note_id
stringlengths
9
12
forum_id
stringlengths
9
13
review_title
stringlengths
0
500
review_body
stringlengths
1
31.1k
review_rating
stringlengths
0
31.1k
review_confidence
stringclasses
38 values
review_rating_integer
int64
-1
106
review_confidence_integer
int64
-1
5
SkSNwcVEl
H1GEvHcee
The authors propose a novel energy-function for RBMs, using the leaky relu max(cx, x) activation function for the hidden-units. Analogous to ReLU units in feed-forward networks, these leaky relu RBMs split the input space into a combinatorial number of regions, where each region defines p(v) as a truncated Gaussian. A further contribution of the paper is in proposing a novel sampling scheme for the leaky RBM: one can run a much shorter Markov chain by initializing it from a sample of the leaky RBM with c=1 (which yields a standard multi-variate normal over the visibles) and then slowly annealing c. In low-dimension a similar scheme is shown to outperform AIS for estimating the partition function. Experiments are performed on both CIFAR-10 and SVHN. This is an interesting paper which I believe would be of interest to the ICLR community. The theoretical contributions are strong: the authors not only introduce a proper energy formulation of ReLU RBMs, but also a novel sampling mechanism and an improvement on AIS for estimating their partition function. Unfortunately, the experimental results are somewhat limited. The PCD baseline is notably absent. Including (bernoulli visible, leaky-relu hidden) would have allowed the authors to evaluate likelihoods on standard binary RBM datasets. As it stands, performance on CIFAR-10 and SVHN, while improved with leaky-relu, is a far cry from more recent generative models (VAE-based, or auto-regressive models). While this comparison may be unfair, it will certainly limit the wider appeal of the paper to the community. Furthermore, there is the issue of the costly projection method which is required to guarantee that the energy-function remain bounded (covariance matrix over each region be PSD). Again, while it may be fair to leave that for future work given the other contributions, this will further limit the appeal of the paper. PROS: Introduces an energy function having the leaky-relu as an activation function Introduces a novel sampling procedure based on annealing the leakiness parameter Similar sampling scheme shown to outperform AIS CONS: Results are somewhat out of date Missing experiments on binary datasets (more comparable to prior RBM work) Missing PCD baseline Cost of projection method
6: Marginally above acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
6
4
Hyd8QeSVl
H1GEvHcee
A new model of RBM is proposed, where the conditional of the hidden is a leaky ReLU. In addition an annealed AIS sampler is also proposed to test the learned models quantifiably
Based on previous work such as the stepped sigmoid units and ReLU hidden units for discriminatively trained supervised models, a Leaky-ReLU model is proposed for generative learning. Pro: what is interesting is that unlike the traditional way of first defining an energy function and then deriving the conditional distributions, this paper propose the forms of the conditional first and then derive the energy function. However this general formulation is not novel to this paper, but was generalized to exponential family GLMs earlier. Con: Because of the focus on specifying the conditionals, the joint pdf and the marginal p(v) becomes complicated and hard to compute. On the experiments, it would been nice to see a RBM with binary visbles and leaky ReLu for hiddens. This would demonstrate the superiority of the leaky ReLU hidden units. In addition, there are more results on binary MNIST modeling with which the authors can compare the results to. While the authors is correct that the annealing distribution is no longer Gaussian, perhaps CD-25 or (Faast) PCD experiments can be run to compare agains the baseline RBM trained using (Fast) PCD. This paper is interesting as it combines new hidden function with the easiness of annealed AIS sampling, However, the baseline comparisons to Stepped Sigmoid Units (Nair &Hinton) or other models like the spike-and-slab RBMs (and others) are missing, without those comparisons, it is hard to tell whether leaky ReLU RBMs are better even in continuous visible domain.
5: Marginally below acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
5
4
r1Cybi8Ex
HyenWc5gx
Review
This paper proposes a method for transfer learning, i.e. leveraging a network trained on some original task A in learning a new task B, which not only improves performance on the new task B, but also tries to avoid degradation in performance on A. The general idea is based on encouraging a model trained on A, while training on the new task B, to match fake targets produced by the model itself but when it is trained only on the original task A. Experiments show that this method can help in improving the result on task B, and is better than other baselines, including standard fine-tuning. General comments/questions: - As far as I can tell, there is no experimental result supporting the claim that your model still performs well on the original task. All experiments show that you can improve on the new task only. - The introduction makes a strong statements about the distilling logical rule engine into a neural network, which I find a bit misleading. The approach in the paper is not specific to transferring from logical rules (as stated in the Sec 2) and is simply relying on the rule engine to provide labels for unlabelled data. - One of the obvious baselines to compare with your approach is standard multi-task learning on both tasks A and B together. That is, you train the model from scratch on both tasks simultaneously (which sharing parameters). It is not clear this is the same as what is referred to in Sec. 8 as "joint training". Can you please explain more clearly what you refer to as joint training? - Why can't we find the same baselines in both Table 2 and Table 3? For example Table 2 is missing "joint training", and Table 3 is missing GRU trained on the target task. - While the idea is presented as a general method for transfer learning, experiments are focused on one domain (sentiment analysis on SemEval task). I think that either experiments should include applying the idea on at least one other different domain, or the writing of the paper should be modified to make the focus more specific to this domain/task. Writing comments - The writing of the paper in general needs some improvement, but more specifically in the experiment section, where experiment setting and baselines should be explained more concisely. - Ensemble methodology paragraph does not fit the flow of the paper. I would rather explain it in the experiments section, rather than including it as part of your approach. - Table 1 seems like reporting cross-validation results, and I do not think is very informative to general reader.
5: Marginally below acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
5
4
ryO3U0GNe
HyenWc5gx
Interesting work, quite domain-specific, suboptimal focus and structure
This paper introduces a new method for transfer learning that avoids the catastrophic forgetting problem. It also describes an ensembling strategy for combining models that were learned using transfer learning from different sources. It puts all of this together in the context of recurrent neural networks for text analytics problems, to achieve new state-of-the-art results for a subtask of the SemEval 2016 competition. As the paper acknowledges, 1.5% improvement over the state-of-the-art is somewhat disappointing considering that it uses an ensemble of 5 quite different networks. These are interesting contributions, but due to the many pieces, unfortunately, the paper does not seem to have a clear focus. From the title and abstract/conclusion I would've expected a focus on the transfer learning problem. However, the description of the authors' approach is merely a page, and its evaluation is only another page. In order to show that this idea is a new methodological advance, it would've been good to show that it also works in at least one other application (e.g., just some multi-task supervised learning problem). Rather, the paper takes a quite domain-specific approach and discusses the pieces the authors used to obtain state-of-the-art performance for one problem. That is OK, but I would've rather expected that from a paper called something like "Improved knowledge transfer and distillation for text analytics". If accepted, I encourage the authors to change the title to something along those lines. The many pieces also made it hard for me to follow the authors' train of thought. I'm sure the authors had a good reason for their section ordering, but I didn't see the red thread in it. How about re-organizing the sections as follows to discuss one contribution at a time? 1,2,4,3,8 including 6, put 9 into an appendix and point to it from here, 7, 5, 10. That would first discuss the transfer learning piece (4, and experiments potentially in a subsection with previous sections 3,8,6), then discuss the distillation of logical rules (7), and then discuss ensembling and experiments for it (5 and 10). One clue that the current structure is suboptimal is that there are 11 sections... I like the authors' idea for transfer learning without catastropic forgetting, and I must admit I would've rather liked to read a paper solely about that (studying where it works, and where it fails) than about the many other topics of the paper. I weakly vote for acceptance since I like the ideas, but if the paper does not make it in, I would suggest that the authors consider splitting it into two papers, each of which could hopefully be more focused.
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
r1ECvAH4g
HyenWc5gx
This paper proposes a regularization technique for neural network training that relies on having multiple related tasks or datasets in a transfer learning setting. The proposed technique is straightforward to describe and can also leverage external labeling systems perhaps based on logical rules. The paper is clearly written and the experiments seem relatively thorough. Overall this is a nice paper but does not fully address how robust the proposed technique is. For each experiment there seems to be a slightly different application of the proposed technique, or a lot of ensembling and cross validation. I can’t figure out if this is because the proposed technique does not work well in general and thus required a lot of fiddling to get right in experiments, or if this is simply an artifact of ad-hoc experiments to try and get the best performance overall. If more datasets or addressing this issue directly in discussion was able to show this the strengths and limitations of the proposed technique more clearly, this could be a great paper. Overall the proposed method seems nice and possibly useful for other problems. However in the details of logical rule distillation and various experiment settings it seems like there is a lot of running the model many times or selecting a particular way of reusing the models and data that makes me wonder how robust the technique is or whether it requires a lot of trying various approaches, ensembling, or picking the best model from cross validation to show real gains. The authors could help by discussing this explicitly for all experiments in one place rather than listing the various choices / approaches in each experiment. As an example, these sorts of phrases make me very unsure how reliable the method is in practice versus how much the authors had to engineer this regularizer to perform well: “We noticed that equation 8 is actually prone to overfitting away from a good solution on the test set although it often finds a pretty good one early in training. “ The introduction section should first review the definitions of transfer learning vs multi-task learning to make the discussion more clear. It also deems justification why “catastrophic forgetting” is actually a problem. If the final target task is the only thing of interest then forgetting the source task is not an issue and the authors should motivate why forgetting matters in their setting. This paper explores sequential transfer so it’s not obvious why forgetting the source task matters. Section 7 introduces the logical rules engine in a fairly specific context. Rather it would be good state more generally what this system entails to help people figure out how this method would apply to other problems.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
Bk3Efq-Nl
HJjiFK5gx
Progress in reducing the supervision required by NPI
Neural Programmer-Interpreters (NPI) achieves greatly reduced sample complexity and better generalization than flat seq2seq models for program induction, but requires program traces at multiple levels of abstraction for training, which is a very strong form of supervision. One obvious way to improve this situation, addressed in this work, is to only train on the lowest-level traces, with a latent compositional program structure. This makes sense because the "raw" low-level traces can be cheaply gathered in many cases just by watching expert demonstrations, without being explicitly told the more temporally abstract structures. This paper shows that a variant of NPI, named NPL, can achieve even better generalization performance with weaker supervision (mostly flat traces), and also extends the model to a new grid world task. Unfortunately, it still requires being told the overall program structure by being given a few *full* execution traces. Still, I see this as important progress. It extends NPI in a quite nontrivial way by introducing a stack mechanism modeling the latent program call structure, which makes the training process much more closely match what the model does at test time. The results tell us that flat execution traces can take us almost all the way toward learning compositional programs from demonstrations - the hard part is of course learning to actually discover the subprogram structure.
7: Good paper, accept
7
-1
rkLqZvB4g
HJjiFK5gx
Review
First I would like to apologize for the late review. This paper proposes an extension of the NPI model (Reed & de Freitas) by using an extension of the probabilistic stacks introduced in Mikolov et al.. This allows them to train their model with less supervision than Reed & de Freitas. Overall the model is a nice extension of NPI. While it requires less supervision than NPI, it still requires "sequences of elementary operations paired with environment observations, and [...] a couple of examples which include the full abstraction hierarchy". This may limit the scope of this work. The paper claims that their "method is leverages stronger supervision in the form of elementary action sequences rather than just input-output examples (sic). Such sequences are relatively easy to gather in many natural settings". It would be great if the authors clarify what they mean by "relatively easy to gather in many natural settings". They also claim that "the additional supervision improves the data efficiency and allow our technique to scale to more complicated problems". However, this paper only addresses two toy problems which are neither "natural settings" nor of a large scale (or at least not larger than those addressed in the related literature, see Zaremba et al. for addition). In the introduction, the author states that "Existing techniques, however, cannot be applied on data like this because it does not contain the abstraction hierarchy." What are the "existing techniques", they are referring to? This work only addresses the problem of long addition and puzzle solving in a block world. Afaik, Zaremba et al. has shown that with no supervision, it can solve the long addition problem and Sukhbaatar et al. ("Mazebase: A sandbox for learning from games") shows that a memory network can solve puzzles in a blockworld with little supervision. In the conclusion, the author states that "remarkably, NPL achieves state-of-the-art performances with much less supervision compared to existing models, making itself more applicable to real-world applications where full program traces are hard to get." However for all the experiments, they "include a small number of FULL samples" (FULL == "samples with full program traces"). Unfortunately even if this means that they need less FULL examples, they still need "full program traces", contradicting their final claim. Moreover, as shown figure 7, their model does not use a "small number of FULL samples" but rather a significantly smaller amount of FULL examples than NPI, i.e., 16 vs 128. "All experiments were run with 10 different random seeds": does the environment change as well between the runs, i.e. are the FULL examples different between the runs? If it is the case and since you select the best run (on a validation set), the NPL model does not consume 16 FULL examples but 160 FULL examples for nanoCraft. Concerning the NanoCraft example, it would be good to have more details about how the examples are generated: how do you make sure that the train/val/test sets are different? How the rectangular shape are generated? If I consider all possible rectangles in a 6x6 grid, there are (6x6)x(6x6)/2 = 648 possibilities, thus taking 256 examples sum up to ~40% of the total number of rectangles. This does not even account for the fact that from an initial state, many rectangles can be made, making my estimate probably lower than the real coverage of examples. Concerning the addition, it would interesting to show what an LSTM would do: Take a 2 layer LSTM that takes the 2 current digits as an input and produce the current output ( "123+45" would be input[0] = [3,5], input[1]=[2,4], input[2]=[1, 0] and output[0] = 8...). I would be curious to see how such baseline would work. It can be trained on input/output and it is barely different from a standard sequence model. Also, would it be possible to compare with Zaremba et al.? Finally, as discussed previously with the authors, it would be good if they discuss more in length the relation between their probabilistic stacks and Mikolov et al.. They have a lot of similarities and it is not addressed in the current version. It should be addressed in the section describing the approach. I believe the authors agreed on this and I will wait for the updated version. Overall, it is a nice extension of Reed & de Freitas, but I'm a bit surprised by the lack of discussion about the rest of the literature (beside Reed & de Freitas, most previous work are only lightly discussed in the related work). This would have been fine if this paper would not suffer from a relatively weak experiment section that does not support the claims made in this work or show results that were not obtained by others before. Missing references: "Learning simple arithmetic procedures", Cottrell et al. "Neural gpus learn algorithms", Kaiser & Sutskever "Mazebase: A sandbox for learning from games", Sukhbaatar et al. "Learning simple algorithms from examples", Zaremba et al.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
SylfwJf4g
HJjiFK5gx
well formulated paper
The paper presents the Neural Program Lattice (NPL), extending the previous Neural Programmer-Interpreters (NPI). The main idea is to generalize stack manipulation of NPI by making it probabilistic. This allows the content of the stack to be stochastic than deterministic, and the paper describes the feed-forward steps of NPL's program inference similar to the NPI formulation. A new objective function is provided to train the model that maximizes the probability of NPL model correctly predicting operation sequences, from execution traces. We believe this is an important extension. The experimental results illustrate that the NPL is able to learn task executions in a clean setting with perfect observations. The paper is clearly presented and its background literature (i.e., NPI) is well covered. We also believe the paper is presenting a conceptually/technically meaningful extension of NPI, which will be of interest to a broad audience. We are still a bit concerned whether the NPL would be directly applicable for noisy observations (e.g., human skeletons) in a continuous space with less explicit structure, so more discussions will be interesting.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
SkyCWALEe
HJOZBvcel
Review
I sincerely apologize for the late-arriving review. This paper proposes to frame the problem of structure estimation as a supervised classification problem. The input is an empirical covariance matrix of the observed data, the output the binary decision whether or not two variables share a link. The paper is sufficiently clear, the goals are clear and everything is well described. The main interesting point is the empirical results of the experimental section. The approach is simple and performs better than previous non-learning based methods. This observation is interesting and will be of interest in structure discovery problems. I rate the specific construction of the supervised learning method as a reasonable attempt attempt to approach this problem. There is not very much technical novelty in this part. E.g., an algorithmic contribution would have been a method that is invariant to data permutation could have been a possible target for a technical contribution. The paper makes no claims on this technical part, as said, the method is well constructed and well executed. It is good to precisely state the theoretical parts of a paper, the authors do this well. All results are rather straight-forward, I like that the claims are written down, but there is little surprise in the statements. In summary, the paper makes a very interesting observation. Graph estimation can be posed as a supervised learning problem and training data from a separate source is sufficient to learn structure in novel and unseen test data from a new source. Practically this may be relevant, on one hand the empirical results are stronger with this method, on the other hand a practitioner who is interested in structural discovery may have side constraints about interpretability of the deriving method. From the Discussion and Conclusion I understand that the authors consider this as future work. It is a good first step, it could be stronger but also stands on its own already.
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
B1aSUyJEg
HJOZBvcel
Advantage of the proposed method
This paper proposes a new method for learning graphical models. Combined with a neural network architecture, some sparse edge structure is estimated via sampling methods. In introduction, the authors say that a problem in graphical lasso is model selection. However, the proposed method still implicitly includes model selection. In the proposed method, $P(G)$ is a sparse prior, and should include some hyper-parameters. How do you tune the hyper-parameters? Is this tuning an equivalent problem to model section? Therefore, I do not understand real advantage of this method over previous methods. What is the advantage of the proposed method? Another concern is that this paper is unorganized. In Algorithm 1, first, G_i and \Sigma_i are sampled, and then x_j is sampled from N(0, \Sigma). Here, what is \Sigma? Is it different from \Sigma_i? Furthermore, how do you construct (Y_i, \hat{\Sigma}_i) from (G_i, X_i )? Finally, I have a simple question: Where is input data X (not sampled data) is used in Algorithm 1? What is the definition of the receptive field in Proposition 2 and Proposition 3?
5: Marginally below acceptance threshold
5
-1
BkF-pCWVl
HJOZBvcel
Interesting algorithm to estimate sparse graph structure
The paper proposes a novel algorithm to estimate graph structures by using a convolutional neural network to approximate the function that maps from empirical covariance matrix to the sparsity pattern of the graph. Compared with existing approaches, the new algorithm can adapt to different network structures, e.g. small-world networks, better under the same empirical risk minimization framework. Experiments on synthetic and real-world datasets show promising results compared with baselines. In general, I think it is an interesting and novel paper. The idea of framing structure estimation as a learning problem is especially interesting and may inspire further research on related topics. The advantage of such an approach is that it allows easier adaptation to different network structure properties without designing specific regularization terms as in graph lasso. The experiment results are also promising. In both synthetic and real-world datasets, the proposed algorithm outperforms other baselines in the small sample region. However, the paper can be made clearer in describing the network architectures. For example, in page 5, each o^k_{i,j} is said be a d-dimensional vector. But from the context, it seems o^k_{i,j} is a scalar (from o^0_{i,j} = p_{i,j}). It is not clear what o^k_{i,j} is exactly and what d is. Is it the number of channels for the convolutional filters? Figure 1 is also quite confusing. Why in (b) the table is 16 x 16 whereas in (a) there are only six nodes? And from the figure, it seems there is only one channel in each layer? What do the black squares represent and why are there three blocks of them. There are some descriptions in the text, but it is still not clear what they mean exactly. For real-world data, how are the training data (Y, Sigma) generated? Are they generated in the same way as in the synthetic experiments where the entries are uniformly sparse? This is also related to the more general question of how to sample from the distribution P, in the case of real-world data.
