d like to clarify that SRWM is indeed a background work (including its role as a replacement to self-attention and the multi-head version). These design details are entirely based on the previous work by Irie et al. ICML 2022 "A Modern Self-Referential Weight Matrix That Learns to Modify Itself". > The details of the training and inference process are not well presented. The corresponding processes are described in Sec 3. and visually highlighted in Figure 1. If the reviewer still thinks these are “not well presented”, we would appreciate it a lot if they could tell more concretely what they find confusing. (A dedicated section in Appendix A.1 “Continual and Meta-learning Terminologies” also helps readers with continual and meta-learning jargon). > Even being a meta-learning approach, the model still needs fine-tuning when given a new task to adapt to a new number of tasks This is a valid point and it remains an open research question e.g., to find how to deal with an unseen maximum number of classes in the class-incremental setting. Here we would like to strongly emphasize that meta-learning is an open research topic. We can not solve all these problems in a single paper. A significant achievement here is that we obtain a successful meta-learned CL algorithm for a fixed maximum number of classes. Extending our work to relax these constraints is an exciting future research direction in meta-learning. > Can the authors explain more on the following claim “The stability-plasticity dilemma are automatically discovered and handled by the gradient-based program search process.“ (line 52)? Yes, our pleasure! When we manually design a CL algorithm (say a regularization method with a hyper-parameter weighting the auxiliary term), we face the problem of deciding how much we allow model parameters to change (plasticity) or not (stability) to balance preservation of current knowledge against learning of new one. In our method, we do not have to deal with this decision ourselves. This process is automated by gradient descent that directly optimizes weight modification rules (Eq. 3) for the ultimate model performance. > What are the advantages of this method compared to previous approaches? > How does the method scale with the number of tasks in terms of performance and computation? The current advantage is that we can achieve a CL algorithm that performs well on classic benchmarks (unlike prior methods that fail). Another potential advantage is that, as a purely data-driven method, more and more improvements are expected by scaling this up with more compute and datasets (Sutton’s bitter lesson). Telling exactly how they scale would require a proper scaling law study, itself requiring more compute. > How do the number of examples per task and order of tasks during the training affect the performance at inference time? An ablation on the number of examples can be found in Table 7 in the appendix (5 vs. 15 examples). Regarding the order, we conducted experiments in the 2-task (A and B) setting: Tables 1 and 2 present results for both A-then-B and B-then-A orders. > It’s unclear how to calculate the loss function in a batch fashion There is no such issue: each sequence in the batch is constructed such that it contains the same number of examples and the task boundaries at the same time steps. We believe our response thoroughly addresses all the concerns raised by the reviewer. Please also refer to our general response highlighting our contributions that we believe are broadly relevant to the NeurIPS community, despite all our limitations. In light of these clarifications, we believe the current rating does not fully reflect the contributions of this paper. If the reviewer finds our response useful/convincing, please consider increasing the score. Thank you very much. --- Rebuttal Comment 1.1: Title: Author-reviewer discussion ending imminently Comment: While we have been respecting the rule of not sending individual reminders to the reviewers ourselves, please allow us to post this single/final message as the end of author-reviewer discussion period is imminent. We would like to thank the reviewer one more time for their valuable time spent on reviewing our work. As all the other reviewers have already responded, we would appreciate it a lot if the reviewer could communicate their perception and rating of this work that take into account our rebuttal. Thank you very much. --- Rebuttal Comment 1.2: Title: Response to Authors' Rebuttal Comment: Thank you for your detailed rebuttal. It really helps clarify my concerns. I agree that the formulation of CL as an ICL problem is interesting and the fact that the trained model can generalize is even more impressive. I am raising the score to 5 and am willing to increase it if more information/discussion is provided. 1. Can you discuss the importance/advantage/limitation of the self-referential weight matrix in facilitating in-context learning? Self-referential is a powerful idea but it is not very clear why you choose that algorithm for CL. 2. At first glance, the update rule for $W_t$ is quite simple, is it general enough to learn all kinds of algorithms? This question might be out of the scope of this paper but I just want to have a conversation. Also, I still maintain that the presentation of this paper could be improved.	Summary: The paper proposes a novel technique to automatically discover in-context continual learning dynamics for image classification task sequences through meta-learning. In order to achieve this purpose, the approach relies on 2 main novelties: * Using self referential weight matrices on top of an image encoder - SRWM, as self-modifying that adapts itself to the stream of inputs, is an natural model for continual learning. * Encoding continual learning desiderata in the meta-objective, i.e. backward and forward transfer. The authors first apply the approach in a classic two-task setting (Split-MNIST) that allows them to showcase and analyse the emergence of in-context catastrophic forgetting phenomena, and to show that using their ACL loss can help reduce it. They further evaluate their method and compare them to replay-free baselines from the CL and meta-CL literature, showing an advantage of their approach in scenarios with up to 3 tasks. The authors further test the limits of their approach by comparing it to more recent learning to prompt techniques for continual learning, leveraging the power of pretrained large models. This scenario show a limitation of the technique in more complex scenarios with more tasks, more diverse and complex data. Strengths: * The paper takes an interesting perspective on continual learning, leveraging the interesting properties of SRWM and the capability of meta-learning to encode the desired behavior in the meta-learning objective. The combination of these two contributions is novel to the best of my knowledge, and lead to interesting insights. * The approach leads to interesting performance in relatively simple scenarios, outperforming some of the existing continual learning techniques. * I also particularly appreciated the authors discussion of the method limitations. Both the experiments with learning to prompts and the discussion provide very valuable insights that can help building on the work in the future. Weaknesses: * In my opinion, the main limitation of the approach is its practicality. From the experiments reported in Table 4, it seems that the approach requires to met-train on a sequence of similar length and/or complexity to provide its potential. This is not possible to know in advance in practice. Moreover, one limitation that the authors have not mentioned is that the meta-objective seems to require keeping in memory a number of copies of the model that is equal to the number of tasks. This can quickly become cumbersome for real applications that can require more complex models and very long sequences of tasks. * While the authors focus on classic benchmarks for continual and meta-learning, these benchmarks are artificial, relatively simple and lack of diversity. Different works highlight the limits of these benchmarks, I invite the authors to look at "Meta-Album: Multi-domain Meta-Dataset for Few-Shot Image Classification" Ullah et al. 2023, and "NEVIS'22: A Stream of 100 Tasks Sampled from 30 Years of Computer Vision Research" Bornschein et al. 2023 for examples of more realistic benchmarks. * It would be interesting to add a discussion of the cost of the approach (computation, memory, ...). Even is it gives a substantial boost in many cases, it would be interesting for practitioners to compare what they gain to what they pay. Technical Quality: 3 Clarity: 3 Questions for Authors: * The approach is focused on the task aware scenario, and rooted in a notions of tasks. In many practical scenarios, the distribution shift occurs in a softer way, with no clear notion of task boundaries. Can the authors comment on the possibility of extend their approach to the task-agnostic scenario? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors provide a detailed discussion of the work limitations, both in the experiments and the discussion sections. Some other limitations are highlighted in the Weaknesses paragraph above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable time reviewing our work and for many positive comments. Thank you very much. == Factual error corrections == Before providing our clarifications to the reviewer’s concerns, we would first like to resolve some factual errors in the review. > in a classic two-task setting (Split-MNIST) > showing an advantage of their approach in scenarios with up to 3 tasks. These statements are not correct. The main results on Split-MNIST (Table 3) correspond to a 5-task setting. (We also report the generalization performance on a 10-task setting by using Split-MINST and Split-FashionMNIST in Table 8 in the appendix). > Moreover, one limitation that the authors have not mentioned is that the meta-objective seems to require keeping in memory a number of copies of the model that is equal to the number of tasks. [also relevant to Reviewer HvJH] This is not correct. We do not need to keep the intermediate copies of the model. The reviewer is right to point out that a naive implementation would require such copy storage. The situation is actually even worse for a naive implementation: it would have to store ALL intermediate copies of the model’s fast weight states (at every step) for backpropagation through time; this would result in impractical memory requirements for ANY experiments we conducted. Instead, the practical code implements a custom memory-efficient algorithm that only stores one copy of the model fast weights in the entire forward pass, and during backpropagation through time, we de-construct the model by applying the reverse fast weight update (Eq.3 in the backward direction) to obtain the model state we need at the corresponding step. This only requires storing the key/value/query/learning-rate activations for all steps in memory (which are much cheaper) and a single copy of the model. We did not describe this implementation trick as this is standard in ANY practical linear Transformer implementations (in case the reviewer is interested, please find example public code below). We will add this discussion in the final version. Thank you for pointing this out. Public code references: - Linear Transformer: https://github.com/idiap/fast-transformers/blob/master/fast_transformers/causal_product/causal_product_cuda.cu - SRWM: https://github.com/IDSIA/modern-srwm/blob/main/supervised_learning/self_ref_v0/self_ref_v0.cu == Clarifications to weaknesses/questions == We believe we have convincing answers to all the main concerns raised by the reviewers. > the main limitation of the approach is its practicality. We agree with the reviewer that our method requires scaling multiple hyper-parameters to be useful in more realistic settings. That said, our main bottleneck is the compute. In the limitation section, we do propose approaches to deal with the algorithmic limitations (e.g., to handle longer lengths, we’ll have to introduce context-carry over as used in language modeling with linear Transformers; see Line 302). There is no reason that better-resourced teams, which can afford to train today's large language models, cannot scale this up to much larger and more diverse datasets. > I invite the authors to look at … examples of more realistic benchmarks. We thank the reviewer for sharing these references. We’d like to emphasize that it’s not that we didn’t know about these more realistic tasks, but rather that we do not have compute to conduct larger scale experiments. Please note that given that our method is novel, we had to allocate a lot of compute for ablation experiments (we have many more experiments in the appendix). As we also emphasized in our general response, we strongly believe that this work has significant contributions by providing several evidences for the promise of this method, even without a large scale continual learning experiments. These contributions are largely overlooked in the current review. > It would be interesting to add a discussion of the cost of the approach (computation, memory, ...). Even is it gives a substantial boost in many cases, it would be interesting for practitioners to compare what they gain to what they pay. Assuming that model hyper-parameters are already available, as we mention in A.3., less than 1 day of compute using a single V100 GPU is enough to produce a single-run experiment reported in this paper (e.g., the Split-MNIST result). We also would like to emphasize that our method does not just provide a “substantial boost” but rather represents a clear-cut switch from a complete failure (hand-crafted approach or some other meta-learning method) to reasonable continual learning in many cases. > The approach is focused on the task aware scenario, and rooted in a notions of tasks. In many practical scenarios, the distribution shift occurs in a softer way, with no clear notion of task boundaries. Can the authors comment on the possibility of extend their approach to the task-agnostic scenario? Thank you for pointing this out. The task awareness is only required for meta-training. At test time, the model is not aware of anything related to the task identities or task boundaries; they are evaluated in a completely task-agnostic manner. For training, it seems reasonable to assume that we do have access to training examples where we know task identities. Therefore, extending this to the task-agnostic setting is rather straightforward. We believe our response thoroughly addresses all the concerns raised by the reviewer, and directly resolves many of them. Please also refer to our general response highlighting our contributions that we believe are highly relevant considering the diverse interests represented in the NeurIPS community. In light of these clarifications, we believe the current rating does not fully reflect the contributions of this paper. If the reviewer finds our response useful/convincing, please consider increasing the score. Thank you very much. --- Rebuttal Comment 1.1: Title: Answer to rebuttal Comment: I thank the authors for their clarifications: * "up to 3 tasks" in my original review is a typo - I meant 5 tasks, and I thank the authors for pointing out the additional results in the appendix. This however doesn't alleviate my main concern here: These validation benchmarks are still of limited length and diversity. * I thank the authors for the clarifications and pointers regarding the memory constraint. This answers my concern. In general, I find that the authors answered most of my questions. Despite the limitation in the evaluation benchmarks mentioned above, I am raising my score. --- Reply to Comment 1.1.1: Comment: Thank you very much for your reply and the increased score. We genuinely thank the reviewer for the effort put into checking our rebuttal.	Rebuttal 1: Rebuttal: == General Response to all the Reviewers == We would first like to sincerely thank all the reviewers for their valuable time reviewing our work. We would like to emphasize that this work has two facets: On the one hand, we explore a novel perspective/approach to continual learning (CL). At the same time, this is also a meta-learning research paper presenting a significant advancement in this domain (learning of learning algorithms). This is a highly relevant topic to the NeurIPS community representing diverse interests, aiming an audience beyond those solely interested in state-of-the-art methods in CL. We believe our contributions in meta-learning are largely overlooked in the current reviews. In fact, we explicitly mention and acknowledge all the current limitations of our method in light of state-of-the-art continual learning methods/tasks. We did this honest exhibition not just to facilitate the reviewer’s job to identify these weaknesses; we did this because we strongly believe that, despite all these limitations, we do have interesting contributions and results for the NeurIPS community (and we sincerely thank Reviewers x6Jn and vTAU for voting toward acceptance). Namely: [Advances in meta-learning learning algorithms] Our meta-learned CL algorithm is successful at Split-MNIST unlike any hand-crafted algorithms and prior meta-learning methods for CL. This is a highly non-trivial achievement and a significant step in meta-learning research of learning algorithms. While Split-MNIST is indeed a toy task compared to other larger scale CL tasks (we also explicitly acknowledge this in the paper), it is not like the standard MNIST which is a truly toy task with many trivial solutions. We are not aware of any trivial solutions for Split-MNIST; it represents non-trivial CL challenges. Regarding “practicality”, our main obstacle for showing results at scale is our compute resource limitation. While we do acknowledge in the paper certain technical challenges to be addressed in the future, much larger scale language modeling experiments are conducted by better-resourced teams. [Novel insights into in-context learning] Furthermore, the phenomena we exhibit and study, called “in-context catastrophic forgetting”, is a new perspective of in-context learning (ICL) that can be found nowhere in prior work. We largely dedicate space in our paper to discuss & analyze this in a comprehensible 2-task CL setting (Sec 4.1 and 4.2) using Mini-ImageNet and Omniglot. Please note that many recent foundational studies on ICL make use of “toy” tasks, such as regression to exhibit the core idea (see, e.g., [0]). We believe this is relevant to anybody in the NeurIPS community interested in ICL. [Novel results on sequence processing using fast weights and linear Transformer-family models] Finally, our work is also relevant to the readers interested in the general idea of sequence processing through “weight modifications” (also called fast weights). Such models are directly related to the linear-complexity variant of Transformers [1], and there is a very recent trend applying such models to language modeling (please see [2] and [3]). In particular, [2] shows that the model called DeltaNet [1] outperforms other popular linear-complexity models such as Mamba at scale. It should be noted that the direct followup of the DeltaNet [1, 2] is the SRWM architecture we use [4], which has been shown to be more expressive than the DeltaNet [4]. Therefore, going beyond the scope of continual and meta-learning, we believe our work also contributes to this emerging research on fast weight architectures for sequence processing [2, 3, 4] as another successful example thereof. For all these reasons, we would really appreciate it a lot if the reviewers could take a look at our rebuttal response and reconsider the true contributions of this work through various angles/interests within the NeurIPS community. We believe our responses provide convincing answers to all the main concerns raised by the reviewers (including a few crucial factual error corrections). We really believe our contributions outweigh the limitations, and we’ll be happy to provide further clarifications if necessary. Once again, thank you very much for your valuable time reviewing our work. References: [0] von Oswald et al. ICML 2023. Transformers Learn In-Context by Gradient Descent. https://arxiv.org/abs/2212.07677 [1] Schlag et al. ICML 2021. Linear transformers are secretly fast weight programmers. https://arxiv.org/abs/2102.11174 [2] Yang et al. arXiv June 2024. Parallelizing Linear Transformers with the Delta Rule over Sequence Length. https://arxiv.org/abs/2406.06484 [3] Sun et al. arXiv July 2024. Learning to (Learn at Test Time): RNNs with Expressive Hidden States. https://arxiv.org/abs/2407.04620 [4] Irie et al. EMNLP 2023. Practical Computational Power of Linear Transformers and Their Recurrent and Self-Referential Extensions. https://arxiv.org/abs/2310.16076 == Correcting one factual error (Reviewers d5XY and HvJH) == While we refer to our responses below for further individual clarifications, here we'd like to correct one factual error made by two reviewers. > (Reviewer d5XY) Moreover, one limitation that the authors have not mentioned is that the meta-objective seems to require keeping in memory a number of copies of the model that is equal to the number of tasks. > (Reviewer HvJH) In each step over a sequence of demonstrations, the method needs to compute and store a new weight matrix in order to perform back-propagation. It might require more memory during training and at inference. These statements are not correct. No such storage/copies is required (there is a well known algorithm for linear Transformers to avoid this). Due to the rebuttal space limitation, we provide a detailed answer in our response to Reviewer d5XY. We kindly ask Reviewer HvJH to refer to the corresponding text.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Binarized Diffusion Model for Image Super-Resolution	Accept (poster)	Summary: The paper introduces BI-DiffSR, a novel binarized diffusion model for image super-resolution, designed to accelerate the inference speed and reduce computational costs of diffusion models while maintaining high performance. It proposes a UNet architecture optimized for binarization, featuring consistent-pixel downsampling/upsampling and channel-shuffle fusion to address dimension mismatch and fusion difficulty, alongside a timestep-aware redistribution and activation function to adapt to varying activation distributions across different timesteps. The model demonstrates superior results over existing binarization methods, approaching the perceptual quality of full-precision models with significantly reduced memory and computational requirements. Strengths: - The paper is well-written and easy to understand. - This paper designs a novelty 1-bit UNet for accurate binarized diffusion model, including: - New downsample module and upsample module for Dimension Consistency. - Channel shuffle module to balance the activation value ranges of two input features. - The timestep-aware redistribution (TaR) and timestep-aware activation function (TaA) - Experiments achieve the state-of-the-art in super resolution with diffusion. Weaknesses: - The basic BI-Conv block lacks novelty, which is as the same as the binarized module in ReActNet that contains RSign and RPReLU. - TaR uses different parameters for different time steps, but in the mean while, the normal time embedding is projected into the resblock, it is also a time-aware on feature maps, what is the differences or why TaR works? - SR3 is not a new diffusion baseline for super resolution, ResShift[1], SinSR[2] should be better, and the metrics of PSNR, SSIM, LPIPS is much old, the CLIPIQA, MUSIQ, MANIQA should be better for evaluating the performance of generative super resolution. - Self-attention and MLP are common modules in diffusion, such as LDM[3] and ResShift[1], which require a lot of computation. How can the method in this paper be extended to self-attention and MLP? [1] Yue, Zongsheng, Jianyi Wang, and Chen Change Loy. "Resshift: Efficient diffusion model for image super-resolution by residual shifting." Advances in Neural Information Processing Systems 36 (2024). [2] Wang, Yufei, et al. "SinSR: diffusion-based image super-resolution in a single step." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. [3] Rombach, Robin, et al. "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022. Technical Quality: 4 Clarity: 4 Questions for Authors: Please refer to the weaknesses above. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The authors have addressed the limitations in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer MRyP (denoted as R4) `Q4-1` The basic BI-Conv block lacks novelty, which is as the same as the binarized module in ReActNet that contains RSign and RPReLU. `A4-1` Thanks for pointing it out. We clarify it below. 1. Indeed, our basic BI-Conv block utilizes RSign and RPReLU for the learnable bias and activation function. However, this is not an innovative aspect of our method. 2. Our innovation lies in timestep-aware operations (i.e., TaR and TaA). Inspired by the mixture of experts (MoE), we employ different RSign and RPReLU according to different timesteps. 3. We demonstrate in Table 1c of the main paper that using timestep-aware operations effectively improves the performance of binarized DM. \| Method \| Params (M) \| OPs (G) \| PSNR (dB) \| LPIPS \| \| :----------- \| :--------: \| :-----: \| :-------: \| :----: \| \| RSign&RPReLU \| 4.30 \| 36.67 \| 31.99 \| 0.0261 \| \| TaA&TaR \| 4.58 \| 36.67 \| 32.66 \| 0.0200 \| 4. To enhance transparency, we will clarify in the manuscript that RSign and RPReLU are from ReActNet. `Q4-2` TaR uses different parameters for different time steps, but in the mean while, the normal time embedding is projected into the resblock, it is also a time-aware on feature maps, what is the differences or why TaR works? `A4-2` We explain the differences below. Differences: 1. Location: Time embedding acts on the ResBlock; TaR operates within the BI-Conv (inside the ResBlock). 2. Purpose: Time embedding operates at the network level, enabling the model to be aware of different timesteps and enhance feature modeling; TaR operates at the module level (such as Conv), adjusting the activation distribution based on the timestep to improve module computations. Functionality of TaR: 1. Adapting Activation Distribution: TaR dynamically adjusts Conv input activation according to different timesteps, adapting changing distributions. 2. Enhancing Conv Representation: TaR divides multiple steps into smaller groups, limiting the changing range of activations, thereby reducing the representational difficulty of BI-Conv and enhancing feature extraction. 3. Complementarity with Time Embedding: TaR and time embedding serve different purposes and are not contradictory. Using TaR (along with TaA) in DM further enhances model performance. This is evidenced in Table 1c (as detailed in `A4-1`). `Q4-3` SR3 is not a new diffusion baseline for super resolution, ResShift[1], SinSR[2] should be better, and the metrics of PSNR, SSIM, LPIPS is much old, the CLIPIQA, MUSIQ, MANIQA should be better for evaluating the performance of generative super resolution. `A4-3` Thanks for your suggestion. We apply the new baseline ResShift [1] (official code), and compare our method, BI-DiffSR, with the full-precision (FP) model and the previous binarized model, BBCU. Setup: 1. We train on the bicubic-×4 task on the DF2K dataset, batch size 32, iteration 400,000, sample step 15, maintaining other settings consistent with the official setup. 2. As our method involves adjusting the channels of UNet, we modify the UNet structure (maintaining the core module unchanged) to create ResShift\. We then apply BBCU* and BI-DiffSR (quantizing both Conv and SA) on ResShift\* to ensure fairness. 3. We measure the FLOPs on an input size of 128×128. All metrics are evaluated on Manga109. \| Method \| Params (M) \| OPs (G) \| CLIPIQA $\uparrow$ \| MANIQA $\uparrow$ \| MUSIQ $\uparrow$ \| \| :-------------------------- \| :--------: \| :-----: \| :----------------: \| :---------------: \| :--------------: \| \| ResShift* (full-precision) \| 58.12 \| 170.77 \| 0.7662 \| 0.5543 \| 67.3729 \| \| BBCU (binarized) \| 3.89 \| 20.19 \| 0.7611 \| 0.5175 \| 64.0867 \| \| BI-DiffSR (binarized, ours) \| 3.65 \| 18.57 \| 0.7659 \| 0.5638 \| 66.5730 \| Results Analyses: 1. Comparison with BBCU: Our method performs better across all new metrics. 2. *Comparison with ResShift\: Our method achieves similar performance while significantly reducing parameters (93.72%) and operations (89.13%). 3. General Effectiveness: These results further demonstrate the generalizability and effectiveness of our method. Further Consideration: 1. Baseline Choice: In our paper, we use SR3 as the baseline, considering it is a normal DM SR method, which can demonstrate the generality of our approach. Classic metrics such as PSNR, SSIM, and LPIPS are used for similar reasons. 2. Inclusion of New Baseline: As the reviewer suggested, using new baselines and metrics better reflects the effect of our method. We will include ResShift-based experiments in our paper for a thorough assessment. `Q4-4` Self-attention and MLP are common modules in diffusion, such as LDM[3] and ResShift[1], which require a lot of computation. How can the method in this paper be extended to self-attention and MLP? `A4-4` The methods are below. Extension to Self-Attention (SA): 1. Linear Layers: SA involves three linear layers for $Q, K, V$, and a linear projection for the output. These parts can be replaced with 1×1 BI-Conv to achieve binarization. 2. Matrix Operations: The matrix operation, $\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$, can be binarized by binarizing the activations: $Q, K, V, \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)$, through the sign function, Sign(·). Extension to MLP: 1. Linear Layers: The MLP consists of several linear layers, which can be directly replaced with 1×1 BI-Conv for binarization. Experiment: We binarize ResShift [1]** with our method, effectively reducing Params and OPs while maintaining performance, as detailed in `A4-3`. --- Rebuttal Comment 1.1: Title: Thank you for the rebuttal Comment: Thank you for the response and additional experiments.BI-DiffSR is insteresting and promising to push the development of diffusion deployment. After reviewing the rebuttal, I decided to raise my score. --- Reply to Comment 1.1.1: Title: Thanks Reviewer MRyP for approving our work Comment: Dear Reviewer MRyP, Thank you for your response. We are pleased that you approve of our work. Best, Authors	Summary: This work present a novel binarized diffusion model for improving the efficiency of super resolution tasks. Compared with the existing works, this work first pointed out the specific challenges of binarized DMs for SR, including the dimension mismatch and fusion difficulty of representations. Then this work present several techniques: consistent-pixel down/upsampleing, channel-shuffle fusion, and Time-step-aware redistribution function for the aforementioned challenges. Comprehensive results show that the provided binarized DMs for SR not only significantly outperform the binarized models with existing SOTA binarization methods, but also achieve floating-point level performance. And for the efficiency, the statistics of params and flops show the advantage of proposed method, and the paper also present the real inference time on edge, which seems important and encouraged in the binarization community. Strengths: 1. As far as I know, this is the first work to present the specific binarization method for diffusion model of SR. Since the good performance has been achieved by DMs in various SR tasks, it’s important to present novel insight to compress these models, especially considering the severe drop still exists after binarizing by existing SOTA methods. 2. The motivation is intuitive and techniques are novelty, especially considering the features of DMs. The proposed CP-Up/down and channel shuffle are highly specified to the architecture of the diffusion models, which is novel and cannot be achieved by previous methods, including binarization function and binarized structures. And the computation is also small, allowing minor burden with significant performance improvement. And the proposed activation function also focus on the high dynamic of activation range during time-step, which is one of the most critical problem for the quantization of DMs. 3. The proposed method achieve SOTA results in accuracy. Comprehensive comparison has been included in this paper, including SOTA binarization methods and various evaluation datasets. The results show that the proposed outperforms than previous binarized DMs for SR with significant improvements. 4. In this paper, diverse analysis, including quantitative, statistical, and visual results are presented in detail. More important, the paper shows the efficiency evaluation based on real inference libraries and edge hardware, which is of great significance for practical application. Weaknesses: Though it’s a good paper, some issues should be addressed. 1. The writing and presentation of the paper should be improved, including but not limited to the grammar and description. For example, some basic knowledge about quantization, SR, and DMs seems to be summarized as a preliminaries section; and let the proposed techniques be highlighted in Figure 2. 2. As for the efficiency, I suggest the authors present the computation more detailed, such as present the computation of each part in the whole network before and after the binarization. This will show the efficiency advantage of the proposed method much clearer. 3. The proposed challenge I and II are insightful, but more further discussion (such as visual, quantitative, or theoretical analysis) are presented after proposing. I suggest authors do more discussion about that. 4. Some recent binarization methods for SR [1] are suggested to be compared and some quantized DMs [2] are suggested to be discussed the differences to make the comparison more comprehensive. [1] Flexible Residual Binarization for Image Super-Resolution. ICML 2024 [2] EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models. ICLR 2024 Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Compared with the quantization of DMs for SR, can authors provide discussion about the advantage and motivation of binarization? 2. What is the type of ARM hardware for the evaluation of inference? 3. If the proposed method have potential generalized to more generative tasks and architectures? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The authors have addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer KWS7 (denoted as R3) `Q3-1` The writing and presentation of the paper should be improved, including but not limited to the grammar and description. For example, some basic knowledge about quantization, SR, and DMs seems to be summarized as a preliminaries section; and let the proposed techniques be highlighted in Figure 2. `A3-1` Thank you for your suggestions. We will check and improve our manuscript. We have added a preliminaries section: >SR Pipeline. SR network aims to reconstruct a low-resolution image into a corresponding high-resolution image. The process can be represented as follows: >$$ >I_{SR} = \mathcal{SR} (I_{LR} ; \Theta), >$$ >where $\mathcal{SR} (\cdot)$ denotes the image SR network, and $\Theta$ represents the network parameters. > >Binarization Framework. In binarized networks, weights and activations are converted using the sign function: >$$ >\operatorname{Sign}\left(x\right)= \begin{cases}+1, & x \geq 0 \\\ -1, & x < 0 \end{cases}. >$$ >As the sign function Sign(⋅) is non-differentiable, we use the straight-through estimator (STE) for backpropagation to train binarized models: >$$ >\frac{\partial \operatorname{sign}}{\partial \boldsymbol{x}}= \begin{cases}1 & \text { if }\|\boldsymbol{x}\| \leq 1 \\\ 0 & \text { otherwise }\end{cases}. >$$ >Binarization reduces storage and accelerates computation through XNOR and bit-counting operations. For Figure 2, we have modified the image to highlight our proposed techniques and added it to the attached PDF (Figure 2). `Q3-2` As for the efficiency, I suggest the authors present the computation more detailed, such as present the computation of each part in the whole network before and after the binarization. This will show the efficiency advantage of the proposed method much clearer. `A3-2` Thank you for your suggestion. We present the Params and OPs for each part of the UNet network: encoder, bottleneck, and decoder, before and after binarization. OPs are tested with an output of 3×256×256. \| Part \| Params$^f$ (M) \| Params$^d$ (M) \| OPs$^f$ (G) \| OPs$^d$ (G) \| \| ---------- \| :------------: \| :------------: \| :---------: \| :---------: \| \| Encoder \| 13.09 \| 0.65 \| 42.66 \| 1.04 \| \| Bottleneck \| 10.59 \| 1.45 \| 13.97 \| 2.31 \| \| Decoder \| 27.30 \| 2.46 \| 79.29 \| 33.31 \| 1. Params$^f$ and OPs$^f$ correspond to the full-precision modules, while Params$^d$ and OPs$^d$ are for the binarized versions. 2. The encoder and bottleneck show high compression ratios. To balance parameters and performance, we use a partial number full-precision ResBlocks in the decoder, which reduces the compression ratio. `Q3-3` The proposed challenge I and II are insightful, but more further discussion (such as visual, quantitative, or theoretical analysis) are presented after proposing. I suggest authors do more discussion about that. `A3-3` Thank you for your suggestion. We add more visual results for the two challenges in the attached PDF (Figure 3). Here are the analyses: Challenge I: Dimension Mismatch. The binarized modules struggle with detailed features. Using full-precision features through residual connections helps, as shown in columns three and four of Figure 3. However, in UNet, up/down sampling makes feature shape mismatches, making it unable to use residual connections. We solve this problem via CP-Down/Up. Challenge II: Fusion Difficulty. The activation distributions of skip connections vary greatly, resulting in information loss in the decoding stage and a lack of details in the restored image. Therefore, we design CS-Fusion to address this problem, as shown in columns four and five of Figure 3. `Q3-4` Some recent binarization methods for SR [1] are suggested to be compared and some quantized DMs [2] are suggested to be discussed the differences to make the comparison more comprehensive. `A3-4` Thanks for your advice. We discuss the differences with FRB [1] and EfficientDM [2] below. FRB [1] 1. Network Design: FRB targets one-step networks; our method includes dimension and timestep considerations in DMs. 2. Training Approach: FRB uses full-precision distillation to guide the training; we use a basic L1 loss to train our model. EfficientDM [2] 1. Quantization Level: EfficientDM implements low-bit (2/4/8) quantization; we consider the extreme case of 1-bit (binarization). 2. Activation Management: EfficientDM employs temporal activation LSQ to handle changes in activation distribution; we propose timestep-aware techniques (TaA and TaR). `Q3-5` Compared with the quantization of DMs for SR, can authors provide discussion about the advantage and motivation of binarization? `A3-5` We discuss them below. Advantage: 1. Higher Compression Ratio: Binarization (1-bit) offers the highest parameter reduction compared to 2/4/8-bit quantizations. 2. Efficient Computation: Binarization allows the model to perform inference via bitwise operations, a capability not inherent to other bit-level models. Motivation: Diffusion models (DMs) have excellent generation abilities but are resource-intensive. Binarization greatly reduces this overhead, enhancing usability on limited-capacity devices. `Q3-6` What is the type of ARM hardware for the evaluation of inference? `A3-6` The evaluation is conducted on a Raspberry Pi 3 Model B+ (BCM2837B0, Cortex-A53 (ARMv8) 64-bit SoC @ 1.4GHz). `Q3-7` If the proposed method have potential generalized to more generative tasks and architectures? `A3-7` Yes, it has the potential. 1. Other Tasks: The model can be directly applied to unconditional generative tasks by removing the LR input. 