Datasets:

yxzwayne

/

abs

like 0

2203.14263.pdf

1 A General Survey on Attention Mechanisms in Deep Learning Gianni Brauwers and Flavius Frasincar Abstract —Attention is an important mechanism that can be employed for a variety of deep learning models across many different domains and tasks. This survey provides an overview of the most important attention mechanisms proposed in the literature. The various attention mechanisms are explained by means of a framework consisting of a general attention model, uniform notation, and a comprehensive taxonomy of attention mechanisms. Furthermore, the various measures for evaluating attention models are reviewed, and methods to characterize the structure of attention models based on the proposed framework are discussed. Last, future work in the field of attention models is considered. Index Terms —Attention models, deep learning, introductory and survey, neural nets, supervised learning ! 1 I NTRODUCTION THEidea of mimicking human attention first arose in the field of computer vision [1], [2] in an attempt to reduce the computational complexity of image processing while improving performance by introducing a model that would only focus on specific regions of images instead of the entire picture. Although, the true starting point of the attention mechanisms we know today is often attributed to originate in the field of natural language processing [3]. Bahdanau et al. [3] implement attention in a machine translation model to address certain issues with the structure of recurrent neural networks. After Bahdanau et al. [3] emphasized the advan- tages of attention, the attention techniques were refined [4] and quickly became popular for a variety of tasks, such as text classification [5], [6], image captioning [7], [8], sentiment analysis [6], [9], and speech recognition [10], [11], [12]. Attention has become a popular technique in deep learn- ing for several reasons. Firstly, models that incorporate attention mechanisms attain state-of-the-art results for all of the previously mentioned tasks, and many others. Fur- thermore, most attention mechanisms can be trained jointly with a base model, such as a recurrent neural network or a convolutional neural network using regular backpropa- gation [3]. Additionally, attention introduces a certain type of interpretation into neural network models [8] that are generally known to be highly complicated to interpret. Moreover, the popularity of attention mechanisms was ad- ditionally boosted after the introduction of the Transformer model [13] that further proved how effective attention can be. Attention was originally introduced as an extension to recurrent neural networks [14]. However, the Transformer model proposed in [13] poses a major development in at- tention research as it demonstrates that the attention mech- anism is sufficient to build a state-of-the-art model. This means that disadvantages, such as the fact that recurrent neural networks are particularly difficult to parallelize, can G. Brauwers and F. Frasincar are with the Erasmus School of Economics, Erasmus University Rotterdam, 3000 DR, Rotterdam, the Netherlands (e- mail:{frasincar, brauwers}@ese.eur.nl). Manuscript received July 6, 2020; revised June 21, 2021; Corresponding author: F. Frasincarbe circumvented. As was the case for the introduction of the original attention mechanism [3], the Transformer model was created for machine translation, but was quickly adopted to be used for other tasks, such as image processing [15], video processing [16], and recommender systems [17]. The purpose of this survey is to explain the general form of attention, and provide a comprehensive overview of attention techniques in deep learning. Other surveys have already been published on the subject of attention models. For example, in [18], a survey is presented on attention in computer vision, [19] provides an overview of attention in graph models, and [20], [21], [22] are all surveys on attention in natural language processing. This paper partly builds on the information presented in the previously mentioned surveys. Yet, we provide our own significant contributions. The main difference between this survey and the previously mentioned ones is that the other surveys generally focus on attention models within a certain domain. This survey, however, provides a cross-domain overview of attention techniques. We discuss the attention techniques in a general way, allowing them to be understood and applied in a variety of domains. Furthermore, we found the taxonomies presented in previous surveys to be lacking the depth and structure needed to properly distinguish the various atten- tion mechanisms. Additionally, certain significant attention techniques have not yet been properly discussed in previ- ous surveys, while other presented attention mechanisms seem to be lacking either technical details or intuitive ex- planations. Therefore, in this paper, we present important attention techniques by means of a single framework using a uniform notation, a combination of both technical and in- tuitive explanations for each presented attention technique, and a comprehensive taxonomy of attention mechanisms. The structure of this paper is as follows. Section 2 in- troduces a general attention model that provides the reader with a basic understanding of the properties of attention and how it can be applied. One of the main contributions of this paper is the taxonomy of attention techniques pre- sented in Section 3. In this section, attention mechanisms are explained and categorized according to the presentedarXiv:2203.14263v1 [cs.LG] 27 Mar 2022

2210.00312.pdf

Published as a conference paper at ICLR 2023 MULTIMODAL ANALOGICAL REASONING OVER KNOWLEDGE GRAPHS Ningyu Zhang1∗Lei Li1∗Xiang Chen1∗Xiaozhuan Liang1Shumin Deng2Huajun Chen1† 1Zhejiang University, AZFT Joint Lab for Knowledge Engine 2National University of Singapore {zhangningyu,leili21,xiang chen,liangxiaozhuan,231sm,huajunsir }@zju.edu.cn ABSTRACT Analogical reasoning is fundamental to human cognition and holds an important place in various fields. However, previous studies mainly focus on single-modal analogical reasoning and ignore taking advantage of structure knowledge. No- tably, the research in cognitive psychology has demonstrated that information from multimodal sources always brings more powerful cognitive transfer than single modality sources. To this end, we introduce the new task of multimodal analogical reasoning over knowledge graphs, which requires multimodal reason- ing ability with the help of background knowledge. Specifically, we construct aMultimodal Analogical Reasoning data Set (MARS ) and a multimodal knowl- edge graph MarKG . We evaluate with multimodal knowledge graph embedding and pre-trained Transformer baselines, illustrating the potential challenges of the proposed task. We further propose a novel model-agnostic Multimodal analogical reasoning framework with Transformer ( MarT ) motivated by the structure map- ping theory, which can obtain better performance. We hope our work can deliver benefits and inspire future research1. 1 I NTRODUCTION Analogical reasoning – the ability to perceive and use relational similarity between two situations or events – holds an important place in human cognition (Johnson-Laird, 2006; Wu et al., 2020; Bengio et al., 2021; Chen et al., 2022a) and can provide back-end support for various fields such as education (Thagard, 1992), creativity (Goel, 1997), thus appealing to the AI community. Early, Mikolov et al. (2013b); Gladkova et al. (2016a); Ethayarajh et al. (2019a) propose visual analogical reasoning aiming at lifting machine intelligence in Computer Vision (CV) by associating vision with relational, structural, and analogical reasoning. Meanwhile, researchers of Natural Language Processing (NLP) hold the connectionist assumption (Gentner, 1983) of linear analogy (Ethayarajh et al., 2019b); for example, the relation between two words can be inferred through vector arithmetic of word embeddings. However, it is still an open question whether artificial neural networks are also capable of recognizing analogies among different modalities. Note that humans can quickly acquire new abilities based on finding a common relational system between two exemplars, situations, or domains. Based on Mayer’s Cognitive Theory of multimedia learning (Hegarty & Just, 1993; Mayer, 2002), human learners often perform better on tests with analogy when they have learned from multimodal sources than single-modal sources. Evolving from recognizing single-modal analogies to exploring multimodal reasoning for neural models, we emphasize the importance of a new kind of analogical reasoning task with Knowledge Graphs (KGs). In this paper, we introduce the task of multimodal analogical reasoning over knowledge graphs to fill this blank. Unlike the previous multiple-choice QA setting, we directly predict the analogical target and formulate the task as link prediction without explicitly providing relations . Specifically, the task can be formalized as (eh,et) : (eq,?)with the help of background multimodal knowledge graph ∗Equal contribution and shared co-first authorship. †Corresponding author. 1Code and datasets are available in https://github.com/zjunlp/MKG_Analogy . 1arXiv:2210.00312v4 [cs.CL] 1 Mar 2023

2310.12397.pdf

GPT-4 Doesn’t Know It’s Wrong: An Analysis of Iterative Prompting for Reasoning Problems Kaya Stechly∗Matthew Marquez∗Subbarao Kambhampati∗ Abstract There has been considerable divergence of opinion on the reasoning abilities of Large Language Models (LLMs). While the initial optimism that reasoning might emerge automatically with scale has been tempered thanks to a slew of counterexamples–ranging from multiplication to simple planning, there is still the wide spread belief that LLMs can self-critique and improve their own solutions in an iterative fashion. This belief seemingly rests on the assumption that verification of correctness should be easier than generation–a rather classical argument from computational complexity, that should be irrelevant to LLMs to the extent what they are doing is approximate retrieval. In this paper, we set out to systematically investigate the effectiveness of iterative prompting of LLMs in the context of Graph Coloring , a canonical NP-complete reasoning problem that is related to proposi- tional satisfiability as well as practical problems like scheduling and allocation. We present a principled empirical study of the performance of GPT4 in solving graph coloring instances or verifying the correctness of candidate colorings–both in direct and iterative modes. In iterative modes, we experiment both with the model critiquing its own answers and an external correct reasoner verifying proposed solutions. In both cases, we analyze whether the content of the criticisms actually affects bottom line performance. The study seems to indicate that (i) LLMs are bad at solving graph coloring instances (ii) they are no better at verifying a solution–and thus are not effective in iterative modes with LLMs critiquing LLM-generated solutions (iii) the correctness and content of the criticisms–whether by LLMs or external solvers–seems largely irrelevant to the performance of iterative prompting. We show that the observed effectiveness of LLMs in iterative settings is largely due to the correct solution being fortuitously present in the top-k completions of the prompt (and being recognized as such by an external verifier). Our results thus call into question claims about the self-critiquing capabilities of state of the art LLMs. 1 Introduction Large Language Models (LLMs), essentially n-gram models on steroids which have been trained on web-scale language corpus, have caught the imagination of the AI research community with linguistic behaviors that no one expected text completion systems to possess. Their seeming versatility has lead many researchers to wonder whether they can also do well on reasoning tasks typically associated with system 2 competency. Initial excitement based on anecdotal performance of LLMs on reasoning tasks has dissipated to some extent by the recent spate of studies questioning the robustness of such behaviors–be it planning [ 17,8], simple arithmetic and logic [ 5], or general mathematical and abstract benchmark[ 14,6]. There still exists considerable optimism that even if LLMs can’t generate correct solutions in one go, their accuracy improves in a iterative prompting regime, where LLMs will be able to "self-critique" their candidate solutions and refine them to the point of correctness [20,19,15,18,7]. This belief seem to rest largely on the assumption that verification of correctness ∗Arizona State University, Tempe. Preprint. Under review.arXiv:2310.12397v1 [cs.AI] 19 Oct 2023

2309.14322.pdf

Small-scale proxies for large-scale Transformer training instabilities Mitchell Wortsman Peter J. Liu Lechao Xiao Katie Everett Alex Alemi Ben Adlam John D. Co-Reyes Izzeddin Gur Abhishek Kumar Roman Novak Jeffrey Pennington Jascha Sohl-dickstein Kelvin Xu Jaehoon Lee*Justin Gilmer*Simon Kornblith* Google DeepMind Abstract Teams that have trained large Transformer-based mod- els have reported training instabilities at large scale that did not appear when training with the same hyperparameters at smaller scales. Although the causes of such instabilities are of scientific interest, the amount of resources required to reproduce them has made investigation difficult. In this work, we seek ways to reproduce and study training stability and instability at smaller scales. First, we focus on two sources of training instability described in pre- vious work: the growth of logits in attention layers (Dehghani et al., 2023) and divergence of the output logits from the log probabilities (Chowdhery et al., 2022). By measuring the relationship between learn- ing rate and loss across scales, we show that these instabilities also appear in small models when training at high learning rates, and that mitigations previously employed at large scales are equally effective in this regime. This prompts us to investigate the extent to which other known optimizer and model interventions influence the sensitivity of the final loss to changes in the learning rate. To this end, we study meth- ods such as warm-up, weight decay, and the µParam (Yang et al., 2022), and combine techniques to train small models that achieve similar losses across orders of magnitude of learning rate variation. Finally, to conclude our exploration we study two cases where instabilities can be predicted before they emerge by examining the scaling behavior of model activation and gradient norms. 1 Introduction Scaling up transformers has led to remarkable progress from chat models to image generation. However, not 104 103 102 101 100 Learning rate2.502.753.003.253.503.754.004.25Final eval loss qk-layernorm = True qk-layernorm = FalseN = 2.4e+06 N = 9.4e+06 N = 1.9e+07 N = 4.2e+07 N = 8.5e+07 N = 1.5e+08 N = 3.0e+08 N = 1.2e+09 107108109 Number of parameters102 101 100LR sensitivityFigure 1: Qk-layernorm [ 11] enables stable training across three orders of magnitude of learning rate (LR) variation. (Top) For transformers with Nparameters, we plot the effect of learning rate on final evaluation loss. (Bottom) We use LR sensitivity to summarize the top plot. LR sensi- tivity measures the expected deviation from optimal when varying learning rate across three orders of magnitude. Qk-layernorm reduces LR sensitivity, but LR sensitivity still increases with model scale. 1arXiv:2309.14322v1 [cs.LG] 25 Sep 2023

2308.05660.pdf

Thermodynamic Linear Algebra Maxwell Aifer, Kaelan Donatella, Max Hunter Gordon, Thomas Ahle, Daniel Simpson, Gavin Crooks, Patrick J. Coles Normal Computing Corporation, New York, New York, USA Linear algebraic primitives are at the core of many modern algorithms in engineering, science, and machine learning. Hence, accelerating these primitives with novel computing hardware would have tremendous economic impact. Quantum computing has been proposed for this purpose, although the resource requirements are far beyond current technological capabilities, so this approach remains long-term in timescale. Here we consider an alternative physics-based computing paradigm based on classical thermodynamics, to provide a near-term approach to accelerating linear algebra. At first sight, thermodynamics and linear algebra seem to be unrelated fields. In this work, we connect solving linear algebra problems to sampling from the thermodynamic equilibrium distri- bution of a system of coupled harmonic oscillators. We present simple thermodynamic algorithms for (1) solving linear systems of equations, (2) computing matrix inverses, (3) computing matrix determinants, and (4) solving Lyapunov equations. Under reasonable assumptions, we rigorously establish asymptotic speedups for our algorithms, relative to digital methods, that scale linearly in matrix dimension. Our algorithms exploit thermodynamic principles like ergodicity, entropy, and equilibration, highlighting the deep connection between these two seemingly distinct fields, and opening up algebraic applications for thermodynamic computing hardware. I. Introduction Basic linear algebra primitives such as solving a linear system of the form Ax=band obtaining the inverse of a matrix are present in many modern algorithms. Such primitives are relevant to a multitude of applications, including for example optimal control of dynamic systems and resource allocation. They are also a common subroutine of many artificial intelligence (AI) algorithms, and account for a substantial portion of the time and energy costs in some cases. The most common method to perform these primitives is LU decomposition, whose time-complexity scales as O(d3). Many proposals have been made to accelerate such primitives, for example using iterative methods such as the conjugate gradient method. In the last decade, these primitives have been accelerated by hardware improvements, notably by their implementation on graphical processing units (GPUs), fueling massive parallelization. However, the scaling of these methods is still a prohibitive factor, and obtaining a good approximate solution to a dense matrix of more than a few tens of thousand dimensions remains challenging. Exploiting physics to solve mathematical problems is a deep idea, with much focus on solving optimization problems [1–3]. In the context of linear algebra, much attention has been paid to quantum computers [4], since the mathematics of discrete-variable quantum mechanics matches that of linear algebra. A quantum algorithm [5] to solve linear systems has been proposed, which for sparse and well-conditioned matrices scales as logd. However, the resource requirements [6] for this algorithm are far beyond current hardware capabilities. More generally building large-scale quantum hardware has remained difficult [7], and variational quantum algorithms for linear algebra [8–10] have battled with vanishing gradient issues [11–13]. Therefore, the search for alternative hardware proposals that can exploit physical dynamics to accelerate linear algebra primitives has been ongoing. Notably, memristor crossbar arrays have been of interest for accelerating matrix-vector multiplications [14, 15]. Solving linear systems has also been the subject of analog computing approaches [16]. Recently, we defined a new class of hardware, built from stochastic, analog building blocks, which is ultimately thermodynamic in nature [17]. (See also probabilistic-bit computers [18–20] and thermodynamic neural networks [21–24] for alternative approaches to thermodynamic computing [25]). AI applications like generative modeling are a natural fit for this thermodynamic hardware, where stochastic fluctuations are exploited to generate novel samples. In this work, we surprisingly show that the same thermodynamic hardware from Ref. [17] can also be used toacceleratekeyprimitivesinlinearalgebra. Thermodynamicsisnottypicallyassociatedwithlinearalgebra, and connecting these two fields is therefore non-trivial. Here, we exploit the fact that the mathematics of harmonic oscillator systems is inherently affine (i.e., linear), and hence we can map linear algebraic primitives onto such systems. (See also Ref. [26] for a discussion of harmonic oscillators in the context of quantum computingspeedups.) Weshowthatsimplybysamplingfromthethermalequilibriumdistributionofcoupled harmonic oscillators, one can solve a variety of linear algebra problems.arXiv:2308.05660v1 [cond-mat.stat-mech] 10 Aug 2023

2309.10150.pdf

Q-Transformer: Scalable Offline Reinforcement Learning via Autoregressive Q-Functions Yevgen Chebotar∗, Quan Vuong∗, Alex Irpan, Karol Hausman, Fei Xia, Yao Lu, Aviral Kumar, Tianhe Yu, Alexander Herzog, Karl Pertsch, Keerthana Gopalakrishnan, Julian Ibarz, Ofir Nachum, Sumedh Sontakke, Grecia Salazar, Huong T Tran, Jodilyn Peralta, Clayton Tan, Deeksha Manjunath, Jaspiar Singht, Brianna Zitkovich, Tomas Jackson, Kanishka Rao, Chelsea Finn, Sergey Levine Google DeepMind Abstract: In this work, we present a scalable reinforcement learning method for training multi-task policies from large offline datasets that can leverage both hu- man demonstrations and autonomously collected data. Our method uses a Trans- former to provide a scalable representation for Q-functions trained via offline tem- poral difference backups. We therefore refer to the method as Q-Transformer. By discretizing each action dimension and representing the Q-value of each ac- tion dimension as separate tokens, we can apply effective high-capacity sequence modeling techniques for Q-learning. We present several design decisions that en- able good performance with offline RL training, and show that Q-Transformer outperforms prior offline RL algorithms and imitation learning techniques on a large diverse real-world robotic manipulation task suite. The project’s website and videos can be found at qtransformer.github.io 1 Introduction Human demonstrationsAutonomousdata Conservative regularizationAutoregressive Q-learningMonte-Carlo returnsMixed quality data environment stepaction dimension ……Q-values per action dimensionQ-Transformer Figure 1: Q-Transformer enables training high- capacity sequential architectures on mixed qual- ity data. Our policies are able to improve upon human demonstrations and execute a variety of manipulation tasks in the real world.Robotic learning methods that incorporate large and diverse datasets in combination with high- capacity expressive models, such as Transform- ers [1, 2, 3, 4, 5, 6], have the potential to acquire generalizable and broadly applicable policies that perform well on a wide variety of tasks [1, 2]. For example, these policies can follow natural language instructions [4, 7], perform multi-stage behaviors [8, 9], and generalize broadly across environments, objects, and even robot morpholo- gies [10, 3]. However, many of the recently pro- posed high-capacity models in the robotic learn- ing literature are trained with supervised learn- ing methods. As such, the performance of the re- sulting policy is limited by the degree to which human demonstrators can provide high-quality demonstration data. This is limiting for two rea- sons. First, we would like robotic systems that aremore proficient than human teleoperators, ex- ploiting the full potential of the hardware to per- form tasks quickly, fluently, and reliably. Second, we would like robotic systems that get better with autonomously gathered experience, rather than relying entirely on high-quality demonstrations. Reinforcement learning in principle provides both of these capabilities. A number of promising recent advances demonstrate the successes of large-scale robotic RL in varied settings, such as robotic grasping and stacking [11, 12], learning heterogeneous tasks with human-specified rewards [13], learning multi-task policies [14, 15], learn- ing goal-conditioned policies [16, 17, 18, 19], and robotic navigation [20, 21, 22, 23, 24]. However, ∗Equal contribution. Corresponding emails: chebotar@google.com, quanhovuong@google.com . 7th Conference on Robot Learning (CoRL 2023), Atlanta, USA.arXiv:2309.10150v2 [cs.RO] 17 Oct 2023

