note_id
stringlengths 9
12
| forum_id
stringlengths 9
13
| invitation
stringlengths 40
95
| content
stringlengths 44
35k
| type
stringclasses 1
value | year
stringclasses 7
values | venue
stringclasses 171
values | paper_title
stringlengths 0
188
| paper_authors
stringlengths 2
1.01k
| paper_abstract
stringlengths 0
5k
| paper_keywords
stringlengths 2
679
| forum_url
stringlengths 41
45
| pdf_url
stringlengths 39
43
| review_url
stringlengths 58
64
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
H11UL9lVg | BJluGHcee | ICLR.cc/2017/conference/-/paper242/official/review | {"title": "The ideas are brilliant, but technical typos exist", "rating": "5: Marginally below acceptance threshold", "review": "The paper provides an interesting use of generative models to address the classification with missing data problem. The tensorial mixture models proposed take into account the general problem of dependent samples. This is an nice extension of current mixture models where samples are usually considered as independent. Indeed the TMM model is reduced to the conventional latent variable models. As much as I love the ideas behind the paper, I feel pitiful about the sloppiness of the presentation (such as missing notations) and flaws in the technical derivations. Before going into the technical details, my high level concerns are as follows:\n(1) The joint density over all samples is modeled as a tensorial mixture generative model. The interpretation of the CP decomposition or HT decomposition on the prior density tensor is not clear. The authors have an interpretation of TMM as product of mixture models when samples are independent, however their interpretation seems flawed to me, and I will elaborate on this in the detailed technical comments below. \n(2) The authors employ convolution operators to compute an inner product. It is realizable by zero padding, but the invariance structure, which is the advantage of CNN compared to feed-forward neural network, will be lost. However, I am not sure how much this would affect the performance in practice. \n(3) The author could comment in the paper a little bit on the sample complexity of this method given the complexity of the model. \n\nBecause I liked the ideas of the paper so much, and the ICLR paper submitted didn't present the technical details well due to sloppiness of notations, so I read the technical details in the arXiv version the authors pointed out. There are a few technical typos that I would like to point out (my reference to equations are to the ones in the arXiv paper). \n(1) The generative model as in figure (5) is flawed. P(x_i|d_i;\\theta_{d_i}) are vectors of length s, there the product of vectors is not well defined. It is obvious that the dimensions of the terms between two sides of the equation are not equal. In fact, this should be a tucker decomposition instead of multiplication. It should be P(X) = \\sum_{d1,\\ldots,d_N} P(d_1,\\ldots,d_N) (P(x_1|d_1;theta_{d_1},P(x_2|d_2;theta_{d_2},\\ldots,P(x_N|d_N;theta_{d_N}), which means a sum of multi-linear operation on tensor P(d_1,\\ldots,d_N), and each mode is projected onto P(x_i|d_i;theta_{d_i}. \n(2) I suspect the special case for diagonal Gaussian Mixture Models has some typos as I couldn't derive the third last equation on page 6. But it might be just I didn't understand this example. \n(3) The claim that TMM reduces to product of mixture model is not accurate. The first equation on page 7 is only right when \"sum of product\" operation is equal to \"product of sum\" operation. Similarly, in equation (6), the second equality doesn't hold unless in some special cases. However, this is not true. This might be just a typo, but it is good if the authors could fix this. I also suspect that if the authors correct this typo,the performance on MNIST might be improved.\n\nOverall, I like the ideas behind this paper very much. I suggest the authors fix the technical typos if the paper is accepted. ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Tensorial Mixture Models | ["Or Sharir", "Ronen Tamari", "Nadav Cohen", "Amnon Shashua"] | We introduce a generative model, we call Tensorial Mixture Models (TMMs) based on mixtures of basic
component distributions over local structures (e.g. patches in an image) where the dependencies between
the local-structures are represented by a "priors tensor" holding the prior probabilities of assigning a
component distribution to each local-structure.
In their general form, TMMs are intractable as the priors tensor is typically of exponential size. However,
when the priors tensor is decomposed it gives rise to an arithmetic circuit which in turn transforms the
TMM into a Convolutional Arithmetic Circuit (ConvAC). A ConvAC corresponds to a shallow (single hidden layer)
network when the priors tensor is decomposed by a CP (sum of rank-1) approach and corresponds to a
deep network when the decomposition follows the Hierarchical Tucker (HT) model.
The ConvAC representation of a TMM possesses several attractive properties. First, the inference is tractable
and is implemented by a forward pass through a deep network. Second, the architectural design of the model
follows the deep networks community design, i.e., the structure of TMMs is determined by just two easily
understood factors: size of pooling windows and number of channels. Finally, we demonstrate the effectiveness
of our model when tackling the problem of classification with missing data, leveraging TMMs unique ability of
tractable marginalization which leads to optimal classifiers regardless of the missingness distribution. | ["Deep learning", "Supervised Learning", "Unsupervised Learning"] | https://openreview.net/forum?id=BJluGHcee | https://openreview.net/pdf?id=BJluGHcee | https://openreview.net/forum?id=BJluGHcee¬eId=H11UL9lVg |
BktskJMEg | BJluGHcee | ICLR.cc/2017/conference/-/paper242/official/review | {"title": "Interesting approach to generative models and missing data", "rating": "7: Good paper, accept", "review": "This paper uses Tensors to build generative models. The main idea is to divide the input into regions represented with mixture models, and represent the joint distribution of the mixture components with a tensor. Then, by restricting themselves to tensors that have an efficient decomposition, they train convolutional arithmetic circuits to generate the probability of the input and class label, providing a generative model of the input and labels. \n\nThis approach seems quite elegant. It is not completely clear to me how the authors choose the specific architecture for their model, and how these choices relate to the class of joint distributions that they can represent, but even if these choices are somewhat heuristic, the overall framework provides a nice way of controlling the generality of the distributions that are represented.\n\nThe experiments are on simple, synthetic examples of missing data. This is somewhat of a limitation, and the paper would be more convincing if it could include experiments on a real-world problem that contained missing data. One issue here is that it must be known which elements of the input are missing, which somewhat limits applicability. Could experiments be run on problems relating to the Netflix challenge, which is the classic example of a prediction problem with missing data? In spite of these limitations, the experiments provide appropriate comparisons to prior work, and form a reasonable initial evaluation.\n\nI was a little confused about how the input of missing data is handled experimentally. From the introductory discussion my impression was that the generative model was built over region patches in the image. This led me to believe that they would marginalize over missing regions. However, when the missing data consists of IID randomly missing pixels, it seems that every region will be missing some information. Why is it appropriate to marginalize over missing pixels? Specifically, $x_i$ in Equation 6 represents a local region, and the ensuing discussion shows how to marginalize over missing regions. How is this done when only a subset of a region is missing? It also seems like the summation in the equation following Equation 6 could be quite large. What is the run time of this? \n\nThe paper is also a bit schizophrenic about the extent to which the results are applicable beyond images. The motivation for the probabilistic model is mostly in terms of images. But in the experiments, the authors state that they do not use state-of-the-art inpainting algorithms because their method is not limited to images and they want to compare to methods that are restricted to images. This would be more convincing if there were experiments outside the image domain.\n\nIt was also not clear to me how, if at all, the proposed network makes use of translation invariance. It is widely assumed that much of the success of CNNs comes from their encoding of translation invariance through weight sharing. Is such invariance built into the authors\u2019 network? If not, why would we expect it to work well in challenging image domains?\n\nAs a minor point, the paper is not carefully proofread. To just give a few examples from the first page or so:\n\n\u201csignificantly lesser\u201d -> \u201csignificantly less\u201d\n\n\u201cthe the\u201d\n\n\u201cprovenly\u201d -> provably\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Tensorial Mixture Models | ["Or Sharir", "Ronen Tamari", "Nadav Cohen", "Amnon Shashua"] | We introduce a generative model, we call Tensorial Mixture Models (TMMs) based on mixtures of basic
component distributions over local structures (e.g. patches in an image) where the dependencies between
the local-structures are represented by a "priors tensor" holding the prior probabilities of assigning a
component distribution to each local-structure.
In their general form, TMMs are intractable as the priors tensor is typically of exponential size. However,
when the priors tensor is decomposed it gives rise to an arithmetic circuit which in turn transforms the
TMM into a Convolutional Arithmetic Circuit (ConvAC). A ConvAC corresponds to a shallow (single hidden layer)
network when the priors tensor is decomposed by a CP (sum of rank-1) approach and corresponds to a
deep network when the decomposition follows the Hierarchical Tucker (HT) model.
The ConvAC representation of a TMM possesses several attractive properties. First, the inference is tractable
and is implemented by a forward pass through a deep network. Second, the architectural design of the model
follows the deep networks community design, i.e., the structure of TMMs is determined by just two easily
understood factors: size of pooling windows and number of channels. Finally, we demonstrate the effectiveness
of our model when tackling the problem of classification with missing data, leveraging TMMs unique ability of
tractable marginalization which leads to optimal classifiers regardless of the missingness distribution. | ["Deep learning", "Supervised Learning", "Unsupervised Learning"] | https://openreview.net/forum?id=BJluGHcee | https://openreview.net/pdf?id=BJluGHcee | https://openreview.net/forum?id=BJluGHcee¬eId=BktskJMEg |
SkGQ_LwEl | BJluGHcee | ICLR.cc/2017/conference/-/paper242/official/review | {"title": "Interesting approach but ...", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes a generative model for mixtures of basic local structures where the dependency between local structures is a tensor. They use tensor decomposition and the result of their earlier paper on expressive power of CNNs along with hierarchical Tucker to provide an inference mechanism. However, this is conditioned on the existence of decomposition. The authors do not discuss how applicable their method is for a general case, what is the subspace where this decomposition exists/is efficient/has low approximation error. Their answer to this question is that in deep learning era these theoretical analysis is not needed. While this claim is subjective, I need to emphasize that the paper does not clarify this claim and does not mention the restrictions. Hence, from theoretical perspective, the paper has flaws and the claims are not justified completely. Some claims cannot be justified with the current results in tensor literature as the authors also mentioned in the discussions. Therefore, they should have corrected their claims in the paper and made the clarifications that this approach is restricted to a clear subclass of tensors.\n\nIf we ignore the theoretical aspect and only consider the paper from empirical perspective, the experiments the appear in the paper are not enough to accept the paper. MNIST and CIFAR-10 are very simple baselines and more extensive experiments are required. Also, the experiments for missing data are not covering real cases and are too synthetic. Also, the paper lacks the extension beyond images. Since the authors repeatedly mention that their approach goes beyond images, and since the theory part is not complete, those experiments are essential for acceptance of this paper.\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Tensorial Mixture Models | ["Or Sharir", "Ronen Tamari", "Nadav Cohen", "Amnon Shashua"] | We introduce a generative model, we call Tensorial Mixture Models (TMMs) based on mixtures of basic
component distributions over local structures (e.g. patches in an image) where the dependencies between
the local-structures are represented by a "priors tensor" holding the prior probabilities of assigning a
component distribution to each local-structure.
In their general form, TMMs are intractable as the priors tensor is typically of exponential size. However,
when the priors tensor is decomposed it gives rise to an arithmetic circuit which in turn transforms the
TMM into a Convolutional Arithmetic Circuit (ConvAC). A ConvAC corresponds to a shallow (single hidden layer)
network when the priors tensor is decomposed by a CP (sum of rank-1) approach and corresponds to a
deep network when the decomposition follows the Hierarchical Tucker (HT) model.
The ConvAC representation of a TMM possesses several attractive properties. First, the inference is tractable
and is implemented by a forward pass through a deep network. Second, the architectural design of the model
follows the deep networks community design, i.e., the structure of TMMs is determined by just two easily
understood factors: size of pooling windows and number of channels. Finally, we demonstrate the effectiveness
of our model when tackling the problem of classification with missing data, leveraging TMMs unique ability of
tractable marginalization which leads to optimal classifiers regardless of the missingness distribution. | ["Deep learning", "Supervised Learning", "Unsupervised Learning"] | https://openreview.net/forum?id=BJluGHcee | https://openreview.net/pdf?id=BJluGHcee | https://openreview.net/forum?id=BJluGHcee¬eId=SkGQ_LwEl |
SkSAkQlHe | H1kjdOYlx | ICLR.cc/2017/conference/-/paper107/official/review | {"title": "Potentially interesting approach but novelty and utility are not clear at this a point", "rating": "4: Ok but not good enough - rejection", "review": "The paper presents an approach to learning shared neural representations of temporal abstractions in hierarchical RL, based on actor-critic methods. The approach is illustrated in two tasks: gridworld with objects and a simplified Minecraft problem). The idea of providing symbolic descriptions of tasks and learning corresponding \"implementations\" is potentially interesting and the empirical results are promising. However, there are two main drawbacks of the current incarnation of this work. First, the ideas presented in the paper have all been explored in other work (symbolic specifications, actor-critic, shared representations). While related work is discussed, it is not really clear what is new here, and what is the main contribution of this work besides providing a new implementation of existing ideas in the context of deep learning. The main contribution if the work needs to be clearly spelled out. Secondly, the approach presented relies crucially on curriculum learning (this is quite clear from the experiments). While the authors argue that specifying tasks in simplified language is easy, designing a curriculum may in fact be pretty complicated, depending on the task at hand. The examples provided are fairly small, and there is no hint of how curriculum can be designed for larger problems. Because the approach is sensitive to the curriculum, this limits the potential utility of the work. It is also unclear if there is a way to provide supervision automatically, instead of doing it based on prior domain knowledge.\nMore minor comments:\n- The experiments are not described in enough detail in the paper. It's great to provide github code, but one needs to explain in the paper why certain choices were made in the task setup (were these optimized? What's this the first thing that worked?) Even with the code, the experiments as described are not reproducible\n- The description of the approach is pretty tangled with the specific algorithmic choices. Can the authors step back and think more generally of how this approach can be formalized? I think this would help relate it to the prior work more clearly as well.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Modular Multitask Reinforcement Learning with Policy Sketches | ["Jacob Andreas", "Dan Klein", "Sergey Levine"] | We describe a framework for multitask deep reinforcement learning guided by
policy sketches. Sketches annotate each task with a sequence of named subtasks,
providing high-level structural relationships among tasks, but not providing the
detailed guidance required by previous work on learning policy abstractions for
RL (e.g. intermediate rewards, subtask completion signals, or intrinsic motivations).
Our approach associates every subtask with its own modular subpolicy,
and jointly optimizes over full task-specific policies by tying parameters across
shared subpolicies. This optimization is accomplished via a simple decoupled
actor–critic training objective that facilitates learning common behaviors from
dissimilar reward functions. We evaluate the effectiveness of our approach on a
maze navigation game and a 2-D Minecraft-inspired crafting game. Both games
feature extremely sparse rewards that can be obtained only after completing a
number of high-level subgoals (e.g. escaping from a sequence of locked rooms or
collecting and combining various ingredients in the proper order). Experiments
illustrate two main advantages of our approach. First, we outperform standard
baselines that learn task-specific or shared monolithic policies. Second, our
method naturally induces a library of primitive behaviors that can be recombined
to rapidly acquire policies for new tasks. | ["Reinforcement Learning", "Transfer Learning"] | https://openreview.net/forum?id=H1kjdOYlx | https://openreview.net/pdf?id=H1kjdOYlx | https://openreview.net/forum?id=H1kjdOYlx¬eId=SkSAkQlHe |
SyJDbsb4e | H1kjdOYlx | ICLR.cc/2017/conference/-/paper107/official/review | {"title": "Review", "rating": "3: Clear rejection", "review": "The paper proposes a new RL architecture that aims at learning policies from sketches i.e sequence of high-level operations to execute for solving a particular task. The model relies on a hierarchical structure where the sub-policy is chosen depending on the current operation to execute in the sketch . The learning algorithm is based on an extension of the actor-critic model for that particular case, and also involves curriculum learning techniques when the task to solve is hard. Experimental results are provided on different learning problems and compared to baseline methods. \n\nThe paper is well-written and very easy to follow. I am not really convinced by the impact of such a paper since the problem solved here can be seen as an option-learning problem with a richer supervision (i.e the sequence of option is given). It thus corresponds to an easier problem with a limited impact. Moreover, I do not really understand to which concrete application this setting corresponds. For example, learning from natural langage instructions is clearly more relevant. So since the model proposed in this article is not a major contribution and shares many common ideas with existing hierarchical reinforcement learning methods, the paper lacks a strong motivation and/or concrete application. So, the paper only has a marginal interest for the RL community\n\n@pros: \n* Original problem with well design experiments\n* Simple adaptation of the actor-critic method to the problem of learning sub policies\n\n\n@cons:\n* Very simple task that can be seen as a simplification of more complex problems like options discovery, hierarchical RL or learning from instructions\n* No strong underlying applications that could help to 'reinforce' the interest of the approach\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Modular Multitask Reinforcement Learning with Policy Sketches | ["Jacob Andreas", "Dan Klein", "Sergey Levine"] | We describe a framework for multitask deep reinforcement learning guided by
policy sketches. Sketches annotate each task with a sequence of named subtasks,
providing high-level structural relationships among tasks, but not providing the
detailed guidance required by previous work on learning policy abstractions for
RL (e.g. intermediate rewards, subtask completion signals, or intrinsic motivations).
Our approach associates every subtask with its own modular subpolicy,
and jointly optimizes over full task-specific policies by tying parameters across
shared subpolicies. This optimization is accomplished via a simple decoupled
actor–critic training objective that facilitates learning common behaviors from
dissimilar reward functions. We evaluate the effectiveness of our approach on a
maze navigation game and a 2-D Minecraft-inspired crafting game. Both games
feature extremely sparse rewards that can be obtained only after completing a
number of high-level subgoals (e.g. escaping from a sequence of locked rooms or
collecting and combining various ingredients in the proper order). Experiments
illustrate two main advantages of our approach. First, we outperform standard
baselines that learn task-specific or shared monolithic policies. Second, our
method naturally induces a library of primitive behaviors that can be recombined
to rapidly acquire policies for new tasks. | ["Reinforcement Learning", "Transfer Learning"] | https://openreview.net/forum?id=H1kjdOYlx | https://openreview.net/pdf?id=H1kjdOYlx | https://openreview.net/forum?id=H1kjdOYlx¬eId=SyJDbsb4e |
HyaORazEe | H1kjdOYlx | ICLR.cc/2017/conference/-/paper107/official/review | {"title": "framework for multiagent hierarchical RL using policy sketches ", "rating": "5: Marginally below acceptance threshold", "review": "This paper studies the problem of abstract hierarchical multiagent RL with policy sketches, high level descriptions of abstract actions. The work is related to much previous work in hierarchical RL, and adds some new elements by using neural implementations of prior work on hierarchical learning and skill representations. \n\nSketches are sequences of high level symbolic labels drawn from some fixed vocabulary, which initially are devoid of any meaning. Eventually the sketches get mapped into real policies and enable policy transfer and temporal abstraction. Learning occurs through a variant of the standard actor critic architecture. \n\nExperiments are provided through a standard game like domain (maze, minecraft etc.). \n\nThe paper as written suffers from two problems. One, the idea of policy sketches is nice, but not sufficiently fleshed out to have any real impact. It would have been useful to see this spelled out in the context of abstract SMDP models to see what they bring to the table. What one gets here is some specialized invocation of this idea in the context of the specific approach proposed here. Second, the experiments are not thorough enough in terms of comparing with all the related work. For example, Ghavamzadeh et al. explored the use of MAXQ like abstractions in the context of mulitagent RL. It would be great to get a more detailed comparison to MAXQ based multiagent RL approaches, where the value function is explicitly decomposed. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Modular Multitask Reinforcement Learning with Policy Sketches | ["Jacob Andreas", "Dan Klein", "Sergey Levine"] | We describe a framework for multitask deep reinforcement learning guided by
policy sketches. Sketches annotate each task with a sequence of named subtasks,
providing high-level structural relationships among tasks, but not providing the
detailed guidance required by previous work on learning policy abstractions for
RL (e.g. intermediate rewards, subtask completion signals, or intrinsic motivations).