7: Good paper, accept
3: The reviewer is fairly confident that the evaluation is correct
7
3
ryztRFW4e
BJVEEF9lx
The paper presents a framework to formulate data-structures in a learnable way. It is an interesting and novel approach that could generalize well to interesting datastructures and algorithms. In its current state (Revision of Dec. 9th), there are two strong weaknesses remaining: analysis of related work, and experimental evidence. Reviewer 2 detailed some of the related work already, and especially DeepMind (which I am not affiliated with) presented some interesting and highly related results with its neural touring machine and following work. While it may be of course very hard to make direct comparisons in the experimental section due to complexity of the re-implementation, it would at least be very important to mention and compare to these works conceptually. The experimental section shows mostly qualitative results, that do not (fully) conclusively treat the topic. Some suggestions for improvements: * It would be highly interesting to learn about the accuracy of the stack and queue structures, for increasing numbers of elements to store. * Can a queue / stack be used in arbitrary situations of push-pop operations occuring, even though it was only trained solely with consecutive pushes / consecutive pops? Does it in this enhanced setting `diverge' at some point? * The encoded elements from MNIST, even though in a 28x28 (binary?) space, are elements of a ten-element set, and can hence be encoded a lot more efficiently just by `parsing' them, which CNNs can do quite well. Is the NN `just' learning to do that? If so, its performance can be expected to strongly degrade when having to learn to stack more than 28*28/4=196 numbers (in case of an optimal parser and loss-less encoding). To argue more in this direction, experiments would be needed with an increasing number of stack / queue elements. Experimenting with an MNIST parsing NN in front of the actual stack/queue network could help strengthening or falsifying the claim. * The claims about `mental representations' have very little support throughout the paper. If indication for correspondence to mental models, etc., could be found, it would allow to hold the claim. Otherwise, I would remove it from the paper and focus on the NN aspects and maybe mention mental models as motivation.
4: Ok but not good enough - rejection
3: The reviewer is fairly confident that the evaluation is correct
4
3
Sy20Q1MNl
BJVEEF9lx
Interesting direction, but not there yet.
A method for training neural networks to mimic abstract data structures is presented. The idea of training a network to satisfy an abstract interface is very interesting and promising, but empirical support is currently too weak. The paper would be significantly strengthened if the method could be shown to be useful in a realistic application, or be shown to work better than standard RNN approaches on algorithmic learning tasks. The claims about mental representations are not well supported. I would remove the references to mind and brain, as well as the more philosophical points, or write a paper that really emphasizes one of these aspects and supports the claims.
4: Ok but not good enough - rejection
3: The reviewer is fairly confident that the evaluation is correct
4
3
ryg9PB-Vg
BJVEEF9lx
Review
The paper presents a way to "learn" approximate data structures. They train neural networks (ConvNets here) to perform as an approximate abstract data structure by having an L2 loss (for the unrolled NN) on respecting the axioms of the data structure they want the NN to learn. E.g. you NN.push(8), NN.push(6), NN.push(4), the loss is proportional to the distance with what is NN.pop()ed three times and 4, 6, 8 (this example is the one of Figure 1). There are several flaws: - In the case of the stack: I do not see a difference between this and a seq-to-seq RNN trained with e.g. 8, 6, 4 as input sequence, to predict 4, 6, 8. - While some of the previous work is adequately cited, there is an important body of previous work (some from the 90s) on learning Peano's axioms, stacks, queues, etc. that is not cited nor compared to. For instance [Das et al. 1992], [Wiles & Elman 1995], and more recently [Graves et al. 2014], [Joulin & Mikolov 2015], [Kaiser & Sutskever 2016]... - Using MNIST digits, and not e.g. a categorical distribution on numbers, is adding complexity for no reason. - (Probably the biggest flaw) The experimental section is too weak to support the claims. The figures are adequate, but there is no comparison to anything. There is also no description nor attempt to quantify a form of "success rate" of learning such data structures, for instance w.r.t the number of examples, or w.r.t to the size of the input sequences. The current version of the paper (December 9th 2016) provides, at best, anecdotal experimental evidence to support the claims of the rest of the paper. While an interesting direction of research, I think that this paper is not experimentally sound enough for ICLR.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
rk_Zn-G4x
S1Jhfftgx
Not very convincing
This paper proposes a way of enforcing constraints (or penalizing violations of those constraints) on outputs in structured prediction problems, while keeping inference unconstrained. The idea is to tweak the neural network parameters to make those output constraints hold. The underlying model is that of structured prediction energy networks (SPENs), recently proposed by Belanger et al. Overall, I didn't find the approach very convincing and the paper has a few problems regarding the empirical evaluation. There's also some imprecisions throughout. The proposed approach (secs 6 and 7) looks more like a "little hack" to try to make it vaguely similar to Lagrangian relaxation methods than something that is theoretically well motivated. Before eq. 6: "an exponential number of dual variables" -- why exponential? it's not one dual variable per output. From the clarification questions: - The accuracy reported in Table 1 needs to be explained. - for the parsing experiments it would be good to report the usual F1 metric of parseval, and to compare with state of the art systems. - should use the standard training/dev/test splits of the Penn Treebank. The reported conversion rate in Table 1 does not tell us how many violations are left by the unconstrained decoder to start with. It would be good to know what happens in highly structured problems where these violations are frequent, since these are the problems where the proposed approach could be more beneficial. Minor comments/typos: - sec.1: "there are" -> there is? - sec 1: "We find that out method is able to completely satisfy constraints on 81% of the outputs." -> at this point, without specifying the problem, the model, and the constraints, this means very little. How many constrains does the unconstrained method satisfies? - sec 2 (last paragraph): "For RNNs, each output depends on hidden states that are functions of previous output values" -- this is not very accurate, as it doesn't hold for general RNNs, but only for those (e.g. RNN decoders in language modeling) where the outputs are fed back to the input in the next time frame. - sec 3: "A major advantage of neural networks is that once trained, inference is extremely efficient." -- advantage over what? also, this is not necessarily true, depends on the network and on its size. - sec 3: "our goal is take advantage" -> to take advantage - last paragraph of sec 6: "the larger model affords us" -> offers?
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
rJfGR9LEg
S1Jhfftgx
This paper attempted to solve an interesting problem -- incorporating hard constraints in seq2seq model. The main idea is to modify the weight of the neural network in order to find a feasible solution. Overall, the idea presented in the paper is interesting, and it tries to solve an important problem. However, it seems to me the paper is not ready to publish yet. Comments: - The first section of the paper is clear and well-motivated. - The authors should report test running time. The proposed approach changes the weight matrix. As a result, it needs to reevaluate the values of hidden states and perform the greedy search for each iteration of optimizing Eq (7). This is actually pretty expensive in comparison to running the beam search or other inference methods. Therefore, I'm not convinced that the proposed approach is a right direction for solving this problem (In table, 1, the authors mention that they run 100 steps of SGD). - If I understand correctly, Eq (7) is a noncontinuous function w.r.t W_\lambda and the simple SGD algorithm will not be able to find its minimum. - For dependency parsing, there are standard splits of PTB. I would suggest the authors follow the same splits of train, dev, and test in order to compare with existing results. Minor comments: several sentences are misleading and should be rewritten carefully. - Beginning of Section 3: "A major advantage of neural network is that once trained, inference is extremely efficient." This sentence is not generally right, and I guess the authors mean if using greedy search as inference method, the inference is efficient. - The description in the end of section 2 is awkward. To me, feed-forward and RNN are general families that cover many specific types of neural networks, and the training procedures are not necessarily to aim to optimize Eq. (2). Therefore, the description here might not be true. In fact, I don't think there is a need to bring up feed-forward networks here; instead, the authors should provide more details the connection between RNN and Eq (2) here. - The second paragraph of section 3 is related to [1], where it shows the search space of the inference can be represented as an imperative program. [1] Credit assignment compiler for joint prediction, NIPS 2016
4: Ok but not good enough - rejection
4
-1
H13e_DgNg
S1Jhfftgx
Reject
This paper proposes a dual-decomposition-inspired technique for enforcing constraints in neural network prediction systems. Many things don't quite make sense to me: 1. Most seq2seq models (such as those used for parsing) have substantially better performance when coupled with beam search than greedy search, and exact search is infeasible. This is because these models are trained to condition on discrete values of past outputs in each timestamp, and hence the problem of finding the highest-scoring total sequence of outputs is not solvable efficiently. It's unclear what kind of model this paper is using which allows for greedy decoding, and how well it compares to the state-of-the-art, specially when constraint-aware beam search is used. This comparison is specially interesting because both constrained beam search and this dual-decomposition-like approach require multiple computations of the model's score. 2. It's unclear (to me at least) how to differentiate the constraint term g() in the objective function in the general case (though the particular example used here is understandable) 3. The paper claims that "Lagrangian relaxation methods for NLP have multipliers for each output variable that can be combined with linear models [...] . Since our non-linear functions and global constraints do not afford us the same ability" but it is possible to add linear terms to the outputs of neural networks, possibly avoiding rerunning all the expensive inference terms. Moreover, the justification for the particular method is hand-wavy at best, with inconvenient terms from equations ignored or changed at will. At this point it might be better to omit the attempted theoretical explanation and just present this method as a heuristic which is likely to achieve the desired result. This, plus the concerns around lack of clear comparisons with baselines on benchmark problems lead me to recommend rejection. Further explanation of how this compares with beam search, how this relates to the state-of-the-art, and a better explanation for how to come up with differentiable constraint sets, are probably required for acceptance.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
SJXnoez4e
BJ46w6Ule
Improve the exposition
The goal of this paper is to learn “ a collection of experts that are individually meaningful and that have disjoint responsibilities.” Unlike a standard mixture model, they “use a different mixture for each dimension d.” While the results seem promising, the paper exposition needs significant improvement. Comments: The paper jumps in with no motivation at all. What is the application, or even the algorithm, or architecture that this is used for? This should be addressed at the beginning. The subsequent exposition is not very clear. There are assertions made with no justification, e.g. “the experts only have a small variance for some subset of the variables while the variance of the other variables is large.” Since you’re learning both the experts and the weights, can this be rephrased in terms of dictionary learning? Please discuss the relevant related literature. The horse data set is quite small with respect to the feature dimension, and so the conclusions may not necessarily generalize.
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
S1zNjzGNg
BJ46w6Ule
Potentially interesting paper, but not clear enough
The paper addresses the problem of learning compact binary data representations. I have a hard time understanding the setting and the writing of the paper is not making it any easier. For example I can't find a simple explanation of the problem and I am not familiar with these line of research. I read all the responses provided by authors to reviewer's questions and re-read the paper again and I still do not fully understand the setting and thus can't really evaluate the contributions of these work. The related work section does not exist and instead the analysis of the literature is somehow scattered across the paper. There are no derivations provided. Statements often miss references, e.g. the ones in the fourth paragraph of Section 3. This makes me conclude that the paper still requires significant work before it can be published.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
HyRxUhRQg
BJ46w6Ule
A type of PoE but the probability seems undefined and the EM algorithms remains obscure. Experiments are illustrative only.
This paper proposes a new kind of expert model where a sparse subset of most reliable experts is chosen instead of the usual logarithmic opinion pool of a PoE. I find the paper very unclear. I tried to find a proper definition of the joint model p(x,z) but could not extract this from the text. The proposed “EM-like” algorithm should then also follow directly from this definition. At this point I do not see if such as definition even exists. In other words, is there is an objective function on which the iterates of the proposed algorithm are guaranteed to improve on the train data? We also note that the “product of unifac models” from Hinton tries to do something very similar where only a subset of the experts will get activated to generate the input: http://www.cs.toronto.edu/~hinton/absps/tr00-004.pdf I tried to derive the update rule on top of page 4 from the “conditional objective for p(x|h)” in sec. 3.2 But I am getting something different (apart form the extra smoothing factors eps and mu_o). Does this follow? (If we define R=R_nk, mu-mu_k and X=X_n, I get mu = (XR)*inv(R^TR) as the optimal solution, which then needs to be projected back onto the probability simplex). The experiments are only illustrative. They don’t compare with other methods (such as an RBM or VAE) nor do they give any quantitative results. We are left with eyeballing some images. I have no idea whether what we see is impressive or not.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
S1qqrWz4l
HJ9rLLcxg
The concept of data augmentation in the embedding space is very interesting. The method is well presented and also justified on different tasks such as spoken digits and image recognition etc. One comments of the comparison is the use of a simple 2-layer MLP as the baseline model throughout all the tasks. It's not clear whether the gains maintain when a more complex baseline model is used. Another comment is that the augmented context vectors are used for classification, just wondering how does it compare to using the reconstructed inputs. And furthermore, as in Table 4, both input and feature space extrapolation improves the performance, whether these two are complementary or not?
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
rk7Sgr-Eg
HJ9rLLcxg
In this paper authors propose a novel data augmentation scheme where instead of augmenting the input data, they augment intermediate feature representations. Sequence auto-encoder based features are considered, and random perturbation, feature interpolation, and extrapolation based augmentation are evaluated. On three sequence classification tasks and on MNIST and CIFAR-10, it is shown that augmentation in feature space, specifically extrapolation based augmentation, results in good accuracy gains w.r.t. authors baseline. My main questions and suggestions for further strengthening the paper are: a) The proposed data augmentation approach is applied to a learnt auto-encoder based feature space termed ‘context vector’ in the paper. The context vectors are then augmented and used as input to train classification models. Have the authors considered applying their feature space augmentation idea directly to the classification model during training, and applying it to potentially many layers of the model? Also, have the authors considered convolutional neural network (CNN) architectures as well for feature space augmentation? CNNs are now the state-of-the-art in many image and sequence classification task, it would be very valuable to see the impact of the proposed approach in that model. b) When interpolation or extrapolation based augmentation was being applied, did the authors also consider utilizing nearby samples from competing classes as well? Especially in case of extrapolation based augmentation it will be interesting to check if the extrapolated features are closer to competing classes than original ones. c) With random interpolation or nearest neighbor interpolation based augmentation the accuracy seems to degrade pretty consistently. This is counter-intuitive. Do the authors have explanation for why the accuracy degraded with interpolation based augmentation? d) The results on MNIST and CIFAR-10 are inconclusive. For instance the error rate on CIFAR-10 is well below 10% these days, so I think it is hard to draw conclusions based on error rates above 30%. For MNIST it is surprising to see that data augmentation in the input space substantially degrades the accuracy (1.093% -> 1.477%). As mentioned above, I think this will require extending the feature space augmentation idea to CNN based models.
6: Marginally above acceptance threshold
6
-1
ByUItzz4g
HJ9rLLcxg
review
TDLR: The authors present a regularization method wherein they add noise to some representation space. The paper mainly applies the technique w/ sequence autoencoders (Dai et al., 2015) without the usage of attention (i.e., only using the context vector). Experimental results show improvement from author's baseline on some toy tasks. === Augmentation === The augmentation process is simple enough, take the seq2seq context vector and add noise/interpolate/extrapolate to it (Section 3.2). This reviewer is very curious whether this process will also work in non seq2seq applications. This reviewer would have liked to see comparison with dropout on the context vector. === Experiments === Since the authors are experimenting w/ seq2seq architectures, its a little bit disappointing they didn't compare it w/ Machine Translation (MT), where there are many published papers to compare to. The authors did compare their method on several toy datasets (that are less commonly used in DL literature) and MNIST/CIFAR. The authors show improvement over their own baselines on several toy datasets. The improvement on MNIST/CIFAR over the author's baseline seems marginal at best. The author also didn't cite/compare to the baseline published by Dai et al., 2015 for CIFAR -- here they have a much better LSTM baseline of 25% for CIFAR which beats the author's baseline of 32.35% and the author's method of 31.93%. The experiments would be much more convincing if they did it on seq2seq+MT on say EN-FR or EN-DE. There is almost no excuse why the experiments wasn't run on the MT task, given this is the first application of seq2seq was born from. Even if not MT, then at least the sentiment analysis tasks (IMDB/Rotten Tomatoes) of the Dai et al., 2015 paper which this paper is so heavily based on for the sequence autoencoder. === References === Something is wrong w/ your references latex setting? Seems like a lot of the conference/journal names are omitted. Additionally, you should update many cites to use the conference/journal name rather than just "arxiv". Listen, attend and spell (should be Listen, Attend and Spell: A Neural Network for Large Vocabulary Conversational Speech Recognition) -> ICASSP if citing ICASSP paper above, should also cite Bahandau paper "End-to-End Attention-based Large Vocabulary Speech Recognition" which was published in parallel (also in ICASSP). Adam: A method for stochastic optimization -> ICLR Auto-encoding variational bayes -> ICLR Addressing the rare word problem in neural machine translation -> ACL Pixel recurrent neural networks -> ICML A neural conversational model -> ICML Workshop
4: Ok but not good enough - rejection
4
-1
H18MIfimg
BJC_jUqxe
Strong, but some framing issues
This paper introduces a sentence encoding model (for use within larger text understanding models) that can extract a matrix-valued sentence representation by way of within-sentence attention. The new model lends itself to (slightly) more informative visualizations than could be gotten otherwise, and beats reasonable baselines on three datasets. The paper is reasonably clear, I see no major technical issues, and the proposed model is novel and effective. It could plausibly be relevant to sequence modeling tasks beyond NLP. I recommend acceptance. There is one fairly serious writing issue that I'd like to see fixed, though: The abstract, introduction, and related work sections are all heavily skewed towards unsupervised learning. The paper doesn't appear to be doing unsupervised learning, and the ideas are no more nor less suited to unsupervised learning than any other mainstream ideas in the sentence encoding literature. Details: - You should be clearer about how you expect these embeddings to be used, since that will be of certain interest to anyone attempting to use the results of this work. In particular, how you should convert the matrix representation into a vector for downstream tasks that require one. Some of the content of your reply to my comment could be reasonably added to the paper. - A graphical representation of the structure of the model would be helpful. - The LSTMN (Cheng et al., EMNLP '16) is similar enough to this work that an explicit comparison would be helpful. Again, incorporating your reply to my comment into the paper would be more than adequate. - Jiwei Li et al. (Visualizing and Understanding Neural Models in NLP, NAACL '15) present an alternative way of visualizing the influence of words on sentence encodings without using cross-sentence attention. A brief explicit comparison would be nice here.
8: Top 50% of accepted papers, clear accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
8
4
ByHCv9b4e
BJC_jUqxe
Interesting embedding method, lacking in analysis of 2d structure
This paper proposes a method for representing sentences as a 2d matrix by utilizing a self-attentive mechanism on the hidden states of a bi-directional LSTM encoder. This work differs from prior work mainly in the 2d structure of embedding, which the authors use to produce heat-map visualizations of input sentences and to generate good performance on several downstream tasks. There is a substantial amount of prior work which the authors do not appropriately address, some of which is listed in previous comments. The main novelty of this work is in the 2d structure of embeddings, and as such, I would have liked to see this structure investigated in much more depth. Specifically, a couple important relevant experiments would have been: * How do the performance and visualizations change as the number of attention vectors (r) varies? * For a fixed parameter budget, how important is using multiple attention vectors versus, say, using a larger hidden state or embedding size? I would recommend changing some of the presentation in the penalization term section. Specifically, the statement that "the best way to evaluate the diversity is definitely the Kullback Leibler divergence between any 2 of the summation weight vectors" runs somewhat counter to the authors' comments about this topic below. In Fig. (2), I did not find the visualizations to provide particularly compelling evidence that the multiple attention vectors were doing much of interest beyond a single attention vector, even with penalization. To me this seems like a necessary component to support the main claims of this paper. Overall, while I found the architecture interesting, I am not convinced that the model's main innovation -- the 2d structure of the embedding matrix -- is actually doing anything important or meaningful beyond what is being accomplished by similar attentive embedding models already present in the literature. Further experiments demonstrating this effect would be necessary for me to give this paper my full endorsement.