2109.01652.pdf

Published as a conference paper at ICLR 2022 FINETUNED LANGUAGE MODELS AREZERO-SHOT LEARNERS Jason Wei∗, Maarten Bosma∗, Vincent Y. Zhao∗, Kelvin Guu∗, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V . Le Google Research ABSTRACT This paper explores a simple method for improving the zero-shot learning abilities of language models. We show that instruction tuning —finetuning language models on a collection of datasets described via instructions—substantially improves zero- shot performance on unseen tasks. We take a 137B parameter pretrained language model and instruction tune it on over 60 NLP datasets verbalized via natural language instruction templates. We evaluate this instruction-tuned model, which we call FLAN, on unseen task types. FLAN substantially improves the performance of its unmodified counterpart and surpasses zero-shot 175B GPT-3 on 20 of 25 datasets that we evaluate. FLAN even outperforms few-shot GPT-3 by a large margin on ANLI, RTE, BoolQ, AI2-ARC, OpenbookQA, and StoryCloze. Ablation studies reveal that number of finetuning datasets, model scale, and natural language instructions are key to the success of instruction tuning. TargetInput (Commonsense Reasoning) keep stack of pillow cases in fridgeInference on unseen task typeFinetune on many tasks (“instruction-tuning”) …Translate this sentence to Spanish: The new office building was built in less than three months.Input (Translation) El nuevo edificio de oficinas se construyó en tres meses.TargetInput (Natural Language Inference) It is not possible to tellFLAN ResponseCoreference resolution tasksSentiment analysis tasksGPT-3 175B zero shotGPT-3 175B few-shotFLAN 137B zero-shotPerformance on unseen task typesNatural language inference42.953.256.2Reading Comprehension63.772.677.4Closed-Book QA49.855.756.6Here is a goal: Get a cool sleep on summer days. How would you accomplish this goal? OPTIONS: -Keep stack of pillow cases in fridge. -Keep stack of pillow cases in oven.Premise: At my age you will probably have learnt one lesson. Hypothesis: It's not certain how many lessons you'll learn by your thirties. Does the premise entail the hypothesis? OPTIONS: -yes -it is not possible to tell -no Figure 1: Top: overview of instruction tuning and FLAN. Instruction tuning finetunes a pretrained language model on a mixture of tasks phrased as instructions. At inference time, we evaluate on an unseen task type; for instance, we could evaluate the model on natural language inference (NLI) when no NLI tasks were seen during instruction tuning. Bottom: performance of zero-shot FLAN, compared with zero-shot and few-shot GPT-3, on three unseen task types where instruction tuning improved performance substantially out of ten we evaluate. NLI datasets: ANLI R1–R3, CB, RTE. Reading comprehension datasets: BoolQ, MultiRC, OBQA. Closed-book QA datasets: ARC-easy, ARC-challenge, NQ, TriviaQA. ∗Lead contributors. Author contributions listed at end of paper. 1arXiv:2109.01652v5 [cs.CL] 8 Feb 2022

1610.06258.pdf

Using Fast Weights to Attend to the Recent Past Jimmy Ba University of Toronto jimmy@psi.toronto.eduGeoffrey Hinton University of Toronto and Google Brain geoffhinton@google.com Volodymyr Mnih Google DeepMind vmnih@google.comJoel Z. Leibo Google DeepMind jzl@google.comCatalin Ionescu Google DeepMind cdi@google.com Abstract Until recently, research on artificial neural networks was largely restricted to sys- tems with only two types of variable: Neural activities that represent the current or recent input and weights that learn to capture regularities among inputs, outputs and payoffs. There is no good reason for this restriction. Synapses have dynam- ics at many different time-scales and this suggests that artificial neural networks might benefit from variables that change slower than activities but much faster than the standard weights. These “fast weights” can be used to store temporary memories of the recent past and they provide a neurally plausible way of imple- menting the type of attention to the past that has recently proved very helpful in sequence-to-sequence models. By using fast weights we can avoid the need to store copies of neural activity patterns. 1 Introduction Ordinary recurrent neural networks typically have two types of memory that have very different time scales, very different capacities and very different computational roles. The history of the sequence currently being processed is stored in the hidden activity vector, which acts as a short-term memory that is updated at every time step. The capacity of this memory is O(H)whereHis the number of hidden units. Long-term memory about how to convert the current input and hidden vectors into the next hidden vector and a predicted output vector is stored in the weight matrices connecting the hidden units to themselves and to the inputs and outputs. These matrices are typically updated at the end of a sequence and their capacity is O(H2) + O(IH) + O(HO)whereIandOare the numbers of input and output units. Long short-term memory networks [Hochreiter and Schmidhuber, 1997] are a more complicated type of RNN that work better for discovering long-range structure in sequences for two main reasons: First, they compute increments to the hidden activity vector at each time step rather than recomputing the full vector1. This encourages information in the hidden states to persist for much longer. Second, they allow the hidden activities to determine the states of gates that scale the effects of the weights. These multiplicative interactions allow the effective weights to be dynamically adjusted by the input or hidden activities via the gates. However, LSTMs are still limited to a short-term memory capacity ofO(H)for the history of the current sequence. Until recently, there was surprisingly little practical investigation of other forms of memory in recur- rent nets despite strong psychological evidence that it exists and obvious computational reasons why it was needed. There were occasional suggestions that neural networks could benefit from a third form of memory that has much higher storage capacity than the neural activities but much faster dynamics than the standard slow weights. This memory could store information specific to the his- tory of the current sequence so that this information is available to influence the ongoing processing 1This assumes the “remember gates ” of the LSTM memory cells are set to one.arXiv:1610.06258v3 [stat.ML] 5 Dec 2016

sciadv.adn0042.pdf

Hikichi et al., Sci. Adv. 10, eadn0042 (2024) 1 March 2024 Science Adv AnceS | ReSeAR cH AR ticle 1 of 20VIROLOGY Epistatic pathways can drive HIV- 1 escape from integrase strand transfer inhibitors Yuta Hikichi1, Jonathan R. Grover2, Alicia Schäfer2, Walther Mothes2, Eric O. Freed1* People living with human immunodeficiency virus (HIV) receiving integrase strand transfer inhibitors (INSTIs) have been reported to experience virological failure in the absence of resistance mutations in integrase. To elucidate INSTI resistance mechanisms, we propagated HIV- 1 in the presence of escalating concentrations of the INSTI dolutegravir. HIV- 1 became resistant to dolutegravir by sequentially acquiring mutations in the envelope glyco - protein (Env) and the nucleocapsid protein. The selected Env mutations enhance the ability of the virus to spread via cell- cell transfer, thereby increasing the multiplicity of infection (MOI). While the selected Env mutations confer broad resistance to multiple classes of antiretrovirals, the fold resistance is ~2 logs higher for INSTIs than for other classes of drugs. We demonstrate that INSTIs are more readily overwhelmed by high MOI than other classes of antiretrovirals. Our findings advance the understanding of how HIV- 1 can evolve resistance to antiretrovirals, including the potent INSTIs, in the absence of drug- target gene mutations. INTRODUCTION Six classes of antiretrovirals (ARVs) have been approved for clinical use by the US Food and Drug Administration: nucleoside reverse transcriptase (RT) inhibitors (NRTIs), nonnucleoside RT inhibitors (NNRTIs), integrase strand transfer inhibitors (INSTIs), protease inhibitors (PIs), entry inhibitors, and a recently approved capsid inhibitor, lenacapavir (LEN) (1 , 2). Combination antiretroviral therapy (cART) has markedly reduced human immunodeficiency virus (HIV)–associated morbidity and mortality. However, resistance to ARVs does arise in some people living with HIV (PLWH), often associated with poor adherence, use of suboptimal drug regimens, and/or lack of viral load monitoring, particularly in poorly re- sourced areas (3). In most cases, drug resistance is caused by muta- tions in the genes targeted by the drugs, often by interfering with the interaction between the drug and the viral target (3). Thus, in the clinical setting, drug resistance monitoring is largely focused on drug- target genes. Recently approved ARVs have been developed with the aim of overcoming resistant variants observed in the clinic. For example, second- generation INSTIs, such as dolutegravir (DTG) and bictegravir (BIC), show some efficacy against IN mutants that are resistant to first- generation INSTIs like raltegravir (RAL) (4). These second- generation INSTIs also exhibit higher genetic barriers to resistance compared to the first- generation INSTIs and RT in- hibitors ( 5). At present, regimens containing DTG are therefore rec- ommended as the preferred first- line regimen for most PLWH (6). Retroviral integration requires two enzymatic reactions catalyzed by IN: 3′ - end processing, during which the enzyme cleaves two nucleotides from the 3 ′ ends of the newly synthesized linear viral DNA, and DNA strand transfer, which entails the insertion of the viral DNA ends into host cell target DNA. The integration reaction takes place in a macromolecular complex known as the intasome, which comprises an IN multimer and the two viral DNA ends (4). INSTIs inhibit the strand transfer reaction by binding IN and the viral DNA ends in the intasome and chelating the Mg++ ions required for IN catalytic activity (4 ). Five INSTIs are currently approved for clinical use: two “first- generation” INSTIs, RAL and elvitegravir (EVG), and three “second- generation” INSTIs, DTG, BIC, and cabotegravir (CAB). Despite the predominant role of drug- target gene mutations in HIV- 1 drug resistance, mutations outside drug- target genes can contribute to drug resistance. Particularly in the case of PIs and INSTIs, some PLWH experience virological failure in the absence of mutations in the target genes (7 –11). Mutations in Gag and the envelope glycoprotein (Env) have been implicated in PI resist- ance ( 12, 13). In vitro studies have reported that mutations in the 3′polypurine tract (3′ PPT) reduce the susceptibility of HIV- 1 to INSTIs (14–16). 3′PPT mutations may lead to the accumulation of unintegrated 1- LTR circles that can support the expression of viral proteins (14, 16) particularly in cell lines that express HTLV- 1 Tax (14). Wijting et al . (11) reported a distinct set of mutations in the 3′PPT from a patient failing DTG monotherapy in the absence of INSTI resistance mutations in IN. However, in other studies, these in vivo–derived 3′ PPT mutations were found not to confer resistance to INSTIs in  vitro (17). It is therefore still unclear whether, or to what extent, 3′PPT mutations contribute to INSTI resistance in vivo. Nevertheless, as more potent inhibitors with higher genetic barriers to resistance are developed, unconventional drug resistance pathways will become important to consider. The Env glycoproteins play a central role in HIV- 1 entry and immune evasion. Env exists as a metastable trimer of three pro- tomers comprising gp120 and gp41 heterodimers on the surface of the virion and the infected cell. The binding of gp120 to CD4 on the target cell triggers conformational rearrangement of the Env trimer that exposes coreceptor (CCR5 or CXCR4) binding sites in gp120. Subsequent binding of gp120 to coreceptor promotes insertion of the gp41 fusion peptide into the target cell membrane, and the refolding of gp41 heptad repeat 1 and 2 (HR1 and HR2) mediates the fusion of viral and cellular membranes, allowing viral entry into the cytosol of the target cell (18). Single- molecule Förster resonance energy transfer (smFRET) analysis has demonstrated that the Env trimer spontaneously transitions between at least three distinct pre- fusion conformations: state 1 (pretriggered, closed conformation), state 2 (necessary, intermediate conformation), and state 3 (fully 1virus- cell interaction Section, Hiv dynamics and Replication Program, center for cancer Research, national c ancer i nstitute, Frederick, Md , USA. 2department of Microbial Pathogenesis, Yale University School of Medicine, new Haven, ct , USA. *corresponding author. email: efreed@ mail. nih. govcopyright © 2024 the Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. no claim to original U.S. Government Works. distributed under a creative c ommons Attribution nonc ommercial license 4.0 ( cc BY- nc ). Downloaded from https://www.science.org on March 26, 2024

10.1016.j.cell.2023.12.034.pdf

Leading Edge Commentary Enabling structure-based drug discovery utilizing predicted models Edward B. Miller,1,*Howook Hwang,1Mee Shelley,2Andrew Placzek,2Joa˜o P.G.L.M. Rodrigues,1Robert K. Suto,3 Lingle Wang,1Karen Akinsanya,1and Robert Abel1 1Schro ¨dinger New York, 1540 Broadway, 24th Floor, New York, NY 10036, USA 2Schro ¨dinger Portland, 101 SW Main Street, Suite 1300, Portland, OR 97204, USA 3Schro ¨dinger Framingham, 200 Staples Drive, Suite 210, Framingham, MA 01702, USA *Correspondence: ed.miller@schrodinger.com https://doi.org/10.1016/j.cell.2023.12.034 High-quality predicted structures enable structure-based approaches to an expanding number of drug dis- covery programs. We propose that by utilizing free energy perturbation (FEP), predicted structures can be confidently employed to achieve drug design goals. We use structure-based modeling of hERG inhibition to illustrate this value of FEP. Introduction Traditional structure-based drug design offers a rational basis to guide the discov-ery of novel chemical matter. Combined with the apparent success of structure- prediction methodology (AlphaFold, Ro-seTTAFold, et al.), the domain of applica- bility of structure-based drug design would, at first glance, appear to havedramatically increased due to the suddenavailability of seemingly high-fidelity pre- dicted structures for any protein seq- uence. However, preliminary evidencesuggests that AlphaFold struggles to reli- ably generate experimentally observed alternative protein conformations. 1Cru- cially, the utility of these predicted struc- tures for atomistic modeling and drug design must be scrutinized before theycan be deployed in lieu of experimental structures. The most direct measurement of a pre- dicted structure’s accuracy is how well itmatches a later solved experimental stru- cture. This metric is crucial for assessing the performance of structure predictionmethods, but within the realm of drug dis- covery, the relevance and value of pre- dicted protein structure models is directlyrelated to their impact on drug design out- comes. Multiple atomic resolution struc- tures, both predicted and experimental,can be used to rationally optimize molec-ular properties, such as on-target po- tency, off-target potency, and absorption, distribution, metabolism, excretion, andtoxicity (ADMET) properties. In this Com- mentary, we explore how predicted struc-tures can be confidently applied to these drug design challenges. We focus on free energy perturbation, a computationalassay, to quantify the accuracy of pre- dicted structures for these purposes. Motivations for structure prediction A structure is most useful when it is of the protein target in the therapeutically rele- vant state. The challenge with structure-based drug design is being able to obtain the right structure in the disease-relevant state bound with project chemical matter.As an example, we point to the experi- mental structural biology pursuits around the leucine-rich repeat kinase 2 (LRRK2).Mutants of LRRK2 have been implicated in Parkinson’s disease. Structures have been obtained of inactive LRRK2 with-out an inhibitor, as a monomer (PDB:7LHW), and as a dimer (PDB: 7LHT), as well as the G2019S mutant (PDB: 7LI3). Later, an active type 1 inhibitor boundstructure was published (PDB: 8TXZ) as well as an inactive state with a type 2 in- hibitor (PDB: 8TZE). Functionally, LRRK2is associated with cellular trafficking, and a structure of microtubule-bound LRRK2 was also recently published(PDB: 7THY). Generally, the demand fora protein structure in various physiologi- cally relevant structural and dynamical states outpaces the supply. From a structure prediction perspec- tive, numerous publications have offered approaches to bias or to explore multiplereceptor states as part of structure pre- diction. 2,3Under favorable conditions, alimited number of predicted structures are presented to the chemist, who must then decide which model or models areworthy of committing resources toward. This is not a trivial commitment—the expectation is that a predicted structureshould precede, if not outright replace, an experimental structure. Therefore, if a predicted structure is considered accu-rate, it should drive consequential deci-sions, among them which compounds to pursue for costly synthesis and to provide a clear, ideally quantitative rationale asto why. Any predicted structure must be judged by its fidelity to reality. Rather than focuson measures of the geometric agreement with some future experimental structure, we propose here that a more meaningfulquestion is to ask the extent to which the predicted structure can be used to model existing structure-activity relation-ships. The expectation is that a modelthat can recapitulate a known structure- activity relationship (SAR) is qualified to make predictions for novel compoundsand to drive synthesis of those com- pounds in response to predicted binding affinity. While a large number of methods ranging from knowledge-based machine learning to physics-based simulationshave shown promises in predicting pro-tein-ligand binding free energies, 4we will focus on the application of one of the most extensively and broadly vali-dated methods, free energy perturbation (FEP), to evaluate a model’s ability to ll Cell 187, February 1, 2024 ª2024 Elsevier Inc. 521

1805.02867.pdf

arXiv:1805.02867v2 [cs.PF] 28 Jul 2018Online normalizer calculation for softmax Maxim Milakov NVIDIA mmilakov@nvidia.comNatalia Gimelshein NVIDIA ngimelshein@nvidia.com Abstract The Softmax function is ubiquitous in machine learning, mul tiple previous works suggested faster alternatives for it. In this paper we propo se a way to compute classical Softmax with fewer memory accesses and hypothesi ze that this reduction in memory accesses should improve Softmax performance on ac tual hardware. The benchmarks confirm this hypothesis: Softmax accelerate s by up to 1.3x and Softmax+TopK combined and fused by up to 5x. 1 Introduction Neural networks models are widely used for language modelin g, for tasks such as machine transla- tion [1] and speech recognition [2]. These models compute wo rd probabilities taking into account the already generated part of the sequence. The probabiliti es are usually computed by a Projection layer, which "projects" hidden representation into the out put vocabulary space, and a following Soft- max function, which transforms raw logits into the the vecto r of probabilities. Softmax is utilized not only for neural networks, for example, it is employed in m ultinomial logistic regression [3]. A number of previous works suggested faster alternatives to compute word probabilities. Differenti- ated Softmax [4] and SVD-Softmax [5] replace the projection layer - which is usually just a matrix multiplication - with more computationally efficient alter natives. Multiple variants of Hierarchical Softmax [6, 7, 8] split a single Projection+Softmax pair int o multiple much smaller versions of these two functions organized in tree-like structures. Sampled- based approximations, such as Importance Sampling [9], Noise Contrastive Estimation [10], and Black out [11] accelerate training by running Softmax on select elements of the original vector. Finally, Self-Normalized Softmax [12] augments the objective function to make the softmax normalization te rm close to 1(and skip computing it during inference). This is not an exhaustive list, but, hopefully, a representa tive one. Almost all of the approaches still need to run the original Softmax function, either on fu ll vector or reduced one. There are two exceptions that don’t need to compute the softmax normal ization term: training with Noise Contrastive Estimation and inference with Self-Normalize d Softmax. All others will benefit from the original Softmax running faster. To the best of our knowledge there has been no targeted effort s to improve the performance of the original Softmax function. We tried to address this shortco ming and figured out a way to compute Softmax with fewer memory accesses. We benchmarked it to see if those reductions in memory accesses translate into performance improvements on a real hardware. Preprint. Work in progress.

