Datasets:

---

spiralworks

/

openreview-structured-forums

like 0

Follow

Spiral Works 3

[ "Kazi Nazmul Haque", "Siddique Latif", "Rajib Rana" ]

[ "Disentangled Representation Learning", "Data Augmentation", "Generative Adversarial Nets", "Unsupervised Learning" ]

{ "note_id": [ "SklqYIhzc4", "H1gXWc7VFN", "SygmXDIftV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555379841658, 1554426363026, 1554306843001 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper26/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper26/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Reject\", \"comment\": \"The reviewers found a number of issues in clarity and were not fully convinced of the experiments\"}", "{\"title\": \"Review: Disentangled Representation Learning with Information Maximizing Autoencoder\", \"review\": \"The introduction lacks some exposition/definition of \\\"disentanglement\\\" in the context of representations. Why are disentangled representations good?\\n\\nPlease remove the nested parentheses within the citations within parentheses (or change the citation style to one that doesn't have nested parentheses).\\n\\nFigure 1 is a difficult to read. It would be easier if each component/input/output were labeled with the name of the component, rather than the variable. It also isn't clear from the figure which variables exist in the \\\"latent variable space\\\" (embedding?).\\n\\nThe notation p(x) almost always denotes a probability (i.e. a real value between 0 and 1) and so the notation x \\\\in p(x) is somewhat confusing, since I assume x is meant to be an event in the probability space whose probability (function/density) is given by p(x).\", \"minor\": \"\\\"Mutual information\\\" need not be capitalized. Same with \\\"convolutional neural network\\\", \\\"batch normalization\\\", etc.\\n\\nThe bibliography items are inconsistently formatted.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Nice contribution, Some clarifications on the main claim and terminology are needed\", \"review\": \"The authors present a framework in which an auto encoder (E, D) is regularized such that its latent representation to share mutual information with a generated latent space representation. This generated latent representation is an output of a generator network (G) that produces its latent variables based on a given random noise variable (z) plus a categorical input (c). The mutual information is maximized using another network (C, classifier network) that tries to map generated latent representations to the categorical input (c) of the generator network. Moreover two discriminator networks (S, D_i) are used to ensure that the generated images come from the same distribution as the data.\\n\\nTo the best of my understanding the presented scheme performs, in an unsupervised manner, a clustering to a pre-defined number of clusters, that are defined by the one hot encoding categorical (c) input of G, while at the same time G disentangles the inter-cluster variability to the random noise variable (z).\\n\\nThe language of the text is clear and the methodology is described in details. I have however a few comments concerning the presentation of the main claim.\\n\\n1. According to the description of the proposed scheme, I find it difficult to understand how a disentangled representation is learned by the encoder (z_e). I understand that it is possible to generate a latent representation using the G network with c and z as inputs, yet this does not prove that the generated representation z_g (~z_e) are also disentangled.\\n\\n2. Unfortunately the experimental setup does also not support the disentanglement claim in the auto-encoder representation space. An experiment on how the generated image looks when interpolating samples within the auto-encoder representation would be more insightful.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Jeffrey Ling", "Nicholas FitzGerald", "Livio Baldini Soares", "David Weiss", "Tom Kwiatkowski" ]

[ "representation learning", "entities", "few-shot", "Wikipedia" ]

{ "note_id": [ "BylJr5hMcE", "S1eSanZ_YN", "rJxY33bSFV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555380790991, 1554681021230, 1554484400692 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper25/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper25/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Small but solid contribution\", \"review\": \"The paper describes a method for embeddings entities, making use of the pre-trained BERT model.\\nEntities are represented either as simple embeddings, by encoding the entity name with BERT, or by encoding the first paragraph of the wikipedia entry with BERT.\\nThe contexts from entity mentions are also encoded with BERT and then optimized to be in the same space as the entity representations.\\nThe model is evaluated on a task of producing more entities for a given category, given some examples. The proposed method performs best with more frequent entities and is outperformed on less frequent entities.\\n\\nThe method is fairly straightforward, with limited novelty, applying the existing BERT model to encode textual representations. \\nHowever, the paper does present a comparison and analysis for different model variations on this task, which provides some useful insight. \\nAlso, the frequency-based analysis of the entities is interesting, showing a clear boundary where the proposed model starts outperforming the baseline.\\n\\nIt is unclear what is the difference between the \\\"small number of exemplars (3-10)\\\" and \\\"one correct candidate entity\\\" and how these are used differently.\\nAlso, for experiments where the candidates do not cover all possible entities, how were the candidates chosen?\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}", "{\"title\": \"Sound approach, but the results are somewhat unconvincing\", \"review\": \"This paper presents an approach to obtaining representations of entities from raw text alone. They compare their method, 'RELIC', with some presumably competitive baselines, i.e., two ways of encoding entities with BERT (Devlin et al. 2018). The authors do a good job of connecting their work to previous work, noting the similarities to entity linking, entity typing, and learning entity representations from knowledge bases.\\n\\nIn terms of the BERT context encoder, I'm not so clear on how these representations are projected into the d-dimensional space. (I assume that d=300, as noted in 3.2?). The authors mention a transformation matrix W, but do not seem to take up how this is learned. Depending on how this is done, this could have a substantial impact on the results reported. What accuracies does BERT get without this dimensionality reduction?\\n\\nThe results are not altogether convincing, as the majority of entities only occur relatively rarely -- more than half have a frequency in the interval [0, 10). I interpret this as showing that the RELIC system is not particularly successful at the few-shot task, as only relatively high frequency entities have better representations with RELIC than with the BERT comparisons. \\n\\nFinally, the conclusions strike me as somewhat odd. What do you mean by that RELIC \\\"[...] allows us to learn highly interesting representations\\\"? An added qualitative analysis of the representations could serve to highlight this further.\", \"minor_comments\": [\"Footnotes should be after punctuation\", \"Is the 'All' row in Table 2 micro or macro averaged?\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Hoang NT", "Jun Jin Choong", "Tsuyoshi Murata" ]

[ "weakly supervised learning", "noisy label data", "graph neural network", "loss correction" ]

{ "note_id": [ "ByxERKnMcV", "BygKOjQKYE", "SJgt2CGwYV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555380683763, 1554754417140, 1554620081080 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper24/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper24/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Paper which tackles an interesting problem, but still needs to be improved.\", \"review\": [\"This work proposes the use of a noise correction loss in the context of graph neural networks to deal with noisy labels. The GNN is implemented following the message passing approach proposed by Xu (2019). The authors compare 3 different noise estimators (namely the conservative, anchors and exact approaches) in the task of graph classification using 9 datasets.\", \"The paper tackles an interesting and relevant problem to the community. The contribution of the proposed loss in real settings is not clear since only experiments with synthetic noise were performed. More importantly, in its current form the paper is not easy to follow and there are missing details and omissions that should be corrected:\", \"Until we read the \\u201cEmpirical results\\u201d section, it is not clear what are the differences between the conservative, anchor and exact methods. The description given in the \\u201cEmpirical results\\u201d section should be part of the \\u201cMethod\\u201d section.\", \"For the conservative approach, it is not clear what is the loss function used to train the first model which estimates C.\", \"In Sec 2.1, what is \\\"m\\\" in 2^m ? Is this the number of possible labels?\", \"What is C^a in Table 2?\", \"Table 3 indicates \\u201cWe calculate the mean and std of accuracy score on test data for 10 runs each configuration\\u201d. However, there is only one value reported in the table which I assume corresponds to the mean value.\", \"Figure 2 indicates \\u201cX-axis presents the test accuracies\\u201d. As far as I understand, test accuracies are indicated in the Y-axis.\"], \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting work but need clear writing\", \"review\": \"Summary: This paper introduces the loss correction for Graph Neural Networks (GNN) to deal with symmetric graph label noise. The paper shows interesting works on GNN, but the writing could be clearer to deliver the proposed idea.\", \"notes\": [\"The paper shows an interesting work on graph datasets. It has the potential for diverse graph-related tasks.\", \"The paper focuses on a graph classification task. It would be better to show performances on a node classification task.\", \"One area which is not clear is the justification of the worst performance of the conservative estimation model. What about the original models? GIN and GraphSAGE also use cross-entropy loss, but they show much better performances than D-CNN-C, A, and E in some datasets. (What is the 'original model'? GIN?)\", \"The writing could be clearer to deliver the proposed idea. Explanations for notations are missing in some parts. Also, the paper needs clearer definitions of each model.\", \"The paper introduces an interesting denoising approach on GNN, and the proposed model shows good performances on datasets. There are some unclear areas in the paper, which should be addressed before final submission.\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Suman Ravuri", "Oriol Vinyals" ]

[ "GANs", "data augmentation", "BigGAN" ]

{ "note_id": [ "HkxRKaJct4", "Hyg8TTYYtV", "BkgnPV1FKV" ], "note_type": [ "official_review", "decision", "official_review" ], "note_created": [ 1554804102045, 1554779582069, 1554736228399 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Paper23/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper23/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Good systematic study of the inadequacy of GAN for data-augmentation\", \"review\": \"The authors train ResNet-50 networks on a mixture of ImageNet data and BigGAN samples and show that replacing ImageNet data with BigGAN samples leads to a decrease in performance.\", \"positives\": [\"This \\\"data-replacement\\\" experiment is very natural to make, therefore I am glad it was done.\", \"The metric for identifying BigGAN failures per class is also an interesting byproduct.\"], \"remarks\": \"a) \\\"per-class FID\\\" is said to be likely \\\"making the per-class estimates unreliable\\\" due to high variance. I would still like to see what the per-class FID gives and how the best and worst-performing classes compare with the one found by this method. Indeed, even if per-class FID could perform worse, it has the great benefit of not necessitating to train new Resnet-50 networks.\\nb) I would be interested to see some samples generated with the best truncation for replacement, and for addition, in order to have a better idea of what kind of \\\"diversity\\\" we are talking of (this could replace one of the two subplots of Fig. 2 since top1/top5 follow the same trends).\\n\\nOverall, I think this paper and its experiments is a nice contribution to the workshop.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Detailed exploration of GAN in a different setting\", \"review\": \"This paper clearly sets out a hypothesis, method to test the hypothesis, and the presents the core (negative) result \\\"BigGANs cannot be currently used for data augmentation and more work is required for it to be\\nused in downstream tasks\\\". This work overall is clear and direct, and I enjoyed reading and reviewing it. I will focus the remainder of the review on suggestions, questions and ideas for perhaps pushing this work in some other directions.\\n\\nThe experiments here are quite detailed, though the nature of the work opens a whole new series of questions that I hope the authors continue exploring. Rather than only using the model for data augmentation, the images / representations for the classifier could also be augmented using discriminator features extracted learned by the learned-and-frozen GAN - this could perhaps still be an improvement even if the direct images / dataset augmentation isn't sufficient to improve classification. These types of \\\"bootstrapping\\\" of the GAN with discriminator features can be seen in papers such as \\\"Denoising Feature Matching\\\", though used in a different setting https://openreview.net/forum?id=S1X7nhsxl .\\n\\nOne big question I have is in regard to the use of GAN (trained on very large datasets, potentially) as a regularization technique for much smaller datasets than ImageNet. This setting might allow the use of much larger models on small data, due to better regularization - though ImageNet has been (slightly) overfit these days, it is still enormous compared to many practical \\\"in-the-wild\\\" datasets and seems a particularly hard test for this data augmentation setup. Out-of-domain generalization is potentially another area to explore with practical application.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Yusuke Iwasawa", "Kei Akuzawa", "Yutaka Matsuo" ]

[ "invariance", "domain generalization" ]

{ "note_id": [ "B1g9v9nGcV", "B1l4dfipYE", "HkeTKMDFYV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555380833818, 1555047020206, 1554768516769 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper22/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper22/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Active enforcing of invariance instead of avoiding unwanted discriminability\", \"review\": \"This paper investigates Attribute Perception Matching as an alternative to Adversarial Feature Learning. While Adversarial Feature Learning attempts to avoid a discriminator being able to discern a certain set of labels that should not be inferrable from a learned representation, the proposed approach attempts to actively generate similarity between the representations learned, across the different unwanted labels. It is also shown that AFL might be unstable.\\n\\nThe example gives some intuition but may not exhibit the same behavior with actual high-dimensional data.\\n\\nNo notation is introduced in the paper, making especially the beginning very hard to follow.\\nOn the other hand the entropy of a uniform distribution is elevated to Theorem status, and proven in the appendix, which is not necessary.\\n\\nThere is a typo in the heading of section 3. Overall the paper is quite hard to understand.\\n\\nThe contribution of APL merits publication, but it seems to this reviewer that it actually misses the topic of this workshop.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"Improving adversarial feature learning by reformulating the conditionally entropy maximisation to a pair-wise distribution matching\", \"review\": \"The authors attempt to improve on Adversarial Feature Learning (AFL) by introducing a distrance between a parametrization of the attribute distribution and the Encoder Distribution. The resulting work appears to be easier to fit and delivers improved results compared to state of the art work.\\nThe work seems relatively incremental, however, the results appear to justify the direction.\\nHowever, the exposition of the work is rather difficult to follow with notation popping without being previously defined, e.g. L_y section 3 last paragraph. I would strongly urge the authors to focus on clarifying the exposition of the work.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}" ] }

[ "Allison Koenecke", "Jordi Feliu-Fabà" ]

[ "Climate Change", "Twitter Data", "Sentiment Analysis", "Automated Labelling", "Cohort Analysis" ]

{ "note_id": [ "Skez8PkYtE", "BklMqjW_F4", "r1xcT5bSYE" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554736970351, 1554680714324, 1554483905929 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper21/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper21/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Good paper, perhaps a better fit for a different workshop\", \"review\": \"The paper investigates sentiment with respect to climate change and how this changes in tweets after specific natural disasters.\\nTweets from hand-picked influencers are collected, such that these tweets can be assumed to have a specific label and are then used for training a climate change sentiment classifier. \\nSeveral well-established models are compared, with the RNN model achieving the best results.\\nThe classifier is then applied on a wider selection of tweets to measure the change in sentiment before and after specific natural disasters.\\nThe results show that the sentiment moves slightly towards the positive after some hurricanes, but is largely not affected by the natural disasters.\\n\\nThe paper is an interesting pilot work and makes some useful contributions, but not necessarily in the target area of the LLD 2019 Workshop. \\nThe workshop focus is on representation learning with limited data, whereas the paper does not address representation learning and the novelty of the limited data part is minimal (selecting specific users whose tweets are assumed to have the same label).\\n\\nThere are definitely interesting contributions here in the area of social sciences and social media research. The analysis of sentiment timed with specific events was interesting to see.\\nThe paper also identifies specific shortcomings of the current work, which is useful as a pilot experiment and can drive future work.\\n\\nIt seems unlikely that out of 500 randomly sampled tweets exactly 50.0% were positive and negative. Was there actually a more imbalanced distribution that was specifically corrected?\\n\\nThe images in Figure 2 are too small to be interpretable.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting twitter analysis of reactions to climate change\", \"review\": \"This paper describes the collection and analysis of tweets expressing denial or acceptance of climate change. Training and development data for supervised training of a binary tweet classifier are collected from influential tweeters with known opinions towards climate change. The trained classifier is then used to classify a larger collection of tweets around natural disasters in the US and its predictions are analyzed, with respect to geolocation and change over time.\", \"pros\": [\"Interesting insights into reaction on twitter towards climate change.\", \"Comparison of a diverse set of classic binary classification algorithms.\", \"Good visualizations and discussion of outcomes, including significance tests.\"], \"cons\": [\"Data: Is the data made publicly available? How were the influencers chosen?\", \"Analysis: It's not clear whether the classifier is reliable enough to draw these conclusions. Examples from the classifications would allow a rough inspection of the quality. If a different classifier would have been chosen, would the conclusions from the analysis still hold?\", \"The trick for dealing with unlimited data is not in the focus of this paper and also not evaluated against any other method.\", \"In summary, I find the analysis of the classified tweets interesting and neat, but my background knowledge on twitter analysis for climate change is too limited to assess whether its novelty. However, possible inaccuracies or biases of the deployed classifier are not discussed and the aspect of learning with limited data is not in the focus of the paper, which makes me question the relevance for this workshop.\"], \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Yasutaka Furusho", "Kazushi Ikeda" ]

[ "Deep learning", "ResNet", "Skip-connection", "Batch-normalization" ]

{ "note_id": [ "B1lKa93zcE", "BkgBC8Z3tV", "B1eJCEuUYV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555380929188, 1554941644613, 1554576583101 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper20/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper20/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Reject\", \"comment\": \"R1 had several issues with the arguments in the paper. The paper is also not a great fit for the workshop topic\"}", "{\"title\": \"Angle preservation analysis in BN and ResNet shortcuts with questionable conclusion\", \"review\": \"Summary:\\n\\nThe authors argue, that BN and ResNet shortcuts encourage angle preservation throughout the layers.\", \"novelty\": \"The analysis of angle preservation seems novel to me.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"A nice paper studying angles of vectors through layers for data separation\", \"review\": \"The authors here present an interesting analysis on how skip-connections in residual networks and batch-normalization affect data separation. Their analysis included observing the transformation of the input vectors through hidden layers of the neural networks, like a standard multilayer perceptron, a resnet and a resnet with batch normalization. They did that by studying the angle and cosine similarity through the layers. This property is critical to decide if the model is able to separate points in different classes.\\n\\nThe paper is well written and easy to read. The settings and configurations are carefully explained and carry the detail needed to understand the analysis. The study of angles and cosine similarities between the layers is very interesting, although its relation to the data separation property (section 3.3) could be written in a more clear way.\", \"pros\": [\"connecting the dynamics of angles with data separation\", \"extensive analysis\"], \"cons\": [\"more settings could be explored\"], \"some_questions_that_the_authors_could_address_as_well\": \"why are the specific resnet models selected by the authors (Yang, Hardt)? what are the effects of using different initialization methods for the internal weights? how much important is the specific initialization chosen by the authors? what happens when the number of training examples increases or decreases?\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Dean Wyatte" ]

[ "weakly", "learning", "prediction entropy", "entropy regularization", "prediction", "effect", "inexact class labels", "data distributions", "specific subclass", "training set" ]

{ "note_id": [ "Hklu2xNFKE", "rygmm14EYV", "ryeFpZd6_V" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554755760113, 1554427674523, 1553985985331 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper19/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper19/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Review: De-biasing Weakly Supervised Learning by Regularizing Prediction Entropy\", \"review\": \"The authors propose using conditional entropy regularization during the training in semi-supervised settings to mitigate the bias of imbalanced data. The idea is elegant and effectively communicated, and the authors demonstrate its effectiveness empirically on MNIST and CIFAR-10/100.\\n\\nI suggest the committee consider this submission for best paper.\\n\\nI'm curious how the weight on the regularization term affects the bias mitigation.\", \"minor_grammatical_errors\": \"\\\"care should be taken *into* account\\\", \\\"a a substantially more diverse ranking\\\".\", \"rating\": \"5: Top 15% of accepted papers, strong accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Promising use of entropy penalty to de-bias noisy datasets, but context could be made clearer\", \"review\": \"Summary: The paper proposes to use the well-known technique of entropy regularization to de-bias the training set. The classes in the training set are assumed to come from biased mixtures of true classes. By penalizing low-entropy prediction of the model, combined with new labels from an oracle or inferred from prediction probabilities, the procedure can reduce the class bias in the noisy dataset. Experiments on MNIST, CIFAR-10 and CIFAR-100 validates the utility of the proposed scheme, showing that the bias decreases after the procedure.\", \"strengths\": \"1. The use of entropy penalty to de-bias noisy dataset is novel to the best of my knowledge.\\n2. Experimental results on MNIST, CIFAR-10, and CIFAR-100 are convincing and validate the claim in the paper.\", \"weaknesses\": \"1. Bias is never defined. I take it to mean class imbalance in the mixture class. I recommending elaborating on the set up and the context where this setting makes sense.\\n2. The experimental procedure isn't clear. What are mixture classes? Why does it make sense to construct mixture classes as such? What kind of real scenario are these experiments simulating?\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Saypraseuth Mounsaveng", "David Vazquez", "Ismail Ben Ayed", "Marco Pedersoli" ]

[ "GAN", "Data Augmentation", "Image Classification" ]

{ "note_id": [ "BkljnWmf5E", "HkePDwl_KN", "SJebvU4Mt4" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555341746889, 1554675551186, 1554298457013 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper18/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper18/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"This paper proposed an interesting data augmentation method using GAN and STN, however the result is not so good.\", \"review\": \"The paper proposed to augment training data by spacial transformer network, which impose a set of simple transformations to the image and is fully differentiable. The method can learn data-specific transformations and outperforms other GAN-based methods in limited-data setting. Compared to Ratner, the method can be trained in an end-to-end manner, but the evaluation result is lower.\\n\\nAs the experiment result is not so good, I'd recommend a weak accept.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"Easy to follow, Interesting approach, Results not so decisive\", \"review\": [\"The authors of this work propose and evaluate a scheme for data augmentation (DA) that is based on rigid and non-rigid transformations of the training samples. The proposed approach is based on a generative adversarial training paradigm that is constrained by two discriminators that regularize the generated(-transformed) images to conform with the class of the input sample and at the same time ensure dissimilarity with input sample. Since the proposed scheme is differentiable it can be trained jointly ans in an end-to-end fashion with a classifier.\", \"# Pros\", \"The text is well written in terms of the language and structure, while the authors adequately describe the proposed scheme.\", \"The contributions have been clearly stated\", \"The results along with the examples in Appendix C look promising\", \"# Cons\", \"One of the claimed contributions (i.e. well suited for low-data regime) is not fully supported by the experimental results (Figure 2, CIFAR10 results)\", \"Some details are missing in the description of the experimental setup. Are the results presented in 3.1, with or without joint training with the classifier?\", \"In Figre 3 (section 3.2), there are no results with the classic baseline-light/strong DA. It would be insightful on how the joint training could affect the performance of the classifier.\", \"[minor] there some grammatical mistakes in the second paragraph of the Introduction\", \"Overall, the text is easy to follow and well written. The combination of the spatial transformer block with a generative model is an interesting approach yet the performance seems to be slightly lower the state of the art. Considering also the fact that such generative schemes are difficult to define and train I wonder if it worth the effort. I would greatly appreciate such a discussion in the text.\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Mikhail Yurochkin", "Sohini Upadhyay", "Djallel Bouneffouf", "Mayank Agarwal", "Yasaman Khazaeni" ]

[ "online learning", "graph convolutional networks", "contextual bandits" ]

{ "note_id": [ "SkghYohf9N", "HyeXCw_jF4", "BkgInDxwtV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381124153, 1554905034903, 1554610094300 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper17/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper17/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Review for paper: Online Semi-Supervised Learning with Bandit Feedback\", \"review\": \"The paper considers a multi-armed bandit problem in which on some iterations, the reward from the environment might be\\ncompletely missing. To the rescue, a modified LinUCB-type algorithm is then proposed. Experimental results seam sound.\\nThe authors show that their proposal outpowers a blind LinUCB algorithm. The t-SNE plots (Figs 1 and 2) also show that\\nthe embeddings learned by the proposed model are more compactly clustered w.r.t the true labels.\\n\\nThe paper is well-written and the movitated problems are well described.\\n\\nThe paper considers a multi-armed bandit problem in which on some iterations, the reward from the environment might be\\ncompletely missing. To the rescue, a modified LinUCB-type algorithm is then proposed. Experimental results seam sound.\\nThe authors show that their proposal outpowers a blind LinUCB algorithm. The t-SNE plots (Figs 1 and 2) also show that\\nthe embeddings learned by the proposed model are more compactly clustered w.r.t the true labels.\\n\\nThe paper is well-written and the movitated problems are well described.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting setting but insufficient positioning in the literature\", \"review\": \"This paper introduces the \\\"Online Partially Rewarded\\\" setting, which appears to be interesting both from a practical and a theoretical point of view. They then propose two approaches based on GCN, which maintain a similarity graph between observations to help predictions. As Rewarded Online GCN only cannot deal with missclassified observations, the authors propose to combine a multi-armed bandit approach with multiple GCN embedding.\\n\\nThe resulting algorithm maintains k different embeddings for the whole dataset, which are recomputed at each iteration (new observation t). This seems to be costly: would it be possible to update only a few last observations? As noted by the authors, this prevents them to apply their algorithm to bigger datasets. A discussion on the performance would still be a plus. \\n\\nThe detail of the algorithm shows that the proposed algorithm does not account for delayed environment response. This seems to be an important limitation, not implied by the OPR setting: at iteration t, the algorithm has to predict y_t and then observes response h_t. If h_t is missing, the corresponding context is set forever, and a late h_t will just be ignored. Recent works in the bandit community propose frameworks which seem to be more general (see e.g. Bandits with Delayed Anonymous Feedback by Pike-Burke et al. or Online Learning under Delayed Feedback by Joulani et al.).\\n\\nExperiments are conducted on real datasets but with a simulated OPR setting, which is a slight limitation. However, only the two proposed methods and a standard algorithm (LINUCB) of the literature are compared. The results are very encouraging, especially, as noted by the authors, when a natural graph structure is present in the data; but the LINUCB baseline is not adapted to the OPR setting and thus the experiments do not provide enough evidence to back the proposed approach. Here again, I would suggest comparisons with more recent works, such as Variational Thompson Sampling for Relational Recurrent Bandits by Lamprier et al.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Giannis Karamanolakis", "Daniel Hsu", "Luis Gravano" ]

[ "aspect extraction", "distillation", "weakly supervised learning" ]

{ "note_id": [ "ryxmfh2f54", "SJx6RN28tN", "Byes1a-St4" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381259343, 1554592981041, 1554484451333 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper16/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper16/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting but not entirely convincing\", \"review\": \"This submission addresses aspect extraction from product reviews by training a neural network (NN) under weak supervision.\\nThis paper seems to me to be an extension of a previous work by Angelidis and Lapata (2018), who use a predefined set of seed words in order to extract aspect of interest from product reviews.\\nHowever, where Angelidis and Lapata fix both the aspect embedding matrix and the word embeddings for training, the authors in this work adopt the knowledge distillation paradigm by combining a bag-of-word model (as a teacher, simple model in which they can encode the domain knowledge), and an embedding-based NN (as a student, more complex but which can generalize better).\\n\\nThe reported results sound convincing, but I would have liked having a bit more discussion with regard to the two variants (BOW and EMB) of the Student model, as well as why, in most cases, dropping the seed words (SWD) does not help and when it does, it does not seem to be significant.\\nThe authors would have strengthened their work by proving the claim that dropping/replacing seed words with the 'UNK' token actually teaches the Student network to associate aspects with non-seed words during training. My assumption is that it could simply be that the model associates aspects with another set of seed words (seed words which are not among the 30 seed words per aspect and domain from MATE, but which occur regularly, along side the UNK token).\\n\\n* Few comments about the paper *\\n- The paper is easy to read;\\n- EMB and CLF abbreviations, although simple, are referred before being defined;\\n- the multi-task learning training objective of the MATE-* models is not introduced (nor why did they authors not considered it);\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Convincing results, although the modelling could be improved\", \"review\": \"This paper presents an approach to classifying online reviews to aspect classes, when no gold labels for aspect labels are available. Their weakly supervised approach involves using a small set of seed words which describe these aspects. Their approach involves first showing that current weakly supervised approaches with neural networks are not suitable for learning the task under the setting used in the paper. To help with this issue, they turn to a variant of the paradigm of knowledge distillation, with the difference that they use a 'teacher' based on a bag-of-words classifier with seed words, and a student which is an embedding-based neural network.\\n\\nIn terms of the methodology used, the motivation for the simplicity of the student is somewhat lacking. While it is, of course, legitimate to leave more complicated formulations to future work, I am not certain of what exactly the authors mean when they find that that representations from the unweighted average of word embeddings is effective. Is the idea here simply that they tried a simple approach, and found the results to be sufficiently good to verify their approach?\\nIt would have been interesting to see some more approaches here, in order to see whether improving the representations could yield better results, or whether perhaps the representations of the student do not matter so much. A relatively fast and simple thing to do, would be to embed the segments using a pre-trained BERT model.\\n\\nNonetheless, the results seem convincing, as their relatively simple distillation approach yields results which seem substantially above previous work.\", \"minor_comments\": [\"The footnotes at ends of sentences should be written after punctuation\"], \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Samuli Laine", "Jaakko Lehtinen", "Timo Aila" ]

[ "denoising", "self-supervised learning" ]

{ "note_id": [ "HkeALwG9KN", "rJepWPJ9t4", "HygkmNAPKV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554814805744, 1554802437350, 1554666519325 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper15/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper15/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"An interesting architecture and training technique for self-supervised denoising\", \"review\": \"The authors propose techniques for training image denoising models without supervision, using only corrupted images and seeing each corrupted image once.\", \"positives\": \"- Interesting blind-spot network architecture with \\\"4-directional\\\" reading head and shared weights;\\n- The Bayesian training is also a nice contribution, and proves to be important in the evaluation;\\n- Concisely written and straight to the point\\n\\nRemarks & questions:\\na) One extra baseline I am curious of: one could train N2N under the same constraint of using each image once using extra synthetic noise. Given a corrupted image, one can add extra corruption to obtain copies of it with different noise levels, and this could be enough supervision to train N2N to reasonable accuracy. It would be an interesting additional baseline, since it would use the same constraint as your work.\\nb) Inference speed: can you compare the inference speed between your network and other baselines e.g. N2N?\\nc) It seems to me that the datasets used in evaluation have standard image artifacts: JPEG, scale, noise... This makes the problematic of designing methods for seeing each image once a bit artificial because N2N is already well suited to the task. I wonder if one could showcase the benefits of the method on datasets with a special type of artifacts, e.g. medical imagery or hyperspectral data.\\n\\nOverall I think the paper is a worthwhile contribution to the workshop.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting method that is well-described; additional experiments could strengthen the paper\", \"review\": \"Paper Summary:\\nThe authors present a method by which a denoising network can be trained using only examples of single noisy images. The technique relies on idea of blind-spot networks, where the center pixel is eliminated from the receptive field, and reasonably strong statistical assumptions about the noise profile. The authors define a network architecture based on a directional CNN applied to four rotations of the same image, with the receptive field of each restricted to a half-plane (just excluding the center row/column), which is then combined via a set of 1-D convolutions. The network is then optimized to learn the parameters of a multivariate Gaussian distribution describing the mean and covariance matrix characterizing each pixel value. Using the assumption of additive Gaussian noise that is IID between the pixels, these parameters are used in a MAP estimation procedure at test time to arrive at the denoised value. Empirical performance on several datasets appears promising.\\n\\nQuality (Pros):\\nThe technical content of the paper represents an interesting approach to the denoising problem, which draws heavy inspiration from the recent Noise2Void technique. Proposed improvements over this previous technique include (1) a more efficiently trainable network architecture (2) utilization of a Bayesian modeling approach and (3) integration of information from the center pixel at test time. The network architecture itself is adequately described in Section 2, and the design choices seem reasonable. Given the length requirements of the paper, it is reasonable that ablations were not performed on network design, but the lack of quantitative training efficiency comparisons to the Noise2Void technique limits this aspect of the contribution. The Bayesian training approach seems a natural fit for this type of denoising problem, where the noise in an image is characterized by a probability distribution that is local to the pixel in question; of course, the assumption the the noise is additive, Gaussian, and IID amongst pixels is strong, but the relatively straightforward MAP estimation procedure these assumptions enable is a strength of the proposed approach. Finally, integration of integration from the center pixel at test time, but not train time, is an important result of this modeling approach.\\n\\nOverall, the paper presents an interesting technique that contains several distinct improvements over previous approaches, and acknowledges that these are applicable only under strong statistical assumptions. The performance experiments appear well-posed to assess how the proposed technique compares to appropriate baselines (Noise2Void, Noise2Noise, Noise2Clean), but comparison to the most similar baseline (Noise2Void) is only presented for one dataset. Though these results appear compelling, the paper could be improved by adding the Noise2Void numbers for other tested datasets.\\n\\nClarity\\nThe paper is clearly written, and at an appropriate level of detail.\\n\\nSignificance\\nIn addition to suggesting the potential for improved image denoising results using only single noisy images for training, this paper presents an interesting application of Bayesian modeling approaches for denoising that may inspire future work that yields similar results without the strong statistical assumptions currently imposed on the noise distribution.\\n\\nLimitations (Cons)\", \"current_limitations_of_this_work_are_as_follows\": \"(1) Lack of experiments quantitatively demonstrating training efficiency relative to Noise2Void\\n(2) Lack of experiments demonstrating comparison to Noise2Void on any dataset but BSD68\\n(3) No description of the various datasets in Table 1 is given; this would be helpful\\n(4) Strong statistical assumptions required on the noise distributions -- to be clear, this is less a limitation of the authors\\u2019 analysis, but rather of the setting they chose to analyze; in particular, the assumption that all images are characterized by additive Gaussian noise with the same standard deviation, and that this standard deviation is known, may be limiting in practice\\n(5) It would be instructive to demonstrate how performance of this method compares to others if statistical assumptions on the noise distribution are violated.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Kei Akuzawa", "Yusuke Iwasawa", "Yutaka Matsuo" ]

[ "Domain Generalization", "Adversarial Training", "Invariant Feature Learning" ]

{ "note_id": [ "BkgyS3nzq4", "S1eX3LR9KV", "SJx4acHDYN" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381303048, 1554863786869, 1554631356020 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper14/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper14/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"A good paper proposing an objective function for domain invariant representation learning under accuracy constraints\", \"review\": \"The paper provides an analysis on the tradeoff between accuracy and domain invariant representation learning, and accordingly proposes an objective function for domain invariant representation learning under accuracy constraints. The paper is well-written and novel.\", \"pros\": [\"The methods are well-grounded mathematically.\", \"Well-selected baselines and ablation studies. It is good to see that the authors included an ablation study on the term D_KL[p(d|y)|qD(d|h)], and verified the effectiveness of the domain-class dependence assumption.\"], \"cons\": [\"The datasets are relatively small.\"], \"others\": \"Please check the Equation 1 in proof 1. I think the final term should have been H(d|y) + H(y|h). It seems like a typo.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Well written and straight to the point paper\", \"review\": \"In this paper, the authors address the problem of potential dependency between classes and domains in domain generalization. Based on remarks from Xie et al. (2017), they propose a new objective function AFLAC, which (provably) enforces accuracy-constrained domain invariance at its optimum.\\n\\nThe positioning towards related works is clear, the technical analysis is brief but precise. The experiments are well conducted and show both an improvement over the state of the art (including recent approaches) and robustness toward the choice of the hyperparameter.\\n\\nA longer version of this work could include a more detailed scenario of the practical application of this method, as well as experiments on more real datasets (unlike BMNISTR) illustrating the precise advantage of AFLAC on this data.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Francesco Locatello", "Michael Tschannen", "Stefan Bauer", "Gunnar R¨¨ätsch", "Bernhard Schölkopf", "Olivier Bachem" ]

[ "labels", "disentangled representations", "factors", "variations", "access", "model selection", "training", "promising research direction", "representation learning", "locatello" ]

{ "note_id": [ "r1lcX2nz94", "ryxgDPLFF4", "H1gboOnDtE" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381282027, 1554765655778, 1554659480927 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper13/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper13/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Semisupervised learning by introducing observed labels in unsupervised learning methods yields non trivial results for few labels\", \"review\": \"The papers propose a model in which some of the latent variables of an unsupervised model are associated with a small number of user defined labels. This increases the discriminativity of the model and enable it to separate factors of variation associated with those labels.\\nThe introduction of the new term in the objective essentially places the model withing the domain of supervised models (or at least semisupervised). However, all the numerical test appear to be against unsupervised models. Improvement by those standards seems trivial to me. I would urge the authors to include comparisons against other semisupervised methods that work on similar numbers of labeled data\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting observation and proof-of-concept\", \"review\": \"The authors of this work consider the problem of learning disentangled representations. They observe that if labeled data are used for model selection, than using them for training can lead to interesting results. They propose to regularize beta-VAE with a cross entropy loss for a small portion of labeled data, and show improvement on the models that use labeled data only for model selection. The experimental results indeed validate the authors hypothesis and show that even a small set of partially of noisily labeled data can have an interesting impact on the training.\\n\\nThe observation and initial results are interesting, and the work is wort to be discussed in this workshop.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Ian Gemp", "Ramesh Nallapati", "Ran Ding", "Feng Nan", "Bing Xiang" ]

[ "topic models", "variational inference", "variational autoencoder", "semi-supervised", "deep learning" ]

{ "note_id": [ "rJlJ8vnzc4", "HkxY9HPIKN", "Bkem3KVBYE" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555380038986, 1554572689453, 1554495914813 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper12/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper12/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"unclear\", \"review\": \"This short paper investigates weak supervision for neural-based topic modelling (NTM) by modifying both the priors and the posteriors, in a VAE setting.\\n\\nThe priors are modified to using the logits of the probability for each topic to appear in a document. The use of the logits is meant to strengthen \\\"ground truth\\\" information given by a user (supervision), while not penalizing the probability for another topics to appear in a document which would not have been labelled by the user (weak supervision).\\nOn the one hand, his part of the work is quite clear, and the authors justify this as a way to make a NMT better align with the semantic of the user.\\nOn the other hand, the explanation about the generative process is unclear, and would benefit of a longer explanation.\\nI understood that, in addition to the normal setting of a VAE, two other posteriors have been tested: the Concrete approximation to Bernouilli, and the use of a logit-normal distribution.\\nThe all section about the comparison of the two is really confusing and it becomes hard to clearly follow authors' reasoning.\\n\\nIn their experiments, the authors have used three sources of multi-labelled dataset. Topic modelling is not my domain of expertise, but could the authors have not used other datasets to evaluate their approach against (e.g. 20NewsGroups and RCV1-v2 datasets, similarly to original NVDM (Miao et al., 2016))? What's the motivation for using BibTeX, Deliicious and AAPD?\\nBased on table 3, both the NVDM-o (with logit-normal posterior), and Bernouilli (columns 2 and 3) are quite similar for at least two of the metrics, but this point is not discussed by the authors, at all.\\n\\n* other comments/questions about the paper *\\n- latent Dirichlet allocation (LDA) => Latent [...];\\n- Often times => Oftentimes;\\n- The integration of table 1 could be better;\\n- Why is Wikipedia mentioned?\\n- what's the difference between P-NPMI and NPMI? It's not clear;\\n- due the => due to the;\\n- W_{K} => W_{k} (Bernouilli decoding computing the word distribution), compared to \\\"W_{k} refers to\\\";\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"A reasonable approach to a practical problem\", \"review\": \"Title:\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\", \"summary\": \"The authors present a method of weakly supervising a topic model to guide it toward topic alignments that better align with user intuition for what topics should be. Their approach allows the user to label a subset of documents with a subset of their topics, making it flexible.\\n\\nI am not in a great position to confirm the soundness of the math, but the problem being solved and the suggested approach seem reasonable and relevant, with practical application.\", \"points\": [\"I appreciate the flexibility of the design; allowing any number of documents to be labeled with any number of topics makes it much more likely to be used in practice, I believe.\", \"The design decision to incorporate this information in the prior is nice and largely decoupled from the VAE in a good way.\", \"I found the explanation of how ground truth was determined to be a little opaque. If this method in used in other works, please cite them. If it is not, more justification should be given here (possibly with a figure or equation for clarification).\", \"Figure 1 is very helpful in justifying why you chose the model you did.\", \"Table 2 does not seem to add much. It is not obvious from looking at the table that any of those methods is better than the others.\", \"What are the confidence intervals on the results in Table 3? Some of this differences are quite small.\"], \"nits\": [\"Separate Table 1 from the text more\", \"You mention in Section 5 that Delicious has 20 topics, but then say that \\\"half\\\" the number of labels is 11, and fully-labeled is up to 28? Please clarify what is being referred to or fix those numbers.\"]}" ] }

[ "Or Litany", "Daniel Freedman" ]

[ "transfer learning" ]

{ "note_id": [ "BJe4ClEFYE", "SkgLM7m_FN", "SyxqSwRvt4" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554755787977, 1554686734301, 1554667329808 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper10/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper10/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Intuitive idea with solid execution. Promising experimental results\", \"review\": \"This paper presents a unified framework for transfer learning and learning with (large amounts of) noisily-labeled (source) data. The authors assume the source and target classifiers share the same \\u201crepresentation layers\\u201d and only differ by the \\u201cclassification layers\\u201d. They then jointly learn these layers as well as the source example weights. Specifically, the authors formulate the problem as a bilevel optimization problem and design the learning process such that the exterior level minimizes the target loss only through the source weights. As a result, the target classifier can only control/adjust the features representation by adjusting the source example weights. Finally, the authors prove that under certain conditions this bilevel optimization procedure will converge and empirically show it helps learning.\\n\\nOverall the paper is in good shape. I would like to see more (empirical) analysis on the convergence condition and understand how likely the condition at the end of page 9 will be satisfied. Also, some hyper-parameter sensitively analysis on lambda_{a} and lambda_{p} will be interesting. Finally, there are some minor mistakes (listed below) that need to be fixed:\\n1. in the equation (1), change min to argmin\\n2. in page 8 appendix A line 5, \\u201cA summer of SOSELETO algorithm appears in (Algorithm) 1\\u201d\\n3. in page 10 Appendix E line 2, Section \\u201c???\\u201d\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"Well motivated work with interesting idea\", \"review\": [\"Pros:\", \"clear and concise writing, clear motivation for the transfer learning part\", \"sufficient detail presented for each experiment\", \"interesting idea to relate an approach for training with noisy labels to one for transfer learning, and show improvement on previous noisy label results\"], \"cons\": [\"Figure 1 text is not readable (axis labels, legends, titles)\"], \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Yifan Wu", "Ezra Winston", "Divyansh Kaushik", "Zachary Lipton" ]

[ "domain adaptation", "distribution alignment", "adversarial training", "theory" ]

{ "note_id": [ "SyxGwnhfcE", "rJl0jWWnYV", "rJgdks9vFV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381337614, 1554940325647, 1554651872225 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper9/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper9/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting method\", \"review\": \"The paper proposes an asymmetrically-relaxed distribution alignment approach, to do unsupervised domain adaptation. For this, they propose 3 different \\\"relaxed\\\" distances.\", \"pros\": [\"The paper is well written and, although dense, quite clear.\", \"The proposed models are a good alternative to the original DANN\", \"The long version of the paper could be submitted to a journal.\"], \"cons\": [\"The paper is very dense.\", \"Experiments section is very short and explanation of results is minimal. We do not know what are the different acronyms because they are not defined. They are only defined in the long version of the paper, that was attached.\", \"Conclusions are lacking.\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Attached paper is good but short version needs rework\", \"review\": \"This paper identifies a limit of the theoretical framework of Ben-David et al. (2010), regarding an upper bound of the target error in the domain adaptation setting, which is the sum of 3 terms. Authors show that Domain adversarial approach from Ganin et al. (2016) focuses on minimizing 2 of these 3 terms but leaving out the third term leads to a lower bound on the target error, which increases when the difference between source and target distributions augment. The paper then suggests to used relaxed metrics to prevent this effect : giving several examples of such metrics, the authors propose new theoretical bounds on the target error.\\n\\nThe attached paper (12 pages, excluding proofs) is very clear and interesting. The framework and the theorems are stated clearly and, despite the technicality of the theorems, the key ideas seem easy to follow. The short (workshop) paper, however, is not as pleasant to read: the authors clearly lacked space, which especially shows in the last section, where no comment or explanations of the experiment are provided. Depending on the LLD organizers, this may be a problem. I suggest that the authors drop a few more results and propositions, for example concentrating on the \\u03b2-f-divergences and the experiments on MNIST-USPS, which seem to be convincing enough.\", \"several_points_could_also_be_detailed\": \"1. Why are WDANN variants absent from the final experiments?\\n2. The new upper bound on target error involves several terms, provided by assumptions on the source and target domains; is there a practical advantage of this formulation? How do these smaller bounds vary when the source and target distributions shift? This seems to be a central point in advocating the relevance of the proposed approach, which could benefit the longer paper.\\n3. The connected assumption 3.2 is intriguing. Does it appear in related works? Did you find datasets from which it is absent, or which provide a bad \\u03b4\\u2083 bound?\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Baihan Lin" ]

[ "MDL", "Universal code", "LLD", "Normalization", "Biological plausibility", "Unsupervised attention", "Imbalanced data" ]

{ "note_id": [ "SkxLzVYy5E", "rJgrqWNnYN", "H1gUBpm8KV" ], "note_type": [ "official_review", "decision", "official_review" ], "note_created": [ 1555170317986, 1554952589225, 1554558270040 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Paper8/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper8/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Review of \\\"Regularity Normalization: Constraining Implicit Space with Minimum Description Length\\\"\", \"review\": \"The authors present a novel normalization procedure based on weighted maximum likelihood. They extend the idea to make it works for feedforward neural networks.\\n\\nGlobally, the paper is clear and well written, and the contribution interesting.\\n\\nThe algorithm is well-motivated, and the related work is explained clearly. I like the flexibility of the algorithm, highlighted in section 3.3. However, the step \\\"increment\\\" in Algorithm 1 could be explained more clearly. I think it is worth to include a quick discussion on the complexity of one iteration, and compare it to other regularisation procedure.\\n\\nThe numerical experiments seem promising, and the combination of the regularisation procedure with layernorm gives impressive results when classes are imbalanced.\\n\\nIn conclusion, I recommend acceptance of this paper.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"Acceptance Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Good introduction but unclear theoretical framework and inconclusive experimental evidence\", \"review\": \"This paper provides an adaptation of the minimum description length (MDL) principle in computational neuroscience to propose a new attention mechanism for deep neural networks (DNN), whether in a supervised, unsupervised, or reinforcement learning setting. The authors name this attention mechanism \\\"regularity normalization\\\".\\n\\nAs of now, the manuscript suffers from many imprecisions and inaccuracies, thus hindering its eloquence.\\n\\nFirst of all, the authors equate MDL with negative log-likelihood (NLL), but nowhere in the text do they make this connection explicit. Yet, there is a considerable body of literature in minimizing negative log-likelihood for unsupervised learning applications. Maybe i am missing an important distinction between MDL and NLL, but even then, the authors should clarify that distinction.\\n\\nThe definition of the universal code \\\\bar{P}(x) is unclear, and there is no justification of the existence of such universal code.\", \"the_following_sentence_is_quite_mysterious\": \"\\\"the compressibility of the model will be unattainable for multiple inputs, as the probability distributions are different\\\". It is impossible to know what these \\\"multiple inputs\\\" and \\\"different probability distributions\\\" are.\\n\\nMany of the claims in the paper lack a proper definition of their terms. An example is: \\\"The normalized maximum likelihood minimizes the worst case regret with the minimax optimal solution\\\". This sentence is very opaque given that the authors haven't explained how they compute regret, what \\\"worst-case\\\" means, and thus what \\\"minimax\\\" corresponds to in their specific setting. The authors should not leave to reader's guesswork the understanding of these technical terms.\\n\\nThe introduction of the paper is well written. The discussion of prior literature, especially in the field of neuroscience, is solid. However, the presentation of the new contribution is too vague to redeem the lack of convincing experimental evidence for its success.\\n\\nFor example, Algorithm 1 (Regularity Normalization) relies heavily on the computation of complexity (COMP in Equation 2), but this equation is not self-contained. On the left-hand side, there is a model class \\\\Theta. On the right-hand side, there are samples x. How does x relate to \\\\Theta? Is the equation integrated over the probability distribution X of the data? None of these questions are properly answered in the text. Therefore, the algorithm is not reproducible. This problem is not limited to Equation (2). Equation (3) is also very difficult to understand, with a dummy variable x_j in the denominator summing up to \\\"0 ... i\\\".\\n\\nAlthough this paper exhibits a relatively original and interesting idea for training deep neural networks, i do not recommend it for publication, because it is still in an immature stage. I would recommend the authors to carry out further research, both theoretical and experimental, on regularity normalization, and submit an updated version of this paper in the near future.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}" ] }

[ "Mostafa Dehghani", "Arash Mehrjou", "Stephan Gouws", "Jaap Kamps", "Bernhard Schölkopf" ]

[ "fidelity-weighted learning", "semisupervised learning", "weakly-labeled data", "teacher-student" ]

{ "note_id": [ "Skx5PD2GqN", "SygPO9VuFN", "HyxSb8QOF4" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555380066515, 1554692719232, 1554687484942 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper7/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper7/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Review of \\\"Learning from Samples of Variable Quality\\\"\", \"review\": \"Summary of the paper:\\nThis paper proposes a technique to tackle the case where the labels of a supervised problem come from different sources and have varying quality. To this end, a three steps approach called FWL (Fidelity Weighted Learning) is proposed. In particular, FWL consists in constructing a \\u201cstudent\\u201d and a \\u201cteacher\\u201d interacting to refine the quality of the prediction.\\n\\nSome numerical experiments supporting the approach are proposed. In particular, the FWL algorithm is tested on the problem of document ranking. \\nAn investigation of the sensitivity of the approach to the quality of the weak annotator is also proposed.\", \"a_few_comments_and_questions\": \"-the fact that the algorithm uses a Neural Net should be emphasized more in the introduction\\n-Can the proposed approach be extended to other learning algorithm (i.e. not only a NN)?\\n-Why specifically use a Gaussian process? Can another approach be used?\\n-concerning the investigation on the \\u201csensitivity of the FWL to the quality of the weak annotator\\u201d (sic),\\ndoes a \\u201cphase transition\\u201d occur? i.e. is there a threshold on the quality of the weak annotator below which FWL does not yield any improvement?\\n\\nReviewer\\u2019s assessment:\\nI found the paper to be fairly well written. The ideas are exposed clearly and the numerical results support the approach. Since the topic of this work clearly falls within the scope of the workshop, I recommend to accept this paper.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"interesting setup, but methods not well motivated and explained\", \"review\": [\"The authors propose a distillation approach to learn from weakly annotated data aside from a strongly annotated data. The general setup and approach of distillation seems interesting, but it is unclear what the resulted model will learn to do. Specifically, I have the following concerns/questions:\", \"the proposed method is kind of hacky, in the sense that it is not well motivated and grounded on any specific intuition, theoretical or not.\", \"it is not clear to me what the student will converge to after training in the step 3. will it just learn to predict whatever the teacher outputs? if that's the case why do we distill it, aside from the obvious fact that now it is parametric rather than a GP?\", \"following the previous point, have you compared the performance of the student with the performance of the teacher?\"], \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Vivek Bakaraju", "Kishore Reddy Konda" ]

[ "Sparsity", "Unsupervised Learning", "Single Layer Models" ]

{ "note_id": [ "SygMao3GcE", "HJgoCcOJ5E", "Skl3LnM_KE" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381177940, 1555167954896, 1554685011814 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper6/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper6/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Review of \\\"ONLY SPARSITY BASED LOSS FUNCTION FOR LEARNING REPRESENTATIONS\\\"\", \"review\": \"The paper discusses a regularisation function that encourages sparsity. Then, the authors use this criterion as a loss to train neural networks in the unsupervised setting. Numerical experiments comparing the approach with other regularization functions are provided.\\n\\nFirst of all, it seems there are strong formatting issues with the paper. The conclusion is on the 5th page, and the margins are significantly larger than other submitted paper.\\n\\nGlobally, all the choices of the paper are poorly motivated, as well as the assumptions. It is hard for the reader to be convinced by the (many) claims in this paper. On top of it, the experiments are not convincing at all, as the dropout seems to be as efficient as the proposed approach (in terms of the tradeoff between accuracy and sparsity).\\n\\nI have many concerns about the author's motivations for the chosen loss. As far as I understand, they penalize the matrix H using a quadratic function (sum of elements in HH^T). However, this seems to me not straightforward why it should encourage sparsity, since\\n1) the diagonal elements are taken into account in the regularisation and\\n2) the function is quadratic.\\n\\nFinally, the writing of the paper seems to have been done in haste, as the explanations are not clear at all, and the math formulas are incomplete.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"The paper proposes an interesting loss ruction that relies on independence of datapoints, unfortunately, it also depends on the data to work.\", \"review\": \"The authors propose a loss function that encourages different samples to be mapped to different feature vectors. The loss is analyzed both in the context of an autoencoder and a single layer supervised neural network.\\nOne issue the authors would consider is introducing more clarity in the exposition of the work. For instance, you never really specify the exact inputs and outputs of your autoencoder. \\nMy main criticism would be that OVR overly relies on the quality of the neighboring samples in a minibatch to identify the loss. if you have a mini-batch of datapoints of the same class it would not work at all. However, even in the case where you have a minibatch of 128 and an expected 12 data points of each class for CIFAR-10 I would expect a lot of \\\"noisy information\\\" steering the loss towards the wrong result. That problem might alleviate in the case of many labels and highly diverse datasets.\\nI think in total the work leaves a lot to be desired. 1. The loss can be broken by the right construction of the dataset. 2. The loss might do better with a rich dataset which is probably not the case for most applications discussed in the LLD workshop.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Mihir Kale", "Aditya Siddhant", "Sreyashi Nag", "Radhika Parik", "Anthony Tomasic", "Matthias Grabmair" ]

[ "transfer learning", "contextual embeddings", "meta embeddings" ]

{ "note_id": [ "r1xMK3nf9E", "Skgkz9GcYN", "S1luM8aHtN" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555381369693, 1554815495461, 1554531855944 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper5/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/LLD/Paper5/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Missing fine-tuning experiments\", \"review\": \"The authors propose a method for combining encoded representations from pretrained supervised models for transfer learning. The authors argue that in low-resource settings, the combined representations can transfer useful supervision signals that unsupervised representations cannot. The authors evaluate their work on an SRL task (cross-task) and NER (cross-domain and cross-lingual).\\n\\nThe authors present a complete piece of work, from a summary of transfer learning techniques, to a description of their convex combination technique, to their experiments. The paper follows a logical structure, and is generally well written.\\n\\nThe authors make a strong argument for the practicality of their technique. They give a thorough description of the off-the-shelf models and architectures they use, as well as a relatively static hyperparameter set. Assuming the relevant pretrained models are available, the projection layers and combination weights are easy to implement. The model indeed differs from ELMo in that the task-specific models are not jointly learned and thus need to be projected into the same space.\\n\\nHowever, I'm concerned some key evaluations are missing. The most critical missing comparison is to ELMo/ULMFiT/BERT fine-tuned using the available labels for the transfer task (especially for cross-task and cross-domain). The most directly comparable one here is ELMo. The authors learn the exact same weights as ELMo fine-tuning would for their task: the decoder, the per-task weight, and the per-representation weight. It's likely that fine-tuning these alone would yield the performance gains obtained from concatenation with the proposed model. In general, a more complete ablation is needed besides [GloVe/ELMo] vs. [GloVe/ELMo] + proposed, e.g. proposed alone, GloVe + ELMo. In addition, the efficacy of the projection layers is not clear. The folk knowledge that a single linear can project unsupervised word vectors between languages likely does not apply here; a more complex projection might be needed to reconcile task-specific models. The authors give little justification for their projection approach and do not validate it experimentally.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Very good paper. Clear and well motivated approach. Detailed experiments and discussion.\", \"review\": \"Summary\\nThe paper proposes a technique to transfer knowledge from trained models to task/domains/languages where such models cannot be trained due to the lack of training data. This is done in contrast to using only pretrained embeddings such as GloVe or ELMO which are trained in an unsupervised manner, and independently from any downstream task. \\nThe proposed approach assumes the existence of pre-trained models either for different tasks to the targeted one, or the same task but on different domains or languages. All these models are trained in a supervised manner where the last layer in the model is an output layer (before applying a Softmax). The authors project the feature representations that are generated from the pre-trained models using the Convex Combination method to unify and combine the features from all the models. Finally, they concatenate the resulting representation with either GloVe or ELMO. \\nIn their experiments on different tasks, domains, and languages they showed that the proposed approach achieves an increase in the F1-Score ranging between 5%-7% compared to only using the pretrained embeddings. \\n\\nPros.\\n1.\\tThe experimental setting fits the workshop. Specifically, training a model on limited data where the results are on par with models trained on all the available data. \\n2.\\tThe proposed approach is sound. Mainly, where it does not require retraining models that were used for other tasks/domains/languages. \\n3.\\tThe experiments are designed to cover the different knowledge transfer settings: different tasks, the same task but different domains, and the same task but different languages. \\n4.\\tOverall, the paper is clear and well-written, and the authors provided a good discussion on the results. Furthermore, the results are well-presented and easy to interpret. \\n5.\\tFrom a reproducibility perspective, the authors provided all the experimental details in their paper. \\n\\nCons.\\n1.\\tIn the Related Work section, it is not clear how the work under Supervised Transfer Learning is different from this work. \\n2.\\tIt is not clear how the GloVe representation was combined from the word level to the document/sentence level. If it was averaged for all the words, please mention that in the Experiments Setup section. \\n\\nFurther comments.\\nIn addition to the proposed future work provided by the authors, I suggest the following:\\n1.\\tIn addition to only using pretrained embedding as a baseline, compare the performance of your model to the performance of SOTA of each task but with limited data. \\n2.\\tRun the experiments on multiple subsets of the data (1k each) and report the average performance with SD. \\n3.\\tFor the cross-task setting: would adding more (models trained on different) tasks always have a positive effect? Would having a large number of tasks eventually lead to a generic representation which is similar to simply using ELMO again? \\n4.\\tMinor formatting-related issues\\n\\u2022\\tCitation. Please note the difference between having the author name inside or outside the brackets: XX (2018) discussed \\u2026 / the model is provided in (XX 2019 \\n\\u2022\\tIn Sec.3 Approach, the acronym SRL was used without providing the full name (Semantic Role Labelling is mentioned in the first line of the introduction).\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Adrià Ruiz", "Oriol Martinez", "Xavier Binefa", "Jakob Verbeek" ]

[ "Disentangled representations", "Weakly-Supervised Learning", "Variational Autoencoders" ]

{ "note_id": [ "rklhyb4Ft4", "SJlO7o_dYE", "BkgNNh1dFE" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554755812390, 1554709280277, 1554672684261 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper4/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper4/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Solid paper, with clear discussion and contributions\", \"review\": \"This paper succinctly describes a VAE based method driven by a \\\"reference set\\\" for factor discovery.\\n\\nI particularly appreciate the detailed experimental section, and extremely thorough comparisons with other architectures. The appendix is also a nice addition, though even the four page version stands on its own merit.\\n\\nThe work itself is quite clear, and the results speak for themselves. They also raise a number of follow-on questions, maybe the authors have explored these in other experiments - none of these are required to be answered, but the author's insights may be useful for other readers of the work (as well as myself).\\n\\nThe selection of the reference set seems important - is it critical that the the factors of interest be constant across every example? How much \\\"noise\\\" in this selection is tolerable? Does choosing different subsets which both hold the variable constant, result in different factors discovered? How does the size of the reference set impact the result generally, is there some threshold below which the method performs far worse?\\nAre there any suggestions for (perhaps) automated discovery of these constant subsets, or ways to bootstrap this labeling via weak learners and pruning? \\n\\nThe primary things that would improve this paper's rating for me are larger scale experiments, or perhaps non-image data. The model, its baselines, the datasets used, and the experiments completed are all thoroughly spelled out in this paper. Will the authors also release code at some point? \\n\\nOne other paper of interest may be GLSR-VAE (https://arxiv.org/abs/1707.04588) - though not directly comparable to this work, the use of a simple \\\"reward\\\" / label to drive factorization of the latent space seems similar in some ways to the requirement of the reference set to contain the factor of interest.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}", "{\"title\": \"Interesting paper but not well explained\", \"review\": \"The authors provides a training paradigm to leverage \\\"weak supervision\\\" signal via a reference set, which they demonstrate empirically to be effective in regularizing the target factors to encode certain features of interests. This line of work seems quite similar in nature to [1,2,3], who propose to regularize the shared features instead of enforcing them to be a constant.\", \"pros\": \"`1. to address the over-regularization problem in traditional VI, the author proposed to use the symmetrized KL loss to encourage the model to make use of the target factor, which is shown to be effective by their experiment. \\n2. the authors have demonstrated the effectiveness of using a reference set to disentangle target factors from common factors. (see point 2 of Cons)\", \"cons\": \"1. the author proposed to use adversarial training to minimize the two KL divergences, but a justification of the necessity of this method is not provided. \\n2. it is not clear how forcing the target factor of a reference set to be a constant among all data points within the set can help with disentanglement in general. It seems to me to be a heuristic that providing some sort of \\\"weak supervision\\\" will impose some structure in the space of target factor, but I'm not sure to what extent this is effective; e.g. whether certain features will be discarded by \\\"e\\\" and be encoded by \\\"z\\\", or the other way around. Also, have you tried to (1) vary the size of the reference set, e.g. making it smaller, or (2) using multiple reference targets, e.g. with more shared features. \\n\\nIn general, this is an interesting direction. But I'd like to see some intuitive explanation of how this specific regularization help with \\\"disentanglement\\\", which is claimed by the authors. \\n\\n\\n[1] Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data\\n[2] Inferring Identity Factors for Grouped Examples\\n[3] Disentangled Sequential Autoencoder\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Jason Wei", "Kai Zou" ]

[ "Data Augmentation", "Text Classification", "Natural Language Processing" ]

{ "note_id": [ "rklN-bVYt4", "rkeqEYSLFV", "SJeisCOTOE" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554755836446, 1554565426079, 1553989282867 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper3/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper3/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"A nice clear paper for text-related fast data augmentation techniques\", \"review\": \"The authors present a framework of fast and easy methods for boosting text classification. The methods include synonym replacement, random insertion of a word, random swap of two words and last random deletion. They empirically prove that their approach can increase accuracy, especially in small training set sizes.\\n\\nThe writing of the paper was very clear and easy to understand. The authors had a very extensive section with experiments, analysis, ablation studies as well as comparison to related previous work. The actual contributions of the paper are random insertions, swaps, and deletions as synonym replacement was previously\", \"pros\": [\"fast and easy methods\", \"compared with training techniques\"], \"cons\": \"- no learning or training\\n- theoretical explanation \\n\\nI really liked the frequently asked questions section in the appendix, where the authors respond to questions that easily arise. \\n\\nBecause the gains are indeed marginal, statistical significance is something that should be added. As the authors state in the conclusion, a small theoretical explanation of the EDA operations could be done here, as the methods themselves did not include something extremely complex. The methods should be first compared with a non-training or learning approach, for example adding knn words as augmentation.\", \"some_questions_that_could_be_answered_as_well_are\": \"did you explore which replacements affected the model more? for example verbs, nouns or what was their POS-tag? Did you observe any pattern concerning that? As the random deletion seems to be the most effective for small training set sizes, what were the words that were deleted and what can you comment about this phenomenon?\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}", "{\"title\": \"Simple data augmentation techniques for text that work well especially on small datasets\", \"review\": \"Summary: The paper proposes 4 data augmentation techniques for text: synonym replacement, random insertion, random swap, and random deletion. These techniques are validated on 5 text classification datasets, showing increased performance of up to 3% when the dataset size is small. Ablation studies show that all four techniques can be helpful, and suggest sensible values of hyperparameters to set.\", \"strengths\": \"1. The proposed techniques are simple and easily implemented. There's no need to train complicated language models.\\n2. The careful ablation studies (Section 3.3 and 3.4) reveals insights about how much these techniques help and in which context.\\n3. Situating the proposed scheme in the context of data augmentation for text (Section 4 and 7) is very helpful.\", \"minor_point\": \"the name EDA (easy data augmentation) might be too generic.\", \"a_reference_for_the_authors\": \"A related technique for data augmentation for text is called data recombination:\\nRobin Jia and Percy Liang. Data Recombination for Neural Semantic Parsing. ACL 2016.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Mennatullah Siam", "Boris Oreshkin" ]

[ "few-shot segmentation" ]

{ "note_id": [ "Bkl7YBwAY4", "Bkxzx1DdKN", "BkedxZGutV" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1555096954541, 1554702058461, 1554682095611 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper2/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper2/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\", \"comment\": \"Interesting idea and good results. The paper needs some thorough writing update.\"}", "{\"title\": \"Important problem and good results but the text needs significant corrections and improvements.\", \"review\": \"This paper proposed a method for training a convolutional neural network (CNN) from few examples (few-shot learning) for the task of semantic segmentation. The proposed technique allows use to incrementally update the weights of the CNN when encountering examples from classes the networks has not seen before. The paper builds on the idea of weight imprinting, introducing an adaptive weight imprinting scheme that enables updating the weights of previously learned classes. The results on the PASCAL-5^i dataset look convincing, however the text lacks clarity and needs to be proofread and corrected. Specifically, I have the following questions/comments:\\n\\n Multiple grammatical and syntactical errors make the text hard to follow, e.g., the last sentence of the first paragraph of the introduction does not make sense, and citations need fixing. \\n\\n \\\"Masked weight imprinting is performed on multiple resolution levels in order to improve the segmentation accuracy\\\" --> the multiscale nature of the algorithm is not captured by eq. 2. It is not clear how weight imprinting is performed at multiple resolutions. \\n\\n What does \\\"interleaved with backpropagation\\\" mean? This should be clarified in the text.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"Authors propose a few-shot semantic segmentation using an adaptive masked imprinting scheme on the Pascal 5^i dataset\", \"review\": [\"Authors use an imagenet pretrained VGG-16 o perform few-shot semantic segmentation scheme where the goal is to adaptively update the weights to perform few shot semantic segmentation with task, image, label pairs. Once a base classifier is trained, the embedding vectors of new of the support set are used to imprint weights (learn a metric w.r.t e novel class as proxies) for new classes in the extended classifier.\", \"What is Wnew, Wold and Wimprinted, and more complete definition would be useful since this is the key contribution.\", \"Results seem quite interesting, though understand why an adaptive update for the weights in a continous stream setup, would help complete the understanding of the paper.\", \"If the continous stream is a video where the frames are correlated temporally would this approach help ?\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Harshita Seth", "Pulkit Kumar", "Muktabh Mayank Srivastava" ]

[ "Audio keyword detection", "prototypical Metric Loss", "Few-shot", "Transfer Learning" ]

{ "note_id": [ "rJx9fSJtYE", "Skl5UDd_tV", "S1eDX4jC_E" ], "note_type": [ "decision", "official_review", "official_review" ], "note_created": [ 1554736402365, 1554708305700, 1554064415303 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/LLD/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/LLD/Paper1/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/LLD/Paper1/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Good starting point, but needs more clarification\", \"review\": \"The task and method used for this paper seem quite promising, and the general task has broad applications as discussed in the paper and background sections. However, the paper seemed to trail off quickly after the first page, and the experiments section ended suddently. With some restructuring of the focus of the text to highlight the actual experiments and results, this paper could be much improved.\\n\\nDetails on the experiments were fairly limited - more discussion of the feature based frontend (were there standard delta and double delta with the mel, or only base mel?) and even some of the examples of the keywords in context would improve the work.\\n\\nRebalancing this work to include a more detailed analysis of the experiments, and reducing or minimizing the first page or so of content would be very helpful. As it stands, the experimental section is basically just the table, having more discussion of that result, and less of the surrounding description of the models, architecture and background would have been better in this 4 page format. \\n\\nIf the results are the focus of the paper, there should also be some testing done on standard benchmarks rather than only demonstrations on this custom dataset. If the focus is on the new dataset, there should be more discussion of the dataset structure, keywords, and some more exploratory analysis of the overall setting in order to motivate why the dataset is special, or different from standard benchmarks.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}", "{\"title\": \"Reject\", \"review\": \"Summary:\\nThe authors propose a method for detecting keywords in continuous speech given a limited number of training examples (120 for each keyword). A pre-trained ASR model is finetuned with a prototypical loss in conjunction with a metric loss. They find that their training method yields improved accuracy on an internal dataset.\", \"pros\": \"Considering the impact of continuous speech on keyword spotting is very important; I would go so far as to say that recognizing words spoken in isolation is not really keyword spotting. It's good to see that this is being considered in this paper.\", \"cons\": \"The claim that \\\"since the HMM techniques use Viterbi algorithms (computationally expensive) a faster approach is required\\\" is not well founded. This depends entirely on how many HMM states there are and how many possible states you can transition into. For example, Apple's \\\"Hey Siri\\\" detector runs the Viterbi algorithm on the outputs of a neural network for a super simple HMM where there are a small number of states and transitions are only allowed between the current phoneme and the next phoneme. They point out in their article about it that this part requires very little computation and almost all the computation is performed running the neural net (https://machinelearning.apple.com/2017/10/01/hey-siri.html).\\n\\nThe paper has a lot of stylistic bugs; consider the very second sentence: \\\"These spotted keyword frequencies can then be used to analyze theme of communication, creating temporal visualizations and word clouds clouds (2019).\\\" This type of thing happens all through the paper.\\n\\nWeird, unnecessary details like \\\"the python code of the model was taken from the open source repository of Convolutional Neural Networks for Keyword Spotting\\\" are included. Unless you plan to release your code (?), this isn't really relevant.\\n\\nThe experiments are run only on an internal dataset, so it's impossible to replicate this paper.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Anonymous" ]

[ "reproducibility", "few-shot", "machine learning", "statistics" ]

{ "note_id": [ "ByeJnHbBYV", "BJgaIUpmY4" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482598776, 1554400853339 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper8/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting Paper but some parts difficult to understand\", \"review\": \"This paper studies reproducibility for few-shot learning.\\n\\nMy general impression after reading it is that it tries to cover a very large amount of ground and different number of subjects, which can make it a bit difficult to understand. This paper doesn't just attempt to reproduce one paper, it attempts to reproduce three different papers and also proposes to use classical statistics tests with a mixed linear regression model to fit accuracies. This breadth makes the paper's impact potentially very high, but also can make it difficult to understand and fully appreciate. \\n\\nSome of the statistical tests could be given more thorough background information, as they may be outside of the range of knowledge of some readers in the ML community.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}" ] }

[ "Anonymous" ]

[ "reproducibility", "ML4H", "health", "systematic review", "replicability", "data access" ]

{ "note_id": [ "Sylucr-rFV", "r1xpLmpJK4" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482576488, 1554137940575 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper7/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting work on ML4H Reproducibility\", \"review\": \"Summary: This paper conducted an excellent quantitative and qualitative review of the state of the reproducibility for ML healthcare applications. I learned a great deal from reading it!\", \"notes\": \"-Reproducibility is especially important in health due to safety concerns. \\n -Review of 100 ML4H research papers relating to reproducibility\\n -ML4H has more issues with data and code access. \\n -Proposes recommendations to make research more reproducible. \\n -In 2018, 12 healthcare tools using ML got FDA clearance. (cool!)\\n -Quantitative and qualitative review showing ML4H has worse code availability data availability and dataset variety than other ML subfields. \\n -The choice of evaluation metrics is reasonable but a bit limited. \\n -Privacy issues make it difficult to release health data publicly. \\n -ML4H papers are more likely to report mean/stdv than other fields in ML. \\n -Only 19% of ML4H studies used multiple datasets. \\n -The issue of preregistration in ML is interesting. \\n \\nComments / questions: \\n -Does failure to reproduce basic research papers in ML4H really lead to problems for production health systems? Presumably there are many steps of validation beyond the basic research papers. \\n -I really like the reproducibility taxonomy: technical, statistical, and conceptual reproducibility. Technical reproducibility refers to getting the exact results, so includes things like implementation. Statistical reproducibility is equivalence but only up to the statistical properties being the same (so the results follow the same distribution). Conceptual reproducibility means that the idea works as long as the concept is preserved. \\n -Figure 1 is quite nice\", \"rating\": \"5: Top 15% of accepted papers, strong accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Anonymous" ]

[ "case study", "reinforcement", "minigo", "research", "research minigo", "reproducibility", "key challenge area", "field", "results", "groundbreaking algorithm" ]

{ "note_id": [ "rklrYS-SFE", "Hke184akY4" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482556570, 1554138183205 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper6/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting Work\", \"review\": \"I enjoyed reading the paper. I wish there are more papers which study how to debug a model which requires lots of computational resources.\\n\\nThis paper focuses on difficulties involved in models which uses large datasets for training. This paper makes the claim that in order to understand better what the model is doing, it's important to have monitoring systems in the discovery process.\\nThis paper does a pretty good job in describing the components involved in a system like \\nAlphaGo (and its variants). Specifically, paper focuses on reproducibility in self-play algorithms.\\nIts interesting to know that conclusion of the paper is to monitor nearly everything one could think of, to make sure that the implementation is bug-free.\", \"rating\": \"5: Top 15% of accepted papers, strong accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Anonymous" ]

[ "reinforcement learning", "python", "experiments", "new library", "open source" ]

{ "note_id": [ "B1ejPHbrKN", "B1eRGEa1tV" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482531431, 1554138134193 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper5/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting Framework\", \"review\": \"Summary: This paper discusses some features of simple_rl, a framework for RL in Python that emphasizes simplicity and tools for reproducibility. This is a nice workshop paper but would benefit from a clearer discussion of the related work.\", \"notes\": \"-SimpleRL is an algorithm for RL experiments in Python. \\n -After creating agents and MDPs, an experiment log as a json is produced. \\n -Practitioners can share a copy of the experiment file to ensure reproducibility. However with docker or containers this should always be achievable, unless there\\u2019s some guarantee that all of the seeds are in the experiment file? \\n -Consists of MDP objects and agent objects. \\n -Main design goal is simplicity. \\n -MDP has \\u201ctransition function\\u201d and \\u201creward function\\u201d objects. I wonder how well the structure generalizes to model-based RL? \\n -Some utilities for reproducing results from the json files. \\n -Plotting utilities are included.\", \"comments\": \"-Section 2 could do a better job of making it clearer how the simplicity of simple_rl isn\\u2019t achieved by the other libraries. Nonetheless, it\\u2019s still a nice overview.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Anonymous" ]

[ "Network Embedding", "Link Prediction", "Edge Sampling", "Evaluation", "Reproducibility" ]

{ "note_id": [ "SygUJS-BtE", "rygrHnxeKV" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482398304, 1554152508709 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper4/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting evaluation\", \"review\": \"The paper presents EvalNE, a framework to evaluate the performance of node embedding methods on link prediction tasks, The framework abstracts away many challenges that naturally arise when evaluating node embedding approaches. For instance, the framework supports various techniques to sample nodes for training the node embedding model. It also comes with different baselines- both learning-based and heuristics-based, thus encouraging reproducibility of prior work on new datasets.\\n\\nOne caveat is that there is no \\\"one-size fits all\\\" solution. For example, I can think of cases where we would want the graph to be sparse (or maybe even disconnected). Nonetheless, the work is a useful contribution for standardizing evaluation of node embedding techniques.\\n\\nThere are minor grammatical errors like \\\"this errors\\\", \\\"samplig\\\", \\\"out knowledge\\\" etc\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Anonymous" ]

[ "reproducibility", "meta-learning", "closed-form", "few-shot", "miniimagenet", "cifar-fs", "deep learning" ]

{ "note_id": [ "rylX8SbBtE", "r1gSl4pJFV" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482507231, 1554138093477 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper3/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting Work\", \"review\": \"Summary: This paper is a reproduction attempt of the Bertinetto et al. 2019 paper \\u201cMeta-learning with differentiable closed-form solvers\\u201d, an improvement over Finn\\u2019s 2018 MAML (Model-Agnostic Meta-Learning) paper. It is a competently-executed reproduction attempt. It has successfully reproduced Bertinetto et al, modulo small, insignificant differences that do not affect the conclusions, provides a TensorFlow implementation of the original authors\\u2019 proposed algorithm, and additionally resulted in the original authors releasing a PyTorch implementation.\\n\\nThe reproduction focuses on Bertinetto\\u2019s proposed Ridge-Regression Differentiable Discriminator (R2D2), leaving aside their Logistic Regression Differentiable Discriminator (LRD2) because its solution is not truly closed-form. The reproduction provides a service to readers of Bertinetto\\u2019s paper by providing the R2D2 algorithm in pseudo-code, rather than textual form.\\n\\nThe reproduction provides some background context for Bertinetto\\u2019s paper, explaining the N-way K-shot (few-shot) meta-learning problem. Figures 1 & 2, however, do not succeed in explaining the problem well; They are as helpful as they are confusing.\\n\\nWhereas Bertinetto\\u2019s paper fails to provide certain parameters of the neural network\\u2019s convolutions layers, among others, the reproduction attempt correctly guesses and publishes them with their reproduction source code. Correspondence with Bertinetto et al via OpenReview has led to improvements to their paper and release of their code.\\n\\nThe reproduction attempt strengthens in some places Bertinetto\\u2019s paper (and the older MAML paper), and is a worthy contribution to the workshop and to science in general. My only comment is that Figures 1 & 2 are not clear, and distract from the background explanation.\", \"rating\": \"5: Top 15% of accepted papers, strong accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Anonymous" ]

[ "statistical testing", "reinforcement learning", "reproducibility", "replicability", "random seeds" ]

{ "note_id": [ "rJlt4BZBFV", "HyeQ676ktV" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482480607, 1554138043148 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper2/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting work, but relevance to RL is somewhat uncertain\", \"review\": \"Summary: This paper has an interesting investigation of significance testing for RL, however I am skeptical about whether significance testing is appropriate for comparing RL experiments.\", \"notes\": \"-Abstract asserts that checking statistical significance is the first step towards reproducibility, which seems somewhat debatable to me. \\n -Paper should probably just say that it\\u2019s about statistical significance in the title.\", \"comments\": \"-This is a matter of judgement, but I feel like statistical significance is useful for cases where there is a tiny effect size, and we want to make sure that it isn\\u2019t just random noise. For example, one of the earliest uses for statistical significance was for studying the sex difference observed in the number of newborn babies (like 50.00001% are male). In this case it\\u2019s important to know if this result is statistically significant or just random noise. In the case of different RL algorithms, I\\u2019d be hard-pressed to imagine a situation where we\\u2019d want to publish a new algorithm where the improvement isn\\u2019t obviously statistically significant, especially because the power of the statistical test can easily be improved just by running more trials. \\n -I can think of one exception to this, which is if we ran our RL algorithm in the real-world, and then we had a breakdown of many different evaluation categories, and then we wanted to identify which categories had significant or non-significant improvements. \\n -One useful outcome of this is determining the sample sizes needed for certain levels of significance, which could be helpful for determining how many trials to run for RL experiments. \\n -Some interesting recommendations are also made regarding the choice of significance tests.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Anonymous" ]

[ "common assumptions", "unsupervised learning", "disentangled representations", "models", "supervision", "disentanglement", "key idea", "data", "explanatory factors", "variation" ]

{ "note_id": [ "BygV7r-HYE", "rygEcQ61K4" ], "note_type": [ "decision", "official_review" ], "note_created": [ 1554482460217, 1554137996412 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/RML/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/RML/Paper1/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Interesting work on representation learning\", \"review\": \"Summary: This paper is an interesting discussion of disentanglement with an experimental study on several VAE variants. I strongly recommend acceptance, however the writing could be a bit more conservative in the claims given that only VAE-based latent variable models are studied experimentally.\", \"notes\": \"-This paper investigates many proposed methods for learning disentangled representations. \\n -Introduction does a good job of laying out the intuition for what we want to get out of \\u201cdisentangled features\\u201d. \\n -There is some intuition that changing some \\u201cdisentangled features\\u201d should only change those factors of variation and not others. \\n -Appendix A proves that unsupervised learning of disentangled representations is impossible without inductive biases (this doesn\\u2019t seem obvious to me!)\\n -Paper measures disentanglement across 12k models. \\n -Releases a \\u201cdisentanglement_lib\\u201d for evaluating disentangled representations. \\n -Study of different models shows that the \\u201caggregated posterior\\u201d is not correlated, but the dimensions of the representation are correlated. In this sentence I\\u2019m a bit confused about whether this is referring to q(z|x) or q(z). On first reading, I find this claim a bit confusing, because if q(z) follows a gaussian distribution, then its dimensions should be disentangled? \\n -All datasets considered work on the assumption that x is a deterministic function of an underlying disentangled z.\", \"comments\": \"-Uses the wrong style sheet. \\n -It\\u2019s a bit weird to play up the \\u201c10k models\\u201d aspect, because presumably this comes from some kind of hyperparameter sweep or combinatorial explosion? \\n -I think it seems like kind of a bad omission to not include ALI (Dumoulin 2016) or any other models in this family. \\n -Beginning of 4.3 has a typo.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Jacob Biloki", "Chen Liang", "Ni Lao" ]

[ "Neural Networks", "Planning", "Reinforcement Learning", "Structured Prediction", "WikiTableQuestions" ]

{ "note_id": [ "Byx9Mg5qK4", "SJg10EirtV", "BkxfaqrNYV", "ryxKyu6mK4" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1554845714049, 1554523334599, 1554434745538, 1554401249484 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper8/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper8/AnonReviewer4" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper8/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\", \"comment\": \"Even though the results are very preliminary we still accept them for the purpose of fostering interesting discussions.\"}", "{\"title\": \"Interesting idea, good results, clarity and presentation needs improvement\", \"review\": [\"Pro:\", \"Neural Program Planner is proposed to evaluate the value of generations by considering structural and context information.\", \"State-of-the-art is achieved on a benchmark dataset WikiTableQuestions.\"], \"con\": [\"The presentation can be improved, in terms of writing and model explanation.\", \"It's unclear if the summation of token-level scores is the best way for aggregation. It's possible to learn another function over the outputs from conv nets (Os in figure 1).\"], \"other_comments\": \"The first time NSM and MAPO are introduced, they should be spelled out.\\nTypos, \\\"considering the its program\\\", \\\"sequence to sequence model\\\" (should be sequence-to-sequence model)...\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"RL-inspired re-ranker in beam search leads to better structure prediction results in one task: WikiTable QA. Writing needs substantial improvement.\", \"review\": \"The authors propose to improve the quality of beam search in structure prediction by using RL-inspired techniques for learning a forward-cost estimator; the forward-cost estimator is used to re-rank hypotheses in a typical beam search. The proposed method is shown to improve the state of the art on the WikiTable Question Answering task [Liang et al, 2018].\\n\\nAlthough the results are interesting, it is not clear to what extent the improvements are explained by the use of RL-inspired techniques or the fact that the forward-cost estimator accesses additional sources of knowledge. These additional sources of knowledge need to be available at training/decoding time, so the approach amounts to a simple re-ranking of hypotheses surfaced via typical beam search \\u2013 as a consequence, the approach may help re-rank search results that are already encoded in the beam, but not search results that fell off the beam.\\n\\nThe paper has some interesting ideas, but the writing is very poor. A substantial rewriting of the submission that goes beyond documenting the authors\\u2019 train of thought would help readers understand better the main claims and limitations of the approach.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Good motivation and experimental setup but results are not convincing enough.\", \"review\": \"The paper proposes a re-ranking function to improve program generation on the task of answering questions that use tables as context.\", \"pros\": \"1- Interesting task that deserves our attention because of its practical applications.\\n2- The ablation study nicely justifies the extra inputs.\", \"cons\": \"1- The paper is hard to understand. Also, it contains some typos and grammatical errors.\\n2- 0.9-1.1% is a small improvement over MAPO.\", \"questions\": \"1- If the final score given by NPP is the sum of token values in the candidate program, how do you prevent longer programs to have higher scores?\\n2- Did you have to subtract the reward by a baseline to train the search policy with REINFORCE?\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Halley Young", "Osbert Bastani", "Mayur Naik" ]

[ "structure", "deep learning", "generative models", "structured prediction" ]

{ "note_id": [ "SJg56IvdFV", "SJlei_HIt4", "B1gdsdortV", "BJg_lNmBKN", "SyxlnxOXFE" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554704065591, 1554565272162, 1554524319524, 1554490351681, 1554378920036 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper7/AnonReviewer5" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper7/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper7/AnonReviewer4" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper7/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Novel approach to improving VAEs\", \"review\": \"This paper is written clearly and is well thought out in its approach and results. Their contributions are clear, although ambitious as they aim to improve image reconstruction by providing a VAE a reconstructed image based on a grid image created via program synthesis. Their work does show improvement over most state of the art based on the Frechet distance as a performance measure in both their synthetic baseline as well as the Facades baseline other than VED which outperforms the proposed model. It can also be argued though that their improvements may be contributed to the augmented input overfitting the original structure. For samples in their work in which they are compared to VED, it is clear that their work does better represent the original structure, but it is not clear if their method has created a faithful reconstruction of the image or simply a grid of an average part of the inputs which overfit to the given input. Overall, I believe this work is interesting and may be moving in the right direction for improving image reconstruction algorithms although it is an early start.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting task.\", \"review\": \"This work consider the problem of generating or completing images with repeated structures. The main idea is to have a generative model which generates image x from hidden programs P, which is in turn generated from hidden embedding z.\\n\\nThe proposed model is mainly a VAE of programs P extended with components for the image x -- 1) domain greedy heuristics to generate P from x, and 1) CycleGAN to generate x from x_structure which is generated from P in a deterministic fashion.\\n\\nI find the task very interesting with potentially practical applications. The result on synthetic data looks pretty good (much better than VAE and SpatialGAN). The result on real dataset (1855 Facades) still need to be improved. It could be that the data set is just too small for the proposed model. It could also be search failure when generating P from x. Or there might be other reasons?\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Interesting technique to incorporate programs representing global structures in generative models\", \"review\": \"This paper presents a technique to incorporate programs (representing global structure in images) into the generative model thereby capturing more global structure in image generation and image completion tasks. For image generation tasks, it first samples a latent vector z that is used to then sample a program (s,c). The sampled program is then executed to generate x_struct, which is afterwards used to generate the image x using image completion techniques such as GLCIC and CycleGAN. For image completion tasks, it first samples a partial program P_part from a partial image x_part, and then extrapolates the partial program to generate a full program P. The full program is generated to obtain x_part+struct, which is then completed to obtain the full image. The approach is evaluated on synthetic and a real-world dataset of building facades, and it outperforms several baseline models.\\n\\nOverall, I really liked the idea of introducing programs as structural biases in the generative models, where the programs comprise of both symbolic (for loops) and neural (sub-images) components. While there are some recent approaches that have looked at learning such neuro-symbolic programs, this seems to be one of the first approaches for learning generative models of images. The idea of conditioning the image completion models on output obtained by evaluating such intermediate programs and extrapolating programs for completion tasks is also quite nice.\\n\\nAlthough this is an interesting first step, many of the design choices seem a bit restrictive for more general image generation tasks. For example, the paper only considers one for-loop template:\\nfor (i,j):\\n draw (a*i+b, a'*i+b', ??)\\nIt might be interesting to extend the class of templates to include nested loops with conditionals to capture more general global structures together with corner cases.\\n\\nThe search algorithm for synthesizing programs performs an exhaustive search and it is not clear if such an enumerative strategy would scale to richer program templates.\\n\\nFor the VAE decoder, is it the case that the sequence decoder decodes the for loop structure program directly?\\n\\nIt would also be good to provide a bit more details about the program extrapolation part for image completion model. For example, is there an assumption that all images are of equal size for the extrapolation model or is the model conditioned on the final expected size of the image?\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"The paper presents an program synthesis method for image generation, experiments focus on domain specific/synthetic datasets where the models are expected to work well, but does have potential to be applied on more general datasets\", \"review\": \"The paper presents an image generation model, focusing on learning a program that explores the regularities within the data, such as cycling patterns. The model aims at learning two hidden variables, one that defines the structure of the program and another that defines how the model fills the patches defined within the program.\\n\\nI believe this is a good direction for research in image generation, or any generation tasks of similar nature. Such models would allow for computationally efficient models, especially for generating higher resolution images, since there is a quadratic increase of the number of parameters needed. It would be great that in the future versions of the paper the authors could show that larger images can be learned with similar numbers of model parameters. \\n\\nFinally, I would like to see results on medium to larger scale datasets, such as CIFAR-10 or Imagenet. Not only samples are more diversified, it would also allow for an analysis of the sample efficiency of the proposed method compared to existing methods.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Layla El Asri", "Adam Trischler" ]

[ "deep reinforcement learning", "structured prediction", "dialogue" ]

{ "note_id": [ "Skx4Ke55KV", "SJgn2Yy_FV", "S1efcAsUK4", "HygpbQI4YE", "BJxd_c3AdV" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554845819801, 1554672051843, 1554591369928, 1554436869479, 1554070127919 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper6/AnonReviewer3" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper6/AnonReviewer5" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper6/AnonReviewer4" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper6/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Good starting point for a discussion of failure modes in RL-RNN training\", \"review\": \"In this well written paper the authors look into a particular failure mode of RNN models trained via RL: the case where the state representation fails to properly encode the history of previous observations because the next correct/optimal output is the same for different histories. Thus the learned model fails to distinguish the states and produces the same, potentially non-optimal policy.\\n\\nThe paper is quite specific as it addresses a problem that occurs only when training RNNs using RL methods such as REINFORCE to learn policies for a problem where states can be easily confused. Nevertheless, because of the prevalence of RL methods and the limited insight into failure modes this is a useful line of investigation.\", \"the_authors_work_through_3_examples_with_growing_complexity\": \"A simple 3-state problem where the agent has to visit 2 states and then gets a reward, another 3 state problem, where the agent has to perform two sub-actions to change the state and thus 4 actions to get a reward, and a text-based game to simulate a dialogue system where the agent has to learn to rank possible actions given a textual representation of the observations.\\n\\nAll these problems should be trivial to learn with even a tiny RNN and indeed the authors show that a supervised training never deteriorates. In contrast, RL often fails to learn the optimal policy. The authors investigave both REINFORCE and DQN, baseline and no baseline, as well as a loss that takes entropy into account. In all cases a number of training runs deteriorate. \\n\\nWhile I don't see any fundamental shortcomings of this paper I am a bit uneasy about the interpretation. Could it be that the observed phenomenon stems from a particular property of the error surface, where a few steps in the wrong direction might be enough to fall for some non-optimal local minimum? That being said I'd recommend this work for publication. The workshop should be the right place to discuss possible reasons for the observed outcome.\", \"some_smaller_details\": \"Tone down the first paragraph of the introduction a bit and define what you mean by a partially observable problem first.\\nMention the number of runs (50) in the caption of Table 3.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Questions regarding the simple maze\", \"review\": \"The authors investigate state aliasing of the learned representations of RNNs trained with policy gradient methods. They start with a simple maze example and claim to show that state aliasing occurs when several states share the same optimal action and the agent is trained via policy gradient. They find that this does not occur when the agent is trained with Q learning. Then they extend this to a more complex text based game and find that entropy bonuses and baselines improve performance.\\n\\nI had great difficulty understanding the conclusions stemming from the simple maze example, which is the core of the paper. McCallum introduces the example to demonstrate that when x1 & x2 are aliased, the optimal policy can be represented, but that learning when specifically using Q learning on the aliased states does not converge to the optimal policy. The issue stems from the fact that Q learning backs up the value from the aliased state. The authors claim that \\\"policy-gradient methods directly estimate a policy instead of estimating a state-action value, [so] we expect models trained with these methods to be sensitive to state aliasing.\\\" This is in opposition to the conclusions from (McCallum 1996).\\n\\nIt is my understanding that policy gradient methods would not suffer from the same issue. Are the authors claiming that PG methods do? Then they should show that with aliased x1 & x2 in the tabular case, it converges to a suboptimal policy. \\n\\nFurthermore, in Fig 2 bottom, it appears that the distance between the representations for x1 and x2 converges to 0 and the policy converges to the optimal policy. This makes sense as the optimal policy can alias x1 & x2, but goes against the conclusions of the author.\\n\\nFinally, entropy bonuses and baselines are known to improve performance. The authors don't clearly explain the connection between them and their hypothesis that RNN state aliasing is a problem.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Decent study, more experimental analysis needed\", \"review\": \"The paper considers reinforcement learning settings where the states are partially observable and an RNN is used as a reinforcement learning agent. The paper studies the learned representations for the agent\\u2019s policy. It shows that when trained with policy gradient methods, the learned policy could be suboptimal when encountering aliased states sharing the same optimal action. To side step this problem, the following is suggested:\\n\\n1) Using value-based learning method such as DQN.\\n2) Regularization through entropy based exploration and using a baseline for the REINFORCE algorithm.\\n\\nThe hypothesis is verified empirically on three simple, yet illustrative reinforcement learning environments: a simple maze, a structured prediction maze setting, and a text based game simulating a dialog setting.\", \"strengths\": \"========\\n\\nThe paper provides more insight into some of the failure cases for training RNN reinforcement learning agents. The paper shows that when using policy gradient methods, the agent may fail to learn the optimal policy perhaps due to state aliasing effect.\", \"weakness\": \"=========\\n\\n1) Many experimental comparisons are missing:\\n\\na) DQN experiments are missing for the structured prediction task and the text-based game setting;\\nb) Results with entropy based regularization were not reported for the structured prediction task.\\nc) What is the effect of different loss-weight settings on the results? Why are the failures in the first row of Table 7 fewer than the other two weight settings?\\n\\n2) The analysis considers a very controlled and simplified setting, but ignores several factors that might affect the hypothesis:\\n\\na) Does the choice of the optimizer affect the number of failures? \\nb) In sequence to sequence tasks, does pre-training with MLE help in learning better state representations?\", \"clarity\": \"======\\n\\nThe paper is mostly clear and well written. Some of the figure captions are short and non informative (e.g. table 7). The description for the text-based game starting from page 7 is not clear. For example:\\n\\n\\u201cThe retrieval model receives a list of valid actions at every time step \\u201c -> where does this list come from?\", \"this_is_only_mentioned_two_paragraphs_later\": \"\\u201cThe Text World game engine returns all valid actions given the current state of the game \\u201d\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Great paper\", \"review\": \"This paper investigates the representations learned by RNN-based RL agents for solving structured prediction problems when trained with both policy gradient and value-based methods. It studies the conditions leading to state aliasing and highlights strategies to prevent this situation. The hypothesis is that state aliasing happens when different states share the same optimal action and can result into a failure to converge to the optimal policy. The authors study this phenomenon using LSTM and GRU networks on synthetic (toy) environments, and validate their hypothesis.\\n\\nThe paper is clear, well written, and very relevant to the workshop.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Zafarali Ahmed", "Arjun Karuvally", "Doina Precup", "Simon Gravel" ]

[ "variational inference", "reinforcement learning", "monte carlo methods", "stochastic processes" ]

{ "note_id": [ "r1lSBzD_Y4", "SygW6LlPtN", "ByxqW0LIK4", "rklpFLlUY4", "SJlZfp6BF4" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554702909467, 1554609849095, 1554570754193, 1554544260861, 1554533640755 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper5/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper5/AnonReviewer5" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper5/AnonReviewer3" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper5/AnonReviewer4" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"interesting proof of concept paper\", \"review\": \"In this paper, the authors propose the use of REINFORCE for selecting proposals in importance sampling. The idea is to write the problem as a variational inference task, and then establish the link between policy optimization and variational lower bound. The paper is relatively well-written, and it seems like it's above the bar of the workshop to me.\", \"a_few_items\": \"1. The experiments are a bit weak, in the sense that the dataset and experimental settings descriptions are a bit vague, and it does not seem that the authors have tried real-world datasets such as vision, games, and language problems. \\n\\n2. It's widely known that MCMC method could be very inefficient when sampling from complex distributions. That's why some people prefer likelihood-free methods such as neural variational inference. It would be great to add some discussions and comparisons. I would appreciate some analysis on runtime since REINFORCE is also bad at sample complexity. \\n\\n3. I haven't seen much interesting theoretical results from this paper. Maybe you can ask more interesting theoretical questions? Such as the sample complexity? Regret bounds? \\n\\nAnyway, I think it's a nice paper showing the connections among VI, REINFORCE, and importance sampling, but the theoretical and empirical results could be strengthened in many different ways.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Improvements in inferred unobserved values in partially observed trajectory\", \"review\": \"This paper is well written and describes their goals and reasons for their decision making clearly. They clearly state their objectives and use of variational inference in as a reinforcement learning problem to improve trajectories. They provide comparisons to Random Monte Carlo methods as well as ISSoft demonstrating improvements of MC methods and in some instances similar performance to ISSoft. The pros are their clear evidence of being able to use variational inference as a means of matching hand crafted baselines as well as a clear reinforcement learning objective to improve the quality of trajectories. The cons are that their proposed method can learn proposals only for a simple stochastic process and do note that they recieve poor results on rare instances from their X distribution as well as their approach not being able to improve the quality of pre-trained proposals.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"1: The reviewer's evaluation is an educated guess\"}", "{\"title\": \"Empirical performance of reinforced variational inference\", \"review\": \"This paper proposes an empirical validation of the reinforced variation inference method proposed by Weber et al. 2015, as a good candidate for transferring existing policy optimization techniques from variational inference.\\nThe results, however, are not consistent, this work reports reasonable results on qualitative experiments, but quantitative experiments suggest that this approach does not perform better than handcrafted proposals on experiments with random walks. Furthermore, experiments with Chi-squared showed that is not a better alternative than KL-divergence.\\n\\nThis paper provides preliminary results on an interesting topic but with inconclusive remarks, I believe further investigation should be done, in particular regarding the effect of performing RVI on real VI problems and comparing against different VI approaches not only random MC. It would be interesting to understand where does this approach fail, for which cases? It is also not clear why the finetuned approach fails when compared with the simple pretrained model. How different is the proposed approach from Weber 2015? I believe a remark in the related work would better help understand the differences.\\nIs RVI leveraging techniques from policy optimization, such as GAE and A3C, it would be interesting to see whether this is helping in the future. \\n\\nThis paper is written clearly and provides a thorough explanation of results, although some details could be clarified, such as more information on how the model is pretrained and fine tuned? why would this objective be a wrong objective function? (5.4), and what are the family of proposal distributions (q_theta) adopted. \\n\\nI believe this paper provides\", \"minor_revisions\": \"equation equation -> equation (2.1)\\namd -> and (5.4)\\naas -> as (FigureS1)\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Preliminary investigation of RVI on toy task\", \"review\": \"This work investigates the connection between RL and variational inference on a toy task. Building upon the RVI (Weber et al., 2015), off-the-shelf RL methods like policy gradient methods can be applied to the variational inference reformulated as RL. The authors attempted to apply this idea to learn a proposal distribution in an importance sampler. Based on RVI, they also proposed alternative objective that replaces the KL divergence with chi-square divergence. An inference problem in simple 1D random walk is used for evaluation. The proposed approach is shown to outperform simple MC baseline, but still worse than a hand-crafted baseline. And the objective using chi-square divergence isn't performing well.\\n\\nThe paper is relatively well-written and easy to follow. This work is an empirical investigation of RVI on a toy task, and proposes the use of chi-square divergence as an alternative to KL-divergence. \\n\\nAlthough the experiment is quite preliminary and the novelty is limited, this is an interesting work in progress and I recommend acceptance. I would suggest exploring some real inference tasks in graphical model as additional benchmark to make the experiments more solid.\", \"minor_issues\": \"Page 1 paragraph 4 \\\"Variational inference (VI, Blei et al. (2006; 2017)) is an alternate technique\\\" --> \\\"alternative technique\\\"\\nThe equation (6) should have \\\"x_T\\\" instead of \\\"x_t\\\".\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}" ] }

[ "Wouter Kool", "Herke van Hoof", "Max Welling" ]

[ "reinforce", "multiple samples", "baseline", "sequence generation", "structured prediction", "travelling salesman problem" ]

{ "note_id": [ "Hylj9ecqtN", "HkxADT9OtV", "rylUMre_YN", "ByehYNJPFV", "SJlR4YwSFE" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554845842996, 1554718054003, 1554674958419, 1554605188405, 1554508085560 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper4/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper4/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper4/AnonReviewer5" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper4/AnonReviewer4" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"REINFORCE loss w/o replacement\", \"review\": \"This paper proposes a method for learning a policy based on a variance reduced REINFORCE loss based on sampling without replacement. Experiments suggest that learning with this approach performs slightly better when compared with a greedy rollout baseline (Kool et al 2019) but it is more sample efficient requiring fewer instances to learn.\\nThe paper provides results in a toy problem using a Travelling Salesman Problem with a few nodes (20), it would be interesting to evaluate the performance on a more complex task to see whether sampling without replacement would still be beneficial over other sampling schemes.\\nIt would also be interesting to evaluate the performance of the normalized and non-normalized version of the loss, to see how much variance is reduce between the two proposed approaches.\\n\\nThis paper provides interesting preliminary results on an important topic for deep structure prediction with RL and therefore I would like to see it presented in this workshop.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Unbiased REINFORCE estimator using sampling without replacement\", \"review\": \"In this work the authors suggest to draw several samples without replacement in REINFORCE training, allowing them to use a local baseline, i.e. a baseline with regard to a single instance as opposed to computing the baseline over the other examples in the batch. Some complexity comes from the fact that the multi-sample-baseline is much easier derived when instead sampling with replacement as common in related work.\\n\\nThis paper relies heavily on previous work that established the \\\"Gumbel top-k trick\\\" which allows sampling without replacement, but self-sufficient given the references. Their new contribution is a gradient estimator based on sampling without replacement, including a variant with a baseline that is based on the other samples produced for the same example (e.g. different translations for the same source sentence).\\n\\nExperiments on a TSP solver show that sampling several times per input example performs better than doing a single sample and on-par with a computationally expensive version where roll-out is used to produce a baseline for the single sample case. Furthermore sampling without replacement seems to work slightly better than with replacement. That being said, the latter seems much easier to implement.\\n\\nThe proposed approach might be a good fit for the somewhat nice setting where one trains on few examples and samples from a peaked distribution where sampling the same sequence repeatedly would be an issue.\\n\\nOverall this paper is very pleasant to read and presents the flow of ideas well.\\n\\nThe work is accompanied by a detailed proof that the derived estimator is unbiased. Skimming the proof revealed no issues but I could have missed something here.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Novelty and experiments\", \"review\": \"This work aims to improve REINFORCE, which is used to train models in structured prediction settings to directly optimize the test-time objective. It shows that by drawing multiple samples (predictions) per datapoint, it can learn with significantly less data, as it freely obtains a REINFORCE baseline to reduce variance.\\n\\nSimilar ideas, drawing multiple samples (predictions) per datapoint, have already been studied in RL applications (see below). Those related works are not reviewed in this paper. It is better to discuss the difference between this work and existing works and highlight the novelty of this work.\\nMinimum Risk Training for Neural Machine Translation, ACL 2016\\nDual Learning for Machine Translation, NIPS 2016\\n\\nThe experiments can be improved from several aspects.\\n1. Compare with recent methods, including non-RL methods such as pointer networks.\\n2. Only very small scale settings are tested (e.g., 20 nodes). It is better to consider larger scale experiments. For example, pointer networks are tested with 50 nodes.\", \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Multi-sample REINFORCE estimators w/ and w/o replacement\", \"review\": \"The authors derive a REINFORCE estimator based on sampling without replacement (building on the recent work of Kool et al.) and show improvements over existing techniques on structured prediction.\\n\\nThe paper is well-written and the experiments are informative. The paper could be improved by theoretically quantifying the improvement gain from sampling w/ replacement. This would help to understand why the benefits of sampling without replacement diminish with larger instances and larger k.\", \"rating\": \"4: Top 50% of accepted papers, clear accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Rodrigo Nogueira", "Jannis Bulian", "Massimiliano Ciaramita" ]

[ "natural language", "reinforcement learning", "structured prediction", "multi-agent learning", "deep learning" ]

{ "note_id": [ "BJxM_lc5YE", "HyxE-hwBtE", "S1eqGbqNtN", "BJg4Bsr4K4", "SkeACccCdE" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554845801579, 1554508795688, 1554452753945, 1554434876395, 1554062037903 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper3/AnonReviewer5" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper3/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper3/AnonReviewer4" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper3/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\"}", "{\"title\": \"Elegant idea, preliminary results\", \"review\": [\"This paper explores whether decomposing an agent-oriented structure prediction task into sub-agents and an aggregator improves performance on query reformulation tasks. The paper considers both decomposing on random sub-sets of the data (bagging), and partitioning into semantically similar classes. The results show that decomposing into sub-agents improves performance.\", \"The method is applied to 2 tasks - document retrieval (on 3 datasets), and question answering. One of the main results is that the same level of performance can be achieved with much less compute.\", \"The basic idea is elegant, but I have 3 concerns. First, I found that some aspects were really unclear. Second, it wasn\\u2019t clear but I inferred that the results were created from a single run, which limits how much can be concluded. Finally, it wasn\\u2019t clear whether the method would generalize to other tasks.\", \"Clarity. I was unclear on the following:\", \"Fig 1: why is there 1 search box in Fig 1b, but N+1 search boxes in Fig 1c? If the idea is that the search in Fig 1b is doing a single search by simultaneously receiving all queries, that wasn\\u2019t at all clear.\", \"Sec 3.4, \\u201cthe aggregator receives as input q_0\\u2026\\u201d: q_0 is not an input into the aggregator in Fig 1. Should q_0 be a_0?\", \"Sec 4.3, \\u201cother metrics \\u2026 resulted in similar or slightly worse performance than Recall@40.\\u201d Does this mean that the proposed method is strongest only when evaluated under certain metrics?\", \"Table 1: I was unclear on how RL-10-Ensemble and RL-10-Full differ, since the descriptions in the text (Sec 4.4 and 4.5) are nearly identical.\", \"Experimental procedure. Were the results the product of a single run, or the results of averaging many runs? I might have missed it but I inferred they were from a single run. As a result, readers won\\u2019t know what the variance of the methods are, and whether the relative orderings of the methods are reliable.\", \"Generality. Query re-writing is a very specific task, and both experiments here relied on an external search engine. I\\u2019m worried that the method may be \\u201cfit\\u201d to specific properties of the task \\u2014 for example, in the current setting, all actions can effectively be executed to produce search results; in other settings, this isn\\u2019t possible. Moreover, because of the search engine, producing a diversity of responses seems to be important; in other settings, having a diverse set of outputs may be less important. The results would be more impactful if other types of tasks were considered \\u2014 even more text generation tasks, like dialog response generation, paraphrasing, translation, etc.\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Simple idea, strong empirical results, but some results are surprising\", \"review\": \"The authors propose a multi-agent approach for query reformulation. A set of sub-agents is trained on disjoint splits of the training data, and a meta-agent (the aggregator) is trained on the whole dataset and learns to combine outputs from the sub-agents. The proposed approach is evaluated on two tasks: document retrieval and question answering. For document retrieval, recall@40 is used as the reward signal. The token level F1 score on the answer is used as the reward signal for question answering.\", \"strength\": [\"========\", \"The proposed approach is simple, easy to parallelize, and hence computational efficiency can be achieved.\", \"Strong empirical results, although no significant improvement was observed when improving over a BERT model on the question answering task.\", \"Paper is clear and easy to follow.\"], \"weakness\": \"=========\\n1) Some results are surprising and counterintuitive, and hence further analysis is needed for a better understanding:\\n\\na) What was the size of the random partitions used for training the sub-agents? I expect the dataset has been split equally between the agents in the RL-N-Sub setting, have the authors experimented with other partition sizes in addition to training the sub-agents on the full dataset?\\nb) The results described in section 5.5 are surprising, perhaps the reason for this results is limited to K-means clustering which is prone to fail when the clusters have non-spherical structures. Have the authors experimented with other density based clustering algorithms (e.g. dbscan)?\\nc) Not being able to significantly improve on BERT in the question answering task is concerning. \\n\\n2) Originality is limited, the main difference between the proposed approach and previous work seems to be the training on disjoint splits for the sub-agents, and then training an aggregator to combine the results.\", \"clarity\": \"======\\n\\nThe paper is clear and well written.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Hierarchical RL applied to document retrieval and QA. Interesting but somewhat disappointing results.\", \"review\": \"This paper instantiates in the context of query reformulation an approach to structure prediction (seq2seq) that is inspired by hierarchical reinforcement learning methods proposed in the early nineties by Singh, Lin, Dietterich, and Hinton.\\n\\nThis is an experimental paper with well executed experimental design and a broad range of comparative results for doc retrieval and question answering. Although experimentally solid, the paper is light in terms of insights.\\n\\nOn the positive side, the approach is easily parallelizable and achieves better generalization performance than a model average ensemble. On the negative side, the approach does not lead to significant improvements.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Clear paper but lack of novelty and marginal improvement\", \"review\": \"This work applies the mixture-of-experts approach to the query reformulation task, where ones goals is to reformulate queries in order to optimize the quality of results returned by an underlying retrieval system. The authors propose to train multiple reformulators (sub-agents) each on a subset of the training dataset, leading to a diverse set of reformulators. Given a new query, each reformulator can thus propose different query reformulations, which are then fed (in addition to the initial query) to an underlying search engine. For each query, the search engine outputs a recommendation. All recommendations are sent to an aggregator (meta-agent), which decides which one to pick. The authors propose an architecture and loss function to train the aggregator.\\n\\nAs pointed out, the idea of using a mixture-of-experts trained on different subsets of data is old and well-known. Results showing that ensembles of such sub-agents produce diverse reformulations are therefore not surprising.\", \"impact_of_aggregator\": \"Appendix C shows the benefits of the ensemble of sub-agents independently from the effect of using an aggregator (Fig.1c) instead of a selector (Fig.1b). However, it seems like anything combined with BERT (alone or as an aggregator) increases the performance so much, therefore making the benefits of using diverse sub-agents appear as marginal. It is not clear that relying on diverse ensembles would even bring any improvement given a strong aggregator.\", \"diversity_of_sub_agents\": [\"How does the size of sub-datasets used by sub-agents affect the performance and diversity? Is there any overlap between sub-datasets used by sub-agents? If not, it means that the number of sub-agents affects the size of dataset available to each sub-agent. How would that affect the performance in return? If sub-datasets are allowed to overlap, should we expect a tradeoff between large sub-datasets (increasing overlap, therefore decreasing diversity) and small sub-datasets (more diversity, maybe worse generalization)?\", \"Overall, although the approach lacks novelty and results are either expected or not impressive, the paper is well written, easy to follow, and presents a decent empirical evaluation. I therefore think that this paper is suitable for a workshop.\", \"Minor questions/comments:\", \"[Sec.2] Small communication overhead: A downside of the approach of Shazeer et al. (2017) is that it requires that \\\"output vectors of experts are exchanged between machines.\\\" The authors claim that their method instead \\\"requires only scalars (rewards) and short strings (original query, reformulations, and answers) to be exchanged.\\\" How are reformulations and answers different from the output vectors?\", \"[Sec.3.3] Training of sub-agents: \\\"At training time, instead of using beam search, we sample reformulations.\\\" From which distribution are reformulation sampled?\", \"[Sec.3.4] Training of aggregator: Why is the loss only using the relevance score (see Eq.3)?\", \"[Sec.4.3] Reward: Any insights regarding why agent trained to optimize Recall@40 perform better than those trained to optimize MAP, even though MAP is used in evaluation?\", \"[Sec.4.5] Proposed model RL-N-Full: What is the \\\"best reformulations of all the agents\\\"?\", \"[Sec.4.6, 2nd paragraph] \\\"...an RL-10-Full with an ensemble of 10 aggregators yields a relative performance...\\\": Should \\\"aggregators\\\" be replaced by \\\"reformulators\\\"?\", \"[Tables 1 and 2] Consistency of training time in sub-agents+BERT: Why does RL-10-Sub + BERT Aggregator require 10 training days (Tab.1) while BERT + AQA-10-Sub require 1 training day (Tab.2)?\"], \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Bowen Tan*", "Zhiting Hu*", "Zichao Yang", "Ruslan Salakhutdinov", "Eric P. Xing" ]

[ "sequence generation", "maximum likelihood learning", "reinforcement learning", "policy optimization", "text generation", "reward augmented maximum likelihood", "exposure bias" ]

{ "note_id": [ "SyxP_aLdFE", "HkxkY2T8KE", "r1gDTrGIKE", "HkxAInwBFV", "B1gGnx_NK4", "SkgJtsr4YN" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554701678969, 1554599031126, 1554552255354, 1554508885863, 1554444458289, 1554434934804 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper2/AnonReviewer4" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper2/AnonReviewer1" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper2/AnonReviewer3" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper2/AnonReviewer2" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper2/AnonReviewer5" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\", \"comment\": \"This paper builds interesting connections of multiple methods (MLE, RL, RAML, etc) used to train model for sequence generation. All reviewers recommend acceptance.\"}", "{\"title\": \"Good work, but not that natrual\", \"review\": \"The authors try to unify different algorithms for sequence generation and present a generalized entropy regularized policy optimization formulation. They show that divergent algorithms can be reformulated as special instances of the framework,\\nwith the only difference being the configurations of reward function and a couple of hyperparameters.\\n\\nThe analysis and proposed framework are interesting. \\n\\nI have several technical questions. \\n\\n1. One cercen about the work is that the formulation in Eq. (1) is not natrual and it is not reasonable intuitively. Which one is the final model, q or \\\\theta? According to previous description, \\\\theta is the model (or model parameters). However, for MLE, when \\\\alpha->0, \\\\theta will not make any impact to the reward defiend in Eq. (1), and only q determines the reward. That is, optimizing L(q,\\\\theta) will only update q but not \\\\theta. \\n\\n2. \\\"Generally, a larger exploration space would lead to a harder training problem.\\\" I don't get this point. \\\"common rewards (e.g., BLEU) used in policy optimization are more smooth than the \\u000e-reward, and permit exploration in a broader space.\\\" If BLEU is more smooth, why leads to a harder training problem?\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\"}", "{\"title\": \"Nice contribution of unifying multiple common objectives in a single framework leading to algorithmic insight\", \"review\": \"This paper takes a number of widely-used algorithms for sequence generation, including maximum-likelihood, RAML, SPG and data noising and shows that they can all be viewed as optimizing a member of a family of objective functions, which is defined through three hyper-parameters that control parts of the objective - the reward, a maximum entropy term, and a KL term between a variational distribution and the model.\\n\\nThe authors show the exact values of the hyper-parameters that lead to these various objectives and also shed light onto how these hyper-parameters correspond to a trade-off between the exploration and difficulty of learning, where more exploration results in a more difficult learning problem. \\n\\nBecause now we have a family of objectives, the authors now naturally propose an algorithm that anneals the values of the hyper-parameters using a curriculum where simple learning without exploration happens at the beginning and more exploration is added later on to avoid local minima.\\n\\nI found the analysis clear and interesting, the insights on the relation between the algorithms to be informative and the final simple algorithmic contribution to be natural and worthwhile. Overall, a nice paper that I think definitely fits well in this workshop.\", \"it_is_worth_noting_that_a_similar_but_different_attempt_has_been_made_in_the_context_of_sequence_generation_when_there_is_no_gold_sequence_given_at_training_time\": \"See Misra et al. 2018\", \"http\": \"//www.cs.cornell.edu/~dkm/papers/mchy-emnlp.2018.pdf\", \"rating\": \"5: Top 15% of accepted papers, strong accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"A nice review of recent approaches (RAML, SPG, data noising)\", \"review\": \"The authors unify several recent frameworks for training sequence generation models and claim that their general framework is principled and provides novel interpretations of previous algorithms. They propose an interpolated model which they evaluate on several sequence generation task and show improved performance.\\n\\nIt was unclear why the unified model was principled (this was not addressed in the paper), and I did not find that it provided novel interpretations beyond the existing literature. The unified objective amounts to adding previously used terms together with weights. The analysis of the objective is mostly qualitative and descriptive, rather than analytical. The paper could benefit from clearly defining terms like \\\"regular shaped rewards\\\", \\\"smoothness\\\" for discrete distributions, and \\\"exploration area\\\", etc.\\n\\nThe authors briefly mention learning to search approaches, but given their recent strong performance (e.g., Sabour et al. ICLR 2019), this is an important comparison that is missing.\\n\\nFor a workshop submission, the review of previous methods and empirical results are sufficiently interesting for acceptance.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"A useful unified formulation for several standard learning algorithms\", \"review\": \"The paper presents a unified objective function for optimizing sequence generation models. The unified formulation includes several standard learning algorithms as special cases. These standard algorithms include training by MLE, RAML, SPG, data noising, and sequential sampling.\\n\\nThe framework is based on policy optimization combined with entropy based regularization. The ground truth sequence y* is perturbed according to a probability distribution q(.|x) resulting in a perturbed target sequence y (i.e. y~q(.|x)). The goal is to maximize the expected reward under the new perturbed distribution of labels q. A KL-divergence penalty is also included to prevent deviation from the model distribution parameterized by theta. The objective function is regularized by imposing a maximum entropy assumption on q. The optimization objective is solved in an expectation maximization fashion. Different choices for the reward function and penalties for the KL-divergence and entropy terms yield standard learning algorithms ranging from MLE to SPG. \\n\\nFinally, a new algorithm is proposed that dynamically interpolates between the different learning algorithms. The algorithm is evaluated in a very limited experimental setting for machine translation and text summarization.\", \"strength\": \"========\\n\\nThe unified formal connection is useful in understanding the difference between the learning algorithms in terms of how different target sequences are rewarded, which sequences are explored, and how the policy is regularized.\", \"weakness\": \"=========\\n\\nThe experimental setting for evaluating the new proposed algorithm is very limited. The dataset used for the machine translation experiment is relatively small. This is a major concern as the exposure bias problem can be mitigated by using larger datasets.\", \"clarity\": \"======\\n\\nThe paper is mostly clear. However, in section 3.1, it wasn\\u2019t clear at the beginning what the distribution q represents. The future reference in the \\u201cmore details below\\u201d was not helpful. In general proving more intuition about the objection function in equation (1) would be helpful and make the presentation more clear.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"A generalized entropy regularized policy formulation encodes a variety of approaches to seq2seq learning, from maximum likelihood estimation to reinforcement learning implemented using the RAML, SPG, and data noising methods and leads to improved results on two tasks: machine translation and summarization.\", \"review\": \"This is a nice submission. The authors show how a generalized entropy regularized policy formulation encodes a variety of approaches to seq2seq learning, from maximum likelihood estimation to reinforcement learning implemented using the RAML, SPG, and data noising methods. Besides the theoretical insight, the authors show how by implementing an easy-to-hard paradigm that resembles curriculum learning leads to improvements on two tasks: machine translation and summarization.\\n\\nThe paper is likely to generate good discussions, especially with respect to other approaches to sequence learning that are not yet encompassed by the proposed framework.\", \"rating\": \"5: Top 15% of accepted papers, strong accept\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Yesmina Jaafra", "Jean Luc Laurent", "Aline Deruyver", "Mohamed Saber Naceur" ]

[ "Neural networks", "Deep reinforcement learning", "Actor-critic model", "Autonomous driving", "Carla simulator" ]

{ "note_id": [ "Skg1WkD_K4", "SkxAvbRIKE", "Hkx0GESBtV", "HJg-dj07tE", "rkgE3D6RO4" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1554702071020, 1554600294382, 1554498581765, 1554406248799, 1554073515663 ], "note_signatures": [ [ "ICLR.cc/2019/Workshop/drlStructPred/Program_Chairs" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper1/AnonReviewer4" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper1/AnonReviewer5" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper1/AnonReviewer3" ], [ "ICLR.cc/2019/Workshop/drlStructPred/Paper1/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Acceptance Decision\", \"decision\": \"Accept\", \"comment\": \"Even though the results are very preliminary we still accept them for the purpose of fostering interesting discussions.\"}", "{\"title\": \"Innovation and contribution\", \"review\": \"This work is about autonoous driving and proposes a deep reinforcement learning (RL) algorithm embedding an actor critic architecture with multi-step returns to achieve a better robustness of the agent learning strategies when acting in complex and unstable environments.\\n\\nThe major problem of this work is that its technical/research innovation is not clear. Actually the proposed algorithm is not new and similar or more advanced algorithms have already been proposed, such as Generalized Advantage Estimation, Schulman et al. 2016(b). This paper is more like a homework in a RL course.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Low improvement, with little evidence.\", \"review\": \"The paper struggles to clearly explain their contribution to current works, and improvements over a baseline. The written quality of the paper shows numerous grammatical errors and careless mistakes. On a positive note they do demonstrate the variance improvement of using multistep rewards over a single step look ahead.\", \"rating\": \"1: Strong rejection\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Nice task but it should have more experiments.\", \"review\": \"This paper evaluates two common RL algorithms in a simulated driving environment.\", \"pros\": \"1- Nice choice of using a more realistic environment.\\n2- The paper is easy to read and understand.\", \"cons\": \"\", \"1__some_details_are_missing\": \"how exactly the reward is computed? How long an episode lasts, on average? What is the average distance a vehicle runs before having a collision?\\n2- Only two RL methods are evaluated (A2C and A3C). Given the abundance of methods in the literature, it would be nice to see comparisons with some other commonly used methods, such as Q-Learning and TRPO.\", \"questions\": \"1- Does the agent use a third person (outside the vehicle) camera view? If yes, the results would be more convincing if the first-person view was used because that is closer to the viewpoint of real vehicle.\\n2- In section 4 you mention that traffic rules are given as input to the agent. How do you extract these rules from the environment? For example, how a stop sign is given as an input to the agent? Also, a more realistic environment should not explicitly give these to the agent. Instead, inferring traffic rules should be a job of the agent as traffic signs often change due to constructions, accidents, etc.\", \"rating\": \"3: Marginally above acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}", "{\"title\": \"Experimental paper with unclear contribution\", \"review\": \"This paper tackles the problem of autonomous driving using deep RL algorithms. More specifically, the authors evaluate the benefit of using a multi-step returns critic in A2C for this task. Experiments are conducted using the realistic driving simulator CARLA.\", \"very_limited_contribution\": [\"Though investigating deep RL approaches on the difficult task of autonomous driving, I find the experiments to be very limited as only one algorithm (A2C) is considered. It would have been interesting to present a broader study with more than one method.\", \"Dosovitskiy et al. (2017) already evaluate A3C for the autonomous driving task using CARLA. The contribution of the current paper therefore seems limited to evaluating the benefit of multi-step returns.\"], \"it_is_not_clear_how_rewards_are_defined\": [\"What are the weights given to each \\\"goal feature\\\"?\", \"How were these weights chosen?\", \"Would a \\\"good policy\\\" in terms of those rewards actually be considered \\\"good\\\" by humans?\"], \"experiments\": [\"How many repetitions were performed?\", \"Does the shaded are on Figs. 2-3 correspond to standard deviation? If so, could you really conclude that there was a difference between the two compared methods?\", \"All experiments are performed on straight roads. It would be interesting to see how different/similar results are on more challenging roads.\", \"What were the environmental conditions during training? If it was always sunny, one cannot really be surprised that the methods do not generalize to different weathers...\", \"Comparison with state-of-the-art results? I understand that existing approaches evaluate performance differently in their papers. The approaches could still be run on the given setting.\", \"The paper is easy to read. The contribution is presented as a study of deep RL techniques for autonomous driving, which is relevant for the workshop. However, this has already been done in the past (e.g. Dosovitskiy et al., 2017), especially using algorithms very close to what is considered here, and the approaches studied previously were not included in the current paper. Moreover, the experiments lack details to actually make the presented comparison meaningful.\"], \"minor_comments\": [\"[Eq.4] R(t) should be R_t.\", \"When citing multiple references one after the other, putting them in the same parenthesis increases readability.\"], \"rating\": \"2: Marginally below acceptance threshold\", \"confidence\": \"2: The reviewer is fairly confident that the evaluation is correct\"}" ] }

[ "Kafeng Wang", "Xitong Gao", "Yiren Zhao", "Xingjian Li", "Dejing Dou", "Cheng-Zhong Xu" ]

[ "transfer learning", "pruning", "faster CNNs" ]

{ "note_id": [ "u4lGEyBWi5n", "EE4Ml5hZtI", "r1e9TnrIsH", "BJxy8nSUiS", "BJlAoqSIsH", "BklJwjhd5B", "BkeROrlU5B", "Ske1UQAX9B", "Skg8EJfGuH" ], "note_type": [ "comment", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1684749831020, 1576798753000, 1573440706488, 1573440582977, 1573440166129, 1572551510593, 1572369782324, 1572229959099, 1570017070423 ], "note_signatures": [ [ "~Gladis_Ne_Limes1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2594/Authors" ], [ "ICLR.cc/2020/Conference/Paper2594/Authors" ], [ "ICLR.cc/2020/Conference/Paper2594/Authors" ], [ "ICLR.cc/2020/Conference/Paper2594/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2594/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2594/AnonReviewer3" ], [ "~Mert_Kilickaya1" ] ], "structured_content_str": [ "{\"title\": \"re\", \"comment\": \"Read more in https://mlsdev.com/services/ui-ux-design\\n 1. UX/UI Design Interviews & Research: Consult with stakeholders and end users to create a thorough understanding of user needs, design goals, and project objectives. 2. User Personas & Use Cases: Generate user personas and use cases to focus the design process and scope requirements for the project. 3. User Interface & Interaction Design: Create high-fidelity wireframes, interface designs, and prototypes to bring user interfaces to life and enable user testing. 4. Interface Testing & Usability: Perform A/B testing, user-testing, and iterative usability testing to ensure design solutions meet user needs. 5. Accessibility & Responsive Design: Ensure design solutions are accessible\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper presents an attention-based approach to transfer faster CNNs, which tackles the problem of jointly transferring source knowledge and pruning target CNNs.\\n\\nReviewers are unanimously positive on the paper, in terms of a well-written paper with a reasonable approach that yields strong empirical performance under the resource constraint.\\n\\nAC feels that the paper studies an important problem of making transfer learning faster for CNNs, however, the proposed model is a relatively straightforward combination of fine-tuning and filter-pruning, each having very extensive prior works. Also, AC has very critical comments for improving this paper:\\n\\n- The Attentive Feature Distillation (AFD) module is very similar to DELTA (Li et al. ICLR 2019) and L2T (Jang et al. ICML 2019), significantly weakening the novelty. The empirical evaluation should consider DELTA as baselines, e.g. AFS+DELTA.\\n\\nI accept this paper, assuming that all comments will be well addressed in the revision.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for your detailed reviews.\", \"comment\": \"Thank you for your comments. We would like to respond to the issues kindly raised by the reviewer:\\n1. In the last sentence of Section 3.5, we mentioned that \\u2018delta_s\\u2019 is set to a value such that 50% of the channel neurons use the predictor function \\u2018h_l\\u2019\\u201d. For this we mean that we first compute the variances of \\u201ch_l(x_{l-1})\\u201d for each channel, and use the median of the channel variances as the value of the threshold \\u201cdelta_s\\u201d. As kindly suggested by the reviewer, we will be updating this section accordingly.\\n2. Thanks for pointing out this to us, it will be fixed in the next revision.\"}", "{\"title\": \"Thank you for your detailed reviews.\", \"comment\": \"Thank you for your comments. We would like to answer your questions:\\n1. Consider a convolution operation with a \\u201ck * k\\u201d kernel, which takes input features with \\u201cCi\\u201d channels, and computes feature maps of shape \\u201cCo * Ho * Wo\\u201d. To evaluate the convolution thus requires \\u201ck^2 * Ci * Co * Ho * Wo\\u201d multiply-accumulate operations (MACs). AFS can reduce the number of MACs required in a coarse-grained manner: before computing the convolution, AFS can predict the importance of each output channel, request the convolution to evaluate only \\u201cceil(d * Co)\\u201d channels, and skip the remaining channels by setting them to zeros. Note that if the preceding layer is also a convolution that produce sparse outputs with only \\u201cd * Ci\\u201d non-zero channels, the input channels can also be skipped, reducing the number of MACs required to \\u201ck^2 * d^2 * Ci * Co * Ho * Wo\\u201d, a quadratic reduction in terms of \\u201cd\\u201d. We will update Section 3.4 to explain this in greater detail.\\n2. As previous work did not examine the opportunity of pruning and transfer learning jointly, In Table 2, we re-implemented L2, L2-SP [1] and LwF [2]. We then used the best they can achieve with any one of the pruning methods, and compared the results against AFDS under 2x, 5x or 10x speedup constraints. We have additionally compared to existing smaller transfer learned models from related works [3, 4] in Table 3.\\n3. We suspect the primary reason for the challenge is with the initial weights used in AFS. As the network depth gets larger, small changes in the variance used in initialization would result in highly sensitive changes in gradient magnitudes. Thanks for pointing out this to us and we will look into this in greater detail and update the paper accordingly.\\n4. The code and accompanying models will be made available soon.\\n\\n[1]: Xuhong Li, et al., Explicit Inductive Bias for Transfer Learning with Convolutional Networks, ICML 2018.\\n[2]: Zhizhong Li, et al., Learning without Forgetting, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.\\n[3]: Sergey Zagoruyko, Nikos Komodakis, Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer, ICLR 2017.\\n[4]: Yunhun Jang, et al., Learning What and Where to Transfer, ICML 2019.\"}", "{\"title\": \"Thank you.\", \"comment\": \"We would like to thank the reviewer for the positive comments.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper presents an improvement to the task of transfer learning by being deliberate about which channels from the base model are most relevant to the new task at hand. It does this by apply attentive feature selection (AFS) to select channels or features that align well with the down stream task and attentive feature distillation (AFD) to pass on these features to the student network. In the process they do channel pruning there by decreasing the size of the network and enabling faster inference speeds. Their major argument is that plain transfer learning is redundant and wasteful and careful attention applied to selection of the features and channels to be transfered can lead to smaller faster models which in several cases presented in the paper provide superior performance.\\n\\nPaper is clear and concise and experimentally sound showing a real contribution to the body of knowledge in transfer learning and pruning.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"In general, I think it is a good paper and I like the contribution of the author. I think they explain in detail the methodology. The results compare the new methodologies with different databases which increase the credibility of the results. However, there is a couple of additional question that is important to manage:\\n\\n1) The paper presents three different contributions. However, it is so clear how this work helps for \\\"By changing the fraction of channel neurons to skip for each convolution, AFDS can further accelerate the transfer learned models while minimizing the impact on task accuracy\\\" I think a better explanation of this part it would be necessary. \\n\\n2) The comparison of the results are very focused on AFDS, Did you compare the results with different transfer learning approach? \\n\\n3) During the training procedure. We need a better explanation of why \\\"we found that in residual networks with greater depths, AFS could become notably challenging to train to high accuracies\\\". Also, the results of the empirical test it would be useful to understand the challenges to train the network. \\n\\n4) I think it would be useful to have the code available for the final version.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a method called attentive feature distillation and selection (AFDS) to improve the performance of transfer learning for CNNs. The authors argue that the regularization should constrain the proximity of feature maps, instead of pre-trained model weights. Specifically, the authors proposes two modifications of loss functions: 1) Attentive feature distillation (AFD), which modifies the regularization term to learn different weights for each channel and 2) Attentive feature selection (AFS), which modifies the ConvBN layers by predicts unimportant channels and suppress them.\\n\\nOverall, this is a good work in terms of theory and experimentation, thus I would recommend to accept it. The approach is well motivated, and the literature is complete and relevant. The author's argument is validated by experiments comparing the proposed AFDS method and other existing transfer learning methods. \\n\\nTo improve this paper, the authors are suggested to address the following issues:\\n1. Section 3.5 is not well organized. Besides, it is not mentioned what value the threshold hyper-parameter delta_m is set. \\n2. In page 9, \\\"MACs\\\" is missing in the sentence \\\"In Figure 3, ...the number of vs. the target...\\\"\"}", "{\"comment\": \"Hello,\\n\\nNice work that learns to select which features from a source model will help improve the fine-tuning performance on a target task! Indeed many times only few features contribute to the recognition in the target domain and others can even act as a confuser. \\n\\nIn structured sparsity section, you mention a few works that try to select which inference path to evaluate conditioned on the input. I thought (Veit and Belongie, ECCV 2018, http://openaccess.thecvf.com/content_ECCV_2018/papers/Andreas_Veit_Convolutional_Networks_with_ECCV_2018_paper.pdf) is also relevant for this list since they are learning to skip or process a layer (not channels) in Residual Networks without introducing architectural changes, and is tested on ImageNet. In this manner this can also be a coarser-grained alternative to \\\"channel skipping\\\" network proposed in this work.\", \"my_question_is\": \"Did you also try to do layer-wise skipping? It could be intuitive that some classes in the source domain may not even exist in the target domain (esp. for Birds-only and Dogs-only datasets) so than deciding to not process a layer on its entirety may yield even bigger drop in number of FLOPS, while maybe even increasing the accuracy?\", \"title\": \"Relevant work of Adaptive Inference Graphs ~ AIG (Veit and Belongie, ECCV 2018) ?\"}" ] }

[ "Vithursan Thangarasa", "Thomas Miconi", "Graham W. Taylor" ]

[ "continual learning", "catastrophic forgetting", "Hebbian learning", "synaptic plasticity", "neural networks" ]