2. Other Architectures: Many generative diffusion models, like stable diffusion, can be binarized by our proposed approach. --- Rebuttal Comment 1.1: Title: After Rebuttal Comment: Thank you to the authors for their detailed response. After reviewing all the explanations, I can confirm that all of my concerns have been fully addressed. I would like to keep my original score. --- Reply to Comment 1.1.1: Title: Thanks Reviewer KWS7 for approving our work Comment: Dear Reviewer KWS7, Thank you for your response. We are honoured that our replies have addressed the reviewer's concerns. We sincerely appreciate your thorough review and valuable suggestions. Best, Authors	Summary: The authors propose BI-DiffSR to binarize diffusion based image super-resolution (SR) model. They design a UNet architecture for the whole binarized model structure. To maintain dimension consistency, they propose two modules, CP-Down and CP-Up, which can further help transfer full-precision information. To enhance feature fusion, they propose the channel-shuffle-fusion (CS-Fusion). They also propose TaR and TaA to dynamically adjust activation distribution cross different timesteps. The authors provide extensice experiments to demonstrate the effectiveness of their proposed method. Strengths: The topic is very important and practical. Diffusion models have shown excellent performance for image super-resolution (SR). It is very practical to quantize the models before deploying them into devices. Binarization is an extreme tool to compress the SR model. Few works have been proposed to investigate such an important problem in image SR. The authors give several insights for the specific topic. Namely, there are some key aspects in diffusion based image SR binarization, like dimension mismatch, fusion difficulty, and activation distraction. Those problems hinder the performance of binarized image SR diffusion models. The observation and analyses given in the introduction section are insightful and motivate readers well. To alliveate the problems in binarized diffuision based SR models, the authors propose consistent-pixel-downsample (CP-Down) and consistent-pixel-upsample (CP-Up) to ensure dimensional consistency. They propose the channel-shuffle-fusion (CS-Fusion) to facilitate the fusion of different features within skip connections and suit binarized modules. They propose the timestep-aware redistribution (TaR) and timestep-aware activation function (TaA) to adjust the binarized module input and output arross different timesteps. They provide extensive ablation study experiments (including quantitative results in Table 1 and visualization analyses in Figures 6 and 7.) to show the effects of each proposed components. Those experiments are convincing. The authors provide comparions with SOTA methods. According to the main quantitive and visual comparisons, they show that their proposed BI-DiffSR achieves superior performance over others. The overall writing and organization are pretty good. I think the work is well-prepared. The supplementary file further provides more details. The paper is easy to follow and they promise to release the code, which makes this work more convincing. Weaknesses: When binarizing full-precision model from 32-bit to 1-bit, ideally we can reduce the parameters by 32 times. But, as shown in Table 2, the authors reduce parameters from 55.41M to 4.58M (for scale 2). There is a gap between ideal case and practical one. Please give some analyese about the reasons for this gap. Also, are there any idea to further narrow the gap? The parameters and Ops are reduced obviously from full-precision to binary one. But the authors did not give results about inference time on real devices or give some analyses. I am curious how fast the binarized model will be. The writing can further refine in some cases. For example, in the abstract part (Line 9-10), “… to maintain dimension consistent” should be changed to “… to maintain dimension consistency”. Technical Quality: 3 Clarity: 4 Questions for Authors: Can this method be applied to other diffusion models, like stable diffusion? If so, can the authors give some suggestions to binarize stable diffusion? Can we apply this binarization method to other related image restoration tasks? Like image denoising, deblurring? How long did the authors to train the models? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Please refer to weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer Z33c (denoted as R2) `Q2-1` When binarizing full-precision model from 32-bit to 1-bit, ideally we can reduce the parameters by 32 times. But, as shown in Table 2, the authors reduce parameters from 55.41M to 4.58M (for scale 2). There is a gap between ideal case and practical one. Please give some analyese about the reasons for this gap. Also, are there any idea to further narrow the gap? `A2-1` Reasons: 1. We use a partial number (i.e., 6) full-precision (FP) ResBlocks in the UNet decoder part to trade off parameters and performance. 2. The proposed timestep-aware redistribution (TaR) and activation function (TaA) add ~0.3M parameters. Strategy to Narrow the Gap: In practice, we can adjust the number of FP ResBlocks according to the situation to suit different environments. For instance, on platforms with significant resource constraints, using fewer FP ResBlocks (and correspondingly more binarized ResBlocks) to reduce Params is acceptable, even if it slightly lowers performance. `Q2-2` The parameters and Ops are reduced obviously from full-precision to binary one. But the authors did not give results about inference time on real devices or give some analyses. I am curious how fast the binarized model will be. `A2-2` Thank you for your suggestion. In Supplementary Material A.2, we have compared the inference time of SR3 (FP) and BI-DiffSR (ours). \| Method \| Params (M) \| OPs (G) \| Simulated Time (s) \| \| ---------------- \| :--------: \| :-----: \| :----------------: \| \| SR3 \| 55.41 \| 176.41 \| 55.37 \| \| BI-DiffSR (ours) \| 4.58 \| 36.67 \| 13.00 \| Our method operates faster compared to the full-precision approach (SR3). Note: Real device testing for binary models is limited by specific hardware requirements. Hence, we estimate inference times using the daBNN inference frame. The correlation between the operation count and running time on an ARM64 CPU is: the FP module costs 313.85 ms per GOP, and the BI module costs 354.56 ms per GOP. `Q2-3` The writing can further refine in some cases. For example, in the abstract part (Line 9-10), “… to maintain dimension consistent” should be changed to “… to maintain dimension consistency”. `A2-3` Thank you for the suggestion. We will check the entire paper and improve the writing. `Q2-4` Can this method be applied to other diffusion models, like stable diffusion? If so, can the authors give some suggestions to binarize stable diffusion? `A2-4` Other diffusion models, such as stable diffusion (SD), can be binarized by our method. The SD also employs the UNet noise estimation network, which is composed of ResBlocks and AttentionBlocks. Methodology: 1. Dimension Matching: Adjust the number of channels in different layers of the UNet to match dimensions via our proposed CP-Down, CP-Up, and CS-Fusion. 2. ResBlock Binarization: Directly binarize it using our proposed binary convolution (BI-Conv). 3. AttentionBlock Binarization: For this block, the core is to binarize self-attention (SA). This can be extended from BI-Conv. The SA obtains $Q, K, V$ by three linear projections, and gets the output by another linear. We can use 1x1 BI-Conv to binarize the linear layer. And for the attention computation $\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$, the activations $Q, K, V$, and the intermediate attention map $\text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)$ can be binarized by the sign function Sign(·). `Q2-5` Can we apply this binarization method to other related image restoration tasks? Like image denoising, deblurring? `A2-5` Yes, our method can be applied to other tasks. 1. Direct Application: Our BI-DiffSR model can be directly trained on specific tasks like deblurring using datasets such as GoPro, without structural modifications. 2. Binarization Task-Specific Model: Our binarization method can also binarize diffusion models specific to these tasks. `Q2-6` How long did the authors to train the models? `A2-6` It takes approximately 84 hours to train the BI-DiffSR (×2) on two NVIDIA A100 GPUs. --- Rebuttal Comment 1.1: Comment: I appreciate your thoughtful and detailed replies to my questions. My concerns have been well solved, thus, I tend to increase my score to 7. --- Reply to Comment 1.1.1: Title: Thanks Reviewer Z33c for approving our work Comment: Dear Reviewer ymRm, Thanks for your response. We are happy to see that our answers can solve your concerns. Best, Authors	Summary: This paper introduce a novel binarized diffusion model, BI-DiffSR, for image SR. A UNet architecture optimized for binarization, channel shuffle fusion, and time-step-aware redistribution and activation functions are designed. The experimental results proved the effectiveness of the method. Strengths: 1. This paper is well written, nicely presented, and well organized. 2. Binarized diffusion networks are promising. 3. The performance improvement over other binary SR networks is significant. Weaknesses: 1. Lack of discussion with some related works[1, 2, 3, 4], in particular [1] which is also for binary SR networks. Please analyze and discuss the differences with [1,2]. 2. Ablation experiments are not convincing enough. Comparisons with some other activation function or fusion methods [1, 2, 3, 4] should be included. 3. It is always known that diffusion models are slow. Although binarization will speed up the operation, can it achieve a better trade-off in performance and efficiency than a real-valued efficient SR network. It is suggested to compare with some efficient SR networks [5, 6, 7] in terms of Params, FLOPs, inference time and performance. > 1. Flexible Residual Binarization for Image Super-Resolution. ICML24. > 2. Q-DM: An Efficient Low-bit Quantized Diffusion Model. NIPS23. > 3. Binarized Low-light Raw Video Enhancement. ICCV23. > 4. Binarized Spectral Compressive Imaging. NIPS23. > 5. Efficient long-range attention network for image super-resolution. ECCV22. > 6. DLGSANet: lightweight dynamic local and global self-attention networks for image super-resolution. ICCV23. > 7. Feature modulation transformer: Cross-refinement of global representation via high-frequency prior for image super-resolution. ICCV23. Technical Quality: 3 Clarity: 4 Questions for Authors: See Weaknesses. Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: Limitations were discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer mEsa (denoted as R1) `Q1-1` Lack of discussion with some related works[1, 2, 3, 4], in particular [1] which is also for binary SR networks. Please analyze and discuss the differences with [1,2]. `A1-1` Thanks for your advice. We add more analyses and discussions of related works and incorporate them into our paper. FRB [1] - Differences: 1. *Application Model:* FRB targets one-step models (e.g., Transformer); our BI-DiffSR suits diffusion models with multiple sampling steps. 2. *Binarization Approach:* FRB proposes a new binarization operation (SRB) to reduce performance gaps; our method uses regular binarization operation, while focusing on dimension matching and timestep awareness. 3. *Training Approach:* FRB uses full-precision distillation (DBT) to guide the binarized network; we train our model using a basic L1 loss for generalizability. - Analysis: FRB and our approach improve binarized SR models from different aspects. The SRB and DBT proposed in FRB could potentially be integrated with our method. Q-DM [2] - Differences: 1. *Task:* Q-DM focuses on low-bit quantization (e.g., 2/4-bit) for generative tasks; we address 1-bit quantization for image SR. 1. *Time Aware Approach:* Q-DM stabilizes activations over timesteps by training adaptations; we design timestep awareness TaA and TaR to adjust activation distributions. 1. *Training Strategy:* Q-DM uses distillation to fine-tune binary DM; we train the network with no special training strategy. - Analysis: Q-DM focuses on different aspects from our method but could be integrated to explore better quantization diffusion models. BRVE [3] - Differences: 1. *Fusion Module:* BRVE uses real-value 1×1 Convs in its fusion block, adding parameters; our CS-Fusion is parameter-efficient. 2. *Temporal Awareness:* BRVE enhances restoration using temporal redundancy; our approach adjusts activation according to different sampling timesteps. BiSRNet [4] - Differences: 1. *Structure:* BiSRNet addresses dimension mismatch; our method further considers activation distribution in the skip connections. 2. *Conv Modules:* BiSR-Conv considers dimensional characteristics in SCI; our TaR&TaA in BI-Conv are timestep-oriented designs. `Q1-2` Ablation experiments are not convincing enough. Comparisons with some other activation function or fusion methods [1, 2, 3, 4] should be included. `A1-2` Thank you for the suggestions. We conduct more ablation studies on activation functions and fusion methods, with settings consistent with Sec. 4.2 (ablation study) of our paper. Activation Function: We compare our timestep-aware activation function (TaA) with RPReLU and LeakyReLU, which are applied in previous methods. \| Method \| Params (M) \| OPs (G) \| PSNR (dB) \| LPIPS \| \| :--------- \| :--------: \| :-----: \| :-------: \| :----: \| \| LeakyReLU \| 4.32 \| 36.67 \| 24.92 \| 0.0903 \| \| RPReLU \| 4.37 \| 36.67 \| 29.27 \| 0.0337 \| \| TaA (ours) \| 4.58 \| 36.67 \| 32.66 \| 0.0200 \| Results demonstrate the superior performance of our TaA. Fusion Method: We compare our channel-shuffle fusion (CS-Fusion) with BFB in BRVE [3] and BiFD in BiSRNet [4]. The structures of these fusion modules are detailed in the attached PDF (Figure 1). \| Method \| Params (M) \| OPs (G) \| PSNR (dB) \| LPIPS \| \| :-------- \| :--------: \| :-----: \| :-------: \| :----: \| \| BFB \| 6.38 \| 43.11 \| 30.53 \| 0.0303 \| \| BiFD \| 4.30 \| 36.67 \| 29.67 \| 0.0384 \| \| CS-Fusion \| 4.30 \| 36.67 \| 31.99 \| 0.0261 \| 1. Compared to BiFD, our CS-Fusion promotes fusion without increasing the Params. 2. Compared to BFB, our CS-Fusion has smaller Params and OPs. `Q1-3` It is always known that diffusion models are slow. Although binarization will speed up the operation, can it achieve a better trade-off in performance and efficiency than a real-valued efficient SR network. It is suggested to compare with some efficient SR networks [5, 6, 7] in terms of Params, FLOPs, inference time and performance. `A1-3` Thank you for your suggestion. We compare our proposed BI-DiffSR with ELAN [5] and DLGSANet [6]. ELAN, lacking official results, is retrained under the same settings as our method, yielding ELAN. We test the OPs and latency (simulated by daBNN) with outputs of 3×256×256, and assess LPIPS on Manga109 (×2). \| Method \| Step \| Params (M) \| Per-Step OPs (G) \| Total OPs (G) \| Per-Step Latency (s) \| Total Latency (s) \| LPIPS \| \| :------------------------------ \| :--: \| :--------: \| :--------------: \| :-----------: \| :------------------: \| :---------------: \| :----: \| \| ELAN (Transformer) \| 1 \| 8.25 \| 161.24 \| 161.24 \| 50.61 \| 50.61 \| 0.0206 \| \| DLGSANet (Transformer) \| 1 \| 4.73 \| 73.52 \| 73.52 \| 20.07 \| 20.07 \| 0.0210 \| \| BI-DiffSR (Diffusion, ours) \| 50 \| 4.58 \| 36.67 \| 1833.44 \| 13.00 \| 650.09 \| 0.0172 \| 1. Compared to real-valued efficient SR methods, our BI-DiffSR obtains superior perceptual performance (LPIPS). 2. Our method achieves comparable parameters. Meanwhile, the per-step OPs (36.67 G) and latency (13.00 s) are lower, although the multi-step nature of DM increases total OPs and latency. 3. Model compression (like binarization) and sampling acceleration are two parallel speedup strategies for DMs. As mentioned in the main paper, our method focuses on the former to accelerate one inference step. Thus, we apply the regular sampler (i.e., DDIM, 50 steps) to show its generality. 4. Furthermore, our method has the potential to be used with more advanced samplers to further enhance the efficiency of DMs. --- Rebuttal Comment 1.1: Title: Further Discussion Comment: I appreciate the author's efforts for rebuttal and some of my concerns were addressed. But I still have some concerns regarding the comparison of this work with real-valued networks. 1. PSNR as an important performance evaluation metric needs to be included in the comparison. 2. It is neither fair nor valuable to compare the per-step computational efficiency. As a binary network, it is natural for it to be more efficient per step than a real-valued network. However, for this method, single-step diffusion completely fails to achieve the claimed performance. As it stands, this work is far less efficient than an efficient real-valued network. In the future, with the development of specific hardware and frameworks, this work may be able to realize significant efficiency improvements. Overall, after rebuttal, I'm willing to raise my score to borderline accept. --- Rebuttal 2: Comment: Dear Reviewer mEsa, Thank you for your response. We are pleased to have addressed some of your concerns. Regarding further questions about comparisons with real-valued networks: 1. We re-evaluate the testing results and add the PSNR values in the following table, where our method does not have higher PSNR values. However, it is important to emphasize that diffusion models (DMs) are generative models, which usually do not achieve very high PSNR values. Due to the perception-distortion trade-off, PSNR does not effectively reflect the performance of DMs. Reviewers can refer to the visual results in our main paper and supplementary file, where our method obtains visually pleasing results consistent with the LPIPS comparisons. \| Method \| Step \| Params (M) \| Per-Step OPs (G) \| Total OPs (G) \| Per-Step Latency (s) \| Total Latency (s) \| PSNR (dB) \| LPIPS \| \| :------------------------------ \| :--: \| :--------: \| :--------------: \| :-----------: \| :------------------: \| :---------------: \| :-------: \| :----: \| \| ELAN* (Transformer) \| 1 \| 8.25 \| 161.24 \| 161.24 \| 50.61 \| 50.61 \| 39.34 \| 0.0206 \| \| DLGSANet (Transformer) \| 1 \| 4.73 \| 73.52 \| 73.52 \| 20.07 \| 20.07 \| 39.57 \| 0.0210 \| \| BI-DiffSR (Diffusion, ours) \| 50 \| 4.58 \| 36.67 \| 1833.44 \| 13.00 \| 650.09 \| 33.99 \| 0.0172 \| 2. Providing the single-step efficiency of DMs is to demonstrate the effectiveness of our proposed method, since our method focuses on compressing single-step sampling. As we mentioned, as the framework evolves (e.g., fewer sampling steps), the efficiency of DMs can be further improved. However, this is not the focus of our research. In the future, we will explore applying our method to more advanced architectures to explore more efficient DM. 3. We totally agree with Reviewer mEsa that with the development of specific quantization-friendly hardware and frameworks, our work may be able to realize significant efficiency improvements. We will explore this direction in the future. Thanks for the suggestions. Best regards, Authors Title: Further Discussion on Comparison with Real-Valued Network	Rebuttal 1: Rebuttal: # Response to all reviewers and area chairs Dear Reviewers and Area Chairs, We thank all reviewers (R1-mEsa, R2-Z33c, R3-KWS7, R4-MRyP) and area chairs for their insightful comments and valuable time. We are pleased that: - R2 and R3 appreciate our intuitive motivation and diverse and insightful analyses. - R2, R3, and R4 acknowledge the novelty of our techniques. - All reviewers recognize the impressive performance of our method, which surpasses existing approaches. We have responded individually to each reviewer to address any concerns. Here, we offer a summary: - We discuss the differences with more related methods, including binarized SR and quantized DM. - We add more experiments, including ablation studies and comparisons on the new baseline (ResShift) with new metrics. - We compare our approach with efficient SR networks and provide an analysis. - We add a preliminary section to our paper and commit to improving our manuscript. - We introduce methods to extend our approach to new modules, models, and tasks. - We clarify the gap between theory and practice, the advantages of the binarization method, the innovative aspects of BI-Conv, and the difference between TaR and time embedding. - We address various detailed questions raised by reviewers, including actual running speeds, training times, module efficiencies, and ARM hardware types. - Finally, we include an attached PDF as a supplement for some questions. Thanks again to all the reviewers and area chairs. We appreciate you taking the time to review our responses and hope to discuss further whether the issues have been resolved. If you need any clarification, please let us know. Best Regards, Authors Pdf: /pdf/a371de219dd8e9bc064c5962380ca4f5430eedbb.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Controlled maximal variability along with reliable performance in recurrent neural networks	Accept (poster)	Summary: The authors propose a principle for selecting actions to drive recurrent neural network activities which aims at maximizing the variability of the neural activity while avoiding unwanted states. They define unwanted states as states where no action is possible, and use a reinforcement learning framework to select the input and analyze the coupling between action choice and network dynamics. They apply their networks and input selection to a few tasks where maximizing the entropy within some boundaries is defined as success, and show that the network performs those tasks. Strengths: The framework is, to the best of my knowledge, original. It is also sound in its analysis, and well explained. Weaknesses: I have two concerns, the first one is (for lack of a better word) teleological, the second more practical. My first concern is that the authors use the word performance often and refer to their networks as solving the task. But from what I can read, the "task" of filling as much of the space (MOP) as possible was never given to the agent/input controller. Thus, if the task of the network is to maximize its entropy while remaining on a bounded region, and the R network is not taught to maximize entropy, we can hardly argue that the R network failed. I get that the point that MOP can be useful for some problems focusing on exploration, but it is a bit odd to talk about solving tasks and performances when such performance was not told to the agent, but rather an agent was build for that task (or for another one in the case of R). Please correct me if I misunderstood. My second concern is that an agent with a binary value function which chooses random actions (value 1) except if those have a punishment (value 0). Such value might be easy to learn (it only requires learning a boundary, which is smooth in the problems presented), and if the state space of the actions is large, random choices could be very close to MOP (because a random action sequence is likely to go through many states). While this is not necessarily true, it is worth checking, as it would make the whole MOP less impactful. Technical remarks: - The addition of stochasticity to the R network is a bit tricky, because the MOP agent never had the problem that it might "accidentally" jump into the terminal state if it did not want to. Thus, for "survival" it might be better to simply stay in some very small region that is as far as possible from the terminal states, because there we find the smallest chance of accidentally going to the wrong region. A better option would be to give R a random action and then ask if it wants to take it. But I suspect that this would give a high variability. Minor issues about literature and references - When the authors mention that usually recurrent neural networks tend to have neurons with saturation, it seems like an unfair comparison. A network trained with a specific task do not have an incentive to maximize the number of states, but this could be added to the loss function (if there is one explicitly) or simply enforced by some intrinsic plasticity rule, as some works in reservoir computing have done (both in Echo State Networks for machine learning and biological models such as SORN). Also, it might simply be the case that for a given task it is better to be close to the saturation. - In the discussion the authors rightfully mention the variability found in songbirds. But they do not note that it agrees with the works that they mention in the introduction (ref 29-35) which use variability only during learning, and that they use as a current limitation motivating their work. It would seem natural from a neuroscience point of view that variability is suppressed during courting by some executive brain area, rather than assuming that the area that generates the song naturally knows when to change behaviors. - The argument that terminal states are those where the agent has no available actions is a bit limiting. Any task that corresponds to reaching a goal (for example a location) leads by definition to a reduced action space, at least in practice (the agent would not leave that position). A note about this would be beneficial. Technical Quality: 3 Clarity: 3 Questions for Authors: Finally, the authors mention that MOP might take deterministic actions to increase entropy later. This would be, in my opinion, an important contribution, but I haven't seen this being done explicitly in the current paper. Why not try to have two circles connected by a very thin line, to showcase this? There is a edge case that bothers me. If we consider the perfect value function, all action paths within the boundary have a limited entropy, but an action path that leads to leave the boundary has technically infinite entropy (it is never taken, thus its probability would be zero, hence infinite entropy). If this is correct, how does the neural network for the values avoid this problem? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: There is no negative societal impact. I think the limitation of learning a policy is fair, but there are more important concerns, namely Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: $\textit{Weaknesses: I have […] controller.}$ The reviewer is correct, the only goal of the MOP agent is to maximize future action-path entropy, Eq. 3. The ‘task’ of filling out space was never given but it emerges naturally as the MOP agent seeks to maximize action-path entropy while avoiding terminal states. Our aim is to demonstrate that MOP agents can generate interesting not-enforced behaviors from a simple objective. We acknowledge that our use of the term 'tasks' may be confusing, and we will clarify it in the next version. $\textit{Thus, […] misunderstood.}$ The reviewer raises a valid point, and we appreciate the opportunity to clarify. The R network indeed does not fail in its task, as it is designed to maximize extrinsic rewards (Eq. 10) and learns that increasing its survival leads to the accumulation of larger cumulative reward. Both MOP and R networks have an intrinsic drive for survival, so it is fair to compare their ability to generate as well emerging behaviors. With equal survival time, MOP agents, driven by an intrinsic occupancy goal, explore states more broadly than the R network. As experimental complexity increases, matching survival times requires to quench the source of variability of R agents. In summary, while R agents achieve their task of maximizing rewards, they lack emerging variability, which is a desired property once performance metrics have been optimized. We will clarify this in the manuscript. $\textit{My second […] impactful.}$ See our comments below. $\textit{Technical remarks: The addition of … }$ We apologize for the misunderstanding. MOP agents act according to policy in Eq. 4, which follows a Boltzmann distribution and assigns a non-zero probability to all actions. Hence, they do face the possibility of accidentally jumping into terminal states. Consequently, the comparison between MOP and R agents is not as unfair as the reviewer implies. We will add paragraph P6 of the global rebuttal to the manuscript to clarify this. $\textit{Thus, for survival …}$ As discussed above, both MOP and R agents can accidentally fall outside the boundary but differ in their risk sensitivity. MOP agents, through action exploration, learn better the value function near boundaries, allowing them to be riskier and increase state space occupation. In contrast, R agents, due to their inherent randomness, play it safer and stay more confined, strongly limiting their learning. Following the reviewer’s suggestion, we created a super-R agent with binary value function, i.e., it takes random actions at any time and it can choose to avoid wrong actions leading to a boundary. Not surprisingly, this agent is riskier and comparable to MOP in terms of average occupancy (Fig. C of the rebuttal) and state occupancy entropy ($H_{MOP}=7.34, H_{super-R}=7.38$, where occupancy probability is approximated with the visitation frequency). We think that this super-R agent does not represent a fair comparison, since it hard-wires the avoidance of terminal states, unlike MOP agents. $\textit{Minor issues ... When […] saturation.}$ We agree that RNNs lack an inherent incentive to maximize the number of states, and hence, exploiting the saturation state could be a beneficial strategy in some tasks. However, this is unlikely to be observed in biological neurons, which benefit from maximum information transmission [47]. The reviewer mentions relevant literature where networks are encouraged to live in a ‘healthy’ regime. While we acknowledge that synaptic plasticity can stabilize the network’s regime, our framework only involves learning the value approximator. Hence, the involved time scales are instantaneous and achieved through active control rather than synaptic change. We comment on this in paragraph P7. $\textit{In the discussion […] behaviors.}$ The reviewer correctly notes that much of the literature, including studies on songbirds, focuses on learning. Nevertheless, flexible switching from stochastic to deterministic behavior is observed also in adult birds, where different levels of variability appear in directed and undirected singing even after learning [57]. While various explanations exist for this phenomenon (e.g., variability suppression by executive areas), this example aligns with our thesis that flexible switching between stochastic and deterministic modes may play a cardinal role in biological networks. We will clarify this by adding paragraph P8. $\textit{The argument […] beneficial.}$ We thank the reviewer for the opportunity to highlight our perspective shift. Our framework does not include the notion of reaching a goal. We hope to address a partial limitation in classical neuroscience tasks, where tasks are designed as to have a fixed point of behavior and episodes terminate upon reaching it. This approach is valuable for specific studies, as characterizing the neural basis of motor reaching, but departs from real-world scenarios where agents’ goals are continuously reallocated. We believe studying this reallocation is crucial and presents future challenges for computational neuroscience. By allowing MOP to dynamically develop its own survival behavior we move towards a more flexible and realistic model. To clarify this, we will add paragraph P9. $\textit{Questions: Finally, […] this?}$ We followed the suggestion and implemented a two circles arena connected by a narrow corridor in the neural space (Fig. D of the figures). We show that the MOP network, starting from the left room of the arena, crosses the narrow corridor by reducing its action entropy, to furtherly increase its action entropy in the right room of the arena. We will add this result to the Appendix of the manuscript. $\textit{There is a edge case […] }$ As commented above, the MOP policy follows a Boltzmann distribution, hence even the perfect value function would not lead to an immediate intrinsic reward of infinity. We will clarify this by adding paragraph P6 to the manuscript. --- Rebuttal Comment 1.1: Title: Appreciated modifications Comment: The authors addressed my concerns. I will update my score --- Reply to Comment 1.1.1: Comment: Thank you very much for your support.	Summary: In natural behaviors, there’s usually variability despite being able to perform tasks with high performance. This paper aims to understand whether it’s possible for neural networks to have high variability while maintaining high task performance and being able to switch to deterministic behavior modes when needed. This paper uses an RNN with fixed weights to model the state of the environment and how it’s affected by the action of the agent. The agent is modeled by a controller which aims at optimizing the occupancy of the action-state space. This is achieved by having a reward that increases when the agent selects an action that is less likely with the current policy. The optimal value function, and in turn the policy, is approximated by using a single hidden layer feedforward NN. Strengths: The paper presents a reward function that maximizes the action path entropy, and provides interesting examples of how this network performs in three different example tasks. The writing is clear and easy to follow. Weaknesses: 1. The tasks are mostly limited to setting several terminating states, so that the MOP network learns to avoid the terminating states while maximizing the entropy. I wonder how general this type of tasks is, or how more interesting RL tasks may or may not be formulated this way. 2. The motivation of this paper is unclear to me. The authors aim to show that NNs can have high variability with good performance, in order to match natural behaviors, and to propose possible mechanisms for neural variability. However, there is no comparison with experiments to show how well the behavior of the NN matches natural behaviors, or to show how the proposed reward may be superior. The authors also do not explain the generation of neural variability, but instead directly enforce variability in the MOP network. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How does the structure (the number of hidden layer neurons, depth) of the NN that serves as value function approximator, affect the results? 2. In Fig 3(c), why is the action entropy low in the lower right and upper left corners but not the other two corners? one would naively expect them to be symmetric. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: limitations are covered in the 'Weakness' section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: $\textit{ Weaknesses: 1. The tasks are …}$ We thank the reviewer for giving us the possibility to clarify the role of terminal states in our algorithm and to highlight the general applicability of our framework to various RL tasks that are not typically formulated in this manner. To illustrate this, we performed an additional simulation where we applied MOP to solve the classical task of balancing the inverted pendulum (Fig. A of the rebuttal figures). We show that the feedforward network quickly learns the approximate the value function and to balance the pole without the need of any extrinsic reward. Furthermore, our framework is compatible with the incorporation of extrinsic rewards. We have already exploited this property in Appendix D by employing an approximation for the state entropy as an extra reward term (Fig. 5). To further clarify this, we added a simulation where the MOP network gets an extrinsic reward $r = 1$ for occupying the top right region of the square in the $(x_1, x_2)$ space (Fig. B of the rebuttal figures). The reward function now consists of two terms, an extrinsic reward component and an entropic one. By adjusting the temperature parameter balancing the two terms, we may observe different behaviors (see left and right panels in Fig. B of the rebuttal figures document), as the network differently balances the internal drive for occupation with the drive to maximize extrinsic reward. We will add these examples in the Appendix to illustrate the generality of our approach. $\textit{ 2. The motivation … }$ The reviewer raises a significant point. Our model employs rate-based recurrent neural networks, a feature that makes the direct comparison with neural data unlikely and of difficult interpretation. A significant direction for future work would be extending our framework to spiking neural networks, where the statistics of the spiking could be easily interpreted and confronted with the Fano Factor and the Coefficient of Variation from experimental data. Nevertheless, with our current setting of rate-based dynamics, we find qualitative similarities by looking at fluctuations, as BOLD signals showcasing large variability at the global scale. Moreover, a state-dependent adaptation of neural variability has been largely reported, with variability (as measured by the Fano Factor) significantly reducing when the stimulus is presented (Churchland et al. 2010). From the perspective of the subjects performing the experiments, switching to a deterministic mode to faithfully encode the sensory stimulus is what guarantees survival. We will add these predictions by adding paragraph P1 of the rebuttal global response to the Discussion section. $\textit{ Questions: 1. How does the structure …}$ We thank the reviewer for giving us the possibility to highlight the robustness of the value approximator with respect to changes in the chosen parameters. In particular, our value approximator is a feedforward network of one hidden layer, i.e., we chose the simplest structure allowing to solve nonlinearly separable problems. We found that this simple feedforward network can correctly approximate the value function for a large set of parameters. In terms of number of hidden neurons, we found that this parameter does not strongly affect the outcome of the simulation even when this number was significantly reduced. More relevantly, we proved the stability of the approximator by feeding into the value network only a subset of the neurons from the recurrent neural network. We will clarify the robustness of our results against hyperparameter choices in the subsection regarding the parameters of simulation in Appendix A by adding the following: We found our algorithm to be robust across various parameter configurations, including the number of neurons in the hidden and input layers, as well as the total number of layers. An exhaustive characterization of parameter behavior was not conducted, as fine-tuning was not necessary to effectively train the approximator. $\textit{ 2. In Fig 3(c), …}$ The reviewer is correct. There is a major observed asymmetry, which could be explained by the fact that all neurons in the network share a common low-dimensional input, i.e. each component of the action signal acts all on the RNN neurons. Therefore, the resulting activities are expected to be highly correlated. We expect these correlations to be attenuated via acting on the read-in matrix K, for instance imposing sparseness on this matrix or increasing the dimensionality of the action signal. This analysis represents a relevant direction for future work which will allow to potentially exclude the presence of other factors (e.g., initialization of the weights in the value function) influencing this asymmetry.	