10.1101.2024.01.02.573943.pdf

De Novo Atomic Protein Structure Modeling for Cryo-EM Density Maps Using 3D Transformer and Hidden Markov Model Nabin Giri1,2and Jianlin Cheng1,2* 1Electrical Engineering and Computer Science, University of Missouri, Columbia, 65211, Missouri, USA. 2NextGen Precision Health Institute, University of Missouri, Columbia, 65211, Missouri, USA. *Corresponding author(s). E-mail(s): chengji@missouri.edu; Contributing authors: ngzvh@missouri.edu; Abstract Accurately building three-dimensional (3D) atomic structures from 3D cryo-electron microscopy (cryo- EM) density maps is a crucial step in the cryo-EM-based determination of the structures of protein complexes. Despite improvements in the resolution of 3D cryo-EM density maps, the de novo con- version of density maps into 3D atomic structures for protein complexes that do not have accurate homologous or predicted structures to be used as templates remains a significant challenge. Here, we introduce Cryo2Struct, a fully automated ab initio cryo-EM structure modeling method that uti- lizes a 3D transformer to identify atoms and amino acid types in cryo-EM density maps first, and then employs a novel Hidden Markov Model (HMM) to connect predicted atoms to build backbone structures of proteins. Tested on a standard test dataset of 128 cryo-EM density maps with varying resolutions (2.1 - 5.6 ˚A) and different numbers of residues (730 - 8,416), Cryo2Struct built substan- tially more accurate and complete protein structural models than the widely used ab initio method - Phenix in terms of multiple evaluation metrics. Moreover, on a new test dataset of 500 recently released density maps with varying resolutions (1.9 - 4.0 ˚A) and different numbers of residues (234 - 8,828), it built more accurate models than on the standard dataset. And its performance is rather robust against the change of the resolution of density maps and the size of protein structures. Keywords: cryo-EM, atomic protein structure modeling, deep learning, transformer, Hidden Markov Model 1 Introduction Determining the three-dimensional (3D) atomic structures of macromolecules, such as protein complexes and assemblies [1–3], is fundamental in structural biology. The 3D arrangement ofatoms provides essential insights into the mecha- nistic understanding of molecular function of pro- teins [4]. In recent years, cryo-electron microscopy (cryo-EM) [5] has emerged as a key technol- ogy for experimentally determining the structures of large protein complexes and assemblies. How- ever, modeling atomic protein structures from 1. CC-BY 4.0 International license made available under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted January 2, 2024. ; https://doi.org/10.1101/2024.01.02.573943doi: bioRxiv preprint

score-matching-denoising.pdf

1 A Connection Between Score Matching and Denoising Autoencoders Pascal Vincent vincentp@iro.umontreal.ca Dept. IRO, Université de Montréal, CP 6128, Succ. Centre-Ville, Montréal (QC) H3C 3J7, Canada. Technical Report 1358 Département d’Informatique et de Recherche Opérationnelle December 2010 THIS IS A PREPRINT VERSION OF A NOTE THAT HAS BEEN ACCEPTED FOR PUBLICATION IN NEURAL COMPUTATION. Keywords: autoencoder, energy based models, score matching, denoising, density estimation. Abstract Denoising autoencoders have been previously shown to be competitive alternatives to Restricted Boltzmann Machines for unsupervised pre-training of each layer of a deep architecture. We show that a simple denoising autoencoder training criterion is equiv- alent to matching the score (with respect to the data) of a specific energy based model to that of a non-parametric Parzen density estimator of the data. This yields several useful insights. It defines a proper probabilistic model for the denoising autoencoder technique which makes it in principle possible to sample from them or to rank examples by their energy. It suggests a different way to apply score matching that is related to learning to denoise and does not require computing second derivatives. It justifies the use of tied weights between the encoder and decoder, and suggests ways to extend the success of denoising autoencoders to a larger family of energy-based models. 1 Introduction This note uncovers an unsuspected link between the score matching technique (Hyväri- nen, 2005; Hyvärinen, 2008) for learning the parameters of unnormalized density mod- els over continuous-valued data, and the training of denoising autoencoders (Vincent et al. , 2008, 2010). Score matching (SM) is an alternative to the maximum likelihood principle suitable for unnormalized probability density models whose partition function is intractable. Its

2202.08371.pdf

arXiv:2202.08371v1 [cs.LG] 15 Feb 2022THE QUARKS OF ATTENTION PIERRE BALDI AND ROMAN VERSHYNIN Abstract. Attention plays a fundamental role in both natural and artifi cial intelligence systems. In deep learning, attention-based neural archite ctures, such as transformer archi- tectures, are widely used to tackle problems in natural lang uage processing and beyond. Here we investigate the fundamental building blocks of atte ntion and their computational properties. Within the standard model of deep learning, we c lassify all possible fundamental building blocks of attention in terms of their source, targe t, and computational mechanism. We identify and study three most important mechanisms: addi tive activation attention, mul- tiplicative output attention (output gating), and multipl icative synaptic attention (synaptic gating). The gating mechanisms correspond to multiplicati ve extensions of the standard model and are used across all current attention-based deep l earning architectures. We study their functional properties and estimate the capacity of se veral attentional building blocks in the case of linear and polynomial threshold gates. Surpri singly, additive activation atten- tion plays a central role in the proofs of the lower bounds. At tention mechanisms reduce the depth of certain basic circuits and leverage the power of quadratic activations without incurring their full cost. Keywords: neural networks; attention; transformers; capacity; comp lexity; deep learning. Contents 1. Introduction 2 2. Sytematic Identification of Attention Quarks: Within and Beyond the Standard Model 3 3. All you Need is Gating: Transformers 10 4. Functional Aspects of Attention 11 5. Cardinal Capacity Review 16 6. Capacity of Single Unit Attention 20 7. Capacity of Attention Layers 28 8. Conclusion 31 9. Appendix: Detailed Proof of Theorem 6.5 33 Acknowledgment 36 References 36 “Everyone knows what attention is... It is the taking possess ion by the mind in clear and vivid form, of one out of what seem several simultaneousl y possible objects or trains of thought...” William James, Principles of Psychology (1890). Date : February 18, 2022. 1

2404.12358.pdf

Preprint From rtoQ∗: Your Language Model is Secretly a Q-Function Rafael Rafailov* Stanford University rafailov@stanford.eduJoey Hejna* Stanford University jhejna@stanford.eduRyan Park Stanford University rypark@stanford.edu Chelsea Finn Stanford University cbfinn@stanford.edu Abstract Reinforcement Learning From Human Feedback (RLHF) has been a critical to the success of the latest generation of generative AI models. In response to the complex nature of the classical RLHF pipeline, direct alignment algorithms such as Direct Preference Optimization (DPO) have emerged as an alternative approach. Although DPO solves the same objective as the standard RLHF setup, there is a mismatch between the two approaches. StandardRLHFdeploysreinforcementlearninginaspecifictoken-levelMDP, while DPO is derived as a bandit problem in which the whole response of the model is treated as a single arm. In this work we rectify this difference, first we theoretically show that we can derive DPO in the token-level MDP as a general inverse Q-learning algorithm, which satisfies the Bellman equation. Using our theoretical results, we provide three concrete empirical insights. First, we show that because of its token level interpretation, DPO is able to perform some type of credit assignment. Next, we prove that under the token level formulation, classical search-based algorithms, such as MCTS, which have recently been applied to the language generation space, are equivalent to likelihood-based search on a DPO policy. Empirically we show that a simple beam search yields meaningful improvement over the base DPO policy. Finally, we show how the choice of reference policy causes implicit rewards to decline during training. We conclude by discussing applications of our work, including information elicitation in multi-tun dialogue, reasoning, agentic applications and end-to-end training of multi-model systems. 1 Introduction Reinforcement Learning from Human Feedback (RLHF) has become the defacto method for aligning large language models (LLMs) with human intent due to its success in a wide range of applications from summarization (Stiennon et al., 2022) to instruction following (Ouyang et al., 2022). By learning a reward function from human-labeled comparisons, RLHF is able to capture complex objectives that are in-describedable in practice. Following the success of (Ziegler et al., 2020), numerous works have considered new algorithms for training and sampling from large models in various domains using techniques from reinforcement learning (RL). In particular direct alignment methods, such as Direct Preference Optimization (DPO) (Rafailov et al., 2023) have gained traction in recent months because of their simplicity (Zhao et al., 2023a; Azar et al., 2023). Instead of learning a reward function and then using RL, direct alignment methods use the relationship between reward functions and policies in the contextual bandit setting to optimize both simultaneously. Similar ideas have since been applied to vision language (Zhao et al., 2023b) and image generation models (Lee et al., 2023). *Denotes equal contribution 1arXiv:2404.12358v1 [cs.LG] 18 Apr 2024

2112.07868.pdf

Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases Shrimai Prabhumoye1, Rafal Kocielnik2, Mohammad Shoeybi1, Anima Anandkumar1,2, Bryan Catanzaro1 1NVIDIA,2California Institute of Technology {sprabhumoye@nvidia.com, rafalko@caltech.edu} Abstract Warning: this paper contains content that may be offensive or upsetting. Detecting social bias in text is challenging due to nuance, subjectivity, and difficulty in ob- taining good quality labeled datasets at scale, especially given the evolving nature of so- cial biases and society. To address these challenges, we propose a few-shot instruction- based method for prompting pre-trained lan- guage models (LMs). We select a few class- balanced exemplars from a small support repository that are closest to the query to be labeled in the embedding space. We then pro- vide the LM with instruction that consists of this subset of labeled exemplars, the query text to be classified, a definition of bias, and prompt it to make a decision. We demon- strate that large LMs used in a few-shot con- text can detect different types of fine-grained biases with similar and sometimes superior ac- curacy to fine-tuned models. We observe that the largest 530B parameter model is signifi- cantly more effective in detecting social bias compared to smaller models (achieving at least 13% improvement in AUC metric compared to other models). It also maintains a high AUC (dropping less than 2%) when the labeled repository is reduced to as few as 100samples. Large pretrained language models thus make it easier and quicker to build new bias detectors. 1 Introduction Detecting social bias in text is of utmost importance as stereotypes and biases can be projected through language (Fiske, 1993). Detecting bias is challeng- ing because it can be expressed through seemingly innocuous statements which are implied and rarely explicit, and the interpretation of bias can be sub- jective leading to noise in labels. In this work, we focus on detecting social bias in text as defined in Sap et al. (2020) using few-shot instruction-based prompting of pre-trained language models (LMs).Current approaches that detect bias require large labeled datasets to train the models (Chung et al., 2019; Waseem and Hovy, 2016; Zampieri et al., 2019; Davidson et al., 2017a). Collecting such labeled sets is an expensive process and hence they are not easily available. Furthermore, most of the prior work relies on finetuning (Sap et al., 2020; Mandl et al., 2019; Zampieri et al., 2019) neural architectures which is costly in case of large LMs (Strubell et al., 2019) and access to finetune large LMs may be limited (Brown et al., 2020). Prior work on bias detection has not fo- cused on modeling multiple types of biases across datasets as it requires careful optimization to suc- ceed (Hashimoto et al., 2017; Søgaard and Gold- berg, 2016; Ruder, 2017). Finetuning a model can also lead to over-fitting especially in case of smaller train sets and to catastrophic forgetting of knowledge present in the pre-trained model (Fatemi et al., 2021). Moreover, finetuning approaches are prone to be affected by noisy labels (Song et al., 2022) which is especially an issue with datasets for bias detection. The human labeling used to an- notate these datasets can introduce bias and noisy labels (Hovy and Prabhumoye, 2021). We harness the knowledge present in large scale pre-trained language models (Davison et al., 2019; Zhou et al., 2020; Petroni et al., 2019; Zhong et al., 2021; Shin et al., 2020) to detect a rich set of bi- ases. Our method prompts the LM with a textual post and labeled exemplars along with instructions to detect bias in the given post. We explore the capabilities of LMs to flexibly accommodate differ- ent dimensions of bias without any finetuning and with limited access to labeled samples (few-shot classification). Prompt-engineering plays a central role in finetuning-free approaches (Liu et al., 2021b). It is the process of creating a prompting function that results in the best performance on the desired down- stream task. Prompt-engineering can be performedarXiv:2112.07868v2 [cs.CL] 15 Apr 2022

2101.03288.pdf

How to Train Your Energy-Based Models Yang Song yangsong@cs.stanford.edu Stanford University Diederik P. Kingma dpkingma@google.com Google Research Abstract Energy-Based Models (EBMs), also known as non-normalized probabilistic models, specify probability density or mass functions up to an unknown normalizing constant. Unlike most other probabilistic models, EBMs do not place a restriction on the tractability of the normalizing constant, thus are more flexible to parameterize and can model a more expressive family of probability distributions. However, the unknown normalizing constant of EBMs makes training particularly difficult. Our goal is to provide a friendly introduction to modern approaches for EBM training. We start by explaining maximum likelihood training with Markov chain Monte Carlo (MCMC), and proceed to elaborate on MCMC-free approaches, including Score Matching (SM) and Noise Constrastive Estimation (NCE). We highlight theoretical connections among these three approaches, and end with a brief survey on alternative training methods, which are still under active research. Our tutorial is targeted at an audience with basic understanding of generative models who want to apply EBMs or start a research project in this direction. 1. Introduction Probabilistic models with a tractable likelihood are a double-edged sword. On one hand, a tractable likelihood allows for straightforward comparison between models, and straightfor- ward optimization of the model parameters w.r.t. the log-likelihood of the data. Through tractable models such as autoregressive (Graves, 2013; Germain et al., 2015; Van Oord et al., 2016) or flow-based generative models (Dinh et al., 2014, 2016; Rezende and Mohamed, 2015), we can learn flexible models of high-dimensional data. In some cases even though the likelihood is not completely tractable, we can often compute and optimize a tractable lower bound of the likelihood, as in the framework of variational autoencoders (Kingma and Welling, 2014; Rezende et al., 2014). Still, the set of models with a tractable likelihood is constrained. Models with a tractable likelihood need to be of a certain form: for example, in case of autoregressive models, the model distribution is factorized as a product of conditional distributions, and in flow-based generative models the data is modeled as an invertible transformation of a base distribution. In case of variational autoencoders, the data must be modeled as a directed latent-variable model. A tractable likelihood is related to the fact that these models assume that exact synthesis of pseudo-data from the model can be done with a specified, tractable procedure. These assumptions are not always natural. Energy-based models (EBM) are much less restrictive in functional form: instead of speci- fying a normalized probability, they only specify the unnormalized negative log-probability, 1arXiv:2101.03288v2 [cs.LG] 17 Feb 2021

2303.07487v2.pdf

Using VAEs to Learn Latent Variables: Observations on Applications in cryo-EM Edelberg, Daniel G. Yale UniversityLederman, Roy R. Yale University May 12, 2023 Abstract Variational autoencoders (VAEs) are a popular generative model used to approximate distributions. The encoder part of the VAE is used in amortized learning of latent variables, producing a latent rep- resentation for data samples. Recently, VAEs have been used to characterize physical and biological systems. In this case study, we qualitatively examine the amortization properties of a VAE used in biological applications. We find that in this application the encoder bears a qualitative resemblance to more traditional explicit representation of latent variables. 1 Introduction Variational Autoencoders (VAEs) provide a deep learning method for efficient approximate inference for problems with continuous latent variables. A brief reminder about VAEs is presented in Section 2.1; a more complete description can be found, inter alia, in [1, 2, 3, 4, 5, 6]. Since their introduction, VAEs have found success in a wide variety of fields. Recently, they have been used in scientific applications and physical systems [7, 8, 9, 10, 11]. Given a set of data x={xi}, VAEs simultaneously learn an encoder Enc ξthat expresses a conditional distribution qξ(z|x) of a latent variable zigiven a sample xi, and a decoder Dec θwhich expresses the conditional distribution pθ(x|z). They are trained using empirical samples to approximate the distribution pθ(x,z). In this work we focus on the properties of the encoder distribution qξ(z|x) that arise as an approximation of the distribution pθ(z|x). A single encoder qξ(z|x) is optimized to be able to produce the distribution of latent variablezfor any input x, which is a form of amortization. Intuitively, one might expect that the encoder qξ(z|x) would generalize well to plausible inputs that it has not encountered during the optimization/training procedure. Indeed, this generalization is observed in many applications, and the ability of the encoder to compute the latent variables for new unseen data points is used in some applications. In addition, the variational construction sidesteps a statistical problem by marginalizing over the latent variables to approximate the maximum-likelihood estimator (MLE) for some parameters θof the distribution pθ(x,z), rather than θandthe latent variables ziassociated with each sample xi. In the latter case, the number of variables grows with the number of samples and the estimates of pθ(x,z) may not converge to the true solution. We present a qualitative case study of the amortization in VAEs in a physical problem, looking at a VAE applied to the problem of continuous heterogeneity in cryo-electron microscopy (cryo-EM), implemented in CryoDRGN [7]. We examine the hypothesis that the encoder in this VAE generalizes well to previously unseen data, and we compare the use of a VAE to the use of an explicit variational estimation of the distribution of the latent variables. In order to study the generalization in a realistic environment, we exploit well-known invariances and approximate invariances in cryo-EM data to produce natural tests. Our case study suggests that in this case the encoder does not seem to generalize well; this can arguably be interpreted as a form of overfitting of the data. Furthermore, we find that using explicit latent variables 1arXiv:2303.07487v2 [stat.ML] 10 May 2023

2205.12365.pdf

Low-rank Optimal Transport: Approximation, Statistics and Debiasing Meyer Scetbon CREST, ENSAE meyer.scetbon@ensae.frMarco Cuturi Apple and CREST, ENSAE cuturi@apple.com Abstract The matching principles behind optimal transport (OT) play an increasingly impor- tant role in machine learning, a trend which can be observed when OT is used to disambiguate datasets in applications (e.g. single-cell genomics) or used to improve more complex methods (e.g. balanced attention in transformers or self-supervised learning). To scale to more challenging problems, there is a growing consensus that OT requires solvers that can operate on millions, not thousands, of points. The low- rank optimal transport (LOT) approach advocated in Scetbon et al. [2021] holds several promises in that regard, and was shown to complement more established entropic regularization approaches, being able to insert itself in more complex pipelines, such as quadratic OT. LOT restricts the search for low-cost couplings to those that have a low-nonnegative rank, yielding linear time algorithms in cases of interest. However, these promises can only be fulfilled if the LOT approach is seen as a legitimate contender to entropic regularization when compared on properties of interest, where the scorecard typically includes theoretical properties (statistical complexity and relation to other methods) or practical aspects (debiasing, hyperparameter tuning, initialization). We target each of these areas in this paper in order to cement the impact of low-rank approaches in computational OT. 1 Introduction Optimal transport (OT) is used across data-science to put in correspondence different sets of observa- tions. These observations may come directly from datasets, or, in more advanced applications, depict intermediate layered representations of data. OT theory provides a single grammar to describe and solve increasingly complex matching problems (linear, quadratic, regularized, unbalanced, etc...), making it gain a stake in various areas of science such as as single-cell biology Schiebinger et al. [2019], Yang et al. [2020], Demetci et al. [2020], imaging Schmitz et al. [2018], Heitz et al. [2020], Zheng et al. [2020] or neuroscience Janati et al. [2020], Koundal et al. [2020]. Regularized approaches to OT. Solving OT problems at scale poses, however, formidable chal- lenges. The most obvious among them is computational: the Kantorovich [1942] problem on discrete measures of size nis a linear program that requires O(n3logn)operations to be solved. A second and equally important challenge lies in the estimation of OT in high-dimensional settings, since it suffers from the curse-of-dimensionality Fournier and Guillin [2015]. The advent of regularized approaches, such as entropic regularization [Cuturi, 2013], has pushed these boundaries thanks for faster algorithms [Chizat et al., 2020, Clason et al., 2021] and improved statistical aspects [Genevay et al., 2018a]. Despite these clear strengths, regularized OT solvers remain, however, costly as they typically scale quadratically in the number of observations. Scaling up OT using low-rank couplings. While it is always intuitively possible to reduce the size of measures (e.g. using k-means) prior to solving an OT between them, a promising line of work proposes to combine both [Forrow et al., 2019, Scetbon et al., 2021, 2022]. Conceptually, these Preprint. Under review.arXiv:2205.12365v2 [stat.ML] 15 Sep 2022