{ "note_id": [ "aC4uAM2KZ", "Skxx5TqhoH", "r1gc-QI3oB", "SkxVhenjiB", "BJelLx3osB", "Hkl9s0sosB", "HyeEBaissH", "HJgxkaiijr", "ByeRgzjkcB", "H1gf5z-pKS", "r1xvrhmUdr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752972, 1573854599535, 1573835521874, 1573793963722, 1573793864331, 1573793442406, 1573793083595, 1573792983836, 1571955189652, 1571783305943, 1570286654887 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2593/Authors" ], [ "ICLR.cc/2020/Conference/Paper2593/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2593/Authors" ], [ "ICLR.cc/2020/Conference/Paper2593/Authors" ], [ "ICLR.cc/2020/Conference/Paper2593/Authors" ], [ "ICLR.cc/2020/Conference/Paper2593/Authors" ], [ "ICLR.cc/2020/Conference/Paper2593/Authors" ], [ "ICLR.cc/2020/Conference/Paper2593/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2593/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2593/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The reviewers agreed that this paper tackles an important problem, continual learning, with a method that is well motivated and interesting. The rebuttal was very helpful in terms of relating to other work. However, the empirical evaluation, while good, could be improved. In particular, it is not clear based on the evaluation to what extent more interesting continual learning problems can be tackled. We encourage the authors to continue pursuing this work.\", \"title\": \"Paper Decision\"}", "{\"title\": \"CLS methods in Related Work and Additional Experiments\", \"comment\": \"Thank you Reviewer 1 for the quick response. We also thank you for validating our observations on this set of class-incremental methods like iCaRL. We would like to acknowledge that in the related work section of the paper, we did include CLS theory inspired strategies based on pseudo-rehearsal, episodic/exact replay, and generative replay. This can be found in the line \\\"Our work draws inspiration from CLS theory which is a powerful computational framework for representing memories with a dual memory system via the neocortex and hippocampus. There have been numerous approaches based on CLS principles involving pseudo-rehersal (Robins, 1995; Ans et al., 2004; Atkinson et al., 2018), episodic replay (Lopez-Paz & Ranzato, 2017; Li & Hoiem, 2018) and generative replay (Shin et al., 2017; Wu et al., 2018).\\\" We also mention that in our work, we are mainly interested in neuroplasticity-inspired CLS techniques for alleviating the catastrophic forgetting problem, where the approaches involving slow and fast weights is analogous to properties of CLS theory. To make things more clear, we have added more formal descriptions of replay- or rehearsal-based techniques in the related work section, as well as an additional reference to iCaRL which would be considered a hybrid technique due to the use of a fixed external memory and a distillation step (that overlaps with the regularization category).\\n\\nIt also to be noted that on Split CIFAR-10/100, the DHP Softmax vastly outperforms fine-tuning and in some cases learning from scratch on each of the splits. Although SI alone outperforms DHP Softmax alone on every split except for the final split, we strongly believe that when DHP Softmax is used in-conjuction with SI or other task-specific synaptic consolidation methods, it will outperform SI on its own -- consistent with our other benchmark results in the paper on PermutedMNIST, Imbalanced PermutedMNIST, SplitMNIST and 5-Vision. We would also like to emphasize that the 5-Vision benchmark involves having to continually learn from 5 different computer vision datasets in a sequential manner. This can be a very challenging problem for neural networks with fixed capacity to have to tackle catastrophic forgetting, and DHP Softmax alone shows noticeable improvement, as well as in-conjunction with other synaptic consolidation methods. We also believe that this paper puts forth an interesting alternative direction of research for continual learning in the context of \\\"continual synaptic plasticity\\\" which involves the dynamic modification of individual synaptic connections in neural networks to help address the overcoming catastrophic forgetting problem.\"}", "{\"title\": \"Additional Experiments\", \"comment\": \"Your observation on this set of class-incremental methods like iCaRL is valid, and would be good to include in the paper integrated with the new text in the related work section. For Split CIFAR-10/100, overall SI performs very well across all splits after continual training, vastly outperforming DHP Softmax on nearly every split until the very end. On the final split, DHP Softmax, finetuning and SI all perform roughly the same. While I sympathise with the amount of time available during the rebuttal, I would like to see a stronger empirical evaluation. Alternative methods such as this work are important to develop, but given how they perform against or even in combination with other methods on these datasets, it remains unconvincing.\"}", "{\"title\": \"Response - additional experiments\", \"comment\": \"We thank Reviewer 1 very much for carefully reading our paper and positive comments on our method and experimental examinations. We answer your question regarding additional experiments on class-incremental learning below:\", \"re\": \"1. \\\"The authors are encouraged to evaluate their method with a large amount of classes, e.g. as done by iCaRL with CIFAR-100 and ILSVRC, with class-incremental training, to show if their method (a) scales (which I would anticipate, given the class-conditional Hebbian update) and (b) can deal with an alternative continual learning setting; a further resource is the work done around CORe50, which I would consider encapsulates more current thinking and practices around continual learning.\\\"\\n\\nWe would like to first emphasize that class-incremental learning methods such as iCaRL maintain exemplar patterns for each observed class and rehearse the model via distillation. Other techniques along these lines employ exact replay or generative replay. Exact replay strategies require storage of the data from previous tasks which are later replayed when learning new tasks. Generative replay strategies train a separate generative model to generate images to be replayed. For these reasons, these replay- or rehearsal-based techniques tend to perform very well in the class-incremental learning scenario. In our work, we focus on task-incremental continual classification, where tasks arrive in a batch-like fashion, and have clear boundaries (as examined through a large body of existing work in the recent review from De Lange et al., (2019) [1]). While we did not explicitly consider rehearsal techniques in our experiments for comparison, we attempt to address your question by evaluating DHP Softmax on the CIFAR-10/100 benchmark introduced by Zenke et al. (2017). Here, we compare our method against finetuning, training from scratch and synaptic intelligence (SI). DHP Softmax outperforms naive finetuning and even exhibits forward transfer of information, thus performing better than learning from scratch on some tasks. We added the additional experiments on Split CIFAR-10/100 to Appendix D of the paper. This experiment showcases the memory capabilities of DHP Softmax when learning new classes consecutively from the CIFAR-100 dataset. While we agree with Reviewer 1 that it would be interesting to evaluate our method with a large amount of classes (e.g., ILSVRC) or CORe50, we have yet to perform these experiments given the limited amount of time for experimentation during the rebuttal phase.\\n\\t\\t \\t \\t \\t\\t\\t\\n[1] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory G. Slabaugh, and Tinne Tuytelaars. Continual learning: A comparative study on how to defy forgetting in classification tasks. CoRR, abs/1909.08383, 2019\"}", "{\"title\": \"Response (part 2) - additional experiments and notation clarification\", \"comment\": \"Re: 2. \\\"The authors should try more experiments. They showed many MNIST variants and a 5-vision dataset. However, results in other papers in the literature are shown in much more sophisticated contexts (e.g. learning to play games). I understand this could exceed the computational capabilities of authors, but why, for example, not to try on a different dataset, as CIFAR? (as the Zenke et al paper did).\\\"\\n\\nWe agree that there have been other methods in the literature which have also been applied to reinforcement learning problems (e.g., learning to play games [1], maze exploration [2], and etc.). However, there are also a number of methods in the recent literature [3, 4, 5, 6] which have evaluated exclusively on the same vision benchmarks we considered (e.g., PermutedMNIST, SplitMNIST, 5-Vision Datasets Mixture). We would also like to emphasize that the 5-Vision benchmark involves having to sequentially learn from 5 different computer vision datasets. This can be a difficult problem for fixed capacity neural network architectures when having to tackle the catastrophic forgetting problem, and DHP Softmax shows noticeable improvement on its own and in-conjunction with other task-specific synaptic consolidation methods. \\n\\nWe did our best to address Reviewer 2\\u2019s concerns by evaluating DHP Softmax on the Split CIFAR-10/100 benchmark introduced by Zenke et al. (2017), which involves training the network on the full CIFAR-10 dataset and then sequentially training on 5 consecutive CIFAR-100 splits, each corresponding to 10 classes from the CIFAR-100 dataset. We followed the same experimental setup as [4] and used the CNN model of Zenke et al. (2017). Inspired by [4], we compare against: finetuning, training from scratch and synaptic intelligence (SI). We observe that DHP Softmax alone achieves an average test accuracy of 61.28% after learning all 6 tasks in the sequence, in comparison to 47.50% from finetuning. On some tasks DHP Softmax outperforms training from scratch, indicating that there is some forward transfer of information between certain tasks. Although SI performs the best, we wanted to demonstrate the memory retention capabilities of DHP Softmax similar to SplitMNIST experiments where DHP Softmax alone significantly outperformed finetuning. We included the additional Split CIFAR-10/100 experiments in Appendix D.\\n\\n[1] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, , and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In International Conference on Learning Representations (ICLR), 2019. \\n[2] Vincenzo Lomonaco, Karen Desai, Eugenio Culurciello, and Davide Maltoni. Continual reinforcement learning in 3d non-stationary environments. arXiv preprint arXiv:1905.10112, 2019. \\n[3] Rahaf Aljundi, Marcus Rohrbach, and Tinne Tuytelaars. Selfless sequential learning. In International Conference on Learning Representations (ICLR), 2019. \\n[4] Johannes von Oswald, Christian Henning, Joa \\u0303o Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. \\n[5] Hongjoon Ahn, Donggyu Lee, Sungmin Cha, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization. In Advances in Neural Information Processing Systems 32. 2019.\\n[6] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In Proceedings of the 36th International Conference on Machine Learning (ICML), pp. 3925\\u20133934, 2019.\", \"re\": \"3. \\\"On section 3, page 4 authors mention the number of labels d but that number appears undefined.\\\"\\n\\nWe defined $d$ as the number of outputs of the last layer and we have provided a comment on the updated manuscript to clarify this.\"}", "{\"title\": \"Response (part 1) - evaluation of CLS-based neuroplasticity methods on continual learning\", \"comment\": \"We appreciate the thoughtful questions from Reviewer 2, positive comments on our method and we are delighted to learn that Reviewer 2 feels that our paper is well written. We answer your questions below:\", \"re\": \"1. \\\"Authors show their method complement improvements due to task-specific approaches (which is great), but proper comparisons to other CLS-based approaches seem missing. In other words, I would hope a much better statement than \\u201dHowever, all of these methods were focused on rapid learning on simple tasks or meta-learning over a distribution of tasks or datasets. Furthermore, they did not examine learning a large number of new tasks while, alleviating catastrophic forgetting in continual learning\\\" Perhaps the authors may try find experimental cases where comparison to such approaches are possible, and show the empirical results? I am concerned the finetuned MLP baseline could be a bit weak.\\\"\\n\\nWe agree that this statement is vague so we added a more detailed explanation on the extension of CLS-based neuroplasticity techniques for continual learning at the end of Section 2. There are several works on neural network models with fast weights as outlined in our paper and to our knowledge, we believe we are the first to propose a slow and fast weights framework for continual learning through the Differentiable Hebbian Consolidation framework. \\n\\nOut of the past works on CLS, the one that is closest to our work is the Hebbian Softmax layer [1]. The Hebbian Softmax modifies the parameters by an interpolation between Hebbian-learning and SGD updates. The authors\\u2019 motivation was to achieve fast binding for rarer classes, especially in the language-modelling setting where they show strong results. Our original intuition was that this would not work well in continual learning setups because this approach relies on an engineered scheduling scheme for annealing between Hebbian plastic and slow fixed weights to enable slow and fast weight updates. The scheduling scheme is designed with a class counter which counts class occurrences and is used for an annealing function. The Hebbian Softmax also has a smoothing limit which is the number of class occurrences before switching completely to SGD updates. The issue here is that if the same classes are observed frequently during training on a task, as in the case for the established continual learning benchmarks (i.e., PermutedMNIST, SplitMNIST, 5-Vision and etc.), the algorithm anneals to mainly SGD updates. Thus, the effect of the fast weights memory storage becomes non-existent and Hebbian Softmax eventually performs like a traditional softmax (i.e., finetuning) in continual learning setups.\\n\\nWe confirmed our intuition by reimplementing the Hebbian Softmax and evaluating it on the same continual learning benchmarks considered for DHP Softmax. We found that DHP Softmax outperforms Hebbian Softmax in all of these benchmarks. Also, other methods such as meta networks [2, 3] combine fast weights based on non-trainable Hebbian learning with regular (slow) weights as well. However, their model was designed for meta-learning and performs very well on one-shot and few-shot learning benchmarks. We felt it was unfair to report and compare against fast weights models that were not designed for continual learning given some of their limitations. In addition to the Finetune baseline, we also evaluate EWC, SI and MAS without differentiable Hebbian plasticity as baselines. With a focus on continual learning, the goal of our work is to metalearn a local learning rule for the fast weights via the fixed (slow) weights and an SGD optimizer.\\n\\nWe have improved the manuscript and provided some more discussion on why other neuroplasticity techniques inspired from CLS theory were not considered as baselines in the continual learning benchmarks we experimented on. \\n\\n\\n[1] Jack W. Rae, Chris Dyer, Peter Dayan, and Timothy P. Lillicrap. Fast parametric learning with activation memorization. In Proceedings of the 35th International Conference on Machine Learning (ICML), pp. 4225\\u20134234, 2018.\\n[2] Tsendsuren Munkhdalai and Hong Yu. Meta networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), pp. 2554\\u20132563, 2017. \\n[3] Tsendsuren Munkhdalai and Adam Trischler. Metalearning with hebbian fast weights. CoRR, abs/1807.05076, 2018.\"}", "{\"title\": \"Response (part 2)\", \"comment\": \"Re: 4. \\\"In terms of motivation, can you explain why this Hebbian strategy is applied only to the final softmax?\\\"\\n\\nWe apply the Hebbian strategy to the final softmax layer because our model explores an optimization scheme where hidden activations from deep representations extracted by the network are accumulated directly into the softmax output layer weights when a class has been seen by the network. This results in better representations in the early stages of learning and can also retain these learned deep representations for a longer timescale. Task-level fast learning, enabled by a highly plastic weight component, are updated within the scope of each task. Between tasks, this plastic component decays to prevent interference. But selective consolidation into a stable component protects old memories, effectively modelling plasticity over a range of timescales to form a learned neural memory (see Section 4.1). In Figure 3, Section 4.1 we observe that the Frobenius norm of the $\\\\alpha$ grows, indicating that the network leverages the Hebbian traces as new tasks are learned in a sequential manner. Thus, by applying the Hebbian learning strategy to the final layer, we enable the network to consolidate a minimum number of synaptic weights while still being able to change enough synaptic weights if required by new task learning.\", \"re\": \"6. \\\"Are there any hyperparameters required for DHS Softmax? It seems to be no?\\\"\\n\\nThe only hyperparameter that DHP Softmax introduces is the initial value of the Hebbian learning rate $\\\\eta$, which also behaves as a decay term. $\\\\eta$ itself is a learned parameter of the network that is optimized by SGD. We spent little to no effort on tuning the initial value of this parameter and we perform a sensitivity analysis to show the effect of the initial value on the final average test accuracy after learning a sequence of tasks on the Permuted MNIST, Imbalanced Permuted MNIST and SplitMNIST continual learning benchmarks (see Appendix A.5).\"}", "{\"title\": \"Response (part 1) - terminology, notations, design decisions and misc clarifications\", \"comment\": \"Thank you Reviewer 3 for the positive comments about our approach and thoughtful questions. As you'll see below, they lead to a meaningful improvement in the framing of our paper. We have provided clarification on some terminology and notation, as well as more intuition on a couple of design decisions of our method. We answer your questions below:\", \"re\": \"3. \\\"Related to above, I am confused what is indexed by i,j in Eq 4. Compared to Eq 1, where theta only has one index (k), in Eq 4 theta has two indices (i,j).\\\"\\n\\nWe apologize for the confusion and want to clarify that the index $k$ in Eq. 1 and $i,j$ in Eq. 4 both refer to the synaptic weight. We used a linear index for a general parameter in Eq. 1 as that is the typical notation in the continual learning literature. However, when we began to talk about synaptic connections in Section 3, we made the change from $k$ to $i,j$ to explicitly represent the weight of the connection from neuron $i$ to neuron $j$. As mentioned in our response to (1), we did this to keep the notation consistent with past works on differentiable plasticity and Hebbian learning-based plasticity rules. Thus, at any time, the total, effective weight of the connection between neurons $i$ and $j$ is the sum of a fixed (slow) component ($\\\\theta_{i,j}$), plus a plastic (fast) component composed of: the Hebbian trace $Hebb_{i,j}$ multiplied by the plasticity coefficient $\\\\alpha_{i,j}$.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": [\"This paper tackles the problems of continual learning and catastrophic forgetting in neural networks. It uses methods inspired by Hebbia learning and complementary learning system (CLS), where there are slow and fast weights in the model.\", \"The paper addresses an important topic. I find the idea of fast/slow weights to be a refreshing and different approach compared to previous work on catastrophic forgetting.\", \"I had trouble following the details of the DHS Softmax, however. This may stem from my inexperience with Hebbian Learning, but here are some questions/suggestions:\", \"I am a bit confused exactly what is referred to as post/pre-synaptic connection, what is penultimate layer, etc. A figure might be helpful.\", \"It also might be helpful to write out an equation for the standard Softmax so it can be compared to Eq 2.\", \"Related to above, I am confused what is indexed by i,j in Eq 4. Compared to Eq 1, where theta only has one index (k), in Eq 4 theta has two indices (i,j).\", \"In terms of motivation, can you explain why this Hebbian strategy is applied only to the final softmax?\"], \"other_questions\": [\"Does it seem like DHS Softmax is not as strong by itself but works best in conjunction with others, such as EWC? I do not quite follow how they complement each other intuitively.\", \"Are there any hyperparameters required for DHS Softmax? It seems to be no?\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors introduce DIFFERENTIABLE HEBBIAN CONSOLIDATION,a new framework for continual learning that can be implemented in the usual differentiable programming setups. This framework is motivated in terms of complementary learning system (CLS) theory which features an episodic memory module. The method is shown to be easily implemented as seen in their pytorch pseudocode (authors also suggest code will be released). Additionally, authors show the method leads to significant improvements over simple baselines, and can complement other task-specific hebbian-based learning paradigms\\n\\n\\nAs an alien to the continual learning literature, my judgment is mostly positive. First, the paper is well-written, and the method is well-motivated in terms of cognitive principles with a nice commentary about underlying neurobiological substrates. The description of alternative methods seems exhaustive (but see below). The resulting algorithm is simple and doesn't seem to entail a significant computational burden. Importantly, it is differentially programing-friendly. Results show improvements over a baseline. However, I have two significant criticisms regarding the experimental section that I hope the authors will address.\\n\\n1)Authors comment on two paradigms to hebbian-based continual learning: the task-specific and theirs, based on CLS. Authors show their method complement improvements due to task-specific approaches (which is great), but proper comparisons to other CLS-based approaches seem missing. In other words, I would hope a much better statement than \\\"\\nHowever, all of these methods were focused on rapid learning on simple tasks or metalearning over a distribution of tasks or datasets. Furthermore, they did not examine learning a large\\nnumber of new tasks while, alleviating catastrophic forgetting in continual learning\\\" Perhaps the authors may try find experimental cases where comparison to such approaches are possible, and show the empirical results? I am concerned the finetuned MLP baseline could be a bit weak.\\n2)The authors should try more experiments. They showed many MNIST variants and a 5-vision dataset. However, results in other papers in the literature are shown in much more sophisticated contexts (e.g. learning to play games). I understand this could exceed the computational capabilities of authors, but why, for example, not to try on a different dataset, as CIFAR? (as the Zenke et al paper did).\", \"minor\": \"On section 3, page 4 authors mention the number of labels d but that number appears undefined\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper addresses the continual learning setting, and aims to mitigate catastrophic forgetting, with results on Permuted MNIST, Split MNIST, Vision Datasets Mixture, and their own class-imbalanced version of the Permuted MNIST dataset. The authors propose to augment differentiable plastic weights - a general neural network component - with class-specific updates (similarly to prior work, such as the Hebbian softmax) at the final layer of a neural network, prior to a softmax. While well-motivated in terms of the background and methodology (indeed, this is a simple way to prevent interference in fast weights), and nicely explored experimentally with lots of examinations into the workings of the method, the weak results on the simpler continual learning settings lead me to consider this a weak reject.\\n\\nThe authors show that fast weights can be applied in the continual learning setting, but alone they do not perform that well on the more challenging datasets, with mixed results on how much better they are as compared to a naive fine-tuning baseline, and they definitely lag behind synaptic consolidation methods. The authors' method combined with synaptic consolidation methods perform the best, but not much beyond the effect of the synaptic consolidation methods themselves. The authors are encouraged to evaluate their method with a large amount of classes, e.g. as done by iCaRL with CIFAR-100 and ILSVRC, with class-incremental training, to show if their method (a) scales (which I would anticipate, given the class-conditional Hebbian update) and (b) can deal with an alternative continual learning setting; a further resource is the work done around CORe50, which I would consider encapsulates more current thinking and practices around continual learning.\"}" ] }

[ "Fei Deng", "Zhuo Zhi", "Sungjin Ahn" ]

[ "parts", "objects", "model", "representation", "generative hierarchical models", "scenes hierarchical structure", "relationship", "scenes", "inherent structure" ]

{ "note_id": [ "2aqzQ-03qO", "SylZZNK2oB", "BkeYkNY3or", "HkgD6XKnoB", "SJgNVQY3iH", "SyxbW7tnjB", "SkgbAzY3jr", "rkxOuHUXcS", "rkgZtVmCtS", "HJxcv96sFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752941, 1573848057037, 1573848032794, 1573847998848, 1573847851583, 1573847801405, 1573847752849, 1572197743997, 1571857529047, 1571703393948 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2592/Authors" ], [ "ICLR.cc/2020/Conference/Paper2592/Authors" ], [ "ICLR.cc/2020/Conference/Paper2592/Authors" ], [ "ICLR.cc/2020/Conference/Paper2592/Authors" ], [ "ICLR.cc/2020/Conference/Paper2592/Authors" ], [ "ICLR.cc/2020/Conference/Paper2592/Authors" ], [ "ICLR.cc/2020/Conference/Paper2592/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2592/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2592/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors proposes a generative model with a hierarchy of latent variables corresponding to a scene, objects, and object parts.\\n\\nThe submission initially received low scores with 2 rejects and 1 weak reject. After the rebuttal, the paper was revised and improved, with significant portions of the paper completely rewritten (the description of the model was rewritten and a new experiment comparing the proposed model to SPAIR was added). While the reviewers acknowledged the improvement in the paper and accordingly adjusted their score upward, the paper is still not sufficiently strong enough to be accepted (it currently has 3 weak rejects).\", \"the_reviewer_expressed_the_following_concerns\": \"1. The experiments uses only a toy dataset that does not convincingly demonstrate the generalizability of the method to more realistic/varied scenarios. In particular, the reviewers voiced concern that the dataset is tailored to the proposed method\\n\\n2. Lack of comparisons with baseline methods such as AIR/SPAIR and other work on hierarchical generative models such as SPIRAL.\\nIn the revision, the author added an experiment comparing to SPAIR, so this is partially addressed. As a whole, the paper is still weak in experimental rigor. The authors argue that as their main contribution is the design and successful learning of a probabilistic scene graph representation, there is no need for ablation studies or to compare against baselines because their method \\\"can bring better compositionality, interpretability, transferability, and generalization\\\". This argument is unconvincing as in a scientific endeavor, the validity of such claims needs to be shown via empirical comparisons with prior work and ablation studies.\\n\\n3. Limited novelty\\nThe method is a fairly straightforward extension of SPAIR with another hierarchy layer. This would not be a concern if the experimental aspects of the work was stronger. \\n\\nThe AC agrees with the issues pointed by the reviewers. In addition, the initial presentation of the paper was very poor. While the paper has been improved, the changes are substantial (with the description of the method and intro almost entirely rewritten). Regardless, despite the improvements in writing, the paper is still not strong enough to be accepted. I would recommend the authors improve the evaluation and resubmit.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"Thank you for the helpful comments. We will respond in order.\\n\\n1) We believe the key to dealing with variable number of objects is the z_pres variable. It can turn on/off individual object representations. Although we use parallel inference, we still have the z_pres variable, so we do not require the number of objects to be fixed. Also, our datasets do contain variable number of objects. We have clarified this in the first paragraph of Section 3.3. We conjecture that our top-down inference approach enables the parallel inference to perform well, since higher-level appearance information can guide lower-level decomposition. Please also kindly refer to general response [A1] - [A2] for the main contribution and novelty of our paper.\\n\\n2) We have updated the paper to clarify our representation of the compositional hierarchy. As shown in Figure 1A, the compositional hierarchy is represented as a latent tree. Both the tree structure and the latent variables are inferred for each static image in the dataset. One way to visualize the tree is to draw bounding boxes around objects and their constituent parts according to the inferred tree structure and pose vectors. Figure 1B illustrates this, and our qualitative results should be interpreted in the same way. Since we recursively apply spatial transformer during inference, we have the implicit constraint that parts should be within the object bounding box.\\n\\n3) Regarding empirical evaluation and ablations / baselines, please kindly refer to general response [A5] - [A8].\\n\\n4) Please kindly refer to general response [A10] - [A11] for a brief comparison.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thank you for your constructive review and for taking the time to provide a reference list. We will respond in order.\\n\\n1a) We agree that it is important to connect to practical applications. However, this line of research has not been close to the level of applying to robotics or real-world datasets. This is due to the challenging problem setting: (i) We use generative modeling, which learns both representation and rendering process; (ii) Our latent representation is hierarchical, compositional, and interpretable at the same time, which no prior work has achieved; (iii) Our model is end-to-end trainable via purely unsupervised learning. In our revised paper, we have introduced a compositional MNIST dataset that presents an additional challenge of shape variation compared to our 2D and 3D datasets. We think this is one characteristic of more realistic data, and we have shown promising results.\\n\\n1b) Regarding empirical evaluation and ablations / baselines, please kindly refer to general response [A5] - [A8].\\n\\n2) Thanks for the suggestion. We have revised our paper to better highlight the significance. Please also refer to general response [A1] - [A2] for a summary of the contribution and novelty of our paper.\\n\\n3) We totally agree and have included more discussion on related work. Please kindly refer to general response [A10] - [A13] for a brief comparison.\\n\\n4a) Thanks for the suggestion. We have updated the notation and equations to make them more accessible. We have also added more intuitive explanation and illustrations to clarify the formulation and contribution of our model.\\n\\n4b) We have included a dedicated paragraph in Section 3.3 to describe the memory. In short, memory is able to capture the canonical pose, which is essential to the scene graph that describes composition based on pose transformations. The learned memory templates also bring additional interpretability to our model.\"}", "{\"title\": \"Response to Review #4\", \"comment\": \"Thank you for the constructive review. We will respond in order.\\n\\n-- Experiments --\", \"our_problem_setting_is_already_challenging\": \"(i) We use generative modeling, which learns both representation and rendering process; (ii) Our latent representation is hierarchical, compositional, and interpretable at the same time, which no prior work has achieved; (iii) Our model is end-to-end trainable via purely unsupervised learning. We believe that research under such problem setting has not been close to the level of applying to real-world datasets. We also would like to note that our datasets, though manually constructed, present challenges of multi-scale and multi-pose entities with occlusion and compositional ambiguity. Specifically, in our 3D dataset, there is occlusion between parts, so parts are not well-separated. In our revised paper, we have shown promising results on a compositional MNIST dataset that presents an additional challenge of shape variation.\\n\\nRegarding empirical evaluation and ablations / baselines, please kindly refer to general response [A5] - [A8]. We have actually provided NLL, counting error, precision, and recall as evaluation metrics. We appreciate your suggestion to evaluate performance as the dataset parameters vary. In the generalization experiments, we have already tested performance on different number of objects.\\n\\n-- Significance --\\nThanks for your suggestion. We have included experiments on downstream tasks in our revised paper. We show that compared to SPAIR, our proposed hierarchical representation approximately doubles the data efficiency on these downstream tasks. Please also refer to general response [A1] - [A2] for the contribution and novelty of our paper.\\n\\n-- Presentation --\\nThanks for the feedback. We have improved our presentation quality and included more discussion on related work. Please kindly refer to general response [A10] for a brief comparison with Xu et al.\"}", "{\"title\": \"General Response to All Reviewers (Part 3 / 3)\", \"comment\": \"3) Incomplete overview of related work.\\n\\n[A9] We agree that we need to discuss more related work, and we thank reviewers for taking the time to list relevant references. Here we briefly compare our model with those pointed by reviewers.\\n\\n[A10] Xu et al [1] focus on motion decomposition from image pairs. They obtain the part hierarchy based on consistency of motion. The latent space of their model contains only the representation for motion. The part appearance and the hierarchical structure are not modeled as latent variables. Also, this latent space does not seem interpretable. In contrast, we model both the hierarchical structure and the object pose and appearance as latent variables. This richer latent representation is learned from single images, i.e., by merely observing the possible static compositions with no motion information.\\n\\n[A11] Sharma et al [2] and Ganin et al [3] learn to parse an image into a program that, when fed to a rendering engine, can generate the image. Their models lack the explicit representation of compositional relationships among primitives. Also, the primitives are pre-defined by the rendering engine. In contrast, our model automatically discovers the primitives and can learn the renderer.\\n\\n[A12] Tulsiani et al [4] and Li et al [5] operate directly in 3D space, which is quite a different problem setting. We appreciate the symmetry modeling in [5], which is complementary to our model. However, we note that neither of these two models has a hierarchically structured latent space.\\n\\n[A13] Kosiorek et al [6] extend AIR to model image sequences. We find the approach inspiring, and also consider extending our model to the sequential setting as an interesting future direction.\\n\\n4) Suboptimal presentation quality.\\n\\n[A14] We have greatly improved our writing. Specifically, we have updated the notation and equations to make them more accessible. We have added more intuitive explanation and illustrations to clarify the formulation and contribution of our model. To make our paper more self-contained, we have also included a paragraph that describes the memory. We sincerely hope that reviewers could take the time to read our paper again, and we would appreciate any additional feedback.\\n\\n\\n[1] Zhenjia Xu, Zhijian Liu, Chen Sun, Kevin Murphy, William T. Freeman, Joshua B. Tenenbaum, Jiajun Wu. Unsupervised Discovery of Parts, Structure, and Dynamics. ICLR 2019.\\n[2] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, Subhransu Maji. CSGNet: Neural Shape Parser for Constructive Solid Geometry. CVPR 2018.\\n[3] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S.M. Ali Eslami, Oriol Vinyals. Synthesizing Programs for Images using Reinforced Adversarial Learning.\\n[4] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. CVPR 2017.\\n[5] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, Leonidas Guibas. GRASS: Generative Recursive Autoencoders for Shape Structures. SIGGRAPH 2017.\\n[6] Adam R. Kosiorek, Hyunjik Kim, Ingmar Posner, Yee Whye Teh. Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects. NeurIPS 2018.\"}", "{\"title\": \"General Response to All Reviewers (Part 2 / 3)\", \"comment\": \"2) Insufficient empirical evaluation: ablations / better baselines are needed.\\n\\n[A5] As we have explained in [A2], our main contribution is the design and successful learning of a probabilistic scene graph representation, which no prior work has accomplished. The main purpose of learning such representation is not to improve detection performance. Instead, the representation itself is of significance, because it can bring better compositionality, interpretability, transferability, and generalization ability. Therefore, in the experiments, we have focused on showing that the representation can be learned properly by our proposed methods, and that the learned representation does bring some of the benefits. We do not agree that lack of comparison with baselines on detection performance can be grounds for insufficient empirical evaluation.\\n\\n[A6] To be more specific, the counting error, precision, and recall are mainly to show that the representation is learned properly, i.e., the nodes in the latent tree associate correctly with the parts and objects. Together with the qualitative results, it can be seen that the compositional hierarchy is well captured by our model, even in the presence of occlusion and compositional ambiguity. In terms of compositionality and interpretability, we are not aware of any baseline that is at the same level as our model. Taking SPAIR as an example, it is clear that SPAIR cannot simultaneously detect parts and objects, let alone model the part-object composition. Besides, SPAIR lacks the discrete interpretable representation that our model provides to capture the identity and canonical pose for individual parts. In terms of transferability, our new experiments on downstream tasks show that the proposed model improves data efficiency over SPAIR.\\n\\n[A7] Regarding the negative log-likelihood (NLL), we include it mainly to show that while the probabilistic scene graph representation brings so many benefits as mentioned above, it can be obtained without impairing NLL or generation quality. This is not trivial as it may appear, because compositionality and interpretability introduce constraints on the model structure and discreteness in the latent variables. Hence, we are faced with a more difficult optimization problem in a more limited model space, compared to non-hierarchical, continuous representations. Therefore, we do not consider improving NLL as our main goal, though our model seems to have achieved better NLL as a side effect. We also believe that AIR and SPAIR would have worse generation quality, because they need to set a very low success probability for the prior of z_pres to encourage sparsity. This can often lead to the generated image being all black, i.e., containing no object. In contrast, our model learns this prior conditioning on higher-level z_appr, and thus avoid this problem.\\n\\n[A8] Similar arguments apply to ablations. Our main contribution is not to improve detection performance or NLL, but to obtain the probabilistic scene graph representation. Hence, we feel it unnecessary to compare detection performance or NLL with ablations that break the scene graph representation. In this sense, the only ablation that may seem reasonable is to remove the memory that learns primitive templates. However, we believe that memory is not redundant. The scene graph relies on the relative pose between entities to describe composition. To obtain the relative pose, the model needs representation of the canonical pose. This is achieved by the memory, as we have shown in the qualitative results. Without memory, there will be no notion of canonical pose, leaving the scene graph not well defined.\"}", "{\"title\": \"General Response to All Reviewers (Part 1 / 3)\", \"comment\": \"We would like to thank all reviewers for their helpful comments and for pointing to relevant literature. Here we address the main concerns shared by all reviewers.\\n\\n1) Limited significance: the proposed method just adds another layer to SPAIR and lacks novelty; hence, experiments on more realistic data / downstream tasks are needed to demonstrate significance.\\n\\n[A1] We have uploaded a revised version of our paper, where we highlight our contributions. We aim to learn interpretable representation for compositional hierarchies. To this end, we propose a probabilistic scene graph representation that describes the compositional hierarchy as a latent tree. The nodes in the tree represent the appearance of entities, and the edges specify the relative pose at each level of composition. To ensure proper learning, we design a decoder for the probabilistic scene graph that closely follows the rendering process in computer graphics. We also present a top-down inference approach that is able to simultaneously infer both the tree structure and the latent appearance and pose vectors.\\n\\n[A2] We believe that our generative modeling of the compositional hierarchy is novel, and that obtaining such interpretable representation from raw pixels through end-to-end purely unsupervised learning is highly nontrivial. Also, our model is fully general and does not have to be limited to three levels. We do not think simple extensions of SPAIR could achieve what we have done, nor are we aware of any prior work that is able to learn hierarchical, compositional, and interpretable latent representation at the same time. Therefore, we believe that our design of the probabilistic scene graph representation and our proposed inference and learning methods constitute significant contributions.\\n\\n[A3] We agree that experiments on more realistic data will add to the significance of our paper. Since our 2D and 3D shape datasets already contain multi-scale and multi-pose entities with occlusion and compositional ambiguity, we think one way to make a more realistic dataset is to include deformable primitives. Hence, in the appendix of our revised paper, we introduce a compositional MNIST dataset where the primitives are handwritten digits with considerable shape variation. We show that with slight modification, our model can successfully deal with such deformable primitives.\\n\\n[A4] We also agree that the benefit of our proposed hierarchical representation would be better demonstrated on downstream tasks. We have included two such tasks in our revised paper. For the 2D and 3D shape datasets, we make a label for each image, and train a classifier on top of the learned representation to predict the label. The label is obtained by first counting the number of distinct parts within each object, and then summing the count over all objects. Similarly, for the compositional MNIST dataset, we seek to predict the sum of max digit within each object. We use SPAIR as a non-hierarchical baseline, and tune it to detect parts. Experimental results show that compared to SPAIR, our proposed hierarchical representation approximately doubles the data efficiency on these downstream tasks.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"-- Update --\\nThank you for the detailed response and updates. The quality has improved but as mentioned in the original review, it would be good to see results on a more significant non-synthetic task before I would argue for acceptance.\\n\\n------------------\\nThe authors propose a generative model with a hierarchy of latent variables corresponding to a scene, objects, and object parts. The method is evaluated on a synthetic dataset with manual inspection, MSE, and object counting performance. The method can form bounding boxes around objects and parts in the examples that are shown, and the MSE is similar to that of a VAE.\\n\\nThe idea of decomposing a probabilistic model hierarchically is potentially interesting, but this paper has drawbacks in terms of experimental quality, significance, and presentation.\\n\\n-- Experiments -- \\nThe experiments are done on a manually constructed synthetic dataset, so it is not clear whether the proposed method would work in more realistic or more challenging settings. For instance, the dataset was constructed using a recursive process which probably does not resemble a realistic distribution of objects and their parts, and the parts are well-separated within the object. \\n\\nThe experiments on the synthetic dataset are not thorough, and could be more interesting with ablations, evaluating performance as the dataset parameters vary (e.g. number of objects, amount of occlusion), better baselines, and an evaluation metric rather than manual inspection.\\n\\n-- Significance --\\nWith respect to significance, the formulation is fairly straightforward and builds off of previous techniques introduced in e.g. SPAIR. The lack of demonstrating usefulness on a downstream application limits this paper's significance; for instance the SPAIR paper used the generative model as a front-end for a card game and Atari game. The paper should more clearly demonstrate the benefits/tradeoffs of the hierarchical aspect.\\n\\n-- Presentation --\\nThe paper leaves many terms undefined so the paper is not self contained, e.g. the 'problem of routing', 'memory buffer', 'memory template', 'variational memory addressing', 'spatial transformer' (missing a reference here). I suspect the related work is missing references, e.g. \\\"Unsupervised Discovery of Parts, Structure, and Dynamics\\\" Xu et al ICLR 2019. Finally, the paper should be proof-read for grammatical errors.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Update: I thank the auhors for the detailed response. The updated paper does look more convincing, but my main concern remains - all considered datasets and tasks are synthetic and specifically made for the method. I think experiments on more real data would be crucial to show the potential of the method. Some examples I can imagine would be part segmentation in images or 3D models, or robotic control.\\n\\n---\\n\\nThe paper proposes an unsupervised approach to learning objects and their parts from images. The method is based on the \\\"Attend, Infer, Repeat\\\" (AIR) line of work and adds a new hierarchy level to the approach, corresponding to object parts. The method is evaluated on two custom synthetic dataset: one composed of simple 2D geometric shapes and another one with 3D geometric shapes. On these datasets the method successfully infers the object-part structure and can parse and reconstruct provided images.\\n\\nWhile the general topic of the paper is interesting, I do not think it is fit for publication. First and foremost, the experiments are very incomplete: the method is only evaluated on two custom synthetic datasets and not compared against any baselines or ablated versions of the method. Moreover, the proposed approach seems like a relatively minor modification of SPAIR (Crawford & Pineau).\", \"pros\": \"1) The topic is interesting and the method generally makes sense.\\n2) The method works on the tasks studied in the paper, including relatively challenging scenarios with 3D occlusions or ambiguity in assigning parts to objects.\\n3) The paper shows generalization to a number of objects different from that seen during training.\", \"cons\": \"1) Experiments are limited:\\n1a) The method is evaluated on custom and relatively toy datasets. It is unclear if it would apply to more practical situations. While I agree that at the high level the question of decomposition of object into parts is interesting, it is still important to connect research in this direction to potential practical applications. Where could the proposed method be used? Perhaps in some control tasks, such as robotics? Could it be applied to more realistic data, such as for instance ShapeNet objects?\\n1b) There are no comparisons to baselines. One might argue that baselines do not exist since there are no methods addressing the same task. Generally, it is the job of the authors to come up with relevant baselines to show that the proposed model actually improves upon some simpler methods. One variant of obtaining baselines is via ablating the proposed model. Another is by taking prior approaches, for instance object-based ones without the part decomposition, and showing that the part decomposition allows to improve upon those in some sense.\\n\\n2) The novelty is somewhat limited: the method seems like a relatively straightforward extension of SPAIR by adding another hierarchy layer. It might be sufficient if the experimental results would be strong, but given that they are not, becomes somewhat concerning. If the method does indeed include significant technical innovation, it might be helpful to better highlight it.\\n\\n3) The related work overview is extremly limited. It is authors' job to provide a comprehensive ovreview of prior literature and I cannot do it for them here, but below are a few papers that come to mind. This is by no means a complete list.\\n\\n[1] Zhenjia Xu, Zhijian Liu, Chen Sun, Kevin Murphy, William T. Freeman, Joshua B. Tenenbaum, Jiajun Wu. Unsupervised Discovery of Parts, Structure, and Dynamics. ICLR 2019\\n[2] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. CVPR 2017\\n[3] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, Leonidas Guibas. GRASS: Generative Recursive Autoencoders for Shape Structures. SIGGRAPH 2017.\\n[4] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, Subhransu Maji. CSGNet: Neural Shape Parser for Constructive Solid Geometry. CVPR 2018.\\n[5] Adam R. Kosiorek, Hyunjik Kim, Ingmar Posner, Yee Whye Teh. Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects. NeurIPS 2018.\\n\\n4) Presentation is at times suboptimal:\\n4a) It would be very helpful to have more visuals and intuitions about the functioning of the method, as opposed to equations. Equations are definitely good to have, but they are not the easiest to parse, especially by those not intimately familiar with this specific line of work. \\n4b) It is quite unclear to me how and why is the memory used in the model. If this is described in another paper, it would be useful to point there, but still briefly summarize in this paper to make it self-contained.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Contributions: this submission proposes a generative framework which employs a hierarchical decomposition of scene, objects and parts. It modifies the SPAIR framework (Crawford & Pineau) and replaces the recurrent generation with parallel generation, by assuming the number of objects is known and fixed. Results are presented on two synthetic datasets with 2D or 3D shapes.\", \"assessment\": \"- The proposed model and learning framework are closely related to previous work (AIR, SPAIR). The authors replaced the \\\"sequential processing ... with single forward pass\\\", which requires the number of objects to be fixed.\\n- In the proposed model, there is no constraint on the part-object hierarchy, and it is unclear whether it can learn to discover such hierarchies directly by reconstructing single static images. It is also not evaluated in the submission.\\n- There are no comparisons with AIR or SPAIR on the two datasets the authors created.\\n- The authors are recommended to compare with previous work on hierarchical generative modelings with objects and parts, such as (Xu et al, ICLR 2019), and other related work, such as SPIRAL (Ganin et al.)\\n\\nDue to the limited contribution, lack of comparison with related work and limited empirical evaluation, I recommend rejection of the submission.\\n\\n[1] Xu el al, Unsupervised Discovery of Parts, Structure, and Dynamics, ICLR 2019.\\n[2] Ganin et al., Synthesizing Programs for Images using Reinforced Adversarial Learning.\\n\\n\\n--------------------------------\", \"post_rebuttal\": \"Thank you for your detailed answers to my questions and updated manuscript. The writing has indeed significantly improved and some questions (e.g. varying number of objects) have been addressed. After reading the rebuttal, my concerns remain that: (1) lack of empirical evaluation on more complex real / synthetic datasets; (2) it is still unclear to me how such hierarchy is inferred from single images.\"}" ] }

[ "Pranav Poduval", "Hrushikesh Loya", "Rajat Patel", "Sumit Jain" ]

[ "uncertainty estimation", "Deep Ensembles", "Adverserial Robustness" ]

{ "note_id": [ "o7Wu2AUY2i", "HkempcshjH", "ryg67R53jH", "ByewHtY3sH", "B1xUVtd3iH", "BJgwyrd3sr", "B1e78-vhoB", "B1eRIM82oH", "H1eQqhE2oB", "Bye1noEnsB", "S1xucsSojS", "rJgAvsSosB", "SJlgDm4iiB", "Hygc1jVwor", "Syg9jcy6cH", "HJee9LcZqS", "rkeat7C6FS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752911, 1573857978911, 1573854757011, 1573849406815, 1573845293600, 1573844190630, 1573839178530, 1573835349554, 1573829771192, 1573829543015, 1573768080061, 1573768037706, 1573761879616, 1573501666511, 1572825761992, 1572083335576, 1571836805469 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/Authors" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2591/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes to use mixture distributions to improve uncertainty estimates in BNNs. Ensemble methods are interpreted as a Bayesian mixture posterior approximation. To reduce the computation, a modification to BBB is provided based on a concrete mixture distribution.\\n\\nBoth R1 and R3 have given useful feedback. It is clear that interpretation of ensemble as a Bayesian posterior is well known, and some of them also have theoretical issues. The experiment to clearly comparing proposed mixture posterior to more commonly used mixture distribution is also necessary. \\n\\nDue to these reasons, I recommend to reject this paper. I encourage the authors to use reviewers feedback to improve the paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to clarification\", \"comment\": \"\\\"Since in VI framework all we need are samples that seem to come from the true posterior.\\\"\\nThis statement is somewhat incorrect - although it is the case that very broadly you only need to come up with an approximation distribution that is \\\"close\\\" to the true posterior, it's not clear if the solution to the optimization problem itself is reasonable in any way.\\n\\n\\\"...rather than minimization of the negative log evidence\\\" Instead, I believe you'd need to be minimizing KL(posterior || g) and would have to justify everything in that sense.\\n\\n\\\"But due to the nature of the bound, the bias will decrease to zero ...\\\" Would it be possible to show this numerically - e.g. via a set of Gaussians?\"}", "{\"title\": \"General Comments for all reviewers\", \"comment\": \"We have extended the proof of Ensembles as described in [1] being bayesian to a more general setting where the architectures of the members in the Ensembles need not be the same, view Appendix 3 for the details.\\n\\nWe have open-sourced the code for two experiments, the uncertainty plots of MoG in case of adversarial attack one, and MoG Reset on CIFAR. We have updated the results of the CIFAR experiment after running each network for 200 epoch and also added the Expected Calibration Error comparison of CMoG Bayesian NN, MoG Bayesian NN (keeping the p's fixed in MoG) and Deep Ensemble.\\n\\n\\n\\n\\n[1] Balaji Lakshminarayanan, Alexander Pritzel, Charles Blundell, NeurIPS 2017 Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles\"}", "{\"title\": \"Do consider ensembles as Bayesian a part of the novelty in our work\", \"comment\": \"\\\"If this is the case, I think this novelty is not significant since it is a straightforward combination.\\\"\\n\\nDo consider our work in linking Deep Ensembles to the Variational Framework as a part of the novelty. We have also successfully extended our proof to an Ensemble of NN's with arbitrary architectures using a slight trick, view Appendix 3 for details, which to the best of our knowledge is novel. We have also argued in Section 3 of our papers ensembles of BBB, dropout too can be seen as a Bayesian approximation which as per our citation [3] is not well known either.\\n\\n\\n[3] Remus Pop and Patric Fulop - Deep Ensemble Bayesian Active Learning: Addressing the Mode Collapse issue in Monte Carlo dropout via Ensembles https://openreview.net/forum?id=Byx93sC9tm\"}", "{\"title\": \"Thank you for the clarifications\", \"comment\": \"Please correct me. Otherwise, I assume q(w_{-K})q(w_K) is the combination of \\\"Convolutional Concrete Dropout and Mixture Distribution\\\". If this is the case, I think this novelty is not significant since it is a straightforward combination.\\n\\n\\n\\\" a categorical mixture distribution will give equal weight\\\"\\nPlease note that q(w) = \\\\int q(z) q(w|z) dz\\nI assume that q(z) is held fixed when q(z) is a categorical distribution. \\nAs mentioned in my initial review, you can *jointly update* the mixing weight q(z) and each component q(w|z) even when q(z) is a categorical distribution instead of a concrete distribution. In the references cited in my initial review, q(z) can be learned even when q(z) is a categorical distribution. \\nI still don't think that this paper can be accepted unless you show that a concrete mixture distribution is better than a mixture distribution with a learned categorical mixture prior.\"}", "{\"title\": \"Clarifications\", \"comment\": \"Indeed our method is a combination of Convolutional Concrete Dropout (for pre-final layers) and Mixture Distribution Variational Inference (for final layers). And the factorization mentioned by you is indeed correct. Thus making it both lightweight and flexible, allowing it to outperform Deep Ensembles that have approximately 3 times its space complexity.\\n\\n\\n\\\"You should show a concrete mixture distribution is better than a categorical mixture distribution. For example, compared to the categorical mixture\\\"\\nTrue a comparison is necessary, we will be adding a comparison of their expected calibration error to the uploaded version of the paper, along with a discussion of the cost.\\n\\nAnother advantage of our method is in Concrete Ensembles, specifically in the case when the architectures of the NN's in the ensemble are not the same. When dealing with Ensembles of NN's like Alexnet and Resnet (architecture of the ensemble members may be different) we know Resnet's are superior and deserve more weight. But a categorical mixture distribution will give equal weight to each member of the ensemble, and here using concrete distribution to learn the importance of each member will become an absolute must.\\n\\nWe are aware there are other alternatives to finding the importance of NN's in an ensemble the most popular being their weighted accuracy on the validation set, but our method of using a concrete mixture distribution provides a principled Bayesian approach for the same.\"}", "{\"title\": \"Request for further clarifications\", \"comment\": \"\\\" another key novelty in our method is using Learnable Dropouts in the Convolutional Network and performing full Bayesian Inference with mixture distribution on only the final layer\\\"\\n\\nIf my understanding is correct, p(w_{-K}, w_K) is parameterized by a NN, where w_K is the weight of the final layer. The authors would like to use a variational distribution q(w_K) such as a Gaussian mixture to perform inference. Is my understanding correct? If this is the case, we can use a variational distribution q(w)=q(w_{-K})q(w_K) where q(w_{-K}) is a dropout distribution and q(w_K) is a Gaussian mixture with a categorical prior. This variational distribution is also \\\"easy to train and scalable to large scale Neural Networks.\\\" Furthermore, we also learn \\\"Dropouts in the Convolutional Network\\\" using q(w_{-K}), and perform \\\"Bayesian Inference with mixture distribution on only the final layer\\\"using q(w_K).\", \"bottom_line\": \"You should show a concrete mixture distribution is better than a categorical mixture distribution. For example, compared to the categorical mixture, the concrete mixture distribution is faster to train while it gives similar performance. You should tell readers the advantages and disadvantages of using the concrete mixture distribution.\"}", "{\"title\": \"Further clarification\", \"comment\": \"\\\"it's still a bit muddled and unclear. I'm not convinced that I fully buy the argument that if you alternately? optimize lower and upper bounds on the KL you will be provably optimizing a lower bound on the log evidence\\\"\\n\\nUnderstanding our proof is much easier if we view training a BNN as minimizing the KL divergence between an \\\"assumed posterior\\\" $q(w|\\\\theta)$ and the true posterior, rather than minimization of the negative log evidence. Since in VI framework all we need are samples that seem to come from the true posterior.\\n\\nThe upper and lower bounds are derived for the variational free energy w.r.t the mixture distribution (i.e. KL divergence between the mixture distribution and the \\\"true posterior\\\" ) and due to the interesting structure of the bounds training a single BNN i.e. minimizing the KL divergence between the individual component of mixture and the true posterior we will be minimizing the Upper and Lower bounds simultaneously ( since both upper and lower bounds differ only by a constant entropy term). Thus as the KL divergence between the mixture distribution and the true posterior is sandwiched between the upper and lower bound, it too would be minimized. Thus we are training a \\\"hypothetical\\\" BNN with mixture distribution posterior by training an ensemble of BNN's.\\n\\nWe understand your concern is that the optimal solution in both the cases, (training a mixture distribution BNN or an ensemble of BNN) will not be the same which is true and that there will be a bias. But due to the nature of the bound, the bias will decrease to zero as the size of the ensemble size grows to infinity ( since $\\\\lim_{N\\\\to\\\\infty} H(\\\\frac{1}{N}) = 0$ ). Even when dealing with finite-sized ensemble the bias will be bounded.\\n\\n\\\"I'd suggest looking back towards the rich statistics literature on the interface of the two methods rather than the modern machine learning literature\\\"\\n\\nThank you for the suggestion we shall look into it.\"}", "{\"title\": \"Thank you for the further clarifications\", \"comment\": \"It's interesting that you __can__ make the MC Dropout approximate posterior work by adding in the very tiny amount of noise. However, that's very much just a mathematical fix and not a real solution to the problem, which is that the optimized solution is a quasi-discrepancy instead.\\n\\nI'd encourage the authors to justify further why they would wish to be Bayesian in the first place. After all, ensembles themselves provide a relatively principled way of reasoning about uncertainty anyways. I think that justifying this would also make the paper stronger.\"}", "{\"title\": \"Thank you for the clarifications\", \"comment\": \"Performance issues: Thank you for pointing out that you were only able to train for 80 epochs. Until this issue is fixed completely, I still don't think that this paper can be accepted. If computational issues are still the concern, I'd look into advanced learning rate schedules (such as super-convergence https://arxiv.org/abs/1708.07120) to speed up the training procedure.\", \"ensembles_being_bayesian\": \"In general, I think that yes, Bayesian interpretations of ensembles are potentially a fruitful avenue of research. However, in my personal (and potentially incorrect) view, I'd suggest looking back towards the rich statistics literature on the interface of the two methods rather than the modern machine learning literature.\", \"bound\": \"Thank you for the clarification - it's still a bit muddled and unclear. I'm not convinced that I fully buy the argument that if you alternately? optimize lower and upper bounds on the KL you will be provably optimizing a lower bound on the log evidence.\"}", "{\"title\": \"Continuation\", \"comment\": \"\\u201cFrom a quick pass through Appendix A.2, the derivation is quite unstructured and difficult to follow. It is tough to see exactly what is being meant by the probabilities in the first line. \\nAdditionally, the phrasing of \\u201cWe further lower bound\\u2026\\u201d seems to suggest that the bound being used is not the well known bound of Durrieu et al, 2012, but a looser bound. If so, what is the approximation accuracy of this bound in relation to the other bounds compared against in this setting?\\u201d, \\n\\n\\nWe apologise for the lack of clarity. Our goal was to show that when training an ensemble of Bayesian NN\\u2019s like BBB or Dropouts we are actually minimizing the variational free energy of a hypothetical Bayesian NN with mixture distribution. The lower bound derived by us is achieved by further lower bounding Durrieu et al. popular approximation to KL divergence between mixture distributions called the Variational Approximation. The lower bound is achieved when the pairwise KL divergence between individual components of the mixture distribution is infinite (i.e. the individual components in the mixture are far apart). The key take away from this was that both the upper and lower bounds of the variational free energy with respect to the mixture distributions had the same form, summation of the variational free energy with respect to individual components in the mixture with only a difference of a constant entropy term $H(\\\\frac{1}{N})$ (N is the number of components in the mixture). This implies that by training an ensemble of BNN\\u2019s individually by minimizing their individual variational free energy we are training a hypothetical BNN with mixture distribution (sandwich theorem).\\n\\n \\u201c Furthermore, in Appendix 2, I see a potential issue in using the KL divergence and would like the authors to clarify.\\nIf the secondary lower bound is being used to approximate the KL divergence between the mixture distributions, it becomes very unclear if what is being optimized is, in fact, a __lower bound__ on the log probability. After all, the standard ELBO is written as:\\n\\tLog p(y) \\\\geq ELBO := E_q(\\\\log p(y | \\\\theta)) \\u2013 KL(q(\\\\theta) || p(\\\\theta)) \\nImplying that another lower bound on the KL would be greater than the ELBO and thus not necessarily a lower bound on the log marginal likelihood.\\u201d\\n\\nWe like to view training a BNN as minimizing the $KL(Q(W||\\\\theta)||P(W||D))$ rather than the perspective of it being an upper bound of the -\\\\log p(D). So all we are trying to do in the Appendix 2 is find an Upper and Lower Bound for $KL(Q(W||\\\\theta)||P(W||D))$, where $Q(W||\\\\theta)$ corresponds to a mixture distribution parameterized by $\\\\theta_1,\\\\theta_2, \\u2026,\\\\theta_N$. The interesting nature of the derived bounds suggests that individual training of BNN by minimizing $KL(Q(W||\\\\theta_i)||P(W||D))$ for i $\\\\subseteq$ [1 to N] also minimizes the $KL(Q(W||\\\\theta)||P(W||D))$, which would imply that we are also minimizing the \\\\log p(y) too as $KL(Q(W||\\\\theta)||P(W||D))$ is it\\u2019s upper bound. \\n\\n\\u201cThe posterior is used only for fully connected layers\\u2026\\u201d Why is this the case? Alternative Bayesian methods can be used on the full network\\u201d\\n\\nOur method too can be used on the full network and as mentioned earlier in a much more general way (beyond MoG). The specific setting we use is a simple trick for providing better uncertainty estimates than deep ensembles even with relatively less no. of parameters. The learnable dropout parameter (Concrete Dropout) on the convolutional layers has negligible cost and a full Bayesian Inference on only the final layer reduces the additional cost of our BNN to negligible. Thus at negligible parametric overhead can beat Ensembles which usually have a huge memory overhead, we believe this is particularly useful in resource-constrained applications. Also, this is the reason why we claimed our method to be far more scalable than standard Bayesian NN that rely on Mixture distributions.\\n\\n\\n\\u201cI am very concerned by the performance of the ResNet20 models trained on CIFAR10 in the paper\\u201d\\n\\nDue to computational constraints we had trained each model for 80 epochs only, and thus our accuracies are lower than the standard ones, we shall fix this issue in the updated version of the paper with an open-sourced code to verify the correctness of our implementation.\\n\\nWe greatly appreciate the detailed reviews and regardless of the final decision would like your professional opinion on whether our work especially on ensembles being Bayesian (with the additional updates on further generalizing the proof to be model agnostic) is worth pursuing? We agree it seems like a low hanging fruit, but to the best of our limited knowledge, these insights are novel.\"}", "{\"title\": \"Thank You for the Detailed Review\", \"comment\": \"\\u201cThe interpretation of ensembling as variational inference is both flawed and well known\\u201d,\\u201dAdditionally, the derivation of dropout as Bayesian inference (which relies on a similar interpretation on taking the variance parameter in the Gaussian to zero) is known to be flawed (Hron et al, 2018); this flaw is directly applicable to the interpretation of ensembling as variational inference as described in Section 3 of this paper. The issue is fundamental: when taking the variance parameter to zero, the supports of the variational distribution and the true distributions have a mismatch, causing the KL divergence to become infinite. As a result, one cannot minimize the objective (as it is infinite by design).\\n\\nIndeed, if we take the posterior as a mixture of impulses then due to pathological KL divergences optimisation is impossible. Thus, we use a trick of keeping sigma sufficiently small. Gal had used this same trick for showing Dropouts as a Bayesian Approximation. So, in our approach all we need to do is keep sigma something as small as 10^(-34) ( for all practical purposes this is a zero for all discrete computers), the final objective function will become -\\\\log(10^(-34)) + required terms = 34 + required terms ( a simplistic argument Hron provides a more rigorous one for the same). Hron et al. [3] had called this approach the convolutional trick and quoting his work directly - \\u201c we can view both the discretisation and convolutional approaches as mere alternative vehicles to derive the same quasi discrepancy measure\\u201d. Hron et al. in their work \\u201cVariational Bayesian dropout: pitfalls and fixes\\u201d had never claimed standard Dropouts were irredeemably non-Bayesian. They had also provided fixes to Gal\\u2019s original arguments by the generalisation of variational inference to a quasi-KL and had shown ties between Gal\\u2019s convolution approach and their work. Thus, by extension, our method too can be considered Bayesian despite seemingly \\u201cinfinite\\u201d KL divergence (Use Quasi KL divergence if uncomfortable with the convolution argument). Hron primary goal was to simply provide better-set arguments for Dropouts being Bayesian and these can be directly extended to our cases of ensembles being Bayesian with a mixture of Gaussian posterior.\\n\\n\\u201cThe Bayesian bootstrap (Rubin, 1983) is a Bayesian interpretation of the bootstrap (identical under many conditions)\\u201d\\nAs pointed out correctly by you Ensembles are indeed very closely related to Rubin\\u2019s work Bayesian Bootstrap, but we will like to make further clarifications. Bayesian Bootstrap demands individual ensembles be trained on subsets of data, but this is not how ensembles are trained in practice. Our paper deals with the specific case of Ensembles mentioned in Lakshminarayanan et al. [1] where each ensemble is trained on the entire dataset, using randomized batch sampling to be data efficient. This \\u201cmathematically incorrect\\u201d form (authors have claimed this to be a non-Bayesian approach) of training the ensemble has been empirically been shown give state of the art uncertainty estimates. Thus, we believe there is merit in our work in reframing this work into a Bayesian framework. We have also gone a step further and extended this to an ensemble of Bayesian NN\\u2019s like Dropouts, BBB etc. We have also added additional proof in the Appendix showing that our proof can be easily extended to an Ensemble of the arbitrary architecture of NN\\u2019s (making our proof model agnostic and proof far more general) which to the best of our knowledge are novel insights, and not well known in the literature. For example, just recently Ensemble of Dropout was proposed as a solution to the mode collapse issue in Active Learning by Remus et al.[ 2], but as pointed out by the reviewers on Openreview, they were unable to provide theoretical justification for this empirical finding.\\n\\n[1] Balaji Lakshminarayanan, Alexander Pritzel, Charles Blundell, NeurIPS 2017 Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles\\n[2] Remus Pop and Patric Fulop - Deep Ensemble Bayesian Active Learning : Addressing the Mode Collapse issue in Monte Carlo dropout via Ensembles https://openreview.net/forum?id=Byx93sC9tm\\n[3]Hron, Matthews & Ghahramani, Variational Bayesian dropout: pitfalls and fixes, ICML, 2018. https://arxiv.org/abs/1807.01969\"}", "{\"title\": \"Thank You for the Pointers\", \"comment\": \"\\u201cI think the main contribution is to use the concrete distribution as a mixture prior q(z) instead of a categorical prior\\u201d, \\u201cSeveral papers consider the problem of learning a mixture of Gaussians in the VI framework ([1,2,3,4,5])\\u201d\\n\\nOur key goal was to make Bayesian Neural Networks (BNN) easily to train and scalable to large scale Neural Networks so another key novelty in our method is using Learnable Dropouts in the Convolutional Network and performing full Bayesian Inference with mixture distribution on only the final layer. This makes our approach several times less costly in terms of memory, while still achieving state of the art uncertainty estimates on out of distribution (OOD). So our work is different from the ones cited by you in the sense we want to design a practical BNN which can replace standard Neural Networks for a limited overhead (besides the MC sampling the complexity is almost the same as standard NN, making this a simple but effective trick).\\n\\n\\u201cThe authors should show why the proposed variational distribution is better than the mixture of Gaussians with a categorical prior\\u201c\\nWe agree a comparison between the two is due and we plan to add a comparison of NLL and Expected Calibration Error comparison for the same in the updated version of the paper, with the open sourced code. An easy intuition as to why our method is guaranteed to be better is that by tweaking the training procedure such that we keep the parameter p\\u2019s of the categorical random variable fixed until convergence and then switch a flag to make p\\u2019s tunable will guarantee a lower ELBO for our case. \\n\\n\\u201cThe deep ensemble method considered in this paper is just one of the existing ensemble approaches\\u201d\\n\\nIndeed there are multiple ways to the ensemble and we are referring to the specific case described by Lakshminarayanan et al. [1]. Links between Ensembles and Bayesian are neither new nor groundbreaking as it is well known that Ensembles are closely linked to the Rubin\\u2019s Bayesian Bootstrap [2]. Rubin\\u2019s method demands each ensemble be trained on an independently sampled subset of data but this is not how ensembles are trained in practice. In order to be data efficient we usually train each ensemble on the entire data albeit using a randomized mini-batch sampler. This \\u201cmathematically incorrect\\u201d form of training causes it to lose the theoretical support that other BNN\\u2019s enjoy. We show that this form of \\u201crandomized mini-batch\\u201d training can be reframed into the VI framework. A merit of this viewpoint is that we can easily generalize our proof to an ensemble of Neural Networks with arbitrary architectures (updated in the Appendix) which is completely novel. We have also shown that ensemble of BNN\\u2019s like Dropouts or BBB too can be considered Bayesian. Just recently Remus et al. had [3] proposed ensemble of Dropouts as a solution to mode collapse issue in Active Learning but as pointed out by reviewers on OpenReview they were unable to provide theoretical justification for their empirical findings.\\n\\n\\n[1] Balaji Lakshminarayanan, Alexander Pritzel, Charles Blundell, NeurIPS 2017 Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles\\n\\n[2] Rubin, The Bayesian Bootstrap, 1981. Annals of Statistics. https://projecteuclid.org/euclid.aos/1176345338\\n\\n[3] Remus Pop and Patric Fulop - Deep Ensemble Bayesian Active Learning : Addressing the Mode Collapse issue in Monte Carlo dropout via Ensembles https://openreview.net/forum?id=Byx93sC9tm\"}", "{\"title\": \"Thank you for the feedback\", \"comment\": \"We thank reviewer for his comments. We try to address his concerns in the best way possible.\\n\\n\\u201cIt is not surprising that Deep ensemble is equivalent to variational inference with learning the mean only. I\\u2019m not sure how useful this argument is since learning the distribution, not only the mean, is the key factor of being Bayesian.\\u201d. \\n\\nDeep Ensembles are well known to have ties with Bayesian Inference since they can be easily related to \\u201cThe Bayesian Bootstrap\\u201d by Rubin [1], where individual ensembles are trained on subsets of data, but this is not how it is used in practice. Lakshminarayanan et al. in their work \\u201cSimple and Scalable Uncertainty Estimation Using Deep Ensembles\\u201d [2] empirically showed that by using a mathematically incorrect form of training all ensembles on the same data, albeit using a randomized batch sampling, we can still obtain state of the art uncertainty estimates. This approach is massively popular for all practical purposes since it has much higher data efficiency, which was the main novelty of their work. The importance of our work comes from the fact that we are able to show that ensembles relying on randomized batch sampling can be viewed as Variational Inference with \\u201cMixture of Gaussian Posterior\\u201d (Not impulses, to avoid pathological KL divergences). We were also able to extend this viewpoint to an Ensemble of Bayesian Neural Networks like an ensemble of Dropout or BBB\\u2019s, which to the best of our knowledge is also novel. Past works have shown ensembles of Dropouts have been shown to be able to be more robust to adversarial perturbations and be able to solve the issue of mode collapse in Active Learning Scenarios [3], but they were unable to provide theoretical justifications for their empirical findings, thus we would argue that our work is useful in the sense it provides theoretical support and is consistent with past works.\\nAnother reason why our view is of importance is because our proof can be easily extended and made model agnostic. In the sense that by using a small trick, we can show that an ensemble of arbitrary architecture is also Bayesian which to the best of our knowledge is not known. We have added this simple extension in the Appendix, view Appendix 3 for details. \\n\\n\\n\\n\\u201ckey factor of being Bayesian\\u201d \\n\\nWe would argue that a key factor to being Bayesian is being able to get samples which seem to be derived from the true posterior, which ensembles are fully capable of and thus are Bayesian in Nature. Do note that Ensembles aren\\u2019t the only particle filter like approach to approximating the True Posterior, Stein Variational Gradient Descent [4] mentioned in our paper are similar to Ensembles and are known to be Bayesian.\\n\\n\\n\\n\\u201cUsing a mixture of Gaussians rather than a Gaussian in Bayes by Backprop is a natural extension and therefore the novelty seems low.\\u201d\\n\\nIndeed our work can be seen as a natural extension of BBB, but the novelty lies in the fact that using our method, we can design arbitrary complex concrete mixture distributions beyond the Gaussian ones used in the paper, e.g. our method can be combined with Normalizing Flows to model fairly complex distributions. We have also proposed and shown a fairly novel way to train Bayesian Neural Networks by treating the Convolution Layers as some type of dimensionality reduction algorithm (learnable dropouts using concrete distribution used here ) and performing Bayesian inference on only the final fully connected layer. This trick will be of exceptional importance for scaling to large scale applications or in memory constraint applications. \\n\\n\\u201cWhy multimodality\\u201d?\\nAs observed in our Fig 2, indeed two modes are very close but still unimodal distributions would not have captured the third distinct mode as in this case. The multimodality advantage will only increase as moving onto more complex regression or other classification tasks. \\nWe shall address your remaining concerns regarding the experiments ( of BBB) soon after open-sourcing the code.\\n\\n\\n\\n[1] Rubin, The Bayesian Bootstrap, 1981. Annals of Statistics.\", \"https\": \"//projecteuclid.org/euclid.aos/1176345338\\n[2] Balaji Lakshminarayanan, Alexander Pritzel, Charles Blundell, NeurIPS 2017. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles\\n[3]Remus Pop and Patric Fulop - Deep Ensemble Bayesian Active Learning : Addressing the Mode Collapse issue in Monte Carlo dropout via Ensembles https://openreview.net/forum?id=Byx93sC9tm\\n[4] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general-purpose Bayesian inference algorithm. In Advances in neural information processing systems\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"Summary\\nIn this works, the authors propose to use a concrete mixture of Gaussians as a variational distribution. The authors show that the deep ensemble method can be viewed as a special case of a mixture of Gaussians with a categorical prior.\\nI think the main contribution is to use the concrete distribution as a mixture prior q(z) instead of a categorical prior.\\nHowever, there are some concerns. The following points should be addressed to get a higher rating. \\n\\n(1) Several papers consider the problem of learning a mixture of Gaussians in the VI framework ([1,2,3,4,5]). The deep ensemble method considered in this paper is just one of the existing ensemble approaches. The related work section should be updated to discuss the novelty of this work.\\n\\n(2) In the deep ensemble method, the categorial prior q(z) is held fixed. However, it is possible to update the categorical prior q(z). \\nIn this work, the proposed distribution is q(w) = sum_{c=1}^{K} Concrete(z=c|\\\\theta_z) Gauss(w|z=c,\\\\theta_{w_c}).\\nBy using the concrete prior, the gradient can be computed by the local reparametrization trick for q(w|z) and the reparametrization trick for q(z). \\nHowever, even when q(z) is the categorical prior, the local reparametrization trick for q(w|z) is still valid if the variational parameters for each mixture component are not tied. The main difference is the reparametrization trick can not be used for q(z). \\nThe authors should show why the proposed variational distribution is better than the mixture of Gaussians with a categorical prior. \\n \\n\\nReferences\\n[1] O. Arenz, M. Zhong, and G. Neumann. \\\"Efficient Gradient-Free Variational Inference using Policy Search.\\\" ICML. 2018.\\n[2] A. C. Miller, N. J. Foti, A. D\\u2019Amour, and R. P. Adams. Variational boosting: Iteratively refining posterior approximations. ICML, 2017. \\n[3] O. Zobay. Variational Bayesian inference with gaussian-mixture approximations. Electronic Journal of Statistics, 8(1):355\\u2013389, 2014.\\n[4] Lin, Wu, Mohammad Emtiyaz Khan, and Mark Schmidt. Fast and Simple Natural-Gradient Variational Inference with Mixture of Exponential-family Approximations. ICML 2019.\\n[5] F. Guo, X. Wang, K. Fan, T. Broderick, and D. B. Dunson. Boosting variational inference. arXiv:1611.05559v2, 2016.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors interpret Deep Ensemble as a special type of variational inference. Based on Bayes by Backprop which uses a Gaussian approximation to the posterior, the authors propose to use a mixture of Gaussians. The proposed methods have been tested on a regression task and Bayesian neural networks.\\n\\nThe authors argue that Deep ensemble is equivalent to variational inference with a mixture of Gaussians approximation with variance going to 0. It is not surprising that Deep ensemble is equivalent to variational inference with learning the mean only. I\\u2019m not sure how useful this argument is since learning the distribution, not only the mean, is the key factor of being Bayesian.\\n\\nUsing a mixture of Gaussians rather than a Gaussian in Bayes by Backprop is a natural extension and therefore the novelty seems low. \\n\\nIn the experiments, I\\u2019m surprised to see that Bayes by Backprop fails to give any uncertainty estimate on the simple regression experiment. From the figure, BBB seems to be very confident about its prediction. However, it has been demonstrated in the literature that BBB is able to perform fairly well on this kind of regression. Can the authors give explanations on why it performs badly here? Also, I do not see why it is important to model multiple modes on this task. I believe a Gaussian approximation is able to work well. The multimodality here is likely induced by the symmetric parameterization of neural networks and trying to capture this kind of multimodality will be meaningless and even problematic. By looking at Figure 2, it seems like the posterior of the proposed method is close to being unimodal. Why is it beneficial to use a mixture of Gaussians under this situation?\\n\\nThe axis and labels in the figures are very small and hard to read.\\nEq. (2) is overlapped with the above text.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: This paper proposes to use either relaxed mixture distributions or relaxed mixtures of matrix Gaussians as the approximate posterior for Bayesian neural networks. Naturally, taking the mixture variance to zero allows a stretched interpretation of ensembling to become Bayesian as well. Experiments are performed on small neural networks on regression tasks, as well as uncertainty estimation experiments on deep networks. Finally, a downstream task of bandit optimization (the mushroom experiment) is performed.\", \"post_rebuttal\": \"I think that this issue was satisfactorily addressed.\\n1)\\tMixture distributions as the variational family have already been proposed \\u2013 at least once \\u2013 see Miller et al, 2017 for an example that additionally uses the reparameterization trick. There, updates to the posterior are performed via selectively adding mixture components when necessary. Given that they also run experiments on neural networks, it would be nice to see a comparison to that method.\", \"tldr\": \"I vote to reject this paper for several reasons. The interpretation of ensembling as variational inference is both flawed and well known. Mixture distributions have been previously proposed for variational inference. There are significant enough weaknesses in the empirical results that make me question the authors\\u2019 own implementations. I will increase my score if these issues are resolved.\", \"originality\": \"1)\\tIt is well known (and immediately obvious) that one can interpret ensembles as a Bayesian method \\u2013 with a mixture of delta functions around the independently trained models. Furthermore, the Bayesian bootstrap (Rubin, 1983) is a Bayesian interpretation of the bootstrap (identical under many conditions), while re-weighted ensembles are Bayesian models (Newton, 1995; Newton, 2018) that converge to the true posterior under many conditions \\u2013 intriguingly, this version of re-weighting could potentially have good uncertainty quantification properties. For the specific case of stacking (which is closely related to ensembling), there are interpretations of stacking as Bayesian model combination (Yao et al, 2018, Clyde & Iversen, 2013). As a result, this view of ensembling as a Bayesian method is not novel.\\n\\n2)\\tAdditionally, the derivation of dropout as Bayesian inference (which relies on a similar interpretation on taking the variance parameter in the Gaussian to zero) is known to be flawed (Hron et al, 2018); this flaw is directly applicable to the interpretation of ensembling as variational inference as described in Section 3 of this paper. The issue is fundamental: when taking the variance parameter to zero, the supports of the variational distribution and the true distributions have a mismatch, causing the KL divergence to become infinite. As a result, one cannot minimize the objective (as it is infinite by design).\", \"clarity\": \"1)\\tThe proposed methods are clearly explained in the setting under which one might be able to re-implement them, particularly in the description of the approximate posteriors.\", \"significance\": \"\", \"quality\": \"1)\\tI am very concerned by the performance of the ResNet20 models trained on CIFAR10 in the paper. The reported results for deep ensembles (an ensemble of _three_ independently trained models) in Appendix A are 85.23% accuracy. By comparison, the number in the original paper (He et al, 2015) is 91.25% for a _single_ model. Furthermore, the first link to a PyTorch repo from a Google search (https://github.com/akamaster/pytorch_resnet_cifar10) comes up with a re-implementation that gets 91.63% (again for a single model).\\n\\n__This issue alone is enough for me to vote to reject the paper until the authors\\u2019 implementations are fixed and run with deep ensembled ResNets that get similar accuracy. I hope to see the implementation issues fixed in the rebuttal period. __\\n\\n2)\\tFurthermore, in Appendix 2, I see a potential issue in using the KL divergence and would like the authors to clarify.\\nIf the secondary lower bound is being used to approximate the KL divergence between the mixture distributions, it becomes very unclear if what is being optimized is in fact a __lower bound__ on the log probability. After all, the standard ELBO is written as:\\n\\tLog p(y) \\\\geq ELBO := E_q(\\\\log p(y | \\\\theta)) \\u2013 KL(q(\\\\theta) || p(\\\\theta)) \\nImplying that another lower bound on the KL would be greater than the ELBO and thus not necessarily a lower bound on the log marginal likelihood.\\n\\n3)\\tBottom of page 7: \\u201cThe posterior is used only for fully connected layers\\u2026\\u201d Why is this the case? Alternative Bayesian methods can be used on the full network.\", \"minor_comments\": [\"Please attempt to use the math_commands.tex in the ICLR style file to write math. This makes the math standardized throughout.\", \"Figure 1 is especially unclear and the legends and captions should be made considerably larger. Zooming in at 800% seems to be necessary to even be able to make out which method is which. Consider, placing all methods on a single plot and then coloring/dashing them separately. Additionally, please compare to HMC on this task (and an equivalent RBF GP) to show the uncertainties from the \\u201cgold standard\\u201d methods.\", \"Please additionally make the figure legends larger for the rest of the figures.\", \"What is the meaning of time in Figure 2? I assume that it means training time, but it is not especially clear from the caption. With that being said, I think that it is quite interesting that the marginals do seem to look multi-modal, even at the end of training.\", \"Line before Eq 7: please use \\\\citet if possible to cite Gupta & Nagar, rather than stating the clunky Gupta et al. \\\\citep{Gupta & Nagar}.\", \"Eq 2: please do not use \\\\hspace{-\\u2026} to save space in equations, so that the equation does not overwrite lines before.\", \"Page 2: \\u201cDeep Ensembles have lacked theoretical support \\u2026\\u201d In order to make this claim, you must first explain why one ought to be Bayesian in the first place. See above for the history of interpreting ensembling from a Bayesian perspective. There is no a priori reason why one would not just wish to have frequentist justifications of ensembling (and potentially even calibration).\", \"Please do not capitalize Ensembling or Variational Inference throughout.\"], \"references\": \"Clyde & Iversen, Bayesian model averaging in the M-open framework. In Bayesian Theory and Applications, 2013. DOI:10.1093/acprof:oso/9780199695607.003.0024\\n\\nHe, Zhang, Ren & Sun, Deep Residual Learning for Image Recognition, CVPR, 2016. https://arxiv.org/abs/1512.03385\\n\\nHron, Matthews & Ghahramani, Variational Bayesian dropout: pitfalls and fixes, ICML, 2018. https://arxiv.org/abs/1807.01969\\n\\nMiller, Foti & Adams, Variational Boosting: Iteratively Refining Posterior Approximations, ICML, 2017. https://arxiv.org/abs/1611.06585\\n\\nNewton & Raftery, Approximate Bayesian Inference with the Weighted Likelihood Bootstrap. JRSS:B, 1994. https://www.jstor.org/stable/2346025?seq=1#metadata_info_tab_contents\\n\\nNewton, Polson & Xu, Weighted Bayesian Bootstrap for Scalable Bayes, 2018; https://arxiv.org/abs/1803.04559\\n\\nRubin, The Bayesian Bootstrap, 1981. Annals of Statistics. https://projecteuclid.org/euclid.aos/1176345338\\n\\nYao, Vehtari, Simpson, & Gelman, Using Stacking to Average Bayesian Predictive Distributions, Bayesian Analysis, 2018; https://projecteuclid.org/euclid.ba/1516093227\"}" ] }

[ "Edward Ombui", "Lawrence Muchemi", "Peter Wagacha" ]

[ "Hate Speech", "Code-switching", "feature selection", "representation learning" ]

{ "note_id": [ "_fHNjvlDFt", "SyxCb3k6FH", "r1gV_dm2KB", "BJgWdenDOr" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752882, 1571777541565, 1571727467906, 1570386024560 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2590/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2590/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2590/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper focuses on hate speech detection and compares several classification methods including Naive Bayes, SVM, KNN, CNN, and many others. The most valuable contribution of this work is a dataset of ~400,000 tweets from 2017 Kenyan general election, although it is unclear whether the authors plan to release the dataset in the future.\\n\\nThe paper is difficult to follow, uses an incorrect ICLR format, and is full of typos.\\n\\nAll three reviewers agree that while this paper deals with an important topic in social media analysis, it is not ready for publication in its current state. The authors did not provide a rebuttal to reviewers' concerns.\\n\\nI recommend rejecting this paper for ICLR.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"I'm sorry to say that this paper is not ready for publication.\\n\\nI think it's an important area and the dataset collected could be quite valuable for tackling hate speech.\\n\\nThe paper does not follow the style guide, is full of typos or 'kkkkk' tokens indicating missing values. The first sentence of the abstract is not grammatical. Codeswitched needs to be mentioned more specifically in the introduction.\\n\\nI couldn't easily find statistics about the dataset, especially in terms of language breakdown. How many of the tweets were multi-lingual? For pure-english tweets, I would be interested in this dataset being split out as a sub-dataset, as I would heavily bet the sota method for classifying hate-speech would be to fine-tune a BERT model on these labels. We need a table of dataset statistics.\\n\\nWe need a mathematical definition of PDC that is made very explicit in the paper, there is too much prose, I did not have time to do the background reading of the linked papers to understand this sociological theory of hate speech.\\n\\nThe paper did not feel sufficiently anonymized. I would anonymize the university used to create the dataset and the funding agencies that supported the research in subsequent submissions.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I made a quick assessment of this paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper compared several classification methods, including deep neural networks (DNN), to identify hate speech texts. Mainly the data was corrected from twitter. The data was prepossessed to deal with by popular classification methods. The LDA and PDA are used to construct the identifier with high accuracy. Finally, Practical data was used to assess the performance of several machine learning algorithms.\\n\\nThe research topic is important from the sociological viewpoint. I'm not sure whether this paper suits to the publication from ICRL. Besides that, the authors did not show any technical insight into the numerical results. That is, can the authors explain the reason why the linear logistic regression with TF-IDF features outperformed all the other methods? Overall, this paper did not provide any useful knowledge, while this paper introduced some statistical methods and showed numerical results. I recommend the authors to add more beneficial insight and to submit the paper to other conferences that deals with sociological issues.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper deals with an important problem in social media analysis. With the spread of hate speech and hate crimes by rioting separatists in Hong Kong as well as equally hateful attacks on Chinese people in the west, I think that this is an issue that deserves more attention in our community. Unfortunately this paper needs more polish to be appropriate for ICLR. It also might be better suited to an NLP focused conference (such as ACL, EMNLP, or NAACL) although I think if the technical contribution is clear enough it could be suitable for ICLR as well. \\n\\nI think the big things to focus on would be including more baselines, improving polish in the paper, and providing a clearer high-level analysis of the dataset (with specific examples mostly left for the appendix).\"}" ] }

[ "Hongbin Pei", "Bingzhe Wei", "Kevin Chen-Chuan Chang", "Yu Lei", "Bo Yang" ]

[ "Deep Learning", "Graph Convolutional Network", "Network Geometry" ]

{ "note_id": [ "N8CgJ43jX18", "BQ-bUBeeDhx", "8tGKV1oSzCr", "hCb9YKDYJq", "dDKYzaVzUJ", "pb92t3wP6R", "4awCnNnpYO", "S1giVoY2iB", "Byx9Lj4hiH", "HklyXj4niS", "SkxLWoVhsr", "Bkgpjq4hsr", "rJxqK542oS", "rkeT4UtIcS", "SJluLVxXqr", "rJgofPeTKr", "rkxArIisYH", "SJg0E9eqKS", "B1lX_zzXdr", "BJx-XvXk_B" ], "note_type": [ "comment", "comment", "comment", "official_comment", "comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "comment", "official_comment", "comment" ], "note_created": [ 1590973723473, 1588278694839, 1587644157273, 1577567648543, 1577014315214, 1576905755572, 1576798752851, 1573849907147, 1573829457832, 1573829398722, 1573829374057, 1573829284769, 1573829250031, 1572406836569, 1572172880107, 1571780371062, 1571694149806, 1571584565574, 1570083435250, 1569826585395 ], "note_signatures": [ [ "~Zhengyang_Wang1" ], [ "~Zhitao_Ying1" ], [ "~Deyu_Bo1" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "~Jiong_Zhu1" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "ICLR.cc/2020/Conference/Paper2589/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2589/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2589/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "~Henry_Blackmore1" ], [ "ICLR.cc/2020/Conference/Paper2589/Authors" ], [ "~Yilun_Jin1" ] ], "structured_content_str": [ "{\"title\": \"MLP outperforms all GNNs on four of the disassortative datasets\", \"comment\": \"We find that a simple 2-layer MLP outperforms all GNNs on four of the disassortative datasets as shown below. More details are provided in our paper (https://arxiv.org/pdf/2005.14612.pdf).\\n\\nDatasets Actor Cornell Texas Wisconsin\\nGCN 30.3 54.2 61.1 59.6\\nGeom-GCN 31.6 60.8 67.6 64.1\\nGeom-GCN-s 34.6 75.4 73.5 80.4\\nMLP 35.1 81.6 81.3 84.9\"}", "{\"title\": \"The experiment is not in semi-supervised setting as in the previous author response\", \"comment\": \"There was a question of whether the semi-supervised setting of GCN and GAT are used for experiments. The authors' answer is contradictory to the description of the paper.\\n\\nThe paper actually uses *Random* split (with 60%, 20%, 20% ratio), rather than the split for semi-supervised setting in GCN (fixed small set of training nodes). \\n\\nWould appreciate if the authors also publish the settings and results in the standard semi-supervised setting, to support the claim and make the results reproducible in the standard setting for Cora, CiteSeer etc. as such split is done for the previous papers.\"}", "{\"title\": \"The error results of Table 3\", \"comment\": \"In the source code of 'utils_data.py', line 96, the adjacency matrix is not symmetrical, which is unfair for GCN and GAT.\\nWhen using a symmetric adjacency matrix, the performances of GCN and GAT are greatly improved.\"}", "{\"title\": \"Thank You for the Bug Report! Increased Lead Over Baseline Methods on the Actor Dataset (Table 3)\", \"comment\": \"Dear Mr. Zhu,\\n\\nThank you very much for your detailed report regarding a bug in the published code for Geom-GCN. We were able to reproduce the overflow in utils_data.py and have fixed the published code accordingly. Please see https://github.com/anonymous-conference-submission/geom-gcn/ for the most up-to-date version of the published code for Geom-GCN.\\n\\nWe have also rerun all affected experimental trials and have found that the performance advantage of Geom-GCN over baseline methods have increased. In particular, the updated results for GCN, GAT, and Geom-GCN on the Actor dataset (Table 3) are as follows:\\n\\n Mean Classification Accuracy (Percent)\\n\\n Model Old -> New Absolute Change\\n\\nGCN 27.92 -> 26.86 -1.06%\\nGAT 28.15 -> 28.45 +0.30%\\n\\nGeomGCN-I 28.65 -> 29.09 +0.44%\\nGeomGCN-P 31.28 -> 31.63 +0.35%\\nGeomGCN-S 29.83 -> 30.30 +0.47%\\n\\nSincerely,\\nGeom-GCN Authors\"}", "{\"title\": \"Bug in the Published Code Repository for Parsing the Actor Dataset\", \"comment\": \"We think there is a bug in the published code which will lead to incorrect parsing of the Actor dataset (in the code it is referred as \\\"film\\\"), and as a result, the performance reported for the Actor dataset in Table 3 of the paper is likely to be incorrect if the experiments are based on the published code.\\n\\nIn file \\\"utils_data.py\\\" in the repo, line 61, at commit bd0acd9:\\n\\nfeature_blank[np.array(line[1].split(','), dtype=np.uint8)] = 1\\n\\nThe \\\"dtype\\\" here should not be \\\"np.uint8\\\" since this array is supposed to contain values larger than 255. From our understanding, since the dimension of feature vectors in this dataset is around 931, the maximum possible value of this array should be around 930. Thus, if \\\"np.uint8\\\" is used as data type, an integer overflow problem will occur and the feature vectors will be incorrectly loaded; it should be fixed to a data type like \\\"np.uint16\\\". \\n\\nThe size of the array allocated in line 60 may also need to be fixed accordingly.\"}", "{\"title\": \"Thanks a lot, and we cited this paper actually\", \"comment\": \"Thanks a lot for your contributions to collecting these great wiki datasets. In fact, we cited the paper you mentioned in this submission (Please see Section 4.1 and the References).\"}", "{\"decision\": \"Accept (Spotlight)\", \"comment\": \"This paper is consistently supported by all three reviewers and thus an accept is recommended.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Reply to Official Review #3 Part 2\", \"comment\": \"Q5. Concerning the Section on empirical analysis, it might be of interest to investigate whether with a proper number of layers a GCN would emulate a geom-GCN.\", \"response\": \"Thanks for the great suggestion. In this paper, we use a two-layer Geom-GCN currently for fair comparisons with GCN and GAT which both adopt a two-layer architecture. The performance of multi-layer GCNs is actually limited, which has been studied in JK-Net [Xu 2018]. Thus, it's a promising direction to design multi-layer graph neural networks. We will try to design deep Geom-GCNs, especially for disassortative graphs, in future work.\\n\\n[Xu 2018] Xu, K., Li, C., Tian, Y., Sonobe, T., Kawarabayashi, K. I., & Jegelka, S. Representation learning on graphs with jumping knowledge networks. International Conference on Machine Learning (ICML). 2018, 5449\\u20135458.\\n\\nWe thank the reviewer once again. And we have released anonymously the code of Geom-GCN, and the particular architectures of Geom-GCN are also reported in the revision. One can access the code via link: https://github.com/anonymous-conference-submission/geom-gcn/.\"}", "{\"title\": \"Reply to Official Review #3 Part 1\", \"comment\": \"We would like to thank the reviewer for their thoughtful comments and appreciation. We would like to answer the reviewer\\u2019s questions as follows:\\n\\n\\nQ1. The strong dependency from the embedding function does not guarantee the discovery of long range connections. It may happen that the proposed embedding does not catch the relevant information for the task; in these cases the a priori knowledge on the task becomes crucial. This potential issue is partially supported by the results presented in the manuscript, where there is a gain only when the correct embedding is chosen.\", \"response\": \"A fundamental weakness of existing message passing neural networks (MPNNs) is lacking the ability to capture long-range dependencies. The reason is exactly what you mentioned, the relevant information is washed out during the many hops.\\n\\nThis weakness can be addressed if the relevant information from distant nodes can be effectively filtered and aggregated in someway. For instance, TO-GCN could automatically connect two disconnected nodes in one class by topology optimization [Yang 2019]. In this paper, we attempt to map the distant nodes with relevant information into a small area in an embedded space. Then the relevant information can be aggregated effectively in the neighborhood defined in the embedded space.\\n\\nWe acknowledge that the efficacy of such aggregation would depend on the selected embedded space. From experiments, we indeed observe that relevant information is aggregated from disconnected nodes in disassortative graphs (see Table 4 in the revision). As future work, we will explore techniques for automatically choosing a right embedding method\\u2013- depending not only on input graphs but also on target applications. \\n\\n[Yang 2019] Yang, L., Kang, Z., Cao, X., Jin, D., Yang, B., & Guo, Y. Topology optimization based graph convolutional network. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI. 2019, 4054-4061.\"}", "{\"title\": \"Reply to Official Review #2 Part 2\", \"comment\": \"Q6. There are various GCN implementations; however, the comparison is performed with only 2 of them. I would like to see either comparison with more implementations, or the explanation why the comparison with the given two suffices.\", \"response\": \"Thanks for the detailed suggestions. All the presentation issues have been modified in the revision carefully. Please see the details in the following.\\n\\n Q8.1: The notation used in the definition of $m_v^l$ is unclear.\\n R8.1: We update the description of $m_v^l$ to clarify it in Section 2C.\\n\\n Q8.2: Why $\\\\tau$ is a part of each node\\u2019s structural neighborhood? It\\u2019s a global function, isn\\u2019t it?\\n R8.2: Yes, $\\\\tau$ is a global function. We add a description for $\\\\tau$ in Section 2B.\\n\\n Q8.3: Introduction: I believe that the exact problems which GCNs solve (e.g. node classification) should be mentioned.\\n R8.3: We describe the exact graph learning task, node classification, in the introduction of the revision.\\n\\n Q8.4: The flow in Section 2.1 is a bit weird. Namely, it says \\u201cTo overcome the first weakness\\u201d, but the first weakness wasn\\u2019t stated in the previous paragraph (of course, one can deduce it, and it also was defined long ago, but it\\u2019s disturbing for a reader).\\n R8.4: We exactly describe the two weaknesses of MPNNs in Section 2.1 of the revision.\\n\\n Q8.5: Figure 1B is confusing: it looks like the nodes from N_g(v) lie in a small region around v.\\n R8.5: We add an explanation for the graph neighborhood in the caption of Figure 1 to clarify this confusion.\\n\\n Q8.6: I think that splitting Figure 1C into 2 figures would make it clearer.\\n R8.6: We modify the organization and caption of Figure 1 to make it clearer to a reader in the revision.\\n \\nWe thank the reviewer once again.\"}", "{\"title\": \"Reply to Official Review #2 Part 1\", \"comment\": \"We thank the reviewer for the helpful comments. We have revised the paper according to the suggestions and would like to answer the reviewer\\u2019s questions as follows:\\n\\nQ1. My main concern is the learning time, which is an issue for a straightforward GCN implementation. There were multiple attempts to decrease it (GraphSAGE, FastGCN, etc.). Therefore, I would like to see running times on the presented graphs as well as on relatively large graphs (see e.g. https://arxiv.org/pdf/1902.07153.pdf for candidates). If some techniques were used to make the implementation faster, I would be good to include them in the paper (or, if they are standard, they should be referenced). At the very least, I believe it should be prioritized as a future direction.\", \"response\": \"In the real implementation, we use the $\\\\tau$ that we defined in Table 1 (section 3). In the revision, we exactly mention which $\\\\tau$ is used in the experiment section.\"}", "{\"title\": \"Reply to Official Review #1 Part 2\", \"comment\": \"Q3. The experimental studies are quite weak. Some ablation studies should be done to evaluate the contribution of each proposed methods. For example, how N_{s}(v) contributes to the performance. This is very important for fully evaluating your methods.\", \"response\": \"We have corrected the formatting of W in the revision according to the standardized notation of ICLR 2020. \\n\\nWe thank the reviewer once again. And we have released anonymously the code of Geom-GCN, and the particular architectures of Geom-GCN are also reported in the revision. One can access the code via link: https://github.com/anonymous-conference-submission/geom-gcn/.\"}", "{\"title\": \"Reply to Official Review #1 Part 1\", \"comment\": \"We thank the reviewer for the thoughtful comments. We have revised and updated the paper according to your suggestions and would like to answer questions as follows:\\n\\n \\nQ1. To overcome the long-term dependency limitation, GEOM-GCN selects some nodes that are close but not directly connected for aggregation. However, the selected nodes in this way may not connect to the center node. This is an issue that if two nodes that are not connected should be aggregated. The authors should clarify this.\", \"response\": \"Indeed, the non-isomorphic example graphs in GIN can be distinguished by applying simple aggregator, mean, or maximum, in a structural neighborhood. We provide and describe a detailed solution in the revision (see Section 2.1.1).\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"GEOM-GCN: GEOMETRIC GRAPH CONVOLUTIONAL NETWORKS\\n\\nThe paper introduces a novel GCN framework, whose purpose is to overcome weaknesses of existing GCN approaches, namely loss of structural neighbor information and failure to capture important dependencies between distant nodes. The paper uses a mapping from nodes to an embedded space and introduces a second type of a neighborhood: a proximity in the embedded space. In the embedded space, a set of relations of nodes is defined. For each node v, the paper uses a 2-stage convolution scheme: 1) for each neighborhood type, the nodes in the same relation with v are combined; 2) the resulting nodes are again combined into a new feature vector. This approach allows one to overcome the issues described above. The experiments show that in most cases the approach outperforms the existing GCN solutions, sometimes with a large gap.\", \"i_have_the_following_concerns_about_the_paper\": \"-- My main concern is the learning time, which is an issue for a straightforward GCN implementation. There were multiple attempts to decrease it (GraphSAGE, FastGCN, etc.). Therefore, I would like to see running times on the presented graphs as well as on relatively large graphs (see e.g. https://arxiv.org/pdf/1902.07153.pdf for candidates). If some techniques were used to make the implementation faster, I would be good to include them in the paper (or, if they are standard, they should be referenced). At the very least, I believe it should be prioritized as a future direction.\\n-- It\\u2019s unclear why we should use the same latent space and the same \\u03c4 for both N_g and N_s. I would expect that mapping into different spaces could provide better results: the two neighborhood types seem very different, and I don\\u2019t see why the neighbors should be aggregated in the same way. If using different spaces doesn\\u2019t provide an improvement, an explanation for this would be very useful.\\n-- \\u03b1 and \\u03b2 are defined and shown in Table 2, but they are never used (as it stands now, \\u03b1 and \\u03b2 can simply be removed). If the results in Table 3 correlate with them, then this dependence should be highlighted. In such case, it would also be better to move \\u03b1 and \\u03b2 to Table 3.\\n-- The paper uses 3 different node embedding strategies. These strategies can be combined in q with different weights (which can be learned as hyperparameters). Will it produce the best of 3 (or better) result?\\n-- \\u201cWe use an embedding space of dimension 2 for ease of explanation\\u201d But what \\u03c4 is used in the real implementation?\\n-- There are various GCN implementations; however, the comparison is performed with only 2 of them. I would like to see either comparison with more implementations, or the explanation why the comparison with the given two suffices.\\n-- Is it possible to make the implementation available?\\n\\nWhile there are a lot of possible improvements, I believe that some of them can be addressed in a future research, and the paper\\u2019s novel approach is noteworthy in itself. My current verdict is 5/10, and I\\u2019ll be happy to improve it if the above issues are fixed.\", \"presentation_issues\": \"-- The notation used in definition of m_v^l is unclear.\\n-- Why \\u03c4 is a part of each node\\u2019s structural neighborhood? It\\u2019s a global function, isn\\u2019t it?\\n-- Introduction: I believe that the exact problems which GCNs solve (e.g. node classification) should be mentioned.\\n-- The flow in Section 2.1 is a bit weird. Namely, it says \\u201cTo overcome the first weakness\\u201d, but the first wickness wasn\\u2019t stated in the previous paragraph (of course, one can deduce it, and it also was defined long ago, but it\\u2019s disturbing for a reader).\\n-- Figure 1B is confusing: it looks like the nodes from N_g(v) lie in a small region around v.\\n-- I think that splitting Figure 1C into 2 figures would make it clearer.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The work is based on the premise that existing MPNNs have two main weaknesses: (i) the loss function doesn't properly capture the spatial information during graph convolution, and (ii) the difficulty to manage the information encoded in the long range connections.\\n\\nThe main contribution of this work is a novel method called geometric aggregation. The proposed method is based on two elements: (i) a latent space mapping to capture spatial information using a new bi-level operator, (ii) an integration of geometric aggregation inside GCN, namely geom-GCN. More in detail the idea reported in this work is to map the input graph into an embedding where the geometric relations between nodes are preserved. The graph is embedded using a usual embedding function that guarantees to preserve some graph property of interest like the hierarchy of nodes. After the node embedding, the authors propose to create a structural neighbourhood both (i) in the latent space, taking the nodes within an arbitrary radius, and (ii) in the original space, by taking the adjacent nodes. The expectation here is that with the proper embedding the latent space can catch connections, which are long range in the original space. The message passing is actuated exploiting first the structural neighbourhood to do low-level aggregation, which aggregates nodes that have the same geometric relationship using permutation invariant operators, and then the result of these aggregations, which are virtual nodes, are aggregated again through high-level aggregation making use of operators like concatenation. \\n\\nThe goal of this work is clearly formulated by posing the proper research questions. The topic is relevant and it is part of the research agenda of ICLR. A key point of the proposed method is the ortogonalithy of geometric aggregation with respect to other aggregators like GAT. The design of the structural neighboorhood allows the network to choose which neighbors are the most important for the learning task.\\n\\nSome minor comments.\\nThe strong dependency from the embedding fuction does not guarantee the discovery of long range connections. It may happen that the proposed embedding does not catch the relevant information for the task; in these cases the a-priori knowledge on the task becomes crucial. This potential issue is partially supported by the results presented in the mauscript, where there is a gain only when the correct embedding is chosen. \\nA further critical point is the choice of the radius. Such a choice can be operated only with an empirical assessment. It is not clear whether it migth be meaningful to choose a radius thatwould encode the same neighbourhood as in the original sapce of data.\\nThe authors claim that even if there are more hops between two nodes the relevant information would arrive from the far node to the target node. Nevertheless we may conceive a situation where the relevant information is washed out during the hops. It may happen when the information of the far node is relevant for the target node, but it is not relevant for the target neighbour nodes.\\nThe use of concatenation as high level operator is critically dependent from the radius and from the number of edges in the graph. In cases of large values for radius or very dense graphs, the concatenation may increase the spatial complexity of the networks.\\nConcerning the Section on empirical analysis, it might be of interest to investigate whether with a proper number of layers a GCN would emulate a geom-GCN.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This work proposes geometric aggregation scheme for GCNs, which aims to overcome the limitations in traditional GCNs; those are lacking long distance dependencies and structure information in nodes. In particular, each node is transformed into a latent space. To overcome the first limitation, some nodes that are not directly connected but fall in a near range are also used in aggregation. A relational operator is used to provide position information for each pair of nodes. In this way, the structure information in graph can be used.\\n\\nThe method proposed in this work is novel and interesting. However, I am confused how this method can overcome the two limitations faced by previous GCNs. To my understanding, GEOM-GCN maps all node in to a 2D latent space. This can be treated as a lower-dimension representation for each node. Based on this, some similar nodes are clustered together. The relational operator is a kind of ranking operator that can rank two nodes based on latent space representations. If my understanding is wrong, please correct me.\\n\\nBased on this understanding, I didn't find this method can solve the two limitations.\\n \\n1. To overcome the long-term dependency limitation, GEOM-GCN selects some nodes that are close but not directly connected for aggregation. However, the selected nodes in this way may not connect to the center node. This is a issue that if two nodes that are not connected should be aggregated. The authors should clarify this.\\n\\n2. The relational operator is used to provide a ranking between two nodes. However, how such kind of operators can be used to aggregate the structure information as described in GIN. For example, how to distinguish those example graphs using this work. I think it would be a plus if authors can make this clear in the paper.\\n\\n3. The experimental studies are quite weak. Some ablation studies should be done to evaluate the contribution of each proposed methods. For example, how N_{s}(v) contributes to the performance. This is very important for fully evaluating your methods.\\n\\n4. More tasks and datasets can be added such as graph classification and social networks.\\n\\n5. Some notations are quite confusing. Like in eq.(1), why m is bold but W is not.\"}", "{\"title\": \"Thanks a lot for your helpful comments!\", \"comment\": \"Thanks a lot for your helpful comments!\\n\\nI try my best to elaborate on your concerns.\", \"q1\": \"Is the size of $m_v^{l+1}$, the output of high-level aggregation function q, fixed?\", \"r1\": \"Yes, the size is fixed. Your concern originates from missing virtual nodes because there may be no node in certain blocks in the neighborhood. And the changing of the number of virtual nodes leads to the different size of input of q, which is a set of features of virtual nodes. To address this issue, we specify a zero-vector as the feature of the missing virtual node in our model, thereby keeping the same size of input/output of q.\", \"q2\": \"To provide experimental results with only low-levlel aggregation in each block\", \"r2\": \"It's a very good suggestion. We did not test Geom-GCN with deep layers. We use a two-layer Geom-GCN for fair comparisons with GCN and GAT which both adopt a two-layer architecture. We will design deep Geom-GCN and test it, especially on disassortative graphs, in future work.\\n\\nThe paper will be modified in the future version according to your comments. Thank you very much!\", \"q3\": \"Did you test Geom-GCN with deep layers (e.g., 5~6 or more) and provide the results to make this paper better?\"}", "{\"comment\": \"Nice work!\\nI have some questions about the details and model.\\n1. If the high-level aggregation function q is concatenation, the dimension of $m^{l+1}_v$ is not fixed, which may change with the number of virtual nodes. So how can you solve this? To my understanding, the input of q is a set of e, while the output should be fixed dimension. Hope you can explain about it.\\n2. The description of bi-level aggegration seems great, but it's better to provide experimental results on each block. I mean how is the performance of your model without high-level aggegration? just do simple aggegration in low-levlel to each node. \\n3. As for me, Geom-GCN has a good use of long-range information, which is not applicable for most existing works. well, i think a main challenge of making GCN deep is the use of long-range information, because former works with deep layers will absorb the same information for each node. What about Geom-GCN? I think with more layers like 5~6 or more, Geom-GCN can achiever better performance. Did you test this or can you provide results with deep layers, making this paper better?\", \"title\": \"Nice work but some detailed results are necessary\"}", "{\"comment\": \"Thanks a lot for your comments and affirmations.\\n\\nFor your concerns about our experiments,\", \"q1\": \"\\\"I suppose you use the semi-supervised setting as in GCN and GAT.\\\"\", \"r1\": \"We will report the exact Geom-GCN architecture on each dataset in the new version. We conduct a fair parameter search for every method (i.e., GCN, GAT, Geom-GCN) on every dataset, thus their architecture is different on each dataset. We will release the code of Geom-GCN after acceptance.\", \"q2\": \"\\\"You did not mention exact GCN architecture, i.e. hidden unit, output size, etc, other than a grid search, which makes it somehow hard to reproduce results.\\\"\", \"title\": \"Thanks a lot for your comments!\"}", "{\"comment\": \"I think the motivation is definitely novel in that it extends the conventional definition of neighbourhoods in graphs and hence addressed the problems of structural information preservation in GCN.\\n\\nHowever, I do find the experiments are not well described. For example, \\n1. I suppose you use the semi-supervised setting as in GCN (Kipf and Welling 2017) and GAT (Velicikovic 2017), but you did not mention it in the paper, is that true? \\n2. You did not mention exact GCN architecture, i.e. hidden unit, output size, etc, other than a grid search, which makes it somehow hard to reproduce results. \\n\\nNonetheless I think it is an interesting and technically inspiring paper on GCN.\", \"title\": \"Novel idea but somehow unclear description of experiments\"}" ] }

[ "Gregoire Morin", "Ryan Razani", "Vahid Partovi Nia", "Eyyub Sari" ]

[ "ternary quantization", "binary", "accuracy", "quantization depth", "resource hungry", "low bit quantization", "common", "resource requirements", "flexible model" ]

{ "note_id": [ "CiCsKPUSHX", "S6PlF27iLaS", "ryxHK6Hhir", "SygAmaS3sH", "B1e4AFrhoH", "r1l2k3kziH", "BJeJHadbcS", "rJgg6GphYH", "r1eEnyfrYS" ], "note_type": [ "decision", "official_review", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752821, 1575560112656, 1573834109494, 1573834021917, 1573833163989, 1573153764147, 1572076854580, 1571766968480, 1571262379955 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2588/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2588/Authors" ], [ "ICLR.cc/2020/Conference/Paper2588/Authors" ], [ "ICLR.cc/2020/Conference/Paper2588/Authors" ], [ "~Lukas_Enderich1" ], [ "ICLR.cc/2020/Conference/Paper2588/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2588/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2588/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper studies mixed-precision quantization in deep networks where each layer can be either binarized or ternarized. The proposed regularization method is simple and straightforward. However, many details and equations are not stated clearly. Experiments are performed on small-scale image classification data sets. It will also be more convincing to try larger networks or data sets. More importantly, many recent methods that can train mixed-precision networks are not cited nor compared. Figures 3 and 4 are difficult to interpret, and sensitivity on the new hyper-parameters should be studied. The use of \\\"best validation accuracy\\\" as performance metric may not be fair. Finally, writing can be improved. Overall, the proposed idea might have merit, but does not seem to have been developed enough.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper studies mixed-precision quantization in deep networks where each layer can be either binarized or ternarized. The authors propose an adaptive regularization function that can be pushed to either 2-bit or 3-bit through different parameterization, in order to automatically determine the precision of each layer. Experiments are performed on small-scale image classification data sets MNIST and CIFAR-10.\\n\\nThe proposed regularization method is simple and straightforward. However, many details are not stated clearly enough for reproduction. E.g, since the proposed regularization already promotes binary or ternary weights, whey is there still a thresholding operation at the end of Section 3? Is it because the proposed regularization can not provide strict binary or ternary weights? Does the method require one more hard binarization/ternarization step after \\\\beta is learned. Indeed, tan(x) is not well-defined when x=pi/2, and the derivative tan'(x)= 1+tan^2(x) can be large when x is near pi/2, and does gradient descent work well in this case?\\n\\nThe experiments are only performed on small-scale data sets. Thus it is hard to tell if the proposed method also works for larger networks or data sets? Moreover, it is not fair to use \\\"best validation accuracy\\\" for comparison with other methods since the validation set is seen during training and it is not clear if the hyper-parameters of the proposed methods are tuned for best performance on the seen validation set. It would be more fair to compare the test accuracy like in the BinaryConnect (BC) paper. Yet another concern is that many recent methods that can train mixed-precision networks are not compared. For instance, the HAQ method [1] searches for precision for each layer using the reinforcement learning method, how does the proposed method perform when compared with it?\\n\\n[1]. Wang, Kuan, et al. \\\"HAQ: Hardware-Aware Automated Quantization with Mixed Precision.\\\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\"}", "{\"title\": \"Answer to Official Blind Review #3\", \"comment\": \"Thank you for your comments. We are running experiments for larger datasets and more architectures.\\n\\nWe will provide run-time performance disucssion in the conclusion section. \\n\\nWe will add sensitivity study about these parameters in the next version or on our github codes. \\n\\nMinor comments will be corrected in the next version of the paper.\"}", "{\"title\": \"Answer to Official Blind Review #2\", \"comment\": \"1) We believe that the reviewer missed the main point of the paper here. Most of the works that have been published on mixed precision training assumed that the quantization depth is known before training. Other works that optimize quantization depth use a separate optimizer. The novelty of our approach lies in the fact that the neural network learns quantization depth for each layer only by a modifed back-propagation. We modify back-propagation only by adding a regularization function, and this allows to train weights, and to estimate quantization depth simultaneously.\\n\\n2) We will be looking at parameter definition and interpretation again. We only introduced two new parameters gamma, and beta in which we provided interpretation. Other equations are only the definition of regularization function and they are discussed in detail in reference [16].\\n\\n3) This is a high-level comment. We did our best to provide enough details to reproduce the work. We can improve the descriptions if the missing points are pinpointed. We provide a github code source after acceptance of the paper.\\n\\n4) We agree. We must avoid using the word 'the best' in scientific articles as 'the best' may change through time. We will correct this in the next version.\\n\\n5) We appreciate this detailed comment. We will add some studies about varying lambda and gamma parameters in the next version, or on our github codes.\\n\\n6) The point of these figures is to see that the weights are indeed pushed on binary or ternary values depending on the shape parameter beta of the regularization function. More clear explanations will be added in the next version of the paper, see reference [16].\\n\\nThanks for the generic comments. They will be considered for the revised version.\"}", "{\"title\": \"Answer to Official Blind Review #1\", \"comment\": \"1. We agree that experiments on networks and datasets are minimal. The main goal of this paper was to demonstrate that it is possible to train a neural network that jump between 1-bit or 2-bits while only adding a regularization function to the loss. In the litterature, mostly tuning quantization depth is performed by an independent optimization algorithm like Bayesian Optimization. This is the first attempt to modify back-propagation using only a regularizer that it tunes quantization depth mutually while training.\\n\\n2. Gamma is a hyperparameter which is fixed before training and controls how aggressive the quantization is towards binary weights. Beta is a trainable parameter which modify the shape of the regularization function and enforce weights to be either binary or ternary. The trained beta values, quantify the relative proportion of binary weights to ternary weights. We did not cover beta in the experimental set up because its value change during training.\\n\\n3. The LR was not modified for the described method. LR was changed between the two experiments (LR = 0.01 for MNIST, LR = 0.001 for CIFAR10) to achieve competitive accuracy.\"}", "{\"title\": \"Request regarding the experimental evaluation, especially concerning the accuracy values in the comparisons.\", \"comment\": \"The authors present their work on learning mixed precision bit-sizes for quantization of neural networks. The article is well structured and well written, so it is very easy to read and easy to follow. We are working in the same domain of network quantization, have also published articles and therefore we would like the authors to comment on the following points:\\n\\n1) We are slightly confused about the accuracy values presented in section 4.1 and 4.2. For example, you compare with Ternary Weight Networks (TWN, Li & Liu 2016) and mention 92.74% accuracy on CIFAR-10 (VGG-7) and 99.38% on MNIST, respectively. However, the accuracy values presented in the original TWN paper are 92.56% and 99.35%, respectively. The same applies for BinaryConnect (BC, 2015). So, we are wondering how the authors came up with the mentioned results (which differ from the values reported in the original papers). Did you use your own implementation? Please, explain in more detail.\\n\\n2) The most recent comparison in your paper, which is the comparison with TWN, originates from 2016. From 2016 to 2019 several new approaches for quantization have been published. Therefore, we would like the authors to compare their method with the most recent methods (for example, Variational Network Quantization [1] or Explicit Loss-error-aware Quantization [2]). Moreover, we published a paper about ternary-valued networks at the ESANN 2019 [3] - we used the same VGG-7 architecture on CIFAR-10 and performed clearly better than TWN (93.73%). We would really appreciate if the authors can also compare with our results.\\n\\n[1] Jan Achterhold, Jan Mathias Koehler, Anke Schmeink, and Tim Genewein. Variational network quantization. ICLR 2018.\\n\\n[2] Aojun Zhou, Anbang Yao, Kuan Wang, and Yurong Chen. Explicit loss-error-aware quantization for low-bit deep neural networks. CVPR 2018.\\n\\n[3] Lukas Enderich, Fabian Timm, Lars Rosenbaum, and Wolfram Burgard. Learning Multimodal Fixed-Point Weights usingGradient Descent. ESANN 2019.\\n\\n3) Additionally, it would be really nice if the authors could compare their method with smaller and more optimized networks like ResNet-56 oder DenseNet (both less than 1M parameters) to investigate whether their approach also yields good results even for networks that a difficult in terms of quantization.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The Paper talks about the Smart Ternary Quantization method that improves the quantization over binary and ternary quantizations by specifying an adaptive quantization. The proposed regularization function is covered in detail and the results are evaluated on MNIST and CIFAR10 datasets\\n\\nThe authors can improve the submission by \\n1. evaluating more modern networks with bigger datasets, as opposed to the ones demonstrated.\\n2. describing gamma (eq 10) , it wasn't clear why that parameter was introduced (in addition to the beta) and it's significance, what was more confusing is coverage for this instead of beta in the experimental setup\\n3. why the LR needed to be modified for the described method\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper presents an approach where the regularisation is used to optimise whether each layer of a DNN is binary or ternary. The paper presents an equation for performing this along with two examples of the process in use.\\n\\nThe paper is inconsistently written with work described at different levels in different sections and has an inconsistent feeling about it. For example the introduction seems to stop abruptly before it describes all the parts of the paper.\\n\\nThe paper seems to contain an idea which might have merit. However, the idea just does not seem to have been developed enough.\", \"major_concerns\": \"1) The authors claim that this is the first attempt at a training algorithm for mixed precision training. However, a simple google search throws up many papers in this area. Many of which are not mentioned in the related work.\\n\\n2) Equations are not discussed in enough detail, nor are the parameters defined. Or if they are defined they are done so much later in the work.\\n\\n3) There doesn\\u2019t seem to be enough material here to reproduce the work.\\n\\n4) In the results you talk about \\u2018the best\\u2019. Given that there has been much criticism over the last two years about good academic practice the fact that you don\\u2019t say at least \\u2018best out of \\u2026\\u2019 is worrying. \\n\\n5) You have magic parameters lambda and gamma. You say that these effect the outcome of the work but in your examples you only state values these are set to. One would expect to see at least some analysis of how varying these values effect the outcome. But better would be to show that you have identified good values for both of these parameters. Ideally would be an evaluation of how others could identify the best values.\\n\\n6) Figures 3 and 4 are difficult to interpret. They need a clear explanation.\", \"some_more_generic_comments\": [\"The abstract seems to assume a huge prior knowledge by the reader.\", \"\\u20181 bit precision\\u2019 - precision seems to have no meaning in this context. Surely just \\u20181 bit\\u2019\", \"The related work contains a lot of equations, but no real explanation of what they are.\", \"\\u2018we let \\u03b2 very per layer\\u2019 -> \\u2018we let \\u03b2 vary per layer\\u2019\", \"In equation 10 what do I and J represent?\", \"\\u201928 \\u00d7 28 gray-scales images\\u2019 -> \\u201928 \\u00d7 28 gray-scale images\\u2019\", \"\\u2018For the training session, we pad the sides of the images with 4 pixels, then a 32 \\u00d7 32 crop is sampled, and flipped horizontally at random.\\u2019 - why?\", \"\\u2018As commonly practiced\\u2019 - by whom?\", \"\\u2018which is costly, specially if\\u2019 -> \\u2018which is costly, especially if\\u2019\"]}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper discusses a generalization to low bit quantization and combines the approaches of binary and ternary quantization methods. Past methods such as Binary Connect and Binary Weights Network have shown that you can train a network efficiently with 1-bit quantization, and methods such as Ternary Weights Network demonstrate 2-bit quantization with weights taking one of {-1, 0, 1} * mu, with mu being a scale computed per weight tensor. The authors generalize these two methods so that the choice of binary vs ternary weights can be made per layer automatically during training. The primary contribution to make that work is by incorporating a generic regularizer with addition hyper-parameters to trade-off between the binary weight regularizer and ternary weight regularizer. In addition to that, the regularization also includes a prior to make the layers prefer binary weights by default. This is done by adding a cost that penalizes the choice of ternary weights for each layer.\\n\\nOverall, the paper is well written and explained, with supporting experiments to show on MNIST and on CIFAR10 that this method performs quite competitively compared to an all-binary or all-ternary weights network. Low-bit quantization is an important research area and this paper makes a strong contribution by studying mixed-precision low-bit quantization. The mathematical explanations of the generalized regularizer are well justified. For instance, the regularization constant lambda is normalized for each layer by the total number of weights to evenly weight all layers.\\n\\nAlthough the experiments cover MNIST and CIFAR10, it's not clear how mixed-precision low-bit methods perform on models more prevalent in the real-world. Supporting the experiments with ResNet variants on ImageNet would help clarify that further. The paper very well explains the fundamentals of 1-bit and 2-bit methods, and the contribution of this paper (generalization of these two methods) is a somewhat natural extension without significant novelty. Moreover, in addition to the accuracy, it would also be better to understand how the run-time performance of such models (on existing software and hardware implementations) compares to pure binary and ternary networks, as knowledge of that would reveal insights into systems optimizations to be made in future work in this field. Given all of this, my rating is a weak accept.\", \"pros\": [\"The problem and the fundamentals (prior work) are very well laid out.\", \"The regularization component that generalizes the two types of quantization is sound.\", \"Experiments on CIFAR10 strongly show improved accuracy and higher compression ratio compared to ternary weights network.\"], \"cons\": [\"Lacking more realistic experiments on larger datasets such as ImageNet and models like ResNets.\", \"The runtime performance of how STQ compares to 1-bit and 2-bit quantization variants isn't shown.\", \"The regularizer introduces more hyper-parameters to tweak and it's not clear how sensitive these are to the choice of the architecture. Different values of gamma are used in the two experiments, and further analysis on how the performance varies for various values of lambda and gamma would shed further insight.\"], \"minor_comments\": [\"In equation (3), the term under argmin should be mu and not alpha.\", \"Section 3, line above equation (9) reads \\\"very per layer\\\" instead of \\\"vary per layer\\\"\"]}" ] }

[ "Samia Mohinta", "Rui Ponte Costa", "Stephane Ciocchi" ]

[ "hippocampus", "continual learning", "navigational strategies", "dpca", "spatial continual learning", "hippocampal neuronal representations", "continual", "spatial memory" ]

{ "note_id": [ "uZDExoLfq", "ryxgFEOjsB", "rkl8ENusor", "S1gJ_Muijr", "ByxQVQroiB", "S1lC7WG5iS", "r1xep1EJ9B", "r1gnCnjpYH", "SyedURFjFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752791, 1573778552077, 1573778477950, 1573778022925, 1573765931247, 1573687589537, 1571925943878, 1571826900397, 1571688016306 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2587/Authors" ], [ "ICLR.cc/2020/Conference/Paper2587/Authors" ], [ "ICLR.cc/2020/Conference/Paper2587/Authors" ], [ "ICLR.cc/2020/Conference/Paper2587/Authors" ], [ "ICLR.cc/2020/Conference/Paper2587/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2587/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2587/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2587/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper analyzes neural recording data taken from rodents performing a continual learning task using demixed principal component analysis, and aims to find representations for behaviorally relevant variables. They compare these features with those of a deep RL agent.\\n\\nI am a big fan of papers like this that try to bridge between neuroscience and machine learning. It seems to have a great motivation and there are some interesting results presented. However the reviewers pointed out many issues that lead me to believe this work is not quite ready for publication. In particular, not considering space when analyzing hippocampal rodent data, as R2 points out, seems to be a major oversight. In addition, the sample size is incredibly small (5 rats, only 1 of which was used for the continual learning simulation). This seems to me like more of an exploratory, pilot study than a full experiment that is ready for publication, and therefore I am unfortunately recommending reject.\\n\\nReviewer comments were very thorough and on point. Sounds like the authors are already working on the next version of the paper with these points in mind, so I look forward to it.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanking the reviewer\", \"comment\": \"We thank you for your time and comments.\"}", "{\"title\": \"Replies to reviewer comments\", \"comment\": \"We would like to thank the reviewer for the thoughtful comments. Our replies to the major issues pointed out are as below:\\n\\n(1) We agree that space is crucial when considering the hippocampus and this is very much an element that we are currently investigating. In our present analysis we only considered activity around the reward zone 2 seconds before and 4 seconds after entering the reward zone (reward sensor point). During this time we think that the importance of any spatial component is reduced as this is only around the reward zone.\\n\\n(2) We agree with the reviewer that our initial results need further analysis. However, we should note that the components found were consistently found across different animals, and changed with learning in a way consistent with theoretical ideas. We are in the process of further testing the statistical relevance of our results.\\n\\n(3) Similar to before, we agree that a closer comparison between RL methods and our observations is needed, and we are working on this. We thought that it would be of interest to the community to share our initial results.\\n\\n(4) The reviewer is right that model-free reinforcement learning models suffer from catastrophic forgetting and fail to adapt to changing environments. However, to our knowledge no previous attempts have been made to directly compare animal behaviour with RL algorithms in a continual learning setting. We are using this as a framework to test different algorithms. We are currently testing recently proposed continual learning algorithms, which are less prone to the catastrophic forgetting problem and are very much interested in comparing these methods with our observations both behaviourally and in terms of the representations learnt.\\n\\n(5) We agree that the current focus of the paper is towards understanding how biological systems solve these tasks. However, if one can provide insights into what elements seem to be relevant in neuroscience, these are also likely to be of relevance for machine learning neural networks. More generally, this can be seem as a framework to contrast biological and artificial algorithms, to identify similarities and differences, which may enable the field to identify new solutions. In a following version of this work, we will try to highlight insights that may be useful for machine learners.\", \"our_replies_to_the_additional_smaller_problems_reported_by_the_reviewer\": [\"Neural activity data from four Long Evans rats were used for the dPCA analysis. However, for the behavioural experiment simulation, we only use behavioural patterns in terms of which trials were allocentric and egocentric during the actual experiment of a single animal SC03. The results are consistent across animals, but we did not have the space to show all these results.\", \"We should have clarified this. The neural recordings were obtained using multiple tetrodes in dorsal and ventral CA1 hippocampus. A standard spike sorting mechanism was used as described in Ciocchi et al. 2015, which for our analysis were converted into firing rates.\", \"The first 250 trials from the animal contained the trial blocks for allocentric and egocentric tasks. Hence, the network was trained to perform the tasks using exactly the same experimental trials used in the animal. We also used a experience replay buffer. We will provide this details in a update version of the ms.\", \"We will try to improve this in the next version. The x-axis shows the trial numbers from 4 trials before the allocentric task switches to an egocentric task at trial 111. In Fig 7B Moving averages of trial outcomes from the next 56 trials after task switch are plotted to show how fast the models or the animals learn to reach correct reward locations in the egocentric task. The colored dashed lines at trial number 106 to 110 show the average performance accuracy values from the end trials for the allocentric task. Similarly in Fig 7C, the relearning rates of the models and the animals after a task switch from an egocentric to an allocentric task is plotted against the trial number around which the task switch takes place. From these curves, we aim to show that an animal can almost immediately recognize that the task type has changed as indicated by performance higher than chance level just after a switch. RL models and neural networks, on the contrary, take significant number of trials to recognize the switch and relearn and act accordingly to the new task type. This points to the belief that the hippocampus CA1 perhaps stores information of allocentric and egocentric tasks separately and can fall back to either of those when required or there is a mechanism to consolidate information in a single neuronal cluster and leverage previously learned information to perform new tasks. Standard neural networks fail to maintain distinct values for the changing tasks even when trained over hundreds of trials.\"]}", "{\"title\": \"Replies to the reviewer comments\", \"comment\": \"We thank the reviewer for the thoughtful comments. Our replies to the major critiques are as below:\\n\\n1. We agree that space is likely to be crucial here, and are in the process of including this in the analysis and see how this changes our components. We have tried to reduce the importance of space by only considering activity around the reward zone.\\n\\n2. We have decided to include our initial analysis of this results, but are now working on quantifying this more closely.\\n\\n3. We thank the reviewer for this comment. But we should note that we only claim that our visual inspection is suggestive, far from conclusive. We are currently doing a quantitative comparison.\\n\\n4. The dPCA results are based on recordings from hippocampus CA1 neurons of all four rodents available. For the continual learning simulation, we use the behaviour of one typical animal SC03 to train and test the Q-learning models (Tab-Q and DQN). We took this approach because of space constraints and also due to the stereotypical behaviour across the four different animals. However, we are currently working on including the continual learning results from all animals and agents.\\n\\n5. Thank you for the comments. We should point out that we do not think that the DQN is a good model. The point we are making is that even though it achieves a relatively good average performance it completely fails in remembering the previous task (i.e. fails in continual learning). We are now implementing a DQN with elastic weight consolidation, which was designed to deal with continual learning problems. Next, we will be comparing the representations of the different methods in a more quantitative/precise way.\"}", "{\"title\": \"Replies inline with the reviewer comments\", \"comment\": \"We would like to thank the reviewer for the helpful comments. Our comments are under the headings as in the detailed comments section.\", \"introduction\": \"\", \"reply\": \"We thank the reviewer for pointing out this lack of clarity, which is partially due to space constrains. Early trials are in the initial stages of learning, while the late trials correspond to after a few days of training.\", \"analysis_of_neural_data\": \"\", \"comparison_between_standard_rl_and_hippocampal_representations\": \"Reply to comments of Fig 4(A)i: Thank you for highlighting this, we agree that some of the elements need a better description. In Fig 4A(i), the x-axis corresponds to discretized timesteps. In this panel we are showing the elements of a simple TD algorithm in a 1D environment, so the x-axis represents timesteps in this 1d environment -- each timestep is also considered to be a new state. This simple setup helps us to demonstrate the key properties (eligibility trace and value function in this case) of a TD model in the presence of positive reinforcement, which could then be compared with the CA1 neural activity patterns. The black curve represents the value fuction over time (each timestep is a different state). We do agree that it would have been better to use the same algorithmic framework throughout (so similar to Q-learning as used in C) and we are currently working on this. The reason for having TD-learning first in 1D environment and then 2D, was to try to keep the model as simple as possible.\", \"reply_to_comments_on_fig_4c\": \"We are going to clarify these points in a next version of the ms. In Fig 4C, the lines are obtained by taking the average of Q-values from allocentric correct/incorrect and egocentric correct/incorrect trials. These average Q-values are then plotted against the four timesteps required to reach the reward/no reward terminal states in the 5x5 grid. Calculated average values from allocentric and egocentric trials turn out to be distinct and hence the separation between the lines.\", \"reply_to_comments_on_policy_information\": \"The starting points have distinct Q-values for the north and south points. However, since the experiment involves allocentric and egocentric tasks where both tasks can start from either north or south, a switch in the task overwrites the Q-values of the start locations pertaining to the previous task. Hence, the model forgets the Q-values of north and south start points learnt for an allocentric task when starts performing an egocentric task from the same start points. We will try to clarify this.\", \"reply_to_comments_of_q_learning\": \"The Q-learning models (DQN and Tab-Q) take about 50 trials with changing start locations an allocentric task and around 65 for an egocentric task to converge to perfect performance. However, when we have these strategies interleaved between each other, the agent finds it difficult to learn both tasks one after another and we can see catastrophic interference.\", \"reply_to_suggestions_on_using_hierarchical_rl\": \"Recently, we have implemented a DQN with Elastic Weight Consolidation as in the paper by Kirkpatrick et al 2017. This model should be able to deal to perform continual learning and model the behavioural experiment with switching allocentric and egocentric tasks. We are currently analysing the results of this model.\", \"comparing_dqn_and_animal_performance\": \"\", \"reply_to_comments_on_dqn\": \"We first tested the network without task-specific information, but in this case the network performance is even worse. We added task-specific information to see whether the network could perform the task. We agree that this is a close representation of the experimental setup, but this further suggests that some other element is needed for these networks to be able to solve continual learning, as not even task-specific information appears to be sufficient. In a future version we will include both models (with and without task-specific information).\", \"reply_to_dqn_animal_comparison\": \"This boils down to the catastrophic interference issue in systems such as neural networks. As the reviewer mentions the network can indeed learn quicker each task, but fails to retain any information about previous tasks.\", \"reply_to_comments_for_fig_7\": \"Thanks for this suggestions. We are already working on this.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"This work sits at the intersection between neuroscience and machine learning. It proposes to use neural recordings from the rodent hippocampus to shed light into how biological agents achieve continual learning. The approach involves (i) analyzing neural data from the rodent hippocampus with the goal of identifying how neurons encode various behavioral variables; (ii) training different RL agents to solve a computer version of the same task, including tabular and DQN agents; (iii) contrasting the performance and representations of the artificial agents with those of animals.\", \"the_paper_has_an_admirable_goal\": \"to find links between neuroscience and machine learning, using tools from one to promote advances in the other. When done correctly, this interaction can be fruitful and produce incredible insights for both fields. However, I believe that this paper does not deliver on either end due to major methodological and analytical flaws, rendering it innapropriate for ICLR. Below I list my major critiques:\\n\\n1. The interpretations of the dPCA components seems very preliminary and lacks methological rigor. In particular, many alternative interpretations of the components are possible. For instance: Component #3 might represent outcome (rather than decision); Component #4 might represent greater engagement of the hippocampus in the egocentric task (e.g. Packard and McGaugh, 1996); etc. Moreover, the reported analyses do not include animal position, which will almost certainly be a major driver of population activity in the hippocampus (see work on place cells). Since the animal position is not reported, it is impossible to know if the dPCA components (e.g. Component #2) are, instead, representing alocentric position. Including animal position (perhaps also as an explanatory variable in the dPCA) might help, but the authors should also do a more thorough job testing their specific hypotheses beyond simply reporting them based on visual inspection.\\n\\n2. The comparison between early and later training sessions is also rather crude and qualitative. The authors makes several claims about differences that are not tested directly with a proper hypothesis test.\\n\\n3. The comparison between RL quantities and dPCA component is also very weak. For instance, the claim that a given dPCA component represents an eligibility trace needs much more evidence than simply showing that this component decays over space when the eligibility trace also decays. With respect to Component #4 (a critical component presumably related to the authors' hypothesis of multi-task representation), the authors report that, for an e-greedy agent, the average value is higher for alocentric vs. egocentric task. Yet, this difference is not investigated further and, again the authors claim victory based on a simple visual comparison between two plots (4C vs. 4D). \\n\\n4. For most of the paper, the authors report the results from a single (typical?) animal. Ideally, the results for all animals would be reported (or some statistically sound aggregate of all animals).\\n\\n5. Finally, the authors compare the performance of animal 3 with the performance of different RL agents. Again, this comparison is incredibly superficial and neglects many confound variables. The fact that the accuracy of both the animal and DQN is around 70% is not sufficient to claim that DQN is a good model for the animal's behavior, or that it is superior to other models that achieve slightly worse or better performance. Much more work needs to be done to properly compare RL and animals (e.g. comparing, in addition to performance, representations across various layers, prediction error signals, reward signals, etc). \\n\\nOverall, while I commend the authors for an intriguing idea, the execution of the idea is so poor that I consider the paper to be of little interest to either the neuroscience or the machine learning communities. Therefore, I cannot recommend the paper for publication at ICLR.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper describes an analysis of hippocampal neuronal activity in rodents during spatial navigation tasks. Features were extracted from the data using a component analysis technique, and these features were then compared to quantities which arise during training of the DQN reinforcement learning algorithm.\\n\\nUnfortunately I am not well versed enough in this literature to assess the merits of this submission. However, I could not find any glaring issues, unclear sentences or any other obvious sign of incompetence. \\n\\nMy apologies for the inadequacy or this review (and/or the paper assignment algorithm).\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This work focuses on the problem of continual learning. It first proposes an analysis of neural recordings from rodents hippocampus that perform a task related to continual learning. The authors then claim to identify similarities in representations of behaviorally relevant variables between biological networks and standard artificial RL systems. Finally the authors propose a DQN implementation of the task reaching performance similar to rodents, although the implementation fail to perform continual learning.\\n\\nI have found the results of the analysis of neural data interesting and well explained. However I haven\\u2019t been able to judge positively the main part of the manuscript pertaining to the comparison with artificial agents, maybe because of the lack of clarity in the exposition of the results (see detailed comments below). As such I think this manuscript is not ready yet for publication in ICLR.\", \"detailed_comments\": \"\", \"introduction\": \"The introduction reads well, it would be useful to have a short paragraph explaining the vast problem of continual learning, and situating the approach of the authors in this vast field. In particular it would be useful to explain in what sense the task considered is related to continual learning (switches between allocentric and egocentric tasks that are not informed by experimenter, but that the rodent has to figure out), because for now it is only briefly stated in the legend of figure 1.\", \"analysis_of_neural_data\": \"Analysis of neural data using dPCA reveals interesting results regarding the representations of correct/incorrect trials, allocentric/egocentric tasks, temporal structure. Globally these results are well presented. Regarding figure 3, it would be nice to define what are early and late trials, is it a distinction inside a block where late trials are just before a switch. Or are late trials after weeks of training, in which case I do not really see the link with continual learning ?\", \"comparison_between_standard_rl_and_hippocampal_representations\": \"The idea to compare hippocampal representations and standard RL implementations is original and interesting. The exposition of the results, however, lacks important information for me to assess the relevance of these comparisons:\\n\\nFig4A(i): could the author explain what is on the x-axis ? What is the 1-D environment mentioned in the caption (in SM a 5*5 grid world is mentioned) ? Why not run TD on the same setting as Q-learning ? What is the black curve \\u00ab\\u00a0value after learning\\u00a0\\u00bb ? Why value depends on the x-axis ?\", \"fig4c\": \"How are the two curves obtained ? As both tasks share the same Q function, it might be worth explaining how the two curves are obtained.\\n\\nThe authors mention that \\u00ab\\u00a0policy information is override as starting location changes over\\u00a0\\u00bb ? Could this fact be explained ? It would seem to me that starting points north s_north and south s_south would have distinct Q(s_north,.) and Q(s_south,.) avoiding overriding.\\n\\nAlso I am quite surprised that Q-learning fails on such a simple task (cf Fig5B), could the authors explain how many trials would be required to reach perfect performance by focusing on either of the two tasks ?\\n\\nIn this section a natural extension could be to use hierarchical RL, or options to model behavioral strategies such as allocentric and egocentric and compare with the hippocampal recordings.\", \"comparing_dqn_and_animal_performance\": \"A first crucial point to clarify is whether the DQN receives information about what task to perform (it is mentioned: \\u00ab\\u00a0task-specific information provided to the neural network as input\\u00ab, \\u00abthis network with task-specific memory\\u00ab, then a DQN \\u00ab\\u00a0without task-specific information\\u00a0\\u00bb is mentioned). This seems rather crucial, because if information about allocentric VS egocentric task is provided, contrary to the behavioral task for rats, then the network is asked to do strategy switches based on some cues, which is a different problem than continual learning.\\n\\nIf the DQN receives information about whether it is in a allocentric or egocentric trial, I am very surprised it is not able to perform this simple task.\\n\\nCould the authors explain why the DQN has similar performance as the animal while it is shown to relearn slower than animals (is it because it learns better before task switch). \\nIn figure 7, the curves for the model are very noisy and hence difficult to interpret, it would be nice to show averages over many same task switches, to get a clearer picture.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper analyses a dataset of representations in the CA1 region of the hippocampus of a rat conducting a spatial plus maze task that switches between allocentric and egocentric versions. In the allocentric version of the task, the rat must always go from north or south arms to the west arm to receive a reward. In the egocentric version the rat must always turn right, regardless of whether they are starting in the north or south arm. These two versions of the task are switched after some period of time.\\n\\nThe researchers conducted demixed principal components analysis (dPCA) on the data, then compared the activity of the components to aspects of a tabular Q-learner. They argue that some of the components match variables used by the Q-learner. Following this, they examined the ability of both the tabular Q-learner and a deep Q-network to perform this task, and compared their performance to the rat. They find that the rat is better than the Q-learners at continually updating from the egocentric to allocentric task, and vice-versa.\\n\\nThe goals of this paper are fantastic. I really like the attempt to link hippocampal activity with RL representations. But, this study is very muddled, and there are very serious problems with the paper that render it inappropriate for acceptance to ICLR. The five most major issues are:\\n\\n(1) The choice of the demixing categories is missing a crucial category: space! Given the importance of the hippocampus for spatial representations, surely a big component of the variance can be explained by the animal's location in space. Why not include this? It might in fact be that some of the \\\"time\\\" components are just reflecting the usual spatial location of the animal during different components of the task.\\n\\n(2) The identification of components from the neural data with things like reward prediction errors, eligibility traces, etc., is all done in a qualitative manner with no statistical controls. This lack of quantitative assessment for some of the key claims in the paper is just not okay for an academic paper. Moreover, even the qualitative claims are underwhelming. The authors' claim, for example, that time component #1 is an eligibility trace is a real stretch, in my opinion (see more below in point 3).\\n\\n(3) The comparison between the animal data and the Q-learner is also done in a qualitative fashion that was extremely underwhelming. Fig 4 A & B were the most egregious. Are the authors really attempting to claim that the curves in 4A are clearly related to the curves in 4B? That's a real stretch, and to provide no quantitative assessment of this claim renders it completely unconvincing.\\n\\n(4) The fact that model-free reinforcement learning algorithms cannot adapt in changing environments/tasks has been known for a long time. As such, the result showing that the Q-learners cannot switch easily between the tasks is not novel. See, for example, this paper: https://www.sciencedirect.com/science/article/pii/S0896627313008052\\n\\n(5) Even if we accept the central claims from this paper, there is very little provided for machine learning researchers at ICLR to benefit from. What about the hippocampal representations makes the rats better at continual learning? What inductive biases or memory mechanisms might we glean from this work? Nothing like that is provided. As it stands, even being charitable, this paper really only speaks to a neuroscience audience, since even the Q-learning components are used only to understand the neuroscience data, not to think about how this could inform new ML systems or theories.\\n\\nIn addition to these problems, there are several small ones:\\n\\n- Was only one animal included in this analysis? That's never stated, but Figs. 5 & 7 seem to suggest that. Not only does this 100% need to be stated, but it's very problematic from a generality standpoint.\\n\\n- What type of recordings were these? How were individual cells identified (e.g. spike sorting)?\\n\\n- How was the Deep Q-net trained exactly? The authors say it was trained on the first 250 trials from the animal, but that's confusing? Was it not trained to perform the task itself? Also, where is all the info on memory buffers, hyperparameters, etc. There is no way to reproduce these simulations given the lack of detail here.\\n\\n- Some of the plots are confusing and hard to follow. For example, in figure 7 B and C, what determines the X-axis? What should I be looking for in the curves?\"}" ] }