Summary: This paper applies the maximum occupancy principle (MOP) -- previously introduced as a normative theory of behavioural variability -- to recurrent neural networks, thereby proposing MOP as a normative theory of neural variability. The MOP postulates that an agent seeks to maximize future occupancy of its state-action space. A key insight of the previously published MOP paper (ref 42) was that an MOP-following agent naturally learns “good” behaviour by actively avoiding terminal states (e.g. death) as those imply very reduced future state occupancy. This effect is demonstrated again here by showing that MOP-following RNNs can be made to avoid specific activity patterns, but remain maximally variable otherwise. In terms of methods, the main challenge here was to approximate the MOP value function for nonlinear RNNs -- the sole determinant of the optimal policy. The authors do so by training a NN value function approximator on a regression objective derived from the self-consistency (Bellman) equation that the value function must satisfy. The framework is then applied to a few toy setups, including a context-dependent drawing task. Some technical limitations are discussed. Strengths: The idea of applying MOP to RNNs is potentially interesting, as it provides a new normative theory of neural variability that will be interesting to confront to neural data -- this paper provides some technical foundations for investigating this hypothesis further. The paper is technically well executed. Weaknesses: There is some, but not an awful lot of, added value relative to ref 42. The MOP is a creative new framework, but the idea that biological agents learn what _not_ to do _instead of_ learning what to do seems a hard sell. The drawing tasks of Figs 3+4 seem carefully designed to demonstrate that MOP-following networks can achieve some “positive” functionality by exclusion, but I have a hard time imagining how the framework would scale to even simple control tasks like swinging a pendulum up; many such tasks are defined by what the agent must do, and many suboptimal states are actually not at all absorbing / terminal. I think the paper probably ought to discuss these limitations in more depth. There is at least one other normative theory of neural variability that ought to be mentioned in the intro more explicitly: sampling-based probabilistic inference, where variability represents uncertainty. (Ref 26 is cited for "nonlinear network interactions leading to variable activity patterns" but has nothing to do with networks. Echeveste et al 2020 by the same group might be more appropriate in this context.) In summary, I think this idea is potentially interesting -- I view it as a putatively useful theoretical framework for studying how brains learn from bad outcomes (which engage a very different system from the brain's dopaminergic reward system). However, the paper as it currently stands is rather incremental and perhaps not of broad appeal to the NeurIPS community; I would strongly encourage the authors to explore the ramifications of the neural-MOP framework for neuroscience, articulating predictions for neural variability in specific setups where neural data is available for confrontation. Technical Quality: 3 Clarity: 4 Questions for Authors: Minor typos I picked up: - sentence on l.54 (“This theory frames...”) is grammatically weird. - l.128: inevitably constraint → constrain - l.135: of parameters → with parameters - l.178: a terminal states → state Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: Some technical limitations are discussed but it would be nice to more clearly spell out the conceptual limitations (c.f. above). EDIT: following discussion with the authors, I am raising my score from 5 to 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: $\textit{There is some, […] discuss these limitations in more depth.}$ We appreciate this criticism, which allows to discuss more deeply some fundamental features of our MOP network. Let us take the example of the inverted pendulum that the reviewer has brought up. Our intrinsic motivation approach with value-function approximation can generate interesting behavior in this example, including balancing the pole upright. To show this, we have substituted the environment in the MOP network, i.e., the RNN, with a cart-pole constituted of a cart moving along a linear track and a vertical pole attached to it (Fig. A of the rebuttal figures). The cart-pole receives five possible actions: an input to the left, to the right, a strong input to the left, a strong input to the right, and no input, all acting directly only on the cart. Terminal states are encountered whenever the angle between the pole and the normal to the cart exceeds a certain threshold ($\|θ_{th} \|=0.74 rad$) or the cart hits any of the borders of the linear track ($\|x_{th} \|=2.4 a.u.$). The value function is approximated via a FFN receiving as input the variables of the cartpole (position, linear velocity, angle, angular velocity) and following the training described in the manuscript. As learning progresses, MOP improves its ability to avoid terminal states (Fig. A). Our algorithm can balance the pole, i.e., avoiding terminal states, while at the same time showing variability in the pole angles and carts positions. This example shows that MOP with terminal states can generate quite interesting behavior, without the need of telling the system what exactly to do. A reward maximizer agent will not lead to variable behavior and instead would only show the up-right position. How far one can go with MOP across different dynamical systems and generate interesting variability is a matter of future exploration. Moreover, we want to remark that we did not intend to say that biological agents learn what not to do instead of learning what to do. As mentioned in lines 127 and 128 of the manuscript, we want to mention the fundamental difference of having extrinsic rewards versus intrinsically motivated agents, with the consequences it has on the generated variability. By introducing extrinsic reward, agents are constrained to choose a single behavior (the optimal one) and follow it, while introducing terminal states we let the agent free to develop all possible behaviors allowed to solve the task. We will add the clarification paragraph P5 of the rebuttal global response to the Discussion section. $\textit{There is at least one other normative theory […] }$ We thank the reviewer for the suggestion, and we will correct the reference in the next version of the manuscript. We will also mention the literature on sampling-based probabilistic inference. $\textit{In summary, I think that the idea is potentially interesting […] }$ We agree with the reviewer that we are building on a previously published work to extend the idea of the Maximum Occupancy Principle to neural activity. However, we believe that both the technical and conceptual contributions of this paper are significant for the NeurIPS community. Firstly, we demonstrate how to control a high-dimensional, chaotic network to perform various tasks using external input currents based on the MOP framework, without the need to train network weights or rely on extrinsic reward signals from the environment. This approach is not limited to simple tasks, as we have shown. Second, we have presented a radically novel theory for neural variability, which is a major theoretical step. There is no precedent in the literature of this new hypothesis and worked out framework. It is possible that these two points were not clearly explained, so we will rewrite old sentences to improve understandability: We show that the principle of maximizing occupancy can be extended beyond behavior and may also underpin neural activity. Our work introduces a radically novel theory of neural activity, where neural variability is a cardinal feature rather than a drawback. To illustrate this, we trained MOP agents to control high-dimensional chaotic RNNs in order to perform different tasks defined by the structure of terminal states and the environment, without modifying the internal weights of the network and or providing any extrinsic reward. $\textit{I would strongly encourage the authors […] }$ We thank the reviewer for the encouraging comments. $\textit{Questions: Minor typos … }$ We thank the reviewer for picking up these typos, we will correct them all in the manuscript. $\textit{Limitations: Some technical limitations are discussed but it would be nice to more clearly spell out the conceptual limitations (c.f. above). }$ We will add the text indicated in the above responses to address this comment. --- Rebuttal Comment 1.1: Comment: Thank you for the thorough rebuttal. I have read the other reviews & associated authors' responses and am inclined to keep my score at 5; it's difficult for me to score higher, because whilst I do like the idea of MOP at the network level and find this paper well executed, I also think that for the NeurIPS high bar perhaps the current version of the theory is not sufficiently testable and it is difficult to gauge the potential impact this paper will have in the field -- in this respect, the rebuttal is not entirely convincing. The addition of the cart-pole system is nice, but not what I was looking for / suggesting. This is _not_ learning to swing a pendulum up -- it's learning to maintain the pole in already upright position. This is actually an example where "what to do" and "what not to do" are very tightly aligned (i.e. staying up = not swinging down), and indeed the way your report performance here is as "survival" (in avoiding bad states), not as "time to success" (actively reaching good states). In other words, this new experiment doesn't address my original concern. What's a lot trickier is to learn to swing the pendulum up: I would imagine that it's hard to encode this in terms of terminal states (I can't imagine any such encoding whereby the initial downright state would not already be labelled as "terminal"). My scoring also reflects a (noisy) comparative assessment of the strength of the field based on my other review assignments. Let me also reiterate the words of encouragement I gave in my original review, as I think that with more work on the neural side, this story could potentially be important for the field if it turns out that it explains aspects of neural variability that are not explained by other theories. --- Reply to Comment 1.1.1: Comment: We sincerely thank the reviewer for their additional comments and for encouraging us to continue developing our theory. We would like to clarify that we do not claim that reward maximization is equivalent to MOP; rather, we view them as complementary approaches. For example, in certain scenarios, such as swinging a pole upright, it may be more straightforward to define a task using a reward function. However, in other cases, defining goals in terms of terminal states may be more appropriate, such as avoiding death in living systems. We believe our approach holds significant value, and we will emphasize this innovative aspect of our theory in the next revision of the paper. In the mentioned example of swinging the pole up, one may choose to define the terminal states differently, by setting different thresholds on the velocities. For instance, we could define terminal states as those where the system moves slowly while in a 'down' position and quickly while in an 'up' position. This would encourage agents to move swiftly while down to reach and maintain a stable up state. Indeed, while ‘down' states with high speeds are allowed, they would be less desirable due to the inherent risks and, thus, not incentivized by MOP. We expect MOP to behave deterministically to swing the pole up, similar to how in Fig. D of the Rebuttal the network learned to behave very deterministically to navigate the narrow corridor and reach the other room in the neural space. In regards to the reviewer’s comments regarding the learning process, we understand their concerns. Nevertheless, their points allow us to emphasize a critical aspect of our framework. By interacting with the environment, agents learn what not to do and to avoid states that are terminal for them. Hence, while it is true we focus on tasks where the agent already starts in a favorable position (e.g., the pendulum with the pole up), we believe this is a fair assumption, as the only requirement is that agents are already alive. Provided that, initial conditions do not play a role. By introducing MOP, we aim to study natural agents and consider tasks that may be relevant for them. Lastly, we would like to highlight that reward maximization and MOP are not mutually exclusive and can indeed be combined, as demonstrated in Fig. B of the Rebuttal. In that instance, we integrated MOP with a simple reward function (r = 1 everywhere within a certain region of the neural space). However, our framework is more general, allowing for an arbitrary reward structure. We thank you again for engaging in this discussion.	Summary: This paper proposes a mechanism to induce variability in "reservoir" recurrent neural networks without impinging upon task performance, by maximising the cumulative entropy of future states and actions/behaviors. These actions are provided by a controller network to the reservoir as input currents. The authors demonstrate through experiments that the induced variability does not come at the cost of adhering to constraints on energy or specific neuronal activities, or task performance (there is no explicit reward in these tasks for the proposed framework, and so this is measured by the survival time, i.e., timesteps until a terminal state is reached or a constraint is violated). Comparisons with networks without the input current modulation or those with explicit rewards show that the former are unable to properly satisfy task constraints, while the latter find overly conservative or "risk-averse" solutions that suppress variability. The demonstrations also show that the proposed framework leads to networks switching between more deterministic modes of computation (lower action entropy) near terminal states/constraints and more stochastic (higher action entropy) ones otherwise. Overall, this paper provides a novel perspective on how neural variability could be maximized while still allowing for accurate performance and avoiding terminal states, especially when an explicit reward function is unavailable or undesirable. Strengths: 1. The paper is well-motivated, clear, and provides a unique perspective on how there could be controlled variability in a system without negatively affecting task performance. The figures and overall presentation are good and clearly demonstrate the validity of the central claims. 2. The experimental results include specific controls for the proposed mechanism – the authors show results for networks without any input current modulation and also networks with explicit constraints imposed by a reward function. 3. In the appendix, the authors show that their central claims are largely valid even when there are other sources of variability such as intrinsic noise in the networks. 4. While the tasks are simplistic, they are well-designed and quite interpretable, allowing the claims to be validated easily through the visualizations. Weaknesses: 1. This is a limitation acknowledged by the authors, but it seems like the computational complexity for the framework is quite high, so it is difficult to evaluate how this would scale to more complex tasks or multi-task settings. 2. To my knowledge, and perhaps I have missed this, but the authors have not provided clear connections to the biological inspirations for the proposed mechanism, such as perhaps neuromodulatory mechanisms, or comments on its biological realism. Could the authors elaborate on this and are there any testable predictions for this model of neural variability? 3. While the simplicity of the tasks is an asset, it would also be important to see what happens when there is a greater diversity among tasks in a multi-task setting (see Yang et al. [1]), and when performance is not only linked to constraint satisfaction or "survival". For example, would this framework impede performance when one of the tasks explicitly requires less variability as is perhaps the case in a memory task (see Yang et al. [1] again for examples)? 4. This is a minor point, and a suggestion rather than a weakness, but there are some interesting recent works that could be mentioned to strengthen the background: 1. Takasu & Aoyagi [2] discussed an input current modulation mechanism and how it affects the Lyapunov exponents of reservoir networks' dynamics – specifically, suppressing chaos and ensuring networks are at the edge of chaos to enable effective information processing (and is thus related to the variability in these networks; also related to [59] from the paper). It would be interesting to briefly compare/contrast the proposed mechanism/goals to that proposed in [2]. 2. In lines 33-37, the authors discuss works where internal synaptic noise is proposed as a mechanism for neural variability, and mention that some works use this assumption to "describe variability during spontaneous activity–in the absence of sensory stimuli". Works such as Asabuki & Fukai [3] and Krishna et al. [4], where such a mechanism is assumed and used to describe properties of spontaneous activity, could be discussed (in addition to [18, 20] from the paper) to provide a better idea of the implications of such mechanisms. References: 1. Yang et al. “Task representations in neural networks trained to perform many cognitive tasks.” Nature neuroscience vol. 22,2 (2019): 297-306. 2. Takasu & Aoyagi. “Suppression of chaos in a partially driven recurrent neural network.” Phys. Rev. Research 6, 013172 (2024). 3. Asabuki & Fukai. “Learning rules for cortical-like spontaneous replay of an internal model.” bioRxiv (2023): 2023-02. 4. Krishna et al. “Sufficient conditions for offline reactivation in recurrent neural networks.” The Twelfth International Conference on Learning Representations (2024). Technical Quality: 4 Clarity: 4 Questions for Authors: See the Weaknesses section. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have adequately discussed limitations related to computational complexity and not learning the policy in the Discussion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: $\textit{Weaknesses: This is […] multi-task settings.}$ We agree with the reviewer’s comments regarding the high computational cost of our framework, as acknowledged in the manuscript. This complexity arises from a specific choice we are committed to in our current approach, which employs an exact representation of the policy (Eq. 4). However, there are several ways to improve computational complexity, by furtherly reducing the dimension of the already low-dimensional action space ($N_c=8$), or by approximating the policy by a neural network. These limitations are already commented in the discussion section. $\textit{To my knowledge, […] neural variability? }$ We thank the reviewer for the question. We will address this comment on these points in the next revision by adding paragraph P1 you will find in the global responses of the rebuttal. $\textit{While the simplicity of the tasks […] for examples)?}$ We thank the reviewer for the question. We predict that our MOP framework will lead to deterministic behavior when required by any task, as it would be in a working memory (WM) task where a memory item needs to be stored. In such cases, we do not expect that a MOP neural network would generate variability that impairs performance. However, consider a more interesting scenario where many memory items need to be stored in a limited WM system, but it is unclear which one should be stored at any time. In this situation, a system like MOP, which promotes variability on the stored items while keeping functionality of the network, might better suited to solve the task. The limited time does not allow us to explore this, but it represents a relevant direction for future work. To complete our response, we have added two additional simulations that partially address the reviewer’s comments. First, we show that in a standard cartpole task (Fig. A of the rebuttal figures) MOP agents are capable of generating variable behavior while controlling non-linear dynamics and confining the corresponding state variables in a desired space region – a reward maximizer agent will lead to just the up-right position without any variability. Second, in a scenario with two arenas connected by a corridor in the neural space (Fig. D of the rebuttal figures), we find that a MOP agent can generate deterministic actions to move neural activity through the narrow corridor. These examples demonstrate once again that MOP networks are capable of deterministic behavior when needed. We will add paragraph P2 of the rebuttal global response to the Discussion section. $\textit{This is a minor point […] background: Takasu and Aoyagi [2] ...}$ We thank the reviewer for the suggestion. The proposed paper investigates the chaoticity in RNNs as a function of the input statistics, using a mean field approach to compute the maximum Lyapunov exponent (MLE). The study finds maximum memory capacity for MLE close to zero, indicating that the network operates at the edge of chaos. This aligns with previous literature suggesting that information transmission can be maximized near the chaotic state boundary. Testing the behavior of the MLE in our framework is a promising direction for future work. Our input current statistics do not allow for a closed mean-field form, hence we propose a numerical computation could be performed as in Laje and Buonomano 2013. We expect MOP networks to maintain consistent neural trajectory distances despite external perturbations, hence approximating an MLE close to zero and supporting the reviewer's points. We will add to the Discussion section paragraph P3. $\textit{In lines 33-37, …}$ We thank the reviewer for pointing out relevant literature on the role spontaneous activity and reactivation play in the brain, which will include in the next version of the manuscript. The mentioned papers offer valuable insights into reactivation patterns using RNNs, driven by observations that spontaneous activity in the brain resembles stimulus-evoked activity. Reactivation is thought to play a relevant role in the brain, such as facilitating the stabilization of memory patterns or the generalization of acquired knowledge. Asabuki et al, utilize reservoir computing and introduce a learning rule based on neuron responses’ prediction, which allow for the formation of an internal model representation of sensory experiences and enabling reactivation without inputs. Relevantly, the learning rule minimizes the Bayesian surprise by reducing the KL divergence between prior and posterior distribution. Krishna et al. show that continuous-time, noisy RNNs trained to track state variables in neuroscience-related tasks (as spatial navigation), learn to match the desired output while compensating for intrinsic noise. The defined dynamics produces patterns of reactivation even in absence of external input. In particular, it is the very presence of noise that fundamentally shapes the dynamics and leads to the emerging property of the reactivation patterns. Both approaches enforce reactivation patterns by indicating the network the ‘favorable’ patterns to be reproduced, either by minimizing the Bayesian surprise or by informing the loss function with a time-dependent target. In our framework, the goal-directed behaviors naturally emerge from the intrinsic motivation of generating action path entropy and the resulting neural patterns are as variable as the terminal states defining the task allow. If we assume that reactivation of patterns is linked with the replay of activity that is relevant for survival, then we may expect MOP networks to favor those states by assigning greater values to those activity states. We appreciate the reviewer’s suggestions for future research directions and will add paragraph P4 to the Discussion section. --- Rebuttal Comment 1.1: Comment: I thank the authors for their rebuttal and appreciate their additional experiments and discussion. A clarification in P4 of the discussion: Asabuki & Fukai (2023) and Krishna et al. (2024) do not utilize reservoir computing but in fact use trained RNNs, my comment was more about how inherent variability/stochasticity in the neural activity has been considered when theoretically/computationally modeling spontaneous activity patterns (in the absence of sensory stimuli, ref. Introduction, line 36). Specifically in Krishna et al., that noise is perhaps directly responsible for such reactivation (which the authors correctly stated in the reviewer-specific rebuttal comment above). I would ask the authors to slightly revise the statement in light of this. Overall, I think this is a good contribution and maintain my positive opinion and score. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for the positive words. The proposed papers employ the learning rules commented above to update the weights of the recurrent network. We apologize for the poor choice of words and we will modify P4 of the discussion as follows: "We postulate variability to play a relevant role in neural activity as visitation of all activity states may increase flexibility and help generalization. The phenomenon of underlying occupancy of previous states via the replay of favorable patterns of activity is commonly observed in the brain. Various mechanisms in recurrent neural networks have been proposed to model the emergence of such reactivation in network activity (Asabuki and Fukai, 2023; Krishna et al., 2024). The defined dynamics produces reactivation patterns even in absence of external inputs. In particular in (Krishna et al., 2024), it is the very presence of noise that fundamentally shapes the dynamics and leads to the emerging property of the reactivation patterns. Within our framework, we anticipate that MOP will facilitate the reactivation of all activity patterns relevant to survival, exhibiting deterministic behavior when required by the task."	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for their insightful comments and for giving us the opportunity to clarify important aspects of our framework. Based on the feedback received, we would propose to modify the Discussion section incorporating the following additional paragraphs. $\textbf{Addition P1}$ : Our model offers several testable predictions regarding the nature of neural variability we should expect in the brain. Firstly, it predicts that neural variability will persist even after extensive training, aligning with studies reporting large spiking variability even in well-trained non-human primates [19, 27]. Despite this persistence, our model also suggests that neural variability may decrease when terminal states are sufficiently close, i.e., when the system enters into a deterministic mode (Churchland et al. 2010, Nogueira et al. 2018). Finally, our model predicts that reward signaling systems in the brain will also signal intrinsic motivation rewards. This is supported by recent studies demonstrating that spontaneous movements elicit dopamine release [11]. $\textbf{Addition P2}$ : We have shown that our MOP network is able to solve different tasks while generating variability. In Appendix, we test our algorithm in tasks where more deterministic modes are required, as when balancing the cartpole or when crossing a narrow corridor in the neural space. $\textbf{Addition P3}$ : We have demonstrated that the MOP network can control high-dimensional RNN activity in the chaotic regime. Future research should investigate how MOP-driven input currents affect the RNN's regime itself. We anticipate MOP currents will stabilize neural trajectories, consistent with operating at the edge of chaos, as indicated by a Maximum Lyapunov Exponent close to zero. Characterizing chaoticity as a function of the input properties (e.g., magnitude) (Takasu & Ayoagi 2024) is a promising direction. $\textbf{Addition P4}$ : We postulate variability to play a relevant role in neural activity as visitation of all activity states may increase flexibility and help generalization. The phenomenon of underlying occupancy of previous states via the replay of favorable patterns of activity is commonly observed in the brain. Various mechanisms in reservoir computing have been proposed to model the emergence of such reactivation in network activity (Asabuki and Fukai, 2023; Krishna et al., 2024). Within our framework, we anticipate that MOP will facilitate the reactivation of all activity patterns relevant to survival, exhibiting deterministic behavior when required by the task. $\textbf{Addition P5}$ : In our framework, we challenge the canonical approach of training agents to solve tasks by telling what to do via extrinsic rewards. By introducing terminal states, we provide the agent solely with the information of what not to do, allowing agents to develop the full behavioral repertoire compatible with the terminal states and the structure of the environment. We show that MOP can solve various tasks, including those typically addressed with extrinsic rewards, such as balancing a cartpole (see Appendix). $\textbf{Addition P6}$ : We emphasize that a MOP policy, by following a Boltzmann distribution (Eq. 4), assigns a non-zero probability to all actions, and therefore agents can inadvertently fall beyond the boundaries that define the terminal states. Therefore, MOP agents face the drawback of stochasticity, but they adapt their randomness via the computation of the value function. This is in contrast with R agents, where the stochasticity parameter $\epsilon$ is state independent. These results suggest that state adaptation of stochasticity is a relevant property we might expect in intelligent systems. $\textbf{Addition P7}$ $\textit{Line 285}$: “[…] uniform occupation as possible, avoiding saturation and encouraging neurons to live in a ‘healthy’ regime, i.e., a regime suitable for computation (Lazar et al. 2009).”. $\textbf{Addition P8}$ $\textit{Line 292-293}$: “[…] Remarkably, it has also been shown that, during courtship, adult zebrafinches significantly reduce their vocal variability compared to their solitary singing [55, 57]. This state dependent variability adaptation persists even after learning. This switch of behavior from random to more deterministic modes aligns with our hypothesis of the existence of directed variability in the brain.”. $\textbf{Addition P9}$ : This work highlights a significant limitation in classical neuroscience studies, where a fixed point of behavior is built into the task and episodes terminate upon reaching the goal. In ecological settings, agents continuously reallocate their goals and generate new behaviors. By allowing MOP to dynamically develop various behaviors, we offer a more flexible and realistic model of natural behavior. Pdf: /pdf/87f6b76d14c7267aab2a412a7a6a79d6a8785b01.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
User-Creator Feature Polarization in Recommender Systems with Dual Influence	Accept (poster)	Summary: This paper models dynamics of both users and creators in a recommender system. The user features shift in the direction of the content recommended to them. The creator dynamics are strategically motivated i.e. they try to align content to attract their audience, to increase profit. The authors then provide sufficient conditions for this model of dual dynamics to converge to polarization under a natural assumption that each creator has some non zero probability of being recommended to every user. The paper then discusses four real world recommendation designs, and whether they cause polarization or multiple clusters etc. They also provide results on synthetic and Movielens data complementing theory results and show that certain recommender designs do lead to polarization vs diverse clusters. Strengths: - This paper is the first to consider dynamics of both users and creators in a recommender systems and provides sufficient analytic conditions for polarization - They apply this theory to 4 natural designs: (1) Top-k, (2) Truncation, (3)Diversity boosting and (4) Lower bounding probability. They show that rules (3, 4) lead to polarization and rule (1) leads to diverse clusters. This section is particularly insightful. - The experimental evaluation with synthetic and Movielens data is also insightful and complements the theory. The softmax probability leads to diminishing creator and recommendation diversity over time. They also study top-k probability and show how lower k is better for higher creator diversity, recommendation relevance. Weaknesses: A criticism I had while reading the paper are gaps in literature for the discussion on dynamics in recommender systems. In addition to [Eilat & Rosenfeld] referenced in the introduction, [1,2,3,4,5,6] consider creator dynamics in recommender systems. These works assume static user features and provide results on content at equilibrium and user welfare. In the context of these works, it would be beneficial to highlight how your work is the first to consider both creator and user dynamics. [1] A Game-Theoretic Approach to Recommendation Systems with Strategic Content Providers (Ben-Porat and Tennenholtz) [2] Supply-side equilibria in recommender systems (Jagadeesan et al) [3] How Bad is Top-k Recommendation under Competing Content Creators? (Yao et al) [4] Modeling content creator incentives on algorithm-curated platforms (Hron et al) [5] Producers Equilibria and Dynamics in Engagement-Driven Recommender Systems (Acharya et al) [6] User Welfare Optimization in Recommender Systems with Competing Content Creators (Yao et al) Technical Quality: 4 Clarity: 2 Questions for Authors: - I understand the motivation for the form of user update in equation (3) . In each recommendation step an item is recommended and the user preference shifts in that direction, this is like in [Dean and Morgenstern]. Can you motivate the update in Eq (4), is this myopically optimal for the creator to do, and how does it generalize [Eilat & Rosenfeld]? - Minor Typo? For Figure 6, larger $\rho$ seems to lead to higher creator diversity (green curve). Confidence: 5 Soundness: 4 Presentation: 2 Contribution: 4 Limitations: The authors discuss limitations of their results in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Q1: Can you motivate the update in Eq (4)? Is this myopically optimal for the creator to do, and how does it generalize [Eilat & Rosenfeld]? [[Eilat & Rosenfeld]](https://arxiv.org/pdf/2302.04336) assumes that creators aim to maximize exposure (defined as the sum of inner products between user embeddings and the creator embedding) minus the content adjustment cost (see their Section 2.1 and equation (6)). They show that such creators will move towards the average of the user embeddings by some step size (their equation (11)). In our notations, their update is $v_i^{t+1} = \mathcal P(v_i^t + \frac{\eta_c}{\|J_i^t\|} \sum_{j\in J_i^t} u_j^t)$, which is the special case of our equation (4) where $g$ is the constant positive function $g(u_j, v_i) = 1$. Nevertheless, as we wrote in Lines 122 - 129, we are motivated by a different type of creators: rating-maximizing creators. This means that the $g(u_j, v_i)$ function should have the same sign as $\langle u_j, v_i\rangle$. Intuitively, for all the users $J_i^t$ being recommended creator $i$, if the user likes the item, $\langle u_j^t, v_i^t\rangle > 0 \implies g(u_j^t, v_i^t) > 0$, then the creator is incentivized to move towards that user to receive more positive rating from that user. Otherwise, $\langle u_j^t, v_i^t\rangle < 0 \implies g(u_j^t, v_i^t) < 0$, the creator is incentivized to move away from the user in order to be recommended less often to that user (to avoid being negatively-rated by the user). Taking both scenarios into account, the creator moves toward the weighted average $\sum_{j\in J_i^t} g(u_j^t, v_i^t) u_j^t$, which gives equation (4). Under this particular assumption on $g$, our equation (4) does not capture [[Eilat & Rosenfeld]](https://arxiv.org/pdf/2302.04336). > Q2: Minor Typo? For Figure 6, larger $\rho$ seems to lead to higher creator diversity (green curve). Yes, this is a minor typo. The figure shows that a higher $\rho$ leads to higher creator diversity but also higher tendency to polarization at the same time. A possible explanation for this phenomenon is (similar to our reasoning in Lines 307 to 309): the system polarizes into two balanced clusters which actually have a large average pairwise distance. So in this case, Tendency to Polarization may be a better measure for diversity loss than the Creator Diversity measure (average pairwise distance). --- Rebuttal Comment 1.1: Title: Acknowledging rebuttal Comment: Thanks for the addressing my questions, I stand with my assessment of accepting this paper.	Summary: The paper explores the dynamics between users and content creators in recommender systems, highlighting the dual influence where users’ preferences are shaped by recommendations and creators modify their content to align with what is more likely to be recommended. The study defines a model called user-creator feature dynamics to capture these interactions, demonstrating that such systems are prone to polarization, resulting in a loss of diversity. The paper then examines various approaches to mitigate polarization and improve diversity, finding that relevancy-optimizing methods, such as top-k recommendations, can prevent polarization more effectively than traditional diversity-promoting approaches. Strengths: The paper provides an interesting perspective by addressing the mutual influence between users and creators in recommender systems. The theoretical results and experimental validation using both synthetic and real-world data look credible. The writing is overall easy to follow. Weaknesses: 1. There are two lines of works focusing on modeling content creator dynamics and user preference evolving dynamics that are neglected by the authors. I listed several representative works and it would be great to include a comprehensive literature review regarding these works in the related work section. 2. One of your main observation (larger $\beta$ leads to higher creator diversity and alleviated polarization) is actually pointed out in [1] under a similar model, where content creators compete for a fixed user population (see section 3.2 in [1]). And another main observation in section 5.3 that smaller $k$ improves diversity does not echo the result in [2], which shows that larger $k$ improves the total creator utilities. It would be better to include some detailed discussions regarding these two works. 3. The user/creator preference updating dynamics need more justifications and empirical evidence. 4. The dynamical model makes some sense to me, but it would be more interesting to understand whether the observations still hold in the presence of noise. If the noisy version is hard to analyze theoretically, additional simulation results could also be valuable. [1]. Modeling Content Creator Incentives on Algorithm-Curated Platforms [2]. How Bad is Top-K Recommendation under Competing Content Creators? [3]. Online recommendations for agents with discounted adaptive preferences [4]. Recommender systems as dynamical systems: Interactions with viewers and creators [5]. Learning from a learning user for optimal recommendations [6]. Supply-side equilibria in recommender systems Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In theorem 3.3, how does the convergence rate depending on the temperature parameter $\beta$? I ask this because when $\beta\rightarrow +\infty$, the softmax recommendation strategy is equivalent to the top-1 recommendation strategy. In this case, Proposition 4.2 predicts that the top-1 recommendation lead to $n$ clusters rather than bi-polarization, which seems to contradict theorem 3.3. 2. I do not fully get why the specific forms of function $f$ and $g$ do not affect the analysis. Is it because your main results only depend on the range of $f$ and $g$? 3. In the experiments, the range of $\beta$ is quite conservative. I'm curious about the result under a larger range of $\beta$. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Weakness 1: related works. Thank you for listing those related works! We will discuss them in the revision. We also provide comparisons between some of those works and our work in a table in our global response. > Weakness 2: It would be better to include some detailed discussions regarding these two works [1] [2]. > > [1] Modeling Content Creator Incentives on Algorithm-Curated Platforms > > [2] How Bad is Top-K Recommendation under Competing Content Creators? [1] indeed also observes that a larger $\beta$ leads to higher creator diversity. We view our work as complementary to [1]. Our findings corroborate those of [1] despite two key differences between our setting and that of [1]: in [1], creators maximize exposure while we consider creators who maximize user engagement; second, [1]'s user population is fixed while ours is adaptive. This helps us understand that there may be some fundamental mitigation strategies (such as selecting $\beta$), which have similar effects regardless of changes to the underlying problem formulation. [2] shows that a larger $k$ improves the social welfare, defined to be the total utility/relevance of the users (which is also the total utility of creators in their model), assuming a fixed user population. In contrast, we focus on the diversity/polarization of creators, and we have adaptive users. So the $k$ plays a different role in our setting: a larger $k$ leads to worse creator diversity. > Weakness 3: the user/creator preference updating dynamics need more justifications and empirical evidence. Our user preference update model is a generalization of [[Dean & Morgenstern, EC 22]](https://arxiv.org/pdf/2205.13026), which models user update as $u^{t+1}_j = \mathcal{P}(u^t_j + \eta \langle v_i^t, u_j^t\rangle v_i^t)$. Our work replaces the inner produce with a general function $f(v_i^t, u_j^t)$ (our Equation 3), with some constraints outlined on lines 113 - 121 of our paper. The motivation for this update rule, as outlined by [[Dean & Morgenstern, EC 22]](https://arxiv.org/pdf/2205.13026), is the "biased assimilation" phenomenon, which is in turn inspired by the opinion polarization literature. (We mentioned this in Lines 107, 115, and 241.) Our creator update model, as we wrote in Lines 122 - 129, is motivated by rating-maximizing creators. This means that the $g(u_j, v_i)$ function has the same sign as $\langle u_j, v_i\rangle$. Intuitively, for all the users $J_i^t$ being recommended creator $i$, if the user likes the item, $\langle u_j^t, v_i^t\rangle > 0 \implies g(u_j^t, v_i^t) > 0$, then the creator is incentivized to move towards that user to receive more positive rating from that user. Otherwise, $\langle u_j^t, v_i^t\rangle < 0 \implies g(u_j^t, v_i^t) < 0$, the creator is incentivized to move away from the user in order to be recommended less often to that user (to avoid being negatively-rated by the user). Taking both scenarios into account, the creator moves toward the weighted average $\sum_{j\in J_i^t} g(u_j^t, v_i^t) u_j^t$, which gives our update rule (4). If any additional details would help improve the justification of the user/creator update rules, please let us know. This is an important aspect of our work and we would like to polish it as much as possible. > Weakness 4: it would be more interesting to understand whether the observations still hold in the presence of noise. In the newly uploaded PDF, we provide simulation results where the user and creator updates include normally distributed unbiased noises inside the projection operator: $u_j^{t+1} = P (u_j^t + \eta_u f( v_{i_j^t}^t, u_j^t) v_{i_j^t}^t + \eta_u \epsilon_j^t)$ where $\epsilon_j^t \sim Normal(0, \sigma^2I)$ and $v_i^{t+1} = P ( v_i^t + \frac{\eta_c}{\|J_i^t\|} \sum_{j \in J_i^t} g(u_j^t, v_i^t) u_j^t + \eta_c \epsilon_i^t)$ where $\epsilon_i^t \sim Normal(0, \sigma^2 I)$. We observe that a small noise still leads to near polarization, while a large noise (large $\sigma$) reduces the tendency to polarization. And the observation that top-k recommendation reduces polarization still holds. (see Figures R2 - R5 in the uploaded PDF.) > Q1: how does the convergence rate depend on the temperature parameter $\beta$? I ask this because when $\beta\rightarrow +\infty$, the softmax recommendation strategy is equivalent to the top-1 recommendation strategy, which by Proposition 4.2 leads to $n$ clusters rather than bi-polarization, which seems to contradict theorem 3.3. Finite $\beta$ and infinite $\beta$ are qualitatively different. Finite $\beta$ leads to bi-polarization, while $\beta = +\infty$ is equivalent to top-1 recommendation and does not lead to bi-polarization. Indeed, the rate of convergence to bi-polarization with a large but finite $\beta$ might be slow (see the "Tendency to Polarization" plot in Figure R1 in the PDF we uploaded during rebuttal). Our Theorem 3.3 is more of an asymptotic result. Analyzing the convergence rate would be an interesting yet challenging direction for future work. > Q2: I do not fully get why the specific forms of function do not affect the analysis. For $f$, our results and analysis only require the assumptions in Lines 114 - 121 that $f(v_i, u_j)$ has the same sign as $\langle v_i, u_j\rangle$ and is two-sided bounded $L_f \le \|f(v_i, u_j)\| \le 1$. For $g$, our current analysis assumes the specific form of $g(u_j, v_i) = sign(\langle u_j, v_i\rangle)$. We believe our analysis (in particular, Lemma E.1) can be generalized to other $g$ functions satisfying similar assumptions as $f$ (but as of now this generalization remains an open problem). > Q3: results under a large range of $\beta$. In Figure R1 of the uploaded PDF, we provide experiment results for $\beta \in [0, 10]$ and $+\infty$. We note that $\beta = 10$, although not very large, has similar effects as $\beta=+\infty$ (top-1 recommendation) in 1000 time steps, because the softmax probability of non-max creators when $\beta =10$ is very close to 0. --- Rebuttal Comment 1.1: Title: Re: Rebuttal Comment: I thank the authors for their detailed response, which addressed most of my concerns. And I appreciate the summarized table which emphasizes the contribution. I decide to maintain my score, leaning towards acceptance.	Summary: This paper studies how recommendations become polarized over the long run when user and creator features dynamically change over time. The authors theoretically prove that, under the assumption that every creator can be recommended to every user with some non-zero probability, recommender systems will eventually converge to polarization. They also simulate some real-world models, including top-k recommendation, truncation, diversity boosting, and lower-bounding probabilities in a long-term setting. The key observation is that top-k recommendation (i.e., only recommending top-k items to users) can reduce polarization to some extent, while existing diversity-boosting methods will worsen polarization when user/creator features dynamically change over time in the system. Strengths: 1. The authors provide both theoretical and empirical evidence showing that relevance-focused recommendations (as opposed to diversity-focused recommendations), which harm diversity in a static setting, are actually effective in improving diversity in the long term. This observation is somewhat counter-intuitive to previous beliefs, making it very interesting. 2. The authors conducted simulations with both synthetic data and real-world data (i.e., Movielens) using four diversity and relevance-related measures. Additionally, the analysis with sensitivity parameters in softmax is insightful and supports the authors' main claim. 3. Studying diversity in a dynamic setting is novel. Weaknesses: 1. Despite the novelty and interestingness, I have concerns about the key assumptions of the theoretical and empirical analyses. The assumption that all items can be recommended to users is not realistic. In practice, almost all recommender systems rely on top-k recommendations for either effectiveness or resource constraints like screen size. For example, on platforms like Netflix or Amazon, customers can only see a certain number of items on the webpage (i.e., p=0 for items that users can't see). Even if they can scroll down and the system continually recommends new items, they cannot physically see all items in the system. Thus, I believe the top-k setting is the most realistic and natural for real-world scenarios, and this seems like a hole in the authors' analyses. In this sense, the measures for empirical analysis should also only consider top-k items, not all items. 2. For the real-world designs, it would be more extensive if users included trustworthiness-aware recommender systems that consider dynamic/continual settings. For example, [1] consider performance difference between two different user groups when the user/item features are continually updated over time in the systems. 3. For the analysis with Movielens, considering the interaction timestamp in the simulation would more accurately reflect real-world scenarios, for example, for determining the true labels. [1] Yoo et al., Ensuring User-side Fairness in Dynamic Recommender Systems, WWW'24 Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Please address the points I raised in Weaknesses. 2. (Minor) Are both consensus and bi-polarization conceptually polarization? 3. (Minor) How are the initial user/creator embeddings initialized in the Movielens experiment? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Q1: Please address the points I raised in Weaknesses. > W1: The assumption that all items can be recommended to users is not realistic. ... Customers can only see a certain number of items on the webpage (i.e., p=0 for items that users can't see). First, we note that customers not seeing some items due to screen size limit does not mean that those items are recommended with probability p=0. The set of items seen by a customer is randomly drawn from a distribution over items which can have positive probability on all items, even if only $k$ items are shown. Second, we argue that non-zero probability of recommendation is realistic, even in large-scale real-world recommendation systems used by Yahoo [[Marlin et al, 2009]](https://dl.acm.org/doi/10.1145/1639714.1639717), [[Li et al, 2010]](https://dl.acm.org/doi/10.1145/1772690.1772758), and Kuaishou [[Gao et al, 2022]](https://arxiv.org/pdf/2208.08696.pdf). As noted in our Section 4.4, practical recommendation systems insert small random traffic (uniformly random recommendations, or missing-at-random MAR [[Yang et al, 2018]](https://dl.acm.org/doi/10.1145/3240323.3240355) data) to improve recommendation diversity [[Section 2.2 of Moller et al, 2018]](https://www.tandfonline.com/doi/full/10.1080/1369118X.2018.1444076) or to explore users' interests for unseen contents [[Gao et al, 2022]](https://arxiv.org/pdf/2208.08696.pdf) or for debias purposes [[Liu et al, 2023]](https://dl.acm.org/doi/10.1145/3582002). These interventions will cause all recommendation probabilities to be non-zero (although they may be very small). > ... In this sense, the measures for empirical analysis should also only consider top-k items, not all items. Note that our measures RD and RR include the recommendation probability $p_{ij}$ (where $p_{ij} = 0$ for non-top-k creators), so RD and RR do consider top-k items. Our other two measures, CD and TP, aim to measure the diversity of the entire creator pool, independent of the recommendation scheme, so they do not consider the recommendation probability for users. > W2: For the real-world designs, it would be more extensive if users included trustworthiness-aware recommender systems that consider dynamic/continual settings. For example, [1] consider performance difference between two different user groups when the user/item features are continually updated over time in the systems. We thank the reviewer for the suggestion and we agree that fairness-aware recommendation systems are important in practice. However, we want to make the best use of the limited space in the main article to introduce the notion of dual influence and outline its relationship with polarization in recommendation systems. Whether a system polarizes its users or creators is by itself an interesting trustworthiness question, separated from fairness considerations. Moreover, [1] does not study strategic content creators in recommendation systems and has fundamental differences with our work. Applying this work in case of adaptive users and creators will require a careful re-designing of the method proposed in [1]. We see the consideration of fairness-aware systems as a deeply important area, and hope that our work on dual influence will inspire future research to consider both fairness-aware recommendation and dual influence (as we fully agree with you that most realistic settings would consider both aspects). We are happy to add further discussions on fairness-aware recommendation systems and how our work relate to them in the appendix in the final version. > W3: For the analysis with Movielens, considering the interaction timestamp in the simulation would more accurately reflect real-world scenarios, for example, for determining the true labels. We agree with the reviewer that this will be an interesting simulation. However, we need to model the dual dynamics (that include both user updates and creator updates) in the system, such dynamics are not currently captured in any publicly available dataset that we are aware of. Since we need to model dual dynamics, solely relying on existing MovieLens data (or any other dataset) is not enough, we need to create pseudo-real-world data which simulates these dual dynamics. > Q2: (Minor) Are both consensus and bi-polarization conceptually polarization? Yes. > Q3: (Minor) How are the initial user/creator embeddings initialized in the Movielens experiment? The initialization is done by fitting a two-tower model [Huang et al.] on the existing MovieLens ratings data and we use the tower tops as the initial user and creator embeddings. ---- [Huang et al.] Learning Deep Structured Semantic Models for Web Search using Clickthrough Data. CIKM, 2013. --- Rebuttal 2: Comment: Thanks for the detailed response. However, I still feel that my question about top-k recommendation has not been adequately answered. Let me clarify the question further. 1. In the paper, lines 192-193 state, "In particular, we consider the top-k recommendation policy where each user is recommended only he k most relevant creators, so pt ij = 0 if i is not one of k creators i′ that maximize ⟨vt i′ , ut j ⟩.", meaning that p=0 in the case of top-K recommendation. My thought was that nearly all recommendation scenarios are essentially top-K recommendation practices, which means p=0 in these cases as well. Could you provide examples where this is not the case? As mentioned in line 188-189, involves filtering out items unlikely to be relevant to a user and then recommending from the remaining items. This filtering process is typically based on relevance scores, where only the highest scores are retained. In this sense, it seems equivalent to "only K items are shown." Therefore, I am unclear on why there would be a non-zero probability when only K items are shown. What is the difference between top-K recommendation in Section 4-1 and these real-world scenarios? 2. Regarding random interventions, are you suggesting or is it possible that these interventions occur on top of top-k recommendations (i.e., so if there are n interventions, there would be K+n items in the list)? I believe this scenario could indeed have a non-zero probability even within the top-k recommendation practice. But without this scenario, I think most recommendation scenarios are basically top-k recommendations in Section 4-1. What are your thoughts on this? --- Rebuttal Comment 2.1: Title: Clarifications on "top-k recommendation", "non-zero probability of not shown items", and "random intervention" Comment: We would like to provide some clarification regarding item delivery. When referring to "top-$k$”, we actually mean “top-$k$ truncation". For a given user $u_i$, we compute probability $p_{ij}$ for each creator $v_j$. Let $p^{(k)}$ be the $k^{\text{th}}$ largest $p_{ij}$, then all $p_{ij} < p^{(k)}$ are set to 0. Creators are then recommended to $u_i$ based on the remaining nonzero probability $p_{ij}$. This filtering process corresponds to the (first) recall stage commonly found in large-scale recommendation systems with two (or more) stages (see, e.g., [[Youtube's DNN recommendation paper]](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45530.pdf)). Top-$k$ truncation does not mean showing $k$ items to the users. Let's say that a user sees $L$ items in a short time frame. In real-world scenarios, the value of $L$ is usually smaller than $k$ and can depend on factors like screen size. Moreover, the set of $L$ items shown to the user is a random sample from a distribution of items. Reviewer EsTG wrote "(p=0 for items that users can't see)" and "Therefore, I am unclear on why there would be a non-zero probability when only K items are shown". That seems to be a misunderstanding. Even if an item $j$ is not shown to the user in one time step, its probability $p_{ij}$ can still be $>0$, and this item can be sampled and shown to the user in the next time step. The random traffic intervention (which has been used in large-scale recommendation systems as we discussed in Lines 225 - 227; see also [[KuaiRand, page 4, left column (ii)]](https://arxiv.org/pdf/2208.08696)) bypasses all the steps (scoring, filter, etc) in the multi-stage recommendation system and replaces the candidates returned by the multi-stage recommendation process with randomly chosen candidates at a low probability. The random traffic intervention does not cause efficiency loss for the system. And with random traffic, a user might be recommended any possible creator (namely, all $p_{ij} > 0$), even if $L$ is small. --- Rebuttal 3: Title: Correct, recommendations are based on probability distribution. Comment: Yes, we are recommending creators based on their probability distribution. Without top-k truncation, each creator $j$ (among the N total creators) is recommended with $p_{ij}$ probability. With top-k truncation, the N creators with the $k$ highest relevance scores pass the filter/truncation stage, and these $k$ creators are then recommended with probability proportional to $p_{ij}$ (calculated and normalized using these top $k$ relevance scores). We are not deterministically selecting the top-$L$ ($L < k$) creators to show to the user. --- Rebuttal 4: Comment: Thank you for the clarifications. Most of my concerns have been resolved, so I'll increase the credits. I hope the authors can include their rebuttal in the final version. --- Rebuttal Comment 4.1: Comment: Thank you! We will surely include our rebuttal in the final version.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for the helpful comments, especially the provided related works. Here, we provide a table to compare our work with those works (and some works that were already cited in our paper). We will add this table to an additional related work section in our appendix. We want to highlight that, while previous works have studied dynamic creators and dynamic users separately, "our work is the first to consider both creator and user dynamics", as pointed out by reviewer JScD. \| Works \| Adaptive Users \| Adaptive Creators \| Creator Reward \| Characterizing Dynamics or Equilibrium Behavior \| Creation Change Model \| \| ---------------- \| ---------\|----- \| -------------- \| --------------------- \| ------------------- \| \| Our Work \| Yes \| Yes \| User Engagement \| Dynamics \| Conditioned on previous time step; implicit cost of content adjustment\| \| Eilat & Rosenfeld [1] \| No \| Yes \| Exposure \| Dynamics \| Conditioned on previous time step; explicit cost of content adjustment \| \| Yao et al [2] \| No \| Yes \| User Engagement \| Dynamics \| Freely choose without cost \| \| Jagadeesan et al [3] \| No \| Yes \| Exposure \| Equilibrium \| Freely choose with cost \| \| Hron and Krauth et al [4] \| No \| Yes \| Exposure \| Equilibrium \| Freely choose without cost \| \| Ben-Porat and Tennenholtz [5] \| No \| Yes \| Exposure \| Equilibrium \| Freely choose without cost \| \| Acharya et al [6] \| No \| Yes \| User Engagement \| Equilibrium \| Freely choose without cost \| \| Yao et al [7] \| No \| Yes \| Reward designed by a social-welfare-maximizing platform \| Dynamics \| Freely choose without cost \| \| Dean and Morgenstern [8] \| Yes \| No* \| N/A \| Dynamics \| N/A \| \| Yao et al [9] \| Yes \| No* \| N/A \| Dynamics \| N/A \| \| Agarwal and Brown [10] \| Adaptive and adversarial \| No* \| N/A \| Dynamics \| N/A \| *: These works study the design of recommendation algorithms for the platform with a fixed set of items, without explicitly modeling the content creators. --- [1] Eilat and Nir Rosenfeld. Performative Recommendation: Diversifying Content via Strategic Incentives. ICML 2023. [2] Yao et al. How Bad is Top-K Recommendation under Competing Conten Creators? ICML 2023. [3] Jagadeesan et al. Supply-Side Equilibria in Recommender Systens. NeurIPS 2023. [4] Hron and Krauth et al. Modeling Content Creator Incentives on Algorithm-Curated Platforms. ICLR 2023. [5] Ben-Porat and Tennenholtz. A Game-Theoretic Approach to Recommendation Systems with Strategic Content Providers. NeurIPS 2018. [6] Acharya et al. Producers Equilibria and Dynamics in Engagement-Driven Recommender Systems. ArXiv 2024. [7] Yao et al. User Welfare Optimization with Competing Ccontent Creators. ArXiv 2024. [8] Dean and Morgenstern. Preference Dynamics Under Personalized Recommendations. EC, 2022. [9] Yao et al. Learning from a Learning User for Optimal Recommendations. ICML 2022. [10] Agarwal and Brown. Online recommendations for agents with discounted adaptive preferences. ALT 2024. Pdf: /pdf/2e22dcc35176d0defd5fcf8b49548123f8e261f4.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Is Mamba Compatible with Trajectory Optimization in Offline Reinforcement Learning?	Accept (poster)	Summary: This paper comprehensively investigates the possibility of leveraging Mamba for trajectory learning. The authors take Decision Mamba as a playground and analyse the performance of this model over trajectory learning scenarios (gym/mujoco) from several aspects. A group of conclusions are attained through rigorous experiments, which is solid and potentially valuable for further researches realted with Mamba. Strengths: 1. Novelty: given that Mamba is still at its exploratory stage, this paper positively probes Mamba's potential for tranjectory learning, with surprising results that with some specific pre-conditions, Mamba is more suited than Transformer. Weaknesses: 1. Most discoveries in this paper have been implicitly discussed for several months within the community, while it is firstly presented officially in this paper. Besides, these dicoveries lean to be emparical evidence, which is relatively shallow. This would make this paper's technical contribution weak. I would appreciate the authors if they could provide more in-depth explanation over these discoveries, in particular: 1) Transfomer-like model favors short sequence. 2) The significant role of the hidden attention. 2. Although the experimental results are solid, I found that this paper is more suitable for Benchmark Track, since the technical novelty revolves around benchmarking Decision Mamba. 3. Figure 1 (the title and the pic) should be improved. For now, it confuses me, especially the corresponding relationship between the text content (title) and the illustration. 3. Minor: line 295: may more suitable -> may be more suitable Technical Quality: 3 Clarity: 2 Questions for Authors: see Weaknesses. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors provide a brief limitation summary in Conclusion. I would appreciate the authors if they could refine this part since the "limitations" listed there do not seem like limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your insightful comments, which have catalyzed numerous enhancements and refinements to the paper. In the following, we reply to the questions one by one for the convenience of checking. --- Weakness 1: Most discoveries in this paper have been implicitly discussed for several months within the community, while it is firstly presented officially in this paper. Besides, these dicoveries lean to be emparical evidence, which is relatively shallow. This would make this paper's technical contribution weak. I would appreciate the authors if they could provide more in-depth explanation over these discoveries, in particular: 1. Transfomer-like model favors short sequence. 2. The significant role of the hidden attention. Response: Thanks for your comment. Past research on trajectory optimization usually approaches RL problems as sequence modeling problems in various ways. However, these studies are often constrained by issues related to large parameter counts and poor scalability, and there has been relatively little investigation into attention mechanisms. This study on DeMa contributes to the development of scalable and efficient decision models. Moreover, our empirical study lays a foundation for the further application of Mamba in RL. More explanations of our discoveries are as follows: 1. Most environments in trajectory optimization are Markovian. In these environments, a model inputs the current state and returns current action can also work(e.g. CQL), which means that excessively long historical information provides limited assistance for current decision-making. As shown in Figures B-C in the one-page PDF file, this result still holds in environments where the Markov property is relaxed. The results from the maze navigation task demonstrate that while DeMa's attention to past information increases, the hidden attention mechanism still primarily focuses on current information. This explains why the Transformer-like model favors short sequences. 2. The significant role of hidden attention lies in its ability to aggregate contextual information efficiently. Hidden attention is the third view of Mamba, which shows that such models can be viewed as attention-driven models[1]. It allows the model to selectively propagate or forget information along the sequence length dimension depending on the current token. This nature significantly reduces the number of parameters and aligns with the Markov property of the RL environment. As a result, DeMa can achieve better performance than DT with fewer parameters. [1] The hidden attention of mamba models. --- Weakness 2: Although the experimental results are solid, I found that this paper is more suitable for Benchmark Track, since the technical novelty revolves around benchmarking Decision Mamba. Response: Thanks for your comment. DeMa proposed in this paper can significantly address the challenges of large parameter counts and limited scalability in transformer-based trajectory optimization methods while maintaining good performance. Furthermore, our analysis of DeMa and the hidden attention mechanism provides valuable practical insights that apply to other sequence-based decision-making models. --- Weakness 3: Figure 1 (the title and the pic) should be improved. For now, it confuses me, especially the corresponding relationship between the text content (title) and the illustration. Response: Thanks for your comment, We have redrawn Figure 1 to improve clarity and readability. We provide it as Figure A in the one-page .pdf file. Overall, we have added numbering to each subplot and revised the descriptions in the title. We have arranged them side by side according to two structures: architecture and block. --- Weakness 4: Minor: line 295: may more suitable -> may be more suitable Response: Thanks for your comment, we have addressed the aforementioned errors. Additionally, we conducted a thorough review of the entire manuscript and corrected similar typos throughout. --- Limitations: The authors provide a brief limitation summary in Conclusion. I would appreciate the authors if they could refine this part since the "limitations" listed there do not seem like limitations. Response: Thanks for your comment, we have refined the limitations at the end of our revised manuscript. Limitations: We investigate the application of Mamba in trajectory optimization and present findings that provide valuable insights for the community. However, there remain several limitations: 1. Trajectory optimization tasks typically involve shorter input sequences, raising questions about how well the RNN-like DeMa performs in terms of memory capacity in RL compared to models such as RNNs and LSTMs. The potential of RNN-like DeMa warrants further exploration, particularly in POMDP environments and tasks that require long-term decision-making and memory. 2. Although we examine the importance of the hidden attention mechanism in Section 4.2, our exploration is still in its nascent stages. Future work could leverage interpretability tools to examine further the causal relationship between memory and current decisions in DeMa, ultimately contributing to the development of interpretable decision models. 3. While we have assessed the properties of DeMa and identified improvements in both performance efficiency and model compactness compared to DT, it remains unclear whether DeMa is suitable for multi-task RL and online RL environments. --- We hope that the above answers can address your concerns satisfactorily. We would be grateful if you could re-evaluate our work based on the above responses. We look forward to receiving your further feedback. --- Rebuttal Comment 1.1: Title: Comments on the authors' response Comment: All my concerns are addressed (at least a majority). I suggest the authors add the refinement into their paper properly. I will rise my score to 6. --- Reply to Comment 1.1.1: Title: Thanks very much for the reviews Comment: We are grateful for the endorsement of the reviewer 85kG. We will carefully follow your constructive comments and include the corresponding contents in the revision to improve our submission. Best, The Author of Submission8092	Summary: This paper investigates how Mamba perform in trajectory optimization in offline RL with ablation analysis on mamba's data input structures and architectural structures and shows Mamba DT can achieve SOTA performance with less parameters. Strengths: 1. The paper writing is good, the visualizations look good. 2. The input concatenation experiments provides useful practical insight also for other sequence-based decision-making models 3. The paper provides a detailed analysis of how various components of Mamba, such as the hidden attention mechanism and different residual structures, influence performance. Weaknesses: 1. Finding 3 is not very surprising on the tested MDP environment, since they by definition should focus only on recent states. It will be interesting to explore how this mechanism might perform in environments with long-term dependencies where the Markov property does not hold strictly. 2. Only tested on standard Atari and mujoco tasks. How would mamba perform on tasks that requires long horizon planning skills? such as maze navigation or tasks with delayed rewards. Technical Quality: 3 Clarity: 3 Questions for Authors: please see weakness. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: please see weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your insightful comments, which have catalyzed numerous enhancements and refinements to the paper. In the following, we reply to the questions one by one for the convenience of checking. --- Weakness 1: Finding 3 is not very surprising on the tested MDP environment. It will be interesting to explore how this mechanism might perform in environments with long-term dependencies where the Markov property does not hold strictly. Response: Thanks for your comments. We conduct additional explorations in environments involving maze navigation(maze2d, antmaze) and delayed rewards(MuJoCo with delayed rewards). The performance of the hidden attention mechanism is illustrated in Figures B-C in the one-page .pdf file, which shows that although DeMa’s attention to past information increases, the hidden attention mechanism still prioritizes the current information when the Markov property of the environment is relaxed. Furthermore, among historical information, the hidden attention mechanism demonstrates a significantly higher focus on states compared to rewards or actions. There remains a significant gap in the exploration of the analysis of Markovian property and attention mechanism. Most studies focus on the attention scores of outputs $y_j$ relative to inputs $x_i$ in the training phase. However, it does not effectively evaluate the context attention of the model during decision-making. Our analysis reveals that the hidden attention mechanism predominantly emphasizes the current token at each decision-making step, even when the Markov property of the environment is relaxed. Thus, our finding (Finding 3) offers valuable insights and guidance for the community. --- Weakness 2: How would mamba perform on tasks that requires long horizon planning skills? such as maze navigation or tasks with delayed rewards. Response: Thanks for your comment. We conduct three new experiments involving maze navigation(maze2d, antmaze) and delayed rewards(MuJoCo with delayed rewards). 1. maze2d: This environment aims at reaching goals with sparse rewards, which is suitable for assessing the model’s capability to efficiently integrate data and execute long-range planning. The objective of this domain is to guide an agent through a maze to reach a designated goal. 2. antmaze: This environment is similar to maze2d, while the agent becomes an ant with 8 degrees of freedom. In these environments, we adjust the hyper-parameters, which will be provided in the Appendix of our revised manuscript. Results in Table B show that DeMa performs better compared to DT in the maze navigation task. Table B: Results for maze2d and antmaze. We report the mean across 3 seeds. \|Dataset\|Env\|DT\|GDT\|DC\|DeMa(Ours)\| \|-\|-\|-\|-\|-\|-\| \|umaze\|maze2d\|31\|50.4\|36.3\|54.5\| \|medium\|maze2d\|8.2\|7.8\|2.1\|16.7\| \|large\|maze2d\|2.3\|0.7\|0.9\|6.6\| \|umaze\|antmaze\|59.2\|76\|85\|96.7\| \|umaze-diverse\|antmaze\|53\|69\|78.5\|96.7\| 3. MuJoCo with delayed rewards: This is a delayed return version of the D4RL benchmarks where the agent does not receive any rewards along the trajectory, and instead receives the cumulative reward of the trajectory in the final timestep. In this environment, we train DeMa using the same hyper-parameters settings, and the results are shown in Table C. Results show that CQL is the most affected, while DT also experiences a certain degree of influence. In contrast, DeMa and GDT are relatively less impacted. The results indicate that DeMa demonstrates effective performance in tasks with delayed rewards. Table C: Results for D4RL datasets with delayed (sparse) reward. The "Origin Average" in the table represents the normalized scores of evaluations across six datasets under the original dense reward setting. We report the mean across 3 seeds. The dataset names are abbreviated as follows: "medium" as "M", "medium-replay" as "M-R". \|Dataset\|Env\|CQL\|DT\|GDT\|DeMa(Ours)\| \|-\|-\|-\|-\|-\|-\| \|M-Delayed\|HalfCheetah\|1.0\|42.2\|43\|42.9\| \|M-Delayed\|Hopper\|23.3\|57.3\|58.2\|69.1\| \|M-Delayed\|Walker\|0.0\|69.9\|78.9\|77.6\| \|M-R-Delayed\|HalfCheetah\|7.8\|33.0\|41\|41.1\| \|M-R-Delayed\|Hopper\|7.7\|50.8\|79.8\|83.8\| \|M-R-Delayed\|Walker\|3.2\|51.6\|70.4\|71.7\| \|Average\|-\|7.2\|50.8\|61.9\|64.4\| \|Origin Average\|-\|65.5\|63.4\|63.8\|66\| Overall, DeMa achieves better performance than DT with fewer parameters in tasks that require long-horizon planning skills. --- We hope that the above answers can address your concerns satisfactorily. We would be grateful if you could re-evaluate our work based on the above responses. We look forward to receiving your further feedback. --- Rebuttal 2: Title: Looking forward to your responses or further suggestions/comments! Comment: Dear Reviewer qC8w, We have carefully considered and addressed your initial concerns regarding our paper. We are happy to discuss them with you in the openreview system if you feel that there still are some concerns/questions. We also welcome new suggestions/comments from you! Best Regards, The authors of Submission 8092 --- Rebuttal Comment 2.1: Comment: Thank you for conducting new experiments on new environments to address my concerns. I have decided to increase score to 6 --- Reply to Comment 2.1.1: Title: Thank you very much for the reviews Comment: Thanks for your recognition of our work and feedback on our response. We will carefully follow your constructive comments and include the corresponding contents in the revision to improve our submission. Best, The Author of Submission8092	Summary: The work introduces Decision Mamba (DeMa) to address the challenges in offline RL posed by the large parameter size and limited scalability of Transformer-based methods. DeMa aims to achieve similar performance to Transformers with significantly fewer parameters. DeMa surpasses the DT with significantly fewer parameters in the benchmarks. Strengths: 1. Extensive evaluations demonstrate the effectiveness of DeMa, highlighting its superior performance and efficiency compared to existing methods. 2. DeMa provides a novel solution to the parameter size and scalability issues in trajectory optimization. Weaknesses: 1. Some symbols are not defined before use. 2. This paper seems to have little relation to RL and appears more like a method applicable to all trajectory optimization. 3. There is too little discussion on the relationship to RL in sections 3.2 and 3.3. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. What is the definition of $L_{MSE/CE}$ and $_{-K:t}$? 2. Is DeMa applicable to all trajectory optimization methods? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The suggestions have been claimed in "Weaknesses" and "Questions". Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your insightful comments, which have catalyzed numerous enhancements and refinements to the paper. In the following, we reply to the questions one by one for the convenience of checking. --- Weakness 1 & Question 1: Some symbols are not defined before use. Response: Thanks for your comment. We have thoroughly reviewed the manuscript and provided definitions for all symbols. Additionally, we have removed some redundant symbols to enhance the article's readability. Major modifications include: 1. $L_{MSE/CE}$ in line 123. It represents the loss function of DeMa, when the output corresponds to continuous actions (as in MuJoCo), the loss function is Mean Squared Error (MSE), while when the output pertains to discrete actions, the loss function is the cross-entropy loss." 2. $s_{-K:t}$ in line 124. Subscript denotes a range from time step t-K+1 to time step t. We have corrected it to $_{t-K+1:t}$ to accurately represent the intended range from time step $t-K+1$ to time step $t$. 3. $A,B,C,D$ in line 133. $A\\in\\mathbb{R}^{N\\times N} , B\\in\\mathbb{R}^{N\\times 1}, C\\in\\mathbb{R}^{1\\times N}, D\\in\\mathbb{R}$ are parameter matrices in State Space Model. 4. $\\bar{K}$ and $\\bar{C}$ in line 136-137. We have removed the redundant symbol L and eliminated the symbol $\\bar{C}$, which was used incorrectly. Now it becomes $$ y_{i}=C\\bar{A}^i\\bar{B}u_1+C\\bar{A}^{i-1}{\\bar{B}}u_2+\\cdots+C\\bar{A}\\bar{B}u_{i-1}+C\\bar{B}u_i,\\quad y=u\\bar{K},\\\\ \\bar{K}=(C\\bar{B},C\\bar{A}\\bar{B},\\ldots,C\\bar{A}^{k}\\bar{B},\\ldots). $$ 5. Line 158-159. To provide a clearer formulation of hidden attention, we have revised it as follows $$ y_i=C_i\sum_{j=1}\^i \\big(\\Pi_{k=j+1}\^i \\bar{A}\_k \\big) \\bar{B}\_j x_j,\\\\ h_i=\sum_{j=1}\^i \\big(\\Pi_{k=j+1}\^i \\bar{A}\_k \\big) \\bar{B}\_j x_j, $$ where $B_i=S_B (\\hat{x}\_i)$, $C_i=S_C(\\hat{x}\_i)$, $\\Delta_i=\\text{softplus}(S_{\\Delta}(\\hat{x}\_i))$ with $S_B$, $S_C$ and $S_{\\Delta}$ are linear projection layers, and SoftPlus is an elementwise function that is a smooth approximation of ReLU. $\\bar{A},\\bar{B}$ is the discretization of $A, B$, that is, $\\bar{A}\_{i}=\\exp(\\Delta\_{i}(A))$ and $\\bar{B}\_{i}=\\Delta_{i}(B_i)$. --- Weakness 2: This paper seems to have little relation to RL and appears more like a method applicable to all trajectory optimization.* Response: Thank you for your comments. 1.Trajectory optimization methods treat RL problems as sequence modeling problems to get better performance and generalization[1]. Decision Mamba (DeMa) proposed in this paper is designed to address the challenges posed by transformer-based trajectory optimization methods, particularly the issues of large parameter sizes and limited scalability, which are long-standing issues that have been largely overlooked. Exploring the potential of DeMa in RL tasks will contribute to the development of scalable and efficient decision models, which will significantly enhance the practical applications of RL. We revise Sections 1-2 to emphasize the significance of DeMa in the context of RL. 2.Our paper primarily focuses on the analysis of DeMa in RL to provide valuable insights for the community. Indeed, DeMa can be combined with other trajectory optimization methods to achieve even better performance. We conduct additional experiments, the results of which are summarized in the table below. By integrating DeMa with QT [2], we develop Q-DeMa, which achieves performance comparable to state-of-the-art (SOTA) models while utilizing less than one-seventh of the parameter size of QT. This finding underscores the significant potential of applying Mamba to RL and emphasizes the critical importance of the research presented in this paper. Table A: Results for D4RL datasets. The dataset names are abbreviated as follows: "medium" as "M", "medium-replay" as "M-R". We report the mean across 3 seeds. \|Dataset\|Env-Gym\|DT\|DeMa (Ours)\|QT\|Q-DeMa (Ours)\| \|-\|-\|-\|-\|-\|-\| \|M\|HalfCheetah\|42.6\|43\|51.4\|51.2\| \|M\|Hopper\|68.4\|74.5\|96.9\|88.1\| \|M\|Walker\|75.5\|76.6\|88.8\|89.1\| \|M-R\|HalfCheetah\|37.0\|40.7\|48.9\|48.6\| \|M-R\|Hopper\|85.6\|90.7\|102.0\|101.5\| \|M-R\|Walker\|71.2\|70.5\|98.5\|99.8\| \|Average\|-\|63.4\|66.0\|81.0\|79.7\| \|All params #\|-\|0.7M/2.6M\|0.2M/0.5M\|3.7M\|0.5M\| [1] On Transforming Reinforcement Learning with Transformers: The Development Trajectory [2] Q-value Regularized Transformer for Offline Reinforcement Learning. --- Weakness 3: There is too little discussion on the relationship to RL in Sections 3.2 and 3.3. Response: Thanks for your comments. We revise Sections 3.2 and 3.3 to enhance the discussion regarding their relation to RL. Specifically, Section 3.2 provides a concise introduction to the two types of Mamba, allowing for the utilization of both types of DeMa in RL. To expand our analysis, we introduce hidden attention in Section 3.3. This enables us to visualize the hidden attention matrices within DeMa, thereby gaining a deeper understanding of the model's internal behaviors. --- Question 2: Is DeMa applicable to all trajectory optimization methods? Response: Thanks for your comments. DeMa is applicable to most transformer-based trajectory optimization methods, as it is designed to address the challenges posed by transformers, particularly the issues of large parameter sizes and limited scalability. By integrating DeMa with QT [2], we develop Q-DeMa, which achieves performance comparable to state-of-the-art (SOTA) models while utilizing only one-seventh of the parameter size of QT. This finding underscores the significant potential of applying Mamba to RL and emphasizes the critical importance of the research presented in this paper. --- We hope that our answers can address your concerns satisfactorily and improve the clarity of our contribution. We would be grateful if you could re-evaluate our paper. We look forward to receiving your further feedback. --- Rebuttal Comment 1.1: Comment: Thank the authors for the clarification and additional experiments. I have decided to increase the score --- Reply to Comment 1.1.1: Title: Thank you very much for the reviews Comment: We really appreciate your further comment and your recognition of our responses. We will carefully follow your constructive comments and include the corresponding contents in the revision to improve our submission. Best, The Author of Submission8092 --- Rebuttal 2: Title: We anticipate your feedback! Comment: Dear Reviewer BojN, The authors really appreciate your time and effort in reviewing this submission, and eagerly await your response. We understand you might be quite busy. However, the discussion deadline is approaching. We have provided detailed responses to every one of your concerns/questions. Please help us to review our responses once again and kindly let us know whether they fully or partially address your concerns and if our explanations are in the right direction. Thanks for your attention. Best regards, The authors of submission 8092.	