2207.06569.pdf

Benign, Tempered, or Catastrophic: A Taxonomy of Over/f_itting Neil Mallinar∗ UC San Diego nmallina@ucsd.eduJames B. Simon∗ UC Berkeley james.simon@berkeley.eduAmirhesam Abedsoltan UC San Diego aabedsoltan@ucsd.edu Parthe Pandit UC San Diego parthepandit@ucsd.eduMikhail Belkin UC San Diego mbelkin@ucsd.eduPreetum Nakkiran Apple & UC San Diego preetum@apple.com Abstract The practical success of overparameterized neural networks has motivated the recent scienti/f_ic study of interpo- lating methods , which perfectly /f_it their training data. Certain interpolating methods, including neural networks, can /f_it noisy training data without catastrophically bad test performance, in de/f_iance of standard intuitions from statistical learning theory. Aiming to explain this, a body of recent work has studied benign over/f_itting , a phenomenon where some interpolating methods approach Bayes optimality, even in the presence of noise. In this work we argue that while benign over/f_itting has been instructive and fruitful to study, many real interpolating methods like neural networks do not /f_it benignly : modest noise in the training set causes nonzero (but non-in/f_inite) excess risk at test time, implying these models are neither benign nor catastrophic but rather fall in an intermediate regime. We call this intermediate regime tempered over/f_itting , and we initiate its systematic study. We /f_irst explore this phenomenon in the context of kernel (ridge) regression (KR) by obtaining conditions on the ridge parameter and kernel eigenspectrum under which KR exhibits each of the three behaviors. We /f_ind that kernels with powerlaw spectra, including Laplace kernels and ReLU neural tangent kernels, exhibit tempered over/f_itting. We then empirically study deep neural networks through the lens of our taxonomy, and /f_ind that those trained to interpolation are tempered, while those stopped early are benign. We hope our work leads to a more re/f_ined understanding of over/f_itting in modern learning. 1 Introduction In the last decade, the dramatic success of overparameterized deep neural networks (DNNs) has inspired the /f_ield to reexamine the theoretical foundations of generalization. Classical statistical learning theory suggests that an algorithm which interpolates (i.e. perfectly /f_its) its training data will typically catastrophically over/f_it at test time, generalizing no better than a random function1 Figure 1c illustrates the catastrophic over/f_itting classically expected of an interpolating method. Defying this picture, DNNs can interpolate their training data and generalize well nonetheless [Neyshabur et al., 2015, Zhang et al., 2017], suggesting the need for a new theoretical paradigm within which to understand their over/f_itting. This need motivated the identi/f_ication and study of benign over/f_itting using the terminology of [Bartlett et al., 2020] (also called “harmless interpolation” [Muthukumar et al., 2020]), a phenomenon in which certain methods that perfectly /f_it the training data still approach Bayes-optimal generalization in the limit of large trainset size. Intuitively speaking, benignly-over/f_itting methods /f_it the target function globally, yet /f_it the noise only locally, and the addition of more label noise does not asymptotically degrade generalization. Figure 1a illustrates a simple method that is ∗Co-/f_irst authors. 1There are various ways to formalize this prediction depending on the setting: it is a consequence of the “bias-variance tradeoff” in statistics, the “bias-complexity tradeoff” in PAC learning, and “capacity control”-based generalization bounds in kernel ridge regression. . 1arXiv:2207.06569v2 [cs.LG] 20 Oct 2022

1909.08593v2.pdf

Fine-Tuning Language Models from Human Preferences Daniel M. Ziegler∗Nisan Stiennon∗Jeffrey Wu Tom B. Brown Alec Radford Dario Amodei Paul Christiano Geoffrey Irving OpenAI {dmz,nisan,jeffwu,tom,alec,damodei,paul,irving}@openai.com Abstract Reward learning enables the application of rein- forcement learning (RL) to tasks where reward is defined by human judgment, building a model of reward by asking humans questions. Most work on reward learning has used simulated environ- ments, but complex information about values is of- ten expressed in natural language, and we believe reward learning for language is a key to making RL practical and safe for real-world tasks. In this paper, we build on advances in generative pretrain- ing of language models to apply reward learning to four natural language tasks: continuing text with positive sentiment or physically descriptive language, and summarization tasks on the TL;DR and CNN/Daily Mail datasets. For stylistic con- tinuation we achieve good results with only 5,000 comparisons evaluated by humans. For summa- rization, models trained with 60,000 comparisons copy whole sentences from the input but skip irrel- evant preamble; this leads to reasonable ROUGE scores and very good performance according to our human labelers, but may be exploiting the fact that labelers rely on simple heuristics. 1. Introduction We would like to apply reinforcement learning to complex tasks defined only by human judgment, where we can only tell whether a result is good or bad by asking humans. To do this, we can first use human labels to train a model of reward, and then optimize that model. While there is a long history of work learning such models from humans through interaction, this work has only recently been applied to mod- ern deep learning, and even then has only been applied to relatively simple simulated environments (Christiano et al., 2017; Ibarz et al., 2018; Bahdanau et al., 2018). By contrast, real world settings in which humans need to specify com- *Equal contribution. Correspondence to paul@openai.com.plex goals to AI agents are likely to both involve and require natural language, which is a rich medium for expressing value-laden concepts. Natural language is particularly im- portant when an agent must communicate back to a human to help provide a more accurate supervisory signal (Irving et al., 2018; Christiano et al., 2018; Leike et al., 2018). Natural language processing has seen substantial recent ad- vances. One successful method has been to pretrain a large generative language model on a corpus of unsupervised data, then fine-tune the model for supervised NLP tasks (Dai and Le, 2015; Peters et al., 2018; Radford et al., 2018; Khandel- wal et al., 2019). This method often substantially outper- forms training on the supervised datasets from scratch, and a single pretrained language model often can be fine-tuned for state of the art performance on many different super- vised datasets (Howard and Ruder, 2018). In some cases, fine-tuning is not required: Radford et al. (2019) find that generatively trained models show reasonable performance on NLP tasks with no additional training (zero-shot). There is a long literature applying reinforcement learning to natural language tasks. Much of this work uses algorithmi- cally defined reward functions such as BLEU for translation (Ranzato et al., 2015; Wu et al., 2016), ROUGE for summa- rization (Ranzato et al., 2015; Paulus et al., 2017; Wu and Hu, 2018; Gao et al., 2019b), music theory-based rewards (Jaques et al., 2017), or event detectors for story generation (Tambwekar et al., 2018). Nguyen et al. (2017) used RL on BLEU but applied several error models to approximate human behavior. Wu and Hu (2018) and Cho et al. (2019) learned models of coherence from existing text and used them as RL rewards for summarization and long-form gen- eration, respectively. Gao et al. (2019a) built an interactive summarization tool by applying reward learning to one ar- ticle at a time. Experiments using human evaluations as rewards include Kreutzer et al. (2018) which used off-policy reward learning for translation, and Jaques et al. (2019) which applied the modified Q-learning methods of Jaques et al. (2017) to implicit human preferences in dialog. Yi et al. (2019) learned rewards from humans to fine-tune dia- log models, but smoothed the rewards to allow supervised learning. We refer to Luketina et al. (2019) for a survey ofarXiv:1909.08593v2 [cs.CL] 8 Jan 2020

1406.2661.pdf

Generative Adversarial Nets Ian J. Goodfellow, Jean Pouget-Abadie∗, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair†, Aaron Courville, Yoshua Bengio‡ D´epartement d’informatique et de recherche op ´erationnelle Universit ´e de Montr ´eal Montr ´eal, QC H3C 3J7 Abstract We propose a new framework for estimating generative models via an adversar- ial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model Dthat estimates the probability that a sample came from the training data rather than G. The train- ing procedure for Gis to maximize the probability of Dmaking a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functionsGandD, a unique solution exists, with Grecovering the training data distribution and Dequal to1 2everywhere. In the case where GandDare defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference net- works during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples. 1 Introduction The promise of deep learning is to discover rich, hierarchical models [2] that represent probability distributions over the kinds of data encountered in artificial intelligence applications, such as natural images, audio waveforms containing speech, and symbols in natural language corpora. So far, the most striking successes in deep learning have involved discriminative models, usually those that map a high-dimensional, rich sensory input to a class label [14, 22]. These striking successes have primarily been based on the backpropagation and dropout algorithms, using piecewise linear units [19, 9, 10] which have a particularly well-behaved gradient . Deep generative models have had less of an impact, due to the difficulty of approximating many intractable probabilistic computations that arise in maximum likelihood estimation and related strategies, and due to difficulty of leveraging the benefits of piecewise linear units in the generative context. We propose a new generative model estimation procedure that sidesteps these difficulties.1 In the proposed adversarial nets framework, the generative model is pitted against an adversary: a discriminative model that learns to determine whether a sample is from the model distribution or the data distribution. The generative model can be thought of as analogous to a team of counterfeiters, trying to produce fake currency and use it without detection, while the discriminative model is analogous to the police, trying to detect the counterfeit currency. Competition in this game drives both teams to improve their methods until the counterfeits are indistiguishable from the genuine articles. ∗Jean Pouget-Abadie is visiting Universit ´e de Montr ´eal from Ecole Polytechnique. †Sherjil Ozair is visiting Universit ´e de Montr ´eal from Indian Institute of Technology Delhi ‡Yoshua Bengio is a CIFAR Senior Fellow. 1All code and hyperparameters available at http://www.github.com/goodfeli/adversarial 1arXiv:1406.2661v1 [stat.ML] 10 Jun 2014

2402.10171.pdf

Data Engineering for Scaling Language Models to 128K Context Yao FuκRameswar PandaηXinyao NiuµXiang YueπHannaneh HajishirziσYoon KimλHao Pengδ κUniversity of EdinburghηMIT-IBM Watson AI LabµUniversity of MelbourneπOhio State University σUniversity of WashingtonλMITδUIUC yao.fu@ed.ac.uk yoonkim@mit.edu haopeng@illinois.edu https://github.com/FranxYao/Long-Context-Data-Engineering Abstract We study the continual pretraining recipe for scal- ing language models’ context lengths to 128K, with a focus on data engineering. We hypoth- esize that long context modeling, in particular the ability to utilize information at arbitrary in- put locations , is a capability that is mostly al- ready acquired through large-scale pretraining, and that this capability can be readily extended to contexts substantially longer than seen during training (e.g., 4K to 128K) through lightweight continual pretraining on appropriate data mix- ture. We investigate the quantity andquality of the data for continual pretraining: (1) for quan- tity, we show that 500 million to 5 billion to- kens are enough to enable the model to retrieve information anywhere within the 128K context; (2) for quality, our results equally emphasize do- main balance andlength upsampling . Concretely, we find that na ¨ıvely upsampling longer data on certain domains like books, a common practice of existing work, gives suboptimal performance, and that a balanced domain mixture is impor- tant. We demonstrate that continual pretraining of the full model on 1B-5B tokens of such data is an effective and affordable strategy for scaling the context length of language models to 128K. Our recipe outperforms strong open-source long- context models and closes the gap to frontier mod- els like GPT-4 128K. 1. Introduction A context window of 128K tokens enables large language models to perform tasks that significantly beyond exist- ing paradigm, such as multi-document question answer- ing (Caciularu et al., 2023), repository-level code under- standing (Bairi et al., 2023), long-history dialog model- ing (Mazumder & Liu, 2024), and language model-powered autonomous agents (Weng, 2023). A popular testbed forwhether models can actually utilize long context length is the recent Needle-in-a-Haystack test (Kamradt, 2023), which asks the model to precisely recite the information in a given sentence where the sentence (the “needle”) is placed in an arbitrary location of a 128K long document (the “haystack”). In the open-source space, although works like LongLoRA (Chen et al., 2023b) and YaRN-Mistral (Peng et al., 2023) theoretically support 100K context, they are not able to pass this test at such context lengths, as shown in Fig. 1. Currently, only closed-source frontier models like GPT-4 128K have demonstrated strong performance on the Needle-in-a-Haystack test. This work investigates data engineering methods for scaling language models’ context lengths. Our objective is to con- tinue pretraining the language model on appropriate data mixtures such that it can pass the Needle-in-a-Haystack test at 128K length. Given that most existing models are trained on less than 4K context length (Touvron et al., 2023a) and that attention has quadratic complexity, continual pretrain- ing with full attention on much longer context lengths (we train on 64K-80K context lengths) may seem prohibitively costly at a first glance. However, we show that this is feasi- ble under academic-level resources (see Table 2). We use LLaMA-2 7B and 13B as our base models. We do not make any significant change to model architecture other than ad- justing the base of RoPE, as in Xiong et al. (2023). Our major focus is the data recipe: what andhow much data is able to well-adapt a model to pass the Needle-in-a-Haystack test at 128K context length. We hypothesize that the capability to utilize information at arbitrary locations within long context length is (mostly) already acquired during pretraining, even for models pre- trained on substantially shorter 4K contexts. This hypothe- sis is in contrast to existing works like Xiong et al. (2023); XVerse (2024), which perform continual pretraining on a large amount of data (400B tokens) to inject long-context- modeling capabilities; in this strategy, the cost can be as high as pre-training from scratch. In this work we show that continual pretraining on a small amount of long-context data, in our case, 1-5B tokens, can “unlock” a 7B model’s 1arXiv:2402.10171v1 [cs.CL] 15 Feb 2024

2402.03175v1.pdf

1 THEMATRIX : A B AYESIAN LEARNING MODEL FOR LLM S Siddhartha Dalal Department of Statistics Columbia University The City of New York sd2803@columbia.eduVishal Misra Department of Computer Science Columbia University The City of New York vishal.misra@columbia.edu ABSTRACT In this paper, we introduce a Bayesian learning model to understand the behavior of Large Language Models (LLMs). We explore the optimization metric of LLMs, which is based on predicting the next token, and develop a novel model grounded in this principle. Our approach involves constructing an ideal generative text model represented by a multinomial transition probability matrix with a prior, and we examine how LLMs approximate this matrix. We discuss the continuity of the mapping between embeddings and multinomial distributions, and present the Dirichlet approximation theorem to approximate any prior. Additionally, we demonstrate how text generation by LLMs aligns with Bayesian learning principles and delve into the implications for in-context learning, specifically explaining why in-context learning emerges in larger models where prompts are considered as samples to be updated. Our findings indicate that the behavior of LLMs is consistent with Bayesian Learning, offering new insights into their functioning and potential applications. 1 Introduction The advent of LLMs, starting with GPT3 [ 2], has revolutionized the world of natural language processing, and the introduction of ChatGPT [ 14] has taken the world by storm. There have been several approaches to try and understand how these models work, and in particular how “few-shot" or “in context learning" works [ 10,11,9], and it is an ongoing pursuit. In our work we look at the workings of an LLM from a novel standpoint, and develop a Bayesian model to explain their behavior. We focus on the optimization metric of next token prediction for these LLMs, and use that to build an abstract probability matrix which is the cornerstone of our model and analysis. We show in our paper that the behavior of LLMs is consistent with Bayesian learning and explain many empirical observations of the LLMs using our model. 1.1 Paper organization and our contributions We first describe our approach at a high level, and in the rest of the paper get into the details of the approach. We focus on the optimization metric of these LLMs, namely, predict the next token, and develop the model from there on. We first describe the ideal generative text model (Section 2.1), and relate it to its representation of an abstract (and enormous) multinomial transition probability matrix. We argue that the optimization metric results in these LLMs learning to represent this probability matrix during training, and text generation is nothing but picking a multinomial distribution from a specific row of this matrix. This matrix, however is infeasible to be represented by the LLMs, even with billions of parameters, so the LLMs learn to approximate it. Further, the training data is a subset of the entire text in the world, so the learnt matrix is an approximation and reflection of the matrix induced by the training data, rather than the a representation of the ideal matrix. Next (Section 3), we relate the rows of this matrix to the embeddings of the prompt and prove (Theorem 3.1) a result on the continuity of the mapping between the embeddings and the multinomial distribution induced by the embedding. We then prove (Theorem 4.1) that any prior over multinomial distribution can be represented as a finite mixture of Dirichlet distributions. We then argue, and demonstrate (Section 5.2) that text ∗The authors are listed in alphabetical order.arXiv:2402.03175v1 [cs.LG] 5 Feb 2024

2402.04845.pdf

AlphaFold Meets Flow Matching for Generating Protein Ensembles Bowen Jing1Bonnie Berger1 2Tommi Jaakkola1 Abstract The biological functions of proteins often de- pend on dynamic structural ensembles. In this work, we develop a flow-based generative mod- eling approach for learning and sampling the conformational landscapes of proteins. We re- purpose highly accurate single-state predictors such as AlphaFold and ESMFold and fine-tune them under a custom flow matching framework to obtain sequence-conditoned generative mod- els of protein structure called Alpha FLOW and ESM FLOW . When trained and evaluated on the PDB, our method provides a superior com- bination of precision and diversity compared to AlphaFold with MSA subsampling. When fur- ther trained on ensembles from all-atom MD, our method accurately captures conformational flexibility, positional distributions, and higher- order ensemble observables for unseen proteins. Moreover, our method can diversify a static PDB structure with faster wall-clock convergence to certain equilibrium properties than replicate MD trajectories, demonstrating its potential as a proxy for expensive physics-based simulations. Code is available at https://github.com/ bjing2016/alphaflow . 1. Introduction Proteins adopt complex three-dimensional structures, often as members of structural ensembles with distinct states, col- lective motions, and disordered fluctuations, to carry out their biological functions. For example, conformational changes are critical in the function of transporters, channels, and enzymes, and the properties of equilibrium ensembles help govern the strength and selectivity of molecular interac- tions (Meller et al., 2023; V ¨ogele et al., 2023). While deep learning methods such as AlphaFold (Jumper et al., 2021) have excelled in the single-state modeling of experimental protein structures, they fail to account for this conforma- tional heterogeneity (Lane, 2023; Ourmazd et al., 2022). 1CSAIL, Massachusetts Institute of Technology2Department of Mathematics, Massachusetts Institute of Technology. Corre- spondence to: Bowen Jing <bjing@mit.edu >.Hence, a method which builds upon the level of accuracy of single-structure predictors, but reveals underlying structural ensembles, would be of great value to structural biologists. Existing machine learning approaches for generating struc- tural ensembles have focused on inference-time interven- tions in AlphaFold that modify the multiple sequence alignment (MSA) input (Del Alamo et al., 2022; Stein & Mchaourab, 2022; Wayment-Steele et al., 2023), resulting in a different structure prediction for each version of the MSA. While these approaches have demonstrated some success, they suffer from two key limitations. First, by operating on the MSA, they cannot be generalized to structure predictors based on protein language models (PLMs) such as ESMFold (Lin et al., 2023) or OmegaFold (Wu et al., 2022), which have grown in popularity due to their fast runtime and ease of use. Secondly, these inference-time interventions do not provide the capability to train on protein ensembles from beyond the PDB—for example, ensembles from molecular dynamics, which are of significant scientific interest but can be extremely expensive to simulate (Shaw et al., 2010). To address these limitations, in this work we combine Al- phaFold and ESMFold with flow matching , a recent genera- tive modeling framework (Lipman et al., 2022; Albergo & Vanden-Eijnden, 2022), to propose a principled method for sampling the conformational landscape of proteins. While AlphaFold and ESMFold were originally developed and trained as regression models that predict a single best protein structure for a given MSA or sequence input, we develop a strategy for repurposing them as (sequence-conditioned) generative models of protein structure. This synthesis relies on the key insight that iterative denoising frameworks (such as diffusion and flow-matching) provide a general recipe for converting regression models to generative models with relatively little modification to the architecture and training objective. Unlike inference-time MSA ablation, this strat- egy applies equally well to PLM-based predictors and can be used to train or fine-tune on arbitrary ensembles. While flow matching has been well established for images, its application to protein structures remains nascent (Bose et al., 2023). Hence, we develop a custom flow matching framework tailored to the architecture and training practices of AlphaFold and ESMFold. Our framework leverages the polymer-structured prior distribution from harmonic diffu- 1arXiv:2402.04845v1 [q-bio.BM] 7 Feb 2024