5: Marginally below acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
5
4
H1zxxEXEg
BJC_jUqxe
Interesting idea but need additional work to be convincing
I like the idea in this paper that use not just one but multiple attentional vectors to extract multiple representations for a sentence. The authors have demonstrated consistent gains across three different tasks Age, Yelp, & SNLI. However, I'd like to see more analysis on the 2D representations (as concerned by another reviewer) to be convinced. Specifically, r=30 seems to be a pretty large value when applying to short sentences like tweets or those in the SNLI dataset. I'd like to see the effect of varying r from small to large value. With large r value, I suspect your models might have an advantage in having a much larger number of parameters (specifically in the supervised components) compare to other models. To make it transparent, the model sizes should be reported. I'd also like to see performances on the dev sets or learning curves. In the conclusion, the authors remark that "attention mechanism reliefs the burden of LSTM". If the 2D representations are effective in that aspect, I'd expect that the authors might be able to train with a smaller LSTM. Testing the effect of LSTM dimension vs $r$ will be helpful. Lastly, there is a problem in the presentation of the paper in which there is no training objective defined. Readers have to read until the experimental sections to guess that the authors perform supervised learning and back-prop through the self-attention mechanism as well as the LSTM. * Minor comments: Typos: netowkrs, toghter, performd Missing year for the citation of (Margarit & Subramaniam) In figure 3, attention plotswith and without penalization look similar.
6: Marginally above acceptance threshold
6
-1
B1v-2iWNx
SJ8BZTjeg
Review
The paper proposes an approach to unsupervised learning based on generative adversarial networks (GANs) and clustering. The general topic of unsupervised learning is important, and the proposed approach makes some sense, but experimental evaluation is very weak and does not allow to judge if the proposed method is competitive with existing alternatives. Therefore the paper cannot be published in its current form. More detailed remarks (many of these are copies of my pre-review questions the authors have not responded to): 1) Realted work overview looks incomplete. There has been work on combining clustering with deep learning, for example [1] or [2] look very related. A long list of potentially related papers can be found here: https://amundtveit.com/2016/12/02/deep-learning-for-clustering/ . From the GAN side, for example [3] looks related. I would like the authors to comment on relation of their approach to existing work, if possible compare with existing approaches, and if not possible - explain why. [1] Xie et al., "Unsupervised Deep Embedding for Clustering Analysis", ICML 2016 http://jmlr.org/proceedings/papers/v48/xieb16.pdf [2] Yang et al., "Joint Unsupervised Learning of Deep Representations and Image Clusters", CVPR 2016 http://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Yang_Joint_Unsupervised_Learning_CVPR_2016_paper.pdf [3] J.T. Springenberg, "Unsupervised and semi-supervised learning with categorical generative adversarial networks", ICLR 2016, https://arxiv.org/pdf/1511.06390v2.pdf 2) The authors do not report classification accuracies, which makes it very difficult to compare their results with existing work. Classification accuracies should be reported. They may not be a perfect measure of feature quality, but reporting them in addition to ARI and NMI would not hurt. 3) The authors have not compared their approach to existing unsupervised feature learning approaches, for example feature learning with k-means (Coates and Ng 2011), sparse coding methods such as Hierarchical Matching Pursuit (Bo et al., 2012 and 2013), Exemplar-CNN (Dosovitskiy et al. 2014) 4) Looks like in Figure 2 every "class" consists essentially of a single image and its slight variations? Doesn't this mean GAN training failed? Do all your GANs produce samples of this quality? 5) Why do you not show results with visual features on STL-10? 6) Supervisedly learned filters in Figure 3 looks unusual to me, they are normally not that smooth. Have you optimized the hyperparameters? What is the resulting accuracy?
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
SynBgHuNx
SJ8BZTjeg
review
The papers investigates the task of unsupervised learning with deep features via k-means clustering. The entire pipeline can be decomposed into two steps: (1) unsupervised feature learning based on GAN framework and (2) k-means clustering using learned deep network features. Following the GAN framework and its extension InfoGAN, the first step is to train a pair of discriminator network and generator network from scratch using min-max objective. Then, it applies k-means clustering on the top layer features from discriminator network. For evaluation, the proposed unsupervised feature learning approach is compared against traditional hand-crafted features such as HOG and supervised method on three benchmark datasets. Normalized Mutual Information (NMI) and Adjusted RAND Index (ARI) have been used as the evaluation metrics for experimental comparison. Although the proposed method may be potentially useful in practice (if refined further), I find the method lacks novelty, and the experimental results are not significant enough.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
BkUsyJGEl
SJ8BZTjeg
review
This paper proposed an unsupervised learning method based on running kmeans on the features learned by a discriminator network in a generative adversarial network setup. Unsupervised learning methods with GANs is certainly a relevant topic but this paper does not propose anything particularly novel as far as I can tell. More importantly, the evaluation methods in this paper are extremely lacking. The authors omit classification results on CIFAR and STL-10 and instead the only quantitative evaluation plot the performance of the clustering algorithm on the features. Not only are classification results not shown, no comparisons are made to the wealth of related work. I list just a few highly related techniques below. Finally, it appear the authors have not train their GANs correctly as the samples in Fig.2 appear to be from a model that has collapsed during training. In summary, the ideas in this paper are potentially interesting but this paper should not be accepted in its current form due to lack of experimental results and comparisons. (non-exhaustive) list of related work on unsupervised learning (with and without GANs): [1] Springenberg. Unsupervised and Semi-supervised Learning with Categorical Generative Adversarial Networks, ICLR 2016 (https://arxiv.org/abs/1511.06390) [2] Salimans et al. Improved Techniques for Training GANs. NIPS 2016 (https://arxiv.org/abs/1606.03498) [3] Dosovitskiy et al. Discriminative unsupervised feature learning with convolutional neural networks, NIPS 2014 (https://arxiv.org/abs/1406.6909)
3: Clear rejection
3
-1
rkNheaUEl
ryxB0Rtxx
a good paper
This paper provides some theoretical guarantees for the identity parameterization by showing that 1) arbitrarily deep linear residual networks have no spurious local optima; and 2) residual networks with ReLu activations have universal finite-sample expressivity. This paper is well written and studied a fundamental problem in deep neural network. I am very positive on this paper overall and feel that this result is quite significant by essentially showing the stability of auto-encoder, given the fact that it is hard to provide concrete theoretical guarantees for deep neural networks. One of key questions is how to extent the result in this paper to the more general nonlinear actuation function case. Minors: one line before Eq. (3.1), U \in R ? \times k
8: Top 50% of accepted papers, clear accept
3: The reviewer is fairly confident that the evaluation is correct
8
3
SJb64ilNl
ryxB0Rtxx
This paper investigates the identity parametrization also known as shortcuts where the output of each layer has the form h(x)+x instead of h(x). This has been shown to perform well in practice (eg. ResNet). The discussions and experiments in the paper are interesting. Here's a few comments on the paper: -Section 2: Studying the linear networks is interesting by itself. However, it is not clear that how this could translate to any insight about non-linear networks. For example, you have proved that every critical point is global minimum. I think it is helpful to add some discussion about the relationship between linear and non-linear networks. -Section 3: The construction is interesting but the expressive power of residual network is within a constant factor of general feedforward networks and I don't see why we need a different proof given all the results on finite sample expressivity of feedforward networks. I appreciate if you clarify this. -Section 4: I like the experiments. The choice of random projection on the top layer is brilliant. However, since you have combined this choice with all-convolutional residual networks, it is hard for the reader to separate the affect of each of them. Therefore, I suggest reporting the numbers for all-convolutional residual networks with learned top layer and also ResNet with random projection on the top layer. Minor comments: 1- I don't agree that Batch Normalization can be reduced to identity transformation and I don't know if bringing that in the abstract without proper discussion is a good idea. 2- Page 5 above assumption 3.1 : x^(i)=1 ==> ||x^(i)||_2=1
This paper investigates the identity parametrization also known as shortcuts where the output of each layer has the form h(x)+x instead of h(x). This has been shown to perform well in practice (eg. ResNet). The discussions and experiments in the paper are interesting. Here's a few comments on the paper: -Section 2: Studying the linear networks is interesting by itself. However, it is not clear that how this could translate to any insight about non-linear networks. For example, you have proved that every critical point is global minimum. I think it is helpful to add some discussion about the relationship between linear and non-linear networks. -Section 3: The construction is interesting but the expressive power of residual network is within a constant factor of general feedforward networks and I don't see why we need a different proof given all the results on finite sample expressivity of feedforward networks. I appreciate if you clarify this. -Section 4: I like the experiments. The choice of random projection on the top layer is brilliant. However, since you have combined this choice with all-convolutional residual networks, it is hard for the reader to separate the affect of each of them. Therefore, I suggest reporting the numbers for all-convolutional residual networks with learned top layer and also ResNet with random projection on the top layer. Minor comments: 1- I don't agree that Batch Normalization can be reduced to identity transformation and I don't know if bringing that in the abstract without proper discussion is a good idea. 2- Page 5 above assumption 3.1 : x^(i)=1 ==> ||x^(i)||_2=1
-1
-1
Hy8H45WVg
ryxB0Rtxx
Paper Summary: Authors investigate identity re-parametrization in the linear and the non linear case. Detailed comments: — Linear Residual Network: The paper shows that for a linear residual network any critical point is a global optimum. This problem is non convex it is interesting that this simple re-parametrization leads to such a result. — Non linear Residual Network: Authors propose a construction that maps the points to their labels via a resnet , using an initial random projection, followed by a residual block that clusters the data based on their label, and a last layer that maps the clusters to the label. 1- In Eq 3.4 seems the dimensions are not matching q_j in R^k and e_j in R^r. please clarify 2- The construction seems fine, but what is special about the resnet here in this construction? One can do a similar construction if we did not have the identity? can you discuss this point? In the linear case it is clear from a spectral point of view how the identity is helping the optimization. Please provide some intuition. 3- Existence of a network in the residual class that overfits does it give us any intuition on why residual network outperform other architectures? What does an existence result of such a network tell us about its representation power ? A simple linear model under the assumption that points can not be too close can overfit the data, and get fast convergence rate (see for instance tsybakov noise condition). 4- What does the construction tell us about the number of layers? 5- clustering the activation independently from the label, is an old way to pretrain the network. One could use those centroids as weights for the next layer (this is also related to Nystrom approximation see for instance https://www.cse.ust.hk/~twinsen/nystrom.pdf ). Your clustering is very strongly connected to the label at each residual block. I don't think this is appealing or useful since no feature extraction is happening. Moreover the number of layers in this construction does not matter. Can you weaken the clustering to be independent to the label at least in the early layers? then one could you use your construction as an initialization in the training. — Experiments : - last layer is not trained means the layer before the linear layer preceding the softmax? Minor comments: Abstract: how the identity mapping motivated batch normalization?
5: Marginally below acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
5
4
r10X7Es4g
B1ckMDqlg
Elegant use of MoE for expanding model capacity, but it would be very nice to discuss MoE alternatives in terms of computational efficiency and other factors.
Paper Strengths: -- Elegant use of MoE for expanding model capacity and enabling training large models necessary for exploiting very large datasets in a computationally feasible manner -- The effective batch size for training the MoE drastically increased also -- Interesting experimental results on the effects of increasing the number of MoEs, which is expected. Paper Weaknesses: --- there are many different ways of increasing model capacity to enable the exploitation of very large datasets; it would be very nice to discuss the use of MoE and other alternatives in terms of computational efficiency and other factors.
6: Marginally above acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
6
4
B1ZFEvR4x
B1ckMDqlg
This paper proposes a method for significantly increasing the number of parameters in a single layer while keeping computation in par with (or even less than) current SOTA models. The idea is based on using a large mixture of experts (MoE) (i.e. small networks), where only a few of them are adaptively activated via a gating network. While the idea seems intuitive, the main novelty in the paper is in designing the gating network which is encouraged to achieve two objectives: utilizing all available experts (aka importance), and distributing computation fairly across them (aka load). Additionally, the paper introduces two techniques for increasing the batch-size passed to each expert, and hence maximizing parallelization in GPUs. Experiments applying the proposed approach on RNNs in language modelling task show that it can beat SOTA results with significantly less computation, which is a result of selectively using much more parameters. Results on machine translation show that a model with more than 30x number of parameters can beat SOTA while incurring half of the effective computation. I have the several comments on the paper: - I believe that the authors can do a better job in their presentation. The paper currently is at 11 pages (which is too long in my opinion), but I find that Section 3.2 (the crux of the paper) needs better motivation and intuitive explanation. For example, equation 8 deserves more description than currently devoted to it. Additional space can be easily regained by moving details in the experiments section (e.g. architecture and training details) to the appendix for the curious readers. Experiment section can be better organized by finishing on experiment completely before moving to the other one. There are also some glitches in the writing, e.g. the end of Section 3.1. - The paper is missing some important references in conditional computation (e.g. https://arxiv.org/pdf/1308.3432.pdf) which deal with very similar issues in deep learning. - One very important lesson from the conditional computation literature is that while we can in theory incur much less computation, in practice (especially with the current GPU architectures) the actual time does not match the theory. This can be due to inefficient branching in GPUs. It would be nice if the paper includes a discussion of how their model (and perhaps implementation) deal with this problem, and why it scales well in practice. - Table 1 and Table 3 contain repetitive information, and I think they should be combined in one (maybe moving Table 3 to appendix). One thing I do not understand is how does the number of ops/timestep relate to the training time. This also related to the pervious comment.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
Hy7i5WXNg
B1ckMDqlg
Nice use of MoE with good results
This paper describes a method for greatly expanding network model size (in terms of number of stored parameters) in the context of a recurrent net, by applying a Mixture of Experts between recurrent net layers that is shared between all time steps. By process features from all timesteps at the same time, the effective batch size to the MoE is increased by a factor of the number of steps in the model; thus even for sparsely assigned experts, each expert can be used on a large enough sub-batch of inputs to remain computationally efficient. Another second technique that redistributes elements within a distributed model is also described, further increasing per-expert batch sizes. Experiments are performed on language modeling and machine translation tasks, showing significant gains by increasing the number of experts, compared to both SoA as well as explicitly computationally-matched baseline systems. An area that falls a bit short is in presenting plots or statistics on the real computational load and system behavior. While two loss terms were employed to balance the use of experts, these are not explored in the experiments section. It would have been nice to see the effects of these more, along with the effects of increasing effective batch sizes, e.g. measurements of the losses over the course of training, compared to the counts/histogram distributions of per-expert batch sizes. Overall I think this is a well-described system that achieves good results, using a nifty placement for the MoE that can overcome what otherwise might be a disadvantage for sparse computation. Small comment: I like Fig 3, but it's not entirely clear whether datapoints coincide between left and right plots. The H-H line has 3 points on left but 5 on the right? Also would be nice if the colors matched between corresponding lines.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
ByGtCUlVl
B1hdzd5lg
SUMMARY. The paper proposes a gating mechanism to combine word embeddings with character-level word representations. The gating mechanism uses features associated to a word to decided which word representation is the most useful. The fine-grain gating is applied as part of systems which seek to solve the task of cloze-style reading comprehension question answering, and Twitter hashtag prediction. For the question answering task, a fine-grained reformulation of gated attention for combining document words and questions is proposed. In both tasks the fine-grain gating helps to get better accuracy, outperforming state-of-the-art methods on the CBT dataset and performing on-par with state-of-the-art approach on the SQuAD dataset. ---------- OVERALL JUDGMENT This paper proposes a clever fine-grained extension of a scalar gate for combining word representation. It is clear and well written. It covers all the necessary prior work and compares the proposed method with previous similar models. I liked the ablation study that shows quite clearly the impact of individual contributions. And I also liked the fact that some (shallow) linguistic prior knowledge e.g., pos tags ner tags, frequency etc. has been used in a clever way. It would be interesting to see if syntactic features can be helpful.
7: Good paper, accept
3: The reviewer is fairly confident that the evaluation is correct
7
3
HJ-dvmfEg
B1hdzd5lg
I think the problem here is well motivated, the approach is insightful and intuitive, and the results are convincing of the approach (although lacking in variety of applications). I like the fact that the authors use POS and NER in terms of an intermediate signal for the decision. Also they compare against a sufficient range of baselines to show the effectiveness of the proposed model. I am also convinced by the authors' answers to my question, I think there is sufficient evidence provided in the results to show the effectiveness of the inductive bias introduced by the fine-grained gating model.
6: Marginally above acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
6
4
SJbtVHfVg
B1hdzd5lg
review
This paper proposes a new gating mechanism to combine word and character representations. The proposed model sets a new state-of-the-art on the CBT dataset; the new gating mechanism also improves over scalar gates without linguistic features on SQuAD and a twitter classification task. Intuitively, the vector-based gate working better than the scalar gate is unsurprising, as it is more similar to LSTM and GRU gates. The real contribution of the paper for me is that using features such as POS tags and NER help learn better gates. The visualization in Figure 3 and examples in Table 4 effectively confirm the utility of these features, very nice! In sum, while the proposed gate is nothing technically groundbreaking, the paper presents a very focused contribution that I think will be useful to the NLP community. Thus, I hope it is accepted.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
HkSTCzENe
ryUPiRvge
Important task but marginal contribution
The authors attempt to extract analytical equations governing physical systems from observations - an important task. Being able to capture succinct and interpretable rules which a physical system follows is of great importance. However, the authors do this with simple and naive tools which will not scale to complex tasks, offering no new insights or advances to the field. The contribution of the paper (and the first four pages of the submission!) can be summarised in one sentence: "Learn the weights of a small network with cosine, sinusoid, and input elements products activation functions s.t. the weights are sparse (L1)". The learnt network weights with its fixed structure are then presented as the learnt equation. This research uses tools from literature from the '90s (I haven't seen the abbreviation ANN (page 3) for a long time) and does not build on modern techniques which have advanced a lot since then. I would encourage the authors to review modern literature and continue working on this important task.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
S1MchhtNe
ryUPiRvge
Learning physical phenomenon
Thank you for an interesting perspective on the neural approaches to approximate physical phenomenon. This paper describes a method to extrapolate a given dataset and predict formulae with naturally occurring functions like sine, cosine, multiplication etc. Pros - The approach is rather simple and hence can be applied to existing methods. The major difference is incorporating functions with 2 or more inputs which was done successfully in the paper. - It seems that MLP, even though it is good for interpolation, it fails to extrapolate data to model the correct function. It was a great idea to use basis functions like sine, cosine to make the approach more explicit. Cons - Page 8, the claim that x2 cos(ax1 + b) ~ 1.21(cos(-ax1 + π + b + 0.41x2) + sin(ax1 + b + 0.41x2)) for y in [-2,2] is not entirely correct. There should be some restrictions on 'a' and 'b' as well as the approximate equality doesn't hold for all real values of 'a' and 'b'. Although, for a=2*pi and b=pi/4, the claim is correct so the model is predicting a correct solution within certain limits. - Most of the experiments involve up to 4 variables. It would be interesting to see how the neural approach models hundreds of variables. - Another way of looking at the model is that the non-linearities like sine, cosine, multiplication act as basis functions. If the data is a linear combination of such functions, the model will be able to learn the weights. As division is not one of the non-linearities, predicting expressions in Equation 13 seems unlikely. Hence, I was wondering, is it possible to make sure that this architecture is a universal approximator. Suggested Edits - Page 8, It seems that there is a typographical error in the expression 1.21(cos(ax1 + π + b + 0.41x2) + sin(ax1 + b + 0.41x2)). When compared with the predicted formula in Figure 4(b), it should be 1.21(cos(-ax1 + π + b + 0.41x2) + sin(ax1 + b + 0.41x2)).