null	null	Rebuttal 1: Rebuttal: We want to thank all the reviewers for their thoughtful suggestions on our submission, and we appreciate that the reviewers have multiple positive opinions of our work, including: * novelty (BojN, 85kG) * good writing, good visualizations (qC8w) * the detailed analysis provides useful practical insight (BojN, qC8w) We provide a summary of our responses, and we will carefully revise our manuscript by adding suggested evaluations, providing more detailed explanations, and fixing the typos. Introduction, Related Works, and Preliminaries: * We strengthen the relationship between DeMa, trajectory optimization, and RL. (for Reviewer BojN) * We refine the notations for improved clarity. (for Reviewer BojN) * We enhance the discussion regarding their relation to RL. (for Reviewer BojN) Experiments: * We redraw Figure 1 to improve clarity and readability. (for Reviewer 85kG) * We explore the applicability of combining DeMa with another method. (for Reviewer BojN) * We analyze the hidden attention in tasks that require long horizon skills and tasks with delayed rewards. (for Reviewer qC8w) * We explore the potential of DeMa in tasks that require long horizon skills and tasks with delayed rewards. (for Reviewer qC8w) * We provide more explanations for our discoveries. (for Reviewer 85kG) Conclusion： * We provide more details about our limitations. (for Reviewer 85kG) We appreciate all the reviewers' time and effort again. All these comments and suggestions are very insightful and beneficial for us to improve the quality of this work. Pdf: /pdf/c2d58a543eb90c7978c5bf62ace59625e279a7cc.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Accelerating ERM for data-driven algorithm design using output-sensitive techniques	Accept (poster)	Summary: This paper addresses the problem of learning optimal parameters for data-driven algorithm design. A characteristic of the problem is that the dual loss function, which measures the performance of an algorithm as a function of parameters, is discontinuous. Nevertheless, the dual loss is typically piecewise structured (constant, linear, etc.) with linear boundaries. Thus, roughly speaking, the problem of finding optimal parameters reduces to exploring polytopic cells partitioned by boundary hyperplanes. The main contribution is a cell enumeration algorithm that runs in output-polynomial time. The algorithm can be seen as a breadth-first search on a cell adjacency graph, where the enumeration of neighbors is done in an output-sensitive manner based on Clarkson's algorithm. The resulting output-sensitive complexity can be significantly better than the worst-case, as demonstrated in Example 1. The authors then instantiate the ERM method based on the cell enumeration for linkage-based clustering and DP-based sequence alignment. The applications involve designing execution graphs, which originate from the execution tree of [BDL20]. Combining appropriate problem-specific execution graphs with the cell enumeration leads to the improved time complexity of ERM in several data-driven algorithm design problems, as in Table 1. Strengths: 1. The paper addresses the important problem of optimizing algorithm parameters in data-driven algorithm design. 2. The theoretical results given in Table 1 appear strong compared with previous ones. 3. The output-sensitive cell enumeration might be of independent interest in the context of computational geometry. Weaknesses: 1. The paper would have been more appealing if implementations of the proposed methods and experimental results were provided. 2. The paper is somewhat dense and it is not easy to follow the technical details. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. I would like to know more intuition about how AugmentedClarkson differs from the original Clarkson and why it is important. 2. While the paper focuses on linear boundaries, some studies on data-driven algorithm design consider parameter spaces partitioned by polynomials: https://proceedings.neurips.cc/paper_files/paper/2022/hash/db2cbf43a349bc866111e791b58c7bf4-Abstract-Conference.html https://proceedings.mlr.press/v178/bartlett22a.html https://proceedings.mlr.press/v206/sakaue23a.html Is there a possibility of applying similar enumeration ideas to such situations? Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and insightful comments. Re experiments, we note that prior empirical work already suggests usefulness of output-sensitive guarantees on typical instances, which we elaborate below. ``Experiments:`` Our work is motivated from prior empirical research (lines 91-104) which indicate that the output cell size on typical problems is empirically much smaller than the worst-case cell size. The missing component is the development of algorithms with theoretical runtime guarantees which are output-sensitive, as previously proposed algorithms scale with the worst-case cell size. We expect implementation of our (LP-based) algorithm to be relatively straightforward, and would like to remark that there are no reasonable baselines known as previous algorithms were intractable. ``Beyond linear boundaries:`` We agree that algebraic boundaries would be the next direction to look at based on our work. One simple way to apply our algorithms to the polynomial boundary case is to use linearization by projecting into a higher dimension corresponding to the monomials (this would work if the polynomials have a small constant degree e.g. Thm 3.1, Lem 4.1 in [1] or Lemma 2 in [2]). A more direct extension which tries to compute cells induced by polynomial boundaries in an output-sensitive way is an interesting direction for future work. ``AugmentedClarkson:`` Clarkson’s Algorithm computes the set of non-redundant hyperplanes in a linear system in an output-sensitive time complexity. Intuitively, our augmentation additionally keeps track of the neighboring cells corresponding to the non-redundant hyperplanes to facilitate the breadth-first search over the cell adjacency graph. This is important as directly applying Clarkson’s and searching for the nearest cell (in a fixed direction) across each bounding hyperplane can add to the runtime complexity. References [1] Balcan, Maria-Florina, Siddharth Prasad, Tuomas Sandholm, and Ellen Vitercik. "Structural analysis of branch-and-cut and the learnability of gomory mixed integer cuts." Advances in Neural Information Processing Systems 35 (2022): 33890-33903. [2] Balcan, Maria-Florina, Travis Dick, and Manuel Lang. "Learning to Link." In International Conference on Learning Representations 2020. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response. I appreciate the answers to my questions. I retain my score.	Summary: In data-driven algorithm design, we are given a collection of problem instances sampled from an unknown distribution, and a family of algorithms for solving the problem, typically parameterized by a real-valued multivariate parameter. The goal is to find a setting of parameters such that the performance of the algorithm they parameterize is close to optimal among the family of algorithms, in expectation over the distribution of instances. Most prior work is firstly focused on the generalization aspect, i.e., on showing that a small number of samples suffices (in the sample complexity sense) for finding approximately optimal parameters for the distribution (these are the ERM parameters for the given sample of instance), and thus the family of algorithms is "learnable". It then (sometimes) proceeds to develop an efficient algorithm for finding those ERM parameters, based on the structure used to prove the generalization bound. This paper focuses more systematically on the ERM efficiency aspect. To this end, it starts with a common theme in many prior works on data-driven algorithms, that had been abstracted out and formulated in a generalized form in Balcan et al. (STOC 2021): for any fixed problem instance, the function that maps a setting of parameters to utility (of invoking their associated algorithm on that instance) admits a simple piecewise structure. Say, there is a small number of "simple" boundary functions (say, linear thresholds) that induce a partition of the parameter space R^d such that the utility function restricted to each piece is "simple" (say, constant). This is helpful in bounding the VC dimension of the utility functions and thus proving generalization bounds, and also potentially for navigating the parameter space efficiently to find the ERM parameters. The novelty in this paper is an attempt to give a more systematic recipe for the second part (navigating the piecewise structure for efficient ERM), with two main advantages -- (1) creating a unified framework that takes care of some parts of the ERM procedure in a general way, thus restricting the portion that needs to be figured out per each problem individually, and (2) obtaining algorithms whose running time depends on the actual number of pieces in the given instances ("output-sensitive") rather than worst-case number of pieces. The per-problem part to figure out is a subroutine that, given a problem instance and parameter setting p, returns a list of candidate boundary functions for the piece that contains p, and this subroutine depends on the specific problem in question. The unified part of the framework uses this subroutine to search through the pieces in an "output-sensitive" running time. __Post-rebuttal__: I appreciate the authors' elaborations on the technical content, and the conceptual aspect of the paper. I have raised my score to support acceptance. Strengths: The strength of this paper is that the matter of efficient ERM in data-driven algorithms indeed merits its own systematic study rather than being left as an afterthought of the generalization bounds. Weaknesses: The main weakness is that the end result isn't very strong: the framework is restricted to linear boundary functions and (more disconcertingly) to a constant number of parameters, and for the most part does not yield improved running times in the worst case, but a different notion of efficiency (output sensitivity). It tends more to systematically organizing ideas that have appeared in the literature in one form or another and less to introducing new algorithmic insights or techniques. I also feel that the presentation and writing could be too opaque for a wide readership like that of NeurIPS. Technical Quality: 3 Clarity: 2 Questions for Authors: N/A Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We respectfully disagree with the reviewer that the main result is not strong. Since its inception (around 2016), the field of theoretical guarantees of data driven algorithm design has been focused on sample complexity results. The major direction that has been left open in this field is to also consider the computational efficiency of data-driven algorithms. Our paper is the first to consider this question at some level of generality. We note that one cannot hope for the same level of generality as for the sample complexity of data driven algorithm design since algorithmic consideration have always been much more problem specific even in the classic (non-data driven) scenarios, that is in classic learning theory: even in classic learning theory we have general sample complexity results for ERM, yet the computational complexity of ERM algorithms depend heavily on the class (and even for basic classes we still have major open questions). Our work contains many subtle aspects, for example: - We make a novel connection with computational geometry to give beyond worst-case improvements on runtime efficiency. It is a priori not clear which tool to use to advance the state-of-the-art here. - The execution tree approach for linkage clustering needs several technical lemmas for establishing soundness of the approach (e.g. lemmas I.5 to I.8 in Appendix I). - The execution DAG approach used in sequence alignment is a completely novel contribution of this paper, and is critical to achieve the desired output-sensitive runtime guarantee. ``Applications of our framework (linear boundaries, constant parameters):`` We show several distinct applications in our work where the linear boundary and constant parameter size is relevant for the data-driven algorithm design problem: linkage clustering - both metric learning and distance function learning; sequence alignment; two-part tariff pricing (App F); item-pricing with anonymous prices (App F). The linear boundary setting appears more frequently than one might initially expect, see e.g. [1] for several applications. Moreover, by a simple linearization argument, our approach yields ERM implementations for polynomial boundaries with small degrees e.g. e.g. Thm 3.1, Lem 4.1 in [2] or Lemma 2 in [3]. Finally, we remark that we initiate a detailed study of computationally efficient implementation of ERM for data-driven algorithm design and tackle the natural first case where we already show significant improvement over previous runtimes (Table 1) provided output size is small. ``the most part does not yield improved running times in the worst case, but a different notion of efficiency (output sensitivity)`` Output-sensitivity is a relevant notion for data-driven algorithm design. We summarize motivation from prior empirical work in lines 91-104, which indicates usefulness of output-sensitivity from a practical perspective. Note that the whole point of data-driven algorithm design is to provide beyond worst-case improvements in algorithm design [4], so it is not unusual to expect a beyond worst-case measure of efficiency (output-sensitivity and fixed parameter tractability). ``It tends more to systematically organizing ideas that have appeared in the literature in one form or another and less to introducing new algorithmic insights or techniques`` We remark that using a computational geometry based technique in the context of learning theory is in itself remarkable and has hardly ever appeared in prior literature. Moreover, a direct application of Clarkson's algorithm itself is not always sufficient. Our “execution DAG” technique proposed for the sequence alignment problem is a novel algorithmic idea not from any previous literature, and is crucial for tracking the partition of the parameter space for each DP sub-problem. In this case the number of alternative alignments (and therefore the number of relevant hyperplanes $t_{LRS}$) for any fixed optimal alignment is exponential, so a direct application of Clarkson’s algorithm would still be exponential runtime. References [1] Balcan, Maria-Florina, Tuomas Sandholm, and Ellen Vitercik. "A general theory of sample complexity for multi-item profit maximization." In Proceedings of the 2018 ACM Conference on Economics and Computation, pp. 173-174. 2018. [2] Balcan, Maria-Florina, Siddharth Prasad, Tuomas Sandholm, and Ellen Vitercik. "Structural analysis of branch-and-cut and the learnability of gomory mixed integer cuts." Advances in Neural Information Processing Systems 35 (2022): 33890-33903. [3] Balcan, Maria-Florina, Travis Dick, and Manuel Lang. "Learning to Link." In International Conference on Learning Representations 2020. [4] Maria-Florina Balcan. Data-Driven Algorithm Design. In Tim Roughgarden, editor, Beyond Worst Case Analysis of Algorithms. Cambridge University Press, 2020. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I have read it and it will be considered.	Summary: The paper explores computational aspects of implementing ERM in data-driven algorithm design. The paper contributes an efficient algorithm to ennumerate cells induced by a collection of hyperplanes. The paper then shows how to utilize this as a subprocedure to solve ERM problems for algorithm design, focusing on linkage-based clustering, sequence alignment, and two-part tariff pricing. Strengths: One of the main interesting things I find about this paper is that the runtime of the ERM implementations is instance dependent, and specifically depends on the number of sum dual class loss function pieces. The paper comments that their runtime bounds imply improvements over prior work in the worst-case R but also can be faster for "typical" R. The paper is well-written and easy to follow. The paper discusses relevant background and related work. Weaknesses: To what extent is the approach generalizable to other data-driven algorithm design problems? Is there a generic principle or a general characterization of the problems for which this approach can be utilized? Technical Quality: 3 Clarity: 3 Questions for Authors: See questions above. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and useful comments. ``Generalizability to other data-driven algorithm design problems:`` Our approach is applicable to a fairly large number of problems, for example the various mechanism design problems in [1] are (d,t)-delineable which is a special case of Definition 1 (We include a couple examples in Appendix F). Our current approach is useful whenever the loss as a function of the algorithmic parameter on a fixed problem instance can be shown to have a piecewise structure with linear transition boundaries. The execution tree/DAG technique enables output-sensitive runtime complexity when the piecewise structure can be viewed as a refinement of pieces induced in successive algorithmic steps (e.g. see Figure 2), each refinement involving linear decision boundaries in the parameter space, and we expect this to be useful beyond the linkage clustering and sequence alignment problems considered. There is also potential for extending our work beyond linear transition boundaries e.g. to algebraic boundaries which would be relevant for several problems of interest in data-driven algorithm design. [1] Balcan, Maria-Florina, Tuomas Sandholm, and Ellen Vitercik. "A general theory of sample complexity for multi-item profit maximization." In Proceedings of the 2018 ACM Conference on Economics and Computation, pp. 173-174. 2018. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: Thank you for addressing my question. I keep my score.	null	null	Rebuttal 1: Rebuttal: Our work is a first major step in making the growing field of data-driven algorithm design a practically feasible endeavor by addressing the frequently noted open question of computational complexity [1, 2]. Our proposed algorithms provide concrete and significant improvements in the running time of ERM from $n^{O(d)}$ (formally XP time) to $f(d).poly(n)$ effectively establish fixed parameter tractability (for parameter $d$, input size $n$, see Table 1) assuming output is polynomially large. Our work is well-motivated by prior empirical work on clustering and sequence alignment (lines 91-104) which show that typical output size is dramatically smaller than worst-case bounds, implying that our methods lead to dramatic speed-up over prior work on typical instances. We have shown four different applications of our results (linear transition boundaries, constant number of parameters) in three different domains (data science, computational biology, computational economics), which is rare for algorithmic results that use the specific structure of the problem. We further expect more direct and indirect applications of our results [3, 4]. Technically, our work involves several novel ideas: - Use of computational geometry in learning theory does not have precedent in the literature as far as we are aware. We identify and adapt the relevant techniques to address the crucial problem of implementing ERM in data-driven algorithm design. - We develop novel techniques (execution tree and DAG) which are critical in addition to comp. geo. tools, e.g. directly using Clarkson’s algorithm would still lead to exponential time in $n$ but our techniques circumvent this. - We present our results in a modular framework, reducing the ERM problem to a simpler problem that can be solved in a problem-specific way, making it easier to apply our results for future applications. References [1] Rishi Gupta and Tim Roughgarden. Data-driven algorithm design. Communications of the ACM, 63(6):87–94, 2020. [2] Avrim Blum, Chen Dan, and Saeed Seddighin. Learning complexity of simulated annealing. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 1540–1548. PMLR, 2021. [3] Balcan, Maria-Florina, Tuomas Sandholm, and Ellen Vitercik. "A general theory of sample complexity for multi-item profit maximization." In Proceedings of the 2018 ACM Conference on Economics and Computation, pp. 173-174. 2018. [4] Balcan, Maria-Florina, Siddharth Prasad, Tuomas Sandholm, and Ellen Vitercik. "Structural analysis of branch-and-cut and the learnability of gomory mixed integer cuts." Advances in Neural Information Processing Systems 35 (2022): 33890-33903.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Excluding the Irrelevant: Focusing Reinforcement Learning through Continuous Action Masking	Accept (poster)	Summary: Learn a state specific mask for actions. Rather than simply a state specific interval, extend the action mask to different convex set representations. Then, derive a policy gradient for each of these masking schemes. The masking schemes are ray masks, hypercube transform mask and distributional masks. Applies action masking to seeker and quadrotor tasks and shows that this action masking improves performance. Strengths: The proposed action masking covers a wide range of possible action mask definitions. The derived policy gradients are relatively straightforward given the definitions of the action boundaries. The derivations appear to be sound when applied empirically. Weaknesses: It is not clear how easy it is to recover the action masking criteria, especially under the more complex generator or distributional schemes, and it seems like this would be rare The experiments are not particularly convincing because they all follow similar control tasks, but it also seems like these are the only tasks for which the action mask could be easily defined. Technical Quality: 3 Clarity: 3 Questions for Authors: It is not clear why related work is not in a separate section, rather than subsection. There does not appear to be a special connection to the introduction. It isn't obvious that if G is square and non-singular, that this does not restrict the space of possible relevant action sets, since this would ensure that the hypercube space had an invertible, i.e. one to one, mapping between itself and the action distribution. It seems like many to one would be preferred if the space of the zonotope's hypercube was higher dimension than the action set. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: While noted in the limitations, deriving a method for identifying not only the policy gradient, but a policy from a learned value function is highly relevant for RL, and it is not clear how these restricted action spaces can be applied to Q value calculations, though it seems reasonable to assume it is possible. Some suggestions in the main work of how the distributional or generator action space restriction could be defined as a function of the dynamics could be relevant since it seems like these functions have to be hand-engineered, and it is not obvious how to do that in domains where the dynamics are less well defined. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your valuable feedback on our manuscript, and for highlighting the broad applicability of our proposed action masking approach. We would like to address your concerns and questions in the following. # Weaknesses ## Action masking criteria Thank you for sharing your assessment. There can be different notions of relevance, which has implications for required task knowledge. A relevant action set could encode a formal specification, guarantee the stability/safety of a dynamic system, or enforce system constraints involving interdependent actions (see summary rebuttal A2 for more details). For an explicit relevant action set, one can determine if it is convex. If so, our continuous action masking can be applied. If not, the relevant action set could be underapproximated with a convex set (see rebuttal for reviewer PubC, Q1 for more details). We will clarify the notion of relevance in the introduction and extend Sec. 4.3 with more details on application limitations originating from the assumptions. ## Unconvincing experiments We appreciate your suggestion to include a different type of task with a more intuitive relevant action set. We will extend our empirical evaluation with the Walker2D MuJoCo environment. Here, the Walker2D is interpreted as a battery-driven robot and the relevant action set is based on the total power constraint of the battery. Such a constraint encodes the interdependence of multiple action dimensions and cannot easily be enforced with vanilla RL. Please refer to the summary rebuttal for more details on the new experiment (see A3) and more examples for relevant action sets (see A2). # Questions ## Related work as separate section Thank you for this remark. We agree and will move the related work subsection to a dedicated section. ## Clarification on the effect of G being square and non-singular Thank you for this question. A generator matrix $G$ that is square and non-singular does restrict the possibilities for relevant action sets, e.g., in two dimensions only rectangles and parallelotopes can be represented but not hexagons. We state in the paper that for this special case of an invertible matrix $G$, the policy gradient for the generator mask does not change (see Proposition 5). You correctly note that it is likely that the generator dimensions $P$ are larger than the action dimensions $N$. Yet, there are cases where relevant action sets with $N=P$ are a valid choice. For example, if the relevant action sets are state-dependent interval sets as for the 3D Quadrotor or if there is a linear dependency between the action dimensions. The latter is the case for the 2D Quadrotor where the force at the left and right rotors should be similar to avoid flipping over. Thus, a parallelotope shape would be a valid choice. Note that existing continuous action masking on interval sets would only poorly cover parallelotope relevant action sets as visualized in the rebuttal PDF, Fig. 2. We will reflect this answer in the paper by clarifying the implications of different choices of $G$ in Sec. 3.2. # Limitations ## Action masking for value-based RL Thank you for your comment. To the best of our knowledge, there is no popular deep RL algorithm for Q-learning in continuous action spaces that is not an actor-critic. For discrete action spaces, previous action masking work has demonstrated its effectiveness [4,6,11,22] and continuous relevant action sets could be discretized to make these approaches applicable. In our limitations, we comment on TD3 and SAC, which both learn a policy and for which future work is required to determine the theoretical and empirical effects of action masking on the policy. Nevertheless, we see that our statement in Line 149 could be misleading: > Since the size of the output layer of the policy or Q-function network cannot change during the learning process We will change it to: > Since the size of the output layer of the policy network cannot change during the learning process ## Obtaining relevant action sets and relevance of action masking Thanks for your comment. We see that our paper lacked a more explicit specification of the notion of action relevance and a clarification of the practical relevance and benefits of the formulation as part of the policy. We will add this in the revised paper and would like to refer you to A1. and A2. of the summary rebuttal for a detailed answer. ## References: [4] Feng et al. 2023 [6] Fulton et al. 2018 [11] Huang et al. 2022 [22] Rudolf et al. 2022 --- Rebuttal Comment 1.1: Title: Response to Authors Comment: I appreciate the detailed response, especially the added experiments in walker. While the changes improve the paper, I think they do not significantly improve the score since the core questions remain. --- Reply to Comment 1.1.1: Title: Response to Reviewer Comment: Could you please reiterate what your remaining concerns are? From our perspective, your detailed review helped us to alleviate the two major weaknesses: > It is not clear how easy it is to recover the action masking criteria, especially under the more complex generator or distributional schemes, and it seems like this would be rare In our new experiments, we now provide an example with a realistic constraint (maximal battery power) that we can very easily represent as a zonotope in the action space. We are convinced that such easy-to-implement constraints are not uncommon in continuous action spaces. > The experiments are not particularly convincing because they all follow similar control tasks, but it also seems like these are the only tasks for which the action mask could be easily defined. We believe that the mujoco continuous control tasks Walker2D, Ant, etc. are the de-facto standard benchmark in continuous control. From our point of view these new experiments provide more than convincing arguments for the use of continuous action masking. We are happy to discuss these points and any other open concerns.	Summary: The paper addresses challenges in RL with continuous action spaces, typically defined as interval sets. These spaces often lead to inefficient exploration due to irrelevant actions. The authors propose three continuous action masking methods to focus learning on relevant actions based on current state, improving predictability and suitability for safety-critical applications. They analyze the implications on policy gradients and evaluate performance using PPO across three control tasks. Results show higher final rewards and faster convergence compared to baseline methods without action masking. Strengths: Originality - The paper presents a unique perspective on action spaces by utilizing the relevance of action utility in tasks to improve performance. Conventional methods are limited to discrete domains (tasks) so applying their methods to continuous environments was interesting to see. Significance - The proposed approach has practical implications, especially in complex environments where distinguishing between relevant and irrelevant actions is crucial. Regardless of the coverage of baseline, their methods significantly outperform it establishing the state of the art performance. Weaknesses: Reinforcement Learning with Policy Gradients (Section 2.1) - L84: "r →: S × A" appears incorrect. Continuous Action Masking (Section 3) - Assumption 1: Clarify the definition of action relevance. Ray Mask (Section 3.1) - L131: Need proof that g(a) is bijective. Generator Mask (Section 3.2) - Why is A(s) suddenly state-dependent? Provide motivation and further description. - In Proposition 2's proof, the matrix multiplication seems infeasible due to mismatched dimensions (C is N x 1 and Ga results in P x 1). Experiment (Section 4) - Justify the rationale behind the design choices for action relevance in each environment. - Compare the chosen action relevance approach to other relevant action settings. Results (Section 4.2) - Why compare to a standard PPO baseline and not to prior relevant works? - Include qualitative results to validate the proposed methods. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: n/a Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your valuable feedback on our manuscript. We are particularly grateful for the acknowledgment of our manuscript's originality and significance. In the following, we outline how we incorporated your feedback and clarify open questions. ## Improvements to the mathematical formulation Thanks for pointing out notation improvements. We reworked the preliminaries and methodology sections to improve notational precision. Specifically, we corrected the following points: - $r \rightarrow: \mathcal{S} \times \mathcal{A}$ to $r : \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ in [L84]. - in the ray mask approach: $g(a) = c + \frac{\lambda_{\mathcal{A}}(a)}{\lambda_{\mathcal{A}^r}(a)} (a - c)$ to $g(a) = c + \frac{\lambda_{\mathcal{A}^r}(a)}{\lambda_{\mathcal{A}}(a)} (a - c)$ in [Eq. 6, L126] - in the generator approach - $g : \mathcal{A} \rightarrow \mathcal{A}^r$ to $g : \mathcal{A}^l \rightarrow \mathcal{A}^r$ in [L163] - $G \in \mathbb{R}^{P \times N}$ to $G \in \mathbb{R}^{N \times P}$ in [L154]. - Now we have $G \in \mathbb{R}^{N \times P}$, so the dimensions in [Eq. 12, L164] fit correctly. ## Clarification of the relevant action space Thank you for highlighting that the notion of action relevance was not properly defined. We will added a formal definition of the relevant action set to the preliminaries: > We further assume that we can compute a state-dependent relevant action set $\mathcal{A}^r(s) \subseteq \mathcal{A}$, which potentially reduces the action space, based on task knowledge. To introduce our notion of relevant actions, we are adding to the first paragraph of the introduction > Irrelevant actions are actions that are either physically impossible, forbidden due to some formal specification, or evidently counterproductive for solving the task. Additionally, we will clarify that only the relevant action set $\mathcal{A}^r(s)$ is state-dependent, the action sets $\mathcal{A}$ and $\mathcal{A}^l$ are not. We would like to refer the reviewer to the summary rebuttal A.2 Computing relevant action sets for a more in-depth discussion on this topic. ## Proof of bijectivity of $g(a)$ in the ray mask Lemma: The mapping function $g: \mathcal{A} \rightarrow \mathcal{A}^r$ of the ray mask $g(a) = c + \frac{\lambda_{\mathcal{A}^r}(a)}{\lambda_{\mathcal{A}}(a)} (a - c)$ is bijective. Proof: We prove that $g(a)$ is bijective by showing that it is both injective and surjective. For any convex set $\mathcal{A}^r$ with center $c$, we can construct a ray from $c$ through any $a^r \in \mathcal{A}^r$ as $r_d(t) = c + d t$, where $d = \frac{a-c}{\Vert a-c \Vert_2}$, and $t \in \left[ 0, t_{max} \right]$ limits the ray to $\mathcal{A}$. Since $\mathcal{A}^r \subseteq \mathcal{A}$, this also holds $\forall a \in \mathcal{A}$. By construction, any two distinct rays only intersect in $c$. For all points on a given ray, the scaling factors $\lambda_\mathcal{A}$ and $\lambda_{\mathcal{A}^r}$ are constants, allowing us to rewrite $g(a)$ for all $a=r_d(t)$ as the linear function: $$ g \left( r_d(t) \right) = c + \frac{\lambda_{\mathcal{A}^r}}{\lambda_{\mathcal{A}}} \left( r_d(t) - c \right). $$ To show that $g(a)$ is injective, i.e. $\forall a_1, a_2 \in \mathcal{A}$, if $a_1 \neq a_2 \implies g(a_1) \neq g(a_2)$, consider the following two cases for $a_1 \neq a_2$. Case 1: If $a_1$ and $a_2$ are on the same ray, then, $a_1 = r_d(t_1)$ and $a_2 = r_d(t_2)$ with $t_1 \neq t_2$. Since $g \left( r_d(t) \right)$ is linear and thereby monotonic, it follows that $g \left( r_d(t_1) \right) \neq g \left( r_d(t_2) \right)$ and consequently $g(a_1) \neq g(a_2)$. Case 2: Otherwise, $a_1$ and $a_2$ are on different rays and $a_1 \neq a_2 \neq c$, so $g(a_1) \neq g(a_2)$ follows directly as the rays only intersect in $c$. Thus, $g(a)$ is injective. For showing $g(a)$ to be surjective, it is sufficient to show that $\forall a^r \in \mathcal{A}^r$, there exists an $a \in \mathcal{A}$, for which $g(a) = a^r$. Consider the ray $r_d(t)$ passing through $a^r$. We know that $g \left( r_d(0) \right) = g(c) = c$, and $g \left( r_d(t_{max}) \right)$ is the boundary point of $\mathcal{A}^r$ from $c$ in the direction $d$. Moreover, $a^r$ lies on the line segment between $c$ and $g \left( r_d(t_{\max}) \right)$. Since $g \left( r_d(t) \right)$ is linear and continuous, according to the intermediate value theorem, $\exists t^* \in \left[ 0, t_{\max} \right]$, for which $g \left( r_d(t^*) \right) = a^r$ for any $a^r \in \mathcal{A}^r$ along the ray $r_d(t)$. As we can construct such a ray through any point in $\mathcal{A}^r$, we have shown that for every $a^r \in \mathcal{A}^r$, there exists an $a \in \mathcal{A}$ such that $g(a) = a^r$, thus proving surjectivity. ## Additional experiments required Based on your recommendations, we added further evaluations as discussed in A3. of the summary rebuttal. Additionally, we will add an illustration of typical rollouts on the Seeker environment as interpretable qualitative result (see rebuttal PDF, Fig. 4). ## Rationale behind the design choices for action relevance in our environments Thank you for your question regarding the design choices of our relevant action space. A key advantage of our masking approaches is that they ensure that only relevant actions are executed and therefore can be used for safety-critical tasks. There are other notions of relevance, which we explore in the summary rebuttal A2. Here, we choose experiments where the relevant action set is a safe action set. In particular, a relevant action is one that avoids collision with an unsafe region. For the Seeker environment, the unsafe region is defined by an obstacle (see Fig. 2). For the quadrotor stabilization experiments, we compute a control-invariant set, which acts as relevant state set $\mathcal{S}^r$. We will clarify in 4.1.1 and 4.1.2 the motivation and computation of the relevant action set. --- Rebuttal Comment 1.1: Title: Response Comment: Thank you for your response and for proving the bijectivity as well as running a new experiment. Unfortunately, I can't seem to find the information about Fig. 4 so could you elaborate on Fig. 4? --- Rebuttal 2: Title: Supplementary figures Comment: Thanks for your question. Fig. 4 shows qualitative deployment behavior for the different RL agents in the Seeker environment. We use the same goal state (gray circle), same obstacle (red circle), and same ten initial states (colored crosses) for the agents for better comparability. The ten trajectories have the same color as their initial state. Note that for the masking approaches the trajectories are almost identical after passing the obstacle and therefore plotted on top of each other. The masking agents (ray, generator and distributional) reach the goal state for all ten initial states, with the generator being the quickest. In contrast, the PPO baseline agent reaches the goal with four out of the ten initial states and otherwise collides with the obstacle. The qualitative behavior is aligned with the quantitative results in Table 1 of the submitted paper meaning that the generator mask agent performs best and the PPO baseline performs worst with respect to reaching the goal efficiently. The goal of the visualization in Fig. 4 is to provide an intuition on the trained agents' behavior. We choose the Seeker environment since we believe it is the most intutive to understand and thus also to interpret the qualitative behavior. Note that there is unfortunately a typo in the subplot titles: distribution should be distributional. --- Rebuttal Comment 2.1: Title: Response Comment: I understand the setup of Fig. 4. However, as a qualitative result, it lacks relevance to action relevance (as you mentioned, this is merely an alternative to the quantitative results in the table). For instance, it’s not clear how this plot demonstrates the role of action relevance in decision-making. --- Reply to Comment 2.1.1: Title: Answer to Response Comment: Thank you for your reply. We are sorry that we misunderstood your review question for qualitative results. Still, we are not entirely sure how to interpret your comment. If we understand you correctly, you are interested in the impact of choosing different relevant action sets on the decision-making of the agent (i.e., the learning process)? Let us briefly provide some intuition on this. While the relevant action sets $\mathcal{A}^r$ in our experiments are all collision-avoidance action sets, their average coverage of $\mathcal{A}$ varies between 25%, 28%, and 70% for the 3D Quadrotor, 2D Quadrotor, and Seeker, respectively. It seems that there is a sweet spot for the restrictiveness of the relevant action set due to two opposing mechanisms. Sample efficiency gains from masking increase, the smaller the relevant action set in relation to the global action set, whereas the exploration capability of the agent decreases. In other words, if the relevant action is almost the entire action space, the sample efficiency gains are likely small compared to not using action masking. If the relevant action set gets too small, the agent cannot learn much due to the very limited amount of possible actions. For example, in our experiments, the reward gain in the Seeker environment is much higher than in the Quadrotor environments, which underscores less constrained exploration. In the extreme case of $\{ a^r \} = \mathcal{A}^r$ always, there is no meaning in using reinforcement learning since the agent cannot explore. Future work will need to investigate these mechanisms in more detail for a multitude of tasks and relevant action sets to provide clearer and more nuanced insights. Yet, we are happy to extend our discussion in Sec. 4.3 (Limitations) in the last paragraph with this intuition. If you are interested in a visual demonstration of the relevant action set during a training episode, we can refer you to Fig. 2 of our original paper. It shows the relevant action set at a certain state in the Seeker environment, which prohibits the agent from colliding with the red obstacle.	