1506.00552.pdf

Coordinate Descent Converges Faster with the Gauss-Southwell Rule Than Random Selection Julie Nutini1, Mark Schmidt1, Issam H. Laradji1, Michael Friedlander2, Hoyt Koepke3 1University of British Columbia,2University of California, Davis,3Dato Abstract There has been significant recent work on the theory and application of randomized coordinate descent algorithms, beginning with the work of Nesterov [ SIAM J. Optim., 22(2), 2012 ], who showed that a random-coordinate selection rule achieves the same convergence rate as the Gauss-Southwell selection rule. This result suggests that we should never use the Gauss-Southwell rule, because it is typically much more expensive than random selection. However, the empirical behaviours of these algorithms contradict this theoretical result: in applications where the computational costs of the selection rules are comparable, the Gauss-Southwell selection rule tends to perform substantially better than random coordinate selection. We give a simple analysis of the Gauss-Southwell rule showing that—except in extreme cases—its convergence rate is faster than choosing random coordinates. We also (i) show that exact coordinate optimization improves the convergence rate for certain sparse problems, (ii) propose a Gauss-Southwell-Lipschitz rule that gives an even faster convergence rate given knowledge of the Lipschitz constants of the partial derivatives, (iii) analyze the effect of approximate Gauss-Southwell rules, and (iv) analyze proximal-gradient variants of the Gauss-Southwell rule. 1 Coordinate Descent Methods There has been substantial recent interest in applying coordinate descent methods to solve large-scale op- timization problems, starting with the seminal work of Nesterov [2012], who gave the first global rate-of- convergence analysis for coordinate-descent methods for minimizing convex functions. This analysis suggests that choosing a random coordinate to update gives the same performance as choosing the “best” coordi- nate to update via the more expensive Gauss-Southwell (GS) rule. (Nesterov also proposed a more clever randomized scheme, which we consider later in this paper.) This result gives a compelling argument to use randomized coordinate descent in contexts where the GS rule is too expensive. It also suggests that there is no benefit to using the GS rule in contexts where it is relatively cheap. But in these contexts, the GS rule often substantially outperforms randomized coordinate selection in practice. This suggests that either the analysis of GS is not tight, or that there exists a class of functions for which the GS rule is as slow as randomized coordinate descent. After discussing contexts in which it makes sense to use coordinate descent and the GS rule, we answer this theoretical question by giving a tighter analysis of the GS rule (under strong-convexity and standard smoothness assumptions) that yields the same rate as the randomized method for a restricted class of functions, but is otherwise faster (and in some cases substantially faster). We further show that, compared to the usual constant step-size update of the coordinate, the GS method with exact coordinate optimization has a provably faster rate for problems satisfying a certain sparsity constraint (Section 5). We believe that this is the first result showing a theoretical benefit of exact coordinate optimization; all previous analyses show that these strategies obtain the same rate as constant step-size updates, even though exact optimization tends to be faster in practice. Furthermore, in Section 6, we propose a variant of the GS rule that, similar to Nesterov’s more clever randomized sampling scheme, uses knowledge of the Lipschitz constants of the coordinate-wise gradients to obtain a faster rate. We also analyze approximate GS rules (Section 7), which 1arXiv:1506.00552v2 [math.OC] 28 Oct 2018

10.1016.j.acha.2021.12.009.pdf

Appl. Comput. Harmon. Anal. 59 (2022) 85–116 Contents lists available at ScienceDirect Applied and Computational Harmonic Analysis www.elsevier.com/locate/acha Loss landscapes and optimization in over-parameterized non-linear systems and neural networks Chaoyue Liua, Libin Zhub,c, Mikhail Belkinc,∗ aDepartment of Computer Science and Engineering, The Ohio State University, United States of America bDepartment of Computer Science and Engineering, University of California, San Diego, United States of America cHalicioğlu Data Science Institute, University of California, San Diego, United States of America a r t i c l e i n f o a b s t r a c t Article history: Received 9 June 2021 Received in revised form 24 December 2021 Accepted 26 December 2021 Available online 10 January 2022 Communicated by David Donoho Keywords: Deep learning Non-linear optimization Over-parameterized models PL∗conditionThe success of deep learning is due, to a large extent, to the remarkable effectiveness of gradient-based optimization methods applied to large neural networks. The purpose of this work is to propose a modern view and a general mathematical framework for loss landscapes and efficient optimization in over-parameterized machine learning models and systems of non-linear equations, a setting that includes over-parameterized deep neural networks. Our starting observation is that optimization landscapes corresponding to such systems are generally not convex, even locally around a global minimum, a condition we call essential non-convexity . We argue that instead they satisfy PL∗, a variant of the Polyak-Łojasiewicz condition [32,25]o n most (but not all) of the parameter space, which guarantees both the existence of solutions and efficient optimization by (stochastic) gradient descent (SGD/GD). The PL∗condition of these systems is closely related to the condition number of the tangent kernel associated to a non-linear system showing how a PL∗-based non-linear theory parallels classical analyses of over-parameterized linear equations. We show that wide neural networks satisfy the PL∗condition, which explains the (S)GD convergence to a global minimum. Finally we propose a relaxation of the PL∗condition applicable to “almost” over-parameterized systems. © 2021 Elsevier Inc. All rights reserved. 1. Introduction A singular feature of modern machine learning is a large number of trainable model parameters. Just in the last few years we have seen state-of-the-art models grow from tens or hundreds of millions parameters to much larger systems with hundreds billion [ 6]o r even trillions parameters [ 14]. Invariably these models are trained by gradient descent based methods, such as Stochastic Gradient Descent (SGD) or Adam [ 19]. Why are these local gradient methods so effective in optimizing complex highly non-convex systems? In the past few years an emerging understanding of gradient-based methods have started to focus on the insight *Corresponding author. E-mail address: mbelkin@ucsd.edu (M. Belkin). https://doi.org/10.1016/j.acha.2021.12.009 1063-5203/© 2021 Elsevier Inc. All rights reserved.

2309.02390.pdf

5 September 2023 Explaining grokking through circuit efficiency Vikrant Varma*, 1, Rohin Shah*, 1, Zachary Kenton1, János Kramár1and Ramana Kumar1 *Equal contributions,1Google DeepMind One of the most surprising puzzles in neural network generalisation is grokking : a network with perfect training accuracy but poor generalisation will, upon further training, transition to perfect generalisation. Weproposethatgrokkingoccurswhenthetaskadmitsageneralisingsolutionandamemorisingsolution, where the generalising solution is slower to learn but more efficient, producing larger logits with the same parameter norm. We hypothesise that memorising circuits become more inefficient with larger training datasets while generalising circuits do not, suggesting there is a critical dataset size at which memorisationandgeneralisationareequallyefficient. Wemakeandconfirmfournovelpredictionsabout grokking, providing significant evidence in favour of our explanation. Most strikingly, we demonstrate two novel and surprising behaviours: ungrokking , in which a network regresses from perfect to low test accuracy, and semi-grokking , in which a network shows delayed generalisation to partial rather than perfect test accuracy. 1. Introduction When training a neural network, we expect that once training loss converges to a low value, the network will no longer change much. Power et al. (2021) discovered a phenomenon dubbed grokking that drastically violates this expectation. The network first “memorises” the data, achieving low and stable training loss with poor generalisation, but with further training transitions to perfect generalisation. We are left with the question: why does the network’s test performance improve dramatically upon continued training, having already achieved nearly perfect training performance? Recent answers to this question vary widely, including the difficulty of representation learning (Liu etal.,2022), thescaleofparametersatinitialisation(Liuetal.,2023), spikesinloss("slingshots")(Thi- lak et al., 2022), random walks among optimal solutions (Millidge, 2022), and the simplicity of the generalising solution (Nanda et al., 2023, Appendix E). In this paper, we argue that the last explanation is correct, by stating a specific theory in this genre, deriving novel predictions from the theory, and confirming the predictions empirically. We analyse the interplay between the internal mechanisms that the neural network uses to calculate the outputs, which we loosely call “circuits” (Olah et al., 2020). We hypothesise that there are two families of circuits that both achieve good training performance: one which generalises well (𝐶gen) and one which memorises the training dataset ( 𝐶mem). The key insight is that when there are multiple circuits that achieve strong training performance, weight decay prefers circuits with high “efficiency” , that is, circuits that require less parameter norm to produce a given logit value. Efficiency answers our question above: if 𝐶genis more efficient than 𝐶mem, gradient descent can reduce nearly perfect training loss even further by strengthening 𝐶genwhile weakening 𝐶mem, which then leads to a transition in test performance. With this understanding, we demonstrate in Section 3 that three key properties are sufficient for grokking: (1) 𝐶gengeneralises well while 𝐶memdoes not, (2)𝐶genis more efficient than 𝐶mem, and (3)𝐶genis learned more slowly than 𝐶mem. Since𝐶gengeneralises well, it automatically works for any new data points that are added to the training dataset, and so its efficiency should be independent of the size of the training dataset. In contrast, 𝐶memmust memorise any additional data points added to the training dataset, and so Corresponding author(s): vikrantvarma@deepmind.com, rohinmshah@deepmind.comarXiv:2309.02390v1 [cs.LG] 5 Sep 2023

10.1016.j.cell.2023.12.035.pdf

Article Brain-wide neural activity underlying memory- guided movement Graphical abstract Highlights dAnatomy-guided activity recordings in multi-regional neural circuits during behavior dMovement encoding is strongest in the medulla, followed bythe midbrain and cortex dChoice coding arises in a specific multi-regional circuitdistributed across the brain dCoding of choice and action exhibit strong correlationsacross brain areasAuthors Susu Chen, Yi Liu, Ziyue Aiden Wang, ...,Shaul Druckmann, Nuo Li, Karel Svoboda Correspondence shauld@stanford.edu (S.D.), nuo.li@bcm.edu (N.L.),karel.svoboda@alleninstitute.org (K.S.) In brief A sparse neural network, distributed across major brain compartments,produces tightly orchestrated activitypatterns underlying decision-making andmovement initiation. Anatomy-guided multi-regional simultaneous recordings Mesoscale activity map data medulla > midbrain > cortexMovement encoding: Choice coding is concentrated in ALM projection zones StriatumThalamusMidbrain SelectivityALM inputALMChoice-related activity is correlated across brain areas Chen et al., 2024, Cell 187, 676–691 February 1, 2024 ª2024 The Authors. Published by Elsevier Inc. https://doi.org/10.1016/j.cell.2023.12.035 ll

2309.14525.pdf

Preprint ALIGNING LARGE MULTIMODAL MODELS WITH FACTUALLY AUGMENTED RLHF Zhiqing Sun∗♠, Sheng Shen∗♣, Shengcao Cao∗♢ Haotian Liu♡, Chunyuan Li♮, Yikang Shen△, Chuang Gan†∇△, Liang-Yan Gui†♢ Yu-Xiong Wang†♢, Yiming Yang†♠, Kurt Keutzer†♣, Trevor Darrell†♣ ♣UC Berkeley,♠CMU,♢UIUC,♡UW–Madison,∇UMass Amherst ♮Microsoft Research,△MIT-IBM Watson AI Lab ABSTRACT Large Multimodal Models (LMM) are built across modalities and the misalign- ment between two modalities can result in “hallucination”, generating textual out- puts that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the task of vision-language alignment, where human annotators are asked to compare two responses and pin- point the more hallucinated one, and the vision-language model is trained to max- imize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with addi- tional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image- text pairs to improve the general capabilities of our model. To evaluate the pro- posed approach in real-world scenarios, we develop a new evaluation benchmark MMH AL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaV A-Bench dataset with the 94% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improve- ment by 60% on MMH AL-BENCH over other baselines. We opensource our code, model, data at https://llava-rlhf.github.io . 1 I NTRODUCTION Large Language Models (LLMs; Brown et al. (2020); Chowdhery et al. (2022); OpenAI (2023)) can delve into the multimodal realm either by further pre-training with image-text pairs (Alayrac et al.; Awadalla et al., 2023) or by fine-tuning them with specialized vision instruction tuning datasets (Liu et al., 2023a; Zhu et al., 2023), leading to the emergence of powerful Large Multimodal Models (LMMs). Yet, developing LMMs faces challenges, notably the gap between the volume and quality of multimodal data versus text-only datasets. Consider the LLaV A model (Liu et al., 2023a), which is initialized from a pre-trained vision encoder (Radford et al., 2021) and an instruction-tuned language model (Chiang et al., 2023). It is trained on just 150K synthetic image-based dialogues, which is much less in comparison to the text-only models (Flan (Longpre et al., 2023) utilizing over 100M examples spanning 1800 tasks. Such limitations in data can lead to misalignment between the vision and language modalities. Consequently, LMMs may produce hallucinated outputs, which are not accurately anchored to the context provided by images. To mitigate the challenges posed by the scarcity of high-quality visual instruction tuning data for LMM training, we introduce LLaVA-RLHF , a vision-language model trained for improved mul- timodal alignment. One of our key contributions is the adaptation of the Reinforcement Learning from Human Feedback (RLHF) (Stiennon et al., 2020; Ouyang et al., 2022; Bai et al., 2022a), a general and scalable alignment paradigm that shows great success for text-based AI agents, to the ∗Equal contribution. Ordering is determined by dice rolling. †Equal advising. 1arXiv:2309.14525v1 [cs.CV] 25 Sep 2023

2306.12672.pdf

From Word Models to World Models: Translating from Natural Language to the Probabilistic Language of Thought Lionel Wong1⋆, Gabriel Grand1⋆, Alexander K. Lew1, Noah D. Goodman2, Vikash K. Mansinghka1, Jacob Andreas1, Joshua B. Tenenbaum1 ⋆Equal contribution. 1MIT,2Stanford Abstract How does language inform our downstream thinking? In particular, how do humans make meaning from language—and how can we leverage a theory of linguistic meaning to build machines that think in more human-like ways? In this paper, we propose rational meaning construction , a computational framework for language-informed thinking that combines neural models of language with probabilistic models for rational inference. We frame linguistic meaning as a context-sensitive mapping from natural language into a probabilistic language of thought (PLoT)—a general-purpose symbolic substrate for probabilistic, generative world modeling. Our architecture integrates two powerful computational tools that have not previously come together: we model thinking with probabilistic programs , an expressive representation for flexible commonsense reasoning; and we model meaning construction with large language models (LLMs), which support broad-coverage translation from natural language utterances to code expressions in a probabilistic programming language. We illustrate our framework in action through examples covering four core domains from cognitive science: probabilistic reasoning, logical and relational reasoning, visual and physical reasoning, and social reasoning about agents and their plans. In each, we show that LLMs can generate context-sensitive translations that capture pragmatically-appropriate linguistic meanings, while Bayesian inference with the generated programs supports coherent and robust commonsense reasoning. We extend our framework to integrate cognitively-motivated symbolic modules (physics simulators, graphics engines, and goal-directed planning algorithms) to provide a unified commonsense thinking interface from language. Finally, we explore how language can drive the construction of world models themselves. We hope this work will help to situate contemporary developments in LLMs within a broader cognitive picture of human language and intelligence, providing a roadmap towards AI systems that synthesize the insights of both modern and classical computational perspectives. 1 Introduction Language expresses the vast internal landscape of our thoughts. We use language to convey what we believe, what we are uncertain about, and what we do not know. We talk about what we see in the world around us, and what we imagine in real or wholly hypothetical futures. We discuss what we want and what we plan to do, and dissect what others want and what we think they will do. We build and pass on new bodies of knowledge in language—we ask questions and offer explanations, give commands and instructions, and propose and refute theories. Some of these ideas can be expressed in part through other means. But language stands apart for its flexibility and breadth, and its seeming proximity to our thoughts. Whatislanguage? How does language get its meaning, and when should we say that a person or machine knows, understands, and can use it? What is the relationship between language and the rest of general cognition—what allows language to inform and support so much of thought? This paper focuses on these questions as they relate to humanlanguage and thought, in computational terms. What integrated cognitive theory can model how language relates to the other core systems of human cognition? If we seek to build AI systems that emulate how humans talk and think, what architecture can integrate language robustly into systems that support the full scope of our thought? Code for the examples in this paper is available at: github.com/gabegrand/world-models . Correspondence: co-primary authors ( zyzzyva@mit.edu, gg@mit.edu ); co-supervisors ( jda@mit.edu, jbt@mit.edu ).arXiv:2306.12672v2 [cs.CL] 23 Jun 2023

2210.17323.pdf

Published as a conference paper at ICLR 2023 GPTQ: A CCURATE POST-TRAINING QUANTIZATION FOR GENERATIVE PRE-TRAINED TRANSFORMERS Elias Frantar∗ IST AustriaSaleh Ashkboos ETH ZurichTorsten Hoefler ETH ZurichDan Alistarh IST Austria & NeuralMagic ABSTRACT Generative Pre-trained Transformer models, known as GPT or OPT, set them- selves apart through breakthrough performance across complex language mod- elling tasks, but also by their extremely high computational and storage costs. Specifically, due to their massive size, even inference for large, highly-accurate GPT models may require multiple performant GPUs, which limits the usability of such models. While there is emerging work on relieving this pressure via model compression, the applicability and performance of existing compression techniques is limited by the scale and complexity of GPT models. In this paper, we address this challenge, and propose GPTQ, a new one-shot weight quantiza- tion method based on approximate second-order information, that is both highly- accurate and highly-efficient. Specifically, GPTQ can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing the bitwidth down to 3 or 4 bits per weight, with negligible accuracy degradation relative to the uncompressed baseline. Our method more than doubles the compression gains rel- ative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU for generative inference. Moreover, we also show that our method can still provide reasonable accuracy in the extreme quantization regime, in which weights are quantized to 2-bit or even ternary quantization levels. We show ex- perimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3.25x when using high-end GPUs (NVIDIA A100) and 4.5x when using more cost-effective ones (NVIDIA A6000). The implemen- tation is available at https://github.com/IST-DASLab/gptq . 1 I NTRODUCTION Pre-trained generative models from the Transformer (Vaswani et al., 2017) family, commonly known as GPT or OPT (Radford et al., 2019; Brown et al., 2020; Zhang et al., 2022), have shown break- through performance for complex language modelling tasks, leading to massive academic and prac- tical interest. One major obstacle to their usability is computational and storage cost, which ranks among the highest for known models. For instance, the best-performing model variants, e.g. GPT3- 175B, have in the order of 175 billion parameters and require tens-to-hundreds of GPU years to train (Zhang et al., 2022). Even the simpler task of inferencing over a pre-trained model, which is our focus in this paper, is highly challenging: for instance, the parameters of GPT3-175B occupy 326GB (counting in multiples of 1024) of memory when stored in a compact float16 format. This exceeds the capacity of even the highest-end single GPUs, and thus inference must be performed using more complex and expensive setups, such as multi-GPU deployments. Although a standard approach to eliminating these overheads is model compression , e.g. (Hoefler et al., 2021; Gholami et al., 2021), surprisingly little is known about compressing such models for inference. One reason is that more complex methods for low-bitwidth quantization or model prun- ing usually require model retraining , which is extremely expensive for billion-parameter models. Alternatively, post-training methods (Nagel et al., 2020; Wang et al., 2020; Hubara et al., 2020; Nahshan et al., 2021), which compress the model in one shot, without retraining, would be very appealing. Unfortunately, the more accurate variants of such methods (Li et al., 2021; Hubara et al., 2021; Frantar et al., 2022) are complex and challenging to scale to billions of parameters (Yao et al., ∗Corresponding author: elias.frantar@ist.ac.at 1arXiv:2210.17323v2 [cs.LG] 22 Mar 2023

10.1016.j.cell.2024.01.026.pdf

Article Cryo-EM structures of the plant plastid-encoded RNA polymerase Graphical abstract Highlights dPlant chloroplast RNA polymerase comprises a catalytic core and four peripheral modules dThe scaffold module stabilizes the catalytic core and bridgesother modules dThe protection module has SOD activity, and the RNAmodule recognizes RNA sequence dThe regulation module likely controls transcription activity ofthe catalytic coreAuthors Xiao-Xian Wu, Wen-Hui Mu, Fan Li, ...,Chanhong Kim, Fei Zhou, Yu Zhang Correspondence zhoufei@mail.hzau.edu.cn (F.Z.), yzhang@cemps.ac.cn (Y.Z.) In brief The cryo-EM structures of Nicotiana tabacum (tobacco) chloroplast RNA polymerase apoenzyme and transcriptionelongation complexes reveal thecomposition, assembly, function, andevolution of the chloroplast transcriptionapparatus. Regulation module Regulation module Wu et al., 2024, Cell 187, 1127–1144 February 29, 2024 ª2024 Elsevier Inc. https://doi.org/10.1016/j.cell.2024.01.026 ll