[ "Lorenzo Luzi", "Randall Balestriero", "Richard Baraniuk" ]

[ "generative adversarial networks", "goodness of fit", "inception score", "empirical approximation error", "validation metric", "frechet inception score" ]

{ "note_id": [ "paDqAd9O-j", "Sklvl9H5sB", "SyeYCUzKsB", "H1x3_8ztjH", "r1lF7UfFsS", "SygOTfanYH", "HJlxTZnitB", "Bkx7i9i1KH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752762, 1573702126983, 1573623504753, 1573623412133, 1573623328993, 1571766976508, 1571697079725, 1570908826631 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2586/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2586/Authors" ], [ "ICLR.cc/2020/Conference/Paper2586/Authors" ], [ "ICLR.cc/2020/Conference/Paper2586/Authors" ], [ "ICLR.cc/2020/Conference/Paper2586/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2586/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2586/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes to measure the distance of the generator manifold to the training data. The proposed approach bears significant similarity to past studies that also sought to analyze the behavior of generative models that define a low-dimensional manifold (e.g. Webster 2019, and in particular, Xiang 2017). I recommend that the authors perform a broader literature search to better contextualize the claims and experiments put forth in the paper.\\n\\nThe proposed method also suffers from some limitations that are not made clear in the paper. First, the measure depends only on the support of the generator, but not the density. For models that have support everywhere (exact likelihood models tend to have this property by construction), the measure is no longer meaningful. Even for VAEs, the measure is only easily applicable if the decoder is non-autoregressive so that the procedure can be applied only to the mean decoding. \\n\\nIn this current state, I do not recommend the paper for submission.\\n\\nXiang (2017). On the Effects of Batch and Weight Normalization in Generative Adversarial Networks\\nWebster (2019). Detecting Overfitting of Deep Generative Networks via Latent Recovery\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks for the response\", \"comment\": \"Thanks for the detailed response.\\n\\nI think it is debatable to claim \\\"generative models should have F = 0\\\", especially when the generative models have covered the high-density areas. \\n\\nOn the other hand, it is crucial to distinguish models which only memorize the training data but generate randomly beyond that and models which cover the data manifold well and generate reasonably. The proposed metric seems to favor the first than the latter, which is counter-intuitive. Therefore, metrics that cover both precision and recall would be more sensible in evaluation.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": [\"Thank you for the feedback on our paper. We address your concerns as follows:\", \"For generative networks that map a latent space of \\u201clow\\u201d dimension to a space of \\u201chigh\\u201d dimension, the range of the network does not cover the entire space, because two spaces are of differing dimensions. Hence, we have that p(x) = 0 can occur for training points. This is true for VAEs as well as GANs as we mentioned in the paper.\", \"If generative models can generate high-quality images but cannot generate the training set, then the learned distribution is likely only focused on a few modes of the true data distribution. This means that the learned distribution might cover a very small part of the data distribution well, which is not desirable since we want to learn the entire data distribution. Our main contribution is the definition of a metric that enables us to measure how well a generative model memorizes the training set.\", \"Mode collapse can be defined with respect to different distributions. With respect to the empirical data distribution, F being 0 is necessary and sufficient to avoid mode collapse. As you mention, with respect to some true data distribution that was sampled to obtain an empirical distribution, F = 0 is only necessary to avoid mode collapse.\", \"If the latent space is much smaller than the output space, then it is not trivial to have F = 0 because the generator cannot reproduce any output image. If F = 0, then you can compare probabilities for generated images against an oracle data distribution. In many toy 2D examples, you have oracle access to the true distributions because the distributions are simulated. However, we are concerned with image datasets where we only have access to an empirical distribution. Hence, mode collapse must be compared with the empirical distribution because we do not have access to an oracle distribution.\", \"The results with the mixture model are not obvious because they imply that using less data can lead to better performance. Specifically, this implies that DCGAN-type architectures are underparameterized.\", \"Let us know if you have any further concerns about our paper, and thank you for the helpful feedback.\"]}", "{\"title\": \"Response to Reviewer #1\", \"comment\": [\"Thank you for the feedback on our paper. We address your concerns as follows:\", \"It is true that if F>0 then we can still have memorization, because we are taking an average. In practice, this does not happen, as we captured in the histograms of Figure 4 and Figure 8 in the paper. We observe that the distribution of distance is relatively symmetric and unimodal, making the average a very informative measure of memorization. In addition, we are mostly concerned with having complete memorization of the data instead of just partial memorization. For partial memorization scenarios, we agree that variations on our metric (such as minimum distance) could be very useful as well.\", \"We consider generating the training set as a first step toward understanding important issues like mode collapse. Of course, our measure alone will not be used to directly evaluate the fidelity of generated samples. High fidelity samples are desirable, but if F > 0 for these models, then that means that they are not learning the simplest distribution of all: the empirical distribution. Hence, generative models should have F = 0, which implies that there is no mode collapse with the empirical distribution.\", \"We are comparing the support of G to the training set only (and not the probability densities), because we are focusing on the simpler, yet necessary, topic of memorization in generative networks. If a generative network cannot learn the training set, then there exists an image x such that the probability of generating x is equal to 0. Thus, a probabilistic distance of x from the distribution of Imag(G) is related to the distance between the image of G and x, which is our approach.\", \"Let us know if you have any further concerns about our paper, and thank you for the helpful feedback.\"]}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thank you for the feedback on our paper. We address your concerns as follows:\\n\\n1. Defining a metric between images is a long-standing problem in image processing, because standard p-norms do not capture image structure well. We use squared distance (2-norm) because it has a long history in not only signal and image processing but beyond. Nevertheless, other metrics can be used within our framework if there is prior information that implies that a particular metric would be superior to others. For this reason, we do not claim that the 2-norm is optimal for images.\\n\\n2. We follow standard methods for optimizing over the latent space of a generative network such those from the GLO paper. This enables us to have a certain confidence in the optimization problem because this method \\u201c... recovers the true latent vector 100% of the time to arbitrary precision.\\u201d [1] Although, this is not theoretical guarantees, their empirical performance is a convincing argument for why they should be used for calculating F in practice. Developing theoretical guarantees for this nonconvex optimization problem is beyond the scope of this paper and would be an interesting paper by itself. \\n\\n3. The robustness of calculating F is measured with different initializations as shown in Figure 6. If we optimize over the latent space with different initial distributions of z, we find that the statistics of the solution z* are the same. Hence, the calculation of F is robust to different initial distributions of z, which means that even unlikely z\\u2019s will converge to a z* that has typical statistics. \\n\\nLet us know if you have any further concerns about our paper, and thank you for the helpful feedback.\\n\\n[1] Zachary C Lipton and Subarna Tripathi. Precise recovery of latent vectors from generative adversarial networks. arXiv preprint arXiv:1702.04782, 2017.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper defines a goodness of fit measure F for generative networks, that reflects how well a model can generate the training data. F allows to detect mode collapse: as long as it is strictly positive, mode collapse is observed as parts of the training data have not been memorized. It aims at providing an alternative to the Fr\\u00e9chet Inception Distance and the Inception Score that rely on pretrained neural networks (whereas this new measure does not). It also provides insight into the DCGAN and WGAN networks in that regard, observing for instance that data subsampling helps decrease F, which motivates the use of a mixture of GANs.\\n\\nThis paper brings an interesting contribution to the evaluation of generative networks. However:\\n\\n1.\\tThe use of the square distance in the image space is not obvious and not justified.\\n2.\\tComputation of this metric is not straightforward: there is no theoretical guarantee and it is computationally expensive.\\n3.\\tThe theoretical properties of this measure and its robustness are not investigated.\\n4.\\tTypos are obscuring the reading of the paper.\\n\\n- Post rebuttal: I have read the authors' response and am maintaining my weak reject rating.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": [\"This work proposed a new goodness of fit measure for generative network evaluations, which is based on how well the network can generate the training data. The measure is zero if the network could perfectly recover the training data, and would represent how far it is from generating the training set in the average manner of the total least square sense, where the one-to-one mapping between the generated data and the training sample is constructed through latent space optimization. Using the proposed measure, the authors showed an interesting trend present in the DCGAN training and the impact of the residual connection. The authors might want to add some discussion in Section 4.2 regarding why the residual connection is detrimental for covering the support. Increasing the model complexity through larger latent space dimension and learning mixtures is proposed as solutions to improve the measure as well.\", \"With all the interesting results presented, I still have the concerns about the sensitivity of the proposed measure:\", \"It is an average over the training data or the selected sample. Above Section 4, the authors argued that \\\"\\\\hat{F}(G) > 0 meaning that we do not observe any memorization\\\". This seems overly assertive. Since the measure is an average over the training data, it has difficulty to differentiate between one network which has almost zero value for part of the training data but large values for the rest, and another network with roughly the same \\\\hat{F}(G) value but small values for all training data. The variance could help, but can not resolve this issue. This would be more important when the training data contains noise or outliers.\", \"It only concerns the generation of the training data, but not the sampled data from the network (at least not directly). Therefore it has no direct control of the fidelity of the generated samples.\", \"As shown by the authors, the proposed measure can be considered as the approximation of the true probability support not covered by the generative models, which also defines a necessary condition to avoid mode collapse. But what about the other part? It would have difficulty comparing two models with the same support but different high-density areas. Indeed, there are existing works which consider both the precision and recall of the generative models [1, 2, 3], and directly work with the generated samples instead of the training data. These should be discussed and compared with, not just the FID scores which have already been shown to have issues [3].\"], \"some_notations\": \"- In the last equation on Page 2, should it be L_{G} instead of L_{D}?\\n- In the first equation on Page 3, should the denominator be N_{B} instead of N_{N}?\\n- \\\"Optimality\\\" in terms of generative models may depend on the downstream tasks. I do not think there exists a universal definition of \\\"optimality\\\" for generative models.\\n\\n[1] M.S.M. Sajjadi, O. Bachem, M. Lucic, O. Bousquet, and S. Gelly. Assessing generative models via precision and recall. NeurIPS 2018.\\n[2] L. Simon, R. Webster, and J. Rabin. Revisiting precision and recall definition for generative model evaluation. ICML 2019.\\n[3] T. Kynkaanniemi, T. Karras, S. Laine, J. Lehtinen, and T. Aila. Improved precision and recall metric for assessing generative models. Arxiv:1904.06991.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposes a new goodness of fit measure for generative models, and uses it to get insight into GAN's. While this is an important topic and a novel approach, I do not think the paper delivers on what it promises.\\n\\nI think this paper should be rejected. First, while it claims to be a general method for generative models it, it is limited to only GANs and even for GANs it is limited. Second, most of the observations are nice but trivial, e.g. larger latent space leads to larger image.\", \"detailed_remarks\": [\"The main point that the training set points x must have p(x)>0 under the model is naturally satisfied for almost all models except GANs such as VAEs, autoregressive model and flow models with standard implementations as the support is the whole space. This is in contrast to the claim in the paper that \\\"its applications can be extended to other generative networks such as Variational Autoencoders.\\\".\", \"Even for GANs as this measure only looks at the support and not the distribution it is not clear if this measure does more then evaluate mode collapse. While this is an important task, it falls short of the promises the authors claim.\", \"The authors claim that \\\"We demonstrate that our measure being minimized is a necessary and sufficient condition to detect mode collapse.\\\" but only show that it is necessary.\", \"Proposition 1 is a trivial statement.\", \"The authors claim that \\\" mode collapse happens if P(x) > 0 but minz ||G(z) \\u2212 x|| > 0\\\". This is a main point by the authors, but it ignores the probability and only looks at the support. It has been shown that mode collapse happens even in 2d distributions, e.g. veegan paper, where it is easy to get the support to be the whole distribution.\", \"The results in sec. 5 are quiet obvious, with a larger latent space you can naturally get a larger support, same as with a mixture model.\", \"In general the method only looks at the support, ignoring the distribution over the support and is therefore very limited in evaluating generative models.\"], \"minor_details\": [\"In eq. 3 the integration should be w.r.t dP(x) for it to be monte-carlo approximated as it is in eq. 4.\", \"Not 100% I understand what the authors try to say here - \\\"we pick the latent variable z and error ||G(z) \\u2212 x||2 that corresponds to the smallest error instead of picking the latent variable that Adam Kingma and Ba (2014) finds.\\\"\"]}" ] }

[ "Fangzhou Mu", "Yingyu Liang", "Yin Li" ]

[ "representation learning", "gradient features", "deep learning" ]

{ "note_id": [ "jt9gk0Eq0", "rJxTJEiojr", "SJlx-Qijjr", "r1l1mGjojH", "SkgVRxoisH", "B1gcXmQr5H", "r1l25t7-9S", "S1xP7SAaFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752733, 1573790692969, 1573790456430, 1573790230803, 1573789900504, 1572315938040, 1572055443927, 1571837214869 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2585/Authors" ], [ "ICLR.cc/2020/Conference/Paper2585/Authors" ], [ "ICLR.cc/2020/Conference/Paper2585/Authors" ], [ "ICLR.cc/2020/Conference/Paper2585/Authors" ], [ "ICLR.cc/2020/Conference/Paper2585/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2585/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2585/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper makes a reasonable contribution to extracting useful features from a pre-trained neural network. The approach is conceptually simple and sufficient evidence is provided of its effectiveness. In addition to the connection to tangent kernels there also appears to be a relationship to holographic feature representations of deep networks. The authors did do a reasonable job of providing additional ablation studies, but the paper would be improved if a clearer study were added to investigate applying the technique to different layers. All of the reviewer comments appear worthwhile, but AnonReviewer2 in particular provides important guidance for improving the paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #2 Part II\", \"comment\": \"[On extra ablation experiments]\\n\\nWe first add an ablation study using an ImageNet pre-trained ResNet-18. The results show that our approach remains effective. More importantly, our model again outperforms fine-tuning, suggesting that the encouraging result on AlexNet is not observed by chance.\\n\\nWe further consider the gradient term alone as another baseline in the two ablation studies. Interestingly, we found that this model is more powerful than the standard logistic regressor on network activation when all network parameters are pre-trained. Moreover, we compare our model against this same baseline in other experimental settings, and find that the full model generally outperforms the gradient model as the dataset and the network grows in complexity. This is highly desired as it makes our method potentially useful in practice.\\n\\nWe present some key results of our ablation study in the supervised setting. Here, gradient feature is w.r.t. the last residual block of an ImageNet pre-trained ResNet-18. We form different combinations of random and pre-trained parameters in the base network for computing the gradient feature, while the activation feature is from the pre-trained network in all scenarios. Please see the appendix of the paper for detailed experimental design and the full set of results.\\n\\nrandom $\\\\theta_1, \\\\theta_2, \\\\omega$: 15.63 (gradient alone) / 82.84 (full)\\npre-trained $\\\\theta_1$, random $\\\\theta_2, \\\\omega$: 62.35 / 82.84\\npre-trained $\\\\theta_1, \\\\theta_2$, random $\\\\omega$: 80.74 / 83.15\\npretrained $\\\\theta_1, \\\\theta_2, \\\\omega$: 83.05 / 83.50\", \"activation_alone\": \"82.65\", \"fine_tuning\": \"82.97\\n\\n[On the difference between our work and the two mentioned by the reviewer]\\n\\nAs can be seen from our general clarification, our work differs significantly from standard final layer transferring approaches, which only use network activation as feature. In fact, we consider it as a baseline in our experiments. Our method, on the other hand, uses additional feature, namely the Jacobian w.r.t. conv layers prior to the fc layer. Similarly, it is easy to distinguish our method from the Bayesian final layer approaches, which again only tune the fc layer.\\n\\n[On the implementation of our algorithm]\\n\\nOur implementation does not require sophisticated software engineering and can be trivially integrated into existing pipelines. Let us consider the toy example above and further assume that $\\\\theta_2$ contains conv3 and conv4. Without loss of generality, we assume ReLU nonlinearity. The PyTorch pseudo-code goes as follows.\\n\\nv3 = Conv2d()\\t\\t\\t# v3 is the counterpart of conv3 in the second term\\nv4 = Conv2d()\\t\\t\\t# v4 is the counterpart of conv4 in the second term\\n\\n# standard forward prop (x is the image input)\\ny1 = relu(conv1(x))\\ny2 = relu(conv2(y1))\\ny3 = relu(conv3(y2))\\ny4 = relu(conv4(y3))\\nz = avgpool2d(y4)\\nout = fc(z)\\n\\n# code for scalable evaluation of the second term in our model\\njvp = v3(y2) * (y3 > 0).float()\\njvp = (conv3(jvp, add_bias=False) + v4(y3)) * (y4 > 0).float()\\njvp = avgpool2d(jvp)\\njvp = fc_copy(jvp, add_bias=False)\\t\\t# fc_copy is a frozen copy of fc at initialization\\n\\nlogits = out + jvp\\t # logits are fed into a softmax for classification\\n\\nIn transfer learning, conv1-4 and fc_copy are kept frozen. fc is the trainable parameter in the first term of our model, and v3, v4 are the trainable parameters in the second term. As the reviewer correctly pointed out, our method does introduce some extra memory overhead because some intermediate output needs to be cached (y3 and y4 in the example). However, it suffices to use gradient from conv4 in practice as we argued previously, hence the memory cost and the number of operations are on par with fine-tuning.\\n\\n[Typos]\\n\\nThank you for pointing out the typos. They will be fixed in our updated draft.\", \"reference\": \"[1] Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent. NeurIPS 19\"}", "{\"title\": \"Response to Reviewer #2 Part I\", \"comment\": \"Thanks for your very detailed comments and for your appreciation of the novelty and significance of our work. We clarify some of the technical details that might have created confusion in the review. We also discuss the choice of layers for computing gradient feature. More importantly, we include additional ablation study on supervised pre-training and demonstrate that our model consistently outperforms all baselines (including using gradient feature only). Finally, we present pseudo-code of our implementation. We hope that our response can address your concerns and kindly ask the reviewer to re-consider the rating.\\n\\nWe present the key results in our response and put the full set of results temporarily in the appendix of the paper. These results will be merged into a future version of the paper. Please do not hesitate to ask if you have further questions.\\n\\n[General clarification]\\n\\nBefore we address any specific concerns, we would like to clarify our approach using a toy example. Recall that our model, which can be intuitively understood as the linearization of a pre-trained network, contains two terms. The first term is the usual logistic regressor on the activation, and the second is linear in the per-sample Jacobian of the activation w.r.t. a set of network parameters, which we denote by $\\\\theta_2$ in the paper. Suppose that a network $F(x)=g(f(x))$ has the architecture (conv1-conv2-conv3-conv4-fc). According to our notation, we refer to (conv1-conv2-conv3-conv4) as the ConvNet $f$ (parameterized by $\\\\theta_1$ and $\\\\theta_2$), and the last fc layer as $g$ (parameterized by $\\\\omega$). When we say \\\"gradient feature from the top layer of $f$\\\", we mean the per-sample gradient w.r.t. the weight and bias of conv4, not the fully connected layer fc. Note, however, that our model still makes use of the gradient w.r.t. fc, which is exactly the activation from $f$ and hence is naturally merged into the logistic regressor term.\\n\\nNow we turn to the concerns raised by the reviewer.\\n\\n[On the choice of the layers for computing gradient feature]\\n\\nIn the experiments, we choose to compute gradient feature from the topmost conv layers (or residual blocks), but not the entirety of the network for the following reasons.\\n\\na) Width requirement. Empirical studies demonstrated that a network is well captured by its linearization only if it is sufficiently wide [1]. The most popular network architectures these days are almost always widest in the top.\\n\\nb) Diminishing gain. According to our ablation studies, the model using gradient from the topmost layer alone is already satisfying. Adding gradient from layers further down the hierarchy results in very little performance gain. This observation can be partially explained by our argument in a). Another sensible explanation is that the layers farther away from the top learned more generic features that are immediately transferable to a new task. Since those layers are already in good shape, one can imagine that their gradients contain very little discriminative power.\\n\\nc) Computational overhead. Similar to fine-tuning, our method becomes increasingly expensive (in terms of both time and space complexity) as we include gradient from more layers. Hence, it helps to restrict ourselves to the topmost layer in order to reduce the computational cost. This choice is further backed by our arguments in a) and b).\\n\\nWe apologize for not sticking with our claim in Sec. 4.1 in some of the succeeding experiments. We have performed additional experiments in which we use gradient only from the topmost layer (or residual block). Below please find a sample of the results.\", \"base_network\": \"ImageNet pre-trained ResNet-18\", \"target_task\": \"VOC07 object classification (mAP)\", \"activation_alone\": \"82.65 (center crop) / 83.59 (ten random crops)\", \"gradient_alone\": \"83.05 / 84.63\\nactivation + gradient: 83.50 / 84.95\", \"fine_tuning\": \"82.97 / 84.14\", \"activation\": \"57.83\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"Thanks for your valuable comments. We discuss additional results of alternative baselines that uses gradient feature alone. These results clearly show the advantage of our method, especially on challenging datasets and large-scale base networks. Moreover, results from our ablation studies rule out the explanation that the performance gain is due to the extra number of parameters. We hope that our response can address your concerns and kindly ask the reviewer to re-consider the rating.\\n\\nWe present the key results in our response and put the full set of results temporarily in the appendix of the paper. These results will be merged into a future version of the paper. Please do not hesitate to ask if you have further questions.\\n\\n[On the alternative baselines]\\n\\n-- Augmenting the baseline with random features of the data.\\n\\nWe indeed have considered this setting in the ablation study of our submission, with the results in the first row of Table 1. Specifically, we augment the activation from a pre-trained network (a BiGAN encoder trained on CIFAR-10) with the gradient from a random network, and the model achieves an accuracy around 63% for all three sizes of $\\\\theta_2$. This is on par with the logistic regressor baseline, which achieves an accuracy of 63.38% by using activation alone from the pre-trained network. We thus conclude that our model's improvement over the standard logistic regressor is NOT a consequence of an increase in the number of parameters. We will revise the description of the experiment to better explain these results. \\n\\n--Just using the gradient as feature.\\n\\nWe agree that using the gradient alone is an important baseline and thank the reviewer for pointing it out. This new baseline is referred to as the gradient model so that it can be distinguished from the full model. Surprisingly, the gradient model achieves impressive results in the case of toy datasets and small networks, sometimes even slightly outperforming the full model. However, the full model consistently outperforms the gradient model in the self-supervised and the transfer learning settings, where large networks (e.g. ResNets) and more challenging datasets (VOC07 and COCO2014) are used. A sample of our results that covers the most interesting scenario (i.e. when all network parameters are pre-trained) is listed below. Please see the appendix of the paper full the full set of results.\", \"base_network\": \"ImageNet pre-trained ResNet-18\", \"target_task\": \"VOC07 object classification (mAP)\", \"activation_alone\": \"82.65 (center crop) / 83.59 (ten random crops)\", \"gradient_alone\": \"83.05 / 84.63\\nactivation + gradient: 83.50 / 84.95\", \"fine_tuning\": \"84.14\\n\\nHence, We conjecture that our method can be more successful as the dataset and the network grow in complexity. This is highly desired as it makes our method potentially useful in practice. Finally, we have to point out that we find using gradient only is less intuitive and hard to interpret in its own right. Instead, the same gradient term acts nicely as a residual term derived from Taylor expansion in our model.\\n\\n[On the ablation study]\\n\\nAs mentioned above, it is evident from our ablation study that our model improves on the baseline not just because it includes more parameters. At the request of Reviewer 2, we include another set of ablation study using a supervised network. The results from this new study further strengthens our conclusion. Please find the new results in the appendix of the paper.\\n\\n[On the discussion of the theoretical connection]\\n\\nWe will follow the reviewer's suggestion to revise this section so that it can fit into two paragraphs. This change will be reflected in a future version of the paper.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thank you for your valuable comments.\\n\\nWe would like to address your concern on the selection of layers for computing gradient feature. In the experiments, we choose to obtain gradient feature from the topmost conv layers (or residual blocks) of the network for the following reasons.\\n\\na) Width requirement. Empirical studies demonstrated that a network is well captured by its linearization only if it is sufficiently wide [1]. The most popular network architectures these days are almost always widest in the top.\\n\\nb) Diminishing gain. According to our ablation studies, the model using gradient from the topmost layer alone is already satisfying. Adding gradient from layers further down the hierarchy results in very little performance gain. This observation can be partially explained by our argument in a). Another sensible explanation is that the layers farther away from the top learned more generic features that are immediately transferable to a new task. Since those layers are already in good shape, one can imagine that their gradient contain very little discriminative power.\\n\\nc) Computational overhead. Similar to fine-tuning, our method becomes increasingly expensive (in terms of both time and space complexity) as we include gradient from more layers. Hence, it helps to restrict ourselves to the topmost layers in order to reduce the computational cost. This decision is further backed by our arguments in a) and b).\\n\\nNevertheless, we admit that our current approach to layer selection is more or less heuristic, and believe that a more principled strategy must exist. Perturbation analysis sounds like a very interesting idea along this line, and we will definitely look into it in our future work.\", \"reference\": \"[1] Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent. NeurIPS 19\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The key idea in this paper is to generate a feature vector based on a Fisher information idea for intermediary levels of a deep neural network. I liked the idea of using some form of linearization of the Fisher information matrix to learn a feature vector. The connection to tangent plane ideas in terms of robustness also made sense. The issue I have is what layer to apply this idea to and formally relating this idea to the variational VAE type framework. Using the top layer seems arbitrary. Also using one layer seems arbitrary. Is there a way to argue via a perturbation analysis of the variational problem of what makes the most sense. I feel the paper hints at this but does not make this idea explicit.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary: This paper considers the use of a neural network's Jacobian as additional features for semi-supervised, unsupervised, and transfer learning of representations. The idea is simple and the authors motivate this choice by connecting it to the literature on the neural tangent kernel (although nothing is proven in this paper). Some experiments are performed in which the authors demonstrate some improvements over the simple baseline of using the neural network's final layer alone.\", \"strengths\": [\"The idea is simple and motivated by the Fisher vector work (Jaakkola & Hausler, 1999), which the authors cite.\"], \"weaknesses\": [\"Comparing to a baseline with half as many features (just using the final layer of the neural network) is not a compelling baseline. The authors could have considered other alternatives: just using the gradients when comparing against the current baseline, augmenting the baseline with random features of the data.\", \"The discussion of the connection to theory added very little to the paper. This is not a theoretical paper, and the hand-wavy connections to the theoretical literature did not bolster the case. Instead, I found it distracting. I agree with the authors that this connection is important to point out, but a paragraph or two suffices.\", \"The ablation study could be more compelling. I'm not particularly surprised that a model improves as one increases the number of parameters that are pre-trained on the same kind of data (even if the pre-trianing objective is somewhat distinct).\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary: This paper proposes to use the gradients of specific layers of convolutional networks as features in a linearized model for transfer learning and fast adaptation. The method is theoretically backed by an appeal to the recently proposed neural tangent kernel and seems like it could be practically useful.\", \"edit\": \"Post rebuttal, I am somewhat satisfied by the authors' response and find their ablation study somewhat compelling hence am increasing my score to a weak accept. However, I'm not completely convinced that their choice of layers to use as gradient features is reasonable. If this paper does end up getting accepted, I'd very strongly advocate for being extremely clear as to\\n1) why the gradient features used are used, \\n2) a good choice of gradient features in practice\\n3) inclusion of all of the relevant ablation studies in the final version.\\n\\nI tend to weakly reject this paper currently despite liking the simplicitly and timeliness of the approach. Specifically, I would like further discussion of the choice of the layers to use as gradient features and an ablation study on supervised trained networks.\", \"originality\": \"I believe that this is one of the first papers to explicitly use the Taylor approximation (neural tangent kernel) in a transfer learning setting, making the approach timely and potentially practically useful.\", \"significance\": \"The approach is a quite nice merging of theoretical insights with a neat practical implementation. Although the main methodological advance is a straightforward application of the thinking behind Jacobian vector products, the method is well described and ought to be practically useful. However, I have a bit of a concern as to its practical necessity in comparison to simple fine-tuning of the final softmax layer (referred to in Table 1 as \\\\theta_2) on a new dataset.\", \"clarity\": \"While the approach described in Section 3 is quite generic \\u2013 theoretically, the method should simply consist of training the final layer of the neural network (w_1) and weights from the Jacobian features around the Taylor expansion. However, the experimental approaches suggest that different layers were used in each experiment \\u2013 see e.g. \\u201c[we] \\u2026 compute our gradient based features from one, two, or all of the top-3 conv layers\\u201d (Section 4.2 for the BiGAN architectures) versus \\u201cour gradient based features are from the last 2 conv layers\\u201d (Table 2 for the AlexNet architectures). Why was the entire network (modulo the last feed forward layer) not simply used for the Jacobian features?\\n\\nSimilarly, why was the entirety of the model (again modulo the last feedforward layer) not used for forwards model in the AlexNet experiments?\\n\\nBased on the ablation study in Section 4.1, the authors find that \\u201cit suffices to set the very top layer as \\\\theta_2 to enjoy a reasonably large performance gain.\\u201d When this is the case, it is a bit tough to distinguish the approach from standard final layer transferring approaches (Sharif Razavian et al, 2014) and indeed Bayesian final layer approaches that simply retrain an output layer (e.g. Perrone et al, 2018). Indeed this seems to be the case in Table 1, if that is so, then why not just re-train \\\\theta_2 as well throughout the experiments as it will typically just be another stochastic convex optimization problem?\\n\\nCould the authors quickly describe an implementation of the scalable Jacobian inner products in Section 3.3? It seems like outputs will have to be cached during the forwards passes, thereby requiring a somewhat significant amount of software engineering (and memory overhead) to be have to do this process for each new layer. Is this the correct understanding of how one would implement the procedure? A quick paste of PyTorch pseudo-code would be sufficient here.\", \"quality\": \"The experiments seem to be relatively convincing \\u2013 it seems exciting that a linear model + the network\\u2019s features itself can typically perform as well as fine-tuning and occasionally even better than fine-tuning itself.\\n\\nHowever, I am a bit concerned by the fact that the ablation studies themselves only utilize models trained in an un-supervised fashion. \\nI\\u2019d instead like the authors to run an ablation study on supervised trained models (perhaps 8 layer conv nets or VGG16) in the same manner as in Table 1 and Section 4.1. Specifically, I\\u2019d like to see this done to see whether the gradients in fact are as interesting as features when they have been trained in a supervised fashion. \\n\\nSimilarly, I\\u2019d like to see the features themselves used as a linear classifier (no network forwards passed) in the same two ablation studies. That is, could the authors use w_1^T J w_2 as the features for their linear classifier. If they have already done so and I\\u2019ve missed that in the tables somehow, I apologize. This experiment should help to test out how _useful_ the features defined by the Jacobian matrix are in comparison to the network\\u2019s forward pass itself.\", \"minor_comments\": \"\", \"introduction\": \"\\u201cAnd the accuracy of the rthe (sic) \\u2026\\u201d typo + please do not start sentences with and if at all possible.\\n\\nSection 3.1 (beneath eq. 2): \\u201cliner\\u201d should be linear.\", \"table_3\": \"\\u201cSelf-supervise\\u201d should be \\u201cSelf-supervised.\\u201d\", \"references\": \"Perrone et al, Scalable Hyperparameter Transfer Learning, NeurIPS, 2018. http://papers.nips.cc/paper/7917-scalable-hyperparameter-transfer-learning.pdf\\n\\nSharif Razavian et al, CNN Features off-the-Shelf: an Astounding Baseline for Image Recognition, CVPRW, 2014. https://arxiv.org/abs/1403.6382\"}" ] }

[ "Junren Luo", "Wei Gao", "Zhiyong Liao", "Weilin Yuan", "Wanpeng Zhang", "Shaofei Chen" ]

[ "deceptive opponent", "proactive network interdiction", "problem", "wcd", "observations", "behaviors", "online", "potential applications" ]

{ "note_id": [ "w-C7OPuhim", "B1eh40NfjH", "HJeoRzyD9r", "BklzD_02Yr", "BJxVeg3kKS" ], "note_type": [ "decision", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752703, 1573174836237, 1572430547094, 1571772505603, 1570910187966 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "~Junren_Luo1" ], [ "ICLR.cc/2020/Conference/Paper2584/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2584/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2584/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper has been withdrawn by the authors.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Submission Withdrawn by the Authors\", \"comment\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper aims to provide a goal recognition framework. The paper reviews previous literature and outlines a model, but it seems to be at the draft stage as it stands as the experiment section is missing (section 4). Also, in some parts the sentences are broken and hard to follow, for example, \\u201cHere we formulate the stochastic goal recognition control (S-GRC) problem as an extension of the GRD problem to allows non-optimal agents and stochastic actions, which means the agent\\u2019s action are not deterministic and the possible successor states are with probability, the measures we can take are interdictions with cost.\\u201d\\nBased on the above comments I think the paper is not ready for publication.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes to use reinforcement learning to model an agent that is reaching goal that it is intended to reach. The authors consider the case where the agent is (1) indifferent to being observed, (2) trying to help an observer reach its goal, and (3) trying to fool the observer into not reach its goal. The paper propose to use a value function to quantify how easy it is to predict where the agent is going (\\\"worst case distinctiveness\\\"). The authors propose to then train an agent to modify the action space to make it difficult for an agent to fool the observer.\\n\\nWhile this is an interesting idea, the paper is clearly incomplete. The experiment section is missing, and much of the method section is partly written.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper studies goal recognition control given a deceptive opponent, who selects actions to intentionally mislead or confusing the learner to learn the true goal. The problem has been studied in the security game, resource allocation game and Stackelberg game, where the defender is to play a resource allocation game given the best response of the attacker. in this paper, the authors use stochastic-shortest-path MDP to model the attacker's planning problem. The defender's objective is stated in eq (2) and (3) but not explained clearly.\\n\\nThe work considers a deceptive attacker whose objective is not to maximize the total reward/minimize the total cost, but rather to balance two objectives: minimizing the total cost while maximizing the KL divergence between its current policy and the optimal goal-achieving policy. This opponent model seems to be simplistic, given security game has studied the opponent using more rigorous equilibrium analysis and with dynamic Bayesian game formulation. \\n\\nsee Manshaei, Mohammad Hossein, et al. \\\"Game theory meets network security and privacy.\\\" ACM Computing Surveys (CSUR) 45.3 (2013): 25.\\n\\nThe practical application is well motivated. \\nThe paper itself is incomplete and contains many typos. The paper is unfinished.\"}" ] }

[ "Xutai Ma", "Juan Miguel Pino", "James Cross", "Liezl Puzon", "Jiatao Gu" ]

[ "Simultaneous Translation", "Transformer", "Monotonic Attention" ]

{ "note_id": [ "TirAOCED-o", "OxjHUKOkHY", "bgMqqD6tkI", "BJeSYds2ir", "SJxMKAc3jB", "H1eys4ajjH", "HygsSKziiS", "H1lRkHlqsH", "ryeQS4x5iB", "SJlFL7lqsB", "r1xlQQxcor", "HJxiJmeciB", "rkeK-Biu9r", "HkxUOzOkqH", "SJg74-IRYB" ], "note_type": [ "comment", "comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1694512912260, 1692551488490, 1576798752674, 1573857404906, 1573854841832, 1573799063358, 1573755202968, 1573680357582, 1573680187170, 1573679952581, 1573679896094, 1573679843105, 1572545793458, 1571943021610, 1571868970990 ], "note_signatures": [ [ "~Gladis_Ne_Limes1" ], [ "~Gladis_Ne_Limes1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2583/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/Authors" ], [ "ICLR.cc/2020/Conference/Paper2583/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2583/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2583/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"rre\", \"comment\": \"EHR software deals with highly sensitive patient information, making data security a top priority. Implementing robust security measures, including encryption, access controls, and audit trails https://mlsdev.com/blog/ehr-development, is essential to protect patient privacy and comply with regulations like HIPAA in the United States.\"}", "{\"title\": \"re\", \"comment\": \"Navigating the world of web design has never been more enlightening https://claspo.io/blog/how-to-avoid-negative-impacts-from-annoying-pop-ups-on-the-website-laspo-tips-and-tools/, thanks to LASPO's guide on avoiding the pitfalls of pesky pop-ups. It's like having a seasoned mentor whispering sage advice in your ear while you build your digital domain. By embracing these tips and tools, you're not just sidestepping annoyances; you're fostering a harmonious user experience that leaves a positive imprint. In an era where user engagement is paramount, LASPO's insights are a compass pointing toward user-centric design and seamless interactions. So let's bid adieu to frustration and welcome a new era of web elegance!\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper extends previous models for monotonic attention to the multi-head attention used in Transformers, yielding \\\"Monotonic Multi-head Attention.\\\" The proposed method achieves better latency-quality tradeoffs in simultaneous MT tasks in two language pairs.\\n\\nThe proposed method is a relatively straightforward extension of the previous Hard and Infinite Lookback monotonic attention models. However, all reviewers seem to agree that this paper is a meaningful contribution to the task of simultaneously MT, and the revised version of the paper (along with the authors' comments) addressed most of the raised concerns.\\n\\nTherefore, I propose acceptance of this paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for your clarifications.\"}", "{\"title\": \"Additional response to question 4 for experiments\", \"comment\": \"We updated the results of average latency regularization for MMA-H in the appendix. We observed than neither the simple average nor the weighted average regularization work on MMA-H models. Even with large weights, we are unable to reduce the latency, while the weighted average severely hurts the translation quality.\\n\\nFor MMA-IL, we also find that we are unable to reduce the latency with the range of weights explored. We are investigating the effect of larger weights for MMA-IL and the simple average loss. For reference, here are the numbers obtained so far:\\n\\nWeight | BLEU |AP | AL | DAL\\n0.05 | 29.9 | 0.98 | 24.50 | 26.24\\n0.1 | 30.0 | 0.95 | 21.16 | 24.01\\n0.2 | 29.8 | 0.94 | 20.20 | 23.09\\n0.4 | 29.9| 0.92 | 18.80 | 22.05\\n1 |28.2 | 0.81 | 11.75 | 14.67\"}", "{\"title\": \"Response to papers with flexible policy\", \"comment\": \"Thank you for pointing out these related papers. We will also cite these two additional references before the rebuttal deadline.\\n\\nIn the second paper (Zheng et al 2019, ACL), the authors propose to use a restricted dynamic oracle, which is controlled by a conservative and an aggressive bound, in order to train their model. The authors also introduced a delay token to give the model information about the READ action. The reviewer is correct in pointing out that this paper is an example of using the Transformer together with a learnable policy. \\n\\nDifferences between (Zheng et al 2019, ACL) and our work include the following:\\n1. In our work, we used a closed form solution of expected alignment for training, while (Zheng et al 2019, ACL) introduced a sampling-based restricted imitation learning.\\n2. (Zheng et al 2019, ACL) modify the model by introducing a special delay token in the target vocabulary while our model introduces read/write probabilities for each attention head.\\n3. In our work, the actions are fully decided by the model, while in (Zheng et al 2019, ACL) hard constraints were applied at test time to control the latency.\\n4. Unfortunately, we are unable to compare our work to (Zheng et al 2019, ACL) within the rebuttal timeframe, especially since the experiments involve the NIST corpus, which is not completely open (it is open to the participants of the OpenMT evaluations).\\n\\nThe first paper (Zheng et al 2019, EMNLP) was just published in EMNLP 2019, which was after the submission deadline for ICLR 2020, and the arxiv version was posted on September 4 2019, which is only 20 days before the ICLR submission deadline so it is effectively concurrent work. Thanks for bringing it to our attention. This paper indeed proposed a trainable policy model for Transformer. First, a sequence of actions were generated based on the rank of the gold target word given partial input given by an offline model. The generated action sequences were then used as labels in order to train an agent in a supervised fashion. At inference time, an offline model was used where actions came from the trained agent. The latency can be controlled by the threshold when generating the action sequence.\\n\\nDifferences between our work and (Zheng et al 2019, EMNLP) include the following:\\n1. In our work, the policy and model were trained jointly while in (Zheng et al 2019, EMNLP), the agent was trained separately from the translation model. Since the translation model is not adapted to different agents, this can potentially introduce a mismatch between training and inference.\\n2. In our work, we trained different translation models for different latency requirements, while (Zheng et al 2019, EMNLP) use one model for all latency regimes, this could also introduce a mismatch and thus hurt the performance.\\n3. In our work, we didn\\u2019t use explicit labels to train the policy, while (Zheng et al 2019, EMNLP) used generated labels to train the policy and the generating actions are controlled by a hyperparameter \\\\rho.\\n4. Our work obtains state-of-the-art performance, while the models in (Zheng et al 2019, EMNLP) underperformed MILk.\"}", "{\"title\": \"papers with flexible policy\", \"comment\": \"Some extra information about your statement \\\"a. To the best of our knowledge, we are the first to enable transformers to have a learnable policy for simultaneous translation.\\\"\\n\\nAt least I know two papers that are using a learnable policy with pure transformer models.\\n\\n[1] \\\"Simpler and Faster Learning of Adaptive Policies for Simultaneous Translation\\\" in EMNLP 2019\\n\\n[2]\\\"Simultaneous Translation with Flexible Policy via Restricted Imitation Learning\\\" in ACL 2019\\n\\ncould you also comment on the difference between those papers as well?\"}", "{\"title\": \"Response to questions from AnonReviewer2\", \"comment\": \"About the method:\\n1. Thank you for the great question. The overall latency are mainly introduced by the fastest head. It is possible when there is a small weight on the latency regularization term. By using proper regularization weight for the latency control method we proposed, we can \\u201cslow down\\u201d the faster heads so even though all heads need to agree to write, we can still achieve a low latency. Moreover, both latency regularization methods punish more on the faster head.\\n\\n2. Thank you for the great question. We carried out an analysis on attention head rank during the inference time. Figure 6 shows that some heads can run faster than others and it is more obvious in MMA-IL. It can also be observed from the case example in Figure 5b. We also find that the higher layers are likely to move faster than lower layers, especially in MMA-IL.\\n\\n3.This is a very good question. We have carried out one simple experiment that only change the inference setting, which is shown in Appendix A.4. However, we observe that this method failed to produce high quality translations because there is a mismatch between training and inference time. This training-inference mismatch was also observed in Ma et al. (2019), where they report a significant difference with wait-k and test wait-k.\\n\\nFinally, would you be able to clarify what you mean by \\u201ctuning the requirements mentioned in (1)\\u201d ?\", \"about_the_experiments\": \"1. This is a very good question. The MMA-IL or MILk model, which has an infinite look back attention mechanism, will have more information when latency increases because of the larger context coverage on source side. An extreme case is that when the model reach the maximum latency (AP=1), the MMA-IL or MILk will behave like the full offline models. However, for MMA-H, because of the hard alignment, a larger latency does not necessarily mean an increase in source information available to the model. In fact, the large latency was introduced by outlier attention heads, which read the entire source sentence and point to the end of the source sentence. These outlier heads not only increase the latency but they also do not provide useful information to the model as they are pointing to the end of sentence token, which is not very informative. We introduce the attention variance loss to eliminate the outliers, as such a loss makes the attention heads focus on the current context for translating the new target token. This is why when the latency decreases (fewer outliers), the MMA-H model has an increase on translation quality. \\n\\n2. Thank you for the suggestion. We have added a visualization of a real sample as well as an analysis on how heads behave at different layers.\\n\\n3. Thank you for pointing this out. We have updated the figure to make it clear. Each subplot now has a title (\\u201cOffline Model\\u201d, \\u201cMMA-H\\u201d, \\u201cMMA-H (latency)\\u201d, \\u201cMMA-IL\\u201d, \\u201cMMA-IL (latency)\\u201d).\\n\\nThank you for the suggestion. We have added a visualization of a real sample as well as an analysis on how heads behave at different layers.\"}", "{\"title\": \"Response to specific comments & suggestion from AnonReviewer1\", \"comment\": \"- There is a typo in the abstract: \\\"multiple attentions heads\\\"\\n\\nThank you, this was corrected.\\n\\n- \\\"the denominator in Equation 5 is clamped into a range of (\\\\eps, 1]\\\" Technically this doesn't need to be an open range on the left.\\n\\nThank you for the comment, we have updated this to \\u201c[\\u201d after checking the implementation.\\n\\n- \\\"which allows the decoder to apply softmax attention over a chunk (subsequence of encoder positions).\\\" It may be clearer if you just say \\\"which allows the decoder to apply to a fixed-length subsequence of encoder states preceding \\\".\\n\\nThank you, the text has been edited accordingly.\\n\\n- I think more could be done to distinguish the behavior of the encoder self-attention, the encoder-decoder attention, and the decoder self-attention. You write \\\"The model will write a new target token only after all the attentions have decided to write\\\" but also \\\"Some heads read new inputs, while the others can stay in the past to retain the source history information.\\\" If an attention head is \\\"staying in the past\\\", then the model will not be able to write a new target token. I think in one of these cases you are referring to (encoder) self-attention and in the other you are referring to decoder attention. Please clarify.\\n\\nWe have provided more details the algorithm 1 on how to update encoder states during the inference. However, both sentences are talking about encoder-decoder attention.\\n\\n\\\"The model will write a new target token only after all the attentions have decided to write\\\" indicate when to start the writing process, which is line 19 in algorithm 1. \\n\\n\\u201cSome heads read new inputs, while others can stay in the past to retain the source history information.\\u201d means that different heads have different speeds when reading the source tokens. If one attention head decides to write, as it shows in line 8-10, the head will stay on this location and wait for other heads finish reading. If an attention head can be \\\"staying in the past\\\" as long as j is smaller than t_max, which is the actually number of tokens read by model.\\n\\n- \\\"Although MMA-H is not quite yet streaming capable since both the encoder and decoder self-attention have an infinite lookback, that model represents a good step in that direction.\\\" I think you should distinguish between online and streaming translation. It sounds like when you say streaming you mean that the utterances may continue arbitrarily, so infinite lookback is impractical. However, one could truncate the source sequence whenever the decoder outputs an end-of-sentence token, or something. I'm not sure people usually make a strong distinction here.\\n\\nWhat we mean here is that hard attention can handle an input stream natively, without having to resort to segmenting the input. However, this is a good point that in practice, both our model and MILk are able to handle arbitrarily long input, provided that we segment the input. We updated the explanation accordingly.\\n\\n- For completeness it would be useful to include at least wait-k, and potentially a line corresponding to offline attention performance, in the plots in Figure 2.\\n\\nWe updated the paper with the offline baseline in Figure 2. We didn\\u2019t include wait-k (Ma et al. 2019) and reinforcement learning models (Gu et al. 2017) because they underperform the MILk model (Arivazhagan et al. (2019)).\\n\\n- Why don't you include MMA-IL in WMT'15 EnDe? (figure 2)\\n\\nThank you for the suggestion. The updated paper now includes the MMA-IL results on the WMT15 De-En dataset.\\n\\n- Are you copying the results from (Arivazhagan et al., 2019) or did you reimplement MILk in your codebase?\\n\\nFor the IWSL15 En-Vi dataset, we implemented monotonic infinite lookback models based on LSTM, following the settings from Colin Raffel et al (2017) and Chiu and Raffel (2018). For WMT15 De-En, we copied the numbers from (Arivazhagan et al., 2019) because they are using a RNMT+ implementation.\\n\\n- The results in Figure 2 could be made much more informative if there was some way of communicating the multiplier (the \\\"lambdas\\\") of the different latency losses for each of the different dots on the plot. This would make it much more obvious how important these losses are and how the effect the quality/latency trade-off.\\n\\nThank you for the suggestion, we updated the detailed results in the appendix.\\n\\n- There are some additional possible references for online seq2seq, like CTC, the Neural Transducer, \\\"Learning Hard Alignments with Variational Inference\\\", \\\"Learning online alignments with continuous rewards policy gradient\\u201d, etc.\\n\\nThank you so much suggestions, we updated the related work.\"}", "{\"title\": \"Response to the comments on the presentation\", \"comment\": \"1) Thank you for pointing this out. We have updated the paper accordingly and added this information in Section 2.2.\\n\\n2) Thank you for the advice, we have updated the offline models in the figures.\\n\\n3) Thank you for pointing this out. We have updated the figure to make it clear. Each subplot now has a title (\\u201cOffline Model\\u201d, \\u201cMMA-H\\u201d, \\u201cMMA-H (latency)\\u201d, \\u201cMMA-IL\\u201d, \\u201cMMA-IL (latency)\\u201d).\\n\\n4) We have updated the results of unidirectional and bidirectional offline model with greedy and beam search. With beam search, the performance is not affected by a unidirectional encoder. Greedy decoding introduces a small drop of 0.4 BLEU and the unidirectional encoder further gives a 0.3 BLEU drop.\\n\\nThank you for the comment. We proofread and edited accordingly.\"}", "{\"title\": \"Response to the comments on the experiments\", \"comment\": \"1)Thank you for the suggestion. The updated paper now includes the MMA-IL results on the WMT15 De-En dataset, which are also positive. We believe that 2 datasets is appropriate to demonstrate the effectiveness of our method, as was done in prior work. However, we will conduct follow up experimentation on WMT14 English-French in order to have an additional comparison datapoint with Arivazhagan et al. (2019) for the camera-ready version, should this paper be accepted.\\n\\n2) This is a very good question. The MMA-IL or MILk model, which has an infinite look back attention mechanism, will have more information when latency increases because of the larger context coverage on source side. An extreme case is that when the model reach the maximum latency (AP=1), the MMA-IL or MILk will behave like the full offline models. However, for MMA-H, because of the hard alignment, a larger latency does not necessarily mean an increase in source information available to the model. In fact, the large latency is introduced by outlier attention heads, which read the entire source sentence and point to the end of the source sentence. These outlier heads not only increase the latency but they also do not provide useful information to the model as they are pointing to the end of sentence token, which is not very informative. We introduce the attention variance loss to eliminate the outliers, as such a loss makes the attention heads focus on the current context for translating the new target token. This is why when the latency decreases (fewer outliers), the MMA-H model has an increase on translation quality. \\n\\n3)Thank you for pointing this out. The purpose of this experiment is to motivate the introduction of the MMA model. The Transformer model usually has multiple encoder-decoder attention heads, either because it has multiple layers (even with single-head attention, there is more than one head: one head per layer), or because it uses multihead attention. We have updated Figure 4 to include the single-head attention case and we also included the results to include the MMA-IL model. The updated figure demonstrates that multiple heads (either coming from multiple layers or from using multihead attention) is beneficial for performance up to a certain point. We also observe that with one layer, increasing the number of heads is beneficial. With more layers, for the baseline and MMA-IL, increasing the number of heads is beneficial up to a certain point, then performance plateaus or degrades. For MMA-H, using more than one head degrades performance initially, then the performance starts recovering. In any case, the single layer/single head case underperforms all other settings, which motivates the introduction of our model.\\n\\n4) Thank you for the great question. The setting in our paper is different from Li et al (2018), especially for MMA-H. The attention is monotonic and hard-aligned to source states in MMA-H, whereas Li et al (2018) uses the full softmax attention. Furthermore, the disagreement of attention in Li et al (2018) is different from the disagreement of attention head positions in our work. In Li et al (2018), disagreement of attention means disagreement of softmax distributions; in our work, the disagreement means the disagreement of attention head position during inference time. These two are not necessarily correlated. When we eliminate the outliers, we find that MMA-H works better because the hard aligned attention heads can focus more on the current source context for translating a new target word. We also add visualization and rank analysis in the paper.\"}", "{\"title\": \"Response to the two reasons not accepting this paper from AnonReviewer4\", \"comment\": \"1) We thank the reviewer for pointing out prior work on monotonic attention based models and latency augmented training for simultaneous machine translation. We\\u2019re convinced that our work has significant differences with prior work (Raffel et al. 2017, Arivazhagan et al. 2019):\\n\\na.To the best of our knowledge, we are the first to enable transformers to have a learnable policy for simultaneous translation. Although monotonic attention and monotonic infinite lookback attention were introduced before this work, it is non-trivial to make them compatible with multihead attention. In fact, Arivazhagan et al. (2019) do not use multihead attention even though the RNMT+ model originally supports multihead attention (https://arxiv.org/abs/1804.09849).\\n\\nb.We introduce two novel models, MMA-H and MMA-IL. MMA-H is designed with the eventual goal of naturally handling arbitrarily long input while the goal of MMA-IL is to retain as much quality as possible with respect to an offline model. These two models require two carefully designed losses, described next, in order to be effective.\\n\\nc.The variance loss we introduce required preliminary experimentation and careful design. This loss encourages different heads to point to positions close to each other. In preliminary experiments, we found that some heads tend to skip every token and move to the end of sentence while some heads tend to stay at the beginning of the sentence. In the case of the MMA-H model, the fast outliers heads increase latency but do not improve quality since they point to an uninformative token (end of sentence) while the slow outlier heads prevent the model to deal with arbitrary long input in a natural way. Our regularization was designed specifically for hard monotonic multihead attention and, to the best of our knowledge, we are the first to propose this idea for latency control.\\n\\nd.Finally, we believe this work will have an impact on the community for three reasons:\\ni. We reach state-of-the-art performance for simultaneous machine translation.\\nii. We are the first to combine the transformer with monotonic attention to introduce a learnable policy transformer model which outperforms the wait-k model (which is also transformer-based).\\niii. We will publish the code to contribute to the community of simultaneous translation. Note that very little prior work did so.\\n\\n\\n2) Thank you for the suggestions.\\n\\nWe picked (Arivazhagan et al. 2019) as our baseline since this represents the current state-of-the-art for simultaneous machine translation. Note that while the monotonic infinite lookback attention (MILk) (Arivazhagan et al. 2019) is an RNN-based model, it still outperforms previous transformer-based approaches such as the wait-k model (Ma et al. 2019, https://www.aclweb.org/anthology/P19-1289/).\\n\\nRegarding the first proposed baseline (\\u201cthe same, but with single-head attention\\u201d), note that as soon as we use multiple layers, our proposed algorithm/model is necessary since there are multiple encoder-decoder attention layers. The only situation for the transformer that does not require a modification of the previously proposed MILk model is when there is single-head attention *and* only one layer. We have also updated Figure 4 with additional experimentation for single-head attention and the MMA-IL model.\\n\\nWe corrected the text of the paper which said \\u201cFor MMA-IL model, we used both loss terms;\\u201d, sorry for the mistake. For that model, we were actually only using the proposed weighted average loss. In summary, we are only using L_var for MMA-H and only using L_avg for MMA-IL. Note that the weighted average loss we introduce is different than the straightforward average loss proposed by the reviewer. The method we proposed automatically assigns more weights on faster heads and reduces overall latency. It also moderately regularizes the slower heads to prevent them for being outliers later. We are doing an additional experiment with this simple average loss.\\n\\nOne of the motivation to use MMA-H is to not have heads which are particularly slow and eventually enable streaming in a natural way. In preliminary experiments, we observe that for MMA-H models, some heads always stays at the beginning of the sentence. This is why we introduced L_var, in order to have different heads stay together. As you can see in Figure 3, L_var is effective from that perspective. We are also doing additional experiments with the simple average loss for MMA-H proposed by the reviewer.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper proposes an approach for simultaneous neural machine translation. While prior works deal with recurrent models, the authors adopt previous approaches for Transformer. Specifically, for decoder-encoder attention they introduce Monotonic Multihead Attention (MMA) designed to deal with several attention heads. Each attention head operates as a separate monotonic attention head similar to the ones from previous works by Raffel et al. (2017) and Arivazhagan et al. (2019). MMA has two versions: MMA-H (hard), where each head attends only to one token, and MMA-IL (infinite lookback), where each head attends to all tokens up to a selected position.\", \"the_main_novelty_of_the_work_in_the_way_of_dealing_with_multiple_heads\": \"(i) average over heads latency instead of single-head latency, (ii) penalty which encourages different heads point to similar positions (L_var).\\n\\nExperimental results on two translation tasks (IWSLT En-Vi and WMT De-En) show better quality-latency trade-offs compared to the recurrent analogs. Experiments with different number of attention heads would be helpful, but in the current state are rather confusing (see the comments below).\\n\\n\\nI can not recommend accepting this paper due to the two main reasons.\\n\\n1) The proposed solution lacks novelty. \\nMMA-H and MMA-IL attentions, introduced in this work, are straightforward applications of previous works by Raffel et al. (2017) and Arivazhagan et al. (2019) respectively to each of the heads. Novelty is in the way of dealing with multiple heads: (i) average over heads latency instead of single-head latency, (ii) penalty which encourages different heads point to similar positions. However, these alone are unlikely to make a large impact.\\n\\n2) The results are experimentally weak, evaluation is questionable. \\nThe current baseline is the recurrent model, but since the main contribution of this work is in how to deal with several heads the proper baselines would be (i) the same, but with single-head attention, (ii) the same, but without L_var (summing latency penalty over heads is straightforward). However, these baselines are absent and the improvement is likely to be due to the replacement of RNN with the Transformer - this would be a limited contribution.\\n\\n\\n\\nOther comments on the experiments.\\n\\n1) MMA-IL was tested only on one small dataset (IWSLT En-Vi), MMA-H only on 2 datasets - more experiments would help.\\n\\n2) Figure 2 shows that when increasing lagging, performance first improves, then drops. While the former is expected, the latter is not: one would expect a model to reach the performance of its offline version when given maximum latency, but not to drop. The MILK model (Arivazhagan et al., 2019) behaves as expected, but the proposed MMA does not: does this mean that there is some problem with the model? An explanation of this behavior would be helpful.\\n\\n3) Number of heads, Figure 4: the results are supposed to show that performance improves when increasing the number of heads, but the figure shows the opposite. If the decoder has more than one layer (which Transformer usually has), for 4 heads perform worse than one, 8 heads (which the standard Transformer setting) - not better than one, and we get the improvement only when using 16 heads with 6 layers. Could you elaborate on this? The text says \\u201cquality generally tends to improve with more layers and more heads\\u201d.\\n\\n4) I can see the motivation for using the L_var loss term in this setting, but forcing heads in MMA-H attend to similar positions seems counterintuitive: there is evidence that making attention heads less correlated improves quality (see for example Li et al, EMNLP 2018 https://www.aclweb.org/anthology/D18-1317/), but L_var may end up doing the opposite. How different is the behavior of the learned heads? Figure 3 shows the average attention span, but this does not tell how many of the heads are doing different things. Some analysis would be nice.\\n\\n\\nComments on the presentation.\\n\\n1) It would be better to mention that encoder self-attention is limited only to previous positions earlier in the text and more prominently (now it\\u2019s on the bottom of the sixth page). A good place would be in Section 2.2 when talking about different attention types in the Transformer and modifications required for simultaneous translation.\\n\\n2) Figure 2: it would be helpful to see offline model performance on these pictures (e.g., a horizontal line showing the BLEU scores). One of the main points of these experiments is to see how much each model drops compared to its offline version.\\n\\n3) Figure 4: it is not clear from the figure which models it corresponds (for example, what is the difference between the first and the second figures: language pairs? models?)\\n\\n4) Table 2 shows the BLEU scores for Transformer with unidirectional encoder and greedy decoding. It would be helpful to see also the BLEU scores of Transformer in the standard setting (full attention in the encoder, beam search inference). For now, it is not clear how much one loses by replacing the encoder to unidirectional: if much, then probably it makes sense to work also on the encoder (read and encode source sentence with a latency).\\n\\nLanguage and typos.\\nThe text is quite raw and could use an extra work; a lot of typos starting from the abstract (e.g., \\u201c...for multiple attentions heads\\u201d).\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"Applying monotonic attention mechanisms to the Transformer architecture is an obvious and necessary idea, and as such this paper constitutes an important contribution. The application of monotonic attention to the Transformer is as I would have expected. The latency-reduction losses are novel, however. I think there are various things the authors could do to clarify their results as well as the presentation of their methods, which I discuss below. Overall, I think this is a solid accept, especially with some improvements to the presentation.\", \"Specific comments & suggestions:\", \"There is a typo in the abstract: \\\"multiple attentions heads\\\"\", \"\\\"the denominator in Equation 5 is clamped into a range of (\\\\eps, 1]\\\" Technically this doesn't need to be an open range on the left.\", \"\\\"which allows the decoder to apply softmax attention over a chunk (subsequence of encoder positions).\\\" It may be clearer if you just say \\\"which allows the decoder to apply to a fixed-length subsequence of encoder states preceding $t_i$\\\".\", \"I think more could be done to distinguish the behavior of the encoder self-attention, the encoder-decoder attention, and the decoder self-attention. You write \\\"The model will write a new target token only after all the attentions have decided to write\\\" but also \\\"Some heads read new inputs, while the others can stay in the past to retain the source history information.\\\" If an attention head is \\\"staying in the past\\\", then the model will not be able to write a new target token. I think in one of these cases you are referring to (encoder) self-attention and in the other you are referring to decoder attention. Please clarify.\", \"\\\"Although MMA-H is not quite yet streaming capable since both the encoder and decoder self-attention have an infinite lookback, that model represents a good step in that direction.\\\" I think you should distinguish between online and streaming translation. It sounds like when you say streaming you mean that the utterances may continue arbitrarily, so infinite lookback is impractical. However, one could truncate the source sequence whenever the decoder outputs an end-of-sentence token, or something. I'm not sure people usually make a strong distinction here.\", \"For completeness it would be useful to include at least wait-k, and potentially a line corresponding to offline attention performance, in the plots in Figure 2.\", \"Why don't you include MMA-IL in WMT'15 EnDe? (figure 2)\", \"Are you copying the results from (Arivazhagan et al., 2019) or did you reimplement MILk in your codebase?\", \"The results in Figure 2 could be made much more informative if there was some way of communicating the multiplier (the \\\"lambdas\\\") of the different latency losses for each of the different dots on the plot. This would make it much more obvious how important these losses are and how the effect the quality/latency trade-off.\", \"There are some additional possible references for online seq2seq, like CTC, the Neural Transducer, \\\"Learning Hard Alignments with Variational Inference\\\", \\\"Learning online alignments with continuous rewards policy gradient\\u201d, etc.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a fully transformer-based monotonic attention framework that extends the idea of MILK. Though the idea of monotonic multi-head attention sounds interesting, I still have some questions below:\", \"about_the_method\": \"1. Is that possible that the MMA would have worse latency than MILK since all the attention heads need to agree to write while MILK only has one attention head?\\n 2. Is there any attention order between different attention head?\\n 3. I think the MMA only could control the latency during training time, which would produce different models with different latency. Is there any way that enables MMA to control the latency during inference time? Can we change the latency for on given model by tuning the requirements mentioned in (1)?\", \"about_the_experiments\": \"1. Do you have any explanation of why both MMA-H and MMA-IL have better BLEU when AL is small? The results in fig 2 seem counterintuitive. \\n 2. I suggest the authors do more analysis of the difference between different attention heads to prove the effectiveness of MMA. \\n 3. For the left two figures in fig 4, which one is the baseline, and which one is the proposed model?\\n\\nI also suggest the authors present more real sample analysis and discussions about the experiments.\"}" ] }

[ "Gábor Berend" ]

[ "sparse word representations", "multilinguality", "sparse coding" ]

{ "note_id": [ "J9krlLXUW1m", "fPjrfR2Q1_", "S1eGTLLnjB", "SJeylM2cjS", "B1l1Gwp_ir", "Byx08zpuiS", "HJlcVa2doH", "SklpIWaPqS", "B1lY96o1cH", "rJe0V3HRKH" ], "note_type": [ "comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1692090858909, 1576798752646, 1573836474160, 1573728742756, 1573603079450, 1573601877731, 1573600562407, 1572487509395, 1571958161509, 1571867701519 ], "note_signatures": [ [ "~Gladis_Ne_Limes1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2582/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2582/Authors" ], [ "ICLR.cc/2020/Conference/Paper2582/Authors" ], [ "ICLR.cc/2020/Conference/Paper2582/Authors" ], [ "ICLR.cc/2020/Conference/Paper2582/Authors" ], [ "ICLR.cc/2020/Conference/Paper2582/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2582/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2582/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"re\", \"comment\": \"In the intricate world of higher education applications, the value of adept Statement of Purpose (SOP) writing services is like discovering a compass for success. Beyond mere words, these services craft narratives that breathe life into aspirations and experiences. As an aspiring student navigating the competitive landscape https://www.sopwriting.org/, I've realized that SOP writing services are more than just a convenience; they're strategic allies. They understand the art of presenting one's journey with authenticity and impact, helping applicants stand out in the admissions process. With dreams to pursue and opportunities to seize, platforms like these, such as SOPWriting.org, are the guiding light for articulating our academic path. Here's to SOP writing services that empower us to share our story with passion and purpose!\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper describes a new method for creating word embeddings that can operate on corpora from more than one language. The algorithm is simple, but rivals more complex approaches.\\n\\nThe reviewers were happy with this paper. They were also impressed that the authors ran the requested multi-lingual BERT experiments, even though they did not show positive results. One reviewer did think that non-contextual word embeddings were of less interest to the NLP community, but thought your arguments for the computational efficiency were convincing.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks!\", \"comment\": \"Thanks very much for adding these two experiments: for adding multilingual BERT to the paper and discussing the relation to BiSparse in the comments above (and presumably later adding it to the paper). I think the conclusion that you actually outperform BiSparse really helps strengthen the paper, and I greatly appreciate your honesty and straightforwardness with the multilingual BERT result.\\n\\nYou have definitely addressed all of my major questions.\"}", "{\"title\": \"Comment related to comparison to BiSparse\", \"comment\": \"As we noted in our general answer, we have experienced that using the same hyperparameters for regularization in the case of BiSparse as for MultiSparse resulted in subpar results in terms of the level of sparsity.\\n\\nAfter increasing the monolingual hyperparameters from $\\\\lambda_s=\\\\lambda_t=2$ to $\\\\lambda_s=\\\\lambda_t=5$, we managed to obtain representations relying on BiSparse that behave comparably to MultiSparse as well.\\nWe created BiSparse representations for the language pairs English--Italian and English--Danish, obtaining approximately 3% of the coefficients becoming nonzero by the end of the optimization.\\n\\nEvaluating the BiSparse representation by QVEC-CCA, we obtained the scores of 0.602, 0.789/0.786, 0.585 for Danish, English and Italian, respectively.\\nWe now have two scores for English this time (0.789/0.786) since BiSparse created separate representations for English when jointly trained with Danish and Italian as well.\", \"multisparse_on_the_other_hand_obtained_qvec_cca_scores_as_follows\": \"0.612, 0.808 and 0.596 for the same languages.\\n\\nThis means that MultiSparse performs comparably (or even better) to BiSparse based on their QVEC-CCA scores. We think that this could potentially be explained by the fact that BiSparse tries to optimize a more difficult objective -- as it involves the joint optimization of two non-convex problems -- hence finding a good solution is more difficult in that case.\"}", "{\"title\": \"Answers to Reviewer #2\", \"comment\": \"We would like to thank the reviewer for the feedbacks provided.\\n\\nWe agree with the review that the investigation of the actual interpretability of the word representations besides their QVEC-CCA scores could provide additional useful insights.\\nConducting such experiments (e.g. performing Word intrusion detection) was currently beyond the scope of the paper, however, marks an interesting future path.\\n\\nBased on the suggestions on how to make the paper more easily accessible, we submitted an updated version of the paper.\"}", "{\"title\": \"Answers to Review #3\", \"comment\": \"We would like to thank the reviewer for the general interest in our work and the constructive feedback provided.\\n\\nWe regard the stability regarding the number of the nonzero coefficients per word forms characterizing MAMUS a useful property, because it means that one can reliably anticipate the level of sparsity that would be induced by a certain choice of the regularization hyperparameter \\\\lambda. Inspecting the average number of nonzero coefficients per word forms in the case of MultiSparse (or BiSparse), we found much higher fluctuations in the sparsity levels obtained for the same choice of hyperparameters.\\n\\nInspired by the question related to the potential usage of multilingual contextual representations, we conducted further XNLI experiments when relying on multilingual BERT representations as inputs. The relative performance of the models trained over MAMUS representations is above 90% to those that are built on top of multilingual BERT embeddings. This suggests that MAMUS representations can serve as a viable alternative to the application of the computationally more demanding contextualized representations.\\nFor further details, please refer to the revised version of the paper.\"}", "{\"title\": \"Answers to Review #4\", \"comment\": \"We would like to thank the reviewer for the insightful review and the useful suggestions for improving the quality of the paper.\\n\\nWe are planning to report evaluation on QVEC-CCA when using BiSparse instead of MultiSparse. There are two difficulties we have encountered regarding this experiment.\\nAs BiSparse involves the optimization of nearly twice as much parameters as MultiSparse (since it learns representations for the source and target languages at the same time), training BiSparse representations for a single pair of languages takes more than 50 hours.\\nFurthermore, as it turned out, using the same hyperparameters in the BiSparse setting could result in substantially different results.\\nFor instance, in the case of Italian embeddings, the BiSparse representations contained more than 25 times as many nonzero coefficients as opposed to the representations obtained by MultiSparse, 1100+ and 44.3, respectively.\\nThis sensitivity of BiSparse for the choice of the regularization hyperparameters, together with its slow running time makes the conduction of the proposed experiment rather cumbersome, given that we would like to compare BiSparse representations of similar sparsity level to the previously calculated MutliSparse representations.\\n\\nWe found the comment on contextual word embeddings a very inspiring one. Hence we conducted further XNLI experiments when relying on multilingual BERT embeddings. We revised the paper with the additional comparisons of multilingual BERT and Mamus. In short, the relative performance of the models that use Mamus instead of multilingual BERT is above 90%. For more details, please refer to the revised version of the paper.\\n\\nFinally, the revised paper also addresses the further points raised in the review.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper describes a method to build sparse multilingual word vectors that is designed to scale easily to many languages. The key idea is to pick one language to be the source language, and then to build word embeddings and then a sparse dictionary + sparse coefficients for that source language monolingually. Other target languages then first align their embedding to the source using a seed list of translations and standard techniques, and then determine their sparse coefficients based on the fixed source sparse dictionary. This latter process is a convex optimization, which improves efficiency and stability. The method is tested with an established correlation-based intrinsic metric (QVEC-CCA) as well as by using the multilingual embeddings to project systems for cross-lingual document classification, dependency parsing and natural language inference.\\n\\nThe core idea of this paper is substantially simpler than the method it compares to (BiSparse), which jointly optimizes dictionaries, coefficients and cross-lingual coefficient alignment. So, the question that immediately comes to mind for me is, how much accuracy am I giving up for improved scalability to multiple languages? This could be easily tested for the two language case, where BiSparse could be directly compared without modification to their proposed method. Instead, BiSparse is modified to fit scale to the multilingual setting (becoming MultiSparse), but since it shares the constraint that the source dictionary and coefficients are fixed, it has already lost a lot of the power of joint optimization. I think the paper would be stronger with a two-language experiment where we would expect the proposed method to lose to BiSparse, but we could begin to understand what has been given up for scalability.\\n\\nI also wonder how relevant bilingual word embeddings are in a world of multilingual BERT and similar approaches. It would be interesting to know how cross-lingual embedding-in-context methods would do on the extrinsic evaluations in this paper, though I also acknowledge that this could be considered beyond the scope of the paper.\\n\\nOtherwise, this is a fine paper. It is well written and easy to follow. The experiments look sane, and the inclusion of both intrinsic and extrinsic tasks makes them fairly convincing. I have only a few remaining nitpicks:\\n\\n(1) As someone relatively unfamiliar with multilingual (as opposed to bilingual) word embedding research, it wasn\\u2019t clear to me how the experiments described here tested the multilinguality (as opposed to bilinguality) of the embeddings. It would be nice to provide an explanation for why (beyond the obvious efficiency gains) one couldn\\u2019t just do the necessary language pairs with bilingual methods for these experiments. And if one could perform the tests with bilingual methods, they should be included as baselines. \\n\\n(2) The discussion of MultiSpare hyper-parameter tuning appears in the Monolingual Experiments section, leading me to wonder what target languages were used for this tuning process.\\n\\n(3) In the second-last paragraph of 3.2.2, there is a sentence fragment that ends in \\u201cnonetheless their multiCluster and multiCCA embedding spaces contain no embeddings for\\u201d\\n\\n(4) The last paragraph before the Conclusion also feels like a fragment, or like two sentences have been spliced together: \\u201cWe have detailed the differences to Upadhyay et al. (2018) extends the previous work by incorporating dependency relations into sparse coding.\\u201d\\n\\nThere are also several places where periods seem to have been left out (such as immediately before the above sentence).\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper proposes a new approach for generating multilingual sparse representations.\\nFor generating such representations, the proposed approach solves a series of convex optimization problems, serially for each language. Compared to previous work for generating sparse cross-lingual representations which is applicable to a pair of languages, the proposed approach is applicable to an arbitrary number of languages.\\n\\nThe paper argues that these sparse representations can lead to better performance for downstream tasks and interpretability. This is demonstrated using experiments on QVEC-CCA (for interpretability analysis), NLI, cross-lingual document classification, and dependency parsing (downstream tasks).\\n\\nOverall, the approach is well-motivated and performs well empirically. The experimental setup is also described in detail. \\n\\nMinor - I did not understand the benefit of MAMUS being \\\"stable\\\" across languages, as argued from Fig 1? The use of the word \\\"cognitively\\\" in the statement \\\"representations determined by MAMUS behave in a cognitively more plausible manner\\\" also seems a stretch to me.\\n \\nOne issue that the authors should discuss is whether such representations hold any extra value over contextual representations like multilingual Elmo, BERT etc. For instance, why would someone use MAMUS representations instead?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a method to generate sparse multilingual embeddings. The key idea is to build only one set of source basis embeddings, and then represent all multilingual embeddings as a linear combination of these source embeddings. I felt the paper is a nice extension of the Vyas 2016 paper. Compared to existing approaches, their method will be faster to train and will need less data (particularly useful for low resource languages). Since I am less aware of work in this area, I cannot comment on whether the evaluation is complete. Particularly, I wonder if there is a qualitative way to show interpretability of the sparse vectors created by the method. We currently only have QVEC-CCA numbers to judge interpretability. A few more suggestions:\\n\\n1. I got confused by the sentence \\\".. over a reduced number of parameters for each target language as it treats D_s as D_t .\\\". Things got clear from the equations, but will be good to fix.\\n\\n2. Figure 1 was very difficult to understand. Ideally your caption should be enough to understand the figure.\"}" ] }

[ "Wanyun Cui" ]

[ "representation learning", "natural language processing", "attention" ]

{ "note_id": [ "aR5hnzfg8o", "SyxgHKR_qS", "rylp9-vO9S", "ryg1sde8qB", "SJxYE6JI9H", "HJg-aRtxqr", "BkxidjcRtB", "rJxkd5v0Fr", "H1lsJqssFB", "BylzE-XI_r", "B1ec1wCzOS", "B1lGqREgur" ], "note_type": [ "decision", "comment", "official_comment", "official_review", "comment", "official_comment", "official_review", "official_review", "comment", "official_comment", "comment", "comment" ], "note_created": [ 1576798752618, 1572559159608, 1572528533102, 1572370582738, 1572367664549, 1572015800734, 1571887987303, 1571875431000, 1571695074649, 1570283818172, 1570068194153, 1569898121864 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "~Hamed_GHAZAVI_KHORASGANI1" ], [ "ICLR.cc/2020/Conference/Paper2581/Authors" ], [ "ICLR.cc/2020/Conference/Paper2581/AnonReviewer3" ], [ "~Hamed_GHAZAVI_KHORASGANI1" ], [ "ICLR.cc/2020/Conference/Paper2581/Authors" ], [ "ICLR.cc/2020/Conference/Paper2581/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2581/AnonReviewer1" ], [ "~Hamed_GHAZAVI_KHORASGANI1" ], [ "ICLR.cc/2020/Conference/Paper2581/Authors" ], [ "~Hao_Zhang1" ], [ "~Sai_Prabhakar_Pandi_Selvaraj1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper incorporates phrases within the transformer architecture.\\n\\nThe underlying idea is interesting, but the reviewers have raised serious concerns with both clarity and the trustworthiness of the experimental evaluation, and thus I cannot recommend acceptance at this time.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for your response\", \"comment\": \"Your answer makes perfect sense as the non-linearity is proposed by your method and does not exist in normal attention. Thanks again for the clarifications.\"}", "{\"title\": \"Response\", \"comment\": \"I think they are different. The information related to phrases in our approach cannot be captured in higher layers in normal attention over words. For k=2, our approach represents the inductive bias and non-linearity of phrases with length 2. For \\\"deep\\\" network structure, such inductive bias and non-linearity of 2-word phrases propagate to more complicated phrases. But in normal attention over words, neither the inductive bias nor the non-linearity are captured. So we cannot expect the similar performance of the normal attention over words if we have more data or deeper structures.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"I think the paper needs a deep review in the English part. For example, in the abstract, they repeat \\\"In this paper\\\" a couple of time and it is complicated to understand the introduction and methodology. Also, I think the paper needs a better structure. The related work should be first in order to understand the relevance of this paper.\\n\\nFrom the experiment point of view, it is necessary a better explanation about the hyperparameters or the experiments which were carried out. In addition, the single database was used to evaluate the technology which is not enough to show the big different respect to the transformed paper. Also, the comparison is not really fair. In each \\\"layer\\\" of the proposed phrase transformer, it has actually two self-attention layers, but the baseline has only one self-attention layer. In addition more methodologies should be necessary to compare the results of the experiment. \\n\\nThe architecture part is complicated to follow and I don't understand the big contribution of this paper. \\n\\nFor that reason, I recommend a reject the paper and work more for the final version\"}", "{\"title\": \"How about attention over words, does that propagate to more complicated compositions in higher layers as well?\", \"comment\": \"Thank you so much for your comprehensive response. If we accept the second explanation (\\\"the information of the 2 word compositions in lower layers will propagate to more complicated compositions in higher layers\\\"), can we also say the information related to phrases (whether 2, 3 or more) would be implicitly captured in higher layers when we use normal attention over words. If this is the case, attention over phrases is only necessary when we don't have enough data or computational resources for extensive training.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for your comments. Very interesting question.\\n\\nOne possible answer is, higher k means more parameters, which leads to overfitting. \\n\\nWe further investigate such phenomenon in PhraseTransformer. Another possible explanation is, as we are using a #deep# neural network, the information of the 2 word composisions in lower layers will propagate to more complicated compositions in higher layers. So we do not need to explicitly use higher k to represent longer phrases. To verify this, we investigate the effectiveness of our proposed method for n_layer=1. In this case, there is no such information propagation. We obtained the results below. In this setting, PhraseTransformer(k=3) performs slightly better than PhraseTransformer(k=2). This is consistent with our above explanation.\\n\\n model BLEU \\n \\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550 \\n PhraseTransformer(k=2,n_layer=1) 29.94 \\n \\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500\\u2500 \\n PhraseTransformer(k=3,n_layer=1) 30.14\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This submission proposes to consider to put attention on \\\"phrases\\\" in NLP. The phrases are generated by taking consecutive words in sentences. Each phrase is treated as a \\\"node\\\" in the same way as words. Then representations of phrases are learned in the network. The algorithm is applied to two applications, translation and pos tagging. The proposed method achieved better performance than transformer.\", \"critics\": \"1. In the abstract and the introduction, the submission argues that usefulness of phrases, which are sementic units represented by word groups. However, in the model development, \\\"phrases\\\" are really bigrams and trigrams. I don't know how much the previous argument is still valid. Particularly, there are so many bigrams and trigrams. The effect from these word combinations should have strong effect on the model, but the effect may not be explained as the argument. \\n\\n2. I think transformer can somewhat capture word combinations in bigrams and trigrams. In higher layers, transformer actually combine words in representations. What is the advantage of the proposed method over the type of combination done in transformer? \\n\\n3. The experiment only compares to transformer in the translation task. It only compares to transformer and semantic phrase transformer. Other SOTA methods (e.g. different versions of transformers) should be compared. \\n\\n4. The comparison is not really fair. In each \\\"layer\\\" of the proposed phrase transformer, it has actually two self-attention layers, but the baseline has only one self-attention layer.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper addresses an issue of compositionality in self-attention models such Transformer. A simple idea of composing multiple words into a phrase as a hypernode and representing it using a non-linear function to capture the semantic mutation is proposed. In the machine translation and PoS tagging tasks, the proposed PhraseTransformer achieves impressive gain, especially +13.7 BLEU score compared to the Transformer.\\n\\nThe motivation of the paper is very clear, and I love this kind of paper; with a simple idea, making a huge impact on the field. I appreciate the real example to compare how word-level self-attention is different from the phrase-level self-attention in Abstract and Figure 1. The problem itself; tackling the semantic compositionality of self-attention, is a very important problem, and I like the part that authors described it as an inductive bias as a model perspective. \\n\\nHowever, this work seems to be problematic in terms of presentation, clarity, and meaningful comparisons. Please see my detailed comments below.\\n\\nFirst, what exactly is \\u201csemantic mutation\\u201d? The term has been used here and there to describe the inductive bias in semantic compositionality and to show how the nonlinearity can effectively capture it. But, I couldn\\u2019t find any definition from the paper, couldn\\u2019t find any formal definition from any ACL papers, and couldn\\u2019t guess myself based on the context. I am guessing it is probably some sort of combination of meaning in phrase-level words. If so, more importantly, how does the simple non-linear function (i.e., sigmoid) can capture such semantic combinations of words? How could it make such a huge gain (PhraseTransformer vs Linear PhraseTransformer) in Table 1 on MT task? This seems to be the most important contribution of this paper, of which I don\\u2019t understand yet. \\n\\n\\nWhat is the \\u201cconcept of hypernodes in hypergraph\\u201d? I think it is not a common word that I can understand it clearly without any references or any background. It would be better to add some references for the concept. Again, I am guessing it is a sort of graph theory that decomposes a large node into small pieces but keeps their connectivity. But, then how is that exactly linked to phrases of words? If you make phrases only on consecutive words, it is basically just chunking. I don\\u2019t find any relevance of phrases in a sentence with the (hyper)graph something. \\n\\nIn Figure 2, how is the bidirectional path made between the word representations and phrase representations? In my understanding of your algorithm based on Equation 2-6, I only see the attention of phrases is computed by word attention, but not the other way. Please clarify this. If so, how does the gradient back-propagate to each other?\\n\\nThe biggest concern of this work is the scores reported in Table 1. I have checked the recent papers which used the Mutli30K(de-en) and other results from WMT[16-18] reports, but the BLEU score reported in Table 1 (20.90) seems way lower than the scores reported by any systems trained by either non-Transformer or Transformer systems. For that reason, it would be fair to include some results from the state-of-the-art systems on the same dataset. \\n\\nA minor point but in the complexity analysis, your m is basically n because you take consecutive words from n length of sentence. You better distinguish which variables are dependent on each other first. \\n\\n\\nThere are MANY typos, missing captions, grammatical errors in the paper. Here are only some of them:\\nThere is no caption for Figure 1. \\nWhen you cite a reference with its model name, you better make the citation under parenthesis. \\n\\u201cequation equation 5\\u201d -? \\u201cequation 5\\u201d\\n\\u201cWee\\u201d -> \\u201cWe\\u201d\"}", "{\"title\": \"k-2 vs k-3 Perfromance\", \"comment\": \"Very interesting work! I think the authors should explain why in all the experiments (Table 1 and Table 2) PhraseTransformer (k=2) outperforms PhraseTransformer (k=3)? If attention over phrases is the answer, shouldn't we get better results with higher k?\"}", "{\"comment\": \"Thanks for your comments.\\n\\nAccording to your first comment, we tried to run the Transformer using OpenNMT over the Multi30k dataset, and got BLEU score of 23.47. Could you specific the detailed hyper-parameters you use in OpenNMT? We are glad to see a better baseline model that you implemented, which may further verify the effectiveness of our method if we apply our optimizations over such a baseline.\\n\\nBy following the parameters in <Attention Is All You Need>, currently we use the command below:\\nonmt_train -encoder_type transformer -decoder_type transformer -gpu_ranks 0 -enc_layers 6 -dec_layers 6 -heads 8 -rnn_size 512 -src_word_vec_size 512 -tgt_word_vec_size 512 -learning_rate 0.001 -optim adam -adam_beta2 0.98\\n\\nAs to the second comment, we are trying to add more experimental analysis. It may take several days.\", \"title\": \"Detailed results by OpenNMT\"}", "{\"comment\": \"A good idea to add phrase information in the transformer.\\nHowever, I have questions about your experiments.\\n\\nI think your results on table 1 for transformer is not good. I run the transformer using OpenNMT code on Multi30K dataset, achieving the Blue value: 34.2, without adjust hyper parameters seriously. So I want to know why you report 20? \\n\\nThe second question is that why do not perform your experiments on standard dataset such as wmt14' or more languages as done on \\\"attention is all your need\\\"?\\n\\nThanks.\", \"title\": \"Results on your table 1\"}", "{\"comment\": \"I liked the overall idea and implementation.\", \"there_are_couple_of_typing_mistakes_in_the_paper\": \"1. In the overall complexity formula on page 4, I think '=' should be replaced with '+'\\n2. In section 2.3 first paragraph, H \\\\n R^{(n+m)xd} instead of H \\\\n R^{(n+m)}\\n3. wee -> we\\n\\nIt would be awesome if you can also show how the hypernode attention helps via visualizations or reduction in errors dealing with phrases.\\nCan you comment on the training time changes?\", \"title\": \"Interesting approach\"}" ] }

[ "Jiawei Du", "Hu Zhang", "Joey Tianyi Zhou", "Yi Yang", "Jiashi Feng" ]

[ "Adversarial attack", "Meta learning" ]

{ "note_id": [ "YVEehT7h70", "S1xaPlJ3jS", "BkegNlL5iS", "ryeEXlnKoS", "HkghA12YsS", "SJgSky2KoS", "H1gbr0itoB", "SkgrvaoKsH", "rkxMuTp4cr", "Hyl8uEoTKr", "SyxMU8GpYr", "Sklxkpz__S" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1576798752587, 1573806181198, 1573703719561, 1573662747605, 1573662675777, 1573662429380, 1573662265001, 1573662044882, 1572293993706, 1571824749936, 1571788361566, 1570413783973 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2580/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2580/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2580/Authors" ], [ "ICLR.cc/2020/Conference/Paper2580/Authors" ], [ "ICLR.cc/2020/Conference/Paper2580/Authors" ], [ "ICLR.cc/2020/Conference/Paper2580/Authors" ], [ "ICLR.cc/2020/Conference/Paper2580/Authors" ], [ "ICLR.cc/2020/Conference/Paper2580/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2580/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2580/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2580/Authors" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper proposes a meta attack approach based on meta learning approaches to learn generalizable prior from the previously observed attack patterns. The proposed approach is able to attack targeted models with much fewer queries. After author response, all reviewers are very positive about the paper. Thus I recommend accept.\", \"title\": \"Paper Decision\"}", "{\"title\": \"raise rating\", \"comment\": \"The authors have successfully addressed the majority of my concerns. The additional experiment of comparing m answers my question well. I would like to raise my rating.\"}", "{\"title\": \"Official Blind Review #4\", \"comment\": \"I have read the response and I think it well-addressed my concerns. I will increase my score.\"}", "{\"title\": \"Response (AnonReviewer3) -Part 2\", \"comment\": \"Cont.\", \"question_4\": \"The authors try to use the gradient from the ZOO to guide the pre-trained meta attacker to adapt to a new classifier in Algorithm 2 (line 3). What if you take multi-step ZOO to give a much stronger prior to fine-tuning?\", \"answer\": \"One significant drawback of stronger prior is that the number of queries will increase substantially. Moreover, a stronger prior may decrease the performance of the meta-attacker. The reason is that our meta-attacker is trained with clean images to achieve fast-adaptive. During the attacking phase, the meta-attacker cannot generate accurate gradients as the magnitude of adversarial perturbation increases. That is why we add the finetuning step for every $m$ iteration in Algorithm 2. We want a not so strong but estimate every $m$ iteration to make the meta-attacker fast adapted to current unseen adversarial examples with less required queries.\"}", "{\"title\": \"Response (AnonReviewer3) -Part 1\", \"comment\": \"Thank you for your thoughtful review. We have conducted three additional experiments to answer your questions; our response is as below.\", \"question_1\": \"The authors miss some important details about training meta attackers like how many images you choose to get the image and its gradient pairs. Are the images from the training set or test set. Is there an overlap between the meta-attacker training set and the test set?\", \"answer\": \"Thanks a lot for the constructive review. We have conducted two experiments in MNIST datasets responding to question 3 and question 5. The first experiment is train a vanilla autoencoder to generate gradients. The training MNIST classification model is different from the target classification model in the testing phase. The second experiment is meta-training a meta-attacker from gradients estimated from ZOO. That is, the training gradients are not generated by the white-box method, but estimated by the black-box ZOO method. The experiment results are presented as follows.\\n\\nUntargeted |Attack Success Rate | L2 Norm | Number of Queries\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nOurs (reported) | 1.000 | 1.78 | 1,103\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nZOO estimated | 1.000 | 1.77 | 1,130 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nVanilla Autoencoder | 1.000 | 1.93 | 1,899\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nInitialised attacker | 0.912 | 1.96 | 1,721\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n\\ntargeted |Attack Success Rate| L2 Norm | Number of Queries\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nOurs (reported) | 1.000 | 2.66 | 1,971\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nZOO estimated | 1.000 | 2.42 | 2,105 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nVanilla Autoencoder | 1.000 | 2.80 | 2,905\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nInitialised attacker | 0.895 | 2.81 | 3,040\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n\\nThe two experiments are conducted on 1000 randomly selected images from the MNIST testing set. As for the four approach settings: \\\"Ours (reported)\\\" is the results we report in our paper; \\\"ZOO estimated\\\" is the meta-attacker trained from ZOO estimated gradients; \\\"Vanilla Autoencoder\\\" is the autoencoder maps image to gradient trained in one different MNIST classification model; \\\"Initialised attacker\\\" is the meta-attacker with randomly initialized weights. We can see the \\\"ZOO estimated\\\" model performs closer to our reported meta-attacker. However, the \\\"Vanilla Autoencoder\\\" performs much worse (with larger L2-norm and more queries), which performs similarly to randomly initialized meta-attacker. The two experiments verify the effectiveness of our first meta-training phase. We have updated the two experiments results in appendix 6.1.2 of our paper. \\n\\n=======================\", \"question_2\": \"What if you try to learn the adversarial perturbation directly instead of learning the gradient of images?\"}", "{\"title\": \"Response (AnonReviewer2)\", \"comment\": \"We want to thank the reviewer for the positive assessment. We will release our codes and pre-trained meta-attacker. Our responses to the weak points and questions are as below.\", \"weak_point\": \"The deploy of the meta-attacker is not as easy as other black-box attack methods. It would not be practical if we need to meta training the attacker before we use it.\", \"answer\": \"Yes, and that is a critical advantage of our proposed meta-attacker. Most of the black-box attack methods are either numerical estimation or gradient optimization methods. Our method is to meta-train a model to predict gradients directly. Therefore, our method could be combined with numerical estimation methods in finetuning and with gradient optimization methods in applying gradients. For instance, we used the Adam-like method to apply predicted gradients.\", \"question_1\": \"Is it hard to meta training the meta-attacker? How long will it take for training?\", \"question_2\": \"Is it necessary to finetune the meta attacker in algorithm 2 every m iteration? How to determine the \\\"m\\\"?\", \"question_3\": \"Could the proposed meta-attacker embedded with other optimized black-box attack methods?\"}", "{\"title\": \"Response (AnonReviewer4) - Part 2\", \"comment\": \"Cont.\\nAnswer to Question 2,3:\\nOne advantage of our proposed meta-attacker is that it can be easily combined with other black-box gradient estimation methods or gradient optimization methods. In the next experiment, we compare with the suggested two baselines. We take the NES as the gradient estimation reference under $L_\\\\infty$ norm constraint.\\n\\nDue to the time limit, we compare our proposed method with the two baselines [2,4]. We compare our proposed method on the same public CIFAR10 model WRN to [2]. We tested the released codes of [4] for comparison with ours under the same experiment. The following results show that our proposed meta-attacker outperforms [2,4]. We have updated [4] results on CIFAR10 in our paper. We will add the mentioned three baselines [2,3,4] in the next version of our paper. \\n\\nUntargeted CIFAR10 , WRN model, $L_\\\\infty$ norm with $\\\\epsilon = 8/255$\\nType |Attack Success Rate | Number of Queries\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nOurs(NES) | 0.991 | 361\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n[2] | 0.997 | 389 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n\\nUntargeted CIFAR10\\nType |Attack Success Rate | L2 Norm |Number of Queries\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nOurs | 0.92 | 0.35 | 2,438\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n[4] | 1.00 | 0.43 | 5,021 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n\\n\\n\\n=======================\", \"question_4\": \"In experiments part, only 100 images from each dataset may not be representative enough. I would suggest the authors to test more samples. Also please consider only the images that can be correctly classified without perturbation, as there is no need to attack those already misclassified images.\", \"answer\": \"All the reported results in our paper are conducted only on the clean images that are correctly classified by the target model. Following the suggestion, we tested our proposed method in 1000 randomly selected images from the testing set, and the results are shown in the following table.\\ntiny-Imagenet \\nUntargeted attack |Attack Success Rate | L2 Norm | Number of Queries\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nResnet34 (reported) | 0.98 | 0.49 | 6,866\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nResnet34 (1000 samples) | 0.993 | 0.53 | 6,498 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nVGG19 (reported) | 0.98 | 0.54 | 6,826\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nVGG19 (1000 samples) | 0.994 | 0.53 | 6,530 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n\\nFrom the above results, our proposed meta-attacker is robust and performs consistently well for large-scale test samples. We have updated the new untargeted results on 1000 samples in our paper, we will update the rest targeted results on 1000 samples once the experiments are finished.\", \"reference\": \"[1] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pp. 15\\u201326. ACM, 2017.\\n[2] Yan, Ziang, Yiwen Guo, and Changshui Zhang. \\\"Subspace Attack: Exploiting Promising Subspaces for Query-Efficient Black-box Attacks.\\\" arXiv preprint arXiv:1906.04392 (2019).\\n[3] Moon, Seungyong, Gaon An, and Hyun Oh Song. \\\"Parsimonious Black-Box Adversarial Attacks via Efficient Combinatorial Optimization.\\\" ICML 2019.\\n[4] Chen, Jinghui, Jinfeng Yi, and Quanquan Gu. \\\"A Frank-Wolfe Framework for Efficient and Effective Adversarial Attacks.\\\" arXiv preprint arXiv:1811.10828 (2018).\"}", "{\"title\": \"Response (AnonReviewer4) - Part 1\", \"comment\": \"Thanks a lot for the review and the helpful advice. We answer the questions as follows:\", \"question_1\": \"The proposed meta learning approach seems to learn a universal gradient direction of an image x for all networks, which seems to be an ambitious goal. The authors did not provide any intuition or demonstrative explanations towards this. I hope the author could provide some more evidence showing that why this is achievable and is this indeed achieved by the meta learner (e.g., comparing the cosine similarity between the true gradient and the meta learner generated gradient?)\", \"answer\": \"Yes, our proposed method can using NES as the reference gradient estimation for fine-tuning and can also perform well in $L_\\\\infty$ norm constraints.\\nSince our proposed meta-attacker derived from ZOO [1], for fair comparison with other methods on query efficiency, we take the same $L_2$ norm constraints as ZOO[1] for ensuring evaluation metric to be consistent. Therefore, we also compare our results with other baselines using $L_2$ norm and attack success rate. We attempted to use NES as the reference gradients estimation for fine-tuning our meta-attacker, but it leads to much higher $L_2$ norm of the perturbation than ZOO [1]. The following results in Resnet34 tiny-Imagenet model are from taking NES for untargeted attack over 1000 randomly selected images. \\n\\nType |Attack Success Rate | L2 Norm | Number of Queries\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nOurs(ZOO)| 0.98 | 0.49 | 6,866\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nOurs(NES) | 0.82 | 1.93 | 5,586 \\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\nZOO | 1.00 | 0.47 | 25,334\\n\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\u2014\\n\\nFrom the results, although NES requires fewer queries, it will lead to lower attack success rate and much larger $L_2$ norm.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper proposed a meta attack approach that is capable of attacking a targeted model with fewer queries. It effectively utilizes meta learning approaches to abstract generalizable prior from previous observed attack patterns. Experiments on several benchmark datasets demonstrates it performance in reducing model queries.\\n\\n- The proposed meta learning approach seems to learn a universal gradient direction of an image x for all networks, which seems to be an ambitious goal. The authors did not provide any intuition or demonstrative explanations towards this. I hope the author could provide some more evidence showing that why this is achievable and is this indeed achieved by the meta learner (e.g., comparing the cosine similarity between the true gradient and the meta learner generated gradient?) \\n\\n- In Algorithm 2, the authors seem to use coordinate-wise zeroth-order gradient estimation as in ZOO. I wonder have the authors considered using NES type Gaussian noise to estimate the gradient? As it has been shown to be more efficient than coordinate-wise zeroth-order gradient estimation in (Ilyas et al. 2018).\\n\\n- I notice that the authors only consider L2 norm attack case and did not include more popular L-infinity norm case. Is there any reason why it can not be applied to L-infinity norm? I didn\\u2019t find anywhere in the algorithm that would restrict the choice of norm type. The author might also need to compare with the following recent papers regarding black-box attacks.\\n\\nYan, Ziang, Yiwen Guo, and Changshui Zhang. \\\"Subspace Attack: Exploiting Promising Subspaces for Query-Efficient Black-box Attacks.\\\" arXiv preprint arXiv:1906.04392 (2019).\\nMoon, Seungyong, Gaon An, and Hyun Oh Song. \\\"Parsimonious Black-Box Adversarial Attacks via Efficient Combinatorial Optimization.\\\" ICML 2019.\\nChen, Jinghui, Jinfeng Yi, and Quanquan Gu. \\\"A Frank-Wolfe Framework for Efficient and Effective Adversarial Attacks.\\\" arXiv preprint arXiv:1811.10828 (2018).\\n\\n- In experiments part, only 100 images from each dataset may not be representative enough. I would suggest the authors to test more samples. Also please consider only the images that can be correctly classified without perturbation, as there is no need to attack those already misclassified images.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors proposed a meta-learning based black-box adversarial attack method to reduce the number of queries. Their proposed meta attacker is trained via a meta-learning approach, and it could utilize historical query information and fast adapted to different target models. The authors conduct experiments in three datasets, and the results show that their proposed meta attacker is more efficient than other black-box attack methods.\", \"strong_points_of_this_paper\": \"1) novelty. The paper combines meta-learning and black-box attack and develops a deep convolutional model (meta attacker) to predict the gradients of another DNN model. Most of the other query-efficient attack methods focus on utilizing estimated gradients, while this paper focuses on predicting accurate gradients. That makes this paper novel. \\n2) The results are good. The proposed method could attain comparable l2 norm and attack success rate with much fewer queries. And the analysis in experiments is interesting. The generalizability experiment of the meta-attacker(meta transfer) shows that gradients in different models and different datasets have some similar patterns.\", \"weak_points_of_this_paper\": \"1) The authors could investigate more in the theoretical part. For example, the authors could gives the theoretical support of why a convolutional network could stimulate other networks' gradient. \\n2) The deploy of the meta-attacker is not as easy as other black-box attack methods. It would not be practical if we need to meta training the attacker before we use it.\", \"questions\": \"1) Is it hard to meta training the meta-attacker? How long will it take for training? \\n2) Is it necessary to finetune the meta attacker in algorithm 2 every m iteration? How to determine the \\\"m\\\"?\\n3) Could the proposed meta-attacker embedded with other optimized black-box attack methods? \\n\\nOverall, this paper is well-structured, novel, and ideas are well motivated. The query-efficient black-box attack problem is practical and meaningful. I would encourage the authors to release their codes. Last, I would recommend acceptance for this paper.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\n\\nThe authors of this paper propose a novel approach for query-efficient black-box adversarial attacks by leveraging meta-learning to approximate the gradient of given clean images.\", \"paper_strength\": \"1.\\tThe overall idea of this paper is interesting to improve the transferability of adversarial examples through meta-learning. \\n2.\\tThe paper is well-organized and easy to follow.\\n3.\\tAdequate experiment results demonstrate that the meta-attacker is efficient to decrease the number of queries for the adversarial attacks. It is interesting to see the meta attacker trained on certain source domain can be transferred to other domains.\", \"paper_weakness\": \"1.\\tThe authors miss some important details about training meta attackers like how many images you choose to get the image and its gradient pairs. Are the images from the training set or test set. Is there an overlap between the meta-attacker training set and the test set? \\n2.\\tWhat if you try to learn the adversarial perturbation directly instead of learning the gradient of images?\\n3.\\tHow about changing the meta learner to vanilla training and learning an autoencoder directly to map the image to its gradient? There could be a problem of how to incorporate the existing classifiers, but I think this could be an important baseline.\\n4.\\tThe authors try to use the gradient from the ZOO to guide the pre-trained meta attacker to adapt to a new classifier in Algorithm 2 (line 3). What if you take multi-step ZOO to give a much stronger prior to finetuning?\\n5.\\tTake a look at the confusion matrix in [1] which studies the transferability of white-box attack and black-box attack. I think it could be more interesting if the authors change the ground truth from gradient to estimated gradient from ZOO for learning meta-attacker. \\n\\n\\nLi, Yandong, et al. \\\"NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks.\\\" arXiv preprint arXiv:1905.00441 (2019).\"}", "{\"comment\": \"Thank you for the comment. Actually, it is the noise level that leads to the difference. In our paper, the noise is constrained to a much smaller level. For example, ours l2 noise is 0.33 for CIFAR10 but the original noise in AutoZoom[1] is 1.99. We compare the #queries at the similar perturbation magnitude, which is a common practice in evaluating efficacy of attacking methods. In order to achieve much smaller noise, AutoZoom consumes more query number.\\n\\n[1] https://arxiv.org/pdf/1805.11770.pdf\", \"title\": \"Response\"}" ] }

[ "Amin Ghiasi", "Ali Shafahi", "Tom Goldstein" ]

[ "certified defenses", "semantic adversarial examples", "adversarial examples", "spoofed robustness certificates", "adversarial attacks", "certified", "certifiable defenses", "networks", "certain" ]

{ "note_id": [ "k5JHM7N2e", "HyxJPqchoB", "r1lSZ9qnir", "rJe2nF5nsB", "SJllHP3yjB", "rJeb2tXkqB", "ryg2OYFAFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752559, 1573853783509, 1573853693112, 1573853620292, 1573009207830, 1571924393063, 1571883380244 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2579/Authors" ], [ "ICLR.cc/2020/Conference/Paper2579/Authors" ], [ "ICLR.cc/2020/Conference/Paper2579/Authors" ], [ "ICLR.cc/2020/Conference/Paper2579/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2579/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2579/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This work presents a \\\"shadow attack\\\" that fools certifiably robust networks by producing imperceptible adversarial examples by search outside of the certified radius. The reviewers are generally positive on the novelty and contribution of the work.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks for your feedback.\", \"comment\": \"Thanks for your constructive feedback. We have modified the paper to clarify some of the terms per your suggestion. Please find our detailed response below:\\n\\n[R1: In Table 1, for ImageNet, Shadow Attack does not always generate adversarial examples that have on average larger certified radii than the natural parallel, at least for sigma=0.5 and 1.0. Could the authors explain the reason?]\\n\\nDuring attack/crafting, we need to make an adversarial example that gets misclassified even after perturbations drawn from a Gaussian distribution centered at zero with scale sigma. During evaluation, while the augmentations are drawn from a similar distribution, the realized random variables are not identical to those used for crafting the adversarial perturbation. In ImageNet, where the dimensionality is high (224X224X3) and for larger sigmas, to have a relatively dense and representative sampling, we need to sample a lot more perturbations during adversarial example crafting. However, in our experiments, we could only sample up to 400 instances per example (the maximum batch-size that could fit on our machine with 4 GPUs is 400). This results in having a sparse sample when the standard deviation is higher. One can potentially improve these results by using larger batch-sizes (i.e., sampling more) or a more powerful GPU or even a TPU, however we do not have the resources for such experiments at this time.\\n\\n[R1: In Table 2, it is not clear to me what is the point for comparing errors of the natural images (which measures the misclassification rate of a natural image) and that of the adversarial images (which measures successful attacks rate), and why this comparison helps to support the claim that the attack results in a stronger certificates. In my opinion, to support the above claim, shouldn\\u2019t the authors provide a similar table as Table 1, directly comparing the certified radii of the natural images and adversarial images?]\\n\\nIn the original submission, we tried to produce tables that look like the tables in papers that we compare to. The randomized smoothing paper reports certified radii and also accuracy (1-error) under various perturbation bounds. However, the CROWN-IBP paper and the improved randomized smoothing paper based on adversarial training of smoothed classifiers (SmoothAdv) only report *error rates* using a fixed distance to the decision boundary. This is done because, unlike the Randomized Smoothing method, the radii are not directly calculated in the CROWN-IBP method and cannot be accessed directly; CROWN-IBP takes a fixed radius chosen by the user, and either produces or fails to produce a certificate for that radius. This is in contrast to randomized smoothing, which outputs different radii for different images (a larger radius means a stronger certificate).\", \"in_regards_to_why_we_compare_the_errors_on_natural_images_and_those_of_our_adversarial_images\": \"Please see the (updated) last paragraph of Section 5, which explains this comparison in detail. In short - we are comparing the rate at which natural images certify to the rate at which adversarial images certify. For the case of large perturbations, we find that our adversarial image produce certificates more often than natural images! For small perturbations, our attack still produces certificates reasonably often, although not quite as frequently as natural images. This shows that certificates alone cannot be used to reliably discern between natural images, and adversarial images produced by the proposed shadow attack.\\n\\n[R1: From Figure 9, we see the certificate radii of the natural have at least two peaks. Though on average the certificate radii of the adversarial attacks is higher than that of the natural images, it is smaller than the right peak. Could the authors elaborate more of the results?]\\n\\nThis happens because, for CIFAR-10, the smoothed classifier is very \\u201cconfident\\u201d on a subset of the validation images which correspond to that right peak. Here, our use of \\u201cconfidence\\u201d should not be confused with the confidence of a network (output of the softmax layer). For the purpose of the certified radii, the \\u201cconfidence\\u201d we are interested in is related to the prediction of the network on the Gaussian perturbed images (i.e., a very high \\u201cconfident\\u201d example is an example where all of the perturbed images get the same label). \\n\\n[R1: Sim(delta) should be Dissim(delta) which measures the dissimilarity between channels. A smaller dissimilarity suggests a greater similarity between channels.]\\nGood point! We have updated this in the revised document, and we think it enhanced clarity.\\n\\n[R1: Lambda sim and lambda s are used interchangeably. Please make it consistent. ]\\nFixed. Thank you.\\n\\n[R1: The caption of Table 1 is a little vague. Please clearly state the meaning of the numbers in the table.]\\nIn the revision, we have described what the numbers are representing in more detail.\"}", "{\"title\": \"Thanks for your feedback.\", \"comment\": \"Thank you for the encouraging review.\\n\\n[R3: Weakness: It would be good to see some comparison to the state of the art ]\\n\\nWith regards to your comment on attacking the current state of the art method for smoothed classifiers, we have added new results to the resubmission (Appendix B), in which we attack the adversarially trained smooth classifier [1].\\n\\n[1]. Salman et al., \\u201cProvably Robust Deep Learning via Adversarially Trained Smoothed Classifiers\\u201d, NeurIPS 2019\"}", "{\"title\": \"Thanks for your feedback.\", \"comment\": \"Thanks for your constructive feedback. We have modified the paper to include some of the experiments you have suggested. Please find our detailed response below:\\n\\n[R4: First, the proposed attack method can yield adversarial perturbations to images that are large in the \\\\ell_p norm. Therefore, the authors claim that the method can attack certified systems. However, attack in Wasserstein distance and some other methods can also do so. They can generate adversarial examples whose \\\\ell_p norm is large.\\nI think the author should have some discussions about these related methods.]\\n\\nThank you for pointing us out to the missing related work which we have included in the revision. Indeed, the Wasserstein attack and the other previously mentioned non-$\\\\ell_p$ bounded attacks are alternatives for producing quasi-imperceptible non-$\\\\ell_p$ bounded adversarial examples. Any of these methods can alternatively be used for generating non $\\\\ell_p$ bounded attacks. However, one major advantage of our attack method over the Wasserstein attack may be its simplicity and scalability. \\n\\nPer your suggestion, we ran experiments using the Wasserstein attack. The authors of [1] suggest that the Wasserstein PGD attack works best when the attacker takes PGD steps in $ell_p$-norm directions and then project the noise back onto the Wasserstein ball. We used their official implementation and adapted it to attack the Randomized Smoothed classifier. Based on the official implementation, after every 10 iterations, if the attack is not successful, we increase the radius of the wasserstein ball in which the noise is projected back onto. Consequently, the attack is always able to reach a comparable, but slightly weaker, spoofed certified radii (~ 67% that of the shadow attack) at the cost of slightly more perceptible adversarial noise in difficult cases. Note that the reason that the examples are more perceptible than those from [1] is that they are made to produce large certified radii and not only cause misclassification (i.e., the entire Gaussian augmented batch needs to get misclassified.) A comparison of the resulting images and average certified radii of those images can be found in the following anonymized link:\", \"https\": \"//docs.google.com/spreadsheets/d/1F0P8aOD_5aiVjW3CrR49fudz4EgrORz7v4t0ZIJEBAo/edit?usp=sharing.\\n\\n[1] Wong et al., \\u201cWasserstein Adversarial Examples via Projected Sinkhorn Iterations\\u201d. \\n\\n\\n[R4: Second, I notice that compared to the result in Table 1, PGD attack can yield better results [1]. I hope to see some discussions about this. Also, Table 1 is really confused. I would not understand the meaning if I am not familiar with the experiment settings.]\\n\\nPer your request, we have attacked the work of [2] and reported results of attacking the pre-trained SmoothAdv classifiers (available in [3]) in Appendix B. Similar to the non-adversarially trained smooth classifier included in the original submission, we can produce adversarial examples for the SmoothAdv classifier which on average produce larger certified radii than their natural example counterpart. Also, in the revised document, we have expanded the caption of Table 1 to make sure that it is clear what a certified is and that a larger radii is better.\\n[2]. Salman et al., \\u201cProvably Robust Deep Learning via Adversarially Trained Smoothed Classifiers\\u201d, NeurIPS 2019 \\n[3]. https://github.com/Hadisalman/smoothing-adversarial\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper presents a new attack: Shadow Attack, which can generate imperceptible adversarial samples. This method is based on adding regularization on total variation, color change in each channel and similar perturbation in each channel. This method is easy to follow and a lot of examples of different experiments are shown.\\nHowever, I have several questions about motivation and method.\\n\\nFirst, the proposed attack method can yield adversarial perturbations to images that are large in the \\\\ell_p norm. Therefore, the authors claim that the method can attack certified systems. However, attack in Wasserstein distance and some other methods can also do so. They can generate adversarial examples whose \\\\ell_p norm is large.\\nI think the author should have some discussions about these related methods. \\n\\nSecond, I notice that compared to the result in Table 1, PGD attack can yield better results [1]. I hope to see some discussions about this. Also, Table 1 is really confused. I would not understand the meaning if I am not familiar with the experiment settings.\\n\\n[1] Salman, Hadi, et al. \\\"Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers.\\\" Neuips (2019).\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper presents a new attack, called the shadow attack, that can maintain the imperceptibility of adversarial samples when out of the certified radius. This work not only aims to target the classifier label but also the certificate by adding large perturbations to the image. The attacks produce a 'spoofed' certificate, so though these certified systems are meant to be secure, can be attacked. Theirs seem to be the first work focusing on manipulating certificates to attack strongly certified networks. The paper presents shadow attack, that is a generalization of the PGD attack. It involves creation of adversarial examples, and addition of few constraints that forces these perturbations to be small, smooth and not many color variations. For certificate spoofing the authors explore different spoofing losses for l-2(attacks on randomized smoothing) and l-inf(attacks on crown-ibp) norm bounded attacks.\", \"strengths\": \"The paper is well written and well motivated. The work is novel since most of the current work focus on the imperceptibility and misclassification aspects of the classifier, but this work addresses attacking the strongly certified networks.\", \"weakness\": \"It would be good to see some comparison to the state of the art\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"The paper proposes a new way to generate adversarial images that are perturbed based on natural images called Shadow Attach. The generated adversarial images are imperceptible and have a large norm to escape the certification regions. The proposed method incorporates the quantities of total variation of the perturbation, change in the mean of each color channel, and dissimilarity between channels, into the loss function, to make sure the generate adversarial images are smooth and natural. Quantitative studies on CIFAR-10 and ImageNet shows that the new attack method can generate adversarial images that have larger certified radii than natural images. To further improve the paper, it would be great if the authors can address the following questions:\", \"In Table 1, for ImageNet, Shadow Attach does not always generate adversarial examples that have on average larger certified radii than the natural parallel, at least for sigma=0.5 and 1.0. Could the authors explain the reason?\", \"In Table 2, it is not clear to me what is the point for comparing errors of the natural images (which measures the misclassification rate of a natural image) and that of the adversarial images (which measures successful attacks rate), and why this comparison helps to support the claim that the attack results in a stronger certificates. In my opinion, to support the above claim, shouldn\\u2019t the authors provide a similar table as Table 1, directly comparing the certified radii of the natural images and adversarial images?\", \"From Figure 9, we see the certificate radii of the natural have at least two peaks. Though on average the certificate radii of the adversarial attacks is higher than that of the natural images, it is smaller than the right peak. Could the authors elaborate more of the results?\", \"Sim(delta) should be Dissim(delta) which measures the dissimilarity between channels. A smaller dissimilarity suggests a greater similarity between channels.\", \"Lambda sim and lambda s are used interchangeably. Please make it consistent.\", \"The caption of Table 1 is a little vague. Please clearly state the meaning of the numbers in the table.\"]}" ] }

[ "Sean M. Hendryx", "Andrew B. Leach", "Paul D. Hein", "Clayton T. Morrison" ]

[ "meta-learning", "image segmentation" ]

{ "note_id": [ "o30pC3H2Uu", "H1gmg3Knir", "S1xgLi0osS", "SkeVF5RoiS", "H1e3cYRsiH", "rkxZZYCiiB", "SJxiZOAiiH", "HklJGPAsoB", "rklbnw5qsr", "HJgcpxe1qH", "H1x8O66AFB", "BJgryYKpFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752530, 1573850090557, 1573804872319, 1573804667726, 1573804435776, 1573804280762, 1573804034648, 1573803783223, 1573722025499, 1571909825596, 1571900781944, 1571817693515 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/Authors" ], [ "ICLR.cc/2020/Conference/Paper2578/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2578/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2578/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The reviewers reached a consensus that the paper was not ready to be accepted in its current form. The main concerns were in regard to clarity, relatively limited novelty, and a relatively unsatisfying experimental evaluation. Although some of the clarity concerns were addressed during the response period, the other issues still remained, and the reviewers generally agreed that the paper should be rejected.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Regarding visualizations\", \"comment\": \"Thank you for noting that visualizations of the EfficientLab architecture and qualitative predictions would help. We have added both. The EfficientLab diagram is included in section 4 and example predictions are included in appendix section A.\"}", "{\"title\": \"Regarding clarity\", \"comment\": \"Thank you for pointing out that the narrative that ties the paper together is unclear. In our revised version of the manuscript, we have made sure the research questions, results, and primary contributions are much more cohesive.\"}", "{\"title\": \"Regarding motivation for scaling few-shot learning systems to many-shot domains\", \"comment\": \"Regarding your concerns: \\\"The paper claims to make several contributions but it\\u2019s hard to concretely understand them in several sections. For instance, the abstract mentions -- 'A natural question that \\u2026.. human level performance in both.' The statement is slightly unclear to me -- is the intended sentiment the fact that the goal should be to develop a single algorithm that works well for both few-shot and many-shot settings?\\nYes, that is exactly the goal.\\n\\\"If so, why should that be the case?\\\"\\nThank you for noting the justification is not self-evident. We have added this sentence to the introduction: \\u201cThis property is desirable because training and maintaining multiple models is more cumbersome than training and maintaining one model.\\u201d\"}", "{\"title\": \"Thank you for your excellent feedback\", \"comment\": \"Firstly, thank you again for taking the time to review our paper and provide helpful feedback. We will address your feedback point by point.\", \"in_regards_to\": \"\\\" Similarly, it\\u2019s unclear what question (3) in the introduction is trying to address. Which \\u201cdata\\u201d (training / testing set of tasks) is the fixed update policy not conditioned on? Could the authors clarify this?\\\"\\nThank you for pointing out the ambiguity here. The hyperparameters of the update routine of MAML-type algorithms are not automatically a function of any data, but are hardcoded by researchers and guessed via trial and error on validation datasets. But, what we specifically are concerned with is that the update routine\\u2019s hyperparameters are not conditioned on the few shot examples for the task, $\\\\mathcal{T}_i$ being adapted to. We utilize random search with successive halving of the search space to estimate the update routine hyerparamaters that maximize the expectation of the intersection over union. We have revised research question 3 to be more explicit: Are MAML-type algorithms hindered by having a fixed update policy for training and testing tasks that is not conditioned on the available labeled examples for a new task?\", \"we_also_discuss_this_clearer_and_more_formally_in_a_new_section\": \"5.3 HYPERPARAMETER SELECTION\"}", "{\"title\": \"Thank you for your excellent feedback\", \"comment\": \"Firstly, thank you for taking the time to review our paper and provide helpful feedback. We have incorporated your feedback and the revised manuscript is much stronger now.\\n\\nRegarding your concerns, thank you for the stating your concerns clearly around novelty. Our novelty proposition is stated in our general response to reviewers above\\n\\n1 and 5 shots are the canonical number of training examples per class in the few shot literature. So for easy comparison, we have included results on the FSS-1000 dataset for those. Furthermore, each class in FSS-1000 only has 10 labeled images total so the absolute maximum you could evaluate via cross-validation would be 9 training shots. One of our central claims is that understanding how meta-learning systems scale to many-shot regimes deserves more attention, so we have released the small FP-k benchmark. We believe that this is also a novel contribution.\\n\\nThank you again for your excellent feedback and we look forward to hearing from you!\"}", "{\"title\": \"continuing previous thread\", \"comment\": \"(2.1)\\nThanks for pointing this out. We agree that \\u201chuman-level\\u201d is an ill-defined term that is both open to subjective interpretation and complex to define precisely. In the few shot domain, accurate segmentation of an object in unseen images requires skills in both recognizing the object and identifying the pixels that belong to that object. How these skills may be implemented in the human perceptual system is out of the scope of this paper. We have removed all references to \\u201chuman level\\u201d segmentation in the paper, noting human performance only in the first sentence with the general statement that humans have a remarkable capability to learn from a small number of labeled examples and gain expertise as more data becomes available.\\n\\n(2.2)\\nWe apologize for the confusion around our use of early stopping. Early stopping is only used inside the UHO algorithm and is entirely an implementation optimization to reduce the runtime of the random search. In our specific application of UHO, it returns the best combination of learning rate and number of iterations seen on the provided tasks. We have provided more details on UHO algorithm and the k-shot learning curves in the appendix and moved the only mention of early stopping to the UHO section in the appendix. We do not use tuned hyperparameters for the k-shot learning curves experiments.\\n\\n(2.3)\\nThank you for noting that important paper. We were aware of it and that it is cited by many of the works we reference. We didn\\u2019t see the direct need to cite it since it is referenced by successor papers that we do cite. Furthermore, we did not experiment with the Meta-SGD methods due to concern regarding an effective doubling in the number of parameters and computational overhead and the lack of a first order approximation. That said, it is important and related work so we have added it to the related work section.\\n\\n(2.4)\\nRuntime is important but was not the focus of this paper. Given that it wasn't the focus we did not have time compute empirical runtimes for both another work and our model. We leave this for future work. Depending on the implementation, RelationNet's can have $O(n^2)$ time costs in the number of training examples.\\n\\n(2.5)\\nThis figure is doing double duty. One point we wanted to show is that the representations are not robust to large changes in the update routine\\u2019s hyperparameters. Thank you again for noting the importance of showing the variance of results. We have since added these for the meta-test set. The lower bound of the 95% confidence interval of mean IoU is greater than the mean IoU from using the meta-training learning rate, so the improvements do appear to be significant.\\n\\n(2.6)\\nIt was noted in the Neural Network Architecture section (which has been renamed to EFFICIENTLAB ARCHITECTURE FOR IMAGE SEGMENTATION) that we use a softmax to constrain the predictions to real numbers in [0, 1].\\n\\n(3.1)\\nThanks for pointing this out. We\\u2019ve corrected the latex.\\n\\n(3.2)\\nThanks for pointing this out. We\\u2019ve corrected the latex.\\n\\n(3.3)\\nWe\\u2019ve added the reference for the Dice score, which is also called the S\\u00f8rensen\\u2013Dice coefficient, or the F1 score in many modern domains.\\n\\nThank you again for your excellent feedback and we look forward to hearing from you!\"}", "{\"title\": \"Thank you for the critical feedback.\", \"comment\": \"Firstly, thank you for taking the time to review our paper and provide this helpful feedback. We have responded to general themes the reviewers had in the above comment but will also drill into some details here.\", \"going_by_your_numbered_feedback\": \"(1.1)\\nThank you for noting the importance of the variance in determining significance. We have rerun the meta-learning experiments with the same test set of tasks from the FSS-1000 authors (this test set hadn\\u2019t been released previously but we emailed the authors and they sent us the test set and then released it on their github page) and have reported mean IoU results with 95% confidence intervals. These results are computed by training and evaluating on each of the 240 held out tasks with 2 different random splits of the 10 examples, yielding 480 results data points per meta-learning method and architecture we experimented with.\\n(1.2)\\nThank you for critically considering this part of the paper. We agree that it should be obvious that the meta-test results of gradient-based meta-learning depend on the update routine\\u2019s learning rate (and other hyperparameters such as number of iterations), but the update routine\\u2019s hyperparameters have not been systematically addressed in previous work (Finn et al. 2017, Rusu et al. 2018, Nichol and Schulman 2018) but rather guessed by the researchers. This is concerning because it does not allow us to know what the upper limits of gradient-based meta-learning are. While we agree that it is trivial to show directly that the test-set loss is dependent on the update routine\\u2019s hyperparameters, we believe that it is a contribution in its own right to mathematically define the generalization error. As far as we can tell from the papers we have cited, the generalization error has not been stated clearly in terms of the non-computable expected loss over the task-generating distribution. We have made this section a subsection of preliminaries and revised it to be more compact and to clearly and concisely show that generalizing to unseen tasks is a function of both the initialization and the update routine\\u2019s hyperparameters.\\n\\n(1.3)\\nThank you for pointing out these details that our paper lacks clarity around. This is really helpful and echoed by reviewer 1. We did mean that with few shots, it is a desirable property for a meta-learning approach to perform as well as existing few-shot methods, but with many shots, it should perform as well as a standard learning algorithm. Performing well here would mean having both high accuracy and comparable inference runtime. We have added additional clarity to the paper and removed ambiguous wording. In particular, \\u201csmoothly\\u201d was a rather ambiguous term, we have revised our submission here to state clearly that we find that our meta-learned initializations continue to provide value as more data becomes available. Regarding providing evidence that other few-shot approaches do not meet this criteria, we did describe general evidence in the Related Work section by mentioning that existing few shot algorithms use model ensembling, relation networks which \\u201ctypically are $O(n^2)$ in memory consumption and[/or] runtime [depending on the implementation] for the update routine when adapting to the new task, where n=number of training examples\\u201d, or iterative optimization modules which require multiple passes through the network per evaluation. We have introduced the FP-k benchmark specifically so that future meta-learnign work can compare the scalability of their approach.\\n\\n(1.4)\\nThis is a good point. We did not provide enough details on Figure 2. We are in complete agreement that is not fair to say the same hyperparameters will yield equally maximal results for meta-learned and imagenet+random initializations. We have provided more details on our hyperparameters in the experiments section. Importantly, we have also reframed the discussion of this experiment as a small benchmark that we hope will attract additional inquiry into the empirical performance of learning algorithms that must scale from few to many-shot settings.\", \"references\": \"Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. arXiv preprint arXiv:1703.03400, 2017.\\nAndrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osin- dero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018.\\nAlex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999, 2018.\"}", "{\"title\": \"Your feedback has been extremely helpful and we have addressed your points in the revised manuscript and below comment\", \"comment\": \"Dear Reviewers,\\n\\nFirst we want to sincerely thank all the reviewers for their critical responses and taking the time to analyze our work in depth. Your feedback is greatly appreciated and has helped us clarify many components and make the paper more cohesive and all-around stronger. Second, we apologize for not responding sooner; you raised many good points that we wanted to address as best as we could in the available time.\\n\\n We have revised the paper throughout such that all concerns that we could attend to in time have been addressed. We have worked to make the primary contributions and the narrative clearer and removed distracting details that were not related to one of our primary contributions. Notably, you will notice that the generalization error section is only a subsection of Preliminaries and is much more compact now. In exchange, we have put additional explanation and experimental ablation results into the proposed architecture EfficientLab.\\n\\nIn response to concerns of the significance of this work, we have rerun all experiments we had time to (including new ablation experiments) with the exact same test set of tasks used by the FSS-1000 authors and have included 95% confidence intervals on our results. We have also provided additional details on the k-shot learning curves and experiments.\\n\\nWe also want to thank you for your concerns regarding novelty and hope that the revised version of the manuscript has a clearer and stronger novelty proposition. We do politely disagree that applying MAML-style algorithms to larger networks and image segmentation is not novel. This point is clearly evidenced by 2 papers from ICLR 2018 stating that MAML-type algorithms are \\u201cunproven for the high dimensionality and skewed distributions of segmentation\\u201d (Rakelly et al., 2018) and even more clearly by Mishra et al. (2018) who found that MAML \\u201coverfit significantly\\u201d when using a deeper neural network. Our novelty proposition in relation to MAML-type algorithms specifically is that they do in fact extend to higher dimensional parameters spaces and the skewed distributions of image segmentation when the hyperparameters of the update routine are tuned to minimize empirical error given a fixed meta-learned initialization, $\\\\theta_0$. In hindsight, this point may seem obvious, but looking at the existing published literature, it is clear that this point has not been investigated in detail. By defining the generalization error of a meta-learning system (which we consider another novel contribution), it is self-evident that the generalization error of meta-learning systems is also a function of how they update based on new data. We also have shown that the EfficientLab neural network architecture is an interesting and useful contribution. In our revised version of the manuscript, we have included ablations around the design choices of the architecture. Lastly, the small FP-k benchmark is a novel contribution which we hope will garner more interest to the study of how meta-learning systems scale as more data becomes available. We will included a table next to figure 3 for clarity as a benchmark tomorrow.\", \"notation_and_terminology_changes\": [\"We have changed notation in a few places to make the paper clearer, easier to read, and more consistent with other key papers in meta-learning. Specifically we have changed:\", \"$\\\\tau$ to $\\\\mathcal{T}_i$ the random variable denoting the generating distribution of a task (such as the generating distribution that would generate image-label pairs for the task black bear).\", \"$\\\\mathbb{H}$ to $\\\\omega$ indicating the hyperparameters that are built into the update routine $U$. We do this because using a set indicator was visually a bit more distracting than the variable $\\\\omega$\", \"minimize to argmin\", \"$\\\\tau_s$ to $\\\\mathcal{D}_s$ denoting an empirical sample of image-mask pairs from a task $\\\\mathcal{T}_i$\", \"Lightweight skip decoder to residual skip decoder (RSD) emphasizing that one of the most meaningful changes in the atrous spatial pyramid pooling block is the residual connection\"], \"additional_gaps_that_have_been_filled\": \"\", \"we_neglected_to_exclude_a_few_details_and_citations_which_we_have_added_to_the_revision\": \"- L2 term to the loss\\n - Simple augmentations of training examples during training\\n - Citation of loss function which we adapted\\n\\nWe will be responding to your individual points in more depth tomorrow and we hope you will have the time to look at the revised version of the paper. Thank you again for your feedback and time.\\n\\nSincerely,\\nThe authors\", \"references\": \"Kate Rakelly, Evan Shelhamer, Trevor Darrell, Alyosha Efros, and Sergey Levine. Conditional networks for few-shot semantic segmentation, 2018. URL https://openreview.net/ forum?id=SkMjFKJwG.\\nNikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations, 2018. URL https: //openreview.net/forum?id=B1DmUzWAW.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes to apply MAML-style meta-learning to few-shot semantic segmentation in images. It argues that this type of algorithm may be more computationally-efficient than existing methods and may offer better performance with a higher number of examples. They further propose to perform hyper-parameter search to choose a new learning rate for the inner learning process after optimizing for the network parameters.\\n\\nAs far as I know, this is the first paper to apply gradient-based (i.e. MAML-style) meta-learning to this specific problem. Existing approaches to few-shot semantic segmentation have mostly used multi-branch conv-networks to condition the output on the training examples. This paper shows that (FO)MAML achieves similar accuracy to the FSS-1000 baseline. This is an empirical contribution in itself. The paper also demonstrates that the EfficientNet architecture can be applied to segmentation.\", \"major_concerns\": \"(1.1) The improvement obtained by the hyper-parameter optimization seems quite marginal (79.0 - 81.4 and 73.3 - 73.9) and there is no study of the variance of the results. The fact that better performance is obtained by tuning the learning rate on the *training* set suggests to me that the improvement might not be significant. You could repeat the experiment by sampling multiple different training and testing sets (with different classes) to estimate (some of) the variance.\\n\\n(1.2) The formalization in Section 4 is mathematically appealing but seems unnecessary. In the end, the paper is essentially arguing that it's better to use different learning rates (for the inner loop) during meta-training and meta-testing. This seems obvious, since this includes equal learning rates as a special case. The paper proposes to optimize the latter learning rate using a gradient-free method. This argument can be made without considering generalization bounds.\\n\\n(1.3) One of the central claims of the paper is that \\\"meta-learned representations smoothly transition as more data becomes available.\\\" It's not entirely clear what this means. I suppose it means that, with few shots, it should perform as well as existing few-shot methods, but with many shots, it should perform as well as a standard learning algorithm. The paper failed to present any evidence that existing algorithms for few-shot segmentation do not satisfy this property. It would strengthen the argument to include an existing few-shot segmentation algorithm in Figure 2.\\n\\n(1.4) The details of the experiment in Figure 2 are not clear. Was a different number of iterations used when there are hundreds of examples? Were different hyper-parameters used when optimizing from a pre-trained checkpoint? It would be unfair to use the same hyper-parameters which had been optimized specifically for the meta-learned initialization.\", \"other_issues\": \"(2.1) It's not clear what it means to achieve human-level performance in the few-shot task and the many-shot task. What is human-level performance at few-shot segmentation? It seems to me that humans are capable of segmenting novel objects (i.e. zero-shot) with almost perfect accuracy. Does this mean that your method should achieve the same accuracy with few- and many-shots?\\n\\n(2.2) The use of early stopping was unclear. Do you use a fixed number of SGD iterations during training and a variable number of iterations (determined by a stopping criterion) during testing? However, this seems to be contradicted by the statement that the UHO algorithm determined an optimal number of iterations (8) for testing? On the other hand, this seems like too few iterations with hundreds of shots. Maybe the automatic stopping criterion was only used with many shots? Or maybe training proceeds until either the stopping criterion is satisfied or the maximum number of iterations is reached? Furthermore, the early stopping criterion was not specified.\\n\\n(2.3) Missing reference: Meta-SGD (arxiv 2017) considers a different learning rate for every parameter and updates the learning rates during meta-training .\\n\\n(2.4) There was no discussion of the running time of different methods. This would be particularly interesting in the many-shot regime. How slow are the RelationNet approaches?\\n\\n(2.5) It is claimed that Figure 1 demonstrates that \\\"the estimated optimal hyperparameters for the update routine ... are not the same as those specified a priori during meta-training\\\". However, it seems that the optimal learning rate is awfully close to the dotted blue line (within the variance of the results).\\n\\n(2.6) For the IOU loss (equation 10), what are the predicted y values? Are they arbitrary real numbers? Do you use a sigmoid to constrain them to real numbers in [0, 1]?\", \"minor\": \"(3.1) It is not worth stating the optimal learning rate to more than 3 or 4 significant figures.\\n(3.2) Use 1 \\\\times 1 instead of 1 x 1.\\n(3.3) Is there a reference for the Dice score? Where does the name come from?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary - The paper first makes the observation that training algorithms and architectures for meta-learning have become increasingly specific to the few-shot set of tasks. Following this, the authors first investigate if it\\u2019s possible to learn good initializations for dense structured prediction tasks such as segmentation (across arbitrary amounts of available input data). Concretely, the claimed contributions of the paper include -- (1) extension and analysis of first order MAML like approaches to image segmentation; (2) using a formalized notion of the generalization error of episodic meta-learning approaches to decrease error on unseen tasks; (3) doing this via a novel neural network parametrically efficient segmentation architecture and (4) empirically comparing meta-learned initializations with ImageNet pre-trained initializations with increasing training set sizes.\\n\\nStrengths\\n\\n- Apart from the flaws mentioned under weaknesses, the paper is generally easy to follow. While it\\u2019s somewhat hard to understand the motivations and the concrete contributions made by the paper, sections are more-or-less well-written.\\nUsing the proposed hyper-parameter search scheme over first-order MAML approaches demonstrates improvements over not baselines which do not use the same and baselines which do not involve meta-learning.\\n\\nWeaknesses\\n\\nThe paper has some major weaknesses that affect the clarity of the points being conveyed in several sections. These weaknesses form the basis of my rating and addressing these would not only help in adjusting the same but would also help in improving the current version of the paper significantly. Highlighting these below:\\n\\n- The paper claims to make several contributions but it\\u2019s hard to concretely understand them in several sections. For instance, the abstract mentions -- \\u2018\\u2019A natural question that \\u2026.. human level performance in both.\\u201d The statement is slightly unclear to me -- is the intended sentiment the fact that the goal should be to develop a single algorithm that works well for both few-shot and many-shot settings? If so, why should that be the case? Essentially, what is the limiting factor being identified that restricts few-shot approaches from performing well in many-shot settings? Maybe the statement could be framed better but in it\\u2019s current form it\\u2019s unclear what is being conveyed. When this is mentioned again in the introductory section, it is followed by a statement indicating that meta-learning an initialization is one solution. Why is this surprising? Maybe I\\u2019m mis-understanding the motivation behind the claim. Could the authors clarify this? \\n\\n- Similarly, it\\u2019s unclear what question (3) in the introduction is trying to address. Which \\u201cdata\\u201d (training / testing set of tasks) is the fixed update policy not conditioned on? Could the authors clarify this?\\n\\n- The description of the single update hyper-parameter optimization (UHO) is hard to understand in Sec. 4. -- specifically the text surrounding eqns (5) and (6). The transition from Eqn (5) -> Eqn (6) is unclear. Could the authors clarify this clearly? This section is further referred to in subsequent sections as a supporting basis for some of the obtained results (specifically, the last para on page 6)\\n\\nReasons for rating\\n\\nI found certain sections of the paper particularly hard to understand and interpret. I would encourage the authors to address these more clearly in the responses. The highlighted strengths and weaknesses of my rating and addressing those clearly would help in improving my current rating of the paper.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper examines the performance of MAML, FOMAML, and Reptile gradient based meta-learning algorithms on the task of semantic image segmentation. The paper proposes to do black box optimization (successive halving) on the hyperparameters of the inner loop of the gradient meta-learners for improved performance. The paper proposes some modification on the segmentation model architecture (no ablation study presented). Finally, it is shown that pre-training using meta-learning on similar segmentation tasks works better then just using ImageNet based model pre-training.\", \"in_its_current_form_i_suggest_to_reject_the_paper_and_urge_the_authors_to_improve_it_according_to_the_following_points\": \"1. In parts (specifically the intro and some other earlier parts) the paper is very well written, but the later parts, the description of the model modifications (did you consider to add an architecture diagram?), the details of the experiments, the punch-line of the theory development that has been attempted, etc are not very clear and hard to follow. I suggest the authors to improve the readability of these parts, add some helpful / motivating diagrams and examples (perhaps some qualitative results too?), state more clearly what is used for meta-training? how it is made sure that meta-testing is done on a separate set of categories? (I did not see this split in the Appendix) and etc\\n\\n2. My main concern is novelty. As it stands, the current novelty proposition is: black-box optimization of LR and number of iterations in MAML style meta learning (hardly novel), architecture modifications (no ablation study if these help or not), small improvement on FSS-1000 5-shot test (what about other shots? still not sure about the splits), and showing meta-training on similar tasks is better then not doing it (that is using ImageNet pre-trained backbone for init) - again hardly a novel insight. For the last point, saying that meta-learned model was initialized from scratch does not cut it, as it was meta-trained on massive data that is more related to the test tasks then the ImageNet.\\n\\nI suggest the authors to mainly focus on 2, although making the writing clearer and better is also very important for a high quality paper.\"}" ] }