Summary: This paper discusses methods for action masking in continuous action spaces to improve convergence stability and sample efficiency in reinforcement learning. The paper introduces three methods for action masking with convex relevant action sets, proves their convergence, and experimentally verifies their effects. Strengths: This paper is excellently written, defines a clear and well-motivated goal, and describes three intuitive and theoretically grounded methods to achieve that goal within well-defined and clearly stated limitations. To my knowledge these approaches are novel (though the distributional mask I suspect has been used as a one-off solution in prior work as it is conceptually very simple), and their definition and analysis are nontrivial. Weaknesses: This paper is pretty solid overall, and I have few major complaints. The one significant issue I see is that I think the distributional mask algorithm is off-policy by nature, meaning it's use with on-policy methods like PPO is biased and will cause performance loss or divergence. This may explain the observed underperformance of this masking method in two of the three experimental tasks, and while the algorithm can clearly converge in some cases it seems like a major issue with that particular mask in the context of PPO (off-policy algorithms could use it without issue, but those are left to future work here) that should be noted, or the mask omitted from this paper and left to future off-policy methods. Beyond that, the experimental evaluation is relatively simple (though I think it is sufficient to validate these algorithms), and more challenging tasks would be useful to demonstrate the limitations of these masking methods. That said, the paper makes it clear that defining a suitable convex relevant action set is a manual process and can be challenging (this is okay as a limitation), so it is understandable why such stress tests are not performed. If there was a way to increase difficulty without major manual action set definition work it would strengthen the evaluation to include it. I have a few other minor issues and questions noted below, but overall this is a paper that is clear in its goals and describes methods that achieve them, validated to a reasonable standard of theoretical and experimental evidence. There's more that could be done on this topic, but the contribution of this paper is significant on its own, so I'm inclined to recommend acceptance (particularly if something is done to address my concern about the distributional mask above). Technical Quality: 3 Clarity: 4 Questions for Authors: -Is assumption 1 (relevant action set is convex) reasonable in most cases? I can imagine disjoint action subsets being relevant in many cases- for example, a self-driving car that needs to steer either left or right to avoid collision, but not go straight ahead or backwards. -I'm not sure it's actually necessary to compute the policy gradient across the action mask (with the exception of the distributional mask). Once an action is sampled from the policy, the mapping to the relevant action set can simply be treated as part of the environment/transition function which the policy can learn to manipulate without gradient flow. Does this simplify things or am I missing something? This would also permit arbitrary nonconvex relevant sets, I believe. -For the gradient of the distributional mask in proposition 4, isn't this affected by off-policyness due to the constrained sampling of actions from the policy distribution? For example, if most of the policy distribution probability mass lies outside the relevant set (e.g. in the event of transfer learning to a different task with a new relevant set) the actions sampled will not follow the policy distribution closely and thus \pi_{\theta}(a\|s) will not be an accurate probability of sampling action a at state s. As noted above, this seems like a big issue that should be noted or addressed, unless I'm missing something that corrects for the off-policyness. -Small quibble: The goal in figure 2 looks black to me on two different monitors, perhaps using a lighter gray would make it more distinct from the agent position? -It would probably be reasonable to move the environment definitions for the two quadrotor tasks to the appendix to save space in the main paper, FWIW. I'm not sure the abbreviated version that's present provides all that much context over the text descriptions of the tasks. -It's not critical to have, and I realize it's a difficult thing to derive a relevant set for by nature, but having an experiment on an environment with 10+ action dimensions would be a nice addition to demonstrate that these masking approach can scale to higher dimensional action spaces tractably. I'd also appreciate some comments on compute cost scaling with action dimension count in the limitations or conclusions sections, if possible, since it seems like compute cost is likely to increase with the dimensionality of the action space. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: This paper does an excellent job of making clear its limitations and scope. I don't see any potential negative societal impacts from this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments and for recognizing the novelty and theoretical grounding of our proposed methods. In the following, we respond to the weaknesses (W1 and W2) stated and questions (Q1 - Q6) raised. # Weaknesses ## W1. Distributional mask being off-policy by nature Thank you for sharing your reading of our distributional mask approach. Could you reiterate on why you would refer to the distributional mask approach as off-policy? We would refer to it as on-policy as we are directly sampling from the relevant policy $\pi_\theta^r(a^r\|s)$, which is normalized through the integral $\int_{\mathcal{A}^r} \pi_\theta(a \| s) \mathrm{d} a$, and derive the corresponding gradient $\nabla_\theta \log \pi_\theta^r(a^r \| s)$. However, our implementation approximates the gradient by treating the integral as a constant, which introduces some off-policyness. However, such off-policyness is also introduced by most stochastic on-policy RL implementations (including PPO) through the clipping of samples from the normal distribution. ## W2. Simplicity of experiments Thank you for your assessment. The main goal of our experiments is to provide initial empirical evidence for continuous action masking and compare the three masking approaches. To strengthen our evaluation, we will add a Walker2D Mujoco environment and an additional baseline to the revised paper as described in A3. of the summary rebuttal and 3. in rebuttal for reviewer 8567. # Questions ## Q1. Convex set assumption Thanks for your question. Note that convex sets are an important generalization of current practice that allows for interval sets only. We believe that assuming relevant action sets are convex is reasonable since exclusion of extreme actions and dependencies between action dimensions can be described by them. Additionally, one can underapproximate a non-convex relevant set with a convex relevant set at the cost of excluding some relevant actions [https://arxiv.org/abs/1909.01778]. For disjoint convex sets, our continuous action masking could be extended with a discrete action that represents a decision variable switching between the disjoint subsets. To find an optimal policy, hybrid RL algorithms could be used [https://arxiv.org/abs/2001.00449]. Yet, this major extension of continuous action masking is subject to future work, and we will add these considerations in Sec. 4.3. ## Q2. Mapping as part of the environment Thank you for your question. Formulating action masking as part of the policy leads to more intuitive action spaces, potential integration of the masking map in the gradient, and more possible masking approaches (see summary rebuttal, A1). Note that implementing the ray mask on the policy or environment side is mathematically identical due to the unchanged gradient. To assess the empirical relevance of our formulation, we added another baseline where as part of the environment dynamics irrelevant actions are replaced with relevant ones. Please refer to 3. in rebuttal for reviewer 8567 for more details. ## Q3. Gradient of the distributional mask You are correct in your assessment that $\pi_\theta(a \| s)$ is not an accurate probability of sampling action $a \in \mathcal{A}^r$ at state $s$. To address this issue, we normalize the constraint distribution with the integral in equation (15) and use the probability of the resulting distribution $\pi_\theta^r(a \| s)$ for sampling and in the objective function. With this formulation, the entire probability mass of the resulting distribution $\pi_\theta^r(a \| s)$ lies within $\mathcal{A}^r$, and the distribution accurately reflects the actual likelihood of sampling action $a^r$. ## Q4. Goal color Thanks for the hint. We will change the color in the revised paper. ## Q5. Dynamics to the appendix Thanks for the suggestion. We will extend Sec. 4 with the additional experiments as described in our summary rebuttal and move the details of the dynamics to the appendix. ## Q6. Experiment with higher action dimensions Thank you for your suggestion to extend our evaluation. To provide empirical insights for higher dimensional action spaces, we will add the Walker2D Mujoco environment as described in the summary rebuttal (see A3). Although the Walker2D has less than ten action dimensions, it has the highest action dimensionality of the four environments with six dimensions. The compute costs are mainly affected by the calculation of the relevant action set in each state and by the additional operations for the mapping function $\mathcal{F}$ of the specific masking approach. The former depends strongly on the notion of relevance. For our experiments with collision avoidance as notion of relevance, we compute the relevant action sets at each state with an exponential cone program, which scales polynomially with system dimensions (i.e., state and action) assuming a suitable interior-point solver [Boyd, 04, Convex Optimization]. The second aspect important for the compute costs is the specific masking approach. For the ray mask, the cost is dominated by computing the boundary points, which is a linear program for zonotopes with polynomial complexity in the state dimension [Kulmburg, 21, On the co-NP-Completeness of the Zonotope Containment Problem]. For the generator mask, the matrix multiplication of $G$ with $a^l$ is the dominating operation. For the distributional mask, sampling with the random-direction hit-and-run algorithm introduces computational cost. With zonotopes, each step involves solving two linear programs to determine the boundary points. As mentioned in line 190 the mixing time, or the number of steps before accepting a sample, is set to $N^3$, where $N$ is the action dimension. We will clarify the computational cost scaling in Sec. 4.3 and extend our runtime table in Appendix A.6 with the Walker2D environment, for which the PPO baseline and the generator mask training run at the same speed. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: Thanks for your detailed comments. -Regarding the off-policyness of the distributional mask, consider the case where the policy is a unit normal distribution and the mask is a truncated unit normal (which I believe is consistent with section 3.3). Actions sampled from this truncated distribution will have a different probability than they would under the original unit normal (e.g. because some actions are excluded, those that are included have a higher probability than otherwise). The exact gradient from Proposition 4 cannot be computed (as noted in lines 198-200), so this algorithm is fundamentally off-policy since it's impossible for it to be on-policy using the truncated distributions discussed in section 3.3 (maybe there's other distributions that would avoid this issue but those aren't discussed and seem like a non-trivial extension to define). Further, I disagree with the assumption that the difference between the truncated and original distributions will be trivial in practice, both because it can be hard to estimate where policy gradients will push the distribution in complex tasks with limited samples per gradient step and because just how truncated the distribution is depends greatly on the task- some tasks could require very aggressive truncation to limit to only the relevant set, while others might be more permissive. The other masks seem fine to me, but this one seems flawed in a way that is hard to solve and will matter in practice as well as theory. I'm not sure what "the clipping of samples from the normal distribution" refers to for PPO/etc. In every PPO/A2C/etc implementation I've worked with the full distribution gets sampled. Can you clarify? Given this, I'd strongly recommend removing the distributional mask method from the paper, or at least making it's off-policyness clear. I'm certainly willing to accept experimental evidence that it works ok in practice, but the experiments presented here don't seem all that reassuring and don't motivate why a theoretically flawed algorithm might be superior in practice (it doesn't outperform the other masking algorithms in the best case and seems to underperform sometimes). -Regarding handling non-convex relevant action sets, I'm sorry but I find the assertion that a convex subset of the full non-convex relevant set is good enough to be unconvincing. When looking at robotics (one of, if not the, biggest application domain for continuous-action RL) one finds non-convex (particularly disjoint- go left or right, but not straight ahead, etc) relevant sets all over the place. I'd need experimental evidence to believe this isn't an issue in practice, but I'm willing to accept this as a topic for future work (not every limitation of an algorithm needs to be solved in one pass). It does seem worth noting as a limitation (sounds like this will be added) and is a mark in favor of environment-masks over policy-masking that it supports disjoint sets by default, however. Considering the above and the discussion/comments of other reviewers, I think I shall maintain my score. I do think this paper could be made stronger, but it's also reasonable to leave these questions to future work. --- Reply to Comment 1.1.1: Title: Answer to Response Comment: > The exact gradient from Proposition 4 cannot be computed (as noted in lines 198-200), so this algorithm is fundamentally off-policy since it's impossible for it to be on-policy using the truncated distributions discussed in section 3.3. You are correct by stating that if one cannot compute the gradient, the distributional mask becomes off-policy, since we update the parameters for a different distribution. However the fact that the integral is intractable does not imply that it cannot be approximated or derived in a different way (granted, our wording of "Since we cannot compute the gradient..." is suboptimal). However, this is certainly not trivial and likely warrants considerable future work. > Further, I disagree with the assumption that the difference between the truncated and original distributions will be trivial ... while others might be more permissive. Saying that the difference is trivial in practice is certainly an understatement, but we believe that we did not make this statement anywhere. We agree that estimating the practical implications is challenging. Our suggestion that approximating the gradient might be valid in practice stems from the theoretical analysis presented in the "Physical implications of ... (2/2)" section of the general discussion. By approximating the gradient, we remove the option 2. for the gradient to increase the log-likelihood of the sample. Since this option can only favor actions close to the boundary of the relevant action set, it might not be applicable often in practice. > In every PPO/A2C/etc implementation I've worked with the full distribution gets sampled. Can you clarify? We acknowledge that our previous answer did not clearly state this point. In nearly all continuous reinforcement learning tasks, sampled actions are clipped to specific ranges (typically $[-1, 1]$) due to a fixed interval action space usually based on physical constraints. This clipping introduces a bias when using a Gaussian distribution as the policy, as its infinite support creates boundary effects. Chou et al. (2017) provide a thorough analysis of this issue in their paper "Improving Stochastic Policy ... " (Section 3.2). Essentially the clipping creates a form of truncated distribution with spikes at the truncation limits. One could argue that neglecting this effect inherently makes most stochastic policy gradient methods off-policy (note that we don't claim this is necessary the case!). PPO does not account for this effect, yet it demonstrates strong performance in practice. Based on this, we infer that our gradient approximation might also have little impact on performance. > Given this, I'd strongly recommend removing the distributional mask method from the paper, or at least making it's off-policyness clear. We believe it's important to include the distributional mask in the paper despite its current limitations. This method employs a novel technique for sampling from the policy distribution, which is rarely used in RL and thus represents a contribution. By publishing this work, even with its current need for gradient approximation, we open the door for other researchers to build upon and potentially solve the existing problems. We recognize that for this justification it is crucial to explicitly state the current limitations of the method, including its off-policy nature. > Regarding handling non-convex relevant action sets, I'm sorry but I find the assertion that a convex subset of the full non-convex relevant set is good enough to be unconvincing. We appreciate your concern about non-convex relevant action sets, particularly in robotics applications. We want to clarify that we didn't intend to suggest that convex sets are universally "good enough". Rather, they represent a significant improvement over the current practice of using only intervals as sets. We fully acknowledge that for many applications, especially in robotics, restricting actions to convex sets can be an oversimplification, as Your example illustrates. Our current work with convex sets is a step towards more flexible action spaces, but we agree that handling non-convex sets would be a valuable extension of our methods. However, this presents significant challenges that we believe warrant separate, focused research. > It does seem worth noting as a limitation (sounds like this will be added) and is a mark in favor of environment-masks over policy-masking that it supports disjoint sets by default, however. We don't really understand why you are certain that environment-masks support convex sets by default. None of our methods can be applied to non-convex sets as is, even if they are implemented in the environment. Can you please elaborate how this would be done? We can imagine that action replacement is implemented with non-convex sets. Yet, this is conceptually different to masking and the newly added action replacement baseline performs worse than masking (see rebuttal for reviewer 8567, 3.).	Summary: This paper proposes mathematical formulations for continuous action masking in reinforcement learning, to incorporate domain-knowledge in the form of state-specific sets of relevant actions. It introduces 3 functional forms to extract relevant actions from the original action space, and consider its effect on the policy gradient. The policy gradient does not change much, and the paper shows that the forms compare similarly and better than learning an agent without any knowledge of the continuous action mask at all. Strengths: - The problem of action masking in continuous action space is an underexplored one, but could have major impact in efficacy of agents and incorporating domain-specific knowledge. - The proposed continuous action masking could potentially be useful for safety demarcations. - The paper provides mathematical frameworks to formulate continuous action masking and also derive their (minimal) effect on the policy gradient. - The paper is mostly well-written and explains the mathematical derivations quite well. Quick note that Section 2.1 and 2.2 could be made more integrated, currently they seem completely disconnected. Weaknesses: ## 1. General applicability of this paper's ideas Obtaining a state-specific relevant action set can be really hard. The paper, however, makes contradictory statements about this: - L5: "little task knowledge can be sufficient to identify significantly smaller state-specific sets of relevant actions." - L284-285: "assume that an appropriate relevant action set can be obtained. Yet, obtaining this can be a major challenge in practice." From the experiments on the 3 environments, it already seems like defining the relevant action set requires a lot of domain knowledge about the state space features and the dynamics function. As of now, there does not seem to be any way the ideas in this paper could be useful for any practical domain. - Can the authors provide some concrete examples of how one can obtain such relevant action sets for problems of practical interest and scale? - Can the authors provide any results on a commonly used continuous action space RL benchmark? ## 2. Gains not coming from the policy gradient, but only because of constraining the action space The paper's proposed formulation is interesting because it uses continuous action masking as part of the learning policy and informs the policy gradient update about the continuous action mask. However, when we look at the resultant policy gradients for each mask in Eq. 10, Line173, and Line199, it seems that the policy gradient simply reduces to $\lambda_\theta log \pi_\theta(a \| s)$ for all cases. So, the effective change in implementation is just how the action to take in the environment is model: $a^r = g(a)$. But, this doesn't utilize continuous action masking to improve the policy learning objective in any meaningful way. Is my understanding correct in this? Another observation that validates the claim that policy learning is not influenced much is seen from the results and L248-249. The initial rewards themselves are significantly higher, which means that the action mask just reduces the space of exploration of the agent so much that, as long as it takes a valid action, it would get a high reward. ## 3. Simpler baselines for continuous action masking Continuing from the above point, if all that needs to be done is to compare different formulations of g(a), there is a much simpler alternative perspective: - Action-Masking as part of environment: Simply, apply the action mask as part of the environment, without changing the PPO objective at all. So, there is an action-masking filter before executing an agent's action in the environment, and ignore if the action is invalid. - Sampling-augmented action-masking: Keep sampling actions from $\pi_\theta$ until you find a valid action that can pass through the known continuous action-masking map. The current PPO baseline is very weak, and does not utilize action-masking at all. It seems most of the learning in PPO is going into the effort of learning the action mask. To really justify this paper's proposed action masking schemes are useful, they must compare against other forms of naive action masking, including the two listed above. This perspective on considering the action mask as part of the environment is also much more generally applicable and does not require any change to the policy gradient update. Technical Quality: 3 Clarity: 3 Questions for Authors: Listed in weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discuss several limitations of the work, but the important ones of applicability and baselines need to be addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your thoughtful and critical comments which helped us to strengthen our arguments for the utility of action masking. We address your questions below. ## 1. General applicability of this paper's ideas Thank you for pointing this out. We agree that the two statements appear contradictory at first glance. However, "can" removes the contradiction in our view. In line 5, “can be” indicates that in some cases, minimal task knowledge might be enough to identify relevant actions. Using "can be obtained" in lines 284-285 highlights that while it is possible to obtain a relevant action set, it may also be quite challenging in many practical situations. Thus, our seemingly contradictory statements highlight the spectrum of difficulties to obtaining a relevant action set, on which we elaborate further in the summary rebuttal (A2). We will highlight this spectrum better in the revised version of this paper. As we mention in the paper, action masking can be especially useful for safety-critical applications, where relevance means collision-avoidance. Regarding your first point, we provided such concrete examples for the application of autonomous driving in A2 of the summary rebuttal. To address the second point, we conducted an additional experiment for the MuJoCo Walker2D task, which we justify and evaluate in A3 of the summary rebuttal. ## 2. Gains not coming from the policy gradient, but only because of constraining the action space Thanks for your question. We defined action masking as transforming the unmasked policy $a \sim \pi_\theta(a \| s)$ to the masked policy $a^r \sim \pi_\theta^r(a^r \| s)$, for which $a^r \in \mathcal{A}^r$ always holds. While most stochastic policy gradient algorithms utilize normal distributions with well-known policy gradients $\nabla_\theta \log \pi$, masked policies generally do not follow standard normal distributions. Consequently, we derived gradients for these masked distributions to provide a mathematically sound method. It's important to note that we do not claim these adapted gradients necessarily improve algorithm performance in general. Nevertheless, it is not true that all masked policy gradients always reduce to the unmasked policy gradient $\nabla_\theta \log \pi_\theta(a \| s)$. This only applies to the ray approach, and the proof for it is not trivial, making it also an important contribution. For the generator mask, the policy gradient remains the same, only if the generator matrix is invertible (line 173), which occurs solely when the zonotope has $2^N$ vertices. The distributional mask yields a substantially different gradient (see equation (17)). In practice, we approximate it with the original gradient, since the integral over $\mathcal{A}^r$ is intractable. We acknowledge this simplification in Sec. 4.3, and suggest potential approximations for future work. Our primary intention was not to claim that the derived gradients universally improve RL algorithm performance. Rather, we emphasize that deriving correct gradients for modified distributions is crucial for establishing a mathematically sound method, thus forming an essential part of our contribution. Regarding your second observation, we agree that most of the learning of vanilla PPO goes into the effort of learning to select relevant actions. However, we see this as a justification for the utility of action masking because it allows us to encode task knowledge directly into the policy and, thereby reducing the exploration space. Nevertheless, we do not intend to claim that action masking is inherently better than standard PPO, which we will make sure to clarify in our revised paper. Yet, our experiments do show that utilizing relevant action sets with action masking can drastically improve the convergence speed, albeit at a higher computational cost. ## 3. Simpler baselines for continuous action masking Thank you for proposing two additional baselines for comparison. However, we do think that our implementation and experiments already cover both of them. The ray mask is a version of the first proposed baseline. As derived for proposition 1, the ray mask does not affect the gradient of the policy distribution (because $g(a)$ is bijective), which makes it mathematically equivalent to being defined as part of the environment. We further address the proposal of generally viewing action masking as part of the environment in A1 of the summary rebuttal. The second proposed baseline is mathematically equivalent to our distributional mask, since it only modifies the sampling procedure to use rejection sampling instead of geometric random walks. We initially used rejection sampling, but quickly found that it carries a significant downside, which makes it not applicable to RL. Consider the case in two dimensions, where $\mathcal{A}^r$ is a small set centered at $\left[ -0.5, -0.5 \right]^T$. If the policy $\pi_\theta(a \| s)$ defines a normal distribution with mean at $\left[ 1.0, 1.0 \right]^T$ and small variance, the likelihood of a sample $a \sim \pi_\theta(a \| s)$ being inside $\mathcal{A}^r$ is almost zero. This can cause the algorithm to get stuck at a sampling step, which we observed in practice. The issue even aggravates in higher dimensions due to the curse of dimensionality. Nevertheless, we acknowledge that the PPO baseline may be considered relatively weak. We add a comparison to another common approach for adapting the actions of RL agents; action replacement [14]. This method substitutes actions outside the relevant action set with randomly sampled actions from within it. We depict the comparison in Fig. 1 of the rebuttal PDF. In the experiment, masking performs as good as action replacement in the 2D Quadrotor task, better in the 3D Quadrotor, and significantly better in the Seeker environment. These result suggests that action masking is a competitive approach with respect to a baseline that also explores the relevant action set only. [14] Krasowski et al. 2023	Rebuttal 1: Rebuttal: Dear reviewers, Thank you for your thoughtful comments and questions. We address general points below. ## A1. Relevance of continuous action masking as part of the policy Action masking enforces task knowledge by focusing learning on relevant actions, thereby increasing sample efficiency and reducing the need for reward tuning - two common practical problems in RL. The existing literature mostly regards discrete action spaces and shows masking is highly effective [4,6,11,22]. However, real-world systems often operate in continuous action spaces, for which only interval action masks have been explored [14]. We develop three methods that enable action masking for arbitrary convex sets. With our experiments, we demonstrate that continuous action masking can increase convergence speed and performance significantly. While reviewers 8567 and PubC suggest to define action masking as part of the environment, we argue for incorporating it into the policy distribution, as done for discrete action spaces [11]. This has three main advantages. First, this formulation is more intuitive since the relevant actions $a^r$ stay interpretable in the original action space $\mathcal{A}$. E.g., for the generator mask, adding the mapping function $g(a)$ to the environment would result in an action space with a dimension for each generator of the relevant action zonotope. An action would be the generator factors, for which the real-world meaning is not intuitive. Second, our formulation allows to incorporate information about the modification into the gradient. This is relevant for the generator and distributional mask where the relevant policy gradient is different to the original policy gradient (see Proposition 3 and 4). Since computing relevant action sets can be costly, it is desirable to use this information in the backward pass as well. Third, action masking as part of the policy can be used for more formulations. Specifically, the distributional mask cannot be moved to the environment since it directly modifies the sampling of the policy distribution. We will clarify these advantages in Sec. 3 of the revised paper. ## A2. Computing relevant action sets We assume that a set of relevant actions is computable. A relevant action set reduces the action space, based on task knowledge. This is a flexible definition that allows for different levels of complexity and required task knowledge. Thus, action relevance is a spectrum and the specific definition is a design choice. For our experiments, we equate relevance with guaranteed collision-avoidance since this is an important feature for safety-critical tasks that standard RL does not achieve [14]. For this highly interpretable and strict notion of relevance, the necessary level of task knowledge (e.g., system dynamics and unsafe sets) and computational cost is high. Yet, this can be a reasonable effort for gaining safety guarantees. However, there are many other notions of relevance with substantially different levels of required task knowledge and compute. Let us illustrate this for autonomous driving. Large steering angles at high velocity potentially destabilize a vehicle, i.e., can be seen as irrelevant actions. Relevant action sets that restrict steering angles depending on the velocity are trivial to derive. Another notion of relevance is compliance with traffic rules [R1], e.g., only turning right in a right-turn lane, or not accelerating in front of a red light. Here, a medium amount of task knowledge (e.g., road map) is required and the relevant action sets for a single rule are still straightforward to compute (e.g., only allowing steering that leads to a right turn). Similar to our experiments, one can also define relevance as collision avoidance for driving. Due to a highly dynamic environment with other participants, this leads to high required task knowledge. Yet, this is not infeasible as demonstrated by [R2]. Note that there is also recent work on obtaining relevant action sets in a data-driven manner [R3]. Nevertheless, we agree that action relevance and resulting implications for computational costs are not described sufficiently, hence we will clarify this in the revised paper. ## A3. Additional experiments While we believe that our experiments provide sufficient initial evidence validating our theoretical contributions to action masking, we agree with the reviewers that an evaluation of our continuous action masking approaches on a more diverse set of tasks would further strengthen our claims. Therefore, we are adding experiments on the Walker2D environments (https://gymnasium.farama.org/environments/mujoco/walker2d/). Our concept here is to define the relevant action space as all actions for which $\|\|a\|\|_1 \leq \alpha_p$. This can be viewed as a maximal power output constraint $\alpha_p$. The results of the experiment are depicted in Fig. 3 of the rebuttal PDF. We observe that the generator and ray mask both learn a performant policy and that the generator mask slightly outperforms standard PPO. Further, encoding the relevant action space as a termination condition prohibits standard PPO to learn. Note that distributional mask is excluded, since it's slow computation time failed to produce results in the available time. In conclusion, this small experiment further highlights the practical utility of action masking. In addition, we add a comparison to an common approach that adapts irrelevant actions in the environment; action replacement [14]. Our experiment shows that action masking performs better or as good as replacement depending on the environment (see Fig. 1 of rebuttal PDF). ## References: [4] Feng et al. 2023 [6] Fulton et al. 2018 [11] Huang et al. 2022 [14] Krasowski et al. 2023 [22] Rudolf et al. 2022 Additional: [R1] Mehdipour et al. 2023. "Formal methods to comply..." [R2] Wang et al. 2023. "Safe Reinforcement Learning for Automated Vehicles..." [R3] Theile et al. 2024 "Learning to Generate..." Pdf: /pdf/a1e177b953061f7b579f8ef0064c939e48e5dfcf.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MAmmoTH2: Scaling Instructions from the Web	Accept (poster)	Summary: The paper proposes a 3-stage pipeline to harvest ex-large-scale instruction data from the pre-training web corpus to enhance LLM reasoning, which involves 1) recalling relevant documents, 2) extracting instruction-response pairs using LLM, and 3) refining the extracted pairs by completing the intermediate reasoning steps using LLM. The paper - proposes an effective pipeline to synthesize large-scale high-quality instruction data, especially for reasonable prompts and reliable answers; - empirically validates the effectiveness of scaling up instruction data for reasoning tasks; - builds `MAmmoTH2-Plus` models, achieving performance superior to or comparable with previous SotA on various reasoning datasets; - provides a ex-large-scale instruction dataset for reasoning tasks, WebInstruct, as unique public data resource; - conducts extensive ablation studies, providing many insights like: - SFT loss is better for LM loss (at least when evaluated on QA tasks); - refining extracted instruction pairs by completing the intermediate reasoning steps is significantly helpful; - using multiple LLMs to refine the instruction data is usually better than a single LLM; - “Education” data (exam-style) are usually better than “Forum” data (discussion-style) (at least when evaluated on QA tasks); - even benchmarks conventionally thought very relevant might conflict with each other (GSM & MATH in Table 5), implying limited generalization of LLMs. Strengths: - The scaling effect of instruction data is an important empirical question. The paper is the first to scale instruction data to 10M pairs, showing the feasibility and effectiveness of scaling up instruction data (for reasoning tasks). - Synthesis of high-quality prompts and answers is important for further data augmentation but rather under-explored. The paper finds an effective method to synthesize reasonable prompt and relatively reliable answers by harvesting from web corpora. - `MAmmoTH2-Plus` models achieve performance superior to or comparable with previous SotA on various reasoning datasets. - Extensive experiments are conducted on various base models and especially diverse challenging reasoning benchmarks, instead of easy ones with limited scope (e.g. many benchmarks similar to GSM8K), convincingly validating the method's effectiveness. - Many insightful and useful observations in ablation studies (as mentioned in the summary). - The paper is generally well written to be clear and detailed. Weaknesses: - It might need further consideration about *whether training on WebInstruct* is compatible with or necessary to be added to existing training pipelines to achieve the best final performance (for reasoning tasks)*. The paper achieves its best performances (`MAmmoTH2-Plus`) with a 2-stage instruction tuning on pre-trained models but doesn’t involve continual pre-training, which should be rather important for models’ reasoning abilities as proved by works like DeepSeek-Math. Pre-training and RL should be out of this work’s scope. But it would be better to further clarify the impacts of 1) continual pre-training, 2) training on WebInstruct, 3) final fine-tuning on additional instruction datasets and their combinations. - Table 7 shows the performance on reasoning benchmarks of applying 2/3/2+3 on Mistral-7B/Mixtral-8x7B. But the comparison might be a little unfair: the domains of the “Public Datasets” are wider than those of the WebInstruct* with the code generation dataset Code-Feedback, but the benchmarks only involve mathematical and scientific reasoning in natural language, which might underestimate the performance of “Public Datasets”, considering the possible confliction between code generation and reasoning in natural language. It might be better to remove Code-Feedback from the “Public Datasets” to compare with WebInstruct. - To consider 1) continual pre-training, it is impossible to conduct by yourselves, but a possible workaround could be to make full use of the resources provided by DeepSeek-Math: DeepSeekMath-7B is continually pre-trained from DeepSeek-Coder-Base-v1.5. By comparing performances on reasoning benchmarks of applying 2/3/2+3 on DeepSeek-Coder-Base-v1.5/DeepSeekMath-7B and the two models themselves, a more comprehensive study on the impacts of these training stages can be done. - Table 7 shows that, for strong Mixtral-8x7B, the gains of adding WebInstruct to “Public Datasets” is marginal, implying that *the effect of WebInstruct* for strong base models might be limited**. --- # After rebuttal and discussion The authors resolved most concerns and validated that MAmmoTH2 can efficiently substitute continual pre-training in the standard SotA pipeline. The limitation is that MAmmoTH2 fails to combine with continual pre-training to effectively push forward the upper limit. I decide to change my score to 8. Technical Quality: 4 Clarity: 4 Questions for Authors: Suggestions: - The refinement step is important and the current setting can be seen as distillation from strong models (Mixtral-22B×8 and Qwen-72B). The method could be more promising if it could help self-improvement/weak-to-strong generalization. I highly recommend adding experiments of training Mixtral-22B×8 and Qwen-72B or stronger models in future versions. Confusions: - Are training data sizes in experiments for Table 5 controlled to be comparable? - What does the Data Source “Base“ mean in Table 5? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The limitations of this work are acceptable and the authors point out potential directions to address the limitations for future works. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback on our effective pipeline, large-scale instruction dataset for reasoning tasks, many useful insights from extensive experiments. > "Compatibility with existing continual-training pipelines and impact investigation" We appreciate this valuable suggestion though continual training is beyond our project's scope. Our focus was on improving reasoning performance through scaled instruction tuning. To address this point, we fine-tuned the Deepseek-Math-Base-7B on WebInstruct and additional public datasets. Results show that WebInstruct can further significantly improve the DeepseekMath (it’s already been continue-pretrained on math documents). After fine-tuning additional public SFT data, our model achieves comparable performance on math reasoning and higher performance on other reasoning benchmarks, demonstrating compatibility with existing continual training pipelines. We will add the additional results and discussions in the revision. \| \| TheoremQA \| MATH \| GSM8K \| GPQA \| MMLU-S \| BBH \| ARC-C \| AVG \| \|---------------------------\|-----------\|-------\|-------\|-------\|--------\|-------\|-------\|------\| \| Deepseek Math 7B Base \| 25.3 \| 34.0 \| 64.2 \| 29.2 \| 56.4 \| 59.5 \| 67.8 \| 48.1 \| \| + WebInstruct \| 30.1 \| 38.2 \| 70.5 \| 33.3 \| 59.5 \| 61.8 \| 76.1 \| 52.8 \| \| + Additional SFT \| 31.5 \| 45.2 \| 80.2 \| 35.2 \| 60.5 \| 62.0 \| 76.4 \| 55.8 \| \| Deepseek Math 7B Instruct \| 23.7 \| 44.3 \| 82.9 \| 31.8 \| 59.3 \| 55.4 \| 70.1 \| 52.5 \| \| Mistral 7B Base \| 19.2 \| 11.2 \| 36.2 \| 24.7 \| 50.1 \| 55.7 \| 74.2 \| 38.8 \| \| + WebInstruct \| 29.0 \| 36.7 \| 68.4 \| 32.4 \| 62.4 \| 58.6 \| 81.7 \| 52.8 \| \| + Additional SFT \| 29.2 \| 45.0 \| 84.7 \| 36.8 \| 64.5 \| 63.1 \| 83.0 \| 58.0 \| > "Self-improvement and weak-to-strong generalization" We agree that exploring self-improvement and weak-to-strong generalization would be valuable. We'll consider experiments with Mixtral-22B×8, Qwen-72B, or stronger models in future work. > "Public Datasets domains wider than WebInstruct" This is not accurate. We evaluated code generation (HumanEval, MBPP) and general chat benchmarks (MT-Bench, AlpacaEval 2.0, Arena Hard) in Table 3. The additional PLUS data training aims to make our models more general and capable of tasks beyond reasoning. > “Confusions” For Table 5, we train all models with the same steps. We will clarify this. “Base" in Table 5 refers to the base model's performance. We'll make this clearer in the table caption or legend. --- Rebuttal Comment 1.1: Comment: Thanks to the authors for your clarifications! However, I still have some concerns as below. > "Compatibility with existing continual-training pipelines and impact investigation" I understand that you focus on scaled instruction tuning. However, from a holistic perspective, it is meaningful to know what components are necessary for a SotA end-to-end pipeline. Despite your new experiments, we still don't know the comparisons between: - CPT + SFT v.s. CPT + Scaled SFT + SFT (e.g. DeepSeekMath-7B-Base + Additional SFT v.s. DeepSeekMath-7B-Base + WebInstruct + Additional SFT) -- It is possible that adding WebInstruct might show few gains, similar to Table 7. - settings above v.s. corresponding ones without CPT (e.g. substituting DeepSeekMath-7B-Base with DeepSeek-Coder-Base-v1.5 to conduct the experiments above and compare all results together) -- It is possible that WebInstruct might efficiently substitute DeepSeekMath corpus without damaging performance. > "Public Datasets domains wider than WebInstruct" I am talking about results from Table 7, where you didn't evaluate coding tasks but only reasoning tasks, instead of Table 3/6. I consider it slightly unfair because Additional SFT contains code-related data, which might damage its reasoning performance. --- Reply to Comment 1.1.1: Title: Follow up - Jcxm Comment: Thanks so much for your follow-up question! We appreciate your constructive comments! To some extent, you share some common concerns as Reviewer k2fv. Do you think it makes sense to you if we have the additional results based on the Deepseek with the following rows: - (a) Deepseek Coder V1.5 (base model) - (b) Deepseek Math (base model + CT/Recall) - (c) Deepseek Coder V1.5 + WebInstruct (base model + Extract + Refine): to verify whether scale-up SFT could be a more cost-effective way than traditional CT. - (d) Deepseek Math + WebInstruct (CT/Recall + Extract + Refine) - (e) Deepseek Math + WebInstruct + Additional SFT (CT/Recall + Extract + Refine + Additional SFT) - (g) Deepseek Math + WebInstruct (without Refine) + Additional SFT (CT/Recall + Extract + Additional SFT) - (f) Deepseek Math + Additional SFT (CT/Recall + Additional SFT) If that makes sense to you, we will try to add the results during the discussion period. Thank you again for the constructive comments!	Summary: The paper introduces MAmmoTH2, a novel approach to instruction tuning for large language models (LLMs) by harvesting naturally existing instruction data from the web. The authors develop a three-step pipeline (recall, extract, refine) to collect 10 million high-quality instruction-response pairs without relying on costly human annotation or GPT-4 distillation. Fine-tuning LLMs with this dataset significantly improves performance on reasoning benchmarks. The MAmmoTH2-Plus model, further tuned on public instruction datasets, achieves state-of-the-art results on multiple benchmarks. Strengths: - Demonstrates a cost-effective way to collect large-scale, high-quality instruction data from the web. - Significant performance gains on reasoning benchmarks, with MAmmoTH2 models outperforming existing models. - Comprehensive evaluation across multiple benchmarks, showing robust improvements. Weaknesses: - The approach primarily combines existing methods (data recall, extraction, refinement) rather than introducing fundamentally new concepts or techniques. - More explicit comparison with prior work is needed to highlight the unique contributions and differences of this approach. - The quality and diversity of the collected data heavily depend on the web sources, which may introduce biases or inconsistencies. Technical Quality: 3 Clarity: 3 Questions for Authors: - How does MAmmoTH2 compare directly with other methods that use synthetic or human-annotated data in terms of data quality and model performance? - What measures were taken to ensure the quality and relevance of the extracted Q-A pairs from the web? - How does the model address potential biases in the web-sourced data, and what steps were taken to mitigate these biases? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors address some limitations of their approach, such as the dependency on web data quality and the challenges in maintaining the diversity and relevance of the instruction data. However, a more detailed discussion on potential biases introduced by web data and the ethical implications of using such data could strengthen the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for positive feedback on our cost-effective approach, significant performance gains, and comprehensive evaluation! > "Novelty of approach" Our method's novelty lies in its unique pipeline to mine naturally existing instruction data at scale, offering a new paradigm for creating large-scale, high-quality instruction datasets without costly human annotation or GPT-4 distillation. To the best of our knowledge, we are the very first paper to formally understand the effect of automatically scaling up SFT data to see its impact, especially on reasoning tasks. Reviewer yEcJ, Reviewer k2fV, and Reviewer Jcxm have acknowledged our approach as novel, simple, and effective. > "More explicit comparison with prior work" In Tables 2 and 3, we provide a comprehensive comparison with prior works, encompassing general pre-trained base LLMs, general instruction-tuned models, and reasoning-specific models such as Deepseek Math, Intern-Math, and Rho-1-Math. Notably, we also include a comparison with Llama-3-8B-Instruct, which utilizes 10 million human-annotated instructions. Our model demonstrates superior reasoning performance to these existing models. > "Web sources have biases and inconsistencies. What steps were taken to mitigate these biases?" We've taken several steps to mitigate biases and ensure data quality: * Using diverse seed data across multiple domains * Employing multiple LLMs in the refinement stage * Implementing a three-step pipeline (recall, extract, refine) to improve data quality > "How does MAmmoTH2 compare directly with other methods that use synthetic or human-annotated data?" * We are pioneers in developing a method for scaling up instruction tuning via data synthesis. * Our approach outperforms models trained on human-annotated datasets of similar size (e.g., Llama-3-8B-Instruct) on reasoning tasks while matching performance on general tasks, suggesting comparable or higher quality for certain tasks. > "What measures were taken to ensure the quality and relevance of the extracted Q-A pairs?" We use a multi-step process to ensure the quality and relevance of our data: * Careful selection of seed data and websites * LLM-based extraction of relevant Q-A pairs * Refinement step to improve formatting and add missing explanations * Human evaluation of a sample set (as shown in Figure 6) --- Rebuttal Comment 1.1: Comment: Thanks for your rebuttal, I have no more questions.	Summary: This paper proposes an approach to automatically harvest large-scale instruction data from pre-training corpora for reasoning tasks. The main steps include: (1) Recall: training a fastText model to recall relevant documents from the pre-training corpus, similar to DeepSeekMath; (2) Extract: using open-source models with few-shot prompting to extract question-answer pairs from the recalled documents; (3) Refine: prompting open-source models to remove noise, adjust formats, and complete the reasoning process for the extracted question-answer pairs. Using this method, the authors harvested 10 million instruction data and trained MAmmoTH2 models. Without relying on closed-source models, MAmmoTH2 achieves excellent performance on various reasoning tasks. Strengths: - Method: The motivation is clear, and the idea of automatically extracting instruction data from web data is novel, simple, and scalable. - Experiments: The experiments and evaluations are comprehensive and achieve good results. - Well Written: The paper is very easy to understand. - Reproducibility: The authors have open-sourced part of the corpus, models, and evaluation scripts to ensure the reproducibility of the results. Weaknesses: 1. Effectiveness: I wonder if the WebInstruct approach can further improve the performance of state-of-the-art domain models. For example, DeepSeekMath achieved good results by only training on recalled documents and fine-tuning on high-quality data (MATH: DeepSeekMath-7B-Instruct 46.8% vs. MAmmoTH2-7B-Plus's 45.0%). Moreover, since the models have already been trained on SFT data, comparing only the few-shot performance is not comprehensive enough. I suggest also comparing the performance of the Plus version trained with high-quality "additional instruction datasets" for most of the experiments. Consider supplementing the following results: - Recall + Plus: Directly train on the 18M recalled documents and fine-tune a Plus version to verify if the "extract + refine" steps have significant benefits. - Recall + Extract + Plus: Directly train on the extracted QA (Fig.5, Extracted QA) with LM/SFT loss and fine-tune a Plus version to verify the benefits of the refine step. - In Fig.5, I also recommend reporting the performance after fine-tuning the Plus version for SFT loss vs. LM Loss. 2. Lack of method details: - For example, the code for the recall stage and the prompts used for extraction and refinement could be included in the repository or appendix. - In Sec. 5.1, I suggest explicitly defining the SFT Loss to help more readers understand it clearly. By "SFT Loss", I understand the authors mean "masking the loss of instruction input", right? 3. Scalability: - The effectiveness of WebInstruct constructed using small models is unknown for larger models; moreover, this approach is difficult to apply to models with hundreds of billions of parameters due to high inference costs. - During refinement, the model generate missing explanations. Have you observed and quantified the hallucination phenomenon? If present, such incorrect reasoning processes can negatively impact model training, such as increasing hallucination/bias, especially if the corpus is used for larger models. 4. Minor points: - Some citations are missing for baselines in Table 2, e.g., Gemma, Abel, and Rho-1. - How can the WebInstruct approach be extended to more general domains? What other issues need to be addressed? - A concurrent work, Jiuzhang3.0 [1], is quite similar in motivation and method. It would be better to discuss and compare with it. What are the advantages and issues of MAmmoTH2 compared to Jiuzhang3.0? --- [1] Zhou, Kun, et al. "JiuZhang3. 0: Efficiently Improving Mathematical Reasoning by Training Small Data Synthesis Models." arXiv preprint arXiv:2405.14365 (2024). Technical Quality: 4 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The authors have addressed most of the limitations. The limitations section can be further improved by referring to the weaknesses part. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive feedback on our work's clarity, novelty, and comprehensive experiments! > “Additional results for the effectiveness of WebInstruct” We've included early-stage results using Qwen-1.5-1.8B to demonstrate the usefulness of our "extraction" and "refinement" steps: \| Model \| MATH \| TheoremQA \| ARC-C \| \|------------------------------------\|----------\|-----------\|-------\| \| Qwen-1.8B Base \| 10.1 \| 11.1 \| 50.11 \| \| Recall \| 11.26 \| 12.38 \| 49.06 \| \| Recall + Extract \| 14.82 \| 13.25 \| 51.19 \| \| Recall + Extract + Refine (WebInstruct) \| 17.18 \| 14.87 \| 53.83 \| These results clearly show the benefits of each step in our pipeline and align with the motivation of the suggested experiments. We will include these results in the Appendix. > “Method details” - We'll add the extraction/refinement prompts to the repository and appendix for transparency. - We'll explicitly define SFT Loss in Sec. 5.1. Yes, it refers to masking the loss of instruction input. > “Scalability” The core idea of WebInstruct is to mine naturally existing high-quality instructions from the Web. Compared with hundreds of billion tokens for continual training, our approach is more cost-effective. Compared with traditional SFT datasets with only hundreds of thousands of examples, WebInstruct with 10M examples is more scalable without requiring any human annotations. > “Quantified Hallucination” We quantified hallucination through human error analysis in Figure 6. Our case study reveals that the harvested instruction tuning dataset is generally accurate with a low error rate: 78% of examples improved after refinement and only 10% introduced hallucinations. Future work includes developing more advanced methods (e.g., training a filtering/reward model) to select the least hallucinated questions. We will strengthen this discussion. > “Minor points” - We'll add missing citations for baselines in Table 2 (e.g., Gemma, Abel, and Rho-1). - We'll discuss potential extensions to more general domains and associated challenges. - Thanks for pointing out the concurrent work Jiuzhang 3.0 and we'll discuss the Jiuzhang 3.0! Both methods leverage synthetic data generation to enhance reasoning capability, but we extract and refine naturally existing instructions on the web rather than synthesizing new ones. Our core claim is that high-quality SFT data for reasoning naturally exist on the web, and our contribution is developing a simple yet effective approach to harvesting this data. --- Rebuttal Comment 1.1: Title: Is the Extract and Refine Process Necessary? Comment: Thank you to the author for supplementing the results and replying. However, my main concern has not been addressed, namely whether the (b) extract and (c) refine processes introduced by MAmmoTH2 provide significant benefit or necessity to the existing data pipeline. As mentioned in my previous comments, DeepSeekMath achieved higher results using only (a) recall and (d) sft (MATH: DeepSeekMath-7B-Instruct 46.8% vs. MAmmoTH2-7B-Plus's 45.0%). Therefore, I once again suggest that the author design controlled experiments and provide the following results (for models >=7B) to demonstrate whether (b) and (c) are necessary: 1. (a) Recall + (d) SFT: Directly train on the 18M recalled documents and fine-tune a Plus version to verify if the "(b) extract + (c) refine" steps have significant benefits. 2. (a) Recall + (b) Extract + (d) SFT: Directly train on the extracted QA (Fig.5, Extracted QA) with LM/SFT loss and fine-tune a Plus version to verify the benefits of the (c) refine step. 3. In Fig.5, I also recommend reporting the performance after fine-tuning the Plus version for SFT loss vs. LM Loss. --- Rebuttal 2: Title: Follow up Comment: Thank you for the follow-up! We really appreciate your constructive comments! To some extent, you share concerns similar to those of the reviewer Jcxm. We added some additional results based on the Deepseek Math 7B base. Do you think it makes sense to you if we have the additional results based on the Deepseek with the following rows: - (a) Deepseek Coder V1.5 (base model) - (b) Deepseek Math (base model + CT/Recall) - (c) Deepseek Coder V1.5 + WebInstruct (base model + Extract + Refine): to verify whether scale-up SFT could be a more cost-effective way than traditional CT. - (d) Deepseek Math + WebInstruct (CT/Recall + Extract + Refine) - (e) Deepseek Math + WebInstruct + Additional SFT (CT/Recall + Extract + Refine + Additional SFT) - (g) Deepseek Math + WebInstruct (without Refine) + Additional SFT (CT/Recall + Extract + Additional SFT) - (f) Deepseek Math + Additional SFT (CT/Recall + Additional SFT) If that makes sense to you, we will try to add the results during the discussion period. Thank you again for the constructive comments! --- Rebuttal Comment 2.1: Comment: Thank you for your response. To demonstrate the necessity of the extract and refine steps, I believe you only need to conduct the two experiments mentioned in my previous reply: 1. (a) Recall + (d) SFT and 2. (a) Recall + (b) Extract + (d) SFT. I think the experiments you planned on DeepSeekMath cannot prove the effectiveness of the extract step because the corpus recalled by MAmmoTH2 is different from that of DeepSeekMath. Therefore, I suggest using your 18M recalled documents as the "Recall" corpus for the experiments. --- Rebuttal 3: Title: Follow up on Ablation Results (Reviewer Jcxm and k2fV) Comment: We appreciate both reviewers' insightful comments and have conducted all the required ablation studies** before the end of the discussion period (bingo!). Our experiments, summarized in the table below, demonstrate the effectiveness and efficiency of our proposed pipeline. ### Experimental Setup - Base: DeepSeek Coder V 1.5 7B - (a) MAmmoTH2's recall documents: 18M documents, 28B (14B tokens * 2 epochs) - (a') DeepSeek Math's CT corpus: 500B tokens (120B math + others for multiple epochs) - (b) Extracted instruction-response pairs from (a): 7B (3.4B tokens * 2 epochs) - (c) Refined instruction-response pairs from (b): 10B (5B tokens * 2 epochs) - (d) MAmmoTH2 Additional public SFT: 2B (1B tokens * 2 epochs) - (d') DeepSeek Math SFT: ~1B tokens \| \| Setting \| Model \| #Train Tokens \| TheorQA \| MATH \| GSM8K \| MMLU-S \| BBH \| ARC-C \| AVG \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| \| All evaluations are held-out \| \| \| \| \| \| \| \| \| \| \| 1 \| Base \| DeepSeek Coder v1.5 7B \| - \| 18.3 \| 22.3 \| 47.9 \| 47.0 \| 53.5 \| 62.4 \| 41.9 \| \| 2 \| Base + (a) \| - \| 28B \| 23.5 \| 30.3 \| 60.3 \| 53.3 \| 55.5 \| 69.4 \| 48.7 \| \| 3 \| Base + (a) + (b) + (c) \| MammoTH2-DS \| 10B \| 27.8 \| 33.8 \| 64.0 \| 56.9 \| 58.5 \| 72.8 \| 52.3 \| \| 4 \| Base + (a’) \| Deepseek Math Base \| 500B \| 25.3 \| 34.0 \| 64.2 \| 56.4 \| 59.5 \| 67.8 \| 51.2 \| \| \| \| \| \| \| \| \| \| \| \| \| \| \| All held-out except GSM and MATH \| \| \| \| \| \| \| \| \| \| \| 5 \| Base + (d) \| - \| 2B \| 23.5 \| 37.2 \| 77.5 \| 52.0 \| 59.8 \| 66.9 \| 52.8 \| \| 6 \| Base + (a) + (d) \| - \| 30B \| 27.2 \| 39.2 \| 79.2 \| 55.6 \| 60.3 \| 71.5 \| 55.5 \| \| 7 \| Base + (a) + (b) + (d) \| - \| 9B \| 27.3 \| 38.6 \| 78.5 \| 54.2 \| 60.5 \| 70.4 \| 54.9 \| \| 8 \| Base + (a) + (b) + (c) + (d) \| MAmmoTH2-DS-Plus \| 12B \| 30.1 \| 43.8 \| 80.1 \| 59.5 \| 61.0 \| 73.2 \| 58.0 \| \| 9 \| Base + (a’) + (d’) \| DeepSeek Math Instruct \| 501B \| 23.7 \| 44.3 \| 82.9 \| 59.3 \| 55.4 \| 70.1 \| 56.0 \| ### Key Findings 1. Cost-Effectiveness: Our pipeline achieves superior overall performance compared to DeepSeek Math models, both before (Row 3 vs Row 4) and after additional SFT (Row 8 vs Row 9), while using significantly fewer tokens. While DeepSeek Math models show slightly higher results on math-specific benchmarks, our approach demonstrates better performance on a broader range of reasoning tasks, including STEM-related benchmarks. 2. “Extraction Step” Efficiency: The "Extract" step in our pipeline leads to a more cost-effective approach, using fewer tokens while maintaining comparable performance (Row 6 vs Row 7). 3. Refinement Importance: The "Refine" step proves to be crucial, significantly improving answer quality by adding missing explanations and chains of thought, resulting in substantially better performance (Row 7 vs Row 8). We believe these comprehensive experiments address the reviewer's concerns and further underscore the merits of our approach, especially the “extract” and “refine” steps. We really appreciate the reviewer's feedback, which has led to these valuable insights and a stronger demonstration of our pipeline's effectiveness. Feel free to let us know if you have further comments!	Summary: This paper proposes a method to synthesize instruction tuning data at scale from the pretraining web corpus. The proposed method first recalls relevant documents from the corpus, and then extracts QA pairs, and finally refines the extracted QA pairs with an LLM. The synthesized instruction data proves to be helpful in enhancing the model’s reasoning abilities compared with instruction tuning data from other sources. Strengths: 1. The proposed method is novel and effective. 2. The authors conduct extensive experiments to demonstrate that it’s possible to synthesize tuning data from unsupervised text corpus to build strong LLMs that outperform models trained with data collected in existing paradigms. 3. The paper is well-written and easy to follow. The code and data are released, which will serve as high-quality resources for research and building strong LLMs. Weaknesses: There lacks a discussion and comparison with a related work “Self-alignment with Instruction Backtranslation” (Li et al., ICLR'24) which also synthesizes instruction tuning data from unlabeled corpus. Technical Quality: 3 Clarity: 3 Questions for Authors: LLMs are used in the “extract” and “refine” steps in the proposed pipeline for generating and editing instruction tuning data. Will the choice of LLMs introduce bias into the synthesized data (especially compared with distillation-based methods)? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations in Appendix H and societal impacts in Appendix I. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive feedback on our work's novelty, comprehensive experiments, and clear writing! > “Lack a discussion and comparison with Humpback [1]” Thanks for the note! Humpback does not release its implementation, data, and models, which makes the replication and head-to-head comparison difficult. Fundamentally, our approach is different from Humpback: - Humpback aims to synthesize instructions by backtranslating the existing documents. We focus on mining naturally existing instruction-response pairs from the web rather than generating new instructions. - Our additional extraction step significantly makes the corpus less redundant and improves corpus quality (see the newly added results in response to Reviewer k2fV). - The "Refine" step further enhances instruction quality. We will add more detailed discussions of Humpback in the related work. > “Will the choice of LLMs introduce bias into the synthesized data (especially compared with distillation-based methods)?” - Thanks for the question! It's important to note that our approach is not distillation in the traditional sense. The LLMs in our pipeline are used solely for extraction and refinement of existing data, not for generating new instructions. MAmmoTH2 essentially learns from a cleaner version of raw web data rather than distilling knowledge from other models. - We acknowledge that the choice of LLM could influence the accuracy of extraction. To address this, we chose two open-source models, Mixtral and Qwen, known for their strong performance and different training approaches. This diversity helps to balance out potential biases from any single model. - Compared to distillation methods, our approach potentially reduces bias by preserving naturally occurring instructions from diverse web sources, while cleaning and structuring them for more effective learning. --- Rebuttal Comment 1.1: Comment: Thank you for the response and it has addressed my concerns. I increased my score to 7.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations	Accept (poster)	Summary: This paper first investigates the effect of class distribution changes on comparative zero-sample learning by proposing and analysing a class distribution shifts parameter model, leading to the idea that loss minimisation leads to poor performance of representations over the class distribution shifts. Based on this finding, the authors utilise hierarchical sub-sampling and OOD environment balancing methods to obtain robust representations and address the poor performance caused by class distribution changes in zero-sample learning, and experimentally validate the effectiveness of the methods. Strengths: 1- This paper studies the distribution bias problem caused by challenging unknown attributes in zero-shot learning and proposes an effective solution, which is important and innovative for solving the distribution bias problem in zero-shot learning. 2- The structure is clear. It enables the reader to quickly follow the research ideas and understand the content of each section. 3- Figures and tables are clear and accurate. The figures and tables in this paper are concise and clear, effectively support the ideas or conclusions, and enable the reader to grasp the critical information quickly. 4- Comparison and ablation studies are comprehensive. The authors demonstrated the superiority of their method through many experiments and analysed various factors. Weaknesses: 1- Authors should describe their proposed soft-AUC trends in detail and analyse the penalty to help readers understand how they play a role. 2- In Experiment 1, the authors used the attribute blonde hair to shift the class distribution, but we know that some people may be hairless, so can the attribute gender be used to shift the class distribution? 3- The language of this paper needs to be scrutinized and improved. For example, redundant phrases such as "with respect to" (line 32) should be avoided. In addition, there are some grammatical errors that need to be improved, such as "assumes" in line 333 should be changed to "assume", and "leverages" in line 350 should be changed to "average". 4- WRITING DETAILS: Abbreviated nouns should be introduced the first time they appear, such as “OOD”. Technical Quality: 3 Clarity: 3 Questions for Authors: See Weaknesses Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Soft-AUC: The use of soft-AUC instead of the standard AUC score is intended to make the complete loss (including the penalty) differentiable, thus enabling gradient-based optimization. As mentioned in lines 218-219, the soft-AUC pointwise converges to the standard AUC score as $\beta$ approaches infinity. For a fixed $\beta$, define $f(t_{1},t_{2})=I[t_{1}<t_{2}]$ and $g_{\beta}(t_{1},t_{2})=\frac{1}{1+e^{-\beta(t_{2}-t_{1})}}$. Denoting $t=t_{2}-t_{1}$ we have $$\left\Vert f(t_{1},t_{2})-g_{\beta}(t_{1},t_{2})\right\Vert ^{2}=\int_{0}^{\infty}\left(1-\frac{1}{1+e^{-\beta t}}\right)^{2}dt+\intop_{-\infty}^{0}\left(0-\frac{1}{1+e^{-\beta t}}\right)^{2}dt=\frac{1}{\beta}\left[2\log2-1\right].$$ We will include this calculation in the appendix, along with a graph illustrating the differences, which we have added as Figure 2 in the rebuttal figures file. 2. Gender attribute for the CelebA dataset: In principle, the male/female attribute could be used instead. However, on the CelebA dataset, representations that distinguish well between male individuals also distinguish well between female ones, and vice versa, for a reasonable representation network. Thus, shifts in gender do not call for interventions on the CelebA dataset, as all methods would be equivalent to ERM. This is not the case for the attribute of being blond. Our experiments show that the representation learned via ERM on mainly non-blond people (including hairless individuals, as this binary indication distinguishes blond people from all others) fails to separate blond people effectively. We thank the reviewer for pointing out the typos in 3 and 4.	Summary: This paper proposes a robust representation learning method that could assume the shift between seen classes and unseen classes. Strengths: Good presentation and sound method. Weaknesses: Lack the experiments on the most popular benchmark of zero-shot learning [1] and comparison to some SOTAs, e.g. [2][3]. [1] Zero-shot learning-the good, the bad and the ugly[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017. [2] Rebalanced zero-shot learning[J]. IEEE Transactions on Image Processing, 2023. [3] Transzero: Attribute-guided transformer for zero-shot learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2022. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their review and the provided references. We did not use the datasets mentioned in your references since they either (i) do not have labeled attributes (e.g., the SUN dataset), (ii) the provided attributes correlate with the data-point label such that shifts in them do not affect ERM, making interventions unnecessary (e.g., the CUBS dataset), or (iii) include too few classes (e.g., the AWA2 dataset with only 50 animal classes). Therefore, we performed the real-data experiments on the CelebA dataset (which is one of the most popular zero-shor benchmarks), and the ETHEC dataset, since both include a large number of classes and labeled attributes in addition to the primary labels (e.g., butterfly family in ETHEC and hair color in CelebA, in addition to species and person identity, respectively).	Summary: Zero-shot learning classifiers face the challenge of distribution shifts, where the distribution of new classes differs significantly from that of the training data. In this paper, the authors introduce a novel algorithm to address this problem by creating robust representations through hierarchical sampling and environment balancing penalization. Experimental results also demonstrate a performance increase compared to the baseline ERM model on several real-world datasets. Strengths: 1. This paper is well-written and easy to understand. 2. The paper proposes a new model that enables handling unknown attributes for distribution shifts and addresses new classes at test time. 3. The method is tested through both simulations and real-world experiments. Weaknesses: 1. Some parameters need to be clearly defined, for example, $\rho_{tr}$, $\rho_{te}$, and $y_{uv}$ in Eq (4). 2. The proposed method creates multiple environments and computes penalties across them. What is the computational complexity? It's also beneficial to discuss the time complexity of Algorithm 1. 3. Figure 5 and Figure 6 do not straightforwardly show the performance. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the Weakness section above. Additional questions: 1. Although the authors mention how to calculate the number of environments, it's better to include an ablation study to test the performance with different numbers of environments. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors acknowledge the limitations in their paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and comments. Below we address the raised weaknesses and questions: Weaknesses: 1. Definition of parameters: $\rho_{tr}$ and $\rho_{te}$ correspond to the the proportion of type $a_1$ classes in train and test sets correspondingly and are defined in lines 139-140, but we will make their definition more clear. $y_{uv}$ is simply the label for the pair of datapoints indexed by u and v: we need two sets of indices since Eq. 4 treats pairs from the same class ($y_{ij}=1$) and those from different classes ($y_{uv}=0$) separately. 2. Complexity: All OOD approaches from the considered family (see Section 2) involve optimization across multiple environments. Therefore, the complexity of each step includes: (i) the complexity of generating the environments, (ii) the original complexity of a single iteration over the representation neural network multiplied by the number of environments, and (iii) the computation of the aggregated penalty. Component (ii) is shared among all approaches from the considered family, while (iii) is negligible compared to training networks, usually involving simple operations like calculating the mean or variance of scalars. Thus, differences between the methods may arise due to (i). However, the additional training time due to the proposed hierarchical sampling, even when compared to training a simple linear representation, is also negligible since hierarchical sampling can be performed offline before training, as shown in the code provided in the supplementary material. We will include a note about this in the manuscript. For example, in the species recognition task, standard sampling takes 1.03 seconds, a naive implementation of hierarchical sampling takes 16.6 seconds, and training the representation takes 1 hour and 6 minutes. Thus, the additional time due to hierarchical sampling is less than 0.4% and similar results (all less than 0.5%) are achieved in all our experiments. 3. Exact performances: Simulation performance is shown in Figure 4, with exact performances reported in Table 1 in the appendix. For real-data experiments, exact performances (and corresponding p-values) are provided in Tables 3 and 4 in the appendix. Figures 5 and 6 highlight other important aspects: the learned feature importance in the simulations, demonstrating that our method relies on the intended features, and the performance of our method compared to ERM (i.e., the improvement over ERM). Question: Unlike the standard OOD setting, our environments are synthetically generated via sampling, resulting in many more environments (Nc over k) than specified in Equation 10. Therefore, there is no reason to use fewer environments, and using more can only enhance the performance of our method. However, we will include an analysis in the appendix that examines performance as a function of the number of environments used for both ERM and our method.	Summary: The paper treats the problem of learning models for zero-shot open-world classification settings (open-world meaning previously unseen classes might appear at test time) that are robust to distribution shifts. The proposed approach consists of two stages. In the first stage, synthetic environments $S_i$ are sampled from the training data following a hierarchical sampling approach, where first classes and then data pairs according to sampled classes are sampled. In the second stage, the model is updated to minimise a loss composed of standard ERM and the variance over environment AUC scores. The benefits of the method are demonstrated on synthetic data, CelebA, and ETHEC (where also on the latter two a distribution shift is introduced synthetically). Strengths: - The proposed approach to generate synthetic environments through hierarchical sampling seems neat and novel (even though the idea of generating synthetic environments for learning robust models is not novel, see weaknesses) - Adjusting the performance metric in the variance regularisation term for zero-shot verification, using AUC on embedding distances instead of loss (like in VaRex) seems like a nice way of avoiding performance plateaus and enables better performance in the conducted experiments - In the experiments, the proposed method shows significant performance gains over ERM - It is good to see a theoretical derivation of the necessary number of sampled environments to achieve a minimum number of examples from each class in at least one environment during the hierarchical sampling with a certain probability (Section 4.4). I believe this result should be made more prominent in form of a Proposition with a proof (in the appendix). Weaknesses: - Lack of baselines wrt generation of synthetic environments: The idea of generating synthetic environments to learn models that are robust to distribution shift is not new, and as such the proposed approach should have been compared to existing methods for this. For example, it would be interesting to see how the approach proposed by the authors compares to the approach of the 'Environment Inference for Invariant Learning' paper by Creager et al. (2021). - The real data experiments are still semi-synthetic in the sense that the distribution shift is introduced synthetically (and is quite stark). I do understand that finding a dataset that has a stark enough distribution shift of one attribute inherently in it is hard or even impossible, and the synthetic shifts are good for highlighting the potential merits of the proposed approach. However, what is missing is a report on the performance of the proposed approach (in comparison to baselines) on the unshifted train and test sets of CelebA and ETHEC, to ensure that there is no performance trade-off. Minor: - Figure 2 is never referred to in the text - as a result it is unclear what its purpose is. - It would be helpful to refer to the result of Section 4.4 already in Section 4.1, when it is claimed that 'hierarchical sampling results in diverse mixtures of any unknown attribute' and that 'smaller subsets with $k < N_c$ classes are likely to exhibit distinct attribute distributions' Section 4.4. backs up these claims. - L 235 typo at the end of the line -> (10) - Line 151: $h$ needs to be defined before referring to it in an equation, or immediately after the equation. Technical Quality: 3 Clarity: 3 Questions for Authors: - How was the number of synthetic environments in Experiments 1 and 2 on real data chosen? What values of $\alpha$ from Section 4.4 do they correspond to? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are briefly discussed in the Discussion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their review and questions. Below we address the raised weaknesses and questions. Weaknesses: 1. Comparison with Creager et al. (2021): Thank you for referring us to Creager et al. (2021). We found their work very interesting and will cite it in our related work. The main difference between our construction of environments and theirs is that Creager et al. (2021) infer the worst-case environments for a fixed classifier (e.g., trained via ERM). In contrast, we consider a shift in an unknown class attribute, which is not necessarily the worst-case and may be uncorrelated or misaligned with the worst-case scenario. However, in our synthetic data simulations (but not real-data experiments), we demonstrated our method on the worst-case shift. Therefore, following your suggestion, we compared the performance of our method using our environments versus those of Creager et al. (2021). The results showed no statistically significant improvement over ERM when using Creager et al.'s environments. We included the results in Figure 1 in the rebuttal figures file. Analyzing the results, we discovered that in the context of contrastive learning, Creager et al.'s assigned environments were almost random, as the optimized soft-assignment q barely changed from the random initialization. We attribute this to Creager et al.'s method being based on the IRM objective, which directly applies to gradients that are known to be noisy in contrastive learning (as discussed at the end of Chapter 5.1). This finding aligns with our broader results, which show that applying the IRM penalty, even on our environments where other penalties provide improvement, does not yield significant improvement over ERM. 2. Performance on unshifted distributions: The performance of the proposed approach on the unshifted train and test sets of CelebA and ETHEC you mentioned is indeed included in figure 4 for the simulations and in table 3 in the Appendix and indeed shows no negative effect on unshifted distribution performance. We agree it might be preferable to move it to the main article. Question: In both experiments, we set the minimal $\rho$ to 0.15. Therefore, the number of synthetic environments in Experiment 1 (CelebA, with 450 classes after filtering) and Experiment 2 (ETHEC, with 117 classes after filtering) corresponds to $\alpha=0.05$ and $\alpha=0.09$, respectively. We will specify this in the manuscript as well. Minor: 1. Figure 2: Thank you for pointing this out, we will add an explanation in the manuscript. Figure 2 shows that the optimal weights of dimensions corresponding to a given type are larger for higher proportions of that type, and that the discrepancy between the weights increases as the ratio of the type variances grows. 2. Location of Section Section 4.4: We accept the recommendation to move Section 4.1 immediately after Section 4.4 and thank the reviewer for the suggestion. --- Rebuttal Comment 1.1: Title: Clarification for rebuttal comment 1 Comment: To clarify this comment regarding Figure 1 in the new figure file: "However, in our synthetic data simulations (but not real-data experiments), we demonstrated our method on the worst-case shift. Therefore, following your suggestion, we compared the performance of our method using our environments versus those of Creager et al. (2021). " Figure 1 shows a comparison of the results of (a) vanilla ERM with (b) the results of training using the VarAUC penalty on synthetic environments that are formed using the method of Creager et al 2021. We can see that using VarAUC on those synthetic environment does not (significantly) improve performance on any of the datasets. This is in contrast to using VarAUC for our hierarchical environments, whose improvement over ERM is shown in the main paper. --- Rebuttal Comment 1.2: Comment: I appreciate the reviewers answers to my questions and concerns and have updated my score accordingly.	Rebuttal 1: Rebuttal: We thank the reviewers for their efforts in reviewing our paper. We address the concerns raised by each reviewer separately. We attach here the rebuttal figures file, which contains the additional figure referenced in our individual responses. Pdf: /pdf/1072b5c61539a6beecd0f9c6340884b543567200.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Learning Bregman Divergences with Application to Robustness	Accept (poster)	Summary: This paper proposes to use input-convex neural networks to learn Bregman divergences as a means to distinguish semantically meaningful image corruptions from random noise perturbations. The approach is linked to classifier robustness by showing how the associated mirror descent algorithm can be used to perform adversarial training against image corruptions coming from a Bregman ball. Experiments on benchmark corruption datasets show that the proposed method outperforms prior learned similarity metrics in distinguishing corruption from noise and in adversarial training. The proposed method is also shown to generalize quite well to corruptions that it is not trained on. Strengths: The paper is very well-written, easy to follow, and has great visualizations. The proposed method appears novel and performative, and is certainly of interest to the ML and robustness communities. Weaknesses: See questions below. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Line 42: "...that has to be convex and with invertible gradient." Do you mean strictly convex? 2. Line 37: The title of this paragraph is "Bregman divergence and mirror descent," yet mirror descent is not discussed at all. Please give a brief description of mirror descent here. 3. Line 64: Please put the footnote number "1" after the punctuation (period). With the footnote number before the punctuation, it makes it look like an exponent. Please also do the same for all other footnotes. 4. Line 73: In (3), do you mean "argmin" instead of "min"? 5. Table 1: Please write out "ICNN" completely as "input convex neural network" here, or define the acronym somewhere in the text before Table 1 for readers unfamiliar with ICNNs. 6. Line 79: This sentence looks strange starting with $\mathbb{B}_h$; I suggest adding "The ball" to the beginning of the sentence. 7. Table 1: Do you mean "strictly convex" instead of "strongly convex" in the text underneath the base function? 8. Line 79: "...but not necessarily convex." This is not true; Bregman balls, as you've defined them, are indeed always convex. Since $h$ is convex, its domain $\mathcal{X}$ is convex, and therefore convexity of the Bregman ball (4) follows from convexity of $\mathcal{X}$ together with convexity of $D_h$ in its first argument. On the other hand, if you were to have defined the ball with respect to a fixed point in the first argument of $D_h$ and varying second arguments, then the set is not necessarily convex. 9. Line 102: I don't recall ever seeing an ICNN defined in terms of Hadamard squares of the input feedthroughs. Can you explain why you are choosing to define your ICNN model using these Hadamard squares, and how the properties of this model might differ from just using the standard linear feedthroughs with arbitrary (not necessarily nonnegative) weights? Your model seems somewhat restrictive in how much influence the feedthrough may have, as its contributions to each preactivation vector are always nonnegative. 10. Line 123: Again, starting a sentence with math looks strange. 11. Figure 4: How come Moco appears to perform on par with or better than the other two prior learning-based methods in Figure 4b, but it's accuracy is shown as 0 across all noise levels in Figure 4a? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the thorough review and for finding our work "novel and performative" and "of interest to the ML and robustness communities". The questions Q2, Q3, Q4, Q5, Q6 and Q10 are directly incorporated for the next revision. We answer the rest of the questions below: Q1 and Q7: In both cases, we mean strongly convex. This will corrected in the next revision. Q8: Bregman balls are not necessary convex [21]. For example, the Itakura-Saito ball (a Bregman ball for $h(x)= − \log x$) is not convex (Nock et al. 2005). Q9: The definition in line 102 is indeed a special case of an ICNN. This can be seen by expanding equation (5) to: $u^1 = q^0\left[W^0x + 0 \right]$ $z^1 = g^0\left[U^1u^1 + V^0x + b^0 \right]$ $u^l = q^{l-1}\left[W^{l-1}x + 0 \right]$ $z^{l} = g^{l-1} \left[U^lu^{l} + V^{l-1}z^{l-1} + b^{l-1} \right]$ here [1, Proposition 1] imposes that the activation $q^l$ and $g^l$ are convex and non-decreasing and the weights $V^1, .., V^{L−1}$ and $U^1, .., U^{L−1}$ are non-negative. In particular, we can choose the activations $q^l$ to be the Hadamard power $()^{\circ 2}$ and the weights $U^1, .., U^{L−1}$ to be the identity matrix to obtain exactly the equation (5). So the non-negativity of the weights is a requirement for convexity that we can not avoid here. The Hadamard square is a choice for an activation function that has practical benefits. As we intend to compute the derivative of this network with respect to the input (to obtain $\Psi$), the derivative of the Hadamard square will be linear feedthroughs. We have tried to define the architecture without these Hadamard squares and the results were not satisfactory. This can be appended as an ablation study for the next revision. Q11: The (positive) ratios of Figure. 4b are an average across all 10000 points. This ratio can be very large (>> 1) for few outliers and just below 1 for the other points. This makes the average of the ratios just above 1 (1.05 for d=1 in Figure 4b) but when we measure the accuracy (number of ratios strictly above 1) we get a negligible number of outliers and thus the accuracy will be close to 0. References: (Nock et al. 2005) Fitting the smallest enclosing Bregman ball. European Conference on Machine Learning, 2005. --- Rebuttal Comment 1.1: Comment: Thank you for your responses and revisions. I maintain my original score. See more discussion below that I hope will help further improve your manuscript. Q1 and Q7: Do you need to assume strong convexity? It is more standard to define Bregman divergences with respect to strictly convex functions, not strongly convex functions (see even your listed reference, Nock et al. 2005). Q8: The Bregman ball, as you defined it in (4), is always convex. As you stated on line 72 of your manuscript, the Bregman divergence $D_h$ is convex in its first argument. Therefore, the set $\mathbb{B}_h(\mathbf{x},\epsilon)$ defined in (4) is a convex set, since it is the $\epsilon$-sublevel set of the convex function $\mathbf{x}' \mapsto D_h(\mathbf{x}',\mathbf{x})$. In your reference Nock et al. 2005, they mention that there are two ways of defining balls from Bregman divergences (equations (3) and (4) in their paper), and they mention that the second way in their equation (4), which coincides with your definition of Bregman ball, leads to a convex set, whereas the first way in their equation (3), does not necessarily lead to convexity. Their example of the nonconvex Itakura-Saito ball is an instantiation of their equation (3), which is different from your definition of Bregman ball (and, correspondingly, their equation (4)). Q9: "here [1, Proposition 1] imposes that the activation and are convex and non-decreasing" ... "we can choose the activations $q^l$ to be the Hadamard power". Technically, [1,Proposition 1] doesn't apply to the expansion of your proposed architecture, since $q^l(\cdot) = (\cdot)^{\circ 2}$ is not nondecreasing (it is decreasing in a given element when that element ranges over negative values). However, it is pretty obvious that your proposed architecture is input-output convex. My original comment was moreso on your choice of using Hadamard square activation functions for the feedthrough terms. In particular, I suggest clearly asserting in your manuscript that your (5) is a particular example of an ICNN where you choose to impose Hadamard square activations on every feedthrough term, in order to clarify to the reader that not all ICNNs can be written in your form. Also, it would be good to discuss why you chose such a parameterization, over the more general ICNNs used in other works. Q11: I see, thanks for the clarification. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for engaging in the rebuttal and for the corrections. Q1 and Q7: Indeed, strict convexity alone is enough to define the divergence but since our ICNNs is $\alpha$-strongly convex (needed later for the conjugate), one can assume strongly convexity from the beginning. Q8: We agree. The Bregman ball as we defined in Eq.4 is what is called the dual Bregman ball which is convex. This convexity explains the success of our projection procedure. > ... since $q^l(\cdot) = (\cdot)^{\circ 2}$ is not nondecreasing (it is decreasing in a given element when that element ranges over negative values). However, it is pretty obvious that your proposed architecture is input-output convex. Indeed, it is always the case that the input range is non-negative due to the CELU activations. > My original comment was moreso on your choice of using Hadamard square activation functions for the feedthrough terms. In particular, I suggest clearly asserting in your manuscript that your (5) is a particular example of an ICNN where you choose to impose Hadamard square activations on every feedthrough term, in order to clarify to the reader that not all ICNNs can be written in your form. Also, it would be good to discuss why you chose such a parameterization, over the more general ICNNs used in other works. We will feature this architectural choice more prominently in the next revision along with an ablation study in the appendix.	Summary: The authors present an approach to learn Bregman divergences that capture perceptual image similarities according to a given dataset. Relying on two input-convex neural networks, they present a procedure that mimics mirror descent over the learned Bregman divergence. The procedure is used to learn networks that are robust to image corruptions. Results on CIFAR-10-C subsets are presented. Strengths: The idea to (attempt to) do mirror descent over learned Bregman divergences in order to train networks robust to corruptions is novel and interesting. Weaknesses: While the idea in itself is interesting, I do not find neither the technical presentation, nor the provided results convincing. Most of the motivation of the work derives from the use of Bergman divergences, which come with an associated mirror descent. However, in practice, I do not think the authors can be claiming to do mirror descent, because of the approximation of the inverse map, and because of the lack of a projection operator. Given the two above limitations, I do not think what the authors do would have convergence guarantees even in the convex case. Taking this into account, the stress on the mathematical motivation of the approach seems to be a bit fragile. Sometimes mathematical concepts are introduced without a clear purpose (for instance, the Legendre type, which is then not really necessary to justify their approximation of the inverse map). I would urge the authors to tone down these claims and refrain from saying they are doing mirror descent: I'd rather call it an approach "inspired by mirror descent". Furthermore, the work assumes that a dataset describing the corruptions is available to learn the divergence: is this a reasonable assumption for datasets such as CIFAR-10-C, which are designed as benchmarks for OOD generalization? Concerning the results: I do not think the proposed comparisons are fair. Both l2 PGD and RLAT use a threat model which is general and not targeted at specific perturbations. As such, they attain good performance over the entirety of the CIFAR-10-C corruptions. The authors, instead, focus on a small set of perturbations and show results for an algorithm that is explicitly aware of the perturbations the network need to be robust against. Technical Quality: 2 Clarity: 3 Questions for Authors: 1) How is the dataset for the training of the Bergman divergence obtained? Is this a holdout from the original CIFAR-10-C? 2) Could the authors add the performance against noise-like perturbations in the comparison against PGD and RLAT? 3) Would the proposed approach scale to ImageNet-C? I do not think scaling to ImageNet-C is necessary, but I think such discussions should be included. 4) The employed ICNN is different from the original work [1] as it displays quadratic terms. I understand that the squared term on x added on top of $z^l$ is useful for strong convexity. What about the quadratic terms in equation (5)? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 1 Limitations: The limitations paragraph is satisfactory. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: The goal is to generate Bregman divergences from learned base functions $\phi$, that are parametrized neural networks. As a result, the gradients of the base functions $\Psi$ and their inverses $\Psi^{-1}$ are also learned and thus approximated by definition. We see your point on calling it mirror descent and will adapt the wording, but, after doing so, we think it is fair to use the name as it closely follows the general structure and despite the absence of convergence guarantees that, to the best of our knowledge, have never been established even for PGD when adopted for adversarial training by the seminal work of [52]. Nevertheless, we believe that the (meanwhile more, see global comment) extensive experimental results show the effectiveness of our method. Finally, we note that our heuristic is mathematically still a projection (idempotent mapping from a set to a subset). > Furthermore, the work assumes that a dataset describing the corruptions is available to learn the divergence: is this a reasonable assumption for datasets such as CIFAR-10-C, which are designed as benchmarks for OOD generalization? > I do not think the proposed comparisons are fair. Both l2 PGD and RLAT use a threat model which is general and not targeted at specific perturbations ... be robust against. To address these two OOD-related concerns, we trained a corruption oblivious Bregman divergence on a different dataset (Berkeley-Adobe Perceptual Patch Similarity) and performed the associated adversarial training on CIFAR-10. The results on CIFAR-10-C outperforms RLAT and PGD especially for the Fog and the Contrast corruptions where PGD and RLAT are known to fail. Please refer to the global rebuttal for details. > How is the dataset for the training of the Bergman divergence obtained? Is this a holdout from the original CIFAR-10-C? The Bregman divergence training algorithm (Sec. 3.2 and Algorithm 1) takes a corruption $\tau$ as an input ($\tau$ is a function that corrupts a given input image). Training is then done on the CIFAR-10 test set, creating pairs, clean and corrupted by $ \tau$. CIFAR-10-C corresponds to the CIFAR-10 test set. Alternatively, it is possible to dircetly train with (clean, corrupted) pairs as we have done in the new experiments on the Berkeley-Adobe Perceptual Patch Similarity dataset (see global rebuttal for details). > Could the authors add the performance against noise-like perturbations in the comparison against PGD and RLAT? The accuracy for noise categories such as Gaussian is low (52.41\%) which a direct consequence of training algorithm: The core idea is to train the Bregman divergence to consider those noisy images far and thus excluded from the Bregman ball and finally not covered during adversarial training. > Would the proposed approach scale to ImageNet-C? I do not think scaling to ImageNet-C is necessary, but I think such discussions should be included. The Bregman divergence (despite being trained on CIFAR-10 32x32 images) performs well for 256x256 ImageNet images as shown in the rebuttal PDF. Unfortunately, the adversarial training on ImageNet to evaluate on ImageNet-C is beyond the computational resources we have access to (one V100 GPU). > The employed ICNN is different from the original work [1] as it displays quadratic terms. I understand that the squared term on x added on top of is useful for strong convexity. What about the quadratic terms in equation (5)? The equation (5) may look a bit different the equations in [1] but it is a special case that conforms to the conditions of the Proposition 1 in [1]. The equation (5) can be expanded to: $u^1 = q^0\left[W^0x + 0 \right]$ $z^1 = g^0\left[U^1u^1 + V^0x + b^0 \right]$ $u^l = q^{l-1}\left[W^{l-1}x + 0 \right]$ $z^{l} = g^{l-1} \left[U^lu^{l} + V^{l-1}z^{l-1} + b^{l-1} \right]$ here [1, Proposition 1] imposes that the activation $q^l$ and $g^l$ are convex and non-decreasing and the weights $V^1, .., V^{L−1}$ and $U^1, .., U^{L−1}$ are non-negative. In particular, we can choose the activations $q^l$ to be the Hadamard power $()^{\circ 2}$ and the weights $U^1, .., U^{L−1}$ to be the identity matrix to obtain exactly the equation (5). --- Rebuttal Comment 1.1: Comment: I thank the authors for their response. I am happy to see the new experiments showing results on CIFAR-10-C with the divergence learned from BAPPS. However, the corruptions in the comparison table (fog, contrast, zoom blur) appear to be even more cherry-picked in this context where the divergence is learned from another dataset. Given that the baselines (PGD and RLAT) typically improve OOD generalisation over a wide range of corruptions, the comparison still feels quite unfair, casting doubts over the overall empirical effectiveness of the proposed approach. In this sense, the marked decrease in the accuracy to noise-based corruptions should be concerning for an approach presented as a tool to increase general robustness. Finally, the authors state that their addition of the quadratic terms to ICNNs was empirically useful. This detail, along with a comprehensive explanation, should have been prominently featured in the submission. --- Reply to Comment 1.1.1: Comment: We are happy that the new experiments (corruption-oblivious Bregman divergence and the associated Mirror Descent adversarial training) addressed the OOD concerns raised earlier and here we answer the new concerns. Before, we want to emphasize that we consider the learning of a meaningful, consistent, applicable Bregman divergence (as confirmed by the experiments and the prototypical and successful application to robustness) a main contribution of this paper and we would imagine that the community could use our pipeline in various settings where non-Euclidean similarity is useful. The new results suggested by you and another reviewer further strengthen our work. > Given that the baselines (PGD and RLAT) typically improve OOD generalisation over a wide range of corruptions ... We are particularly focusing here to present the results on fog and contrast corruptions only because these corruptions are problematic for both PGD and RLAT: From (Kireev et al., 2022): > Interestingly, for the fog and contrast corruptions, the performance degrades for all methods (see Table 10 in App. H), consistently with the observation made in Ford et al. (2019). From (Ford et al., 2019): > Interestingly, both methods performed much worse than the clean model on the fog and contrast corruptions. For example, the adversarially trained model was 55% accurate on the most severe contrast corruption compared to 85% for the clean model. We have shown that our method (trained on a different dataset and oblivious to the corruptions) excels particularly for these problematic corruptions, the improvement in accuracy is about 27%. Also, our method achieves comparable performance on other non-problematic corruptions such zoom blur. This is a notable strength of our method. > In this sense, the marked decrease in the accuracy to noise-based corruptions should be concerning for an approach presented as a tool to increase general robustness. We respectfully disagree. Robustness against noise is, in this case, neither very important $i)$ nor expected $ii)$: * $i)$ Robustness against real-world corruptions such as contrast is more important than robustness against Gaussian noise (that is equivalent to robustness on $L^2$ balls), e.g.: > Goodfellow, Shlens, and Szegedy intended $l_p$ adversarial examples to be a toy problem where evaluation would be easy, with the hope that the solution to this toy problem would generalize to other problems. (Gilmer et al., 2018) * $ii)$ We train the Bregman divergence by forcing it to consider noisy images far from the original (divergence values for noisy images are in 10000 order of magnitude while divergence to corrupted images are typically below 100). So when we pick a Bregman ball radius (of 100) for adversarial training, the noisy images (gaussian noise) will be automatically excluded. Thus, model trained to be robust in these Bregman balls are not even expected to be robust on noisy images. > Finally, the authors state that their addition of the quadratic terms to ICNNs was empirically useful. This detail, along with a comprehensive explanation, should have been prominently featured in the submission. These quadratic terms are not a result of random tuning. They are a choice of activation functions obeying the conditions of the ICNN. In fact, the quadratic function is the simplest choice that obeys them (apart from the identity). We will make this clearer in the next revision and add an ablation study in the appendix.	Summary: The authors propose a new method to learn Bregman divergences from raw, high-dimensional data. This method measures similarity between images in pixel space, and considers two images as similar even if one image is corrupted by real-world corruptions, such as blur, changes in contrast, or weather conditions such as fog. The method does this in-part by simultaneously considering real-world corruptions as close to the original image, while noisy perturbations as far from the original image, even when the $L^p$ distance considers noisy perturbations as close. The authors then define adversarial attacks by replacing the projected gradient descent with mirror descent using the learned Bregman divergence. Through adversarial training on this new learned Bregman divergence, they improve the state-of-the-art in robustness. Strengths: - The authors clearly explain the pipeline of the algorithm and give great explanations for the choices they made (e.g. using equation (7) to approximate $\nabla \bar{\phi}$.) - The authors make a good case for using Bregman divergences and learning the metric, and it seems like an interesting direction. - The algorithm seems well-motivated at each step of the pipeline, and it looks like the authors took care to make sure each step follows theory. Figures 1, 2, and 3 are helpful in explaining the motivation. Weaknesses: - A big weakness is using only one dataset for comparison. Perhaps the authors could show more experiments on ImageNet-C, and/or the Berkeley-Adobe Perceptual Patch Similarity (BAPPS) that was introduced with one of the methods the authors compared against, LPIPS. - On the above note, there are a lot of parts to the algorithm, and it's unclear how hard one has to tune the algorithm to make sure each approximation lines up to get an overall, well-performing model. I would have wanted to see a stress test on the pipeline with larger images. - If the authors had comparisons on more datasets and of more difficulty than CIFAR10-C, then I would be inclined to raise the score to an accept. - The proposed method takes longer to train than other standard adversarial training methods, as mentioned in the appendix. EDIT: After considering the author responses and reading all the reviewers, I have raised my score from a 4 to a 5. Technical Quality: 4 Clarity: 4 Questions for Authors: - Can we see some examples of noisy images that are considered not semantically the same as the original? Under some noise threshold, it seems reasonable for a human to classify a noisy image as the original label, right? - In Table 3, when using the proposed method, training for one corruption doesn't necessarily perform the best for that corruption (as acknowledged by the authors in the paper). For example, $\text{MD} \thinspace D_{\phi}^{\text{contrast}}$ does not perform the best on contrast corruptions, but rather $\text{MD} \thinspace D_{\phi}^{\text{zoom-blur}}$ does the best. Do you have an explanation for that? I would have liked to see this explored and explained, as it goes against intuition. - There are a lot of approximations, like computing the inverse map. How do you expect your algorithm to perform with higher resolution images? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 2 Limitations: The authors appropriately acknowledge limitations where appropriate. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for acknowledging the "great explanations" of the choices we made (especially the approximation of the conjugate $\nabla \overline \phi$), for finding the use of the Bregman divergences for metric learning "interesting direction", and for acknowledging that each step of the pipeline is "well-motivated". We have performed the suggested experiments, and we answer the other concerns below: > A big weakness is using only one dataset for comparison. Perhaps the authors could show more experiments on ImageNet-C, and/or the Berkeley-Adobe Perceptual Patch Similarity (BAPPS) that was introduced with one of the methods the authors compared against, LPIPS. Done, we have done two new experiments on BAPPS and evaluation on 256x256 ImageNet images. Please refer to the global rebuttal. > How do you expect your algorithm to perform with higher resolution images? The performance on higher resolution (up to 256x256) matches that presented for lower resolution (32x32) as shown by the new experiments on BAPPS and ImageNet. The Bregman divergence successfully distinguishes corrupted from noisy 256x256 ImageNet images despite being trained on 32x32 CIFAR-10 images. The inference is possible because the convolutional layers are followed by an average pooling to size 1x1 before the final feedforward layers. > there are a lot of parts to the algorithm, and it's unclear how hard one has to tune the algorithm to make sure each approximation lines up to get an overall, well-performing model. The part that needs a tuning effort is the architecture of the base function $\phi$ which is typical for deep learning pipelines. The other parts are fairly standard such as the widely used Adam optimizer. > The proposed method takes longer to train than other standard adversarial training methods, as mentioned in the appendix. Yes, it requires roughly twice the runtime as the standard AT when implemented in PyTorch and run on a single V100 GPU. > Can we see some examples of noisy images that are considered not semantically the same as the original? Under some noise threshold, it seems reasonable for a human to classify a noisy image as the original label, right? Please refer to the rebuttal PDF. Those are 256x256 ImageNet clean samples in the first column, corrupted versions in the second column, and noisy versions (with different noise thresholds) thereafter. Unlike the $L^2$, the trained Bregman divergence considers the corrupted version closer than the noisy versions (as extensively evaluated in the plot and table in Figure 4 of the paper). > In Table 3, when using the proposed method, training for one corruption doesn't necessarily perform the best for that corruption (as acknowledged by the authors in the paper). [...] Do you have an explanation for that? I would have liked to see this explored and explained, as it goes against intuition. This behavior is investigated and to some degree explained in the paragraph "Cross-corruption generalization" of Sec.7 with further experiments reported in Tab.4. For example, the fact that classification models trained with mirror descent under the divergence $D_\phi^{zoom-blur}$ are performing well on contrast can be traced back to the fact that the divergence $D_\phi^{zoom-blur}$ does consider Contrast corruptions close the original images. Why this is the case is not clear but one reason may be that fog/blur/zoom-blur are somewhat similar corruptions. > There are a lot of approximations, like computing the inverse map Our approach includes two heuristics: Approximate inverse map. This one we consider strong. The approximation is not only theoretically principled but also empirically precise. It is based on the Fenchel conjugate (Fenchel, 1949) as detailed in Section 4. The resulting approximation has satisfactory empirical precision: the mean square error between an image $\boldsymbol{x}$ and its reconstruction $\overline\Psi(\Psi(\boldsymbol{x}))$ (after undergoing the map $\Psi$ and the inverse map $\overline\Psi$) is less than 0.001 on the test set. Approximate projection. Indeed, the exact projection to the closest point with respect to Bregman divergence is known to be an open problem (Dhillon & Tropp, 2008) and our approach would immediately benefit from any progress, exact or heuristic. However, our proposed heuristic is successful: * it yields valid results (inside the Bregman ball), * it results in images with high semantic similarity (as illustrated in Figures 7 and 8), * it is fast enough to be incorporated in training, * it provides good results in the down-stream task (an accuracy improvement up to 27.16%), * mathematically, it is still a projection (idempotent mapping from a set to a subset). --- Rebuttal Comment 1.1: Comment: I thank the authors very much for their hard work in addressing my concerns and in the additional experiments. I just have a couple more questions. - Can you provide more context on the rebuttal PDF, such as procedures used to generate those images, where the images came from, etc.? I'm assuming they came from BAPPS? - In the rebuttal pdf, I see that in the 4th row (image of a parrot's right face), under the proposed metric $D_\phi$ the black-and-white image is much closer to the original image than the image in the 3rd column (slight noise). But, in my opinion, most humans would make the the opposite judgement. How would you address such concerns, as I imagine the metric $D_\phi$ is meant to be more "semantic" than the simple L2 distance? Thank you again in working to address my concerns. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for engaging in the rebuttal. We are happy to answer the new questions: > Can you provide more context on the rebuttal PDF, such as procedures used to generate those images, where the images came from, etc.? I'm assuming they came from BAPPS? Those are 256x256 images from the ImageNet validation set. The first column is for clean samples. The second column are corrupted versions of clean images. The chosen corruption here is Gray color conversion. Thereafter, we generate noisy images (based on the clean image) according to Equ.8 of the paper. For each image (corrupted or noisy) we report the $L^2$ distance to its clean image along with the Bregman divergence $D_\phi$. This Bregman divergence is trained on 32x32 CIFAR-10 training set for the corruption $\tau$ as Gray color conversion following the scheme detailed in Sec. 3.2 and Algorithm 1. > In the rebuttal pdf, I see that in the 4th row (image of a parrot's right face), under the proposed metric $D_\phi$ the black-and-white image is much closer to the original image than the image in the 3rd column (slight noise). But, in my opinion, most humans would make the the opposite judgement. How would you address such concerns, as I imagine the metric $D_\phi$ is meant to be more "semantic" than the simple L2 distance? We train the Bregman divergence by forcing it to consider corrupted images (gray images in the second column) close to the clean images, and the noisy images as far. The threshold to generate noisy images is the adjustable hyperparameter, $d>0$ in Equ.6. For this particular experiment the noise threshold $d$ is chosen to be relatively low: randomly sampled from $[10^{-7}, 0.99 ]$ (see paragraph "Training details" of Sec. 6 for more details). The noise parameter $d$ corresponding to the image in the fourth row, third column is $d\approx0.4$ (which is the $L^2$ of this noisy image divided by the $L^2$ of the corrupted image in the second column). This value of $d$ falls into what we forced the divergence to consider far from the image ($10^{-7}<d<0.99$) so the divergence is behaving as intended. We have chosen these low values of $d$ because they gave the best result once the divergence is used for Mirror Descent adversarial training. Nevertheless, in another use-case, the threshold noise $d$ can be increased (for example above 1), so the divergence will consider even images in the seventh column close to the original image if that is what the user wants.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their comments. We responded to each concern in detail in our individual responses. These discussions and also corrections will be incorporated in the next revision. We have strengthened the results of our work by performing an evaluation on the BAPPS dataset (suggested by Reviewer DULG) and further trained a corruption-oblivious Bregman divergence to address the Out-Of-Distribution concerns raised by Reviewer AVtS. These experiments are detailed below: A. Performance on BAPPS. We applied Bregman learning to the Berkeley-Adobe Perceptual Patch Similarity (BAPPS) dataset (Zhang et al. 2018). Particularly, we considered the two alternative forced choice (2AFC) test. The dataset is a collection of image triplets (reference, distortion 1, distortion 2) and a human judgment stating which of distortions is similar to the reference (so no classification labels). The images and the distortions are diverse from 6 categories. We train our Bregman divergence to mimic the human judgement, more similar distortion closer to the reference. Since we are not using additional data, we compare against the VGG version of LPIPS that does not use ImageNet pre-training. The training pipeline (loss, optimizers, batch size, ... etc.) is similar to that in (Zhang et al. 2018). We report the accuracy results on the 6 test categories of 2AFC below. Our method outperformed LPIPS in all categories except for Frame Interp: \| \| Traditional \| CNN-based \| Superres \| Video Deblur \| Colorization \| Frame Interp \| \| :--- \| :----: \| :----: \| :----: \|:----: \|:----: \|:----: \| \| LPIPS \| 51.41 \| 72.10 \| 60.46 \| 54.25 \| 55.18 \| 55.55 \| \| Bregman (ours) \| 63.65 \| 79.57 \| 61.04 \| 56.95 \| 61.63 \| 53.73 \| B. The corruption-oblivious Bregman divergence. Inspired by the BAPPS experiments, we extended our robustness pipeline and experiment to be oblivious to the corruptions in CIFAR-10-C. Namely, we trained a Bregman divergence on the entire BAPPS that considers distorted versions close to the original and noisy far, exactly following our scheme described in Sec. 3.2 and Algorithm 1 except that the corrupted image is not generated by a corruption $\tau$ but rather it is one of the distortions from BAPPS. BAPPS considers a few dozens of distortions and the divergence is thus not corruption specific, and oblivious to the corruptions in CIFAR-10-C. We then re-executed adversarial training on CIFAR-10 with this divergence. The results on CIFAR-10-C show again that our method outperforms RLAT and PGD especially for the Fog and the Contrast corruptions where PGD and RLAT are known to fail (e.g., [25] and [41]): \| \| Clean \| Fog \| Contrast \| Zoom Blur \| \| :--- \| :----: \| :----: \| :----: \|:----: \| \| PGD [41] \| 93.65 \| 77.18 \| 63.19 \| 86.08 \| \| RLAT [41] \| 93.28 \| 77.01 \| 62.87 \| 85.89 \| \| Bregman (ours) \| 93.61 \| 88.00 \| 77.70 \| 87.12 \| References: (Zhang et al. 2018) The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR 2018. Pdf: /pdf/9a2a4d582aae6f87496d09a8a6f35fb898f97554.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.