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
BJAY0Y07g
ryUPiRvge
An interesting paper for domain adapatation with NO target domain data
Thank you for an interesting read. To my knowledge, very few papers have looked at transfer learning with **no** target domain data (the authors called this task as "extrapolation"). This paper clearly shows that the knowledge of the underlying system dynamics is crucial in this case. The experiments clearly showed the promising potential of the proposed EQL model. I think EQL is very interesting also from the perspective of interpretability, which is crucial for data analysis in scientific domains. Quesions and comments: 1. Multiplication units. By the universal approximation theorem, multiplication can also be represented by a neural network in the usual sense. I agree with the authors' explanation of interpolation and extrapolation, but I still don't quite understand why multiplication unit is crucial here. I guess is it because this representation generalises better when training data is not that representative for the future? 2. Fitting an EQL vs. fitting a polynomial. It seems to me that the number of layers in EQL has some connections to the degree of the polynomial. Assume we know the underlying dynamics we want to learn can be represented by a polynomial. Then what's the difference between fitting a polynomial (with model selection techniques to determine the degree) and fitting an EQL (with model selection techniques to determine the number of layers)? Also your experiments showed that the selection of basis functions (specific to the underlying dynamics you want to learn) is crucial for the performance. This means you need to have some prior knowledge on the form of the equation anyway! 3. Ben-David et al. 2010 has presented some error bounds for the hypothesis that is trained on source data but tested on the target data. I wonder if your EQL model can achieve better error bounds? 4. Can you comment on the comparison of your method to those who modelled the extrapolation data with **uncertainty**?
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
SyoOxS0Xl
SJCscQcge
Blackbox adversarial examples
The authors propose a method to generate adversarial examples w/o relying on knowledge of the network architecture or network gradients. The idea has some merit, however, as mentioned by one of the reviewers, the field has been studied widely, including black box setups. My main concern is that the first set of experiments allows images that are not in image space. The authors acknowledge this fact on page 7 in the first paragraph. In my opinion, this renders these experiments completely meaningless. At the very least, the outcome is not surprising to me at all. The greedy search procedure remedies this issue. The description of the proposed method is somewhat convoluted. AFAICT, first a candidate set of pixels is generated by using PERT. Then the pixels are perturbed using CYCLIC. It is not clear why this approach results in good/minimal perturbations as the candidate pixels are found using a large "p" that can result in images outside the image space. The choice of this method does not seem to be motivated by the authors. In conclusion, while the authors to an interesting investigation and propose a method to generate adversarial images from a black-box network, the overall approach and conclusions seem relatively straight forward. The paper is verbosely written and I feel like the findings could be summarized much more succinctly.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
S1RZxo-Hg
SJCscQcge
Too verbose for little insight
Paper summary: This work proposes a new algorithm to generate k-adversarial images by modifying a small fraction of the image pixels and without requiring access to the classification network weight. Review summary: The topic of adversarial images generation is of both practical and theoretical interest. This work proposes a new approach to the problem, however the paper suffers from multiple issues. It is too verbose (spending long time on experiments of limited interest); disorganized (detailed description of the main algorithm in sections 4 and 5, yet a key piece is added in the experimental section 6); and more importantly the resulting experiments are of limited interest to the reader, and the main conclusions are left unclear. This looks like an interesting line of work that has yet to materialize in a good document, it would need significant re-writing to be in good shape for ICLR. Pros: * Interesting topic * Black-box setup is most relevant * Multiple experiments * Shows that with flipping only 1~5% of pixels, adversarial images can be created Cons: * Too long, yet key details are not well addressed * Some of the experiments are of little interest * Main experiments lack key measures or additional baselines * Limited technical novelty Quality: the method description and experimental setup leave to be desired. Clarity: the text is verbose, somewhat formal, and mostly clear; but could be improved by being more concise. Originality: I am not aware of another work doing this exact same type of experiments. However the approach and results are not very surprising. Significance: the work is incremental, the issues in the experiments limit potential impact of this paper. Specific comments: * I would suggest to start by making the paper 30%~40% shorter. Reducing the text length, will force to make the argumentation and descriptions more direct, and select only the important experiments. * Section 4 seems flawed. If the modified single pixel can have values far outside of the [LB, UB] range; then this test sample is clearly outside of the training distribution; and thus it is not surprising that the classifier misbehaves (this would be true for most classifiers, e.g. decision forests or non-linear SVMs). These results would be interesting only if the modified pixel is clamped to the range [LB, UB]. * [LB, UB] is never specified, is it ? How does p = 100, compares to [LB, UB] ? To be of any use, p should be reported in proportion to [LB, UB] * The modification is done after normalization, is this realistic ? * Alg 2, why not clamping to [LB, UB] ? * Section 6, “implementing algorithm LocSearchAdv”, the text is unclear on how p is adjusted; new variables are added. This is confusion. * Section 6, what happens if p is _not_ adjusted ? What happens if a simple greedy random search is used (e.g. try 100 times a set of 5 random pixels with value 255) ? * Section 6, PTB is computed over all pixels ? including the ones not modified ? why is that ? Thus LocSearchAdv PTB value is not directly comparable to FGSM, since it intermingles with #PTBPixels (e.g. “in many cases far less average perturbation” claim). * Section 6, there is no discussion on the average number of model evaluations. This would be equivalent to the number of requests made to a system that one would try to fool. This number is important to claim the “effectiveness” of such black box attacks. Right now the text only mentions the upper bound of 750 network evaluations. * How does the number of network evaluations changes when adjusting or not adjusting p during the optimization ? * Top-k is claimed as a main point of the paper, yet only one experiment is provided. Please develop more, or tune-down the claims. * Why is FGSM not effective for batch normalized networks ? Has this been reported before ? Are there other already published techniques that are effective for this scenario ? Comparing to more methods would be interesting. * If there is little to note from section 4 results, what should be concluded from section 6 ? That is possible to obtain good results by modifying only few pixels ? What about selecting the “top N” largest modified pixels from FGSM ? Would these be enough ? Please develop more the baselines, and the specific conclusions of interest. Minor comments: * The is an abuse of footnotes, most of them should be inserted in the main text. * I would suggest to repeat twice or thrice the meaning of the main variables used (e.g. p, r, LB, UB) * Table 1,2,3 should be figures * Last line of first paragraph of section 6 is uninformative. * Very tiny -> small
4: Ok but not good enough - rejection
3: The reviewer is fairly confident that the evaluation is correct
4
3
S1_Yb5N4g
SJCscQcge
review: incremental
The paper presents a method for generating adversarial input images for a convolutional neural network given only black box access (ability to obtain outputs for chosen inputs, but no access to the network parameters). However, the notion of adversarial example is somewhat weakened in this setting: it is k-misclassification (ensuring the true label is not a top-k output), instead of misclassification to any desired target label. A similar black-box setting is examined in Papernot et al. (2016c). There, black-box access is used to train a substitute for the network, which is then attacked. Here, black-box access in instead exploited via local search. The input is perturbed, the resulting change in output scores is examined, and perturbations that push the scores towards k-misclassification are kept. A major concern with regard to novelty is that this greedy local search procedure is analogous to gradient descent; a numeric approximation (observe change in output for corresponding change in input) is used instead of backpropagation, since one does not have access to the network parameters. As such, the greedy local search algorithm itself, to which the paper devotes a large amount of discussion, is not surprising and the paper is fairly incremental in terms of technical novelty.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
rktOx2WNl
SyJNmVqgg
Review
This work proposes to augment normal gradient descent algorithms with a "Data Filter", that acts as a curriculum teacher by selecting which examples the trained target network should see to learn optimally. Such a filter is learned simultaneously to the target network, and trained via Reinforcement Learning algorithms receiving rewards based on the state of training with respect to some pseudo-validation set. Stylistic comment, please use the more common style of "(Author, year)" rather than "Author (year)" when the Author is *not* referred to or used in the sentence. E.g. "and its variants such as Adagrad Duchi et al. (2011)" should be "such as Adagrad (Duchi et al., 2011)", and "proposed in Andrychowicz et al. (2016)," should remain so. I think the paragraph containing "What we need to do is, after seeing the mini-batch Dt of M training instances, we dynamically determine which instances in Dt are used for training and which are filtered." should be clarified. What is "seeing"? That is, you should mention explicitly that you do the forward-pass first, then compute features from that, and then decide for which examples to perform the backwards pass. There are a few choices in this work which I do not understand: Why wait until the end of the episode to update your reinforce policy (algorithm 2), but train your actor critic at each step (algorithm 3)? You say REINFORCE has high variance, which is true, but does not mean it cannot be trained at each step (unless you have some experiments that suggest otherwise, and if so they should be included or mentionned in the paper). Similarly, why not train REINFORCE with the same reward as your Actor-Critic model? And vice-versa? You claim several times that a limitation of REINFORCE is that you need to wait for the episode to be over, but considering your data is i.i.d., you can make your episode be anything from a single training step, one D_t, to the whole multi-epoch training procedure. I have a few qualms with the experimental setting: - is Figure 2 obtained from a single (i.e. one per setup) experiment? From different initial weights? If so, there is no proper way of knowing whether results are chance or not! This is a serious concern for me. - with most state-of-the-art work using optimization methods such as Adam and RMSProp, is it surprising that they were not experimented with. - it is not clear what the learning rates are; how fast should the RL part adapt to the SL part? Its not clear that this was experimented with at all. - the environment, i.e. the target network being trained, is not stationnary at all. It would have been interesting to measure how much the policy changes as a function of time. Figure 3, could both be the result of the policy adapting, or of the policy remaining fixed and the features changing (which could indicate a failure of the policy to adapt). - in fact it is not really adressed in the paper that the environment is non-stationary, given the current setup, the distribution of features will change as the target network progresses. This has an impact on optimization. - how is the "pseudo-validation" data, target to the policy, chosen? It should be a subset of the training data. The second paragraph of section 3.2 suggests something of the sort, but then your algorithms suggest that the same data is used to train both the policies and the networks, so I am unsure of which is what. Overall the idea is novel and interesting, the paper is well written for the most part, but the methodology has some flaws. Clearer explanations and either more justification of the experimental choices or more experiments are needed to make this paper complete. Unless the authors convince me otherwise, I think it would be worth waiting for more experiments and submitting a very strong paper rather than presenting this (potentially powerful!) idea with weak results.
6: Marginally above acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
6
4
HyoMSTSVl
SyJNmVqgg
Final Review
Final review: The writers were very responsive and I agree the reviewer2 that their experimental setup is not wrong after all and increased the score by one. But I still think there is lack of experiments and the results are not conclusive. As a reader I am interested in two things, either getting a new insight and understanding something better, or learn a method for a better performance. This paper falls in the category two, but fails to prove it with more throughout and rigorous experiments. In summary the paper lacks experiments and results are inconclusive and I do not believe the proposed method would be quite useful and hence not a conference level publication. -- The paper proposes to train a policy network along the main network for selecting subset of data during training for achieving faster convergence with less data. Pros: It's well written and straightforward to follow The algorithm has been explained clearly. Cons: Section 2 mentions that the validation accuracy is used as one of the feature vectors for training the NDF. This invalidates the experiments, as the training procedure is using some data from the validation set. Only one dataset has been tested on. Papers such as this one that claim faster convergence rate should be tested on multiple datasets and network architectures to show consistency of results. Especially larger datasets as the proposed methods is going to use less training data at each iteration, it has to be shown in much larger scaler datasets such as Imagenet. As discussed more in detail in the pre-reviews question, if the paper is claiming faster convergence then it has to compare the learning curves with other baselines such Adam. Plain SGD is very unfair comparison as it is almost never used in practice. And this is regardless of what is the black box optimizer they use. The case could be that Adam alone as black box optimizer works as well or better than Adam as black box + NDF.
4: Ok but not good enough - rejection
4
-1
SyBXdRUEx
SyJNmVqgg
data filtering for faster sgd
Paper is easy to follow, Idea is pretty clear and makes sense. Experimental results are hard to judge, it would be nice to have other baselines. For faster training convergence, the question is how well tuned SGD is, I didn't see any mentioning of learning rate schedule. Also, it would be important to test this on other data sets. Success with filtering training data could be task dependent.
7: Good paper, accept
7
-1
SySmUNZNg
ByOK0rwlx
Novel quantization method to reduce memory and complexity of pre-trained networks, but benefit over other methods is unclear
This paper explores a new quantization method for both the weights and the activations that does not need re-training. In VGG-16 the method reaches compression ratios of 20x and experiences a speed-up of 15x. The paper is very well written and clearly exposes the details of the methodology and the results. My major criticisms are three-fold: for one, the results are not compared to one of the many other pruning methods that are described in section 1.1, and as such the performance of the method is difficult to judge from the paper alone. Second, there have been several other compression schemes involving pruning, re-training and vector-quantization [e.g. 1, 2, 3] that seem to achieve much higher accuracies, compression ratios and speed-ups. Hence, for the practical application of running such networks on low-power, low-memory devices, other methods seem to be much more suited. The advantage of the given method - other then possibly reducing the time it takes to compress the network - is thus unclear. In particular, taking a pre-trained network as a starting point for a quantized model that is subsequently fine-tuned might not take much longer to process then the method given here (but maybe the authors can quantify this?). Finally, much of the speed-up and memory reduction in the VGG-model seems to arise from the three fully-connected layers, in particular the last one. The speed-up in the convolutional layers is comparably small, making me wonder how well the method would work in all-convolutional networks such as the Inception architecture. [1] Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding, https://arxiv.org/abs/1510.00149 [2] Compressing Deep Convolutional Networks using Vector Quantization, https://arxiv.org/abs/1412.6115 [3] XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks, https://arxiv.org/abs/1603.05279
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
rk9ryJzNx
ByOK0rwlx
Clarify my comments
I do need to see the results in a clear table. Original results and results when compression is applied for all the tasks. In any case, i would like to see the results when the compression is applied to state of the art nets where the float representation is important. For instance a network with 0.5% - 0.8% in MNIST. A Imagenet lower that 5% - 10%. Some of this results are feasible with float representation but probably imposible for restricted representations.
5: Marginally below acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
5
3
HJ5-4JL4e
ByOK0rwlx
Review
This paper addresses to reduce test-time computational load of DNNs. Another factorization approach is proposed and shows good results. The comparison to the other methods is not comprehensive, the paper provides good insights.
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
SyxnNWM4e
HJ7O61Yxe
Interesting idea but formulation and experiments not convincing
This manuscript proposes an approach for modeling correlated timeseries through a combination of loss functions which depend on neural networks. The loss functions correspond to: data fit term, autoregressive latent state term, and a term which captures relations between pairs of timeseries (relations have to be given as prior information). Modeling relational timeseries is a well-researched problem, however little attention has been given to it in the neural network community. Perhaps the reason for this is the importance of having uncertainty in the representation. The authors correctly identify this need and consider an approach which considers distributions in the state space. The formulation is quite straightforward by combining loss functions. The model adds to Ziat et al. 2016 in certain aspects which are well motivated, but unfortunately implemented in an unconvincing way. To start with, uncertainty is not treated in a very principled way, since the inference in the model is rather naive; I'd expect employing a VAE framework [1] for better uncertainty handling. Furthermore, the Gaussian co-variance collapses into a variance, which is the opposite of what one would want for modelling correlated time-series. There are approaches which take these correlations into account in the states, e.g. [2]. Moreover, the treatment of uncertainty only allows for linear decoding function f. This significantly reduces the power of the model. State of the art methods in timeseries modeling have moved beyond this constraint, especially in the Gaussian process community e.g. [2,3,4,5]. Comparing to a few of these methods, or at least discussing them would be useful. References: [1] Kingma and Welling. Auto-encoding Variational Bayes. arXiv:1312.6114 [2] Damianou et al. Variational Gaussian process dynamical systems. NIPS 2011. [3] Mattos et al. Recurrent Gaussian processes. ICLR 2016. [4] Frigola. Bayesian Time Series Learning with Gaussian Processes, University of Cambridge, PhD Thesis, 2015. [5] Frigola et al. Variational Gaussian Process State-Space Models. NIPS 2014 One innovation is that the prior structure of the correlation needs to be given. This is a potentially useful and also original structural component. However, it also constitutes a limitation in some sense, since it is unrealistic in many scenarios to have this prior information. Moreover, the particular regularizer that makes "similar" timeseries to have closeness in the state space seems problematic. Some timeseries groups might be more "similar" than others, and also the similarity might be of different nature across groups. These variations cannot be well captured/distilled by a simple indicator variable e_ij. Furthermore, these variables are in practice taken to be binary (by looking at the experiments), which would make it even harder to model rich correlations. The experiments show that the proposed method works, but they are not entirely convincing. Importantly, they do not shed enough light into the different properties of the model w.r.t its different parts. For example, the effect and sensitivity of the different regularizers. The authors state in a pre-review answer that they amended with some more results, but I can't see a revision in openreview (please let me know if I've missed it). From the performance point of view, the results are not particularly exciting, especially given the fact that it's not clear which loss is better (making it difficult to use the method in practice). It would also be very interesting to report the optimized values of the parameters \lambda, to get an idea of how the different losses behave. Timeseries analysis is a very well-researched area. Given the above, it's not clear to me why one would prefer to use this model over other approaches. Methodology wise, there are no novel components that offer a proven advantage with respect to past methods. The uncertainty in the states and the correlation of the time-series are the aspects which could add an advantage, but are not adequately researched in this paper.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
rkHCIUMVg
HJ7O61Yxe
Important line of research, muddled presentation and unconvincing empirical results
Because the authors did not respond to reviewer feedback, I am maintaining my original review score. ----- This paper proposes to model relational (i.e., correlated) time series using a deep learning-inspired latent variable approach: they design a flexible parametric (but not generative) model with Gaussian latent factors and fit it using a rich training objective including terms for reconstruction (of observed time series) error, smoothness in the latent state space (via a KL divergence term encouraging neighbor states to be similarly distributed), and a final regularizer that encourages related time series to have similar latent state trajectories. Relations between trajectories are hard coded based on pre-existing knowledge, i.e., latent state trajectories for neighboring (wind speed) base stations should be similar. The model appears to be fit using gradient simple descent. The authors propose several elaborations, including a nonlinear transition function (based on an MLP) and a reconstruction error term that takes variance into account. However, the model is restricted to using a linear decoder. Experimental results are positive but not convincing. Strengths: - The authors target a worthwhile and challenging problem: incorporating the modeling of uncertainty over hidden states with the power of flexible neural net-like models. - The idea of representing relationships between hidden states using KL divergence between their (distributions over) corresponding hidden states is clever. Combined with the Gaussian distribution over hidden states, the resulting regularization term is simple and differentiable. - This general approach -- focusing on writing down the problem as a neural network-like loss function -- seems robust and flexible and could be combined with other approaches, including variants of variational autoencoders. Weaknesses: - The presentation is a muddled, especially the model definition in Sec. 3.3. The authors introduce four variants of their model with different combinations of decoder (with and without variance term) and linear vs. MLP transition function. It appears that the 2,2 variant is generally better but not on all metrics and often by small margins. This makes drawing a solid conclusions difficult: what each component of the loss contributes, whether and how the nonlinear transition function helps and how much, how in practice the model should be applied, etc. I would suggest two improvements to the manuscript: (1) focus on the main 2,2 variant in Sec. 3.3 (with the hypothesis that it should perform best) and make the simpler variants additional "baselines" described in a paragraph in Sec. 4.1; (2) perform more thorough experiments with larger data sets to make a stronger case for the superiority of this approach. - The authors only allude to learning (with references to gradient descent and ADAM during model description) in this framework. Inference gets its one subsection but only one sentence that ends in an ellipsis (?). - It's unclear what is the purpose of introducing the inequality in Eq. 9. - Experimental results are not convincing: given the size of the data, the differences vs. the RNN and KF baselines is probably not significant, and these aren't particularly strong baselines (especially if it is in fact an RNN and not an LSTM or GRU). - The position of this paper is unclear with respect to variational autoencoders and related models. Recurrent variants of VAEs (e.g., Krishnan, et al., 2015) seem to achieve most of the same goals as far as uncertainty modeling is concerned. It seems like those could easily be extended to model relationships between time series using the simple regularization strategy used here. Same goes for Johnson, et al., 2016 (mentioned in separate question). This is a valuable research direction with some intriguing ideas and interesting preliminary results. I would suggest that the authors restructure this manuscript a bit, striving for clarity of model description similar to the papers cited above and providing greater detail about learning and inference. They also need to perform more thorough experiments and present results that tell a clear story about the strengths and weaknesses of this approach.