Our approach associates every subtask with its own modular subpolicy,
and jointly optimizes over full task-specific policies by tying parameters across
shared subpolicies. This optimization is accomplished via a simple decoupled
actor–critic training objective that facilitates learning common behaviors from
dissimilar reward functions. We evaluate the effectiveness of our approach on a
maze navigation game and a 2-D Minecraft-inspired crafting game. Both games
feature extremely sparse rewards that can be obtained only after completing a
number of high-level subgoals (e.g. escaping from a sequence of locked rooms or
collecting and combining various ingredients in the proper order). Experiments
illustrate two main advantages of our approach. First, we outperform standard
baselines that learn task-specific or shared monolithic policies. Second, our
method naturally induces a library of primitive behaviors that can be recombined
to rapidly acquire policies for new tasks. | ["Reinforcement Learning", "Transfer Learning"] | https://openreview.net/forum?id=H1kjdOYlx | https://openreview.net/pdf?id=H1kjdOYlx | https://openreview.net/forum?id=H1kjdOYlx¬eId=HyaORazEe |
ry65wbbSx | r1w7Jdqxl | ICLR.cc/2017/conference/-/paper449/official/review | {"title": "", "rating": "5: Marginally below acceptance threshold", "review": "The authors proposed to learn embeddings of users and items by using deep neural network for a recommendation task. The resulting method has only minor differences from the previous CDL, in which neural networks were also used for recommendation tasks. In the experiments, since the proposed method, DualNets have use more item features than WMF and CDL, the comparisons are unfair. ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Collaborative Deep Embedding via Dual Networks | ["Yilei Xiong", "Dahua Lin", "Haoying Niu", "JIefeng Cheng", "Zhenguo Li"] | Despite the long history of research on recommender systems, current approaches still face a number of challenges in practice, e.g. the difficulties in handling new items, the high diversity of user interests, and the noisiness and sparsity of observations. Many of such difficulties stem from the lack of expressive power to capture the complex relations between items and users. This paper presents a new method to tackle this problem, called Collaborative Deep Embedding. In this method, a pair of dual networks, one for encoding items and the other for users, are jointly trained in a collaborative fashion.
Particularly, both networks produce embeddings at multiple aligned levels, which, when combined together, can accurately predict the matching between items and users. Compared to existing methods, the proposed one not only provides greater expressive power to capture complex matching relations, but also generalizes better to unseen items or users. On multiple real-world datasets, this method outperforms the state of the art. | ["users", "dual networks", "items", "collaborative deep embedding", "difficulties", "deep embedding", "long history", "research", "recommender systems", "current approaches"] | https://openreview.net/forum?id=r1w7Jdqxl | https://openreview.net/pdf?id=r1w7Jdqxl | https://openreview.net/forum?id=r1w7Jdqxl¬eId=ry65wbbSx |
SJddLfmNx | r1w7Jdqxl | ICLR.cc/2017/conference/-/paper449/official/review | {"title": "review for Collaborative Deep Embedding via Dual Networks", "rating": "5: Marginally below acceptance threshold", "review": "This paper provides a minor improvement paper of DeepRS. The major improvement comes from the coupling of user-item factors in prediction. While the motivation is clear, the improvement of the model architecture is minor. \nI think the author should improve the paper to discuss more on the impact of introduction of coupling, which might make this paper stronger. Specifically, conduct isolate experiment to change loss, architecture gradually, from a non-coupled network to a final proposed coupled network to demonstrate the importance of coupling.\nAnother important missing part of the paper seems to be time complexity. Since coupled net would be much more costly to generate recommendations, a discussion on how it would impact real world usages should be added.\nOverall, I think this is a paper that should be improved before accepted.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Collaborative Deep Embedding via Dual Networks | ["Yilei Xiong", "Dahua Lin", "Haoying Niu", "JIefeng Cheng", "Zhenguo Li"] | Despite the long history of research on recommender systems, current approaches still face a number of challenges in practice, e.g. the difficulties in handling new items, the high diversity of user interests, and the noisiness and sparsity of observations. Many of such difficulties stem from the lack of expressive power to capture the complex relations between items and users. This paper presents a new method to tackle this problem, called Collaborative Deep Embedding. In this method, a pair of dual networks, one for encoding items and the other for users, are jointly trained in a collaborative fashion.
Particularly, both networks produce embeddings at multiple aligned levels, which, when combined together, can accurately predict the matching between items and users. Compared to existing methods, the proposed one not only provides greater expressive power to capture complex matching relations, but also generalizes better to unseen items or users. On multiple real-world datasets, this method outperforms the state of the art. | ["users", "dual networks", "items", "collaborative deep embedding", "difficulties", "deep embedding", "long history", "research", "recommender systems", "current approaches"] | https://openreview.net/forum?id=r1w7Jdqxl | https://openreview.net/pdf?id=r1w7Jdqxl | https://openreview.net/forum?id=r1w7Jdqxl¬eId=SJddLfmNx |
SkS6BflNl | r1w7Jdqxl | ICLR.cc/2017/conference/-/paper449/official/review | {"title": "official review", "rating": "4: Ok but not good enough - rejection", "review": "The responses to the pre-review questions are not strong; especially w.r.t. the question about dataset density and why the dataset had to be subsampled, the authors responded that subsampling is common in recommender systems work, including the papers cited. This isn't a particularly strong justification of why subsampling is a good idea, and in particular doesn't answer the question of \"how would the results look without subsampling,\" which I think is a question that could easily have been answered directly.\n\nEspecially given that the goal of dealing with the cold-start issue is so heavily emphasized in the paper, in seems odd to sample the data to reduce sparsity.\n\nOther than that, the pre-review questions seem to have been answered satisfactorily.\n\nThe contribution of the paper is to propose user and item embedding methods, as a means of learning complex non-linear interactions between users and items. This is fairly similar to recent work on deep RS, though the network formulation has some differences.\n\nOverall this is an reasonably put together paper that makes a contribution in an important area, though there are still some shortcomings that should be addressed, namely:\n1) The evaluation is unusual. Recall@M is the only result reported, though this is not usually an evaluation seen in recommender systems research. At the very least other performance measures (rmse or AUC) should be reported for completeness, even if the results are not strong\n2) Given that the contribution is fairly simple (i.e., the \"standard\" recommender systems task, but with a new model) it's a shame that unusual data samples have to be taken. This should be a case where it's possible to report results against competing methods using *exactly* the same data they used, and exactly the same error measure, for the fairest comparison possible.\n\nWithout the above it's hard to tell how much the performance improvements are really due to the method being better, versus the choice of datasets and the choice of loss functions.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Collaborative Deep Embedding via Dual Networks | ["Yilei Xiong", "Dahua Lin", "Haoying Niu", "JIefeng Cheng", "Zhenguo Li"] | Despite the long history of research on recommender systems, current approaches still face a number of challenges in practice, e.g. the difficulties in handling new items, the high diversity of user interests, and the noisiness and sparsity of observations. Many of such difficulties stem from the lack of expressive power to capture the complex relations between items and users. This paper presents a new method to tackle this problem, called Collaborative Deep Embedding. In this method, a pair of dual networks, one for encoding items and the other for users, are jointly trained in a collaborative fashion.
Particularly, both networks produce embeddings at multiple aligned levels, which, when combined together, can accurately predict the matching between items and users. Compared to existing methods, the proposed one not only provides greater expressive power to capture complex matching relations, but also generalizes better to unseen items or users. On multiple real-world datasets, this method outperforms the state of the art. | ["users", "dual networks", "items", "collaborative deep embedding", "difficulties", "deep embedding", "long history", "research", "recommender systems", "current approaches"] | https://openreview.net/forum?id=r1w7Jdqxl | https://openreview.net/pdf?id=r1w7Jdqxl | https://openreview.net/forum?id=r1w7Jdqxl¬eId=SkS6BflNl |
H1Sfbi-Nl | Hk8rlUqge | ICLR.cc/2017/conference/-/paper284/official/review | {"title": "", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes an alternative to Conditional Variational Auto-Encoders and Conditional MultiModal Auto-Encoders to perform inference of missing modalities in dataset with multiple modalities. The proposed approach is a Variational Auto-Encoder jointly on all the modalities with additional KL divergence penalties between the approximate posterior given all the modalities and the approximate posterior given a subset of the modalities. The approach is named Joint Multimodal Variational Auto-Encoder.\nThe authors make a connection between this approach and the Variation of Information. It is unclear why the authors chose the JMVAE approach instead of a more elegant Variation of Information approach.\nAnother unaddressed issue is the scalability of the method. As far as I can tell (given that no code is provided and the specification of the encoder is missing), this approach requires a new encoder per subset of missing modalities. Right now this approach seems to scale since there are only two modalities.\nThe fact that the estimating the log-likelihood log(p(x)) using multiple modalities provide a lower value than with just one in Table 1 is a bit odd. Could you explain that ?\nThe comparison with between the representation learned by JMVAE and CVAE might be unfair given that the representation of CVAE is learned conditionally, on the label in the case of MNIST, and should therefore not consider the label in this representation. Intuitively, this representation could represent \"style\" as shown in (Kingma et al., 2014) in their conditional generation figure.\nFor CelebA, comparing log-likelihood on models that use GANs is probably not significant since GAN does not optimizes log-likelihood. \nOverall this is an interesting problem and there are also interesting ideas worth exploring further, but the execution of the paper requires more work.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Joint Multimodal Learning with Deep Generative Models | ["Masahiro Suzuki", "Kotaro Nakayama", "Yutaka Matsuo"] | We investigate deep generative models that can exchange multiple modalities bi-directionally, e.g., generating images from corresponding texts and vice versa. Recently, some studies handle multiple modalities on deep generative models, such as variational autoencoders (VAEs). However, these models typically assume that modalities are forced to have a conditioned relation, i.e., we can only generate modalities in one direction. To achieve our objective, we should extract a joint representation that captures high-level concepts among all modalities and through which we can exchange them bi-directionally. As described herein, we propose a joint multimodal variational autoencoder (JMVAE), in which all modalities are independently conditioned on joint representation. In other words, it models a joint distribution of modalities. Furthermore, to be able to generate missing modalities from the remaining modalities properly, we develop an additional method, JMVAE-kl, that is trained by reducing the divergence between JMVAE's encoder and prepared networks of respective modalities. Our experiments show that our proposed method can obtain appropriate joint representation from multiple modalities and that it can generate and reconstruct them more properly than conventional VAEs. We further demonstrate that JMVAE can generate multiple modalities bi-directionally.
| ["modalities", "deep generative models", "multiple modalities", "jmvae", "models", "joint representation", "joint multimodal", "multimodal", "images", "texts"] | https://openreview.net/forum?id=Hk8rlUqge | https://openreview.net/pdf?id=Hk8rlUqge | https://openreview.net/forum?id=Hk8rlUqge¬eId=H1Sfbi-Nl |
HJrLdzMEl | Hk8rlUqge | ICLR.cc/2017/conference/-/paper284/official/review | {"title": "Some interesting ideas, but has major problems", "rating": "3: Clear rejection", "review": "The proposed method of modeling multimodal datasets is a VAE with an inference network for every combination of missing and present modalities. The method is evaluated on modeling MNIST and CelebA datasets.\n\nMNIST is hardly a multimodal dataset. The authors propose to use the labels as a separate modality that gets modeled with a variational autoencoder. The reviewer finds this choice perplexing.\nEven then the modalities are never actually missing, so the applicability of the suggested method is questionable.\nIn addition the differences in log-likelihoods between different models are tiny, and likely to be due to noise.\n\nThe other experiment reports log-likelihood of models that were not trained to maximize log-likelihood. It is not clear what conclusions can be drawn from such comparison.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Joint Multimodal Learning with Deep Generative Models | ["Masahiro Suzuki", "Kotaro Nakayama", "Yutaka Matsuo"] | We investigate deep generative models that can exchange multiple modalities bi-directionally, e.g., generating images from corresponding texts and vice versa. Recently, some studies handle multiple modalities on deep generative models, such as variational autoencoders (VAEs). However, these models typically assume that modalities are forced to have a conditioned relation, i.e., we can only generate modalities in one direction. To achieve our objective, we should extract a joint representation that captures high-level concepts among all modalities and through which we can exchange them bi-directionally. As described herein, we propose a joint multimodal variational autoencoder (JMVAE), in which all modalities are independently conditioned on joint representation. In other words, it models a joint distribution of modalities. Furthermore, to be able to generate missing modalities from the remaining modalities properly, we develop an additional method, JMVAE-kl, that is trained by reducing the divergence between JMVAE's encoder and prepared networks of respective modalities. Our experiments show that our proposed method can obtain appropriate joint representation from multiple modalities and that it can generate and reconstruct them more properly than conventional VAEs. We further demonstrate that JMVAE can generate multiple modalities bi-directionally.
| ["modalities", "deep generative models", "multiple modalities", "jmvae", "models", "joint representation", "joint multimodal", "multimodal", "images", "texts"] | https://openreview.net/forum?id=Hk8rlUqge | https://openreview.net/pdf?id=Hk8rlUqge | https://openreview.net/forum?id=Hk8rlUqge¬eId=HJrLdzMEl |
rJ3O99HEe | Hk8rlUqge | ICLR.cc/2017/conference/-/paper284/official/review | {"title": "The JMVAE is rather straightforward extension of VAE.", "rating": "5: Marginally below acceptance threshold", "review": "The paper introduces the joint multimodal variational autoencoder, a directed graphical model for modeling multimodal data with latent variable. the model is rather straightforward extension of standard VAE where two data modalities are generated from a shared latent representation independently. In order to deal with missing input modalities or bi-directional inference between two modalities the paper introduces modality-specific encoder that is trained to minimize the KL divergence of latent variable distributions between joint and modality-specific recognition networks. The paper demonstrates its effectiveness on MNIST and CelebA datasets, both in terms of test log-likelihoods and the conditional image generation and editing.\n\nThe proposed method is rather straightforward extension of VAE and therefore the model should inherent the probabilistic inference methods of VAE. For example, for missing data modalities, the model should be able to infer joint representation as well as filling in the missing modalities via iterative sampling as introduced by Rezende et al. (2014). Given marginal improvement, I am not convinced by the contribution of modality-specific encoders in Section 3.3. In addition, the inference methods introduced for generating Figure 5 looks somewhat unprincipled; I am wondering the conditional image generation results by following more principled approach (e.g., iterative sampling). Experimental results on joint image-attribute generation is also missing.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Joint Multimodal Learning with Deep Generative Models | ["Masahiro Suzuki", "Kotaro Nakayama", "Yutaka Matsuo"] | We investigate deep generative models that can exchange multiple modalities bi-directionally, e.g., generating images from corresponding texts and vice versa. Recently, some studies handle multiple modalities on deep generative models, such as variational autoencoders (VAEs). However, these models typically assume that modalities are forced to have a conditioned relation, i.e., we can only generate modalities in one direction. To achieve our objective, we should extract a joint representation that captures high-level concepts among all modalities and through which we can exchange them bi-directionally. As described herein, we propose a joint multimodal variational autoencoder (JMVAE), in which all modalities are independently conditioned on joint representation. In other words, it models a joint distribution of modalities. Furthermore, to be able to generate missing modalities from the remaining modalities properly, we develop an additional method, JMVAE-kl, that is trained by reducing the divergence between JMVAE's encoder and prepared networks of respective modalities. Our experiments show that our proposed method can obtain appropriate joint representation from multiple modalities and that it can generate and reconstruct them more properly than conventional VAEs. We further demonstrate that JMVAE can generate multiple modalities bi-directionally.
| ["modalities", "deep generative models", "multiple modalities", "jmvae", "models", "joint representation", "joint multimodal", "multimodal", "images", "texts"] | https://openreview.net/forum?id=Hk8rlUqge | https://openreview.net/pdf?id=Hk8rlUqge | https://openreview.net/forum?id=Hk8rlUqge¬eId=rJ3O99HEe |
Hy5E206El | B1IzH7cxl | ICLR.cc/2017/conference/-/paper192/official/review | {"title": "A relevant contribution to time series modelling with flexible nonlinear models", "rating": "6: Marginally above acceptance threshold", "review": "The authors propose a recurrent variational neural network approach to modelling volatility in financial time series. This model consists of an application of Chung et al.\u2019s (2015) VRNN model to volatility forecasting, wherein a Variational Autoencoder (VAE) structure is repeated at each time step of the series. \n\nThe paper is well written and easy to follow (although this reviewer suggests applying a spelling checking, since the paper contains a number of harmless typos). \n\nThe paper\u2019s main technical contribution is to stack two levels of recurrence, one for the latent process and one for the observables. This appears to be a novel, if minor contribution. The larger contribution is methodological, in areas of time series modelling that are both of great practical importance and have hitherto been dominated by rigid functional forms. The demonstration of the applicability and usefulness of general-purpose non-linear models for volatility forecasting would be extremely impactful.\n\nI have a few comments and reservations with the paper:\n1) Although not mentioned explicitly, the authors\u2019 framework are couched in terms of carrying out one-timestep-ahead forecasts of volatility. However, many applications of volatility models, for instance for derivative pricing, require longer-horizon forecasts. It would be interesting to discuss how this model could be extended to forecast at longer horizons.\n \n2) In Section 4.4, there\u2019s a mention that a GARCH(1,1) is conditionally deterministic. This is true only when forecasting 1 time-step in the future. At longer horizons, the GARCH(1,1) volatility forecast is not deterministic. \n\n3) I was initially unhappy with the limitations of the experimental validation, limited to comparison with a baseline GARCH model. However, the authors provided more comparisons in the revision, which adds to the quality of the results, although the models compared against cannot be considered state of the art. It would be well advised to look into R packages such as `stochvol\u2019 and \u2018fGarch\u2019 to get implementations of a variety of models that can serve as useful baselines, and provide convincing evidence that the modelled volatility is indeed substantially better than approaches currently entertained by the finance literature.\n\n4) In Section 5.2, more details should be given on the network, e.g. number of hidden units, as well as the embedding dimension D_E (section 5.4)\n\n5) In Section 5.3, more details should be given on the data generating process for the synthetic data experiments. \n\n6) Some results in the appendix are very puzzling: around jumps in the price series, which are places where the volatility should spike, the model reacts instead by huge drops in the volatility (Figure 4(b) and (c), respectively around time steps 1300 and 1600). This should be explained and discussed.\n\nAll in all, I think that the paper provides a nice contribution to the art of volatility modelling. In spite of some flaws, it provides a starting point for the broader impact of neural time series processing in the financial community.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | A Neural Stochastic Volatility Model | ["Rui Luo", "Xiaojun Xu", "Weinan Zhang", "Jun Wang"] | In this paper, we show that the recent integration of statistical models with recurrent neural networks provides a new way of formulating volatility models that have been popular in time series analysis and prediction. The model comprises a pair of complementary stochastic recurrent neural networks: the generative network models the joint distribution of the stochastic volatility process; the inference network approximates the conditional distribution of the latent variables given the observable ones.