10.1038.s41467-021-26529-9.pdf

ARTICLE The generative capacity of probabilistic protein sequence models Francisco McGee1,2,3, Sandro Hauri4,5, Quentin Novinger2,5, Slobodan Vucetic4,5, Ronald M. Levy1,3,6,7, Vincenzo Carnevale2,3✉& Allan Haldane1,7✉ Potts models and variational autoencoders (VAEs) have recently gained popularity as gen- erative protein sequence models (GPSMs) to explore fitness landscapes and predict mutation effects. Despite encouraging results, current model evaluation metrics leave unclear whetherGPSMs faithfully reproduce the complex multi-residue mutational patterns observed innatural sequences due to epistasis. Here, we develop a set of sequence statistics to assessthe “generative capacity ”of three current GPSMs: the pairwise Potts Hamiltonian, the VAE, and the site-independent model. We show that the Potts model ’s generative capacity is largest, as the higher-order mutational statistics generated by the model agree with thoseobserved for natural sequences, while the VAE ’s lies between the Potts and site-independent models. Importantly, our work provides a new framework for evaluating and interpretingGPSM accuracy which emphasizes the role of higher-order covariation and epistasis, withbroader implications for probabilistic sequence models in general.https://doi.org/10.1038/s41467-021-26529-9 OPEN 1Center for Biophysics and Computational Biology, Temple University, Philadelphia 19122, USA.2Institute for Computational Molecular Science, Temple University, Philadelphia 19122, USA.3Department of Biology, Temple University, Philadelphia 19122, USA.4Center for Hybrid Intelligence, Temple University, Philadelphia 19122, USA.5Department of Computer & Information Sciences, Temple University, Philadelphia 19122, USA.6Department of Physics, Temple University, Philadelphia 19122, USA.7Department of Chemistry, Temple University, Philadelphia 19122, USA.✉email: vincenzo.carnevale@temple.edu ; allan.haldane@temple.edu NATURE COMMUNICATIONS | (2021) 12:6302 | https://doi.org/10.1038/s41467-021-26529-9 | www.nature.com/naturecommunications 11234567890():,;

2205.11916.pdf

Large Language Models are Zero-Shot Reasoners Takeshi Kojima The University of Tokyo t.kojima@weblab.t.u-tokyo.ac.jpShixiang Shane Gu Google Research, Brain Team Machel Reid Google Research∗Yutaka Matsuo The University of TokyoYusuke Iwasawa The University of Tokyo Abstract Pretrained large language models (LLMs) are widely used in many sub-fields of natural language processing (NLP) and generally known as excellent few-shot learners with task-specific exemplars. Notably, chain of thought (CoT) prompting, a recent technique for eliciting complex multi-step reasoning through step-by- step answer examples, achieved the state-of-the-art performances in arithmetics and symbolic reasoning, difficult system-2 tasks that do not follow the standard scaling laws for LLMs. While these successes are often attributed to LLMs’ ability for few-shot learning, we show that LLMs are decent zero-shot reasoners by simply adding “Let’s think step by step” before each answer. Experimental results demonstrate that our Zero-shot-CoT, using the same single prompt template, significantly outperforms zero-shot LLM performances on diverse benchmark reasoning tasks including arithmetics (MultiArith, GSM8K, AQUA-RAT, SV AMP), symbolic reasoning (Last Letter, Coin Flip), and other logical reasoning tasks (Date Understanding, Tracking Shuffled Objects), without any hand-crafted few-shot examples, e.g. increasing the accuracy on MultiArith from 17.7% to 78.7% and GSM8K from 10.4% to 40.7% with large-scale InstructGPT model (text-davinci- 002), as well as similar magnitudes of improvements with another off-the-shelf large model, 540B parameter PaLM. The versatility of this single prompt across very diverse reasoning tasks hints at untapped and understudied fundamental zero-shot capabilities of LLMs, suggesting high-level, multi-task broad cognitive capabilities may be extracted by simple prompting. We hope our work not only serves as the minimal strongest zero-shot baseline for the challenging reasoning benchmarks, but also highlights the importance of carefully exploring and analyzing the enormous zero-shot knowledge hidden inside LLMs before crafting finetuning datasets or few-shot exemplars. 1 Introduction Scaling up the size of language models has been key ingredients of recent revolutions in natural language processing (NLP) [Vaswani et al., 2017, Devlin et al., 2019, Raffel et al., 2020, Brown et al., 2020, Thoppilan et al., 2022, Rae et al., 2021, Chowdhery et al., 2022]. The success of large language models (LLMs) is often attributed to (in-context) few-shot or zero-shot learning. It can solve various tasks by simply conditioning the models on a few examples (few-shot) or instructions describing the task (zero-shot). The method of conditioning the language model is called “prompting” [Liu et al., 2021b], and designing prompts either manually [Schick and Schütze, 2021, Reynolds and McDonell, 2021] or automatically [Gao et al., 2021, Shin et al., 2020] has become a hot topic in NLP. ∗Work done while at The University of Tokyo. 36th Conference on Neural Information Processing Systems (NeurIPS 2022).arXiv:2205.11916v4 [cs.CL] 29 Jan 2023

2308.06259v3.pdf

Published as a conference paper at ICLR 2024 SELF-ALIGNMENT WITH INSTRUCTION BACKTRANS - LATION Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Omer Levy, Luke Zettlemoyer Jason Weston &Mike Lewis Meta {xianl,jase,mikelewis}@meta.com ABSTRACT We present a scalable method to build a high quality instruction following language model by automatically labelling human-written text with corresponding instruc- tions. Our approach, named instruction backtranslation , starts with a language model finetuned on a small amount of seed data, and a given web corpus. The seed model is used to construct training examples by generating instruction prompts for web documents ( self-augmentation ), and then selecting high quality examples from among these candidates ( self-curation ). This data is then used to finetune a stronger model. Finetuning LLaMa on two iterations of our approach yields a model that outperforms all other LLaMa-based models on the Alpaca leaderboard not relying on distillation data, demonstrating highly effective self-alignment. 1 I NTRODUCTION Aligning large language models (LLMs) to perform instruction following typically requires finetuning on large amounts of human-annotated instructions or preferences (Ouyang et al., 2022; Touvron et al., 2023a; Bai et al., 2022a) or distilling outputs from more powerful models (Wang et al., 2022a; Honovich et al., 2022; Taori et al., 2023; Chiang et al., 2023; Peng et al., 2023; Xu et al., 2023). Recent work highlights the importance of human-annotation data quality (Zhou et al., 2023; Köpf et al., 2023). However, annotating instruction following datasets with such quality is hard to scale. In this work, we instead leverage large amounts of unlabelled data to create a high quality instruction tuning dataset by developing an iterative self-training algorithm. The method uses the model itself to both augment and curate high quality training examples to improve its own performance. Our approach, named instruction backtranslation , is inspired by the classic backtranslation method from machine translation, in which human-written target sentences are automatically annotated with model-generated source sentences in another language (Sennrich et al., 2015). Our method starts with a seed instruction following model and a web corpus. The model is first used toself-augment its training set: for each web document, it creates an instruction following training example by predicting a prompt (instruction) that would be correctly answered by (a portion of) that document. Directly training on such data (similarly to Köksal et al. (2023)) gives poor results in our experiments, both because of the mixed quality of human written web text, and noise in the generated instructions. To remedy this, we show that the same seed model can be used to self-curate the set of newly created augmentation data by predicting their quality, and can then be self-trained on only the highest quality (instruction, output) pairs. The procedure is then iterated, using the improved model to better curate the instruction data, and re-training to produce a better model. Our resulting model, Humpback , outperforms all other existing non-distilled models on the Alpaca leaderboard (Li et al., 2023). Overall, instruction backtranslation is a scalable method for enabling language models to improve their own ability to follow instructions. 2 M ETHOD Our self-training approach assumes access to a base language model, a small amount of seed data, and a collection of unlabelled examples, e.g. a web corpus. The unlabelled data is a large, diverse set 1arXiv:2308.06259v3 [cs.CL] 12 Mar 2024

2209.12892.pdf

LEARNING TO LEARN WITH GENERATIVE MODELS OF NEURAL NETWORK CHECKPOINTS William Peebles∗Ilija Radosavovic∗Tim Brooks Alexei A. Efros Jitendra Malik University of California, Berkeley ABSTRACT We explore a data-driven approach for learning to optimize neural networks. We construct a dataset of neural network checkpoints and train a generative model on the parameters. In particular, our model is a conditional diffusion transformer that, given an initial input parameter vector and a prompted loss, error, or return, predicts the distribution over parameter updates that achieve the desired metric. At test time, it can optimize neural networks with unseen parameters for downstream tasks in just one update. We find that our approach successfully generates parameters for a wide range of loss prompts. Moreover, it can sample multimodal parameter solutions and has favorable scaling properties. We apply our method to different neural network architectures and tasks in supervised and reinforcement learning. 1 I NTRODUCTION Gradient-based optimization is the fuel of modern deep learning. Techniques of this class, such as SGD (Robbins & Monro, 1951) and Adam (Kingma & Ba, 2015), are easy to implement, scale reasonably well and converge to surprisingly good solutions—even in high-dimensional, non-convex neural network loss landscapes. Over the past decade, they have enabled impressive results in computer vision (Krizhevsky et al., 2012; Girshick et al., 2014), natural language processing (Vaswani et al., 2017; Radford et al., 2018) and audio generation (Van Den Oord et al., 2016). While these manual optimization techniques have led to large advances, they suffer from an important limitation: they are unable to improve from past experience. For example, SGD will not converge any faster when used to optimize the same neural network architecture from the same initialization the 100th time versus the first time. Learned optimizers capable of leveraging their past experiences have the potential to overcome this limitation and may accelerate future progress in deep learning. Of course, the concept of learning improved optimizers is not new and dates back to the 1980s, if not earlier, following early work from Schmidhuber (1987) and Bengio et al. (1991). In recent years, sig- nificant effort has been spent on designing algorithms that learn via nested meta-optimization, where the inner loop optimizes the task-level objective and the outer loop learns the optimizer (Andrychow- icz et al., 2016; Li & Malik, 2016; Finn et al., 2017). In some instances, these approaches outperform manual optimizers. However, they are challenging to train in practice due to a reliance on unrolled optimization and reinforcement learning. Taking a modern deep learning perspective suggests a simple, scalable and data-driven approach to this problem. Over the past decade, our community has trained a massive number of checkpoints. These checkpoints contain a wealth of information: diverse parameter configurations and rich metrics such as test losses, classification errors and RL returns that describe the quality of the checkpoint. Instead of leveraging large-scale datasets of images or text, we propose learning from large-scale datasets of checkpoints recorded over the course of many training runs. To this end, we create a dataset of neural network checkpoints (Figure 1, left). Our dataset consists of 23 million checkpoints from over a hundred thousand training runs. We collect data from supervised learning tasks (MNIST, CIFAR-10) as well as reinforcement learning tasks (Cartpole), and across different neural network architectures (MLPs, CNNs). In addition to parameters, we record relevant task-level metrics in each checkpoint, such as test losses and classification errors. *Equal contribution. Code, data and pre-trained models are available on our project page. 1arXiv:2209.12892v1 [cs.LG] 26 Sep 2022

2023.findings-acl.426.pdf

Findings of the Association for Computational Linguistics: ACL 2023 , pages 6810–6828 July 9-14, 2023 ©2023 Association for Computational Linguistics “Low-Resource” Text Classification: A Parameter-Free Classification Method with Compressors Zhiying Jiang1,2, Matthew Y.R. Yang1, Mikhail Tsirlin1, Raphael Tang1, Yiqin Dai2and Jimmy Lin1 1University of Waterloo2AFAIK {zhiying.jiang, m259yang, mtsirlin, r33tang}@uwaterloo.ca quinn@afaik.io jimmylin@uwaterloo.ca Abstract Deep neural networks (DNNs) are often used for text classification due to their high accu- racy. However, DNNs can be computationally intensive, requiring millions of parameters and large amounts of labeled data, which can make them expensive to use, to optimize, and to trans- fer to out-of-distribution (OOD) cases in prac- tice. In this paper, we propose a non-parametric alternative to DNNs that’s easy, lightweight, and universal in text classification: a combi- nation of a simple compressor like gzip with ak-nearest-neighbor classifier. Without any training parameters, our method achieves re- sults that are competitive with non-pretrained deep learning methods on six in-distribution datasets. It even outperforms BERT on all five OOD datasets, including four low-resource lan- guages. Our method also excels in the few-shot setting, where labeled data are too scarce to train DNNs effectively. Code is available at https://github.com/bazingagin/npc_gzip. 1 Introduction Text classification, as one of the most fundamen- tal tasks in natural language processing (NLP), has improved substantially with the help of neu- ral networks (Li et al., 2022). However, most neu- ral networks are data-hungry, the degree of which increases with the number of parameters. Hyper- parameters must be carefully tuned for different datasets, and the preprocessing of text data (e.g., tokenization, stop word removal) needs to be tai- lored to the specific model and dataset. Despite their ability to capture latent correlations and rec- ognize implicit patterns (LeCun et al., 2015), com- plex deep neural networks may be overkill for sim- ple tasks such as topic classification, and lighter alternatives are usually good enough. For exam- ple, Adhikari et al. (2019b) find that a simple long short-term memory network (LSTM; Hochreiter and Schmidhuber, 1997) with appropriate regular- ization can achieve competitive results. Shen et al.(2018) further show that even word-embedding- based methods can achieve results comparable to convolutional neural networks (CNNs) and recur- rent neural networks (RNNs). Among all the endeavors for a lighter alternative to DNNs, one stream of work focuses on using com- pressors for text classification. There have been several studies in this field (Teahan and Harper, 2003; Frank et al., 2000), most of them based on the intuition that the minimum cross entropy be- tween a document and a language model of a class built by a compressor indicates the class of the document. However, previous works fall short of matching the quality of neural networks. Addressing these shortcomings, we propose a text classification method combining a lossless compressor, a compressor-based distance metric with a k-nearest-neighbor classifier ( kNN). It uti- lizes compressors in capturing regularity, which is then translated into similarity scores by a compressor-based distance metric. With the re- sulting distance matrix, we use kNN to perform classification. We carry out experiments on seven in-distribution datasets and five out-of-distribution ones. With a simple compressor like gzip, our method achieves results competitive with those of DNNs on six out of seven datasets and outperforms all methods including BERT on all OOD datasets. It also surpasses all models by a large margin under few-shot settings. Our contributions are as follows: (1) we are the first to use NCD with kNN for topic classifica- tion, allowing us to carry out comprehensive ex- periments on large datasets with compressor-based methods; (2) we show that our method achieves results comparable to non-pretrained DNNs on six out of seven in-distribution datasets; (3) on OOD datasets, we show that our method outperforms all methods, including pretrained models such as BERT; and (4) we demonstrate that our method ex- cels in the few-shot setting of scarce labeled data.6810

1911.00172.pdf

Published as a conference paper at ICLR 2020 GENERALIZATION THROUGH MEMORIZATION : NEAREST NEIGHBOR LANGUAGE MODELS Urvashi Khandelwal†∗, Omer Levy‡, Dan Jurafsky†, Luke Zettlemoyer‡& Mike Lewis‡ †Stanford University ‡Facebook AI Research {urvashik,jurafsky }@stanford.edu {omerlevy,lsz,mikelewis }@fb.com ABSTRACT We introduce kNN-LMs, which extend a pre-trained neural language model (LM) by linearly interpolating it with a k-nearest neighbors ( kNN) model. The near- est neighbors are computed according to distance in the pre-trained LM embed- ding space, and can be drawn from any text collection, including the original LM training data. Applying this augmentation to a strong W IKITEXT -103 LM, with neighbors drawn from the original training set, our kNN-LM achieves a new state- of-the-art perplexity of 15.79 – a 2.9 point improvement with no additional train- ing. We also show that this approach has implications for efficiently scaling up to larger training sets and allows for effective domain adaptation, by simply varying the nearest neighbor datastore, again without further training. Qualitatively, the model is particularly helpful in predicting rare patterns, such as factual knowl- edge. Together, these results strongly suggest that learning similarity between se- quences of text is easier than predicting the next word, and that nearest neighbor search is an effective approach for language modeling in the long tail. 1 I NTRODUCTION Neural language models (LMs) typically solve two subproblems: (1) mapping sentence prefixes to fixed-sized representations, and (2) using these representations to predict the next word in the text (Bengio et al., 2003; Mikolov et al., 2010). We present a new language modeling approach that is based on the hypothesis that the representation learning problem may be easier than the prediction problem. For example, any English speaker knows that Dickens is the author of andDickens wrote will have essentially the same distribution over the next word, even if they do not know what that distribution is. We provide strong evidence that existing language models, similarly, are much better at the first problem, by using their prefix embeddings in a simple nearest neighbor scheme that significantly improves overall performance. We introduce kNN-LM, an approach that extends a pre-trained LM by linearly interpolating its next word distribution with a k-nearest neighbors ( kNN) model. The nearest neighbors are computed according to distance in the pre-trained embedding space and can be drawn from any text collec- tion, including the original LM training data. This approach allows rare patterns to be memorized explicitly, rather than implicitly in model parameters. It also improves performance when the same training data is used for learning the prefix representations and the kNN model, strongly suggesting that the prediction problem is more challenging than previously appreciated. To better measure these effects, we conduct an extensive empirical evaluation. Applying our kNN augmentation to a strong W IKITEXT -103 LM using only the original dataset achieves a new state- of-the-art perplexity of 15.79 – a 2.86 point improvement over the base model (Baevski & Auli, 2019) – with no additional training. We also show that the approach has implications for efficiently scaling up to larger training sets and allows for effective domain adaptation, by simply varying the nearest neighbor datastore. Training a model on 100-million tokens and using kNN search over a 3-billion token dataset can outperform training the same model on all 3-billion tokens, opening a ∗Work done while the first author was interning at Facebook AI Research. 1arXiv:1911.00172v2 [cs.CL] 15 Feb 2020

2024.03.18.585544v1.full.pdf

1 Towards Interpretable Cryo-EM: Disentangling Latent Spaces of Molecular Conformations David A. Klindt1,2,∗, Aapo Hyv ¨arinen3, Axel Levy1,4, Nina Miolane2and Fr´ed´eric Poitevin1 1LCLS, SLAC National Accelerator Laboratory, Stanford University, CA, USA 2Department of Electrical and Computer Engineering, UCSB, CA, USA 3Department of Computer Science, University of Helsinki, Finland 4Department of Electrical Engineering, Stanford, CA, USA Correspondence*: David A. Klindt klindt.david@gmail.com ABSTRACT2 Molecules are essential building blocks of life and their different conformations (i.e., shapes) 3 crucially determine the functional role that they play in living organisms. Cryogenic Electron4 Microscopy (cryo-EM) allows for acquisition of large image datasets of individual molecules.5 Recent advances in computational cryo-EM have made it possible to learn latent variable models6 of conformation landscapes. However, interpreting these latent spaces remains a challenge7 as their individual dimensions are often arbitrary. The key message of our work is that this8 interpretation challenge can be viewed as an Independent Component Analysis (ICA) problem9 where we seek models that have the property of identifiability. That means, they have an10 essentially unique solution, representing a conformational latent space that separates the 11 different degrees of freedom a molecule is equipped with in nature. Thus, we aim to advance 12 the computational field of cryo-EM beyond visualizations as we connect it with the theoretical 13 framework of (nonlinear) ICA and discuss the need for identifiable models, improved metrics, and 14 benchmarks. Moving forward, we propose future directions for enhancing the disentanglement 15 of latent spaces in cryo-EM, refining evaluation metrics and exploring techniques that leverage 16 physics-based decoders of biomolecular systems. Moreover, we discuss how future technological 17 developments in time-resolved single particle imaging may enable the application of nonlinear ICA 18 models that can discover the true conformation changes of molecules in nature. The pursuit of 19 interpretable conformational latent spaces will empower researchers to unravel complex biological 20 processes and facilitate targeted interventions. This has significant implications for drug discovery 21 and structural biology more broadly. More generally, latent variable models are deployed widely 22 across many scientific disciplines. Thus, the argument we present in this work has much broader 23 applications in AI for science if we want to move from impressive nonlinear neural network models 24 to mathematically grounded methods that can help us learn something new about nature. 25 Keywords: cryo-EM, machine learning, ICA, AI for science, disentanglement, physics-based models 26 1(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint this version posted March 19, 2024. ; https://doi.org/10.1101/2024.03.18.585544doi: bioRxiv preprint