4: Ok but not good enough - rejection
3: The reviewer is fairly confident that the evaluation is correct
4
3
ryC8AjbVx
HJ7O61Yxe
Interesting model, further experiments required
In absence of authors' response, the rating is maintained. --- This paper introduces a nonlinear dynamical model for multiple related multivariate time series. It models a linear observation model conditioned on the latent variables, a linear or nonlinear dynamical model between consecutive latent variables and a similarity constraint between any two time series (provided as prior data and non-learnable). The predictions/constraints given by the three components of the model are Gaussian, because the model predicts both the mean and the variance or covariance matrix. Inference is forward only. The model is evaluated on four datasets, and compared to several baselines: plain auto-regressive models, feed-forward networks, RNN and dynamic factor graphs DFGs, which are RNNs with forward and backward inference of the latent variables. The model, which introduces lateral constraints between different time series, and which predicts both the mean and covariance seems interesting, but presents two limitations. First of all, the paper should refer to variational auto-encoders / deep gaussian models, which also predict the mean and the variance during inference. Secondly, the datasets are extremely small. For example, the WHO contains only 91 times series of 52*10 = 520 time points. Although the experiments seem to suggest that the proposed model tends to outperform RNNs, the datasets are very small and the high variance in the results indicates that further experiments, with longer time series, are required. The paper could also easily be extended with more information about the model (what is the architecture of the MLP) as well as time complexity comparison between the models (especially between DFGs and this model). Minor remark: The footnote 2 on page 5 seems to refer to the structural regularization term, not to the dynamical term.
4: Ok but not good enough - rejection
4
-1
Hkxf8DNNe
BJ6oOfqge
simple approach showing some decent results
This paper presents a model for semi-supervised learning by encouraging feature invariance to stochastic perturbations of the network and/or inputs. Two models are described: One where an invariance term is applied between different instantiations of the model/input a single training step, and a second where invariance is applied to features for the same input point across training steps via a cumulative exponential averaging of the features. These models evaluated using CIFAR-10 and SVHN, finding decent gains of similar amounts in each case. An additional application is also explored at the end, showing some tolerance to corrupted labels as well. The authors also discuss recent work by Sajjadi &al that is very similar in spirit, which I think helps corroborate the findings here. My largest critique is it would have been nice to see applications on larger datasets as well. CIFAR and SVHN are fairly small test cases, though adequate for demonstration of the idea. For cases of unlabelled data especially, it would be good to see tests with on the order of 1M+ data samples, with 1K-10K labeled, as this is a common case when labels are missing. On a similar note, data augmentations are restricted to only translations and (for CIFAR) horizontal flips. While "standard," as the paper notes, more augmentations would have been interesting to see --- particularly since the model is designed explicitly to take advantage of random sampling. Some more details might also pop up, such as the one the paper mentions about handling horizontal flips in different ways between the two model variants. Rather than restrict the system to a particular set of augmentations, I think it would be interesting to push it further, and see how its performance behaves over a larger array of augmentations and (even fewer) numbers of labels. Overall, this seems like a simple approach that is getting decent results, though I would have liked to see more and larger experiments to get a better sense for its performance characteristics. Smaller comment: the paper mentions "dark knowledge" a couple times in explaining results, e.g. bottom of p.6. This is OK for a motivation, but in analyzing the results I think it may be possible to have something more concrete. For instance, the consistency term encourages feature invariance to the stochastic sampling more strongly than would a classification loss alone.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
B1u6EURmg
BJ6oOfqge
This paper presents a semi-supervised technique for “self-ensembling” where the model uses a consensus prediction (computed from previous epochs) as a target to regress to, in addition to the usual supervised learning loss. This has connections to the “dark knowledge” idea, ladder networks work is shown in this paper to be a promising technique for scenarios with few labeled examples (but not only). The paper presents two versions of the idea: one which is computationally expensive (and high variance) in that it needs two passes through the same example at a given step, and a temporal ensembling method that is stabler, cheaper computationally but more memory hungry and requires an extra hyper-parameter. My thoughts on this work are mostly positive. The drawbacks that I see are that the temporal ensembling work requires potentially a lot of memory, and non-trivial infrastructure / book-keeping for imagenet-sized experiments. I am quite confused by the Figure 2 / Section 3.4 experiments about tolerance to noisy labels: it’s *very* incredible to me that by making 90% of the labels random one can still train a classifier that is either 30% accurate or ~78% accurate (depending on whether or not temporal ensembling was used). I don’t see how that can happen, basically. Minor stuff: Please bold the best-in-category results in your tables. I think it would be nice to talk about the ramp-up of w(t) in the main paper. The authors should consider putting the state of the art results for the fully-supervised case in their tables, instead of just their own. I am confused as to why the authors chose not to use more SVHN examples. The stated reason that it’d be “too easy” seems a bit contrived: if they used all examples it would also make it easy to compare to previous work.
8: Top 50% of accepted papers, clear accept
8
-1
SyezfkfEg
BJ6oOfqge
Review
This work explores taking advantage of the stochasticity of neural network outputs under randomized augmentation and regularization techniques to provide targets for unlabeled data in a semi-supervised setting. This is accomplished by either applying stochastic augmentation and regularization on a single image multiple times per epoch and encouraging the outputs to be similar (Π-model) or by keeping a weighted average of past epoch outputs and penalizing deviations of current network outputs from this running mean (temporal ensembling). The core argument is that these approaches produce ensemble predictions which are likely more accurate than the current network and are thus good targets for unlabeled data. Both approaches seem to work quite well on semi-supervised tasks and some results show that they are almost unbelievably robust to label noise. The paper is clearly written and provides sufficient details to reproduce these results in addition to providing a public code base. The core idea of the paper is quite interesting and seems to result in higher semi-supervised accuracy than prior work. I also found the attention to and discussion of the effect of different choices of data augmentation to be useful. I am a little surprised that a standard supervised network can achieve 30% accuracy on SVHN given 90% random training labels. This would only give 19% correctly labeled data (9% by chance + 10% unaltered). I suppose the other 81% would not provide a consistent training signal such that it is possible, but it does seem quite unintuitive. I tried to look through the github for this experiment but it does not seem to be included. As for the resistance of Π-model and temporal ensembling to this label noise, I find that somewhat more believable given the large weights placed on the consistency constraint for this task. The authors should really include discussion of w(t) in the main paper. Especially because the tremendous difference in w_max in the incorrect label tolerance experiment (10x for Π-model and 100x for temporal ensembling from the standard setting). Could the authors comment towards the scalability for larger problems? For ImageNet, you would need to store around 4.8 gigs for the temporal ensembling method or spend 2x as long training with Π-model. Can the authors discuss sensitivity of this approach to the amount and location of dropout layers in the architecture? Preliminary rating: I think this is a very interesting paper with quality results and clear presentation. Minor note: 2nd paragraph of page one 'without neither' -> 'without either'
9: Top 15% of accepted papers, strong accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
9
4
B1atIp-Ve
BJuysoFeg
An interesting paper that shows improvements, but I am not sure about its technical advantage
Overall I think this is an interesting paper which shows empirical performance improvement over baselines. However, my main concern with the paper is regarding its technical depth, as the gist of the paper can be summarized as the following: instead of keeping the batch norm mean and bias estimation over the whole model, estimate them on a per-domain basis. I am not sure if this is novel, as this is a natural extension of the original batch normalization paper. Overall I think this paper is more fit as a short workshop presentation rather than a full conference paper. Detailed comments: Section 3.1: I respectfully disagree that the core idea of BN is to align the distribution of training data. It does this as a side effect, but the major purpose of BN is to properly control the scale of the gradient so we can train very deep models without the problem of vanishing gradients. It is plausible that intermediate features from different datasets naturally show as different groups in a t-SNE embedding. This is not the particular feature of batch normalization: visualizing a set of intermediate features with AlexNet and one gets the same results. So the premise in section 3.1 is not accurate. Section 3.3: I have the same concern as the other reviewer. It seems to be quite detatched from the general idea of AdaBN. Equation 2 presents an obvious argument that the combined BN-fully_connected layer forms a linear transform, which is true in the original BN case and in this case as well. I do not think it adds much theoretical depth to the paper. (In general the novelty of this paper seems low) Experiments: - section 4.3.1 is not an accurate measure of the "effectiveness" of the proposed method, but a verification of a simple fact: previously, we normalize the source domain features into a Gaussian distribution. the proposed method is explicitly normalizing the target domain features into the same Gaussian distribution as well. So, it is obvious that the KL divergence between these two distributions are closer - in fact, one is *explicitly* making them close. However, this does not directly correlate to the effectiveness of the final classification performance. - section 4.3.2: the sensitivity analysis is a very interesting read, as it suggests that only a very few number of images are needed to account for the domain shift in the AdaBN parameter estimation. This seems to suggest that a single "whitening" operation is already good enough to offset the domain bias (in both cases shown, a single batch is sufficient to recover about 80% of the performance gain, although I cannot get data for even smaller number of examples from the figure). It would thus be useful to have a comparison between these approaches, and also a detailed analysis of the effect from each layer of the model - the current analysis seems a bit thin.
Overall I think this is an interesting paper which shows empirical performance improvement over baselines. However, my main concern with the paper is regarding its technical depth, as the gist of the paper can be summarized as the following: instead of keeping the batch norm mean and bias estimation over the whole model, estimate them on a per-domain basis. I am not sure if this is novel, as this is a natural extension of the original batch normalization paper. Overall I think this paper is more fit as a short workshop presentation rather than a full conference paper. Detailed comments: Section 3.1: I respectfully disagree that the core idea of BN is to align the distribution of training data. It does this as a side effect, but the major purpose of BN is to properly control the scale of the gradient so we can train very deep models without the problem of vanishing gradients. It is plausible that intermediate features from different datasets naturally show as different groups in a t-SNE embedding. This is not the particular feature of batch normalization: visualizing a set of intermediate features with AlexNet and one gets the same results. So the premise in section 3.1 is not accurate. Section 3.3: I have the same concern as the other reviewer. It seems to be quite detatched from the general idea of AdaBN. Equation 2 presents an obvious argument that the combined BN-fully_connected layer forms a linear transform, which is true in the original BN case and in this case as well. I do not think it adds much theoretical depth to the paper. (In general the novelty of this paper seems low) Experiments: - section 4.3.1 is not an accurate measure of the "effectiveness" of the proposed method, but a verification of a simple fact: previously, we normalize the source domain features into a Gaussian distribution. the proposed method is explicitly normalizing the target domain features into the same Gaussian distribution as well. So, it is obvious that the KL divergence between these two distributions are closer - in fact, one is *explicitly* making them close. However, this does not directly correlate to the effectiveness of the final classification performance. - section 4.3.2: the sensitivity analysis is a very interesting read, as it suggests that only a very few number of images are needed to account for the domain shift in the AdaBN parameter estimation. This seems to suggest that a single "whitening" operation is already good enough to offset the domain bias (in both cases shown, a single batch is sufficient to recover about 80% of the performance gain, although I cannot get data for even smaller number of examples from the figure). It would thus be useful to have a comparison between these approaches, and also a detailed analysis of the effect from each layer of the model - the current analysis seems a bit thin.
4: The reviewer is confident but not absolutely certain that the evaluation is correct
-1
4
rkpVV6H4l
BJuysoFeg
trivially simple yet effective
This paper proposes a simple domain adaptation technique in which batch normalization is performed separately in each domain. Pros: The method is very simple and easy to understand and apply. The experiments demonstrate that the method compares favorably with existing methods on standard domain adaptation tasks. The analysis in section 4.3.2 shows that a very small number of target domain samples are needed for adaptation of the network. Cons: There is little novelty -- the method is arguably too simple to be called a “method.” Rather, it’s the most straightforward/intuitive approach when using a network with batch normalization for domain adaptation. The alternative -- using the BN statistics from the source domain for target domain examples -- is less natural, to me. (I guess this alternative is what’s done in the Inception BN results in Table 1-2?) The analysis in section 4.3.1 is superfluous except as a sanity check -- KL divergence between the distributions should be 0 when each distribution is shifted/scaled to N(0,1) by BN. Section 3.3: it’s not clear to me what point is being made here. Overall, there’s not much novelty here, but it’s hard to argue that simplicity is a bad thing when the method is clearly competitive with or outperforming prior work on the standard benchmarks (in a domain adaptation tradition that started with “Frustratingly Easy Domain Adaptation”). If accepted, Sections 4.3.1 and 3.3 should be removed or rewritten for clarity for a final version.
This paper proposes a simple domain adaptation technique in which batch normalization is performed separately in each domain. Pros: The method is very simple and easy to understand and apply. The experiments demonstrate that the method compares favorably with existing methods on standard domain adaptation tasks. The analysis in section 4.3.2 shows that a very small number of target domain samples are needed for adaptation of the network. Cons: There is little novelty -- the method is arguably too simple to be called a “method.” Rather, it’s the most straightforward/intuitive approach when using a network with batch normalization for domain adaptation. The alternative -- using the BN statistics from the source domain for target domain examples -- is less natural, to me. (I guess this alternative is what’s done in the Inception BN results in Table 1-2?) The analysis in section 4.3.1 is superfluous except as a sanity check -- KL divergence between the distributions should be 0 when each distribution is shifted/scaled to N(0,1) by BN. Section 3.3: it’s not clear to me what point is being made here. Overall, there’s not much novelty here, but it’s hard to argue that simplicity is a bad thing when the method is clearly competitive with or outperforming prior work on the standard benchmarks (in a domain adaptation tradition that started with “Frustratingly Easy Domain Adaptation”). If accepted, Sections 4.3.1 and 3.3 should be removed or rewritten for clarity for a final version.
3: The reviewer is fairly confident that the evaluation is correct
-1
3
rJW8h4GEl
BJuysoFeg
Final review
Update: I thank the authors for their comments. I still think that the method needs more experimental evaluation: for now, it's restricted to the settings in which one can use pre-trained ImageNet model, but it's also important to show the effectiveness in scenarios where one needs to train everything from scratch. I would love to see a fair comparison of the state-of-the-art methods on toy datasets (e.g. see (Bousmalis et al., 2016), (Ganin & Lempitsky, 2015)). In my opinion, it's a crucial point that doesn't allow me to increase the rating. This paper proposes a simple trick turning batch normalization into a domain adaptation technique. The authors show that per-batch means and variances normally computed as a part of the BN procedure are sufficient to discriminate the domain. This observation leads to an idea that adaptation for the target domain can be performed by replacing population statistics computed on the source dataset by the corresponding statistics from the target dataset. Overall, I think the paper is more suitable for a workshop track rather than for the main conference track. My main concerns are the following: 1. Although the main idea is very simple, it feels like the paper is composed in such a way to make the main contribution less obvious (e.g. the idea could have been described in the abstract but the authors avoided doing so). 2. (This one is from the pre-review questions) The authors are using much stronger base CNN which may account for the bulk of the reported improvement. In order to prove the effectiveness of the trick, the authors would need to conduct a fair comparison against competing methods. It would be highly desirable to conduct this comparison also in the case of a model trained from scratch (as opposed to reusing ImageNet-trained networks).
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
Hk4IU9t4g
SJJN38cge
review
This work proposes to use basic probability assignment to improve deep transfer learning. A particular re-weighting scheme inspired by Dempster-Shaffer and exploiting the confusion matrix of the source task is introduced. The authors also suggest learning the convolutional filters separately to break non-convexity. The main problem with this paper is the writing. There are many typos, and the presentation is not clear. For example, the way the training set for weak classifiers are constructed remains unclear to me despite the author's previous answer. I do not buy the explanation about the use of both training and validation sets to compute BPA. Also, I am not convinced non-convexity is a problem here and the author does not provide an ablation study to validate the necessity of separately learning the filters. One last question is CIFAR has three channels and MNIST only one: How it this handled when pairing the datasets in the second set of experiments? Overall, I believe the proposed idea of reweighing is interesting, but the work can be globally improved/clarified. I suggest a reject.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
S1m50VGEl
SJJN38cge
Final review.
Update: I thank the author for his comments! At this point, the paper is still not suitable for publication, so I'm leaving the rating untouched. This paper proposes a transfer learning method addressing optimization complexity and class imbalance. My main concerns are the following: 1. The paper is quite hard to read due to typos, unusual phrasing and loose use of terminology like “distributed”, “transfer learning” (meaning “fine-tuning”), “softmax” (meaning “fully-connected”), “deep learning” (meaning “base neural network”), etc. I’m still not sure I got all the details of the actual algorithm right. 2. The captions to the figures and tables are not very informative – one has to jump back and forth through the paper to understand what the numbers/images mean. 3. From what I understand, the authors use “conventional transfer learning” to refer to fine-tuning of the fully-connected layers only (I’m judging by Figure 1). In this case, it’s essential to compare the proposed method with regimes when some of the convolutional layers are also updated. This comparison is not present in the paper. Comments on the pre-review questions: 1. Question 1: If the paper only considers the case |C|==|L|, it would be better to reduce the notation clutter. 2. Question 2: It is still not clear what the authors mean by distributed transfer learning. Figure 1 is supposed to highlight the difference from the conventional approach (fine-tuning of the fully-connected layers; by the way, I don’t think, Softmax is a conventional term for fully-connected layers). From the diagram, it follows that the base CNN has the same number of convolutional filters at every layer and, in order to obtain a distributed ensemble, we need to connect (for some reason) filters with the same indices. This does not make a lot of sense to me but I’m probably misinterpreting the figure. Could the authors revise the diagram to make it clearer? Overall, I think the paper needs significant refinement in order improve the clarity of presentation and thus cannot be accepted as it is now.
Update: I thank the author for his comments! At this point, the paper is still not suitable for publication, so I'm leaving the rating untouched. This paper proposes a transfer learning method addressing optimization complexity and class imbalance. My main concerns are the following: 1. The paper is quite hard to read due to typos, unusual phrasing and loose use of terminology like “distributed”, “transfer learning” (meaning “fine-tuning”), “softmax” (meaning “fully-connected”), “deep learning” (meaning “base neural network”), etc. I’m still not sure I got all the details of the actual algorithm right. 2. The captions to the figures and tables are not very informative – one has to jump back and forth through the paper to understand what the numbers/images mean. 3. From what I understand, the authors use “conventional transfer learning” to refer to fine-tuning of the fully-connected layers only (I’m judging by Figure 1). In this case, it’s essential to compare the proposed method with regimes when some of the convolutional layers are also updated. This comparison is not present in the paper. Comments on the pre-review questions: 1. Question 1: If the paper only considers the case |C|==|L|, it would be better to reduce the notation clutter. 2. Question 2: It is still not clear what the authors mean by distributed transfer learning. Figure 1 is supposed to highlight the difference from the conventional approach (fine-tuning of the fully-connected layers; by the way, I don’t think, Softmax is a conventional term for fully-connected layers). From the diagram, it follows that the base CNN has the same number of convolutional filters at every layer and, in order to obtain a distributed ensemble, we need to connect (for some reason) filters with the same indices. This does not make a lot of sense to me but I’m probably misinterpreting the figure. Could the authors revise the diagram to make it clearer? Overall, I think the paper needs significant refinement in order improve the clarity of presentation and thus cannot be accepted as it is now.