[ "Tassilo Klein", "Moin Nabi" ]

[ "representation learning", "performance", "data", "priori", "disentanglement", "disentanglement deep learning", "latest machine", "technology", "era", "success" ]

{ "note_id": [ "LTAafnEG6", "Hkg5Zu9Q5S", "HJxNUdZ6Fr", "HJl6GIKeFB", "SyegKnHROH", "Hke4nKuFur" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798752502, 1572214786335, 1571784779802, 1570965013259, 1570819191694, 1570503084038 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2577/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2577/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2577/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2577/Authors" ], [ "~Jay_Pavagadhi1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper leverages variational auto-encoders (VAEs) and disentanglement to generate data representations that hide sensitive attributes. The reviewers have identified several issues with the paper, including its false claims or statements about differential privacy, unclear privacy guarantee, and lack of related work discussion. The authors have not directly addressed these issues.\", \"title\": \"Paper Decision\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper investigates the use of variational auto-encoders (VAEs) and disentanglement to create high quality data representations that hide sensitive attributes. The authors consider two settings: (1) a supervised setting where both the sensitive labels and the downstream machine learning task labels are available to the data holder, and (2) a weakly supervised setting where only sensitive attributes are available. For both settings, the authors propose creating data representations that have 2 components: component 1 captures any information in the data bout the sensitive attributes, and component 2 captures everything else (especially the information useful for a downstream machine learning task). The goal is to ensure that these two components are disentangled. This is done by training classifiers on each component: 2 classifiers that try to reconstruct the sensitive attributes from each component separately (and in the supervised setting, 2 other classifiers that guess the downstream ML task from each component separately). The overall loss captures all components and ensures that one cannot reconstruct the sensitive attributes from component 2.\\n\\nOverall, the paper addresses an interesting and timely problem that is of great relevance to this community. However, despite being generally clear, the paper has many grammatical errors and typos. In its current shape, the paper cannot be published. \\n\\nMore importantly, the paper has a number of issues that need to be improved:\\n\\n1. The introduction makes a number of claims that are incorrect. For instance, the introduction claims that under differential privacy (and refers to Abadi et al.), machine learning models are learned from anonymized data and that complex training procedures run the risk of exhausting the budget before convergence. This is not true because under the classical setting of differential privacy, the models are trained on clean data BUT the process of learning is anonymized (see DP-SGD from Abadi et al.). More importantly, recent research and results in this space indicate that one can train high quality machine learning models with strong differential privacy guarantees (check McMahan et al. ICLR18: Learning Differentially Private Recurrent Language Models). Further, the authors claim that Federated Learning does not allow for the use of the trained model in a central setting. This is also not true because federated learning is an approach to train models on massively decentralized data -- in a way that is orchestrated by a service provider. The learned model can be used in a centralized or decentralized service. Moreover, the communication cost of training models under federated learning may be lower than communicating the data (several hundreds if not thousands of high resolution images) to a central service provider, so the claims about communication are not always correct. Finally, the authors claim that federated learning and differential privacy are useful when the private attributes are known a priori and these approaches fail to provide privacy protections when the private information contained in the dataset isn't identified. This is wrong. Both approaches do not require the knowledge of private information. They, however, require the knowledge of the downstream machine learning task -- something that may not always defined a priori.\\n\\n2. The privacy guarantees provided by the authors' approach are rather weak and unclear. (1) They assume that a trusted data holder already access to the dataset and wants to release private representations. In the supervised setting, why not just revealing the learned model? In the weakly supervised setting, where will the downstream machine labels come from? Are these representations released only for unsupervised learning tasks? This should be clarified -- and the experiments at the end should reflect this. (2) The privacy guarantees are with respect to (a) a computationally and statistically bounded adversary (i.e. there may very well be stronger adversaries with access to side information that can perfectly reconstruct the sensitive attributes), and (b) pre-defined sensitive attributes (i.e. there may be other sensitive attributes that one can learn from the published representations that are not captured by this framework). \\n\\n3. The paper makes no attempt to properly survey the literature on learning representations under censorship and fairness constraints. For example, they do not reference and compare against: (a) Censoring Representations with an Adversary (https://arxiv.org/abs/1511.05897), (b) Learning Adversarially Fair and Transferable Representations (https://arxiv.org/abs/1802.06309), (c) Context-Aware Generative Adversarial Privacy (https://arxiv.org/abs/1710.09549), (d) Learning Generative Adversarial RePresentations (GAP) under Fairness and Censoring Constraints (https://arxiv.org/abs/1910.00411), and many others. Without a clear (empirical) comparison to these works, the benefits of the proposed approach are unclear.\\n\\n4. The proposed method for disentanglement does not scale to non-binary sensitive attributes (because of the blowup in the number of classifier pairs that would need to be trained to ensure disentanglement). Thus, this approach may be limited to cases with a few sensitive attributes. \\n\\n5. The authors are encouraged to show the learned representations (so that one could verify with the naked eye that the representations are indeed disentangling the sensitive attributes).\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper brings to our attention the problem of privacy-preserving representation learning.\", \"the_assumption_in_this_work_is_that_the_data_feature_can_be_split_into_two_parts\": \"private and non private.\\n\\nThe authors propose to use VAE with additional regularizer terms to learn hidden representations that would encode both private and non-private part as independent as possible.\", \"there_are_two_scenarios_considered\": \"1) supervised disentanglement, total of 4 terms (two for public part, two for private part) are added to the VAE loss.\\n2) weakly-supervised disentanglement: in this setting, downstream task is unknown and only two terms are added tot he VAE loss: one to penalize the classifiability using private information in the public z space, and other to promote the classifiably using private information in the private z space.\\n\\nOverall, I think it is a well-stated problem and this work seems to get some positive results on CelebA dataset. However, a few questions must be clearly addressed:\\n1. Eqn 6 is derived from Eqn 2 and Eqn 5. please clean up the notation on L terms. Currently they are not consistent and L_i and L_j are not defined.\\n\\n2. In section 3.3, the authors mention the optimization is nontrivial. Can you expand on that? current section 3.3 is too short for the readers to appreciate that non-triviality.\", \"other_comments\": \"On Page 1, \\ngrowing privacy concerns will entails \\u2014> entail\\nthe \\\"Glass / gender\\\" example doesn\\u2019t make sense to me. why glass is considered non-private while gender is private? maybe we should use another example here.\\n\\nOn Page 2,\\n penultimate paragraph, where it used as \\u2014> where it is used as \\u2026\\nWhile the latter encourage preserving \\u2014> encourages\\nThis two additional terms \\u2014> these two additional terms\\n\\nPage 7, Figure 2 caption needs rewriting.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"PRIVACY-PRESERVING REPRESENTATION LEARNING BY DISENTANGLEMENT\\n\\nSummary\\nThis paper introduces a method to disentanglement the private and public attribute information in representation learning.\", \"strength\": \"1. The idea of introducing the confusion term to disentanglement private and public information seems novel. \\n2. The problem studied in this paper is very important.\", \"comments\": \"1. The existing results are not sufficient to validate the effectiveness of the method to prevent privacy leakage. More comparison with other previous methods should be conducted. For example, the previous work [1] has both theoretically and experimentally validated the effectiveness of DP based methods for preventing attribute attack. Recent work[2] also tries to reduce information leakage in representation learning.\\n2. Some important related works are missing. The difference between the proposed method and previous works with the same purpose should be made more clearly. See comment1 for some concrete examples of previous works. \\n3. The notation of $z$ is confusing. The author denotes $z=(z_{*}, z_{.})$, however, these three vectors' dimensions are the same, which really confuses me.\\n4. Some details of $L_{VAE}$ are missing. According to Equation~1., the author uses $z$ to build the reconstruct loss. However, based on previous notations, $z$ comprises of both private and public representations. So, what's the strategy to combine these two representations (concat?)?\\n5. Typos. For example, blue-> read in the caption of Figure 1. \\n6. The threat model in this paper needs to be made more clearly. The method proposed in this paper can only be used in the context where the private attribute is well-defined. There are many other threat models in secure machine learning research, such as membership attack and adversarial attack, which are not covered by this paper (My suggestion is adding a specific sub-section of threat model and do not use \\\"privacy-preserving\\\" in the original title ). \\n[1] Privacy Risk in Machine Learning: Analyzing the Connection to Over\\ufb01tting\\n[2] Mitigating Information Leakage in Image Representations: A Maximum Entropy Approach\"}", "{\"comment\": \"We thank the reviewer for the interest in our work and we appreciate the very insightful comments.\\n\\ni) Federated Learning: Learning disentangled representations in the federated setting is a challenging endeavor. That is why we pose it as a future research direction in the paper. Extending the proposed method to a federated setup can be achieved on different levels of decentralization. \\nIn the simplest case, the objective is to learn a downstream task in a decentralized fashion on feature representation. This scenario assumes that the disentangled feature representation is learned beforehand via a trusted curator. Then the downstream task is learned decentralized via \\u201cfederated averaging\\u201d [1]. \\nThe more challenging scenario entails not only learning the downstream task in a decentralized fashion but also the disentangled representation. The challenge arises here as the domain confusion classifier w.r.t. sensitive attributes has to be learned locally on each client. Consequently, as the client-level supervision (i.e., IDs) is not accessible in a federated optimization. One possible solution is to divide the model parameters into public and private parameters. \\nConsequently, the public parameters can be trained in a decentralized fashion by federated averaging [1], while the private parameters have to be optimized locally on each client. Eventually, both the aggregated public and local private parameters are used for the downstream task. A similar strategy is proposed in the work of [2]. It should be noted that such a solution based on parameter splitting does not guarantee any sorts of semantic disentanglement between the public and the private parts of the representation. This contrasts with the underlying idea of the proposed approach, where disentanglement is achieved through employing the confusion loss at the client-level.\\n\\n\\nii) VAE reconstruction visualization: The VAE is used as a backbone for learning disentangled representation, with the goal of utilizing the representation for a downstream task. Therefore, the visual aspect of reconstruction itself is of minor importance. However, we agree that the visualizations provide further insights into the disentanglement behavior. Specifically, we noticed that visualizations from private latent and public parts confirm the disentanglement visually. More concretely, reconstructions from the private part of the representation result in a sharpening of identity revealing properties such as skin color, eye style. Simultaneously, these visualizations feature vanishing of public attributes such as sunglasses, jewelry. In contrast to that, reconstructions from the public representation part contain generic facial features, e.g., smiles, teeth, and face outlines. We plan to add visualizations in the final version of the manuscript.\\n\\niii) ID classification: We analyzed the predictive pattern of the ID association. We did not observe any pattern related to, i.e., swapping pairs. We would add further detail on that in the final version of the manuscript to give more insight in this regard.\\n\\nWe thank the reviewer for hinting at the typos. The final version of the manuscript will accommodate these along with suggestions made.\\n\\n[1] \\u201cCommunication-Efficient Learning of Deep Networks from Decentralized Data\\u201d, McMahan et al., AISTAT 2016\\n\\n[2] \\u201cFederated User Representation Learning\\u201d, under review, ICLR 2020, https://openreview.net/forum?id=Syl-_aVtvH\", \"title\": \"Extension to federated setup; Visualization of public/private representation.\"}", "{\"comment\": [\"Strength:\", \"Excellent clarity in presentation & well-written paper.\", \"The idea behind paper is important, especially the learning of publicly shareable representation that doesn't compromise privacy.\", \"Also, I liked the fact that grouping samples by source can easily be used to learn private attributes and can boost overall accuracy. In many classification tasks, this is often overlooked feature of dataset.\"], \"weakness\": [\"The author mentions about federated learning and claims that proposed methods in paper can be extended to those settings. However, if I am not mistaken, the basic premise of federated learning is to have small number of samples on very large number of devices and learning representation from those small samples may not be plausible.\", \"I wish author could have added some visualization of sample images and their reconstructions for VAE and proposed methods.\", \"Also, as desired, ID classification accuracy dropped from 77% to 0, I was wondering if model just swapped ID of one actor with other's creating sort of a pair. It would be really interesting to see, how many images of an actor misclassified to another specific actor.\"], \"typos\": [\"where the first time is the reconstruction loss -> where the first term is the reconstruction loss\", \"the ever increasing performance -> the ever-increasing performance\", \"in the cases where the the downstream -> in the cases where the downstream\", \"this two additional terms -> these two additional terms\", \"where it is not know a priori -> where it is not known a priori\", \"In this experiments -> In this experiment\", \"accuracy drops on the the public -> accuracy drops on the public\", \"task is a priori know -> task is a priori known\", \"in the supervised setup explain in -> in the supervised setup explained in\", \"adding the confusion team leads to -> adding the confusion term leads to\", \"(red and blue) - (red and green)\", \"That his, these approaches -> That is, these approaches\"], \"title\": \"Review\"}" ] }

[ "Zhiyuan Li", "Sanjeev Arora" ]

[ "batch normalization", "weight decay", "learning rate", "deep learning theory" ]

{ "note_id": [ "0lFT19rXkk", "rklU4eINhr", "HygVdTqLir", "r1e-C3cUor", "SJl1vjc8oB", "Hyle1ic8sS", "rJxVktSXsr", "HJxfYraZsB", "HkxH37tRFH", "rJlhm_ETtH" ], "note_type": [ "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752473, 1574359085715, 1573461355523, 1573461193425, 1573460823484, 1573460695594, 1573243099615, 1573143930411, 1571881901034, 1571797027576 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2575/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2575/Authors" ], [ "ICLR.cc/2020/Conference/Paper2575/Authors" ], [ "ICLR.cc/2020/Conference/Paper2575/Authors" ], [ "ICLR.cc/2020/Conference/Paper2575/Authors" ], [ "~Micah_Goldblum1" ], [ "ICLR.cc/2020/Conference/Paper2575/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2575/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2575/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Spotlight)\", \"comment\": \"After the revision, the reviewers agree on acceptance of this paper. Let's do it.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"UPDATE TO MY EARLIER REVIEW\\n===========================\\n\\nThe authors improved the writing of the paper substantially relative to the first version they submitted, and fixed minor issues. I am changing my rating from \\\"Weak accept\\\" to \\\"Strong accept\\\". It is now a must-read paper for anyone doing research on deep learning. \\n\\n\\n\\nMY EARLIER REVIEW\\n=================\\n\\nThis exciting and insightful paper presents theorems (and illustrating examples and experiments) describing an equivalence of commonly used learning rate schedules and weight decay settings with an exponentially increasing learning rate schedule and no weight decay, for neural networks with scale-invariant weights. Hence, the results apply to a large set of commonly employed settings. The paper contains an interesting example of a neural network for which gradient descent converges if with batch normalization as well as with L2 regularization, but not when both are used. \\n\\nFrom a theory viewpoint, the paper offers new insights into Batch normalization and other normalization schemes. From a practical viewpoint, the paper suggests a way to speed up hyper-parameter search, effectively allowing to consider learning rate and weight decay as one parameter.\", \"a_small_gripe\": [\"this paper is a bit rough around the edges, and reads a bit like a draft (see comments on details below).\", \"Detailed Comments / advice / questions\", \"==================================\", \"It often takes a bit of searching to figure out what proof goes with what theorem / fact. I recommend to add to each occurrence of \\u201cProof.\\u201d (before a proof) with a reference to the theorem or fact that is being proved, e.g. \\u201cProof of Theorem 2.6\\u201d.\", \"The authors state that theorem B2 applies to \\u201cgeneral deep nets\\u201d. In this theorem, the limit R_\\\\infty could very well be zero (e.g. for networks with weights that are not scale-invariant), in which case the statement contains a division by zero. I wonder if the authors overlooked this or forgot to state an assumption in the theorem. Maybe I am missing something. Since this theorem does not appear to be critical to the main contributions of the paper, it may be easiest to remove the theorem if the division by zero is indeed a problem.\", \"On page 6, if c_w(x) is independent of w, would c(x) be more suitable?\", \"At the bottom of page 6, the authors state that \\u201cAs a result, for any fixed learning rate, ||w_{t+1}|| ^4= \\u2026\\u201d. It appears that the authors get the expression for ||w_{t+1}||^4 from the expression for ||w_{t+1}||^2 above (maybe by multiplying by ||w_{t+1}||^2?), but I don\\u2019t see how. Could the authors explain this? Maybe authors mixed up w_t and w_{t+1} by accident?\", \"The authors claim to \\u201cconstrict better Exponential learning rate schedules\\u201d. Since the authors only perform a limited evaluation of their proposed learning rate schedule on CIFAR10, I suggest qualifying this statement.\", \"Theorem 1.1 does not introduce what \\\\tilde{\\\\gamma} is. It\\u2019s somewhat obvious, but I would state it nonetheless.\", \"The authors state that \\u201c\\u2026the exponent trumps the effect of initial lr very fast, which serves as another explanation of the standard wisdom that initial lr is unimportant when training with BN\\u201d. I don\\u2019t think that this constitutes a full explanation without further argument. I am also wondering if \\u201canother\\u201d is appropriate here: I am not aware of any (other) mathematically precise explanations of why initial learning rates do not matter in this set-up. If the authors are, they should cite it.\", \"The authors often forgot spaces before \\\\cite commands. If you are trying to avoid a line break before the \\\\cite command, you can use a tilde ~ , like this \\u201cGroup Normalization~\\\\cite{somepaper}\\u201d.\", \"Please introduce the abbreviation LR for \\u201cLearning Rate\\u201d, and always use the all-upper-case version (not \\u201clr\\u201d).\", \"Definition 1.2 has a broken \\\\cite .\", \"Theorem 2.7 should introduce \\\\hat{eta}_t, but it doesn\\u2019t.\", \"Appendix A.1 contains a broken sentence (\\u201cas a function of\\u2026\\u201d)\", \"It\\u2019s odd that the proof for theorems B1 and B2 appear before theorems B1 and B2. I would restructure the appendices to improve this.\", \"Theorem B2 contains a stray \\u201cwhen\\u201d, and \\u201cexists\\u201d should be \\u201cexist\\u201d\", \"What the authors call \\u201cProof of Theorem 2.4\\u201d in the appendix is really a proof of the rigorous version of Theorem 2.4 - Theorem 2.7! The proof should in my opinion be labeled \\u201cProof of Theorem 2.7\\u201d\", \"The typo \\u201ceventsequation\\u201d should be replaced with something like \\u201cthe events from equations\\u201d\", \"Replace the colloquialism \\u201cnets\\u201d with \\u201cnetworks\\u201d.\", \"Replace \\u201cBatchNorm\\u201d with \\u201cBatch Normalization\\u201d\", \"\\u201cCOvariate\\u201d has casing issues\", \"\\u201cRiemmanian\\u201d should be \\u201cRiemannian\\u201d\", \"\\u201cBNhas\\u201d should be \\u201cBN has\\u201d\", \"The paper\\u2019s title states that the results are for batch-normalized networks, while the analysis appears to be more generally for networks with scale-invariant weights, which as the authors point out can arise from mechanisms other than batch normalization. Have the authors considered changing the paper\\u2019s title to better capture what their work applies to? In terms of discoverability, the authors would do the community a service by titling the paper in such a way that it captures the set-up well.\"]}", "{\"title\": \"Response\", \"comment\": \"Thanks for your thoughtful and valuable reviews!\\n\\n-We\\u2019ve improved the writing and fixed the typos in the paper. Regarding your suggestion about convergence analysis for routine smooth convex stochastic optimization problem, we do know how to get such a bound for the case of SGD + momentum + constant LR, which could be seen as a generalization of the convergence result in [Arora, Li, Lyu\\u20192019] where momentum is turned off. The main idea is the same, because of the norm of the weight monotone increases (as shown in this paper), we can use the objective as a potential to bound the sum of the squared norm of gradients along the trajectory. Here the assumption is that the smoothness of scale invariant loss L(w) needs to be bounded outside the unit ball. Due to the space limit, we didn\\u2019t put this result in the paper. But given your suggestion, we will try to put this result in the appendix of the next revision if time permits.\\n\\nHowever, it seems to be impossible to get a convergence rate for the case of SGD + momentum + constant lr + weight decay, as we\\u2019ve already given a non-convergence result for the toy example. If we further allowing step decay LR schedule, then when LR is sufficiently small, the dynamics is like gradient flow and it converges to stationary points of course. \\n\\n-The main point about the connection to other LR schedules is that we can get rid off WD by adopting the exponential version of the original schedule, which could accelerate the hyperparameter search.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for your detailed and thoughtful reviews as well as the helpful suggestions!\\n\\n1. Our new revision tries to improve the writing and fixed the typos you have mentioned. We specified which theorem/lemma each proof corresponds to as you suggested. We also reorganized the sections in appendix.\\n\\n2. You\\u2019re correct on that $R_\\\\infty$ could be 0, thus making the statement meaningless. When $R_\\\\infty= 0$, $D_\\\\infty$ is also equal to 0 since $D_\\\\infty <= 4* R_\\\\infty$. A quick fix is just to replace the original statement by $D_\\\\infty = \\\\frac{2\\\\eta\\\\lambda}{1+\\\\gamma}R_\\\\infty$. \\n\\n3. Random variable $c_w$ is actually a function mapping input data $x$ to a real number. This function indeed depends on $w$, but for different $w$, the random variable $c_w$ follows the same law(distribution). \\nTo avoid confusion, we modified this argument in the new revision and now we used the fact for each w, with constant probability, $\\\\|\\\\nabla_w L\\\\|> \\\\frac{C}{\\\\|w\\\\|}$, where $C$ is some other constant. \\n\\n4. We indeed missed some subscription in the sentence \\u201cAs a result, for any fixed learning rate, $\\\\|w_{t+1}\\\\|^4=$ \\u2026\\u201d. The reason that $\\\\|w_{t+1}\\\\|^4>= \\\\|w_t\\\\|^4+2\\\\eta^2c_w^2$ holds is that we can expand $\\\\|w_{t+1}\\\\|^4$ as$ (\\\\|w_{t+1}\\\\|^2)^2$, which is larger than $(\\\\|w_t\\\\|^2+\\\\eta^2*c_w^2/\\\\|w_t\\\\|^2)^2 = \\\\|w_t\\\\|^4+2\\\\eta^2c_w^2 + \\\\eta^4*c_w^4/\\\\|w_t\\\\|^4$.\\n\\n5. We\\u2019re sorry for the typo in the statement \\u201c\\u2026the exponent trumps the effect of initial LR very fast, which serves as another explanation of the standard wisdom that initial LR is unimportant when training with BN\\u201d. It should be that the initial scale of initialization is unimportant, and we give further explanation for this in the new revision. We don\\u2019t have a theory for this but we verified this intuition in experiments.\\n\\nThank you so much for pointing this out!! And actually, the initial LR does matter a lot in practice. \\n\\n6. As you suggested, we changed the title of the paper into \\u201cAn Exponential Learning Rate Schedule for Deep Learning\\u201d.\\n\\nThanks again for your helpful and thoughtful suggestions!! Given the new revision addressing all your concerns, would you consider raising your rating?\"}", "{\"title\": \"Response\", \"comment\": \"We thank the reviewer for the careful reviews.\\n\\nIn the new revision, we added a new section in appendix explaining the scale invariance in the modern network architectures. We also improved the writing of the proofs and fixed the typos.\\n\\nThanks again for your appreciation.\"}", "{\"title\": \"A new revision is uploaded\", \"comment\": \"Besides improving writing and fixing the typos in the paper, we made the following major changes in the new revision:\\n\\n1. We changed the title to \\u201cAn Exponential Learning Rate Schedule for Deep Learning\\u201d for better discoverability as suggested by Reviewer #3.\\n\\n2. We added a warm-up subsection in Section 2, proving the equivalence between exp LR and weight decay in the momentum-free case. \\n\\n3. We provided a new perspective for understanding the almost equivalence between Step Decay + WD and our TEXP LR schedule, which says that with additional one-time momentum correction per phase, TEXP is **exactly** equivalent to Step Decay+WD. This explanation is conceptually easier than the original one(convergence of the LR growth rate) and thus we put this explanation in the main paper and defer the original one into the appendix. \\n\\n4. We reorganized the experiment section and improved the writing.\\n\\n5. In response to the reviews of Reviewer#2, we added a new section in the appendix named \\u201cScale Invariance in Modern Network Architectures\\u201d explaining why various network architectures are scale invariant, including feedforward CNN and FC nets, ResNet, PreResNet. In detail, we gave an efficient algorithm for checking the sufficient condition of scale invariance and demonstrate how we could apply it onto the aforementioned architectures. The scale invariance of some architectures, like ResNet and PreResNet, is not entirely trivial to show.\\n\\n6. We add a section named in appendix named \\u201cViewing EXP LR Via Canonical Optimization Framework\\u201d, aiming at explaining why the efficacy of exponential LR in deep learning is mysterious to us, at least as viewed in the canonical framework of optimization theory.\"}", "{\"title\": \"An Interesting Connection\", \"comment\": \"Hi Authors,\\nThank you for your interesting paper. I noticed that your work concerning learning rates for neural networks with batch normalization is related to our paper, which shows that an alternative to weight decay, which may stabilize effective learning rate, can improve performance for networks, especially those with batch norm.[1] Please consider mentioning the relationship with our work in your next version.\\n\\n[1] https://arxiv.org/abs/1910.00359\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This work makes an interesting observation that it is possible to use exponentially growing learning rate schedule when training with neural networks with batch normalization. This paper provides both theoretical insights and empirical demonstration of this remarkable property. In detail, the authors prove that for stochastic gradient descent (SGD) with momentum, this exponential learning rate schedule is equivalent to constant learning rate + weight decay, for any scale invariant networks, including networks with Batch Normalization and other normalization methods. This paper also contains an interesting toy example where gd converges when normalization or weight decay is used alone while not when normalization and weight decay are used together.\", \"pros\": \"1. This paper gives new and important insight to the complex interplay between the tricks of network training, such as weight decay, normalization and momentum. The assumption and derivation are simple but the result is quite surprising. In classical optimization framework, it is common to keep the learning rate smaller than the 1/smoothness such that gd decreases the loss. However, the connection between exponential learning rate schedule and weight decay in common practice built by this paper suggests that the current neural net training recipe may be inherently non-smooth.\\n\\n2. The experiment of this paper also suggests that in practice (with normalization layer), learning rate and weight decay coefficient can be packed into a single parameter, which reduces the effort needed for hyper-parameter tuning.\", \"cons\": \"1. Though it's obvious for the feedforward networks with normalization layers to be scale invariant, it's not the case for ResNet ( and the authors use this for experiment). And this needs to be clarified.\\n2. The writing of the proofs should be imporved.\", \"typos\": \"1. Definition 1.2 broke citation\\n2. Equation (1) \\\\eta_t should be \\\\eta_{t-1}\\n3. Some facts about Equation 4, incomplete sentence\\n4 In thm B.2,R_\\\\infty might be 0. So the authors can just delete the last equation on page 12 and use the equation above as the statement of the lemma.\\n5. In the first line of Equation (13), the appearance of ( \\\\beta * e^{1-\\\\beta} )^{ k/2 } seems to be a mistake\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This exciting and insightful paper presents theorems (and illustrating examples and experiments) describing an equivalence of commonly used learning rate schedules and weight decay settings with an exponentially increasing learning rate schedule and no weight decay, for neural networks with scale-invariant weights. Hence, the results apply to a large set of commonly employed settings. The paper contains an interesting example of a neural network for which gradient descent converges if with batch normalization as well as with L2 regularization, but not when both are used.\\n\\nFrom a theory viewpoint, the paper offers new insights into Batch normalization and other normalization schemes. From a practical viewpoint, the paper suggests a way to speed up hyper-parameter search, effectively allowing to consider learning rate and weight decay as one parameter.\", \"a_small_gripe\": [\"this paper is a bit rough around the edges, and reads a bit like a draft (see comments on details below).\", \"Detailed Comments / advice / questions\", \"==================================\", \"It often takes a bit of searching to figure out what proof goes with what theorem / fact. I recommend to add to each occurrence of \\u201cProof.\\u201d (before a proof) with a reference to the theorem or fact that is being proved, e.g. \\u201cProof of Theorem 2.6\\u201d.\", \"The authors state that theorem B2 applies to \\u201cgeneral deep nets\\u201d. In this theorem, the limit R_\\\\infty could very well be zero (e.g. for networks with weights that are not scale-invariant), in which case the statement contains a division by zero. I wonder if the authors overlooked this or forgot to state an assumption in the theorem. Maybe I am missing something. Since this theorem does not appear to be critical to the main contributions of the paper, it may be easiest to remove the theorem if the division by zero is indeed a problem.\", \"On page 6, if c_w(x) is independent of w, would c(x) be more suitable?\", \"At the bottom of page 6, the authors state that \\u201cAs a result, for any fixed learning rate, ||w_{t+1}|| ^4= \\u2026\\u201d. It appears that the authors get the expression for ||w_{t+1}||^4 from the expression for ||w_{t+1}||^2 above (maybe by multiplying by ||w_{t+1}||^2?), but I don\\u2019t see how. Could the authors explain this? Maybe authors mixed up w_t and w_{t+1} by accident?\", \"The authors claim to \\u201cconstrict better Exponential learning rate schedules\\u201d. Since the authors only perform a limited evaluation of their proposed learning rate schedule on CIFAR10, I suggest qualifying this statement.\", \"Theorem 1.1 does not introduce what \\\\tilde{\\\\gamma} is. It\\u2019s somewhat obvious, but I would state it nonetheless.\", \"The authors state that \\u201c\\u2026the exponent trumps the effect of initial lr very fast, which serves as another explanation of the standard wisdom that initial lr is unimportant when training with BN\\u201d. I don\\u2019t think that this constitutes a full explanation without further argument. I am also wondering if \\u201canother\\u201d is appropriate here: I am not aware of any (other) mathematically precise explanations of why initial learning rates do not matter in this set-up. If the authors are, they should cite it.\", \"The authors often forgot spaces before \\\\cite commands. If you are trying to avoid a line break before the \\\\cite command, you can use a tilde ~ , like this \\u201cGroup Normalization~\\\\cite{somepaper}\\u201d.\", \"Please introduce the abbreviation LR for \\u201cLearning Rate\\u201d, and always use the all-upper-case version (not \\u201clr\\u201d).\", \"Definition 1.2 has a broken \\\\cite .\", \"Theorem 2.7 should introduce \\\\hat{eta}_t, but it doesn\\u2019t.\", \"Appendix A.1 contains a broken sentence (\\u201cas a function of\\u2026\\u201d)\", \"It\\u2019s odd that the proof for theorems B1 and B2 appear before theorems B1 and B2. I would restructure the appendices to improve this.\", \"Theorem B2 contains a stray \\u201cwhen\\u201d, and \\u201cexists\\u201d should be \\u201cexist\\u201d\", \"What the authors call \\u201cProof of Theorem 2.4\\u201d in the appendix is really a proof of the rigorous version of Theorem 2.4 - Theorem 2.7! The proof should in my opinion be labeled \\u201cProof of Theorem 2.7\\u201d\", \"The typo \\u201ceventsequation\\u201d should be replaced with something like \\u201cthe events from equations\\u201d\", \"Replace the colloquialism \\u201cnets\\u201d with \\u201cnetworks\\u201d.\", \"Replace \\u201cBatchNorm\\u201d with \\u201cBatch Normalization\\u201d\", \"\\u201cCOvariate\\u201d has casing issues\", \"\\u201cRiemmanian\\u201d should be \\u201cRiemannian\\u201d\", \"\\u201cBNhas\\u201d should be \\u201cBN has\\u201d\", \"The paper\\u2019s title states that the results are for batch-normalized networks, while the analysis appears to be more generally for networks with scale-invariant weights, which as the authors point out can arise from mechanisms other than batch normalization. Have the authors considered changing the paper\\u2019s title to better capture what their work applies to? In terms of discoverability, the authors would do the community a service by titling the paper in such a way that it captures the set-up well.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"%%% Update to the review %%%\\nThanks for your clarification and the revision - the paper looks good! With regards to your comment on accelerating hyper-parameter search, note that there are fairly subtle issues owing to the use of SGD - refer to a recent work of Ge et al \\\"The Step Decay Schedule: A Near Optimal, Geometrically Decaying Learning Rate Procedure For Least Squares\\\" (2019). \\n%%%\\n\\nThe paper makes an interesting observation connecting the use of weight decay + normalization to training the same network (without regularization) using an exponentially increasing learning rate schedule, under an assumption of scale invariance that is satisfied by normalization techniques including batch norm, layer norm and other variants. An example is also provided where the joint use of batch norm and weight decay can lead to non-convergence to a global minimum whereas the use of one (without the other) converges to a global minimum - serving to indicate various interdependencies between the hyper-parameters that one needs to be careful about.\\n\\nWhile the connection of scale invariant models to novel schemes of learning rates is interesting (and novel), the paper will benefit quite a bit in its contributions through attempting a convergence analysis towards a stationary point even for solving a routine smooth non-convex stochastic optimization problem. Owing to the equivalence described in the paper, this enables us to understand the behavior of the combination of batch norm (or some scale invariance property) + weight decay + momentum (+ step decay of the learning rate), which, to my knowledge isn\\u2019t present in the literature of non-convex optimization. \\n\\nThe paper is reasonably written (the proof of the main claim is fairly easy to follow), but needs to be carefully read through because I see typos and ill-formed sentences that should be rectified - e.g. see point 3. in appendix A.1 - some facts about equation 4, missing citation in definition 1.2 amongst others. I went over the proof of the main result and this appears to be correct. Furthermore, I find the connections to other learning rates (such as the cosine learning rate) to be rather hard to understand/interpret, in the current shape of the paper.\"}" ] }

[ "Ronan Fablet", "Lucas Drumetz", "François Rousseau" ]

[ "end-to-end-learning", "irregularly-sampled data", "energy representations", "optimal interpolation" ]

{ "note_id": [ "4lVMROrsN7", "B1lNNWYdcr", "Byetnaz8cr", "rJx-I2HJ9S" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752445, 1572536619841, 1572380080643, 1571933257501 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2574/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2574/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2574/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This work looks at ways to fill in incomplete data, through two different energy terms.\\nReviewers find the work interesting, however it is very poorly written and nowhere near ready for publication. This comes on top of poorly stated motivation and insufficient comparison to prior work.\\nAuthors have chosen not to answer the reviewers' comments.\\nWe recommend rejection.\", \"title\": \"Paper Decision\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper discusses various approaches to predict missing values in the input (filling/inpainting task).\\nThey define an energy function equal to the squared L2 distance between the input and its reconstruction by various kinds of neural nets. They show slightly better performance compared to a PCA-based baseline.\\n\\nPositive things about this work\\n- the topic is interesting\\n- the last application is interesting (water temperature prediction) \\n\\nNegative things about this work\\n- this work is very poorly written and lacks sufficient clarity. This work needs a major rewrite. I do not even know where to start, but to give an example:\", \"1st_sentence__of_the_abstract_reads\": \"\\u201cFor numerous domains, including for instance earth observation, medical imaging, astrophysics,..., available image and signal datasets often irregular spacetime sampling patterns and large missing data rates.\\u201d, a sentence that misses the verb.\", \"2nd_sentence_of_the_abstract_reads\": \"\\\"These sampling properties is\\u201d which is not grammatically correct\\nAnd so on so forth. Speaking of which, the Authors make excessive use of \\\"....\\\".\\n- because of the lack of clarity throughout the paper, I have had hard time figuring out what exactly the authors do. From the limited understanding I got after reading this draft twice, I think they consider a few different variants of auto-encoder and a \\\"log-prior\\\"/energy function equal to the squared L2 distance between input and its reconstruction. However, this formulation has barely any novelty. For instance, see old work like:\\nS. Roth and M. J. Black, \\u201cFields of experts: A framework for learning image priors,\\u201d in Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol. 2, San Diego, California, June 2005, pp. 860\\u2013867\\nwhere they used a different log-prior, but essentially the same optimization process. If the novelty is the use of auto-encoders, then the comparison should be against methods that do not auto-encode (like the above or more modern versions of it).\\n- several choices made by the Authors seem not well motivated. For instance, it's written that computing gradients is too expensive and therefore these are replaced by the G network. However, doesn't the G network also need gradients to be updated?\\n- the terminology is not standard and confusing. Hidden state usually refers to the encoder output, not to the decoder output.\\n- the Authors never introduce the metrics they use, reconstruction and interpolation scores.\\n- in general, the motivation is unclear. Why is it a problem if the data does not come from a grid? In vision, inpainting on unconstrained masks has been standard for decades. \\nMore recently, transformer architectures and GNNs seem quite good at representing sets and graph structured data.\\nSo considering this context, the current motivation provided by the authors needs some refinement. In fact, the authors could consider building on top of these other more modern approaches.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper proposes a framework to learn representations of signals when they the signals are irregularly sampled. They propose to do this by using some modified iteration steps from DINEOF algorithm. In addition to this, they propose a new energy function which is inspired by the idea of Gibbs distribution. The parametrize the energy function by some convolutional filters with a constraint. As claimed by the paper in the introduction, they only use the under sampled signals to learn as opposed to using the fully sampled ground truth signals in the previous deep learning basic approaches.\\n\\nI choose to reject this paper. 1) The paper doesn't justify what these representations are or where these representations will be used 2) There is no theoretical or extensive empirical motivation that the representation that is coming out of this network is actually useful for downstream tasks of interest. 3) Ill-defined evaluation metrics 4) Missing comparisons with basic models - why not compare with in-painting models for images and time series in experiments? 5) The paper is hard to read, not well organized and missing a lot of details.\", \"main_argument\": \"It is not clear to me the motivation for learning these representations or how the usefulness of this representation will be evaluated. The only place where I see a use for representation other than reconstruction of the signal is in Table 1. They use a 3 layered MLP on top of the representation from auto encoder. Why is this interesting? If you're trying to claim that the representation is powerful enough, shouldn't you be able to do the classification with just a linear classifier? Can you justify the usefulness of representation for some downstream tasks in your other experiments? \\n\\nIn any case, I don't understand why the authors don't compare with a simple auto-encoder. The authors assumes that they know where the signal is sampled and where is is not. Why not train an auto-encoder with MSE loss, but calculating MSE over only those pixels which the authors know are sampled. Comparison with such a model will be helpful in understanding if the framework is adding anything of value.\\n\\nAuthors useful I-Score, R-Score, AE-Score and C-Score to compare across models. These metrics are not defined in the main text (or any reference) but there is a loose (english) definition in the caption of Table 1. For example, R-Score, is the \\\"reconstruction performance of known image areas\\\". Why are these numbers in percentages? Shouldn't you be using MSE, PSNR or SSIM? Why are some of scores negative?\\n\\nIf you are comparing how well you're interpolating for natural images or time series, you should compare with standard interpolation techniques in image processing and time series analysis. \\n\\nThe paper have a lot of missing details, including simple things like not linking to a Figure or Section the authors are referring to. Some examples are:\\n1. Authors motivate using Gibbs energy by saying that in auto encoders you have to project to a lower dimensional representation and gibbs energy based solution has no such dimensionality reduction constraint. There are no such constraint in AE as well - you can vey well have your hidden representation to be over complete - so this is not a good motivation for Gibbs energy\\n2. Main section of 4.1 says that they use MNIST but the caption of Table 1 says they use Fashion MNIST\\n3. The naming convention of the model is hard to follow and the authors notation itself is inconsistent.\", \"some_other_problems\": \"1. In Abstract, they say 2. images. Why 2?\\n2. Introduction, the citation for Gaussian priors is empty\\n3. Last word before beginning of section to, broken link\\n4. In the second para of 3.2 did you mean to say solve for (2) instead of (3)?\\n5. What is \\\"XXX\\\" matrix in second para of 3.2?\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes an end-to-end learning framework for interpolation problems, motivated by problems such as irregularly-sampled images or time-series.\\n\\nIt was not clear after reading the paper where the key novelty of the proposal lies. The energy formulations for U\\u03b8, namely an autoencoder and Gibbs model, are not to my knowledge new in this context. I gather the novelty must then be in Section 3.2, which uses a neural-network based interpolation scheme. This builds on an existing update scheme (Equation 8), but replaces the full operator \\u03c8 with an iterative update. The key idea is then to replace the full gradient for the energy function with a learned function, based on a CNN. While this is not my area of expertise, I am not sure as to technical significance or novelty of such a proposal.\\n\\nThe paper emphasises their view of an interpolation operator I, which is a little confusing in an autoencoder context. Typically, we pick \\\\hat{X} to be the solution attaining minimal squared error compared to the reference Y. It was not clear why one needs a further invocation of a minimisation between Y^i and I(U\\u03b8, Y^i, \\u03a9^i).\\n\\nThe writing could be improved, as there are a number of grammatical issues, as well as missing section or equation #'s. In terms of presentation, the Introduction delves far too soon into a detailed discussion of related work in the area. I would suggest instead to provide a crisp high-level overview of the limitations of existing work, and how they are overcome in the present paper.\"}" ] }

[ "Nitin Rathi", "Gopalakrishnan Srinivasan", "Priyadarshini Panda", "Kaushik Roy" ]

[ "spiking neural networks", "ann-snn conversion", "spike-based backpropagation", "imagenet" ]

{ "note_id": [ "i1zKlyyBEw", "BkgpDWL5sB", "H1xRH-89sr", "B1gEUxI5iB", "Hke4VJI5iS", "SklebdhWoB", "Bye9F-4CFS", "S1gQnUOTFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752414, 1573704037146, 1573704006066, 1573703756147, 1573703467705, 1573140472444, 1571860865906, 1571813034690 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2573/Authors" ], [ "ICLR.cc/2020/Conference/Paper2573/Authors" ], [ "ICLR.cc/2020/Conference/Paper2573/Authors" ], [ "ICLR.cc/2020/Conference/Paper2573/Authors" ], [ "ICLR.cc/2020/Conference/Paper2573/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2573/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2573/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"After the rebuttal, all reviewers rated this paper as a weak accept.\\nThe reviewer leaning towards rejection was satisfied with the author response and ended up raising their rating to a weak accept. The AC recommends acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thank you for your remarks. You have summarized our paper very nicely and we appreciate your time and effort for reviewing our work. Please let us know if you have any suggestions that can help to increase the score of the paper. We have added Appendix A and B to compare different pseudo-derivatives and the training effort for training an SNN from scratch compared to hybrid conversion-and-STDB training.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"We thank the reviewer for their time and effort to review our work. Please find below the explanation for individual comments.\\n\\n1. Reviewer's comment: This paper proposes methods to initialize and train spiking NNs (SNNs) as an alternative to ANNs, not driven primarily by improved loss or generalization, but by energy efficiency improvements derived from timing-event based sparse operation rather than asynchronous sweeps. The backpropagation method introduced is sensible, as are the experiments on known datasets to show its effectiveness.\\nThe paper is well written (apart from the miniscule Figure 3 containing the main result).\\n\\nAuthor's response: Thank you for pointing this out. We have edited Figure 3.\\n\\n2. Reviewer's comment: I recommend acceptance, with caveats: the energy performance is actually not directly calculated, but speculatively estimated, it depends on the computational architecture chosen to implement the respective networks. I point out that ANNs need to be trained first to properly initialize an SNN, so the relative training effort claimed is less impressive, but energy performance does count in actual operational practice - training is (or should) be a small fraction of that.\\n\\nAuthor's response: We absolutely agree with the reviewer that the energy estimate is dependent on the network architecture as the number of spikes will vary with architecture. We compute the efficiency of the network in terms of latency (number of time-steps) and the number of spikes per image during inference. We have added Appendix B to compare the training and testing effort for both ANN and SNN.\"}", "{\"title\": \"Response to Review #4 Part 2/2\", \"comment\": \"4. Reviewer's comment: \\\"The authors argue for SOTA performance in Table 2, but the comparison to other work doesn\\u2019t clearly separate their performance from the other listed works; For example the accuracy gain against Lee et al.,2019 only comes from the architecture being VGG16 as opposed to VGG9, as can be seen from comparing with the VGG9 architecture from Table 1, furthermore they take the same amount of timesteps, which is supposed to be the principle gain of this work\\\".\\n\\nAuthor's response: Thank you for mentioning this. In Table 2, we compare the best results (over all architectures) reported in various works and compare it with our best results for same datasets. As correctly pointed out by the reviewer, the hybrid approach performs at par (for similar architecture) with the results reported by Lee et al. (2019). Lee et al. (2019) employed a backpropagation mechanism to train the SNN from scratch. As we show in Appendix B, the effort for training SNN from scratch is 10X more compared to hybrid training. Therefore, for most practical purposes it is not possible to train deeper SNNs from scratch. In our approach, we start with a network that performs well with large number of time-steps and train the network with spike-based BPTT to reduce the number of inference time-steps. The entire goal of training in spiking regime is to reduce the number of time-steps and therefore it is very natural to achieve similar performance if the same network is trained from scratch with same number of time-steps. The benefit we achieve with hybrid training is convergence within few epochs because we start with a good initialization and therefore this technique can scale to deeper networks and larger datasets.\\n\\n5. Reviewer's comment: \\\"Some small suggestions that are independent from the above: 1. The most similar or relevant version of equation (2) in previous work could be referenced nearby for context\\\". \\n\\nAuthor's response: We have added a reference for the iterative modeling of the LIF neuron in Section 2.1\\n\\n6. Reviewer's comment: \\\"The last sentence of the first paragraph on p.4 \\u201cthe outputs from each copy...\\u201d is confusing. Are you just meaning to describe BPTT?\\\" \\n\\nAuthor's response: We have edited the sentence to be more concise and clearer. Yes, we are referring to the BPTT mechanism.\\n\\n7. Reviewer's comment: \\\"Typos: sec7 4th line \\u201cneruons\\u201d, sec 2.2 \\u201cboth the credit\\u201d (remove \\u201cthe\\u201d)\\\"\\n\\nAuthor's response: Thank you for pointing these out. We have made the corrections in the paper.\"}", "{\"title\": \"Response to Review #4 Part 1/2\", \"comment\": \"We thank the reviewer for the detailed comments. Please find below the explanation for individual comments.\\n\\n1. Reviewer's comment: \\\"This paper examines combining two approaches of obtaining a trained spiking neural network (SNN). The first approach of previous work is converting the weights of a trained artificial neural network (ANN) with a given architecture, to the weights and thresholds of a SNN, and the second approach uses a surrogate gradient to train an SNN with backpropagation. The true novelty of the paper seems to be in showing that combining the two approaches sequentially, trains a SNN that requires fewer timesteps to determine an output which achieves state of the art performance. This is summarized by Table 1.\\\"\\n\\nAuthor's response: The proposed hybrid technique trains an SNN that can process the input information in fewer time-steps compared to purely ANN-SNN conversion methods as well as with this technique the training effort in the spiking domain is reduced. Backpropagation training with spikes is expensive because multiple iterations are performed in the forward pass and it requires larger memory to store all the activations for backpropagation. The training effort for ANN and SNN is compared in Appendix B of the revised manuscript. The high training effort for SNN hindered its application for large and complex datasets like ImageNet. The hybrid approach makes it possible to train SNNs for large datasets with a reasonable amount of time and memory by having a good initialization and only fine-tuning in spiking domain for few epochs to reduce the number of time-steps. \\n\\n2. Reviewer's comment: \\\"However, it does not mention how many epochs it takes to train an SNN from scratch, nor compare this to the total training time (ANN training + SNN fine-tuning) of their approach\\\".\\n\\nAuthor's response: Thank you for pointing this out. We performed simulations to quantify the effort of training SNN from scratch compared to the hybrid training technique with similar hardware constraints. The results are shown in Appendix B of the revised manuscript.\\n\\n3. Reviewer's comment: \\\"They also claim a novel spike-time based surrogate gradient function (eq. 11), but it is very practically similar to the ones explored in the referenced Wu. et al 2018 (eq. 27 for instance), and these should be properly contrasted showing that this novel surrogate function is actually helpful (the performance/energy efficiency might only come from the hybrid approach)\\\".\\n\\nAuthor's response: Thank you for pointing this out. Wu et al. (2018) mentioned several approximations for the derivative of the spike function. Similar approximations have been proposed in other works as well (Bellec et al., 2018; Zenke & Ganguli 2018; Shrestha & Orchard, 2018). These approximations are either a linear or exponential function of (u-Vt) where u is the membrane potential and Vt the threshold voltage. In our work, we propose to replace the quantity (u- Vt) with \\u0394t, where \\u0394t is the time from the last spike. \\u0394t and (u- Vt) have similar behavior, i.e., their value is small close to the spike event and the value increases as we move away from the spike event. Therefore, \\u0394t works as a good replacement for (u- Vt) and it may lead to energy/computation benefits. All possible values of \\u0394t is bounded and known once the number of time-steps is fixed (Equation 12), therefore we can pre-compute the gradient values and store it in a look-up table for faster access during backpropagation. On the other hand, (u- Vt) is a dynamically changing real number and the gradient must be computed during the simulation. The gradient is computed at every time-step for BPTT and the look-up table may provide faster access compared to computing the gradient based on (u- Vt). The exact benefit in energy is dependent on the overall system architecture and evaluating it is beyond the scope of this paper. We tested our hybrid technique with other approximations based on (u- Vt) and achieved similar performance. The proposed hybrid technique is a general methodology for training deep SNNs and any existing spike-based backpropagation mechanism can be used in this technique. We have added Appendix A to compare various approximations of the gradient of spike function.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper examines combining two approaches of obtaining a trained spikingneural network (SNN). The first approach of previous work is converting the weights of a trained artificial neural network (ANN) with a given architecture, to the weights and thresholds of a SNN, and the second approach uses a surrogate gradient to train an SNN with backpropagation. The true novelty of the paper seems to be in showing that combining the two approaches sequentially, trains a SNN that requiresfewer timesteps to determine an output which achieves state of the art performance. This is summarized by Table 1. However, it does not mention how many epochs it takes to train an SNN from scratch, nor compare this to the total training time (ANN training + SNN fine-tuning) of their approach. They also claim a novel spike-time based surrogate gradient function (eq. 11), but it is very practicallysimilar to the ones explored in the referenced Wu. et al 2018 (eq. 27 for instance), and these should be properly contrasted showing that this novel surrogate function is actually helpful (the performance/energy efficiency might only come from the hybrid approach). The authors argue for SOTA performance in Table 2, but the comparison to other work doesn\\u2019t clearly separate their performance from the otherlisted works; For example the accuracy gain against Lee et al.,2019 only comes from the architecture being VGG16 as opposed to VGG9, as can be seen from comparing with the VGG9 architecture from Table 1, furthermore they take the sameamount of timesteps, which is supposed to be the principle gain of this work.\", \"some_small_suggestions_that_are_independent_from_the_above\": \"1.The most similar or relevant version of equation (2) in previous work could be referenced nearby for context.\\n\\n2.The last sentence of the first paragraph on p.4 \\u201cthe outputs from each copy...\\u201d is confusing. Are you just meaning to describe BPTT? \\n\\n3.Typos: sec7 4th line \\u201cneruons\\u201d, sec 2.2 \\u201cboth the credit\\u201d (remove \\u201cthe\\u201d)\\n\\n---------------\\nFollowing the author response I have upgraded my rating.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes methods to initialize and train spiking NNs (SNNs) as an alternative to ANNs, not driven primarily by improved loss or generalization, but by energy efficiency improvements derived from timing-event based sparse operation rather than asynchronous sweeps. The backpropagation method introduced is sensible, as are the experiments on known datasets to show its effectiveness. The paper is well written (apart from the miniscule Figure 3 containing the main result).\\nI recommend acceptance, with caveats: the energy performance is actually not directly calculated, but speculatively estimated, it depends on the computational architecture chosen to implement the respective networks. I point out that ANNs need to be trained first to properly initialize an SNN, so the relative training effort claimed is less impressive, but energy performance does count in actual operational practice - training is (or should) be a small fraction of that.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper presents a fine-tuning method of models converted from standard encoding and SGD training to Spike/NN's.\\nThe key point of the paper is that directly training S/NN's with spike-back-prop is slow and inefficient, while directly inferencing with converted models is also inefficient due to the large integration window required to get a good estimate of the spiking neuron potential. The authors claim, and to a good extent show that, their proposed method is best of both worlds: train the models efficiently with standard encoding / SGD, this is something we know works and scale well, then convert and fine-tune with spike-backprop to get models that perform well under a shorter integration window, and thus are more efficient at inference time. The intuition is that models can achieve shorter integration windows while keeping good results because, under the assumptions made by the proposed algorithm, the fine-tuning is effectively unrolling neuron dynamics that can be trained with BPPT, in a way similar to LSTM/Recurrent models. In that case, since model dynamics are taken into account during fine-tuning, it results in better performance even under shorter time-windows. This is an interesting concept, since the training doesn't only consider a mean-field estimate of the spike-activation, but it looks at spiking neuron dynamics with an higher granularity. The paper is well written, clear and easy to understand. Results are comparatively competitive and code is made available.\"}" ] }

[ "Sanghyun Hong", "Michael Davinroy", "Yiǧitcan Kaya", "Dana Dachman-Soled", "Tudor Dumitraş" ]

[ "Reconstructing Novel Deep Learning Systems" ]

{ "note_id": [ "K1gHRlmgml", "Hygvy9O3ir", "B1ekcK5tir", "HJlPSt5YoH", "H1xcou5YoS", "HygKud9KoS", "SkxmSd5KiS", "rJgGcD4RtS", "SkgYKSmAtS", "rkxi9slCKr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752386, 1573845470690, 1573656967104, 1573656894726, 1573656737688, 1573656689290, 1573656635035, 1571862409595, 1571857792948, 1571847058615 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2572/Authors" ], [ "ICLR.cc/2020/Conference/Paper2572/Authors" ], [ "ICLR.cc/2020/Conference/Paper2572/Authors" ], [ "ICLR.cc/2020/Conference/Paper2572/Authors" ], [ "ICLR.cc/2020/Conference/Paper2572/Authors" ], [ "ICLR.cc/2020/Conference/Paper2572/Authors" ], [ "ICLR.cc/2020/Conference/Paper2572/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2572/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2572/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper proposes using Flush+Reload to infer the deep network architecture of another program, when the two programs are running on the same machine (as in cloud computing or similar).\\n\\nThere is some disagreement about this paper; the approach is thoughtful and well executed, but one reviewer had concerns about its applicability and realism. Upon reading the author's rebuttal I believe these to be largely addressed, or at least as realistically as one can in a single paper. Therefore I recommend acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Summary of Our Responses and Changes to the Manuscript\", \"comment\": \"We thank our reviewers again for taking the time to read, evaluate our work, and provide constructive feedback. We have uploaded a revised version of our paper, with edits to address the concerns raised. Here, we summarize our responses and updates below:\\n\\n[Reviewer 1]\\nQ1. We provide our answer to the concerns about the motivation (in our comments).\\n\\nQ2. We clarify our attack can work with GPUs (in our comments) and highlight this in Sec 3, with the detailed discussion in Appendix A (of our revised paper).\\n\\nQ3. We clarify our attacker and the victim use the same framework version (in our comments) and clearly state this assumption in Sec 3.1 (of our revised paper). \\n\\nQ4. We provide the answers to the questions/concerns about our experiments (in our comments).\\n\\n[Reviewer 2]\\nQ1. We clarify the knowledge of our attacker in reconstruction (in our comments) and include this discussion in the first paragraph of Sec 4 (of our revised paper).\\n\\nQ2. We discuss the potential defense mechanisms (in our comments) and include this discussion in Sec 5 (of our revised paper).\\n\\nQ3. We explain the reason we choose ProxylessNAS-CPU for evaluation (in our comments)\\n\\n[Reviewer 3]\\nQ1. We provide our attacker\\u2019s gain in terms of time and resources compared to running NAS from scratch (in our comments) and include this information in Sec 4 (of our revised paper).\\n\\nQ2/3. We provide a discussion about the limitations of our attack (in our comments).\\n\\nPlease see our replies to each reviewer for our detailed responses to individual points.\"}", "{\"title\": \"Clarification Regarding The Threat Model, Applicability, and Experiments (cont'd)\", \"comment\": \"[Continued discussion about the reviewer's questions in the previous comment]\\n\\n2) We also assume that the time taken for a tensor multiplication will depend on the size of tensor operands, and the computations will be performed in sequential order. If the computation time of an operation is not proportional to the size of operands, or the operations are performed out-of-order, the reconstructed architecture and its parameters will not match the victim\\u2019s architecture. However, we believe no common deep learning framework breaks these properties, and implementing them would likely be computationally inefficient.\\n\\n3) A network with an excessive number of branches can explode the number of architecture candidates and increase the time taken for the reconstruction significantly. However, this type of architecture does not make our reconstruction attack impossible, and having this property is not common due to the significant addition to the computational overhead introduced.\\n\\n[Question about Using Custom Layers]\\n\\nUnless the custom layers are implemented by using a non-shared library, our attacker can monitor the function invocations that make up the custom layer and reconstruct the correct architecture. For instance, the authors of [10] implemented \\u201cSwish\\u201d activation in PyTorch by utilizing existing tensor operations, i.e., Swish(x) := x * Sigmoid(x). Since our attacker can monitor the multiplication (*) and Sigmoid function, the reconstruction algorithm can recover the swish activation as a sequence of a sigmoid function and a tensor multiplication.\\n\\n[References]\\n[1] So et. al, The Evolved Transformer, ICML\\u201919\\n[2] Zoph et. al, Learning Transferable Architectures for Scalable Image Recognition, CVPR\\u201918\\n[3] Pham et. al, Efficient Neural Architecture Search via Parameter Sharing, ICML\\u201918\\n[4] Tan et. al, MnasNet: Platform-Aware Neural Architecture Search for Mobile, CVPR\\u201919\\n[5] Cai et. al, ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware, ICLR\\u201919\\n[6] 6th ICML Workshop on Automated Machine Learning, ICML\\u201919: https://sites.google.com/view/automl2019icml/\\n[7] Customization of operations in TensorFlow: https://www.tensorflow.org/guide/create_op\\n[8] Amazon SageMaker: https://aws.amazon.com/sagemaker/\\n[9] Google AutoML: https://cloud.google.com/automl/\\n[10] Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. \\\"Swish: a Self-Gated Activation Function.\\\" arXiv preprint arXiv:1710.05941 7 (2017).\"}", "{\"title\": \"Clarification Regarding The Threat Model, Applicability, and Experiments\", \"comment\": \"We thank the reviewer for the constructive feedback. Here, we provide answers to your questions and concerns, and we updated our paper for clarification.\\n\\n(1) Concerns about the Motivations of Our Paper\\n\\nSeveral custom architectures have outperformed the standard architectures, such as VGGs, ResNets, InceptionNets, or Transformer. For example, the Evolved Transformer [1] has achieved the state-of-the-art BLEU score of 29.8 on WMT\\u201914 English-German translation. Also, the same architecture shows the consistent improvement over the previous architectures in four well-established language benchmarks\\u2014i.e., the novel architecture performs better on multiple datasets. In consequence, neural architecture search (NAS) is an active research topic in the deep learning community [6], focusing on automating the process of inventing near-optimal neural network architectures [1, 2, 3, 4, 5]. The substantial computational resources required for running NAS algorithms provide an incentive for companies and individuals to keep the networks secret to obtain a competitive advantage and therefore for attackers to steal the result of the search.\\n\\n(2) Applicability to GPUs\\n\\nOur attack is not fundamentally different for GPUs. In most deep learning frameworks, when a network performs a computation, it invokes the same function implemented in C++ and the function decides whether the back-end computation can use GPUs or not. This practice [7] maximizes the hardware compatibility of a framework; however, this also makes the framework vulnerable to our attacker who can still observe the common functions listed in Table 3 by monitoring the shared cache. On GPUs the timings would be different, so we would have to profile the computational times, e.g., the time taken for the matrix multiplication with various sizes of tensor operands. However, on both CPUs and GPUs, the computation time is proportional to the size of tensor operands, which enables our attacker to estimate the architecture parameters with timing observations. We include this information in Sec 3.1 and Appendix A of our revised paper.\\n\\n(3) Applicability: The Attacker and Victim Uses the Different Versions of a Framework \\n\\nIndeed, we implicitly assume that the attacker and victim use the same framework. We believe this is a reasonable assumption since in Machine-Learning as-a-Service scenarios (e.g., Amazon SageMaker [8] or Google\\u2019s AutoML [9]), cloud providers enforce anyone who uses these services to have the same framework version. Cloud providers want to increase the compatibility so that any network and model implemented by users (clients) can run on their environment without issues. Also, when the attacker and victim use the same framework, they are likely to use the same backend matrix multiplication library. For instance, TensorFlow uses MKL-DNN (i.e., Intel\\u2019s linear algebra library) v0.18 from Mar. 4th 2019 up to now while TensorFlow has been updated from v1.12.2 to v2.0.0. We also include this information in Sec 3.1.\\n\\n(4) Questions about Our Experiments\\n\\n[Question about Reconstructing N Random Architectures]\\n\\nWe agree with the reviewer that generating N random architectures would be a valid experiment to perform. However, we would like to note that, even with the small search space on CIFAR10 that ENAS uses, the number of candidate networks (N) that we need to consider is 6^L times 2^(L(L\\u22121)/2) for the network\\u2019s length (L). If we choose L=12, the total candidates for the reconstruction are over 1.6 times 10^29. Thus, even if we choose a reasonable length (L) of the victim network, reconstructing all of them would be too computationally difficult to perform.\\n\\n[Questions about The Attack Failure Cases]\\n\\nThis reconstruction attack relies on several assumptions; thus, our attack successes when the assumptions are met, otherwise, the attacker fails. Here, we listed the circumstances when our assumptions break.\\n\\n1) We assume that the victim architecture is implemented using the computations supported by an open-source deep learning framework. This includes that any layers a victim may create are based on these supported computations. However, if a victim uses a closed-source framework (i.e. a framework developed by and/or only available to them), it disables our attacker from monitoring computations via Flush+Reload. We believe, due to the popularity of open-source deep learning frameworks and the difficulty in developing and maintaining them, this is a rare scenario.\"}", "{\"title\": \"Clarifications Regarding The Attacker\\u2019s Gain and The Limitations\", \"comment\": \"We thank the reviewer for their constructive feedback. Here, we provide answers to your questions and concerns, and we updated our paper for clarification.\\n\\n(1) The Gain of the Attacker in terms of the Time and Resources\\n\\nOur reconstruction algorithm saves a significant amount of time and resources over running NAS search algorithms because the attack does not involve training of candidate architectures\\u2014i.e., the attacker only requires a CPU to run the reconstruction process, whereas running NAS demands multiple GPUs, and does the reconstruction in a shorter time. In our experiments, we reconstructed MalConv in a few minutes and ProxylessNAS-CPU in 12 hours on a CPU. Compared to the victim\\u2019s search time for one architecture, e.g., Proxyless-G (mobile) that takes approximately 40k GPU hours*, our attacker can steal the same architecture by only investing 0.003% of this time. We acknowledge the reviewer pointed out our attacker\\u2019s strength, and we incorporated this information in Sec 4 in our revised paper.\\n\\n(*) ProxylessNAS starts its searching process from a backbone architecture such as NASNet; thus, even if the paper reported a search took 200 GPU hours, this number does not include the time spent searching a backbone architecture, i.e., the 40k GPU hours to find NASNet.\\n\\n(2)/(3) The Limitations of Our Attack.\\n\\nOur attacker can reconstruct the architectures when they are implemented by using the computations (i.e., tensor operations and layers) in an open-source deep learning framework. For instance, EvolvedTransformer [1] is implemented in TensorFlow and utilizes the conventional layers and tensor operations; thus, our attacker can monitor all the computations via Flush+Reload. Additionally, an example GCN in PyTorch* uses the sequence of tensor multiplications and a bias addition to implement a new GraphConvolution layer. Therefore, we believe our attacker will be able to observe the sequence of these computations (e.g., torch.nn, torch.spmm, and a \\u2018+\\u2019 operation) via Flush+Reload and reconstruct networks using this layer.\\n\\n(*) https://github.com/tkipf/pygcn/blob/master/pygcn/layers.py#L9 (line 31)\\n\\nOur reconstruction attack relies on several assumptions, and we agree that our algorithm may not always find an exact match (0% error) quickly if these assumptions are not met. The limitations are as follows:\\n\\n1) We assume that the victim architecture is implemented using the computations supported by an open-source deep learning framework. This includes that any layers a victim may create are based on these supported computations. However, if a victim uses a closed-source framework (i.e. a framework developed by and/or only available to them), it disables our attacker from monitoring computations via Flush+Reload. We believe, due to the popularity of open-source deep learning frameworks and the difficulty in developing and maintaining them, this is a rare scenario.\\n\\n2) We also assume that the time taken for a tensor multiplication will depend on the size of tensor operands, and the computations will be performed in sequential order. If the computation time of an operation is not proportional to the size of operands, or the operations are performed out-of-order, the reconstructed architecture and its parameters will not match the victim\\u2019s architecture. However, we believe no common deep learning framework breaks these properties, and implementing them would likely be computationally inefficient.\\n\\n3) A network with an excessive number of branches can explode the number of architecture candidates and increase the time taken for the reconstruction significantly. However, this type of architecture does not make our reconstruction attack impossible, and having this property is not common due to the significant addition to the computational overhead introduced.\\n\\n[References]\\n[1] The Evolved Transformer, ICML\\u201919\"}", "{\"title\": \"Clarifications Regarding Our Threat Model and The Potential Defenses (cont'd)\", \"comment\": \"[Continued discussion about the potential defenses in the previous comment]\\n\\n4) Moreover, we can run separate networks in parallel on the same physical host. The network obfuscates what our attacker will observe via Flush+Reload. In this case, our attacker may not be possible to reconstruct the victim architecture by monitoring a single query. However, if our attacker can observe multiple queries, the attacker can use the frequent sequence mining (FSM)\\u2014that we used in the block identification\\u2014to identify repeated components over the observations and can reconstruct the victim architecture.\\n\\n(3) Our Choice of ProxylessNAS-CPU Architecture\\n\\nWe selected ProxylessNAS because the Proxyless (CPU) architecture provides the state-of-the-art top-1 accuracy on the ImageNet classification task, at the time of our experiments. We expect to obtain similar reconstruction results for MNAS or ENAS since they use the same tensor operations and layers in TensorFlow or PyTorch to express and run their architecture.\\n\\n[1] Yan et. al, Cache Telepathy: Leveraging Shared Resource Attacks to Learn DNN Architecture, ArXiv\\u201918\\n[2] Duddu et. al, Stealing Neural Networks via Timing Side Channels, ArXiv\\u201918\\n[3] Kim et. al, STEALTHMEM: System-Level Protection Against Cache-Based Side-Channel Attacks in the Cloud, USENIX\\u201912\\n[4] Liu et. al, Catalyst: Defeating Last-level Cache Side-channel Attacks in Cloud Computing, HPCA\\u201916\"}", "{\"title\": \"Clarifications Regarding Our Threat Model and The Potential Defenses\", \"comment\": \"We thank the reviewer for the constructive feedback. Here, we provide answers to your questions and concerns, and we updated our paper for clarification.\\n\\n(1) About the Knowledge of Our Attacker in Reconstruction\\n\\nWe acknowledge the reviewer\\u2019s concern that in Sec 4 where we discuss the reconstruction of MalConv and ProxylessNAS-CPU, the search space of our attacker is not clear. Here, we elaborated on the attacker\\u2019s search space, and we updated the section in our paper.\\n\\n[For Reconstructing Data Pre-processing Pipelines]\\nOur attacker knows what tensor operations and functions to monitor in the victim\\u2019s open-source deep learning framework. These functions are model-independent; they correspond to architectural attributes designated by the deep learning framework. In Table 3, we provide a complete list of the operations and functions used for composing a data-preprocessing pipeline or a deep neural network from two popular frameworks: PyTorch and TensorFlow. We show that this knowledge is sufficient to reconstruct novel data-preprocessing pipelines, such as MalConv, that are usually shallower than the network architectures.\\n\\n[For Reconstructing Deep Neural Network Architectures]\\nTo extract the deeper network architecture automatically designed by NAS algorithms, we assume our attacker has some knowledge about the NAS search space\\u2014e.g., NASNet search space\\u2014the victim\\u2019s search process relies on. This information includes knowledge about the list of layers used and the fact that a set of layers (known as blocks) are repeatedly used\\u2014e.g., in NASNet, Normal and Reduction Blocks are used to build the architecture. We make this assumption because, from the sequence of computations observed via Flush+Reload, our attacker can easily identify a set of layers and the repetitions of the layers. However, we don\\u2019t assume how each block is composed by using the layer observations directly. Instead, we identify candidate blocks by using frequent sequence mining (FSM). We demonstrate that, under these assumptions, our attack reconstructs the ProxylessNAS architecture in 12 CPU hours rather than running a NAS algorithm from scratch that takes 40k GPU hours.\\n\\nWe do not assume our attacker knows the victim\\u2019s architecture family. This differs from prior work [1, 2] in which the adversary already knows the victim\\u2019s network is one of the VGG family and aims to identify whether it is VGG11, 13, 16, or 19. We also do not assume our attacker is able to extract the number of for-loop iterations without noise like [1] assumed. We instead propose a technique to reduce the noise in the attacker\\u2019s observation effectively.\\n\\n(2) Potential Defense Mechanisms\\n\\nHere, we proposed four defense mechanisms that obfuscate our attacker\\u2019s observation in the Flush+Reload traces. We think this is the key idea behind what the reviewer suggested by adding null/useless operations. All the proposing defenses can apply to the deep learning framework easily; however, we also mention that the obfuscations come with a cost. We include the detailed discussion in Sec 5 of our revised paper.\\n\\n1) We can add a small amount of noise (null/useless information) to the matrix multiplication to make it difficult for the attacker to estimate the computational parameters such as kernel sizes or strides. We propose to achieve it by padding zeros to tensor operands randomly. However, if our attacker can observe the same computation multiple-times, our attacker can cancel-out the blended noise and estimate the parameters correctly.\\n\\n2) We can modify the victim\\u2019s architecture so that it includes the identity layers or the branches whose outputs are not used. We think a small number of null/useless operations will not increase the attacker\\u2019s computational burden significantly. This addition will increase the time needed to reconstruct the victim\\u2019s architecture by a few hours. When the defender adds an excessive amount of null/useless layers or branches, this can significantly increase the time taken for the reconstruction. However, the defense may not make the reconstruction impossible, but will increase network evaluation time significantly.\\n\\n3) We can also shuffle the computation orders of a victim network; a network computes the layers sequentially as they are defined in the source code. Here, we can identify the layers that can be computed independently, and shuffle the computation orders randomly each time when the network processes an input. This will make the attacker\\u2019s observation via Flush+Reload inconsistent. However, to hold the output of independent computations, the defender may use memory more.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This work proposed a method to reconstruct machine learning pipelines and network architectures using cache side-channel attack. It is based on a previous proposed method Flush+Reload that generates the raw trace of function calls. Then the authors applied several techniques to rebuild the computational graph from the raw traces. The proposed method is used to reconstruct MalConv which is a data pre-processing pipeline for malware detection and ProxyLessNas which is a network architecture obtained by NAS.\\n\\nOverall, the paper is well-written and easy to read. The problem of stealing machine learning pipelines/architectures is interesting and important, since it enables an attacker to actually know the private networks that are being used for prediction. Therefore, I think this is a promising direction for future work.\", \"i_hope_the_authors_can_address_my_concerns_as_follows\": \"\", \"q1\": \"What is the knowledge of the attacker? The authors should be explicit in summarizing the detailed search space of the attacker. Currently i found it very hard to understand the capability of attacker. This is important in evaluating this work.\", \"q2\": \"Can the authors add some discussion on how to defend against the proposed attack? For example, one can add some null/useless operation during execution to make the reconstruction process harder?\", \"q3\": \"I am curious why the authors choose ProxylessNAS-CPU for evaluation. There is a bunch of other architectures found by NAS, e.g. MNas, ENas?\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary\\n---\\nThis paper proposes to use a computer security method, \\\"Flush+Reload\\\" to infer the DNN architecture of a victim in the setting where both the attacker and the victim share the same machine in a cloud computing scenario. This does not require any physical access to the machine, however it does require that a CPU is shared, and the inference of the architectural details is based on the time it takes to reload computations from cache. \\nThe paper is overall clear and well written.\\n\\nMotivations of the paper\\n---\\nHowever, concerning the motivations of the paper, I'd like some clarifications. As far as I know, in the deep learning community, the most effective architectures are published and public (VGG, Inception, ResNet, Transformer...).\\nI am a bit confused by the sentence \\\"As a result, in the industry such novel DL systems are kept as trade secrets or\\nintellectual property as they give their owners a competitive edge (Christian & Vanhoucke, 2017).\\\" which justifies that architectures are kept secret and thus may be prone being stolen.\\nThis US patent is public and explains the method. As far as I know, it has never been enforced. Furthermore, this patent is associated with the paper \\\"Going deep with convolutions\\\", Szegedy et al. which introduced the Inception architecture, is public, very well-known, and thus I do not believe anyone would have any commercial interest in stealing it. \\nFurthermore, I do have the impression that the edge many companies have over their competitors is the private datasets they own much more than the architectural details.\\n\\nMethod and applicability\\n---\\nWhile the method Flush+Reload itself is not novel, its application to the DNNs case and the way to reconstruct the architecture (generating the candidates, pruning) is.\\nHowever, I do have some practical concerns about the applicability of the method.\\n\\nAs far as I understand, it can only work on one CPU. Most DNNs, even for inference, are run on (one or multiple) GPUs. Can the method be extended to work on GPUs?\\n\\nAlso, while the assumption that both the attacker and the victim use the same framework is realistic to me, I believe, they should also both use the same version of the said library, no? Otherwise some operations might be faster in some versions and slower in others, this is thus an additional and much stronger assumption to make.\\n\\nAt last, this would require the victim to use a public cloud service. However, as far as I know, many of the companies who could potentially design new architectures have their own private cloud. I am not certain that someone disposing of a new, private, and powerful architecture would use it on a public cloud service. \\n\\nExperiments\\n---\\nThe experimental section seems very limited to me. The authors show that they are able to reconstruct perfectly 2 architectures. While this is encouraging, I would like to see the limits of the proposed method.\\nWhy not generate N random (or not so random) architectures and try to reconstruct them? Where does the method fail, where does it succeed?\\nWhat if the victim used a custom layer that the method could not recover? Does it still recover a similar architecture?\\n\\n\\nConclusion\\n---\\nWhile the paper, proposed to use Flush+Reload for recovering DNNs architectures and succeeds for at least 2 non trivial architectures, I do not recommend acceptance.\\nFirst I am concerned by the problem this paper is tackling. Can this realistically happen in a real-life scenario?\\nSecond, I am worried that the method suffers from very strong limitations in practice (eg the usage of a CPU for both victim and attacker).\\nFinally, and importantly, while the experiments show some interesting first results, they are limited, I am not able to judge the strengths and weaknesses of the method, and thus I cannot assess the usefulness of the proposed method.\", \"note\": \"I have to say that this paper is definitely out of my area of expertise, even though I am confident in my understanding of the paper, it may be that some of my concerns are unfounded. If this is the case I will adjust my score accordingly.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a way to attack and reconstruct a victim's neural architecture that is co-located on the same host. They do it through cache side-channel leakage and use Flush+Reload to extract the trace of victim's function call, which tells specific network operations. To recover the computational graph, they use the approximate time each operation takes to prune out any incompatible candidate computation graph. They show that they can reconstruct exactly the MalConv and ProxylessNAS.\\n\\nThe paper looks very interesting but also alarming -- more research should be done to countermeasure this attack. I have the following questions: \\n\\n1. To reconstruct the network, you need to generate potentially exponentially number of candidates and do some pruning based on the estimated parameters. This also looks very expensive. I am wondering compared to just doing NAS yourself, how much gain in terms of resources and time this attack can give? \\n\\n2. What is the limitation of the proposed approach, i.e., does it work on any network structures, e.g., sequence networks, graph convolutional networks, etc. \\n\\n3. In the experiments shown, you can reconstruct MalConv and ProxylessNAS with zero error, does the proposed approach alway find the exact match? Under what circumstances can you find the exact match?\"}" ] }

[ "Congzheng Song", "Shanghang Zhang", "Najmeh Sadoughi", "Pengtao Xie", "Eric Xing" ]

[ "Generalized Zero-shot Learning", "ICD Coding", "NLP", "Generative Model", "Deep Learning" ]

{ "note_id": [ "yjaXmBk1rN", "Bkeaav2isr", "S1gmBDhjsr", "HkxGbwhsiH", "B1gDaU2isr", "Sygm8F2p5S", "rJeTy-FAYS", "HkxLy93pFB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752357, 1573795781211, 1573795642856, 1573795577705, 1573795518948, 1572878666895, 1571881189090, 1571830238066 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2571/Authors" ], [ "ICLR.cc/2020/Conference/Paper2571/Authors" ], [ "ICLR.cc/2020/Conference/Paper2571/Authors" ], [ "ICLR.cc/2020/Conference/Paper2571/Authors" ], [ "ICLR.cc/2020/Conference/Paper2571/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2571/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2571/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a method to do zero-shot ICD coding, which involves assigning natural language labels (ICD codes) to input text. This is an important practical problem in healthcare, and it is not straightforward to solve, because many ICD codes have none or very few training examples due to the long distribution tail. The authors adapt a GAN-based technique previously used in vision to solve this problem. All of the reviewers agree that the paper is well written and well executed, and that the results are good. However, the reviewers have expressed concerns about the novelty of the GAN adaptation step, and left this paper very much borderline based on the scores it received. Due to the capacity restrictions I therefore have to recommend rejection, however I hope that the authors resubmit elsewhere.\", \"title\": \"Paper Decision\"}", "{\"title\": \"List of changes in the revised version\", \"comment\": \"Dear reviewers,\\nWe uploaded a revised version of our paper and below is the list of changes based on the suggestions from the reviews.\\n\\nReviewer #1, #2, #3\\nWe revised the introduction to emphasize the background of generalized zero-shot ICD coding problem and the novelty of our framework.\\n\\nReviewer #1\\nWe modified Figure 1 and added Label Encoder to keep it consistent with the text.\\nWe revised and added a subsection 3.1 as an overview and global picture of our framework. \\n\\nReviewer #2\\nWe clarified definition of nearest sibling in Section 3.3.\\n\\nWe slightly revised the experimental section to make the explanation of each experiment more organized. We removed the misleading citation and caption in Table 1, which shows the results of our modified feature extractor on seen ICD codes. We added the comparison of our modified feature extractor with the original work [1] in Table 5 of Appendix E. We emphasized the captions in Table 2 and 4 so that the difference is clearer. \\n\\nReviewer #3\\nWe added the precision-recall curves in Appendix D.\", \"reference\": \"[1] Rios and Kavuluru, Few-Shot and Zero-Shot Multi-Label Learning for Structured Label Spaces, EMNLP 2018\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We thank Reviewer 1 for the insightful comments and would like to address the specific questions below.\\n\\nComment #1. Description of the complete system:\\nThanks. We have added more detailed description of the complete system in the revised version. Please kindly refer to the second paragraph of Section 1 and Section 3.1 for a description of our complete system and global picture, including classification part. Detailed discussion on classification was included in the \\u201cMulti-label classification\\u201d paragraph of Section 3.2 and \\u201cFine-tuning on generated features\\u201d paragraph of Section 3.3.\\n\\nComment #2. Emphasis that the classes come with a short text description:\\nThanks. We have added the emphasis of this fact in both Section 1 and Section 3.1 in the revision. \\n\\nComment #3. \\u201c3.1 is related to feature extraction... is also apply to class description.\\u201d\\nAs explained in \\u201cLabel-wise feature extraction\\u201d paragraph of Section 3.2 in the revised version (3.1 in the previous version), we process the label description in the feature extractor so as to extract label-wise features for each document. For an input document, a total of |L| feature vectors are extracted and each feature vector contains the relevant information to each label description in the set of labels L. \\n\\nComment #4. \\u201c3.2 label encoder : if the notations... help understanding.\\u201d\\nThanks. We have updated figure 1 with the Label encoder. Other notations in the figure are consistent with the text.\\n\\nComment #5. \\u201cResults : the claim \\u2018our methods improve the F1 score from nearly 0 to 20.91% for the zero-shot codes\\u2019 is not true\\u201d\\nWe would like to point out that Xian2018 and Felix2018 are published works on image classification in the computer vision research community, not on ICD coding. The only existing baseline on the generalized zero-shot ICD coding is Rios and Kavuluru\\u2019s work and their F1 score is nearly zero. We are the first to propose the adversarial generative model with hierarchical label structure for the GZSL text classification problem and achieved ~21% F1.\\n\\nEven compared with Xian2018 and Felix2018, our final model has achieved significant improvement. We have performed student-t test on micro F1 scores. The p-values are 0.00018 and 0.014 comparing to Xian2018 and Felix2018, indicating the improvement between our final model and prior works are statistically significant.\\n\\nComment #6. \\u201cWhen looking at Figure 3 and the analysis... impact on the classification results (F1 20.48 versus F1 20.30)\\u201d\\nWe conducted student-t test on micro F1 scores from WGAN-Z and WGAN and the p-value is 0.01 which demonstrates that the improvement of WGAN-Z is statistically significant.\\nThe goal of Figure 3 is to show that features generated for zero-shot codes using WGAN-Z are closer to the real features from the nearest sibling codes, which indicates that using WGAN-Z can generate features that better preserve the ICD hierarchy. \\n\\nComment #7. \\u201cIt seems that the proposed methods make the model more complex without bringing a significant improvements\\u201d.\\nAs mentioned above, via student-t test, we verified that the new modules proposed in our model yield statistically significant improvement over the baselines adapted from Xian2018 and Felix2018. Xian2018 and Felix2018 focused on image classification, instead of ICD coding in multi-label text classification, therefore they did not address several key challenges in zero-shot ICD coding: 1) The class label distribution is extremely long-tailed and there are many codes without labeled data. 2) The label space has complex hierarchical tree structure. 3) The input clinical documents are very long and noisy and the key information corresponding to each assigned code is quite sparse. \\n\\nTo solve these challenges, it is necessary to propose new ideas. Specifically, we proposed a novel conditional generative model with hierarchical label structure, which has three novel modules to address the three challenges. 1) our framework adversarially generates features for zero-shot codes and finetune the classifiers with the synthetic features, 2) we exploit the hierarchical tree of the ICD codes by encouraging the features of zero-shot codes to be similar to their siblings, 3) we reconstruct the code-relevant keywords in the input documents from the synthetic features to guarantee the semantic consistency.\\n\\nMore importantly, via significance tests, each of these proposed modules has demonstrated to be effective in improving ICD-coding performance. These modules together frame our novel approach, which, to the best of our knowledge, is the first adversarial generative model with hierarchical label structure for generalized zero-shot ICD coding problem. Besides, compared with the only existing baseline on the generalized zero-shot ICD coding (Rios and Kavuluru\\u2019s work), we improve the F1 score from nearly zero to ~21%.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We thank Reviewer 2 for the constructive suggestions and sincerely appreciate your positive feedback on the novelty and effectiveness of the proposed methods. We would like to address the specific comments below.\\n\\nComment #1. Relationship with previous work by Rios and Kavuluru:\\nThanks for your detailed comment and we made it clearer about the relationship with the previous work by Rios and Kavuluru in the revised version (uploaded in OpenReview). Please kindly refer to the discussion of the limitations of Rios and Kavuluru\\u2019s work in Section 2, and the discussion of the difference between ours and Rios and Kavuluru\\u2019s work in the first paragraph of Section 3.2. We only use the modified Rios and Kavuluru\\u2019s model to serve as the feature extractor and pretrained classifier for our proposed approach. As acknowledged by Reviewer 2, the modifications include: 1) replace the graph convolutional networks with graph recurrent neural network; 2) adopting the label-distribution-aware margin loss for training. The performance gain from our modification is added to Table 5 in Appendix E.\\n\\nBesides the above modification, we would like to highlight other differences of our framework to Rios and Kavuluru\\u2019s: 1) our framework adversarially generates features for zero-shot codes to finetune the classifiers. 2) We exploit the tree structure of ICD codes by encouraging the synthetic features of zero-shot codes to be similar to their siblings 3), we reconstruct the code-relevant keywords in the input documents from the synthetic features to guarantee the semantic consistency. To conclude, we are the first to propose the adversarial generative model with hierarchical label structure for the generalized zero-shot text classification problem and the whole framework is brand new compared with Rios and Kavuluru\\u2019s work.\\n\\nComment #2. Clarify the comment regarding the independent training of the zero-shot cases vs the rest:\\nBy independent training we mean that we only fine-tune the binary classifiers for zero-shot codes with the synthetic features while keeping the pretrained classifier for the rest of the codes fixed. Please kindly refer to the explanation in the last paragraph of Section 3.1 and Section 3.3, and \\u201cresults on seen codes\\u201d paragraph of Section 4.2. \\n\\nComment #3. Definition of siblings ICD codes:\\nSibling codes are codes that share an immediate parent. Since each code can have more than one sibling, we select the nearest sibling code measured by the cosine similarity in the encoded label space. Please refer to \\u201cDiscriminating zero-shot codes using ICD hierarchy\\u201d paragraph in Section 3.3 for the detailed explanation.\\n\\nComment #4. Results in Table 1:\\nThanks for your detailed comments and we are sorry for any confusion caused by the citation. Table 1 shows the performance of our modified feature extractor on the SEEN codes and these performance will remain the same after fine-tuning on zero-shot codes as we do not finetune on SEEN codes (explained in the last paragraph of Section 3.1 and Section 3.3, and \\u201cresults on seen codes\\u201d paragraph of Section 4.2). As explained for Comment #1, we modified Rios and Kavuluru\\u2019s work to serve as the feature extractor and pretrained classifier for our generative framework. The first row in Table 1 shows the results for modified ZAGRNN, and the second row shows the results for modified ZAGRNN trained with label-distribution-aware margin loss. We will clarify this in the final version. \\n\\nComment #5. In Table 2, for the cases where precision and recall are 0, are the AUC numbers correct? \\nPrecision and recall are 0 because the baseline model assign a very low probability to zero-shot codes and thus never predicts these codes on any input text given the default threshold of 0.5. AUC can be larger than the random guessing baseline of 0.5 as AUC is measured with all possible thresholds. \\n\\nComment #6. Explanation on what you considered few-shot cases for those results. \\nThanks for your suggestion. Please find our detailed explanation about few-shot cases in the last paragraph \\u201cResults on few-shot codes\\u201d of Section 4.2.\\n\\nComment #7. I think the paper could improve the explanation of the components of the model and the presentation of the evaluation results.\\nThanks for your suggestion. We will carefully improve the explanation of the model components and the presentation of the evaluation results in the final version.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We thank Reviewer 3 for the insightful comments. We appreciate your positive feedback and address the specific questions below.\\n\\nComment #1 & #2. Degree of (effective) novelty:\\nWe would like to emphasize that prior works on generalized zero-shot learning (GZSL) mainly focused on visual tasks, the study of GZSL for multi-label text classification is largely under-explored. Generalized zero-shot ICD coding is a much harder task and we cannot simply apply GAN-based method as in vision problem (e.g., Xian et al. 2018) or label-distribution-aware margin loss to solve it, due to the following aspects: 1) The label distribution is extremely long-tailed and it is challenging to perform multi-label classification on both seen codes and zero-shot codes. 2) The label space has complex hierarchical tree structure. 3) The input clinical documents are very long and noisy and the key information corresponding to each assigned code is quite sparse. \\n\\nTo solve the above challenges, we proposed a novel conditional generative model with hierarchical label structure. Our model has three new components to address the aforementioned three challenges. For the first challenge, our framework adversarially generates features for zero-shot codes to finetune the classifiers. For the second challenge, we exploit the tree structure of ICD codes by encouraging the features of zero-shot codes to be similar to their siblings. For the third challenge, we reconstruct the code-relevant keywords in the input documents from the synthetic features to guarantee the semantic consistency. These three model components were not contained in the model of Xian et al. (2018).\\n\\nEach module contributed to the improved performance, and all of them together frame our novel approach, which, to the best of our knowledge, is the first adversarial generative model with hierarchical label structure for generalized zero-shot ICD coding problem. \\n\\nRegarding the effectiveness, we have performed student-t test comparing our final model against previous GAN based approach on micro F1 scores. The p-values are 0.00018 and 0.014 comparing to Xian2018 and Felix2018, indicating the differences between our results and prior works are statistically significant.\\n\\nQuestion #1. Could the model benefit from having a self-attention model, like a transformer? \\nThe clinical documents are very long and each document can have thousands of tokens. We use 1D convolution for feature extraction mainly because it is most efficient to train for long text. Current state of the art transformer models (BERT, GPT-2) are all trained on sentence level inputs and are extremely memory consuming for longer sequences[1], mainly due to the n x n self-attention matrix in each layer where n is the number of tokens. It is an interesting future direction to explore whether we can perform efficient training and feature extraction with more advanced encoder models like transformers on long clinical texts. \\n\\nQuestion #2. Could the precision-recall curves be added to the supplementary material?\\nThanks for your constructive suggestion. We added the Precision-recall curves to Appendix D of the revised version (uploaded in OpenReview).\\n\\nReference\\n[1] https://github.com/google-research/bert/blob/master/README.md#out-of-memory-issues\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper proposes a method to do zero-shot ICD coding, which comes down to determining which elements from a set of natural language labels (ICD codes) apply to a given text (diagnostic summary). The main problem is that many codes have zero or very few examples. The proposed solution for this problem is to learn a feature-space generator for examples which can be conditioned on a code. This generator is trained using a GAN, which moves the few-/zero-shot problem to training the discriminator. Here, the tree structure of the ICD codes is used, by using examples with sibling labels as approximate examples for codes with few or on labelled data points. Additionally, a keyword reconstruction loss is used, based on the idea that the keywords of the corresponding ICD code can be reconstructed from a good feature vector.\\n\\nThe paper is written and executed well; the setup, which involves a fair number of components, is described and motivated clearly. The experiments include comparisons to state of the art results, finding that applying the GAN-based technique they choose (introduced in Xian et al., 2018 for computer vision problems) produces significant gains in recall of low-data classes (few or zero examples). The precision decreases, probably due to a shift from false negatives to false positives, while the AUC shows modest gains.\\n\\nThe main potential issue with the paper is the degree of (effective) novelty. The bulk of the gain relative to the SotA seems to be achievable by applying the GAN-based method, which is described in the Xian et al. (2018) reference, or by applying the label-distribution-aware margin, which are known techniques even if they have not been applied to this particular dataset yet. The additional elements introduced by the authors - the use of sibling codes and the keyword reconstruction loss - are good ideas, and it is worthwhile having a documented test of the benefit they provide, but they don\\u2019t seem to have a major influence on the quality of the model.\\n\\nAll in all, I would argue for the paper to be accepted. The work done here is valuable to have on record, and the presentation and execution are well done.\", \"two_questions_for_the_authors\": \"1. Could the model benefit from having a self-attention model, like a transformer? This applies mostly to the diagnostic text encoder, as interpreting the ICD code descriptions seems not to depend strongly on structure or context. From the text of the paper it appears that a 1D convolution was used to process the diagnostic texts, but I would expect that longer-distance links between words can be quite relevant there.\\n\\n2. Could the precision-recall curves be added to the supplementary material?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2571\", \"review\": \"This paper addresses the problem of zero-shot prediction for ICD codes for medical notes. The main idea of the paper is to use GANs to generate latent features for the text conditioned on the ICD codes and taking into account the tree-like structure of the ICD codes themselves. The classification of ICD codes is trained on the GAN-extracted features.\\n\\nThe input clinical document is represented as a concatenation of its word embeddings. The representation of an ICD code is realized by averaging the embeddings of the words in its corresponding description. The feature extraction model takes the word embeddings from the input text and passes them through a CNN; in addition, attention is used for the encodings of all the ICD codes.\\n\\nFor further encoding the ICD codes, the tree-like organization of the codes is exploited by using GRNNs. The generated features together with the graph-encodings of the ICD codes are used to classify the clinical text.\", \"some_suggestions_for_improvement\": [\"It was not clear the relationship with previous work by Rios and Kavuluru. I had to refer to the paper to be able to appreciate the differences. The feature extraction network seems to be similar (Rios & Kavuluru used GCNNs instead of GRNNs to encode the ICD codes). The GAN-based feature generation seems to be new in this paper.\", \"I did not understand the comment regarding the independent training of the zero-shot cases vs the rest.\", \"Siblings ICD codes mean codes that share an immediate parent?\", \"Table 1. seems to cite the related work results, but it doesn't seem to include the results of the proposed method.\", \"In Table 2, for the cases where precision and recall are 0, are the AUC numbers correct? Table 2 also shows the results for the ablation studies that are discussed later in the section. It took me a while to understand the difference between Table 2 and Table 4.\", \"For Table 4, you could add the explanation on what you considered few-shot cases for those results.\", \"I think the paper could improve the explanation of the components of the model and the presentation of the evaluation results.\", \"I think the idea of using GANs to generate features that can help with the classification + taking advantage of the relationship/structure among the ICD codes are both interesting ideas.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper deals with the problem of text classification when the number of class is large (17000) and most of the classes do not have examples in the training set. This problem is known as Zero-shot learning. The paper proposes to use adversarial methods and the hierarchical organisation of the classes to improve current models.\", \"the_paper_lacks_a_clear_description_of_the_complete_system\": \"Figure 1 seems to be the one but there is no mention of the classification part. The description of each bloc is clear enough independently but many question remains on the global picture.\\n\\nThe author should for example emphasis the fact that in their problemn the classes come with a short text description. This is somehow unususal for a text classification problem where usually the classes are not defined by a description. In the figure, the difference of the processing of the text from the clinical document and from the class description is not clear. \\n\\n3.1 is related to feature extraction for the clinical document, but it is also said that this processing is also apply to class description.\\n\\n3.2 label encoder : if the notations were the same as on figure A, it would help understanding.\", \"results\": \"the claim \\\"our methods improve the F1 score from nearly 0 to 20.91% for the zero-shot codes\\\" is not true : state-of-the-art models such as Xian2018 and Felix2018 are already over 20% F1.\\n \\nWhen looking at Figure 3 and the analysis, WGAN-Z seem to provide better representation that WGAN with it has almost no impact on the classification results (F1 20.48 versus F1 20.30)\\n\\nIn conclusion, it seems that the proposed methods make the model more complex without bringing a significant improvements.\"}" ] }

[ "Dongsung Huh" ]

[ "dynamics", "curvature", "gradient problems", "natural gradient descent", "gradient descent", "path", "approximations", "exact analysis" ]

{ "note_id": [ "QfJjGq7LS", "S1ehohihsB", "HJgZu9j2oH", "HJgx6PjhoB", "r1ltmPsniB", "Syx3zXsnsr", "BJx3CVotqH", "BJxrxSCgqS", "Bkgq6wO6KS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752328, 1573858467879, 1573857897516, 1573857207845, 1573857057240, 1573856020022, 1572611283565, 1572033772856, 1571813314039 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2570/Authors" ], [ "ICLR.cc/2020/Conference/Paper2570/Authors" ], [ "ICLR.cc/2020/Conference/Paper2570/Authors" ], [ "ICLR.cc/2020/Conference/Paper2570/Authors" ], [ "ICLR.cc/2020/Conference/Paper2570/Authors" ], [ "ICLR.cc/2020/Conference/Paper2570/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2570/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2570/AnonReviewer4" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper aims to study the effect of curvature correction techniques on training dynamics. The focus is on understanding how natural gradient based methods affect training dynamics of deep linear networks. The main conclusion of the analysis is that it does not fundamentally affect the path of convergence but rather accelerates convergence. They also show that layer correction techniques alone do not suffice. In the discussion the reviewers raised concerns about extrapolating too much based on linear networks and also lack of a cohesive literature review. One reviewer also mentioned that there is not enough technical detail. These issues were partially addressed in the response. I think the topic of the paper is interesting and timely. However, I concur with Reviewer #2 that there are still lots of missing detail and the connection with the nonlinear case is not clear (however the latter is not strictly necessary in my opinion if the rest of the paper is better written). As a result I think the paper in its current form is not ready for publication.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for your comment\", \"comment\": \"Thank you for your comment on scaling up the analysis.\\n\\nTo numerically confirm the predictions of our theoretical results, our simulation required estimating and inverting the full Hessian of the system (without approximations). Also, the continuous-time dynamics analysis required using very small learning rate for accurate comparison. The computational cost of full Hessian and the small update size were prohibitive to use real world dataset in this work. We are indeed implementing application of the new proposed method for real world dataset in our coming work!\"}", "{\"title\": \"Thank you for the comment\", \"comment\": \"Dear reviewer,\\nThank you for your comment. \\nWe have indeed started analyzing more complicated cases of nonlinear networks and other architectures, and discovered that the results indeed generalize beyond the linear, fully connected networks. \\n1. NGD (natural gradient descent) perfectly smoothes out the learning dynamics of input-output map making it independent of jacobian, and 2 partial NGD (sqrt-NGD) has the equivalent effect but on the parameter dynamics. The insight from the simple system indeed helps understanding these results for the complex, but they are outside the scope of this paper.\"}", "{\"title\": \"Part 2-\", \"comment\": \"The numerical simulations simply used direct estimation of Hessian, which is numerically inverted to estimate natural gradient. The reference for the detailed methods for Hessian estimation of Hessian is now included.\"}", "{\"title\": \"Thank you for the feedback.\", \"comment\": \"Dear reviewer,\\nThank you for the feedback. We included major updates and clarifications per your comments. \\n\\n1. The results of the paper are limited to specific deep linear neural networks which are not practical. Unrealistic/Unexplained assumptions. \\n\\nOur work aims to provide a novel contribution to extending our understanding of deep network's learning behavior beyond the usual gradient descent update regime. The focus on linear networks is an important mathematical tool/framework that allows insightful understanding of the complex phenomenon. It is a necessary limitation as a first step, upon which we hope to develop more generalized theory. \\n\\nAll the assumptions are borrowed from other theoretical works in the field, which are now more clearly referenced. They used for deriving the exact analytical formulas, but they not critical for the general qualitative results. we now include numerical results that supports the generality of the predictions. \\n\\n1. Lacks a cohesive introduction:\\n\\nWe have extensively improved our introduction section. Especially we explain that many of the theoretical analysis of deep learning depends on the specific example of gradient descent, and raise the challenge whether those theories generalize under a wider spectrum of learning rules. \\n\\n2.\", \"algorithms\": \"\"}", "{\"title\": \"General Comments\", \"comment\": \"We thanks all reviewers for their comments. We have extensively updated our manuscript per reviewers\\u2019 suggestions. All major updates are marked in blue. The main summary of the updates are:\\n\\n1. We included extended literature review and detailed motivation in the introduction. We also clarified our contribution and further polished the overall writing as suggested by the reviewers. \\n\\n2. We clearly state the assumptions and limitations relevant for the analysis.\\n\\n3. We then included additional numerical simulation results that shows our prediction is robust and generalizes beyond the assumption that allowed exact theoretical analysis. (Figure 3 D, Fig SI1, Fig SI2)\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this manuscript, the authors analyze the dynamics of training deep linear neural networks under a generalized family of natural gradient methods that apply curvature corrections. They first show that the learning trajectory (direction of singular mode dynamics) in natural gradient descent follows the same path as gradient descent, while only accelerating the temporal dynamics along the path. Moreover, the authors show that the learning trajectory in layer-restricted approximations of natural gradient descent can significantly differ from the true natural gradient. Also, the authors proposed a fractional natural gradient that applies partial curvature correction which in addition to faster convergence, neutralizes vanishing/exploding gradient problems.\\n\\nI vote to reject this paper due to the following major concerns.\\n1.\\tThe results of the paper are limited to specific deep linear neural networks which are not practical. More specifically, in addition to only covering the deep linear case, the results hold under the following unexplained assumptions: (a) the correlation matrix of the input data is assumed to be whitened; (b) the singular vectors of adjacent weight matrices are well aligned; (c) Section 4, the singular values of all the weight matrices are the same.\\n\\n2.\\tThe paper lacks a cohesive introduction that includes a literature review on this very rich area of research.\\n\\n3.\\tThe paper is not clearly written and is missing many details.\\n\\nI discuss my comments on the manuscript in details next.\\n\\n\\u2022\\tWhile the assumption on the input correlation matrix being whitened can be dropped, other assumptions in the paper are unrealistic and are stated without any intuition or explanation (the singular vectors of adjacent weight matrices are well aligned; (c) Section 4, the singular values of all the weight matrices are the same.)\\n \\n\\u2022\\tThe analysis performed in the paper cover the deep linear case and no intuition is provided on how these results can help in understanding the more general non-linear case. By citing [Saxe et al. 2013], the authors claim that the effect of curvature of loss landscape on deep learning dynamics can be well captured by deep linear networks. However, in [Saxe et al. 2013] this was only shown empirically on a specific example. In recent years, we have encountered many results that hold in deep linear networks but fail to hold when including simple non-linearities. For instance, when the target matrix Y is full rank, every local minimum of the studied problem is global in the deep linear case. However, this result does not hold even with the simplest non-linearities.\\n\\n\\u2022\\tThe paper is unpolished and the objectives and contributions are not clearly stated. Moreover, the results are not rigorously stated and proved. Moreover, the algorithms stated and proposed in the paper are directly used without any introduction or explanation (how they work? What update rules do they use? \\u2026)\\n\\n\\u2022\\tThe introduction in the paper is very short and does not include a literature review on this rich area of research. Hence, the problem is not motivated by a gap in the literature and not linked to other results that study the behavior of algorithms used for training deep neural networks.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Authors analyse curvature corrected optimization methods in the context of deep learning. They build their analysis on Saxe et.al.s work. They show that curvature corrected methods preserve properties of SGD. They also show the disadvantages of layer restricted approximations. They show the importance of time scales in optimization. The paper looks to deep learning from a dynamical systems perspective and hence their experiments are fitting to this framework.\\nI have checked the analysis and it seems solid. However, my only concerns it the use case. I would like to see such methods show difference in real world datasets. I believe this is a big missing part of the paper.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"In this paper, the authors study the training dynamics of natural gradient descent with linear DNNs.\\nSpecifically, they showed that the curvature corrected natural gradient descents preserve the learning trajectory of plain gradient descent and only affect the temporal dynamics. A fractional natural gradient \\ndescent method that only applies partial curvature correction is proposed to address the numerical stability \\nissue of vanilla natural gradient descent.\\n\\nThe paper is well presented and the derivations of analytical solutions are clear. Although the analysis is \\nlimited to the linear case, it does give some insight into the behaviors of curvature corrected gradient descents.\\nHowever, it would be interesting to see how these algorithms would perform under the no-linear case. Also, \\nwill these results hold for other neural network structures? For example, the Hessian of each layer of\\nResNet will be close to orthogonal and how will that affect the NGD?\\n\\nOverall I think this is an interesting paper with strong results and vote for accepting.\"}" ] }

[ "Jung-Su Ha", "Young-Jin Park", "Hyeok-Joo Chae", "Soon-Seo Park", "Han-Lim Choi" ]

[ "Reinforcement learning", "Self-supervised learning", "unsupervised learning", "representation learning" ]

{ "note_id": [ "iV_Cewp2j-", "rkl_IbunsB", "H1gZ-iP3sB", "BJxY32vjjH", "r1lStIxijB", "Hyx-WQZvsr", "HJlhYAV8iH", "BygMuAE8jS", "Skee2aV8sH", "BkxYlaN8jB", "B1exE3ELiH", "SJgWiAJCFr", "r1erI4tctH", "Byl6ccPwFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752299, 1573843279787, 1573841657129, 1573776560768, 1573746300865, 1573487352783, 1573437060199, 1573437033979, 1573436839791, 1573436657437, 1573436456061, 1571843736532, 1571619916753, 1571416724663 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/Authors" ], [ "ICLR.cc/2020/Conference/Paper2565/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2565/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2565/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors present a self-supervised framework for learning a hierarchical policy in reinforcement learning tasks that combines a high-level planner over learned latent goals with a shared low-level goal-completing control policy. The reviewers had significant concerns about both problem positioning (w.r.t. existing work) and writing clarity, as well as the fact that all comparative experiments were ablations, rather than comparisons to prior work. While the reviewers agreed that the authors reasonably resolved issues of clarity, there was not agreement that concerns about positioning w.r.t. prior work and experimental comparisons were sufficiently resolved. Thus, I recommend to reject this paper at this time.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Final comment to the reviewers and AC\", \"comment\": \"We are really grateful for the reviewers' valuable and constructive comments/discussions. The followings are the issues raised by the reviewers and our responses.\\n==============================================================================================\\n- The reviewers mentioned that our draft was hard to follow, mainly because the experiment section was missing several details. We've revised the whole manuscript to clearly show our contribution and included all the experiment details. We will also release the code with the final manuscript for the reproducibility issue.\\n\\n- The reviewers pointed out the lack of comparison with the existing works. As our main contribution is a method to construct an internal model, we considered the baseline models with the different design choices of a model and now, in Section 4.1 and with Table 1, we clearly state the correspondence of each choice to the existing works. That being said, we'd say the ablation study in Section 4.1 shows comparisons of our framework with the existing works, providing valuable implications and showing the novelty of this work, as the reviewer2 agreed.\\n\\n- The reviewer3 worried that the usage of the terminology of \\\"self-supervised\\\" might cause confusion (i.e., self-supervsied learning of skills vs. self-supervised learning of a latent model). We agree with his/her opinion and we will modify the title and the statements about that properly.\\n\\n- Based on the reviewer3's comment, We will include the quantitative results for Section 4.2 in the final manuscript, to show the generalization ability of an internal model more clearly. The reviewer 2 and 3 had a question whether this learning is still valid even without imitation tasks; we will conduct the additional experiments on this for the final manuscript.\\n============================================================================================\\nThanks again for all the comments and we hope the revisions have made our work more solid.\"}", "{\"title\": \"Response to reviewer3\", \"comment\": \"Thanks a lot for your constructive comments. We respect your opinion, it really helped us to figure out our contribution more clearly.\\n\\nPrimitive labels $\\\\mathbf{h}\\\\in\\\\{-1, 0, 1\\\\}$ are only used for the initial policy learning ($L=1$) and such labels act just as prior for the internal model learning. For later iterations ($L>1$), a planning module then outputs any commands in $\\\\mathcal{R}$, thus the policy is trained to perform any command in addition to $\\\\mathbf{h}\\\\in\\\\{-1, 0, 1\\\\}$. The internal model also learns a low-dimensional dynamics of such command executions. By \\u201cself-supervised\\u201d or \\u201cunsupervised\\u201d, we only meant the algorithm constructs its latent dynamics and mappings without supervision. There is no notion of \\u201cskills\\u201d in our argument and we only assume that states that a planning module should consider lie on a low-dimensional manifold and our internal model provides dynamics on that manifold in an unsupervised manner, enabling an efficient planning algorithm on that. There are many design choices for constructing internal models, and our ablation study provides a comparison between such choices. As also shown in Table 1 and 2, only the proposed framework considers both a latent state $\\\\mathbf{z}$ and a high-level command $\\\\mathbf{h}$, which outperforms the baseline models (many existing works are already categorized as one of those); high-dimensional states don't allow for learning exact dynamics and high-dimensional actions prohibit the effective planning.\\n\\nTo show our contribution more clearly without confusion and to address your comments, we will revise our paper in the three directions.\\n\\n(1) About the terminology: We can change the title of our paper to something like \\u201cLearning to Reason: Distilling Hierarchy for Planning and Control with Internal Model\\u201d and modify the statements about self-supervised learning to be clearer (by explicitly restricting our discussion to internal model learning). Would you be satisfied with such modifications?\\n\\n(2) Discovering skills without imitation: Questions on which tasks to choose for model learning (either imitation or intrinsic motivation) are actually orthogonal to our work. Better tasks help the algorithm to find the low-dim space more quickly, but even without them, the algorithm should be able to construct an internal model as long as our assumption on a low-dim manifold is true. To see that, we can consider the following minimal scenario: The system has a 2-dim state $s=(s_1, s_2)$ and a 2-dim action, and the dynamics is given by s\\u2019 = s+a. Suppose that a set of tasks that the agent should solve is given by those reward functions that strongly penalize differences between $s_1$ and $s_2$ and distance from the target that is only defined with $s_1$, i.e. $r=100*(s_1-s_2)^2 + (s_1-target)^2$ (the manifold $s_1-s_2=0$ here corresponds to the manifold of locomotions in the humanoid example.). Then, a lower-level policy should stabilize the states w.r.t. the manifold $s_1-s_2=0$ as well as execute commands, while a high-level planner only needs to plan state trajectories on that manifold with the reduced DOF along the manifold. That means, our framework will learn an internal model with a 1-dim latent state a 1-dim command; the internal model would think deviations of the states from the manifold (which have stable dynamics) as nothing but noise. We will include this toy example in the final manuscript.\\n\\n(3) Quantitative results: Section 4.2 shows the hierarchical policy learned with imitation directly can be applied to navigation tasks, and by generalization, we meant it is because of the generalization ability of having an internal model. In the final manuscript, we will also include the quantitative results/analysis of this example to support our claim.\"}", "{\"title\": \"Response to Rebuttal\", \"comment\": \"Thank you to the authors for revising the draft and responding to the comments. The new draft is certainly more polished. However, I think there are still major issues with the method presentation and comparisons, so I am keeping my score as is.\\n\\nIn the arguments below, the authors write \\\"the low-level policy learns locomotions via the imitation learning with the manually set reasoning module and the internal model is trained via self-supervised learning.\\\" If this is the case, then the proposed approach amounts to using ground truth information (demonstrations and primitive labels [-1,0,1] to learn a set of motion primitives in the form of the conditioned policy, then doing planning in this high level 'primitive' space. Thus even though the internal model is learned on just its own experience, this is dependent on having a good set of primitives, which would not be possible without the imitation. So as currently formulated, it seems like the success of the method as a whole only would work given the strong supervision of demonstrations and primitive labels, so I still think calling it \\\"self-supervised\\\" is a stretch. \\n\\nHowever regardless of terminology - I still feel the lack of comparisons hasn't been addressed. The results still only consist of a set of ablations. There should be comparisons to approaches which use similar amounts of supervision. So given that the method is using imitation learning, there should be comparisons to methods which also use demonstrations. For example there is work on learning 'skills' from demonstrations [1] which seems especially relevant. Since the proposed approach also has labels for the skills, it should be able to outperform this style of approach. I even think showing experiments against the fully unsupervised approaches would be good, since the proposed approach should work quite a bit better, and would also better motivate using demonstrations. But without any comparisons to prior work its hard to assess how well the method works.\\n\\nAlso minor point, the authors write \\\"...showing powerful generalization ability of reasoning/planning process\\\". I don't this can be claimed without quantitative numbers indicating the generalization ability is better than existing work. As it stands - the results are only for 3 tasks - left, forward, right, all of which are seen during training.\\n\\n[1] Roy Fox, Sanjay Krishnan, Ion Stoica, and Ken Goldberg. Multi-level discovery of deep options.\", \"arxiv_preprint_arxiv\": \"1703.08294, 2017.\"}", "{\"title\": \"Further revisions in experiment section\", \"comment\": \"Thank you for your positive comments and suggestions. We've revised the experiment section to deliver our setting and results more clearly.\\n\\n- We explain the model, simulation environment, learning algorithms, and hyper-parameters with explicit references to the appendices at the beginning of the section.\\n- The subsections are then started with the task specifications and the settings for the baseline algorithms, with the proper references to appendices.\\n\\nWe hope the experiment section has now become much clearer and we will continue to revise the entire manuscript.\"}", "{\"title\": \"Authors meaningfully improved draft + score update.\", \"comment\": \"As of the revision visible at the present moment, I remain not entirely happy with the presentation of the experiments, but these concerns are now essentially just issues related to what I believe to be clear writing. I encourage a few more read-throughs by the authors with an eye towards editing for clarity. In particular, I still think the authors don't clearly define their setting succinctly at the beginning of the \\\"Experiment\\\" section. And in the subsections of that section, they should open by very briefly defining the tasks. The appendices provide additional information, but they should be more explicitly referenced, with an indication of what they contain.\\n\\nThe above comments notwithdstanding, the authors have made meaningful revisions, and the content required to fully interpret the results is now present. I think this work is both interesting and contains valuable contributions. So I will update my score to a weak accept (6). Ultimately, the novel contributions of this work involve the ability to perform planning using the low-dimensional space to reuse the humanoid motor skills, and these contributions warrant acceptance to the conference. This is a challenging problem, so this work does provide a somewhat satisfying demonstration. The ablation experiments make the work valuable to build upon.\"}", "{\"title\": \"Initial response to reviewer2 (1/2)\", \"comment\": \"Thank you for your kind review and suggestions. We admit that many parts of our draft were incomplete (due to hurriedness in writing\\u2026), but we\\u2019ve made a big revision of the manuscript especially in the main algorithm and experiment sections and hope that the paper has been improved a lot particularly in terms of presentation.\\n\\n==============================================================================================================================\", \"re\": \"about the dimensionality of the latent space\\nYes, it is actually a coincidence that the (initial) intentions are one-of-three and the latent state is the 3-dimensional. We empirically found that the latent dynamics is encoded best in the 3-dimensional space. We believe that, ultimately, the dimensionality of the latent space also should be learned but the current gradient-based learning algorithms seem not very suitable for doing so.\"}", "{\"title\": \"Initial response to reviewer2 (2/2)\", \"comment\": \"Re: about the experiment section\\n(1) We\\u2019ve clearly stated what \\u201cref\\u201d, \\u201cplan\\u201d, and \\u201ctrue\\u201d mean in the revised paper.\\n(2) We\\u2019ve added much more detailed descriptions of the different structures in 4.1 as well as in the appendix. There is actually no depicted graphical model for $zaz\\u2019$ so we stated the additional explanation for this model (as LVM counterpart of Fig. 2(a)).\\n(3) Yes, your understanding is correct. After the internal model trained (which operates with the full state space), the reasoning is performed there. Since the dynamics should be learned in the full state space, it fails to capture the valid stochastic transition according to $\\\\mathbf{h}$, thereby making the planning become worse. We\\u2019ve added the statements on this as well.\\n(4) We\\u2019ve included the detailed experimental setting for 4.2, such as planning horizon and frequency, used reward functions, and the implication of the results.\\n(5) For the internal model learning, we used 32 samples. It works with so few samples because the learning should be parallelized, but the minibatch training seems to alleviate the problem from the small sample size. We used 1024 samples for the planning and observed that larger sample sizes result in better planning results. A particle filter is just one instance of a reasoning module; if it iterates the whole procedure multiple times to compute a better command, then the planning algorithm becomes the adaptive path integral method. And if the resampling procedure is eliminated, it is equivalent to the widely-used CEM. CEM would work here as well here. We chose a particle filter because we thought its resampling procedure could be very useful for our task; during forward recursion, the procedure greedily gets rid of candidate motions in the wrong direction for imitation tasks and motions colliding with an obstacle for navigations.\", \"re\": \"minor comments\\n- We explicitly state the dimensionality of data, as 197-dimensional state feature and 36-dimensional action space. The 90-dimensional state was just another form of data representation (position/orientation of each link vs. quaternion representation of each joint), and we omitted it for clarity.\\n- f() is a nonlinear function which is parameterized as a neural network.\\n\\n==============================================================================================================================\\nThank you again for your constructive comments. All of them were helpful to revise our manuscript. We would be happy to discuss further if you have any more comments. Please adjust your score if you think the revision has been done properly.\"}", "{\"title\": \"Initial response to reviewer3 (1/2)\", \"comment\": \"Thank you for your critical and constructive comments. As you pointed out, our draft was missing several important details about the experimental setup, descriptions, and discussions, so we\\u2019ve extensively revised the manuscript to strengthen them. Followings are our response to your comments:\", \"re\": [\"about the overall learning algorithm especially regarding self-supervised learning\", \"I think our ideas on how to utilize self-supervised learning are somewhat different from yours. By self-supervised learning, we mean that the agent acquires the low-dimensional representations of observation sequences and its stochastic dynamics only to maximize the data likelihood (e.g., Eq.(7)) without any supervision about any forms of latent space [1,2,3,4].\", \"Imitation learning tasks are considered just to boost the procedure of identifying the low-dimensional internal model (and also because the Mocap data as well as a suitable imitation learning algorithm, DeepMimic, is available. DeepMimic is almost the only algorithm making humanoid walk like a human, in my personal opinion.). In the first main iteration where the agent doesn\\u2019t have any internal models, the low-level policy learns locomotions via the imitation learning with the manually set reasoning module and the internal model is trained via self-supervised learning.\", \"Besides imitation, we might be able to design a proper curriculum of tasks in different outer-loop iterations from crawling, sitting, standing to walking, turning, running, just like how a human learns to walk. Instead of a curriculum, an RL objective function for intrinsic motivation also can be introduced to acquire skills ([5]), which in your perspective is \\u201ctotally unsupervised\\u201d learning; we consider this work complementary to ours (we think imitations are also equally important objectives as intrinsic motivations for acquiring skills) and we\\u2019ve included the statements in the related work and the future work sections.\"]}", "{\"title\": \"Initial response to reviewer3 (2/2)\", \"comment\": [\"RE: about comparisons\", \"Our initial descriptions about the experiment were insufficient and now we\\u2019ve updated them properly. For the ablation study, some numbers for $sas\\u2019$ and $zaz\\u2019$ are missing because these models failed to let the humanoid walk at all. For $sas\\u2019$ and $zaz\\u2019$, the reasoning modules should search over the 36-dimensional action space with the limited number of particles (1024 in our case). It is infeasible due to the curse of dimensionality and, given the fact that the humanoid has very unstable dynamics, the inaccurately planed action can\\u2019t make the humanoid walk.\", \"In this work, we propose a hierarchical policy that, in a high-level, performs planning using a learned internal model. Section 4.2 demonstrates that the proposed policy learned via imitations directly can be applied to unseen navigation tasks, showing powerful generalization ability of reasoning/planning process. We don\\u2019t want to argue that our framework outperforms existing model-free HRL algorithms; even on top of the HRL algorithms, a reasoning module can be introduced (e.g., having more than two levels). What we try to argue in this paper is that, in order to construct such a reasoning module, low-dimensional latent state $\\\\mathbf{z}$ as well as low-dimensional command $\\\\mathbf{h}$ should be considered together. While most model-based RL methods with planning module in a higher-level utilize only one of $\\\\mathbf{z}$ and $\\\\mathbf{h}$, our ablation study shows that the scalable reasoning can be performed only through having both $\\\\mathbf{z}$ and $\\\\mathbf{h}$. We believe that this is a valid and meaningful argument, given the fact that high-dimensional motion planning and control is a longstanding challenging problem in robotics literature [6,7].\", \"We will add the quantitative numbers/analysis for Section 4.2 as well.\"], \"re\": \"relatively minor comments\\n- We\\u2019ve addressed all the other comments, thanks. (fixed typos, strengthened captions, emphasized our contribution, cited and discussed the existing work)\\n\\n==============================================================================================================================\\nWe hope our revision is valid enough so that you can reconsider the score of this submission. We would be happy to have further discussions as well.\\n\\n\\n[1] Karl, Maximilian, et al. \\\"Deep variational bayes filters: Unsupervised learning of state space models from raw data.\\\"\\n[2] Krishnan, Rahul G., Uri Shalit, and David Sontag. \\\"Structured inference networks for nonlinear state space models.\\\"\\n[3] Fraccaro, Marco, et al. \\\"A disentangled recognition and nonlinear dynamics model for unsupervised learning.\\\"\\n[4] Ha, Jung-Su, et al. \\\"Adaptive Path-Integral Autoencoders: Representation Learning and Planning for Dynamical Systems.\\\"\\n[5] Sharma, Archit, et al. \\\"Dynamics-aware unsupervised discovery of skills.\\\"\\n[6] Vernaza, Paul, and Daniel D. Lee. \\\"Learning and exploiting low-dimensional structure for efficient holonomic motion planning in high-dimensional spaces.\\\"\\n[7] Ha, Jung-Su, Hyeok-Joo Chae, and Han-Lim Choi. \\\"Approximate inference-based motion planning by learning and exploiting low-dimensional latent variable models.\\\"\"}", "{\"title\": \"Initial response to reviewer1\", \"comment\": [\"Thank you for the constructive comments. There seem to be some misunderstandings because of the uncleanness of our initial draft. We\\u2019ve extensively revised the manuscript and let us clarify some points here as well:\", \"This work proposes a framework to train a hierarchical policy for multi-task RL problems (imitation learning is considered just to learn such a policy), where a high-level reasoning module plans the agent\\u2019s motion and a low-level policy generates actual control inputs. We introduce a latent state, $\\\\mathbf{z}$, and a latent command, $\\\\mathbf{h}$, with which the reasoning problem can be formulated. A variational distribution over $\\\\mathbf{h}$, which the reasoning module tries to compute, is parameterized by $\\\\mathbf{u}$ that acts to the same subspace as $\\\\mathbf{h}$ in the latent space. Therefore, the reasoning becomes nothing but the inference problem of low-dimensional $\\\\mathbf{u}$ with the low-dimensional stochastic dynamics of $\\\\mathbf{z}$.\", \"The experiments show that one hierarchical policy learned by the proposed framework can perform various tasks such as imitations of human locomotion and navigations through complex environments. The ablation study in Section 4.1 tells us that, without having either $\\\\mathbf{h}$ or $\\\\mathbf{z}$ (which is the case of most HRL frameworks), the reasoning module can\\u2019t predict its future properly or the plan can\\u2019t be executed well by the low-level policy. Section 4.2 demonstrates that the policy trained for imitation directly can be applied to the unseen (navigation) tasks without any additional training, which shows the generalization power of reasoning.\", \"Regarding the citations, we consider [1] closely related to our work so we\\u2019ve heavily mentioned this work in the manuscript. Thank you for letting us know. The focus on the others [2,3] is somewhat addressing partial observability rather than planning, so we haven\\u2019t included these works.\", \"==============================================================================================================================\"], \"reply_to_the_detailed_comments\": \"\", \"re\": \"about Nit-picks\\nThanks for pointing those out. We fixed typos and will revise further for the final manuscript.\\n\\n==============================================================================================================================\\nWe would be happy to discuss more and please consider to re-score the submission if you agree things become clearer.\\n\\n[1] Danijar Hafner et al. Learning Latent Dynamics for Planning from Pixels.\\n[2] Maximilian Igl et al. Deep Variational Reinforcement Learning for POMDPs.\\n[3] Alex Lee at al. Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable Model.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes a latent variable model to perform imitation learning. The authors propose the model in the control-as-inference framework and introduce two additional latent variables: one that represents a latent state (z) and another that represents a latent action (h). For the generative model, the authors use a sequence latent variable model. For inferring the latent action, the authors use a particle filter. For inferring the states, the authors use an \\\"Adaptive path-integral autoencoder,\\\" though it was unclear where the controls \\\"u\\\" come from. (I assume u is the same as the actions, at which point inferring the states amounts to rollout the policy in the sequence latent variable model). The authors compare to not having the latent states and/or not having the latent actions, and demonstrate that they get better imitation learning scores.\\n\\nOverall, I found the paper difficult to follow and some of the reasoning a bit unclear. The experiments seemed limited in scope, given that the authors discuss reinforcement learning in general, but only provide results on the reconstruction error when doing imitation learning. It is also unclear to me whether the gains from the experiments are from their model, or from the fact that their model probably has more parameters since it has more components. It would be good for the authors to compare to existing work that uses sequential latent variables models for deep RL, such as [1,2,3]\", \"more_detailed_comments\": \"It would be good for the authors to substantiate statements like, \\\"Since training sas? is just a simple supervised learning problem, it had the lowest reconstruction error but the computed action from such the internal model couldn\\u2019t make the humanoid walk.\\\" with plots.\\n\\nThe statement, \\\"zaz' also failed to let the robot walk, because reasoning of the high-dimensional action can\\u2019t be accurate enough.\\\" seems similarly unjustified. If the authors wanted to test this, they could train a zaz' model with some low-dimensional action (e.g. left, straight right) and verify that this works.\\n\\nWhile the authors state that \\\"h can be interpreted as high-level commands,\\\" but if it is inferred at every time step, why is this a \\\"high-level\\\" command?\", \"nit_picks\": \"- \\\"The procedure consists of outer internal\\\" --> \\\"The procedure consists of *an* outer internal\\\"\\n- \\\"via via\\\"\\n- \\\"Such the sophisticated separation was\\\"\\n\\n[1] Danijar Hafner et al. Learning Latent Dynamics for Planning from Pixels.\\n[2] Maximilian Igl et al. Deep Variational Reinforcement Learning for POMDPs.\\n[3] Alex Lee at al. Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable Model.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper presents a framework for learning hierarchical policies using a latent variable conditioned policy operating at the low level, with model based planning at the high level. Unlike prior work which does hierarchical reinforcement learning, the key technical contribution of this work is that they use planning with a latent dynamics model as their high level policy. They demonstrate the method on a humanoid walking task in the DeepMimic [1] environment.\\n\\nWhile the idea is well motivated, this paper should be rejected, primarily due to a lack of experimental results. In particular, the experiments (1) are missing several critical details about the experimental setup, (2) the experimental setup differs significantly from the claims of self-supervision and multi-task RL made in the introduction/method, and (3) there is no comparison to any prior work. Without these, it is impossible to determine if any of the claims made about the proposed method are empirically true.\\n\\nFirst, no information is provided about the reward function used, the horizon of the tasks, or about the planning parameters or policy learning parameters. As currently stated, I don't think any of the results in the paper could be reproduced. Furthermore, the reward function and task horizon used are necessary to determine the difficulty of the proposed tasks. \\n\\nSecond, the title and method section would imply that the method is self-supervised, specifically in how the latent dynamics model is learned. While the samples used to train the latent dynamics model are taken from the agent's experience, the latent conditioned policy is trained (1) with ground truth task reward, and (2) is actually pre-trained on demonstrations of the tasks with ground truth skill labels. This suggests that much of the actual skill learning is done offline in this pre-training stage - with full supervision. As a result, this would make the learning of the latent dynamics model much easier, since the sub-policies have already converged to different behaviors for different values of h. Without this pre-training, learning the low level policies and the LVM jointly would be much more challenging. Hence it seems that the \\\"self-supervised\\\" learning of the LVM is actually heavily dependent on the full supervision used in the pre-training stage. Additionally, the demonstrations used for the pre-training correspond to the same 3 tasks that the agent is later evaluated on (moving forward, left, right). So the method receives full supervision on the test tasks, so the experiments do not actually reflect generalization in multi-task RL as claimed.\\n\\nLastly, and most importantly, there are no comparisons to prior work. The only result shown is the trajectory error against 3 ablations of the proposed method. The reported numbers are error between the reference trajectory and ground truth, predicted plan and reference, and predicted plan and ground truth. First, it seems like the most important number here is the difference between the predicted plan and ground truth, for which numbers are missing for 2/3 ablations. Why is task success or reward not the reported number, and why is performance for 2/3 ablations missing? Additionally, there should be comparisons to existing work both in terms of hierarchical model free RL (for example Nachum et al [2]) and model based RL with latent dynamics models (Hafner et al [3]). The results as presented do not actually support that the proposed method performs better than existing work. The final result of the video of the agent doing a long horizon task also has no quantitative numbers, so again does not support that the proposed method is better.\", \"some_other_less_significant_points\": \"- there are typos throughout (for example \\\"Figure 3: (a) Leanred latent model.\\\").\\n- the tables and figures have very limited or no captions.\\n- The method section is difficult to follow and could use some figures which demonstrate the key technical contribution.\\n- Also from the method section it seems like the novelty is combining the latent conditioned policy learning from Haarnoja et al [4] and the latent dynamics learning/planning from Ha et al [5]. Is there an additional technical contribution beyond combining these two existing works? If so the method section should more clearly show it.\\n- There is a recent work (Sharma et al [6]), which also learns skill conditioned low level policies, and does model based planning in the space of skills to reach previously unseen goals. In this work the skill discovery is also totally unsupervised. The authors should add citation to this paper and clarify the differences between their work and this work.\\n\\n[1] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. DeepMimic: Example guided deep reinforcement learning of physics-based character skills.\\n[2] Ofir Nachum, Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning\\n[3] Hafner, D., Lillicrap, T. P., Fischer, I., Villegas, R., Ha, D., Lee, H., and Davidson, J. Learning latent dynamics for planning from pixels\\n[4] Tuomas Haarnoja, Kristian Hartikainen, Pieter Abbeel, and Sergey Levine. Latent space policies for hierarchical reinforcement learning.\\n[5] Jung-Su Ha, Young-Jin Park, Hyeok-Joo Chae, Soon-Seo Park, and Han-Lim Choi. Adaptive path integral autoencoders: Representation learning and planning for dynamical systems.\\n[6] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamicsaware unsupervised discovery of skills.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"*Edit: original score was a weak reject (3), updating to a weak accept (6) in light of revisions.*\\n\\nThis work implements a hierarchical control scheme for a high-dimensional control problem (locomotion using a humanoid body). The hierarchy consists of a high-level module that plans in an abstract space of \\\"intention\\\", and the intention variables serve as inputs, along with state, to a low-level controller that actually executes the movements. The premise is that a lower-level controller should be usable for multiple tasks, and should be able to be commanded by a lower-dimensional intention input. I find the basic ideas presented clear, the literature reviewed reasonably well, and the motivation and setting to be very interesting. The video summary is valuable.\\n\\nMy main concerns have to do with presentation, but I think they are relatively significant concerns. As the draft currently stands, I would, somewhat regrettably, be inclined to reject the submission (marginally). I think revisions could seriously improve this paper and incline me towards acceptance.\\n\\nAlgorithm 1 indicates that the learning of the low-level controller will be done jointly with the learning of the latent model and planning using the high-level, learned intention space. In the experiment, it is indicated that the low-level controller is pretrained. This points to a couple issues that are not clear in the draft:\\n(1) Presumably this pretraining is necessary and things do not work without it. Indeed, it is hard to imagine that the movements will be well grounded to human motion capture movements without this pretraining. Does the algorithm work as written or is pretraining a fundamentally essential step? There are no settings, even toy settings, where the algorithm as written is shown to be effective.\\n(2) The authors should be clearer how they conduct the pretraining which involves learning the low-level controller. \\n(3) I'm not clear how updating the low-level controller is effective in the algorithm. While I understand why it makes sense to plan in the intention space of the pre-trained controller, and I understand why learning a model is a core part of planning, it would seem like fine-tuning the low-level controller could make the movements deviate considerably from the initial movement space and maybe even eliminate the ability of the low-level policy to express movements that are not used early in training. So essentially, while the planning in the low-D space makes sense and the learning of the model makes sense, the low-level controller update seems possibly to not make sense, and there aren't experiments showing that step helps. \\n\\nIs it just a coincidence that the intention space (h) is one-of-three and the low-d state space (z) is 3-dimensional as well? Or are these both selected with sort of going straight vs turning left or right in mind?\\n\\nThe experiment section is generally very unclear, though details are made a little clearer from the video. In the paper, there are a few points that need to be clearer: \\n(1) \\\"ref\\\", \\\"plan\\\", and \\\"true\\\" are not well defined and it is unclear what these distances in Table 1 refer to precisely. Clearly introduce what each of these refers to. The authors simply say that there are imitation tasks but do not walk through what these terms refer to. \\n(2) In 4.1, the different structures are not adequately introduced. The pointers to the figure 2 diagrams are essential, but there is no pointer for zaz', the pointers are only in the table (not in the text), there are grammar issues in the text and the text could be verbally clearer about the variants.\\n(3) shs' setting is a bit unclear. Basically, clarify briefly how planning is performed in this case. Is a forward model still trained, but the model opperates with the full state space? If so, presumably the forward model is much worse and then the planning approach is correspondingly bad, hence the poor rollouts? \\n(4) 4.2 is I think obviously inadequately described in the text and I can only assume was the result of rushing for the deadline? The second experiment is essentially not presented in the text all aside from a still image. \\n(5) And for all of the experiments that rely on planning with a particle filter, details such as how reliable the filter is in generating useful control, how many samples are required, and possibly elements of compute speed would make much clearer how well the approach actually works. Does the choice of planner matter at all? A common, albeit relatively weak, baseline planning approach is CEM...would CEM work here? I'd like to understand if the choice of particle filter is the author's default choice, which is fine if so, or if there is a positive assertion being made that the particle filter is particularly valuable.\\n\\nI hope the authors will generally improve the exposition in the experiment section (4) during the revisions.\\n\\nOverall, I find the paper well motivated in framing the problem (i.e. using model-based approaches to control the latent space of a low-level controller). I also appreciate the scale of the problem (humanoid control is challenging, so this is not a toy problem). I find the results a bit unclear, perhaps due to hurriedness in writing, so I find them a bit difficult to fully appreciate. Nevertheless, the core contribution that I take away from this work is that there is a value to learning the low-dimensional state representation (z, via the LVM), relative to planning using a forward model on the full state (?...I'm still unclear on the presentation of this result, due to unclear exposition). Slightly more broadly, this is a good demonstration of using a planner jointly with a learned high-level command/intention representation, for a high-dimensional problem. \\n\\nIf I've understood this correctly, I'd be reasonably interested in this result. If the authors can both clarify the core results and communicate that the choices made in the algorithm are well thought through, I would be happy to adjust my score.\", \"relatively_minor\": \"Abstract says 90-dimensional humanoid system, but later it is stated \\\"34 degrees of freedom,\\n197 state features, and 36 action parameters\\\". Where the 90 dimensions comes from is unclear. Often people refer to number of actuators or DoFs. Please adjust this or be more explicit.\\n\\nIn equation 9, f() is not very clearly specified. Is f() a nonlinear function (e.g. a neural network) or is it a linear function? It seems like it might as well be a linear function, since the authors propose to learn a latent dynamics model that is nonlinearly related to the state.\\n\\nTypos in Fig 3 caption.\"}" ] }

[ "Anton Tsitsulin", "Marina Munkhoeva", "Davide Mottin", "Panagiotis Karras", "Alex Bronstein", "Ivan Oseledets", "Emmanuel Mueller" ]

[ "Deep Learning", "Generative Models", "Nonlinear Dimensionality Reduction", "Manifold Learning", "Similarity and Distance Learning", "Spectral Methods" ]