Our focus in this paper is on the formulation of temporal dynamics of volatility over time under a stochastic recurrent neural network framework. Our derivations show that some popular volatility models are a special case of our proposed neural stochastic volatility model. Experiments demonstrate that the proposed model generates a smoother volatility estimation, and largely outperforms a widely used GARCH model on several metrics about the fitness of the volatility modelling and the accuracy of the prediction. | ["Deep learning", "Supervised Learning"] | https://openreview.net/forum?id=B1IzH7cxl | https://openreview.net/pdf?id=B1IzH7cxl | https://openreview.net/forum?id=B1IzH7cxl¬eId=Hy5E206El |
S1149YZNl | B1IzH7cxl | ICLR.cc/2017/conference/-/paper192/official/review | {"title": "Lack of smoothing for learning and lack of novelty", "rating": "5: Marginally below acceptance threshold", "review": "\nThe authors propose a recurrent neural network approach for constructing a\nstochastic volatility model for financial time series. They introduce an\ninference network based on a recurrent neural network that computes the\napproximation to the posterior distribution for the latent variables given the\npast data. This variational approximation is used to maximize the marginal\nlikelihood in order to learn the parameters of the model. The proposed method\nis validated in experiments with synthetic and real-world time series, showing\nto outperform parametric GARCH models and a Gaussian process volatility model.\n\nQuality:\n\nThe method proposed seems technically correct, with the exception that in\nequation (19) the inference model is doing filtering and not smoothing, in the\nsense that the posterior for z_t' only depends on those other z_t and x_t\nvalues with t<t', but in the true posterior p(Z|X) each z_t depends on all the\nX. This means the proposed learning method is inefficient. The reference below\nshows how to perform smoothing too and the results in that paper show indeed that\nsmoothing produces better results for learning the model.\n\nSequential Neural Models with Stochastic Layers Fraccaro, Marco and S\\o nderby,\nS\\o ren Kaae and Paquet, Ulrich and Winther, Ole In NIPS 2016.\n\nIt is not clear if the method proposed in the above reference would perform better \njust because of using smoothing when learning the model parameters.\n\nClarity:\n\nThe paper is clearly written and easy to read.\n\nFor the results on real-world data, in Table 1, how is the NLL computed? Is the\naverage NLL across the 162 time series?\n\nOriginality:\n\nThe method proposed is not very original. Previous work has already used the\nvariational approach with the reparametrization trick to learn recurrent neural\nnetworks with stochastic units (see the reference above). It seems that the main\ncontribution of the authors is to apply this type of techniques to the problem\nof modeling financial time series.\n\nSignificance:\n\nThe results shown are significant, the method proposed by the authors\noutperforms previous approaches. However, there is a huge amount of techniques\navailable for modeling financial time-series. The number of GARCH variants is\nprobably close to hundreds, each one claiming to be better than the others.\nThis makes difficult to quantify how important the results are.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | A Neural Stochastic Volatility Model | ["Rui Luo", "Xiaojun Xu", "Weinan Zhang", "Jun Wang"] | In this paper, we show that the recent integration of statistical models with recurrent neural networks provides a new way of formulating volatility models that have been popular in time series analysis and prediction. The model comprises a pair of complementary stochastic recurrent neural networks: the generative network models the joint distribution of the stochastic volatility process; the inference network approximates the conditional distribution of the latent variables given the observable ones.
Our focus in this paper is on the formulation of temporal dynamics of volatility over time under a stochastic recurrent neural network framework. Our derivations show that some popular volatility models are a special case of our proposed neural stochastic volatility model. Experiments demonstrate that the proposed model generates a smoother volatility estimation, and largely outperforms a widely used GARCH model on several metrics about the fitness of the volatility modelling and the accuracy of the prediction. | ["Deep learning", "Supervised Learning"] | https://openreview.net/forum?id=B1IzH7cxl | https://openreview.net/pdf?id=B1IzH7cxl | https://openreview.net/forum?id=B1IzH7cxl¬eId=S1149YZNl |
rytIA0A7e | B1IzH7cxl | ICLR.cc/2017/conference/-/paper192/official/review | {"title": "An interesting application of RNN generative model to financial data", "rating": "5: Marginally below acceptance threshold", "review": "Thank you for an interesting read.\n\nI found the application of VRNN type generative model to financial data very promising. But since I don't have enough background knowledge to judge whether the performance gap is significant or not, I wouldn't recommend acceptance at this stage. \n\nTo me, the biggest issue for this paper is that I'm not sure if the paper contains significant novelty. The RNN-VAE combination has been around for more than a year and this paper does not propose significant changes to it. Maybe this paper fits better to an application targeting conference, rather than ICLR. But I'm not exactly sure about ICLR's acceptance criteria, and maybe the committee actually prefer great performances and interesting applications?", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | A Neural Stochastic Volatility Model | ["Rui Luo", "Xiaojun Xu", "Weinan Zhang", "Jun Wang"] | In this paper, we show that the recent integration of statistical models with recurrent neural networks provides a new way of formulating volatility models that have been popular in time series analysis and prediction. The model comprises a pair of complementary stochastic recurrent neural networks: the generative network models the joint distribution of the stochastic volatility process; the inference network approximates the conditional distribution of the latent variables given the observable ones.
Our focus in this paper is on the formulation of temporal dynamics of volatility over time under a stochastic recurrent neural network framework. Our derivations show that some popular volatility models are a special case of our proposed neural stochastic volatility model. Experiments demonstrate that the proposed model generates a smoother volatility estimation, and largely outperforms a widely used GARCH model on several metrics about the fitness of the volatility modelling and the accuracy of the prediction. | ["Deep learning", "Supervised Learning"] | https://openreview.net/forum?id=B1IzH7cxl | https://openreview.net/pdf?id=B1IzH7cxl | https://openreview.net/forum?id=B1IzH7cxl¬eId=rytIA0A7e |
rkqw-ubNg | BycCx8qex | ICLR.cc/2017/conference/-/paper287/official/review | {"title": "Final review", "rating": "5: Marginally below acceptance threshold", "review": "Overall, this is a nice paper. Developing a unifying framework for these newer\nneural models is a worthwhile endeavor.\n\nHowever, it's unclear if the DRAGNN framework (in its current form) is a\nsignificant standalone contribution. The main idea is straightforward: use a\ntransition system to unroll a computation graph. When you implement models in\nthis way you can reuse code because modules can be mixed and matched. This is\nnice, but (in my opinion) is just good software engineering, not machine \nlearning research.\n\nMoreover, there appears to be little incentive to use DRAGNN, as there are no\n'free things' (benefits) that you get by using the framework. For example:\n\n- If you write your neuralnet in an automatic differentiation library (e.g.,\n tensorflow or dynet) you get gradients for 'free'.\n\n- In the VW framework, there are efficiency tricks that 'the credit assignment\n compiler' provides for you, which would be tedious to implement on your\n own. There is also a variety of algorithms for training the model in a\n principled way (i.e., without exposure bias).\n\nI don't feel that my question about the limitations of the framework has been\nsatisfactorily addressed. Let me ask it in a different way: Can you give me\nexamples of a few models that I can't (nicely) express in the DRAGNN framework?\nWhat if I wanted to implement https://openreview.net/pdf?id=HkE0Nvqlg or\nhttp://www.cs.jhu.edu/~jason/papers/rastogi+al.naacl16.pdf? Can I implement the\ndynamic programming components as transition units and (importantly) would it be\nefficient?\n\n disagree that the VW framework is orthogonal, it is a *competing* way to\nimplement recurrent models. The main different to me appears to be that VW's\nimperative framework is more general, but less modular.\n\nThe experimental contribution seems useful as does the emphasis on how easy it\nis to incorporate multi-task learning.\n\nMinor:\n\n- It would be useful to see actual code snippets (possibly in an\n appendix). Otherwise, its unclear how modular DRAGNN really are.\n\n- The introduction states that (unlike seq2seq+attention) inference remains\n linear. Is this *necessarily* the case? Users define a transition system that\n is quadratic, just let attention be over all previous states. I recommend that\n authors rephrase statement more carefully.\n\n- It seems strange to use A() as in \"actions\", then use d as \"decision\" for its\n elements.\n\n- I recommend adding i as an argument to the definition of the recurrence\n function r(s) to make it clear that it's the subset of previous states at time\n i, otherwise it looks like an undefined variable. A nice terse option is to\n write r(s_i).\n\n- Real numbers should be \\mathbb{R} not \\mathcal{R}.\n\n- It's more conventional to use t for a time-step instead of i.\n\n- Example 2: \"52 feature embeddings\" -> did you mean \"52-DIMENSIONAL feature\n embeddings\"?\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | DRAGNN: A Transition-Based Framework for Dynamically Connected Neural Networks | ["Lingpeng Kong", "Chris Alberti", "Daniel Andor", "Ivan Bogatyy", "David Weiss"] | In this work, we present a compact, modular framework for constructing new recurrent neural architectures. Our basic module is a new generic unit, the Transition Based Recurrent Unit (TBRU). In addition to hidden layer activations, TBRUs have discrete state dynamics that allow network connections to be built dynamically as a function of intermediate activations. By connecting multiple TBRUs, we can extend and combine commonly used architectures such as sequence-to-sequence, attention mechanisms, and recursive tree-structured models. A TBRU can also serve as both an {\em encoder} for downstream tasks and as a {\em decoder} for its own task simultaneously, resulting in more accurate multi-task learning. We call our approach Dynamic Recurrent Acyclic Graphical Neural Networks, or DRAGNN. We show that DRAGNN is significantly more accurate and efficient than seq2seq with attention for syntactic dependency parsing and yields more accurate multi-task learning for extractive summarization tasks.
| ["Natural language processing", "Deep learning", "Multi-modal learning", "Structured prediction"] | https://openreview.net/forum?id=BycCx8qex | https://openreview.net/pdf?id=BycCx8qex | https://openreview.net/forum?id=BycCx8qex¬eId=rkqw-ubNg |
H1qE6OVVg | BycCx8qex | ICLR.cc/2017/conference/-/paper287/official/review | {"title": "", "rating": "7: Good paper, accept", "review": "The authors present a general framework for defining a wide variety of recurrent neural network architectures, including seq2seq models, tree-structured models, attention, and a new family of dynamically connected architectures. The framework defines a new, general-purpose recurrent unit called the TBRU, which takes a transition system, defining and constraining its inputs and outputs, and input function which defines the mapping between raw inputs and fixed-width vector representations, and recurrence function that defines the inputs to each recurrent step as a function of the current state, and an RNN cell that computes the output from the input (fixed and recurrent). Many example instantiations of this framework are provided, including sequential tagging RNNs, Google\u2019s Parsey McParseface parser, encoder/decoder networks, tree LSTMs and less familiar examples that demonstrate the power this framework. \n\nThe most interesting contribution of this work is the ease by which it can be used to incorporate dynamic recurrent connections through the definition of the transition system. In particular, this paper explores the application of these dynamic connections to syntactic dependency parsing, both as a standalone task, and by multitasking parsing with extractive summarization, using the same compositional phrase representations as features for the parser and summarization (previous work used discrete parse features), which is particularly simple/elegant in this framework. In experimental results, the authors demonstrate that such multitasking leads to more accurate summarization models, and using the framework to incorporate more structure into existing parsing models also leads to increased accuracy with no big-oh efficiency loss (compared with e.g. attention). \n\nThe \u201craison d\u2019etre,\u201d in particular the example, perhaps described even more thoroughly/explicitly, should be made as clear as possible as soon as possible. This is the most important contribution, but it gets lost in the description and presentation as a framework \u2014 emphasizing that attention, seq2seq, etc can be represented in the framework is distracting and makes it seem less novel than it is. AnonReviewer6 clearly missed this point, as did I in my first pass over the paper. To get this idea across and to emphasize the benefits of this representation, I\u2019d love to see more detailed analysis of these representations and their importance to achieving your experimental results. I think it would also be helpful to emphasize the difference between a stack LSTM and Example 6. \n\nOverall I think this paper presents a valuable contribution, though the exposition could be improved and analysis of experimental results expanded. ", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | DRAGNN: A Transition-Based Framework for Dynamically Connected Neural Networks | ["Lingpeng Kong", "Chris Alberti", "Daniel Andor", "Ivan Bogatyy", "David Weiss"] | In this work, we present a compact, modular framework for constructing new recurrent neural architectures. Our basic module is a new generic unit, the Transition Based Recurrent Unit (TBRU). In addition to hidden layer activations, TBRUs have discrete state dynamics that allow network connections to be built dynamically as a function of intermediate activations. By connecting multiple TBRUs, we can extend and combine commonly used architectures such as sequence-to-sequence, attention mechanisms, and recursive tree-structured models. A TBRU can also serve as both an {\em encoder} for downstream tasks and as a {\em decoder} for its own task simultaneously, resulting in more accurate multi-task learning. We call our approach Dynamic Recurrent Acyclic Graphical Neural Networks, or DRAGNN. We show that DRAGNN is significantly more accurate and efficient than seq2seq with attention for syntactic dependency parsing and yields more accurate multi-task learning for extractive summarization tasks.
| ["Natural language processing", "Deep learning", "Multi-modal learning", "Structured prediction"] | https://openreview.net/forum?id=BycCx8qex | https://openreview.net/pdf?id=BycCx8qex | https://openreview.net/forum?id=BycCx8qex¬eId=H1qE6OVVg |
BkAyn-8Ex | BycCx8qex | ICLR.cc/2017/conference/-/paper287/official/review | {"title": "official review", "rating": "6: Marginally above acceptance threshold", "review": "The paper proposes a new neural architecture, called DRAGNN, for the transition-based framework. A DRAGNN uses TBRUs which are neural units to compute hidden activations for the current state of a transition-based system. The paper proves that DRAGNNs can cover a wide range of transition-based methods in the literature. In addition, one can easily implement multitask learning systems with DRAGNNs. The experimental results shows that using DRAGNNs the authors built (near) state-of-the-art systems for 2 tasks: parsing and summarization. \n\nThe paper contains two major parts: DRAGNN and demonstrations of its usages. \n\nRegarding to the first part, the proposed DRAGNN is a neat tool for building any transition-based systems. However, it is difficult to say whether the DRAGNN is novel. Transition-based framework is already well defined and there's a huge trend in NLP using neural networks to implement transition-based systems. In my opinion, the difference between the Stack-LSTM (Dyer et al., 2015) and DRAGNN is slight. Of course, the DRAGNN is a powerful architecture but the contribution here should be considered mainly in terms of software engineering.\n\nIn the second part, the authors used DRAGNN to implement new transition-based systems for different (multi-)tasks. The implementations are neat, confirming that DRAGNN is a powerful architecture, especially for multitask learning. However, we should bear in mind that the solutions employed are already there in the literature, thus making difficult to judge the novelty of this part w.r.t. the theme of the conference. ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | DRAGNN: A Transition-Based Framework for Dynamically Connected Neural Networks | ["Lingpeng Kong", "Chris Alberti", "Daniel Andor", "Ivan Bogatyy", "David Weiss"] | In this work, we present a compact, modular framework for constructing new recurrent neural architectures. Our basic module is a new generic unit, the Transition Based Recurrent Unit (TBRU). In addition to hidden layer activations, TBRUs have discrete state dynamics that allow network connections to be built dynamically as a function of intermediate activations. By connecting multiple TBRUs, we can extend and combine commonly used architectures such as sequence-to-sequence, attention mechanisms, and recursive tree-structured models. A TBRU can also serve as both an {\em encoder} for downstream tasks and as a {\em decoder} for its own task simultaneously, resulting in more accurate multi-task learning. We call our approach Dynamic Recurrent Acyclic Graphical Neural Networks, or DRAGNN. We show that DRAGNN is significantly more accurate and efficient than seq2seq with attention for syntactic dependency parsing and yields more accurate multi-task learning for extractive summarization tasks.