3: The reviewer is fairly confident that the evaluation is correct
-1
3
HJm7-kf4g
SJJN38cge
This paper proposed to use the BPA criterion for classifier ensembles. My major concern with the paper is that it attempts to mix quite a few concepts together, and as a result, some of the simple notions becomes a bit hard to understand. For example: (1) "Distributed" in this paper basically means classifier ensembles, and has nothing to do with the distributed training or distributed computation mechanism. Granted, one can train these individual classifiers in a distributed fashion but this is not the point of the paper. (2) The paper uses "Transfer learning" in its narrow sense: it basically means fine-tuning the last layer of a pre-trained classifier. Aside from the concept mixture of the paper, other comments I have about the paper are: (1) I am not sure how BPA address class inbalance better than simple re-weighting. Essentially, the BPA criteria is putting equal weights on different classes, regardless of the number of training data points each class has. This is a very easy thing to address in conventional training: adding a class-specific weight term to each data point with the value being the inverse of the number of data points will do. (2) Algorithm 2 is not presented correctly as it implies that test data is used during training, which is not correct: only training and validation dataset should be used. I find the paper's use of "train/validation" and "test" quite confusing: why "train/validation" is always presented together? How to properly distinguish between them? (3) If I understand correctly, the paper is proposing to compute the BPA in a batch fashion, i.e. BPA can only be computed when running the model over the full train/validation dataset. This contradicts with the stochastic gradient descent that are usually used in deep net training - how does BPA deal with that? I believe that an experimental report on the computation cost and timing is missing. In general, I find the paper not presented in its clearest form and a number of key definitions ambiguous.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
B1F8hRVEg
Hkg8bDqee
novel idea but requires more details / experimentation
The paper reads well and the idea is new. Sadly, many details needed for replicating the results (such as layer sizes of the CNNs, learning rates) are missing. The training of the introspection network could have been described in more detail. Also, I think that a model, which is closer to the current state-of-the-art should have been used in the ImageNet experiments. That would have made the results more convincing. Due to the novelty of the idea, I recommend the paper. I would increase the rating if an updated draft addresses the mentioned issues.
8: Top 50% of accepted papers, clear accept
8
-1
SktFDJDNl
Hkg8bDqee
Review
In this paper, the authors use a separate introspection neural network to predict the future value of the weights directly from their past history. The introspection network is trained on the parameter progressions collected from training separate set of meta learning models using a typical optimizer, e.g. SGD. Pros: + The organization is generally very clear + Novel meta-learning approach that is different than the previous learning to learn approach Cons: - The paper will benefit from more thorough experiments on other neural network architectures where the geometry of the parameter space are sufficiently different than CNNs such as fully connected and recurrent neural networks. - Neither MNIST nor CIFAR experimental section explained the architectural details - Mini-batch size for the experiments were not included in the paper - Comparison with different baseline optimizer such as Adam would be a strong addition or at least explain how the hyper-parameters, such as learning rate and momentum, are chosen for the baseline SGD method. Overall, due to the omission of the experimental details in the current revision, it is hard to draw any conclusive insight about the proposed method.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
Bkm_OazEx
Hkg8bDqee
Valuable insight but needs careful analysis
EDIT: Updated score. See additional comment. I quite like the main idea of the paper, which is based on the observation in Sec. 3.0 - that the authors find many predictable patterns in the independent evolution of weights during neural network training. It is very encouraging that a simple neural network can be used to speed up training by directly predicting weights. However the technical quality of the current paper leaves much to be desired, and I encourage the authors to do more rigorous analysis of the approach. Here are some concrete suggestions: - The findings in Section 3.0 which motivate the approach, should be clearly presented in the paper. Presently they are stated as anecdotes. - A central issue with the paper is that the training of the Introspection network I is completely glossed over. How well did the training work, in terms of training, validation/test losses? How well does it need to work in order to be useful for speeding up training? These are important questions for anyone interested in this approach. - An additional important issue is that of baselines. Would a simple linear/quadratic model also work instead of a neural network? What about a simple heuristic rule to increase/decrease weights? I think it's important to compare to such baselines to understand the complexity of the weight evolution learned by the neural network. - I do not think that default tensorflow example hyperparameters should be used, as mentioned by authors on OpenReview. There is no scientific basis for using them. Instead, first hyperparameters which produce good results in a reasonable time should be selected as the baseline, and then added the benefit of the introspection network to speed up training (and reaching a similar result) should be shown. - The authors state in the discussion on OpenReview that they also tried RNNs as the introspection network but it didn't work with small state size. What does "didn't work" mean in this context? Did it underfit? I find it hard to imagine that a large state size would be required for this task. Even if it is, that doesn't rule out evaluation due to memory issues because the RNN can be run on the weights in 'mini-batch' mode. In general, I think other baselines are more important than RNN. - A question about jump points: The I is trained on SGD trajectories. While using I to speed up training at several jump points, if the input weights cross previous jump points, then I gets input data from a weight evolution which is not from SGD (it has been altered by I). This seems problematic but doesn't seem to affect your experiments. I feel that this again highlights the importance of the baselines. Perhaps I is doing something extremely simple that is not affected by this issue. Since the main idea is very interesting, I will be happy to update my score if the above concerns are addressed.
9: Top 15% of accepted papers, strong accept
9
-1
BkjrLVG4x
HyWDCXjgx
Contribution not clear enough; concerns about data set itself
The manuscript is a bit scattered and hard to follow. There is technical depth but the paper doesn't do a good job explaining what shortcoming the proposed methods are overcoming and what baselines they are outperforming. The writing could be improved. There are numerous grammatical errors. The experiments in 3.1 are interesting, but you need to be clearer about the relationship of your ResCeption method to the state-of-the-art. The use of extensive footnotes on page 5 is a bit odd. "That is a competitive result" is vague. A footnote links to "http://image-net.org/challenges/LSVRC/2015/results" which doesn't seem to even show the same task you are evaluating. ResCeption: "The best validation error is reached at 23.37% and 6.17% at top-1 and top-5, respectively". Single model ResNet-152 gets 19.38 and 4.49, respectively. Resnet-34 is 21.8 and 5.7, respectively. VGGv5 is 24.4 and 7.1, respectively. [source: Deep Residual Learning for Image Recognition, He et al. 2015]. I think it would be more honest for you to report results of competitors and say that your model is worse than ResNet and slightly better than VGG on ImageNet classification. 3.5, retrieval on Holidays, is a bit too much of a diversion from the goal of this paper. If this paper is more about the novel architecture and less about the particular fashion attribute task then the narrative needs to change accordingly. Perhaps my biggest concern is that this paper is missing baselines (e.g. non recurrent models, attribute classification instead of detection) and comparisons to prior work by Berg et al. "Our policy restricts to reveal much more details about the internal dataset" This is a significant issue. The dataset used in this work cannot be shared? How are future works going to compare to your benchmark?
3: Clear rejection
3: The reviewer is fairly confident that the evaluation is correct
3
3
rJJNlzU4g
HyWDCXjgx
interesting exploration but several major concerns
The paper presents a large-scale visual search system for finding product images given a fashion item. The exploration is interesting and the paper does a nice job of discussing the challenges of operating in this domain. The proposed approach addresses several of the challenges. However, there are several concerns. 1) The main concern is that there are no comparisons or even mentions of the work done by Tamara Berg’s group on fashion recognition and fashion attributes, e.g., - “Automatic Attribute Discovery and Characterization from Noisy Web Data” ECCV 2010 - “Where to Buy It: Matching Street Clothing Photos in Online Shops” ICCV 2015, - “Retrieving Similar Styles to Parse Clothing, TPAMI 2014, etc It is difficult to show the contribution and novelty of this work without discussing and comparing with this extensive prior art. 2) There are not enough details about the attribute dataset and the collection process. What is the source of the images? Are these clean product images or real-world images? How is the annotation done? What instructions are the annotators given? What annotations are being collected? I understand data statistics for example may be proprietary, but these kinds of qualitative details are important to understand the contributions of the paper. How can others compare to this work? 3) There are some missing baselines. How do the results in Table 2 compare to simpler methods, e.g., the BM or CM methods described in the text? While the paper presents an interesting exploration, all these concerns would need to be addressed before the paper can be ready for publication.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
B1Mp8grVl
HyWDCXjgx
Good practical visual search system but lack novelty
This paper introduces a pratical large-scale visual search system for a fashion site. It uses RNN to recognize multi-label attributes and uses state-of-art faster RCNN to extract features inside those region-of-interest (ROI). The technical contribution of this paper is not clear. Most of the approaches used are standard state-of-art methods and there are not a lot of novelties in applying those methods. For multi-label recognition task, there are other available methods, e.g. using binary models, changing cross-entropy loss function, etc. There aren't any comparison between the RNN method and other simple baselines. The order of the sequential RNN prediction is not clear either. It seems that the attributes form a tree hierarchy and that is used as the order of sequence. The paper is not well written either. Most results are reported in the internal dataset and the authors won't release the dataset.
3: Clear rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
3
4
r1UHA4XNg
rJEgeXFex
Thorough empirical investigation of an interesting and (to my knowledge) novel application area
This is a well written, organized, and presented paper that I enjoyed reading. I commend the authors on their attention to the narrative and the explanations. While it did not present any new methodology or architecture, it instead addressed an important application of predicting the medications a patient is using, given the record of billing codes. The dataset they use is impressive and useful and, frankly, more interesting than the typical toy datasets in machine learning. That said, the investigation of those results was not as deep as I thought it should have been in an empirical/applications paper. Despite their focus on the application, I was encouraged to see the authors use cutting edge choices (eg Keras, adadelta, etc) in their architecture. A few points of criticism: -The numerical results are in my view too brief. Fig 4 is anecdotal, Fig 5 is essentially a negative result (tSNE is only in some places interpretable), so that leaves Table 1. I recognize there is only one dataset, but this does not offer a vast amount of empirical evidence and analysis that one might expect out of a paper with no major algorithmic/theoretical advances. To be clear I don't think this is disqualifying or deeply concerning; I simply found it a bit underwhelming. - To be constructive, re the results I would recommend removing Fig 5 and replacing that with some more meaningful analysis of performance. I found Fig 5 to be mostly uninformative, other than as a negative result, which I think can be stated in a sentence rather than in a large figure. - There is a bit of jargon used and expertise required that may not be familiar to the typical ICLR reader. I saw that another reviewer suggested perhaps ICLR is not the right venue for this work. While I certainly see the reviewer's point that a medical or healthcare venue may be more suitable, I do want to cast my vote of keeping this paper here... our community benefits from more thoughtful and in depth applications. Instead I think this can be addressed by tightening up those points of jargon and making the results more easy to evaluate by an ICLR reader (that is, as it stands now researchers without medical experience have to take your results after Table 1 on faith, rather than getting to apply their well-trained quantitative eye). Overall, a nice paper.
7: Good paper, accept
7
-1
S11T5vW4e
rJEgeXFex
Good medical application paper for a medical or data science venue
This is a well-conducted and well-written study on the prediction of medication from diagnostic codes. The authors compared GRUs, LSTMs, feed-forward networks and random forests (making a case for why random forests should be used, instead of SVMs) and analysed the predictions and embeddings. The authors also did address the questions of the reviewers. My only negative point is that this work might be more relevant for a data science or medical venue rather than at ICLR.
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
rJEGyBz4g
rJEgeXFex
Strong application work, very important problem
In light of the detailed author responses and further updates to the manuscript, I am raising my score to an 8 and reiterating my support for this paper. I think it will be among the strongest non-traditional applied deep learning work at ICLR and will receive a great deal of interest and attention from attendees. ----- This paper describes modern deep learning approach to the problem of predicting the medications taken by a patient during a period of time based solely upon the sequence of ICD-9 codes assigned to the patient during that same time period. This problem is formulated as a multilabel sequence classification (in contrast to language modeling, which is multiclass classification). They propose to use standard LSTM and GRU architectures with embedding layers to handle the sparse categorical inputs, similar to that described in related work by Choi, et al. In experiments using a cohort of ~610K patient records, they find that RNN models outperform strong baselines including an MLP and a random forest, as well as a common sense baseline. The differences in performance between the recurrent models and the MLP appear to be large enough to be significant, given the size of the test set. Strengths: - Very important problem. As the authors point out, two the value propositions of EHRs -- which have been widely adopted throughout the US due to a combination of legislation and billions of dollars in incentives from the federal government -- included more accurate records and fewer medication mistakes. These two benefits have largely failed to materialize. This seems like a major opportunity for data mining and machine learning. - Paper is well-written with lucid introduction and motivation, thorough discussion of related work, clear description of experiments and metrics, and interesting qualitative analysis of results. - Empirical results are solid with a strong win for RNNs over convincing baselines. This is in contrast to some recent related papers, including Lipton & Kale et al, ICLR 2016, where the gap between the RNN and MLP was relatively small, and Choi et al, MLHC 2016, which omitted many obvious baselines. - Discussion is thorough and thoughtful. The authors are right about the kidney code embedding results: this is a very promising result. Weaknesses: - The authors make several unintuitive decisions related to data preprocessing and experimental design, foremost among them the choice NOT to use full patient sequences but instead only truncated patient sequences that each ends at randomly chosen time point. This does not necessarily invalidate their results, but it is somewhat unnatural and the explanation is difficult to follow, reducing the paper's potential impact. It is also reduces the RNN's potential advantage. - The chosen metrics seem appropriate, but non-experts may have trouble interpreting the absolute and relative performances (beyond the superficial, e.g., RNN score 0.01 more than NN!). The authors should invest some space in explaining (1) what level of performance -- for each metric -- would be necessary for the model to be useful in a real clinical setting and (2) whether the gaps between the various models are "significant" (even in an informal sense). - The paper proposes nothing novel in terms of methods, which is a serious weakness for a methods conference like ICLR. I think it is strong enough empirically (and sufficiently interesting in application) to warrant acceptance regardless, but there may be things the authors can do to make it more competitive. For example, one potential hypothesis is that higher capacity models are more prone to overfitting noisy targets. Is there some way to investigate this, perhaps by looking at the kinds of errors each model makes? I have a final comment: as a piece of clinical work, the paper has a huge weakness: the lack of ground truth labels for missing medications. Models are both trained and tested on data with noisy labels. For training, the authors are right that this shouldn't be a huge problem, provided the label noise is random (even class conditional isn't too big of a problem). For testing, though, this seems like it could skew metrics. Further, the assumption that the label noise is not systemic seems very unlikely given that these data are recorded by human clinicians. The cases shown in Appendix C lend some credence to this assertion: for Case 1, 7/26 actual medications received probabilities < 0.5. My hunch is that clinical reviewers would view the paper with great skepticism. The authors will need to get creative about evaluation -- or invest a lot of time/money in labeling data -- to really prove that this works. For what it is worth, I hope that this paper is accepted as I think it will be of great interest to the ICLR community. However, I am borderline about whether I'd be willing to fight for its acceptance. If the authors can address the reviewers' critiques -- and in particular, dive into the question of overfitting the imperfect labels and provide some insights -- I might be willing to raise my score and lobby for acceptance.
8: Top 50% of accepted papers, clear accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
8
4
BytMf6WVe
rJ8uNptgl
interesting experimental evaluation of variable bit-rate CNN weight compression scheme
This paper proposes a novel neural network compression technique. The goal is to compress maximally the network specification via parameter quantisation with a minimum impact on the expected loss. It assumes pruning of the network parameters has already been performed, and only considers the quantisation of the individual scalar parameters of the network. In contrast to previous work (Han et al. 2015a, Gong et al. 2014) the proposed approach takes into account the effect of the weight quantisation on the loss function that is used to train the network, and also takes into account the effect on a variable-length binary encoding of the cluster centers used for the quantisation. Unfortunately, the submitted paper is 20 pages, rather than the 8 recommended. The length of the paper seems unjustified to me, since the first three sections (first five pages) are very generic and redundant can be largely compressed or skipped (including figures 1 and 2). Although not a strict requirement by the submission guidelines, I would suggest the authors to compress their paper to 8 pages, this will improve the readability of the paper. To take into account the impact on the network’s loss the authors propose to use a second order approximation of the cost function of the loss. In the case of weights that originally constitute a local minimum of the loss, this leads to a formulation of the impact of the weight quantization on the loss in terms of a weighted k-means clustering objective, where the weights are derived from the hessian of the loss function at the original weights. The hessian can be computed efficiently using a back-propagation algorithm similar to that used to compute the gradient, as shown in cited work from the literature. The authors also propose to alternatively use a second-order moment term used by the Adam optimisation algorithm, since it can be loosely interpreted as an approximate Hessian. In section 4.5 the authors argue that with their approach it is more natural to quantise weights across all layers together, due to the hessian weighting which takes into account the variable impact across layers of quantisation errors on the network performance. The last statement in this section, however, was not clear to me: “In such deep neural networks, quantising network parameters of all layers together is more efficient since optimizing layer-by-layer clustering jointly across all layers requires exponential time complexity with respect to the number of layers.” Perhaps the authors could elaborate a bit more on this point? In section 5 the authors develop methods to take into account the code length of the weight quantisation in the clustering process. The first method described by the authors (based on previous work), is uniform quantisation of the weight space, which is then further optimised by their hessian-weighted clustering procedure from section 4. For the case of nonuniform codeword lengths to encode the cluster indices, the authors develop a modification of the Hessian weighted k-means algorithm in which the code length of each cluster is also taken into account, weighted by a factor lambda. Different values of lambda give rise to different compression-accuracy trade-offs, and the authors propose to cluster weights for a variety of lambda values and then pick the most accurate solution obtained, given a certain compression budget. In section 6 the authors report a number of experimental results that were obtained with the proposed methods, and compare these results to those obtained by the layer-wise compression technique of Han et al 2015, and to the uncompressed models. For these experiments the authors used three datasets, MNIST, CIFAR10 and ImageNet, with data-set specific architectures taken from the literature. These results suggest a consistent and significant advantage of the proposed method over the work of Han et al. Comparison to the work of Gong et al 2014 is not made. The results illustrate the advantage of the hessian weighted k-means clustering criterion, and the advantages of the variable bitrate cluster encoding. In conclusion I would say that this is quite interesting work, although the technical novelty seems limited (but I’m not a quantisation expert). Interestingly, the proposed techniques do not seem specific to deep conv nets, but rather generically applicable to quantisation of parameters of any model with an associated cost function for which a locally quadratic approximation can be formulated. It would be useful if the authors would discuss this point in their paper.
7: Good paper, accept
3: The reviewer is fairly confident that the evaluation is correct
7
3
ryJNu_b4x
rJ8uNptgl
Effective quantization
The paper has two main contributions: 1) Shows that uniform quantization works well with variable length (Huffman) coding 2) Improves fixed-length quantization by proposing the Hessian-weighted k-means, as opposed to standardly used vanilla k-means. The Hessian weighting is well motivated, and it is also explained how to use an efficient approximation "for free" when using the Adam optimizer, which is quite neat. As opposed to vanilla k-means, one of the main benefits of this approach (apart from improved performance) is that no tuning on per-layer compression rates is required, as this is achieved for free. To conclude, I like the paper: (1) is not really novel but it doesn't seem other papers have done this before so it's nice to know it works well, and (2) is quite neat and also works well. The paper is easy to follow, results are good. My only complaint is that it's a bit too long. Minor note - I still don't understand the parts about storing "additional bits for each binary codeword for layer indication" when doing layer-by-layer quantization. What's the problem of just having an array of quantized weight values for each layer, i.e. q[0][:] would store all quantized weights for layer 0, q[1][:] for layer 1 etc, and for each layer you would have the codebook. So the only overhead over joint quantization is storing the codebook for each layer, which is insignificant. I don't understand the "additional bit" part. But anyway, this is really not a important as I don't think it affects the paper at all, just authors might want to additionally clarify this point (maybe I'm missing something obvious, but if I am then it's likely some other people will as well).