{ "note_id": [ "kQujRaJ0z", "SJxlZk-hoH", "HJe2GlmisH", "H1gtFkmiiS", "rkg6ARzssr", "SylsznxZcB", "HJxW01GvYB", "HklyYEYltS", "ryeqp63PuH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1576798752268, 1573814008219, 1573756947550, 1573756800703, 1573756629083, 1572043794612, 1571393480735, 1570964599034, 1570389442145 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2564/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2564/Authors" ], [ "ICLR.cc/2020/Conference/Paper2564/Authors" ], [ "ICLR.cc/2020/Conference/Paper2564/Authors" ], [ "ICLR.cc/2020/Conference/Paper2564/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2564/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2564/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2564/Authors" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper introduces a way to measure dataset similarities. Reviewers all agree that this method is novel and interesting. A few questions initially raised by reviewers regarding models with and without likelihood, geometric exposition, and guarantees around GW, are promptly answered by authors, which raised the score to all weak accept.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Rebuttal acknowledged\", \"comment\": \"Thanks for the detailed feedback, which cleared up a few things for me (in particular #4 was a big deal for me). I will change my score to a weak accept.\"}", "{\"title\": \"Response to Official Blind Review #3\", \"comment\": \"We thank the reviewer for the detailed comments. Please find our responses below.\\n\\nWith respect to bounding the Gromov-Wasserstein distance, the bound is the loosest for isospectral manifolds / graphs. While there are known isospectral constructions for both Riemmanian manifolds [1,2] and graphs [3], it is conjectured that the proportion of isospectral graphs drops to zero with increasing number of nodes [4].\\n\\nRegarding the restrictions of scale due to the normalization term $e^{-2(t+t^{-1})}$, this term prevents the blowup of heat kernel trace as the scale parameter $t$ approaches 0 from the right. In practice, this temperature scale ($t \\\\approx 0$) does not bring any information except the difference in the number of nodes in the graph (samples in the data), i.e. $\\\\sum_i^n e^{-0\\\\lambda_i} = n$. Going further away from 0, we can observe this behaviour from the Taylor expansion of the matrix exponential of the heat kernel trace for kNN graphs $h_t \\\\approx n - tn + t^2n$. We empirically observe that when $t<0.5$ first two terms of the Taylor expansion approximate the heat trace with a very small error ($<10^{-3}$) with no information that is \\\"interesting\\\" in terms of the shape.\", \"references\": \"[1] https://en.wikipedia.org/wiki/Language_convergence\\n\\n[2] https://en.wikipedia.org/wiki/Ural\\\\%E2\\\\%80\\\\%93Altaic_languages\\n\\n[3] https://en.wikipedia.org/wiki/Ottoman_Hungary\\n\\n[4] https://en.wikipedia.org/wiki/Balkan_sprachbund\\n\\n[5] https://en.wikipedia.org/wiki/Holy_Crown_of_Hungary\"}", "{\"title\": \"Response to Official Blind Review #1\", \"comment\": \"We thank the reviewer for the detailed questions and comments. We hope that our response and improved exposition of the paper will address the concerns.\\n\\n[#1] We consider a general problem of comparing the geometry of samples drawn from unknown data distributions, including cases without any model and, consequently, without likelihood. In our experiments, we study a diverse set of applications, focusing on the scenarios that were not previously possible in the literature. For example, in the experiment in Section 4.2 we compare word embeddings across languages with no model attached to it and unaligned vocabularies. In a different vein, in Section 4.3 we consider similarities in representations for networks *across datasets*, this scenario is not possible to study using traditional methods such as [2, 3]. Besides, even for likelihood-based models, the objective has been shown to not always be sensible for out-of-distribution examples [1].\\n\\n[#2,#3] We have improved the text for a better exposition. The discussion in Section 3.1 starts with heat diffusion in the Euclidean space to present the most common space in the machine learning literature. We shift the focus to graph Laplacians constructed from the samples from general manifolds since they are known to approximate their diffusion properties (see Theorem 1). We then approximate graphs\\u2019 spectrum with stochastic Lanczos quadrature method and use the approximated spectral descriptors to compute the lower bound to the Gromov-Wasserstein distance between the underlying manifolds.\\n\\n[#4] The diffusion process does **not** occur between manifolds. Instead, to compare manifolds we compute the distance **between the descriptors** of these manifolds obtained by the diffusion inside each of them. Such a distance lower bounds an optimal matching of the heat kernel defined on each manifold and thus the Gromov-Wasserstein distance between manifolds.\\n\\n[#5] IMD inherits several guarantees from past literature, and, in addition, we provide approximation error guarantees in Theorem 2. IMD is supported by multiple theoretical statements, while not all of them are error guarantees, we make sure all the steps in the algorithm are principled. IMD is an approximation of the lower bound of the spectral Gromov-Wasserstein distance, as per bounds of (M\\u00e9moli, 2011). We use an OR-graph construction to approximate the Laplace-Beltrami operator of the underlying data manifold, as per (Ting et al., 2010). Then, we approximate the lower bound with Stochastic Lanczos Quadrature, and do so with error controlled by our Theorem 2.\\n\\n[1] E. Nalisnick, et al. \\\"Do deep generative models know what they don't know?.\\\" ICLR (2019)\\n\\n[2] A. Morcos, M. Raghu, and S. Bengio. \\\"Insights on representational similarity in neural networks with canonical correlation.\\\" NeurIPS (2018)\\n\\n[3] S. Kornblith, et al. \\\"Similarity of neural network representations revisited.\\\" ICML (2019)\"}", "{\"title\": \"Response to Official Blind Review #2\", \"comment\": \"We thank the reviewer for the positive feedback and constructive suggestions. We have incorporated points #4-#13 into the updated version of the paper, and will continue to work on better presentation of the work after this comment is posted, i.e. we will include an algorithm and update the introduction as requested. We discuss some of the points below for better exposition.\\n\\n[#1] In machine learning it is common to treat the data as points sampled from some unknown distribution in $\\\\mathbb{R}^d$. We mention data moments to contrast with existing techniques for measuring closeness between two samples. We propose to make use of the geometry of samples instead. We treat samples as coming from some manifold embedded in $\\\\mathbb{R}^d$ (possibly equipped with its own distribution). \\n\\n[#4] An **extrinsic property** undergoes a change when transformations like rotation of shifts are applied to data. For example, the mean of a Gaussian is not invariant under shift, but variance is. An **intrinsic property** stays invariant under **all** isometric transformations (dependent only on the metric tensor). The intuition behind **multi-scale property** of the heat kernel is closely connected with time parameter t. As t grows larger, heat kernel depends on larger neighborhoods of points, so the heat kernel values can serve as a description/summary of the neighborhood at different scales. This is similar to the high-order moments of the data if we treat it in a traditional way as having an unknown distribution in $\\\\mathbb{R}^d$ from which it is sampled. The more moments of a distribution we know, the more we know about its shape. In a similar way the heat kernel accumulates the shape information of the manifold covering all the possible moments of the data seen as a distribution.\\n\\n[#14] We mention that eigenvalues of the graph/shape can be inferred from the spectrum. The exact procedure is defined in the remark 4.8 in (M\\u00e9moli, 2011); we have added this reference to the main body of the paper. In short, we iteratively seek eigenvalues $\\\\lambda_i$ such that $\\\\lim_{t \\\\rightarrow \\\\infty}{e^{\\\\lambda_i t} hkt(t)} \\\\not= 0$, starting from $\\\\lambda_0$. This allows to establish an equivalence relation between the spectrum and the heat kernel trace, allowing us to make precise statements about the informativeness of our metric.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"1. Define/explain data manifold: in the abstract, ``data moments\\\" suggests the data is treated as random variables; the authors could explain more on how they make the connection between random variables and manifolds\\n2. introduction: include an example of ``models having different representation space\\\" \\n3. introduction: elaborate more on ``unaligned data manifolds\\\" \\n4. introduction: the distinction between ``intrinsic\\\" and ``extrinsic\\\" can be made more explicit; as currently written, the difference between ``extrinsic\\\"/``multi-scale\\\"/``utilizing higher moments\\\" isn't clear \\n5. the authors mentioned the hypothesis that high-dimensional data lies on a low-dimensional manifold, but in (3.1) they only considered $\\\\mathbb{R}^d$ without justifying this restriction \\n6. (3.1) ``$u(t, \\\\cdot) \\\\to \\\\delta(\\\\cdot -x)$\\\" change the dot to a variable \\n7. (3.1) ``$u(t, x) = 0 \\\\forall x \\\\in \\\\partial X$ and for all $t$\\\" change to ``for all $t$ and $x \\\\in \\\\partial X$ \\n8. equation 1, write ``we restrict our exposition to Euclidean spaces $X = \\\\mathbb{R}^d$\\\" and remove ``$\\\\forall x, x' \\\\in \\\\mathbb{R}^d, t \\\\in \\\\mathbb{R}^+$\\\" \\n9. (3.1) the phrase ``a compact X including $\\\\mathbb{R}^d$\\\" is confusing\\n as $\\\\mathbb{R}^d$ is not compact under Euclidean topology \\n10. (3.1) ``for a local domain with Dirichlet condition $D$\\\" change to ``for a local domain $D$ with Dirichlet condition\\\" \\n11. (3.1) ``Definition 1\\\" is more of a proposition/theorem. \\n12. (3.1) notation: in the body paragraph the authors use $k(x, x', t)$ but in ``definition 1\\\" they used $k(x, y, t)$ \\n13. (3.1) ``$c$ disconnected components\\\" to ``$c$ connected components\\\" \\n14. (3.1) the authors could elaborate more about how HKT contains all the information in the graph's spectrum. Currently, they only stated the limit equal to the number of connected components. \\n15. Github code is helpful but it would be better if they can include an algorithm section\\n16. A figure of the architecture could be included under the ``putting IMD together\\\" section\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper propose a novel way to measure similarity between datasets, which e.g. is useful to determine if samples from a generative model resembles a test dataset. The approach is presented as an efficient numerical algorithm for working with diffusion on data manifolds.\\n\\nI am very much in doubt about this paper as there are too many aspects of the work I do not (currently) understand. I suspect that the underlying idea is solid, but in its current form I cannot recommend publication of the paper.\", \"detailed_questions_and_comments\": \"*) It seems that the focus is on GANs and related models that do not come with a likelihood. In my reading, the practical problem that the paper address is model comparison in models that do not have a likelihood. Is that correct? If so, I am left to wonder why models with a likelihood a not discussed, and why the such models aren't included in the experiments? Wouldn't the likelihood be a sensible baseline?\\n\\n*) The term \\\"multi-scale\\\" is not defined until page 3. I'd recommend defining this term much more early or avoid its use until its defined.\\n\\n*) I found the geometric exposition to be rather confusing. In Sec. 3.1 emphasis seem to initially be on Euclidean diffusion, and at some point emphasis changes to diffusion on graphs. Sometimes the paper seem focused on diffusion on general manifolds. I found this exposition to be rather chaotic, and I suspect that this is the root cause of me not understanding many aspects of the work.\\n\\n*) In Sec. 3.2 the notion of diffusion *on* manifolds is discussed. At some point (not quite clear to me when) the results of this discussion is applied to diffusion *between* manifolds. I don't quite understand what it means to diffuse between manifolds. I would have appreciated a more explicit exposition here.\\n\\n*) In the concluding remarks it is stated that IMD provides guarantees. Which guarantees are that?\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I made a quick assessment of this paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"IMD is a lower bound to Gromov-Wasserstein distance. Which implies that when IMD is small, this does not guarantee that GW will be small. It is interesting how large can be that gap (and when typically the gap increases).\\nThe term e{\\u22122(t+t^\\u22121)} in the definitions of GW and IMD, obviously imposes some \\\"scale\\\" for both manifolds M and N, so it is hard to call IMD multi-scale.\\n\\nExperiments with language affinities are not convincing to me. How is armenian and hebrew are close, or hungarian and turkish (if \\\"el\\\" is greek, then greek, hungarian and turkish are close somehow?).\"}", "{\"comment\": \"Due to an error in figure generation, Figure 4(b) presents incorrect relative accuracy for VGG-16 trained on the CIFAR-100 dataset.\", \"correct_figure_4_is_available_at_the_following_url\": \"https://i.imgur.com/bMi0tl8.png\\nWe apologize for the error.\", \"title\": \"Formatting corrections to Figure 4\"}" ] }

[ "Hiroaki Yamane", "Chin-Yew Lin", "Tatsuya Harada" ]

[ "numerical common sense", "word embedding", "semantic representation" ]

{ "note_id": [ "Ni5eKNNvjY", "SyxG5xWcqB", "Skl7ei7gcH", "S1gt14y3YS" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752235, 1572634761974, 1571990251274, 1571709920812 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2563/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2563/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2563/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors tackle an interesting and important problem, developing numerical common-sense. They use a crowdsourcing service to collect a dataset and use regression from word embeddings to numerical common sense.\\n\\nReviewers were concerned with the size and quality of the dataset, the quality of the prediction methods used, and the analysis of the experimental results.\\n\\nGiven the many concerns, I recommend rejecting the paper, but I encourage the authors to revise the paper to address the concerns and resubmit to another venue.\", \"title\": \"Paper Decision\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper attempts to study if learned word embeddings for common objects contain information about \\\"numerical common sense\\\". The hypothesis is that certain numerical information may co-occur with the words for certain objects/measurement units within their context windows. To verify this hypotheses, the authors have created a dataset through a crowd-sourcing service which represents \\\"numerical common sense\\\". Using this dataset, the authors examine the predict abilities of regressors trained on learned word embeddings and the aforementioned crowd-sourced dataset. The hypothesis is that if the regressors demonstrate good accuracy, then the word embeddings contained information relevant to \\\"numerical common sense\\\". To the best of my knowledge, this is the first paper that attempts to analyze learned word embeddings in the context of numerical common sense.\\n\\nThis paper should be rejected because (1) the NCS datasets are too small to represent \\\"numerical common sense\\\" (2) the NCS datasets contain faulty data points and (3) the results from the experiments conducted are not sufficient to accept or refute the hypotheses.\\n\\nMain argument\\n\\nThe first question that we must ask is - are the NCS-50x1 and NCS-60x3 datasets reliable for experiments on \\\"numerical common sense\\\". No, because of two flaws:\\n\\n(1) The number of samples in the dataset is too small to represent \\\"numerical common sense\\\". Consider the histogram for object \\\"dog\\\" in Figure 2. If the largest data point in this plot was absent, the average of the distribution would be smaller by several orders of magnitude. Perhaps there are other objects in the dataset which are missing samples from the tail end of the distribution that could have large effects on the mean of the collected dataset. \\n\\n(2) Some data points in the dataset don't make sense to me. For example, Fig 2 represents the \\\"small\\\" dataset, yet I see samples like 400m long dogs, 40m long cats, 150m long monitors and 20m long mice?\\n\\nAlso, it is not clear how the confidence scores of the participants were taken into account when training the regressors or if they were used at all.\\n\\nIf the NCS dataset does not represent \\\"numerical common sense\\\", it invalidates all experimental results from the paper.\\n\\nMy second issue with the paper is that it is not possible to conclude if the experimental results support or refute the hypothesis (ignoring the issue with the dataset):\\n\\n1. In tables 2 and 3, the correlation coefficients were quite low and and the MAEs were pretty large. In Table 3, rows 1 and 2, even though the correlation is 0.57 and 0.48, the MAE is 100 million yen and 7.5 million yen respectively which is quite large. To me this suggests that just because the correlation is larger we cannot conclude mean that the model is performing well.\\n\\n2. It is unclear why the correlation coefficient was chosen to decide that ARD is the superior model in experiment 1. The MAE for random forests with concatenated feature vectors was an order of magnitude smaller than that of the ARD model.\\n\\n3. Why are the correlation coefficients missing for the unit-only experiment in Table 2? The LS model shows very good MAE relative to the other models and perhaps the correlation should have been measured for that as well? In fact, if the correlation coefficients for this case is comparable to the case with concatenated features, it would mean that the word embedding for the object is not helping at all! Moreover, I find it surprising that the LS model with concatenated features performs worse than the unit-only features. We cannot conclude if paper's interpretation about the results is correct unless this missing information is provided.\\n\\n4. It is hard to judge what a correlation coefficient of 0.57 means. Why didn't you provide a scatter plot of the predictions vs targets as well? It often happens that even noisy plots demonstrate good correlations.\\n\\n5. The paper should have additional ablation studies - for example, what would happen in the concatenated feature vector experiment if you trained the regressors using randomly initialized word embeddings instead of the trained word embeddings? Do you get the same performance as learned word embeddings?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper describes the collection and analysis of a numerical common-sense dataset. The paper also states that a novel regression method is presented for quantifying numerical common sense.\", \"strengths\": [\"the numerical common sense task and its related challenges are presented and motivated clearly\", \"the data/annotations and their analysis are presented clearly. The experiment can be replicated reasonably (provided that the data is released, as the authors state)\"], \"weaknesses\": [\"the dataset is too small (230 pairs of units and values). Two observations follow from this: (1) such small-scale data is of limited use to state of the art (deep learning) data-hungry approaches; (2) the relatively low cost of crowdsourcing typically results in much bigger datasets.\", \"some decisions or definitions seem a bit ad hoc and are not convincing. For instance: in Table 1, why is the apple measured in centimetres but the coffee cup measured in meters? Another example from section 1: why are temperature and weight described as nonphysical scales, but money described as a subjective scale? Another example of a definition: \\\"similar words share similar units and thus concrete words contains numerical information\\\". Statements like these are problematic because one could easily think of counter examples where this is not the case.\", \"the experimental findings are not based on state on the art methods applied in a setup where the robustness and transferability of the methods on other data and domains has been analysed and discussed.\", \"Overall this work is interesting as a starting point. My advice is to consider scaling up both the data and the methods, and to accompany this by a more comprehensive analysis of limitations.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"I have read the author response. Thank you for responding to my questions.\\n\\nThis paper aims to predict typical \\u201ccommon sense\\u201d values of quantities using word embeddings. It includes the construction of a data set and some experiments with regression models. The general direction of this work is worthy of study, but the paper needs additional justification for its task, better discussion of recent related work, and more development of its regression models.\\n\\nThe work starts by describing the construction of interesting crowdsourced data sets that include people\\u2019s estimates of typical quantities, what they would consider to be low or high values for a given object in given units e.g. the temperature of a hot spring or the height of a giraffe. Overall, the data sets are interesting but are not especially large (2300 total [low, high] pairs of 230 different quantities). Further, the particular task formulation here needs more justification. I think most of us would agree that common sense is a critical AI challenge, and that the question of whether embeddings reflect typical quantities is important. But, in a paper where the data set is considered a primary contribution, I would expect more justification for exactly how this task is formulated, and which objects and units were selected. As one example, why ask about \\u201clarge\\u201d and \\u201csmall\\u201d values rather than something with more precise semantics (like the 10th and 90th percentile, for example)? I also felt that the introduction could be improved to provide more convincing motivation. E.g., the first paragraph only says that humans apply different adjectives (like \\u201chefty\\u201d and \\u201ccheap\\u201d) to different things depending on their numerical attributes (weight, cost), but does not argue why teaching AI systems to use those adjectives is a priority.\\n\\nRegarding related work, the paper is missing a discussion of several relevant papers that use embeddings to obtain relative comparisons or estimates of commonsense properties of objects, including:\\n\\nForbes, Maxwell, and Yejin Choi. \\\"Verb physics: Relative physical knowledge of actions and objects.\\\" ACL 2017\\n\\nYang, Yiben, et al. \\\"Extracting commonsense properties from embeddings with limited human guidance.\\\" ACL 2018\\n\\nElazar, Yanai, et al. \\\"How Large Are Lions? Inducing Distributions over Quantitative Attributes.\\\" ACL 2019\\n\\nThe paper then presents the performance of some regression models. These models are standard existing techniques, and given the relatively low performance I would have liked more development of the models and more analysis of the performance. For a conference like ICLR I would expect to see a more thorough exploration and analysis of possible models for the task. Looking at more powerful neural regressors (perhaps using contextual embeddings rather than just fixed word embeddings) might be one option. Offering an explanation for why ARD seems to work better than the other approaches would be helpful.\", \"minor\": \"In Table 3, the way that small and large are interleaved makes it hard to compare systems, I think presenting all the small results together, and large results together may help. In Figure 2, it would be helpful to see the histogram for size-large within the same plots here, so we could see how far apart they are.\\n\\n\\u201cBecause Skip-gram has to handle more words to predict words, we assume Skip-gram will obtain more information about numerical values.\\u201d\\n-- I didn\\u2019t understand what you meant about skip-gram having to \\u201chandle more words to predict words.\\u201d Also, I did not understand how this entails that skip-gram would obtain more info about numerical values.\"}" ] }

[ "Michal Rozenwald", "Aleksandra Galitsyna", "Ekaterina Khrameeva", "Grigory Sapunov", "Mikhail S. Gelfand" ]

[ "Machine Learning", "Recurrent Neural Networks", "3D chromatin structure", "topologically associating domains", "computational biology." ]

{ "note_id": [ "0emF5aKs-X", "r1e_byc2sH", "rkeXVCKhsH", "HygflpY3sH", "H1lpF642sS", "rJe0SjlScS", "ryg8ROfaYH", "HylET_-0OB", "HylmAQEc_B", "B1gJ-kcNdH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_review", "comment" ], "note_created": [ 1576798752203, 1573850880125, 1573850667370, 1573850346161, 1573830021437, 1572305734385, 1571789006502, 1570801852097, 1570550731116, 1570180855427 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2562/Authors" ], [ "ICLR.cc/2020/Conference/Paper2562/Authors" ], [ "ICLR.cc/2020/Conference/Paper2562/Authors" ], [ "ICLR.cc/2020/Conference/Paper2562/Authors" ], [ "ICLR.cc/2020/Conference/Paper2562/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2562/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2562/Authors" ], [ "ICLR.cc/2020/Conference/Paper2562/AnonReviewer2" ], [ "~Maria_Anna_Rapsomaniki1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors consider the problem of predicting DNA folding patterns.\\nThey use a range of simple, linear models and find that a bi-LSTM architecture \\nyielded best performance. \\n \\nThis paper is below acceptance. \\nReviewers pointed out strong similarity to previously published work. \\nFurthermore the manuscript lacked in clarity, leaving uncertain eg details about \\nexperimental details.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Blind Review #1\", \"comment\": \"Thank you very much for your thoughtful review! Comments below:\\n\\na) We have expanded the description for the wMSE in the paper. The parameter K in the wMSE definition stands for the number of input objects (in test/val/train sets). \\n\\nb) To increase clarity we have updated the definition of the loss function to be called modified wMSE. \\n\\nc) Fixed the citation.\\n\\nd) We have rephrased the suggested paragraph about the DNA properties.\\nDNA is a long structured molecule formed out of nucleotides arranged in a linear sequence. DNA is double-stranded which means each nucleotide has a complementary pair, together called a base pair. DNA molecule might be several million base pairs (Mb) long and serves as the storage and the means of the utilization of genetic information. The information content of DNA is equivalent if read in forward and reverse direction, thus all local properties of its sequence should be independent of the selected direction. To use this property of DNA molecule, we selected bidirectional LSTM RNN architecture \\n\\ne) Corrected the x-axis label.\\n\\nf) We have expanded the discussion about the models choice and the advantage of using Neural networks. The NNs keep the information about the local processes. While the linear regression model looks only at one data point. The LSTM keeps in memory information about the bins that are surrounding the central bin. In comparison to other models for RNNs the sequential order models meters mostly.\\n\\ng) We have extended the description of the evaluation of the experiments.\"}", "{\"title\": \"Response to Blind Review #3\", \"comment\": \"Thank you very much for your thoughtful review! Comments below:\\n\\nIn the current work, we define an architecture of a machine learning algorithm that will effectively solve the problem of prediction of DNA structure (TADs) formation as assessed by gamma transitional from the binding of well-studied factors of chromatin in Drosophila. \\nDue to biological reasons, we do not expect the underlying mechanism of TADs formation to be the same for other cell types and organisms, thus we didn\\u2019t test our trained models on other Hi-C datasets. Due to the fundamentally different nature of experiments, we also didn\\u2019t try to test our models on ATAC-Seq data. Of note, ATAC-Seq contains information about DNA binding of all the factors and does not allow an explanation of what factors are explanatory for the structure formation. However, these ideas are highlighting new directions for further research.\\n\\nThe value 11 comes from the distribution of the target value Transitional Gamma. Thus, in order to give higher value to the smaller true values, we divide the difference between the maximum value and the true value of the bin by the maximum value. In order to avoid division and multiplication by 0 instead of the maximum value we max_value + 1. The updated description of the weighted MSE function can be found in the new version of the paper.\\n\\nIn order to show the advantage of using the LSTM model, we added the results of more experiments for a broader set of evaluation metrics.\"}", "{\"title\": \"Response to Blind Review #2\", \"comment\": \"Thank you very much for your detailed review! Comments below:\\n\\n1. The main focused of this work is to predict the information that characterizes the 3D chromatin structure instead of the HI-C full reconstruction. We do not use the Hi-C map as input to our models, on the other hand, we are interested in exploring what other biological experiments can bring insides into the formation of chromatin folding. No other full published work was performed to predict Topologically Associated Domain characteristics from ChIP-seq data, to the best of our knowledge. \\n\\nWe have expanded the literature review and the description of the differences between the published papers.\\n\\n2. DNA molecule is folded in a complex 3D architecture that contains local regions of higher number of contacts called Topologically Associating Domains (TADs), that affect the regions from 20 000 to 200 000 DNA base pairs (20 Kb and 200 Kb for brevity, according to the naming conventions). We sought for a measure that will represent the presence of TADs in the genome from Hi-C input data. Gamma transitional is a good candidate for that, although the process of its calculation is a complicated procedure. \\n\\nThe input features of our model were the ChIP-Seq properties for sets of chromatin marks. \\nThe resolution of the Hi-C map that we use is 20 000 basepairs (20 Kb). This required the ChIP-Seq features to be calculated as the integrated value from the available ChIP-Seq signal for the corresponding 20 Kb. Bins from the end of the chromosomes were taken together with parts of the ends of the next chromosome. Each feature is a float number that characterizes each DNA bin. Thus, the features were represented directly after the pre-processing that is described in section 3.1. This data was not binarized.\\n\\n3. DNA is a long structured molecule formed out of nucleotides arranged in a linear sequence. \\nDNA is double-stranded meaning each nucleotide has a complementary pair, together called a base pair. DNA molecule might be several million base pairs (Mb) long and serves as the storage and the means of the utilization of genetic information. The information content of DNA is equivalent if read in forward and reverse direction, thus all local properties of its sequence should be independent of the selected direction.\\nThe input of the models are features that characterize the DNA segments which encourages us to us DNA sequentiality properties. Regions of the DNA that are physically close to each other are more likely to have similar gamma transitional scores. \\n\\nWe selected bidirectional LSTM RNN to use this property of DNA molecule in the architecture itself. Similarly to the usage of recurrent neural networks for text application where the sequential order matters. \\n\\n4. For all the models the wMSE was used as training loss, which is important for the comparison of the models.\\n\\n5. Currently, we perform the features importance extraction using features drop out. Correlating the activations of LSTM units with input features and also analyze if importances depend on the position of features is a good direction for further research.\\n\\n6. Section 4.3 Methods contain information about the hyper-parameters of the biLSTM that we evaluated (page 4). For instance, number of LSTM Units, number of training Epochs, the sequence length of input RNN objects of consecutive DNA bins.\\nMain results of the experiment are shown on Figure 5. \\n\\n7. The wMSE was chosen as it is a straightforward approach for the weighting of the samples. \\nHowever, we have calculated the values of the metrics to compare the results and added them to the paper.\\n\\n8. DNA from the left and the right to the central bin could potentially have the same level of importance for predicting the characteristic of the 3D structure of the chromatin. The Bidirectional LSTM kept information from both sides of the central bin. One output value is produced for each set of DNA bing and no hidden states are ignored, all the stated features contribute to the final prediction.\", \"as_a_result\": \"i) We have extended the literature review section and added the suggested papers. And pointing out the difference between the paper\\u2019s focus.\\nii) We have expanded the description of the data and methods to increase the clarity of the work.\\niii) We have also added a broader set of evaluation matrices to prove the performance quality of our research.\"}", "{\"title\": \"Thank you for the reviews and updates based on the comments\", \"comment\": \"Thank you very much for your thoughtful review! We have revised the paper according to the suggestions and would like to answer the reviewer\\u2019s questions as follows:\\n\\n1. Our work is focused on predicting the information that characterizes the 3D chromatin structure. We do not use the Hi-C map as input to our models, on the other hand, we are interested in exploring what other biological experiments can bring insides into the formation of chromatin folding. \\nTo the best of our knowledge, no other full published work predict Topologically Associated Domain characteristic Transitional Gamma from ChIP-seq data.\\n\\n2. We have expanded the literature review and the description of the differences between the published papers.\\n\\n3. The information content of DNA molecule is equivalent if read in forward and reverse direction, thus all local properties of its sequence should be independent of the selected direction. We sought for a model that will use this property of the DNA and thus selected bidirectional LSTM RNN architecture. \\n\\n4. All of the models were trained using the custom weighed MSE metric as loss function, to present a fair comparison of the results. \\n\\n5. We have added a table with a broader list of models and calculated additional evaluation matrices as was suggested by the reviewers. Now you can find the MSE, MAE, R^2 and weighted MSE score for the training and test sets of different types of the explored models. \\n\\n6. The best results were achieved using the bidirectional LSTM RNN model. It proves that the utilization of the linear sequence properties of the DNA brings valuable improvement.\\n\\n7. Moreover, there was no other baseline model ever provided for this biologically meaningful dataset. As a result, now we have a set of models with full evaluation scores. \\n\\nWe have made some changes to our manuscript according to the suggestions, expanded the description of the models as well as the data and added a set of evaluations.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper predicts DNA folding using different machine learning methods. The authors show that LSTM out performs other methods. They attribute this to the memory of sequential DNA and LSTM model structure. They also propose a weighted mean squared error that improves the performance of the proposed model.\\n\\nThe authors compare the LSTM model with other classical approaches showing better performance based on predictive and quality metrics, applied to Hi-C data for drosophila, for predicting TADs.\\n\\nMy major concern is that it is not clear if the improvement is a by-product of LSTM without the proposed new metric. A fair comparison would also consider similar loss function designs for other approaches or at least comparing to a vanilla LSTM model. \\n\\nAlso the approach lacks illustration of generalizability. The definition of the loss function is also very specific (why 11?) and I wonder if this is generalizable to other Hi-C datasets or predictions based on other epigenetic features beyond ChIP-seq, e.g. ATAC-seq.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors utilise DNA spatial structures called Topologically-associative domains (TADs) from Hi-C data (of the Drosophila fly) and epigenetic marks (such as binding factors) from Chiq-seq data, thereby making use of physical proximity, to predict DNA folding patterns. The authors use a bidirectional LSTM RNN further emphasising that memory of the DNA states contributes to chromatin folding structures.\\n\\nThe paper offers for a good read. \\n\\nBelow are comments for improvement and clarification.\\n\\na) There is only one equation in the paper and this is also not given clearly. What is the K being summed over? \\n\\nb) wMSE is an old concept and depending on the objective function, the equation can vary. Therefore, change the sentence to read that the authors have a \\u2018modified' wMSE instead. \\n\\nc) Section 3.2, Page 3 last line: what is [5]?\\n\\nd) Section 4.3: 3rd paragraph, last line. Consider formalising the sentence, it will make it more readable. \\n\\ne) Figure 5: Right panel (top and bottom): correct the x-axis label\\n\\nf) There is no discussion explaining why regression is better or worse than a neural network, in this application setting. \\n\\ng) Figure 6 and associated experiment is very interesting and important. The authors should elaborate why, in certain cases, there were huge negative errors in training whereas the test error was positive.\"}", "{\"comment\": \"Dear Maria Anna Rapomaniki, thank you for your feedback.\\nIn our paper, we sought to base our literature review on indexed full-size papers published in journals or main track of conferences. We are aware of the works you mention but we decided not to include them in our review also because they were presented at workshops, not the main tracks. The authors of these works might be expected to develop their work further. For instance, the work https://ieeexplore.ieee.org/document/8621486 was presented at the Analysis and modeling of the three-dimensional structure of chromatin Workshop http://mccmb.belozersky.msu.ru/2018/chromatin.html. Moreover, this paper is a short abstract and it does not contain a detailed description of what exactly was done.\", \"as_for_the_difference_between_the_work_https\": \"//arxiv.org/abs/1811.09619, we predict the folding pattern characteristic obtained directly from the Hi-C map, but not the 3D structure using ChIP-seq data. This is the major reason why we didn't cite this paper in the first place. However, this work looks promising and worth mentioning, thus we will be happy to cite this work in the next iteration of paper improvement if supported by reviewers.\\nKind regards.\", \"title\": \"Addressing the related work comment\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors compared a bidirectional LSTM to traditional machine learning models for genomic contact prediction. Although the paper presents an interesting application of ML, I vote for rejection since i) the paper is similar to previously published works and lacks methodological novelty, ii) the description about the data and methods is not written clearly enough, and iii) the evaluation needs to be strengthened by additional baseline models and evaluation metrics.\\n\\nMajor comments\\n=============\\n1. As pointed out by Maria Anna Rapsomaniki, the paper is similar to previously published papers, which are not cited.\\n\\n2. It is unclear which features were used as inputs. How were the features that are described in section 3.1 represented? Are these binary values or the mean of the Hi-C signal over the genomic region? What does 20kb mean? How were genomic segments of different chromosomes treated?\\n\\n3. The motivation of using a biLSTM in section 4.3 is unclear. What has the choice of a biLSTM to do with the fact that DNA is double stranded? The DNA is not used as input to the model and genomic contacts can be formed between non-adjacent segments. Please compare recurrent models to non-recurrent models such as fully connected or convolutional networks.\\n\\n4. The authors used the weighted MSE as evaluation metric, which was used as training loss of the biLSTM, but it is unclear if the same loss was used for linear and gradient boosting regression. To understand if the performance differences are due to the loss function or model architecture, the authors should use the same loss function for training all models, and use additional evaluation metrics such as the MSE and R^2 score.\\n\\n5. The authors used linear regression weights to quantify feature importance (figure 4). It is unclear if the biLSTM assigned the same importance to features. To quantify the importance of features learned by the biLSTM, the authors could consider correlating the activations of LSTM units with input features, and also analyze if importances depend on the position of features.\\n\\n6. Which hyper-parameters of the biLSTM and baseline method did the authors tune and how?\\n\\n7. The authors should compare the described weighted MSE to the mean squared logarithmic error loss, which is commonly used for penalizing large errors less.\\n\\n8. Why did the authors use the activation of the center LSTM hidden state instead of concatenating the last hidden state of the forward and reverse LSTM? By using the center hidden state, half of the forward and reverse LSTM activations are ignored.\"}", "{\"comment\": \"Although the topic is very interesting, there are some issues with its novelty. We noticed that the proposed article is very similar the work published in 2018 at the IEEE International Conference on Bioinformatics and Biomedicine (BIBM) which is indexed: https://ieeexplore.ieee.org/document/8621486. The authors should cite this article and clearly state what are the improvements. Moreover, the idea of predicting genome folding via an LSTM model has been explored even before the BIBM conference here: https://arxiv.org/abs/1811.09619 and also presented at the NeurIPS 2018 Workshop: http://www.quantum-machine.org/workshops/nips2018/. Since the two models are strikingly similar, the authors should again cite this very relevant work and provide a discussion over the improvements. Thanks!\", \"title\": \"Already published and related work\"}" ] }

[ "Diego Klabjan", "Baiyang Wang" ]

[ "GAN", "NLP", "embeddings" ]

{ "note_id": [ "SPoGClV2DZ", "H1eZpcn45H", "H1gUv3cg5r", "S1eUbIW3YH" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752170, 1572289208746, 1572019293972, 1571718653528 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2561/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2561/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2561/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The general consensus amongst the reviewers is that this paper is not quite ready for publication. The reviewers raised several issues with your paper, which I hope will help you as you work towards finding a home for this work.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper proposes to use Generative Adversarial Networks (GANs) in the context of natural language processing and introduces two models for generating document embeddings. The first model, weGAN, aggregates multiple sets of single-corpus word2vec embeddings into one set of cross-corpus word representations; document embeddings are a function of these updated word embeddings. The second model, deGAN, side-steps word-level embeddings and directly generates document-level representations. For both models, the real examples come from word2vec and tf-idf, while the artificial examples are the ones generated by the network. The authors show that their document embeddings are better than the word2vec/tf-idf baseline at clustering documents according to the corpus they originate from.\\n\\nWhile the context in which GANs are used is novel and creative, the reasons for rejection outweigh the positives. The main reason is the fact that the paper is outdated by two/three years. There is no mention of the shift towards contextual embeddings that has been happening in the last few years (InferSent, ELMo, BERT) and that became ubiquitous in NLP. Instead, all comparisons are against word2vec (published in 2013) that is no longer a relevant baseline, even for non-contextual word embeddings. In fact, the newest cited papers are from 2017. The Github link in the paper was last updated on 23rd December 2017. My objection is not that this work was started a long time ago, but that it has not been updated in years. In the lack of proper comparison against more novel techniques, it is hard to estimate whether the reported gains over word2vec are meaningful.\\n\\nAnother argument for rejection is the fact that the experimental section does not make a convincing case that the positive results matter beyond the very specific datasets and tasks selected.\\n\\nBefore commenting in more detail on the selection of datasets, tasks and metrics, I want to document my understanding of the experimental setup; I found it non-trivial to understand what constitutes a \\\"single corpus\\\" in the experiments. My mental model is the following: the paper operates on four datasets (CNN, TIME, 20 Newsgroups, and Reuters-21578). These are four different experiments, and the four datasets never interact. The cross-corpus property of the model is exercised within each of these datasets, which have a natural grouping of topics; for instance, CNN is divided into \\\"US\\\", \\\"world\\\", and \\\"politics\\\". In other words, a cross-corpus CNN model is trained on three corpora: US, world, and politics. With this mental model of the experimental setup, the experiments raise a series of questions:\\n\\n1) Do the selected datasets / classification tasks matter in the grad scheme of NLP?\\nThe end-task is to classify the corpus of origin for each document (for instance, for the CNN experiment, the model has to decide between US, world, and politics). If I understand correctly, the weGAN model jointly trained its word embeddings with this objective. As a final step, the classifier was additionally fine-tuned on this task. So it is not very surprising that weGAN is better at clustering documents than word2vec, since it was explicitly trained to do so. It would have been more informative to see weGAN performing better on a task that was solely fine-tuned on (e.g. some sentiment classification task that was not part of the training objective when embeddings were trained). In that case, one could argue that the enhanced embeddings can be used as an out-of-the box tool for a lot of different tasks. Another option would be evaluating on the same document membership task, but on datasets that were not seen during training.\\n\\nAdditionally, the fact that the authors evaluated on *new* datasets with a *new* task distances them even further from any potential comparisons against prior work. Even for the two already-existing datasets with results included in the appendix, there are no references to previous SoTA.\\n\\n2) Are the reported positive results meaningful?\\nAnother problem with not showing results from previous work is that it is hard to understand whether the gains over the word2vec baseline are meaningful. For instance, the classification accuracy for the CNN corpus is increased from 92.05% (word2vec) to 92.29% (deGAN) and 92.36% (weGAN). The authors show that this is *statistically significant*, but is it meaningful? Is a 0.24% or 0.31% increase in accuracy worth the increased complexity of the model? For reference, what ratio of gains/computational cost would BERT incur?\\n\\n3) What is the message of qualitative evaluation?\\nThe authors offer qualitative analysis of their model. For instance, Table 2 shows the top 10 synonyms for words \\\"Obama\\\", \\\"Trump\\\" and \\\"US\\\", as produced by word2vec and weGAN. The lists of synonyms are extremely similar and don't seem to convey much more information besides the fact that weGAN makes minimal changes. I find the particular examples that the authors choose to underline the superiority of their algorithm unconvincing. For instance, the fact that, in the list of synonyms for US, the word British (selected by word2vec as the 9th closest word) is turned into American (on the same 9th position) loses its significance when you also consider the fact that weGAN did not derank any of the arguably less similar terms from higher positions (e.g. Turkish remains the 6th most similar term to US).\", \"other_nit_picks\": [\"Formatting issues: In Sections 3.1 and 3.2, the numerical references to equations are broken. Figures 3 and 4 cover the text above. Tables 2, 4, 6, 7 are hard to read. It almost feels like the paper was updated from one conference format to another, without the necessary Latex adjustments.\", \"The paper has no \\\"Conclusion\\\" section\", \"Improper use of singular / plural of \\\"corpus\\\" / \\\"corpora\\\"\", \"To end on a more positive note, here are some reasons why I did appreciate the paper:\", \"The usage of GANs in this context is very creative; the models do not generate *raw* text, but rather intermediate representations; this might be a promising direction for using GANs in NLP.\", \"The description of the models in sections 3.1 and 3.2 is rigorous, with useful explanations that break down the reasoning behind the loss function choices.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": [\"Summary\", \"The paper proposes extensions of Generative Adversarial Networks to modeling multiple text corpora. Concretely, the paper looks at two problems: 1) given independently pretrained word embeddings from K corpora, finding a common word embdding, 2) extracting document representations from a discriminator of a GAN trained to generate tf-idf vectors. Preliminary experiments show that the proposed approaches outperform baseline classifiers trained with word2vec.\", \"Strengths\", \"The paper clearly mentions all the experimental details\", \"The paper has a nice set of qualitative examples that probe what the proposed model is learning\", \"Weaknesses\", \"It is not clear what problem the paper is trying to solve. Is the goal to get better document representations, in which case why do we think that using a GAN is a good idea. Learning a generative model to use representations for a downstream task seems like a pretty roundabout way of doing things (if we dont care about generating anything in the first place). Further, if better document representations are desired then the paper should compare to works like Bert (Devlin et.al. [1]).\", \"Sec. 3.1: Calling the proposed weGAN model a GAN seems a bit inconsistent/ wrong. The proposed model is not a generative adversarial network, there is no sampling of a noise or a notion of a generative distribution. The real data, similar to the image case is a bunch of samples from the data distribution (where the stochasticity comes not from the word embeddings but from the tf-idf of thte document), but the generator also uses the tf-idf representation, so essentially both the generator as well as the real world data use the document as an input. Hence it feels like a somewhat odd model formulation, which would be nice to clarify / explain.\", \"Further, it is not clear why one would want to train \\u201ccross-corpus\\u201d embeddings in the way describeed in Sec. 3. 1. Why not just train word2vec on the union of all the corpora (instead of per-corpus) and use it as the word representation? What are the conditions in which one would not want to do the common word representation?\", \"Sec. 3.2: Why not formulate modeling multiple corpora in the spirit of CoGAN (Liu and Tuzel, 2016), by getting M discriminators (potentially with parameter sharing) to solve binary classification tasks as opposed to the 2M way classification problem? Atleast a comparison to an approach like this seems warranted. Further, it is not clear why one would want to generate tf-idf vectors. It seems like all the experiments are around representation learning and classification as opposed to generating or evaluating tf-idf vectors for documents.\", \"As a side note, for the experiments concerning deGAN vs the baseline, it would be nice to check what happens when the last classification layer is not initialized with LDA parameters. Since that is an additional source of information.\", \"References\", \"[1]: Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. \\u201cBERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding.\\u201d arXiv [cs.CL]. arXiv. http://arxiv.org/abs/1810.04805.\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": [\"This work proposes two models: (1) A model to learn word embeddings from different corpus. (2) A model to generate robust bag-of-words document embeddings. This topic may not be novel enough, considering the current development in GAN and domain adaptation. Some questions:\", \"For the word embedding model, why it is called GAN? It seems that nothing is generated.\", \"Further, for the word embedding model, the notation is over-complicated. This model is very similar to the domain adversarial network (DANN). This is my understanding and please correct me if I am inaccurate: A classifier is trained to distinguish the domains/corpus, while the word encoder tries to learn shared features across domains/corpus. The sub- and super-scripts are confusing.\", \"For the deGAN, only generating bag-of-words may not be that attracting. Further, the idea is also similar to many existing works in domain adaptation. Each domain/corpus has unique components, while all domains/corpus also share similar features. The final output is just a combination of the shared components and the unique features. This idea is straightforward. It would be much better if the author can generate natural language, instead of just bag-of-words.\", \"Please fix the formatting issues. For instance, Figure 3 and Figure 4 cover the text. \\u2018To explain (10), first we consider\\u2026\\u2019. I think here should be (6).\", \"More baseline competitors are preferred.\"]}" ] }

[ "Zhuozhuo Tu", "Fengxiang He", "Dacheng Tao" ]

[ "generalization", "recurrent neural networks", "learning theory" ]

{ "note_id": [ "rEmwTBSm2J", "SyeQvOGKiS", "Hygpk4fFoH", "H1lNcs-Kjr", "Byl1NFbKjr", "H1lnk5Sp5H", "rJeKFZ8aFH", "rJxdMvq3KB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752139, 1573623898923, 1573622756993, 1573620620475, 1573620007407, 1572850148258, 1571803521258, 1571755792178 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2560/Authors" ], [ "ICLR.cc/2020/Conference/Paper2560/Authors" ], [ "ICLR.cc/2020/Conference/Paper2560/Authors" ], [ "ICLR.cc/2020/Conference/Paper2560/Authors" ], [ "ICLR.cc/2020/Conference/Paper2560/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2560/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2560/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper presents a generalization bound for RNNs based on matrix-1 norm and Fisher-Rao norm. As the initial bound relies on non-signularity of input covariance, which may not always hold in practice, the authors present additional analysis by noise injection to ensure covariance is positive definite. Through the resulted bound, the paper discusses how weight decay and gradient clipping in the training can help generalization. There were some concerns raised by reviewers, including rigorous report of the experiment results, claims on generalization in IMDB experiment, claims of no explicit dependence on the size of networks, and the relationship of small eigenvalues in input covariance to high frequency features. The authors responded to these and also revised their draft to address most of these concerns (in particular, authors added a new section in the appendix that includes additional experimental results). Reviewers were mainly satisfied with the responses and the revision, and they all recommend accept.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to all reviewers\", \"comment\": \"We appreciate the insightful and constructive comments by all reviewers. We have revised the manuscript as suggested by the reviewers, and summarize the major changes as follows:\\n\\n> Update all misleading sentences pointed out by the reviewers.\\n\\n> Replot Figure 1 with error bars.\\n\\n> Add a new Appendix D which includes additional experimental results.\\n\\nApart from the above changes, we have also fixed some typos. Please see answers to individual reviewer below, for particular comments. We thank the reviewers again for their effort in reviewing the manuscript.\"}", "{\"title\": \"Response to Reviewer #4\", \"comment\": \"Thank you for the detailed reviews. We have updated the experimental section according to your suggestions and answer specific questions below.\\n\\n> I am confused by the following sentence: \\u201cgeneralization errors \\u2026 for different combinations of $L$ and $\\\\sigma_\\\\epsilon$ are shown in Figure 1.\\u201d\\n\\nWe apologize for the confusion caused. We have changed the sentence to \\u201c\\u2026 for $L=100$ and different values of $\\\\sigma_\\\\epsilon$ is shown in Figure 1 (results for other values of $L$ in Appendix D) \\u201d and added more experiments for various values of $L$ in new Appendix D.\\n\\n> The experimental results should be reported as a scatter plot, preferably with error bars.\\n\\nThanks. We have replotted the empirical results with error bars in Figure 1.\\n\\n> If space limitations prevent reporting the experimental results more rigorously, I would prefer to see the experiment results reported in an appendix.\\n\\nThanks. We have added a new Appendix D which includes additional experimental results.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We thank the reviewer for the comments. We have modified our manuscript to address your question.\\n\\n> In the experiments on the IMDB dataset, the authors claim that the generalization error is worst at $\\\\sigma_{\\\\epsilon} = 0$, but it appears that the error is actually larger for $\\\\sigma_{\\\\epsilon} = 0.4$? \\n\\nThanks for pointing out this. We have changed the sentences to \\u201cWe observe that as we start injecting noise, the generalization error becomes better. But when the deviation of noise keeps growing, the generalization error shows an increasing tendency\\u201d.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We thank the reviewer for the comments and would like to address the reviewer\\u2019s concerns as follows.\\n\\n> I find that the bounds still depend on the size of the network but implicitly. Then what is the value of this contribution?\\n\\nThanks. Our bound has no explicit dependence on the size of networks. But the norm itself used in our bound may still relate to the network size implicitly as the reviewer might concern. This is why we claim \\u201cexplicit\\u201d in the paper. In fact, obtaining a generalization bound which is completely independent of the size of networks is an important future direction since such a result might answer a long standing question of why deep neural networks generalize so well in practice. Despite that it is still a long way to go, we already take a step further and give a bound without any explicit dependence on the network size by using a new proof technique in contrast to existing results which either have a factor of max{d,m,k} (Zhang et al., 2018) or the total number of parameters of the network (Chen et al., 2019) (see Comparison with existing results paragraph in Section 3.2). This new technique can potentially be applied to other neural networks architectures such as convolutional neural networks, which might be of independent interest. Besides, from our empirical results, the implicit dependency of norm on the network size might be much weaker than what we expect. For example, in our experiments, the matrix 1-norm of the weight matrix after training is around 2, which is much smaller than the size of matrix 128.\\n\\n> Why smaller eigenvalues usually contribute to high frequency components of the input signal and what the frequency components of the input signal is. What is the formulation of the high frequency components of the input signal?\\n\\nThanks. Formally, the frequency component of an input signal can be defined via Fourier transform as $X(w) = \\\\int_{-\\\\infty}^{\\\\infty} x(t) \\\\cdot e^{-2 j \\\\pi w t}dt$ where w stands for the frequency, x(t) is original input signal in time domain and X denotes the signal in frequency domain, and a signal x(t) has a high frequency component if X(w) has a relatively large value at a high frequency w. We refer the reviewer to [1] for more details. \\n\\nFor eigenvalues, it corresponds to the variance explained by its eigenvectors. Not strictly speaking, the smaller eigenvalues contribute less to the energy in the data which usually correspond to the high frequency information, i.e., the details of an image. For example, in the application of Singular Value Decomposition to noise reduction in images, SVD can remove the noise and maintain most of the information in the original image by discarding those small singular values. Since most of the noise elements contribute to high frequency components of the image, we can think that the smaller eigenvalues correspond to the high frequency components and SVD acts like a low-pass filter. \\n\\nTheir relationship might also be understood from a perspective of basis transformation. In short, both the eigenvectors and the Fourier transform can be seen as a change of basis. If we consider the eigenbasis of the space formed by all eigenvectors, we can show that the eigenbasis vectors corresponding to smaller eigenvalue are more sensitive to small changes in the data, i.e., a large condition number (Section 4.4, [3]). The same applies to the Hilbert space basis vectors $e^{j t w}$ with a high frequency w under the Fourier transform (Chapter 4, [2]). In this sense, smaller eigenvalues correspond to high frequency. But we have to mention that the true relationship between them might be much more complicated than what we claim in the paper. Therefore, for precision, we have replaced the word \\u201cusually\\u201d with \\u201cmay\\u201d in the paper. We apologize for the misleading information and thank the reviewer again for pointing out this.\\n\\n[1]. R. Polikar, The Wavelet Tutorial. [Online]. Available: http://users.rowan.edu/~polikar/WTpart2.html\\n[2]. Bertero, Mario, and Patrizia Boccacci. Introduction to inverse problems in imaging. CRC press, 1998.\\n[3]. Workalemahu, Tsegaselassie. \\\"Singular value decomposition in image noise filtering and reconstruction.\\\" (2008).\\n\\n> I think the first extending work is to study the generalization bound of vanilla RNN with tangent or sigmoid activation. \\n\\nThanks for the suggestion. For the future work, we agree that it is valuable to first extend our results to vanilla RNNs with tangent or sigmoid activation function.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #4\", \"review\": \"This is an interesting and very well-written paper. I read the paper carefully but I don\\u2019t have sufficient expertise to determine whether all the proof steps are correct.\\n\\nThis paper builds on recent works giving generalization bounds on RNNs.\\n\\nThe primary contribution is that they are able to bound the generalization error for RNNs without a dependence on the network size parameters d and m.\", \"their_proof_is_roughly_as_follows\": \"1. They decompose the RNN + ReLU activation into the sum of a linear network and difference terms. This step is key because it lets you treat each term independently when estimating the Rademacher complexity.\\n2. The linear network term can be bounded directly with their Fisher-Rao norm.\\n3. They make a second decomposition: the difference term can be written as a sum of simpler terms\\n4. For the simpler terms, their Rademacher complexity can be bounded independently using matrix-1 norm. Using the matrix-1 norm instead of the spectral norm means their bounds won\\u2019t depend on the network size parameters.\\n5. Then they combine these bounds to give the Rademacher complexity bounds for RNNs.\\n6. Lastly, they combine the Rademacher complexity bound with Kuznetsov et al 2015\\u2019s multiclass margin bound to give the generalization bound.\\n\\nThe downside to their generalization bound is that it requires the covariance matrix of the input data must be positive definitive, and it explodes when the smallest eigenvalue is close to zero.\\n\\nTheir second contribution attempts to address these downsides. They prove another generalization bound for RNNs when training with random noise, which has the effect of increasing the term containing the smallest eigenvalue of the input covariance matrix.\", \"they_remark_on_several_empirical_phenomena_that_are_consistent_with_their_results\": [\"Correlation of features in the input data makes it harder for RNNs to generalize\", \"Weight decay could help by decreasing the relevant gradient terms in their bounds\", \"Gradient clipping could help when the smallest eigenvalue of the input covariance is very small\", \"Their third contribution is a single experiment. This contribution is fairly weak and the practical value of their theoretical work would be much more convincing if they were to put more effort into this section.\", \"They use IMDB data set (50k movie reviews + binary sentiment classification task)\", \"They add Gaussian noise to the input data with four different values.\", \"They plot the generalization error, which is the difference between the test error without noise and the training error with noise\"], \"specific_comments_to_improve_their_experiment_section\": [\"I\\u2019m confused by the following sentence: \\u201cgeneralization errors \\u2026 for different combinations of L and \\\\sigma_epsilon are shown in Figure 1.\\u201d However, in Figure 1 I only see different values of sigma. I don\\u2019t see anything about using various values for L. This should be clarified.\", \"Figure 1 draws linear interpolation between data points - there\\u2019s no evidence for those interpolations. It should be reported as a scatter plot, preferably with error bars.\", \"If space limitations prevent reporting the experiment results more rigorously, I would prefer to see the experiment results reported in an appendix, with a brief comment on their significance in the main paper.\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"In this paper, the authors investigate the topic of theoretical generalizability in recurrent networks. Specifically, they extend a generalization bound using matrix 1-norm and the Fisher-Rao norm, proving the effectiveness of adding noise to input data for generalizability. These bounds have no dependence on the size of networks being investigated. The authors also propose a technique for representing RNNs as a decomposition into a sum of linear networks with a difference term, which allows for easier estimation of Rademacher complexity. The authors claim this is a representation that can be extended to other neural network architectures such as convolutional networks.\\n\\nThis work has the potential to be of interest for the learning theory community on theoretical properties of recurrent neural networks.\", \"one_question_i_have_for_the_authors\": \"in the experiments on the IMDB dataset, the authors claim that the generalization error is worst at \\\\sigma_{\\\\epsilon} = 0, but it appears that the error is actually larger for \\\\sigma_{\\\\epsilon} = 0.4?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes a new generalization bound for vanilla RNN with ReLU activation in terms of matrix-1 norm and Fisher-Rao norm. This bound has no explicit dependence on the size of networks.\\n\\nI am actually not familiar with the generalization theorem on RNN. Nevertheless, according to the demonstration of the authors, I can understand the results of the theorems in this paper. I think the analysis on the property of the covariance matrix of the input data are valuable. However, I still have some concerns as follows.\\n\\n1.\\tIt is interesting if the bound can be independent on the size of networks. However, according to the bounds, I find that the bounds still depend on the size of the network but implicitly. Thus, I understand why the authors claim that the bound they provide has no \\u201cexplicit\\u201d dependence on the size of networks. Then what is the value of this contribution?\\n2.\\tIn Section 3.4.1, the authors state that \\u201cSince the smaller eigenvalues usually contribute to high frequency components of the input signal, our bound suggests that high frequency information is often more difficult to generalize, which is consistent with intuition.\\u201d. Can the authors provide more explanations on this since I do not understand why smaller eigenvalues usually contribute to high frequency components of the input signal and what the frequency components of the input signal is. What is the formulation of the high frequency components of the input signal?\\n3.\\tIt seems that ReLU activation is not widely used in RNN. Instead, tangent and sigmoid function are more prevalent. The authors mention in Conclusion that the extending the results to other variants of RNNs like LSTM and MGU might be an interesting topic for future research. I think the first extending work is to study the generalization bound of vanilla RNN with tangent or sigmoid activation.\"}" ] }

[ "Parsa Mahmoudieh", "Trevor Darrell", "Deepak Pathak" ]

[ "skills", "demonstration", "agent", "sub-task", "primitives", "robot learning", "manipulation" ]

{ "note_id": [ "XrNsBg-gYo", "B1eCKJo2or", "BJe6m0c2oH", "Bygjj6choH", "BJxlyIDAYB", "ryx5EZ5htB", "r1lG65CZKB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752106, 1573855110403, 1573854757090, 1573854627361, 1571874263909, 1571754290383, 1571052218067 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2558/Authors" ], [ "ICLR.cc/2020/Conference/Paper2558/Authors" ], [ "ICLR.cc/2020/Conference/Paper2558/Authors" ], [ "ICLR.cc/2020/Conference/Paper2558/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2558/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2558/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors present a multiple instance learning-based approach that uses weak supervison (of which skills appear in any given trajectory) to automatically segment a set of skills from demonstrations. The reviewers had significant concerns about the significance and performance of the method, as well as the metrics used for analysis. Most notably, neither the original paper nor the rebuttal provided a sufficient justification or fix for the lack of analysis beyond accuracy scores (as opposed to confusion matrices, precision/recall, etc), which leaves the contribution and claims of the paper unclear. Thus, I recommend rejection at this time.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to AnonReviewer3\", \"comment\": \"We thank the reviewer for the constructive feedback and are glad that the reviewer finds our problem statement proposal interesting and \\u201cof clear value\\u201d. We address the concerns in detail below.\", \"r3\": \"\\u201cHaving the per time-step labelling of the trajectory provides no indication that the data could be useful for learning reusable skills for downstream tasks later.\\u201d\\n=> In our setting, we assume that downstream tasks can be defined by a new sequence of skills not seen in the training data. Having a good segmentation model will lead to less noisy separation of the data into subsets where each skill is trained with independently. These trained skills are then re-usable for a new task given the sequence in which they should be used at test-time.\"}", "{\"title\": \"Response to AnonReviewer2\", \"comment\": \"We thank the reviewer for the constructive feedback and are glad that the reviewer considers our method a good application of the MIL approach. We address the concerns in detail below.\", \"r2\": \"\\u201cHow does this compare to a fully-supervised oracle method trained with per-timestep labels?\\u201d\\n=> A fully supervised oracle method should achieve very-high accuracy with the same amount of data given to the weakly supervised approach. We will add this.\"}", "{\"title\": \"Response to AnonReviewer1\", \"comment\": \"We thank the reviewer for the constructive feedback and are glad that the reviewer likes the idea of identifying skills in the way we proposed and considers our problem statement important to address. We address the concerns in detail below.\", \"r1\": \"\\u201cI think comparisons with unsupervised clustering methods would also be useful\\u201d\\n=> In our preliminary experiments, we couldn\\u2019t get clustering to work well because it is unclear in which space one should compare two images in a video (raw pixels doesn\\u2019t work well). We will add feature-based clustering comparisons in the final version.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper presents a weakly supervised method for segmentation of trajectories into sub-skills inspired by multi-instance learning (MIL) in image classification by Andrews et al. (2002). This is done via training a classifier to label each observation per time-step with the probability of skills corresponding to that observation. These predictions are then accumulated throughout the trajectory to compute the probability of the skill in that trajectory. There is only a trajectory level supervision provided which specifies which skills are present with no specification of the order in which they appear. They empirically show that their model can achieve decent skill level classification scores on multiple environments provided that there is a large variety of demonstrations provided.\\n\\nIn its current form, I would recommend this paper to be rejected because 1) the framing and motivation of the paper does not correspond to the results and experiments reported and hence seems misleading 2) the paper is limited in scope 3) further experiments and comparisons to relevant baselines are needed to support the claims made in the paper.\\n\\nThe problem that they are proposing is interesting and is of clear value, however, the paper falls short in addressing this problem. In particular, the paper is framed as a way to learn re-usable useful sub-skills that can help generalize to new situations in control. However, the method presented provides a per time-step labelling of each observation with the associated most likely sub-skill. Having the per time-step labelling of the trajectory provides no indication that the data could be useful for learning reusable skills for downstream tasks later. One very basic experiment could be to train a behaviour cloning (BC) agent on the observation-action pairs conditioned on the sub-skill (or a separate network per sub-skill) and show that the learned policies can be leveraged in solving the tasks presented. For instance, one can train a meta-controller that can switch between these learned sub-skills to successfully perform the task. Training such sub-skills from weakly supervised skill annotations has been successfully done by Shiarlis et al. (2018). It should be noted that in their setting, the annotations are ordered which simplifies the problem to some extent.\\n\\nIgnoring the motivation and focusing only on the problem setting addressed which is annotation of trajectories with skill labels, the experiments seem very restrictive to me. To my understanding there are at most 4-6 primitive skills present in each environment investigated. Looking at the videos linked, it feels like there is some overfitting to the trajectories provided as for example in the case of the video with the red bowl (Reach and Stir inside Cup), the classifier is predicting 'Stirring' from the first frames where there is not much information present in the scene and 'reach to object' could also be plausible for example. It would have been nice to see the logits of predictions per classes for these examples to understand the confidence of the model particularly when some of the skills could be equally likely (e.g. first few frames).\\n\\nI found the experimental setup unclear. Particularly, details regarding the task setup, how many demonstrations are needed per task/skill, architectural choices and hyper-parameters are lacking and makes experiments hard to follow and understand. I would have also liked to see more analysis regarding the segmented trajectories, particularly how consistent these predictions are through time. To my understanding there is nothing that keeps a skill annotation consistent over some period of time (e.g. the model could keep switching its prediction every time-step). This would be quite unsatisfactory if one would like to use this to actually segment sub-trajectories associated with a given skill that could be useful for training policies. They report applying Gaussian smoothing to filter out noisy predictions but there are no details provided on how this is tuned and how much this affects the quality of segmentations.\\n\\nOverall, the paper seems to me very limited in its scope and experimental results. The claims made throughout the paper are not supported empirically or theoretically. There is not enough evidence for me to assess the significance of the proposed method and know whether this is indeed useful in practice.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper tackles the problem of learning to label individual timesteps of sequential data, when given labels only for the sequence as a whole. The authors take an approach derived from the multiple-instance learning (MIL) literature that involves pooling the per-timestep predictions into a sequence-level prediction and to learn the per-timestep predictions without having explicit labels. They evaluate several pooling techniques and conclude that the log-sum-exp pooling approach is superior. The learned segmentations are used to train policies for multiple control skills, and these are used to solve hierarchical control tasks where the correct skill sequence is known.\\n\\nThis is a good application of the MIL approach. However, I have settled on a weak reject because in my view, the novelty and results are minor.\\n\\nThe main point of comparison is the log-sum-exp() pooling as compared to max() and neighborhood-max() pooling. However, if I understand correctly, the log-sum-exp() approach has been used successfully in several other domains including its original domain of semantic image segmentation. So I view the novelty of the approach to be fairly low.\\n\\nIn addition, although the superior pooling method (which already exists in the literature) does outperform the alternatives evaluated here, the results are somewhat underwhelming, at only ~35-60% validation accuracy. How does this compare to a fully-supervised oracle method trained with per-timestep labels?\\n\\nThe behavioral cloning results are also fairly underwhelming, and the experiments are not very clearly described. Am I correct in my understanding that the learned skills are composed to solve a task where the correct sequence of skills is known, but is longer than the training sequences? A success rate of 50% on this task seems rather low. How does this compare, as above, to a fully-supervised oracle baseline? Why is there no success rate reported for the CCNN baseline?\\n\\nI think this is a good application of weakly-supervised MIL, but I find the specific contributions to be lacking in novelty and impressiveness of results. There are several directions that I think could improve the work:\\n- oracle fully-supervised results, to indicate the gap between the fully- and weakly-supervised case\\n- more thorough baselines on the behavior task, such as Policy Sketches [1]\\n- perhaps the temporal aspect of the problem could be incorporated into the pooling approach more directly to produce a more novel algorithmic contribution\\n\\n[1] Andreas, Jacob, Dan Klein, and Sergey Levine. \\\"Modular multitask reinforcement learning with policy sketches.\\\" Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 2017.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper introduces a weakly supervised learning approach for trajectory segmentation, which relies on coarse labelling about the occurrence of a skill in a demonstration to segment trajectories. This is accomplished through a recurrent model predicting skill categories for each step in a trajectory, and a trajectory level loss function that penalises the probability of seeing a given skill in a trajectory.\\n\\nOverall, I like the idea of identifying skills in this manner, and think that this is an important problem to address. However, I have concerns about it's feasibility when a large number of skills are present. It seems that there are certain requirements of skill occurrences in datasets that need to be met if this approach is to be feasible. For example, consider the dataset of skill sequences:\\n\\n[111 222 333]\\n[444 222 333]\\n[222 333 555]\\n\\nIt seems that it will never be possible to learn to identify skills 2 and 3 from this dataset. This paper would be greatly strengthened if the minimum dataset requirements to learn all skills were enumerated, or some theoretical bounds provided around when this loose labelling could possibly be successful provided. The paper glosses over this point by suggesting that \\\"data\\\" makes this a non-issue, but the paper would be much stronger if these limitations were confronted and some bounds on the chances of meeting the requirements needed for learning provided. \\n\\nThe classification results seem extremely poor, and it is hard to assess these on the basis of accuracy alone. For example, in the Robosuit example, the test results could potentially have been obtained by simply making the same prediction throughout the test, and there is no indication that anything sensible was actually learned. Confusion matrices, or precision and recall metrics, are required to avoid this. At present it is impossible to know if these errors are due to noisy predictions, the dataset limitations described above, or simply a poor classifier.\\n\\nAlong these lines, qualitative results in the form of trajectory classification (say following the presentation conventions in \\n\\nBolanos, 15, https://arxiv.org/pdf/1505.01130.pdf \\nRanchod, 15, https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7353414&tag=1\\n\\nwould also help to address these concerns. \\n\\nA question regarding the dial jaco videos experiments, why does the classifier not predict a 5 when moving from 4 to 6? I would expect a very jumpy prediction here, but the prediction looks very smooth - is this a filtered result?\\n\\nFinally, baseline experiments are limited to other weakly supervised learning segmentation approaches, but I think comparisons with unsupervised clustering methods would also be useful.\\n\\nUnfortunately, due to the lack of evidence that the proposed approach is able to learn effectively, I am inclined to reject this paper. Clarifying the bounds, and presenting stronger evidence (beyond accuracy) would make this paper stronger.\"}" ] }

[ "Minghong Yao", "Liansheng Zhuang", "Houqiang Li", "Jian Yang", "Shafei Wang" ]

[ "word embedding", "natural language processing" ]

{ "note_id": [ "g-YhRcFdh", "ryg1XGKOcB", "rkxk7rD6tS", "Byg0bMQ5YS" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752074, 1572536855299, 1571808534616, 1571594758132 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2557/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2557/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2557/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper presents an approach to learning interpretable word embeddings. The reviewers put this in the lower half of the submissions. One reason seems to be the size of the training corpora used in the experiments, as well as the limited number of experiments; another that the claim of interpretability seems over-stated. There's also a lack of comparison to related work. I also think it would be interesting to move beyond the standard benchmarks - and either use word embeddings downstream or learn word embeddings for multiple languages [you should do this, regardless] and use Procrustes analysis or the like to learn a mapping: A good embedding algorithm should induce more linearly alignable embedding spaces.\", \"nb\": \"While the authors cite other work by these authors, [0] seems relevant, too. Other related work: [1-4].\\n\\n[0] https://www.aclweb.org/anthology/Q15-1016.pdf\\n[1] https://www.aclweb.org/anthology/Q16-1020.pdf\\n[2] https://www.aclweb.org/anthology/W19-4329.pdf\\n[3] https://www.aclweb.org/anthology/D17-1198/\\n[4]\\u00a0https://www.aclweb.org/anthology/D15-1183.pdf\", \"title\": \"Paper Decision\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper is about learning word embeddings. In contrast to previous work (word2vec's Skipgram) where a word is represented by a fixed word and context vector, here each word is represented by a von Mises-Fisher distribution and a context vector.\\nThe paper claims that the resulting embeddings are more interpretable, and that the inner product of two context vectors represents their point-wise mutual information.\", \"my_main_concerns_are_the_following\": [\"Representing words by distributions is not a novel idea; it was previously done by Brazinskas et al. (2017): Embedding words as distributions with a bayesian skip-gram model (BSG). They however used Gaussian distributions and not von Mises-Fisher. BSG is acknowledged in the paper, but only small scale comparisons are performed (15 million) while the BSG paper uses a lot larger data sets. There is therefore not a meaningful comparison to the most relevant previous work.\", \"Word similarity experiments are not enough to justify this approach. BSG at least showed the strength of using distributions to represent words by showing that different samples could constitute different word senses/meanings. There is no such analysis here.\", \"The claim that the resulting representations are more \\\"interpretable\\\" is not backed up by any evidence at all, even though the word \\\"interpretable\\\" is in the title and the list of contributions.\", \"Generally this paper could benefit from proof reading and editing.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper addresses the problem the problem in word embeddings where \\\"word-word\\\" or \\\"context-context\\\" inner products are relied upon in practice when the embeddings are usually only optimized for good properties of \\\"word-context\\\" inner products. The new approach to training word embeddings addresses interpretability in the sense of having a theoretically grounded interpretation for the good properties that the inner products should possess -- namely that they capture the PMI between words. By changing the embedding training procedure (to include a step that samples vectors from each word's distribution representation) this interpretation is made possible without problematic theoretical assumptions. Now trained for the purpose they are intended to be used for, the approach yields strong SOTA results on the non-small challenge datasets.\\n\\nThe word is well exceptionally well motivated (we should directly optimize for the properties we want!) and situated in the literature. The opposition of Arora 2016 (getting at \\\"interpretable\\\" word embeddings with a latent variable model) and Mimno & Thompson 2017 (problematic assumptions in skip-gram embeddings) is particularly convincing. The vMF distribution is underexamined as tool for analyzing high-dimensional semantic vector distributions, and it is an excellent fit for the purposes of this project.\\n\\nTo back up claims of theoretical interpretability, a derivation (building on Ma 2017) proceeds without problematic leap thanks to a property introduced by a carefully selected (if straightforward) regularization term. This reviewer did not read the appendix (not sure where this would be located -- first time reviewing with OpenReview here), but the intuition doesn't seem to be much different than Ma.\\n\\nTo back up claims of applied performance, appropriate experiments are conducted on multiple evaluation datasets. The results seem to be fairly interpreted.\\n\\nThis reviewer decides to accept this paper for its balance of theoretical and empirical contributions and for the role it might play in reducing dependence on mysticisms in word embeddings (relying on accidental / uncharacterized properties of previous embedding strategies).\", \"suggestions_for_improvement\": [\"Run a spell checker. It will catch a large number of problems that weren't ever big enough to hurt my appreciation of the paper.\", \"Consider the creation of an adversarial dataset (of ~3 words in ~3 contexts) where techniques that optimize for the wrong thing will catastrophically fail but the proposed approach succeeds.\", \"Write one or two more sentences about the fate of \\\\kappa_c -- is it ever updated or am I just missing something? Making a bigger point about leaving \\\\kappa un-optimized shows there is room for additional depth in this line of thinking. (If you start learning concentration parameters, check out the Kent distribution for an even more expressive distribution on the unit hypersphere: https://en.wikipedia.org/wiki/Kent_distribution)\", \"Figure out what happened to your appendix.\", \"Watch out that readers/reviewers of similar work may try to read the word \\\"interpretable\\\" in the title as a reference to the much broader topic of interpretability in ML related to explaining a model's behavior. Is there a synonym for \\\"interpretable\\\" that won't raise this false link?\"]}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: This paper proposed a variational word embedding method, by using vMF distribution as prior and adding an entropy regularization term.\", \"strengths\": \"[+] In the experiment parts, the authors describe the experiment settings in detail.\\n[+] Detailed descriptions for each equation is given.\", \"weaknesses\": \"[-] Template: It seems that the authors use the wrong template, which is not the template for ICLR2020.\\n[-] Appendix: Although the author mentioned 'appendix' many times, I cannot see the appendix.\\n[-] Motivation: The connection between motivation and proposed method seems weak. The authors argue 'interpretable' in their title and abstract, but their method and experiments do not show this point explicitly. \\n\\nQuestions\\n[.] The experiments seem a little weak. The vocabulary size is 10K and the corpus is not so big, and I wonder whether the performance of the proposed method will be better for large corpus and longer training time.\\n[.] For Equation 8, we should have a guarantee on the concentration of partition functions. Is it still true for vMF distribution? \\n[.] What is the advantage of vMF distribution?\"}" ] }

[ "Lucien K.L. Ng", "Sherman S.M. Chow", "Anna P.Y. Woo", "Donald P. H. Wong", "Yongjun Zhao" ]

[ "machine learning", "security", "privacy", "TEE", "trusted processors", "Intel SGX", "GPU", "high-performance" ]

{ "note_id": [ "ju25JPc_-a", "S1gAleIhiS", "rklXMaVnsH", "rygS-TNhsS", "HkekgaE3oB", "BklsAhN3sB", "H1x7p2Nnjr", "SJgFs242jH", "BkehxucWcH", "HkguQA3BYB", "rkgoPyQVFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752044, 1573834742344, 1573829898784, 1573829884552, 1573829863276, 1573829842725, 1573829818858, 1573829792792, 1572083699632, 1571307039988, 1571200866607 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2556/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2556/Authors" ], [ "ICLR.cc/2020/Conference/Paper2556/Authors" ], [ "ICLR.cc/2020/Conference/Paper2556/Authors" ], [ "ICLR.cc/2020/Conference/Paper2556/Authors" ], [ "ICLR.cc/2020/Conference/Paper2556/Authors" ], [ "ICLR.cc/2020/Conference/Paper2556/Authors" ], [ "ICLR.cc/2020/Conference/Paper2556/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2556/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2556/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a framework for privacy-preserving training of neural networks within a Trusted Execution Environment (TEE) such as Intel SGX. The reviewers found that this is a valuable research directions, but found that there were significant flaws in the experimental setup that need to be addressed. In particular, the paper does not run all the experiments in the same setup, which leads to the use of scaling factor in some cases. The reviewers found that this made it difficult to make sense of the results. The writing of this paper should be streamlined, along with the experiments before resubmission.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to author comments\", \"comment\": \"I thank the authors for the detailed rebuttal.\\nNevertheless, I think that many misconceptions remain, which I describe below:\\n\\n* On the LAN vs WAN setting:\\nWhile one issue in a WAN is bandwidth, the much bigger problem is *latency*. After every layer, the server hosting the SGX enclave has to send data to the server hosting the second GPU, and wait for the GPU's reply before moving on to the next layer. In a WAN, the roundtrip latency will be in the 5-150 millisecond range depending on how geographically close the servers are. Now, multiply this by the number of layers. Running the entire computation inside SGX will always be faster.\\n\\nYour rebuttal mentions the need of a good network line. Paying more money will indeed get you more bandwidth, but latency is fundamentally limited by the speed of light.\\n \\n* On non-collusion in a LAN:\\nThe real-world example you give with Facebook's password protections doesn't really relate to your case. Here, Facebook is trying to protect against an outsider adversary. But Facebook obviously has access to the contents of both servers, so this doesn't protect the data from Facebook.\\nThe challenge in Goten is to make sure that no single entity can see the data processed by the two GPUs. If these two GPUs both belong to the same company (e.g., Facebook), then you need to trust that company for privacy. But if you do trust this company for privacy, you can just have them run the entire computation in the clear.\\n \\n* On access control:\\nYou mention that one way to ensure non-collusion within an organization is to enforce access-control of the servers. Again, this doesn't actually provide any security from the organization itself, which is in charge of the access control. But even so, at this point you could also just run the whole computation in the clear in a single GPU (without SGX) and restrict access to this GPU using the same access control strategies. Using two servers and SGX doesn't buy you anything in this scenario.\\n \\n* On optimizing CaffeScone:\\nEven if optimizing this baseline is not the main goal of your paper, it is still necessary to make the point that Goten is practical. The question is basically: if a company wants to spend resources to deploy private ML, should they implement an optimized Goten, or an optimized CaffeScone? If the optimized CaffeScone achieves similar performance, they'll be more likely to use that as it is much simpler to deploy and doesn't require non-colluding servers.\"}", "{\"title\": \"6/6\", \"comment\": \"R1/R2. [Experiment] Simulation mode.\\n\\nPrograms in hardware (HW) mode has negligible overhead as long as no paging is triggered. Specifically, according to the experimental results in Privado [Tople et al., 2018], the neural networks which do not trigger page-fault do not have any performance overhead.\\n\\n[Tople et al., 2018] Tople, S., Grover, K., Shinde, S., Bhagwan, R., & Ramjee, R. (2018). Privado: Practical and secure DNN inference. arXiv preprint arXiv:1810.00602.\\n\\nMoreover, based on our experimental data of non-linear layers, namely, the data for computing scaling factors, the performance of programs in simulation mode is similar to that in HW mode when no page-fault is triggered. Indeed, our SGX-aware chunked operations can totally prevent page-fault in enclaves, and thus the performance overhead due to HW mode is negligible in our case. Furthermore, the values of scaling factors support our claim. The scaling factors are the ratio of 1) the running time of a kind of non-linear layers running in HW mode on a machine equipped with SGX to 2) that in simulation mode on a Google Cloud VM. According to our experiment results, all these values are less than 1, meaning that the running time of the simulation mode on Google Cloud VM overestimates the running time of the full-fledged system.\\n\\nThat said, we plan to perform experiments for upper-bounding Goten\\u2019s running time. To address the concern on the simulation mode, we will run our experiments (of linear layers) on a machine equipped with the SGX module and a weak GPU. Then, we benchmark as a baseline timing the running time in a fine-grained manner, where we record the running time on the CPU, GPU, and the transfer time between CPU and GPU and the networks. To provide a more accurate estimation for the GPU part, we will record its running time and use it to replace the GPU running time in the baseline timing. This evaluation ensures that the running time on the CPU and SGX is authentic. However, such fine-grained timing degrades the measured perform (compared with the true performance) because we synchronize different computing units before any of them starts transferring data or computation. This does not fully utilize the computational resources and the network.\\n\\n\\nR1. [Technical Contributions] Quantization.\\nIt is true that most modern GPU are optimized for single-precision float operation, but single-precision floats are not enough for our secret sharing scheme since it only has 23 bits of significand precision for integer operations. Specifically, to achieve both high performance and high accuracy, 23 bits is not enough because integers with 23 bits will easily overflow after multiplication and a bunch of additions, while modulo operations are slow in GPU. Therefore, a sensible choice is to perform computation with double-precision floats, which have 53 bit of significand precision. We use a modern GPU which is also optimized for double-precision float operation.\\n\\nA high-level idea that we do not run into the overflow problem is that we have a very careful treatment of the available bits (as explained above and Section 3.3).\\n\\nR3. How does the inference performance of Goten compare to Slalom given the same privacy and correctness guarantees? This isn't clear from the experiments section.\\n\\nAs argued above (R1/R3), Slalom needs to pre-compute f(r) for computing online f(x). We achieve higher throughput for inference as ours is a real outsourcing solution which the pre-computation is not as expensive as the online computation for the real problem instance.\\n\\nA primary motivation of Goten is to provide a more comprehensive privacy coverage which was not known to be possible (by Slalom) before, namely, to further achieve data privacy (confidentiality) for the model's parameters. For such protection in training (and inference), Goten does introduce overhead (just like all the subtleties we discussed above). Hence, we found it natural that Slalom has better performance on *inference*, because its design inherently cannot support privacy for model parameter, and it is tailor-made for its weaker privacy guarantee confined to only inference. Simply put, again, it fails to support privacy-preserving training.\"}", "{\"title\": \"5/n\", \"comment\": \"R1. [Experiment] Optimizations should also be applied to CaffeScone.\\n\\nFirst, we provide CaffeSCONE as a baseline approach if one applies a generic solution (SCONE) for making use of SGX for training (not supported by Slalom). Further optimizing it is *not* our goal (for one, it does not use GPU at all). It requires significant engineering effort to implement all our optimization tricks (those without using GPU) to CaffeSCONE. We believe further optimizing it has its own impact as a system research work. \\n\\n(Until now, no results have applied any optimization on paging to any SGX-protected DNN framework -- a side-evidence that such an optimized framework is not as trivial as one might think, despite the potential impact in the system research community.)\\n\\nMore importantly, our experiments for comparing CaffeSCONE and Goten is done with *low* paging overhead because these experiments are done in the simulation mode (where programs can run as fast as if they are executed in single-thread normal execution environments). In other words, our results are *not* derived from an experiment which gives Goten unfair advantage. Yet, our experiments showed that Goten can still outperform CaffeSCONE by 2.5x to 18x in different linear layers. (Also see Figure 5(a).) \\n\\nHence, even in an ideal case where paging overhead is no longer a problem (either by applying our optimization techniques on both frameworks; or by using a non-existent, future version of SGX), Goten can still significantly outperform pure-SGX solutions, specifically, for privacy-preserving training.\\n\\nR1. [Experiment] Running the baseline [CaffeSCONE] and Goten in different CPUs and re-normalizing throughputs is convoluted and prone to mistakes.\\n\\nThe experiment setting is disadvantageous to us because CaffeScone ran in the CPU with a higher clock rate (4.2GHz of Intel i7-7700 vs. 2.0GHz of Google VM's), meaning our evaluation overestimates the running time of Goten. Moreover, we only re-normalize the non-linear layers, and thus this evaluation method does not affect the timing on the linear layers, whose performance improvement is our main contribution.\\n\\nThat said, we felt sorry that we employed different CPUs since we do not have easy access to production-scale hardware resources (GPU) with SGX on the same platform.\\n\\nR1. [Technical Contribution/Experiment] \\\"Fundamentally incorrect\\\"\\n\\nThe said statement focused on the code compilation and the protection of the algorithms; it is true that it does not use encryption, which we are glad to fix this \\\"incomplete\\\" statement (in Section 4.1).\\n\\nThat said, \\\"simulation mode will not be affected by the overhead of SGX's paging, as the memory is never encrypted\\\" does not apply here since we use simple exclusive-or (or one-time pad with pseudorandom sequence) which introduce no space overhead. Specifically, no matter if it is encrypted or not (i.e., no matter which mode we consider), such padding does not elongate the data to be processed. Hence the paging aspect is also not affected. (Also, see the discussion above and below.)\"}", "{\"title\": \"4/n\", \"comment\": \"R2. [Experiment] The connection between the evaluation (which mostly focuses on the speed benefits) and the claimed contributions is tenuous.\\n\\n(Section 4.2) First of all, we show the training throughput of CaffeSCONE in Fig. 3, by which we emphasize that using more cores on CPU cannot improve the performance of such a pure-SGX approach. Moreover, we benchmark the throughput with batch sizes of 128 (a common setting in plaintext setting) and 512 (the setting we adopted for Goten). We confirmed that the former one has better performance for VGG11 in CaffeScone, and thus we adopt it in the latter experiments. Note that we adopt batch size to 512 in Goten because with which Goten has better performance.\\n\\nTable 1 illustrates the speed up of Goten compared to CaffeSCONE in training phrase.\\n\\nWe first explain the experimental settings. Goten ran with simulation mode on Google VMs and employed memory-aware measures to reduce the overhead of paging. Moreover, we rescale the running time on non-linear layer, which bases on the running time with the real SGX setting, i.e., the hardware mode on the experimental machine equipped with Intel i7-7700. \\n\\nAccording to the experimental results on non-linear layers, program running with the real setting are faster than those on Google VMs. Hence, we believe that linear layers in the real setting are also faster as both kinds of layers have similar operation and (linear) access patterns.\\n\\nAccording to Table 1, in VGG11, Goten outperforms CaffeSCONE by about 8 times on linear layers and non-linear layers.\\n\\nFurthermore, Table 2 demonstrates how the performance speed up leads to higher convergence rate. The training methods for CaffeSCONE and Goten are different.\\nThe former adopts the most common approach, which uses plain single-precision floats. The latter one adopt employs the dynamic quantization scheme SWALP, explained in Section 3.3. Hence, it is natural to wonder whether Goten can attain a higher convergence rate. Our experimental result is confirmative. We record the converge trajectory of both training methods, which was captured in an unprotected setting on GPU, and then rescale the time axis according to the timing from Table 1. The results show that Goten can converge much faster.\\n\\nTo better emphasize our advantage on the convergence rate, Table 2 lists the speed up (at different levels of accuracy). We can attain 0.88 accuracy by about 7 times faster.\\n\\nAs our main contribution is the performance speed up on linear layers, we further isolate the performance gain of them and show it in Fig. 5. Moreover, because in Section XX we propose that the performance gain is portional to the arithmetic intensity, we hope to confirm it by more fine-grained benchmark. Hence, we record the performance gain with respect to the shapes of linear layers, which determines the arithmetic intensity. \\n\\nFig. 5 (a) is derived from the experiments that both CaffeScone and Goten ran on simulation mode and on Google VMs, and Goten does employ the chunked-operations. As we mentioned above, in simulation mode programs run as fast as those in single-thread normal environment, meaning that the paging overhead of SGX does not affect the performance of both of them.\\n\\nBelow we also provide a partial list of the figures and tables in our paper accompanied by the different settings.\\n[Figure 3: Training Throughput of CaffeSCONE]:\\nCaffeSCONE runs in HW mode\\n\\n[Figure 4: Accuracy Convergence in VGG11], [Table 2: Attaining accuracy using GPU-powered Scheme]. [Table 2: Attaining accuracy using GPU-powered Scheme]:\\nCaffeSCONE in HW mode. Goten in mixed mode (i.e., linear layers in sim mode and non-linear layers are rescaled to hw mode)\\n\\n[Figure 5 (a) With Low Paging Overhead] CaffeSCONE: Sim mode. Goten: Sim mode (without memory-aware measures)\\n\\n[Figure 5 (b) SGX Hardware Mode] CaffeSCONE: HW mode. Goten: Sim mode (with memory-aware measures)\"}", "{\"title\": \"3/n\", \"comment\": \"R1. [Technical Contributions] In the considered setup, the cloud provider (Google in this case), could just observe the communication between all servers, thereby breaking privacy.\\n\\nStrictly speaking, observing the communication alone is not enough to break privacy, but one needs to further look into internal memory. Of course, when the cloud provider controls all the servers, it can as well look into internal memory. This comes to a fine-grained treatment of the concept of \\\"trust\\\" beyond a boolean yes/no discussion. Thanks for raising this concern. \\n\\nIn response, we will make this threat model clear. In practice, abstracting some technical details, access control system can also help in terms of security. Assuming an insider attacker does not have the super-user privilege, and the audit log function of the access control system remains secure, it can then hold the entity with special access-permissions accountable when the access-rights are abused. We add this discussion to Appendix C.2.\\n\\nAgain, we took a simple treatment with our experiment, but this does not totally nullify the security of our system. As mentioned above, we will also enrich our experiment along this dimension.\\n\\nR1/R2. [Technical Contributions] Lack of computation integrity (a.k.a. correctness guarantee)\\n\\nIt is not our aim to provide integrity specifically, which we will acknowledge. Freivald's algorithm (IFIP Congress 1977) used by Slalom provides verifiability for an interactive outsourcing protocol of linear layers. Freivald's technique allows detection with sound error according to the space of the underlying finite field. This is a classical generic technique, which we can use as Slalom since we overate over finite field elements as well. We are happy to discuss more in the appendix.\\n\\nOur work focuses on privacy, yet, it does not mean we achieve that at the cost of integrity. In our case, privacy and integrity are orthogonal security goals and not inherently conflicting with each other.\\n\\n(In the \\\"secure outsourcing\\\" literature from the cryptography/security community, many major works, for example, searchable encryption, have been studied extensively only in the \\\"privacy only\\\" setting. There are also generic transformations that equip integrity to a searchable encryption scheme that does not ensure integrity.)\\n\\nIn practice, the servers may not have high incentive to jeopardize the training as they invested much cost to establish a system and gather the data to support training (which is different from the federated learning scenario or distributed learning scenario). We will also add this \\\"motivation/justification.\\\"\"}", "{\"title\": \"2/n\", \"comment\": \"R1. [Technical Contributions / Experiment] Ignoring the high network communication.\\n\\nOur experiment (despite being in the LAN setting) did not ignore the communication overhead among the servers. We took a simplistic view focusing on computation and accuracy. In the revision, we will provide experiments to shed light on how fast is \\\"enough,\\\" i.e., to estimate the effect of network speed of various WAN settings.\\n\\nIt is our premise to rely on communicating servers to achieve something which was not possible with the pure cryptographic approach, pure SGX approach, or \\\"simple\\\" SGX+GPU approach. In retrospect, Slalom achieves better performance than pure cryptographic approaches, by assuming the existence of a trusted processor (even though security researchers have demonstrated some form of side-channel attacks). In a sense, it further motivates research works (that have been) trying to make the assumption needed by Slalom more realistic.\\n\\nIn our case, we solved what Slalom failed to solve, i.e., privacy-preserving training. Further, it is efficient (and accurate) with a fast enough connection between the servers. Our suggestion is thus to use our system when the premise (again, we will quantify in our experiments) is satisfied. \\n\\nWe stress that we are not solving the problem by merely relying on such an assumption. Many challenges remain, as discussed in the paper and our responses (and some of them are mentioned by the reviews as well). For one, our outsourcing protocol minimizes (sequential) communication needed in the original protocol (see \\\"Parallelizable Pre-Processing without Communication\\\" in Section 3.2).\\n\\nPlease see below for further discussions on our experiment setting and how can we realize our assumptions placed on the servers in practice.\\n\\nR1. [Technical Contributions] In [LAN] setting, it is hard to argue that non-collusion is a valid security assumption as the cloud provider controls all servers\\n\\nWe believe that non-colluding servers do not *have to* be geographically located far apart. Instead, the crucial point is whether they are really *owned* by the same entity. Moreover, as the network bandwidth is increasing continuously, the communication overhead may not be a bottleneck in certain situations.\\n\\nEven if one indeed place both servers in the same LAN, there can be other justifications for the security. Let us justify with a real-world example. For user authentication of Facebook, they also maintain an internal server that is \\\"more secure\\\" than the public-facing server. E.g., see \\\"an adversary that compromises the web server and the password hashes it stores must still mount an online attack against the PRF service to compromise accounts,\\\" as described by https://eprint.iacr.org/2015/644.pdf. Simply put, security there is also \\\"strengthened\\\" for relying on the \\\"PRF secret key\\\" of another server.\\n\\nFor sure, when both servers are compromised, there is no security. It then comes to the classical discussion from Security 101 on how to fortify a selected machine. We got many possible ways, including but not limited to, i) using different hardware and software configuration from the other server (so a vulnerability in one platform will not lead to an easy compromise of both machines), ii) placing it within the network-level security parameter behind the DMZ, iii) limited access to the machine instead of public-facing, etc. We add this discussion to Appendix C.2.\\n\\nLet us elaborate more on how non-colluding servers assumption are usually realized in existing works (which is briefly mentioned at the end of \\\"Secure Outsourcing to GPU\\\" in Section 1.2). To perform training, companies may join forces and have the motivation to dedicate a better network line between them. A similar argument also applies to the setting in which one of the parties is generally trusted would not collude with others, say, the government. To provide machine learning as a service in the application context of electronic healthcare (e.g., precision medicine), the Department of Health & Human Services (or alike) can take the effort. The better network line is the cost we need to pay to support practical enough privacy-preserving machine-learning training (and inference), something that was not known to be possible (with a comparable level of efficiency and accuracy) before our work.\"}", "{\"title\": \"Responses to the Official Reviews\", \"comment\": \"Meta comment.\\nWe thank all the reviewers for the detailed review, which allows us to have concrete revision plan. We are glad that our technical contributions are rightfully observed and appreciated by reviewer 3. Below, we clarify our technical contributions, in particular, how our approach is innovative or at least new, that we can supplement the \\\"fundamentally incorrect\\\" statement (because we focused two aspects but missed one) but the missing part does not affect our experiment. We also further explain our experimental settings and lay out our plan for some more experiments. (Thanks for the inputs!) \\n\\nNo matter whether this is accepted or not eventually, we would appreciate further comments (if possible/applicable) on how well we responded to each issue, and would definitely appreciate and felt encouraged if the score can be at least slightly adjusted (of course, when deemed appropriate).\\n\\nFor sure, we will revise the paper according to the reviews and the responses below.\\n\\nR1/R3. [Technical Contributions] The system builds heavily on the prior Slalom system.\\n\\n1. We share the same general idea of SGX+GPU, but we deviated by just considering one-level lower already. We use 2 SGXes. Our tailor-made outsourcing protocol in Section 3.2 leverages the best of SGX (deriving randomness) and GPU (for batch processing). Existing use of the \\\"well-known trick\\\" just consider outsourcing in general and does not take into account the unique characteristics of SGX and GPU.\\n\\ni. Our protocol thus achieves *real* outsourcing (in terms of total computation time saved). This is different from Slalom's approach, in which the pre-computation time is as much as if it were computed online locally without outsourcing.\\n\\n(We briefly reiterate Section 1.3 and what is explained by Reviewer 3: Slalom precomputes f(r), outsources the computation of f(x + r), and gets back f(x) by f(x + r) - f(r). The precomputation of f(r) is as slow as computing the real problem instance f(x).)\\n\\nii. Further, to better pinpoint a benefit of our new design, even if we stick with the same deployment assumption as Slalom, i.e., we do the offline precomputation online, the whole protocol is faster than running by the SGX alone. In this sense, saying we used \\\"well-known primitive\\\" is a bit more accurate, but we achieved more.\\n\\niii. In the literal sense, it did use \\\"three non-colluding servers, multiplication triplets,\\\" like some other existing works. From a bird-eye view, it is still \\\"some form\\\" of \\\"outsourcing\\\" and secure multi-party computation, but we use new design principles (e.g., minimizing the use of encryption to minimal, correlated randomness). We stress that the specific goals we aim to achieve, and hence, the technical details of the seemingly similar conceptual building blocks are different.\\n\\n2. After all, both works process the neural network. We understand that, from a high-level point of view, they must share a somewhat similar template conceptually to deal with different layers of the neural network. We stress that, in terms of the contribution of the new technical design, we specifically address all the open problems left by Slalom, which hindered them from also supporting training.\\n\\nSpecifically, our design works under several conflicting constraints to support training for the first time, while Slalom failed to resolve (with an explicit discussion on the challenges, so they are not something some \\\"easy\\\" extensions can solve). For one, Slalom's approach inherently leaks the (dynamically changing) weights. An overview is given in the second paragraph of Section 1.3, and details are discussed in Sections 3.3 and 3.4.\\n\\nTo conclude, we think it may be a bit oversimplifying to say our system relies heavily on the prior *system* and we merely use \\\"well-known trick,\\\" at least for the facts that i) existing such protocols were not designed for outsourcing from SGX to GPU but just treating them as \\\"generic processors\\\" and ii) we tackled all the challenges left behind by Slalom.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": [\"The paper proposes a method for privacy-preserving training and evaluation of DNNs. The method is based on a combination of hardware support from a trusted execution enclave (Intel SGX) and an algorithm for offloading intensive computation to unsecure GPU devices and communicating with the trusted environment without losing security guarantees during communication. Compared to related work on a similar system (Slalom), the proposed system enables secure training in addition to inference. The approach is based on the use of additive secret sharing to relegate chunks of computation to independent GPU servers.\", \"The evaluation presents experiments that report timings of the proposed system against a baseline. Throughput (images/second) improvements of 8-9x are reported, but the contrast point is unclear (it appears to be CaffeSCONE, but there are several unclear points, summarized below). In addition, a set of results reporting a speed up ratio against attained accuracy of a trained VGG11 network, and a set of results reporting speed up against arithmetic intensity of the workload are given.\", \"I lean towards rejection of this draft, as it has several weaknesses:\", \"The connection between the evaluation (which mostly focuses on the speed benefits) and the claimed contributions is tenuous at best. This issue is further compounded by clarity issues in the experiments and their description\", \"The empirical results are unclear due to differences between simulation of SGX capability vs hardware support of SGX capability. It is not clear what part of the results is influenced significantly by this disparity, and more importantly whether all the comparisons are done in an equal footing (for example the reported results comparing CaffeSCONE with Goten are performed in two different regimes). As a byproduct, there is a confusing \\\"scaling factor\\\" described by the authors that is applied to the timings.\", \"A brief mention is made of the fact that the proposed system does not in fact provide correctness guarantees (unlike CaffeSCONE), but this is dismissed by reference to utilizing the same trick used by Slalom. However, this trick is not described or motivated.\", \"The writing in the current draft is of relatively low quality, significantly impacting the readability of the paper and making it hard to understand the contributions and whether they are backed by the presented results.\"]}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper builds a privacy-preserving training framework within a Trusted Execution Environment (TEE) such as Intel SGX. The work is heavily inspired from Slalom, which does privacy-preserving inference in TEEs. The main drawbacks of Slalom when extending to training are (1) weight quantization needs to be dynamics as they change during training, and (2) pre-processing step of Slalom to compare u = f(r) isn't effective as the weights change, and running this within TEE is no better than running the full DNN within TEE. In addition, Goten also makes the weights private as opposed to Slalom. Overall, this is a very important contribution towards privacy preserving training and the paper takes a strong practical and implementation-focused approach by considering issues arising due to memory limitations in TEE and the performance implications of default Linux paging.\\n\\nThe paper comes up with a novel outsourcing protocol with two non-colluding servers for offloading linear operations in a fully privacy-preserving way and does detailed analysis of the performance implications. Similar to a lot of other methods for training with quantization, the weights are stored and updated in floats while the computation is performed using quantized values. The experimental results suggest a strong improvement over the CaffeSCONE baseline. One drawback with experiments is the lack of comparison with Slalom for inference if Goten is assumed to be a framework for both training and prediction in a privacy-preserving way.\\n\\nAnother downside of the paper is that a few sections could be improved with their explanation, and there is quite a bit of redundancy in going over the downsides of Slalom and why it can't be used for secure training. For instance,\\n- Section 1.1: \\\"Our results (referring to Section 4.2) show that CaffeSCONE\\u2019s performance greatly suffer from the enclave\\u2019s memory limit as it needs an inefficient mechanism to handle excessive use of memory not affordable by the enclave\\\". Here, it's not clear which mechanism is inefficient. Are we talking about mechanisms in CaffeSCONE for reducing memory usage while training and if so, are they somehow inefficient? Or does it mean to imply that we can't train a DNN fully within an enclave due to memory limits?\\n- Last paragraph of section 2.2 is unclear. \\\"CaffeSCONE guarantees the correctness of both training and prediction. Goten does not provide it as we present it due to page limitation, but we can resort to the trick used by Slalom\\\". What does the last sentence mean? Does Goten guarantee correctness during training and prediction or not? And what trick from Slalom are we referring to? The blinding trick used for privacy or the Freivalds' algorithm used for correctness?\\n\\nOverall a strong contribution with supporting experimental results, but the certain parts need further explanation or rewriting for higher rating.\", \"pros\": [\"An important contribution in the direction of fully private DNN training and inference within a TEE. Draws inspirations from Slalom and mainly addresses the challenges left to extend the approach to training.\", \"Motivation and reasons for why Slalom can't be used for training is very well laid out.\", \"In addition to input and output activations, Goten also preserves the privacy of the weights.\", \"Good baseline for comparison using CaffeSCONE.\", \"Implementation factors considered and analyzed such as tricks as using SGX-aware paging instead of naive Linux paging.\", \"Strong experiments and benchmarks\"], \"cons\": [\"Some sections are not explained well and unclear as mentioned earlier.\", \"How does the inference performance of Goten compare to Slalom given the same privacy and correctness guarantees? This isn't clear from the experiments section.\"], \"minor_comments\": [\"\\\"Slalom\\\" is mis-spelt in line 4 of the abstract.\", \"There appear to be typos and grammatical errors at many places in the paper. Further proof-reading might be helpful.\"]}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary\\n========\\nThis paper proposes a framework for privacy-preserving training of neural networks, by leveraging trusted execution environments and untrusted GPU accelerators.\\nThe system builds heavily on the prior Slalom system, and uses standard MPC techniques (three non-colluding servers, multiplication triplets) to extend Slalom's inference-only protocol to privacy-preserving training.\\nThis is a valuable and hard to reach goal. Unfortunately, the paper's evaluation fails to deliver on its strong promises, by ignoring the high network communication between the non-colluding servers.\\nSpecifically, all experiments were conducted with three servers co-located in a public cloud's LAN. In this setting, it is hard to argue that non-collusion is a valid security assumption as the cloud provider controls all servers (alternatively, if the cloud provider is trusted, then there is no need for any trusted execution or cryptography). If the same experiments were conducted on a WAN, the communication costs would alleviate any savings in computation time.\\n\\nFor these reasons, I lean strongly towards rejection of this paper. \\n\\nDetailed comments\\n=================\\nExtending the ideas in Slalom to support privacy-preserving training is a good research question, and Tramer and Boneh had discussed some of the challenges and limitations towards this in their original paper.\\nGetting rid of the pre-processing stage for blinding factors by leveraging non-colluding servers is a well-known trick from the MPC literature, but it does not seem easily applicable here. \\nThe problem is that the servers need to communicate an amount of data proportional to the size of each internal layer of the network, for each forward and backward pass. If the servers communicate over a standard WAN, the communication time will be much too high to be competitive with the CaffeScone baseline.\\nIn a LAN, as in this paper's experiments, the network latency is low enough for the network costs to be dominated by computation. But this begs the question of whether servers running in a same LAN (e.g., hosted by a single cloud provider) can really be considered non-colluding. In the considered setup, the cloud provider (Google in this case), could just observe the communication between all servers, thereby breaking privacy.\\n\\nAnother security issue with proposed scheme is the lack of computation integrity. This corresponds to the so-called \\\"honest-but-curious\\\" threat model which often appears in the MPC literature, and this should be acknowledged and motivated.\\n\\nOn the experimental side, the considered baseline, CaffeScone, seems pretty weak. In particular, any optimizations that the authors implement for Goten (e.g., better paging) should also be added to their baseline for a fair comparison.\\nThe numbers in Figure 3 show that the baseline could be optimized a lot further. A gap between hardware/simulation modes of ~6x seems indicative of sub-optimal paging. Even the single-core, simulation mode throughput numbers seem low for CIFAR10.\\n\\nThe experimental setup is quite confusing. Running the baseline and Goten in different environments (e.g., different CPUs) and then re-normalizing throughputs is somewhat convoluted and prone to mistakes. Why not run all experiments on the same setup?\\nSimilarly, converting between results in SGX's hardware and simulation modes is also not very indicative. The authors note (p. 8) that in SGX's simulation mode \\\"code compilation is almost the same as hardware mode except that the program is not protected by SGX, which is fine for our purpose since the DNN training and prediction algorithms are publicly known\\\". This is fundamentally incorrect!\\nSGX's simulation mode provides absolutely no security guarantees. It simply compiles the code using the SGX libraries and ensures that the enclaved code performs no untrusted operations, but it does not provide any hardware protections whatsoever. In particular, code running in simulation mode will not be affected by the overhead of SGX's paging, as the memory is never encrypted.\\nAs a result, performance results in simulation mode are usually not indicative of performance in hardware mode. Trying to convert runtimes from simulation mode to hardware mode by comparing times of specific layers is also prone to many approximation errors. \\n\\nFinally, I had some trouble understanding the way in which Goten quantization works. Section 3.3. mentions that values are treated as floats, but then mentions the use of 53 bits of precision. Did you mean double-precision floats here? But then, aren't modern GPU optimized mainly for single-precision float operations? Section 3.3. also says that the quantization ensures that there are nearly no overflows. What happens when an overflow occurs? I guess that because of the randomized blinding, a single overflow would result in a completely random output. How do you deal with this during training?\\n\\nMinor\\n=====\\n- Typo in abstract: Slaom -> Slalom\\n- I don't understand the purpose of footnote 3 in Appendix B.2. First, the bibliographic entry for (Volos et al. 2018) explicitly says that the paper was published in OSDI 2018, a top-tier peer-reviewed conference. Regardless, claiming a date for a first unpublished draft of your paper is a little unusual and somewhat meaningless. I'm sure Volos et al. had a draft of their paper ready in late 2017 or even earlier if they submitted to OSDI in XXX 2018. If you want to timestamp your paper, post in to arXiv or elsewhere online.\"}" ] }

[ "Christian Bueno", "Alan G. Hylton" ]

[ "universal approximation", "point clouds", "deep learning", "hausdorff metric", "wasserstein metric" ]

{ "note_id": [ "dWBjn-o9m", "Syx6NYS3oB", "Hyx7lxrhsB", "H1g7UpV3iH", "B1eCTtV2oB", "rJgDrkcXqr", "BJl_MskAFS", "rJxCrFiDKH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798752014, 1573833013246, 1573830635317, 1573829963411, 1573829061940, 1572212543152, 1571842831870, 1571432774434 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2555/Authors" ], [ "ICLR.cc/2020/Conference/Paper2555/Authors" ], [ "ICLR.cc/2020/Conference/Paper2555/Authors" ], [ "ICLR.cc/2020/Conference/Paper2555/Authors" ], [ "ICLR.cc/2020/Conference/Paper2555/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2555/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2555/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The present paper establishes uniform approximation theorems (UATs) for PointNet and DeepSets that do not fix the cardinality of the input set.\\n\\nTwo nonexperts read the paper and came away not understanding what this exercise has taught us and why the weakening of the hypotheses was important. The authors made no attempt to argue these points in their rebuttals and so I went looking at the paper to find the answer in their revisions, but did not find it after scanning through the paper. I think a paper like this needs to explain what is gained and what obstructions earlier approaches met, and why the current techniques side step those. One of the reviewers felt that the fixed cardinality assumption was mild. I'm really not sure why the authors didn't attack this idea. Maybe it is mild in some technical sense?\\n\\nWhat I read of the paper seemed excellent in term of style and clarity. I think the paper simply needs to make a better case that it is not merely an exercise in topology. I think the result here is publishable on its own grounds, but for the paper to effectively communicate those findings, the authors should have revised it to address these issues. They chose not to and so I recommend ICLR take a pass. Once the reviewers revised the framing and scope/impact, provided it doesn't sound trivial, I think it'll be ready for publication.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to some of the above concerns\", \"comment\": \"Thank you for your detailed review. Below we attempt to address some of your points:\\n\\n\\\"However, for the two base architectures that are examined in this work, UATs where already provided in the original manuscripts.\\\"\\n\\nThough it is true that the original manuscripts provide UATs, it is not a completely apples-to-apples comparison in the case of DeepSets. As mentioned in the paper, we made a slight adjustment to the form of the DeepSets architecture by replacing summation with averaging. We believe this variation is not directly addressed by the theorem presented in the original manuscript. We also believe replacing summation with averaging provides a theoretical improvement by allowing one to interpret DeepSets as acting continuously on probability measures, which lends some intuition to its behavoir when applied to larger point clouds.\\n\\nOur work also elucidates how simplistic the architecture can be while retaining UATs.\\n\\n\\\"The paper shows two specific functions, where one can be approximated by PointNet but not by DeepSets and vice-versa. The authors again note that this might vanish when a fixed cardinality of the input point cloud is assumed. Again, it seems that this is a mild restriction and I'd like the authors to elaborate on the importance of removing this restriction.\\\"\\n\\nWe added new results that demonstrate this problem does not go away when cardinality is bounded. At least not for the case of PointNet trying to approximate the center-of-mass. We are able to extract lower bounds on the uniform-norm error (independent of the details of the PointNet implementation) and are able to generate as many examples as we desire that achieve error above a theoretical guarantee. This forms the basis of our empirical verification of our results.\\n\\n\\\"The paper is clearly targeted at a specialist audience and invokes dense and advanced mathematical concepts. It might find a better audience at a venue that is more specialized in this type of work\\\"\\n\\nWe hope that the addition of an explicit error lower bound and experiment confirming it will broaden the appeal. \\n\\n\\\"UATs are important and interesting, however, they do exist for the architectures that are targeted in this paper. It is not clear what is gained by the main difference, i.e. removing the cardinality. I'd be grateful if the authors could comment on this. \\\"\\n\\nWe believe that in addition removing the cardinality assumption, and elucidating how simple these networks can get away with being, the generalization of the classical UAT to compact Hausdorff spaces could prove useful to other researchers working with other non-traditional data types. In particular, the proofs in the original manuscripts of PoitnNet and DeepSets use techniques that don't seem to readily apply to the other, let alone scale to arbitrary finite cardinality. The introduction of this fairly easy generalization of the classical UAT is what allowed us to prove multiple universal approximation theorems with one method of attack.\"}", "{\"title\": \"Improvements to results allow for the generation of adversarial examples to PointNet when learning center-of-mass\", \"comment\": \"Thank you for your review.\\n\\nBefore we were unable to provide experiments due to the difficulty of algorithimically finding examples that achieve L^\\\\infty errors promised by the inability to uniformly approximate. This was made even more difficult because point cloud cardinality was free to be unbounded. However we now have results that apply to the bounded cardinality case for PointNet and show that this still cannot uniformly approximate the center-of-mass. We provide an explicit error bound and empirical test of this bound in the paper. We found these results to be robust in the sense that we could randomly sample pairs of points, and use those as a seed to find sets that yield the predicted worst case error. In a sense, this demonstrates a method to produce adversarial point cloud examples with theoretical guarantees.\\n\\nWe hope this further elucidates the nature of these function classes and improves the practicality of these results.\"}", "{\"title\": \"Regarding some of these issues\", \"comment\": \"Thank you for careful review and insights.\\n\\nDue to time constraint in obtaining newly proven error bound (added in revision) and its experimental verification for in the bounded cardinality case -- which we hope addresses your concerns in 3) -- we were not able to fully address 1). If there remains time to edit after this comment we would certainly like to expand on this.\\n\\nRegarding 2), we carefully considered a change of notation but found that it would make the management of the proofs in the paper more unwieldy and possibly more opaque in other regards. The notation is also there to help illustrate how sparse the architecture can be while preserving the UAT and to distinguish between the two types of layers involved with neural networks on these more difficult spaces. \\n\\nRegarding 3) we made an official comment at the top that we believe addresses these concerns.\"}", "{\"title\": \"Addition of Theorem & Explicit Error Bound in Case of Bounded Cardinality and Empirical Verification\", \"comment\": \"We thank the reviewers for their thoughtful and constructive comments. We address this comment to the three reviewers who shared similar concerns regarding the practicality of the results for unbounded cardinality and lack of experiments. We took these concerns seriously and pushed the methods to arrive at no-go theorems for the case of bounded point cloud size. We prove that even with cloud size bounded to k or less points, PointNet cannot uniformly approximate the center-of-mass function. Moreover we prove an explicit worst case error lower bound (worst case error is at least (k-2)*Diam(domain)/k) and empirically confirm the robustness of these results. We also add images to aid in the understanding of these results.\\n\\nWe hope that these changes will alleviate some of the concerns of the reviewers. We will individually address the remaining concerns in the comments.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper aims to establish novel theoretical properties of known point-based architectures for deep learning, PointNet and DeepSets. To this end, the authors prove a series of theoretical results and establish limitations of these architectures for learning from point clouds.\\n\\nUnfortunately, due to a delay with my review, I cannot afford the joy of carefully examining the proofs. \\n\\nI do not find significant practical value in the obtained results. While the theorems proved in the paper are original and novel, they are a refinement of the already known results regarding approximation theorems for PointNet and DeepSets, respectively, hence only a marginal improvement in understanding these function classes.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"This work examines the fundamental properties of two popular architectures -- PointNet and DeepSets -- for processing point clouds (and other unordered sets). The authors provide a new universal approximation theorem on real-valued functions that doesn't require the assumption of a fixed cardinality of the input set. They further provide examples of functions that can't be mutually approximated by PointNets and DeepSets.\", \"It is important in order to know the representational power and fundamental limitations of basic algorithmic building blocks. However, for the two base architectures that are examined in this work, UATs where already provided in the original manuscripts. The presentation in this paper does remove the assumption of a fixed cardinality, but since this seems to be a mild assumption, it is not clear what is gained by this (beyond mathematical elegance). The paper doesn't give any hints here.\", \"The paper shows two specific functions, where one can be approximated by PointNet but not by DeepSets and vice-versa. The authors again note that this might vanish when a fixed cardinality of the input point cloud is assumed. Again, it seems that this is a mild restriction and I'd like the authors to elaborate on the importance of removing this restriction.\", \"The paper is clearly targeted at a specialist audience and invokes dense and advanced mathematical concepts. It might find a better audience at a venue that is more specialized in this type of work (e.g. applied mathematics or more theory-focused machine learning venues).\"], \"summary\": \"UATs are important and interesting, however, they do exist for the architectures that are targeted in this paper. It is not clear what is gained by the main difference, i.e. removing the cardinality. I'd be grateful if the authors could comment on this.\", \"disclaimer\": \"While I believe to have a reasonable mathematical background, I'm not an expert in this field and my assessment is primarily based on the bottom line of these proofs.\\n\\n=== Post rebuttal update ===\\nI'd like to thank the authors for their efforts and additional insights. The additional illustration improves the accessibility of the paper. However, the rebuttal does not alleviate my concerns about additional impact beyond the UATs in the original paper. I thus maintain my recommendation of weak reject\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"PointNet (Qi et al, 2017) and Deep sets (Zaheer et al, 2017) have allowed to use deep architectures that deal with point clouds as inputs, taking into account the invariance in the ordering of points. However, existing results on their approximation abilities are limited to fixed cardinalities. This paper removes the cardinality limitation and gives two kinds of results:\\n\\n1.\\tPointNet (resp. Deep sets) can approximate uniformly real-valued functions that are uniformly continuous with respect to the Hausdorff (resp. Wasserstein) metric;\\n2.\\tOnly constant functions can be uniformly approximated by both (PointNet, Hausdorff metric) and (Deep sets, Wasserstein).\\n\\nThis paper brings a valuable theoretical contribution to the existing state of the art of their approximation abilities. With some improvements, I am willing to increase the score.\\n\\n1.\\tThe introduction lacks insight into the literature on point cloud or measure networks, including in practice, which would motivate the subject and place it more precisely within the literature.\\n2.\\tNotations in the section 2.4 make the reading particularly unclear. Notations should showcase the result that theoretically, only two hidden layers (with appropriate definitions) are needed.\\n3.\\tThe paper lacks an experimental section. It would be interesting to investigate empirically the limitations of these architectures, for instance by playing with the diameter and center of mass functions as suggested in 3.3.\\n\\n- Post rebuttal: the authors have addressed 3., therefore I am leaning towards accept.\"}" ] }

[ "Kamil Ciosek", "Vincent Fortuin", "Ryota Tomioka", "Katja Hofmann", "Richard Turner" ]

[ "uncertainty quantification", "deep learning", "Gaussian process", "epistemic uncertainty", "random network", "prior", "Bayesian inference" ]

{ "note_id": [ "PugTQsmZ1p", "HylZ0QwnjH", "HkxA9Xw2iB", "SkxSCMv2sr", "SJxttMvhiH", "rylzpBMioS", "BJeSRf3qoH", "BkxsBgKdoH", "Byl7Pu4QiB", "rkeEsDE7jB", "Hkl9YrEXjr", "SJxrqE4Xor", "B1gsx74XoS", "HJeO7lNQsS", "B1xsPiVxsH", "HkgK3JURYB", "B1xA7Po2tS", "HJgnXst2tr", "HJecBlyaur", "Bklx07aldH", "rJlP-g5JuH", "HJleGHsoDH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review", "official_comment", "comment", "official_comment", "comment" ], "note_created": [ 1576798751981, 1573839816840, 1573839765647, 1573839565140, 1573839489027, 1573754298186, 1573728973172, 1573584962970, 1573238875197, 1573238683556, 1573238146288, 1573237900786, 1573237491070, 1573236768035, 1573043043045, 1571868592861, 1571759910515, 1571752740000, 1570725954340, 1569932232204, 1569853439459, 1569596679698 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "~Khurai_Kim1" ], [ "ICLR.cc/2020/Conference/Paper2553/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2553/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2553/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "~Pranav_Poduval1" ], [ "ICLR.cc/2020/Conference/Paper2553/Authors" ], [ "~Anthony_Wittmer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper provides theoretical justification for a previously proposed method for uncertainty estimation based on sampling from a prior distribution (Osband et al., Burda et al.).\\n\\nThe reviewers initially raised concerns about significance, clarity and experimental evaluation, but the author rebuttal addressed most of these concerns.\\n\\nIn the end, all the reviewers agreed that the paper deserves to be accepted.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks again for the review - we summarise the major relevant changes in the final revision.\", \"comment\": \"Our conservatism guarantee holds for finite bootstraps as well as in expectation. We clarified this by explicitly adding Corollary 1. We believe Proposition 1 is significant because uncertainty estimates that avoid overconfidence .\\nWe fixed the clarity issues you mentioned. We enhanced the experimental evaluation by adding a new experiment on CIFAR100 (appendix B.4). We feel this is sufficient for the paper to meaningfully contribute to the community.\"}", "{\"title\": \"Thanks again for the review. We summarise the relevant changes made in the final revision.\", \"comment\": \"Relative the the original submission. We added an extended discussion of the prior and extended the comparison with deep ensembles, describing an explicit situation when they fail (each model in ensemble convex in weights). We shortened the preliminaries and addressed the clarity issues. Our key contribution over from prior work by Osband is that we deal with non-linear function approximation, a key requirement for many practical tasks (Osband's work only had theory for linear regression).\"}", "{\"title\": \"Thanks again for the detailed review - we have uploaded the final version.\", \"comment\": \"Accuracy figures for the classification task are now in Appendix B.3. The training error is included in Appendix A.2 and accuracy figures for the OOD task are in Appendix B.2. The paper also now includes an ablation on the initialization scale. We tried training a BNN using the state-of-the art VOGN BNN optimizer (https://github.com/team-approx-bayes/dl-with-bayes) but couldn't get it to work with our ResNet architecture, despite very low learning rates, long training, applying several settings for the prior precision and generally putting as much effort into the BNN as the other methods. We understand this may not be everyone's opinion, but this may be taken as confirmation that current BNNs are just very hard to get to work with architectures combining residual connections and convolutions, necessary for modern vision tasks. In the end, we did not include the BNN results in the paper because we felt the architecture needs to be kept the same across methods for the experiments to be meaningful. We feel that thanks to our theoretical contributions, our paper still represents a useful contribution to the community.\"}", "{\"title\": \"Final version of paper.\", \"comment\": \"Thanks again to all reviewers. We have now uploaded the final version of the paper.\\n\\nOverall, we feel that main contribution of the paper - a sound theoretical framework for what was previously an ad-hoc method lacking principled justification will prove useful for the community. \\n\\nRelative to the original submission, we added a new experiment on CIFAR100 (appendix B) and an ablation on initialization scale (page 10). We now include details of the classification and OOD accuracy as well as the other technical details requested by reviewer 1. We also made many clarifications in the text, requested by the reviewers. Particularly, we provide additional background on when deep ensembles can fail as well as on the used prior. \\n\\nWe provide individual summary response messages below each review.\"}", "{\"title\": \"Thanks for the comment.\", \"comment\": \"The accuracy numbers have now been added (Appendix B2). We are working on a BNN example.\"}", "{\"title\": \"Response to rebuttal\", \"comment\": \"Thank you for addressing thoroughly my comments and I appreciate the revised version of the manuscript. The clarity has improved and the extra sections help in providing more insight about potential edge cases of this method. One thing I noticed is that the classification accuracy numbers for the in-distribution test set for both the RP and baselines are still missing (there is Figure 4 but that is kind of coarse and I am not sure if the images are from the test set).\\n\\nFor now, I am in general in favor of accepting this work so I will maintain my current grade. Should the authors manage to also get some convincing results against a simple variational BNN I will consider increasing my score.\"}", "{\"title\": \"We have uploaded a revised version.\", \"comment\": [\"We have uploaded a revised version of the paper. We made many clarifications throughout the paper, particularly in Section 4. We list the main points below.\", \"In response to review #1:\", \"filled in the missing technical details, in particular:\", \"we added details of the architecture (Appendix A2).\", \"we added information about training error (Appendix A2).\", \"we added a description of weight cloning (last paragraph, Section 5)\", \"added clarifications concerning dropout\", \"In response to reviewer #2:\", \"we extended our discussion of the prior\", \"we enhanced a discussion of deep ensembles, adding an example where they can fail (each model in an ensemble convex in weights, which means all models converge to the same solution).\", \"we shortened the preliminaries and used the space for a clearer description of our method\", \"In response to review #3:\", \"we added an explicit statement (Corollary 1) that shows conservatism for finite number of bootstraps, including just one bootstrap.\", \"we added an additional appendix that lists the variances of random variables used in the paper (Appendix E).\", \"we changed the notations to be more readable ('v' symbol for variance).\", \"Y axis in Figure 3,6 has been changed as requested\", \"We are still working on the paper and will keep updating it.\"]}", "{\"title\": \"Thanks for the pointers.\", \"comment\": \"Thanks for mentioning these. We will reference them in the revised paper.\"}", "{\"title\": \"Thanks for the review.\", \"comment\": \"Below, we address your points in detail.\\n\\n\\\"[in the Strength Section] Proposition 1 is an interesting result, although the paper does not seem to discuss its significance and implications enough\\\".\\nOverall, the main contribution of the paper is the we provide a sound theoretical framework for what was previously an ad-hoc method lacking principled justification. We will provide more background specifically about the usefulness Proposition 1 in a revised version of the paper. The main justification is that Proposition 1 guarantees that we won't become overconfident. We show empirically (Figures 3 and 6) that this problem can happen when using deep ensembles.\\n\\n\\\"Proposition 1 is described as 'our uncertainty estimates are never too small'. However, as the name of the proposition suggests, it seems to only hold in expectation over models trained, which is a quite different statement.\\\"\\nYou are right when you say that Proposition 1 on its own holds in expectation. However, we also provide a result for a finite number of trained models in Lemmas 2 and 3 in the \\\"Finite Bootstraps\\\" section of the paper. Combining these two Lemmas with Proposition 1 gives a conservatism guarantee that is valid even for just one trained model. We agree that it is not very clearly presented - we will will add an explicit corollary that provides a conservatism guarantee with a finite number of bootstraps / models.\\n\\n\\\"Proposition 2 seems to simply state that a small network can be distilled into a large network. Maybe I missed part of the reasoning here? Otherwise, it should probably not be highlighted as a major contribution.\\\"\\nConcentration is a crucial property for uncertainty estimates and we feel that a paper without this result would be incomplete. Proposition 2 shows that, given enough data, uncertainties become small on *unseen* points. This is not the same as simply distilling the smaller \\\"prior\\\" network on some dataset. In other words, in order for the concentration result to hold, the capacity of the larger (predictor) network has to be controlled (which we do by the Lipschitz constant). While the Rademacher complexity tools we use to show this result are standard, to the best of our knowledge, they have not been used in a similar context. \\n\\n\\\"The experimental evaluation is very limited, training only on CIFAR 10. While the experiments add little value to the community, this may be acceptable for a mostly theoretical paper.\\\"\\nAs you say, the experiments are meant to lend support to the theoretical part of the paper. The experimental evaluation has been carefully selected to focus on how conservative uncertainty estimates avoid overconfidence. For example, Figures 3 and 6 show that uncertainty estimates obtained from random priors show a better separation between seen and unseen points. They are also better calibrated - Figure 5 shows that the curves for random priors are closer to monotonic than for competing approaches. We believe that our experiments do provide a value for the community in the sense of validating theory. \\n\\n\\\"The paper was difficult to read and unclear in explanations. It could help to define notation upfront instead of introducing shorthands on the way.\\\"\\nThanks for pointing this out. We will upload a revised submission where the notation is clearer.\\n\\n\\\"In Figure 3, the Y axis limits should be fixed across seen and unseen histograms for the same method. The current presentation is a bit misleading here, as the presented method seems to have moved the most mass under this chart scaling\\\"\\nWe will revise Figure 3 (and also Figure 6) to allow a direct comparison. We agree that it will make the figures easier to interpret.\\n\\n\\\"[In the comments] As found in prior work cited in the submission, the method tends to perform well with just one network pair. This raises the question whether the contribution of the paper that holds in expectation over many pairs and the empirical success of the approach are connected.\\\"\\nThanks to Lemmas 2 and 3, out theory carries over to a finite number of prior-predictor pairs (including just one pair). We will add a corollary to make this explicit. \\n\\n\\\"The marginal posterior variance \\\\sigma^2(x_\\\\star) appears in various forms with hat, tilde, and different subscripts. It maybe worth assigning different letters to these to avoid confusion.\\\"\\nWe will provide a revised version of the paper with an updated notation. \\n\\nOverall, thanks a lot for the feedback. Please let us know if you have other suggestions for improving the paper.\"}", "{\"title\": \"Thanks for the review.\", \"comment\": \"We address your points below.\\n\\n\\\"Overall, the paper is relatively well-written, although it might be at times hard to follow, especially for someone who is not familiar with the original work that used randomized prior functions (Burda\\u201918, Osband \\u201818, \\u201819).\\\"\\nThanks for the kind words! We will shortly upload a revised version of the paper with improved clarity. \\n\\n\\\"I find the core idea behind the paper quite interesting, however, as indicated by authors themselves, it has already been studied in a slightly different context (RL, works by Burda et. al, Osband et. al). That said, authors do provide additional insides for the supervised settings, and also analyse theoretically the behaviour of uncertainty estimates.\\\"\\nThanks for pointing this out. The main contribution of the paper is a theoretical analysis of uncertainties obtained from fitting random priors for *arbitrary deep networks*. This is different from Osband's work, which discusses Bayesian linear regression. We believe that the difference in setting is hugely significant, because non-linear function approximation is necessary for many applications of Machine Learning. The work by Burda et al. provides great empirical results in RL, but no theoretical justification for the uncertainties.\\n\\n\\\"(somewhat minor) p1: 'While deep ensembles \\u2026, where the individual ensembles are trained on different data' - here and related text, it should probably be 'individual models' / 'individual networks'. Generally, I am not convinced that these are strong arguments against deep ensembles.\\\"\\nYou are right about the wording - we meant 'individual network'. In terms of comparing our work to deep ensembles, we wanted to make two points. First, our work can be understood as complementary, to be used in situations where the conservatism property is desired. Second, our method has some practical advantages over deep ensembles. In practice, deep ensembles can also become overconfident (Figure 6) and poorly calibrated (Figure 5). Also, deep ensembles require more bootstraps (we will describe this more explicitly in the updated paper). We will provide a more detailed comparison with deep ensembles in the revised version of the paper. Also, the main theoretical argument against deep ensembles is that the uncertainties obtained from them in practice (when each ensemble was trained on the whole dataset) do not have a connection to a Bayesian posterior. \\n\\n\\\"(minor) p2-p3: \\u201c2. Preliminaries\\u201d - I am not sure if this section adds much to the understanding, it would seem more natural to spend more time explaining the intuitions behind the net\\\"\\nIn the revised submission, we will move some of the preliminary material to the appendix and use the space to provide more intuitions.\\n\\n\\\"The explanation of why using a randomly initialized network makes sense is not very strict. I kind of get the general idea, but it is not clear to me why not use something less expensive, e.g. just random projections, and why do we actually need a full network. Intuitively it seems quite strange to waste a lot of capacity to fit to essentially fit a set of random weights: is it something that allows the network to avoid easily learning the \\u201crandom prior\\u201d? And, more generally, can this also be considered as a \\u201ctrick\\u201d to de-correlate individual predictors? I believe these points should be discussed in more detail.\\\"\\nThanks for pointing this out. On the formal side, the justification for using random priors is that our uncertainty estimates overestimate the Gaussian posterior. Deep ensembles do not have a comparable guarantee. Intuitively, the role that the random networks play is to avoid spurious confidence. We would not say that random priors waste capacity since the predictor network still depends on the x's (training points) and that is very useful for OOD detection. The fact that the predictors are trained independently means, as you say, that uncertainty estimates obtained from each predictor-prior pair are independent (de-correlated). We will include these insights in the revised version of the paper.\\n\\nOverall, thanks for the detailed comments. If you have other suggestions for improving the paper, please share!\"}", "{\"title\": \"Response to \\\"Other Comments\\\" and \\\"Misc remarks\\\".\", \"comment\": \"Thanks again for the feedback.\\n\\n\\\"It is worth pointing out that [1] showed that Monte-Carlo dropout performs approximate MAP inference, which seems more plausible than the approximate Bayesian inference perspective of [2].\\\"\\nThanks for pointing this out, we will mention the interpretation of Dropout in terms of MAP inference and include [1 - Nalisnick et al.] in the revised version of the paper. We want to point out an importance difference in the setting - while our work talks about uncertainty of the posterior, MAP inference considers the effect of the prior on weights estimates.\\n\\n\\\"In the introduction you argue that Bayesian neural networks rely on procedures different from standard supervised learning and thus most ML pipelines are not optimized for them in practice. Could you elaborate a bit about this statement? Variationally trained BNNs with e.g. the reparametrization trick [3, 4] are straightforward since you can just use backpropagation to update their (variational) parameters.\\\"\\nWhile Bayesian Neural Networks provide a link between deep learning and Bayesian inference, they are very slow to train. Even recent tuned implementations of BNNs (Osawa et al.) are several times slower than supervised learning (see Table 1 in Osawa et al. for benchmarks in a paper that argues for BNNs). This happens despite using a battery of technical optimizations, including distributed training and batch normalization. Moreover, modern convolutional BNNs still carry a significant accuracy penalty when deployed with realistic settings of prior variance (see appendix E in Osawa et al.). We will include this point in the revised paper.\\n\\n\\\"What is the x-axis for Figure 3 for the baselines? (I take it that for RP it is the \\\\hat{sigma}^2(x)).\\\"\\nWe use the method for obtaining uncertainties recommended by the \\\"deep ensembles\\\" paper. The x axis shows (1 - p_max), where p_max is the probability of the most probable class. The probability vectors from each bootstrap are averaged over before performing the max. We will clarify the description of the figure. \\n\\n\\\"I believe that a comparison against a simple variationally trained BNN would make the results more convincing.\\\"\\nWe are not sure we would be able to run this experiment on time. We will report back.\\n\\nAlso, thanks for pointing out the typos - we will fix this in the revised version.\\n\\n\\\"Sixth page, \\u201cCorollary 1 and proposition 2\\u201d; where is corollary 1? Do you mean Proposition 1?\\\"\\nYes, we meant Proposition 1. We will fix this.\\n\\n\\\"Overall, I tend to accept this work\\\"\\nThanks a lot for the detailed feedback. If you have any other suggestions you believe may improve the paper, please share!\\n\\nReferences (for this post):\\nOsawa, K., Swaroop, S., Jain, A., Eschenhagen, R., Turner, R. E., Yokota, R., & Khan, M. E. (2019). Practical Deep Learning with Bayesian Principles. \\nEric Nalisnick, Jos\\u00e9 Miguel Hern\\u00e1ndez-Lobato, Padhraic Smyth, Dropout as a Structured Shrinkage Prior, 2019\"}", "{\"title\": \"Thanks for the detailed review!\", \"comment\": \"We answer your individual major points below.\\n\\n\\\"This work is interesting as it seems to provide a simple way to obtain reasonable uncertainty estimates. For this reason it can potentially serve as a strong baseline for this field\\\"\\nThanks a lot. We appreciate. \\n\\n\\\"the writing could use some more work in order to make things more clear\\\"\\nWe will provide a revised version of the paper which we hope is clearer. \\n\\n\\\"How exactly do you apply your method on the classification scenario? Do you select an arbitrary hidden layer of the classification model for the prior and predictor network architectures or the output logits / softmax probabilities?\\\" \\nFor classification, the architecture for the prior network was the same as the classification network, but using a final linear layer instead of the softmax layer. We used squared error on that last layer to get the uncertainties. We will clarify this in the revised paper.\\n\\n\\\"What is the average training error of the predictor networks for the out-of-distribution task and subsampling ablation task, i.e. how far away from concentration were the priors?\\\"\\nThe histogram of the training error is the same as the histogram of the uncertainties (Figures 3 and 6), first column. We will provide a summary of training / test errors in the revised version of the paper.\\n\\n\\\"An effect that I found weird is the following: what happens for the out-of-distribution examples when the predictor networks can perfectly predict the prior network outputs? Wouldn\\u2019t that then imply that the uncertainty would be zero for any input (even an out-of-distribution one), as the prior network and predictor network always agree? One could imagine that for e.g. simple priors and with sufficiently dense sampling of the domain of the function this can happen in practice.\\\"\\nOur experiments show that exact network cloning does not happen in practice. On the theoretical side, a useful edge case to consider is if the priors are just linear functions and we are solving a one-dimensional regression problem. In this case, (assuming no observation noise), evaluating the prior at two points is enough to fit the linear function and completely know the value of the prior everywhere. As you say, afterwards, the prior and predictor will always agree and uncertainty estimates will be zero. However, this is still consistent with our theory! The reason for this is once we assume such a linear prior, we are comparing to a GP with a linear kernel. But a GP *with that kernel* will also have zero uncertainty after seeing two samples, hence the conservatism guarantee still holds. In practice, this means that we have to be careful when choosing the architecture of the prior networks However, this difficulty is no different from performing Bayesian inference with a prior inappropriate for the problem. Empirically, reasonable network architectures do not show such cloning behavior.\\n\\n\\\"For the conservatism you show that your uncertainty estimate is higher, on average, than the posterior variance when you sample points from the model itself. In a sense this guarantee translates to the actual data when the prior is \\u201ccorrect\\u201d. How do those conservatism guarantees translate to the case when there is model misspecification, i.e. when the prior is not correct? Perhaps a small toy example would be informative.\\\"\\nOur method is affected by model misspecification in the same way as Gaussian Processes are. In particular, our conservatism results guarantee that the uncertainty obtained from random priors are greater or equal than uncertainties obtained from a GP which uses the same prior. Also, since in practice the priors are neural networks, chosen to have large capacity, we are unlikely to end up in a situation where the prior has no support on the ground truth function. \\n\\n\\\"For the predictor networks as described in figure 2; do you train both the green and red parts of the network or only the red parts and keep the green part fixed to the values you used for the prior f? (This helps in understanding how easy / difficult is the task of the predictor network).\\\"\\nIn figure 2, we train the entire predictor networks (both parts). We will clarify the caption.\\n\\n\\\"What is the accuracy on the actual in-distribution prediction task for the RP and baselines?\\\"\\nIn the paper, we reported the AUROC as it is a more robust. measure of classification quality. In the revised version, we will additionally provide accuracy numbers in the appendix. \\n\\n\\\"What did 'B' correspond to for the dropout networks? Was it the number of dropout samples you averaged over to get the final predictive?\\\"\\nAs you say, B was the number of samples averaged over at test time. We will make this clearer in the revised submission. \\n\\n\\\"How sensitive are the results on the actual initialization strategy of the prior network? It would be good to see e.g. some form of performance / init variance curve in order to decipher the sensitivity.\\\" \\nWe will try to do an ablation on the initialization scale.\"}", "{\"title\": \"Thanks for the reviews! We will provide a revised version of the paper.\", \"comment\": \"Thanks to all reviewers for providing feedback!\\n\\nWe address the main points below. We also provide a detailed response to each review individually.\\n\\n1. [Significance] The main contribution of the paper is that we provide a sound theoretical framework for what was previously an ad-hoc method lacking principled justification, but was found to work really well in practice (Burda et al). Osband et al. only justified the method for linear regression, while our theory works for arbitrary deep networks.\\n\\n2. [Comparison with deep ensembles] We do not aim do replace deep ensembles, but to provide a method which is complementary to them in settings where conservative uncertainties are desired. Having said that, our it also has practical advantages. In particular, in practice, deep ensembles can become overconfident and poorly calibrated. Also, deep ensembles require using more bootstraps (we will describe this more explicitly in the updated paper). We will provide a more detailed comparison with deep ensembles in a revised version of the paper.\\n\\nWe agree with the reviewers that the clarity of the submission can be improved. \\nWe will provide a revised version of the paper.\\n\\nReferences (for this post):\\nYuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation.\\nIan Osband, John Aslanides, and Albin Cassirer. Randomized prior functions for deep reinforcement learning. NeurIPS 2019 \\nIan Osband, Benjamin Van Roy, Daniel J. Russo, and Zheng Wen. Deep exploration via randomized value func-tions. Journal of Machine Learning Research, 2019.\"}", "{\"title\": \"About structural priors implied by the architectures\", \"comment\": \"I think you could include [1][2] when you say \\\"we are implicitly making the assumption that the network architecture is appropriate for the dataset anyway.\\\"\\n\\n[1] Ulyanov, Dmitry, Andrea Vedaldi, and Victor Lempitsky. \\\"Deep image prior.\\\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.\\n[2] Zezhou Cheng, Matheus Gadelha, Subhransu Maji, Daniel Sheldon. \\\"A Bayesian Perspective on the Deep Image Prior\\\". CVPR 2019.\\n\\nThese explicitly discuss the subject of structural priors of network architectures.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper shows that the MSE of a deep network trained to match fixed random network is a conservative estimate of uncertainty in expectation over many such network pairs. An experiment compares this previously proposed method to other approaches for uncertainty estimation on CIFAR 10.\", \"strengths\": [\"Obtaining uncertainty estimates for predictions of deep neural networks is an important and open research question.\", \"Proposition 1 is an interesting result, although the paper does not seem to discuss its significance and implications enough.\"], \"weaknesses\": [\"Proposition 1 is described as \\\"our uncertainty estimates are never too small\\\". However, as the name of the proposition suggests, it seems to only hold in expectation over models trained, which is a quite different statement.\", \"Proposition 2 seems to simply state that a small network can be distilled into a large network. Maybe I missed part of the reasoning here? Otherwise, it should probably not be highlighted as a major contribution.\", \"The experimental evaluation is very limited, training only on CIFAR 10. While the experiments add little value to the community, this may be acceptable for a mostly theoretical paper.\"], \"clarity\": [\"The paper was difficult to read and unclear in explanations. It could help to define notation upfront instead of introducing shorthands on the way.\", \"In Figure 3, the Y axis limits should be fixed across seen and unseen histograms for the same method. The current presentation is a bit misleading here, as the presented method seems to have moved the most mass under this chart scaling.\"], \"comments\": [\"As found in prior work cited in the submission, the method tends to perform well with just one network pair. This raises the question whether the contribution of the paper that holds in expectation over many pairs and the empirical success of the approach are connected.\", \"The marginal posterior variance \\\\sigma^2(x_\\\\star) appears in various forms with hat, tilde, and different subscripts. It maybe worth assigning different letters to these to avoid confusion.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Overview:\\nThis paper introduces a new method for uncertainty estimation which utilizes randomly initialized networks. Essentially, instead of training a single predictor that outputs means and uncertainty estimates together, authors propose to have two separate models: one that outputs means, and one that outputs uncertainties. The later one consists of two networks: a randomly initialized \\u201cprior\\u201d which is fixed and is not trained, and a \\u201cpredictor\\u201d, which is then trained to predict the output of the randomly initialized \\u201cprior\\u201d applied to the training samples. \\nAuthors show that under some reasonable assumptions the resulting estimates are conservative and concentrated (i.e. bounded and converge to zero with more data).\", \"writing_quality\": \"Overall, the paper is relatively well-written, although it might be at times hard to follow, especially for someone who is not familiar with the original work that used randomized prior functions (Burda\\u201918, Osband \\u201818, \\u201819).\", \"evaluation\": \"The method is experimentally evaluated on a task of out-of-distribution detection on CIFAR+SVHN, and seems to perform on-par or better than the baselines (including \\u201cstandard\\u201d deep ensembles and dropout networks). In addition, there are experiments that demonstrate that the model is performing relatively well in terms of calibration (whether the model predictive behaviour makes sense as the model confidence changes).\", \"decision\": \"I find the core idea behind the paper quite interesting, however, as indicated by authors themselves, it has already been studied in a slightly different context (RL, works by Burda et. al, Osband et. al). That said, authors do provide additional insides for the supervised settings, and also analyse theoretically the behaviour of uncertainty estimates. \\nOverall, I cannot say I am fully convinced that the paper should be accepted as is (also see questions below), but generally I am positive about this work, and hence the final score: \\u201cweak accept\\u201d.\\n\\nAdditional comments / questions:\\n(somewhat minor) p1: \\u201cWhile deep ensembles \\u2026, where the individual ensembles are trained on different data\\u201c - here and related text, it should probably be \\u201cindividual models\\u201d / \\u201cindividual networks\\u201d. Generally, I am not convinced that these are strong arguments against deep ensembles.\\n\\n(minor) p2-p3: \\u201c2. Preliminaries\\u201d - I am not sure if this section adds much to the understanding, it would seem more natural to spend more time explaining the intuitions behind the net\\n\\n(kind of major) p3. \\u201cprior\\u201d - The explanation of why using a randomly initialized network makes sense is not very strict. I kind of get the general idea, but it is not clear to me why not use something less expensive, e.g. just random projections, and why do we actually need a full network. Intuitively it seems quite strange to waste a lot of capacity to fit to essentially fit a set of random weights: is it something that allows the network to avoid easily learning the \\u201crandom prior\\u201d? And, more generally, can this also be considered as a \\u201ctrick\\u201d to de-correlate individual predictors? I believe these points should be discussed in more detail.\\n\\n<update>\", \"i_would_like_to_thank_authors_for_verbose_response_and_the_revised_version\": \"it is a bit more clear.\\nI stand by my original rating.\\n</update>\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": [\"This work introduces a simple technique to obtain uncertainty estimates for deep neural networks. This is achieved by having a set of random networks (i.e. neural networks where their parameters are randomly initialized) and then computing an uncertainty value based on the difference in the predictions between those random networks and networks that are trained to mimic them on a finite collection of points. The authors further show that this method results into uncertainties that are conservative, meaning that they are higher than the uncertainty of a hypothetical posterior, and concentrate, i.e. they converge towards zero when we get more and more data. The authors further draw connections to ensemble methods and discuss how such a method can be effectively realized in practice. They then evaluate their approach on an out-of-distribution detection task, they measure the calibration of their uncertainty estimates and finally perform a small ablation study for their concentration result.\", \"This work is interesting as it seems to provide a simple way to obtain reasonable uncertainty estimates. For this reason it can potentially serve as a strong baseline for this field. The theoretical considerations also help in providing some guarantees about such an approach. Having said that, in my opinion the writing could use some more work in order to make things more clear as some critical experimental details and baselines are missing and thus do not make the method as convincing. Furthermore, I also believe that some clarifications on the theoretical aspects of this work, will help in boosting its quality. More specifically:\", \"How exactly do you apply your method on the classification scenario? Do you select an arbitrary hidden layer of the classification model for the prior and predictor network architectures or the output logits / softmax probabilities? In appendix A you mention the architecture but not precisely how it is employed. I believe this can be an important piece of information in order to decipher the importance of e.g. the output dimensionality on the uncertainty quality, as higher dimensional outputs might be harder to approximate thus could induce a larger squared error and hence uncertainty.\", \"What is the average training error of the predictor networks for the out-of-distribution task and subsampling ablation task, i.e. how far away from concentration were the priors?\", \"An effect that I found weird is the following: what happens for the out-of-distribution examples when the predictor networks can perfectly predict the prior network outputs? Wouldn\\u2019t that then imply that the uncertainty would be zero for any input (even an out-of-distribution one), as the prior network and predictor network always agree? One could imagine that for e.g. simple priors and with sufficiently dense sampling of the domain of the function this can happen in practice.\", \"For the conservatism you show that your uncertainty estimate is higher, on average, than the posterior variance when you sample points from the model itself. In a sense this guarantee translates to the actual data when the prior is \\u201ccorrect\\u201d. How do those conservatism guarantees translate to the case when there is model misspecification, i.e. when the prior is not correct? Perhaps a small toy example would be informative.\", \"For the predictor networks as described in figure 2; do you train both the green and red parts of the network or only the red parts and keep the green part fixed to the values you used for the prior f? (This helps in understanding how easy / difficult is the task of the predictor network).\", \"What is the accuracy on the actual in-distribution prediction task for the RP and baselines? What did \\u201cB\\u201d correspond to for the dropout networks? Was it the number of dropout samples you averaged over to get the final predictive?\", \"How sensitive are the results on the actual initialization strategy of the prior network? It would be good to see e.g. some form of performance / init variance curve in order to decipher the sensitivity.\", \"Other comments\", \"It is worth pointing out that [1] showed that Monte-Carlo dropout performs approximate MAP inference, which seems more plausible than the approximate Bayesian inference perspective of [2].\", \"In the introduction you argue that Bayesian neural networks rely on procedures different from standard supervised learning and thus most ML pipelines are not optimized for them in practice. Could you elaborate a bit about this statement? Variationally trained BNNs with e.g. the reparametrization trick [3, 4] are straightforward since you can just use backpropagation to update their (variational) parameters.\", \"What is the x-axis for Figure 3 for the baselines? (I take it that for RP it is the \\\\hat{sigma}^2(x)).\", \"I believe that a comparison against a simple variationally trained BNN would make the results more convincing.\", \"Misc\", \"Second page, \\u201cFigure 1, top two plota\\u201d -> \\u201cFigure 1, top two plots\\u201d\", \"Third page, \\u201c[\\u2026] introduced in equation 2 denotes the posterior covariance [\\u2026.]\\u201d -> \\u201c[\\u2026] introduced in equation 2 denotes the posterior variance [\\u2026]\\u201c\", \"Fifth page, \\u201c[\\u2026] this makes it is reasonable for W large enough [\\u2026]\\u201d -> \\u201c[\\u2026] this makes it reasonable for W large enough [\\u2026]\\u201d\", \"Sixth page, \\u201cCorollary 1 and proposition 2\\u201d; where is corollary 1? Do you mean Proposition 1?\", \"Seventh page, \\u201c[\\u2026] inspired by, an builds on, [\\u2026]\\u201d -> \\u201cinspired by, and builds on, [\\u2026]\\u201d\", \"Ninth page \\u201cmontonicity\\u201d -> \\u201cmonotonicity\\u201d\", \"Overall, I tend to accept this work, although, depending on the author rebuttal and other discussions, I am willing to change my rating accordingly.\", \"[1] Eric Nalisnick, Jos\\u00e9 Miguel Hern\\u00e1ndez-Lobato, Padhraic Smyth, Dropout as a Structured Shrinkage Prior, 2019\", \"[2] Yarin Gal, Zoubin Ghahramani, Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning, 2016\", \"[3] Diederik P. Kingma, Max Welling, Auto-Encoding Variational Bayes, 2014\", \"[4] Danilo Jimenez Rezende, Shakir Mohamed, Daan Wierstra, Stochastic Backpropagation and Approximate Inference in Deep Generative Models, 2014\"]}", "{\"comment\": \"Thanks for the feedback on our paper's introduction Pranav!\\n\\nYou are right that the Bayes by backprop approach can be extended to use approximate posterior distributions which are non-Gaussian. Although Monte Carlo dropout can be viewed in this light, to do so requires a rather unnatural approximating family from the perspective approximate inference. It then requires a limit to be taken (like the convolutional one you and Hron point out) or Hron's generalisation of variational inference to a quasi-KL. An alternative view of MC dropout is as an ensemble method in which the ensemble members have shared parameters (meaning that they should therefore be trained together) and where the ensembling is applied at test time too. This latter view is arguably as natural as the Bayesian interpretation. For this reason we set MC dropout apart in the discussion of prior work.\\n\\nWe take your valuable feedback on board and will make these points -- including the connection between MC dropout and Bayes by backprop -- explicit in the updated version of the paper.\", \"title\": \"Thanks for the feedback.\"}", "{\"comment\": \"Hron et al. pointed out when trying to prove standard Dropouts as Bayesian Approx. we cannot take mixture of Impulses as the posterior, Gal et al. got around this problem by taking mixture of Gaussian Posterior albeit with very small sigma's ( of the order 10^(-34) ), Hron et al. called this trick, convolutional approach, so I don't think there is any difference between Dropouts and Bayes By Backprop. I hope you can clear this up in your paper. Good Work otherwise\", \"title\": \"Dropouts are still Bayesian\"}", "{\"comment\": \"Wow, thanks for the pointer. You are right that a discussion of [1] is due. Both works use the term \\\"prior networks\\\", but they use it to mean different things - [1] learns a network that parameterizes a Dirichlet prior, while our work uses fixed prior networks to obtain an upper bound on the Bayesian posterior.\\n\\nWe will add [1] to the prior work section when updates to the manuscript become possible again. \\n\\n[1] Predictive Uncertainty Estimation via Prior Networks, NeurIPS 2018\", \"title\": \"Thanks for the pointer.\"}", "{\"comment\": \"Great work and I really enjoy reading it.\\n\\nHowever, previous work has also studied the uncertainty estimation via prior networks. Please check out this paper [1]\\n\\nIn my opinion, a discussion/comparison seems due.\\n\\n\\n[1] Predictive Uncertainty Estimation via Prior Networks, NeurIPS 2018\", \"title\": \"A closely related paper\"}" ] }