| ["Natural language processing", "Deep learning", "Multi-modal learning", "Structured prediction"] | https://openreview.net/forum?id=BycCx8qex | https://openreview.net/pdf?id=BycCx8qex | https://openreview.net/forum?id=BycCx8qex¬eId=BkAyn-8Ex |
SybLfCSEe | HJIY0E9ge | ICLR.cc/2017/conference/-/paper228/official/review | {"title": "Interesting results and the start of a good paper", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes and tests two ideas. (1) a method of pruning networks by identifying highly correlated neuron pairs, pruning one of the pair, and then modifying downstream weights to compensate for the removal (which works well if the removed neurons were highly correlated). (2) a method, dubbed NoiseOut, for increasing neuron correlation by adding auxiliary noise target outputs to the network during training.\n\n\nThe first idea (1) is fairly straightforward, and it is not clear if it has been tried before. It does seem to work.\n\n\nThe second idea (2) is of unclear value and seems to this reviewer that it may merely add a regularizing effect. Comments in this direction:\n - In Fig 4 (right), the constant and Gaussian treatments seem to produce the same effect in both networks, right? And the Binomial effect seems the same as No_Noise. If this is true, can we conclude that the NoiseOut targets are simply serving to regularize the network, that is, to reduce its capacity slightly?\n - To show whether this effect is true, one would need to compare to other methods of reducing the network capacity, for example: by reducing the number of neurons, by applying L2 regularization of various values, or by applying Dropout of various strengths. Fig 7 makes an attempt at this direction, but critically misses several comparison treatments: \u201cPruned without any regularization\u201d, \u201cPruned with only L2\u201d, and \u201cPruned with only DropOut\u201d. Have these experiments been run? Can their results be included and used to produce plots like Fig 5 and Fig 7?\n\nWithout these comparisons, it seems impossible to conclude that NoiseOut does anything but provide similar regularization to DropOut or L2.\n\n\nThe combined ideas (1) + (2) DO produce a considerable reduction in parameters, but sadly the experiments and exposition are somewhat too lacking to really understand what is going on. With a little more work the paper could be quite interesting, but as is it should probably not be accepted.\n\n\nAdditional comments:\n - Section 4 states: \u201cIn all of these experiments, the only stop criteria is the accuracy decay of the model. We set the threshold for this criteria to match the original accuracy; therefore all the compressed network have the same accuracy as the original network.\u201d Is this accuracy the train accuracy or test accuracy? If train, then test accuracy needs to be shown (how much test performance is lost when pruning?). If test, then this would typically be referred to as \u201ccheating\u201d and so the choice needs to be very clearly stated and then defended.\n - Lowercase rho is used to indicate correlation but this is never actually specified, which is confusing for. Just state once that it indicates correlation.\n - How do these results compare to other pruning methods? No numerical comparison is attempted.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | A Simple yet Effective Method to Prune Dense Layers of Neural Networks | ["Mohammad Babaeizadeh", "Paris Smaragdis", "Roy H. Campbell"] | Neural networks are usually over-parameterized with significant redundancy in the number of required neurons which results in unnecessary computation and memory usage at inference time. One common approach to address this issue is to prune these big networks by removing extra neurons and parameters while maintaining the accuracy. In this paper, we propose NoiseOut, a fully automated pruning algorithm based on the correlation between activations of neurons in the hidden layers. We prove that adding additional output neurons with entirely random targets results into a higher correlation between neurons which makes pruning by NoiseOut even more efficient. Finally, we test our method on various networks
and datasets. These experiments exhibit high pruning rates while maintaining the accuracy of the original network. | ["Deep learning"] | https://openreview.net/forum?id=HJIY0E9ge | https://openreview.net/pdf?id=HJIY0E9ge | https://openreview.net/forum?id=HJIY0E9ge¬eId=SybLfCSEe |
B1zUKEd4x | HJIY0E9ge | ICLR.cc/2017/conference/-/paper228/official/review | {"title": "Good proof of concept, but more experimental evidence needed", "rating": "5: Marginally below acceptance threshold", "review": "Summary:\nIn this paper, the authors introduce NoiseOut, a way to reduce parameters by pruning neurons from a network. \nThey do this by identifying pairs of neurons produce the most correlated outputs, and replacing the pair by one neuron, and then appropriately adjusting weights.\nThis technique relies on neurons having high correlations however, so they introduce an additional output neuron -- a noise output, which results in the network trying to predict the mean of the noise distribution.\nAs this is a constant, it increases correlation between neurons.\nExperiments test this out on MNIST and SVHN\n\nComments:\nThis is an interesting suggestion on how to prune neurons, but more experiments (on larger datasets) are probably need to be convincing that this is an approach that is guaranteed to work well. \n\nEquation (5) seems to be very straightforwards?\n\nIt seems like that for larger datasets, more noise outputs might have to be added to ensure higher correlations? Is there a downside to this in terms of the overall accuracy?\n\nThe paper is presented clearly, and was definitely interesting to read, so I encourage the authors to continue this line of work.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | A Simple yet Effective Method to Prune Dense Layers of Neural Networks | ["Mohammad Babaeizadeh", "Paris Smaragdis", "Roy H. Campbell"] | Neural networks are usually over-parameterized with significant redundancy in the number of required neurons which results in unnecessary computation and memory usage at inference time. One common approach to address this issue is to prune these big networks by removing extra neurons and parameters while maintaining the accuracy. In this paper, we propose NoiseOut, a fully automated pruning algorithm based on the correlation between activations of neurons in the hidden layers. We prove that adding additional output neurons with entirely random targets results into a higher correlation between neurons which makes pruning by NoiseOut even more efficient. Finally, we test our method on various networks
and datasets. These experiments exhibit high pruning rates while maintaining the accuracy of the original network. | ["Deep learning"] | https://openreview.net/forum?id=HJIY0E9ge | https://openreview.net/pdf?id=HJIY0E9ge | https://openreview.net/forum?id=HJIY0E9ge¬eId=B1zUKEd4x |
By14J3MNx | HJIY0E9ge | ICLR.cc/2017/conference/-/paper228/official/review | {"title": "Intriguing idea, but theory and experiments are lacking.", "rating": "3: Clear rejection", "review": "The paper proposes to prune a neural network by removing neurons whose operation is highly correlated with other neurons. The idea is nice and somewhat novel - most pruning methods concentrate on removal of individual weights, however I haven't done a through research on this topic. However, the experimental and theoretical justification of this method need to be improved before publication:\n\n1. Experiments. The authors do not report accuracy degradation while pruning in the tables, laconically stating that the networks did not degrade. This is not convincing. The only details are given in Figure 5, however this Figure disagrees with Table 2: in the Table, the number of parameters ranges from 40k-600k, while the Figure pictures the range 12k-24k. Unless more details are provided, simply claiming that a network can remove 50% neurons with no number on the degradation of accuracy is not convincing.\n\n2. Theory. The proofs do not match the experimental conditions and make unreasonable assumptions. The proofs show that in the absence of biases a network with a constant output will have two correlated neurons that generate the output offset. However, this is exactly why networks have biases and doesn't explain why noise injection helps (the proof suggests that all should be fine with deterministic auxiliary neuron). My interpretation is that the noisy output injects gradient noise (see e.g. the concurrent ICLR submission https://openreview.net/forum?id=rkjZ2Pcxe). As such the proof muddies the picture more than it helps in understanding what is happening.\n\nVerdict:\nReject and resubmit. \nThe pruning idea has potential, however its efficiency must be more soundly demonstrated (please provide network accuracies at various pruning levels, the method removes one neuron at a time, this allows making of nice plots) rather than laconically stating that a degradation on mnist from 97% accuracy to 92% is not significant (Figure 5.). Please provide Figures and Tables that agree with the text in terms of numbers provided.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | A Simple yet Effective Method to Prune Dense Layers of Neural Networks | ["Mohammad Babaeizadeh", "Paris Smaragdis", "Roy H. Campbell"] | Neural networks are usually over-parameterized with significant redundancy in the number of required neurons which results in unnecessary computation and memory usage at inference time. One common approach to address this issue is to prune these big networks by removing extra neurons and parameters while maintaining the accuracy. In this paper, we propose NoiseOut, a fully automated pruning algorithm based on the correlation between activations of neurons in the hidden layers. We prove that adding additional output neurons with entirely random targets results into a higher correlation between neurons which makes pruning by NoiseOut even more efficient. Finally, we test our method on various networks
and datasets. These experiments exhibit high pruning rates while maintaining the accuracy of the original network. | ["Deep learning"] | https://openreview.net/forum?id=HJIY0E9ge | https://openreview.net/pdf?id=HJIY0E9ge | https://openreview.net/forum?id=HJIY0E9ge¬eId=By14J3MNx |
rJpkoyG4g | r1S083cgx | ICLR.cc/2017/conference/-/paper540/official/review | {"title": "Interesting, but low novelty, and sub-standard evaluation", "rating": "3: Clear rejection", "review": "The paper presents a method for sequence generation with a known method applied to feature extracted from another existing method. The paper is heavily oriented towards to chosen technologies and lacks in literature on sequence generation. In principle, rich literature on motion prediction for various applications could be relevant here. Recent models exist for sequence prediction (from primed inputs) for various applications, e.g. for skeleton data. These models learn complex motion w/o any pre-processing. \n\nEvaluation is a big concern. There is no quantitative evaluation. There is no comparision with other methods.\n\nI still wonder whether the intermediate representation (developed by Plamondon et al.) is useful in this context of a fully trained sequence generation model and whether the model could pick up the necessary transformations itself. This should be evaluated.\n\nDetails:\n\nThere are several typos and word omissions, which can be found by carefully rereading the paper.\n\nAt the beginning of section 3, it is still unclear what the application is. Prediction of dynamic parameters? What for? Section 3 should give a better motivation of the work.\n\nConcerning the following paragraph\n\n\"While such methods are superior for handwriting analysis and biometric purposes, we opt for a less precise method (Berio & Leymarie, 2015) that is less sensitive to sampling quality and is aimed at generating virtual target sequences that remain perceptually similar to the original trace. \n\"\nThis method has not been explained. A paper should be self-contained.\n\nThe authors mentioned that the \"V2V-model is conditioned on (...)\"; but not enough details are given. \n\nGenerally speaking, more efforts could be made to make the paper more self-contained.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Sequence generation with a physiologically plausible model of handwriting and Recurrent Mixture Density Networks | ["Daniel Berio", "Memo Akten", "Frederic Fol Leymarie", "Mick Grierson", "R\u00e9jean Plamondon"] | The purpose of this study is to explore the feasibility and potential benefits of using a physiological plausible model of handwriting as a feature representation for sequence generation with recurrent mixture density networks. We build on recent results in handwriting prediction developed by Graves (2013), and we focus on generating sequences that possess the statistical and dynamic qualities of handwriting and calligraphic art forms. Rather than model raw sequence data, we first preprocess and reconstruct the input training data with a concise representation given by a motor plan (in the form of a coarse sequence of `ballistic' targets) and corresponding dynamic parameters (which define the velocity and curvature of the pen-tip trajectory). This representation provides a number of advantages, such as enabling the system to learn from very few examples by introducing artificial variability in the training data, and mixing of visual and dynamic qualities learned from different datasets. | ["Deep learning", "Supervised Learning", "Applications"] | https://openreview.net/forum?id=r1S083cgx | https://openreview.net/pdf?id=r1S083cgx | https://openreview.net/forum?id=r1S083cgx¬eId=rJpkoyG4g |
By5zswDVl | r1S083cgx | ICLR.cc/2017/conference/-/paper540/official/review | {"title": "Nice paper, but no machine learning contribution or evaluation.", "rating": "3: Clear rejection", "review": "This paper has no machine learning algorithmic contribution: it just uses the the same combination of LSTM and bivariate mixture density network as Graves, and the detailed explanation in the appendix even misses one key essential point: how are the Gaussian parameters obtained as a transformation of the output of the LSTM.\nThere are also no numerical evaluation suggesting that the algorithm is some form of improvement over the state-of-the-art.\n\nSo I do not think such a paper is appropriate for a conference like ICLR. The part describing the handwriting tasks and the data transformation is well written and interesting to read, it could be valuable work for a conference more focused on handwriting recognition, but I am no expert in the field.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Sequence generation with a physiologically plausible model of handwriting and Recurrent Mixture Density Networks | ["Daniel Berio", "Memo Akten", "Frederic Fol Leymarie", "Mick Grierson", "R\u00e9jean Plamondon"] | The purpose of this study is to explore the feasibility and potential benefits of using a physiological plausible model of handwriting as a feature representation for sequence generation with recurrent mixture density networks. We build on recent results in handwriting prediction developed by Graves (2013), and we focus on generating sequences that possess the statistical and dynamic qualities of handwriting and calligraphic art forms. Rather than model raw sequence data, we first preprocess and reconstruct the input training data with a concise representation given by a motor plan (in the form of a coarse sequence of `ballistic' targets) and corresponding dynamic parameters (which define the velocity and curvature of the pen-tip trajectory). This representation provides a number of advantages, such as enabling the system to learn from very few examples by introducing artificial variability in the training data, and mixing of visual and dynamic qualities learned from different datasets. | ["Deep learning", "Supervised Learning", "Applications"] | https://openreview.net/forum?id=r1S083cgx | https://openreview.net/pdf?id=r1S083cgx | https://openreview.net/forum?id=r1S083cgx¬eId=By5zswDVl |
HkrSAC-Eg | r1S083cgx | ICLR.cc/2017/conference/-/paper540/official/review | {"title": "New representation space for a cut down version of an old model, with no quantitative results", "rating": "3: Clear rejection", "review": "This paper takes a model based on that of Graves and retrofits it with a representation derived from the work of Plamondon. \npart of the goal of deep learning has been to avoid the use of hand-crafted features and have the network learn from raw feature representations, so this paper is somewhat against the grain. \n\nThe paper relies on some qualitative examples as demonstration of the system, and doesn't seem to provide a strong motivation for there being any progress here. \nThe paper does not provide true text-conditional handwriting synthesis as shown in Graves' original work. \n\nBe more consistent about your bibliography (e.g. variants of Plamondon's own name, use of \"et al.\" in the bibliography etc.) ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Sequence generation with a physiologically plausible model of handwriting and Recurrent Mixture Density Networks | ["Daniel Berio", "Memo Akten", "Frederic Fol Leymarie", "Mick Grierson", "R\u00e9jean Plamondon"] | The purpose of this study is to explore the feasibility and potential benefits of using a physiological plausible model of handwriting as a feature representation for sequence generation with recurrent mixture density networks. We build on recent results in handwriting prediction developed by Graves (2013), and we focus on generating sequences that possess the statistical and dynamic qualities of handwriting and calligraphic art forms. Rather than model raw sequence data, we first preprocess and reconstruct the input training data with a concise representation given by a motor plan (in the form of a coarse sequence of `ballistic' targets) and corresponding dynamic parameters (which define the velocity and curvature of the pen-tip trajectory). This representation provides a number of advantages, such as enabling the system to learn from very few examples by introducing artificial variability in the training data, and mixing of visual and dynamic qualities learned from different datasets. | ["Deep learning", "Supervised Learning", "Applications"] | https://openreview.net/forum?id=r1S083cgx | https://openreview.net/pdf?id=r1S083cgx | https://openreview.net/forum?id=r1S083cgx¬eId=HkrSAC-Eg |
rk8jJ0VNx | ryMxXPFex | ICLR.cc/2017/conference/-/paper103/official/review | {"title": "clever and useful contribution; clear and thorough exposition", "rating": "9: Top 15% of accepted papers, strong accept", "review": "This paper presents a way of training deep generative models with discrete hidden variables using the reparameterization trick. It then applies it to a particular DBN-like architecture, and shows that this architecture achieves state-of-the-art density modeling performance on MNIST and similar datasets. \n\nThe paper is well written, and the exposition is both thorough and precise. There are several appendices which justify various design decisions in detail. I wish more papers in our field would take this degree of care with the exposition!\n\nThe log-likelihood results are quite strong, especially given that most of the competitive algorithms are based on continuous latent variables. Probably the main thing missing from the experiments is some way to separate out the contributions of the architecture and the inference algorithm. (E.g., what if a comparable architecture is trained with VIMCO, or if the algorithm is applied to a previously published discrete architecture?)\n\nI\u2019m a bit concerned about the variance of the gradients in the general formulation of the algorithm. See my comment \u201cvariance of the derivatives of F^{-1}\u201d below. I think the response is convincing, but the problem (as well as \u201cengineering principles\u201d for the smoothing distribution) are probably worth pointing out in the paper itself, since the problem seems likely to occur unless the user is aware of it. (E.g., my proposal of widely separated normals would be a natural distribution to consider until one actually works through the gradients \u2014 something not commonly done in the age of autodiff frameworks.)\n\nAnother concern is how many sequential operations are needed for inference in the RBM model. (Note: is this actually an RBM, or a general Boltzmann machine?) The q distribution takes the form of an autoregressive model where the variables are processed one at a time. Section 3 mentions the possibility of grouping together variables in the q distribution, and this is elaborated in detail in Appendix A. But the solution requires decomposing the joint into a product of conditionals and applying the CDFs sequentially. So either way, it seems like we\u2019re stuck handling all the variables sequentially, which might get expensive. \n\nMinor: the second paragraph of Section 3 needs a reference to Appendix A.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Discrete Variational Autoencoders | ["Jason Tyler Rolfe"] | Probabilistic models with discrete latent variables naturally capture datasets composed of discrete classes. However, they are difficult to train efficiently, since backpropagation through discrete variables is generally not possible. We present a novel method to train a class of probabilistic models with discrete latent variables using the variational autoencoder framework, including backpropagation through the discrete latent variables. The associated class of probabilistic models comprises an undirected discrete component and a directed hierarchical continuous component. The discrete component captures the distribution over the disconnected smooth manifolds induced by the continuous component. As a result, this class of models efficiently learns both the class of objects in an image, and their specific realization in pixels, from unsupervised data; and outperforms state-of-the-art methods on the permutation-invariant MNIST, Omniglot, and Caltech-101 Silhouettes datasets. | ["Deep learning", "Unsupervised Learning"] | https://openreview.net/forum?id=ryMxXPFex | https://openreview.net/pdf?id=ryMxXPFex | https://openreview.net/forum?id=ryMxXPFex¬eId=rk8jJ0VNx |
Bkf8UgS4l | ryMxXPFex | ICLR.cc/2017/conference/-/paper103/official/review | {"title": "Review", "rating": "8: Top 50% of accepted papers, clear accept", "review": "Paper proposes a novel Variational Encoder architecture that contains discrete variables. Model contains an undirected discrete component that captures distribution over disconnected manifolds and a directed hierarchical continuous component that models the actual manifolds (induced by the discrete variables). In essence the model clusters the data and at the same time learns a continuous manifold representation for the clusters. The training procedure for such models is also presented and is quite involved. Experiments illustrate state-of-the-art performance on public datasets (including MNIST, Omniglot, Caltech-101). \n\nOverall the model is interesting and could be useful in a variety of applications and domains. The approach is complex and somewhat mathematically involved. It's not exactly clear how the model compares or relates to other RBM formulations, particularly those that contain discrete latent variables and continuous outputs. As a prime example:\n\nGraham Taylor and Geoffrey Hinton. Factored conditional restricted Boltzmann machines for modeling motion style. In Proc. of the 26th International Conference on Machine Learning (ICML), 1025\u20131032, 2009.\n\nDiscussion of this should certainly be added. \n", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"} | review | 2017 | ICLR.cc/2017/conference | Discrete Variational Autoencoders | ["Jason Tyler Rolfe"] | Probabilistic models with discrete latent variables naturally capture datasets composed of discrete classes. However, they are difficult to train efficiently, since backpropagation through discrete variables is generally not possible. We present a novel method to train a class of probabilistic models with discrete latent variables using the variational autoencoder framework, including backpropagation through the discrete latent variables. The associated class of probabilistic models comprises an undirected discrete component and a directed hierarchical continuous component. The discrete component captures the distribution over the disconnected smooth manifolds induced by the continuous component. As a result, this class of models efficiently learns both the class of objects in an image, and their specific realization in pixels, from unsupervised data; and outperforms state-of-the-art methods on the permutation-invariant MNIST, Omniglot, and Caltech-101 Silhouettes datasets. | ["Deep learning", "Unsupervised Learning"] | https://openreview.net/forum?id=ryMxXPFex | https://openreview.net/pdf?id=ryMxXPFex | https://openreview.net/forum?id=ryMxXPFex¬eId=Bkf8UgS4l |