2309.03649.pdf

Exploring kinase DFG loop conformational stability with AlphaFold2-RAVE Bodhi P. Vani,†Akashnathan Aranganathan,‡and Pratyush Tiwary∗,¶,§ †Institute for Physical Science and Technology, University of Maryland, College Park, Maryland 20742, USA ‡Biophysics Program and Institute for Physical Science and Technology, University of Maryland, College Park 20742, USA ¶Department of Chemistry and Biochemistry and Institute for Physical Science and Technology, University of Maryland, College Park 20742, USA §Corresponding author E-mail: ptiwary@umd.edu Abstract Kinases compose one of the largest fractions of the human proteome, and their misfunction is implicated in many diseases, in particular cancers. The ubiquitousness and structural similarities of kinases makes specific and effective drug design difficult. In particular, conformational variability due to the evolutionarily conserved DFG mo- tif adopting in and out conformations and the relative stabilities thereof are key in structure-based drug design for ATP competitive drugs. These relative conformational stabilities are extremely sensitive to small changes in sequence, and provide an impor- tant problem for sampling method development. Since the invention of AlphaFold2, the world of structure-based drug design has noticably changed. In spite of it being limited to crystal-like structure prediction, several methods have also leveraged its underlying 1arXiv:2309.03649v1 [physics.bio-ph] 7 Sep 2023

NIPS-2007-active-preference-learning-with-discrete-choice-data-Paper.pdf

Active Preference Learning with Discrete Choice Data Eric Brochu, Nando de Freitas and Abhijeet Ghosh Department of Computer Science University of British Columbia Vancouver, BC, Canada {ebrochu, nando, ghosh}@cs.ubc.ca Abstract We propose an active learning algorithm that learns a continuous valuation model from discrete preferences. The algorithm automatically decides what items are best presented to an individual in order to find the item that they value highly in as few trials as possible, and exploits quirks of human psychology to minimize time and cognitive burden. To do this, our algorithm maximizes the expected improvement at each query without accurately modelling the entire valuation sur- face, which would be needlessly expensive. The problem is particularly difficult because the space of choices is infinite. We demonstrate the effectiveness of the new algorithm compared to related active learning methods. We also embed the algorithm within a decision making tool for assisting digital artists in rendering materials. The tool finds the best parameters while minimizing the number of queries. 1 Introduction A computer graphics artist sits down to use a simple renderer to find appropriate surfaces for a typical reflectance model. It has a series of parameters that must be set to control the simulation: “specularity”, “Fresnel reflectance coefficient”, and other, less-comprehensible ones. The parame- ters interact in ways difficult to discern. The artist knows in his mind’s eye what he wants, but he’s not a mathematician or a physicist — no course he took during his MFA covered Fresnel reflectance models. Even if it had, would it help? He moves the specularity slider and waits for the image to be generated. The surface is too shiny. He moves the slider back a bit and runs the simulation again. Better. The surface is now appropriately dull, but too dark. He moves a slider down. Now it’s the right colour, but the specularity doesn’t look quite right any more. He repeatedly bumps the specularity back up, rerunning the renderer at each attempt until it looks right. Good. Now, how to make it look metallic...? Problems in simulation, animation, rendering and other areas often take such a form, where the desired end result is identifiable by the user, but parameters must be tuned in a tedious trial-and- error process. This is particularly apparent in psychoperceptual models, where continual tuning is required to make something “look right”. Using the animation of character walking motion as an example, for decades, animators and scientists have tried to develop objective functions based on kinematics, dynamics and motion capture data [Cooper et al., 2007 ]. However, even when expen- sive mocap is available, we simply have to watch an animated film to be convinced of how far we still are from solving the gait animation problem. Unfortunately, it is not at all easy to find a mapping from parameterized animation to psychoperceptual plausibility. The perceptual objective function is simply unknown. Fortunately, however, it is fairly easy to judge the quality of a walk — in fact, it is trivial and almost instantaneous. The application of this principle to animation and other psychoper- ceptual tools is motivated by the observation that humans often seem to be forming a mental model of the objective function. This model enables them to exploit feasible regions of the parameter space where the valuation is predicted to be high and to explore regions of high uncertainty. It is our the- 1

2206.14858.pdf

Solving Quantitative Reasoning Problems with Language Models Aitor Lewkowycz∗, Anders Andreassen†, David Dohan†, Ethan Dyer†, Henryk Michalewski†, Vinay Ramasesh†, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur∗, Guy Gur-Ari∗, and Vedant Misra∗ Google Research Abstract Language models have achieved remarkable performance on a wide range of tasks that require natural language understanding. Nevertheless, state-of-the-art models have generally struggled with tasks that require quantitative reasoning, such as solving mathematics, science, and engineering problems at the college level. To help close this gap, we introduce Minerva, a large language model pretrained on general natural language data and further trained on technical content. The model achieves state-of-the-art performance on technical benchmarks without the use of external tools. We also evaluate our model on over two hundred undergraduate-level problems in physics, biology, chemistry, economics, and other sciences that require quantitative reasoning, and find that the model can correctly answer nearly a third of them. 1 Introduction Artificial neural networks have seen remarkable success in a variety of domains including computer vision, speech recognition, audio and image generation, translation, game playing, and robotics. In particular, large language models have achieved excellent performance across a variety of natural language tasks including common-sense reasoning, question answering, and summarization (Raffel et al., 2019; Brown et al., 2020; Rae et al., 2021; Smith et al., 2022; Chowdhery et al., 2022). However, these models have struggled with tasks that require quantitative reasoning, such as solving mathematics, science, and engineering problems (Hendrycks et al., 2021; Cobbe et al., 2021). Quantitative reasoning problems are an interesting domain of application for language models because they test the capability of models on several fronts. They require the solver to correctly parse a natural language input, potentially recall world knowledge that pertains to the problem, and apply an algorithm or series of computations to the information provided in order to arrive at a correct solution. They also require that the solver is able to correctly parse and generate precise sequences of mathematical tokens, as well as apply a computational procedure to tokens via symbolic and numerical manipulation. Finally, such problems are a proving ground for research toward robust quantitative reasoning solvers that are useful in supporting the work of humans in scientific and technical fields. Previous research has shown that large language models achieve impressive performance on math and programming questions after training on domain specific datasets (Chen et al., 2021; Austin et al., 2021; ∗Equal leadership and advising contribution †Equal contribution 1arXiv:2206.14858v2 [cs.CL] 1 Jul 2022

1909.12264.pdf

Quantum Graph Neural Networks Guillaume Verdon X, The Moonshot Factory Mountain View, CA gverdon@x.teamTrevor McCourt Google Research Venice, CA trevormccrt@google.com Enxhell Luzhnica, Vikash Singh, Stefan Leichenauer, Jack Hidary X, The Moonshot Factory Mountain View, CA {enxhell,singvikash, sleichenauer,hidary}@x.team Abstract We introduce Quantum Graph Neural Networks ( QGNN ), a new class of quantum neural network ansatze which are tailored to represent quantum processes which have a graph structure, and are particularly suitable to be executed on distributed quantum systems over a quantum network. Along with this general class of ansatze, we introduce further specialized architectures, namely, Quantum Graph Recurrent Neural Networks ( QGRNN ) and Quantum Graph Convolutional Neural Networks (QGCNN ). We provide four example applications of QGNN s: learning Hamiltonian dynamics of quantum systems, learning how to create multipartite entanglement in a quantum network, unsupervised learning for spectral clustering, and supervised learning for graph isomorphism classification. 1 Introduction Variational Quantum Algorithms are a promising class of algorithms that are rapidly emerging as a central subfield of Quantum Computing [ 1,2,3]. Similar to parameterized transformations encountered in deep learning, these parameterized quantum circuits are often referred to as Quantum Neural Networks (QNNs). Recently, it was shown that QNNs that have no prior on their structure suffer from a quantum version of the no-free lunch theorem [ 4] and are exponentially difficult to train via gradient descent. Thus, there is a need for better QNN ansatze. One popular class of QNNs has been Trotter-based ansatze [ 2,5]. The optimization of these ansatze has been extensively studied in recent works, and efficient optimization methods have been found [ 6,7]. On the classical side, graph-based neural networks leveraging data geometry have seen some recent successes in deep learning, finding applications in biophysics and chemistry [ 8]. Inspired from this success, we propose a new class of Quantum Neural Network ansatz which allows for both quantum inference and classical probabilistic inference for data with a graph-geometric structure. In the sections below, we introduce the general framework of the QGNN ansatz as well as several more specialized variants and showcase four potential applications via numerical implementation. Preprint. Under review.arXiv:1909.12264v1 [quant-ph] 26 Sep 2019

2403.08763.pdf

Simple and Scalable Strategies to Continually Pre-train Large Language Models Adam Ibrahim∗†⊚ibrahima@mila.quebec Benjamin Thérien∗†⊚benjamin.therien@mila.quebec Kshitij Gupta∗†⊚kshitij.gupta@mila.quebec Mats L. Richter†⊚mats.richter@mila.quebec Quentin Anthony♢†⊚qubitquentin@gmail.com Timothée Lesort†⊚t.lesort@gmail.com Eugene Belilovsky‡⊚eugene.belilovsky@concordia.ca Irina Rish†⊚irina.rish@umontreal.ca Department of Computer Science and Operation Research, Université de Montréal, Montréal, Canada † Department of Computer Science and Software Engineering, Concordia University, Montréal, Canada ‡ Mila, Montréal, Canada ⊚ EleutherAI ♢ Abstract Large language models (LLMs) are routinely pre-trained on billions of tokens, only to start the process over again once new data becomes available. A much more efficient solution is to continually pre-train these models – saving significant compute compared to re-training. However, the distribution shift induced by new data typically results in degraded performance on previous data or poor adaptation to the new data. In this work, we show that a simple and scalable combination of learning rate (LR) re-warming, LR re-decaying, and replay of previous data is sufficient to match the performance of fully re-training from scratch on all available data, as measured by final loss and language model (LM) evaluation benchmarks. Specifically, we show this for a weak but realistic distribution shift between two commonly used LLM pre-training datasets (English →English) and a stronger distribution shift (English →German) at the 405M parameter model scale with large dataset sizes (hundreds of billions of tokens). Selecting the weak but realistic shift for larger-scale experiments, we also find that our continual learning strategies match the re-training baseline for a 10B parameter LLM. Our results demonstrate that LLMs can be successfully updated via simple and scalable continual learning strategies, matching the re-training baseline using only a fraction of the compute. Finally, inspired by previous work, we propose alternatives to the cosine learning rate schedule that help circumvent forgetting induced by LR re-warming and that are not bound to a fixed token budget. 1 Introduction Over the past few years, large pre-trained models have enabled massive performance improvements in language modeling (Brown et al., 2020; Zhao et al., 2023), visual understanding (Radford et al., 2021; Alayrac et al., 2022; Kirillov et al., 2023), text-to-image generation (Rombach et al., 2022; Pernias et al., 2024), and text-to-video generation (Brooks et al., 2024)—to name a few. Large language models (LLMs) are at the center of all these improvements, providing an intuitive means for humans to interface with machine learning algorithms through language. ∗Equal contribution; authorship order within equal contributors was randomized. 1arXiv:2403.08763v1 [cs.LG] 13 Mar 2024

2310.02226.pdf

Think before you speak: Training Language Models With Pause Tokens Sachin Goyal∗ Machine Learning Department Carnegie Mellon University sachingo@andrew.cmu.eduZiwei Ji Google Research, NY ziweiji@google.comAnkit Singh Rawat Google Research, NY ankitsrawat@google.com Aditya Krishna Menon Google Research, NY adityakmenon@google.comSanjiv Kumar Google Research, NY sanjivk@google.comVaishnavh Nagarajan Google Research, NY vaishnavh@google.com Abstract Language models generate responses by producing a series of tokens in immediate succession: the (K+ 1)thtoken is an outcome of manipulating Khidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, K+10 hidden vectors, before it outputs the (K+1)thtoken? We operationalize this idea by performing training and inference on language mod- els with a (learnable) pause token, a sequence of which is appended to the input prefix. We then delay extracting the model’s outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate pause-training on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact re- call. Our main finding is that inference-time delays show gains on our tasks when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on eight tasks, most prominently, a gain of 18% EM score on the QA task of SQuAD, 8%on CommonSenseQA and 1%accuracy on the reason- ing task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm. 1 Introduction Transformer-based causal language models generate tokens one after the other in immediate succes- sion. To generate the (K+ 1)thtoken, the model consumes the Kprevious tokens, and proceeds layer by layer, computing Kintermediate vectors in each hidden layer. Each vector in itself is the output of a module (consisting of self-attention and multi-layer-perceptrons) operating on the pre- vious layer’s output vectors. However sophisticated this end-to-end process may be, it abides by a peculiar constraint: the number of operations determining the next token is limited by the number of tokens seen so far. Arguably, this was the most natural design choice when the Transformer was first conceived by Vaswani et al. (2017). But in hindsight, one may wonder whether for some inputs, the(K+ 1)thtoken demands K+MTransformer operations in each layer (for M > 0), which cannot be met by the arbitrarily constrained Koperations per layer. This paper explores one way to free the Transformer of this arbitrary per-layer computational constraint. The approach we study is to append dummy tokens into a decoder-only model’s input, thereby de- laying the model’s output. Specifically, we select a (learnable) pause token (denoted <pause> ) and append one or more copies of <pause> as a sequence to the input. We simply ignore the model’s cor- responding outputs until the last <pause> token is seen, after which we begin extracting its response. ∗Work done in part as a Student Researcher at Google. 1arXiv:2310.02226v1 [cs.CL] 3 Oct 2023

2212.00178.pdf

Open Relation and Event Type Discovery with Type Abstraction Sha Li, Heng Ji, Jiawei Han University of Illinois Urbana-Champaign {shal2, hengji, hanj}@illinois.edu Abstract Conventional “closed-world" information ex- traction (IE) approaches rely on human ontolo- gies to define the scope for extraction. As a result, such approaches fall short when ap- plied to new domains. This calls for systems that can automatically infer new types from given corpora, a task which we refer to as type discovery . To tackle this problem, we intro- duce the idea of type abstraction, where the model is prompted to generalize and name the type. Then we use the similarity between in- ferred names to induce clusters. Observing that this abstraction-based representation is of- ten complementary to the entity/trigger token representation, we set up these two represen- tations as two views and design our model as a co-training framework. Our experiments on multiple relation extraction and event extrac- tion datasets consistently show the advantage of our type abstraction approach. 1 Introduction Information extraction has enjoyed widespread suc- cess, however, the majority of information extrac- tion methods are “reactive”, relying on end-users to specify their information needs in prior and pro- vide supervision accordingly. This leads to “closed- world” systems (Lin et al., 2020; Du and Cardie, 2020; Li et al., 2021; Zhong and Chen, 2021; Ye et al., 2022) that are confined to a set of pre-defined types. It is desirable to make systems act more “proactively” like humans who are always on the lookout for interesting new information, generalize them into new types, and find more instances of such types, even if they are not seen previously. One related attempt is the Open Information Ex- traction paradigm (Banko et al., 2008), which aims at extracting all (subject, predicate, object) triples from text that denote some kind of relation. While OpenIE does not rely on pre-specified relations, its exhaustive and free-form nature often leads to noisy and redundant extractions. <h>John</h> earned a bachelor’s degree from the <t>University of Wollongong</t>.Token ViewUniversity of Wollongong is the [MASK] of John. Mask ViewRelation: School_AttendedFigure 1: For each instance, the token view is computed from the pre-trained LM embedding of the first token in entity/trigger. The mask view is computed from the [MASK] token embedding in the type prompt. To bridge the gap between closed-world IE and OpenIE, a vital step is for systems to possess the ability of automatically inducing new types and extracting instances of such new types. Under vari- ous contexts, related methods have been proposed under the name of “relation discovery” (Yao et al., 2011; Marcheggiani and Titov, 2016),“open rela- tion extraction” (Wu et al., 2019; Hu et al., 2020) and “event type induction” (Huang and Ji, 2020; Shen et al., 2021). In this paper, we unify such terms and refer to the task as type discovery . Type discovery can naturally be posed as a clus- tering task. This heavily relies on defining an appro- priate metric space where types are easily separable. The token embedding space from pre-trained lan- guage models is a popular choice, but as observed by (Zhao et al., 2021), the original metric space derived from BERT (Devlin et al., 2019) is often prone to reflect surface form similarity rather than the desired relation/event-centered similarity. One way to alleviate this issue is to use known types to help learn a similarity metric that can also be applied to unknown types (Wu et al., 2019; Zhao et al., 2021; Huang and Ji, 2020). In this paper we introduce another idea of ab- straction : a discovered type should have an ap- propriate and concise type name. The human vo- cabulary serves as a good repository of concepts that appear meaningful to people. When we assign a name to a cluster, we implicitly define the com-arXiv:2212.00178v1 [cs.CL] 30 Nov 2022

10.1016.j.cell.2023.12.037.pdf

Article Xist ribonucleoproteins promote female sex-biased autoimmunity Graphical abstract Highlights dTransgenic mouse models inducibly express Xist in male animals dXist expression in males induces autoantibodies andautoimmune pathology dXist in males reprograms T and B cell populations to female-like patterns dAutoantibodies to Xist RNP characterize female-biasedautoimmune diseases in patientsAuthors Diana R. Dou, Yanding Zhao,Julia A. Belk, ..., Anton Wutz, Paul J. Utz,Howard Y. Chang Correspondence howchang@stanford.edu In brief The Xist RNA protein complex, presentonly in females, is immunogenic and mayunderlie female-biased autoimmunity. Dou et al., 2024, Cell 187, 733–749 February 1, 2024 ª2024 The Authors. Published by Elsevier Inc. https://doi.org/10.1016/j.cell.2023.12.037 ll