[ "Benjamin Letham", "Roberto Calandra", "Akshara Rai", "Eytan Bakshy" ]

[ "Bayesian optimization", "high-dimensional", "Gaussian process" ]

{ "note_id": [ "hYLvNQJ6XW", "S1xgsdcnjS", "B1l_lGq3iH", "HJlVOb5njB", "r1xHz3YnoH", "SJg7FudRYr", "BJlTr1daYB", "Syli-Rq2Fr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751950, 1573853336124, 1573851632257, 1573851500477, 1573850125427, 1571879034626, 1571811141320, 1571757571138 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2552/Authors" ], [ "ICLR.cc/2020/Conference/Paper2552/Authors" ], [ "ICLR.cc/2020/Conference/Paper2552/Authors" ], [ "ICLR.cc/2020/Conference/Paper2552/Authors" ], [ "ICLR.cc/2020/Conference/Paper2552/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2552/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2552/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper explores the practice of using lower-dimensional embeddings to perform Bayesian optimization on high dimensional problems. The authors identify several issues with performing such an optimization on a lower-dimensional projection and propose solutions leading to better empirical performance of the optimization routine. Overall the reviewers found the work well written and enjoyable. However, the reviewers were concerned primarily about the connection to existing literature (R2) and the empirical analysis (R1, R3). The authors claim that their method outperforms state-of-the-art on a range of problems but the reviewers did not feel there was sufficient empirical evidence to back up this claim.\\n Unfortunately, as such the paper is not quite ready for publication. The authors claim to have significantly expanded the experiments in the response period, however, which will likely make it much stronger for a future submission.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Reply to Review 3\", \"comment\": \"Thank you for your in depth review of the paper. We especially appreciate the thoughts around improving aspects of modeling and interesting extensions. Some of the concerns with the paper are around modeling decisions. We made clarifying edits in the paper, and address each question below. The primary concern is around the empirical analysis. This was admittedly weak, so in the revision it has been significantly expanded. We now evaluate 5 distinct problems, with D from 72 to 1000. We compare a total of 14 methods. We added a non-synthetic problem to our experimental results. We feel the breadth of our experiments now exceeds that typical of HDBO papers in top conferences. We have also added the specific empirical results requested below. We hope that with these improvements you will be able to improve your score. If there are additional concerns that prevent you from recommending acceptance, please let us know so we can make appropriate adjustments.\\n\\n1) We are not exactly sure which previous work is being described. Binois et al. (2018) altered the REMBO projection to define a polytope that projects up to the same space as REMBO. However this was not directly incorporated into a constrained optimization problem. The reason (and perhaps the impracticality you reference) is possibly because acquisition function optimization used a genetic algorithm, which does not naturally support constraints. However, standard gradient optimizers naturally support linear constraints with little additional cost (we use Scipy SLSQP). Table 1 shows there was only a 10% increase in ALEBO running time when increasing D from 100 to 1000, which increases the linear constraints from 200 to 2000. With gradients, there are no issues with handling thousands of linear constraints.\\n\\n2) We originally selected MSE because it is typical for constructing GP learning curves. Per your request we have replaced that with the test-set log marginal prob. (Fig. 9), which did not alter the conclusions.\\n\\nTo better clarify, we added a new plot directly showing GP predictions on a test set (Fig. 8). The ARD RBF kernel simply predicts the mean, while the Mahalanobis kernel makes good (as in, useful for BO) predictions.\\n\\nWe are not aware of prior work on embeddings for HDBO that has shown GP fits within the embedding. We were surprised to discover that ARD RBF performs so poorly in the embedding, which led to the derivation of the appropriate kernel. The new figures and discussion in A.2 will make this more clear.\", \"ard_rbf_with_marginalization\": \"Here we show the behavior of the kernel typically used for BO in a linear embedding. Based on Fig. 8 and the theoretical result in Prop. 1 that ARD kernels are not appropriate for this setting, we don\\u2019t expect hyperparameter marginalization to have much impact.\\n\\n3) This was done to maintain Gaussian posteriors and thus analytic tractability of the acquisition function. This is a good idea and could be implemented using a fully MC-based acquisition while maintaining differentiability. We have prioritized improvements to exposition and additional experiments for this revision, but can add further discussion of this if needed.\", \"choosing_b\": \"The idea of selecting B to maximize Popt is very interesting. The LP of A.4 will be an important tool for this optimization. This is an interesting area of research for HDBO, though beyond this paper.\", \"why_laplace\": \"We initially tried HMC, which was far too slow to be practical. We then turned to the Laplace approximation, which is computationally efficient and provides uncertainty estimates good enough for BO. The goal of our work here is to investigate the apparent lack of robustness in linear embedding-based approaches to BO and address the failure modes. Using slice sampling is a great idea, and would provide better uncertainty estimates than the Laplace approximation with less cost than HMC. We hope to test this in the future, though we note that even with the Laplace approximation ALEBO still consistently performs at least as well as the other HDBO methods.\", \"log_regret\": \"We have added figures with log regret as requested (Fig. 12). We left \\u201craw\\u201d values in the main text because in real BO tasks the value of a solution is typically the function value. The utility behind acquisition functions like EI and KG is the function value. Log regret allows for identifying small differences in performance that may not be practically significant. With log regret, the improvement of ALEBO over other methods appears even larger than in Fig. 4.\", \"standard_test_functions\": \"We agree it can be tiresome to see the same few benchmark tasks in every paper, though there is value in standard problems making it easier to compare and understand results across papers. To address these concerns, we have included a more realistic example in the robot locomotion task.\\n\\nThank you again for taking the time to understand the paper at a technical level and for the quality of your feedback.\"}", "{\"title\": \"Reply to Review 2\", \"comment\": \"Thank you for your review of our paper, and in particular for recommending some additional methods to include in the benchmark experiments. Per your request, we have included LineBO (3 variants), the method of Binois et al. (2018), and also the method of Binois et al. (2015). The results of these methods can be seen in Figs. 4 and 11, and we give some discussion of them in response to your comments below. We hope that as a result of these improvements you will be able to raise your rating as indicated. If there are any additional items that prevent you from recommending acceptance, please let us know as we continue improving the paper.\\n\\n###\\n\\u201cHaving said that, I am a bit disappointed that this paper does not talk about LineBO (ICML 2019). LineBO is a good solution for high dimensions without any assumption on structure like low effective dimensionality. [\\u2026] Thus, LineBO is a stronger contender to the proposed algorithm. I will lift my rating if the author provide their response to this point.\\u201d\\n\\nThank you for pointing out the absence of LineBO in our discussion of HDBO methods. The LineBO paper primarily addresses the issue of performing \\u201csafe\\u201d BO in more than 1-2 dimensions: while previous methods (i.e., SafeOpt) involved discretizing the input space, LineBO is able to move along 1D subspaces within the search space.\\u00a0 In this sense, LineBO solves the problem of *high dimensional safe bayesian optimization*. However, the underlying probabilistic model in LineBO is still a standard GP, and so it does not help alleviate issues with estimating GPs with >20 dimensions, which is the subject of our work. Since they are focused on Safe BO, the LineBO methods are not necessarily appropriate for regular (non-safe) HDBO, and in fact most of the LineBO paper primarily focuses on problems that would not be considered high-dimensional for regular BO (but are for Safe BO). For instance the highest dimensional problem is only 24D; compare to 100D, the lowest in our paper.\\n\\nWe added LineBO to our experiments per your request, and you can see the results in Fig. 12, in the appendix, for the Hartmann6 D=100 problem. LineBO performs very poorly on this task, and in fact performs much worse than random. This result is consistent with those in the LineBO paper: The LineBO paper uses Hartmann6 D=20 as a benchmark problem, and even at D=20 they show that CoordinateLineBO requires about 400 iterations to perform better than random search. With 1200 iterations RandomLineBO and DescentLineBO still did not perform better than random. Increasing the dimensionality from 20 to 100 can only further decrease performance.\\n\\nUltimately, LineBO is meant for a different type of problem (Safe BO, not generic HDBO). We do not want to give readers the impression that LineBO is a bad method (since it is not, it is just not meant for this problem) so we included the LineBO comparison only in the appendix, where we give a paragraph to explain why the performance on this problem is as poor as it is (the last paragraph of A.6.2).\\n\\n###\\n\\u201cAdditionally, the author should compare their method to the algorithm of Binois et al (2018) that solved very well disadvantages of REMBO by setting bounds to avoid (1). Moreover, because the problem of the paper is high-dimensional Bayesian optimization under the assumption of low effective dimensionality, they should compare to other strong algorithms under the same assumption such as SI-BO algorithm (NIPS 2013) that used active learning to learn the low-dimensional subspace instead of using random embedding like REMBO.\\u201d\\n\\nWe have added a comparison to the algorithm of Binois et al. 2018 (REMBO-gamma k_Psi), as well as the algorithm of Binois et al. 2015 (REMBO-phi k_Psi) as requested. On the Branin problem these methods have similar performance as HeSBO and REMBO. On the Hartmann6 problem, REMBO-gamma k_Psi could not scale to D=1000, but for D=100 it was tied with ALEBO for the best-performing method.\\n\\nThe Djolonga et al. 2013 SI-BO paper is very interesting and we do give a brief description of it in the related work. It uses an active subspace method, with low-rank matrix active learning methods to approximate gradients. The work is oriented around bandit problems with much larger budgets than typical BO problems. In their 100D benchmarks, over 500 points are used just for the random initialization, and subsequent optimization occurred over thousands, rather than hundreds of iterations. Djolonga et al. do benchmark experiments with the same 100D augmented Branin function (p17 of the extended version) that we used in Figs. 4 and 12. They use 2500 iterations, vs. 50 in our experiments. The simple regret reported in their figure is higher than that seen for ALEBO in Fig. 11. Ultimately, SI-BO is a very interesting method with advantages in learning the subspace as you note, but is meant for problems with a sample budget orders of magnitude larger.\\n\\nThank you again for your review!\"}", "{\"title\": \"Reply to Review 1\", \"comment\": \"Thank you for your careful review of the paper. Points 1-5 primarily deal with aspects of the technical contribution of the paper that were unclear. We have updated the text to clarify these important points, and respond to each below. Point 6 deals largely with the empirical analysis. We now have 4 separate experiments in the main text that compare a total of 11 different methods for HDBO. These include a robot locomotion task. We also have 3 additional experiments in the supplement, with a total of 14 methods compared.\\n\\nThe improvements in the exposition and clarity of the text should resolve points 1-5, and the greatly expanded experiments resolve point 6. With these points addressed, we hope you will be able to increase your rating for the paper; please let us know of any additional issues that prevent you from recommending acceptance of the paper.\\n\\n1. Thm. 3 of the REMBO paper states that there exists an optimum in the embedding. However, there is no guarantee that that optimum will project up inside the box bounds! This means that if you are restricted to evaluating the function within the bounds (which is generally their purpose), REMBO has *no guarantee* that an optimum can be found in the embedding. We have better clarified this point by discussing it in Sec. 5.3. Why it matters that most points project to facets: Sec. 4 shows that points projected to the facet undergo a nonlinear projection that introduces kernel non-stationarity, an issue for GPs. The fact that nearly the entire volume of the space undergoes this nonlinear projection shows the severity of the issue. We have clarified this in the paper by re-ordering the first two points of Sec. 4, so we first describe the issue of nonlinear projections and then provide these results to indicate the extent of the issue. Number of samples is 1000, now in the text.\\n\\n2. The reason the GP is inappropriate is because of the nonlinear projection. ALEBO restricts BO to the portion of the embedding where the projection is linear (no clipping), allowing for GP modeling. We now discuss this in Sec. 5.2 to make it more clear. The paper actually had a visualization of the ALEBO embedding, in the appendix due to space constraints; it is Fig. 10. For the constrained space, the constraints form a convex, bounded polytope; since it is convex, we can be sure there are no discontinuities. We have clarified this by expanding the discussion in Section 5.2.\\n\\n3. We added to the text of that section (Sec. 4, \\u201cLinear projections...\\u201d) to describe this mathematically. We have also changed the reference to directly point to Prop. 1, which shows mathematically that a product kernel in the true space (ARD RBF) produces a not-product-kernel (Mahalanobis) in the embedding.\\n\\n4. The dagger denotes matrix pseudoinverse; we now state this in Sec. 5.\\n\\n5. We expanded Sec. 5.3 to derive this result.\\n\\n6. We have greatly expanded the set of empirical experiments. We promoted two experiments from the appendix to the main text so there are now benchmark results shown for three synthetic problems. One of these has D=1000, as requested. Another includes black-box constraints, so they constitute a diverse set of problems. We also added experiments for a 72-D robot locomotion task.\", \"guarantees_of_an_optimum_in_the_constraint_space\": \"let us re-emphasize that no linear embedding method has any guarantee of the embedding containing an optimizer when evaluations are restricted to box bounds. Thm 3 of the REMBO paper does not hold once you clip to box bounds. The probability the embedding contains an optimum will depend on the problem: D, d_e, and where the optimizer is in the embedding. In our paper we derive an unbiased estimator for this probability, under a prior for the location of the embedding (Sec. 5.3).\", \"all_linear_embedding_methods_inherently_have_rotational_invariance\": \"rotate the true subspace T to RT, and then rotate the embedding B to RB to produce the same problem.\", \"overall_conclusion\": \"In points 1-5 there were many aspects of the technical contributions of this paper that were not clear. We have added additional text to address all of these points, which we expect will help readers to understand these subtle technical issues.\\n\\nThe problem of HDBO is real and important, as it enables sample-efficient black-box optimization of any type of system. Many optimization problems involve more parameters than can be handled by a standard GP, and if work published in ICLR is any indication, it is plausible to expect that the underlying process can be represented within a low-dimensional subspace. We have also added a real HDBO problem to the paper, optimizing a 72-D controller for robot locomotion. These are real problems and the work we do in this paper produces a real advance in both technical understanding of the problem and in HDBO optimization performance.\\n\\nWe expect you will find the expanded experiment section as described above to be more compelling.\\n\\nThank you again for your review!\"}", "{\"title\": \"Revised version\", \"comment\": \"We are pleased to post a revised version of the paper with several improvements to address the points raised by the reviewers. We have significantly expanded the empirical evaluation of the method. The revision includes an additional benchmark experiment, 5 new benchmark comparison methods, empirical experiments with a robotics application problem, and improved exposition to clarify areas as indicated by the reviews. Thank you for your reviews and for the improvements they have brought to the paper.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This is a well-written paper and I enjoyed reading it. In summary the paper tries to address the following shortcomings of REMBO:\\n\\n(1)\\tREMBO uses a random embedding to project a point in high dimensional space to a lower dimensional embedding space basing on the relation f(x) = f(Ay) with high probability, where x in D dimensions and y in d dimensions and d<<D. The problem of REMBO is that the relation f(x) = f(Ay) is only guaranteed with high probability and so, the embedding space may not contain an optimum. Second, When a point y where Ay is outside the search space X, REMBO uses a projection to map Ay to its nearest point in X. This projection is not enough good. These observation are identified in paper of Binois et al (2018)as well.\\n(2)HESBO is an extension of REMBO that avoid above restrictions by proposing a new random projection. However, as the paper mentioned, HeSBO have a limitation is that the probability that the embedding will contain an optimum can be quite low! \\n(3)Besides, the paper also identify a new observation that linear projections do not preserve product kernels. \\n\\nThen, the paper proposes a new solution of BO to overcome these restrictions by using a Mahalanobis kernel to avoid (3). This kernel is a replace of ARD Euclidean distance to a Mahalanobis distance. To avoid (1), the paper use equation 1 (please find in the paper). To avoid (2), they use the projection P_opt. In all , I think the theoretical contribution is good enough.\\n\\nHaving said that, I am a bit disappointed that this paper does not talk about LineBO (ICML 2019). LineBO is a good solution for high dimensions without any assumption on structure like low effective dimensionality. It uses even one-dimensional subspaces to solve high dimensional problem with the strong theoretical guarantee. It do not need to learn subspace, and so it avoids disvantages (1), (2) and (3) that cause due to the fact that the embedding may not contain an optimum as mentioned above. Thus, LineBO is a stronger contender to the proposed algorithm. I will lift my rating if the author provide their response to this point.\\n\\nAdditionally, the author should compare their method to the algorithm of Binois et al (2018) that solved very well disadvantages of REMBO by setting bounds to avoid (1). Moreover, because the problem of the paper is high-dimensional Bayesian optimization under the assumption of low effective dimensionality, they should compare to other strong algorithms under the same assumption such as SI-BO algorithm( NIPS 2013) that used active learning to learn the low-dimensional subspace instead of using random embedding like REMBO.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper criticizes existing High Dimensional BO (HDBO) via linear embedding literature for the following reasons:\\n\\n- Points in the embedded space projected mostly to the facet of the bounding box in the original space.\\n- The projection induces a distorted space which is not fit to be modeled by a GP.\\n- Linear projections do not preserve product kernels.\\n- Linear embeddings have low probability of containing an optimum.\\n\\nThe paper then proposes ALEBO which supposedly improves these aspects over other linear embedding BO techniques.\\n\\nI have the following concerns regarding the authors' criticisms above:\\n\\n1. What is wrong with the points being projected mostly to the facet of the original bounding box?\\nIf I understand correctly, Theorem 3 of the REMBO paper proved that there exists y* \\\\in R^{d_e} such that f(Ay*) = f(x*)\\n(i.e., the projected space contains the optimum) with high probability so to me, it does not really matter if the projection does not include the interior of the bounding box. Fig. 1 seems to make a point that most projections seem to indeed land on the facet of the bounding box (to be thorough, how many points did the authors sample to make this plot) but given what I said above, I do not think there is much of a point in Fig. 1 here. \\n\\n2. The authors claimed that it is not appropriate to model the distorted space with GP, but ended up using GP \\nfor ALEBO anyway (although with a different kernel). I understand that the authors did not project Ay back to B\\nlike REMBO did, but the authors also gave me no reason to believe that this will improve things either. In fact, I think \\nthe authors should show the space induced by ALEBO embedding as a comparison. I suspect that with the imposed \\nconstraint -1 <= (B^T)y <= 1 the space will have discontinuous regions and is also not fit to be modeled \\nwith any GP. \\n\\n3. The authors stated that \\\"A product kernel in the true subspace will not produce a product kernel in the embedding; \\nwe will see this more explicitly in Sec. 5.1\\\" but I did not see it explicitly in Sec. 5.1. At the very\\nleast, I do not see how REMBO fails to do the same thing. Also, the authors claimed that \\\"Inside the embedding, \\nmoving along a single dimension will move across all dimensions of the true subspace, at rates depending \\non the angles between the embedding and the true subspace\\\". This seems like a very qualitative claim.\\nCan the authors formally define what this statement means, and prove it or at least provide some backup citations?\", \"other_comments\": \"4. Why do the authors use conjugate transpose (if B^T means what i think it means) instead of normal transpose \\nwhen B is drawn from R^{d_e \\\\times d}? Shouldn't they be the same?\\n\\n5. Please explain the choice of \\\\Epsilon(B) = {x: B^T B x = x} \\n\\n6. The experiments provided are very limited. There is only one set of experiments showing performance of\\nALEBO against other methods and it was done on a very small extrinsic dimension too (D = 100). I would like to see how \\nALEBO scales with truly large dimension (REMBO also claimed that it could scale up to much higher extrinsic dimension). What about other important properties like does it guarantee that an optimum lies in the constraint space? What about rotational invariance? There are so many elements missing from the analysis.\", \"overall_conclusion\": \"This paper is largely empirical and lacks technical depth. It is not at all convincing that the problem it\\naddresses is real, much less important. It also does not offer strong empirical evidence (too few experiments). \\nGiven these reasons, I do not think the paper is not ready to be published as it is.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors investigate pitfalls common to random embedding-based approaches to high-dimensional Bayesian optimization (HDBO). Each of several practical shortcomings is separately analyzed and, subsequently, addressed in straightforward fashion:\\n\\n a. Large-scale distortions caused by clipping are handled by generalizing box constraints\\n (in the embedded space) to the polytope corresponding to the set of points that project\\n to the interior of the original search space.\\n\\n b. Local distortions caused by defining squared Euclidean distances in embedded spaces\\n are handled by substitution for Mahalanobis distances.\\n\\n c. Embedded spaces potentially failing to contain optima are handled by constructing an\\n estimator for the probability of this happening, which is then be used to pick better\\n embeddings.\", \"feedback\": \"To the extent that I enjoyed reading this piece, I am not sure that it warrants publication at this time. Specifically, the degree of novelty on offer seems minimal and the empirical results are underwhelming.\\n\\n 1) Constraints on embedded candidate $\\\\mathbf{y}$ are naively defined in terms of a polytope, but this formulation has previously been deemed impractical. If nothing else, it would be good to clarify this matter: why did preceding works chose not to explore this direction and/or how did you make it work here?\\n\\n 2) Regarding use of Mahalanobis distances, evidence here (as provided in A.2) seems thin. Firstly, predictive MSE (Figure 6) seems like an odd choice of metric; log marginal probabilities would seemingly be more natural for GPs. Reporting of MSE is particularly suspect when results shown in Figure 7 indicate that the Mahalanobis distance based GPs are (markedly) overconfident. This issue is allegedly improved by marginalizing Mahalanobis parameters $\\\\Gamma$; however, the appropriate baseline here would be ARD with marginalized lengthscales (which, to my knowledge, is not shown).\\n\\n 3) Regarding ALEBO itself, Algorithm 1 states that acquisition functions were expressed in terms of an approximate posterior formed via moment matching against the Gaussian mixture formed by $m$ different samples of hyperparameters $\\\\Gamma$? One usually pushes this uncertainty through the acquisition function as, e.g., $\\\\mathbb{E}_{\\\\Gamma}[\\\\alpha(\\\\mathbf{x}; \\\\Gamma)]$. What motivated this design choice?\", \"questions\": [\"Does $\\\\mathcal{B}$ need to be sampled or can it chosen to maximize $P_{opt}$?\", \"Marginalization of hyperparameters\", \"Did you jointly marginalize over all hyperparamerters or just Mahalanobis parameters $\\\\Gamma$?\", \"Why was a Laplace approximation used lieu of, e.g., slice sampling?\", \"Nitpicks, Spelling, & Grammar:\", \"Various figures: 'NewMethod' -> 'ALEBO'\", \"Please report log immediate regret along with error bars\", \"To the extent that testing on e.g. \\\"high-dimensional\\\" variants of Branin and Hartmann-6 is standard, it isn't particularly convincing.\"]}" ] }

[ "Xin Zhang", "Weixiao Huang", "Renjie Liao", "Yanhua Li" ]

[ "Multi-agent", "Imitation Learning", "Inverse Reinforcement Learning" ]

{ "note_id": [ "GBLwLAMRgN", "rygutwUsjr", "Syx_TR9toB", "r1xeSirOiH", "ByxxFiMPsH", "SyxV3t78ir", "Syl8OYmLjB", "HylJxYmUsH", "ryesBumLoS", "r1e_Q73NsS", "rkeEN05Vor", "SkguwdqNoS", "SJxncUYNsS", "rklpwou4iH", "BJgnR9OVir", "HJegpGI7iB", "HJgcPaH7jH", "HJxEBSSXiS", "SylxzBS7iB", "BJeioPE7sH", "SkgrhmORFB", "SJgQFmBRYH", "HJeIPYyAtB", "rJxTbuHqKr", "S1gfZMOa_r", "r1l9cYFddS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798751920, 1573771135750, 1573658303532, 1573571384198, 1573493623554, 1573431724046, 1573431661760, 1573431527050, 1573431363189, 1573335840352, 1573330476380, 1573328991692, 1573324435974, 1573321572691, 1573321428202, 1573245623697, 1573244257884, 1573242171998, 1573242119596, 1573238691219, 1571877805241, 1571865467509, 1571842397577, 1571604484978, 1570763257619, 1570441617621 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2551/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2551/Authors" ], [ "~Anonymous_Someone1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper extends multi-agent imitation learning to extensive-form games. There is a long discussion between reviewer #3 and the authors on the difference between Markov Games (MGs) and Extensive-Form Games (EFGs). The core of the discussion is on whether methods developed under the MG formalism (where agents take actions simultaneously) naturally can be applied to the EFG problem setting (where agents can take actions asynchronously). Despite the long discussion, the authors and reviewer did not come to an agreement on this point. Given that it is a crucial point for determining the significance of the contribution, my decision is to decline the paper. I suggest that the authors add a detailed discussion on why MG methods cannot be applied to EFGs in the way suggested by reviewer #3 in the next version of this work and then resubmit.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Acknowledging here that I read the updated paper.\\n\\nI appreciate that the terminology has been cleared up, but unfortunately my issues with this work are not about terminology. \\n\\nThe authors are claiming that MAGAIL would model a turn based game in a grid world by assuming that an agent is \\\"standing still\\\" when really it's not their turn to act (and thus end up with agents standing still too much as shown in their video), when of course the loss should only look at agents when it *is* their turn. Even if you agree with this interpretation of the existing work, I think this \\\"extension\\\" is pretty obvious and is not sufficiently novel for publication at ICLR.\\n\\nUnfortunately I will not be able to continue this discussion any further, as I don't think we will reach agreement.\"}", "{\"title\": \"A kind reminder that an updated paper is available for your review.\", \"comment\": \"This is just a kind reminder to reviewer #3 that our updated paper is available for your further review. Please refer to the previous response for more details. We believe that we have addressed your concerns, i.e., the inconsistency of game-theory terminology with the literature.\\n\\nWe appreciate your detailed comments and constructive suggestions. Please let us know if you have any further questions.\"}", "{\"title\": \"An updated paper is available for your review.\", \"comment\": \"An updated paper is available for your review. Please refer to the previous response for more details. We appreciate your detailed comments and constructive suggestions.\"}", "{\"title\": \"Response to Turn-Based SGs\", \"comment\": \"Thanks for sharing the two references with us. We agree that turn-based games (or in general asynchronous decision-making games) can all be modeled (mapped) as Strategic Games/Markov Games (in a simultaneous decision-making fashion), by exactly the mapping mechanism you mentioned earlier, say, allowing multiple action sets for each agent. This is clear from the definition of $n$-person stochastic game $G$ in Sec 2 on page 5 in [1], which has a Selector function $\\\\sigma_i(s)$ (governed by the environment/game), indicating which action set to use in a particular round/turn. This Selector function is essentially the player function $Y$ we introduced in our work.\\n\\nHowever, we are solving multi-agent GAIL problems. The previous MA-GAIL work [RRR1] did not model the Selector Function (or our player function), thus it can only solve a subset of Markov Games (though it claimed Markov Games), where each player has only one action set and all players simultaneously play in each turn. \\n\\nGeneralizing MA-GAIL [RRR1] to AMA-GAIL to tackle this specific problem, we make two contributions: 1) explicitly modeling the Selector/Player function for Multi-agent GAIL problems under general asynchronous Markov Game settings, and 2) developing theoretical and algorithmic solutions to solve the problem. As a result, our contributions are novel and significant in advancing imitation learning techniques for multi-agent settings. \\n\\nAn updated paper is available now, where we have cited references you provided, and rephrased the game setting to emphasize our focus as (general) Markov Games, allowing asynchronous decision-making processes. MA-GAIL [RRR1] did not define Markov Games rigorously. \\nIn this sense, it is important for us to make it clear, so the readers in the community will have a clearer view of various game settings. We believe that now our terminology is consistent with the literature. \\n\\nFor your ease of reviewing the updates, we highlighted the game setting related changes in blue color texts. After your review, we will remove the text color by 11/15. Please kindly review the updated version, and let us know if you have any questions. \\n\\n[1] Chatterjee, Krishnendu, Rupak Majumdar, and Marcin Jurdzi\\u0144ski. \\\"On Nash equilibria in stochastic games.\\\" International Workshop on Computer Science Logic. Springer, Berlin, Heidelberg, 2004.\\n\\n[RRR1] Jiaming Song, Hongyu Ren, Dorsa Sadigh, and Stefano Ermon. Multi-agent generative adversarial imitation learning. In Advances in Neural Information Processing Systems, pp. 7461\\u20137472, 2018.\"}", "{\"title\": \"Turn-Based SGs\", \"comment\": \"The mapping of a turn-based game to an equivalent simultaneous-move SG is completely generic, and as far as I understand has nothing to do with the GAIL algorithm in particular. So if you're claiming that this reduction is novel/non-obvious, then does that mean you're claiming that it's not known that turn-based games can be modeled as a subclass of SGs via this (or a similar) reduction ? If so, I point you to e.g. [1] Section 5, or [2] Section 2, which both define turn-based SGs in their preliminaries in the same way I describe, and noting that turn-based SGs are a subclass of all SGs.\\n\\n\\n[1] Chatterjee, Krishnendu, Rupak Majumdar, and Marcin Jurdzi\\u0144ski. \\\"On Nash equilibria in stochastic games.\\\"\\u00a0International Workshop on Computer Science Logic. Springer, Berlin, Heidelberg, 2004.\\n\\n[2] Hansen, Thomas Dueholm, Peter Bro Miltersen, and Uri Zwick. \\\"Strategy iteration is strongly polynomial for 2-player turn-based stochastic games with a constant discount factor.\\\"\\u00a0Journal of the ACM (JACM)\\u00a060.1 (2013): 1.\"}", "{\"title\": \"A Modified Manuscript has been Uploaded\", \"comment\": \"We have added a few sentences at the end of the last section (Section 6 on page 8), to discuss the interesting idea you suggested, where our proposed player function $Y$ can in fact capture more general asynchronous game-playing scenarios, with multiple actions from each player. We also added a footnote on the last page (page 8) to acknowledge our reviewers for their contribution in bringing up this interesting point.\\n\\nWe greatly appreciate you and all reviewers for the detailed comments on our paper. We have carefully revised our paper based on these comments. As a result, we believe that the quality of the paper has been considerably improved. The new manuscript is now available. \\n\\nMoreover, we are still wondering if our previous response to your new mapping idea makes sense to you or not. Please let us know if you have any further questions.\"}", "{\"title\": \"A Modified Manuscript has been Uploaded\", \"comment\": \"We have updated Section 4 & 5 to address your concerns:\\n1) We modified eq. (14) in Section 4.1 on page 6, and the algorithm in Appendix B. on page 14. Now, the differences between MAGAIL and AMAGAIL are clearly highlighted.\\n2) We clearly explained how we generate and utilize expert trajectories for BC and MAGAIL, at the beginning of the third paragraph in Section 5 on page 7.\\n3) We cited [Ross and Bagnell, \\u201cEfficient reductions for imitation learning\\u201d], and explained why the performances of BC in Figure 2 does not increase as more demonstration data are used. (See the third paragraph of Sec 5.1 on page 7.)\\n\\nFor the minor comments, we have \\n1) We included the definition of initial state distribution as is highlighted in Section 2.1 on page 2, as \\u201cThe initial states are determined by a distribution $\\\\eta: \\\\mathcal{S} \\\\mapsto [0, 1]$.\\u201d\\n\\n2) We clarified the definition of each agent\\u2019s reward in the footnote of Section 2.1 on page 2, as \\u201cBecause of the asynchronous setting, the rewards only depend on agents' own actions.\\u201d\\n\\n3) Corrected the notation of \\u201cno-participaton\\u201d in Definition 1 in Section 3.2 on page 5. Now, we consistently use $\\\\phi$, rather than $0$. \\n\\n4) We corrected the typo of $\\\\zeta(i) = 1$ below Definition 1 in Section 3.2 on page 5.\\n\\n5) We updated the description that we don't have knowledge of transition $P$, but we can sample from it as a blackbox. (See our update in the last paragraph in Section 4.1 on page 6.)\\n\\nAll in all, we think that we have thoroughly answered and addressed your questions, and according to your constructive suggestions, we have updated our paper. As a result, we sincerely wish you could improve your rating for this work. Many thanks!\"}", "{\"title\": \"A Modified Manuscript has been Uploaded\", \"comment\": \"We have carefully revised our paper based on all reviewers\\u2019 comments. We believe that the quality of the paper has been considerably improved. The new manuscript is now available.\"}", "{\"title\": \"Author Response to Reviewer #3\", \"comment\": \"Thanks again for the response, and we sincerely appreciate your constructive suggestion. In fact, what you just suggested is identical to what we proposed in our paper.\\n\\nFirst, for the reward function, and policy function of each agent, yes, our proposed AMA-GAIL introduces the player function $Y$ as the gate function controlling whether the policy and reward functions ignore a particular agent at a certain round or not. Please refer to Figure 1 (b) and (c) in our submission for a visual illustration of AMA-GAIL, which matches the idea you outlined. \\n\\nSecond, when the multi-agent games involve asynchronous decision-making processes, the transition probability function $P$ (governed by the environment or the game settings) works exactly as you outlined, say, ignoring the non-acting agents, where the input is also controlled by the player function $Y$. \\n\\nOf course, in this case, each agent keeps only one action set, rather than multiple action sets.\\n\\nNote that MA-GAIL does not have the player function $Y$ (given its simultaneous decision-making nature) to enable the steps of ignoring agents in asynchronous games. \\n\\nDoes this match what is in your mind? If yes, that is great, since this is exactly what we proposed.\"}", "{\"title\": \"Response\", \"comment\": \"If I understand the author response correctly here, it is that the equivalence I proposed doesn't work for MA-GAIL because \\\"MA-GAIL as written only allows one action set for each agent\\\". In that case, how about the following mapping:\\n\\nAt each state, each player i has the same set of actions, but if an agent is not acting at a particular state then the transition and reward functions simply ignore that player's actions, e.g. if only player 1 acts at s_t, then P(s_{t+1} | s_t, a_1, a_2) = P(s_{t+1} | s_t, a_1).\\n\\nDoes that work?\"}", "{\"title\": \"Author Response to Reviewer #3\", \"comment\": \"Thanks again for your prompt response.\\n\\nFirst, [RR1] we provided clearly discusses and distinguishes MGs (Section 6.2 Stochastic games on page 159) and EFGs (Section 5.1 on page 117, especially, in definition 5.1.1 on page 118), where MGs are simultaneous decision-making processes, and EFGs are asynchronous decision-making, with a player function $\\\\rho$ defined on page 118.\\n\\n[RR1] Shoham, Yoav, and Kevin Leyton-Brown. Multiagent systems: Algorithmic, game-theoretic, and logical foundations. Cambridge University Press, 2008.\", \"http\": \"//www.eecs.harvard.edu/cs286r/courses/fall08/files/SLB.pdf\\n\\nMoreover, you mentioned that MGs and EFGs are *functionally* equivalent because of your statement in Response #2, and there exist mappings between MG <-->EFG. We have responded to it in *Author Response to Reviewer #3 (Part 2/2)*, which clearly explains why the simple mapping you outlined would not work for our AMA-GAIL problem. It seems that you overlooked it. For your ease of accessing it, we re-attach our response below. Please let us know if you have any questions:\\n\\n________________________________________\\nQuestion #2: Sorry for the confusion. I did not mean that no-op is an \\\"additional action\\\" for the agent, but rather that agents *only have a single no-op action* at decision points where they do not act. E.g. to model chess as a SG, the specification of the *environment* is such that the agent has a choice of moves when it is their turn, and has only a single no-op move when it is the other player's turn.\\n\\n________________________________________\\nResponse #2: Thanks for the clarification. It is an interesting idea to model the no-op move as an action set (with only one choice though). In this case, the agent has two action sets for different rounds (participation rounds vs no-participation rounds). Such a model is still an asynchronous decision-making case, and cannot be handled by MA-GAIL. Please find the detailed explanations below.\\n\\nSuch modeling matches a general asynchronous decision-making scenario, because though each agent makes an action at each round, the action set the agent uses is still governed/chosen in an ASYNCHRONOUS fashion by either a turn-based player function (in deterministic game), or more generally, by a stochastic player function, e.g., the action set is chosen by a (conditional) distribution defined by the environment. As a result, using such a multi-action-sets modeling, the multi-agent imitation learning with asynchronous decision-making processes cannot be simply solved by MA-GAIL [RR4], because MA-GAIL only allows one action set for each agent, namely, each agent takes an action at each round from the same action set. To allow multi-action-sets, an environment-defined function needs to be introduced, which is exactly our proposed player function. \\n\\nWe like this discussion. Thank you for bringing up interesting ideas. This reminds us that our player function can not only model whether an agent participates (i.e., choosing an action between a single no-op action set and a regular action set), but also support switching between multiple (N>=2) action sets. We will try to incorporate this interesting discussion into our paper (and acknowledge the reviewers). \\n\\n[RR4] Song, Jiaming, et al. \\\"Multi-agent generative adversarial imitation learning.\\\" Advances in Neural Information Processing Systems. 2018. https://arxiv.org/pdf/1807.09936.pdf\"}", "{\"title\": \"Response\", \"comment\": \"I was not claiming that the formal definition of MG and EFGs are identical; the definitions are slightly different, I think due to historical reasons of which games each community was looking at, but they are *functionally* equivalent because as described in Response #2, any MG can be mapped to an equivalent EFG and vice versa. I described the EFG --> MG mapping in Response #2, and there is also a (slightly less trivial but still simple) mapping from a MG --> EFG. Since you can map any EFG to an equivalent MG, an algorithm for one also solves the other.\\n\\nThe references you provide each give formal definitions for one class or the other, but notice that there is no reference that discusses *both* MGs and EFGs and points out the differences between them (because there is not a functional difference, therefore authors choose to work with one or the other).\"}", "{\"title\": \"Author Response to Reviewer #3 (Part 1/2)\", \"comment\": \"We appreciate your prompt reply. Please find our point-by-point responses below.\\n\\nQuestion #1: The main criticism is that I don't think that MDPs and EFGs are really two separate classes of problem, they're just different terminology used by different communities, with a slightly different way of representing the same class of problems. In [R3] for example, notice in slide 8 they say \\\"One example of a two-player zero-sum stochastic game is Backgammon\\\", which has \\\"two agents who take turns\\\". So it seems that SG includes turn-based games. I can find nowhere in the literature that talks about SGs and EFGs being distinct classes of problem based on whether actions are simultaneous, but I'd be happy to be pointed to it.\\n\\nResponse #1: In fact, MDPs and EFGs are two classes of problems: MDPs model SINGLE agents\\u2019 decision making processes, and EFGs characterize games involving MULTIPLE agents. We guess you meant to say \\u201cMarkov games (MGs)/Stochastic Games (SGs)\\u201d, rather than \\u201cMDPs\\u201d. \\n\\nMGs/SGs generalize MDPs from a single agent to multi-agent scenarios [RR4],[RR5]. However, MGs and EFGs are still two classes of problems, with MGs for simultaneous decision-making and EFGs for more general cases allowing asynchronous decision-making. Please find strong evidence below.\\n\\n1. In [RR2] (Page 1) by Prof Ramesh Johari, stochastic games are clearly defined as follows:\\n``(Stochastic Games) The game starts in state $x^0$ . At each stage t, all players simultaneously choose (possibly mixed) actions $a^t_ i$ , with possible pure actions given by the set $A_i(x^t)$.\\u2019\\u2019\\n\\n2. In Section 6.2 on Page 159 in [RR1], stochastic games are defined as below, namely, a collection of normal-form games, where each normal-form game (as defined in earlier chapters in [RR1] and the Wikipedia page [RR3]) involve all agents making decisions simultaneously. \\n\\n\\u201cA stochastic game is a collection of normal-form games; the agents repeatedly play games from this collection, and the particular game played at any given iteration depends probabilistically on the previous game played and on the actions taken by all agents in that game.\\u201d\\n\\n3. In MA-GAIL [RR4] and MA-AIRL [RR5], the authors explicitly define MGs with all players' simultaneous decision-making at each round. Please refer to Sec 2.1 on page 2 and Sec. 4.1 on page 8 in [RR4] and Sec 2.1 on page 2 and Sec. 3.3 on page 5 in [RR5] for more details.\\n\\n[RR1] Shoham, Yoav, and Kevin Leyton-Brown. Multiagent systems: Algorithmic, game-theoretic, and logical foundations. Cambridge University Press, 2008.\", \"http\": \"//www.eecs.harvard.edu/cs286r/courses/fall08/files/SLB.pdf\\n[RR2] Ramesh Johari. Lecture 4: Stochastic games. 2007. http://web.stanford.edu/~rjohari/teaching/notes/336_lecture4_2007.pdf\\n[RR3] Normal-Form Games, Wiki page: https://en.wikipedia.org/wiki/Normal-form_game\\n[RR4] Song, Jiaming, et al. \\\"Multi-agent generative adversarial imitation learning.\\\" Advances in Neural Information Processing Systems. 2018. https://arxiv.org/pdf/1807.09936.pdf\\n[RR5] Yu, Lantao, Jiaming Song, and Stefano Ermon. \\\"Multi-Agent Adversarial Inverse Reinforcement Learning.\\\" arXiv preprint arXiv:1907.13220 (2019). http://proceedings.mlr.press/v97/yu19e/yu19e.pdf\"}", "{\"title\": \"Author Response to Reviewer #3 (Part 2/2)\", \"comment\": \"Question #2: Sorry for the confusion. I did not mean that no-op is an \\\"additional action\\\" for the agent, but rather that agents *only have a single no-op action* at decision points where they do not act. E.g. to model chess as a SG, the specification of the *environment* is such that the agent has a choice of moves when it is their turn, and has only a single no-op move when it is the other player's turn.\\n\\nResponse #2: Thanks for the clarification. It is an interesting idea to model the no-op move as an action set (with only one choice though). In this case, the agent has two action sets for different rounds (participation rounds vs no-participation rounds). Such a model is still an asynchronous decision-making case, and cannot be handled by MA-GAIL. Please find the detailed explanations below.\\n\\nSuch modeling matches a general asynchronous decision-making scenario, because though each agent makes an action at each round, the action set the agent uses is still governed/chosen in an ASYNCHRONOUS fashion by either a turn-based player function (in deterministic game), or more generally, by a stochastic player function, e.g., the action set is chosen by a (conditional) distribution defined by the environment. As a result, using such a multi-action-sets modeling, the multi-agent imitation learning with asynchronous decision-making processes cannot be simply solved by MA-GAIL [RR4], because MA-GAIL only allows one action set for each agent, namely, each agent takes an action at each round from the same action set. To allow multi-action-sets, an environment-defined function needs to be introduced, which is exactly our proposed player function $Y$. \\n\\nWe like this discussion. Thank you for bringing up interesting ideas. This reminds us that our player function can not only model whether an agent participates (i.e., choosing an action between a single no-op action set and a regular action set), but also support switching between multiple (N>=2) action sets. We will try to incorporate this interesting discussion into our paper (and acknowledge the reviewers). \\n\\n[RR4] Song, Jiaming, et al. \\\"Multi-agent generative adversarial imitation learning.\\\" Advances in Neural Information Processing Systems. 2018. https://arxiv.org/pdf/1807.09936.pdf\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thanks for the response.\\n\\nRe Response #1: The authors are correct that I put an incorrect emphasis on the distinction between Perfect Info and Imperfect Info being the \\\"fundamental difference\\\" between stochastic games / EFGs. I was talking about the typical way these terms are used, but you're right that people study \\\"perfect info EFGs\\\" and \\\"partially observed MGs\\\". Arguing over this distinction is a distraction from my main criticism.\\n\\nThe main criticism is that I don't think that MDPs and EFGs are really two separate classes of problem, they're just different terminology used by different communities, with a slightly different way of representing the same class of problems. In [R3] for example, notice in slide 8 they say \\\"One example of a two-player zero-sum stochastic game is Backgammon\\\", which has \\\"two agents who take turns\\\". So it seems that SG includes turn-based games. I can find nowhere in the literature that talks about SGs and EFGs being distinct classes of problem based on whether actions are simultaneous, but I'd be happy to be pointed to it.\\n\\nRe Response #2: Sorry for the confusion. I did not mean that no-op is an \\\"additional action\\\" for the agent, but rather that agents *only have a single no-op action* at decision points where they do not act. E.g. to model chess as a SG, the specification of the *environment* is such that the agent has a choice of moves when it is their turn, and has only a single no-op move when it is the other player's turn.\"}", "{\"title\": \"Author Response to Reviewer #3\", \"comment\": \"Thank you for your review and the time taken for it, below we address each of your questions in turn.\\n\\nQuestion #1: After reading Sections 1 and 2, I believe that the authors have a misunderstanding of the relationship between Markov Games and Extensive Form Games. The fundamental difference between these two formalisms is that Markov Games aka Stochastic Games (https://en.wikipedia.org/wiki/Stochastic_game) are fully observed while EFGs in general are not fully observed. EFGs are called Partially-Observed Markov Games in the RL literature. \\\"Go\\\" is actually a MG, in contradiction with the authors' statement in the intro.\\n\\nResponse #1: Thanks for pointing this out, but we do not agree with your statement. The difference between Markov games vs. Extensive-form games has nothing to do with full vs partial observability.\\n\\nMarkov games (aka stochastic game) are normal-form games with multiple stages, and at each stage all agents need to make simultaneous decisions. In Extensive-form games (or extensive games), agents make asynchronous decisions over game stages (See Page 89 Chapter 6.1.1, line 1 in [R1], and [R3]). On the other hand, fully vs. partially observable games are also called perfect information games vs. imperfect information games (See [R1]). For a game, being a perfect or imperfect information game is totally orthogonal to whether it is a Markov game or an extensive-form game. For example, [R2] discussed that Markov games can be perfect information or imperfect information; [R1] discussed that extensive games can be perfect information (Part 2 in [R1]) or imperfect information (Part 3 in [R1]). Chess is an extensive game with perfect information (Page 100 in [R1]), so is Go. \\n\\nBTW, we checked the wiki link provided by the reviewer, which does not seem to support the statement of \\u201cstochastic games are fully observed\\u201d at all.\\n\\n[R1] Osborne, Martin J., and Ariel Rubinstein. A course in game theory. MIT press, 1994. http://ebour.com.ar/pdfs/A%20Course%20in%20Game%20Theory.pdf\\n[R2] Hansen, Eric A., Daniel S. Bernstein, and Shlomo Zilberstein. \\\"Dynamic programming for partially observable stochastic games.\\\" In AAAI, vol. 4, pp. 709-715. 2004. https://www.aaai.org/Papers/Workshops/2004/WS-04-08/WS04-08-005.pdf\\n[R3] Introduction to Game, University of Maryland, https://www.cs.umd.edu/users/nau/game-theory/8%20Stochastic%20games.pdf\\n\\nQuestion #2: In fact, turn-based games (even when the turn order is dynamic/stochastic) is easily handled by the MG formalism, by assigning \\\"no-op\\\" moves to players who are not active at this decision point.\\n\\nResponse #2: Good point, but you cannot simply model \\u201cno-op\\u201d move (i.e., no-participation) as an additional action that the agent can choose in AMA-GAIL problem, because \\u201cno-op\\u201d (i.e., no-participation) itself is out of control of agents, and it is purely controlled/governed by the environment (e,g., in a stochastic turn-based game, the environment may by chance block some agents from participating in the game in certain rounds). In fact, when we implemented MA-GAIL in evaluations, we took the \\u201cno-participation\\u201d as an action for agents, and Fig 2 (a)-(c) show the comparison results with AMA-GAIL, and BC, and our AMA-GAIL outperforms the other baselines.\\n\\nQuestion #3: I don't understand the necessity of (t+1)-step constraints in order the achieve a SPE.\\n\\nResponse #3: Thanks for pointing this out. Given the AMA-RL problem with Subgame Perfect Equilibrium (SPE) constraints defined in eq.(8)-(9), we are not using (t+1)-step constraints to achieve the SPE. Instead, we use (t+1)-step constraints to find the corresponding AMA-IRL problem eq.(12), in a consistent form as MA-IRL in MA-GAIL (Song et al.2018) and IRL in GAIL (Ho et al. 2016). With eq.(12), we can further formulate the AMA-RL $\\\\circ$ AMA-IRL problem in (Theorem 3 and eq.(13)) and employ the GAN framework to solve it.\"}", "{\"title\": \"Author Response to Reviewer #1 (Part 2/2)\", \"comment\": [\"Question #6: There are some minor comments:\", \"In 2.1., $\\\\eta$(initial state distribution) is not explicitly defined.\", \"In 2.1., MGs assume each agent\\u2019s reward function depends on other agents\\u2019 actions as well as agents\\u2019 own actions, but in the submission, rewards only depend on agents\\u2019 own actions. This may be due to the asynchronous setting, but I think it should be mentioned in the paper.\", \"In Definition 1, the null action is described as 0, whereas it was defined as $\\\\phi$ in 2.1.\", \"In a sentence below Definition 1, $\\\\eta (i) = 1$-> $\\\\zeta (i)=1$.\", \"Below (14), we don\\u2019t have full knowledge of transition P, but we can sample from it (like black-box model).\", \"Response #6: Thanks for pointing these out. We will update our paper accordingly and upload the new version by this weekend.\", \"The initial states are determined by a distribution $\\\\eta : \\\\cal{S} \\\\mapsto [0, 1]$ like in MA-GAIL (Song et al.2018).\", \"Yes. It is due to the asynchronous setting. We will make it clear in our updated version.\", \"We will correct the expression in Definition 1 and consistently use $\\\\phi$.\", \"We will fix it. Thank you for pointing it out.\", \"Yes. This is exactly what we are doing. We will surely make it clear in the next version.\"]}", "{\"title\": \"Author Response to Reviewer #1 (Part 1/2)\", \"comment\": \"We sincerely appreciate your careful review of our work and the precise summarization of the paper. We are also grateful for your comments and have tried very hard to address your and other reviewers\\u2019 concerns, as detailed in our point-by-point responses below.\\n\\nQuestion #1: I think section 4 (Practical Asynchronous Multi-Agent Imitation Learning) and section 5 (Experiments) should be much clearly written.\\n\\nResponse #1: Thank you for your suggestion. We are revising our section 4 and 5 to make them more clear. We will make the updated version available over the weekend. \\n\\nQuestion #2: It seems that the key difference between MA-GAIL and AMA-GAIL is whether we consider the cost function when the indicator is 0 or not, but it\\u2019s difficult to figure out just by comparing (3) (MA-GAIL objective) and (14) (AMA-GAIL objective) at the first glance. Similarly in Appendix B, it\\u2019s difficult to figure out the difference between MA-GAIL algorithm and AMA-GAIL except for the fact that eMGs are assumed. I think some additional explanation is needed.\\n\\nResponse #2: Thank you for pointing out our carelessness. The proposed AMA-GAIL is a more general problem to MA-GAIL, where diverse player participation scenarios are modeled by the introduced player function $Y$. In eq. (14), the two expectations $\\\\mathbb{E}$ should both be with a subscript of $Y$ as $\\\\mathbb{E}_{\\\\pi_{\\\\theta},Y}$, and $\\\\mathbb{E}_{\\\\pi_{E},Y}$, respectively. Such an expectation definition is clearly defined in Sec 2.1 (Extensive Markov Games), i.e., the last sentence in Sec 2.1. We will fix this in the updated version.\\n\\nQuestion #3: How did you generate expert trajectories in eMGs setting? Suppose there is an agent taking an action \\u201c1\\u201d at time t, but it was not applied to the dynamics because the indicator variable of the agent is equal to 0 at time t. In this case, what would be stored in the expert trajectories? \\u201cNull\\u201d or \\u201c1\\u201d? If the agent cannot take an action in advance (before looking at its indicator variable), I think adding indicator variables in a condition of policy, e.g., $\\\\pi(a|s,i) = \\\\pi(a|s)$ if $i=1$ , otherwise $a=$\\\"Null\\\" , is mathematically rigorous.\\n\\nResponse #3: Thanks for your questions and suggestions. It is \\u201cnull\\u201d. Yes. As you mentioned, the agent cannot take an action in advance (before looking at its indicator variable). Yes. Adding indicator variables in a condition of policy is absolutely correct mathematically. We still prefer to separate them 1) for the consistency with MA-GAIL paper (Song et al.2018), using the same form in $\\\\pi$, and 2) for highlighting the difference from MA-GAIL, by the player function $Y$.\\n\\nQuestion #4: Assuming that experts\\u2019 trajectories include \\u201cNull\\u201d actions, how did you use MA-GAIL and BC with those trajectories? \\n\\nResponse #4: In the implementation, we added \\u201cnull\\u201d (no-participation) as an additional action to each agent\\u2019s action set in MA-GAIL and BC experiments. The detailed results are shown in Figure 2, Table 1, and Appendix C.1 & C.2. They all show that our AMA-GAIL outperforms other baselines, i.e., MA-GAIL and BC. This is because \\u201cno-participation\\u201d itself is out of control of agents, and it is purely controlled/governed by the environment (e,g., in a stochastic turn-based game, the environment may by chance block some agents from participating in the game in certain rounds).\\n\\nQuestion #5: The performance of BC seems weird to me since adding lots of training data reduces supervised learning errors and can also reduce covariate shift problems in BC since the theorem tells us that the regret is bounded by (error) * (episode length) ^ 2 in the worst case [Ross and Bagnell, \\u201cEfficient reductions for imitation learning\\u201d]. Such a tendency is empirically shown in MA-GAIL paper as well, i.e., the performance of BC increases as the amount of expert trajectories increases. Is there any reason, BC shows poor performance in eMGs?\\n\\nResponse #5: Thanks for your comments and for providing the reference link to us. We will cite it in our updated version. As for the performance of BC, in Figure 2(b) the performance is poor at the beginning but increases rapidly and then converges at around 0.65 with 300 demonstrations. This is, in fact, consistent with MA-GAIL work, because in MA-GAIL, they only evaluated up to 400 demonstrations, where we evaluated from 200 to 1000 demonstrations.\\n\\nMoreover, in Figure 2(a), deterministic cooperative navigation is easier to learn compared with the stochastic cooperative navigation game shown in Figure 2(b), since there is no randomness in the player function. The performance from the beginning (200 demonstrations) has already stabilized at 0.7.\"}", "{\"title\": \"Author Response to Reviewer #2\", \"comment\": \"We really appreciate your precise summarization along with positive remarks about our paper. We are encouraged to work hard to improve the quality of the paper.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The submission extends the MARL\\u25e6MAIR to the extensive Markov game case, where the decisions are made asynchronously. As a result, a stronger equilibrium SPE is becomes the target of the proposed method. To this end, the submission takes advantage of the previous game theory results, to formulate the problem, and transform the model to a MAGAIL form. The empirical performance of the proposed method is demonstrated using experiments.\\n\\nI believe the submission considers an interesting and challenging problem, and has extended the existing multi-agent IRL methods to the extensive Markov game case.\"}", "{\"title\": \"A short video showing AMAGAIL and MAGAIL agents performing tasks considered in this work\", \"comment\": \"We would like to share a video showing AMAGAIL and MAGAIL agents performing tasks considered in this work:\", \"https\": \"//youtu.be/xojMJvAYB-g .\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"In this work, a multi-agent imitation learning algorithm for extensive Markov Games is proposed. Compared to Markov Games (MGs), extensive Markov Games (eMGs) introduces indicator variables, which means whether agents will participate in the game at the specific time step or not, and player function, which is a probability distribution of indicator variables given histories and assumed to be governed by the environment, not by the agents. Such a model allows us to consider asynchronous participation of agents, whereas MGs only consider synchronous participation, which is assumed in the existing multi-agent imitation learning algorithms such as MA-GAIL and MA-AIRL.\", \"The contribution of this submission can be summarized as follows. From a theoretical perspective, the submission extends the theorems in MA-GAIL to those in eMGs, where most of them deal with Lagrange multiplier, its meaning, and properties. Followed by Theorem1 and 2, authors define an extensive occupancy measure, a natural extension of occupancy measures in MGs, and cast a multi-agent imitation learning problem into extensive occupancy measure matching problem in Theorem 3. For a practical algorithm, AMA-GAIL is proposed and shown to have a performance gain relative to BC and MA-GAIL.\", \"The submission is highly interesting, but I think section 4 (Practical Asynchronous Multi-Agent Imitation Learning) and section 5 (Experiments) should be much clearly written. The followings are comments regarding those sections:\", \"It seems that the key difference between MA-GAIL and AMA-GAIL is whether we consider the cost function when the indicator is 0 or not, but it\\u2019s difficult to figure out just by comparing (3) (MA-GAIL objective) and (14) (AMA-GAIL objective) at the first glance.\", \"Similarly in Appendix B, it\\u2019s difficult to figure out the difference between MA-GAIL algorithm and AMA-GAIL except the fact that eMGs are assumed. I think some additional explanation is needed.\", \"How did you generate expert trajectories in eMGs setting? Suppose there is an agent taking an action \\u201c1\\u201d at time t, but it was not applied to the dynamics because the indicator variable of the agent is equal to 0 at time t. In this case, what would be stored in the expert trajectories? \\u201cNull\\u201d or \\u201c1\\u201d? If the agent cannot take an action in advance (before looking at its indicator variable), I think adding indicator variables in a condition of policy, e.g., $\\\\pi(a|s, i)=pi(a|s)$ if $i=1$, otherwise $\\\\mathbb{I}\\\\{a=\\\"Null\\\"\\\\}$, is mathematically rigorous.\", \"Assuming that experts\\u2019 trajectories include \\u201cNull\\u201d actions, how did you use MA-GAIL and BC with those trajectories?\", \"The performance of BC seems weird to me since adding lots of training data reduces supervised learning errors and can also reduce covariate shift problems in BC since the theorem tells us that the regret is bounded by (error) * (episode length) ^ 2 in the worst case [Ross and Bagnell, \\u201cEfficient reductions for imitation learning\\u201d]. Such a tendency is empirically shown in MA-GAIL paper as well, i.e., performance of BC increases as the amount of expert trajectories increases. Is there any reason, BC shows poor performance in eMGs?\"], \"there_are_some_minor_comments\": [\"In 2.1., $\\\\eta$ (initial state distribution) is not explicitly defined.\", \"In 2.1., MGs assume each agent\\u2019s reward function depends on other agents\\u2019 actions as well as agents\\u2019 own actions, but in the submission, rewards only depend on agents\\u2019 own actions. This may be due to the asynchronous setting, but I think it should be mentioned in the paper.\", \"In Definition 1, null action is describe as $0$, whereas it was defined as $\\\\phi$ in 2.1.\", \"In a sentence below Definition 1, $\\\\eta(i)=1$ -> $\\\\zeta(i)=1$.\", \"Below (14), we don\\u2019t have full knowledge of transition P, but we can sample from it (like black-box model).\"]}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I made a quick assessment of this paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"After reading Sections 1 and 2, I believe that the authors have a misunderstanding of the relationship between Markov Games and Extensive Form Games. The fundamental difference between these two formalisms is that Markov Games aka Stochastic Games (https://en.wikipedia.org/wiki/Stochastic_game) are fully observed while EFGs in general are not fully observed. EFGs are called Partially-Observed Markov Games in the RL literature. \\\"Go\\\" is actually a MG, in contradiction with the authors' statement in the intro.\\n\\nThe authors make an artificial distinction that \\\"turn-based\\\" games cannot be handled under the MG formalism. In fact, turn-based games (even when the turn order is dynamic/stochastic) is easily handled by the MG formalism, by assigning \\\"no-op\\\" moves to players who are not active at this decision point. Therefore, I don't see why any additional mechanics are needed to apply MA-GAIL to turn-based games.\\n\\nSimilarly, I don't understand the necessity of (t+1)-step constraints in order the achieve a SPE, which appears to be the central contribution of this work. The one-shot deviation principle still holds even if some players' actions are \\\"no-ops\\\" at certain decision points.\\n\\nI did not read the proofs or experiments closely since I believe there are flaws in the central idea of this work. I'd be happy to do a more thorough review if I am incorrect about these central points.\"}", "{\"comment\": \"We sincerely appreciate your constructive comments. Given our AMA-RL problem with Subgame Perfect Equilibrium (SPE) constraints defined in eq.(8)-(9), our goal is to find the formulation of AMA-IRL problem, that has a consistent format to MA-IRL in MAGAIL (Song et al. 2018) and IRL in GAIL (Ho et al. 2016). This is challenging for the non-convex AMA-RL problem. As a result, we \\\"relax\\\" the primal problem by choosing a particular \\\\lambda as shown in Theorem 2, which leads to the consistent AMA-IRL formulation (in eq.(12)) to that of MA-IRL and single agent IRL. Such relaxation enables us to solve AMA-RL \\u25e6 AMA-IRL with GANs models (as outlined in Sec 4). Again, we never tried to use the specific $\\\\lambda$ to solve the dual, nor the original AMA-RL problem.\\n\\n(Song et al. 2018) Jiaming Song, Hongyu Ren, Dorsa Sadigh, and Stefano Ermon. Multi-agent generative adversarial imitation learning. In Advances in Neural Information Processing Systems, pp. 7461\\u20137472, 2018. \\n\\n(Ho et al. 2016) Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In Advances in Neural Information Processing Systems, pp. 4565\\u20134573, 2016.\", \"title\": \"Clarification for the choice of \\\\lambda\"}", "{\"comment\": \"This paper proposed AMAGAIL, which extends MAGAIL into extensive Markov game. While it is a good idea to solve the imitation learning problem in extensive Markov game with good writing, I find some weaknesses which can be improved.\\n\\nThis paper follows the structure of MAGAIL, which firstly presents the original problem of asynchronous multi-agent adversarial reinforcement learning and then shows its dual. Then get the AMAIRL objective function by a specific choice of $\\\\lambda$.\", \"as_the_paper_says\": \"\\\"Theorem 2 illustrates that a specific $\\\\lambda$ is able to recover the difference of the sum of expected rewards\\nbetween not all optimal and all optimal policies.\\\" and\\n\\\"Theorem 2 (proved in Appx A.3) provides a horizon to establish AMAIRL objective function with\\nregularizer .\\\"\\n\\nThen it is meaningless to propose the dual problem since this specific Lagrange multiplier do not solve the dual problem, not to say the original problem. All the derivations which cover most of the methodology just become the \\\"horizon to establish\\\" the objective function. And to be honest, this specific choice can be revealed through a reversed derivation starting from the final objective, which makes no difference with the MAGAIL objective.\", \"title\": \"A specific choice of $\\\\lambda$ makes the dual problem meaningless.\"}" ] }

[ "Xingyu Lu", "Pieter Abbeel", "Stas Tiomkin" ]

[ "reinforcement learning", "representation learning", "reward shaping", "predictive coding" ]

{ "note_id": [ "-Ug2fCPXLB", "rkgOPynosB", "Hkg0pRsjiS", "BklKzCojsS", "SkgcmjEaFH", "Byeo7VT3tS", "HJg3kV22FS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751884, 1573793631885, 1573793477635, 1573793296833, 1571797793552, 1571767330552, 1571763172086 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2550/Authors" ], [ "ICLR.cc/2020/Conference/Paper2550/Authors" ], [ "ICLR.cc/2020/Conference/Paper2550/Authors" ], [ "ICLR.cc/2020/Conference/Paper2550/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2550/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2550/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes to use the representation learned via CPC to do reward shaping via clustering the embedding and providing a reward based on the distance from the goal.\\n\\nThe reviewers point out some conceptual issues with the paper, the key one being that the method is contingent on a random policy being able to reach the goal, which is not true for difficult environments that the paper claims to be motivated by. One reviewer noted limited experiment runs and lack of comparisons with other reward shaping methods.\\n\\nI recommend rejection, but hope the authors find the feedback helpful and submit a future version elsewhere.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Thank you very much for your insightful review. Below are our responses to your questions:\\n\\n1. How is the \\u2018success rate\\u2019 computed (e.g. in figure 7 and table 1).\\n\\n\\u201cSuccess\\u201d in the grid word domains means the agent is able to reach the goal position (randomly set each run) within a certain time step limit (typically 100). \\n\\n\\n2. How were the parameters selected (e.g. table 5 in the appendix). Why did you use the default the parameters?\\n\\nThe parameters were selected after hyperparameter searches. We did not exactly use any default parameters except for the architectural choice of using a GRU, which we found working quite well for all experiments. \\n\\n\\n3. How many runs are the curve averaged over and what\\u2019s the shaded region (e.g. one standard error)? Most of the results in the paper seem not statistically significant. \\n\\nMost of the experiments have 3 runs for each setting, and the shaded region represents one standard error. Our main goal was to compare our results against the standard setting (i.e. only with sparse rewards), and the comparisons for this pair are mostly statistically significant.\\n\\n\\n4. In figure 4, five domains are mentioned but only three of them are tested in the section 5.\\n\\nWe decided to move 2 experiments to the appendix, and keep the 3 domains that best represent the power of the embedding.\\n\\n\\n5. Section 6.2 seems irrelevant to the paper. What\\u2019s the purpose of this section?\\n\\nThis section aims to differentiate predictive coding from other unsupervised learning methods such as an autoencoder or VAE. Specifically, we wanted to show that predictive coding is more suitable for RL tasks than an autonecoder or VAE because the embedding can be trained to be background-agnostic. This is advantageous because in robotic tasks the background often will not stay stationary, and we could filter out less important information through predictive coding.\\n\\n\\n6. Figure 10 shows the result of using CPC feature directly vs. reward shaping. Are both feature using the same NN architecture, same PPO parameters, and same control setting? Moreover, the reward shaping method assumes we know the goal state but using CPC feature does not. Is it a fair comparison?\\n\\nBoth settings use the same architectural and algorithmic settings. For both methods, the embedding is trained from states that include information about the goal (i.e. the goal is a part of the state), so features from both methods contain the goal states.\"}", "{\"title\": \"Response\", \"comment\": \"Thank you very much for your insightful review. Below are our responses to some of your concerns:\\n\\n* The clustering bonus how do you prevent it from staying at the edge of the goal region and deriving infinite rewards from that ?\\n\\nThe cluster reward is one time only. With the cluster reward, the MDP is no longer time-homogeneous, so we convert it to a time-homogeneous MDP by treating entering the cluster as the first goal, and entering the environment goal as the second goal.\\n\\n\\n* Contrastive learning has been used to find goal embeddings before Warde-Farley et al. Unsupervised control through non-parametric discriminative rewards. In that paper they don't need the CPC future state predictors but instead contrast the goal and the final state of the trajectory. They use the resulting embedding to learn a reward function and ignore the extrinsic reward. Interestingly, they show that the rewards can be learnt online (maybe ideas from that paper can be applied here). \\n\\nAlthough our approach is currently off-line, we discussed near the end of our paper that converting the scheme to an online fashion is straightforward. Our goal in this paper is to focus on the properties of predictive embedding, and how they may contain information useful for RL; our contribution is largely on the finding that the predictive embedding flattens out the original state space and provide straightforward paths to any goal in the environment. We believe an online approach would work with our scheme, but that was not the focus of our paper.\\n\\n\\n* I am very surprised that CPC does such a good job given that the main object for learning in CPC is the distribution of trajectories which should be quite different in random exploration and the optimal policy. I presume this is because the environments are quite simple. This issue is of interest to the readers so an example of a simple failure case should make that point. \\n* Secondly, as nice as those embeddings in figure 5 look I wonder what happens in larger mazes with more structure i.e. somewhere where random walk will not be a uniform distribution and thus CPC will (most likely) not work as intended.\\n\\nWith our scheme, having good exploration of the state space is key to training good predictive embedding. This is more challenging for more difficult environments such as a large maze. We believe that the need for good exploration is not unique to our approach; our focus of the paper is to provide a method to leverage good exploration data without reward signals, and this method is to train predictive embedding.\\n\\nFor environments where random exploration is not enough (including the AntMaze environment), we overcame the issue by a) sampling from a robust distribution of initial states in the simulator b) training goal-conditioned policies for pure exploration strategies. These two approaches allowed us to collect good exploration data and train high-quality predictive embedding in more difficult environments \\n\\n\\n* Since these are not potential functions how do you prevent the rewards from biasing learning ? Ng et al. -- Policy invariance under reward transformations. This should be discussed in the paper because it is of practical importance.\\n\\nA rigorous investigation of the mathematical properties of the latent space is out of the scope of this applicative paper. From an empirical perspective, however, the ability for an agent to learn an optimal policy in various experimental settings with our scheme reflect that the negative distance scheme is likely a potential function. Ultimately, this depends on whether the latent space induced by the encoder preserves the metric property of the original state space, and one could train an invertible map to enforce this.\"}", "{\"title\": \"Response\", \"comment\": \"Thank you very much for your insightful review. Below are our responses to your questions:\", \"q1\": \"What constitutes success in the grid world domains for table 1?\\n\\n\\u201cSuccess\\u201d in the grid word domains means the agent is able to reach the goal position (randomly set each run) within a certain step limit (typically 100). We have added this detail to the paper.\", \"q2\": \"If you were to train a single policy using your reward shaping scheme (i.e. not the tiered approach currently employed) how would the new policy perform? Are there problems with the current scheme where you have to have the tiered policy?\\n\\nA single policy works well for the negative distance scheme.\\n\\nFor the clustering scheme, some simple changes to the state observations to account for whether the cluster reward has been received would allow the training of a single policy. An example is to just augment the state with a Boolean variable indicating whether the cluster reward has been received.\\n\\nSince the cluster reward is one time, the original MDP is no longer time-homogeneous, so a single policy naively learned with the added cluster reward will mistake the boundary of the cluster as a high reward zone.\", \"q3\": \"Why CPC and not say an autoencoder or VAE? A comparison over many types of unsupervised embedding learning algorithms could be interesting and make your method more general than currently presented. It could also strengthen your argument for using CPC.\\n\\nAs we motivated in discussion, we found that predictive coding is more suitable for RL tasks than an autonecoder or VAE because the embedding can be trained to be background-agnostic, as predictive coding only extracts information from states that are helpful for summarizing the past and predicting the future. The goal of predictive information is to extract the valuable (for predicting the future) information from the past, rather than to reconstruct the samples. Autoencoders, in general, will reconstruct the entire scene (both background and robotic arm), which makes their latent space representations redundant and not task-specific.\", \"q4\": \"How long were the trajectories for training CPC?\\n\\nWe have added this information to table 5 in the appendix. Each trajectory corresponds to one episode in the environment.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"*Synopsis*:\\n This paper proposes using the features learned through Contrastive Predictive Coding as a means for reward shaping. Specifically, they propose to cluster the embedding using the clusters to provide feedback to the agent by applying a positive reward when the agent enters the goal cluster. In more complex domains they add another negative distance term of the embedding of the current state and goal state. Finally, they provide empirical evidence of their algorithm working in toy domains (such as four rooms and U-maze) as well as a set of control environments including AntMaze and Pendulum.\", \"main_contributions\": \"- Using the embedding learned through Contrastive Predictive Coding for reward shaping.\\n - A reward shaping scheme that seems generally applicable to any embedding.\\n\\n\\n *Review*:\\n\\n I think the paper provides a compelling motivation, and is well written. I think using the embedding learned through CPC could provide meaningful semantics for representation learning and for reward shaping (as done in the current paper), and encourage the authors to continue down this line of inquiry. Unfortunately, I have several concerns over the method as currently implemented and the empirical comparisons (specifically with the chosen competitors) which I detail below. Given these concerns I am unwilling to recommend accepting this paper, unless several of these concerns are addressed.\\n\\n 1. This algorithm, by nature, is purely offline as the CPC and clustering all are currently done offline. Furthermore, the clustering portion of this approach requires states to be randomly sampled from the environment to create a nice set of clusters which are representative of the environment's full state space. These two requirements significantly limit this approach, especially when looking at domains where simulation is not possible, or the underlying state distribution is unknown. By and all, I don't think this means we should completely discount this method entirely and the authors do mention this as a detriment to their algorithm in section 6.3. I'm wondering if this paper should look at implementing an online version before publication, but think this is less prescient to the other concerns.\\n\\n 2. I am concerned about the current policy learning scheme (i.e. tiered policy (1) go to the correct cluster (2) go to the goal) which seems to only be used by your approach. This invalidates the comparisons made with the other reward shaping schemes as this goes beyond reward shaping (you are learning two separate policies).\\n\\n 3. The current competitors are unsatisfactory as they don't include other reward shaping techniques from the literature. Also the related works section seems to completely disregard this part of the literature. I would recommend comparing your approach to (methods you even cite!):\\n - \\\"Reward Shaping via Meta Learning\\\" by Zou et. al https://arxiv.org/pdf/1901.09330.pdf\\n - \\\"Potential based reward shaping\\\" by Gao et. al. https://www.aaai.org/ocs/index.php/IJCAI/IJCAI15/paper/viewPaper/10930\\n (I'm sure there are others beyond these)\\n\\n 4. I'm also concerned with the current significance of the results, particularly AntMaze, Half Cheetah, and Reacher. I'm especially concerned because I don't feel your competitors are not relevant/representative of the current state of reward shaping.\\n\\n 5. It would be worthwhile to try this method on several RL algorithms (i.e. Q-learning, TRPO, SAC, DDPG, etc...). This will help readers understand if this approach is a general method, or only applicable to PPO.\\n\\n 6. I quite like the idea of predictive coding (albeit the original scheme presented by Rao and Ballard 1999) as an unsupervised representation learning scheme, but am unsure this is critical for your method and the current approach is not really predictive coding in a sense (or at least the ideas I'm familiar with from cognitive computational neuroscience). I am concerned with the predictive coding idea being highlighted here as a key ingredient, but none of the papers containing the originating idea of predictive coding are mentioned. Rao and Ballard is one, but there is a rich literature following from this work into the free energy formulation (Friston) and active learning. These ideas should appear in your introduction, as you heavily rely on them. Also there are other predictive coding schemes for unsupervised representation learning (such as PredNets from David Cox https://coxlab.github.io/prednet/), which I believe should be mentioned. In light of these other methods, I'm not sure your discussion on only using predictive features for reward shaping is accurate, and instead these claims should be softened for only features learned through CPC.\", \"missing_experimental_settings\": \"- Number of runs tested\\n - What are the error bars in your plots?\\n\\n I would also like to recommend \\\"Deep Reinforcement Learning that Matters\\\" by Henderson for a reference on how to conduct meaningful Deep RL experiments using policy gradient methods (https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/viewPaper/16669). I think this paper could benefit from the recommendations made there.\\n\\n\\n More questions/clarifications:\", \"q1\": \"What constitutes success in the grid world domains for table 1?\", \"q2\": \"If you were to train a single policy using your reward shaping scheme (i.e. not the tiered approach currently employed) how would the new policy perform? Are there problems with the current scheme where you have to have the tiered policy?\", \"q3\": \"Why CPC and not say an autoencoder or VAE? A comparison over many types of unsupervised embedding learning algorithms could be interesting and make your method more general than currently presented. It could also strengthen your argument for using CPC.\", \"q4\": \"How long were the trajectories for training CPC?\\n\\n *Other minor comments not taken into account in the review*\\n - I think your labels are backwards in table 1.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper presents a method to derive shaping rewards from a representation learnt with CPC. They propose learning a CPC representation from some random data and fix it. They assume exploration is not too difficult so that rewards are achievable without additional mechanisms. Once a reward is achieved they can compute the embedding of the corresponding state as a goal under the CPC learnt representation either directly ie. distance or via a clustering step. The paper is clearly written and presents a nice idea. The paper is correct and part of the approach seems novel. I find the analysis of the results well executed though I think that they should be improved for publication.\", \"major_points\": [\"I think this is a valid contribution and it seems like something the RL audience might be interested in.\", \"I am very surprised that CPC does such a good job given that the main object for learning in CPC is the distribution of trajectories which should be quite different in random exploration and the optimal policy. I presume this is because the environments are quite simple. This issue is of interest to the readers so an example of a simple failure case should make that point.\", \"Secondly, as nice as those embeddings in figure 5 look I wonder what happens in larger mazes with more structure i.e. somewhere where random walk will not be a uniform distribution and thus CPC will (most likely) not work as intended.\", \"The clustering bonus how do you prevent it from staying at the edge of the goal region and deriving infinite rewards from that ?\", \"Since these are not potential functions how do you prevent the rewards from biasing learning ? Ng et al. -- Policy invariance under reward transformations. This should be discussed in the paper because it is of practical importance.\", \"Contrastive learning has been used to find goal embeddings before Warde-Farley et al. Unsupervised control through non-parametric discriminative rewards. In that paper they don't need the CPC future state predictors but instead contrast the goal and the final state of the trajectory. They use the resulting embedding to learn a reward function and ignore the extrinsic reward. Interestingly, they show that the rewards can be learnt online (maybe ideas from that paper can be applied here).\"], \"minor_points\": [\"please make legends on plots readable size.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper proposes a reward shaping method which aim to tackle sparse reward tasks. The paper first trains a representation using contrastive predictive coding and then uses the learned representation to provide feedback to the control agent. The main difference from the previous work (i.e. CPC) is that the paper uses the learned representation for reward shaping, not for learning on top of these representation. This is an interesting research topic.\\n\\nOverall, I am leaning to reject this paper because (1) the main contribution of the paper is not clear (2) the experiments are missing some details and does not seem to support the claim that the proposed methods can tackle the sparse reward problem. \\n\\nFirst of all, it would\\u2019ve been better to have a conclusion section, so the readers can see the contributions of the paper. After reading the paper, I still do not understand what are the contributions of the paper and what're from the previous works. The paper does not provide well justification why CPC feature can provide useful information for reward shaping. The paper does not provide a new method to learn predictive coding. It does not provide a novel reward shaping method (the \\u201cOptimizing on Negative Distance\\u201d method is very similar to [1]). So, I am not sure what\\u2019re the contributions of this paper. \\n\\nMoreover, I am not convinced that the proposed method can tackle long horizon and sparse reward problems. As the paper discuss in introduction, learning in sparse reward environment is hard because it relies on the agent to enter the goal during exploration. However, the proposed approach seems only able to work in environments where exploration with random policy can generate trajectories that contain sufficient environment dynamics (e.g. dynamics near the goal states). How can the method learn that information without entering the goal? \\n\\nFurthermore, it seems that the proposed approach only works for goal-oriented tasks (since we need to know the goal state for reward-shaping). I think this should be clearly stated in the paper.\", \"there_are_some_missing_details_which_makes_it_difficult_to_draw_conclusions\": \"1. How is the \\u2018success rate\\u2019 computed (e.g. in figure 7 and table 1).\\n2. How were the parameters selected (e.g. table 5 in the appendix). Why did you use the default the parameters?\\n3. How many runs are the curve averaged over and what\\u2019s the shaded region (e.g. one standard error)? Most of the results in the paper seem not statistically significant. \\n4. In figure 4, five domains are mentioned but only three of them are tested in the section 5.\\n5. Section 6.2 seems irrelevant to the paper. What\\u2019s the purpose of this section?\\n6. Figure 10 shows the result of using CPC feature directly vs. reward shaping. Are both feature using the same NN architecture, same PPO parameters, and same control setting? Moreover, the reward shaping method assumes we know the goal state but using CPC feature does not. Is it a fair comparison?\", \"the_paper_has_some_imprecise_parts\": \"1. The definition of MDPs (in section 2) is imprecise. For example, how is the expectation defined? how is the initial state sampled? What does $p\\\\in\\\\mathcal{P}$ (last line in the first paragraph) mean where $\\\\mathcal{P}$ is the state transition function?\", \"minor_comments_which_do_not_impact_the_score\": \"1. Figure 1 should come before figure 2. \\n2. It would have been better if there is a short description of how the hand-shaped rewards is designed for each domain in the main text. \\n\\n[1] The Laplacian in RL: Learning Representations with Efficient Approximations\"}" ] }

[ "David van Niekerk" ]

[ "meta-learning", "learning to learn", "few-shot classification", "memory-based optimization" ]

{ "note_id": [ "a0sO1Ldrx1", "H1eDLjf4cr", "rkgxcuLcYr", "ByeW_VS5Kr" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751846, 1572248398603, 1571608711693, 1571603561394 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2549/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2549/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2549/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes a LSTM-based meta-learning approach that learns how to update each neuron in another model for best few-shot learning performance.\\n\\nThe reviewers agreed that this is a worthwhile problem and the approach has merits, but that it is hard to judge the significance of the work, given limited or unclear novelty compared to the work of Ravi & Larochelle (2017) and a lack of fair baseline comparisons.\\n\\nI recommend rejecting the paper for now, but encourage the authors to take the reviewers' feedback into account and submit to another venue.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a meta-learner that learns how to make parameter updates for a model on a new few-shot learning task. The proposed meta-learner is an LSTM that proposes at each time-step, a point-wise multiplier for the gradient of the hidden units and for the hidden units of the learner neural network, which are then used to compute a gradient update for the hidden-layer weights of the learner network. By not directly producing a learning rate for the gradient, the meta-learner\\u2019s parameters are only proportional to the square of the number of hidden units in the network rather than the square of the number of weights of the network. Experiments are performed on few-shot learning benchmarks. The first experiment is on Mini-ImageNet. The authors build upon the method of Sun et al, where they pre-train the network on the meta-training data and then do meta-training where the convolutional network weights are frozen and only the fully-connected layer is updated on few-shot learning tasks using their meta-learner LSTM. The other experiment is on Omniglot 20-way classification, where they consider a network with only full-connected layers and show that their meta-learner LSTM performs better than MAML.\\n\\nThe closest previous work to this paper is by Ravi & Larochelle, who also propose a meta-learner LSTM. The submission states about this work that \\u201cA challenge of this approach is that if you want to optimize tens of thousands of parameters, you would have a massive hidden state and input size, and will therefore require an enormous number of parameters for the meta-learner\\u2026In Andrychowicz at al. an alternative approach is introduced that avoids the aforementioned scaling problem by individually optimizing each parameter using a separate LSTM\\u2026\\u201d I don\\u2019t believe this is true.\\n\\nAs stated in the work by Ravi & Larochelle, \\u201cBecause we want our meta-learner to produce updates for deep neural networks, which consist of tens of thousands of parameters, to prevent an explosion of meta-learner parameters we need to employ some sort of parameter sharing. Thus as in Andrychowicz et al. (2016), we share parameters across the coordinates of the learner gradient. This means each coordinate has its own hidden and cell state values but the LSTM parameters are the same across all coordinates.\\u201d Thus, Ravi & Larochelle employ something similar to Andrychowicz at al., meaning that the number of parameters in the LSTM meta-learner there is actually a constant relative to the size of the learner network. \\n\\nThus, the paper\\u2019s contribution relative to Ravi & Larochelle is to propose a LSTM meta-learner with more parameters relative to the learner model. Firstly, I think this comparison should be stated and explained clearly in the paper. Additionally, in order to prove the benefit of such an approach, I think a comparison to the work of Ravi & Larochelle with the exact experimental settings used in the submission (pre-training the convolutional network and only using the meta-learner LSTM to tune the last fully-connected layer) would be helpful in order to validate the usefulness of the extra set of parameters in their proposed meta-learner LSTM.\\n\\nThe submission also states that \\u201cIn many cases it is preferred to have an inner loop that consists of multiple sequential updates. However the inner loop\\u2019s computational graph can become quite large if too many steps are taken. This often results in exploding and vanishing gradients since the outer loop still needs to differentiate through the entire inner loop (Aravind Rajeswaran (2019), Antoniou et al. (2018)). This limits MAML to domains where a small amount of update steps are sufficient for learning. The LSTM-based meta-learner proposed in this work, allow gradients to effectively flow through a large number of update steps. NORML can therefore be applied to a wide array of domains.\\u201d I think this statement should be validated if it is stated. Can it be shown that when making a lot of inner-loop updates, the LSTM meta-learner performs better than MAML because of overcoming the stated issues with differentiation through a long inner loop? The experiments done in the paper involve very few inner loop steps and so I don\\u2019t believe the claim is supported.\\n\\nLastly, the experimental results are not very convincing. Though the authors say they modify the method proposed in Sun et al, the results from Sun et al. are not shown in the paper. Sun et al actually seem to achieve better results than the submission with the same 1-shot and better 5-shot accuracy. Was there a reason these results are not shown in the submission? Additionally, for the Omniglot experiment, is there any reason why it was not performed with the typical convolutional network architecture? Since the original MAML results on 20-way Omniglot are with a convolutional network, using the convolutional network would make the results more meaningful relative to previous work and show that the method is more broadly applicable to convolutional networks.\\n\\nI believe there are several issues with the paper as stated above. Because of these issues, it is hard to evaluate whether the idea proposed is of significant benefit.\\n\\nReferences\\nAndrychowicz et al. Learning to learn by gradient descent by gradient descent. NIPS 2016.\\nRavi & Larochelle. Optimization as a Model for Few-Shot Learning. ICLR 2017.\\nSun et al. Meta-transfer learning for few-shot learning.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": [\"The paper targets the scalability issue of certain meta-learning frameworks, and is therefore addressing an important and interesting topic.\", \"Theoretical and technical novelty is rather minimal.\", \"The paper writing is well beyond the ICLR level, and is honestly beyond the level required by a scientific manuscript in general. Writing needs a massive overhaul.\", \"There are way too many grammatical mistakes. In addition, there are citations written the wrong way -e.g. Rajeswaran (2019)-, and also the flow of the ideas has room for improvement.\", \"For the aforementioned reference, the author ordering is wrong, and not consistent with the way it is cited in the text.\", \"page 1: \\\"while the learner can learn a set of initial task parameters are easily optimized for a new task. \\\": Please fix and/or clarify.\", \"page 2: \\\"where the classes contained in each meta-set is disjointed\\\".\", \"page 3: \\\"The LSTM-based meta-learner proposed in this work, allow gradients to\\\"\", \"\\\"Memory-based Under review as a conference paper at ICLR 2020 methods (Ravi & Larochelle (2017)) that use all of the learner\\u2019s parameters as input to a meta-learner tend to break down when using a learner with a large number of parameters (Andrychowicz et al., 2016).\\\": Just to clarify: The latter paper, which is criticising the former category, is older than the paper representing the former category. Is that right?\", \"page 2: \\\"superior parameter updates\\\". I see that the result has been based on classification accuracy. Notwithstanding the ablative study in the experiments secion, maybe superior parameter updates can be either replaced by a more unequivocal description of the mentioned comparison, or formally defined from there onwards.\", \"At the ICLR level, I do not think that all this detailed description of backpropagation would be necessary.\", \"Equation 1 and the description that follows: \\\"a_l is the layer\\u2019s pre-activation output\\\". Is a_l the post-activation output as well? Equation 1 implies so, doesn't it?\", \"page 3: \\\"This limits MAML to domains where a small amount of update steps are sufficient for learning.\\\": What do you mean by \\\"This\\\"? Is it to have inner loops consisting of multiple sequential updates, which is what was referred to as \\\"preferable\\\" in the beginning of the same paragraph? i.e. Is the MAML limitation noted in this paragraph a limitation of the vanilla MAML (plus the other versions) as well or solely due to the \\\"preferred, yet detrimental\\\" extension of adopting multiple sequential updates?\"], \"minor\": [\"page 1: Supervised few-shot learning \\\"aims to challenge machine learning models to ...\\\": It does not challenge ML models; it is a specific ML paradigm.\"]}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"This submission proposes NORML, a meta-learning method that 1) learns initial parameters for a base model that leads to good few-shot learning performance and 2) where a recurrent neural network (LSTM) is used to control the learning updates on a small support set for a given task. The method is derived specifically for full connected neural networks, where the meta-learner produces gating factors on the normal gradients (one for each neuron that the parameter is connecting). The method is compared with various published few-shot learning methods on miniImageNet, and an ablation study and detailed comparison with MAML is presented on Omniglot.\\n\\nUnfortunately, I believe this submission should be rejected, for the following reasons:\\n\\n1. Limited novelty: it is a fairly incremental variation compared to the Meta-Learner LSTM (Ravi & Larochelle). I actually kind of like the proposed variant, but then I would at least expect a direct comparison with Meta-Learner LSTM. And though the authors try to make a case for why NORML is better than Meta-Learner LSTM, unfortunately they don't actually provide a fair comparison. Indeed, the results for Meta-Learner LSTM on miniImageNet use a very different base learner (i.e. not a ResNet-12) which wasn't pretrained. Same in fact can be said about many of the results reported in Table 1.\\n\\n2. Missing baselines: though the comparison in Table 2 with MAML is a good step, I actually believe the 3 versions of MAML considered aren't appropriate. Instead, I believe that 1) at a minimum, a comparison with a version of MAML where a separate learning rate is learned \\\"per hidden unit\\\" AND \\\"per inner loop step\\\" is necessary, as it is closer to what NORML can achieve, and 2) a comparison with MAML++ (Antoniou et al.) would be ideal. Finally, I think this study should also be done on miniImageNet, not only Omniglot.\\n\\n3. Limited applicability: NORML assumes that the base learner is a fully connected network. That strongly limits the applicability of the method, given that most few-shot learners usually have convolutional layers.\\n\\n4. Inaccurate descriptions of some prior work: The related work includes certain statements that I believe aren't accurate. For example, it is stated that the Meta-Learner LSTM requires \\\"an enormous number of parameters\\\". I believe this is not true: there is a single, small LSTM that is used for all parameters of the base learner (however, the authors are right that running this LSTM requires a very large cell state and input size, since the batch size of the Meta-Learner LSTM is essentially the number of parameters of the base learner). Similarly, it is stated that Andrychowicz et al. proposes to \\\"individually [optimize] each parameter using a separate LSTM\\\". I believe this is also not true, and that a single LSTM is used for all parameters. In fact, ironically, NORML on the contrary appears to be using different LSTMs for each layer of the base learner (at least based on Equations 14-19, where the LSTM parameters are indexed by layer ID l, thus implying different LSTMs), unlike what is claimed in the Relate Work section (\\\"this work uses a single meta-learner\\\").\\n\\nFor me to consider increasing my rating, I would expect all points above to be addressed. \\n\\nFinally, here are a few minor issues I've found:\\n- In Equation 16, \\\"b_{i,\\\\tilde{c}}\\\" should be \\\"b_{l,\\\\tilde{c}}\\\" (i.e. i should be l)\\n- The submission doesn't mention whether gradients are propagated into the base learner gradients (i.e. whether a first-order version of the method is used)\\n- 4th paragraph of the Relate Work section has a missing reference (\\\"?\\\")\\n- Table 2 is missing confidence intervals\"}" ] }

[ "Benjamin Akera", "Joyce Nakatumba-Nabende", "Ali Hussein", "Daniel Ssendiwala", "Jonathan Mukiibi" ]

[ "keyword spotter", "radio data", "crop pest and disease", "agriculture" ]

{ "note_id": [ "Br72Ua5za1", "Skeu3xh0KH", "SyeDlJt6FB", "BklSdgXQtH" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751813, 1571893424472, 1571815150679, 1571135596934 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2548/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2548/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2548/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"Main summary: Design an effective and economical model which spots keywords about pests and disease from community radio data in Luganda and English.\", \"discussions\": \"all reviewers vote on rejecting the paper, due to lack of generalizability, training and evaluation discussion need work\", \"recommendation\": \"Reject\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper describes an approach to analyze radio data with ML-based speech keyword techniques. The authors identify keywords related to agriculture and build a model that can automatically detect these keywords of interest. Their contribution is the proposed model relies on relatively simple neural networks (15-layer 1-D CNNs) to achieve keyword detection of low resourced language (e.g., Luganda). Therefore, they are capable of monitoring food security concerns in rural areas.\\n\\nThis paper should be rejected because (1) no novel algorithm is proposed and only straightforward solution is used (2) the paper never clearly demonstrates the existence of problem they are trying to solve, (3) the experiments are difficult to understand and missing many details, and (4) the paper is incomplete and unpolished.\\n\\nMain argument\\nThe paper does not do a great job of demonstrating that current model development and drawbacks of dealing with low resourced language (e.g., Luganda). They said they solve the problem by the Keywords Spotting system (KWS), but they did not give a comprehensive review of the current progress of KWS. For example, they should include and compare with the paper, called \\u201cSpeech recognition and keyword spotting for low-resource languages: Babel Project Research at CUED\\u201d by Dr. Mark Gales.\\n\\nWhen it comes to methodology, the authors use unnecessary pages to describe how they collect Luganda audios and keyword dataset. It would be appropriate to illustrate their data flow or detection pipeline instead of irrelevant engineering details. In addition, SIAMESE CNN is used to compare performance, but the authors use a relatively large portion of the paper to explain its architecture. Did their method developed beyond SIAMESE CNN? Otherwise, it would be inappropriate to describe it in methodology and model design. For the 1D convolution model, the model is demonstrated in an extremely simple way. They did not even mention their target data (Y), please clarify it in detail.\\nThe experiment did not provide convincing evidence of the utility of the proposed method (1-D convolution). They only demonstrated results in 1 dataset, which did not indicate the effectiveness of the model. There are so many missing details. How many times the experiment they did? How much size is an audio file? Why 1d-conv show the same values of accuracy, precision, and recall? What\\u2019s the model drawn in Figure 1? The paper lacks complete discussion on their empirical results. We did not know how to interpret their results.\\n\\nConsidering the above, there are many drawbacks to the paper. The paper seems incomplete. Therefore, we highly recommend rejecting the paper.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper tries to design an effective and economical model which spots keywords about pests and disease from community radio data in Luganda and English. The author collected keywords from both radio and online newspapers, set up a Siamese convolutional neural network, which includes input 1 and input 2, Resnet 50, Flatten, five-layer dense and L1 distance, trained the model on a total of 18426 samples and validated on 4607 and tested on 5759 samples via a K80 GPU on Google Colab.\", \"the_paper_should_be_rejected_because\": \"(1) The generalizability to different use cases needs to be shown. The approach is too applied.\\n(2) The fit to ICLR is not perfect. I would expect a stronger focus on representation learning.\\n(3)\\tModel\\nThe paper lacks necessary explanation and clarification for the model, which is the core of the paper. For example, the part \\u201cDense\\u201d in Siamese convolutional model is not well explained. The number of layers is up to 5, if there is no dropout, the model may have overfitting and gradient vanishing problems. \\n(4)\\tResult\\nAs can be seen from both model accuracy and model loss curves in Figure 1, the part of 21st -30th epochs shows an increasing accuracy of training data while that of validation data does not increase anymore and a declining loss of training data while that of validation does not decrease. This is a tendency of overfitting, which means the model should adopt 20 epochs rather than 30 epochs. \\n(5)\\tIncompleteness \\nCompared with the detailed depiction of the first three parts, part 4, part 5 and part 6 are insufficient. It is necessary to complement these three parts with interpretation, explanation and discussion.\", \"things_to_improve_the_paper_that_did_not_impact_the_score\": \"(1)\\tLanguage problems such as in page 3 paragraph 2: \\u201cTo compliment the keywords captured from radio talkshows, we also scrapped an online local newspaper.\\u201d Here \\u201ccompliment\\u201d should be \\u201ccomplement\\u201d, and \\u201cscrap\\u201d is semantically incorrect. \\n(2)\\tPage 4 part 4.1: \\u201c\\u2026according to a hierarchical feature extraction given by equation 3.\\u201d It should be equation 2. \\n(3)\\tFormat in references: fonts are not uniformed.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"Overview\\n\\nThis paper presents a very interesting application of speech keyword spotting techniques; the aim is to listen to continuous streams of community radio in Uganda in order to spot keywords of interest related to agriculture to monitor food security concerns in rural areas. The lack of internet infrastructure results in farmers in rural areas using community radio to share concerns related to agriculture. Therefore accurate keyword spotting techniques can potentially help researchers flag up areas of interest. The main engineering challenge that the study deals with is the fact that there isn\\u2019t a lot of training data available for languages like Lugandu.\\n\\nDetailed Comments\\n\\n Section 3\\n\\n1. The word \\u201cscrapped\\u201d should be replaced with \\u201cscraped\\u201d.\\n2. Its not entirely clearly what is meant by \\u201cas well as an alternative keyword in form of a stem\\u201d. What is meant by \\u201cstem\\u201d in this context? Maybe a citation or an example would make it more clear. \\n3. The corpus of keywords supposedly contains 193 distinct keywords however the models in Section 4 are only trained to discriminate between 10 randomly sampled keywords. I don\\u2019t understand why this is. Training on the full corpus would allow the network to see more training data and consequently might result in more accurate models. \\n\\nSection 4\\n\\n1. The authors refer to Equation 3 in Section 4.1, but I think the reference is to Equation 2.\\n2. The 1-D CNN has 14 layers but it is not clear to me what the size of the final network is. It would be useful to provide more details of the architecture and a comment about the network size either in terms of number of parameters, number of computations or size of the network weights on disk. \\n3. The figure for the Siamese network should be numbered. \\n4. The authors say that the inputs to the ResNets are of shape 32x100. What do these dimensions refer to? Are their 32 frequency bins? Are their 100 frames as inputs? What are the parameters of the FFT computation? Do 100 frames correspond to a second of audio? \\n5. How many parameters are their in the Siamese network? \\n6. Its not clear to me how the Siamese networks are used during inference. The authors say they use the L1 distance to determine if a given test example is similar to a given keyword. How are the examples for the keyword of interest selected? Are all training examples used and the scores average? Are the training examples averaged first to form a centroid vector for the keyword? \\n\\nComments on the methodology\\n\\nOne of the main challenges in this work is the fact that there isn\\u2019t a large number of training examples available. However the authors still train relatively large acoustic models with 14 layers for 1-D CNN and a ResNet for the Siamese architecture. There are several studies for keyword spotting on mobile devices that aim to train tiny networks that consume very little power [1] [2]. I think it would be more appropriate to start with architectures similar to these due to the small size of the training dataset. Additionally, it would also be very useful to try and identify languages that are phonetically similar to Lugandu and that have training datasets available for speech recognition. Acoustic models trained on this data can then be adapted to the keyword spotting task using the training set collected for this task. Note that the languages need to have only a small amount of phonetic overlap in order for these acoustic models to be useful starting points for training keyword spotting models. \\n\\nComments on the evalation\\n\\nThe keyword spotting models presented here are trained with the intention of applying them to streaming radio data. However, the models are trained and tested on fixed chunks of audio. A big problem with this experimental design is that the models can overfit to confounding audio cues in the training data. For example, if all training data are in the form of 1 second audio chunks and all chunks have at least some silence at the start, then the models learn that a keyword is always preceded by a certain duration of silence. This is not the case when keywords occur in the middle of sentences in streaming audio data. The fact that the evaluation is also performed on individual chunks of audio fails to evaluate how the trained models would behave when presented streaming audio. I am certain that when applied to streaming audio these models would false trigger very often, however this fact isn\\u2019t reflected by measuring precision and recall on fixed chunks of test audio. \\n\\nA more principled evaluation strategy would be to present results in the form of detection-error tradeoff (DET) curves [1], [2]. Here the chunks with examples of the keyword in question can be used to measure the number of false rejects by the system. And long streams of audio data that do not contain the keyword should be used to measure the false alarms. Given that its relatively easy to collect streams of radio data, the false alarms can be measured in terms of the number of false alarms per unit time (minutes, hours or days). This evaluation strategy roughly simulates the streaming conditions in which this model is intended to be deployed. Furthermore, DET curves present the tradeoff between false alarms and false rejects as a function of the operating threshold, which provides much more insight into the performance/accuracy of the trained models. \\n\\nSummary\\n\\nI think the area of application of this work is extremely interesting, however the training and evaluation methodologies have to be updated in order to realistically measure the way this system might perform in real-world test conditions. \\n\\nReferences\\n\\n[1] Small-footprint Keyword Spotting Using Deep Neural Networks\\n[2] Efficient Voice Trigger Detection for Low Resource Hardware\"}" ] }

[ "Arber Zela", "Julien Siems", "Frank Hutter" ]

[ "Neural Architecture Search", "Deep Learning", "Computer Vision" ]

{ "note_id": [ "wGKhoi28OiB", "fhaP5EbL15", "BJgfr4RFoS", "BkeNfNCFjB", "HyxD2ZRFjH", "BJgaF-RFsr", "SyxkV1RtsH", "Byxt3MX9qH", "BkeRiU01qB", "SkljhB5ndB" ], "note_type": [ "comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1703607948683, 1576798751784, 1573671994162, 1573671947575, 1573671342949, 1573671300919, 1573670694599, 1572643504862, 1571968678452, 1570706866639 ], "note_signatures": [ [ "~Gladis_Ne_Limes1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2547/Authors" ], [ "ICLR.cc/2020/Conference/Paper2547/Authors" ], [ "ICLR.cc/2020/Conference/Paper2547/Authors" ], [ "ICLR.cc/2020/Conference/Paper2547/Authors" ], [ "ICLR.cc/2020/Conference/Paper2547/Authors" ], [ "ICLR.cc/2020/Conference/Paper2547/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2547/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2547/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"re\", \"comment\": \"Protecting sensitive data is a top priority. Establishing stringent security protocols, including secure communication channels and data encryption, is crucial to mitigate data security risks.\", \"https\": \"//mlsdev.com/services/it-staff-augmentation\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"The authors present a new benchmark for architecture search. Reviews were somewhat mixed, but also with mixed confidence scores. I recommend acceptance as poster - and encourage the authors to also cite https://openreview.net/forum?id=HJxyZkBKDr\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #1 (2/2)\", \"comment\": \"-- References --\\n[1] Chris Ying, Aaron Klein, Esteban Real, Eric Christiansen, Kevin Murphy, Frank Hutter, NAS-Bench-101: Towards Reproducible Neural Architecture Search, ICML 2019\\n[2] Hieu Pham, Melody Y. Guan, Barret Zoph, Quoc V. Le, Jeff Dean, Efficient Neural Architecture Search via Parameter Sharing, ICML 2018\\n[3] Hanxiao Liu, Karen Simonyan, Yiming Yang, DARTS: Differentiable Architecture Search, ICLR 2019\\n[4] LIAM LI, AMEET TALWALKAR. Random Search and Reproducibility for Neural Architecture Search, ArXiv 2019\\n[5] Han Cai, Ligeng Zhu, Song Han, ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware, ICLR 2019\\n[6] Sirui Xie, Hehui Zheng, Chunxiao Liu, Liang Lin, SNAS: Stochastic Neural Architecture Search, ICLR 2019\\n[7] Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, Ameet Talwalkar, Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization, JMLR 2018\\n[8] Stefan Falkner, Aaron Klein, Frank Hutter, BOHB: Robust and Efficient Hyperparameter Optimization at Scale, ICML 2018\"}", "{\"title\": \"Response to Reviewer #1 (1/2)\", \"comment\": \"We would like to thank the reviewer for his/her feedback. We address the concerns raised by the reviewer in the following paragraphs:\\n\\n> The main contribution of this submission is the benchmark NAS-Bench-1Shot1, which can benefit future research for one-shot NAS. However, this benchmark is based on NAS-Bench-101 [1], and the difference is that NAS-Bench-101 only contains discrete architecture, while the presented NAS-Bench-1Shot1 has a component to discretize architectures and then match to NAS-Bench-101. The novelty may not be enough. \\n\\nWe agree with the reviewer that NAS-Bench-101 is not new, but we argue that our paper is nevertheless novel, because:\\n- Finding a way to reuse NAS-Bench-101 is one of the strengths of our approach, since the immense computational resources required to generate NAS-Bench-101 (120 TPU years!) can be efficiently reused this way. We believe this is indeed *more* novel than to simply create another benchmark, and it is far more environmentally friendly since it comes with zero additional computational burden. As we demonstrate in our paper our benchmark allows for a quick and thorough analysis of different NAS methods, which would otherwise be too time consuming. \\n\\n- The mapping between different search space representations is novel to the best of our knowledge. Most previous work in NAS represents the architecture search space as a directed acyclic graph, however they differ by where the operation choices are encoded in the graph -- either nodes (e.g. ENAS [2]) or edges (e.g. DARTS [3]). By introducing a framework that allows us to create search spaces that automatically satisfy all the constraints in NAS-Bench-101, one can clearly investigate the contribution that each specific algorithm has on the achieved performance, without adding a confounding factor due to a different search space.\\n\\n- A correlation analysis between the stand-alone architectures evaluated with the one-shot weights and retrained from scratch has not been done before for all the architectures in such a large search space (more than 360k architectures).\\n\\n- As the reviewer already pointed out, the tunability of one-shot NAS optimizers has not been addressed before.\\n\\n> The authors claim that they introduce a general framework for one-shot NAS methods. Personally, I think this is over-claimed. There are several variants of DARTS, and the authors just implement these variants and DARTS itself in a unified code base. It would be a true general framework if it can also work with other one-shot NAS methods such as ENAS and the rest ones.\\n\\nWe acknowledge that at the time of submission we had focused on implementing various NAS methods based on DARTS, due to the interest which it has created in the community. However, while the code base of Random NAS with weight sharing [4] is based on DARTS, its search method differs and already demonstrates that similar sampling-based search methods can also fit into our framework. (Indeed, our framework allows to give further clues on why the results of RandomWS were subpar, by demonstrating the weak correlation between one-shot model and final architecture performance.) We have now integrated ENAS (Algorithm 5 in the paper) in our framework and updated the figures in the paper accordingly. Thanks for this suggestion! We are currently working on integrating other methods, such as ProxylessNAS [5] or SNAS [6].\\n\\nNote that our claim that NAS-Bench-1Shot1 is a general framework for one-shot NAS methods (and not only; also for NAS methods using black-box optimizers) is also due to our novel unification between different graph representations in existing NAS papers.\\n\\nOne other advantage of using the NAS-Bench-101 benchmark is the evaluation of all the architectures using different budgets. This allows evaluating the sensitivity of NAS algorithms during search with respect to the number of epochs used to train individual architectures for estimating the validation error returned to the NAS optimizer (e.g., one can expect a different behavior of ENAS if the sampled architectures from the controller would be evaluated using the lowest or the highest training budget in NAS-Bench-101). This also allows running multi-fidelity black-box optimizers, such as Hyperband [7] or BOHB [8], in our framework.\\n\\n> One interesting thing is that the authors try to use HPO method such as BOHB to tune the hyperparameter of NAS methods.\\n\\nThanks! We also updated the results with more HPO runs. \\n\\nWe hope that with this response we addressed all your concerns and that you will consider updating your assessment, particularly seeing that we have now demonstrated that our framework also encompasses ENAS. Thank you for your time and effort.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We greatly appreciate the reviewer's effort and the very positive review. We share the reviewer\\u2019s hope that NAS-Bench-1Shot1 will serve the community to prototype and benchmark NAS algorithms without confounding factors.\"}", "{\"title\": \"Response to Reviewer #4\", \"comment\": \"We thank the reviewer for reading our paper thoroughly and for the very positive feedback.\\n\\n> the background section might be improved\\nBased on the reviewer\\u2019s suggestion we updated Section 2 in the paper and added some more details for each respective NAS method we used throughout the paper in Appendix B. We hope that now the concepts will be easier to understand, also for a reader not familiar with the different NAS paradigms, and without going through the literature in detail.\"}", "{\"title\": \"General comment to the AC and to reviewers\", \"comment\": [\"We thank the reviewers for reading our paper thoroughly and for their feedback. We updated the paper as follows:\", \"More independent runs for each method in all the experiments in the paper.\", \"Added ENAS [1] (Algorithm 5 in the paper) to the codebase and benchmarked it in all our search spaces. The results for this optimizer are shown in Figure 2 and Figure 7 in the paper.\", \"Added more detailed information to the background section.\", \"Fixed a small bug in the codebase and updated the plots with the new results. The paper\\u2019s analysis and conclusions of course remain unchanged. In order to make the results reproducible, we provide all the seeds and settings with the code release.\"], \"we_would_like_to_note_that_there_was_a_parallel_concurrent_work_submitted_to_iclr_2020_named\": \"\\u201cAn Algorithm-Agnostic NAS Benchmark\\u201d. We believe this is also a very valuable benchmark, and it also received very positive scores (8,8,6); the reason we point this out here is that we believe the reviewers\\u2019 arguments for giving high scores also apply to our paper. In particular, they emphasize the importance of good benchmarks like these for one-shot NAS algorithms in the reproducibility crisis. The main differences between the two benchmarks are as follows:\\n\\n- They use extensive computation to create a *new* benchmark (with 15.625 architectures), while we devise a novel reformulation to reuse the even much more extensive computation of the NAS-Bench-101 dataset (~120 TPU years) to create three new one-shot search spaces with 6240, 29.160, and 363.648 architectures each. \\n- We believe that our reuse of the extreme computation of NAS-Bench-101 is environmentally friendly, since it provides similar value as a new benchmark at zero computational cost.\\n- We show that (contrary to their claim), it *is* possible to reuse the graph representation in NAS-Bench-101 to run one-shot NAS methods; this requires changes to the one-shot search space, but allows a mapping which can be used for architecture evaluation. \\n- They evaluate their search space on 3 image classification datasets, while we introduce 3 different search spaces (as sub-spaces of NAS-Bench-101) with growing complexity.\\n- Our paper does not only serve as a benchmark, but also provides additional insights into the NAS optimization process, such as:\\n 1. Investigating the correlation between the stand-alone models evaluated with the one-shot weights vs. retrained from scratch (to the best of our knowledge, the first time this is done for a large search space (more than 360k architectures).\\n 2. Analyzing the anytime performance of one-shot NAS algorithms\\n 3. Tuning the hyperparameters these optimizers such that they perform comparable with black-box NAS optimizers with the same number of function evaluations.\\n\\nWe thank the reviewers and the AC for their time and effort for ICLR.\\n\\n-- References --\\n[1] Hieu Pham, Melody Y. Guan, Barret Zoph, Quoc V. Le, Jeff Dean, Efficient Neural Architecture Search via Parameter Sharing, ICML 2018\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper introduces a benchmarking framework to evaluate a class of one-shot NAS methods by exploiting the NAS-Bench-101 database. In order to do so, the authors apply techniques from Bender et al. (2018) to match the search spaces in NAS-Bench to one-shot scenarios and weight edges in the search graph allowing for a variable number of edges per cell. These weights are then used to query a larger, discrete architecture from NAS-Bench for the corresponding evaluation errors. The authors motivate the soundness of a unique framework by showing minimal differences between three NAS optimizers.\\nIn their analysis, they also corroborate the findings of Sciuto et al. (2019) on a larger scale, showing little generalization of the ranking from one-shot validation to the final architecture. Finally, the authors also use this framework to show that a fine-tuned baseline compares favorably to SOTA methods, and that large regularization factors lead to more robust configurations.\\n\\nThe paper is well-written and introduces a framework to cheaply compare one-shot NAS optimizers based on expensive computations behind the NAS-Bench-101 benchmark. In addition, the authors provide empirical analyses that reflect generalization and regularization issues with current methods and that could lead towards designing more robust algorithms. \\n\\nIn my opinion, the the background section might be improved by presenting higher level notions that would lead to an easier understanding by a wider audience. Also, although comprehensible due to the nature of this study, readers not familiar with a variety of previous work might find difficult to parse the paper as many concepts are just referenced.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a benchmark dataset for evaluating One-Short Neural Architecture Search models. It extends on the idea of the NASBench-101 (Ying et. al., 2019) into One-Shot models. \\n\\nI think this is important work. With all the development of neural archictecture search methods, reproducible research is paramount. This is even more timely considering the fact that these problems are computationally expensive, so it is even more computationally difficult than usual to run exhaustive comparisons among proposed methods. \\n\\nThe paper is well-written and the design decisions are clearly explained. The comparison of NAS methods is also interesting to read. In general, I think this is a worthy publication.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this submission, the authors present a benchmark NAS-Bench-1Shot1 for one-shot Network Architecture Search. The presented benchmark reuses the existing NAS-Bench-101 that only contains discrete architectures, and the authors give a method to discretize architectures to match NAS-Bench-101. Also, the authors claim that they introduce a general framework for one-shot NAS methods.\\n\\nOverall speaking, the technique contribution and novelty of this submission requires further improvement, and I suggest to rejecting this submission. The reasons are detailed as follows:\\n\\n1) The main contribution of this submission is the benchmark NAS-Bench-1Shot1, which can benefit future research for one-shot NAS. However, this benchmark is based on NAS-Bench-101, and the difference is that NAS-Bench-101 only contains discrete architecture, while the presented NAS-Bench-1Shot1 has a component to discretize architectures and then match to NAS-Bench-101. The novelty may not be enough.\\n\\n2) The authors claim that they introduce a general framework for one-shot NAS methods. Personally, I think this is over-claimed. There are several variants of DARTS, and the authors just implement these variants and DARTS itself in a unified code base. It would be a true general framework if it can also work with other one-shot NAS methods such as ENAS and the rest ones.\\n\\nGiven these, this submission requires further improvement, especially in terms of technique contribution and novelty. \\n\\nOne interesting thing is that the authors try to use HPO method such as BOHB to tune the hyperparameter of NAS methods.\"}" ] }

[ "Wenjing Huang", "Shikui Tu", "Lei Xu" ]

[ "Adversarial Defense", "Auto-encoder", "Adversarial Attack", "GAN" ]