r16qB307e | ryMxXPFex | ICLR.cc/2017/conference/-/paper103/official/review | {"title": "Rich set of ideas on how to make VAEs work better. ", "rating": "8: Top 50% of accepted papers, clear accept", "review": "This is an interesting paper on how to handle reparameterization in VAEs when you have discrete variables. The idea is to introduce a smoothing transformation that is shared between the generative model and the recognition model (leading to cancellations). \nA second contribution is to introduce an RBM as the prior model P(z) and to use autoregressive connections in generative and recognition models. The whole package becomes a bit entangled and complex and it is hard to figure out what causes the claimed good performance. Experiments that study these contributions separately would have been nice. \nThe framework does become a little complex but this should not be a problem if nice software is delivered that can be used in a plug and play mode.\nOverall, the paper is very rich with ideas so I think it would be a great contribution to the conference. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Discrete Variational Autoencoders | ["Jason Tyler Rolfe"] | Probabilistic models with discrete latent variables naturally capture datasets composed of discrete classes. However, they are difficult to train efficiently, since backpropagation through discrete variables is generally not possible. We present a novel method to train a class of probabilistic models with discrete latent variables using the variational autoencoder framework, including backpropagation through the discrete latent variables. The associated class of probabilistic models comprises an undirected discrete component and a directed hierarchical continuous component. The discrete component captures the distribution over the disconnected smooth manifolds induced by the continuous component. As a result, this class of models efficiently learns both the class of objects in an image, and their specific realization in pixels, from unsupervised data; and outperforms state-of-the-art methods on the permutation-invariant MNIST, Omniglot, and Caltech-101 Silhouettes datasets. | ["Deep learning", "Unsupervised Learning"] | https://openreview.net/forum?id=ryMxXPFex | https://openreview.net/pdf?id=ryMxXPFex | https://openreview.net/forum?id=ryMxXPFex¬eId=r16qB307e |
H1KxcSgVg | ByG8A7cee | ICLR.cc/2017/conference/-/paper201/official/review | {"title": "Review", "rating": "5: Marginally below acceptance threshold", "review": "This paper introduces pointer-network neural networks, which are applied to referring expressions in three small-scale language modeling tasks: dialogue modeling, recipe modeling and news article modeling. When conditioned on the co-reference chain, the proposed models outperform standard sequence-to-sequence models with attention.\n\nThe proposed models are essentially variants of pointer networks with copy mechanisms (Gulcehre et al., 2016; Gu et al., 2016; Ling et al., 2016), which have been modified to take into account reference chains. As such, the main architectural novelty lies in 1) restricting the pointer mechanism to focus on co-referenced entities, 2) applying pointer mechanism to 2D arrays (tables), and 3) training with supervised alignments. Although useful in practice, these are minor contributions from an architectural perspective.\n\nThe empirical contributions are centred around measuring perplexity on the three language modeling tasks. Measuring perplexity is typical for standard language modeling tasks, but is really an unreliable proxy for dialogue modeling and recipe generation performance. In addition to this, both the dialogue and recipe tasks are tiny compared to standard language modeling tasks. This makes it difficult to evaluate the impact of the dialogue and recipe modeling results. For example, if one was to bootstrap from a larger corpus, it seems likely that a standard sequence-to-sequence model with attention would yield performance comparable to the proposed models (with enough data, the attention mechanism could learn to align referring entities by itself). The language modeling task on news article (Gigaword) seems to yield the most conclusive results. However, the dataset for this task is non-standard and results are provided for only a single baseline. Overall, this limits the conclusions we can draw from the empirical experiments.\n\n\nFinally, the paper itself contains many errors, including mathematical errors, grammatical errors and typos:\n- Eq. (1) is missing a sum over $z_i$.\n- \"into the a decoder LSTM\" -> \"into the decoder LSTM\"\n- \"denoted as his\" -> \"denoted as\"\n- \"Surprising,\" -> \"Surprisingly,\"\n- \"torkens\" -> \"tokens\"\n- \"if follows that the next token\" -> \"the next token\"\n- In the \"COREFERENCE BASED LANGUAGE MODEL\" sub-section, what does $M$ denote?\n- In the sentence: \"The attribute of each column is denoted as $s_c, where $c$ is the c-th attribute\". For these definitions to be make sense, $s_c$ has to be a one-hot vector. If yes, please clarify this in the text.\n- \"the weighted sum is performed\" -> \"the weighted sum is computed\"\n- \"a attribute\" -> \"an attribute\"\n- In the paragraph on Pointer Switch, change $p(z_{i,v} |s_{i,v}) = 1$ -> $p(z_{i,v} |s_{i,v}) = 0$.\n- In the \"Table Pointer\" paragraph, I assume you mean outer-product instead of cross-product? Otherwise, I don't see how the equations add up.\n\n\nOther comments:\n- For the \"Attention based decoder\", is the attention computed using the word embeddings themselves or the hidden states of the sentence encoder? Also, it applied only to the previous turn of the dialogue or to the entire dialogue history? Please clarify this.\n- What's the advantage of using an \"Entity state update\" rule, compared to a pointer network or copy network, which you used in the dialogue and recipe tasks? Please elaborate on this.\n- In the Related Work section, the following sentence is not quite accurate: \"For the task oriented dialogues, most of them embed the seq2seq model in traditional dialogue systems while our model queries the database directly.\". There are task-oriented dialogue models which do query databases during natural language generation. See, for example, \"A Network-based End-to-End Trainable Task-oriented Dialogue System\" by Wen et al.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Reference-Aware Language Models | ["Zichao Yang", "Phil Blunsom", "Chris Dyer", "Wang Ling"] | We propose a general class of language models that treat reference as an explicit stochastic latent variable. This architecture allows models to create mentions of entities and their attributes by accessing external databases (required by, e.g., dialogue generation and recipe generation) and internal state (required by, e.g. language models which are aware of coreference). This facilitates the incorporation of information that can be accessed in predictable locations in databases or dis- course context, even when the targets of the reference may be rare words. Experiments on three tasks show our model variants outperform models based on deterministic attention. | ["Natural language processing", "Deep learning"] | https://openreview.net/forum?id=ByG8A7cee | https://openreview.net/pdf?id=ByG8A7cee | https://openreview.net/forum?id=ByG8A7cee¬eId=H1KxcSgVg |
BJ_SQ64Vg | ByG8A7cee | ICLR.cc/2017/conference/-/paper201/official/review | {"title": "", "rating": "6: Marginally above acceptance threshold", "review": "This paper presents a new type of language model that treats entity references as latent variables. The paper is structured as three specialized models for three applications: dialog generation with references to database entries, recipe generation with references to ingredients, and text generation with coreference mentions.\n\nDespite some opaqueness in details that I will discuss later, the paper does a great job making the main idea coming through, which I think is quite interesting and definitely worth pursuing further. But it seems the paper was rushed into the deadline, as there are a few major weaknesses.\n\nThe first major weakness is that the claimed latent variables are hardly latent in the actual empirical evaluation. As clarified by the authors via pre-review QAs, all mentions were assumed to be given to all model variants, and so, it would seem like an over-claim to call these variables as latent when they are in fact treated as observed variables. Is it because the models with latent variables were too difficult to train right?\n\nA related problem is the use of perplexity as an evaluation measure when comparing reference-aware language models to vanilla language models. Essentially the authors are comparing two language models defined over different event space, which is not a fair comparison. Because mentions were assumed to be given for the reference-aware language models, and because of the fact that mention generators are designed similar to a pointer network, the probability scores over mentions will naturally be higher, compared to the regular language model that needs to consider a much bigger vocabulary set. The effect is analogous to comparing language models with aggressive UNK (and a small vocabulary set) to a language models with no UNK (and a much larger vocabulary set).\n\nTo mitigate this problem, the authors need to perform one of the following additional evaluations: either assuming no mention boundaries and marginalizing over all possibilities (treating latent variables as truly latent), or showing other types of evaluation beyond perplexity, for example, BLEU, METEOR, human evaluation etc on the corresponding generation task.\n\nThe other major weakness is writing in terms of technical accuracy and completeness. I found many details opaque and confusing even after QAs. I wonder if the main challenge that hinders the quality of writing has something to do with having three very specialized models in one paper, each having a lot of details to be worked out, which may have not been extremely important for the main story of the paper, but nonetheless not negligible in order to understand what is going on with the paper. Perhaps the authors can restructure the paper so that the most important details are clearly worked out in the main body of the paper, especially in terms of latent variable handling \u2014 how to make mention detection and conference resolution truly latent, and if and when entity update helps, which in the current version is not elaborated at all, as it is mentioned only very briefly for the third application (coreference resolution) without any empirical comparisons to motivate the update operation.\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Reference-Aware Language Models | ["Zichao Yang", "Phil Blunsom", "Chris Dyer", "Wang Ling"] | We propose a general class of language models that treat reference as an explicit stochastic latent variable. This architecture allows models to create mentions of entities and their attributes by accessing external databases (required by, e.g., dialogue generation and recipe generation) and internal state (required by, e.g. language models which are aware of coreference). This facilitates the incorporation of information that can be accessed in predictable locations in databases or dis- course context, even when the targets of the reference may be rare words. Experiments on three tasks show our model variants outperform models based on deterministic attention. | ["Natural language processing", "Deep learning"] | https://openreview.net/forum?id=ByG8A7cee | https://openreview.net/pdf?id=ByG8A7cee | https://openreview.net/forum?id=ByG8A7cee¬eId=BJ_SQ64Vg |
BJBPqNGNg | ByG8A7cee | ICLR.cc/2017/conference/-/paper201/official/review | {"title": "", "rating": "5: Marginally below acceptance threshold", "review": "This paper explores 3 language modeling applications with an explicit modeling of reference expressions: dialog, receipt generation and coreferences. While these are important tasks for NLP and the authors have done a number of experiments, the paper is limited for a few reasons:\n\n1. This paper is not clearly written and is pretty hard to follow some details. In particular, there are many obvious math errors, such as missing the marginalization sum in Eq (1), and P(z_{i,v}...) = 1 (should be 0 here) on page 5, pointer switch section.\n\n2. The major novelty seems to be the 2-dimensional attention from the table and the pointer to the 2-D table. These are more of a customization of existing work to a particular task with 2-D tables as a part of the input to seq2seq model with both attentions and pointer networks.\n\n3. The empirical results are not very conclusive yet, limited by either the relatively small data size, or the lack of well-established baseline for some new applications (e.g., the recipe generation task).\n\nOverall, this paper, as it is for now, is more suitable for a workshop rather than for the main conference.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Reference-Aware Language Models | ["Zichao Yang", "Phil Blunsom", "Chris Dyer", "Wang Ling"] | We propose a general class of language models that treat reference as an explicit stochastic latent variable. This architecture allows models to create mentions of entities and their attributes by accessing external databases (required by, e.g., dialogue generation and recipe generation) and internal state (required by, e.g. language models which are aware of coreference). This facilitates the incorporation of information that can be accessed in predictable locations in databases or dis- course context, even when the targets of the reference may be rare words. Experiments on three tasks show our model variants outperform models based on deterministic attention. | ["Natural language processing", "Deep learning"] | https://openreview.net/forum?id=ByG8A7cee | https://openreview.net/pdf?id=ByG8A7cee | https://openreview.net/forum?id=ByG8A7cee¬eId=BJBPqNGNg |
r1dPtY0me | HyxQzBceg | ICLR.cc/2017/conference/-/paper241/official/review | {"title": "A paper with interesting methods, but the presentation is a bit confusing", "rating": "6: Marginally above acceptance threshold", "review": "Thank you for an interesting read. I personally like the information bottleneck principle and am very happy to see its application to deep neural networks. To my knowledge, this is the first paper that applies IB to train deep networks (the original papers only presented the concept), but see below for the note of independent work claim. \n\nThe derivation of the variational lowerbound is very clear, even for those who are not very familiar with variational inference. Also the explanation of the IB principle is clear. Experimental results seem to be very promising.\n\nI found the presentation for the model a bit confusing. In variational inference/information maximisation, p usually denotes the model and q represents the \"inference engine\". This means the choice of inference method is independent to the modelling procedure. However the presented VIB assumed p(x, y) as the **underlying data distribution** (and approximated by the empirical distribution), thus here the model is actually q(y|z)p(z|x). Then the authors presented p(y|x) as the **predictive distribution** in page 8, paragraph 2 of section 4.2.3. Predictive in what sense? I guess you meant p(y|x) = \\int q(y|z) p(z|x) dz in this case, but this makes the two definitions contradict to each other!\n\nThe authors have made an interesting connection to variational auto-encoder and the warm-up training (by tuning beta). However, even when the loss function formula is the same to the variational lowerbound used in VAE (in this case beta = 1), the underlying model is different! For example, r(z) in VIB is the variational approximation to p(z) (which means r(z) is not a component in the model), while in VAE it is the prior distribution which is actually defined in the modelling procedure. Similaly p(z|x) in VIB is included in the model, while in VAE that is the approximate posterior and can be independently chosen (e.g. you can use p(x|z) as a deep NN but p(z|x) as a deep NN or a Gaussian process).\n\nIn summary, I think the presentation for the modelling procedure is unclear. I hope these point would be made clearer in revision since the current presentation makes me uncomfortable as a Bayesian person. In the VAE part, it's better to clearly mention the difference between VIB and VAE, and provide some intuitions if the VIB interpretation is preferred.\n\n\nTypos:\nEq. 9-11: did you mean q(y|z) instead of q(z|y)?\nFig 2 \"as beta becomes smaller\": did you mean \"larger\"?\n\n**claim for independent work**\nThe authors claimed that the manuscript presented an independent work to Chalk et al. 2016 which is online since May 2016. It seems to me that nowadays deep learning research is very competitve that many people publish the same idea at the same time. So I would trust this claim and commend the authors' honesty, but in case this is not true, I would not recommend the manuscript for acceptance.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Variational Information Bottleneck | ["Alexander A. Alemi", "Ian Fischer", "Joshua V. Dillon", "Kevin Murphy"] | We present a variational approximation to the information bottleneck of Tishby et al. (1999). This variational approach allows us to parameterize the information bottleneck model using a neural network and leverage the reparameterization trick for efficient training. We call this method “Deep Variational Information Bottleneck”, or Deep VIB. We show that models trained with the VIB objective outperform those that are trained with other forms of regularization, in terms of generalization performance and robustness to adversarial attack. | ["Theory", "Computer vision", "Deep learning", "Supervised Learning"] | https://openreview.net/forum?id=HyxQzBceg | https://openreview.net/pdf?id=HyxQzBceg | https://openreview.net/forum?id=HyxQzBceg¬eId=r1dPtY0me |
rJ_3n_ZNx | HyxQzBceg | ICLR.cc/2017/conference/-/paper241/official/review | {"title": "Great idea, lacking empirical section", "rating": "6: Marginally above acceptance threshold", "review": "Summary:\nThe paper \u201cDeep Variational Information Bottleneck\u201d explores the optimization of neural networks for variational approximations of the information bottleneck (IB; Tishby et al., 1999). On the example of MNIST, the authors show that this may be used for regularization or to improve robustness against adversarial attacks.\n\nReview:\nThe IB is potentially very useful for important applications (regularization, adversarial robustness, and privacy are mentioned in the paper). Combining the IB with recent advances in deep learning to make it more widely applicable is an excellent idea. But given that the theoretical contribution is a fairly straight-forward application of well-known ideas, I would have liked to see a stronger experimental section.\n\nSince the proposed approach allows us to scale IB, a better demonstration of this would have been on a larger problem than MNIST. It is also not clear whether the proposed approach will still work well to regularize more interesting networks with many layers.\n\nWhy is dropout not included in the quantitative comparison of robustness to adversarial examples (Figure 4)?\n\nHow was the number of samples (12) chosen?\n\nWhat are the error bars in Figure 1 (a)?\n\nOn page 7 the authors claim \u201cthe posterior covariance becomes larger\u201d as beta \u201cdecreases\u201d (increases?). Is this really the case? It\u2019s hard to judge based on Figure 1, since the figures are differently scaled.\n\nIt might be worth comparing to variational fair autoencoders (Louizos et al., 2016), which also try to learn representations minimizing the information shared with an aspect of the input.\n\nThe paper is well written and easy to follow.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Variational Information Bottleneck | ["Alexander A. Alemi", "Ian Fischer", "Joshua V. Dillon", "Kevin Murphy"] | We present a variational approximation to the information bottleneck of Tishby et al. (1999). This variational approach allows us to parameterize the information bottleneck model using a neural network and leverage the reparameterization trick for efficient training. We call this method “Deep Variational Information Bottleneck”, or Deep VIB. We show that models trained with the VIB objective outperform those that are trained with other forms of regularization, in terms of generalization performance and robustness to adversarial attack. | ["Theory", "Computer vision", "Deep learning", "Supervised Learning"] | https://openreview.net/forum?id=HyxQzBceg | https://openreview.net/pdf?id=HyxQzBceg | https://openreview.net/forum?id=HyxQzBceg¬eId=rJ_3n_ZNx |
By6srMMVx | HyxQzBceg | ICLR.cc/2017/conference/-/paper241/official/review | {"title": "Review", "rating": "7: Good paper, accept", "review": "Update: raised the score, because I think the arguments about adversarial examples are compelling. I think that the paper convincingly proves that this method acts as a decent regularizer, but I'm not convinced that it's a competitive regularizer. For example, I don't believe that there is sufficient evidence that it gives a better regularizer than dropout/normalization/etc. I also think that it will be much harder to tune than these other methods (discussed in my rebuttal reply). \n\n----\n\nSummary: If I understand correctly, this paper proposes to take the \"bottleneck\" term from variational autoencoders which pulls the latent variable towards a noise prior (like N(0,1)) and apply it in a supervised learning context where the reconstruction term log(p(x|z)) is replaced with the usual supervised cross-entropy objective. \n\nThe argument is that this is an effective regularizer and increases robustness to adversarial attacks. \n\nPros: \n\n-The presentation is quite good and the paper is easy to follow. \n\n-The idea is reasonable and the relationship to previous work is well described. \n\n-The robustness to adversarial examples experiment seems convincing, though I'm not an expert in this area. Is there any way to compare to an external quantitative baseline on robustness to adversarial examples? This would help a lot, since I'm not sure how the method here compares with other regularizers in terms of combatting adversarial examples. For example, if one uses a very high dropout rate, does this confer a comparable robustness to adversarial examples (perhaps at the expense of accuracy)? \n\nCons: \n\n-MNIST accuracy results don't seem very strong, unless I'm missing something. The Maxout paper from ICML 2013 listed many permutation invariant MNIST results with error rates below 1%. So the 1.13% error rate listed here doesn't necessarily prove that the method is a competitive regularizer. I also suspect that tuning this method to make it work well is harder than other regularizers like dropout. \n\n-There are many distinct architectural choices with this method, particularly in how many hidden layers come before and after z. For example, the output could directly follow z, or there could be several layers between z and the output. As far as I can tell the paper says that p(y | z) is a simple logistic regression (i.e. one weight matrix followed by softmax), but it's not obvious why this choice was made. Did it work best empirically? \n\nOther: \n\n-I wonder what would happen if you \"trained against\" the discovered adversarial examples while also using the method from this paper. Would it learn to have a higher variance p(z | x) when presented with an adversarial example? ", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Variational Information Bottleneck | ["Alexander A. Alemi", "Ian Fischer", "Joshua V. Dillon", "Kevin Murphy"] | We present a variational approximation to the information bottleneck of Tishby et al. (1999). This variational approach allows us to parameterize the information bottleneck model using a neural network and leverage the reparameterization trick for efficient training. We call this method “Deep Variational Information Bottleneck”, or Deep VIB. We show that models trained with the VIB objective outperform those that are trained with other forms of regularization, in terms of generalization performance and robustness to adversarial attack. | ["Theory", "Computer vision", "Deep learning", "Supervised Learning"] | https://openreview.net/forum?id=HyxQzBceg | https://openreview.net/pdf?id=HyxQzBceg | https://openreview.net/forum?id=HyxQzBceg¬eId=By6srMMVx |