2012.02296v2.pdf

Generative Capacity of Probabilistic Protein Sequence Models Francisco McGee1,2,4, Quentin Novinger2,5, Ronald M Levy1,3,4,6, Vincenzo Carnevale2,3,*, and Allan Haldane1,6,* 1Center for Biophysics and Computational Biology, Temple University, Philadelphia, 19122, USA 2Institute for Computational Molecular Science, Temple University, Philadelphia, 19122, USA 3Department of Biology, Temple University, Philadelphia, 19122, USA 4Department of Chemistry, Temple University, Philadelphia, 19122, USA 5Department of Computer & Information Sciences, Temple University, Philadelphia, 19122, USA 6Department of Physics, Temple University, Philadelphia, 19122, USA *Corresponding authors: vincenzo.carnevale@temple.edu, allan.haldane@temple.edu ABSTRACT Potts models and variational autoencoders (VAEs) have recently gained popularity as generative protein sequence models (GPSMs) to explore fitness landscapes and predict the effect of mutations. Despite encouraging results, quantitative characteri- zation and comparison of GPSM-generated probability distributions is still lacking. It is currently unclear whether GPSMs can faithfully reproduce the complex multi-residue mutation patterns observed in natural sequences arising due to epistasis. We develop a set of sequence statistics to assess the “generative capacity” of three GPSMs of recent interest: the pairwise Potts Hamiltonian, the VAE, and the site-independent model, using natural and synthetic datasets. We show that the generative capacity of the Potts Hamiltonian model is the largest; the higher order mutational statistics generated by the model agree with those observed for natural sequences. In contrast, we show that the VAE’s generative capacity lies between the pairwise Potts and site-independent models. Importantly, our work measures GPSM generative capacity in terms of higher-order sequence covariation statistics which we have developed, and provides a new framework for evaluating and interpreting GPSM accuracy that emphasizes the role of epistasis. Introduction Recent progress in decoding the patterns of mutations in protein multiple sequence alignments (MSAs) has highlighted the importance of mutational covariation in determining protein function, conformation and evolution, and has found practical applications in protein design, drug design, drug resistance prediction, and classification1–3. These developments were sparked by the recognition that the pairwise covariation of mutations observed in large MSAs of evolutionarily diverged sequences belonging to a common protein family can be used to fit maximum entropy “Potts” statistical models4–6. These contain pairwise statistical interaction parameters reflecting epistasis7 between pairs of positions. Such models have been shown to accurately predict physical contacts in protein structure6,8–10, and have been used to significantly improve the prediction of the fitness effect of mutations to a sequence compared to site-independent sequence variation models which do not account for covariation11,12. They are “generative” in the sense that they define the probability, p(S), that a protein sequence Sresults from the evolutionary process. Intriguingly, the probability distribution p(S)can be used to sample unobserved, and yet viable, artificial sequences. In practice, the model distribution p(S)depends on parameters that are found by maximizing a suitably defined likelihood function on observations provided by the MSA of a target protein family. As long as the model is well specified and generalizes from the training MSA, it can then be used to generate new sequences, and thus a new MSA whose statistics should match those of the original target protein family. We refer to probabilistic models that create new protein sequences in this way as generative protein sequence models (GPSMs). The fact that Potts maximum entropy models are limited to pairwise epistatic interaction terms and have a simple functional form for p(S)raises the possibility that their functional form is not flexible enough to describe the data, i.e. that the model is not well specified. While a model with only pairwise interaction terms can predict complex patterns of covariation involving three or more positions through chains of pairwise interactions, it cannot model certain triplet and higher patterns of covariation that require a model with more than pairwise interaction terms13. For example, a Potts model cannot predict patterns described by an XOR or boolean parity function in which the 1arXiv:2012.02296v2 [cs.LG] 15 Mar 2021

2401.00368.pdf

Improving Text Embeddings with Large Language Models Liang Wang∗, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, Furu Wei Microsoft Corporation https://aka.ms/GeneralAI Abstract In this paper, we introduce a novel and simple method for obtaining high-quality text embeddings using only synthetic data and less than 1k training steps. Unlike existing methods that often depend on multi-stage intermediate pre-training with billions of weakly-supervised text pairs, followed by fine-tuning with a few labeled datasets, our method does not require building complex training pipelines or relying on manually collected datasets that are often constrained by task diversity and language coverage. We leverage proprietary LLMs to generate diverse synthetic data for hundreds of thousands of text embedding tasks across nearly 100languages. We then fine-tune open-source decoder-only LLMs on the synthetic data using standard contrastive loss. Experiments demonstrate that our method achieves strong performance on highly competitive text embedding benchmarks without using any labeled data. Furthermore, when fine-tuned with a mixture of synthetic and labeled data, our model sets new state-of-the-art results on the BEIR and MTEB benchmarks. 1 Introduction Text embeddings are vector representations of natural language that encode its semantic information. They are widely used in various natural language processing (NLP) tasks, such as information retrieval (IR), question answering, semantic textual similarity, bitext mining, item recommendation, etc. In the field of IR, the first-stage retrieval often relies on text embeddings to efficiently recall a small set of candidate documents from a large-scale corpus using approximate nearest neighbor search techniques. Embedding-based retrieval is also a crucial component of retrieval-augmented generation (RAG) [ 21], which is an emerging paradigm that enables large language models (LLMs) to access dynamic external knowledge without modifying the model parameters. Source attribution of generated text is another important application of text embeddings [ 14] that can improve the interpretability and trustworthiness of LLMs. Previous studies have demonstrated that weighted average of pre-trained word embeddings [ 35,1] is a strong baseline for measuring semantic similarity. However, these methods fail to capture the rich contextual information of natural language. With the advent of pre-trained language models [11], Sentence-BERT [ 37] and SimCSE [ 13] have been proposed to learn text embeddings by fine- tuning BERT on natural language inference (NLI) datasets. To further enhance the performance and robustness of text embeddings, state-of-the-art methods like E5 [ 46] and BGE [ 48] employ a more complex multi-stage training paradigm that first pre-trains on billions of weakly-supervised text pairs, and then fine-tunes on several labeled datasets. Existing multi-stage approaches suffer from several drawbacks. Firstly, they entail a complex multi-stage training pipeline that demands substantial engineering efforts to curate large amounts ∗Correspondence to {wangliang,nanya,fuwei}@microsoft.com Technical Report.arXiv:2401.00368v2 [cs.CL] 19 Jan 2024

More-Is-Different-Anderson.pdf

The reductionist hypothesis may still lbe a topic for controversy among phi- losophers, but among the great majority of active scientists I think it is accepted without question The workings of our minds and bodles, and of all the ani- mate or lnanimate matter of which we have any detailed knowledges are as sumed to be controlled by the same set o£ fundamental laws which except under certain extreme conditions we feel we know pretty well. It seems inevitable to go on unerit- ically to what appears at first sight to be- an obvious corollary of reduction ism: that if everything obeys the same fundamental laws, then the only sci entists who are studying anything really fundamental are those who are working on those laws. In practice, that amounts to some astrophysicists, some elemen- tary particle physicists, some logicians and other mathematicians, and few others. This point of view, which it is the main purpose of this article to oppose, is expressed in a rather well- known passage by Weisskopf (1): Looking at the development of science in thP Twentieth' Century one can dis tinguish two trends, which I will call sSintensive and "extensive" research, lack- ing a better 'terminology. In short: in- tensive research goes for the fundamental laws, extensive research goes for the ex- The author is a member of the technlical staff of the Bell Telephone Laboratories, Murray Hill, New Je1 sey 07974, and visiting professor of theoretical physics at Cavendish Laboratory, Cambridge, England. This article is an expanded version of a Regents' Lecture given in 1967 at the University of California, La Jolla. 4 AUGUST 1972 4 August 1972, Volume 177, Number 4047 less relevance they seem to have to the very real problems of the rest of sci- ence, much less to those of society. The constructionist hypothesis breaks down when confronted with the twin difficulties of scale and complexity. The behavior of large and complex aggre- gates of elementary particles, it turns out, is not to be understood in terms of a simple extrapolation of the prop- erties of a few particles. Instead, at each level of complexity entirely new properties appear, and the understand- ing of the new behaviors requires re- search which I think is as fundamental in its nature as any other. That is, it seems to me that one may array the sciences roughly linearly in a hierarchy, according to the idea The elementary entities of science X obey the laws of science Y planatlon - of phenomena ;n terms of lnown fundamental laws. As always, dis- tinotions of this kind are not unambiguous, but they are clear in most cases. Solid state physics, plasma physics, and perhaps also biology are extensivee High energy physics and a good part of nuclear physics are intensive. There is always much less intensive research going on than extensive. Once new fundamental laws are discov- ereds a large and ever increasing activity begins in order to apply the discoveries to hitherto unexplained phenomena. Thus, there are two dimensions to basic re- search The frontier of science extends all along a long line from the newest and most modern intenslve research5 over the ex- tensive research recently spawned by the intensive research of yesterday, to the broad and well developed web of exten- sive research activities based on mtensive research of past decades. The effectiveness of this message may be indicated by the fact that I heard it quoted recently by a leader in the field of materials science, who urged the participants at a meeting dedicated to "fundamental problems in condensed matter physics" to accept that there were few or no such problems and that nothing was left but extensive scienceS which he seemed to equate with device . @ englneerlng. The main fallacy in this kind of thinking is that the reductionist hypoth- esis does not by any rneans imply a "constructionist" one: The ability to reduce everything to simple fundamen- tal laws does not imply the ability to start from those laws and reconstruct the universe. In fact, the more the ele-- mentary particle physicists tell us about the nature of the fundamental laws the Xsolid state or many-body physics chemistry mo-lecular biology cell biology . . * psychology * . . soclal sclences y elementary particle physics many-body physics chemistry molecular biology * physlology psychology But this hierarchy does not imply that science X is "just applied Y*" At each stage entirely new laws, concepts, and generalizations are necessary, re- qulring inspiration and creativity to just as great a degree as in the previous one. Psychology is not applied biology, nor s biology applied chemistry. In my own field of many-body physB ics, we are, perhaps, closer to our fun damental, intensive underpinnings than in any other science in which non- trivial complexities occur, and as a re- sult we have begun to formulate a general theory of just how this shift from quantitative to qualitative differ- entiation takes place. This formulation, called the theory of "broken sym- metry," may be of help in making more generally clear the breakdown of the constructionist converse of reduction- ism. I will give an elementary and in complete explanation of these ideas, and then go on to some more general spec- ulative comments about analogies at 393 SCIE:NC1S More Is Different Broken symmetry and the nature of the hierarchical structure of science P. W. Anderson

2002.11557v1.pdf

Query-Efficient Correlation Clustering David García–Soriano d.garcia.soriano@isi.it ISI Foundation Turin, ItalyKonstantin Kutzkov kutzkov@gmail.com Amalfi Analytics Barcelona, Spain Francesco Bonchi francesco.bonchi@isi.it ISI Foundation, Turin, Italy Eurecat, Barcelona, SpainCharalampos Tsourakakis ctsourak@bu.edu Boston University USA ABSTRACT Correlation clustering is arguably the most natural formulation of clustering. Given nobjects and a pairwise similarity measure, the goal is to cluster the objects so that, to the best possible extent, similar objects are put in the same cluster and dissimilar objects are put in different clusters. A main drawback of correlation clustering is that it requires as input the Θ(n2)pairwise similarities. This is often infeasible to compute or even just to store. In this paper we study query- efficient algorithms for correlation clustering. Specifically, we devise a correlation clustering algorithm that, given a budget of Qqueries, attains a solution whose expected number of disagreements is at most 3·OPT+O(n3 Q), where OPT is the optimal cost for the instance. Its running time is O(Q), and can be easily made non-adaptive (meaning it can specify all its queries at the outset and make them in parallel) with the same guarantees. Up to constant factors, our algorithm yields a provably optimal trade-off between the number of queries Qand the worst-case error attained, even for adaptive algorithms. Finally, we perform an experimental study of our proposed method on both synthetic and real data, showing the scalability and the accuracy of our algorithm. CCS CONCEPTS •Theory of computation →Graph algorithms analysis ;Fa- cility location and clustering ;Active learning ; KEYWORDS correlation clustering, active learning, query complexity, algorithm design ACM Reference Format: David García–Soriano, Konstantin Kutzkov, Francesco Bonchi, and Char- alampos Tsourakakis. 2020. Query-Efficient Correlation Clustering. In Pro- ceedings of The Web Conference 2020 (WWW ’20), April 20–24, 2020, Taipei, Taiwan. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3366423. 3380220 This paper is published under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license. Authors reserve their rights to disseminate the work on their personal and corporate Web sites with the appropriate attribution. WWW ’20, April 20–24, 2020, Taipei, Taiwan ©2020 IW3C2 (International World Wide Web Conference Committee), published under Creative Commons CC-BY 4.0 License. ACM ISBN 978-1-4503-7023-3/20/04. https://doi.org/10.1145/3366423.33802201 INTRODUCTION Correlation clustering [3] (or cluster editing ) is a prominent cluster- ing framework where we are given a set V=[n]and a symmetric pairwise similarity function sim:V 2
→{0,1}, whereV 2
is the set of unordered pairs of elements of V. The goal is to cluster the items in such a way that, to the best possible extent, similar ob- jects are put in the same cluster and dissimilar objects are put in different clusters. Assuming that cluster identifiers are represented by natural numbers, a clustering ℓis a function ℓ:V→N, and each cluster is a maximal set of vertices sharing the same label. Correlation clustering aims at minimizing the following cost: cost(ℓ)=Õ (x,y)∈(V 2), ℓ(x)=ℓ(y)(1−sim(x,y))+Õ (x,y)∈(V 2), ℓ(x),ℓ(y)sim(x,y).(1) The intuition underlying the above problem definition is that if two objects xandyare dissimilar and are assigned to the same cluster we should pay a cost of 1, i.e., the amount of their dissimi- larity. Similarly, if x,yare similar and they are assigned to different clusters we should pay also cost 1, i.e., the amount of their similarity sim(x,y). The correlation clustering framework naturally extends to non-binary, symmetric function, i.e., sim:V 2
→[0,1]. In this paper we focus on the binary case; the general non-binary case can be efficiently reduced to this case at a loss of only a constant factor in the approximation [ 3, Thm. 23]. The binary setting can be viewed very conveniently through graph-theoretic lenses: the n items correspond to the vertices of a similarity graph G, which is a complete undirected graph with edges labeled “+” or “-”. An edge e causes a disagreement (ofcost1) between the similarity graph and a clustering when it is a “+” edge connecting vertices in different clusters, or a “–” edge connecting vertices within the same cluster. If we were given a cluster graph [22], i.e., a graph whose set of positive edges is the union of vertex-disjoint cliques, we would be able to produce a perfect (i.e., cost 0) clustering simply by computing the connected components of the positive graph. However, similarities will generally be inconsistent with one another, so incurring a cer- tain cost is unavoidable. Correlation clustering aims at minimizing such cost. The problem may be viewed as the task of finding the equivalence relation that most closely resembles a given symmetric relation. The correlation clustering problem is NP-hard [3, 22].arXiv:2002.11557v1 [cs.DS] 26 Feb 2020

10.1093.gbe.evad084.pdf

Unsupervised Deep Learning Can Identify Protein Functional Groups from Unaligned Sequences Kyle T. David 1,* and Kenneth M. Halanych 2 1Department of Biological Sciences, Auburn University, Auburn, Alabama, USA 2Center for Marine Sciences, University of North Carolina Wilmington, Wilmington, North Carolina, USA *Corresponding author: E-mail: kzd0038@auburn.edu . Accepted: 13 May 2023 Abstract Interpreting protein function from sequence data is a fundamental goal of bioinformatics. However, our current understand - ing of protein diversity is bottlenecked by the fact that most proteins have only been functionally validated in model organ - isms, limiting our understanding of how function varies with gene sequence diversity. Thus, accuracy of inferences in clades without model representatives is questionable. Unsupervised learning may help to ameliorate this bias by identifying highly complex patterns and structure from large data sets without external labels. Here, we present DeepSeqProt, an unsupervised deep learning program for exploring large protein sequence data sets. DeepSeqProt is a clustering tool capable of distinguish - ing between broad classes of proteins while learning local and global structure of functional space. DeepSeqProt is capable of learning salient biological features from unaligned, unannotated sequences. DeepSeqProt is more likely to capture complete protein families and statistically significant shared ontologies within proteomes than other clustering methods. We hope this framework will prove of use to researchers and provide a preliminary step in further developing unsupervised deep learning in molecular biology. Key words: machine learning, protein annotation, bioinformatics. Introduction As sequencing technology continues to improve, there is an ever-increasing need to adequately annotate and charac - terize novel protein sequences and their predicted func- tions. With thousands of new sequences being uploaded every day, predicting the function of every protein directly with conventional experimental studies such as gene knockouts or assays is not possible. Thus, attempting to in- fer protein function automatically is necessary. Many such methods exist but fundamentally operate the same way: by matching the sequence of a protein with unknown func- tion to a reference sequence of a protein with known func- tion and then assuming that functions are the same. These Significance In this manuscript, we report the results of a new unsupervised machine learning software, DeepSeqProt. Unsupervised methods offer several advantages which can help escape longstanding pitfalls and biases pervading computational mo- lecular biology. DeepSeqProt learns from and processes unaligned protein sequences with the goal of clustering them into informative groups with regard to protein family and function, as well as distributing the clusters themselves in a lower dimension space. We discovered that unsupervised deep learning is capable of recognizing patterns shared among proteins of similar families and functional affinities, exceeding conventional sequence similarity-based clustering in some scenarios. DeepSeqProt has broad applications for computational molecular biology and may be especially use- ful for nonmodel organisms. © The Author(s) 2023. Published by Oxford University Press on behalf of Society for Molecular Biology and Evolution. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (https://creativecommons.org/licenses/by-nc/4.0/ ), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact journals.permissions@oup.comGBE Genome Biol. Evol. 15(5) https://doi.org/10.1093/gbe/evad084 Advance Access publication 22 May 2023 1Downloaded from https://academic.oup.com/gbe/article/15/5/evad084/7175204 by guest on 19 February 2024

10.1038.s41467-024-46631-y.pdf

Article https://doi.org/10.1038/s41467-024-46631-y Alignment of brain embeddings and arti ficial contextual embeddings in natural languagepoints to common geometric patterns Ariel Goldstein1,2, Avigail Grinstein-Dabush2,8, Mariano Schain2,8, Haocheng Wang3, Zhuoqiao Hong3, Bobbi Aubrey3,4, Mariano Schain2, Samuel A. Nastase3,Z a i dZ a d a3,E r i cH a m3, Amir Feder2, Harshvardhan Gazula3, Eliav Buchnik2, Werner Doyle4,S a s h aD e v o r e4, Patricia Dugan4, Roi Reichart5,D a n i e lF r i e d m a n4, Michael Brenner2,6, Avinatan Hassidim2, Orrin Devinsky4, Adeen Flinker4,7&U r iH a s s o n2,3 Contextual embeddings, derived from deep language models (DLMs), provide a continuous vectorial representati on of language. This embedding space differs fundamentally from the symbolic representations posited by tradi-tional psycholinguistics. We hypoth esize that language areas in the human brain, similar to DLMs, rely on a con tinuous embedding space to represent language. To test this hypothesis, we densely record the neural activity pat-terns in the inferior frontal gyrus ( IFG) of three participants using dense intracranial arrays while they listen ed to a 30-minute podcast. From these fine- grained spatiotemporal neural recordi ngs, we derive a continuous vectorial representation for each word (i.e., a br a i ne m b e d d i n g )i ne a c hp a t i e n t .U s i n g stringent zero-shot mapping we demonstrate that brain embeddings in the IFGand the DLM contextual embedding sp ace have common geometric patterns. The common geometric patterns allow us to predict the brain embedding in IFG of a given left-out word based sole ly on its geometrical relationship to other non-overlapping words in the podcast. Furthermore, we show that contextual embeddings ca pture the geometry of IFG embeddings better than static word embeddings. The continu ous brain embedding space exposes a vector-based neural code for natural la nguage processing in the human brain. Deep language models (DLMs) trained on massive corpora of natural text provide a radically different framework for how language isrepresented in the brain. The recent success of DLMs in modelingnatural language can be traced to the gradual development of threefoundational ideas in computational linguistics.Thefirst key innovation was to (1) embed words in continuous vector space: Traditionally, words in language were viewed as discretesymbolic units in a lexicon 1,2. Early work in distributional semantics demonstrated that the meaning of words could instead be capturedby geometric relationships in a continuous vector space based onReceived: 24 July 2022 Accepted: 4 March 2024 Check for updates 1Business School, Data Science department and Cognitive Department, Hebrew University, Jerusalem, Israel.2Google Research, Tel Aviv, Israel.3Department of Psychology and the Neuroscience Institute, Princeton University, Princeton, NJ, USA.4New York University Grossman School of Medicine, New York, NY, USA.5Faculty of Industrial Engineering and Management, Technion, Israel Institute of Technology, Haifa, Israel.6School of Engineering and Applied Science, Harvard University, Cambridge, MA, USA.7New York University Tandon School of Engineering, Brooklyn, NY, USA.8These authors contributed equally: Avigail Grinstein-Dabush, Mariano Schain. e-mail: ariel.y.goldstein@mail.huji.ac.il Nature Communications | (2024) 15:2768 11234567890():,; 1234567890():,;