{ "note_id": [ "msEYgWjTQ", "BygTjO82iB", "Byxhy4L3ir", "BJgXYMU3ir", "SJleI4SniB", "BJgDflrhjr", "rkxeACN3jH", "rygeKoE2iS", "rJxMJaDpcr", "Bkxa-E0hcr", "H1lFTg4w9B", "B1lfFS_qFr", "r1eK-R1aOH", "HklJqt5huS", "H1lYkuFiDB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_review", "comment", "official_comment", "comment" ], "note_created": [ 1576798751755, 1573836964997, 1573835747541, 1573835386566, 1573831751524, 1573830670873, 1573830343903, 1573829496130, 1572859097703, 1572819973440, 1572450497207, 1571616121569, 1570729472569, 1570707847007, 1569589217272 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "ICLR.cc/2020/Conference/Paper2546/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2546/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2546/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2546/AnonReviewer2" ], [ "~Anthony_Wittmer1" ], [ "ICLR.cc/2020/Conference/Paper2546/Authors" ], [ "~Anthony_Wittmer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes a defense for adversarial attacks based on autoencoders that tries to find the closest point to the natural image in the output span of the decoder and \\\"purify\\\" the adversarial example. There were concerns about the work being too incremental over DefenseGAN and about empirical evaluation of the defense. It is crucial to test the defense methods against best available attacks to establish the effectiveness. Authors should also discuss and consider evaluating their method against the attack proposed in https://arxiv.org/pdf/1712.09196.pdf that claims to greatly reduce the defense accuracy of DefenseGAN.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to evaluation questions and obfuscated gradients\", \"comment\": \"Thanks for your reply and your suggestions.\\n\\nAccording to your suggestion, we take PGD adversarial training (Madry et al. 2018) into comparisons under the same experimental settings. Here we list the comparisons on MNIST.\\n*****************************************************\\n None PGD adv.tr Our\", \"white_box______fgsm\": \"0.144 0.949 0.984\", \"pgd\": \"0.007 0.920 0.982\", \"cw\": \"0.008 0.773 0.979\", \"bpda\": \"0.000 0.924 0.810\", \"black_box\": \"A/B 0.701 0.971 0.935\\n******************************************************\\nAs shown in the above table, PGD adv.tr achives better performance than our method under BPDA[1], which is consistent with the conclusions in [1] and [2]. Hope this will help to answer your confusions.\\n\\nMeanwhile, we have added additional results on the Large-scale CelebFaces Attributes (CelebA) dataset in the appendix of the paper. Since it takes more time to prepare the results on CIFAR-10, we will include the results on CIFAR-10 later. The results on CelebA show that AE-GAN+rs is still effective on more complex datasets with large RGB images. For further details, please refer to Appendix G in the revised version. \\n\\n[1] Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML 2018\\n[2] NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. ICML 2019\"}", "{\"title\": \"Response to Review #4 (part 2/2)\", \"comment\": \"Q2. The methods also rely on the capacity of the auto-encoder to learn good representations/reconstructions. It will be useful to show the performance on more complicated datasets was learning a good AE model is more challenging.\\nA1. Thanks for the suggestions. We agree with the reviewer that our method relies on the capacity of the auto-encoder to learn good representations/reconstructions and it is useful to evaluate the method on more complex datasets. Following the suggestion, we have added results on the Large-scale CelebFaces Attributes (CelebA) dataset in the appendix of the paper. The results show that AE-GAN+rs is still effective on complex datasets with large RGB images. For further details, please refer to Appendix G in the revised version. Here we list some results on FGSM ($\\\\epsilon=0.3$) white-box attacks, gray-box attacks as well as black-box attacks.\\n********************************************************************\\n No Defense PGD adv.tr. DefenseGAN Our\\nWhite-box 0.0416 0.6119 0.8879 0.9364\\nGary-box 0.0416 --- 0.8559 0.8897\\nBlack-box 0.0404 0.6121 0.8517 0.8839 \\n********************************************************************\\n\\nQ3. The autoencoder is also equipped with an adversarial loss such that the decoder produces realistic images. Is there any assurance that the latent distribution F(x) from the data and the prior p(z) will be similar. It might be useful to explicitly ensure this. The discriminator currently never sees images that re reconstructions of the true data samples. So the decoder behavior could be different for different regions of the input space.\\nA3. Thanks for the advice. There is no assurance that the latent distribution F(x) from the data and the prior p(z) will be similar. In AE-GAN, the prior p(z) is a multidimensional normal distribution, while F(x) is derived by the encoder and bounded in $[-1,1]^n$. Based on the advice, we have made the following efforts\\uff1a(1) Adding a KL-loss between F(x) and p(z) to push F(x) close to p(z). However, the results are not good. One possible reason is that the auto-encoder is trained to self-organize the internal codes of the input patterns, which might not be consistent with the KL loss pushing the codes as a single Gaussian distribution. It may deserve further investigations on the balance between representative power and distribution consistency in the training stage. (2) Construct p(z) to get close to F(x). We use another GAN (a GAN built in the latent space) to train a generator p(z), which is trained to get close to F(x). However, it is hard to train the two GANs simultaneously and we are not able to make p(z) similar to F(x). Meanwhile, the performance of (2) is not so good as AE-GAN.\\nSo far, we have not found an effective way to ensure F(x) and p(z) similar. It may be a good topic in future research.\\n\\n Q4. It is not clear why the proposed method performs better on stronger white-box attacks compared to much weaker black-box attacks.\\nA4. The main reason for strong defense on white-box attacks lies in the searching process which effectively blocks the gradients. According to the definition of obfuscated gradients in [2], the obfuscated gradients has happened in the proposed AE-GAN+rs, because AE-GAN+rs includes an optimization process when reconstructing the input image, similar to Defense-GAN [1]. Athalye, A., Carlini, N., & Wagner, D. (2018) have proposed a new attack method BPDA [2] to tackle obfuscated gradients, and we also carry experiments on BPDA attacks. We implement the BPDA attack with $\\\\epsilon=0.3$ on the test set of MNIST (same as the setting in Table 1 of the manuscript). The classification accuracy of the baseline classifier on the MNIST test set is 99% without any attacks. The accuracy of our model against BPDA attacks is 81%, lower than 98.2% by PGD, which shows the BPDA attack is more effective than PGD when attacking our method. Under BPDA attacks, our method performs worse on white-box than black-box (93.5%).\", \"references\": \"[1]Samangouei, P., Kabkab, M., & Chellappa, R. (2018). Defense-gan: Protecting classifiers against adversarial attacks using generative models. arXiv preprint arXiv:1805.06605.\\n[2] Athalye, A., Carlini, N., & Wagner, D. (2018). Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. arXiv preprint arXiv:1802.00420.\\n\\nWe hope that this response addresses your concerns, and we would like to answer any other questions you may have.\"}", "{\"title\": \"Reponse to Review #4 (part 1/2)\", \"comment\": \"We thank you for the comments and suggestions, and answer your questions as follows:\\n\\nQ1. The methods seem to heavily rely on the auto-encoder's ability to detect adversarial samples. The detection performance can be still susceptible to white-box attacks.\\nA1. To defend an adversarial example, our method can always activate the searching process by AE-GAN+rs, i.e., use the code determined by the searching process to reconstruct the example, and then feed the reconstruction into the classifier for the final classification result. To be more efficient, we also design a detection mechanism (in Sec. 4.4) to avoid the unnecessary searching process for the clean examples or the adversarial examples that are affected by small perturbations. If the detection mechanism is used, the adversarial example may cheat the detection step and attack the classifier by the reconstruction from AE-GAN+r. Moreover, as you mentioned in the above comments, the detection mechanism can be susceptible to white-box attacks, which design adversarial examples by attacking the stacked model, i.e., concatenating the classifier with AE-GAN+r as a new classifier. In the following, we conduct experiments to demonstrate that the proposed method is still effective against the above mentioned white-box adversarial attacks.\\n*****************************************************************************\\nUsing MSE as the indicator with threshold=0.01 \\n*****************************************************************************\\n AE-GAN+r AE-GAN+rs AE-GAN+rs(with detection) Detection \\nFGSM 3.0: 0.691 0.8526 0.8565 100% \\nFGSM 1.0: 0.8943 0.9721 0.9721 99.62%\", \"cw\": \"0.1182 0.7346 0.7275 99.82%\\n*****************************************************************************\\nIn the above table, the first column denotes the classification accuracy when using AE-GAN+r to purify the perturbed images under white-box attacks on the stacked model (i.e., concatenating the classifier with AE-GAN+r). That is, no searching process is activated for any input.\\nThe second column denotes the classification accuracy when using AE-GAN+rs to purify the adversarial examples, i.e., the searching process is activated for every input.\\nThe third column denotes the classification accuracy when incorporating the detection mechanism to determine whether the searching process is activated for the input.\\nMoreover, among the examples that successfully attack the stacked model (i.e., concatenating the classifier with AE-GAN+r), we calculate the percentage of detected ones, as given in the fourth column.\\nBy comparing the first column and the third column, we can see that the detection mechanism is still effective in detecting the adversarial examples generated by the white-box attacks on the stacked model (i.e., concatenating the classifier with AE-GAN+r). Take the first row for example, when the AE-GAN+r can not diminish the adversarial perturbations for the 1-0.691=30.9% examples, these 30.9% examples can be 100% detected by the detection mechanism (as shown in the last column). Then, they will trigger the searching process and be purified by AE-GAN+rs for higher accuracy. Thus, the final performance of the proposed method (AE-GAN+rs incorporated with detection) on the generated adversarial examples achieves 85.65%, close to AE-GAN+rs (without detection). Therefore, the above white-box attacks didn't affect the performance of the proposed method.\\nBy comparing the first column and the second column, we can see that the searching process can achieve a higher accuracy on the generated adversarial examples, which demonstrates the effectiveness of the searching process. When the detection mechanism is incorporated, there is no decline in performance. In other words, when the detection helps to save time, it didn\\u2019t affect the performance, even under the strongest white-box attacks. We have conducted more experiments to investigate the performance of the detection stage in the appendix in our revised version. For more details, please refer to Appendix F in the revised paper.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"Thanks for reading our paper thoroughly and the helpful comments. We would like to address your concerns one by one.\\n\\nQ1. The comparison should include the state of the art adversarial training defenses, PGD Adversarial Training (Madry et al; you might want to cite the ICLR 18 paper) and TRADES (Zhang et al., Interpreting adversarially trained convolutional NN, ICML 2019)\\nA1. Thanks for providing the information about the two papers. According to your suggestion, we take PGD adversarial training [1] into comparisons under the same experimental settings. Here we list the comparisons with PGD adversarial training. Due to the time limit, we have not finished the comparisons with TRADES. We will include the comparisons with TRADES later.\\n*****************************************************\\nOn MNIST\\n*****************************************************\\n None PGD adv.tr Our\", \"white_box______fgsm\": \"0.144 0.949 0.984\", \"pgd\": \"0.028 0.717 0.816\", \"cw\": \"0.062 0.224 0.767\", \"black_box\": \"A/B 0.227 0.784\\t 0.603\\n******************************************************\\nThe results in the above tables indicate that our method can consistently provide effective defense on various attacks, while PGD adversarial training can not generalize to CW attacks. However, the performances of defenses on the F-MNIST dataset shows noticeably lower than that on MNIST, this is due to the large $\\\\epsilon =0.3$ in the FGSM attack. The qualitative examples in Appendix H can show that $\\\\epsilon=0.3$ represents large perturbations.\\n\\nQ2. Fig. 4 is unclear: the percentage of adversarial examples detected by the detector, but quid of the false alarms. \\nA2. According to your suggestion, we revised Figure 4 by adding Receiver Operating Characteristic (ROC) curves as well as the Area Under the Curve (AUC) metric. Also, we conduct more experiments on the detection performance and visualize how False Negative Rate (FNR, the percentage of missed diagnosis) and False Positive Rate (FPR, the percentage of false alarms) changes with the threshold increases. The added visualizations on MSNIT are shown in Figure 3 in the revised version, and the results on FashionMNIST are included in Appendix F. Here we give some examples. For more details, please refer to Figure 3 and Appendix F in our revised version. \\n*******************************************************\\nUsing MSE as the indicator \\n*******************************************************\\n FPR FNR \\n $\\\\epsilon=0.1$ $\\\\epsilon=0.3$ \\nThreshold=0.01 73.37% 4.34% 0%\\nThreshold=0.015 44.01% 30.18% 0% \\nThreshold=0.02 23.25% 59.81% 0%\\n*******************************************************\\nUsing L1 distance as the indicator , $\\\\epsilon=0.1$\\n*******************************************************\\n FPR FNR \\nThreshold=0.02 20.15% 0% \\nThreshold=0.025 9.65% 0% \\nThreshold=0.03 4.35% 0% \\n********************************************************\", \"white_box_____fgsm\": \"0.073 0.739 0.804\", \"notation\": \"$\\\\epsilon$ is the strength of attacks.\\n\\nQ3. There are quite some typos (nosie; iterarion; tabel; unsafety) and missing words. The paper has been hastily written; and it includes one page more than allowed.\\nA3. Thanks for carefully checking our presentation and pointing out our typos. We have carefully revised our paper and improved the quality and readability. To meet the page limit, we tried to revise the paper accordingly but the paper is still a little more than the recommended paper length, but we do not exceed the strict upper limit of the length (10 pages).\", \"references\": \"[1] Madry, A., Makelov, A., Schmidt, L., Tsipras, D., & Vladu, A. (2017). Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083.\"}", "{\"title\": \"Reponse to Review #2 (part 2/2)\", \"comment\": \"Q5. Given that the BiGAN also incorporates an encoder (and they were also benchmarked against in the experiments), where do the advantages of the AE-GAN+rs come from? I think making this point in the text would also be helpful as well.\\nA5. Thanks for the valuable advice. Some related analysis was given in Sec 4.5 (Ablation Study), and here we carry out more experiments to improve the analysis. As shown in Table 3 (page 9) in the original version, although BiGAN also incorporates an encoder, it shows poor reconstruction performance which degrades its defense results.\\n************************************************************* \\n MNIST FashionMNIST\\n BiGAN AE-GAN+r BiGAN AE-GAN+r\", \"clean\": \"0.6854 0.9708 0.5538 0.8267\\nFGSM ($\\\\epsilon$=0.3): 0.4593 0.8568 0.3366 0.6148\\n*************************************************************\\nThe above table is provided with further experimental results in addition to Table 3 of the original version. The accuracies of the target classifier when using BiGAN or AE-GAN+r defense strategy are reported. It could be observed that BiGAN suffers from the problem of poor reconstruction on both clean and perturbed images, and AE-GAN+r (without searching) can increase the accuracy by 30%.\\nThe discriminator in BiGAN discriminates jointly the data and latent code ( (X, E(x)) versus (G(z), z) ), while the discriminator in AE-GAN is adopted to enhance the generative ability of the decoder such that it can generate realistic images. Thus, the encoder-decoder system is still good at reconstruction in our framework, which is important for the encoder to provide a good initialization for the searching process. BiGAN itself performs badly in purifying a perturbed image, because it may reconstruct an image inconsistent with the input. For a fair comparison, we equip BiGAN with the searching process. As the encoder in BiGAN cannot provide a good start point for the searching process, BiGAN suffers the same problem as Defense-GAN does. In practice, we set the steps=60 for BiGAN and steps=15 for ours, and from Table 1, we can see that AE-GAN+rs still performs much better than BiGAN.\"}", "{\"title\": \"Response to Review #2 (part 1/2)\", \"comment\": \"Thanks for giving us valuable advice. We would like to address your points one by one as follows.\\n\\nQ1. There are a lot of phrases that should be reworded in the text for better clarity. The writing was pretty hard to follow and the paper would really benefit overall after it has been improved.\\nA1. We admit that there are confusing phrases in the original version. We will address your concern by revising the paper thoroughly.\\n\\nQ2. It was hard for me to tell from the text how exactly the models were implemented and trained. Making this point more clear, in either the main text or the Appendix, would be really helpful.\\nA2. Thanks for the suggestion. We would like to provide pseudo-codes for the training and testing stage in Appendix C in our revised version. The testing stage was implemented differently from the training stage. In Sec. 3.3 of the paper, the method for training, denoted as AE-GAN+r, does not include the searching process. When defending against adversarial samples in the inference process, the trained AE-GAN+r becomes AE-GAN+rs (see Sec 3.2 and Sec 3.4) by including the searching process which is denoted as {+s} (see Sec 3.4). More details are given below:\\n(1) For the training stage, AE-GAN+r has no searching process and is trained with legitimate samples. As a modified auto-encoder, AE-GAN+r adds a discriminator to AE and is equipped with an adversarial loss such that the decoder produces realistic images. AE-GAN+r is trained to optimize the min-max loss in Eq.(4) in the manuscript. The training strategy is the same as GANs [1], where the parameters in Encoder-Decoder and the parameters in discriminator are updated alternatively to minimize \\u201cGenerative Loss + Reconstruction Loss\\u201d and \\u201cD_Loss\\u201d respectively. \\n(2) In the inference stage, the AE-GAN+r has been trained and we want to use it to defend against adversarial samples. To get better reconstruction, the searching process is activated, i.e., we use gradient descent to find the best encoding in the latent space for a corrupted input. This defense strategy is denoted as AE-GAN+rs. In practice, we take 15 steps for GD iterations in the searching process, and the learning rate is set to 0.01 (multiplied by 0.5 every 5 steps).\\n\\nQ3. The performance of AE-GAN+rs as evaluated on MNIST and FashionMNIST did not provide convincing evidence that it outperformed baselines such as Defense-GAN or adversarial training. The results would also have been more compelling had the authors evaluated their method on more complex datasets such as CIFAR-10.\\nA3. Thanks for your suggestion. Based on the suggestion, we have added additional results on the Large-scale CelebFaces Attributes (CelebA) dataset in the appendix of the paper. Since it takes more time to prepare the results on CIFAR-10, we will include the results on CIFAR-10 later. The results on CelebA show that AE-GAN+rs is still effective on more complex datasets with large RGB images. For further details, please refer to Appendix G in the revised version. Here we list some results on FGSM ($\\\\epsilon=0.3$) white-box attacks, gray-box attacks as well as black-box attacks.\\n************************************************************************\\n No Defense Adv.tr. Defense-GAN Our\\nWhite-box 0.0416 0.6119 0.8879 0.9364\\nGary-box 0.0416 --- 0.8559 0.8897\\nBlack-box 0.0404 0.6121 0.8517 0.8839 \\n************************************************************************\\n\\nQ4. It wasn\\u2019t clear to me how the encoder-assisted search process (Section 3.4) works in practice. When you use gradient descent to find the best encoding for a corrupted input, how many steps do you need to take? Does this step happen while the auto-encoder and GAN are being trained, or are those pre-trained and you just take additional gradient steps in the latent space?\\nA4. Please see the above A2 for the details.\", \"references\": \"[1]Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information processing systems (pp. 2672-2680).\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We thank reviewer #1 very much for their positive and encouraging comments. You are right that the proposed method can be regarded as an alternative to the Defence-GAN method of (Samangouei et. al. 2018) which employed a computational expensive search due to multiple starting points and many iterations. Our method is much faster with comparable performance, and it significantly outperforms Defense-GAN for a single starting point during search and fewer iterations. We have also carefully revised our paper and improved the quality of the paper.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"Authors propose to use a modified autoencoder at the input of a network to ward off adversarial samples by purifying them. The encoder-decoder is trained with a discriminator which can identify whether the input is malicious or normal. A malicious input image is purified with the help of a search procedure in the latent space of the auto-encoder. The gradient based search obtains a latent code that corresponds to the reconstructed image that is closest to the input. The algorithm uses output of the decoder as a starting point in the search. Experiments show a reasonable performance against adversarial attacks on MNIST and F-MNIST images. The proposal seems to be an alternative to the Defence-GAN method of (Samangouei et. al. 2018) where search is computationally expensive - due to multiple starting points and many iterations. The ease of optimization in the proposed method comes at the cost of accuracy (Table 1). However, for a single starting point during search and fewer iterations, it outperforms Defense-GAN by a significant margin (Table 2).\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper aims to refine DefenseGAN (ICLR 18), where an autoencoder is used to initialize the search for projecting an adversarial examples to the manifold of real examples.\\n\\nThe main contribution is to reduce the computational cost of DefenseGAN, the claim is \\\"by an order of magnitude\\\". \\n\\nThough the idea is good, I found the contribution to be too incremental for the paper to be accepted:\\n* The comparison should include the state of the art adversarial training defenses, PGD Adversarial Training (Madry et al; you might want to cite the ICLR 18 paper) and TRADES (Zhang et al., Interpreting adversarially trained convolutional NN, ICML 2019);\\n* I would consider the Szegedy baseline as obsolete;\\n* Fig. 4 is unclear: the percentage of adversarial examples detected by the detector, but quid of the false alarms; \\n\\nThere are quite some typos (nosie; iterarion; tabel; unsafety) and missing words. The paper has been hastily written; and it includes one page more than allowed.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper combines an auto-encoder (AE) based approach to correct the adversarially perturbed samples with a GAN based approach for the same task. More specifically, the encoder of the AE is used to provide a better initialization for the latent space optimization employed in the Defense-GAN approach.\\n\\nThe authors propose a two-stage inference process. First, the autoencoder is used to detect if an input sample is from the natural image distribution or adversarial distribution. In the latter case, Defense-GAN is employed that uses gradient descent in the latent space, with encoder output as the initialization, to find a natural or non-adversarial counterpart of the input sample. Results are reported for MNIST and F-MNIST that show that the proposed method is computationally cheaper than Defence-GAN.\\n\\nQuestions / Concerns:\\n\\n - The methods seem to heavily rely on the autoencoder's ability to detect adversarial samples. The detection performance can be still susceptible to white-box attacks. The methods also relies on the capacity of the auto-encoder to learn good representations/reconstructions. It will be useful to show the performance on more complicated datasets was learning a good AE model is more challenging. \\n - The autoencoder is also equipped with an adversarial loss such that the decoder produces realistic images. Is there any assurance that the latent distribution F(x) from the data and the prior p(z) will be similar. It might be useful to explicitly ensure this. The discriminator currently never sees images that re reconstructions of the true data samples. So the decoder behavior could be different for different regions of the input space. \\n- It is not clear why the proposed method performs better on stringer white-box attacks compared to much weaker black-box attacks.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary: This paper proposes AE-GAN+sr, an auto-encoder based GAN for equipping neural networks with better defenses against adversarial attacks. The authors evaluate their method on black-box attacks, white-box attacks, and gray-box attacks on MNIST and Fashion-MNIST, and show decent empirical results when compared to baselines.\", \"decision\": [\"Reject. The writing was hard to follow and the experimental evaluations could have been stronger.\", \"Supporting Arguments/Feedback:\", \"There are a lot of phrases which should be reworded in the text (e.g. \\u201cit requires attention to\\u201d in paragraph 1, \\u201cdetect if sample is from a normal distribution\\u201d at the end of Section 1, etc.) for better clarity. The writing was pretty hard to follow and the paper would really benefit overall after it has been improved.\", \"It was hard for me to tell from the text how exactly the models were implemented and trained (see Question #1). Making this point more clear, in either the main text or the Appendix, would be really helpful.\", \"The performance of AE-GAN+sr as evaluated on MNIST and FashionMNIST did not provide convincing evidence that it outperformed baselines such as Defense-GAN or adversarial training, though I did appreciate the authors\\u2019 additional analysis on the tradeoffs between computational cost and performance for the Defense-GAN. The results would also have been more compelling had the authors evaluated their method on more complex datasets such as CIFAR-10.\"], \"questions\": [\"It wasn\\u2019t clear to me how the encoder-assisted search process (Section 3.4) works in practice. When you use gradient descent to find the best encoding for a corrupted input, how many steps do you need to take? Does this step happen while the autoencoder and GAN are being trained, or are those pre-trained and you just take additional gradient steps in the latent space?\", \"Given that the BiGAN also incorporates an encoder (and they were also benchmarked against in the experiments), where do the advantages of the AE-GAN+sr come from? I think making this point in the text would also be helpful as well.\"]}", "{\"comment\": \"Thanks for the reply and additional results. However, I still have some confusion on the results of Adv.Tr., which reveals that the methods based on adversarial training are weaker than the methods based on obfuscated gradients. But actually it is not true[1][2].\\n\\nWhy the authors use FGSM adversarial training as the baseline rather than PGD adversarial training (Madry et al. 2018), which is a necessary baseline. I would have expected the authors provide a comparison to Madry et al. (2018) which provides the strongest white-box robustness to date. \\n\\nIn addition, have the authors tried to apply the proposed methods on a much larger dataset, such as CIFAR-10? \\nAs we know, MNIST is too easy and over used. More importantly, [3] suggests that MNIST is peculiar in that there exists a simple \\u201cclosed-form\\u201d denoising procedure, which leads to similarly robust models without adversarial training. For an average MNIST image, over 80% of the pixels are in {0, 1}, so the adversarial perturbation are easily clipped.\\n\\n[1] Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML 2018\\n[2] NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. ICML 2019\\n[3] Ensemble Adversarial Training: Attacks and Defenses. ICLR 2018\", \"title\": \"Thanks for the reply\"}", "{\"comment\": \"Hi Anthony, thanks for your comments.\\n\\nAccording to the definition of obfuscated gradients in [2], the obfuscated gradients has happened in the proposed AE-GAN{+sr}, because the proposed method also includes an optimization process when removing the adversarial perturbations in the adversarial examples, similar to Defense-GAN [1]. To investigate the performance of our method against the BPDA attack [2], we use the official implementation of BPDA (https://github.com/anishathalye/obfuscated-gradients) to generate adversarial examples for evaluation. Since our method was implemented in PyTorch, so we rewrite the source code of BPDA in PyTorch to use it for our method. Our implementation is available at https://github.com/Annonymous-repos/attacks-in-pytorch . With all the settings in the source code kept unchanged, we implement the BPDA attack with $\\\\epsilon=0.3$ on the test set of MNIST (same as the setting in Table 1 of the manuscript). The classification accuracy of the baseline classifier on the MNIST test set is 99.6% without any attacks. The accuracy of our model against BPDA attacks is 81%, lower than 98.2% by PGD, which shows the BPDA attack is more effective than PGD.\\n\\nOn page 12 (Appendix B) of [2], another attack method in \\u201cEvaluation A\\u201d was presented to attack Defense-GAN [1]. This attack was implemented as \\u201cl2_attack\\u201d in the source code (https://github.com/anishathalye/obfuscated-gradients). We find that similar to the results reported for Defense-GAN in [2], the adversarial samples generated by l2_attack in our model can mislead the classifier with nearly 100% success. We admit that the adversarial examples exist directly on the projection of the generator/decoder of our model (Generator can directly generate adversarial samples). However, all adversarial examples should be purified by AE-GAN{+sr} to remove the adversarial perturbations before being fed into the classifier. Since the attacking process takes a long time, we compute l2_attack on the first 1000 samples of test set on MNIST, and our method defends the attack with accuracy 92.4%.\\n\\nTherefore, we admit that there still exist \\u201cbad points\\u201d in the latent space, which could be used to generate adversarial samples through the Generator/Decoder, but the encoder in our method can encode these samples into \\u201cgood points\\u201d on the manifold of natural images. As a result, the outputs of encoder are no longer the original bad points. When we start the search process to find the closet reconstruction, the bad points can hardly be found in few steps with the guidance of MSE reconstruction error, where the number of epochs for the search process is set to 15 in our model.\\n\\nAccording to the above experiments, our method seems robust against the two attack methods performed for Defense-GAN in [2].\\n\\n[1]Defense-gan: Protecting classifiers against adversarial attacks using generative models.\\u00a0ICLR2018\\n[2]Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples.\\u00a0ICML2018\", \"title\": \"Response to evaluation questions and obfuscated gradients\"}", "{\"comment\": \"Hi, I have comments about the experimental results:\\n\\n1. In table 1, for the proposed model, it seems that the black-box attacks are better than the white-box attacks, which reveals that the defense is obfuscating gradients[1]. The results of Defense-GAN have the same behavior, and Defense-GAN has been broken by [1]. However, for Adv. Tr., the white-box attacks are better than the black-box attacks.\\n\\n2. In order to check whether obfuscated gradients has happened, BPDA[1] or Nattack[2] is a better choice to evaluate the models.\\n\\n[1] Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML 2018\\n[2] NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. ICML 2019\", \"title\": \"Evaluation questions and obfuscated gradients?\"}" ] }

[ "Amina Mollaysa", "Brooks Paige", "Alexandros Kalousis" ]

[ "molecule generation", "disentangling" ]

{ "note_id": [ "a7t7RRFwj-", "BJeynP12or", "S1eljSynoS", "BJewJEJ2iH", "H1lScMJniH", "B1l9h1k3sB", "HJeJEM-0tS", "Byo_bZRtB", "HJlbqymTYH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751726, 1573808039244, 1573807511606, 1573807070913, 1573806732716, 1573806001655, 1571848742770, 1571848562514, 1571790728967 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2544/Authors" ], [ "ICLR.cc/2020/Conference/Paper2544/Authors" ], [ "ICLR.cc/2020/Conference/Paper2544/Authors" ], [ "ICLR.cc/2020/Conference/Paper2544/Authors" ], [ "ICLR.cc/2020/Conference/Paper2544/Authors" ], [ "ICLR.cc/2020/Conference/Paper2544/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2544/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2544/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper aims to generate molecules with desired properties using a variant of supervised variational auto-encoders. Disentanglement is encouraged among the style factors. The reviewers point out that the idea is nice, but authors avoid quantitative comparison with SotA Graph-based generative models. Especially, the JT-VAE is acknowledged as a strong baseline widely in the community and is a VAE-based model, it is important to do these comparisons.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer#3 (continuation)\", \"comment\": \"# Figure 6 and overfitting\\nThe step-like nature of this figure is actually primarily due to the fact that not every \\u201cstructure\\u201d is compatible with different target values of the property. It is true that for a z that is encoded from certain x, when we combine it with different ys, for many values of y, it tends to reconstruct back the original molecule that is being encoded ( first block in figure 6). However, this is not the same as overfitting. First, note that the x that is used to encode is taken from the held-out test set. Second, the model also generated many other molecules that are different from the original x ( the other blocks in figure 6 ). The only thing in the graph is that when 100 different ys are paired with a fixed z, the model only generates 19 unique molecules (among around 90 valid molecules). It shows that for a certain combination of z and x, physically there is no molecules that that matches such combination or our model did not learn yet to generate molecules for such combination. This is due to the fact that molecule space is discrete, and some combinations of \\u201cstyle\\u201d with particular target properties may not be possible. It also shows that the effect of z is stronger than the effect of y in the generation phase: we can see that for a given z, with a very small change of y, one often end up generating the same molecules. For example for z that is encoded from a molecule whose property is around -0.9 (after scaling to -1 and 1), if we change the property towards 0, we either get invalid molecules or the same molecule; only once we change the property to around 0.1 do we get different molecules.\\n\\n# Correlation as a measure when the model overfit\\nBefore proceeding to the conditional generation, we also checked the novelty of the generated molecules when compared against the training set. As table 1 shows the novelty is pretty high (100% for ZINC and 98 % for QM9 ). Therefore, when we sample a property value from the testset and generate molecules from that particular properties, most of the generated molecules are novel. \\n\\n#The literature references seem to omit all molecular generation work since 2018 \\nWe did, in fact, include quite a few papers from 2018 in the literature review. Due to the space limit, we took the most relevant ones, but we can add a more recent literature review as part of an additional discussion on using our conditional generation alongside other decoder architectures.\\n\\n#Stacked LSTM works better\\nIt is also to our surprise that a simple conditional LSTM on SMILES works so well for conditional generation. Perhaps even more surprising is that such a strong and simple model has not seen more widespread use as a baseline in work on conditional generation (\\u201cLearning Deep Generative Models of Graphs\\u201d, Li et al 2018, is the only example we could find). However, as it does not produce a latent representation of the molecule the way a VAE model does, direct style transfer is not possible. The setting we focus on has the potential of both doing direct conditional generation and style transfer, and it also produces continuous latent representations of the molecules which could be useful for other downstream tasks. We certainly agree that investigating ways for directly enabling style transfer on such conditional LSTMs is an interesting direction for future work.\"}", "{\"title\": \"Response to Reviewer#3\", \"comment\": \"We would like to thank you for the detailed comments. We addressed each question as follows:\\n\\n# The dependency of the latent states on the property predictor (\\u201cThe first change consists of introducing a property prediction part directly into ELBO\\u2026\\u201d)\\nThe ELBO in equation 6 is lower bound for p(x|y), while the extended ELBO introduced in equation 10 is lower bound for p(x, f(x) |y). This extension models not only generating molecules x for a given y, but also the corresponding property f(x) of the generated molecule \\u2014 this does have the effect of explicitly encouraging the generated property f(x) to match the value of y passed to the generative model. Regarding the structure of the posterior itself, it is true in eq (10) we write in terms of a posterior purely over z, as q(z | x), with y as an additional observed latent variable in the generative model. However, one could also view the \\u201cjoint\\u201d posterior q(z,y | x) as defined via q(z|x) and a delta function on the \\u201ctrue\\u201d property function f evaluated on the molecule x. Adding y as part of the posterior over z, i.e. with a functional form of q(z | x, y), does not really provide much additional context since y is assumed to be a deterministic property of x and the encoder network is quite powerful. In practice we saw no discernible difference between models trained with encoders q(z | x) and q(z | x, y) and thus went with the simpler option; of course, it is possible if more or different properties y are chosen this may change.\\n\\n# The prior and averaged posterior over the training encodings\\nWe would like to try to clear up some possible confusion here. We believe there is some misunderstanding between section 3.4 (describing the evaluation of the gradient) with the experiment section (describing how we sample z when doing generation at test time). Throughout training, as stated in section 3.4, instead of sampling z from regions of the prior that may correspond to invalid molecules, we instead sample from the aggregate posterior; this is motivated by the style transfer definition, and handled in practice by sampling z from the approximate posterior q(z|x_j) for x_j not paired with the target y_i given to the decoder. During testing, for generation, we found that sampling z from the approximated aggregated posterior where it is estimated using simple Gaussian with a variance that matches the average posterior variance has better performance in terms of validity. The restriction to the simple Gaussian is made not for computational reasons (it is simple to instead refer to the original dataset to sample from the aggregate posterior) but rather for a fairer comparison with sampling from the prior.\\n\\n# Limited to LogP\\n Our method by design is not restricted to specific properties; adapting to other properties require no further adjustment. We wanted to make a \\u201creal\\u201d conditional generation model for molecules such that for a certain target value we can sample molecules that has properties match. Many existing models that can do \\u201cconditional\\u201d generation are actually targeted to optimize certain properties, which means, the model can only generate molecules with the target property raised. Those models make sense for, say, QED as a higher QED score is clearly better. However, for generating molecules with certain values, LogP is a more sensible choice of property, as particular values may be desired, rather than simply \\u201cas large / small as possible\\u201d. The model itself is not conceptually restricted to certain properties; all we need to do is to replace the LogP with other properties at train time. \\n\\n# Validity measure\\nValidity is defined as an actually valid molecule, confirmed using RDkit (not just proper syntax).\\n\\n# Is table 1 reconstruction really autoencoder reproduction percentage? Is the correct property for the molecule being reconstructed offered as an additional input to the authors' model? isn't this a bit unfair?\\nThe base VAE model is in a supervised VAE, by default it takes x and y as input and try to reconstruct x \\u2014 so, in this case, they both have the same information (both molecule embedding z and property y) at generation time. This highlights that the reconstruction performance of our model does not degrade, despite more successful incorporation of the property into the generative model.\\nHowever, even for unsupervised standard VAE, it takes as input x, and tries to reconstruct x. In both model information about x is given to reconstruct x. The only difference is the first model tries to encode all information about x into z and the second model tries to encode in z all information about x beside the y.\"}", "{\"title\": \"Response to Review #2(continuation)\", \"comment\": \"#On the disentanglement behavior of our model\\nWe believe that there is a confusion in the understanding of our model and objective function. Let us consider the following three generative models (similar to the discussion with reviewer 1)\", \"standard_vae\": \"x -> z -> \\\\hat{x}\", \"supervised_vae\": \"x -> z, (z,y) -> \\\\hat{x}\\nOur model (eq 3):\\t\\tx->z (z,y) -> \\\\hat{x}, f(\\\\hat{x}) = y\\n\\nFirst, a comment on the disentangling properties of Standard VAE and Supervised VAE. Indeed the KL divergence of the standard ELBO (eq. 2) has a disentangling-promoting effect, due to the fact that it pushes the learned posterior to be close to an isotropic Gaussian. Nevertheless, what is disentangled in the standard ELBO (eq 2) is only the dimensions of the latent variable z. Instead what we want to achieve is the disentanglement of the learned latent variable z from the property y. So let us take a look on how the three models listed above operate, repeating a part of our answer to reviewer 1. \\n\\nIf the property y is a function of x, and we have a perfect generative model (and corresponding encoder), i.e. one that gives us a perfect reconstruction, then the reconstructed molecule \\\\hat{x} will match x and thus will have the property y. This means that one can just use 1) standard VAE and the reconstructed x will have the appropriate property. It also means that in 2) supervised VAE there is no reason necessarily for the decoder to consider the information brought by y, since everything is inside z. Of course we will not be able to do conditional generation and control y at will, since in the case 1) standard VAE we have no y, while in the case of 2) supervised VAE the y is potentially ignored by the decoder. \\n\\nWe can speculate that this trivial behavior for 2) Supervised VAE is indeed what happens. If we look at the results we give in Table 1, where SD VAE is what we call above the standard VAE, and Sup-VAE-X-GRU is the supervised VAE, we see that these two models have almost an identical reconstruction performance which provides some evidence towards the fact that the Supervised VAE does ignore the label information y. \\n\\nIn our model we add the constraint that the reconstruction \\\\hat{x} should have the y property, i.e. f(\\\\hat{x}) = y. However if we only train with z,y pairs, where z is an encoding of a true x and y is the property of x, i.e. what reviewer 1 calls keeping the y fixed during training, then the model can still collapse in the trivial case discussed above in which it only considers z; since z can contain all the information needed to reconstruct x and then trivially match the constraint f(\\\\hat{x}) = y. Thus any model that has a perfect reconstruction will trivially satisfy the requirement that \\\\hat{x} has the property y, including our model and the supervised VAE. \\nBy arbitrarily switching y (last equation section 3.4) we make sure that the decoder cannot have access to the label information through z, since z and y are now by construction conditionally independent given x. As a result, we force the decoder to rely on y. This switch originates in our disentanglement regulariser (eq 5) were one should note that y_i and z are independent. y_i comes from the training data but z is randomly sampled from its p(z) prior. Thus there is no relation between y_i and z and no way for the decoder to get the correct y from z since these are now independent by the very definition of our disentanglement regulariser (eq 5)\\n\\n#Figure 5 and figure readability\", \"regarding_figure_5\": \"To produce the figure we randomly sampled 9 QM9 molecules, the ones surrounded by the dotted rectangles. For each one of them we retrieve its latent representation z and combine it with 11 different property values and then generate the respective new molecules. For each original molecule we give the generated molecules ordered along the y axis according to the y property that they actually exhibit. The x-axis does not provide an ordering of the original molecules according to z, in fact we have ordered the original molecules by their y property. The graph should be seen column-wise. We will clarify this. And we will correct the readability of the other figures. We updated the figure and tables in the new version to make it more readable.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"We would like to thank you for the detailed comments. We addressed each question as follows:\\n\\n#Relation to work on disentanglement\\nWe want to emphasize that there is a very important difference between what we do and the disentanglement work that the reviewer mentions. These works are all doing unsupervised disentanglement, i.e. they seek to disentangle the components of the learned z latent variable. Instead what we do is disentangling the latent variable z from the observed property y, as explained above. We want to make sure that z contains no information about y. These are two very different settings, a fact that guided our decision not to make a more extensive literature review on the unsupervised disentanglement work. We can add discussion which contrasts this paper (and other semi-supervised disentanglement work) to the references suggested here.\\nBy disentangling z and y we are able to do directly conditional generation and style transfer, making our work the first that can directly address conditional generation tasks. Previous work that seeks to generate molecules with certain properties require additional optimization and search steps, or can only generate molecules were a target property is maximized. \\n\\nThe work of [Ryu19] can do style transfer as well, as they present in Fig 2 in their paper. However, the problem setting is totally different from our Their work aims to learn a common representation z as a shared concept given two correlated data variables (x,y), i.e., a pair of images, and a local representation to capture the remaining randomness. As a result, it allows them to do conditional generation and style transfer, as the shared components are isolated into a single latent variable. However, the model consider a pair of correlated data and aim to learn a common representation. This would be useful if we were aiming to extract commonality from, say, a dataset of pairs of similar molecules; however, this is not applicable to our setting, where the property y is not \\u201cdata\\u201d which we could posit is generated in parallel to other data x, but rather an essential property which would be an input into the generative process for x. In this case there is no \\u201cshared\\u201d representation between x and y beyond the value of y itself! We assume that the property y is a part of the main factor that generates the input data x and try to learn a representation z that captures the rest of the factor other than the properties y.\\n\\n#On the final objective (eq 11) and the non-differentiable property objective\\nThe final objective in Equation 11 does not contain a standard VAE ELBO. Its first term already derived in equation 9 does both reconstruction and property prediction jointly. The second term does the disentangling by independently sampling z for any given y_i (eq 5, see also discussion above). Moreover, we never use RDkit in our optimization and learning procedure. We operate over the decoder output, the logits, and train the property predictor f_\\\\omega over that decoder output to map to the correct property y. This allow us to circumvent the non-differentiability issues that the discrete molecule structure and RDkit raise. \\n\\n#Experiments and relation to graph-based approaches\\nThe primary contribution of our paper is the adaptation of an existing deep generative model for molecules from an \\u201cunconditional\\u201d setting with an unstructured latent space, to a model which has structured latent variables that disentangle target properties from the remainder of the representation.\\nWe chose a particular SMILES-based model for these experiments, but in fact the entire process is modular. The encoder model can trivially be replaced with any other choice of encoder from other graph-based VAE-style models for molecules. The options for the decoder are only slightly less flexible, and only require a structure where the decoder is composed of real-valued neural network layers (analogous to the function g() in the paper) which are then fed into a discrete sampling process. This structure is used by most if not all recently proposed deep generative models for molecular graphs and SMILES strings which we are aware of.\\nOur model can do direct conditional generation of molecules that exhibit a desired property value; neither [Jin18] nor [You18] can do direct property optimization. More specifically, [Jin18] every time they want to obtain a molecule with a specific property value they have to go through an additional optimization/search step, e.g. Bayesian optimization, in the latent space. [You18] follow an RL approach to molecule generation, which by construction is more difficult to train. Moreover, their reward considers whether the produced molecules match a desired target, such as a desired property value range. Every time then, that one wants to generate molecules with a different property value range, one has to retrain the model from scratch.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"Thank you for the constructive feedback. We addressed each question as follows:\\n\\n#Why keeping the y fixed will not work:\\nLet us consider the three following generative models\", \"standard_vae\": \"x -> z -> \\\\hat{x}\", \"supervised_vae\": \"x -> z, (z,y) -> \\\\hat{x}\\nOur model (eq 3):\\t\\tx->z (z,y) -> \\\\hat{x}, f(\\\\hat{x}) = y\\nIf the property y is a function of x, and we have a \\u201cperfect\\u201d generative model and encoder pair, i.e. one that gives us a perfect reconstruction, then the reconstructed molecule \\\\hat{x} will match x and thus will have the property y. This means that one can just use 1) standard VAE and the reconstructed x will have the appropriate property. It also means that in 2) supervised VAE there is no reason the decoder *needs* to consider the information brought by y, since everything is inside z. Of course, we will not be able to do conditional generation and control y at will, since in the case of 1) standard VAE we have no y, while in the case of 2) supervised VAE the value of y may be mostly ignored by the decoder.\\n\\nWe can speculate that this trivial behavior for 2) Supervised VAE is indeed what happens. If we look at the results we give in Table 1, where SD VAE is what we call above the standard VAE, and Sup-VAE-X-GRU is the supervised VAE, we see that these two models have almost an identical reconstruction performance which provides some evidence towards the fact that the Supervised VAE does ignore the label information y. \\n\\nFurthermore even if in our model we keep y fixed to its true value the model can still collapse in the trivial case discussed above in which it only considers z; since z can contain all the information needed to reconstruct x and then trivially match the constraint f(\\\\hat{x}) = y. Thus any model that has a perfect reconstruction will trivially satisfy the requirement that \\\\hat{x} has the property y, including our model and the supervised VAE. \\n\\nBy arbitrarily switching y (last equation section 3.4) we make sure that the decoder cannot have access to the label information through z, since z and y are now by construction conditionally independent given x. As a result, we force the decoder to rely on y. We will elaborate more in the text on why y cannot be held fixed.\\n\\n#\\u201dActivity cliff property\\u201d\\nIndeed the \\u201cactivity cliff\\u201d property is very pertinent. First, we note here that our model imposes no constraint on what y should be; thus we can train our model on problems that more drastically exhibit such behavior. However, we also would contest that LogP is sufficiently discontinuous with respect to molecule structure, and particularly to single-molecule (or single-character for SMILES) modifications, as to demonstrate a relevant proof of concept. In particular, note that a diverse array of structures is generated by our model for any given target LogP \\u2014 the important aspect, as also related to the discussion above, is that the property y is only weakly informative as to the molecule structure.\\n\\n#Relation to \\u201cpre-image\\u201d related work\\nDefinitely the problem of finding pre-images is challenging! Part of the goal of this setting (and indeed, autoencoder setting in general) is that we do not have to explicitly perform a search to find the pre-image of the computation of an embedding, as done in e.g. [1]. Instead, the generative model / decoder network is trained to reconstruct the original input molecule from the feature embedding / latent space. While this may still not a perfect inverse mapping of the latent space embedding (i.e. the encoder network), we note that (a) we measure this as the reconstruction accuracy in Table (1) and find it to be quite high ( 99 % for QM9 and 88.6 % for ZINC) and (b) the hope is that the inaccuracy is well-captured by the stochasticity in the generative model, as we do not in general try to deterministically invert the feature mapping but instead provide a distribution over molecules. We\\u2019d be happy to add a bit of discussion along these lines.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"# Summary\\nThe paper considers the problem of generating molecules with desired properties using a variant of supervised variational auto-encoders. The key novelty is the idea to learn a representation with disentangled molecular properties which can be modified during generation of novel molecules.\\n\\nTo account for the diversity of molecules in a particular target class the conditional probability of a molecule x given a target property y is modelled with a latent variable z, i.e., p(x | y) = \\\\int p(x | z, y) p(z) dz.\\n\\nThe \\\"style transfer\\\" is achieved by taking an initial molecule x and computing the posterior of novel molecule x' with modified property y' via p(x' | y', x) = \\\\int p(x' | y', z) p(z | x) dz.\\n\\nThe latent representation is learned using a supervised auto-encoder (Kingma et al., 2014), formally specified in Eq. (2). The optimization problem is then further extended by imposing a soft constraint that the molecules from a particular class exhibit a certain property (Eq. 3). The reason for this constraint is to enforce that the conditional probability p(x | y, z) takes into account the information in label y. This part might not be sufficiently well-motivated and would benefit from strengthening the argument for the property predictor and soft constraint (would an illustration be possible here?). In particular, why should it not be possible to encode (and keep fixed during training) the information present in y into the sufficient statistics of p(x | y, z)?\\n\\nThe soft constraint involves an oracle function (see Eq. 3), which evaluates the designed molecules. Section 3.2 deals with the approximation of the oracle via a property predictor function to side-step the fact that the oracle is not necessarily differentiable or CPU-bound. Section 3.3 and 3.4 deal with joint optimization of the supervised variational auto-encoder and gradient estimates.\\n\\nThe approach is evaluated on two datasets, relative to state-of-the-art baselines based on variational auto-encoders and deep learning. The main goals of the experiments are to establish that the approach can learn useful disentangled conditional distributions over molecules and to assess the extent of improvement in the data generation process as a result of using the soft constraint term. The focus of the experiment is on generating molecules with certain range of logP values. The data generating process is simulated/evaluated while controlling for the logP value range and molecular structure.\\nThe experiments provide an objective assessment of the merits of the approach and indicate clear advantages over prior work. \\n\\n\\n# Recommendation\\nThe paper is very well written, easy to follow, and properly structured. The work is novel and the idea to disentangle molecular properties from the target property is quite interesting. The experiments are detailed and compare to baselines based on variational auto-encoders and deep learning. The part that could be improved is the choice of the target property. In particular, it would be great to evaluate the approach on the problem of finding molecules binding to a certain protein site. Such a problem is likely to exhibit scaffold hops and activity cliffs which make drug design problems very difficult (e.g., see [6-8]).\\n\\n\\n# Related work (additional references)\\n- While mapping of the discrete space of molecules to a continuous latent space (in active learning approaches combined with VAE) allows for gradient computation and the use of active learning/optimization techniques, the very difficult problem of finding latent space pre-images persists (in general, the problem should be in the NP class). The latter refers to finding a molecule that maps to a fixed point in the latent space. There have, however, been some efficient approximations in special cases [1-3].\\n- In [4] and [5], an approach for generation of molecules with desired properties has been pursued with posterior sampler based on a conditional exponential family model. The representation is based on a tuple kernel mapping mapping (x, y) to some reproducing kernel Hilbert space (without disentanglement). The approach provides a consistent data generation process and a mean to deal with the fact that designs are not independent and identically distributed (e.g., Eq. 2 assumes that examples are IID).\\n\\n\\n# Black-box target property\\n- The logP property might not be a good proxy for the effectiveness in generating molecules with desired binding activities (to some protein site) because (to the best of my knowledge) it does not exhibit the \\\"activity cliff property\\\" characteristic to drug design. This refers to the property that a small change in molecular structure results in a completely different activity level (e.g., see [6-8] and [4]). In this sense, the \\\"style transfer\\\" might be beneficial for finding molecules binding to a certain protein site that are structurally similar to some available molecules with bad activity levels (for which a synthesis path might be known or easy to derive).\\n- An actual black-box property (realized via a docking program) that is expensive to evaluate served as a target property in [5]. That docking program (supplied with suitable binding/docking constraints) can provide a nice proxy for the actual generation of molecules.\\n\\n\\n# References\\n[1] G. H. Bakir, J. Weston, and B. Schoelkopf. Learning to find pre-images, NIPS 2004.\\n[2] C. Cortes, M. Mohri, and J. Weston. A general regression technique for learning transductions, ICML 2005.\\n[3] S. Giguere, A. Rolland, F. Laviolette, and M. Marchand. On the string kernel pre-image problem with applications in drug discovery, ICML 2015.\\n[4] D. Oglic, R. Garnett, and T. Gaertner. Active search in intensionally specified structured spaces, AAAI 2017.\\n[5] D. Oglic, S. Oatley, S. Macdonald, T. Mcinally, R. Garnett, J. Hirst, and T. Gaertner. Active search for computer-aided drug design, Molecular Informatics 2018.\\n[6] G. Schneider and U. Fechner. Computer-based de novo design of drug-like molecules, Nature Reviews Drug Discovery, 2005.\\n[7] P. Schneider and G. Schneider. De novo design at the edge of chaos. Journal of Medicinal Chemistry, 2016.\\n[8] J. Scannell, A. Blanckley, H. Boldon, and B. Warrington. Diagnosing the decline in pharmaceutical R&D efficiency. Nature Reviews Drug Discovery, 2012.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a VAE-based conditional molecular graph generation model. For that purpose, the disentanglement approach is adopted in this paper: learn to separate property information from the structure representation of a molecular graph. The authors use the supervised VAE objective since the KL regularizer in the objective has reportedly disentanglement-promoting effect. The final objective function is a standard VAE ELBO plus a penalty term of the property value prediction error. Non-differentiable property estimation is conducted via stochastic sampling expectation with the help of an external program (RDKit).\\n\\nThe proposed model directly optimizes the generation model conditioned on the property values in a single objective function. This is preferable compared to many existing molecular graph generations, where property optimization is often carried out after the core generative models are trained and fixed. \\nDerivation of a stable lowerbound (Eq. 11) is another plus. \\n\\nMy main concern about the paper is a weak survey for disentanglement researches. Since the core of graph generation model is not original of the authors (adopted from Dai+, 2018), the main technical advancement of the paper should be related to the disentangling VAE modeling. \\nCurrent researches in disentangling VAEs go beyond the InfoGAN and beta-VAE. I present a part of the must-referred papers below:\\n\\n[Gao19] Gao+, \\u201cAuto-Encoding Total Correlation Explanation\\u201d, AISTATS, 2019. \\n[Kim_Mnih18] Kim and Mnih, \\u201cDisentangling by Factorising\\u201d, ICML, 2018. \\n[Ryu19] Ryu+, \\u201cWyner VAE: Joint and Conditional Generation with Succinct Common Representation Learning\\u201d, arxiv:1905.10945, 2019. \\n\\nAmong them, [Ryu19] is closely related to the proposed framework. In my understanding, the variable dependency structure studied in this paper (Fig. 1) is studied by [Ryu19]. The authors should clearly state the novelty of the proposed work in the literature of these disentanglement VAE works. \\n\\nI think the L_disent (Eq.5) is not a penalty for disentanglement: it enforces the model to correctly predict target property values, and do not say anything for factor disentanglement, correct? \\n\\nI have a few concerns about the experiment designs. \\nIn the table 1, reconstruction performance evaluations, the authors chose string(SMILE)-based molecular graph generation models. However, it is largely admitted in the molecular graph generation studies that the graph-based generative models generally performed better. I want justifications for the choice of string-based generation models. Extending the table 1 with graph?base SotA generation models will strengthen the manuscript greatly.\", \"for_example\": \"[Jin18] Jin+, \\u201cJunction Tree Variational Autoencoder for Molecular Graph Generation\\u201d, ICML, 2018. \\n[You18] You+, \\u201cGraph Convolutional Policy Network for Goal-Directed Molecular Graph Generation\\u201d, NeurIPS, 2018. \\n\\nEspecially, [You18] directly optimize the property values of the generated graph. This is closely related to the goal of this paper, thus a proper reference and discussions are strongly expected. \\n\\n\\nFigure 5. is an attractive visualization of the latent (z, y) vectors. However, the authors do not provide how to understand the figure. I GUESS that the authors want to show that learned z and y are somehow disentangled: the structural changes of molecular graphs are less sensitive to the vertical axis (property value) than the horizontal column (different). Please clearly state key messages of each figure and table. \\n\\nThere are several presentation issues. They are details, but fairly degrades the readability of the manuscript. \\n- Too small letters (alphabets) in Figure 3, 4, 6, it is simply unreadable so I cannot tell main messages of these figures (So I do not evaluate these figures positively). \\n- In table 1, what are the significance figures of the presented scores? Some scores have 3 digits, others have 4 digits. Some are rounded at 0.01 precision but others are round at 0.1 precision. \\n\\nEvaluation summaries\\n+ A graph generation model combining conditional property optimization in a single objective function.\\n+ A new lowerbound for stable training (Eq.11)\\n+ The property and the structure of the graph are somehow disentangled in the experiment (Fig. 5). \\n+- Disentanglement effect is not modeled directly in the proposal. L_disent is not for disentanglement but for property regression.\\n-- Survey for disentangling VAEs is not enough. This makes the proposal less convincing in terms of the novelty and the contribution compared to the recent dissent-VAEs. \\n- Not compared with SotA graph-based molecular graph generation models. \\n- Key messages of figures and tables are not clearly stated (e.g. Fig.5). \\n-- Some figures are simply unreadable. The significance figure of the table 1 is unclear. These are formatting details but essential for readability.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors introduce a variational autoencoder for conditional generation of molecules. The model is borrowed from text-based style transfer, applied here on sequence (SMILES) representation of molecules rather than viewing molecules as graphs (as more recent approaches). From a modeling point of view, the main new part is an additional regularizer whose role is to 1) ensure that the property used as input during generation matches the property derived from the generated molecule, and 2) to dissociate the latent molecule representation in the autoencoder (loosely speaking, its overall structure) from the property being controlled. This regularizer is just a squared difference between predicted and actual properties, averaged over independent samples from latent states and properties which are parametrically mapped to predicted properties via the decoder state (so as to be able to backprop).\\n\\nThe resulting ELBO criterion for the VAE, together with the regularizer, could be in principle directly used to train all the model parameters (encoder, decoder, property mapping) but apparently this criterion does not result in a \\\"stable\\\" estimation procedure. The model is then amended in various ways. \\n\\nThe first change consists of introducing a property prediction part directly into ELBO so that the posterior over latent states will also be affected by property prediction (previously property prediction would only appear in the additional regularizer). Beyond stability that the authors cite as the reason, this is likely also needed because in eq (2) the posterior over latent states given the molecule does not depend on the property. The posterior is derived from a v-structure where latent states and properties appear as independent parents of the molecule. Since for the training cases the property is fixed, this would be fine with an arbitrarily complex posterior encoding. However, when the encoder is defined parametrically, this dependence may be needed, and may relate to the stability issues. It seems a bit roundabout way of putting the dependence back in, if so. The authors should comment on this further. \\n\\nThe other simplification that is introduced is that the prior over the latent states in the regularizer is modified to be the average posterior of training encodings, and later further collapsed (for computational reasons) to an adaptively estimated simple Gaussian with variance that matches the average posterior variance.\\n\\nIt's a bit unsatisfying that the authors only use and experiment with a simple logP as the property to control in the model. This is not really a particularly relevant property to control. The approach should in principle generalize to other properties but the many adjustments needed to make it work makes this a little questionable. Also, the results are a bit limited (only logP) and not encouraging (see comments below). \\n\\n- how is validity defined? SMILES syntax or validity as a molecule? \\n\\n- is table 1 reconstruction really autoencoder reproduction percentage? Is the correct property for the molecule being reconstructed offered as an additional input to the authors' model? isn't this a bit unfair?\\n\\n- the correlation between property given as input and the property derived from the generated molecule does not seem like a good metric. Wouldn't a model that overfits (essentially spitting out the nearest property neighbor from the training set) do particularly well according to this metric?\\n\\n- it seems Figure 6 can be interpreted as also indicating some degree of overfitting?\\n\\n- the stacked LSTM molecular generator baseline that takes the property as input seems to do much better in terms of maintaining the property in its generated molecules and it also realizes a more diverse collection of molecules. It's true that it doesn't support incremental molecular manipulation directly, eg, if one is keen on keeping the modified molecule close to a starting point. But it is so much better at maintaining the input property in its diverse realizations that perhaps one should instead invest in a tailored sampling procedure to obtain realizations close to a chosen reference.\\n\\nThe literature references seem to omit all molecular generation work since 2018\"}" ] }

[ "Edo Cohen-Karlik", "Amir Globerson" ]

[ "rnns", "permutation invariant", "permutation invariance", "commutative state transitions", "tasks", "analysis", "set", "inputs", "learned functions" ]

{ "note_id": [ "Mv3l-P3-C", "BJgfw1hYiS", "H1ebzJ2FsB", "B1xsq6jYsH", "SkeVXO9RKH", "ByxIDYM0KS", "rye8e_rTFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751697, 1573662554067, 1573662473381, 1573662099400, 1571887132477, 1571854686341, 1571801069978 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2542/Authors" ], [ "ICLR.cc/2020/Conference/Paper2542/Authors" ], [ "ICLR.cc/2020/Conference/Paper2542/Authors" ], [ "ICLR.cc/2020/Conference/Paper2542/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2542/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2542/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper examines learning problems where the network outputs are intended to be invariant to permutations of the network inputs. Some past approaches for this problem setting have enforced permutation-invariance by construction. This paper takes a different approach, using a recurrent neural network that passes over the data. The paper proves the network will be permutation invariant when the internal state transition function is associative and commutative. The paper then focuses on the commutative property by describing a regularization objective that pushes the recurrent network towards becoming commutative. Experimental results with this regularizer show potentially better performance than DeepSet, another architecture that is designed for permutation invariance.\\n\\nThe subsequent discussion of the paper raised several concerns with the current version of the paper. The theoretical contributions for full permutation-invariance follow quickly from the prior DeepSet results. The paper's focus on commutative regularization in the absence of associative regularization is not compelling if the objective is really for permutation invariance. The experimental results were limited in scope. These results lacked error bars and an examination of the relevance of associativity. The reviewers also identified several related lines of work which could provide additional context for the results that were missing from the paper.\\n\\nThis paper is not ready for publication due to the multiple concerns raised by the reviewers. The paper would become stronger by addressing these concerns, particularly the associativity of the transition function, empirical results, and related work.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We thank the reviewer for the comments and points raised. We answer each point below.\\n\\n(1) Results are mostly trivial - we believe the results are important for context and completeness. These results should be interpreted with the limitations of DeepSets to efficiently represent certain functions such as the max function.\\n\\n(2) I am concerned the final learned networks have deepset-like architecture - Our network will be equivalent to a \\nDeepSet only if the transition rule suffices st+1=A(W(xt)+st)=(xt)+stwhich requires very specific structure over the weight matrices. We have verified this is not the case in our experiments.\\n\\n(3) Extensive description on network architecture is lacking - We will add a complete description of the configurations used.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We thank the reviewer for their detailed feedback and insights as well as the important remarks regarding prior work which we did not cite. Below we add clarifications to each point raised by the reviewer.\\n\\n(1) The authors focus on commutative without clear justification - We identify commutativity and associativity as two components of permutation invariance. In the paper we show how to regularize towards commutativity and show this is empirically effective for achieving permutation invariance. Indeed we will add quantitative evaluation to show that associativity is also achieved in these cases. Regarding setting \\\\Theta=W, this leads to a permutation invariant network (i.e., it is associative not only commutative), but one which is less expressive and may require more parameters to fit a given function.\\n\\n(2) Jannosy pooling gives an alternative way to use RNNs - Indeed Janossy gives another approach to permutation invariance using RNNs via canonical representations or sampling. We expect our approach to outperform it when sampling has high variance or the canonical representation is hard for the RNN to classify.\\n\\n(3) The proof of f(x_1,...,x_n)=max_i x_i does not show necessity and is informal - We will clarify this.\\n\\n(4) Missing work: We will add the reference and discuss relation to our approach.\\n\\n(5) Concerns regarding the operator not being associative - We will add an empirical evaluation of the associativity of the network as a function of its commutativity.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We thank the reviewer for the comments and suggestions, and address specific points below.\\n\\n(1) Section 3 is rather trivial - we believe it is important for completeness. We will consider rephrasing.\\n\\n(2) Theorem 3.6 is hard to connect to an RNN and the rest of the paper - The theorem states that any permutation invariant architecture can be implemented by an associative-commutative RNN. The proof is simply to note that the DeepSet architecture is also associative-commutative. We understand this is a bit confusing, since the DeepSet is a \\u201cdegenerate\\u201d RNN. Note however, that for a given permutation-invariant function (e.g. the parity in our example), there could be RNN implementations with much fewer parameters than the smallest DeepSet implementation.\\n\\n(3) The other explanations make things confusing - We will revise the writing style.\\n\\n(4) Could it be that the network indeed learns addition operator? See answer (2).\\n\\n(5) The approach will not work for LSTM - Indeed the regularization expression is specific to the architecture described. We agree that it would be nice to obtain a similar expression for LSTMs.\\n\\n(6) The plots are missing error bars - we\\u2019ll add them in future versions.\\n\\n(7) It\\u2019s unclear whether the authors have done cross validation - We performed cross validation, and will add the missing details and exact settings used.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper starts with presenting an RNN formulation and essentially writing out the sequence of RNN applications. Not surprisingly, if these applications were associative and commutative the RNN would be permutation invariant. Then a condition for commutativity is formulated in terms of an expectation of a difference. Based on a prior result, it is shown that the expectation can be computed in closed form. Although it is not shown if an RNN regularized that way is permutation invariant, since the associativity is not demonstrated, empirically it is shown that it may be already of use.\", \"contributions\": \"1. A regularizer for RNNs that enforces commutativity\\n 2. A closed form for computing it\\n 3. A fully learnable permutation invariant \\\"deep\\\" network, per an empirical demonstration\\n\\n The main contribution is the empirical demonstration of the learnable nature of the obtained function unlike the prior art (e.g. DeepSets) where a choice of the aggregation function severily affects the results.\", \"the_theoretical_component_of_the_paper_is_unclear\": \"1. Section 3 is rather trivial.\\n 2. Theorem 3.6 is hard to connect to an RNN and the rest of the paper. Unclear why bother learning the RNN at all if it needs to converge to addition of the input and hidden state to be universal anyway.\\n 3. In essence, the result of the paper is a way to encourage commutativity in an RNN and a demonstration that it works in practice for encouraging permutation invariance. The other explanations make things confusing and do not seem to contribute to the rest of the paper.\\n\\nCould it be that the network indeed learns addition operator? RNNs usually are only able to operate on very small sequences because of the vanishing gradient problem, yet the proposed approach will not directly work on the more robust LSTM.\\n\\nSignificance or lack of the difference between the proposed method and DeepSets is unclear as the plots are missing the error bars.\\n\\n The table and the accuracies reported in Section 8 are impossible to interpret. It is unclear whether the authors done cross validation. If so, it would be helpful to see standard deviations of the reported numbers\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The rebuttal did not address my concerns convincingly. There were also simple fixes that the authors could have implemented but they decided not to update the paper. I will keep my original assessment.\\n\\n--------------\", \"the_premise_of_the_work_is_very_interesting\": \"RNNs that are permutation-invariant. Unfortunately, the paper seems rushed and needs a better justification for not having a RNN memory that is associative. It also should cast the contributions in light of other existing work (not cited). The paper says \\\"In this section and the remainder of the paper, we focus on the latter [commutative RNN memory operator], namely introducing a constraint (or equivalently, regularizer) that is commutative\\\", but it never talks about the impact of a RNN memory using a non-associative operator. Being commutative is easy, isn't Equation (2.4) commutative if \\\\Theta = W? Being associative is hard, since non-linear activations are not easily amenable to associativity.\", \"section_4\": \"\\\"The above example demonstrates that RNNs can in some cases be a natural computational model for permutation invariant functions.\\\" => Janossy pooling (Murphy et al., 2019) gives an alternative way to use RNNs, with a way to make their method tractable. Actually, my guess to why the RNNs experiments work well, even without an associative memory, is because the training examples come in multiple permuted forms, which is the data-augmentation version of the pi-SGD optimization described in Janossy pooling.\\n\\nOn page 1, \\\"consider the problem of computing the permutation invariant function f(x_1, . . . , x_n) = max_i x_i\\\", what follows is not a proof of necessity. It is an informal argument that either should be made formal or should be described as informal.\", \"there_is_a_lot_of_missing_related_work_for_sets\": \"Murphy, Ryan L., Balasubramaniam Srinivasan, Vinayak Rao, and Bruno Ribeiro. \\\"Janossy pooling: Learning deep permutation-invariant functions for variable-size inputs.\\\" ICLR 2019.\\nWagstaff, Edward, Fabian B. Fuchs, Martin Engelcke, Ingmar Posner, and Michael Osborne. \\\"On the limitations of representing functions on sets.\\\" ICML 2019.\\nLee, Juho, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. \\\"Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks.\\\" ICML 2019.\", \"also_missing_related_work_for_graphs\": \"Bloem-Reddy, Benjamin, and Yee Whye Teh. \\\"Probabilistic symmetry and invariant neural networks.\\\" arXiv:1901.06082 (2019).\\nMurphy, Ryan L., Balasubramaniam Srinivasan, Vinayak Rao, and Bruno Ribeiro. \\\"Relational Pooling for Graph Representations.\\\" ICML 2019.\\n\\nThe paper has an interesting question but needs to build on prior work. As of now, I am unconvinced that not having an associative operator for the RNN memory will lead to a good nearly permutation invariance function (unless there is data augmentation, per Janossy pooling).\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary: this paper proposes a new principled methodology for deriving and training RNN neural networks for prediction of permutation invariant functions. Authors show on simple tasks their method may outperform DeepSets, the state of the art.\\n\\n\\nAlthough the idea is interesting and the paper reflects thorough work, I believe in its current form results are too weak to deserve publication. More specifically.\\n\\n1)Mathematical results and statements are mostly trivial and may well be omitted or included as an appendix. They don't seem to convey anything profound (with the exception of theorem 3.6, but this follows from results on deepsets paper). Some of these results are also mostly anectodal \\n\\n2)The regularization idea seems interesting, but I am concerned it is showing that the final learned networks have a deepset-like architecture: more specifically, theorem 3.6 shows RNN can implement permutation invariant functions by making identifying the parameters with the ones of deepperm. Also, as the authors mentioned, when learning a permutation invariant function then for any degree of regularization the regularization loss can be made zero. So for me, results seem to indicate that the network might have learned a deepperm kind of representation, which equivalently can be expressed as a RNN. Authors should make clear there are fundamental differents between both frameworks\\n\\n3)Overall, the experimental validation section is weak and an extensive description of network architectures is lacking. Without them it is hard to resolve my concerns on 2).\"}" ] }

[ "Maksym Del", "Mark Fishel" ]

[ "cross-lingual transfer", "sentence embeddings", "polyglot language models", "knowledge distillation", "natural language inference", "embedding alignment", "embedding mapping" ]

{ "note_id": [ "ZbmJAKoT_f", "rkeSEwFnjS", "SJxsnUY3iH", "ByeBtItnor", "r1xdX8FhoH", "rJgnL8waqB", "S1e0ZJj35B", "BygAwA40FH", "rJewS5lLtB", "Skxg25ViDB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1576798751666, 1573848876798, 1573848755279, 1573848700519, 1573848607765, 1572857427762, 1572806406386, 1571864165748, 1571322430934, 1569569448356 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2541/Authors" ], [ "ICLR.cc/2020/Conference/Paper2541/Authors" ], [ "ICLR.cc/2020/Conference/Paper2541/Authors" ], [ "ICLR.cc/2020/Conference/Paper2541/Authors" ], [ "ICLR.cc/2020/Conference/Paper2541/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2541/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2541/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2541/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2541/Authors" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a method for transferring an NLP model trained on one language a new language, without using labeled data in the new language.\\n\\nReviewers were split on their recommendations, but the reviews collectively raised a number of concerns which, together, make me uncomfortable accepting the paper. Reviewers were not convinced by the value of the experimental setting described in the paper\\u2014at least in the experiments conducted here, the claim that the model is distinctively effective depend on ruling out a large class of models arbitrarily. it would likely be valuable to find a concrete task/dataset/language combination that more closely aligns with the motivations for this work, and to evaluate whether the proposed method is genuinely the most effective practical option in that setting. Further, the reviewers raise a number of points involving baseline implementations, language families, and other issues, that collectively make me doubt that the paper is fully sound in its current form.\", \"title\": \"Paper Decision\"}", "{\"title\": \"We will work on the presentation!, comments on novelty, motivation and other details below:\", \"comment\": \"This was very insightful, thank you very much for your review! Sorry for the hasty submission, we will work on the presentation and improve the writing and clarity of the paper.\\n\\nYou are right about Urdu and Swahili, thank you for correcting us. Instead the higher translation quality (BLEU score) in XNLI for Swahili and lower score for Urdu are the more correct justification for the choices concerning these two languages.\\n\\nIn our view the novelty here is that we present 2 approaches that achieve state-of-the-art results on XNLI (at the moment of the ICRL submission deadline): one using labels for all the languages, another one using only English labels. The combination between the multilingual approach and XD gives a further small improvement, probably because they are based on the same XNLI inputs, but we still added this as an additional experiment. \\n\\nThe main motivation for our knowledge distillation approach is that it can be potentially applied to any unannotated data, as long as the data has some translations available (from a parallel corpus OR an from an MT engine). Although we did not test it on a different dataset, in that scenario it does require any labels on the new data -- also not for English, just the English labels on the original dataset (used for creating the multilingual based model).\\n\\nWe will add the individual-tuning results done on our own runs, as you suggested. We will also add statistical significance results via bootstrapping and re-run the key models of the paper to estimate the variance, as you proposed.\\n\\nConcerning the choice of languages, it was mostly guided, not ad-hoc: for example, the removal of Urdu was driven by its lowest BLEU score of the MT system used to create the XNLI dataset; the 4 languages of the combination were indeed chosen without a thorough comparison (why not 3? 5?), but again based on their BLEU score and NLI performance of the baseline model. The idea was just to illustrate the sensitivity of XD to the quality of multilingual \\u201calignment\\u201d data.\\n\\nWe agree that a more detailed investigation of language combinations would be exciting, we did our best to explain the motivation for the choices we did in this paper and we hope to provide a more thorough comparison in a followup paper.\"}", "{\"title\": \"Will improve the writing; some comments on significance, language choices and method combination:\", \"comment\": \"Thank you for your review! We will make sure to improve the writing and clarity of the paper, sorry for making you read a hurried submission. We will definitely incorporate your smaller comments.\\n\\nWe did not run any significance test for the submitted version; since it was suggested by several reviewers, we will do it via bootstrapping, as well as re-run the key models to estimate the variance of their results.\\n\\nWe agree that a more detailed investigation of language combinations would be exciting, we did our best to explain the motivation for the choices we did in this paper. We hope to provide a more thorough comparison in the next paper.\\n\\nAs for \\u201cSimilarly: A combination of MTL and XD doesn't seem straightforward. Why? What is learned?\\u201d:\\nIn the case of MTL, data from low-resource unrelated (to English) languages were used just as well as data for resource-rich high-translation-quality languages. Our case studies with particular languages suggest that the quality of the parallel signal matters for XD. By using XD with only a high-quality parallel signal from high-resource languages we are able to further improve the system learned with multilanguage finetuning that used data from all sources.\"}", "{\"title\": \"Will improve the writing; comments on novelty and language choices below:\", \"comment\": \"Thank you for your review and sorry for making you read a rushed submission -- we will definitely improve the writing and clarity of the paper.\\n\\nIn our view the novelty here is that we present 2 approaches that achieve state-of-the-art results on XNLI (at the moment of the ICRL submission deadline): one using labels for all the languages, another one using only English labels. The combination between the multilingual approach and XD gives a further small improvement, probably because they are based on the same XNLI inputs, but we still added this as an additional experiment. \\n\\nConcerning the choice of languages, it was mostly guided, not ad-hoc: for example, the removal of Urdu was driven by its lowest BLEU score of the MT system used to create the XNLI dataset; the 4 languages of the combination were indeed chosen without a thorough comparison (why not 3? 5?), but again based on their BLEU score and NLI performance of the baseline model. The idea was just to illustrate the sensitivity of XD to the quality of multilingual \\u201calignment\\u201d data.\"}", "{\"title\": \"Will add significance and improve the writing; some comments on zero-shot inside:\", \"comment\": \"Thank you for your review!\\n\\nWe will definitely improve the writing and clarity of the paper, sorry for making you read a rushed submission. Thank you for the specific improvement suggestions, we will surely incorporate them.\\n\\nYou are correct that our method has its differences in comparison to other zero-shot approaches in terms of the parallel signal we are using. However, our alignment approach (XD) does not use human labels in the XNLI dataset -- neither English, nor other languages, only a model trained on the annotated English data. Thus, the method can also be potentially applied to any unannotated data, as long as the data has some translations available (from a parallel corpus OR an from an MT engine). So it still belongs to the family of methods that don't use labels for target languages. We already highlighted the differences with other zero-shot methods in the results table and in the text and we will make sure to make it even more clear in the paper.\\n\\nAs for the significance of the comparisons, we initially used the term generally, not in the statistical sense, when referring to the larger gap comparisons (like individually tuned vs our final approach combination). However, we will add statistical significance tests via bootstrapping, and we will also do a couple of random restarts for the key models of the paper to estimate the variance of their results.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper addressed multilingual natural language inference. The motivations of the authors are two folds: 1/ to have only one model for all the languages instead of one model per language and 2/ achieving good results in a zero shot setup where only the english labels are available.\\n\\nPrevious work proposed first to learn multilingual language model in a self-supervised manner for the initialization. Then, fine-tuning it on the final task, separately for each different language. It results in multiple models, one for each language.\\n\\nThe authors proposed to fine-tune the model with all the data from the different languages simultaneously, resulting in only one model with comparable results.\\n\\nIn addition, they also proposed a method based on distillation where only the english targets are used. While the model doesn't require the label data for the other languages, the scores remained similar to the previous experiments.\\n\\nFinally the authors proposed to combine both methods. It compares favorably, obtaining 4 points of improvement over SOTA in the zeroshot setup.\", \"pros\": \"-the motivation for having only one model is interesting\\n-the results are promising\", \"cons\": \"-one of main motivation of the paper is to achieve zeroshot as opposed to previous work. To that purpose, the authors chose to keep only the translated non-en input and not the non-en target. In XNLI, all the non-en training data, including target, are automatically translated from the English data. Therefore, the authors did not use less human annotated data and their approche still requires automatic translation. Hence, the motivation to perform the task in a zero-shot scheme, as opposed to previous work, doesn't seem correct. \\n-the paper was not always easy to follow and would benefit from more clarity.\\n'large margin of 5.9 and 4.2 points' in 3.2.1, please add the reference to table 6.\\n-the method 'significantly' improved results. I didn't see any significance measurements and it would be important to add them.\\n\\nOverall, using all the data together seems like a natural and effective approach to and achieves good results through only one model. \\nHowever, the motivation behind 'distillation' is to perform in a zeroshot scheme. It seems abusive to me since it actually requires the exact same amount of human labels than the previous works.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"What is the task?\\nMultilingual natural language inference (NLI)\\n\\nWhat has been done before?\\nCurrent state-of-the-art results in multilingual natural language inference (NLI) are based on tuning XLM (a pre-trained polyglot language model) separately for each language involved, resulting in multiple models.\\n\\nWhat are the main contributions of the paper?\\n[Not novel] Significantly higher average XNLI accuracy with a single model for all 15 languages.\\n[Moderately novel] Cross-lingual knowledge distillation approach that uses one and the same XLM model to serve both as teacher (for English sentences) and student (for their translations into other languages). The approach does not require end-task labels and can be applied in an unsupervised setting\\n\\nWhat are the main results? \\n A single model trained for all 15 languages in the XNLI dataset can achieve better results than 15 individually trained models, and get even better when unrelated poorly-translated languages are removed from the multilingual tuning scheme.\\n Using XD they outperformed the previous methods that also do not use target languages labels.\", \"weaknesses\": \"1. Combining XD with multilingual tuning is not effective in improving average results or even in case of target languages\\n2. Final system is adhoc as experiments on a particular set of languages have been used to support claims. For example, Urdu was excluded to get the best MLT model. Only 4 languages were used while combining XD and MLT\\n3. Findings, methods and experiments are not strongly novel.\\n4. Paper was not an easy read.\", \"strengths\": \"Using XD they outperformed the previous methods that also do not use target languages labels.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"First, the authors propose to train a model for natural language inference (NLI) on multiple languages simultaneously. In particular, they translate English examples into all target languages and fine-tune a pretrained language model on all thereby obtained data at once. This is different from the previous state-of-the-art approach which consisted of, after translating from English into target languages, fine-tuning one NLI model for each language individually. The authors show that their approach is superior to training individual models for each language. For evaluation, XNLI is used.\\n\\nSecond, they introduce cross-lingual knowledge distillation (XD), where the same polyglot model is used both as teacher and student across languages to improve its sentence representations without using the target task labels. The main idea is that the same sentence in all languages should receive output representations as similar as possible.\\n\\nThe paper seems okay to me and the experiments seem solid. However, the results are not particularly surprising and the methods are not very innovative. The writing could be improved.\", \"this_paper_could_further_be_improved_in_the_following_ways\": [\"A more detailed investigation which combination of languages improve performance (and why?).\", \"Similarly: A combination of MTL and XD doesn't seem straightforward. Why? What is learned?\"], \"smaller_comments\": [\"Articles are missing frequently (e.g., \\\"we substitute the word prediction head with classification layer\\\" -> \\\"we substitute the word prediction head with a classification layer\\\")\", \"Table 5: \\\"w/0\\\" -> \\\"w/o\\\"?\", \"Have you run any significance tests?\"]}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": [\"This paper proposes two improved strategies for fine-tuning XLM (a multilingual variant of BERT) for cross-lingual NLI. First of all, it shows that fine-tuning a single model on the combination of all languages (the original English data from MultiNLI and their MT translation into the rest of languages) performs better than fine-tuning a separate model for each language. Furthermore, they show that minimizing the L2 distance between the English training sentences and their MT translation into the rest of languages, which does not explicitly use any labels in the foreign languages and is presented as a way of performing cross-lingual knowledge distillation, also performs better than zero-shot transferring a regular model fine-tuned in English.\", \"I think that the paper makes some interesting contributions and, in particular, I think that the finding that multilingual fine-tuning performs better than the standard approach of fine-tuning a separate model for each language is important. Nevertheless, I am not convinced that there is enough novelty and substance on this, I have some concerns on the evaluation, and I think that the overall presentation should also be improved:\", \"I am not convinced by the \\\"knowledge distillation\\\" approach. First, although I see the connection, I do not think that presenting this as \\\"knowledge distillation\\\" is consistent with the common use of this term in the literature. More importantly, I do not see what is the value of this approach considering that multilingual fine-tuning performs better, and combining them both does not bring any clear improvement. The authors motivate it as a form of performing zero-shot cross-lingual transfer as, unlike the multilingual fine-tuning, this approach does not use any label in the foreign languages. However, I am not convinced at all by this reasoning, as it still relies on the translation of the English labeled data into the other languages. So, from a practical perspective, it requires the exact same resources as the other approach, as you would always be able to use the English labels for the rest of the languages, while being more complex and worse.\", \"It looks like the IndT results, which is the real baseline, are taken from the original XLM paper, while the rest of the results come from the authors' own runs, who use a different implementation. I think that you should also report IndT results from your own runs to make sure that your improvements come from the actual method, and not from small implementation details.\", \"You are trying small variations of your method (e.g. removing a particular language from the multilingual training) to support your claims, and it is not clear if the (rather small) differences in the results are significant. It would be good if you at least run the baseline multiple times and show the variance.\", \"It is unfair to try so many variants of your method in the test set, and then pick the best and claim SOTA as done in Table 6. Your final system looks rather ad-hoc and arbitrary: it is doing multilingual fine-tuning in all languages but Urdu, and cross-lingual knowledge distillation in a subset of 4 languages out of 15. It might get SOTA results in this particular scenario, but what if we move to a different set of languages, a different task, or even a different test set?\", \"The authors claim that \\\"Urdu (ur) is an unrelated language\\\" and \\\"Swahili (sw) is loosely between French and Urdu in terms of relatedness to English\\\", which they use to justify why Urdu behaves differently in their experiments. I do not speak neither Swahili nor Urdu, but from what I know this statement looks wrong. Swahili and English belong to completely different language families, and from what I know their grammar is very different. In contrast, Urdu at least belongs to the Indoeuropean language family.\", \"This is not relevant at all, but I would suggest the authors to find a different acronym instead of XD, which happens to be a widely used emoticon. I assume that the authors deliberately made this choice thinking that it would be funny, but I just find it confusing to see XD all over the place in a formal paper.\"]}", "{\"comment\": \"We could not guarantee anonymity in the shared code (license, Google Drive user visibility) so we had to switch the sharing off till de-anonymization.\", \"title\": \"Code and model sharing\"}" ] }

[ "Andrea Dittadi", "Ole Winther" ]

[ "structured scene representations", "compositional representations", "generative models", "unsupervised learning" ]

{ "note_id": [ "Cz0iRDKaaz", "HyeSe_lhoS", "HJeeQSlnsS", "SygdvztCKB", "H1xyP11TYH", "B1xFs1anFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751635, 1573812205219, 1573811479846, 1571881568102, 1571774295312, 1571766176880 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2540/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2540/Authors" ], [ "ICLR.cc/2020/Conference/Paper2540/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2540/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2540/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper presents a VAE approach where the model learns representation while disentangling the location and appearance information. The reviewers found issues with the experimental evaluation of the paper, and have given many useful feedback. None of the reviewers were willing to change their score during the discussion period. with the current score, the paper does not make the cut for ICLR, and I recommend to reject this paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"I'm keeping my score\", \"comment\": \"Thanks for the response. I appreciate that you are going to compare your method with other baselines. I cannot take your word for it when evaluating the paper, though, and it has no influence on my score.\", \"other_remarks\": [\"I do realize that SCALOR was made available only after the ICLR deadline. I cited it because it solves a very similar problem.\", \"Even if the other works do not report marginal likelihoods or ELBOs, I don't think that merely reporting them does constitute a big contribution. Moreover, Sequential AIR of Kosiorek et. al. does report ELBOs and estimates of the log-ikelihood for AIR.\"]}", "{\"title\": \"Response to reviewers\", \"comment\": \"We thank the reviewers for their thorough and helpful comments. The reviewers agree that our approach is interesting and qualitative results are impressive, but also deem the experimental evaluation insufficient, and suggest that an experimental comparison with related methods is needed. We agree with these general points, and we plan to extend the experimental section and compare our results with methods such as AIR [1], SPAIR [2], and SuPAIR [3] in the future.\\n\\nHowever, we would like to make a few remarks:\\n\\n* SCALOR [4] is a concurrent submission to ICLR, and was uploaded to arxiv after the deadline (October 6).\\n\\n* The papers cited by the reviewers are not really concerned about the generative aspect of the models, and they do not show generated samples or report log-likelihood scores. Comparison with previous methods is mostly qualitative and not necessarily very thorough. For example, in the AIR paper the ELBO is only shown in plots and not explicitly reported, and estimates of the marginal log-likelihood are not mentioned. The SuPAIR paper reports the ELBO for SuPAIR and AIR, but because of the choice of data likelihood, these scores are not comparable, as also stated by the authors. The performance evaluation of SPAIR is entirely focused on counting accuracy and related downstream tasks, overlooking typical generative modeling metrics.\\n\\n* Real-world examples are rarely considered as they present a different set of difficulties. SCALOR is tested on crowd tracking from surveillance cameras, but the fact that qualitative results on that data set are good is debatable (predicted frames after only 1-2 time steps).\\n\\n* \\\"Has been achieved previously and in a very similar fashion by SPAIR\\\": there are in fact different mechanisms involved, and experimental results (which are now missing) should show to what extent SPAIR and our model can solve the problem under consideration.\\n\\n* MNIST digits are rescaled so that more of them can fit in an image, as done in related works. They are binarized so the data can be modeled as a set of independent Bernoulli variables -- a common approach in generative modeling on simple data sets.\\n\\n* We use the definition of disentanglement used for example in [5] and [6].\\n\\n\\n\\n[1] Eslami et al., \\u201cAttend, Infer, Repeat:...\\u201d, NIPS 2016.\\n[2] Crawford and Pineau, \\u201cSpatially Invariant Unsupervised Object Detection with Convolutional Neural Networks\\u201d, AAAI 2019.\\n[3] Stelzner et al., \\u201cFaster Attend-Infer-Repeat with Tractable Probabilistic Models\\u201d, ICML 2019.\\n[4] Jiang et al., \\u201cScalable Object-Oriented Sequential Generative Models\\u201d, arXiv 2019.\\n[5] Higgins et al., \\\"beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework\\\", ICLR 2017.\\n[6] Locatello et al., \\\"Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations\\\", ICML 2019.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"After reading reviews and comments I have decided to confirm the initial rating.\\n\\n===================\\n\\nThe work presents an approach to encode latent representations of objects such that there are separate and disentangled representations for location and appearance of objects in a scene. The work presents impressive qualitative results which shows the practical use of the proposal on multi-mnist and multi-dSprites. \\n\\nWhile the use of inference networks proposing positions for the network as a means of improving the disentanglement is clever and seems novel, though not unlike inference sub-networks which are well-known in conditional generation, the evaluation is not up to a standard I can endorse, resulting in a recommendation to reject. \\n\\nDespite the interesting qualitative results, I will have to quote the work in saying, \\u201cAll methods cited here are likelihood based so they can and should be compared in terms of test log likelihood. We leave this for future work.\\u201d. Indeed the cited works should have been evaluated against, especially Greff et al. 2019, Nash et al, 2017, and Eslami et al, 2016, which are all very similar. As written it\\u2019s impossible to tell whether this work actually improves over the state of the art, we only have the constructed baseline (which as a community we all know clearly would not have worked). \\n\\nA figure showing the relevant submodules of the network architecture and what they do in relation to the overall method would be helpful to understand the pipeline and how the inference network relates to the whole.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper presents a probabilistic generative model for identifying the location, type and colour of images placed within a grid. The paper is in general well written and presents a large number of visual images demonstrating that the approach works. The main concerns with the paper are as follows:\\n\\n1) The implementation details for the work are relegated to an appendix. As this is a core concept for the work one would expect this to be presented in the main body of the work.\\n\\n2) Although there are a large number of visual images there is little in the way of analytical analysis of the work. As a particular concern the authors claim that figure 3 \\u2018prove\\u2019 that objects are disentangled. From Maths 101 I remember it being drummed into us that \\u2018proof by example is not a proof\\u2019.\\n\\n3) The are parts of the process which are not explained for example, why is the following conducted? \\u2018The digits are first rescaled from their original size (28 \\u00d7 28) to 15 \\u00d7 15 by bilinear interpolation, and finally binarized by rounding.\\u2019\\n\\n4) I would also like to know why this is called disentangling as the whole process prevents the \\u2018icons\\u2019 from overlapping. Disentangle normally refers to things which are at least overlapping. There are examples of works in this area.\\n\\n5) The example cases are simple. One would expect at least one real-world example.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper introduces a compositional generative model of images, where the image is described by a variable number of latent variables. Moreover, the latent variables are disentangled, in the sense that they represent different parts of the scene, and where appearance and location are described separately. While the generative model is very similar to AIR [1], the central claim of the paper is that the proposed inference scheme can generalize to a much higher number of objects than seen during training, which is demonstrated empirically, and with which AIR struggles. Better generalization is achieved by removing the recurrent core of AIR and replacing it with a fully-convolutional inference network, which predicts appearance vectors at every location in the feature map. The appearance vectors are then stochastically sampled without replacement according to a location distribution. Unlike AIR, this approach does not model the object scale.\\n\\nI recommend REJECTing this paper. While the improved generalization performance is a useful property, it has been achieved previously and in a very similar fashion by SPAIR [2], which follows a similar model design. SPAIR still uses an RNN, while the proposed approach does not, but this is only a small simplification and does not warrant publication at a top tier conference. There are no other contributions in this paper. Additionally, on the one hand, the experimental evaluation is insufficient: the proposed approach is compared only against a fully-convolutional VAE, while very similar models like AIR [1], SPAIR [2], SuPAIR [3] are not considered. On the other hand, the experimental section focuses on the disentanglement of representations, which is a) evident from the model construction and b) achieved in all previous models. \\n\\nThe paper is clearly written, and the presented generative model is interesting. Having said that, the problem that this paper addresses is mostly solved in [2] and [4]; also both [2] and [4] scale the general approach introduced in [1] to much higher number of objects (in the hundreds) and more general settings (real images, atari games).\\n\\n[1] Eslami et. al., \\u201cAttend, Infer, Repeat:...\\u201d, NIPS 2016.\\n[2] Crawford and Pineau, \\u201cSpatially Invariant Unsupervised Object Detection with Convolutional Neural Networks\\u201d, AAAI 2019.\\n[3] Stelzner et. al., \\u201cFaster Attend-Infer-Repeat with Tractable Probabilistic Models\\u201d, ICML 2019.\\n[4] Jiang et. al., \\u201cScalable Object-Oriented Sequential Generative Models\\u201d, arXiv 2019.\"}" ] }

[ "Sophia Sanborn", "Michael Chang", "Sergey Levine", "Thomas Griffiths" ]

[ "hierarchical reinforcement learning", "unsupervised learning", "compression" ]

{ "note_id": [ "aN2RvgJtU9", "ByggG2tniH", "rye7qxJAKS", "SJlAblT6YB", "HklhbwPTYr" ], "note_type": [ "decision", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751605, 1573850119869, 1571840138574, 1571831814097, 1571809028164 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2539/Authors" ], [ "ICLR.cc/2020/Conference/Paper2539/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2539/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2539/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes an interesting idea of identifying repeated action sequences, or behavioral motifs, in the context of hierarchical reinforcement learning, using sparsity/compression. While this is a fresh and useful idea, it appears that the paper requires more work, both in terms of presentation/clarity and in terms of stronger empirical results.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for the thoughtful feedback and suggestions\", \"comment\": \"We appreciate the thoughtful feedback on our paper from all 3 reviewers. We believe that substantial revisions to the exposition and experiments are required to properly communicate and demonstrate our approach. This includes 1) comparisons to alternative sequence compression based approaches, 2) a clearer analysis of macro actions learned with our method and others, 3) experiments that more clearly demonstrate the extent of our approach\\u2019s generality and limitations, and 4) substantial changes to the exposition. We plan to address these concerns in a future version that we will submit elsewhere. Thank you to the reviewers for your time and careful reading.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper discusses identifying motifs for aiding in the solving of cognitive tasks when using Reinforcement Learning. The idea seems quite novel, but the presentation seems to be more complicated than it needs to be for the idea. For example the introduction is quite hard to parse and when you get down to it, the ideas don\\u2019t seem that complex.\\n\\nThe discussion of the technique seems to be lacking in detail. I would be hard-pushed to reproduce the work from the material presented.\\n\\nFigure 2 is complex and lacks enough discussion in the text.\\n\\nFigure 3 is likewise complex and is not mentioned at all in the text.\\n\\nFigure 4 needs more discussion.\\n\\nThe results presented are quite minimal and don\\u2019t fully explore and evaluate the approach taken.\", \"specific_issues\": [\"Page 2: broken citation: \\u201cstate space [cite], \\u201c\", \"\\u201cLightbot: The Lightbot domain \\u2026 a positive reward of only if it successfully turns off all lights.\\u201d - this seems to be the opposite of all previous statements which talked about Turing lights on.\", \"\\u201cWe also model each Fractal Lightbot puzzle \\u2026 and a reward of 100 for successfully transferring the tower of disks.\\u201d - this sounds more like the reward for the tower.\"]}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes a method that aims at encoding trajectories (described as a sequence of actions) into a set of discrete codes with a hierarchical structure. The principle of the algorithm (as far as the article allows me to understand) is to apply multiple iterations of classical sparse coding over the trajectories. The experimental section on simple (deterministic) tasks shows that the SSC method is able to extract interesting options, which can then be used to learn faster on some close domains.\\n\\nIn terms of positioning, I find the idea of the paper interesting (i.e encoding trajectories through discrete symbols) since it uses sparse coding approaches which, as far as I know, are not classical in the RL domain. This type of approach can give us both a meaningful insight about the \\\"nature\\\" of the learned policy (as it is the case in the paper that compresses expert trajectories), and can also become a manner to constraint an RL algorithm to force it to exhibit behaviors that could seem more natural to humans. \\n\\nBut the way it is done in this article is disappointing. First of all, the article is badly written, and I am still not sure to fully understand how the algorithm exactly works. Indeed, many notations are not well defined (see at the end of the review), and it makes the algorithm 1 difficult to catch. Then, the authors consider that trajectories are represented as sequences of actions (using one-hot encoding) and do not discuss this hard choice: representing trajectories as a sequence of actions usually rely on the assumption that both the environment is deterministic, and the initial state is always the same. Is it the case in this paper? If it is, it clearly restricts the applicability of the technique. If it is not, then I don't see how it could work well... As far as I understand, all experimental environments are deterministic. So the algorithm description would clearly need to be rewritten, and the authors have to discuss the assumptions they are doing mainly: deterministic environments and also the fact that the \\\"options\\\" can only be extracted once a first policy has be learned (or by using expert traces) which limits its applicability.\\n\\nIn terms of experiments, the assumption made is that we have access to a set of 'good' trajectories (which is easy in the proposed environments, but may be difficult in the real-life). It is compared to the option-critic architecture which simultaneously learns the options and the policy and I think that the comparison is somehow unfair. Since SSC is more a \\\"sequence compression\\\" algorithm, I would prefer to compare with existing sequence compression algorithms like hierarchical recurrent neural networks for instance. The results are illustrated in very simple environments and the article would gain by using more complex ones (for instance the Atari grand challenge dataset could be used for such a study). So it is difficult to understand if the approach as it is is really interesting and efficient for general RL purposes.\", \"summary\": \"A good idea, but not well described, with strong assumptions not discussed, and with low-quality experimental results.\", \"some_other_minor_remarks\": \"The introduction is a little bit messy and does not well allow one to understand the focus on the paper, mixing some notions of neuroscience with classical reinforcement learning aspects, the connection between the two domains being not trivial.\", \"equation_2_versus_equation_3\": \"What is the difference?\\ns notation appears in 2.1 and 2.2 while it corresponds to different things. The variables are not defined and we don't know in which domain they rely on. \\nArticulation between sparse coding and MDL not clear (since sparse coding is directly a way to minimize the MDL). MDL never used after that.\\nsection 3, paragraph 3: I do not understand what is described here. The description has to be rewritten to allow the readers to understand the algorithm e.g \\\"the size of the dictionary elements is set to 2-timesteps. \\\" ? \\\"The dictionary element a which has the highest explained variance is then selected and assigned an integer code n + 1 \\\" Variance on what ? what is T_i ?\\n[cite] appears in the introduction\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper aims to propose a new way of discovering a specific type of temporal abstraction, the pattern of actions. To achieve this goal, it applies the sparse coding method to discover an efficient encoding of the action sequences generated by the agent's interaction with the environment. The filters/dictionaries discovered by the method then represent certain patterns of actions.\\n\\nIn general, I think it proposes an interesting view of the temporal abstraction. Although this kind of temporal abstraction is only valid in the certain types of environment (For example, as it only represents certain patterns of action sequence, it suffers from non-optimality in stochastic environment where no fixed action sequence would be optimal for solving the problem), the paper, especially the experiment part, clearly tells its readers in what situation can we expect it to perform well. In this sense, I think it provides some scientific insights that benefit my understanding.\\n\\nHowever, there are many places in the paper making me confused. Therefore I can not fully understand the paper and can not accept it. My concerns are:\\n\\n1. Equations 1, 2, and 3 are loosey-goosey. For example, there is no definition of x_i, is it a vector or a scalar? Similarly, there is no definition of W and s.\\n2. paragraph 3 in section 3 is extremely hard to understand. For example, the first sentence: \\\"At all stages, the size of the dictionary elements is set to 2-timesteps\\\". Where does the \\\"2-timesteps\\\" come? I have no idea what it is talking about.\\n3. in the algorithm, T_i is not defined.\\n4. in the experiment part, figure 4 only provides the termination of options/skills but doesn't provide corresponding policies/action sequence, which makes me hard to evaluate the result.\\n5. figure 7 and 8 seems to be contradicted with the paper's claim. And the author didn't give a reasonable explanation.\"}" ] }

[ "Katsuki Ohto" ]

[ "Reinforcement Learning", "Monte Carlo Tree Search", "Alpha Zero" ]

{ "note_id": [ "MX1NRIMyEa", "r1e8or7xir", "SylnaCGliB", "SklCBsGeoS", "B1g9iFA2KH", "r1xxIGWhFB", "rklDTxPsKB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751575, 1573037470049, 1573035716105, 1573034822179, 1571772834287, 1571717703840, 1571676350826 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2538/Authors" ], [ "ICLR.cc/2020/Conference/Paper2538/Authors" ], [ "ICLR.cc/2020/Conference/Paper2538/Authors" ], [ "ICLR.cc/2020/Conference/Paper2538/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2538/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2538/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper is a clear reject. The paper is very poorly written and contains zero citations. Also, the reviewers have a hard time understanding what the paper is about.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for your review\", \"comment\": \"Thank you deeply for your reviewing my paper and for your kindness.\\nI want to improve my paper writing skills and I hope to earn more money to feel free to ask for proofreading before the next chance.\\n\\nI hope to know whether there has been other publishing for this topic (meta procedure for MCTS based reinforcement learning).\\nAfter registration of this paper, I finally found the almost same idea (meta-tree) in my country, but I could not find international papers so far.\\nThough AlphaZero itself was amazing algorithm, I think there might be several possible updates for robustness or speedup.\"}", "{\"title\": \"Thank you for your review\", \"comment\": \"Thank you deeply for your reviewing my paper.\\nI am so sorry that even the proposed method itself is hard to be understood.\\nI hope I can write better one in the next time.\"}", "{\"title\": \"Thank you for your review\", \"comment\": \"Thank you deeply for your reviewing my paper.\\nI didn't know how to use .bib file, then I had no choice but to remove all references.\\nI would like to improve my paper writing skills and to do more research on this topic.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper tries to improve exploration performed by alphazero in a game of tic-tac-toe using MCTS. The paper proposes to use master game tree (which is MCTS as well) to control the generation of episodes when solving a game using MCTS.\\nThis is a clear case of less than half-baked paper. The paper does not cite any previous research (no references) and is poorly written. I believe this is sufficient ground for not recommending accept. I would suggest the authors to re-submit when the paper is complete, maybe, start with related work and references.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"Badly and unprofessionally written\\n\\nI had and still have a hard time to report what this paper is about. Generally, I can tell probably this paper study the Monte Carlo Tree Search-based methods in the field of Reinforcement Learning. \\n\\nThe paper is badly written with grammatical issues almost in all the sentences. I definitely urge the author(s) to take the writing significantly more seriously. I almost all the time do not end up rejecting a paper due to the lack of clear writing, but for this paper, unfortunately, despite spending time, I could not understand the paper.\", \"comments_beside_the_grammatical_issues\": \"I am sorry but I could not understand the abstract.\\nWhat does the author(s) mean by \\\"In this research, the\\ncoordination of episode generation was focused on.\\\" Same issue with almost all the sentences in the abstract. Unfortunately, I could not follow any of them. Moreover, this paper is full of inaccurate and unjustified statements. Furthermore, again, unfortunately, this paper does not represent even the preliminary elements of scientific writing.\\n\\n\\nI decided to provide further comments, but sadly I had to change my mind because otherwise, I had to write at least one paragraph of explanation for each paragraph of this paper. With that regard, I do not think it is an appropriate way to utilize the review process. But I would be happy to provide more abstract comments upon request.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper appears to be attempting to improve Monte Carlo tree search, though it's not clear how. The paper is very poorly written, with many incomprehensible passages. Moreover there are no references at all. It does not appear to be a serious academic endeavor, merely a set of notes. There are many mathematical symbols not defined, much of the paper consists of blank space yet it goes up to 10 pages, the algorithm boxes are very difficult to follow, the writing is so poor that I cannot decipher what the intent even was. A significant amount of work would be required to make this acceptable for publication in any venue.\"}" ] }

[ "John Wieting", "Graham Neubig", "Taylor Berg-Kirkpatrick" ]

[ "sentence embedding", "semantic similarity", "multilingual", "latent variables", "vae" ]

{ "note_id": [ "Tn0JpKszF", "rkgXEC3ojr", "HygUn6niiB", "Byxm962sjH", "BJxQNnnjjH", "r1gT3o3iiS", "rJxpiIYr5r", "SJeqGhkAYr", "SJgs3cCaFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751545, 1573797419074, 1573797293845, 1573797259005, 1573796907438, 1573796789415, 1572341413506, 1571843090230, 1571838642617 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2536/Authors" ], [ "ICLR.cc/2020/Conference/Paper2536/Authors" ], [ "ICLR.cc/2020/Conference/Paper2536/Authors" ], [ "ICLR.cc/2020/Conference/Paper2536/Authors" ], [ "ICLR.cc/2020/Conference/Paper2536/Authors" ], [ "ICLR.cc/2020/Conference/Paper2536/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2536/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2536/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper presents a model for building sentence embeddings using a generative transformer model that encoders separately semantic aspects (that are common across languages) and language-specific aspects. The authors evaluate their embeddings in a non-parametric way (i.e., on STS tasks by measuring cosine similarity) and find their method to outperform other sentence embeddings methods. The main concern that both reviewers (and myself) have about this work relates to its evaluation part. While the authors present a set of very interesting difficult evaluation and probing splits aiming at quantifying the linguistic behaviour of their model, it is unsatisfying the fact that the authors do not evaluate their model extensively in standard classification embedding benchmarks (e.g., as in GLUE). The authors comment: \\u201c[their model in producing embeddings] it isn\\u2019t as strong when using classification for final predictions. This indicates that the embeddings learned by our approach may be most useful when no downstream training is possible\\u201d. If this is true, why is it the case and isn\\u2019t it quite restrictive? I think this work is interesting with a nice analysis but the current empirical results are borderline (yes, the model is better on STS, but this is quite limited of an idea compared to using these embeddings as features in a classification tasks). As such, I do not recommend this paper for acceptance but I do hope that authors will keep improving their method and will make it work in more general problems involving classification tasks.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Thanks you for the review and comments! We have addressed them below:\\n\\n\\\"compare their model with many state-of-the-art models that could produce sentence embeddings. However, how they produce the sentence embeddings with existing models is not convincing. For example, why using the hidden states of the last four layers of BERT?\\\"\\n\\nThanks for drawing our attention to this point of confusion! For all the approaches we compare with, apart from BERT, we produce the sentence embeddings using the exact implementations described in the corresponding paper using the released code. For BERT, as far as we know, there is no set way of creating sentence embeddings for directly computing sentence similarity. But, to help address your question, we have added comparisons to Sentence-BERT [1] (a model that fine-tunes BERT on SNLI+MNLI data) to give a stronger baseline for a BERT-based approach. In results, we find that our proposed models substantially outperform this fine-tuned BERT-based approach. While this is not the main evaluation in our paper, we believe it indicates that BERT is not appropriate for this type of unsupervised sentence similarity task without substantial modification / further innovation. Finally, we want to briefly clarify: our reasoning behind concatenating the last four layers of BERT is based on the positive results in [1] and that this approach was recommended in the original BERT paper [2].\\n\\n\\\"However, I expected to see more analysis or experimental results to show why it is better than a monolingual variational sentence embedding framework.\\\"\\n\\nWe completely agree that further comparisons of this type are of general interest. Thanks for the comment! We have since added more analysis to the paper in the form of probing experiments and ablations showing the benefits of each component of our model. We also added a monolingual autoencoder and VAE to better understand the impact the parallel data has on the performance of our models (in Table 2). We hope that you find the analysis and new experiments helpful.\\n\\n[1] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bertnetworks. EMNLP, 2019.\\n[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019.\"}", "{\"title\": \"Response to Reviewer #1 (Part 2)\", \"comment\": \"\\\"I\\u2019m a bit unclear how these sentence embeddings are translated into a score that decides the degree to which sentences have the same meaning. Is it just cosine similarity of two sentence embedding vectors?\\\"\\n\\nYes it is just cosine similarity. We did experiment with some others like L2 or L1, but cosine seems to work best (and not just for our models -- for all the models we compare with on STS).\\n\\n\\\"While the purpose of these references is to generate sentences instead of building a sentence embedding, the method is related and comparison and discussion would be worthwhile.\\\"\\n\\nThanks for pointing these out! This was a point also raised by R2. We added experiments (Section 4.4) with an autoencoder and a monolingual VAE (just using the English side of the en-fr bitext) to help better understand how much performance is due to training on bitext (they are about 7+ points behind BGT on the STS datasets). We hope that this helps better motivate our approach.\"}", "{\"title\": \"Response to Reviewer #1 (Part 1)\", \"comment\": \"Thank you for the feedback! We address your comments below.\\n\\n\\\"While I\\u2019m not totally convinced this distinction between language-specific characteristics and semantics of the sentence, it makes it easier to understand what\\u2019s going on in the model.\\\"\\n\\nWe hope that the experiments with the model ablations and the newly added probing experiments make this distinction more convincing, clearer and better motivated.\\n\\n\\\"One of my question is, why not test this method in more popular benchmark such as MNLI or other classification tasks?\\\"\\n\\nThank you for this question. There are a few reasons for this (and we have added an experiment in Appendix D). The main reason for not comparing on MNLI specifically is that all of the pretrained baseline methods are created by training on SNLI + MNLI (either in full as in Infersent and Sentence-BERT, or are one of the objectives as in USE or GenSen), so they have a big advantage since they use direct supervision.\\n\\nHowever, we agree that evaluation on additional datasets and tasks is of general interest -- and further, since different tasks use sentence embeddings in different ways, how our embeddings are most useful in practice is worth further investigation. Throughout STS, cosine distance between embeddings is used to define similarity. But in some related tasks, regression models are fit on top of embeddings, or classifiers are trained on top of similarity scores, etc.., in order to make task predictions. Are our embeddings specially suited for cosine distance evaluations (which requires no further training)? Do they also work when supervised models are trained to predict similarity scores based on the embeddings? \\n\\nTo explore these questions, we have added the Quora Question Pairs dataset (QQP) to our evaluations. QQP is a paraphrase classification task that is also part of GLUE [1]. Since the test set of QQP is private, we deviated slightly from the standard evaluation protocol and split the development set into 2 halves of 20,215 examples each -- one half for model selection and the other for evaluation. We evaluated in two ways: cosine similarity (score all pairs with cosine similarity and then find the threshold that gives the best accuracy) and classification by training a logistic regression model on top of embeddings. It\\u2019s worth noting that the pretrained baseline models on this task were directly trained to produce the feature set used by the downstream classifier, while our embeddings are trained without this supervision. We describe the results in Appendix D and show that our model outperforms our baseline models, SP, EnglishTranslation, and BilingualTranslation for both evaluations, and compares favorably to the pretrained models when using cosine (only behind USE which was trained on Quora data in an unsupervised way and Sentence-BERT which uses BERT), but it isn\\u2019t as strong when using classification for final predictions. This indicates that the embeddings learned by our approach may be most useful when no downstream training is possible -- though semi-supervised objectives that consider the downstream task might aid our approach, like the baselines, if downstream training is the goal. \\n\\nA few more comments about our paper and related work. \\n 1. Our model is unique to most of the popular sentence embeddings models as our core contribution is proposing a new generative model. Most prior work uses NLI data or encoder-decoder models to predict different phenomena (i.e. a translation, the next sentence, etc.). \\n 2. Much of this related work is trained on substantially more data than our model, or, in the case of Infersent/Sentence-Bert, use word embeddings/BERT which are trained massive data. In contrast, we used just 1M samples from OpenSubtitles/Gigaword which represents a substantially smaller (and noisier) training set -- a more practical use case for most downstream applications. In principle, the use of larger / cleaner data and the incorporation of BERT\\u2019s contextualized embeddings is somewhat orthogonal and might stack with our results. \\n\\n[1] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. ICLR, 2019.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thank you for the comments! We hope we addressed your concerns below.\\n\\n\\u201cit is stated repeatedly that one of the strengths of the proposed model is that it is sensitive to word order on a sentence level, this particular aspect of the model is neither evaluated nor analysed\\u201d\\n\\nThe motivation for these claims was in our performance on the STS Hard datasets (and other data splits like the negation split). These splits were designed to mine sentences from the STS datasets where being able to model ordered word interactions is necessary for strong performance. Specifically, we filtered for examples that had either low symmetric word error rates and low similarity scores (i.e. similar structure and word choice, but low scores) or high symmetric word error rates and high similarity scores. This is described in Section 4.3, and Table 2 and Table 3 show results where our models (and their ablations) achieve much higher performance on these splits relative to prior work. \\n\\n\\u201cThe analysis of all these three aspects is superficial\\u201d\\n\\t\\nWe have added probing experiments to make this analysis more precise (see our comment to all reviewers). \\nThese experiments helped us better understand what exactly is being learned by the semantic and language-specific variables and how this changes with language choice. This is described in more detail in 5.2. Note that we also added some style transfer experiments that corroborate the probing results and show an interesting application of our model.\\n\\n\\u201cThe extent to which these gains are due to tuning as opposed to the inherent design of the model is not clear\\u201d\\n\\nWe hope that our new experiments with different ablations (see comment to all reviewers) illustrate that each component of our model design is contributing to the overall performance, and that all of them together produce the best results. Another addition to our revised version is that we also simplified our model -- there is now no pretraining, no freezing of encoders, and we only have a single hyperparameter (the annealing rate - which just linearly increases). We found our model robust to the annealing rate as well, and we are therefore confident that our results are not due to tuning.\"}", "{\"title\": \"Paper Revision\", \"comment\": \"We would like to thank all the reviewers for their thoughtful and constructive comments. We have posted a revised version of our paper to both address these comments and also to add additional experiments to further analyze and evaluate our models. Here are some of the highlights:\\n\\n1. (Section 2,3) We have simplified our model reducing it to a single semantic encoder that is shared between the languages We removed all pretraining as well (as we found that it only led to marginal gains) and now our best model trains from scratch with only a single hyperparameter: the annealing rate of the KL term. Our results are also improved over our original version.\\n2. (Section 4.4) We added ablations, EnglishTrans, BilingualTrans, BGT w/o LangVars, and BGT w/o Prior to our paper to better understand how each of our design choices cause further improvement over the EnglishTrans baseline. More specifically, BilingualTrans vs. EnglishTrans shows the effect of bilingual training, BGT w/o Prior vs BilingualTrans shows the effect of the KL term regularizing the embeddings space, and BGT w/o LangVars vs BilingualTrans shows the effect of adding language-specific encoders. When compared to BGT, BGT w/o LangVars and BGT w/o Prior show the effect of having both language-specific variables and priors over the embedding space.\\n3. (Section 4.4) We added an English-English autoencoder and VAE using our dataset as recommended by R1 and R2. These experiments show the contribution that parallel text gives and better motivates our model.\\n4. (Section 5.2) We added further analysis through probing experiments to better quantify the behavior of the semantic and language-specific encoders. These probing tasks cover semantic, syntactic properties in addition to length, word content, punctuation, and classifying correct/incorrect gender (in Romance languages). These results corroborate the analysis in the initial version of the paper, and we hope that you find them more convincing.\\n5. (Section 5.2) We also trained bilingual models in other languages (Spanish, Arabic, Turkish, and Japanese) and found consistently better results with our model over strong baselines. We also analyze the effect of language choice on information encoded by the semantic and language-specific encoders, showing that it can have a pronounced impact.\\n6. (Section 5.3) Lastly, we also added some qualitative examples of language generation from our generative model. These results give another view of the information captured by the semantic and language-specific variables. We feed different sentences in each encoder (semantic and language-specific), and then see how the output relates to the source sentences. The results indicate that content is encoded by the semantic variable, while roughly \\u201cstylistic\\u201d attributes are encoded by the language-specific variable -- drawing a connection with results from style transfer.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper presents a model that, given parallel bilingual data, separates the common semantics from the language-specific semantics on a sentence level.\\n\\nOverall the presentation is clear and the experiments show gains over the baselines. One major point of confusion however is that, while early on in the paper (introduction), it is stated repeatedly that one of the strengths of the proposed model is that it is sensitive to word order on a sentence level, this particular aspect of the model is neither evaluated nor analysed. Instead the empirical analysis focuses on sentence length, punctuation and semantics. The analysis of all these three aspects is superficial: for sentence length, it consists of computing the sentence mean and median; for punctuation, it consists of masking punctuation; and the last part just computes vectors of nouns only (and states that this is \\\"semantics\\\"). But there is no analysis per se.\\n\\nOverall, the paper presents a model and shows gains over baselines. The extent to which these gains are due to tuning as opposed to the inherent design of the model is not clear. The analysis is superficial.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper addresses the problem of constructing a sentence embedding using a generative transformer model which encodes semantic aspects and language-specific aspect separately. They use transformers to encode and decode sentence embedding, and the objective reconstructs input with a latent variables (language variables for each language and semantic language). These latent variables are sampled from multivariate Gaussian prior, and the learning uses evidence lower bound (ELBO) for variational approximation of the joint distribution of latent variables and input.\", \"the_method_is_evaluated_on_two_tasks\": \"sentence similarity task and machine translation evaluation metric tasks. Both tasks evaluates how similar are two sequences, and the metric is correlation score with score\\u2019s from human judge. The model shows promising results on the first task, but weaker results on the second task, especially when compared against pretty naively built sentence embedding from BERT model. I\\u2019m not expert in sentence embedding literature, so a bit tricky to evaluate, but baselines seem strong and experimental results on semantic textual similarity task.\\n\\nIn terms of evaluation, I appreciated how they defined harder subset of the evaluation dataset and showed a larger improvements on those portions of the dataset. The The paper also includes analysis on what is captured by their language-specific latent vector and semantic latent vector. While I\\u2019m not totally convinced this distinction between language-specific characteristics and semantics of the sentence, it makes it easier to understand what\\u2019s going on in the model. \\n\\nOne of my question is, why not test this method in more popular benchmark such as MNLI or other classification tasks? MNLI evaluates how each sentence pair relates to one another, thus would be a good benchmark for sentence embeddings as well. Having to encode all the information about a sentence into a single vector will make these sentence embedding model weaker than other models which can do cross sentence attentions and etc, but I think that\\u2019s the genuine limitation of sentence embedding research and has to be clarified as such. I recommend discussing and clarifying these points. \\n\\nI\\u2019m a bit unclear how these sentence embeddings are translated into a score that decides the degree to which sentences have the same meaning. Is it just cosine similarity of two sentence embedding vectors?\\n\\nWhile the purpose of these references is to generate sentences instead of building a sentence embedding, the method is related and comparison and discussion would be worthwhile. \\n\\nGenerating Sentences from a Continuous Space\\nSamuel R. Bowman,\\u00a0Luke Vilnis,\\u00a0Oriol Vinyals,\\u00a0Andrew M. Dai,\\u00a0Rafal Jozefowicz,\\u00a0Samy Bengio\", \"https\": [\"//arxiv.org/abs/1703.00955\", \"Comments & Questions:\", \"Methods using a large amount of unsupervised monolingual data shows very strong performance in a panoply of NLP tasks these days. If I understand correctly, this model is constrained by the amount of bitext \\u2014 some analysis on this would be interesting.\", \"Figure 1 mentions about \\u201cSection 3, 4\\u201d but I don\\u2019t think they are correct references?\", \"BERT baseline seemed not to allow fine-tuning of the LM parameters. I think this makes the baseline significantly weaker?\", \"It seems odd that only English semantic encoder is used to downstream application.\", \"Does table 3 covers all the data? What proportion of the data is covered by each row?\", \"Given the similarity of English and French, I\\u2019m not sure how \\u201clanguage-specific\\u201d such latent vectors are. It would be much more interesting analysis if it studies distant language pairs.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper presents a bilingual generative model for sentence embedding based variational probabilistic framework. By separating a common latent variable from language-specific latent variables, the model is able to capture what's in common between parallel bilingual sentences and language-specific semantics. Experimental results show that the proposed model is able to produce sentence embeddings that reach higher correlation scores with human judgments on Semantic Textual Similarity tasks than previous models such as BERT.\", \"strength\": \"1) the idea of separating common semantics and language-specific semantics in the latent space is pretty neat; 2) the writing is very clear and easy to follow; 3) the authors explore four approaches to use the latent vectors and four approaches to merge semantic vectors, makes the final choices reasonable.\", \"weakness\": \"1) Experiments: \\n\\nMy major concern is the fairness of the experiments. The authors compare their model with many state-of-the-art models that could produce sentence embeddings. However, how they produce the sentence embeddings with existing models is not convincing. For example, why using the hidden states of the last four layers of BERT? Moreover, the proposed model is trained with parallel bilingual data, while the BERT model in comparison is monolingual. Also, the proposed deep variational model is close to an auto-encoder framework. You can also train a bilingual encoder-decoder transformer model (perhaps with pre-trained BERT parameters) with auto-encoder objective using the same parallel data set. It seems to be a more comparable model to me. \\n\\nAlthough the proposed model is based on variational framework, there's no comparison with previous neural variational models that learn encodings of texts as well such as https://arxiv.org/abs/1511.06038. \\n\\n2) Ablation study and analysis\\n\\nI really like the idea of separating common semantic latent variables with language-specific latent variables. However, I expected to see more analysis or experimental results to show why it is better than a monolingual variational sentence embedding framework.\"}" ] }

[ "Youngwoon Lee", "Jingyun Yang", "Joseph J. Lim" ]

[ "reinforcement learning", "hierarchical reinforcement learning", "modular framework", "skill coordination", "bimanual manipulation" ]

{ "note_id": [ "bqtUAvmNo8", "HygPavW2iB", "Byenowbnjr", "SJggRL-3iS", "SJljR-WnjS", "S1lSlx-noB", "r1eDu0Dk5S", "H1eyGV_aYB", "HJxR4YcnYS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751514, 1573816255491, 1573816228232, 1573816008406, 1573814739202, 1573814252663, 1571941999203, 1571812359385, 1571756341646 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2535/Authors" ], [ "ICLR.cc/2020/Conference/Paper2535/Authors" ], [ "ICLR.cc/2020/Conference/Paper2535/Authors" ], [ "ICLR.cc/2020/Conference/Paper2535/Authors" ], [ "ICLR.cc/2020/Conference/Paper2535/Authors" ], [ "ICLR.cc/2020/Conference/Paper2535/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2535/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2535/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper deals with multi-agent hierarchical reinforcement learning. A discrete set of pre-specified low-level skills are modulated by a conditioning vector and trained in a fashion reminiscent of Diversity Is All You Need, and then combined via a meta-policy which coordinates multiple agents in pursuit of a goal. The idea is that fine control over primitive skills is beneficial for achieving coordinated high-level behaviour.\\n\\nThe paper improved considerably in its completeness and in the addition of baselines, notably DIAYN without discrete, mutually exclusive skills. Reviewers agreed that the problem is interesting and the method, despite involving a degree of hand-crafting, showed promise for informing future directions. \\n\\nOn the basis that this work addresses an interesting problem setting with a compelling set of experiments, I recommend acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"References\", \"comment\": \"[1] Eysenbach et al. \\u201cDiversity is All You Need: Learning Skills without a Reward Function\\u201d, ICLR 2019\\n[2] Frans et al. \\u201cMeta Learning Shared Hierarchies\\u201d, ICLR 2018\\n[3] Co-Reyes et al. \\u201cSelf-consistent trajectory autoencoder: Hierarchical reinforcement learning with trajectory embeddings\\u201d, ICML 2018\\n[4] Nachum et al. \\u201cData-efficient hierarchical reinforcement learning.\\u201d NeurlPS 2018\\n[5] Lee et al. \\u201cComposing Complex Skills by Learning Transition Policies\\u201d, ICLR 2019\\n[6] Bacon et al. \\u201cThe Option-Critic Architecture\\u201d, AAAI 2017\\n[7] Andreas et al. \\u201cModular Multitask Reinforcement Learning with Policy Sketches\\u201d, ICML 2017\\n[8] Oh et al. \\u201cZero-shot Task Generalization with Multi-Task Deep Reinforcement Learning\\u201d, ICML 2017\\n[9] Levy et al. \\u201cHierarchical Reinforcement Learning with Hindsight\\u201d, ICLR 2019\\n[10] Lee et al. \\u201cTo Follow or not to Follow: Selective Imitation Learning from Observations\\u201d, CoRL 2019\\n[11] Sukhbaatar et al. \\u201cLearning Multiagent Communication with Backpropagation\\u201d, NIPS 2016\\n[12] Peng et al. \\u201cMultiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games\\u201d, arXiv 2017\\n[13] Jiang et al. \\u201cLearning Attentional Communication for Multi-Agent Cooperation\\u201d, NIPS 2018\\n[14] Tan et al. \\u201cMulti-agent reinforcement learning: Independent vs. cooperative agents,\\u201d ICML 1993\\n[15] Sunehag et al. \\u201cValue-Decomposition Networks For Cooperative Multi-Agent Learning\\u201d, AAMAS 2018\\n[16] Rashid et al. \\u201cQMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning\\u201d, ICML 2018\\n[17] Foerster et al. \\u201cCounterfactual Multi-Agent Policy Gradients\\u201d, AAAI 2018\\n[18] Tang et al. \\u201cHierarchical Deep Multiagent Reinforcement Learning with Temporal Abstraction\\u201d, arXiv 2018\\n[19] Han et al. \\u201cMulti-Agent Hierarchical Reinforcement Learning with Dynamic Termination\\u201d, AAMAS Extended Abstract 2019\\n[20] Ahilan et al. \\u201cFeudal Multi-Agent Hierarchies for Cooperative Reinforcement Learning\\u201d, arXiv 2019\"}", "{\"title\": \"Response to Reviewers\", \"comment\": \"We thank all the reviewers for providing constructive feedback and have updated our paper accordingly. We first summarize concerns from reviewers here and then respond to individual reviews.\\n\\n\\n- DIAYN only baseline without a priori subtasks to help evaluate the importance of expert knowledge (predefined subtasks) to the final performance. [R2]\\n\\nThe tasks in this paper deal with coordination between complex skills of multiple agents. As shown in Figure 4 and Table 1, we found that the baseline suggested by R2 fails to solve these tasks because DIAYN skills [1] acquired without a priori subtasks do not capture complicated interactions with the environment required for the final tasks. This result shows that prior knowledge about subtasks takes a critical role in tackling complex tasks.\\n\\nTo show this concretely, we have added two baselines: centralized policy and decentralized policy with Skill Behavior Diversification (centralized SBD and decentralized SBD). The centralized SBD baseline has a single low-level policy trained with DIAYN while the decentralized SBD baseline has multiple low-level policies trained with DIAYN. Both baselines have a meta policy, which controls low-level policies by generating behavior embeddings, and do not have an external skill-specific reward. Both baselines fail to learn useful skills for the final tasks, while our method can successfully solve the tasks with pre-trained primitive skills. This shows that domain expert knowledge becomes more critical as tasks become more complicated. \\n\\n\\n- Can this method work with other skill discovery methods? [R2, R3]\\n\\nDiverse skills discovered by any skill discovery method can be used by our method as long as acquired skills are controllable and diverse enough to complete the final task. Please note that our method is orthogonal to skill discovery methods. \\n\\n\\n- Design choice about the fixed horizon of the low-level skills $T_{low}$. [R1, R3]\\n\\nA fixed horizon $T_{low}$ has been actively used in HRL [2-5] due to its simplicity and efficient implementation. Instead of using a fixed $T_{low}$, skill termination function [6-8] can be learned to support variable lengths of primitive skills and flexible skill selection. There are also works [9,10] that vary the horizon for the low-level policy with the goal specified by the meta policy. However, joint learning of which skill to select and when to switch skills becomes difficult for the meta policy due to the credit assignment problem.\\n\\n\\n- The meta-controller is treating skills as primitive actions rather than temporally extended behavior with $T_{low}=1$. [R1]\\n\\nIn this paper, we focus on behavioral coordination of skills rather than exploiting temporal abstraction of skills, and we found that adjusting the behavior of skills frequently is critical to the performance of the final tasks.\\n\\nWe let $T_{low}$ as a tunable hyperparameter according to tasks. If primitive skills have a long horizon and not sensitive to temporal alignment, larger $T_{low}$ can provide faster learning by utilizing temporal abstraction. On the other hand, if the task requires sharp changes over primitive skills, small $T_{low}$ is more suitable. Additional experiments (Figure 7) show that the agent can realize more flexible skill coordination with small $T_{low}$ by adjusting the behavior embedding more frequently for our tasks, while switches skills only when required.\\n\\n\\n- More discussion regarding related multi-agent methods. [R1, R3]\\n\\nTo the best of our knowledge, there is no prior work about tackling the coordination problem between multiple agents with skills learned in isolation. Existing multi-agent reinforcement learning methods either learn communication mechanisms [11-13] or attempt to resolve the credit assignment problem [14-17] between agents. However, communication mechanisms cannot be utilized when primitive skills are pre-trained separately without communication. Also, credit assignment problems become more challenging when the complexity of cooperative tasks increases and all agents need to learn completely from scratch. There are a few works [18-20] utilizing hierarchies for multi-agent RL. However, the applications of these methods are limited to simple 2d tasks since all skills need to be learned from scratch.\"}", "{\"title\": \"Response to Reviewer 3\", \"comment\": \"We thank the reviewer for the constructive feedback and address the concerns in detail below.\\n\\n\\n- Should we treat temporal abstraction under the multi-agent setting different from it under the single-agent setting?\\n\\nWe treat temporal abstraction (TA) under the multi-agent system the same as TA under the single-agent setting. However, utilizing temporally-extended actions (primitive skills) in the multi-agent system introduces the coordination problem between agents where skills from different agents disturb each other or are not able to complete collaborative tasks. Please note that this paper focuses on tackling this problem by learning to coordinate skills. \\n\\n\\n- Why should we consider the DIAYN method, instead of other skills discovery methods?\\n\\nDiverse skills discovered by any skill discovery methods can be used by our method as long as acquired skills are controllable and diverse enough to complete the final task. In this paper, we choose DIAYN for discovering diverse behaviors and corresponding latent representations given a reward function.\\n\\n\\n- \\u201cHowever, all these approaches are focused on working with a single end-effector or agent with learned primitive skills, and learning to coordinate has not been addressed.\\u201d Does the author mean there is no temporal abstraction method for multiple collaborative agents? \\n\\nThe sentence R3 mentioned refers to the modular approaches which suggest to learn reusable skills and combine them to solve more complex tasks. As far as we know, there is no prior work about tackling the coordination problem between multiple agents with skills learned in isolation.\\n\\nThere are a few works [2-4] utilizing temporal abstraction for multi-agent RL. However, the applications of these methods are limited to simple 2d tasks since all skills need to be learned from scratch. Instead, our method utilizes primitive skills to solve complex tasks and tackles the coordination problem which happens when multiple agents perform their skills without considering each other.\\n\\n\\n- Analyze the design choice \\u201ca skill, once being chosen, will be executed for $T_{low}$ steps, where $T_{low}$ is fixed and pre-defined by the algorithm designer\\u201d.\\n\\nA fixed horizon $T_{low}$ has been actively used in HRL [11-14] due to its simplicity and efficient implementation. Instead of using a fixed $T_{low}$, skill termination function [6-8] can be learned to support variable lengths of primitive skills and flexible skill selection. There are also works [9,10] that vary the horizon for the low-level policy with the goal specified by the meta policy. However, joint learning of which skill to select and when to switch skills becomes difficult for the meta policy due to the credit assignment problem.\\n\\nWe let $T_{low}$ as a tunable hyperparameter according to tasks. If primitive skills have a long horizon and not sensitive to temporal alignment, larger $T_{low}$ can provide faster learning by utilizing temporal abstraction. On the other hand, if the tasks require sharp changes over primitive skills, small $T_{low}$ is more suitable. Additional experiments (Figure 7) show that the agent can realize more flexible skill coordination with small $T_{low}$ by adjusting the behavior embedding more frequently for our tasks, while switches skills only when required.\\n\\n\\n- In section 3.3, how does the method learn $q(z|s)$ to approximate $p(z|s)$?\\n\\nWe learn $q(z|s)$ to approximate $p(z|s)$ by maximizing the reward (Equation 4) that includes the variational lower bound [1,5]. Maximizing the variational lower bound minimizes KL-divergence between $q(z|s)$ and $p(z|s)$, which means $q(z|s)$ approximates $p(z|s)$. Please refer to [1] for more details.\\n\\n\\n\\n[1] Eysenbach et al. \\u201cDiversity is All You Need: Learning Skills without a Reward Function\\u201d, ICLR 2019\\n[2] Tang et al. \\u201cHierarchical Deep Multiagent Reinforcement Learning with Temporal Abstraction\\u201d, arXiv 2018\\n[3] Han et al. \\u201cMulti-Agent Hierarchical Reinforcement Learning with Dynamic Termination\\u201d, AAMAS Extended Abstract 2019\\n[4] Ahilan et al. \\u201cFeudal Multi-Agent Hierarchies for Cooperative Reinforcement Learning\\u201d, arXiv 2019\\n[5] Kingma et al. \\u201cAuto-Encoding Variational Bayes\\u201d, ICLR 2014\\n[6] Bacon et al. \\u201cThe Option-Critic Architecture\\u201d, AAAI 2017\\n[7] Andreas et al. \\u201cModular Multitask Reinforcement Learning with Policy Sketches\\u201d, ICML 2017\\n[8] Oh et al. \\u201cZero-shot Task Generalization with Multi-Task Deep Reinforcement Learning\\u201d, ICML 2017\\n[9] Levy et al. \\u201cHierarchical Reinforcement Learning with Hindsight\\u201d, ICLR 2019\\n[10] Lee et al. \\u201cTo Follow or not to Follow: Selective Imitation Learning from Observations\\u201d, CoRL 2019\\n[11] Frans et al. \\u201cMeta Learning Shared Hierarchies\\u201d, ICLR 2018\\n[12] Co-Reyes et al. \\u201cSelf-consistent trajectory autoencoder: Hierarchical reinforcement learning with trajectory embeddings\\u201d, ICML 2018\\n[13] Nachum et al. \\u201cData-efficient hierarchical reinforcement learning.\\u201d NeurlPS 2018\\n[14] Lee et al. \\u201cComposing Complex Skills by Learning Transition Policies\\u201d, ICLR 2019\"}", "{\"title\": \"Response to Reviewer 2\", \"comment\": \"We thank the reviewer for the constructive feedback and address the concerns in detail below.\\n\\n\\n- Subtasks must be specified in advance. \\u2026 Did the authors consider using the DIAYN objective on its own to encourage sufficiently diverse behaviors? If this didn't work, would perhaps a larger latent skill vector, or a large discrete set of DIAYN skills, have made it work?\\n\\nThank you for suggesting an interesting idea! Solving complex tasks using DIAYN skills [1] conditioned on a large latent skill vector without a priori knowledge is an interesting direction. If DIAYN skills without a priori subtasks have sufficiently diverse behaviors, our method can coordinate those skills by selecting the latent skill vector. However, we feel that it is out of the scope of this project.\\n\\nAs Reviewer 2 (R2) noted in \\u201clearning individual skills in isolation is generally more tractable than learning their combined application from scratch\\u201d, the focus of this paper is to independently learn flexible skills that can be composed with other skills of other end-effectors/agents, and to coordinate those skills to solve more complicated and cooperative tasks. \\n\\nPlease note that our method is complementary to the suggestion R2 pointed out. As R2 suggested, unsupervisedly acquiring skills for more complicated tasks is an important problem yet orthogonal and complementary to our method. We will leave this problem as a future work of the community.\\n\\n\\n- The paper would be improved if it included a strong baseline that uses DIAYN only (no a priori subtasks), so we can evaluate how important that expert knowledge is to the final performance.\\n\\nThe tasks discussed in this paper deal with coordination between complex primitive skills of several agents or agent body parts. We found that the baseline R2 suggested cannot solve these tasks mainly because DIAYN skills [1] acquired without a priori subtasks do not capture complicated interactions with the environment required for the final tasks. This result shows that prior knowledge about subtasks takes a critical role in tackling complex and cooperative tasks.\\n\\nTo show this concretely, we have added two baselines (Section 4.1) \\u2014 centralized policy and decentralized policy with Skill Behavior Diversification (centralized SBD and decentralized SBD), as suggested by R2. The centralized SBD baseline has a single low-level policy trained with DIAYN while the decentralized SBD baseline has multiple low-level policies trained with DIAYN. Both baselines have a meta policy, which controls low-level policies by generating behavior embeddings, and do not have an external skill-specific reward.\\n\\nThese baselines have a hard time learning useful skills for skill coordination towards the final composite task, which indicates domain expert knowledge becomes more critical as tasks become more complicated. On the other hand, our proposed method can successfully solve the final tasks with pre-trained primitive skills. We have added these results in Figure 4 and Table 1. \\n\\n\\n- The size of latent skill embedding is not reported. \\n\\nWe thank R2 for pointing out missing information about the size of latent skill embedding. We used 5 for the size of latent behavior embedding. We have added this information in Section 3.5 and supplementary material (Appendix C.1).\\n\\n\\n[1] Eysenbach et al. \\u201cDiversity is All You Need: Learning Skills without a Reward Function\\u201d, ICLR 2019\"}", "{\"title\": \"Response to Reviewer 1\", \"comment\": \"We thank the reviewer for the constructive feedback and address the concerns in detail below.\\n\\n\\n- The meta-controller is treating skills as primitive actions rather than temporally extended behavior with $T_{low}=1$.\\n\\nAs Reviewer 1 (R1) mentioned, with $T_{low}=1$, the meta policy can treat primitive skills as primitive actions rather than temporally extended behaviors. In this paper, we focus on behavioral coordination of skills rather than exploiting temporal abstraction of skills, and we found that adjusting the behavior of skills frequently is critical to the performance of the final tasks.\\n\\nTo investigate the effect of skill selection interval $T_{low}$, we conducted an additional experiment on the Jaco environments with different $T_{low}$ values (1, 2, 3, 5, 10) and added the results in Figure 7 (Appendix A.2). The results show that smaller $T_{low}$ values (1, 2, 3) perform better than larger $T_{low}$ values (5, 10). This can be because the agent can realize more flexible skill coordination by adjusting the behavior embedding more frequently.\\n\\nTo verify our reasoning above, we came up with a variation of our method in which the meta policy can choose a primitive skill to execute every time step but choose a behavior embedding only when a primitive skill changes. In this setting, the meta policy at times switches back and forth between two skills in consecutive time steps to change the behavior embedding. This results in worse performance compared to our method with a small $T_{low}$ which switches skills only when required, but adapts behavior embeddings frequently. This indicates that the meta policy needs to frequently adjust the behavior embedding in order to optimally coordinate the skills of the different agents.\\n\\n\\n- More in-depth discussion regarding alternative multi-agent methods.\\n\\nTo the best of our knowledge, there is no prior work about tackling the coordination problem between multiple agents with skills learned in isolation. Existing multi-agent reinforcement learning methods either learn communication mechanisms [1-3] or attempt to resolve the credit assignment problem during training [4-7] between agents. However, communication mechanisms cannot be utilized when primitive skills are pre-trained separately without communication. Also, credit assignment problems become more challenging to resolve when the complexity of cooperative tasks increases and all agents need to learn completely from scratch. \\n\\nThere are a few works [8-10] utilizing multiple layers of hierarchies for multi-agent RL. However, the applications of these methods are limited to tasks in simple 2d environments given that all skills need to be learned from scratch. Instead, we propose to utilize primitive skills to solve complex tasks and tackle the coordination problem which happens when multiple agents perform their skills without considering other agents.\\n\\n\\n- High variance in performance of the method.\\n\\nAs R1 pointed out, Figure 4 shows a high variance since the curves are plotted without moving average. We updated all curves (Figure 4-7) with moving average of 10 evaluations. Moreover, we ran experiments with more seeds, so all Jaco experiments are reported using 6 different random seeds.\\n\\n\\n- Analysis plot in Section 4.5 shows episode rewards but not success rates.\\n\\nWe choose the episode reward over the success rate since most experiments fail to succeed on the task, which makes the differences between experiments less noticeable. As suggested by R1, we have also added a plot of success rates in Figure 6 in the supplementary material (Appendix A.1).\\n\\n\\n- Notations at times inconsistent and confusing.\\n\\nWe thank for pointing out inconsistent notations and suggesting an idea of displaying notations in Figure 2. We revised the paper for consistent notations and Figure 2 to provide a better understanding of our framework. \\n\\n\\n[1] Sukhbaatar et al. \\u201cLearning Multiagent Communication with Backpropagation\\u201d, NIPS 2016\\n[2] Peng et al. \\u201cMultiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games\\u201d, arXiv 2017\\n[3] Jiang et al. \\u201cLearning Attentional Communication for Multi-Agent Cooperation\\u201d, NIPS 2018\\n[4] Tan et al. \\u201cMulti-agent reinforcement learning: Independent vs. cooperative agents,\\u201d ICML 1993\\n[5] Sunehag et al. \\u201cValue-Decomposition Networks For Cooperative Multi-Agent Learning\\u201d, AAMAS 2018\\n[6] Rashid et al. \\u201cQMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning\\u201d, ICML 2018\\n[7] Foerster et al. \\u201cCounterfactual Multi-Agent Policy Gradients\\u201d, AAAI 2018\\n[8] Tang et al. \\u201cHierarchical Deep Multiagent Reinforcement Learning with Temporal Abstraction\\u201d, arXiv 2018\\n[9] Han et al. \\u201cMulti-Agent Hierarchical Reinforcement Learning with Dynamic Termination\\u201d, AAMAS Extended Abstract 2019\\n[10] Ahilan et al. \\u201cFeudal Multi-Agent Hierarchies for Cooperative Reinforcement Learning\\u201d, arXiv 2019\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper presents a hierarchical reinforcement learning method for coordination of multiple cooperative agents with pre-learned adaptable skills. These skills are learned via a maximum entropy objective where diversity of behaviour given each skill is maximised and controlled via a latent conditioning vector. This allows controllability of the variation in skill execution by the meta-policy via changing the skill-specific latent vectors. The paper presents empirical results in manipulation (pick-push-place and moving a long bar by coordinating two Jaco arms) and locomotion (two Ants pushing a large block to a goal location). The method proposed outperforms the baselines reported.\\n\\nOverall, this paper addresses an interesting problem and can be impactful with the caveat for some clarifications and analysis. Given that the authors address my concerns, I would be willing to increase my score.\\n\\nThe main novelty of this work lies in how one can learn sub-skills that can be leveraged and adapted for down-stream tasks. The problem setting used to test the method is a multi-agent setting where it is crucial that skills are adapted to enable cooperation. I found the environments and the problem setting generally interesting and important for testing the proposed method. I have a few concerns that I listed below:\\n\\n\\n1) I found the notations at times inconsistent and confusing. It would have helped to see some more details on the diagram (Figure 2) to understand how everything fits together. \\n\\n2) The set of skills for the two agents are selected by the meta-policy in every T_low steps. It looks like in the Jaco environments T_low = 1. Can you comment on this? This seems slightly concerning since it seems like the meta-controller is treating these skills as primitive actions rather than temporally extended behaviour.\\n\\n3) Looking at the training curves in Figure 4, there seems to be a really high variance in performance of the method. Can you comment on this as this seems concerning. Could you run more seeds to improve this?\\n\\n4) It is nice to see in section 4.5 how the hyper-parameters balancing diversity in combination with external reward (equation 4) is tuned and how sensitive that is to achieving adaptability for downstream tasks. The only criticism I have is that it is difficult to understand from \\\"Episode reward\\\" on y-axis what the success rate is (similar to Figure 4)? It would\\u2019ve been nice to report results in a consistent way throughout the paper for these environments. \\n\\n5) Given that all the tasks in the experiments are cooperative multi-agent settings, I would have liked to see more in depth discussion regarding alternative multi-agent methods. The multi-agent baseline provided (which is using a decentralized policy with a shared critic, inspired by Lowe et al., 2017) seems fair, but I wonder if there has been more recent work in this direction that could have been highlighted?\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper provides a specific way of incorporating temporal abstraction into the multi-agent reinforcement learning (MARL) setting. Specifically, this method first discovers diversified skills for every single agent and then train a meta-policy to choose among skills for all agents.\\n\\nOverall, this paper is well presented so I can understand it well. Unfortunately, this paper didn't give me too much scientific insight. As maybe this is because I don't know too much about MARL, I would like to ask the author to help me address the following questions. \\n\\nMy first key question is, should we treat temporal abstraction (TA) under the multi-agent setting different from it under the single-agent setting? If they are the same, why do we bother discussing TA under the multi-agent setting? Why not just discuss it under the simpler single-agent setting? If they are not, what are the differences? \\n\\nThe second key question is if TA under the multi-agent system is special, then why the DIAYN method, which is proposed under the single-agent setting, could be directly used in the multi-agent setting? Why should we consider the DIAYN method, instead of other skills discovery methods? \\n\\nFurthermore, I would also like the author to help me address three more concrete questions. \\n\\n1. Section 1, paragraph 2, the author wrote: \\\"However, all these approaches are focused on working with a single end-effector or agent with learned primitive skills, and learning to coordinate has not been addressed.\\\" Does the author mean there is no temporal abstraction method for multiple collaborative agents? \\n\\n2. Section 3.3, the author wrote, \\\" the prior distribution p(z) is Gaussian.\\\" I.e., Z is continuous r.v. I would like to know how the author could learn q(z|s) to approximate p(z|s), which is an arbitrary continuous distribution. Maybe I am wrong, but I don't see a way to do this. \\n\\n3. In algorithm 1, a skill, once being chosen, will be executed for T_{low} steps, where T_{low} is fixed and pre-defined by the algorithm designer. I would like to hear to author analyzing the pros and cons of this critical design choice.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper aims to achieve multi-agent coordination by composing diverse skills learned by augmenting individual subtask objectives with DIAYN-style diversity bonuses. Once individual diverse skills are learned for the subtasks, the agents are combined by a meta-agent to coordinate multiple distinct robots to achieve a shared goal.\\n\\nThis is a good application of low-level skill learning to multi-agent coordination. I have settled on a weak acceptance, because the approach is simple and seems scalable, but the acceptance is weak because the method relies on specifying the subtasks in advance.\\n\\nThe approach is well-motivated in that learning individual skills in isolation is generally more tractable than learning their combined application from scratch, and the building blocks of this system are well-chosen. The results demonstrate the importance of the diversity objective, and find a good sweet spot for the diversity weight.\\n\\nI do have some criticisms, related primarily to the decision to pre-train with both a continuously-parameterized diversity conditioning as well as a discrete set of concrete subtasks. Because these subtasks must be specified in advance, this limits the wide applicability of the resulting approach to those that can be broken down a-priori into components. Did the authors consider using the DIAYN objective on its own to encourage sufficiently diverse behaviors? If this didn't work, would perhaps a larger latent skill vector, or a large discrete set of DIAYN skills, have made it work?\\n\\nI also don't see the size of the latent skill embedding reported anywhere. How big is this vector; that information should be added to the paper.\\n\\nHowever, the approach is generally good. I think the paper would be improved if it included a strong baseline that uses DIAYN only (no a-priori subtasks), so we can evaluate how important that expert knowledge is to the final performance.\"}" ] }

[ "Paul Kim", "Robin Winter", "Djork-Arné Clevert" ]

[ "Unsupervised Representation Learning", "Computational biology", "computational chemistry", "protein-ligand binding" ]