B1atHaUEg | HycUbvcge | ICLR.cc/2017/conference/-/paper361/official/review | {"title": "Review", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes a deep extension of generalized CCA. The main contribution of the paper is deriving the gradient update for the GCCA objective.\n\nI disagree with the claim that \u201cthis is the first Multiview representation learning technique that combines the flexibility of nonlinear representation learning with the statistical power of incorporating information from many independent resources or views\u201d. [R1] proposes a Multiview representation learning method which is both non-linear and capable of handling more than 2 views. This is very much relevant to what authors are proposing. The objective function proposed in [R1] maximizes the correlation between views and minimizes the self and cross reconstruction errors. This is intuitively similar to nonlinear version of PCA+CCA for multiple views. Comparing these 2 methods is crucial to prove the usefulness of DGCCA and the paper is incomplete without this comparison. Authors should also change their strong claim.\n\nRelated work section is minimal. There are significant advances in 2-view non-linear representation learning which are worth mentioning. \n\nReferences:\n[R1] Janarthanan Rajendran, Mitesh M. Khapra, Sarath Chandar, Balaraman Ravindran: Bridge Correlational Neural Networks for Multilingual Multimodal Representation Learning. HLT-NAACL 2016: 171-181\n\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Deep Generalized Canonical Correlation Analysis | ["Adrian Benton", "Huda Khayrallah", "Biman Gujral", "Drew Reisinger", "Sheng Zhang", "Raman Arora"] | We present Deep Generalized Canonical Correlation Analysis (DGCCA) – a method for learning nonlinear transformations of arbitrarily many views of data, such that the resulting transformations are maximally informative of each other. While methods for nonlinear two-view representation learning (Deep CCA, (Andrew et al., 2013)) and linear many-view representation learning (Generalized CCA (Horst, 1961)) exist, DGCCA is the first CCA-style multiview representation learning technique that combines the flexibility of nonlinear (deep) representation learning with the statistical power of incorporating information from many independent sources, or views. We present the DGCCA formulation as well as an efficient stochastic optimization algorithm for solving it. We learn DGCCA representations on two distinct datasets for three downstream tasks: phonetic transcription from acoustic and articulatory measurements, and recommending hashtags and friends on a dataset of Twitter users. We find that DGCCA representations soundly beat existing methods at phonetic transcription and hashtag recommendation, and in general perform no worse than standard linear many-view techniques. | ["Unsupervised Learning", "Deep learning", "Multi-modal learning"] | https://openreview.net/forum?id=HycUbvcge | https://openreview.net/pdf?id=HycUbvcge | https://openreview.net/forum?id=HycUbvcge¬eId=B1atHaUEg |
BkPCVXwEl | HycUbvcge | ICLR.cc/2017/conference/-/paper361/official/review | {"title": "Deep Generalised Canonical Correlation Analysis", "rating": "6: Marginally above acceptance threshold", "review": "The authors propose a method that extends the non-linear two-view representation learning methods, and the linear multiview techniques, and combines information from multiple sources into a new non-linear representation learning techniques. \n\nIn general, the method is well described and seems to lead to benefits in different experiments of phonetic transcription of hashtag recommendation. Even if the method is mostly a extension of classical tools (the scheme learns a (deep) network for each view essentially), the combination of the different sources of information seems to be effective for the studied datasets. \n\nIt would be interesting to add or discuss the following issues:\n\n- what is the complexity of the proposed method, esp. the representation learning part?\n- would there by any alternative solution to combine the different networks/views? That could make the proposed solution more novel.\n- the experimental settings, especially in the synthetic experiments, should be more detailed. If possible, the datasets should be made available to encourage reproducibility. \n- the related work is far from complete unfortunately, especially from the perspective of the numerous multiview/multi-modal/multi-layer algorithms that have been proposed in the literature, in different applications domaines like image retrieval or classification, or bibliographic data for example (authors like A. Kumar, X. Dong, Ping-Yu Chen, M. Bronstein, and many others have proposed works in that direction in the last 5 years). No need to compare to all these works obviously, but a more complete description of the related could help appreciating better the true benefits of DGCCA.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Generalized Canonical Correlation Analysis | ["Adrian Benton", "Huda Khayrallah", "Biman Gujral", "Drew Reisinger", "Sheng Zhang", "Raman Arora"] | We present Deep Generalized Canonical Correlation Analysis (DGCCA) – a method for learning nonlinear transformations of arbitrarily many views of data, such that the resulting transformations are maximally informative of each other. While methods for nonlinear two-view representation learning (Deep CCA, (Andrew et al., 2013)) and linear many-view representation learning (Generalized CCA (Horst, 1961)) exist, DGCCA is the first CCA-style multiview representation learning technique that combines the flexibility of nonlinear (deep) representation learning with the statistical power of incorporating information from many independent sources, or views. We present the DGCCA formulation as well as an efficient stochastic optimization algorithm for solving it. We learn DGCCA representations on two distinct datasets for three downstream tasks: phonetic transcription from acoustic and articulatory measurements, and recommending hashtags and friends on a dataset of Twitter users. We find that DGCCA representations soundly beat existing methods at phonetic transcription and hashtag recommendation, and in general perform no worse than standard linear many-view techniques. | ["Unsupervised Learning", "Deep learning", "Multi-modal learning"] | https://openreview.net/forum?id=HycUbvcge | https://openreview.net/pdf?id=HycUbvcge | https://openreview.net/forum?id=HycUbvcge¬eId=BkPCVXwEl |
H1GCKjB4l | HycUbvcge | ICLR.cc/2017/conference/-/paper361/official/review | {"title": "Good paper", "rating": "7: Good paper, accept", "review": "The proposed method is simple and elegant; it builds upon the huge success of gradient based optimization for deep non-linear function approximators and combines it with established (linear) many-view CCA methods. A major contribution of this paper is the derivation of the gradients with respect to the non-linear encoding networks which project the different views into a common space. The derivation seems correct. In general this approach seems very interesting and I could imagine that it might be applicable to many other similarly structured problems.\nThe paper is well written; but it could be enhanced with an explicit description of the complete algorithm which also highlights how the joint embeddings G and U are updated. \n \nI don\u2019t have prior experience with CCA-style many-view techniques and it is therefore hard for me to judge the practical/empirical progress presented here. But the experiments seem reasonable convincing; although generally only performed on small and medium sized datasets.\n \nDetailed comments: \n\nThe colours or the sign of the x-axis in figure 3b seem to be flipped compared to figure 4.\n \nIt would be nice to additionally see a continuous (rainbow-coloured) version for Figures 2, 3 and 4 to better identify neighbouring datapoints; but more importantly: I\u2019d like to see how the average reconstruction error between the individual network outputs and the learned representation develop during training. Is the mismatch between different views on a validation/test-set a useful metric for cross validation? In general, it seems the method is sensitive to regularization and hyperparameter selection (because it has many more parameters compared to GCCA and different regularization parameters have been chosen for different views) and I wonder if there is a clear metric to optimize these.\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Generalized Canonical Correlation Analysis | ["Adrian Benton", "Huda Khayrallah", "Biman Gujral", "Drew Reisinger", "Sheng Zhang", "Raman Arora"] | We present Deep Generalized Canonical Correlation Analysis (DGCCA) – a method for learning nonlinear transformations of arbitrarily many views of data, such that the resulting transformations are maximally informative of each other. While methods for nonlinear two-view representation learning (Deep CCA, (Andrew et al., 2013)) and linear many-view representation learning (Generalized CCA (Horst, 1961)) exist, DGCCA is the first CCA-style multiview representation learning technique that combines the flexibility of nonlinear (deep) representation learning with the statistical power of incorporating information from many independent sources, or views. We present the DGCCA formulation as well as an efficient stochastic optimization algorithm for solving it. We learn DGCCA representations on two distinct datasets for three downstream tasks: phonetic transcription from acoustic and articulatory measurements, and recommending hashtags and friends on a dataset of Twitter users. We find that DGCCA representations soundly beat existing methods at phonetic transcription and hashtag recommendation, and in general perform no worse than standard linear many-view techniques. | ["Unsupervised Learning", "Deep learning", "Multi-modal learning"] | https://openreview.net/forum?id=HycUbvcge | https://openreview.net/pdf?id=HycUbvcge | https://openreview.net/forum?id=HycUbvcge¬eId=H1GCKjB4l |
HJod5BH4l | rkYmiD9lg | ICLR.cc/2017/conference/-/paper422/official/review | {"title": "", "rating": "6: Marginally above acceptance threshold", "review": "The paper describes how to use a tensor factorization method called Tensor Train for modeling the interactions between features for supervised classification tasks. Tensor Train approximates tensors of any dimensions using low rank products of matrices. The rank is used as a parameter for controlling the complexity of the approximation. Experiments are performed on different datasets for binary classification problems.\n\nThe core of the paper consists in demonstrating how the TT formalism developed by one of the authors could be adapted for modeling interactions between features. Another contribution is a gradient algorithm that exploits the geometrical structure of the factorization. These ideas are probably new in Machine learning. The algorithm itself is of reasonable complexity for the inference and could probably be adapted to large size problems although this is not the case for the experiments here.\n\nThe experimental section is not well structured, it is incomplete and could be improved. We miss a description of the datasets characteristics. The performance on the different datasets are not provided. Each dataset has been used for illustrating one aspect of the model, but you could also provide classification performance and a comparison with baselines for all the experiments. The experiments on the 2 UCI datasets show optimization performance on the training set, you could provide the same curves on test sets to show how the algorithm generalizes. The comparison to other approaches (section 8.4) is only performed on artificial data, which are designed with interacting features and are not representative of diverse situations. The same holds for the role of Dropout. The comparison on the Movielens dataset is incomplete. Besides, all the tests are performed on small size problems.\n\nOverall, there are original contributions which could be worth a publication. The experiments are incomplete and not conclusive. A more detailed comparison with competing methods, like Factorization Machines, could also improve the paper. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Exponential Machines | ["Alexander Novikov", "Mikhail Trofimov", "Ivan Oseledets"] | Modeling interactions between features improves the performance of machine learning solutions in many domains (e.g. recommender systems or sentiment analysis). In this paper, we introduce Exponential Machines (ExM), a predictor that models all interactions of every order. The key idea is to represent an exponentially large tensor of parameters in a factorized format called Tensor Train (TT). The Tensor Train format regularizes the model and lets you control the number of underlying parameters. To train the model, we develop a stochastic Riemannian optimization procedure, which allows us to fit tensors with 2^160 entries. We show that the model achieves state-of-the-art performance on synthetic data with high-order interactions and that it works on par with high-order factorization machines on a recommender system dataset MovieLens 100K. | ["Supervised Learning", "Optimization"] | https://openreview.net/forum?id=rkYmiD9lg | https://openreview.net/pdf?id=rkYmiD9lg | https://openreview.net/forum?id=rkYmiD9lg¬eId=HJod5BH4l |
Sy9wO0tSl | rkYmiD9lg | ICLR.cc/2017/conference/-/paper422/official/review | {"title": "Nice idea, but experiments are very preliminary", "rating": "5: Marginally below acceptance threshold", "review": "The paper presents an application of a tensor factorization to linear models, which allows to consider higher-order interactions between variables in classification (and regression) problems, and that maintains computational feasibility, being linear in the dimension. The factorization employed is based on the TT format, first proposed by Oseledests (2011). The authors also propose the adoption of a Riemannian optimization scheme to explicit consider the geometry of the tensor manifold, and thus speed up convergence.\n\nThe paper in general is well written, it presents an interesting application of the TT tensor format for linear models (together with an application of Riemannian optimization), which in my opinion is quite interesting since it has a wide range of possible applications in different algorithms in machine learning.\n\nOn the other side, I have some concerns are about the experimental part, which I consider not at the level of the rest of the paper, for instance in terms of number of experiments on real datasets, role of dropout in real datasets, comparison with other algorithms on real datasets. Moreover the authors do not take into account explicitly the problem of the choice of the rank to be used in the experiments. In general the experimental section seems a collection of preliminary experiments where different aspects have been tested by not in a organic way.\n\nI think the paper is close to a weak acceptance / weak rejection, I don't rate it as a full acceptance paper, mainly due to the non-satisfactory experiment setting. In case of extra experiments confirming the goodness of the approach, I believe the paper could have much better scores.\n\nSome minor comments:\n-formula 2: Obvious comment: learning the parameters of the model in (1) can be done as in (2), but also in other ways, depending on the approach you are using.\n-the fact that the rank is bounded by 2r, before formula 9, is explained in Lubich et al., 2015?\n-after formula 10: why the N projections in total they cost O(dr^2(r+N)), it should be O(Ndr^2(r+1)), no? since each of the elements of the summation has rank 1, and the cost for each of them is O(dr^2(r+TT_rank(Z)^2)), where TT-rank(Z)=1. Am I wrong?\n-section 6.2: can you explain why the random initialization freezes the convergence? This seems interesting but not motivated. Any guess?\n-section 6.3: you adopt dropout: can you comment in particular on the advantages it gives in the context of the exponential machines? did you use it on real datasets?\n-how do you choose r_0 in you experiments? with a validation set?\n-in section 7: why you don't have x_1 x_2 among the variables?\n-section 8: there is a typo in \"experiments\"\n-section 8.1: \"We simplicity, we binarized\" I think there's a problem with the English language in this sentence\n-section 8.3: \"we report that dropout helps\".. this is quite general statement, only tested on a synthetic dataset\n-section 8.5: can you provide more results for this dataset, for instance in terms of training and inference time? or test wrt other algorithms?", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Exponential Machines | ["Alexander Novikov", "Mikhail Trofimov", "Ivan Oseledets"] | Modeling interactions between features improves the performance of machine learning solutions in many domains (e.g. recommender systems or sentiment analysis). In this paper, we introduce Exponential Machines (ExM), a predictor that models all interactions of every order. The key idea is to represent an exponentially large tensor of parameters in a factorized format called Tensor Train (TT). The Tensor Train format regularizes the model and lets you control the number of underlying parameters. To train the model, we develop a stochastic Riemannian optimization procedure, which allows us to fit tensors with 2^160 entries. We show that the model achieves state-of-the-art performance on synthetic data with high-order interactions and that it works on par with high-order factorization machines on a recommender system dataset MovieLens 100K. | ["Supervised Learning", "Optimization"] | https://openreview.net/forum?id=rkYmiD9lg | https://openreview.net/pdf?id=rkYmiD9lg | https://openreview.net/forum?id=rkYmiD9lg¬eId=Sy9wO0tSl |
BylWr2YHg | rkYmiD9lg | ICLR.cc/2017/conference/-/paper422/official/review | {"title": "Interesting idea, analysis could be improved", "rating": "6: Marginally above acceptance threshold", "review": "This paper introduces a polynomial linear model for supervised classification tasks. The model is based on a combination of the Tensor Train (TT) tensor decomposition method and a form of stochastic Riemannian optimization. A few empirical experiments are performed that demonstrate the good performance of the proposed model relative to appropriate baselines.\n\nFrom a theoretical standpoint, I think the approach is interesting and elegant. The main machinery underlying this work are the TT decomposition and the geometric structure of the manifold of tensors with fixed TT-rank, which have been established in prior work. The novelty of this paper is in the combination of this machinery to form an efficient polynomial linear model. As such, I would have hoped that the paper mainly focused on the efficacy of this combination and how it is superior to obvious alternatives. For example, I would have really appreciated seeing how FMs performed when optimized over the manifold of positive definite matrices, as another reviewer mentioned. Instead, there is a bit too much effort devoted to explaining prior work.\n\nI think the empirical analysis could be substantially improved. I am particularly puzzled by the significant performance boost obtained from initializing with the ordinary logistic regression solution. I would have liked some further analysis of this effect, especially whether or not it is possible to obtain a similar performance boost with other models. Regarding the synthetic data, I think an important baseline would be against a vanilla feed forward neural network, which would help readers understand how complicated the interactions are and how difficult the dataset is to model. I agree with the previous reviewer regarding a variety of other possible improvements to the experimental section.\n\nA few typos: 'Bernoulli distrbution', 'reproduce the experiemnts', 'generilize better'.\n\nOverall, I am on the fence regarding this paper. The main idea is quite good, but insufficient attention was devoted to analyzing the aspects of the model that make it interesting and novel.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Exponential Machines | ["Alexander Novikov", "Mikhail Trofimov", "Ivan Oseledets"] | Modeling interactions between features improves the performance of machine learning solutions in many domains (e.g. recommender systems or sentiment analysis). In this paper, we introduce Exponential Machines (ExM), a predictor that models all interactions of every order. The key idea is to represent an exponentially large tensor of parameters in a factorized format called Tensor Train (TT). The Tensor Train format regularizes the model and lets you control the number of underlying parameters. To train the model, we develop a stochastic Riemannian optimization procedure, which allows us to fit tensors with 2^160 entries. We show that the model achieves state-of-the-art performance on synthetic data with high-order interactions and that it works on par with high-order factorization machines on a recommender system dataset MovieLens 100K. | ["Supervised Learning", "Optimization"] | https://openreview.net/forum?id=rkYmiD9lg | https://openreview.net/pdf?id=rkYmiD9lg | https://openreview.net/forum?id=rkYmiD9lg¬eId=BylWr2YHg |
BJMAByP4l | rkYmiD9lg | ICLR.cc/2017/conference/-/paper422/official/review | {"title": "A good paper ", "rating": "7: Good paper, accept", "review": "This paper proposes to use the tensor train (TT) decomposition to represent the full polynomial linear model. The TT form can reduce the computation complexity in both of inference and model training. A stochastic gradient over a Riemann Manifold has been proposed to solve the TT based formulation. The empirical experiments validate the proposed method.\n\nThe proposed approach is very interesting and novel for me. I would like to vote acceptance on this paper. My only suggestion is to include the computational complexity per iteration.\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Exponential Machines | ["Alexander Novikov", "Mikhail Trofimov", "Ivan Oseledets"] | Modeling interactions between features improves the performance of machine learning solutions in many domains (e.g. recommender systems or sentiment analysis). In this paper, we introduce Exponential Machines (ExM), a predictor that models all interactions of every order. The key idea is to represent an exponentially large tensor of parameters in a factorized format called Tensor Train (TT). The Tensor Train format regularizes the model and lets you control the number of underlying parameters. To train the model, we develop a stochastic Riemannian optimization procedure, which allows us to fit tensors with 2^160 entries. We show that the model achieves state-of-the-art performance on synthetic data with high-order interactions and that it works on par with high-order factorization machines on a recommender system dataset MovieLens 100K. | ["Supervised Learning", "Optimization"] | https://openreview.net/forum?id=rkYmiD9lg | https://openreview.net/pdf?id=rkYmiD9lg | https://openreview.net/forum?id=rkYmiD9lg¬eId=BJMAByP4l |