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
Syu1KV9rg
rJ8uNptgl
Review for "Towards the Limit of Network Quantization"
This paper proposes a network quantization method for compressing the parameters of neural networks, therefore, compressing the amount of storage needed for the parameters. The authors assume that the network is already pruned and aim for compressing the non-pruned parameters. The problem of network compression is a well-motivated problem and of interest to the ICLR community. The main drawback of the paper is its novelty. The paper is heavily built on the results of Han 2015 and only marginally extends Han 2015 to overcome its drawbacks. It should be noted that the proposed method in this paper has not been proposed before. The paper is well-structured and easy to follow. Although it heavily builds on Han 2015, it is still much longer than Han 2015. I believe that there is still some redundancy in the paper. The experiments section starts on Page 12 whereas for Han 2015 the experiments start on page 5. Therefore, I believe much of the introductory text is redundant and can be efficiently cut. Experimental results in the paper show good compression performance compared to Han 2015 while losing very little accuracy. Can the authors mention why there is no comparison with Hang 2015 on ResNet in Table 1? Some comments: 1) It is not clear whether the procedure depicted in figure 1 is the authors’ contribution or has been in the literature. 2) In section 4.1 the authors approximate the hessian matrix with a diagonal matrix. Can the authors please explain how this approximation affects the final compression? Also how much does one lose by making such an approximation? minor typos (These are for the revised version of the paper): 1) Page 2, Parag 3, 3rd line from the end: fined-tuned -> fine-tuned 2) Page 2, one para to the end, last line: assigned for -> assigned to 3) Page 5, line 2, same as above 4) Page 8, Section 5, Line 3: explore -> explored
7: Good paper, accept
3: The reviewer is fairly confident that the evaluation is correct
7
3
rkqq9Mime
BJh6Ztuxl
Interesting analytic results on unsupervised sentence encoders
This paper presents a set of experiments investigating what kinds of information are captured in common unsupervised approaches to sentence representation learning. The results are non-trivial and somewhat surprising. For example, they show that it is possible to reconstruct word order from bag of words representations, and they show that LSTM sentence autoencoders encode interpretable features even for randomly permuted nonsense sentences. Effective unsupervised sentence representation learning is an important and largely unsolved problem in NLP, and this kind of work seems like it should be straightforwardly helpful towards that end. In addition, the experimental paradigm presented here is likely more broadly applicable to a range of representation learning systems. Some of the results seem somewhat strange, but I see no major technical concerns, and think that that they are informative. I recommend acceptance. One minor red flag: - The massive drop in CBOW performance in Figures 1b and 4b are not explained, and seem implausible enough to warrant serious further investigation. Can you be absolutely certain that those results would appear with a different codebase and different random seed implementing the same model? Fortunately, this point is largely orthogonal to the major results of the paper. Two writing comments: - I agree that the results with word order and CBOW are surprising, but I think it's slightly misleading to say that CBOW is predictive of word order. It doesn't represent word order at all, but it's possible to probabilistically reconstruct word order from the information that it does encode. - Saying that "LSTM auto-encoders are more effective at encoding word order than word content" doesn't really make sense. These two quantities aren't comparable.
8: Top 50% of accepted papers, clear accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
8
4
HkHqRoIEe
BJh6Ztuxl
Review
The authors present a methodology for analyzing sentence embedding techniques by checking how much the embeddings preserve information about sentence length, word content, and word order. They examine several popular embedding methods including autoencoding LSTMs, averaged word vectors, and skip-thought vectors. The experiments are thorough and provide interesting insights into the representational power of common sentence embedding strategies, such as the fact that word ordering is surprisingly low-entropy conditioned on word content. Exploring what sort of information is encoded in representation learning methods for NLP is an important and under-researched area. For example, the tide of word-embeddings research was mostly stemmed after a thread of careful experimental results showing most embeddings to be essentially equivalent, culminating in "Improving Distributional Similarity with Lessons Learned from Word Embeddings" by Levy, Goldberg, and Dagan. As representation learning becomes even more important in NLP this sort of research will be even more important. While this paper makes a valuable contribution in setting out and exploring a methodology for evaluating sentence embeddings, the evaluations themselves are quite simple and do not necessarily correlate with real-world desiderata for sentence embeddings (as the authors note in other comments, performance on these tasks is not a normative measure of embedding quality). For example, as the authors note, the ability of the averaged vector to encode sentence length is trivially to be expected given the central limit theorem (or more accurately, concentration inequalities like Hoeffding's inequality). The word-order experiments were interesting. A relevant citation for this sort of conditional ordering procedure is "Generating Text with Recurrent Neural Networks" by Sutskever, Martens, and Hinton, who refer to the conversion of a bag of words into a sentence as "debagging." Although this is just a first step in better understanding of sentence embeddings, it is an important one and I recommend this paper for publication.
8: Top 50% of accepted papers, clear accept
8
-1
H1rEX6WNl
BJh6Ztuxl
Experimental analysis of unsupervised sentence embeddings
This paper analyzes various unsupervised sentence embedding approaches by means of a set of auxiliary prediction tasks. By examining how well classifiers can predict word order, word content, and sentence length, the authors aim to assess how much and what type of information is captured by the different embedding models. The main focus is on a comparison between and encoder-decoder model (ED) and a permutation-invariant model, CBOW. (There is also an analysis of skip-thought vectors, but since it was trained on a different corpus it is hard to compare). There are several interesting and perhaps counter-intuitive results that emerge from this analysis and the authors do a nice job of examining those results and, for the most part, explaining them. However, I found the discussion of the word-order experiment rather unsatisfying. It seems to me that the appropriate question should have been something like, 'How well does model X do compared to the theoretical upper bound which can be deduced from natural language statistics?' This is investigated from one angle in Section 7, but I would have preferred to the effect of natural language statistics discussed up front rather than presented as the explanation to a 'surprising' observation. I had a similar reaction to the word-order experiments. Most of the interesting results, in my opinion, are about the ED model. It is fascinating that the LSTM encoder does not seem to rely on natural-language ordering statistics -- it seems like doing so should be a big win in terms of per-parameter expressivity. I also think that it's strange that word content accuracy begins to drop for high-dimensional embeddings. I suppose this could be investigated by handicapping the decoder. Overall, this is a very nice paper investigating some aspects of the information content stored in various types of sentence embeddings. I recommend acceptance.
8: Top 50% of accepted papers, clear accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
8
4
Hys6iL27x
HyFkG45gl
A nice approach to this problem, but inputs seem too artificial
The paper uses neural networks to answer falling body physics questions by 1. Resolving the parameters of the problem, and 2. Figure out which quantity is in question, compute it using a numerical integrator and return it as an answer. Learning and inference are performed on artificially generated questions using a probabilistic grammar. Overall, the paper is clearly written and seems to be novel in its approach. The main problems I see with this work are: 1. The task is artificial, and it's not clear how hard it is. The authors provide no baseline nor do they compare it to any real world problem. Without some measure of difficulty it's hard to tell if a much simple approach will do better, or if the task even makes sense. 2. The labler LSTM uses only 10 hidden units. This is remarkably small for language modeling problems, and makes one further wonder about the difficulty of the task. The authors provide no reasoning for this choice.
5: Marginally below acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
5
4
Sk0Aqs-Vg
HyFkG45gl
An interesting paper to read but could be made better
This paper build a language-based solver for simple physics problems (a free falling object under constant velocity). Given a natural language query sampled from a fixed grammar, the system uses two LSTM models to extract key components, e.g., physical parameters and the type of questions being asked, which are then sent to a numerical integrator for the answer. The overall performance in the test set is almost perfect (99.8%). Overall I found this paper quite interesting to read (and it is well written). However, it is not clear how hard the problem is and how much this approach could generalize over more realistic (and complicated) situations. The dataset are a bit small and might not cover the query space. It might be better to ask AMT workers to come up with more complicated queries/answers. The physics itself is also quite easy. What happens if we apply the same idea on billiards? In this case, even we have a perfect physics simulator, the question to be asked could be very deep and requires multi-hop reasoning. Finally, given the same problem setting (physics solver), in my opinion, a more interesting direction is to study how DNN can take the place of numerical integrator and gives rough answers to the question (i.e., intuitive physics). It is a bit disappointing to see that DNN is only used to extract the parameters while still a traditional approach is used for core reasoning part. It would be more interesting to see the other way round.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
BJDe0lzNl
HyFkG45gl
Rich data generation procedure but system specific and not well motivated
The authors describe a system for solving physics word problems. The system consists of two neural networks: a labeler and a classifier, followed by a numerical integrator. On the dataset that the authors synthesize, the full system attains near full performance. Outside of the pipeline, the authors also provide some network activation visualizations. The paper is clear, and the data generation procedure/grammar is rich and interesting. However, overall the system is not well motivated. Why did they consider this particular problem domain, and what challenges did they specifically hope to address? Is it the ability to label sequences using LSTM networks, or the ability to classify what is being asked for in the question? This has already been illustrated, for example, by work on POS tagging and by memory networks for the babi tasks. A couple of standard architectural modifications, i.e. bi-directionality and a content-based attention mechanism, were also not considered.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
H1th_uZNg
HJTzHtqee
A solid empirical study
This paper proposes a compare-aggregate framework that performs word-level matching followed by aggregation with convolutional neural networks. It compares six different comparison functions and evaluates them on four datasets. Extensive experimental results have been reported and compared against various published baselines. The paper is well written overall. A few detailed comments: * page 4, line5: including a some -> including some * What's the benefit of the preprocessing and attention step? Can you provide the results without it? * Figure 2 is hard to read, esp. when on printed hard copy. Please enhance the quality.
7: Good paper, accept
7
-1
ryR1LZoQe
HJTzHtqee
Effective model design, great evaluation
The paper presents a general approach to modeling for natural language understanding problems with two distinct textual inputs (such as a question and a source text) that can be aligned in some way. In the approach, soft attention is first used to derive alignments between the tokens of the two texts, then a comparison function uses the resulting alignments (represented as pairs of attention queries and attention results) to derive a representations that are aggregated by CNN into a single vector from which an output can be computed. The paper both presents this as an overall modeling strategy that can be made to work quite well, and offers a detailed empirical analysis of the comparison component of the model. This work is timely. Language understanding problems of this kind are a major open issue in NLP, and are just at the threshold of being addressable with representation learning methods. The work presents a general approach which is straightforward and reasonable, and shows that it can yield good results. The work borders on incremental (relative to their earlier work or that of Parikh et al.), but it contributes in enough substantial ways that I'd strongly recommend acceptance. Detail: - The model, at least as implemented for the problems with longer sequences (everything but SNLI), is not sensitive to word order. It is empirically competitive, but this insensitivity places a strong upper bound on its performance. The paper does make this clear, but it seems salient enough to warrant a brief mention in the introduction or discussion sections. - If I understand correctly, your attention strategy is based more closely on the general/bilinear strategy of Luong et al. '15 than it is on the earlier Bahdanau work. You should probably cite the former (or some other more directly relevant reference for that strategy). - Since the NTN risks overfitting because of its large number of parameters, did you try using a version with input dimension l and a smaller output dimension m (so an l*l*m tensor)? - You should probably note that SubMultNN looks a lot like the strategy for *sentence*-level matching in the Lili Mou paper you cite. - Is there a reason you use the same parameters for preprocessing the question and answer in (1)? These could require different things to be weighted highly.
8: Top 50% of accepted papers, clear accept
8
-1
H1rX01G4e
HJTzHtqee
Official Review
This paper proposed a compare-aggregate model for the NLP tasks that require semantically comparing the text sequences, such as question answering and textual entailment. The basic framework of this model is to apply a convolutional neural network (aggregation) after a element-wise operation (comparison) over the attentive outputs of the LSTMs. The highlighted part is the comparison, where this paper compares several different methods for matching text sequences, and the element-wise subtraction/multiplication operations are demonstrated to achieve generally better performance on four different datasets. While the weak point is that this is an incremental work and a bit lack of innovation. A qualitative evaluation about how subtraction, multiplication and other comparison functions perform on varied kinds of sentences would be more interesting.
6: Marginally above acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
6
4
SyF1qboVg
ry18Ww5ee
interesting extension to successive halving, still looking forward to the parallel asynchronous version
This was an interesting paper. The algorithm seems clear, the problem well-recognized, and the results are both strong and plausible. Approaches to hyperparameter optimization based on SMBO have struggled to make good use of convergence during training, and this paper presents a fresh look at a non-SMBO alternative (at least I thought it did, until one of the other reviewers pointed out how much overlap there is with the previously published successive halving algorithm - too bad!). Still, I'm excited to try it. I'm cautiously optimistic that this simple alternative to SMBO may be the first advance to model search for the skeptical practitioner since the case for random search > grid search (http://www.jmlr.org/papers/v13/bergstra12a.html, which this paper should probably cite in connection with their random search baseline.) I would suggest that the authors remove the (incorrect?) claim that this algorithm is "embarrassingly parallel" as it seems that there are number of synchronization barriers at which state must be shared in order to make the go-no-go decisions on whatever training runs are still in progress. As the authors themselves point out as future work, there are interesting questions around how to adapt this algorithm to make optimal use of a cluster (I'm optimistic that it should carry over, but it's not trivial). For future work, the authors might be interested in Hutter et al's work on Bayesian Optimization With Censored Response Data (https://arxiv.org/abs/1310.1947) for some ideas about how to use the dis-continued runs.
8: Top 50% of accepted papers, clear accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
8
4
HyOACaZEx
ry18Ww5ee
Good extension of successive halving and random search
This paper presents Hyperband, a method for hyperparameter optimization where the model is trained by gradient descent or some other iterative scheme. The paper builds on the successive halving + random search approach of Jamieson and Talwalkar and addresses the tradeoff between training fewer models for a longer amount of time, or many models for a shorter amount of time. Effectively, the idea is to perform multiple rounds of successive halving, starting from the most exploratory setting, and then in each round exponentially decreasing the number of experiments, but granting them exponentially more resources. In contrast to other recent papers on this topic, the approach here does not rely on any specific model of the underlying learning curves and therefore makes fewer assumptions about the nature of the model. The results seem to show that this approach can be highly effective, often providing several factors of speedup over sequential approaches. Overall I think this paper is a good contribution to the hyperparameter optimization literature. It’s relatively simple to implement, and seems to be quite effective for many problems. It seems like a natural extension of the random search methodology to the case of early stopping. To me, it seems like Hyperband would be most useful on problems where a) random search itself is expected to perform well and b) the computational budget is sufficiently constrained so that squeezing out the absolute best performance is not feasible and near-optimal performance is sufficient. I would personally like to see the plots in Figure 3 run out far enough that the other methods have had time to converge in order to see what this gap between optimal and near-optimal really is (if there is one). I’m not sure I agree with the use of random2x as a baseline. I can see why it’s a useful comparison because it demonstrates the benefit of parallelism over sequential methods, but virtually all of these other methods also have parallel extensions. I think if random2x is shown, then I would also like to see SMAC2x, Spearmint2x, TPE2x, etc. I also think it would be worth seeing 3x, 10x, and so forth and how Hyperband fares against these baselines.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
ryVZxPfVe
ry18Ww5ee
A nice paper, just needs to relate to the existing literature better
This paper discusses Hyperband, an extension of successive halving by Jamieson & Talwalkar (AISTATS 2016). Successive halving is a very nice algorithm that starts evaluating many configurations and repeatedly cuts off the current worst half to explore many configuration for a limited budget. Having read the paper for the question period and just rereading it again, I am now not entirely sure what its contribution is meant to be: the only improvement of Hyperband vs. successive halving is in the theoretical worst case bounds (not more than 5x worse than random search), but you can (a) trivially obtain that bound by using a fifth of your time for running random configurations to completion and (b) the theoretical analysis to show this is said to be beyond the scope of the paper. That makes me wonder whether the theoretical results are the contribution of this paper, or whether they are the subject of a different paper and the current paper is mostly an empirical study of the method? I hope to get a response by the authors and see this made clearer in an updated version of the paper. In terms of experiments, the paper fails to show a case where Hyperband actually performs better than the authors' previous algorithm successive halving with its most agressive setting of bracket b=4. Literally, in every figure, bracket b=4 is at least as good (and sometimes substantially better) than Hyperband. That makes me think that in practice I would prefer successive halving with b=4 over Hyperband. (And if I really want Hyperband's guarantee of not being more than 5x worse than random search I can run random search on a fifth of my machines.) The experiments also compare to some Bayesian optimization methods, but not to the most relevant very closely related Multi-Task Bayesian Optimization methods that have been dominating effective methods for deep learning in that area in the last 3 years: "Multi-Task Bayesian Optimization" by Swersky, Snoek, and Adams (2013) already showed 5x speedups for deep learning by starting with smaller datasets, and there have been several follow-up papers showing even larger speedups. Given that this prominent work on multitask Bayesian optimization exists, I also think the introduction, which sells Hyperband as a very new approach to hyperparameter optimization is misleading. I would've much preferred a more down-to-earth pitch that says "configuration evaluation" has been becoming a very important feature in hyperparameter optimization, including Bayesian optimization, that sometimes yields very large speedups (this can be quantified by examples from existing papers) and this paper adds some much-needed theoretical understanding to this and demonstrates how important configuration evaluation is even in the simplest case of being used with random search. I think this could be done easily and locally by adding a paragraph to the intro. As another point regarding novelty, I think the authors should make clear that approaches for adaptively deciding how many resources to use for which evaluation have been studied for (at least) 23 years in the ML community -- see Maron & Moore, NIPS 1993: "Hoeffding Races: Accelerating Model Selection Search for Classification and Function Approximation" (https://papers.nips.cc/paper/841-hoeffding-races-accelerating-model-selection-search-for-classification-and-function-approximation). Again, this could be done by a paragraph in the intro. Overall, I think for this paper having the related work section at the end leads to many concepts appearing to be new in the paper that turn out not to be new in the end, which is a bit of a let-down. I encourage the authors to prominently discuss related work, including the recent trends in Bayesian optimization towards configuration evaluation, in the beginning, and then clearly state the contribution of this paper by positioning it in the context of that related work and saying what exactly is new. (I think the answer is "very simple method", "great empirical results for several deep learning tasks" and "much-needed new theoretical results", which is a very nice contribution.) I'm giving an accepting score trusting that the authors will follow this suggestion. I have some responses to some of the author responses: 1) "In response to your question, we ran an experiment modeled after the empirical studies in Krueger et al tuning 2 hyperparameters of a kernel SVM to compare CVST (Krueger et al 2015) and Hyperband. Hyperband is 3-4x faster than CVST on this experiment and the two achieve similar test performance. Notably, CVST was only 50% faster than standard holdout. For the experiments in our paper, we excluded CVST due to the aforementioned theoretical differences and because CVST is not an anytime algorithm, but as we perform more experiments, we will update the draft to reflect this comparison." Great, I am looking forward to seeing the details on these experiments before the decision phase. 2) "Hyperband makes no assumptions on the shape or rate of convergence of the validation error, just that it eventually converges." It's only the worst-case analysis that makes no assumption, but of course one would not be happy with that worst-case performance of being 5x worse than random search. (The 5x is what the authors call "modestly worse, by a log factor"; it's the logarithm of the dataset size or of the number of epochs, both of which tend to be large numbers). I think this number of 5x should be stated explicitly somewhere for the authors choice of Hyperband parameters. (E.g., at the beginning of the experiments, when Hyperband's parameters are stated.) 3) "Like random search, it is also embarrassingly parallel." I think this is not quite correct. Let's say I want to tune hyperparameters on ImageNet and each hyperparameter evaluation takes 1 week, but I have 100 GPUs, then random search will give a decent solution (the best of 100 random configurations) after 1 week. However, Hyperband will require 5 weeks before it will give any solution. Again, the modest log factor is a factor of 5. To me, "embarassingly parallel" would mean making great predictions after a week if you throw enough resources at it.
7: Good paper, accept
7
-1
SydvmezVe
SkXIrV9le
.
This paper proposes a generative model of videos composed of a background and a set of 2D objects (sprites). Optimization is performed under a VAE framework. The authors' proposal of an outer product of softmaxed vectors (resulting in a 2D map that is delta-like), composed with a convolution, is a very interesting way to achieve translation of an image with differentiable parameters. It seems to be an attractive alternative to more complicated differentiable resamplers (such as those used by STNs) when only translation is needed. Below I have made some comments regarding parts of the text, especially the experiments, that are not clear. The experimental section in particular seems rushed, with some results only alluded to but not given, not even in the appendix. For an extremely novel and exotic proposal, showing only synthetic experiments could be excused. However, though there is some novelty in the method, it is disappointing that there isn't even an attempt at trying to tackle a problem with real data. I suggest as an example aerial videos (such as those taken from drone platforms), since the planar assumption that the authors make would most probably hold in that case. I also suggest that the authors do another pass at proof-reading the paper. There are missing references ("Fig. ??"), unfinished sentences (caption of Fig. 5), and the aforementioned issues with the experimental exposition.