{ "note_id": [ "QGVQ7oSjL", "BkxnneP9cS", "H1lbzsATKB", "HJxiMPIitS", "Bylv0OAStB", "SJg0owRHYS", "r1xmGnv7KH" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_comment", "official_comment", "comment" ], "note_created": [ 1576798751483, 1572659379756, 1571838728572, 1571673874932, 1571313870708, 1571313574105, 1571154954733 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2534/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2534/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2534/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2534/Authors" ], [ "ICLR.cc/2020/Conference/Paper2534/Authors" ], [ "~Christian_Dallago1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper uses unsupervised learning to create useful representations to improve the performance of models in predicting protein-ligand binding. After reviewers had time to consider each other's comments, there was consensus that the current work is too lacking in novelty on the modeling side to warrant publication in ICLR. Additionally, current experiments are lacking comparisons with important baselines. The work in its current form may be better suited for a domain journal.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"The authors present a model with state-of-the-art performance for predicting protein-ligand affinity and provide a thorough set of benchmarks to illustrate the superiority of combining learned low-dimensional embedding representations of both ligands and proteins. The authors then show that these learned representations are more powerful than handcrafted features such as circular fingerprints, etc. when combined into a model that jointly takes as input both the ligand and protein.\\n\\nI originally suggested to accept the paper, but agree with the other reviewers that the novelty of the work in this paper likely doesn't meet the bar for acceptance given that the most significant contributions of this paper are around combining good ideas from other papers without much additional novelty.\\n\\nMajor comments\\n\\nUnless I'm misunderstanding the naming conventions used in Table 1, it seems like an omission not to include the performance of DeepPCM+HP+HC in Table 1 to more fully decouple the performance due to the DeepPCM architecture versus the contribution due to using both unsupervised descriptors in combination.\\n\\nI might expect the performance of the DeepPCM + HP + HC model (not currently shown unless I'm missing it) to exceed the performance of the NIB + HP + HC model based upon the reasoning given in the discussion (i.e., usefulness of joint input training), even if the individual representations were inferior to the unsupervised counterparts.\\n\\nSince the primary point of the paper is to illustrate the power of combining the unsupervised representations, I'm surprised the aforementioned performance comparison is not prominent. That is, the performance comparison of DeepPCM+HP+HC versus DeepPCM+UP+UC seems like a very central quantity to present and discuss, but appears to be missing currently.\\n\\n\\nMinor comments\\n\\nTable 1 was a bit confusing to me at first, because it appears that UD = UP + UC, and HD = HP + HC, but this wasn't obvious to me initially; I'd change the NIB rows to be NIB + HP + HC and NIB + UP + UC, and then you don't even need to define the UD and HD acronyms, which simplifies the table's cognitive load for me and also makes the relationship between the DeepPCM variants more obvious.\\n\\n> \\\"On the low-coverage-split, we also find that our method significantly outperforms the benchmark.\\\"\\n\\nIt would be good to add the %improvement inline here to parallel the other dataset splits listed just before so that you can directly compare the magnitudes.\\n\\n> \\\"using unsupervised-learned descriptors than when using handcrafted descriptors\\\"\\n\\nWould be more compelling if you added the DeepPCM + HC + HP performance to Table 1.\\n\\n> \\\"All hyperparameter optimization of our model was performed on the temporal split\\\"\\n\\nI think it would improve the paper if the extent to which various architecture choices were optimized over could be included as well; for example, which types, layer sizes, and depths of network architectures were considered in the hyperparam tuning, and any accompanying justification for these design choices.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper tries to solve the protein-legend binding prediction problem in the computational biology field. It uses the learned embedding for protein and legend, separately, from two published papers. Then those two embeddings were inputted to another deep learning model, performing the final prediction. Tested on one dataset, it shows the proposed method can outperform the other baseline methods.\", \"the_paper_should_be_rejected_for_the_following_reasons\": \"1. The idea of the paper is not interesting and novel enough. It only used the results from two published papers and then applied another deep learning model on them. The novelty of the paper is limited.\\n2. The experiment part is not that comprehensive. It indeed performs enough ablation studies. However, all the legendary methods in the computational biology field are not included in the comparison. \\n3. The experiments are only performed on one dataset. Usually, for application papers, the experiments should be performed against at least 2 datasets to avoid bias.\\n4. The discussion part is not well-developed. For the paper which focuses on only one task, the more in-depth discussion is expected beyond the simple discussion of the performance. For example, the authors may try to explain the result: why the model used embedding from unsupervised learning is better than the hand-crafted features. Since they shared the same model, the unsupervised embedding should contain more information. Then, what is the additional information?\", \"some_further_questions_and_comments\": \"1. What's the sequence similarity of the 1226 proteins?\\n2. Can the model generalize well to a completely new protein? \\n3. What's the detailed performance of the model on proteins belonging to different families? \\n4. I guess if the authors check the detailed performance, they will find nonuniform performance across different proteins. I think the authors can also further investigate that.\\n5. Can the authors train the embedding model as well as the classification model in an end-to-end fashion? This can be more interesting.\\n6. One big problem of the assay data is that it would not be able to provide the structure information of the interaction between the protein and the legend. Usually, it is a very important piece of information for the biology people. The computational methods based on the assay data will also inherit the flaw.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes to learn representations of protein and molecules for the prediction of protein-ligand binding prediction.\\n\\nThe presentation of this paper is a bit lengthy and repetitive in some cases. The long descriptions of protein/drug descriptors are a nice overview, but it may be unnecessary as the authors in the end use other works\\u2019 embedding.\\n\\nThe author points out that there are interpretability issues & the inability to capture the shape of the substructures with previous ligand descriptors, however, it seems that CDDD also is not interpretable and could not capture the shape as it operates on SMILES strings, although seems to have better predictive performance.\\n\\nFor the protein descriptor, the author is missing several important descriptors that may not have the issues mentioned such as Protein Sequence Composition descriptor. \\n\\nThe technical novelty is very limited. It seems the usage of CDDD and UniRep are its only difference from previous works such as DeepDTA, WideDTA, DeepConv-DTI, PADME and CDDD and UniRep are also from other works. It may be more suitable for a domain journal instead of ICLR which focuses on method innovation. \\n\\n\\nThe experimental setup is solid with realistic considerations. However, it is missing many baselines such as DeepDTA, WideDTA, DeepConv-DTI, PADME and more classic methods such as SimBoost and KronRLS.\"}", "{\"comment\": \"- Writing: Table 1 should appear in the results instead of the methods, and the discussion should not introduce new methods (e.g. evaluation of the individual impact of each descriptor type, in the second paragraph of discussion)\\n\\nWe were not sure where to put the evaluation of the individual impact of each descriptor type, since it is a method but is more for investigative purposes rather than a key point of our results. It seemed redundant to explain it in both the methods and then in the discussion. We will make changes in the revision to improve the flow while ensuring the discussion does not introduce new methods. \\n\\n\\n* UniRep was used for the embedding of proteins, but another method is available (SeqVec). We suggest also trying this encoder and see if there is any improvement over using UniRep. We acknowledge a conflict of interest, as SecVec was produced by our lab (but in another ICLR open review publication (paper) SeqVec seemed to spawn better results then UniRep for the problem of GO annotation prediction) \\n\\nThank you for this suggestion. We were unaware of the SeqVec \\u2013 there seem to be several new protein embeddings coming out in the past year, and we felt that benchmarking across all these different embeddings would be out of the scope of this paper. We will evaluate on SeqVec and if the results on SeqVec are better, we will use these results for the paper instead of UniRep. \\n\\n\\n* For the test sets: create three different test sets based on difficulty / novelty of the proteins, e.g. based on similar to the different test sets evaluated on protein-protein interaction prediction information (Evolutionary profiles improve protein\\u2013protein interaction prediction from sequence, Hamp&Rost, 2015)\\n\\nWe thank the commenter for this suggestion \\u2013 these types of test sets would make the evaluation much stronger. We will explore the addition of more strictly designed test-sets as demonstrated in the paper.\", \"title\": \"Thank you for you helpful comments (2)\"}", "{\"comment\": \"We thank the commenters for their input and for their helpful suggestions for further directions to improve the paper. We address the specific comments below.\\n\\n- The authors claim that there might be similar data items in the random train and test splits (as similar chemicals or proteins might not be accounted for randomly). However, in the temporal splits there is no way of telling that there is no similarity between a compound being analyzed in the past w.r.t. more recently, leading us to believe that temporal splits address only partly the bias between training and testing\\n\\nWhile the commenters are correct that the temporal split may not fully address the bias between training and testing, we include the temporal split in the results for two reasons\\n 1. The temporal split creates a more realistic test set, since in real-world usage of PCM models one would train on all currently available data. This split simulates this case. \\n 2. The temporal split was used by Lenselink et al to benchmark their handcrafted-descriptor models \\u2013 for the sake of comparability we stick to the same evaluation schema. \\nOn revision, we will include these details \\u2013 in particular to state that the temporal split may not fully address the bias between train and test, but that we include it because it is a realistic split. \\nWe note that we are aware of more rigorous splitting schemes based on protein sequence similarity, and of the contribution of Rost & Sander to these ideas in their seminal work from 1993 (https://www.ncbi.nlm.nih.gov/pubmed/8356056)\\n\\n\\n- Validation set for temporal split: it's not clear how the authors split the temporal splits for validation, training and testing. Especially, as this split will very likely have the same bias as the authors describe for the random split\\n\\nWe apologize for not clearly explaining the construction of the validation set for the temporal split. We used the same split procedure as Lenselink et al, where the test set contains data generated from assays done in 2013 and later, while the training set and validation set is constructed by randomly splitting the pre-2013 data. By ensuring both training and validation is done on pre-2013 data, we seek to closely mimic the real-world case of training on all currently available bioactivity data, and to avoid the bias from random splitting as much as possible. We will make this clearer in revision.\\n\\n\\n- It is not clear how the authors try to calculate the standard error and how they apply bootstrapping. This might be due to unclear wording or lack of details \\n\\nWe apologize for our unclear writing in this section. For the random split, each bootstrap set is constructed by separating a random test set, then separating a random validation set from the remaining data, then resampling with replacement from the training set to create our bootstrap training set. For the temporal split, the test set is fixed, containing all bioactivity data generated from assays done in 2013 and later. Then, in each set we randomly partition the pre-2013 data into a training and validation set, then resample with replacement from the training set to create our bootstrap training set. In all cases, we also save our random seeds for future reproducibility. The standard error we report is then computed from the standard deviation of the performance of the bootstrap samples. In revision we will make this section more clear and include the details in the main body or in the appendix. \\n\\n\\n- Low coverage splits: why wasn't a traditional approach (e.g. homology reduction) used for this? \\n\\nWe ask the commenter to elaborate what they are referring to as homology reduction, other than the mathematical concept in algebraic topology. Do they mean some technique involving construction of a hold-out set where the hold-out set comprises a distinct family of proteins excluded from the training set?\\n\\n\\n- Table 1: the results on the low coverage splits are missing, making it difficult to easily compare across different approaches \\n\\nThe low-coverage split does not lend itself to computing standard errors in the same manner as the temporal or random splits, since we already create several splits of the data \\u2013 based on which five proteins have been assigned to the test set. Also, in this case, we find the average performance is not as useful as visualizing the difference in performance with respect to each protein.\", \"title\": \"Thank you for the helpful comments! (1)\"}", "{\"comment\": [\"Summary:\", \"This work describes a novel way to combine data-driven representations for small molecules as well as proteins for the prediction of protein-ligand binding. The authors show that these novel representations, which are learned in an unsupervised fashion, improve over traditionally crafted features.\", \"# Things that we appreciated:\", \"Network graphs: clear and easy to follow\", \"Approach: novel and interesting, laying out well the disadvantages of traditional chemical/protein encoding\", \"# A few notes:\", \"The authors claim that there might be similar data items in the random train and test splits (as similar chemicals or proteins might not be accounted for randomly). However, in the temporal splits there is no way of telling that there is no similarity between a compound being analyzed in the past w.r.t. more recently, leading us to believe that temporal splits address only partly the bias between training and testing\", \"Validation set for temporal split: it's not clear how the authors split the temporal splits for validation, training and testing. Especially, as this split will very likely have the same bias as the authors describe for the random split\", \"It is not clear how the authors try to calculate the standard error and how they apply bootstrapping. This might be due to unclear wording or lack of details\", \"Low coverage splits: why wasn't a traditional approach (e.g. homology reduction) used for this?\", \"Table 1: the results on the low coverage splits are missing, making it difficult to easily compare across different approaches\", \"Writing: Table 1 should appear in the results instead of the methods, and the discussion should not introduce new methods (e.g. evaluation of the individual impact of each descriptor type, in the second paragraph of discussion)\", \"# Suggestions:\", \"UniRep was used for the embedding of proteins, but another method is available (SeqVec). We suggest also trying this encoder and see if there is any improvement over using UniRep. We acknowledge a conflict of interest, as SecVec was produced by our lab (but in another ICLR open review publication (https://openreview.net/forum?id=Skx73lBFDS), SeqVec seemed to spawn better results then UniRep for the problem of GO annotation prediction)\", \"For the test sets: create three different test sets based on difficulty / novelty of the proteins, e.g. based on similar to the different test sets evaluated on protein-protein interaction prediction information (Evolutionary profiles improve protein\\u2013protein interaction prediction from sequence, Hamp&Rost, 2015)\"], \"this_comment_was_co_authored_by_members_of_rostlab\": \"Michael Heinzinger, Tobias Olenyi, Oscar Llorian, Maria Littmann, Christian Dallago\", \"title\": \"Nice work, needs some clarifications\"}" ] }

[ "Dibakar Gope", "Jesse G Beu", "Urmish Thakker", "Matthew Mattina" ]

[ "Model compression", "ternary quantization", "energy-efficient models" ]

{ "note_id": [ "fmCYK3XXCA", "BygTB6qnsr", "B1g-Xoq2sr", "ByxO8q53oS", "S1g1j74b9r", "Bkxmu76RYr", "ryxEfElTYB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751452, 1573854533071, 1573853977132, 1573853775692, 1572058006859, 1571898218774, 1571779595558 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2533/Authors" ], [ "ICLR.cc/2020/Conference/Paper2533/Authors" ], [ "ICLR.cc/2020/Conference/Paper2533/Authors" ], [ "ICLR.cc/2020/Conference/Paper2533/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2533/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2533/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper presents a quantization method that generates per-layer hybrid filter banks consisting of full-precision and ternary weight filters for MobileNets. The paper is well-written. However, it is incremental. Moreover, empirical results are not convincing enough. Experiments are only performed on ImageNet. Comparison on more datasets and more model architectures should be performed.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"We thank the reviewer for the thoughtful feedback. Please find the responses inline below.\\n\\n(1) I think the authors over-state their claims of no loss in accuracy, in Table 2 we see a clear loss in accuracy from MobileNets to MobileNets + Hybrid Filter Banks.\\n\\nPlease note that the results reported for hybrid filter banks are from a single run of our training procedure, we believe the 0.51% accuracy drop can be bridged with more training hyperparameter exploration. We will add the updated results with more hyperparameter exploration in the final version. In any case we will change the text appropriately to reflect the reviewer\\u2019s concerns.\\n\\n(2) I think a better venue for this research may be a more systems-focused conference or journal.\\n\\nWhile we agree this paper has a system-focused contribution in terms of proposing a novel hardware accelerator consisting of both MAC and adder units for efficient execution of ternary weight networks, the primary focus throughput the paper is the model compression of MobileNets\\u2019 already highly optimized network, the discussion of issues around quantizing MobileNets to ternary weights and subsequent development of a ML solution to address that. Hence, we submitted the work to a ML conference where we feel attendees will find more value and interest in this work.\\n\\n(3) I think this research is quite incremental over MobileNets.\\n\\nPlease see answer to Q1 of Review #1. Approximately 2x compression in model size and 28% saving in energy with no degradation in inference throughput or little degradation in accuracy is not an incremental contribution for an already compact and compute-efficient architecture like MobileNets. This paper provides the new best-in-class result for this task.\\n\\n(4) I think this research is unlikely to spur further research strains.\", \"we_strongly_believe_that_this_work_will_spur_further_research_strains_because_of_the_following_reasons\": \"a)\\tThe research community has put significant effort into quantizing ResNet architecture to ternary values while preserving the accuracy of full-precision model [1][2][3][4][5][6][7]. Considering the fact that MobilNet family of networks (1x1 convolutions) have fueled tremendous progress in the design and organization of resource-efficient architectures in recent years, it is highly likely that the MobileNet-like compute-efficient networks will be the next target for ternary quantization. \\nb)\\tTo the best of our knowledge, the hybrid filter banks proposed in this work is a first step towards quantizing the already compute-efficient MobileNets architecture to ternary values with a negligible loss in accuracy on a large-scale dataset, such as ImageNet. The Table 2 in Page 13 of reference [4] paper further shows the evidence of our claim.\", \"references\": \"(1)\\tKuan Wang et al., \\u201cHAQ: Hardware-Aware Automated Quantization with Mixed Precision\\u201d, CVPR 2019\\n(2)\\tRon Banner et al., \\u201cPost-training 4-bit quantization of convolution networks for rapid-deployment\\u201d, NeurIPS 2019\\n(3)\\tYoni Choukroun et al., \\u201cLow-bit Quantization of Neural Networks for Efficient Inference\\u201d, ICCV 2019 workshop\\n(4)\\tChristos Louizos et al., \\u201cRelaxed Quantization for Discretized Neural Networks\\u201d, ICLR 2019\\n(5)\\tJiwei Yang et al., \\u201cQuantization Networks\\u201d, CVPR 2019\\n(6)\\tRuihao Gong et al., \\u201cDifferentiable Soft Quantization: Bridging Full-Precision and Low-Bit Neural Networks\\u201d, ICCV 2019\\n(7)\\tSangil Jung et al., \\u201cLearning to Quantize Deep Networks by Optimizing Quantization Intervals with Task Loss\\u201d, CVPR 2019\\n\\n\\n(5) There is a significant amount of compute and training complexity required to reduce the model size, e.g. versus model pruning or tensor decomposition. It seems this research would be incredibly difficult to reproduce.\\n\\n(a)\\tThe training of hybrid filter banks requires finding the appropriate division of output channels to be generated from either full-precision or ternary weight filters at each layer to explore the best cost-accuracy tradeoffs. Model pruning and tensor decomposition techniques also must find either the appropriate sparsity factors or the appropriate sizes of the tensor decomposed matrices during training to meet the required cost-accuracy tradeoffs. From our experience hybrid filter banks require similar training complexity to that of the training complexity of model pruning and tensor decomposition techniques.\\n(b)\\tFurthermore, note that the model pruning and the tensor decomposition techniques (few cited in section 5, Related Work) are orthogonal to our compression scheme, they can be used in conjunction with hybrid filter banks to further reduce model size and computational complexity (likely at the cost of accuracy).\\n(c)\\tHybrid filter banks compression scheme is profitable and generalizable to other network architectures (ResNet dominated with 3x3 convolutional layers) as well (please see answer to Q1 to Review# 3).\\n(d)\\tWe are happy to release the training infrastructure for hybrid filter banks to ease reproducibility of our work.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We thank the reviewer for the thoughtful feedback. Please find the responses inline below.\\n\\n1. This paper is incremental in nature, with a natural generalization of (Tschannen et al.(2018)).\\n\\nWhile our work can be considered an extension of the state-of-the-art StrasseNets technique, its key contributions are as follows:\\n(a)\\tObservation of the difficulty with quantizing depthwise separable convolutional layers (dominated by 1x1 pointwise convolutions) using state-of-the-art techniques\\n(b)\\tObservation of the variance in the sensitivity (listed below) of filters to ternary quantization and subsequent exploitation of this to develop hybrid filter banks to quantize already highly optimized 1x1 pointwise layers of MobileNets architecture.\\n(1) Different sensitivity of individual filters to quantization. We empirically showed evidence of that in Figure 1(a)\\n(2) Different sensitivity for groups of filters under StrassenNets quantization. We gave a mathematical evidence/proof of that in Figure 1(b) and later in Appendix A.\\n(c)\\tFurthermore, we believe that ~2x compression in model size and 28% saving in energy with no degradation in inference throughput or accuracy is not an incremental contribution for an already compact and compute-efficient architecture like MobileNets. This paper provides the new best-in-class result for this task.\\n\\n\\n2. For this kind of paper, I would like to see a more complete set of empirical results. However, The experiments only perform comparison on ImageNet dataset. Though this dataset has a reasonable size, however, as in many cases, different datasets can bias the performance of different models. To strengthen the results, could the experiments be conducted on multiple datasets as in (Tschannen et al.(2018))? The proposed method is only designed for MobileNets. Is it possible to apply the hybrid filter banks to other neural network.\\n\\n\\nPlease see answer to Q1 of Review #3 for the additional data points we\\u2019ve collected in response to these rebuttals. We will add these performance results for ResNet, MobileNetV2 on the ImageNet dataset in the final version.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"We thank the reviewer for the thoughtful feedback. Please find the responses inline below.\\n\\n1. Comparison to prior work:\\n\\n(a)\\tWe haven\\u2019t found a work that attempt to quantize the weights of MobileNet to ternary or binary values for either CIFAR10, ImageNet or any other dataset. Table 2 in Page 13 of reference [1] further supports this claim. Prior works on ternary quantization [1,2,3,4] primarily evaluate the ResNet architecture on different datasets. To the best of our knowledge, the hybrid filter banks proposed in this work is a first step towards quantizing the already compute-efficient MobileNets architecture to ternary values with a negligible loss in accuracy on a large-scale dataset, such as ImageNet.\\n(b)\\tTo address your concern, we have since evaluated the ResNet-20 architecture with hybrid filter banks on the CIFAR-10 dataset to demonstrate its efficacy over other state-of-the-art ternary quantization techniques and its generalizability to other neural network architectures, especially to architectures dominated with 3x3 convolutional layers. ResNet-20 has 19 3x3 convolutional layers. Hybrid filter banks for ResNet-20 consistently achieves a better accuracy for different hidden layer widths when compared to the accuracy reported in StrassenNets [5] while ensuring improvement or no degradation in inference throughput. For example, the Hybrid ResNet-20 with the Alpha = 0.25 and the r = 0.75*cout configuration achieves an accuracy of 91.55% when compared to the accuracy 90.62% observed by StrassenNets with r = 0.75*cout. The accuracy of full-precision ResNet-20 is 92.1. All other configurations consistently outperform the state-of-the-art StrassenNets, demonstrating the generalizability and effectiveness of hybrid filter banks to 3x3 convolutional layers. We will add the performance results of hybrid filter banks for ResNet-20 on CIFAR-10 in the final version.\\n\\nAlpha = The fraction of channels generated in an output feature map from the full-precision weight filters\\nr = The hidden layer width of a strassenified convolution layer (ternary weight filters)\\n\\n(c)\\tIn this work, we compare hybrid filter banks against StrassenNets and TWN. We will compare hybrid filter banks to the other ternary quantization techniques (such as, TTQ [6] and ABC-Net which are published before StrassenNets [5]) and add their performance results in the final version. As StrassenNets already showed better performance results over the prior ternary quantization methods [6], we concentrated our effort to demonstrate improvement over StrassenNets for MobileNet-like already compact and compute-efficient architecture on a challenging dataset. Furthermore, we believe that when the prior ternary quantization schemes such as TWN, TTQ, ABC-Net could not preserve the accuracy for over-parameterized and less compute-efficient ResNet (in comparison to MobileNet), we should not expect preservation of accuracy for the already compact and compute-efficient MobileNets for these techniques.\", \"references\": \"(1)\\tChristos Louizos et al., \\u201cRelaxed Quantization for Discretized Neural Networks\\u201d, ICLR 2019\\n(2)\\tJiwei Yang et al., \\u201cQuantization Networks\\u201d, CVPR 2019\\n(3)\\tRuihao Gong et al., \\u201cDifferentiable Soft Quantization: Bridging Full-Precision and Low-Bit Neural Networks\\u201d, ICCV 2019\\n(4)\\tSangil Jung et al., \\u201cLearning to Quantize Deep Networks by Optimizing Quantization Intervals with Task Loss\\u201d, CVPR 2019\\n(5)\\tMichael Tschannen et al., \\u201cStrassenNets: Deep learning with a multiplication budget\\u201d, ICML 2018\\n(6)\\tChenzhuo Zhu et al., \\\"Trained Ternary Quantization\\\", ICLR 2017\\n\\n\\n2. Little blurry figures.\\nThank you for pointing to this out; we will fix that in the final version.\\n\\n\\n3. Current framework to MobileNetV2?\\n\\n(a)\\tYes, it is absolutely possible to apply the current framework and Hybrid Filter Banks to MobileNetV2. In the limited rebuttal time, we could only evaluate MobileNetV2 on the ImageNet dataset with one initial set of model hyperparameters associated with hybrid filter banks approach. The initial results with MobileNetV2 is promising, incurring as little as 2.62% accuracy loss compared to the uncompressed MobilNetV2 *with only the first set of hyperparameters we chose *. We believe the small accuracy drop can be bridged with more model hyperparameters (e.g., appropriate division of output channels to be generated from either full-precision or ternary weight filters at each layer, appropriate value of hidden layer width for ternary weight filters, etc.) exploration associated with hybrid filter banks approach\\n(b)\\tFurthermore, due to the rebuttal time constraint and long training time of MobileNetV2 on ImageNet we could not apply knowledge distillation (as exploited by the StrasseNets baseline and our hybrid filter banks, mentioned at the end of Section 3). Knowledge distillation historically improves accuracy by another 1-2%. We will add the performance results for MobileNetV2 with hybrid filter banks in the final version.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper presents a quantization method that generates per-layer hybrid filter banks consisting of full-precision and ternary weight filters for MobileNets.\", \"strength\": \"(1)\\tThe paper proposes to only quantize easy-to-quantize weight filters of a network layer to ternary values while also preserving the representational ability of the overall network by relying on few full-precision difficult-to-quantize weight filters. \\n(2)\\tThe proposed method maintains a good balance between overall computational costs and predictive performance of the overall network. Experimental results show that the proposed hybrid filter banks for MobileNets achieves savings in energy and reduction in model size while preserving comparable accuracy. \\n(3)\\tThe description is clear in general.\", \"weakness\": \"(1)\\tThough the paper claims that recent works on binary/ternary quantization either do not demonstrate their potential to quantize MobileNets on ImageNet dataset or incur modest to significant drop in accuracy while quantizing MobileNets with 4-6-bit weights, it may worth comparing to the methods that achieved start-of-art results on other datasets to demonstrate the efficiency of the proposed method. \\n(2)\\tFigure 1 and Figure 2 is a little blurry. \\n(3). How is about the performance compared to latest work? Is it possible to apply current framework to MobileNetV2 ? If can, what's performance?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes a novel quantization method towards the MobileNets architecture, with the consideration of balance between accuracy and computation costs. Specifically, the paper proposes a layer-wise hybrid filter banks which only quantizes a fraction of convolutional filters to ternary values while remaining the rest as full-precision filters. The method is tested empirically on ImageNet dataset.\\n\\nThis paper is generally well written with good clarity. Several concerns are as follows:\\n\\n1. This paper is incremental in nature, with a natural generalization of (Tschannen et al.(2018)). But it is still an interesting contribution. For this kind of paper, I would like to see a more complete set of empirical results. However, The experiments only perform comparison on ImageNet dataset. Though this dataset has a reasonable size, however, as in many cases, different datasets can bias the performance of different models. To strengthen the results, could the experiments be conducted on multiple datasets as in (Tschannen et al.(2018))?\\n\\n2. The proposed method is only designed for MobileNets. Is it possible to apply the hybrid filter banks to other neural network architecture?\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors focus on quantizing the MobileNets architecture to ternary values, resulting in less space and compute.\\n\\nThe space of making neural networks more energy efficient is vital towards their deployment in the real world.\\n\\nI think the authors over-state their claims of no loss in accuracy, in Table 2 we see a clear loss in accuracy from MobileNets to MobileNets + Hybrid Filter Banks.\\n\\nI think this research is quite incremental over MobileNets and is unlikely to spur further research strains. I think a better venue for this research may be a more systems-focused conference or journal.\\n\\nThere is a significant amount of compute and training complexity required to reduce the model size, e.g. versus model pruning or tensor decomposition. It seems this research would be incredibly difficult to reproduce.\"}" ] }

[ "Gregor Bachmann", "Gary Bécigneul", "Octavian-Eugen Ganea" ]

[ "graph convolutional neural networks", "hyperbolic spaces", "gyrvector spaces", "riemannian manifolds", "graph embeddings" ]

{ "note_id": [ "hODE9LtIBl", "HJlp1gQsjr", "H1lgGrMjiH", "ByetjhULsH", "BJlcvhU8iS", "H1xjf2ILjr", "HyxrcqSgsH", "H1eQwEeRYS", "B1lrh6UpYS", "rJxr_fmNFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751422, 1573756900943, 1573754120184, 1573444769403, 1573444706027, 1573444626876, 1573046925494, 1571845211042, 1571806637241, 1571201645363 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2532/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2532/Authors" ], [ "ICLR.cc/2020/Conference/Paper2532/Authors" ], [ "ICLR.cc/2020/Conference/Paper2532/Authors" ], [ "ICLR.cc/2020/Conference/Paper2532/Authors" ], [ "~Jialin_Liu3" ], [ "ICLR.cc/2020/Conference/Paper2532/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2532/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2532/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes using non-Euclidean spaces for GCNs, leveraging the gyrovector space formalism. The model allows products of constant curvature, both positive and negative, generalizing hyperbolic embeddings.\\n\\nReviewers got mixed impressions on this paper. Whereas some found its methodology compelling and its empirical evaluation satisfactory, it was generally perceived that this paper will greatly benefit from another round of reviewing. In particular, the authors should improve readability of the main text and provide a more thorough discussion on related recent (and concurrent) work.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Thank you for the response/revision. It clarifies all my questions\"}", "{\"title\": \"Thank you for taking another look at our improved paper\", \"comment\": \"We would love to hear your feedback on our substantially improved paper (based on your suggestions). Unfortunately, ICLR's tight schedule would prevent us to answer any additional questions after 15th of November. Thank you!\"}", "{\"title\": \"Paper improvement\", \"comment\": \"We would like to thank reviewer #1 for the constructive critics.\\n\\nWe have significantly improved the exposition in this paper, ameliorated the discussion of the related work, added new theoretical results in the main text, compressed the appendix and enhanced the description of the experimental setup. We have updated the paper pdf and kindly ask the reviewer to take another look.\", \"theory\": \"Section 2 now formalizes our contribution of unifying Riemannian geometry and gyrovector spaces for all stereographic spaces of constant curvature of any sign, formally proving that the interpolation is smooth at 0. Please note that gyrovector spaces and their properties were stated only for negative curvature spaces prior to our work.\\nTheorem 1 formally states when k-addition is defined for spherical spaces. \\nTheorem 2 formally derives formulas of exponential map, logarithmic map, geodesics and distance w.r.t. Gyro-operations in the setting of positive curvature, by making use of the Egregium theorem.\\nTheorem 3 proves smoothness of the main quantities appearing in Theorem 2 w.r.t. curvature, i.e. that they are differentiable at 0, by showing that left and right derivatives from both models actually match. It also gives the derivative of the distance function w.r.t. curvature around 0.\\nTheorem 4 describes two simple properties of our proposed left-matrix-multiplication, in particular that it is Mobius-scalar-associative.\\nTheorem 5 shows that our proposed left-matrix-multiplication is an intrinsic averaging for Riemannian manifolds of constant curvature, i.e. it does commute with isometries when the matrix A is right-stochastic, which is not the case of right-matrix-multiplication. This is a desirable property for Riemannian averaging.\\nTheorem 6 shows that weighted combinations in the tangent space used in the very recent works of [1,2] (appeared after ICLR submission deadline) is also an intrinsic averaging for Riemannian manifolds of constant sectional curvature, i.e. it does commute with isometries.\", \"literature\": \"A more careful review of the literature has been incorporated into the introduction section.\\nWe have also incorporated the recent works [1,2] to appear soon at Neurips\\u201919 (their text became available very recently and after the ICLR submission deadline).\", \"writing\": \"\", \"we_have_unified_all_notations\": \"bold for multidimensional quantities, i.e. vectors, matrices, embeddings.\\nWe have significantly improved the writing, re-done the bibliography and citing, and organized the most important theorems and definitions into a clearer presentation.\", \"motivation\": \"We have rewritten the introduction and detailed the motivation of using spaces of constant curvature, i.e. they better represent certain classes of data such as hierarchical structures, scale-free graphs, complex networks and cyclical or spherical data. This is in accordance with the previous related work on non-Euclidean embeddings. Moreover, we have detailed some limitations of Euclidean spaces in Appendix B.\\n\\nEXPERIMENTAL SETTINGS & HYPERPARAMETERS:\\nAre added to section 4 and appendix E\", \"experimental_results\": \"Our methods outperform the Euclidean GCN models for the synthetic datasets for minimizing distortion, while being better or competitive for node classification. We believe further experimental investigations are needed to better train non-Euclidean graph neural network models.\\n\\n\\n[1] Hyperbolic Graph Neural Networks, I. Chami et al., Neurips\\u201919\\n[2] Hyperbolic Graph Convolutional Networks, Q. Liu et al., Neurips\\u201919\"}", "{\"title\": \"Paper update and comments\", \"comment\": \"We thank reviewer #3 for the extensive feedback. We have incorporated it as stated below.\\n\\nWe have significantly improved the exposition in this paper, ameliorated the discussion of the related work and added the suggested new interesting references, added theorems and more formal statements in the main text, compressed the appendix and enhanced the description of the experimental setup. \\n\\nProofs that the space \\\"interpolates smoothly with curvature\\\": we added formal proofs (see theorems 2 and 3) that all the operations are differentiable, i.e. the gradients are equal from both the left and right at 0, w.r.t. curvature, for the chosen models of hyperbolic and spherical geometry.\", \"k_addition_definiteness_in_the_spherical_setting\": \"we have added the formal condition that the k-addition be well-defined, and a proof that for two points this condition indeed recovers x != y / (k ||y||^2) - see Theorem 1.\\n\\nTheorem 5 shows that our proposed left-matrix-multiplication is an intrinsic averaging for Riemannian manifolds of constant curvature, i.e. it does commute with isometries when the matrix A is right-stochastic, which is not the case of right-matrix-multiplication. This is a desirable property for Riemannian vector averaging.\\n\\nTheorem 6 shows that weighted combinations in the tangent space used in the very recent works of [1,2] (appeared after ICLR submission deadline) is also an intrinsic averaging for Riemannian manifolds of constant sectional curvature, i.e. it does commute with isometries.\", \"other_comments\": \"-Synthetic tree: contains |V| - 1 edges. We corrected this mistake.\\n-Curvatures: are learned as we state in the paper section 4 and appendix F.\\n-Citations: fixed. \\n-Working with the Poincare ball as opposed to the hyperboloid model: this allows us to use gyrovector spaces which are defined either for the Poincare ball or the Klein model, as well as to connect those with the Riemannian geometry of the space. Moreover, as we show in the paper, we can now smoothly interpolate between all constant curvature spaces which is beneficial for learning curvatures without a priori deciding on their signs.\\n-Statement about \\u201cgeneral class of trees\\u201d replaced by \\u201call weighted or unweighted trees\\u201d. \\n\\n[1] Hyperbolic Graph Neural Networks, I. Chami et al., Neurips\\u201919\\n[2] Hyperbolic Graph Convolutional Networks, Q. Liu et al., Neurips\\u201919\"}", "{\"title\": \"Thanks & paper update\", \"comment\": \"We thank reviewer #2 for the useful feedback.\\n\\nWe have significantly improved the exposition in this paper, ameliorated the discussion of the related work, added theorems and more formal statements in the main text, compressed the appendix and enhanced the description of the experimental setup. We have updated the paper and kindly ask the reviewer to take another look.\", \"robustness_of_the_curvature_sampling_method\": \"we provide confidence intervals in Table 2 of our results. We learn curvatures for each of the component spaces and show learned values together with confidence intervals in Appendix E. It can be seen that these variances are in the low regime.\"}", "{\"title\": \"GCN's weight sharing mechanism understanding\", \"comment\": \"Really interesting work. GNN formed in Non-Euclidean Space is a promising topics.\\nI don't understand why the author said \\\"The proposed architecture ... while this weight sharing can be understood as an efficient diffusion-like regularizer.\\\" in part 1 introduction, page 2. What does \\\"diffusion-like regularizer\\\" mean?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this paper, the authors address representation learning in non-Euclidean spaces. The authors are motivated by constant curvature geometries, that can provide a useful trade-off between Euclidean representations and Riemannian manifolds, i.e. arriving at more suitable representations than possible in the Euclidean space, while not sacrifising closed-form formulae for estimating distances, gradients and so on.\\n\\nThe authors point out that an extension of the gyrovector space formalization to spaces of constant positive curvature (spherical) is required, and with the corresponding formalization for hyperbolic spaces, one can arrive at a unified formalism that can interpolate smoothly between all geometries of constant curvature. The authors propose to do so by replacing the curvature while flipping the sign in the standard Poincare model. This is a strong point reagarding this work, as it seems that no such unification has been attempted in the past (although simply replacing the curvature in the Poincare model seems a bit too straightforward to not have been attempted, it seems to be the case). \\n\\nThe authors also provide extensive supplementary material (around 20 pages) with detailed derivations and descriptions of experiments. This also makes me wonder if this paper is more suitable for a journal - both in terms of the extensive supplementary material (e.g., curvature sampling algorithm and other details can be found only in supplementary), as well as the more rigorous review process that a journal paper goes through. \\n\\nIn the main paper, only proof of concept experiments are provided (one experiment), that nevertheless show competitive performance under varying settings. However, it seems to me that such contributions in the rising field of geometric deep learning, where several challenges are yet to be overcome, can be beneficial for future research. \\n\\nSince in the supplementary experiments also, it seems that curvature does have small variance in the results, how would the authors assess the robustness of the curvature sampling method with respect to the results?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThe authors propose using non-Euclidean spaces for GCNs. This is inspired by the recent work into non-Euclidean, and especially hyperbolic, embeddings. A few papers have recently tried to go past embeddings into building non-Euclidean models, requiring the lifting of standard operations in Euclidean space to non-Euclidean settings. This has been done in particular in hyperbolic space, but some datasets benefit from more complex spaces. The authors combine the mixed-curvature product formalism that uses products of Euclidean, hyperbolic, and spherical spaces for embeddings, but use these for GCN operations. \\n\\nDoing this requires, in particular, developing a reasonable way to perform these operations in spherical space (since Euclidean is trivial and hyperbolic has been recently worked on). The authors do a nice lifting via complex operations, and both the hyperbolic and spherical spaces can devolve into the flat Euclidean space when their curvature goes to 0. The authors implement these GCNs, train the curvatures, and demonstrate performance improvements over Euclidean only versions on node classification on benchmark datasets. They also give a fairly nice introduction to all of these ideas in an extended appendix.\\n\\n\\nStrengths, Weaknesses, Recommendation:\\nThis paper is reasonably interesting---it joins an effort to produce non-Euclidean models in a tractable way, which is fairly challenging, but could have a good impact. On the plus side, it's great that the authors added the nice development for the spherical operations, since that will come in handy for many models. The experiments are also good. On the downside, everything here is an extension of existing work, and the body of the paper is hard to read (though this may be inevitable, there's a lot of background to go over here). Overall I recommend accepting it; I think it's a solid contribution.\", \"comments\": [\"I don't understand why the authors say that their space \\\"interpolates smoothly\\\" just because the limit in the curvature is the same from the left and right side. For example, the absolute value function has the same limit from the left and the right at 0, but it's not differentiable there. Is it actually true that if we take the derivatives of the piecewise hyperbolic/spherical distance function that it's differentiable at c=0?\", \"There are a couple of recent papers that also consider hyperbolic GCNs, and in fact use similar ideas for the aggregation and update steps (i.e., same lift to hyperbolic space). However, these were recently NeurIPS papers, and the text is not yet out, so I don't think this should affect the authors' independent work (and also the product part is new). I do recommend that the authors compare against those results in a future update of this work. The papers are \\\"Hyperbolic Graph Convolutional Neural Networks\\\" by Chami et al and Hyperbolic Graph Neural Networks by Liu et al.\", \"One thing that I didn't see discussed by the authors is that there are subtle difference between hyperbolic and spherical spaces. For example, the weighted midpoint of Def. 3.2 doesn't immediately extend to spherical space (or at least won't be unique). As an example, consider S^2 and the mean of two antipodal points on it---there's many choices for the midpoint. You probably have to limit the operation to a half-sphere (there's some ideas for this in Gu et al).\", \"For the synthetic tree, why is the number of edges 2(|V|-1) rather than |V|-1?\", \"Are the curvatures the same for each layer for the GCNs? This is an interesting point to discuss (some of the NeurIPS papers I mentioned train the curvature for each layer). Also, how do you select the number of factors of each type?\", \"Minor, but some of these citations can be updated. The \\\"De Sa\\\" et al 2018 arxiv citation is really Sala et al and is an ICML '18 paper. Similarly, Gulcehre et al is a 2019 ICLR paper, and so on. It's always good to get these right.\", \"Is there any actual empirical importance from recovering the Euclidean case exactly for 0 curvature? The reason I ask is that my experience is that the hyperboloid is typically easier to work with.\", \"One useful thing to point out in B.3.3 is that in general, it need not be a diffeomorphism for all of M for any manifold, which leads to non-uniqueness. In differential geometry, the \\\"cut locus\\\" is the region beyond which there is this non-uniqueness.\", \"In the appendix, the statement \\\"Sarkar (2011) show that a similar statement as in Theorem 2 holds for a very general class of trees\\\" is confusing to me. The \\\"general class\\\", as far as I know, is actually *all* trees, weighted or unweighted.\"]}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper builds a new graph convolutional network (GCN) based on hyperbolic representations of the graph nodes: all latent representations of the nodes are points on Poincare disks. The authors adapted the Hyperbolic Neural Networks by Ganea et al. (2018) into the Kipf & Welling's (2016) version of GCN. Specifically, the authors variated the right matrix multiplication in GCN with Ganea et al.'s Mobius matrix-vector multiplication (that can be regarded as a transformation between two Poincare disks). Moreover, as a non-trivial adaptation, the author defined the left matrix multiplication in GCN (that can be regarded as a weighted linear combination of several points on the same Poincare disk) with Ungar's (2010) weighted barycenter, which is from computational geometry but not the machine learning community. The resulting method is tested on a toy problem and semi-supervised node classification of commonly used datasets, showing the possibility of improvement.\\n\\nThis is a practical contribution but not a theoretical contribution, as there is no main theorem (or equivalent statements), and the main novelty is on using Ungar's (2010) weighted barycenter to perform neighborhood features aggregations. There are some general discussions on how to adapt gyrovector space theory into spherical spaces. This is interesting but no formal results are presented.\\n\\nI vote for rejection for four major weaknesses explained as follows.\\n\\n(1) The experimental results cannot show the usefulness of the proposed GCN. The results on real datasets are similar to the regular GCN. As the authors themselves remarked, \\\"it can be seen that our models sometimes outperform the two Euclidean GCN\\\". The experimental settings, e.g. how the train:test datasets are split, and hyperparameter settings, are not clearly given.\\n\\n(2) The method is not well motivated. The motivation, based on the writing, is that constant curvature spaces are more general where the computation is easy to handle. This is too general and not enough as a motivation. After reading, the reviewer cannot understand why the user should bother to use hyperbolic representations that are more complex to compute in GCNs, given that the experimental results are roughly the same.\\n\\n(3) A large body of graph neural network literature is omitted. The authors start from a very high-level description of machine learning in constant curvature spaces. Such high-level introductions require more comprehensive literatures to support. In the first mentioning of Graph Neural Networks, the authors only cited Kipf & Welling's GCN. This is misleading. For example, the original GNN, ChebNet, etc. that leads to GCN can be mentioned. The reviewer recommends the authors to read some literature review of the topic of GCN and re-organize the references, and use search engines to have a better view on the state of the art of (hyperbolic) geometric theories and graph convolutional networks. This is a thriving area that requires a careful literature review.\\n\\n(4) The writing quality is not satisfactory. Here are a few examples: The ICLR citation style needs to use sometimes \\\\citep. The authors instead used \\\\cite everywhere, making the paper hard to read. The authors are suggested to use unified notations to denote vectors/matrices (e.g. all bold). The introduction can start at a lower level (such as flat/hyperbolic neural networks). Section 3.4, as the main technical innovation, can be extended and includes some demonstrations.\"}" ] }

[ "Yigit Ugur", "George Arvanitakis", "Abdellatif Zaidi" ]

[ "clustering", "Variational Information Bottleneck", "Gaussian Mixture Model" ]

{ "note_id": [ "DCJH7cc14u", "rkgw2B4Jqr", "Byx_WBDRtr", "HJl3w-spFH", "BJx9ZbkrdS", "Byl8Wur9Dr" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798751390, 1571927471262, 1571874048187, 1571823972290, 1570201858485, 1569507326434 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2531/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2531/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2531/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2531/Authors" ], [ "~Pranav_Poduval1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes to use a mixture of Gaussians to variationally encode high-dimensional data through a latent space. The latent codes are constrained using the variational information bottleneck machinery.\\n\\nWhile the paper is well-motivated and relatively well-written, it contains minimal novel ideas. The consensus in reviews and lack of rebuttal make it clear that this paper should be significantly augmented with novel material before being published to ICLR.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper purposes to cluster data in an unsupervised manner that estimates the distribution with GMM in latent space instead of original data space. Also, to better describe the distribution of latent code, the authors involve VIB to constraint the latent code.\\n\\nThe whole pipeline is clear and makes sense. And the experiment proves it's effective. \\n\\nHowever, in my eyes, it's almost like an extension of VaDE which combines VIB and GMM too. The main difference is the authors use some hyperparameters to control the optimization of the whole model and make some more proper hypothesis. All of those make sense but may not contribute much to the research field.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The author(s) posit a Mixture of Gaussian's prior for a compressed latent space representation of high-dimensional data (e.g. images and documents). They propose fitting this model using the Variational Information Bottleneck paradigm and explicate its derivation and tie it to the variational objective used by similar models. They empirically showcase their model and optimization methodology on the MNIST, STL-10, and Reuters10k benchmarks.\\n\\nThe idea of using a latent mixture of Gaussian's to variationally encode high-dimensional data is not new. The author(s) appropriately cite VaDE (Jiang, 2017) and DEC (Xie, 2016), \\\"Deep Unsupervised Clustering with Gaussian Mixture Variational Autoencoders\\\" (Dilokthanakul, 2017). However, the author(s) unfortunately did not mention \\\"Approximate Inference for Deep Latent Gaussian Mixtures\\\" (Nalisnick, 2016). Nalisnick et al. consider the exact generative process as the proposed by the author(s), but with the addition of a Dirichlet prior on the Categorical distribution. Nalisnick et al. fit their model with a VAE and circumvent Dirichlet intractability by positing a Kumarswamy stick-breaking variational posterior (the resulting distribution is on the simplex). They achieve 91.58% accuracy, but do so after fitting a KNN to the latent space. New techniques such as \\\"Implicit Reparameterization Gradients\\\" (Figurnov, 2018) and \\\"Pathwise Derivatives Beyond the Reparameterization Trick\\\" (Jankowiak, 2018) allow direct use of a Dirichlet posterior in VAEs. These methods are respectively implemented in the TensorFlow and PyTorch APIs. My point here is that there are many ways to fit this pre-existing model. From my review of these other works we have, for \\\"best run\\\" on MNIST, that author(s) > VaDE > Nalisnick > Dilokthanakul > DEC. Thus, the author(s) are SoTA for this particular generative process for their best run. It would be nice to see their standard deviation to gauge how statistically significant their results are. However, I am dubious of the SoTA claim as I detail later.\\n\\nThe author(s) derivation was sound, but a bit confusing for me. In particular, I found keeping track of P's and Q's very burdensome after reading sections 2.1 and 2.2. In my experience, Q is typically reserved for a variational distribution that approximates some intractable P distribution, while P is used to describe the generative process (i.e. likelihood) and other exact distributions. Furthermore, one typically introduces the generative process first using P distributions. Once, I got past this confusion everything else made sense. I might suggest introducing the generative process first and with P distributions instead of Q's. The variational model can follow with the Q distributions as the author(s) have it. Equations 4, 5, and 6 all exactly match the unsupervised information bottleneck objective (Alemi, 2017)--see appendix B. I am therefore confident in those equations. I carefully checked their derivation of the VADe comparison (equation 10 and appendix B). Their derivation is straightforward. The principle trick is using the MC assumption C->X->U for P distributions to claim p(u|x,c) = p(u|x). Equations 12-16 all follow naturally from equation 11. If equation 17 is indeed an approximation or bound approximation, I would suggest not using the equals sign. Instead, consider another appropriate operator or rename Dkl to indicate it is the approximated version (just as in equation 24).\\n\\nIf the author(s) like my suggestion regarding generative vs variational nomenclature, I would also change the second sentence of page 6 to something like \\\"We use our variational Qc|u distribution to compute assignments.\\\" Thereafter, I would drop the star indicator for optimal parameters. These parameters are not necessarily optimal given the non-convexity of the DNN. Replace with, \\\"after convergence, we ...\\\" Similarly, drop optimal from line 2 of algorithm 1.\\n\\nCircling back to the experiments, the author(s) use reported values from DEC and VaDE. Those works compute cluster assignments using a KNN classifier on the latent space. This paper however uses the arg max of equation 19. I much prefer this article's method, but for comparison purposes, the author(s) should similarly use a KNN classifier on their latent space to compute accuracy in the same manner. The use of KNN in these other works allows them to consider a number of latent clusters larger than the number of true classes (e.g. 20 clusters for MNIST). I like that the author(s) stick to 10 clusters for MNIST, but for comparison I would have liked to see a KNN generated accuracy alongside their equation 19 based accuracy. It looks like the author(s) implemented VADe for STL-10 based on table 1. If so, it seems they could easily implement equation 19 for their STL-10 value for VADe. If not, please correct this. Not to belabor further, but I would really like to see a table 2 that reports both equation 19 and KNN accuracies when available. Namely, the author(s) should report both values for their model and can leave equation 19 accuracies blank for reported values.\\n\\nLastly, I always raise an eyebrow when I see tSNE latent space representations. STL-10 was used to generate figure 6, where one needs dim(u) >> 2. However, MNIST can be well reconstructed using just 2 latent dimensions. In this case, tSNE is unnecessary. The author(s) state \\\"Colors are used to distinguish between clusters.\\\" This statement is unclear as to whether the author(s) are using the class label or learned latent cluster. If it is the former, figure 6 makes sense in that it shows misclassifications (i.e. red dots in the green cluster). However, if it is the latter then I am concerned tSNE is doing something weird.\\n\\nTo summarize, I enjoyed the paper. My only concern is novelty. Being the first to pair an existing model with an existing method, in my eyes, does not necessarily meet the ICLR bar. The author(s) seemingly achieve SoTA, but without KNN-based accuracies for their model it is hard to compare to cited works. Having these KNN results and error bars would strengthen their case substantially to: achieving SoTA for an existing model by being the first to pair it with an existing method.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": [\"This paper considers the autoencoder model combining the usual information bottleneck and the Gaussian mixture model (GMM). Using an approximation to deal with GMMs, the authors derive a bound on the cost function generalizing the ELBO. The performance of the proposed method is tested on three benchmark datasets and compared with existing methods combining VAE with GMM.\", \"While the framework and the performance of the proposed method are interesting and promising, some of its main parts are unclearly explained.\", \"Although Remark 1 well explains the difference between VaDE and the proposed objective functions, little is discussed how this difference affects the learnt model.\", \"I wonder why Q_\\\\phi(x|u) in Eq. (11) doesn\\u2019t have to be a distribution. What does the expression [\\\\hat{x}] mean?\", \"Eq. (17) is not explained clearly. Since the equality symbol is used, it is unclear where is approximation. Although this approximation is one of the main parts of the proposed method, little is discussed on the influence of this approximation.\", \"Are the information plane and latent representations in Figs 5 and 6 also available for DEC and VaDE and not limited to the proposed method?\"], \"minor_comments\": \"p.5, the math expression between Eqs. (16) and (17): The distribution q(x_i|u_i,m) is undefined.\\np.6, l.4 from the bottom: ACC is defined later.\\np.7, l.14: Should J be n_u?\"}", "{\"comment\": \"Since there is no exact calculation of the KL divergence between two Gaussian mixtures (in our case between a single component multivariate Gaussian and a Gaussian Mixture Model) we have to use an approximation. Our requirements for this approximation are 1) to be closed-form (and differentiable), in order to be able to back-propagate on the cost function by using stochastic gradient descent and 2) the approximation should not depend on the assignment probability $P_{C|X}$ because this probability is not known and as we mentioned in the paper the assumption $P_{C|X} = Q_{C|U}$, affects negatively the training process. The lower bound approximation of Hershey & Olsen (2007) full fills our requirements and works well enough in our experiments.\\n\\nIf there is another approximation of KL divergence between Gaussian mixtures, that satisfies the aforementioned requirements is definitely worth to give it a try to see how the performance of our proposed algorithm changes. \\n\\nFor sure the framework could be extended to other mixtures of distributions such as Bernoulli mixtures. The main modification will be the calculation of the new KL divergence term.\\n\\nI hope that we cover your questions.\", \"title\": \"Reply regarding KL approximation and extendibility of the model\"}", "{\"comment\": \"There must be other methods to approximate KL divergence b/w mixture distributions, is Variational the best? The Variational Approximation to KL divergence proof seemed to be true for general distributions, so can ur work be extended to a mixture distributions other than Gaussians, like Mixture of Bernoulli's (just an example)\", \"title\": \"Great Work, just a simple doubt\"}" ] }

[ "Hassan Kané", "Mohamed Coulibali", "Pelkins Ajanoh", "Ali Abdalla" ]

[ "NLP", "Protein", "Representation Learning" ]

{ "note_id": [ "_mmUJ7DqbB", "S1gcZeCCFB", "S1x_LgRTtr", "HJl3xM52tr", "BJgIvFnadr" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1576798751359, 1571901442227, 1571835983643, 1571754484243, 1570781533521 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2530/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2530/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2530/AnonReviewer3" ], [ "~Christian_Dallago1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper presents a linear classifier based on a concatenation of two types of features for protein function prediction. The two features are constructed using methods from previous papers, based on peptide sequence and protein-protein interactions.\\n\\nAll the reviewers agree that the problem is an important one, but the paper as it is presented does not provide any methodological advance, and weak empirical evidence of better protein function prediction. Therefore the paper would require a major revision before being suitable for ICLR.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"In this study, the authors develop a method to predict the function of proteins from their structure as well as the network of proteins with which they interact in a given tissue. The method consists in training a linear classifier on the output of two existing embedding methods, UniRep/SeqVec and OhmNet, respectively embedding the amino acid sequences and the tissue-specific protein-protein interaction networks. This method improves prediction of protein function by 19% compared to OhmNet alone.\\n\\nAlthough the topic is important and the article clearly written, I would tend to reject this article because there is no innovation in ML that would justify presentation at ICLR.\", \"strengths\": [\"the article is well-written and straight-forward. Prior art is well-described.\", \"timely and important topic (prediction of protein function), where ML is likely to have an big impact.\", \"positive scientific result (prediction is improved compared to prior art).\"], \"weakness\": [\"the ML aspect of this work is entirely based on prior art, the main innovation consisting in fitting a linear classifier on concatenated features extracted by two existing embedding methods (UniRep/SeqVec and OhmNet).\"], \"additional_feedback\": [\"In the ablation studies, why not include the condition SeqVec-Random and UniRep-random?\", \"-\\\"The average AUROC score from Random is a big higher than what could be expected from such representations thanks to the spikes (Placenta, Epidermis) which might also result from the huge functional class imbalance within those two tissues which, given the uniformity of the data, gets them more often than not on the right side of the hyperplane. \\\"\", \"=> unclear sentence.\", \"\\\"is a big higher\\\" => typo\", \"\\\"beta sheets .\\\" => typo\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper introduces a method to incorporate both sequence information and graph information to learn the protein representations. The idea is very straightforward. Basically, it used the embedding from OhmNet [Marinka et al, 2017] for the graph information and used the sequence information from UniRep [Ethan et al, 2019] or SeqVec [Michael et al, 2019]. It uses one experiment to show the performance of the combination of the two pieces of information.\", \"this_paper_should_be_rejected_for_the_following_reasons\": \"(1) The paper is obviously in the preliminary form without too much polish.\\n (2) The simple combination of the results from two published articles is not that interesting\\n(3) the presentation of the paper and idea is not in an acceptable form (the authors should at least draw a figure to show the big idea of the paper).\\n(4) the experiment is not convincing (there is only one experiment and it is not compared with the other state-of-the-art methods; since an embedding of a protein can be of broad usage, the authors should give its performance on four tasks: protein function prediction (GO term) [Maxat et al, 2018], enzyme function prediction (EC number) [Yu et al, 2018], protein secondary structure prediction [Sheng et al, 2016], protein contact map prediction [Jinbo Xu, 2019]) \\n(5) The learned embedding is not well discussed. The author should at least visualize the embeddings and check the physical and biological meaning of those embeddings, if possible. \\n\\nSince this manuscript would be for sure and have to be largely rewritten in the future, I would not give too many detailed suggestions but some high-level suggestions if the authors would like to refine this manuscript further and submit it somewhere else or ICLR next year:\\n(1) Further improve the idea of combining different sources of information. Combining different pieces of information will definitely be helpful but the authors should figure out a way to use them in a more natural way.\\n(2) Compared with other methods, which can combine different sources of information.\\n(3) Run more experiments on various tasks instead of one: protein function prediction (GO term), enzyme function prediction (EC number), protein secondary structure prediction, protein contact map prediction\\n(4) Refine the representation of the paper.\", \"references\": \"[Marinka et al, 2017] Predicting multicellular function through multi-layer tissue networks, 2017, https://arxiv.org/abs/1707.04638\\n[Ethan et al, 2019] Unified rational protein engineering with sequence-based deep representation learning, 2019, Nature Methods\\n[Michael et al, 2019] Modeling the Language of Life \\u2013 Deep Learning Protein Sequences, 2019, https://www.biorxiv.org/content/10.1101/614313v2\\n[Maxat et al, 2018] DeepGO: predicting protein functions from sequence and interactions using a deep ontology-aware classifier, 2018, Bioinformatics\\n[Yu et al, 2018] DEEPre: sequence-based enzyme EC number prediction by deep learning, 2018, Bioinformatics\\n[Sheng et al, 2016] Protein Secondary Structure Prediction Using Deep Convolutional Neural Fields, 2016, Scientific Reports\\n[Jinbo Xu, 2019] Distance-based protein folding powered by deep learning, 2019, PNAS\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This work tries to predict the protein functional activation on a tissue by combining the information from amino acid sequence, and tissue-specific protein-protein interaction network. The authors claim that with this joint representation, their model outperforms current methods (Omhnet) on 10 out of 13 tissues by a larger margin(19% on average).\", \"notations\": \"The notations in experiment is a little bit confusing. In Table 1, the authors refer to different representations with Ohmnet128, Ohmnet64, Ohmnet-Unirep, etc. However, these are not consistent to the ones introduced in Section 4.1: Ohmnet, Ohmnet64-Unirep64, etc. And \\\"0-pad\\\" is introduced in section 3.3 while they denote one method as \\\"Ohmnet64-0Padded\\\" in section 4.1. It would be difficult for the reader to infer the meaning of these abbreviations.\", \"method\": \"--amino acid sequence representation:\\nIt would be better to report the explained variance when using Principle Component Analysis (PCA) to project the 1024-dimensional output vector of SeqVec to 64 dimensional space. And the authors can show us more results of different projected dimensions (with different explained variance of the PCA).\", \"experiments\": \"--model:\\nMaybe the authors can provide us more information about the model they use. For classification, what exactly the linear model is? For learning representation, is there any modification of the structure and hyperparameter of UniRef, SeqVec and OhmNet? And is there any regularization? Showing training details like batch size, epochs would be helpful, too.\\n\\n--data:\\nIt would be better to show the details of the data this paper uses, like what the data looks like, what is the size, the distribution, and the pre-processing. What's more, since validation set is used for tuning, it would be better to report the results on test set.\\n\\n--result:\\nIn the second paragraph of Section 4.1, it would be more clear to use a table instead of words to show the results. What's more, what's exactly the 13 tissues this paper is using? Why they are chosen? Exactly what is the AUROC of each protein in each tissue? What the learning curves look like?\\n\\nAnother big issue is, what \\\"current methods\\\" is this paper comparing its result with? It seems like the authors are comparing their implementation of Ohmnet-SeqVec + linear model with Ohmnet + linear model, and report that the former one is of 19% higher AUROC than the latter. But how about the results of other models/methods on the same task in the literature. Is there anyone using similar joint representation and what is their results? \\n\\n--conclusion:\\nSince the proposed methods only achieve best results in 10 out of 13 tissues, it is improper to claim \\\"\\u2026 we make consistently better tissue-specific function predictions in 13 complex tissues \\u2026\\\".\\n\\nIn conclusion, I find this is an interesting paper, that the authors tries to combine amino acid sequence representation and tissue information to predict the activation of protein on specific tissue. However, the authors should perform more rigorous experiment, and show us more implementation details. What's more, comparing results with the start-of-art methods on the same task setting is important, too.\"}", "{\"comment\": \"The approach described in this work fuses protein information (obtained via embeddings) and network information (obtained via observed protein-protein interactions) to predict protein function according to GO. The proposed method betters the existing ones by a margin of 19%.\\n\\n- About the writing:\\n\\nIn general, I find the manuscript is written in good English. Nevertheless, at times, I feel more consideration should have been put in the choice of words in order to ease a non biological audience into the manuscript. En example of this is \\\"Proteins are generally understood through four levels of structures\\\", and following use of the word \\\"structure\\\" (e.g. in primary structure) in the Introduction. In protein space, structure has a very precise meaning. I would have personally preferred the word representation, or alike. On this note: sometimes the writing gets rapidly technical, and I fear that the audience might not fully grasp important concepts. An example of this is the description of quaternary structure in the introduction, which could be easily explained by simply describing proteins forming interactions amongst themselves or other proteins. Furthermore, the authors seem to sometimes loose consistency, e.g. UniRep becomes Unirep, SeqVec becomes Seqvec, OhmNet becomes Ohmnet. There also seem to be artifacts of editing, e.g.:\\n\\n... thanks to spikes (Placenta, Epidermis) ....\\n\\nbut never before have actual tissues been discussed and there is no reference to a figure which might display said spikes. More on this to follow.\\n\\nAnother example are the two last paragraphs of the Introduction, which share the same beginning \\\"In this work,..\\\", where instead I would have expected a natural following of the last paragraph to the second last.\\n\\n- Background/related work:\\n\\nWhile this particular approach (embeddings+graph) is the first I've heard of, I can hardly imagine there not existing any other approach using a mix of protein-level (1D) and network-level (4D) information. It would have been nice to see at least one paragraph spent on any previous work that aims at solving protein function prediction using these information.\\n\\nI find the manuscript is generally lacking in citations, breath and detail. For example, the authors describe sequence homology (as a way to infer function), but never reference important work introducing or exploring this concept (from the top of my head: Sander 1991, Rost 1992). Another example, in the Introduction, relates to the concept of secondary structure, which is described in Kabsch&Sander 1983.\\n\\nSome claims (e.g. in the Introduction) lack proper validation (either external or in the manuscript), e.g. \\\"Recent availability of high-throughoutput experimental data and machine learning based computational methods can be useful for unveiling and understanding such patterns.\\\": Why? Where is this shown?\\n\\n- Science & results:\\n\\nWhile I see the appeal of the approach, I am not convinced that the authors are currently able to prove this. I am lacking many things: a proper description of the goal (I personally get confused about tissue-specific functions vs. GO, vs. prediction of the tissue,...), hard numbers on the targets (binary classification of GO terms: how many? which ones? Supplementary table),...\\n\\nIn my general confusion, I stumbled across a giant red flag for protein-related prediction tasks: the split of test and training sets. The authors describe this as a stratified random split. I sincerely hope that the stratification has NOT been made by looking at homology, and instead I wish the authors had discussed reduction of homology in their training and test sets, potentially picking very far related proteins between the different sets. Not to mention: I'm missing the size of test and train, and what \\\"stratification\\\" means in this context, how many samples per label,...\", \"dummy_vectors\": \"I'm not sure that the boundaries in the vector spaces of OhmNet and SecVeq are [-1, 1].\\n\\nThe authors at some point introduce a \\\"Ohmnet64-0Padded\\\" and \\\"Ohmnet64-Random64\\\" feature vectors, but I was unable to find any results for using these in the manuscripts.\", \"in_table_1\": \"I'm missing something that approximates an error estimate. I would have preferred to see the actual curves, in order to gaze an understanding about how their behavior in different conditions of sensitivity/specificity.\\n\\nAs mentioned above, \\\"spikes\\\" are mentioned, but no graph is presented. I'm missing the exact tissues for which each predictor was better (e.g. 6/13 and 4/13 --> hard numbers don't tell me if they are not overlapping).\\n\\nThese last considerations make me personally very suspicious about the results presented.\\n\\n- Conclusion\\n\\nI find the manuscript explores an interesting idea, but I *urge* the authors to rework the manuscript from top to bottom to (i) explain the background better, (ii) explain the problem they are trying to solve better, (iii) present the datasets, labels, classes better, and (iv) the results more clearly.\", \"title\": \"Interesting idea but needs major rework\"}" ] }

[ "Dogu Araci", "Zulkuf Genc" ]

[ "Financial sentiment analysis", "financial text classification", "transfer learning", "pre-trained language models", "BERT", "NLP" ]

{ "note_id": [ "Ykoc8KR8VD", "r1gkpwkhcH", "H1xnFZHVqH", "HkeWvWtCtr", "r1g2yTe2Yr" ], "note_type": [ "decision", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751327, 1572759479504, 1572258180266, 1571881305068, 1571716324197 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2529/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2529/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2529/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2529/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper presents FinBERT, a BERT-based model that is further trained on a financial corpus and evaluated on Financial PhraseBank and Financial QA. The authors show that FinBERT slightly outperforms baseline methods on both tasks.\\n\\nThe reviewers agree that the novelty is limited and this seems to be an application of BERT to financial dataset. There are many cases when it is okay to not present something entirely novel in terms of model as long as a paper still provides new insights on other things. Unfortunately, the new experiments in this paper are also not convincing. The improvements are very minor on small evaluation datasets, which makes the main contributions of the paper not enough for a venue such as ICLR.\\n\\nThe authors did not respond to any of the reviewers' concerns. I recommend rejecting this paper.\", \"title\": \"Paper Decision\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper described the application of BERT in the field of financial sentiment analysis. Authors find that when fine-tuned with in-domain data, BERT outperforms the state-of-the-art, demonstrating that language model pre-training can transfer knowledge learned from unsupervised large corpus to new domain with minimum effort. Experiments are conducted to explore 1) the utility of different in-domain dataset for further pre-training; 2) strategies to avoid catastrophic forgetting, and 3) effectiveness of fine-tuning a subset of the full model.\", \"i_am_in_favor_of_rejecting_this_paper_and_my_reasons_are_as_follows\": \"First, this paper may lack deeper innovation, although it demonstrates a good application of the BERT models in financial domain. For example, the framework of general-domain LM pretraining, to in-domain LM pretraining and finally in-domain classifier fine-tuning, as well as techniques of catastrophic forgetting were already proposed in Howard & Ruder 2018. Therefore, I think this paper may be more suitable for other (finance) application-oriented venues.\\n\\nSecond, the dataset used in evaluation is of small size (for example, Financial PhraseBank test set has one 1K). Thus, even though the paper is about transfer learning to domains without large data, I find it might be more convincing to draw a solid conclusion with a larger test set.\\n\\nThis paper is well organized and easy to follow. It may be beneficial to clarify in a few places (if space permits):\\n1) Some description or statistics of the data may be helpful (e.g., average sentence length or some examples);\\n2) Citations to Elmo and ULMFit can be made more explicit. Authors did cite Peters 2018 and Howard 2018 at the beginning of the paper, but may want to explicitly associate them with \\u2018Elmo\\u2019 and \\u2018ULMFit\\u2019 when these two terms first occur respectively;\\n3) For table 2, does the \\u2018all data\\u2019 or \\u2018data with 100% agreement\\u2019 include training data (80%) or just the test data (20%)?\\nThe difference between FinBERT(-domain) and ULMFit can be explicitly contrasted in the paper. Is the former initialized with BERT while latter with ULMFit?\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper presents a method for financial sentiment analysis based on the texts obtained from news. The method is based on an existing method BERT (Devlin et al. 2018). The authors have performed thorough experimental studies of the BERT method on an existing dataset TRC2-financial, a subset of TRC2 consisting of 1.8M news articles. Although the results may be of interest to communities working in this area, there are no or little novel contributions. By reading section 2 (only one page), which describes the method used, I have the impression that the authors took the method BERT and then applied this to the TRC2-financial dataset and then reported the results and also discussed some parameter choices in the BERT method. Therefore, the only value about this paper is the experimental results. Apart from this, there are no other contributions or insights to the methods/problems. In addition, section 2 is over-brief and very unclear, and it only contains a brief summary of the BERT method. For these reasons, I think the paper should be rejected for lacking novelty and writing quality.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2529\", \"review\": \"This paper presents an analysis of the BERT language model on financial text. FinBERT is evaluated on two datasets from the financial domain: a sentiment prediction dataset (classification with 3 different classes) and a sentiment score prediction (the score is a float number between -1 and 1).\\n\\nI find the phrasing \\\"FinBERT is a language model based on BERT\\\" misleading; I think FinBERT is BERT trained on financial text. There is no modification that is done to the original BERT model.\\n\\nThe paper presents several experiments using BERT as the language model and fine-tuning for the financial tasks. FinBERT is compared to a few baselines such as LSTMs with ElMO embeddings and ULMfit. I find interesting that the model performs better on the subset of the dataset for which there is perfect agreement between the annotators.\\n\\nI also find the results on training on financial data interesting. The results seem to indicate that further training on financial text does not seem to result in additional improvement when compared to original BERT.\\n\\nWhile I find the analysis and the experiments presented in the paper interesting, the novelty of the paper is rather low. There is no new idea introduced in this paper, it contains a series of experiments with BERT on financial text and tasks.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a domain adaptation type of task via proposing fine-tuning of pre-trained models such as BERT on data from financial domains. The paper starts off with a good motivation about requiring some kind of domain adaptation particularly when performing tasks such as sentiment analysis on data sets from the financial domain. However, there is not much novelty in this paper.\\n\\n1)The authors do not propose any new model architectures. Even if we were to argue the novelty is in terms of their empirical work, there are some flaws/missing details in the experiments.\\n2)In table 1 authors present agreement amongst annotators, it would be nice if in addition to mentioning the source of the data, the authors included what metric was used to attain agreement. I had to read the original paper releasing the data set to figure this out.\\n3)Table 4 presents results that do not seem significant. It is hard to conclude if a certain pre-training strategy worked for sure.\\n\\nOn the whole I am very lukewarm on this paper. I find this paper lacking in novelty. Seems like an ambitious class project turned into an ICLR submission.\"}" ] }

[ "Chieh-Hsin Lai", "Dongmian Zou", "Gilad Lerman" ]

[ "robust subspace recovery", "unsupervised anomaly detection", "outliers", "latent space", "autoencoder" ]

{ "note_id": [ "GQzgiTL8Y", "rJe8kNCjsr", "r1l1nslHiH", "BylbIigSiB", "rJxzJoxBsH", "rylGn9lSir", "S1gfPcgSiB", "ryl4G2X15S", "HkxvUjYsKS", "HyxIlEaNFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751297, 1573802973773, 1573354406904, 1573354313310, 1573354201836, 1573354154432, 1573354074008, 1571925004365, 1571687246616, 1571242989912 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2527/Authors" ], [ "ICLR.cc/2020/Conference/Paper2527/Authors" ], [ "ICLR.cc/2020/Conference/Paper2527/Authors" ], [ "ICLR.cc/2020/Conference/Paper2527/Authors" ], [ "ICLR.cc/2020/Conference/Paper2527/Authors" ], [ "ICLR.cc/2020/Conference/Paper2527/Authors" ], [ "ICLR.cc/2020/Conference/Paper2527/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2527/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2527/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"Three reviewers have assessed this paper and they have scored it as 6/6/6/6 after rebuttal. Nonetheless, the reviewers have raised a number of criticisms and the authors are encouraged to resolve them for the camera-ready submission.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to all reviewers with summary of changes\", \"comment\": \"Thanks again for the comments and suggestions.\\n\\nWe have modified our manuscript in order to address the concerns we responded to on November 10. The changes in the paper are indicated in red.\", \"the_major_revisions_are_as_follows\": [\"We clarified our contribution in Section 2.2. In particular, we distinguished our method from a direct application of RSR to a regular autoencoder (in response to R1 and R3). We also emphasized that our method is designed to handle the unsupervised setting and not the semi-supervised or supervised ones (in response to R2).\", \"We added the definition of AUC and AP in Appendix E and mentioned it in Section 4.2 (addressing R2).\", \"We added Section 5.1 in order to provide more intuition for RSRAE. We explained why an RSR layer is needed instead of a direct application of an $l_{2,1}$ within a regular autoencoder. A proposition for linear autoencoders with different loss functions (that include the cases of PCA and RSR) is proved in Appendix D.1. (Section 5.1 is motivated by various claims of R1, which we disagreed with in our earlier response. It tries to further clarify our work and avoid some misunderstandings).\", \"We added Appendix F, which compares RSRAE with (1) RSR (using both the FMS and SFMS algorithms) and (2) RCAE by Chalapathy, Menon and Chawla (2017). (This addressed a request by R1).\", \"We added Appendix G, which includes sensitivity analysis. Section G.1 addresses sensitivity to the intrinsic dimension (addressing R2 and R3). Section G.2 addresses sensitivity to the learning rate (addressing R3) and Section G.3 addresses sensitivity of RSRAE+ to the lambda\\u2019s (addressing R3).\", \"We added runtime comparison in Appendix H (addressing R3).\"]}", "{\"title\": \"Response to Review #3\", \"comment\": \"Thank you for your comments and constructive suggestions. We respond before the deadline, so we may communicate if needed.\\n\\n\\u201c\\nThe proposed approach is a simple combination of existing approaches.\\n\\u201d\", \"response\": \"When we tested our method on the deep features of Imagenet and the features of Reuters and 20 Newsgroups, we worked with a set of feature vectors. In those examples, we didn\\u2019t treat the features like images or sequences, and the network was simply a fully-connected network instead of a convolutional one. Please let us know if we missed anything.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thank you for your comments and constructive suggestions. We respond before the deadline, so we may communicate if needed.\\n\\n\\u201c\\n* There is a serious problem in the results (Figure 1) as the AP curves show better scores for larger corruption factors. Are the AP-score graphs flipped ? Please explain.\\n* The AUC and AP scores need to be defined. \\n\\u201d\", \"response\": \"We plan to add some experiments addressing the subspace dimension. We chose 10 because we were interested in a single parameter that may work for a wide range of real datasets, and we generally noticed some stability to changes of this dimension. However, we agree that we need to report the experiments that demonstrate this stability.\"}", "{\"title\": \"Response to Review #1 (part 3)\", \"comment\": \"\\u201cThe matrix A and the parameters of the AE are trained jointly. So, it can be seen that two processes can occur:\\n- The AE in order to minimize the reconstruction error would learn such latent space z, that would fit into the subspace of A^TA, so that projection \\\\tilde z =Az doesn\\u2019t cause data loss.\\n- The AE in order to minimize the reconstruction error would learn such A, so that the subspace that z approximates is the best possible.\\nIt is not clear, which of the two cases would take place. If the first one would dominate, then it is not clear if such method would have any discriminating capabilities.\\n\\u201d\", \"response\": \"RCAE in \\u201cRobust, Deep and Inductive Anomaly Detection\\u201d (ECML, 2017) is indeed another unsupervised method for anomaly detection. However, RCAE is related to RPCA, where one assumes entry-wise corruption of the data matrix. On the other hand, we discuss here the case where whole data points are outliers (that is, rows of the data matrix are corrupted), instead of few entries of a matrix are corrupted. Therefore, we don\\u2019t expect that RCAE will work well for our model of outliers. We can make few comparisons with RCAE in the appendix or supplemental material.\\n \\nAll the other methods you mentioned above are state-of-the-art for the semi-supervised setting (novelty detection), but here we emphasize a completely unsupervised setting (anomaly detection). Note that some of those methods (the ICLR 2018 and CVPR 2017 papers you mentioned) claim to be unsupervised anomaly detection / outlier detection but they are actually performing semi-supervised tasks, where one trains a model for the inliers. We already included GT from NIPS 2018 and DAGMM from ICLR 2018 and we do not find a reason to add more baseline methods, which are not unsupervised. \\n\\nAs for the claim that \\u201csome citations are missing\\u201d, we will be happy to cite whatever relevant references we missed. Please let us know specific relevant papers that we failed to cite.\"}", "{\"title\": \"Response to Review #1 (part 2)\", \"comment\": \"\\u201cIn addition to that, all the following discussion and proofs are limited to the linear case.\\u201d\", \"response\": \"We don\\u2019t agree that \\u201cautoencoders can potentially learn any, arbitrary entangled latent space\\u201d. The number of neurons in the autoencoder (and in particular, the dimension of the latent code) imposes a serious constraint. Actually, you mentioned that there is a bottleneck layer in a regular autoencoder. Due to this bottleneck layer, it is impossible to learn an autoencoder that produces zero error for all the data points. We try to focus only on the inliers and have a latent subspace for the inliers within the ambient space of the encoder. If outliers are mapped to this subspace, as you may worry, then these outliers will have huge reconstruction errors, which will easily help distinguish them and thus RSRAE should perform well in this case. That is, the answer to your point \\u201cit is not clear why outliers should necessarily have such embedding that is outside of the learned subspace\\u201d, is that there is no reason for the outliers to be outside the learned inliers\\u2019 subspace and it will be actually easier to distinguish them if they also happen to lie in this subspace. However, we wish to carefully learn such a subspace, while putting special emphasis on the inliers, instead of mapping all points to a low-dimensional subspace without recognizing properties of the inliers. The loss function of RSRAE does not care much about the outliers, but takes careful care of the inliers. It tries to carefully learn their structure and use it to distinguish between inliers and outliers.\"}", "{\"title\": \"Response to Review #1 (part 1)\", \"comment\": \"Thank you for your comments. We respond before the deadline, so we may communicate if needed.\\n\\n\\u201cThe paper claims to generalize the existing RSR framework to the nonlinear case.\\u201d\", \"response\": \"The linear layer is in the middle of an autoencoder and an RSR loss is part of the total loss function. It is not correct to think of two separate entities: an autoencoder and a linear RSR applied to the latent space of the autoencoder. That is, we do not try to first get the latent code form AE and then apply RSR to the latent code. Note that a general autoencoder tries to parametrize the structure of the data with a latent code within a low-dimensional linear space. Here we try to parametrize only the structure of the inliers, where the low-dimensional subspace lies within the ambient space of the encoder, we then map the output of the encoder onto a low-dimensional space corresponding to the latter subspace. This mapping is natural to the inliers, but not to the outliers, which are expected to have high reconstruction error.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"After reading all the reviews and the comments, I feel more positive about the paper. I appreciate the feedback of the Authors and I have decided to increase the rating.\\n\\n============================\\n\\nThe paper proposes using Robust Subspace Recovery in combination with an autoencoder (and possibly GANs) for anomaly detection. The encoder maps input data to the latent space of dimensionality D, which then is linearly projected to a subspace of dimensionality d (d < D). The projection of the latent space then goes to a decoder that reconstructs the input.\\nA transformation matrix A is trained jointly with the autoencoder. Two additional terms are added to the loss: one to encourage the subspace of A^TA to approximate the latent space z and the second one to force it to be an orthogonal projector.\\n \\nThe paper claims to generalize the existing RSR framework to the nonlinear case. However, the linear RSR is applied to the latent space of the autoencoder. In addition to that, all the following discussion and proofs are limited to the linear case. \\n\\nSince the proposed method is using RSB as it\\u2019s core part, and claims to be a non-linear extension of it, it would be crucial to have a comparison with RSB, at least on those experimental setups, where high-level features are used (Tiny Imagenet with ResNET features, Reuters-21578, and 20 Newsgroups). However, there is no such comparison.\\n\\nSince autoencoders can potentially learn any, arbitrary entangled latent space, it is not clear why outliers should necessarily have such embedding that is outside of the learned subspace. In the case of the original RSR it happens due to the dimensionality reduction by the orthogonal projector. However, autoencoders already perform dimensionality reduction at each layer down to the bottleneck layer.\\n\\nThe matrix A and the parameters of the AE are trained jointly. So, it can be seen that two processes can occur:\\n- The AE in order to minimize the reconstruction error would learn such latent space z, that would fit into the subspace of A^TA, so that projection \\\\tilde z =Az doesn\\u2019t cause data loss.\\n- The AE in order to minimize the reconstruction error would learn such A, so that the subspace that z approximates is the best possible.\\n\\nIt is not clear, which of the two cases would take place. If the first one would dominate, then it is not clear if such method would have any discriminating capabilities.\\n \\nMy point is mainly that the presented work is not really a generalization of RSR as it claims to be, but rather it is just using RSR on a leaned embedding of the data. \\n\\nSome citations are missing, as well as it is missing a comparison to some state-of-the art methods such as OCNN \\u2018Robust, Deep and Inductive Anomaly Detection\\u2019 ECML 2017; \\u2018Adversarially Learned One-Class Classifier for Novelty Detection\\u2019 CVPR 2017; DSVDD \\u2018Deep one-class classification.\\u2019 ICML, 2018; ODIN \\u2018Enhancing The Reliability Of Out-of-distribution Image Detection In Neural Networks\\u2019 ICLR 2018; \\u2018Generative Probabilistic Novelty Detection with Adversarial Autoencoders\\u2019 NeurIPS 2018.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper adapts the concept of Robust Subspace Recovery (RSR) as a layer in an auto-encoder model for anomaly detection. A loss function is proposed that combines reconstruction error and a regularizer that enforces robustness against outliers. The reconstruction error expresses the accuracy of the nonlinear dimensionality reduction imposed by the autoencoder. The regularizer is the sum of absolute deviations from the latent subspace that represents a linear structure robust against outliers. An alternative procedure is applied where the loss terms are applied iteratively during training. Once trained, the reconstruction error is used directly for anomaly detection with a threshold. The AUC is used as a performance measure. The method is compared against 6 other methods (LOF, OCSVM, IF, DESBM, GT, DAGMM). The setting is fully unsupervised, meaning that the training data contains various amounts of anomalies, and the results are parametrized with the amount of corruption. The results show that the proposed approach outperforms the other methods in most cases, especially for larger amounts of corruption. An ablation study compares the approach with auto-encoder-only and a non-alternating gradient descent (fixed factors for each part of the loss function) and shows that the alternating method outperfroms all by a wide margin.\", \"pros\": [\"A novel approach to fully unsupervised anomaly detection that beats the state of the art.\", \"The RSR layer is a simple fully connected layer and the loss function is simple to calculate, making the approach computationally efficient.\", \"A pseudo-code algorithm is provided in the appendix, which should help reproducibility.\", \"The paper is well written and the math is clearly laid out.\", \"The result benchmarks are sufficiently exhaustive in both the methods that are compared and the datasets used.\", \"The ablation study is informative and shows the effect of the regularization term of the loss function as well as the effect of alternating the gradient descent with the separate losses.\"], \"cons\": [\"There is a serious problem in the results (Figure 1) as the AP curves show better scores for larger corruption factors. Are the AP-score graphs flipped ? Please explain.\", \"The AUC and AP scores need to be defined.\", \"The results should include the case where the training data is not contaminated with outliers (c=0). This would correspond to the semi-supervised scenario and it would be very interesting to see how the method compares to DAGMM and GT which are build for that scenario.\", \"It would be interesting to see the effect of varying the subspace dimension. The authors chose 10 for all experiments, why is this number chosen, what would be the effect of choosing a smaller one ? This is a key parameter as it defines the structure of the projection subspace. Should this parameter be systematically tuned for each dataset ?\", \"Overall this is a good paper proposing a novel approach to fully unsupervised anomaly detection with state-of-the art results.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes to use the robust subspace recovery layer (RSR) in the autoencoder model for unsupervised anomaly detection.\\nThis paper is well written overall. Presentation is clear and it is easy to follow.\\nThe proposed approach is a simple combination of existing approaches.\\nAlthough its theoretical analysis with respect to the performance of anomaly detection is limited, experiments show that the proposed method is effective and superior to the existing anomaly detection methods.\", \"i_have_the_following_comments\": [\"Parameter sensitivity should be examined.\", \"The proposed method has the number of parameters including \\\\lambda_1, \\\\lambda_2, and parameters in neural networks.\", \"Since parameter tuning is fundamentally difficult in the unsupervised setting, the sensitivity of the proposed method with respect to changes of such parameters should be examined.\", \"Since the efficiency is also an important issue for anomaly detection methods, runtime comparison would be interesting.\", \"It would be also interesting whether the proposed method is also effective for non-structured data, where a dataset is given as just a set of (real-valued) feature vectors, with its comparison to the standard anomaly detection methods such as LOF and iForest.\"]}" ] }

[ "Tonghan Wang*", "Jianhao Wang*", "Chongyi Zheng", "Chongjie Zhang" ]

[ "Multi-agent reinforcement learning", "Nearly decomposable value function", "Minimized communication", "Multi-agent systems" ]

{ "note_id": [ "SJAwYKJaDC", "BkxAqrOhsr", "HJlk7y82ir", "HJgCxDMviS", "rJeQuBzPjH", "SJgHHSfPsr", "rylDKVfwiS", "HyxgLm1q9B", "SygnB7pe5S", "HJgpxkoRFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798751263, 1573844373941, 1573834519374, 1573492470397, 1573492075104, 1573492028731, 1573491839229, 1572627271980, 1572029251578, 1571888885153 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2526/Authors" ], [ "ICLR.cc/2020/Conference/Paper2526/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2526/Authors" ], [ "ICLR.cc/2020/Conference/Paper2526/Authors" ], [ "ICLR.cc/2020/Conference/Paper2526/Authors" ], [ "ICLR.cc/2020/Conference/Paper2526/Authors" ], [ "ICLR.cc/2020/Conference/Paper2526/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper2526/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2526/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper extends recent value function factorization methods for the case where limited agent communication is allowed. The work is interesting and well motivated. The reviewers brought up a number of mostly minor issues, such as unclear terms and missing implementation details. As far as I can see, the reviewers have addressed these issues successfully in their updated version. Hence, my recommendation is accept.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Clarifications of point 2 and 2.1\", \"comment\": \"Thank you for your feedback.\", \"for_point_2\": \"In our formulation, we use the action selection variable $A_j$ of agent $j$, which is deterministic only when the histories of *all* agents are given. However, in the definition of the mutual information $I(A_j; M_{ij} | \\\\mathrm{T}_j, M_{(-i)j})$, the history of agent $i$ *is not given*, and the action selection of agent $j$ is stochastic for its local history random variable $\\\\mathrm{T}_j$ and the message random variables $M_{(-i)j}$. Thus, $A_j$ is a random variable.\\n\\nFor point 2.1:\\nIn the value function factorization framework, optimal value functions of agents satisfy the Individual-Global-MAX (IGM) property:\\n\\n$$ argmax_{\\\\mathbf{a}} Q_{tot}(\\\\mathbf{\\\\tau}, \\\\mathbf{a})=\\n\\\\mathit{<}argmax_{a_1} Q_1(\\\\tau_1, a_1), argmax_{a_2} Q_2(\\\\tau_2, a_2), \\\\dots, argmax_{a_n} Q_n(\\\\tau_n, a_n) \\\\mathit{>}\\n$$\\n(please refer to Eq.4 in QMIX [1] and Eq. 1 in QTRAN [2]). All value function factorization methods fulfill this property (QMIX [1], QTRAN [2], and VDN [3]). Our framework is based on QMIX, so the IGM property is guaranteed by the monotonic mixing network.\\n\\n\\n\\n\\n[1] Rashid, T., Samvelyan, M., Witt, C.S., Farquhar, G., Foerster, J. and Whiteson, S., 2018, July. QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning. In International Conference on Machine Learning (pp. 4292-4301).\\n\\n[2] Son, K., Kim, D., Kang, W.J., Hostallero, D.E. and Yi, Y., 2019, May. QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning. In International Conference on Machine Learning (pp. 5887-5896).\\n\\n[3] Sunehag, P., Lever, G., Gruslys, A., Czarnecki, W.M., Zambaldi, V., Jaderberg, M., Lanctot, M., Sonnerat, N., Leibo, J.Z., Tuyls, K. and Graepel, T., 2018, July. Value-decomposition networks for cooperative multi-agent learning based on team reward. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems (pp. 2085-2087).\"}", "{\"title\": \"Final\", \"comment\": \"Thank you for your rebuttal and for partially addressing my comments. Point 2 and 2.1 are still in place. My assessment remains unchanged.\"}", "{\"title\": \"Manuscript updated\", \"comment\": \"Dear reviewers,\\n\\nWe have updated the manuscript to improve its presentation based on your comments. Thank you again for your feedback.\"}", "{\"title\": \"Thanks for your comments. We have run all the SC2 experiments with more random seeds. Here we provide explanations to clarify your questions.\", \"comment\": \"Thanks for your comments. Here we provide explanations to clarify your questions. In addition, please feel free to refer to our response to reviewer #1, which summarizes novelties of our paper.\", \"q\": \"The experimental results section is not well organized. The authors mention five questions on page 6, but it is not very clear which examples/set of experiments address which question.\", \"a\": \"Performance comparison and visualization results of didactic experiments can clarify all the questions. Our SC2 experiments further prove that the miscoordination problem of full decomposition methods is quite common and that our method can outperform QMIX and a state-of-the-art attentional communication algorithm.\\n \\nWe hope that our clarifications can address your questions. Please let us know if you have any other questions.\\n \\n[Foerster et al., AAAI 2018] Foerster, J.N., Farquhar, G., Afouras, T., Nardelli, N. and Whiteson, S., 2018, April. Counterfactual multi-agent policy gradients. In Thirty-Second AAAI Conference on Artificial Intelligence.\\n\\n[Rashid et al., ICML 2019] Rashid, T., Samvelyan, M., Witt, C.S., Farquhar, G., Foerster, J. and Whiteson, S., 2018, July. QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning. In International Conference on Machine Learning (pp. 4292-4301).\\n \\n[Singh et al., ICLR 2019] Singh, A., Jain, T. and Sukhbaatar, S., 2019. Learning when to communicate at scale in multiagent cooperative and competitive tasks. In International Conference on Learning Representations.\\n \\n[Das et al., ICML 2019] Das, A., Gervet, T., Romoff, J., Batra, D., Parikh, D., Rabbat, M. and Pineau, J., 2019, May. TarMAC: Targeted Multi-Agent Communication. In International Conference on Machine Learning (pp. 1538-1546).\"}", "{\"title\": \"Thanks for your comments and interest in our paper.\", \"comment\": \"Thanks for your comments and interest in our paper.\\n \\nRecently, the learning paradigm of value function factorization stands out as an effective approach for multi-agent deep reinforcement learning. [Samvelyan et al., AAMAS 2019] shows that its performance is much better than multi-agent policy gradients and independent value-based algorithm on StarCraft II micromanagement tasks.\\n \\nHowever, current value function factorization methods *fully* decompose the Q-functions among agents, and, as a result, each agent acts solely on its own local observations during execution. This full decomposition is not effective in many cases, because it is quite common that an agent\\u2019s optimal decision makings sometimes depend on information from other agents.\\n \\nTo address these limitations, in this paper, we develop a novel function factorization method that learns nearly decomposable value functions. Our model learns not only Q-functions for agents but also communication protocols for their improved coordination. Different from other communication approaches, we explicitly formalize novel optimization objectives for minimizing communication so that only useful and necessary information (those can reduce uncertainty in value functions of other agents) can be exchanged between agents. These novelties explain why our method can significantly improve the performance of QMIX on StarCraft II micromanagement tasks even when 80% of the messages are cut off while other communication methods, e.g., TarMAC [Das et al., ICML 2019], can not.\\n \\n[Samvelyan et al., AAMAS 2019] Samvelyan, M., Rashid, T., Schroeder de Witt, C., Farquhar, G., Nardelli, N., Rudner, T.G., Hung, C.M., Torr, P.H., Foerster, J. and Whiteson, S., 2019, May. The starcraft multi-agent challenge. In Proceedings of the 18th International Conference on Autonomous Agents and Multi-Agent Systems (pp. 2186-2188). International Foundation for Autonomous Agents and Multi-agent Systems.\\n \\n[Das et al., ICML 2019] Das, A., Gervet, T., Romoff, J., Batra, D., Parikh, D., Rabbat, M. and Pineau, J., 2019, May. TarMAC: Targeted Multi-Agent Communication. In International Conference on Machine Learning (pp. 1538-1546).\"}", "{\"title\": \"Thanks for your detailed and constructive comments. Here we provide explanations to clarify your questions.\", \"comment\": \"Thanks for your detailed and constructive comments. Here we provide explanations to clarify your questions.\", \"q1\": \"Precise definition of the message.\", \"a1\": \"A message $m_{ij}$ from agent $i$ to agent $j$ encodes necessary information of the local observation-action history of agent $i$, which is helpful for the decision making of agent $j$. Precisely, a message $m_{ij}$ is defined as a sample drawn from a multivariate Gaussian distribution, whose covariance is a unit matrix and whose mean is given by an encoder $f_m(\\\\tau_i, j;\\\\theta_c)$, where $\\\\tau_i$ is the local observation-action history of agent $i$ and $\\\\theta_c$ are parameters of the encoder $f_m$. As the reviewer suggested, we have refined and moved this definition to the beginning of Sec. 3 in the updated version of our paper.\", \"q2\": \"Confusion about \\\"optimal action policy of agent $j$\\\" in Sec 3.1.\", \"a2\": \"Sorry, this is a typo. It should be \\u201caction selection of agent $j$\\u201d. As shown in Fig. 1, the mutual information is defined between messages and action selection. The variable $A_j$ used in the definition of mutual information is a random variable, whose probability distribution describes how the actions are selected by agent $j$. As action selection may depend on other agents, action selection can be stochastic and thus the mutual information is still well-defined. We have added more descriptions in the updated version of our paper.\", \"q3\": \"Stability issue of our model.\", \"a3\": \"Like QMIX, during the training, our model is just one neural network (the message encoder can be regarded as some hidden layers), which is trained in an end-to-end fashion. Therefore, the message encoder will not introduce extra instability, and our model is faced with similar stability issue with QMIX. As in QMIX (and many other works in deep reinforcement learning), we use target networks, for both value functions and the message encoder, to stabilize training.\", \"q4\": \"The same history shape of agents limits the domains that our method can be applied.\", \"a4\": \"Like most algorithms in MARL, we share network structure and parameters among agents. Such sharing is feasible because in common MARL benchmarking environments, such as SC2 unit micromanagement [Samvelyan et al., AAMAS 2019] and MPE [Lowe et al., NeurIPS 2017], heterogeneous agents have observation and action spaces with the same shape. In our SC2 experiments, results on three tasks (1o2r_vs_4r, 1o10b_vs_1r, and MMM) verify that our approach works well with heterogeneous agents.\\n \\nWe hope that our clarification can address your questions. Please let us know if you have any other questions.\\n \\n[Samvelyan et al., AAMAS 2019] Samvelyan, M., Rashid, T., Schroeder de Witt, C., Farquhar, G., Nardelli, N., Rudner, T.G., Hung, C.M., Torr, P.H., Foerster, J. and Whiteson, S., 2019, May. The starcraft multi-agent challenge. In Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems (pp. 2186-2188). International Foundation for Autonomous Agents and Multiagent Systems.\\n \\n[Lowe et al., NeurIPS 2017] Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, O.P. and Mordatch, I., 2017. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems (pp. 6379-6390).\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper tackles the collaborative multi-agent RL problem as the problem of finding almost-decentralized value functions, where the loss tries to minimize communications between the agents. The core idea is around maximizing the mutual information between each massage and the existing knowledge of its receiver. This way, redundant messages are naturally removed. The authors then assert an entropy regularization to (almost) prevent the agents from *cheating*. The paper is in general well-written and motivated. There are however certain issues that should be addressed. [The second one is my main issue.]\\n\\n1. The term *message* has been used repeatedly but not defined. You should define precisely what you mean by a message in the background section. In particular, it is not clear from the text what a message looks like mathematically. You may also consider giving intuitive examples of how different ways of designing a message alters the behaviour in a given context before start talking about how to optimise them. \\n\\n2. Section 3.1 needs a revisit. I assume by \\\"optimal action policy of agent j\\\" you mean \\\"optimal policy of agent j\\\". Formally, an optimal policy is greedy to the optimal value function and is deterministic unless there exist multiple actions with same optimal value at a given state. Therefore, the mutual information is not mathematically well-defined since most of the time the optimal policy is deterministic and does not induce a probability distribution. \\n\\n2.1 What is the optimal value function of an agent? The agents are not optimizing their own value function, so even if your complex model converges, it does not imply optimality of agent-level value functions (and they shouldn\\u2019t be locally optimal). If by that you mean the agent-level value functions *after* convergence of $Q_tot$, then you need to clarify it to avoid the confusion. Even so, it is still not clear how you conceive the agent-level policies from these value functions. \\n\\n2.2. Minor: As for the notation, I found $A_j$ quite strange for a policy; you may want to consider something like $\\\\pi_j$. \\n\\n3. If $f_m$ is learnt, then message encoding is not stationary at least in the training time. It may make the training potentially become unstable. Specifically, there is no condition to assure stability of your model. Suggestion: your entropy regularization might be sufficient to induce stability. More discussion on this (or a simple experiment to show if an instability exists and will be alleviated with regularization) would be quite helpful. \\n\\n4. If I understand it correctly [and it wasn't clear from the text], the encoder $f_m$ is conditioned on local history, which induces that all the agents must share same history shape (tensor-wise), which in turn means that all the local agents has to have same local state shape (computationally, they should for example have same output/internal-state shape in their neural networks). This sounds like a limitation, and it may only be OK in domains where agents are homogenous.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Authors propose a new method for multi-agent reinforcement learning by using nearly decomposable value functions. The main idea is to have local (agent specific) value function and one global value function (for all agents). They also try minimizing the required communication need for multi-agent setup. To this end they deploy variational inference tools and support their method with an experimental study.\\n\\nI found the paper interesting and empirical evaluation is good. However, my knowledge in the field is quite limited.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": [\"The authors propose a framework for combining value function factorization and communication learning in a multi-agent setting by introducing two regularizers, one for maximizing mutual information between decentralized Q functions and communication messages and the other for minimizing the entropy of messages between agents. The authors also discuss a method for dropping non-informative messages. They illustrate their approach on sensor and hallway tasks and evaluate their method on the decentralized StarCraft II benchmark. The paper addresses an interesting problem, and the authors show that their approach gives good performance compared to alternative approaches even when a large percentage of communication is cut off between the agents.\", \"Questions/Comments:\", \"Implementation details (e.g., hyper-parameters, model size) are missing from the paper.\", \"The results are average over only 3 seeds, is this enough to compare different algorithms?\", \"How should beta should be determined?\", \"The authors present results when 80% of messages are cut off. What is the performance of the model when all communication is cut off for comparison?\", \"How does the approach work in competitive environments?\", \"The experimental results section is not well organized. The authors mention five question on page 6, but it is not very clear with examples/set of experiments address which question.\", \"There are many spelling/grammar errors in the paper.\"]}" ] }

[ "Angus Galloway", "Anna Golubeva", "Thomas Tanay", "Medhat Moussa", "Graham W. Taylor" ]

[ "batch normalization", "adversarial examples", "robustness" ]

{ "note_id": [ "42owlipPkB", "H1xR93zcoB", "S1xIO7xciH", "B1xtz9k9sB", "SyxyABk9ir", "rkgmZfkciB", "Sygo0J1ciH", "rkx6F6AFoS", "SJeMYKKyiS", "HkxOtjiG5r", "S1lNkzJ2YB", "rJg7I5MztH", "rJxofaHa_r", "HygzLHdjuS", "rygmCcRcOH", "BylPxYLmuH", "rkloiiNGuH", "SygA8UybOr", "HkeHazWxdr", "SklNaNu2vr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review", "official_comment", "comment", "comment", "official_comment", "comment", "official_comment", "comment" ], "note_created": [ 1576798751232, 1573690518397, 1573679981670, 1573677584730, 1573676486593, 1573675515327, 1573674963332, 1573674372879, 1572997497793, 1572154240346, 1571709403655, 1571068491488, 1570753810672, 1570633034141, 1570593482935, 1570101486663, 1570028451026, 1569941078094, 1569882813262, 1569649852225 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "~Rui_Wang1" ], [ "ICLR.cc/2020/Conference/Paper2525/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper2525/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper2525/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "~Anthony_Wittmer1" ], [ "~Antoine_Labatie1" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "~Antoine_Labatie1" ], [ "ICLR.cc/2020/Conference/Paper2525/Authors" ], [ "~Anthony_Wittmer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This article studies the effects of BN on robustness. The article presents a series of experiments on various datasets with noise, PGD adversarial attacks, and various corruption benchmarks, that show a drop in robustness when using BN. It is suggested that a main cause of vulnerability is the tiling angle of the decision boundary, which is illustrated in a toy example.\\nThe reviewers found the contribution interesting and that the effect will impact many DNNs. However, they the did not find the arguments for the tiling explanation convincing enough, and suggested more theory and experimental illustration of this explanation would be important. In the rebuttal the authors maintain that the main contribution is to link BN and adversarial vulnerability and consider their explanation reasonable. In the initial discussion the reviewers also mentioned that the experiments were not convincing enough and that the phenomenon could be an effect of gradient masking, and that more experiments with other attack strategies would be important to clarify this. In response, the revision included various experiments, including some with various initial learning schedules. The revision clarified some of these issues. However, the reviewers still found that the reason behind the effect requires more explanations. In summary, this article makes an important observation that is already generating a vivid discussion and will likely have an impact, but the reviewers were not convinced by the explanations provided for these observations.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Summary of changes\", \"comment\": \"Dear reviewers,\\n\\nThank you for your thought provoking reviews which have helped us improve the work. Here we briefly recap the main points expressed in the reviews, and where changes have been made to accommodate the concerns raised.\\n\\nReviews 2 and 3 are encouraged by the contribution and its significance, but with legitimate concerns about the \\\"gradient masking\\\" phenomenon (R3), and to what extent the results apply to adversarially trained models (R2).\", \"we_have_demonstrated_absence_of_gradient_masking_through_the_following_best_practices\": [\"Additive white Gaussian noise (AWGN), and common corruption benchmarks that don't require gradients\", \"Accuracy versus perturbation magnitude curves (Fig. 5, all go to zero accuracy)\", \"Unbounded white-box attacks all reach 100% success (fooling images, PGD, CWL2)\", \"Single step attacks perform worse than iterative attacks (wrt the attacker)\", \"Black-box transferability analysis on ImageNet (Table 12)\", \"Qualitative evidence, i.e., adversarial examples that contain semantic features\", \"Regarding adversarial training, we have added commentary and more detailed results showing the limitations of the PGD training approach on common corruptions.\", \"R1 was more sceptical, expressing concerns that the training procedure and effective learning rate may have confounded the results. R1 may not have seen some important details regarding the initial learning rate and schedules that were considered originally, but we can see how this could have been easily missed and have reorganized the section in question. Also, based on their feedback and after reading the linked Li et al., 19 work, we made some interesting discoveries while evaluating robustness on CIFAR-10 every ten epochs during training for different learning rate schedules, which shows more precisely at which times the baselines outperform their BN equivalent.\", \"To more easily parse through the changes, we provide a detailed changelog below:\", \"Re-organize Section 4 \\\"Empirical Results\\\" (per R1, R3), defer implementation details to Appendix.\", \"Section 5, clarify relevance of input dimension experiments (per R3)\", \"Discussed the work of Labatie 2019 where relevant in the Introduction and Section 3.\", \"Appendix B \\\"PGD Implementation Details\\\" centralize and clarify all implementation details related to PGD and preprocessing (per R2).\", \"Appendix D \\\"PGD Training Yields Unfortunate Robustness Trade-Offs\\\". Supplementary explanations and full breakdown of MNIST-C results with multiple runs (Table 9), showing that a reasonable baseline with similar clean test accuracy outperforms: BN, PGD, and PGD + BN.\", \"Appendix I \\\"Alternative Explanations of the Vulnerability\\\". We place the results of a concurrent submission in the context of our work and discussion of BN's numerical stability constant. Qualitatively and quantitatively shows what happens when we \\\"fold\\\" the BN statistics into the weights post-training.\", \"Appendix J \\\"Adversarial Examples\\\". We craft fooling images and Carlini & Wagner L2 examples on MNIST and SVHN, providing qualitative and quantitative evidence that BN degrades robustness.\", \"Appendix K \\\"On The Initial Learning Rate\\\". We place our results in the context of Li et al., showing that *at no time during training* does the *best robustness of BN* exceed that obtainable by an equivalent unnormalized model. The BN model's PGD (40-step) accuracy is consistently below 40%, whereas the baseline achieves 52%. We support these results with two standard tests of gradient masking by comparison with AWGN, and FGSM (per R3 concern). Perhaps most interesting is that the PGD accuracy of the BN models declines during training, whereas that of the baseline increases monotonically or plateaus. The use of early stopping is therefore much more important when BN is used, but this still does not recover the robustness of the unnormalized model.\", \"We sincerely hope to have the opportunity to engage in discussions with the reviewers prior to the platform closing in case anything remains unclear, and we thank them for their effort.\"]}", "{\"title\": \"Robustness generally decreases with many BN layers; effect is slower in residual networks\", \"comment\": \"Thank you for the question and kind words.\\n\\nThe experiment depicted in Figure 2 examines the impact of increasing the number of layers (where every layer is batch-normalized) from 10-60 in a vanilla feedforward network. There, we see consistency on train-ability with the upper bound from Yang et al., (2019), however in terms of robustness to perturbations/noise there appears to be a sweet spot around the point where accuracy begins to degrade @ ~20-30 layers. \\n\\nIn a separate experiment (not reported in the paper), we train constant width ($n=384$) ReLU networks of 1-10 layers ($L$) with vanilla GD (LR=0.1) on the 3 vs 7 testbed from Section 3. For BNGD over five random seeds, at $L=1$ we have for Clean, AWGN, PGD test sets respectively, absolute test accuracies of $98.4 \\\\pm 0.1\\\\%$, $69 \\\\pm 1\\\\%$, $15.1 \\\\pm 0.8\\\\%$. For $L=10$, this is reduced to $93.1 \\\\pm 0.2\\\\%$, $52.2 \\\\pm 0.8\\\\%$, and $0.05 \\\\pm 0.01\\\\%$. Without BN, at $L=1$ we have $98.0 \\\\pm 0.1\\\\%$, $84.7 \\\\pm 0.2\\\\%$, $50.7 \\\\pm 0.7\\\\%$, and for $L=10$, $98.0 \\\\pm 0.1$, $82 \\\\pm 2$, $40 \\\\pm 5\\\\%$.\\n\\nFor BN + SGD (mini batch size 50), the trend is similar but accuracy decays more slowly than for BNGD over the range of 1-10 layers. E.g. for the AWGN test accuracy, instead of dropping from $69 \\\\pm 1\\\\%$ to $52.2 \\\\pm 0.8\\\\%$, dropping from $70 \\\\pm 2\\\\%$ to $66 \\\\pm 4\\\\%$. The biggest factor over this range is between using BN or not, the unnormalized model consistently yields higher AWGN and PGD robustness by a difference of high 10s, and high 20s absolute percent, respectively.\\n\\nFor residual networks, Figure 4 of the paper evaluates the robustness of ResNet{20, 32, 44, 56, 110}, which shows that the robustness gap generally widens with depth between BN and fixed-update initialization (Fixup - Zhang et al., ICLR 2019) variants. Section 7 of Labatie (ICML 2019) sheds theoretical insight as to how the combined action of BN and skip-connections slows down the exploding gradients/sensitivity.\"}", "{\"title\": \"Response (part 2) - Implementation details and references\", \"comment\": \"\\\"Could you provide more implementation details about the adversarial training and attacker?\\\"\\n\\nThe meta-parameters used for the PGD training of the WideResNet 28-10 on CIFAR-10-C were $\\\\epsilon_{\\\\max} = 4/255$, 5 iterations, step size of 1. This was the model for which accuracy on the contrast CIFAR-10-C corruption is degraded by over 20% absolute for Fixup and BN variants. These details were also in a Figure caption, but we have made the connection more explicit in text. We also PGD train normal ResNets {32, 110} using $\\\\epsilon_{\\\\max}=8/255$, 7 iterations, and a step size of 2 (as in Madry et al., 18) and observe similar trends.\\n\\nFor PGD training on MNIST we use 20 iterations with a step size $\\\\epsilon / 10$.\\n\\n--\", \"re\": \"Question about 20 or 40 iterations of PGD for eval\\n\\nWe agree this wasn\\u2019t clear from our language. We report 20 iterations in most cases, but confirmed that 40 iterations did not significantly improve upon this, i.e., degrade the accuracy much further to within the measurement random error. For example, for VGG16 on CIFAR-10 evaluated using 40 iterations of PGD with a step size of $\\\\epsilon_\\\\infty / 20$, instead of 20 iterations with $\\\\epsilon_\\\\infty / 10$, reduced accuracy from $28.9 \\\\pm 0.2\\\\%$ to $28.5 \\\\pm 0.3\\\\%$, a difference of $0.4 \\\\pm 0.5\\\\%$.\\n\\nAlso, to err on the side of caution, we use 40 iterations of PGD for evaluation on CIFAR-10 in the new Appendix \\\"On The Initial Learning Rate\\\".\\n\\n--\\n\\nLastly, as suggested we have discussed the work of Labatie 2019 in several places where relevant. \\n\\nReferences\\n\\nYash Sharma and Pin-Yu Chen. Attacking the Madry Defense Model with $L_1$-based Adversarial Examples. In ICLR Workshop, 2018.\\n\\nLukas Schott, Jonas Rauber, Matthias Bethge, and Wieland Brendel. Towards the first adversarially robust neural network model on mnist. International Conference for Learning Representations, 2019.\"}", "{\"title\": \"Response (part 1) - on visualizations and PGD training\", \"comment\": \"Thank you for your feedback, and for taking the time to reproduce the main result on several architectures, this effort goes above and beyond, we appreciate it. We respond to each concern below.\\n\\n--\\n\\n\\\"Why do not visualize the decision boundary of networks (used in this work) to valid the boundary tilting \\\"theory\\\". The toy example is clear but not convincing enough. There exist several techniques may be helpful to the visualization. (i.e. Robustness via curvature regularization, and vice versa). I think it is one of the important parts of this work.\\\"\\n\\nAside from space limitations, we wanted to restrict visualizations to scenarios where they are faithful to the underlying model. Although such visualizations would be nice to have for further intuition, we believe the experiments provide sufficient evidence to support the main hypothesis. Also, beyond the toy model, we computed the boundary tilting angle of linear models w.r.t. the nearest centroid classifier on MNIST. Any dimensionality reduction technique used to visualize a high dimensional decision boundary introduces trade-offs that we didn\\u2019t feel were necessary to support the main message. We may opt to add such visualizations in a subsequent blog article as you suggest.\\n\\n--\\n\\n\\\"The observation of BN causes adversarial vulnerability is interesting, but the main focus should be offering more convincing explanations.\\\"\\n\\nThank you for stating that you believe the main finding of this work is interesting. We believe we have provided a reasonably accessible explanation, but wish to emphasize that our main contribution is that, to the best of our knowledge, this is the first work to explicitly link batch norm and adversarial vulnerability. This has major implications for state-of-the-art networks, and we believe this connection was previously unknown in the adversarial examples / robustness to distribution shift literature. Many laws of physics that we now take for granted were initially discovered via systematic experiment, and subsequently formalized by others. Other works, e.g., Yang et al., ICLR 2019, Labatie, ICML 2019 examine batch norm at length from a theoretical perspective which supports our conclusions, and a concurrent submission provides further insight as to why the vulnerability we discovered occurs. The official reviews of Yang et al., (https://openreview.net/forum?id=SyMDXnCcF7) expressed concern that their approach was not particularly intuitive or reflecting of popular practice, hence we made an effort to cast the limitations of BN in a practical light for practitioners and researchers of broad backgrounds to understand.\\n\\n--\\n\\n\\\"I do run experiments \\u2026 There exist ~20 ATA performance gaps between networks with BN and networks without BN. But for adversarial trained models, the gaps don't exist anymore, at least for VGG11,13,16,19 on cifar10 with PGD 3 attack. (The performance gap is less than 0.5). In other words, without the BN layers, the robustness of adversarially trained models will not increase in my experiments. I see you report some results in Appendix C. But it is not enough to convince me. If adversarial training can fix the vulnerability by BN and BN can give a TA boost, there is no reason we need to remove BN in our adversarial training setting.\\\"\\n\\nWe agree that the adversarial training results originally presented in Appendix C were perhaps a bit too informal. We have tightened up this section, including a full breakdown showing test accuracy and variance for each corruption, instead of simply stating the mean test accuracy over all corruptions.\\n\\nAs we motivated in the main text, it is imperative to consider robustness to unseen adversaries. Thus, it is unfair to benchmark the robustness of natural and adversarially trained networks using the same procedure, when one approach directly optimizes performance w.r.t. one of the evaluations. As you found, in some circumstances the performance degradation of BN seems small if we train on PGD and evaluate on the same, but this no longer holds if we consider other more realistic threat models and common corruptions. \\n\\nTo be more clear, we have renamed Appendix C to \\u201cPGD Training Yields Unfortunate Robustness Trade-offs\\u201d, as PGD can fail to yield a model with semantically meaningful gradients or convincingly broad robustness even for MNIST, due to the thresholding operation that is learned to favor $\\\\ell_{\\\\infty}$ robustness (see the discussion of Appendix E \\u201cMNIST Inspection\\u201d of Madry et al., (ICLR 2018)). These limitations have been widely discussed, e.g., Sharma & Chen., (ICLR Workshop 2018), Schott et al., (ICLR 2019), Jacobsen et al., (ICLR 2019, ICLR Workshop 2019), Mu & Gilmer., (ICML Workshop 2019) and are not too surprising given Goodhart\\u2019s law: \\u201cWhen a measure becomes a target, it ceases to be a good measure\\u201d. The \\u201cmeasure\\u201d in this case being the $\\\\ell_{\\\\infty}$ norm of the perturbations, which has become, somewhat arbitrarily, a focal point in the adversarial examples literature.\"}", "{\"title\": \"Response (part 2)\", \"comment\": \"\\u201cthere is nothing \\\"special\\\" nor unique about BN-networks; instead, the BN factorization merely permits accelerated training efficiency. Consider the fact that a BN model may be re-expressed by merely folding in the parameters (i.e. applying the matrix multiplications) into the MLP weights or CNN filters. Thus, the numerical function approximated by the BN and the \\\"folded\\\" non-BN model is identical.\\u201d\\n\\nAs outlined by existing theoretical work, training with BN is unique in several respects. To start, it imposes a hard upper limit on maximum trainable depth that is solely a function of the mini-batch size due to gradient explosion (Thms 3.9 & 3.10, Yang et al., 2019), and leads to exploding sensitivity wrt the input (S6 & Fig 4, Labatie et al., 2019).\", \"note\": \"in response to a concurrent submission which suggests that the use of tracked statistics are the source of adversarial vulnerability, we\\u2019ve added an Appendix H \\u201cAlternative Explanations of the Vulnerability\\u201d in which we isolate the effect of folding the BN statistics into the weights post-training on MNIST. The weights learned under BN, either with or without plugging in the tracked statistics at test time, differ substantially both qualitatively, and quantitatively in terms of boundary tilting angle wrt unnormalized weights.\\n\\n--\\n\\n\\u201cadversarial vulnerability predates BN. Likewise, non-BN models exhibit adversarial vulnerability. Thus, this title is not a great reflection of the findings of the paper. I would strongly suggest replacing \\\"is a cause\\\" with \\\"increases\\\" or \\\"exacerbates\\\". \\n\\nWe are happy to discuss the title of this work. The publication of adversarial examples predating that of batch normalization does not imply that batch normalization cannot act as one of many contributory causes of adversarial vulnerability. Qualitatively, we demonstrated that for MNIST, where a reasonably robust model can be obtained via per-image normalization and L2 regularization (see the adversarial examples of Appendix I which support this), the addition of BN in this case is sufficient to induce adversarial vulnerability. In terms of an apparent relationship between input dimension and vulnerability, which intuitively should not affect robustness as it does not affect the signal-to-noise ratio (Shafahi et al., ICLR 2019), the introduction of batch norm is again sufficient for this adverse relationship to exist. Ultimately, given that robustness for natural datasets is usually a trade-off, e.g., with accuracy (Tsipras et al., 2019), or between minimum and mean margin (Wu & Yu, 2019), we can accept your suggestion to replace \\\"is a cause\\\" with \\\"increases\\\" or \\\"exacerbates\\\".\\n\\nReferences\\n\\nDimitris Tsipras and Shibani Santurkar and Logan Engstrom and Alexander Turner and Aleksander Madry. Robustness May Be at Odds with Accuracy. In International Conference on Learning Representations, 2019.\\n\\nKaiwen Wu and Yaoliang Yu. Understanding Adversarial Robustness: The Trade-off between Minimum and Average Margin. In NeurIPS 19 Workshop on Machine Learning with Guarantees.\"}", "{\"title\": \"Response (part 1) - misc clarifications, new experiments on initial learning rate and folding BN statistics into weights\", \"comment\": \"We thank the reviewer for their frank comments and pointing us to the work on explaining the effect of using a large initial learning rate. Responding to this review has helped us improve the work, as well as our own understanding. We aim to satisfy the reviewer\\u2019s concerns with a new section titled \\u201cOn The Initial Learning Rate\\u201d in which we evaluate robustness on CIFAR-10 during training under various initial learning rates and schedules. In-line responses/clarifications to the review follow:\\n\\n--\\n\\n\\\"The authors demonstrate their results on SVHN, CIFAR-10, CIFAR-100 CIFAR-10.1\\\"\\n\\nTo clarify, we evaluated on CIFAR-{10, 10.1, 10-C}, but not on CIFAR-100. We also evaluated pre-trained models on ImageNet, as well as the Adversarial Spheres dataset (Gilmer et al., ICLR Workshop 2018) (albeit in an Appendix).\\n\\n--\\n\\n\\\"I do not know how much of the changes in adversarial vulnerability are due to batch normalization as opposed to other facets of the training procedure that may have changed in their BN vs no-BN experiments.\\\"\\n\\nNo aspect of the training procedure was changed for BN vs no-BN, unless explicitly stated otherwise, e.g., to allow BN a larger initial learning rate, see next bullet.\\n\\n--\\n\\n\\u201cbatch normalization usually accommodates higher learning rates and it is not clear if the authors adjusted the initial learning rate\\u201d\", \"quoting_from_the_paper_at_the_time_of_submission\": \"(In ref Table 4): \\u201cIt has been suggested that one of the benefits of BN is that it facilitates training with a larger learning rate (Ioffe & Szegedy, 2015; Bjorck et al., 2018). We test this from a robustness perspective in an experiment summarized in Table 4, where the initial learning rate is increased to 0.1 when BN is used.\\u201d\\n\\nTo prevent this from being missed and improve the clarity of our work, we can summarize the scenarios considered at the beginning of Section 4, e.g., \\u201cCase 1) both models get small LR\\u201d, \\u201cCase 2) BN gets high initial LR\\u201d. If the reviewer previously read this and still believes it to be unclear, please let us know.\\n\\nNote that we also provide a counterexample to the conventional wisdom that BN allows a higher learning rate in Appendix G on the Adversarial Spheres dataset.\\n\\n--\\n\\n\\\"By swapping in batch normalization, the authors may just be altering the norm of the weight change in the (re-parameterized) weights. In this scenario, the gains of removing batch normalization could just as well be explained by the effective change in the learning rate, and not about batch normalization itself, c.f. (Li et al., 2019).\\\"\\n\\nAs indicated by the two cases (on the synthetic and MNIST dataset) where we compute the boundary tilting angle, BN affects not only the norm (invariance) of the weights, but also the angle. Consider the first thing that happens when we feed forward a batch of inputs after random initialization: the weights are rescaled by the inverse of the standard deviation (and numerical stability const) along each dimension. Thus, the weights corresponding to low variance features increase in value, while those corresponding to high variance features shrink, and the resulting batch-normalized weight vectors can be nearly orthogonal to those without BN. Thus, the angle is a critical difference between batch-normalized and unnormalized weights. Although we motivated and explicitly characterized this for linear models, Section 6 of Labatie, (ICML 2019) shows that deep batch-normalized networks accordingly suffer from increased sensitivity w.r.t. the input as a result, and similarly remark: \\\"[under BN] directions of high signal variance are dampened, while directions of low signal variance are amplified. This preferential exploration of low signal directions naturally deteriorates the signal-to-noise ratio and amplifies $\\\\chi^l$\\\". Where $\\\\chi^l$ is defined as the normalized sensitivity from layer 0 to $l$, such that $\\\\chi^l > 1$ degrades the signal-to-noise ratio. \\n\\nWe found Li et al., 2019 and the learning order concept interesting. We agree that in the context of this work, it is imperative to consider the case where a higher initial learning rate is used with BN given that it does usually facilitate this. Please see Table 4, and the new section \\u201cOn The Initial Learning Rate\\u201d regarding this point which shows that over the course of 150 epochs of training using several learning rate schedules, BN obtains at most 40% accuracy to 40-step PGD while the baseline exceeds 50% accuracy on the same.\"}", "{\"title\": \"References\", \"comment\": \"References\\n\\nHuan Xu, Constantine Caramanis, and Shie Mannor. Robustness and Regularization of Support VectorMachines. In Journal of Machine Learning Research, 10:1485\\u20131510, 2009\\n\\nAnh Nguyen, Jason Yosinki, Jeff Clune. Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 427-436, 2015.\\n\\nYusuke Tsuzuku, Issei Sato, and Masashi Sugiyama. Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks. In Advances in Neural Information Processing Systems 31, pp. 6541-6650, 2018.\"}", "{\"title\": \"Clarification about gradient masking\", \"comment\": \"We thank the reviewer for recognising the positive aspects of this work, and for stating that they believe the work to be theoretically sound. We agree that it's imperative to ensure results are not tainted by gradient masking; we have taken care to ensure this did not play a role here and wish to alleviate this natural concern.\\n\\n--\", \"re\": \"\\u201cApply cw-l2 attack, and show batch norm has forced large perturbation.\\u201d\\n\\nThank you for this suggestion. We have added adversarial examples crafted by the CWL2 method in Figure 13 of an Appendix I titled \\u201cAdversarial Examples\\u201d. The L2 distortion required to achieve a fixed misclassification confidence threshold is 0.95 in the case of batch norm, and 2.89 for the baseline, which represents a three fold improvement on the relevant performance metric for this attack.\\n\\nThe white-box procedures would not work without access to a clean gradient signal, yet all reach 100% success, or confidence, in all cases. \\n\\nPlease let us know if these clarifications address your concerns. We agree that our work would be far less impactful if our result could be reduced to gradient masking. We hope that the additional experiments and analysis we have performed lays this concern to rest.\"}", "{\"title\": \"The number of batchnorm layers applied\", \"comment\": \"Really enjoyed reading your paper, inspired a lot! One question though.\\n\\nThere are models employed not only one but more layers of batch normalizations, have you considered making some evaluation on the relationship between the number of BNs used and the degradation (if any) seen in the same setting as you have demonstrated in the paper?\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"Overview:\\nThis is an interesting work. The paper is dedicated to studying the effect of BN to network robustness. The author shows that BN can reduce network robustness to small adversarial input perturbations and common corruptions by double-digit percentages. Then, they use a linear \\\"toy model\\\" to explain the mechanism that the actual cause is the tilting of the decision boundary. Moreover, the author conducts extensive experiments on popular datasets to show the robustness margin with or without the BN module. Finally, the author finds that substituting weight decay for BN is good enough to nullify a relationship between adversarial vulnerability and the input resolution.\", \"strength_bullets\": \"1. I like the linear toy example. For that binary classification example, the author explicitly explains the boundary tilting, which increases the adversarial vulnerability of the model. It is clear.\\n2. The paper conducts extensive experiment on SVHN, MNIST, CIFAR10 (C) datasets. And they show performance margin with or without the BN module. And for the attacker setting, they do use the popular setting (i.e. Mardy's PGD setting) in this field which makes the results more convincing.\", \"weakness_bullets\": \"1. Why do not visualize the decision boundary of networks (used in this work) to valid the boundary tilting \\\"theory\\\". The toy example is clear but not convincing enough. There exist several techniques may be helpful to the visualization. (i.e. Robustness via curvature regularization, and vice versa). I think it is one of the important parts of this work. The observation of BN causes adversarial vulnerability is interesting but the main focus should be offering more convincing explanations.\\n2. I do run experiments for VGG11,13,16,19 on cifar10 with PGD 3 attack (Mardy's setting). There exist ~20 ATA performance gaps between networks with BN and BN networks without BN. But for adversarial trained models, the gaps don't exist anymore, at least for VGG11,13,16,19 on cifar10 with PGD 3 attack. (The performance gap is less than 0.5). In other words, without the BN layers, the robustness of adversarially trained models will not increase in my experiments. I see you report some results in Appendix C. But it is not enough to convince me. Could you provide more implementation details about the adversarial training and attacker? And more experiment results about this point are needed. If adversarial training can fix the vulnerability by BN and BN can give a TA boost, there is no reason we need to remove BN in our adversarial training setting. I see there are similar concerns in the OpenReview.\\n3. [Minior] The experiments need to be organized better. Especially for section 3, it will be better to divide different experiments or observations into the different subsections.\", \"recommendations\": \"For the above weakness bullets, this is a week reject.\", \"suggestions\": \"1. To solve the weakness bullets;\\n2. minor suggestion: add the reference mention in the OpenReview, they are related to this work.\", \"questions\": \"1. You mention that you run PGD for 20-40 iterations in the experiment at the bottom of page three. But at each table, you only report one number. So my question is for that accuracy number, you run how many iterations for PGD?\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary:\\nIn this empirical study, the authors identify that batch normalization -- a common technique for accelerating training -- leads to brittle representations that exhibit a lack of robustness and are more susceptible to adversarial attacks. The authors demonstrate their results on SVHN, CIFAR-10, CIFAR-100 CIFAR-10.1 using a variety of network architectures including VGG, BagNet, WideResNet, AlexNet, etc.\", \"major_concerns\": \"1. As presented, the experiments are not convincing.\\n\\nI do not know how much of the changes in adversarial vulnerability are due to batch normalization as opposed to other facets of the training procedure that may have changed in their BN vs no-BN experiments.\\n\\nFor instance, batch normalization usually accommodates higher learning rates and it is not clear if the authors adjusted the initial learning rate, learning rate schedule or training schedule accordingly. If so, it would be important to run a set of experiments with these parameters fixed as per the baseline no-BN models. \\n\\nThat said, even if the authors did run these experiments, it is still not clear if the cause of adversarial vulnerability is due to BN. Consider that what is truly important in model training is not the learning rate (i.e. step size), but rather the magnitude of the changes in each weight (or the ratio of weight change to the weight). By swapping in batch normalization, the authors may just be altering the norm of the weight change in the (re-parameterized) weights. In this scenario, the gains of removing batch normalization could just as well be explained by the effective change in the learning rate, and not about batch normalization itself, c.f.\\n Towards Explaining the Regularization Effect of Initial Large Learning Rate in Training Neural Networks\\n Yuanzhi Li, Colin Wei, Tengyu Ma\", \"https\": \"//arxiv.org/abs/1907.04595\\nIf the differences in the adversarial vulnerability could be ascribed to effective changes in gradient updates, then this would change the interpretation of these results notably.\\n\\n2. The underlying hypothesis is specious.\\n\\nI have several reservations about the underlying hypothesis that requires stronger evidence to overcome. In particular, I have reservations in believing that BN itself is a cause of adversarial vulnerability because BN is just a factorization of a network's weights. That is, there is nothing \\\"special\\\" nor unique about BN-networks; instead, the BN factorization merely permits accelerated training efficiency.\\n\\nConsider the fact that a BN model may be re-expressed by merely folding in the parameters (i.e. applying the matrix multiplications) into the MLP weights or CNN filters. Thus, the numerical function approximated by the BN and the \\\"folded\\\" non-BN model is identical. What would it mean to say that the BN is \\\"causing\\\" adversarial vulnerability in the BN model given that both the BN and non-BN model perform the identical function?\\n\\nAnother way to say this is to pretend we train a non-BN MLP or CNN model. After training the model, we could apply a BN factorization of the weights. Thus, the non-BN model may be factorized into a BN model. If the resulting BN model were adversarial vulnerable (which I suspect is the case), it would seem very hard to believe that BN was the cause of the vulnerability given it was a post-hoc factorization of the weights.\\n\\nThat said, I could definitely imagine that the training procedure itself could lead to adversarial vulnerability (e.g. citation above) and by employing a BN factorization, one may be encouraged to use a training procedure which leads to increased vulnerability. I would encourage the authors to consider this line of attack and thus, re-orient their analysis and discussion accordingly.\\n\\n3. The title is poorly worded.\\n\\nNot withstanding the point above, adversarial vulnerability predates BN. Likewise, non-BN models exhibit adversarial vulnerability. Thus, this title is not a great reflection of the findings of the paper. I would strongly suggest replacing \\\"is a cause\\\" with \\\"increases\\\" or \\\"exacerbates\\\".\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper identifies an important weakness of batch normalization: it increases adversarial vulnerability. It is very well written and the claims are theoretically sound. In the experiments, the authors demonstrated a significant difference in robustness between networks with or without batch normalization layers, in varies settings against both random input noise and adversarial noise. This weakness of batch norm was explained due to the \\\"decision boundary tilting\\\" effect caused by the normalization. Overall, this paper has done solid work to reveal an interesting phenomenon. If it is true, this finding will impact almost all DNN models.\\n\\nMy concern is that this phenomenon is just another effect of \\\"gradient masking \\\" (as pointed out by Athalye, et al.). Batch norm is a well-known technique to avoid overfitting, without batch norm the network can be easily trained to be saturated with almost zero gradients, demonstrating a false signal of \\\"robustness\\\" to noise. The random noise and real-world corruption experiments are definitely helpful to clear this doubt, but only partially. My concern remains because of two obvious signs of gradient masking: \\n1. The accuracy on PGD-li (epsilon=0.031) attacks are suspiciously too high (20% - 40% Table 3/4). For this level of attack, the acc should be nearly zero. This is likely caused by the gradient masking effect, considering the cifar-10 networks were trained for longer time with larger learning rate (150 epochs, fixed lr 0.01). Training on MNIST is much easier to get zero gradients. \\n2. The weight decay discussion is not helpful at all, on the contrary, it confirms my concern on the gradient masking effect. In Table 8, the robustness was increased ~40% by just using large weight decay. This is not the \\\"real robustness\\\", and can be easily evaded by adaptive attack (see Athalye's paper).\\nWith the above two concerns in mind, I doubt the phenomenon revealed in this paper is just \\\"one can easily train a saturated model without batch norm\\\" or equivalently \\\"it's hard to train a saturated model with batch norm\\\". It is hard to say if this is a bad thing for batch norm.\\n\\nI am quite surprised that the authors ignore this completely. Here are a few things that can be done to rule out the possibility of gradient masking. The masked gradient can be identified by: 1) One-step attacks perform better than iterative attacks; 2) Unbounded attacks do not reach 100% success., etc (see Section 3.1 of Athalye's paper).\\n1. Including FGSM in the experiments and show the same trends as PGD-li. \\n2. Show two networks have similar gradient norms.\\n3. Apply cw-l2 attack, and show batch norm has forced large perturbation.\", \"two_other_suggestions\": \"1. Summarize the different angles/steps taken to verify the phenomenon, somewhere before the experiments.\\n2. Cannot see why the input dimension discussion contribute to explanations of the batch norm weakness.\\n\\n============\\nMy rating stays the same after rebuttal.\", \"my_original_concerns_are_like_the_other_reviewers\": \"why BN, not other techniques such as structure of DNNs MLP vs CNN vs ResNet, activation functions, weight decay, learning rates, softmax etc. My initial suspect was that it is caused by gradient masking likely caused by the l2 weight regularization, so asked the authors to look at the gradient norms and run some testes to rule this out. Yes, the weight norm is directly related to the Lipschitz continuity of the function represented by the network, but it often becomes more complicated on complex nonlinear neural networks.\\n\\nAccording to the new experiment results, the vulnerability is indeed not an effect of gradient masking, thanks for the clarification. However, the new results also indicate that the finding is susceptible to both weight decay and learning rate: in Figure 16 (a): \\\"Un PGD\\\" < \\\"BN PGD\\\" before learning rate decay, andFigure 17 (a) vs (b), doubling the weight decay penalty to 1e-3 also increases the vulnerability of BN. Overall, I believe the phenomenon exists, but the reasons behind requires more explanations, at least not just the batch norm.\"}", "{\"comment\": \"Thanks for catching this, VGG8 is indeed a non-standard architecture. To reduce capacity, we remove layers from VGG11 such that there is only one convolution layer between each 2x2 max pooling (indicated 'M').\", \"vgg8\": \"[64, 'M', 128, 'M', 256, 'M', 512, 'M', 512, 'M']\", \"vgg11\": \"[64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M']\", \"title\": \"Architecture\"}", "{\"comment\": \"What is VGG8?\\n\\nI did not find any clue about the VGG version with 8 weight layers in the paper of VGG.\", \"title\": \"Minor question\"}", "{\"comment\": \"I also believe that the discussion from the ICML 2019 paper on the effect of BN is very relevant. Thank you very much for your quick response.\", \"title\": \"Thank you for your response\"}", "{\"comment\": \"Yes, your interpretation seems valid, although we do not exactly make this argument in the text. This is worth formalizing and could make our study that considers the input dimension and vulnerability more precise for the case with BN.\\n\\nWe're sorry to have missed the reference you mention from ICML 2019. On first pass, it is indeed highly relevant to our discussion in Section 3, particularly how BN reduces signal-to-noise in Section 6, where you mention \\\"directions of high signal variance are dampened, while directions of low signal variance are amplified\\\". We're looking forward to discussing theses connections in our next revision.\", \"title\": \"Thank you for pointing us to this work\"}", "{\"comment\": \"Hi,\\n\\nI enjoyed reading your paper. I just have two comments.\\n\\nFirst, the discussion and the experiments of Section 3 are very interesting. Do you think that the following interpretation would be valid:\\n- Since unnormalized MNIST images \\\"live\\\" in a subspace of small dimensionality, many directions do not play any role in the max-margin solution\\n- On the other hand, normalized MNIST images \\\"live\\\" in a subspace of larger dimensionality, so that many more dimensions play a role in the\\u00a0max-margin solution\\n\\nSecond, I wanted to mention a closely related paper from ICML 2019 precisely showing that\\u00a0batch normalization causes exploding sensitivity to input perturbations at initialization [1].\\n\\n[1] Characterizing Well-Behaved vs. Pathological Deep Neural Networks. ICML 2019.\", \"title\": \"Two comments\"}", "{\"comment\": \"1. Thanks for pointing us to these relevant concurrent submissions. More discussion follows.\\n\\n2. Regarding \\\"Adversarially trained models also have the adversarial vulnerability for strong attack\\\", I'm sorry but I don't quite understand the connection to our claim about BatchNorm (BN). All models are vulnerable to strong attacks if \\\"strong\\\" means large perturbation budget. Our claim is simply that using BN makes models more vulnerable. We report results for PGD trained models on MNIST and CIFAR-10 in Appendix C. We take a broad view of robustness, but considering the limited \\\"max-norm\\\" attack model we observe a non-trivial improvement in robustness using Fixup in place of BN for the same WideResNet as in Madry et al. One of the concurrent submissions you reference that aims to fix the vulnerability induced by BN reports an even larger 14% gap in terms of PGD test accuracy for PGD trained vanilla ResNets.\\n\\nAs an aside, I stress that the current form of PGD training with a one-size-fits all epsilon (i.e. the same constant is applied to all examples regardless of their intrinsic noise) can yield excessive invariance that reduces robustness more broadly, see e.g., Jacobsen et al., (2019), Mu & Gilmer, (2019). In the case where epsilon is adapted to each training example, the procedure is equivalent to penalizing the parameters' dual norm for linear models (Xu et al., (2009)). We perform several experiments on the interaction of BN with parameter norm regularization in the text.\\n\\n3. We believe it would be valuable to study the effect of these other normalization layers on robustness and would be excited if someone else does so. Due to widespread use of BN and space constraints, these are out of scope for the present submission. \\n\\nWe tested the hypothesis that the vulnerability of BN could arise from using the tracked mean and variance of the training data at test time. This is an insightful observation, indeed not using these statistics at test time does increases robustness, but the same can be achieved through increasing the numerical stability constant (while preserving the ability to test with arbitrary batch sizes). Neither of these mechanisms fully account for the vulnerability, due to decision boundary tilting inherent in the normalization procedure itself. We've added an Appendix to discuss this alternate hypothesis which will become available when updates to all papers are enabled.\", \"title\": \"Thank you for the comments\"}", "{\"comment\": \"Hi, it is an interesting work.\\n\\n1. In intuition, the mean and the variance in the BN layer are strongly related with the clean training data, which leads to the unsuitable normalization to the adversarial testing data. I see two related submission, namely https://openreview.net/forum?id=HyxJhCEFDS and https://openreview.net/forum?id=BJlEEaEFDS&noteId=BylDR1y3Pr , where the authors propose MBN to disentangle the distribution for clean and adversarial data to estimate normalization statistics.\\n\\n2. Adversarially trained models also have the adversarial vulnerability for strong attack. So I have a little doubt whether batch normalization is a cause of adversarial vulnerability for adversarially trained models. In other words, without the BN layers, the robustness of adversarially trained models will increase? In my opinion, the answer is \\\"no\\\", because a stronger neural network architecture can help to increase the robustness of the adversarially trained model.\\n\\n3. Have the authors tried other normalization layers, which do not need to store the mean and the variance of training data, such as instance normalization, layer normalization and so on?\", \"title\": \"Nice work and some questions\"}" ] }