SJuj1-NSe | Sy6iJDqlx | ICLR.cc/2017/conference/-/paper349/official/review | {"title": "", "rating": "7: Good paper, accept", "review": "This paper studies the problem of transferring solutions of existing tasks to tackle a novel task under the framework of reinforcement learning and identifies two important issues of avoiding negative transfer and being selective transfer. The proposed approach is based on a convex combination of existing solutions and the being-learned solution to the novel task. The non-negative weight of each solution implies that the solution of negative effect is ignored and more weights are allocated to more relevant solution in each state. This paper derives this so-called \"A2T\" learning algorithm for policy transfer and value transfer for REINFORCE and ACTOR-CRITIC algorithms and experiments with synthetic Chain World and Puddle World simulation and Atari 2600 game Pong. \n+This paper presents a novel approach for transfer reinforcement learning.\n+The experiments are cleverly designed to demonstrate the ability of the proposed method.\n-An important aspect of transfer learning is that the algorithm can automatically figure out if the existing solutions to known tasks are sufficient to solve the novel task so that it can save the time and energy of learning-from-scratch. This issue is not studied in this paper as most of experiments have a learning-from-scratch solution as base network. It will be interesting to see how well the algorithm performs without base network. In addition, from Figure 3, 5 and 6, the proposed algorithm seems to accelerate the learning speed, but the overall network seems not better than the solo base network. It will be more convincing to show some example that existing solutions are complementary to the base network.\n-If ignoring the base network, the proposed network can be considered as ensemble reinforcement learning that take advantages of learned agents with different expertise to solve the novel task. ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Attend, Adapt and Transfer: Attentive Deep Architecture for Adaptive Transfer from multiple sources in the same domain | ["Janarthanan Rajendran", "Aravind Lakshminarayanan", "Mitesh M. Khapra", "Prasanna P", "Balaraman Ravindran"] | Transferring knowledge from prior source tasks in solving a new target task can be useful in several learning applications. The application of transfer poses two serious challenges which have not been adequately addressed. First, the agent should be able to avoid negative transfer, which happens when the transfer hampers or slows down the learning instead of helping it. Second, the agent should be able to selectively transfer, which is the ability to select and transfer from different and multiple source tasks for different parts of the state space of the target task. We propose A2T (Attend Adapt and Transfer), an attentive deep architecture which adapts and transfers from these source tasks. Our model is generic enough to effect transfer of either policies or value functions. Empirical evaluations on different learning algorithms show that A2T is an effective architecture for transfer by being able to avoid negative transfer while transferring selectively from multiple source tasks in the same domain. | ["Deep learning", "Reinforcement Learning", "Transfer Learning"] | https://openreview.net/forum?id=Sy6iJDqlx | https://openreview.net/pdf?id=Sy6iJDqlx | https://openreview.net/forum?id=Sy6iJDqlx¬eId=SJuj1-NSe |
Sy-SiOZNe | Sy6iJDqlx | ICLR.cc/2017/conference/-/paper349/official/review | {"title": "Review", "rating": "7: Good paper, accept", "review": "The paper tackles important problems in multi-task reinforcement learning: avoid negative transfer and allow finer selective transfer. The method is based on soft attention mechanism, very general, and demonstrated to be applicable in both policy gradient and value iteration methods. The introduction of base network allows learning new policy if the prior policies aren't directly applicable. State-dependent sub policy selection allows finer control and can be thought of assigning state space to different sub policies/experts. The tasks are relatively simplistic but sufficient to demonstrate the benefits. One limitation is that the method is simple and the results/claims are mostly empirical. It would be interesting to see extensions to option-based framework, stochastic hard attention mechanism, sub-policy pruning, progressive networks. \n\nIn figure 6, the read curve seems to perform worse than the rest in terms of final performance. Perhaps alternative information to put with figures is the attention mask activation statistics during learning, so that we may observe that it learns to turn off adversarial sub-policies and rely on newly learned base policy mostly. This is also generally good to check to see if any weird co-adaptation is happening. ", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Attend, Adapt and Transfer: Attentive Deep Architecture for Adaptive Transfer from multiple sources in the same domain | ["Janarthanan Rajendran", "Aravind Lakshminarayanan", "Mitesh M. Khapra", "Prasanna P", "Balaraman Ravindran"] | Transferring knowledge from prior source tasks in solving a new target task can be useful in several learning applications. The application of transfer poses two serious challenges which have not been adequately addressed. First, the agent should be able to avoid negative transfer, which happens when the transfer hampers or slows down the learning instead of helping it. Second, the agent should be able to selectively transfer, which is the ability to select and transfer from different and multiple source tasks for different parts of the state space of the target task. We propose A2T (Attend Adapt and Transfer), an attentive deep architecture which adapts and transfers from these source tasks. Our model is generic enough to effect transfer of either policies or value functions. Empirical evaluations on different learning algorithms show that A2T is an effective architecture for transfer by being able to avoid negative transfer while transferring selectively from multiple source tasks in the same domain. | ["Deep learning", "Reinforcement Learning", "Transfer Learning"] | https://openreview.net/forum?id=Sy6iJDqlx | https://openreview.net/pdf?id=Sy6iJDqlx | https://openreview.net/forum?id=Sy6iJDqlx¬eId=Sy-SiOZNe |
S1CEWJdSl | Sy6iJDqlx | ICLR.cc/2017/conference/-/paper349/official/review | {"title": "Final Review: Learned convex combination of many fixed and a jointly learned expert is used to represent action policies in proof-of-concept, transfer/hierarchical RL, settings.", "rating": "7: Good paper, accept", "review": "In this paper a well known soft mixture of experts model is adapted for, and applied to, a specific type of transfer learning problem in reinforcement learning (RL), namely transfer of action policies and value functions between similar tasks. Although not treated as such, the experimental setup is reminiscent of hierarchical RL works, an aspect which the paper does not consider at length, regrettably.\nOne possible implication of this work is that architecture and even learning algorithm choices could simply be stated in terms of the objective of the target task, rather than being hand-engineered by the experimenter. This is clearly an interesting direction of future work which the paper illuminates.\n\n\nPros:\nThe paper diligently explains how the network architecture fits in with various widely used reinforcement learning setups, which does facilitate continuation of this work.\nThe experiments are good proofs of concept, but do not go beyond that i.m.h.o. \nEven so, this work provides convincing clues that collections of deep networks, which were trained on not entirely different tasks, generalize better to related tasks when used together rather than through conventional transfer learning (e.g. fine-tuning).\n\nCons:\nAs the paper well recounts in the related work section, libraries of fixed policies have long been formally proposed for reuse while learning similar tasks. Indeed, it is well understood in hierarchical RL literature that it can be beneficial to reuse libraries of fixed (Fernandez & Veloso 2006) or jointly learned policies which may not apply to the entire state space, e.g. options (Pricop et. al). What is not well understood is how to build such libraries, and this paper does not convincingly shed light in that direction, as far as I can tell.\nThe transfer tasks have been picked to effectively illustrate the potential of the proposed architecture, but the paper does not tackle negative transfer or compositional reuse in well known challenging situations outlined in previous work (e.g. Parisotto et. al 2015, Rusu el. al 2015, 2016).\nSince the main contributions are of an empirical nature, I am curious how the results shown in figures 6 & 7 look plotted against wall-clock time, since relatively low data efficiency is not a limitation for achieving perfect play in Pong (see Mnih. et al, 2015). It would be more illuminating to consider tasks where final performance is plausibly limited by data availability. It would also be interesting if the presented results were achieved with reduced amounts of computation, or reduced representation sizes compared to learning from scratch, especially when one of the useful source tasks is an actual policy trained on the target task.\nFinally, it is perhaps underwhelming that it takes a quarter of the data required for learning Pong from scratch just to figure out that a perfect Pong policy is already in the expert library. Simply evaluating each expert for 10 episodes and using an average-score-weighted majority vote to mix action choices would probably achieve the same final performance for a smaller fraction of the data. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Attend, Adapt and Transfer: Attentive Deep Architecture for Adaptive Transfer from multiple sources in the same domain | ["Janarthanan Rajendran", "Aravind Lakshminarayanan", "Mitesh M. Khapra", "Prasanna P", "Balaraman Ravindran"] | Transferring knowledge from prior source tasks in solving a new target task can be useful in several learning applications. The application of transfer poses two serious challenges which have not been adequately addressed. First, the agent should be able to avoid negative transfer, which happens when the transfer hampers or slows down the learning instead of helping it. Second, the agent should be able to selectively transfer, which is the ability to select and transfer from different and multiple source tasks for different parts of the state space of the target task. We propose A2T (Attend Adapt and Transfer), an attentive deep architecture which adapts and transfers from these source tasks. Our model is generic enough to effect transfer of either policies or value functions. Empirical evaluations on different learning algorithms show that A2T is an effective architecture for transfer by being able to avoid negative transfer while transferring selectively from multiple source tasks in the same domain. | ["Deep learning", "Reinforcement Learning", "Transfer Learning"] | https://openreview.net/forum?id=Sy6iJDqlx | https://openreview.net/pdf?id=Sy6iJDqlx | https://openreview.net/forum?id=Sy6iJDqlx¬eId=S1CEWJdSl |
Hy3_VU-Eg | S1jmAotxg | ICLR.cc/2017/conference/-/paper129/official/review | {"title": "Good paper", "rating": "8: Top 50% of accepted papers, clear accept", "review": "Summary: This is the first work to investigate stick-breaking priors, and corresponding inference methods, for use in VAEs. The background material is explained clearly, as well as the explanation of the priors and posteriors and their DNCP forms. The paper is really well written.\n\nIn experiments, they find that stick-breaking priors does not generally improve upon spherically Gaussian priors in the completely unsupervised setting, when measured w.r.t. log-likelihood. The fact that they do report this 'negative' result suggests good scientific taste. In a semi-supervised setting, the results are better.\n\nComments:\n- sec 2.1: There is plenty of previous work with non-Gaussian p(z): DRAW, the generative ResNet paper in the IAF paper, Ladder VAEs, etc.\n- sec 2.2: two comma's\n- text flow eq 6: please refer to appendix with the closed-form KL divergence\n- \"The v's are sampled via\" => \"In the posterior, the v's are sampled via\". It's not clear you're talking about the posterior here, instead of the prior.\n- The last paragraph of section 4 is great.\n- Sec 7.1: \"Density estimation\" => Technically you're also doing mass estimation.\n- Sec 7.1: 100 IS samples is a bit on the low side. \n- Figure 3(f). Interesting that k-NN works so well on raw pixels.\n\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Stick-Breaking Variational Autoencoders | ["Eric Nalisnick", "Padhraic Smyth"] | We extend Stochastic Gradient Variational Bayes to perform posterior inference for the weights of Stick-Breaking processes. This development allows us to define a Stick-Breaking Variational Autoencoder (SB-VAE), a Bayesian nonparametric version of the variational autoencoder that has a latent representation with stochastic dimensionality. We experimentally demonstrate that the SB-VAE, and a semi-supervised variant, learn highly discriminative latent representations that often outperform the Gaussian VAE’s. | ["Deep learning", "Unsupervised Learning", "Semi-Supervised Learning"] | https://openreview.net/forum?id=S1jmAotxg | https://openreview.net/pdf?id=S1jmAotxg | https://openreview.net/forum?id=S1jmAotxg¬eId=Hy3_VU-Eg |
BkgfmZf4l | S1jmAotxg | ICLR.cc/2017/conference/-/paper129/official/review | {"title": "Motivation is not very clear but good paper overall", "rating": "8: Top 50% of accepted papers, clear accept", "review": "This paper presents an approach which modifies the variational auto-encoder (VAE) framework so as to use stochastic latent dimensionality. This is achieved by using an inherently infinite prior, the stick-breaking process. This is coupled with inference tailored to this model, specifically the Kumaraswamy distribution as an approximate variational posterior. The resulting model is named the SB-VAE which also has a semi-supervised extension, in similar vein to the original VAE paper.\n\nThere's a lot of interest in VAEs these days; many lines of work seek to achieve automatic \"black-box\" inference in these models. For example, the authors themselves mention parallel work by Blei's lab (also others) towards this direction. However, there's a lot of merit in investigating more bespoke solutions to new models, which is what the authors are doing in this paper. Indeed, a (useful) side-effect of providing efficient inference for the SB-VAE is drawing attention to the use of the Kumaraswamy distribution which hasn't been popular in ML.\n\nAlthough the paper is in general well structured, I found it confusing at parts. I think the major source of confusion comes from the fact that the model specification and model inference are discussed in a somehow mixed manner. The pre-review questions clarified most parts.\n\nI have two main concerns regarding the methodology and motivation of this paper. Firstly, conditioning the model directly on the stick-breaking weights seems a little odd. I initially thought that there was some mixture probabilistic model involved, but this is not the case. To be fair, the authors discuss about this issue (which became clearer to me after the pre-review questions), and explain that they're investigating the apparently challenging problem of using a base distribution G_0. The question is whether their relaxation is still useful. From the experiments it seems that the method is at least competitive, so the answer is yes. Hopefully an extension will come in the future, as the authors mention.\n\nThe second concern is about the motivation of this method. It seems that the paper fails to clearly explain in a convincing way why it is beneficial to reformulate the VAE as a SB-VAE. I understand that the non-parametric property induced by the prior might result in better capacity control, however I feel that this advantage (and potentially others which are still unclear to me) is not sufficiently explained and demonstrated. Perhaps some comparison with a dropout approach or a more thorough discussion related to dropout would make this clearer.\n\nOverall, I found this to be an interesting paper, it would be a good fit for ICLR.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Stick-Breaking Variational Autoencoders | ["Eric Nalisnick", "Padhraic Smyth"] | We extend Stochastic Gradient Variational Bayes to perform posterior inference for the weights of Stick-Breaking processes. This development allows us to define a Stick-Breaking Variational Autoencoder (SB-VAE), a Bayesian nonparametric version of the variational autoencoder that has a latent representation with stochastic dimensionality. We experimentally demonstrate that the SB-VAE, and a semi-supervised variant, learn highly discriminative latent representations that often outperform the Gaussian VAE’s. | ["Deep learning", "Unsupervised Learning", "Semi-Supervised Learning"] | https://openreview.net/forum?id=S1jmAotxg | https://openreview.net/pdf?id=S1jmAotxg | https://openreview.net/forum?id=S1jmAotxg¬eId=BkgfmZf4l |
HkqgbsZEl | S1jmAotxg | ICLR.cc/2017/conference/-/paper129/official/review | {"title": "", "rating": "4: Ok but not good enough - rejection", "review": "The paper attempts to combine Variational Auto-Encoders with the Stick-Breaking process. The motivation is to tackle the component collapsing and have a representation with stochastic dimensionality. To demonstrate the merit of their approach, the authors test this model on MNIST and SVHN in an unsupervised and semi-supervised fashion.\nAfter reading the paper in more detail, I find that the claim that the dimensionality of the latent variable is stochastic does not seem quite correct: all latent variables are \"used\" (which actually enable backpropagation) but the latent variables are parametrized differently (into $\\pi$) and the decoding process is altered as to give the impression of sparsity. The way all these latent variables are used does not involve any marginalization but is very similar to the common soft-gating mechanism already used in LSTM or attentional model.\nWith respect to the Figure 5b showing the decoder input weights: component collapsing probably does not have the same effect as Gaussian prior. $\\pi$ is positive therefore having a very small average value might mean that its value is close to zero most of the time, not requiring any update on the weight. For the standard Gaussian prior, component collapsing means having a very noisy input with no signal involved, which forces the decoder to shut down this channel, i.e. have small incoming weights from this collapsed variable.\nAdding a histogram of the latent variables in addition to that might help decide if the associated weights are relatively large because they are actually used or if it's because the inputs are zero anyway.\nThe semi-supervised results are better than a weaker version of the model used in (Kingma et al., 2014), but as to have a fairer comparison, the results should be compared with the M1+M2 model in that paper, even if that requires also using two VAEs.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Stick-Breaking Variational Autoencoders | ["Eric Nalisnick", "Padhraic Smyth"] | We extend Stochastic Gradient Variational Bayes to perform posterior inference for the weights of Stick-Breaking processes. This development allows us to define a Stick-Breaking Variational Autoencoder (SB-VAE), a Bayesian nonparametric version of the variational autoencoder that has a latent representation with stochastic dimensionality. We experimentally demonstrate that the SB-VAE, and a semi-supervised variant, learn highly discriminative latent representations that often outperform the Gaussian VAE’s. | ["Deep learning", "Unsupervised Learning", "Semi-Supervised Learning"] | https://openreview.net/forum?id=S1jmAotxg | https://openreview.net/pdf?id=S1jmAotxg | https://openreview.net/forum?id=S1jmAotxg¬eId=HkqgbsZEl |
HkicR4ZEg | S1LVSrcge | ICLR.cc/2017/conference/-/paper246/official/review | {"title": "Review: Variable Computation in Recurrent Neural Networks", "rating": "7: Good paper, accept", "review": "This paper describes a simple but clever method for allowing variable amounts of computation at each time step in RNNs. The new architecture seems to outperform vanilla RNNs on various sequence modelling tasks. Visualizations of the assignment of computational resources over time support the hypothesis that the model is able to learn to assign more computations whenever longer longer term dependencies need to be taken into account.\n\nThe proposed model is evaluated on a multitude of tasks and its ability to outperform similar architectures seems consistent. Some of the tasks allow for an interesting analysis of the amount of computation the model requests at each time step. It\u2019s very interesting to see how the model seems to use more resources at the start of each word or ASCII character. I also like the investigation of the effect of imposing a pattern of computational budget assignment which uses prior knowledge about the task. The superior performance of the architecture is impressive but I\u2019m not yet convinced that the baseline models had an equal number of hyperparameters to tune. I\u2019ll come back to this point in the next paragraph because it\u2019s mainly a clarity issue.\n\nThe abstract claims that the model is computationally more efficient than regular RNNs. There are no wall time measurements supporting this claim. While the model is theoretically able to save computations, the points made by the paper are clearly more conceptual and about the ability of the model to choose how to allocate its resources. This makes the paper interesting enough by itself but the claims of computational gains are misleading without actual results to back them up. I also find it unfortunate that it\u2019s not clear from the text how the hyperparameter \\bar{m} was chosen. Whether it was chosen randomly or set using a hyperparameter search on held-out data influences the fairness of a comparison with RNNs which did not have a similar type of hyperparameter for controlling regularization like for example dropout or weight noise (even if regularization of RNNs is a bit tricky). I don\u2019t consider this a very serious flaw because I\u2019m impressed enough by the fact that the new architecture achieves roughly similar performance while learning to allocate resources but I do think that details of this type are too important to be absent from the text. Even if the superior performance is due to this extra regularization controlling parameter it can actually be seen as a useful part of the architecture but it would be nice to know how sensitive the model is to its precise value.\n\nTo my knowledge, the proposed architecture is novel. The way the amount of computation is determined is unlike other methods for variable computation I have seen and quite inventive. Originality is one of this paper\u2019s strongest points. \n\nIt\u2019s currently hard to predict whether this method for variable computation will be used a lot in practice given that this also depends on how feasible it is to obtain actual computational gains at the hardware level. That said, the architecture may turn out to be useful for learning long-term dependencies. I also think that the interpretability of the value m_t is a nice property of the method and that it\u2019s visualizations are very interesting. It might shed some more light into what makes certain tasks difficult for RNNs. \n\nPros:\nOriginal clever idea.\nNice interesting visualizations.\nInteresting experiments.\n\nCons:\nSome experimental details are not clear.\nI\u2019m not convinced of the strength of the baseline.\nThe paper shouldn\u2019t claim actual computational savings without reporting wall-clock times.\n\nEdit:\nI'm very positively impressed by the way the authors ended up addressing the biggest concerns I had about the paper and raised my score. Adding an LSTM baseline and results with a GRU version of the model significantly improves the empirical quality of the paper. On top of that, the authors addressed my question about some experimental detail I found important and promised to change the wording of the paper to remove confusion about whether the computational savings are conceptual or in actual wall time. I think it's fine that they are conceptual only as long as this is clear from the paper and abstract. I want to make clear to the AC that since the changes to the paper are currently still promises, my new score should be assumed to apply to an updated version of the paper in which the aforementioned concerns have indeed been addressed. \n\nEdit: \nSince I didn't know that the difference with the SOTA for some of these tasks was so large, I had to lower my score again after learning about this. I still think it's a good paper but with these results I cannot say that it stands out.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Variable Computation in Recurrent Neural Networks | ["Yacine Jernite", "Edouard Grave", "Armand Joulin", "Tomas Mikolov"] | Recurrent neural networks (RNNs) have been used extensively and with increasing success to model various types of sequential data. Much of this progress has been achieved through devising recurrent units and architectures with the flexibility to capture complex statistics in the data, such as long range dependency or localized attention phenomena. However, while many sequential data (such as video, speech or language) can have highly variable information flow, most recurrent models still consume input features at a constant rate and perform a constant number of computations per time step, which can be detrimental to both speed and model capacity. In this paper, we explore a modification to existing recurrent units which allows them to learn to vary the amount of computation they perform at each step, without prior knowledge of the sequence's time structure. We show experimentally that not only do our models require fewer operations, they also lead to better performance overall on evaluation tasks. | ["Natural language processing", "Deep learning"] | https://openreview.net/forum?id=S1LVSrcge | https://openreview.net/pdf?id=S1LVSrcge | https://openreview.net/forum?id=S1LVSrcge¬eId=HkicR4ZEg |