4: Ok but not good enough - rejection
3: The reviewer is fairly confident that the evaluation is correct
4
3
HJUQPFZNl
SkXIrV9le
Mostly incremental generative model of video data with preliminary experimental results
This paper presents a generative model of video sequence data where the frames are assumed to be generated by a static background with a 2d sprite composited onto it at each timestep. The sprite itself is allowed to dynamically change its appearance and location within the image from frame to frame. This paper follows the VAE (Variational Autoencoder) approach, where a recognition/inference network allows them to recover the latent state at each timestep. Some results are presented on simple synthetic data (such as a moving rectangle on a black background or the “Moving MNIST” data. However, the results are preliminary and I suspect that the assumptions used in the paper are far too strong too be useful in real videos. On the Moving MNIST data, the numerical results are not competitive to state of the art numbers. The model itself is also not particularly novel and the work currently misses some relevant citations. The form of the forward model, for example, could be viewed as a variation on the DRAW paper by Gregor et al (ICML 2014). Efficient Inference in Occlusion-Aware Generative Models of Images by Huang & Murphy (ICLR) is another relevant work, which used a variational auto-encoder with a spatial transformer and an RNN-like sequence model to model the appearance of multiple sprites on a background. Finally, the exposition in this paper is short on many details and I don’t believe that the paper is reproducible from the text alone. For example, it is not clear what the form of the recognition model is… Low-level details (which are very important) are also not presented, such as initialization strategy.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
ByfxSqbNx
SkXIrV9le
Experimental results are too preliminary
This paper presents an approach to modeling videos based on a decomposition into a background + 2d sprites with a latent hidden state. The exposition is OK, and I think the approach is sensible, but the main issue with this paper is that it is lacking experiments on non-synthetic datasets. As such, while I find the graphics inspired questions the paper is investigating interesting, I don't think it is clear that this work introduces useful machinery for modeling more general videos. I think this paper is more appropriate as a workshop contribution in its current form.
4: Ok but not good enough - rejection
4: The reviewer is confident but not absolutely certain that the evaluation is correct
4
4
BkjpniLEg
B184E5qee
Review
The authors present a simple method to affix a cache to neural language models, which provides in effect a copying mechanism from recently used words. Unlike much related work in neural networks with copying mechanisms, this mechanism need not be trained with long-term backpropagation, which makes it efficient and scalable to much larger cache sizes. They demonstrate good improvements on language modeling by adding this cache to RNN baselines. The main contribution of this paper is the observation that simply using the hidden states h_i as keys for words x_i, and h_t as the query vector, naturally gives a lookup mechanism that works fine without tuning by backprop. This is a simple observation and might already exist as folk knowledge among some people, but it has nice implications for scalability and the experiments are convincing. The basic idea of repurposing locally-learned representations for large-scale attention where backprop would normally be prohibitively expensive is an interesting one, and could probably be used to improve other types of memory networks. My main criticism of this work is its simplicity and incrementality when compared to previously existing literature. As a simple modification of existing NLP models, but with good empirical success, simplicity and practicality, it is probably more suitable for an NLP-specific conference. However, I think that approaches that distill recent work into a simple, efficient, applicable form should be rewarded and that this tool will be useful to a large enough portion of the ICLR community to recommend its publication.
7: Good paper, accept
7
-1
H1YGZUMNx
B184E5qee
Review
This paper not only shows that a cache model on top of a pre-trained RNN can improve language modeling, but also illustrates a shortcoming of standard RNN models in that they are unable to capture this information themselves. Regardless of whether this is due to the small BPTT window (35 is standard) or an issue with the capability of the RNN itself, this is a useful insight. This technique is an interesting variation of memory augmented neural networks with a number of advantages to many of the standard memory augmented architectures. They illustrate the neural cache model on not just the Penn Treebank but also WikiText-2 and WikiText-103, two datasets specifically tailored to illustrating long term dependencies with a more realistic vocabulary size. I have not seen the ability to refer up to 2000 words back previously. I recommend this paper be accepted. There is additionally extensive analysis of the hyperparameters on these datasets, providing further insight. I recommend this interesting and well analyzed paper be accepted.
9: Top 15% of accepted papers, strong accept
9
-1
SJv48zBNl
B184E5qee
review
This paper proposes a simple extension to a neural network language model by adding a cache component. The model stores <previous hidden state, word> pairs in memory cells and uses the current hidden state to control the lookup. The final probability of a word is a linear interpolation between a standard language model and the cache language model. Additionally, an alternative that uses global normalization instead of linear interpolation is also presented. Experiments on PTB, Wikitext, and LAMBADA datasets show that the cache model improves over standard LSTM language model. There is a lot of similar work on memory-augmented/pointer neural language models, and the main difference is that the proposed method is simple and scales to a large cache size. However, since the technical contribution is rather limited, the experiments need to be more thorough and conclusive. While it is obvious from the results that adding a cache component improves over language models without memory, it is still unclear that this is the best way to do it (instead of, e.g., using pointer networks). A side-by-side comparison of models with pointer networks vs. models with cache with roughly the same number of parameters is needed to convincingly argue that the proposed method is a better alternative (either because it achieves lower perplexity, faster to train but similar test perplexity, faster at test time, etc.) Some questions: - In the experiment results, for your neural cache model, are those results with linear interpolation or global normalization, or the best model? Can you show results for both? - Why is the neural cache model worse than LSTM on Ctrl (Lambada dataset)? Please also show accuracy on this dataset. - It is also interesting that the authors mentioned that training the cache component instead of only using it at test time gives little improvements. Are the results about the same or worse?
5: Marginally below acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
5
4
rkrSDoZ4e
ryHlUtqge
Review
The paper proposes to study the problem of semi-supervised RL where one has to distinguish between labelled MDPs that provide rewards, and unlabelled MDPs that are not associated with any reward signal. The underlying is very simple since it aims at simultaneously learning a policy based on the REINFORCE+entropy regularization technique, and also a model of the reward that will be used (as in inverse reinforcement learning) as a feedback over unlabelled MDPs. The experiments are made on different continous domains and show interesting results The paper is well written, and easy to understand. It is based on a simple but efficient idea of simultaneously learning the policy and a model of the reward and the resulting algorithm exhibit interesting properties. The proposed idea is quite obvious, but the authors are the first ones to propose to test such a model. The experiments could be made stronger by mixing continuous and discrete problems but are convincing.
8: Top 50% of accepted papers, clear accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
8
4
BknFD3GVg
ryHlUtqge
An approach to semi supervised RL using inverse RL
In supervised learning, a significant advance occurred when the framework of semi-supervised learning was adopted, which used the weaker approach of unsupervised learning to infer some property, such as a distance measure or a smoothness regularizer, which could then be used with a small number of labeled examples. The approach rested on the assumption of smoothness on the manifold, typically. This paper attempts to stretch this analogy to reinforcement learning, although the analogy is somewhat incoherent. Labels are not equivalent to reward functions, and positive or negative rewards do not mean the same as positive and negative labels. Still, the paper makes a worthwhile attempt to explore this notion of semi-supervised RL, which is clearly an important area that deserves more attention. The authors use the term "labeled MDP" to mean the typical MDP framework where the reward function is unknown. They use the confusing term "unlabeled MDP" to mean the situation where the reward is unknown, which is technically not an MDP (but a controlled Markov process). In the classical RL transfer learning setup, the agent is attempting to transfer learning from a source "labeled" MDP to a target "labeled" MDP (where both reward functions are known, but the learned policy is known only in the source MDP). In the semi-supervised RL setting, the target is an "unlabeled" CMP, and the source is both a "labeled" MDP and an "unlabeled" CMP. The basic approach is to use inverse RL to infer the unknown "labels" and then attempt to construct transfer. A further restriction is made to linearly solvable MDPs for technical reasons. Experiments are reported using three relatively complex domains using the Mujoco physics simulator. The work is interesting, but in the opinion of this reviewer, the work fails to provide a simple sufficiently general notion of semi-supervised RL that will be of sufficiently wide interest to the RL community. That remains to be done by a future paper, but in the interim, the work here is sufficiently interesting and the problem is certainly a worthwhile one to study.
6: Marginally above acceptance threshold
6
-1
B13U25zEx
ryHlUtqge
Review
This paper formalizes the problem setting of having only a subset of available MDPs for which one has access to a reward. The authors name this setting "semi-supervised reinforcement learning" (SSRL), as a reference to semi-supervised learning (where one only has access to labels for a subset of the dataset). They provide an approach for solving SSRL named semi-supervised skill generalization (S3G), which builds on the framework of maximum entropy control. The whole approach is straightforward and amounts to an EM algorithm with partial labels (: they alternate iteratively between estimating a reward function (parametrized) and fitting a control policy using this reward function. They provide experiments on 4 tasks (obstacle, 2-link reacher, 2-link reacher with vision, half-cheetah) in MuJoCo. The paper is well-written, and is overall clear. The appendix provides some more context, I think a few implementation details are missing to be able to fully reproduce the experiments from the paper, but they will provide the code. The link to inverse reinforcement learning seems to be done correctly. However, there is no reference to off-policy policy learning, and, for instance, it seems to me that the \tau \in D_{samp} term of equation (3) could benefit from variance reduction as in e.g. TB(\lambda) [Precup et al. 2000] or Retrace(\lambda) [Munos et al. 2016]. The experimental section is convincing, but I would appreciate a precision (and small discussion) of this sentence "To extensively test the generalization capabilities of the policies learned with each method, we measure performance on a wide range of settings that is a superset of the unlabeled and labeled MDPs" with numbers for the different scenarios (or the replacement of superset by "union" if this is the case). It may explain better the poor results of "oracle" on "obstacle" and "2-link reacher", and reinforce* the further sentences "In the obstacle task, the true reward function is not sufficiently shaped for learning in the unlabeled MDPs. Hence, the reward regression and oracle methods perform poorly". Correction on page 4: "5-tuple M_i = (S, A, T, R)" is a 4-tuple. Overall, I think that this is a good and sound paper. I am personally unsure as to if all the parallels and/or references to previous work are complete, thus my confidence score of 3. (* pun intended)
7: Good paper, accept
3: The reviewer is fairly confident that the evaluation is correct
7
3
rkgMSRKrx
SkkTMpjex
Review - Distributed K-FAC
In this paper, the authors present a partially asynchronous variant of the K-FAC method. The authors adapt/modify the K-FAC method in order to make it computationally tractable for optimizing deep neural networks. The method distributes the computation of the gradients and the other quantities required by the K-FAC method (2nd order statistics and Fisher Block inversion). The gradients are computed in synchronous manner by the ‘gradient workers’ and the quantities required by the K-FAC method are computed asynchronously by the ‘stats workers’ and ‘additional workers’. The method can be viewed as an augmented distributed Synchronous SGD method with additional computational nodes that update the approximate Fisher matrix and computes its inverse. The authors illustrate the performance of the method on the CIFAR-10 and ImageNet datasets using several models and compare with synchronous SGD. The main contributions of the paper are: 1) Distributed variant of K-FAC that is efficient for optimizing deep neural networks. The authors mitigate the computational bottlenecks of the method (second order statistic computation and Fisher Block inverses) by asynchronous updating. 2) The authors propose a “doubly-factored” Kronecker approximation for layers whose inputs are too large to be handled by the standard Kronecker-factored approximation. They also present (Appendix A) a cheaper Kronecker factored approximation for convolutional layers. 3) Empirically illustrate the performance of the method, and show: - Asynchronous Fisher Block inversions do not adversely affect the performance of the method (CIFAR-10) - K-FAC is faster than Synchronous SGD (with and without BN, and with momentum) (ImageNet) - Doubly-factored K-FAC method does not deteriorate the performance of the method (ImageNet and ResNet) - Favorable scaling properties of K-FAC with mini-batch size Pros: - Paper presents interesting ideas on how to make computationally demanding aspects of K-FAC tractable. - Experiments are well thought out and highlight the key advantages of the method over Synchronous SGD (with and without BN). Cons: - “…it should be possible to scale our implementation to a larger distributed system with hundreds of workers.” The authors mention that this should be possible, but fail to mention the potential issues with respect to communication, load balancing and node (worker) failure. That being said, as a proof-of-concept, the method seems to perform well and this is a good starting point. - Mini-batch size scaling experiments: the authors do not provide validation curves, which may be interesting for such an experiment. Keskar et. al. 2016 (On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima) provide empirical evidence that large-batch methods do not generalize as well as small batch methods. As a result, even if the method has favorable scaling properties (in terms of mini-batch sizes), this may not be effective. The paper is clearly written and easy to read, and the authors do a good job of communicating the motivation and main ideas of the method. There are a few minor typos and grammatical errors. Typos: - “updates that accounts for” — “updates that account for” - “Kronecker product of their inverse” — “Kronecker product of their inverses” - “where P is distribution over” — “where P is the distribution over” - “back-propagated loss derivativesas” — “back-propagated loss derivatives as” - “inverse of the Fisher” — “inverse of the Fisher Information matrix” - “which amounts of several matrix” — “which amounts to several matrix” - “The diagram illustrate the distributed” — “The diagram illustrates the distributed” - “Gradient workers computes” — “Gradient workers compute” - “Stat workers computes” — “Stat workers compute” - “occasionally and uses stale values” — “occasionally and using stale values” - “The factors of rank-1 approximations” — “The factors of the rank-1 approximations” - “be the first singular value and its left and right singular vectors” — “be the first singular value and the left and right singular vectors … , respectively.” - “\Psi is captures” — “\Psi captures” - “multiplying the inverses of the each smaller matrices” — “multiplying the inverses of each of the smaller matrices” - “which is a nested applications of the reshape” — “which is a nested application of the reshape” - “provides a computational feasible alternative” — “provides a computationally feasible alternative” - “according the geometric mean” — “according to the geometric mean” - “analogous to shrink” — “analogous to shrinking” - “applied to existing model-specification code” — “applied to the existing model-specification code” - “: that the alternative parametrization” — “: the alternative parameterization” Minor Issues: - In paragraph 2 (Introduction) the authors mention several methods that approximate the curvature matrix. However, several methods that have been developed are not mentioned. For example: 1) (AdaGrad) Adaptive Subgradient Methods for Online Learning and Stochastic Optimization (http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) 2) Stochastic Quasi-Newton Methods for Nonconvex Stochastic Optimization (https://arxiv.org/abs/1607.01231) 3) adaQN: An Adaptive Quasi-Newton Algorithm for Training RNNs (http://link.springer.com/chapter/10.1007/978-3-319-46128-1_1) 4) A Self-Correcting Variable-Metric Algorithm for Stochastic Optimization (http://jmlr.org/proceedings/papers/v48/curtis16.html) 5) L-SR1: A Second Order Optimization Method for Deep Learning (https://openreview.net/pdf?id=By1snw5gl) - Page 2, equation s = WA, is there a dimension issue in this expression? - x-axis for top plots in Figures 3,4,5,7 (Updates x XXX) appear to be a headings for the lower plots. - “James Martens. Deep Learning via Hessian-Free Optimization” appears twice in References section.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4
H1gDgMH4e
SkkTMpjex
Official Review
The paper proposes an asynchronous distributed K-FAC method for efficient optimization of deep networks. The authors introduce interesting ideas that many computationally demanding parts of the original K-FAC algorithm can be efficiently implemented in distributed fashion. The gradients and the second-order statistics are computed by distributed workers separately and aggregated at the parameter server along with the inversion of the approximate Fisher matrix computed by a separate CPU machine. The experiments are performed in CIFAR-10 and ImageNet classification problems using models such as AlexNet, ResNet, and GoogleReNet. The paper includes many interesting ideas and techniques to derive an asynchronous distributed version from the original K-FAC. And the experiments also show good results on a few interesting cases. However, I think the empirical results are not thorough and convincing enough yet. Particularly, experiments on various and large number of GPU workers (in the same machine, or across multiple workers) are desired. For example, as pointed by the authors in the answer of a comment, Chen et.al. (Revisiting Distributed Synchronous SGD, 2015) used 100 workers to test their distributed deep learning algorithm. Even considering that the authors have a limitation in computing resource under the academic research setting, the maximum number of 4 or 8 GPUs seems too limited as the only test case of demonstrating the efficiency of a distributed learning algorithm.
6: Marginally above acceptance threshold
3: The reviewer is fairly confident that the evaluation is correct
6
3
HkYCUhr4g
Byk-VI9eg
This work brings multiple discriminators into GAN. From the result, multiple discriminators is useful for stabilizing. The main problem of stabilizing seems is from gradient signal from discriminator, the authors motivation is using multiple discriminators to reduce this effect. I think this work indicates the direction is promising, however I think the authors may consider to add more result vs approach which enforce discriminator gradient, such as GAN with DAE (Improving Generative Adversarial Networks with Denoising Feature Matching), to show advantages of multiple discriminators.
6: Marginally above acceptance threshold
4: The reviewer is confident but not absolutely certain that the evaluation is correct
6
4
r1D00RZNl
Byk-VI9eg
Review
In this interesting paper the authors explore the idea of using an ensemble of multiple discriminators in generative adversarial network training. This comes with a number of benefits, mainly being able to use less powerful discriminators which may provide better training signal to the generator early on in training when strong discriminators might overpower the generator. My main comment is about the way the paper is presented. The caption of Figure 1. and Section 3.1 suggests using the best discriminator by taking the maximum over the performance of individual ensemble members. This does not appear to be the best thing to do because we are just bound to get a training signal that is stricter than any of the individual members of the ensemble. Then the rest of the paper explores relaxing the maximum and considers various averaging techniques to obtain a ’soft-discriminator’. To me, this idea is far more appealing, and the results seem to support this, too. Skimming the paper it seems as if the authors mainly advocated always using the strongest discriminator, evidenced by my premature pre-review question earlier. Overall, I think this paper is a valuable contribution, and I think the idea of multiple discriminators is an interesting direction to pursue.
7: Good paper, accept
3: The reviewer is fairly confident that the evaluation is correct
7
3
B1Ob_V4Ne
Byk-VI9eg
Interesting ideas, needs more empirical results.
The paper extends the GAN framework to accommodate multiple discriminators. The authors motivate this from two points of view: (1) Having multiple discriminators tackle the task is equivalent to optimizing the value function using random restarts, which can potentially help optimization given the nonconvexity of the value function. (2) Having multiple discriminators can help overcome the optimization problems arising when a discriminator is too harsh a critic. A generator receiving signal from multiple discriminators is less likely to be receiving poor gradient signal from all discriminators. The paper's main idea looks straightforward to implement in practice and makes for a good addition to the GAN training toolbelt. I am not very convinced by the GAM (and by extension the GMAM) evaluation metric. Without evidence that the GAN game is converging (even approximately), it is hard to make the case that the discriminators tell something meaningful about the generators with respect to the data distribution. In particular, it does not inform on mode coverage or probability mass misallocation. The learning curves (Figure 3) look more convincing to me: they provide good evidence that increasing the number of discriminators has a stabilizing effect on the learning dynamics. However, it seems like this figure along with Figure 4 also show that the unmodified generator objective is more stable even with only one discriminator. In that case, is it even necessary to have more than one discriminator to train the generator using an unmodified objective? Overall, I think the ideas presented in this paper show good potential, but I would like to see an extended analysis in the line of Figures 3 and 4 for more datasets before I think it is ready for publication. UPDATE: The rating has been revised to a 7 following discussion with the authors.
7: Good paper, accept
4: The reviewer is confident but not absolutely certain that the evaluation is correct
7
4