rJ0I-lfEe | S1LVSrcge | ICLR.cc/2017/conference/-/paper246/official/review | {"title": "review", "rating": "4: Ok but not good enough - rejection", "review": "TLDR: The authors present Variable Computation in Recurrent Neural Networks (VCRNN). VCRNN is similar in nature to Adaptive Computation Time (Graves et al., 2016). Imagine a vanilla RNN, at each timestep only a subset (i.e., \"variable computation\") of the state is updated. Experimental results are not convincing, there is limited comparison to other cited work and basic LSTM baseline.\n\n=== Gating Mechanism ===\nAt each timestep, VCRNN generates a m_t vector which can be seen as a gating mechanism. Based off this m_t vector, a D-first (D-first as in literally the first D RNN states) subset of the vanilla RNN state is gated to be updated or not. Extra hyperparams epsilon and \\bar{m} are needed -- authors did not give us a value or explain how this was selected or how sensitive and critical these hyperparms are.\n\nThis mechanism while novel, feels a bit clunky and awkward. It does not feel well principled that only the D-first states get updated, rather than a generalized solution where any subset of the state can be updated.\n\nA short section in the text comparing to the soft-gating mechanisms of GRUs/LSTMs/Multiplicative RNNs (Wu et al., 2016) would be nice as well.\n\n=== Variable Computation ===\nOne of the arguments made is that their VCRNN model can save computation versus vanilla RNNs. While this may be technically true, in practice this is probably not the case. The size of the RNNs they compare to do not saturate any modern GPU cores. In theory computation might be saved, but in practice there will probably be no difference in wallclock time. The authors also did not report any wallclock numbers, which makes this argument hard to sell.\n\n=== Evaluation ===\nThis reviewer wished there was more citations to other work for comparison and a stronger baseline (than just a vanilla RNN). First, LSTMs are very simple and quite standard nowadays -- there is a lack of comparison to any basic stacked LSTM architecture in all the experiments.\n\nThe PTB BPC numbers are quite discouraging as well (compared to state-of-the-art). The VCRNN does not beat the basic vanilla RNN baseline. The authors also only cite/compare to a basic RNN architecture, however there has been many contributions since a basic RNN architecture that performs vastly better. Please see Chung et al., 2016 Table 1. Chung et al., 2016 also experimented w/ PTB BPC and they cite and compare to a large number of other (important) contributions.\n\nOne cool experiment the authors did is graph the per-character computation of VCRNN (i.e., see Figure 2). It shows after a space/word boundary, we use more computation! Cool! However, this makes me wonder what a GRU/LSTM does as well? What is the magnitude of the of the change in the state vector after a space in GRU/LSTM -- I suspect them to do something similar.\n\n=== Minor ===\n* Please add Equations numbers to the paper, hard to refer to in a review and discussion!\n\nReferences\nChung et al., \"Hierarchical Multiscale Recurrent Neural Networks,\" in 2016.\nGraves et al., \"Adaptive Computation Time for Recurrent Neural Networks,\" in 2016.\nWu et al., \"On Multiplicative Integration with Recurrent Neural Networks,\" in 2016.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Variable Computation in Recurrent Neural Networks | ["Yacine Jernite", "Edouard Grave", "Armand Joulin", "Tomas Mikolov"] | Recurrent neural networks (RNNs) have been used extensively and with increasing success to model various types of sequential data. Much of this progress has been achieved through devising recurrent units and architectures with the flexibility to capture complex statistics in the data, such as long range dependency or localized attention phenomena. However, while many sequential data (such as video, speech or language) can have highly variable information flow, most recurrent models still consume input features at a constant rate and perform a constant number of computations per time step, which can be detrimental to both speed and model capacity. In this paper, we explore a modification to existing recurrent units which allows them to learn to vary the amount of computation they perform at each step, without prior knowledge of the sequence's time structure. We show experimentally that not only do our models require fewer operations, they also lead to better performance overall on evaluation tasks. | ["Natural language processing", "Deep learning"] | https://openreview.net/forum?id=S1LVSrcge | https://openreview.net/pdf?id=S1LVSrcge | https://openreview.net/forum?id=S1LVSrcge¬eId=rJ0I-lfEe |
SJpiju-Ve | S1LVSrcge | ICLR.cc/2017/conference/-/paper246/official/review | {"title": "Interesting exploratory work.", "rating": "7: Good paper, accept", "review": "This is high novelty work, and an enjoyable read.\n\nMy concerns about the paper more or less mirror my pre-review questions. I certainly agree that the learned variable computation mechanism is obviously doing something interesting. The empirical results really need to be grounded with respect to the state of the art, and LSTMs are still an elephant in the room. (Note that I do not consider beating LSTMs, GRUs, or any method in particular as a prerequisite for acceptance, but the comparison nevertheless should be made.)\n\nIn pre-review responses the authors brought up that LSTMs perform more computation per timestep than Elman networks, and while that is true, this is an axis along which they can be compared, this factor controlled for (at least in expectation, by varying the number of LSTM cells), etc. A brief discussion of the proposed gating mechanism in light of the currently popular ones would strengthen the presentation.\n\n---\n2017/1/20: In light of my concerns being addressed I'm modifying my review to a 7, with the understanding that the manuscript will be amended to include the new comparisons posted as a comment.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Variable Computation in Recurrent Neural Networks | ["Yacine Jernite", "Edouard Grave", "Armand Joulin", "Tomas Mikolov"] | Recurrent neural networks (RNNs) have been used extensively and with increasing success to model various types of sequential data. Much of this progress has been achieved through devising recurrent units and architectures with the flexibility to capture complex statistics in the data, such as long range dependency or localized attention phenomena. However, while many sequential data (such as video, speech or language) can have highly variable information flow, most recurrent models still consume input features at a constant rate and perform a constant number of computations per time step, which can be detrimental to both speed and model capacity. In this paper, we explore a modification to existing recurrent units which allows them to learn to vary the amount of computation they perform at each step, without prior knowledge of the sequence's time structure. We show experimentally that not only do our models require fewer operations, they also lead to better performance overall on evaluation tasks. | ["Natural language processing", "Deep learning"] | https://openreview.net/forum?id=S1LVSrcge | https://openreview.net/pdf?id=S1LVSrcge | https://openreview.net/forum?id=S1LVSrcge¬eId=SJpiju-Ve |
rJIjX1MEx | r1aGWUqgg | ICLR.cc/2017/conference/-/paper288/official/review | {"title": "Unsupervised Learning of State Representations for Multiple Tasks", "rating": "6: Marginally above acceptance threshold", "review": "The paper presents a method to learn a low-dimensional state representations from raw obervation for multi-task setting. In contrast to classic multi-task learning setting where a joint representation is usually learned by exploring the transferable information among different tasks, the method aims to identify individual task and solve them separately. To this end, the authors extend the learning with robotic priors approach by extending the loss function with additional term for task coherence, i.e., a task only changes representation between training episodes. The method has been evaluated on two tasks, multi-task slot-car racing and mobile navigation to prove its efficacy.\n\nthere were several unclear issues:\n\n1. The first question is that if the method is only appealing on the scenario like the slot-car racing, otherwise it should be benchmarked with mutli-task learning. While the author made the argument in the related work, the proposed method is orthogonal to multi-task learning they did admit both explore shared knowledge between tasks. What's the advantage and disadvantage for the proposed method for general mutiple task setting, in particular over the multi-task learning?\nThe reply of the authors was not fully satisfactory. The argument did not support the lack of comparison to multi-task joint-learning. It seems they don't plan to include any comparison neither. I think it's important for the fundamental motivation for the work, without such comparison, the method seems to be purely an alternative to multi-task joint-learning without any(or much) practical advantage.\n\n2.Following up to the previous question, please clarify the results on the mobile navigation scenario. It's not clear how the plot on the right indicates MT-LRP identifies all tasks as the author claimed and and seems very weak to support the method, in particular compared to the multi-slot car-racing driving experiment, there is too little results to make sound argument (almost no comparison to alternative methods, i.e. no baseline method, is that true for the problem).\nThe explanation of the authors did provide more details and more explicit information. \n\n3. The proposed gated neural network architecture seems to be a soft gated structure(correct me if I am wrong), a possible baseline would be a hard gated unit, how would this affect the conclusion. This is particularly interesting as the authors reflect on the constraint that the representation should stay consistent during the training.\nThe author simply stated again what they did for the modeling without counter the comparison to hard-gating, but it's probably less an issue compared to Question 1.\n\nIn summary, while there are remaining concerns about lacking comparisons, the is a weak tendency towards accepting the submission.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Unsupervised Learning of State Representations for Multiple Tasks | ["Antonin Raffin", "Sebastian H\u00f6fer", "Rico Jonschkowski", "Oliver Brock", "Freek Stulp"] | We present an approach for learning state representations in multi-task reinforcement learning. Our method learns multiple low-dimensional state representations from raw observations in an unsupervised fashion, without any knowledge of which task is executed, nor of the number of tasks involved.
The method is based on a gated neural network architecture, trained with an extension of the learning with robotic priors objective. In simulated experiments, we show that our method is able to learn better state representations for reinforcement learning, and we analyze why and when it manages to do so. | ["Reinforcement Learning", "Unsupervised Learning"] | https://openreview.net/forum?id=r1aGWUqgg | https://openreview.net/pdf?id=r1aGWUqgg | https://openreview.net/forum?id=r1aGWUqgg¬eId=rJIjX1MEx |
HJkwiyG4e | r1aGWUqgg | ICLR.cc/2017/conference/-/paper288/official/review | {"title": "Review", "rating": "5: Marginally below acceptance threshold", "review": "This paper builds upon the method of Jonschkowski & Brock to learn state representations for multiple tasks, rather than a single task. The research direction of learning representations for multiple tasks is an interesting one, and largely unexplored. The approach in the paper is to learn a different representation for each task, and a different policy for each task, where the task is detected automatically and built into the neural network.\n\nThe authors state that the proposed method is orthogonal to multi-task learning, though the end goal of learning to solve multiple tasks is the same. It would be interesting and helpful to see more discussion on this point in the paper, as discussed in the pre-review question phase. References to other multi-task learning works, e.g. policy distillation and actor-mimic (both ICLR \u201916), may be appropriate as well.\n\nThe method proposes to jointly learn a task classifier with a state representation learner, by using a differentiable gating mechanism to control the flow of information. The paper proposes a task coherence prior for this gating mechanism to ensure that the learned task classifier is temporally coherent. Introducing this structure is what enables the method to improve performance over the standard, non-multitask approach.\n\nThe evaluation involves two toy experimental scenarios. The first involves controlling one of two cars to drive around a track. In this task, detecting the \u201ctask\u201d is very easy, and the learned state representation is linear in the observation. The paper evaluates the performance of the policies learned with the proposed approach, and shows sufficient comparisons to demonstrate the usefulness of the approach over a standard non-multitask set-up.\n\nIn the second navigation scenario, only the state representation is qualitatively shown, not the resulting control policy nor any other learned state representations for comparison. Since the multi-task state representation learning approach is only useful if you can also learn control better, the paper should also evaluate on control, with the same comparisons as in the first experiment. Without this evaluation, the experiment is incomplete.\n\nLastly, to be on par with publications at a venue like ICLR, the method should be evaluated more thoroughly, on a wider range of set-ups, to demonstrate the generality of the approach and show that the method applies to more complex tasks. While in theory, the method should scale, the experiments do not demonstrate that it can handle more realistic scenarios, such as scaling beyond MNIST-level images, to 3D or real images, or higher-dimensional control tasks. Evaluating the method in this more complex scenario is important, because unexpected issues can come up when trying to scale. If scaling-up is straight-forward, then running this experiment (and including it in the paper) should be straight-forward.\n\nIn summary, here are the pros and cons of this paper:\nCons\n- The approach does not necessarily share information across tasks for better learning, and requires learning a different policy for each task\n- Only one experimental set-up that evaluates learned policy with multi-task state representation\n- No experiments on more realistic scenarios, such 3D environments or high-dimensional control problems\nPros: \n- This approach enables using the same network for multiple tasks, which is often not true for transfer and multi-task learning approaches\n- Novel way to learn a single policy for multiple tasks, including a task coherence prior which ensures that the task classification is meaningful\n- Experimentally validated on two toy tasks. One task shows improvement over baseline approaches\n\nThus, my rating would be higher if the paper included an evaluation of the control policy for navigation and included another more challenging and compelling scenario.\n\n\nLastly, here are some minor comments/questions on how I think the paper could be improved, but are not as important as the above:\n\nApproach:\nCould this approach be combined with other state representation learning approaches? e.g. approaches that use an autoencoder.\n\nExperiments:\nOne additional useful comparison would be to evaluate performance in the single-task setting (e.g. only controlling the red car), as an upper bound on how well the policy should be able to perform. Does the learned multi-task policy reach the same level of performance? This upper bound will be tighter than the \u201cknown car position\u201d baseline (which is also useful in its own right).\n\nDoes the \u201cobservations\u201d baseline eventually reach the performance of the LRP approach? It would be useful to know if this approach simply speeds up learning (significantly) or if it enables better performance.\n\nIf there are aliasing issues with the images, why not just use higher resolution images?", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Unsupervised Learning of State Representations for Multiple Tasks | ["Antonin Raffin", "Sebastian H\u00f6fer", "Rico Jonschkowski", "Oliver Brock", "Freek Stulp"] | We present an approach for learning state representations in multi-task reinforcement learning. Our method learns multiple low-dimensional state representations from raw observations in an unsupervised fashion, without any knowledge of which task is executed, nor of the number of tasks involved.
The method is based on a gated neural network architecture, trained with an extension of the learning with robotic priors objective. In simulated experiments, we show that our method is able to learn better state representations for reinforcement learning, and we analyze why and when it manages to do so. | ["Reinforcement Learning", "Unsupervised Learning"] | https://openreview.net/forum?id=r1aGWUqgg | https://openreview.net/pdf?id=r1aGWUqgg | https://openreview.net/forum?id=r1aGWUqgg¬eId=HJkwiyG4e |
SkFuoeMNl | r1aGWUqgg | ICLR.cc/2017/conference/-/paper288/official/review | {"title": "Unsupervised Learning of State Representations for Multiple Tasks", "rating": "6: Marginally above acceptance threshold", "review": "This paper is about learning unsupervised state representations using multi-task reinforcement learning. The authors propose a novel approach combining gated neural networks with multitask learning with robotics priors. They evaluated their approach on two simulated datasets and showed promising results. The paper is clearly written and is theoretically sound.\n\nPositives:\n+ Gating to enable learning a joint representation\n+ Multi-task learning extended from a single task in prior work\n+ Combining multiple types of losses to learn a strong representation (Coherence, Proportionality, Causality, Repeatability, Consistency and Separation)\n\nNegatives:\n- Parameters choice is arbitrary (w parameters)\n- Limiting the multi-task learning to be different to individual tasks rather than sharing and transferring knowledge between tasks\n- The experiments could have been conducted using a standardized simulation tool such as OpenAI Gym to make it easy to compare.\n\nI would recommend that the authors consider a more standardized way of picking the model parameters and evaluate on a more standard and high-dimensional datasets.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Unsupervised Learning of State Representations for Multiple Tasks | ["Antonin Raffin", "Sebastian H\u00f6fer", "Rico Jonschkowski", "Oliver Brock", "Freek Stulp"] | We present an approach for learning state representations in multi-task reinforcement learning. Our method learns multiple low-dimensional state representations from raw observations in an unsupervised fashion, without any knowledge of which task is executed, nor of the number of tasks involved.
The method is based on a gated neural network architecture, trained with an extension of the learning with robotic priors objective. In simulated experiments, we show that our method is able to learn better state representations for reinforcement learning, and we analyze why and when it manages to do so. | ["Reinforcement Learning", "Unsupervised Learning"] | https://openreview.net/forum?id=r1aGWUqgg | https://openreview.net/pdf?id=r1aGWUqgg | https://openreview.net/forum?id=r1aGWUqgg¬eId=SkFuoeMNl |
HJLhbHQ4x | SJBr9Mcxl | ICLR.cc/2017/conference/-/paper177/official/review | {"title": "Review of \"indexing neuron selectivity\"", "rating": "7: Good paper, accept", "review": "This paper attempts to understand and visualize what deep nets are representing as one ascends from low levels to high levels of the network. As has been shown previously, lower levels are more local image feature based, whereas higher levels correspond to abstract properties such as object identity. In semantic space, we find higher level nodes to be more semantically selective, whereas low level nodes are more diffuse.\n\nThis seems like a good attempt to tease apart deep net representations. Perhaps the most important finding is that color figures prominently into all levels of the network, and that performance on gray scale images is significantly diminished. The new NF measure proposed here is sensible, but still based on the images shown to the network. What one really wants to know is what function these nodes are computing - i.e., out of the space of *all* possible images, which most activate a unit? Of course this is a difficult problem, but it would be nice to see us getting closer to understanding the answer. The color analysis here I think brings us a bit closer. The semantic analysis is nice but I'm not sure what new insight we gain from this. \n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Understanding trained CNNs by indexing neuron selectivity | ["Ivet Rafegas", "Maria Vanrell", "Lu\u00eds A. Alexandre"] | The impressive performance and plasticity of convolutional neural networks to solve different vision problems are shadowed by their black-box nature and its consequent lack of full understanding. To reduce this gap we propose to describe the activity of individual neurons by quantifiyng their inherent selectivity to specific properties. Our approach is based on the definition of feature selectivity indexes that allow the ranking of neurons according to specific properties. Here we report the results of exploring selectivity indexes for: (a) an image feature (color); and (b) an image label (class membership). Our contribution is a framework to seek or classify neurons by indexing on these selectivity properties. It helps to find color selective neurons, such as a red-mushroom neuron in layer conv4 or class selective neurons such as dog-face neurons in layer conv5, and establishes a methodology to derive other selectivity properties. Indexing on neuron selectivity can statistically draw how features and classes are represented through layers at a moment when the size of trained nets is growing and automatic tools to index can be helpful. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=SJBr9Mcxl | https://openreview.net/pdf?id=SJBr9Mcxl | https://openreview.net/forum?id=SJBr9Mcxl¬eId=HJLhbHQ4x |
rkn9pmD4e | SJBr9Mcxl | ICLR.cc/2017/conference/-/paper177/official/review | {"title": "review of \"UNDERSTANDING TRAINED CNNS BY INDEXING NEURON SELECTIVITY\"", "rating": "3: Clear rejection", "review": "The authors analyze trained neural networks by quantifying the selectivity of individual neurons in the network for a variety of specific features, including color and category. \n\nPros:\n * The paper is clearly written and has good figures. \n * I think they executed their specific stated goal reasonably well technically. E.g. the various indexes they use seem well-chosen for their purposes. \n\nCons:\n * I must admit that I am biased against the whole enterprise of this paper. I do not think it is well-motivated or provides any useful insight whatever. What I view their having done is produced, and then summarized anecdotally, a catalog of piecemeal facts about a neural network without any larger reason to think these particular facts are important. In a way, I feel like this paper suffers from the same problem that plagues a typical line of research in neurophysiology, in which a catalog of selectivity distributions of various neurons for various properties is produced -- full stop. As if that were in and of itself important or useful information. I do not feel that either the original neural version of that project, or this current model-based virtual electrophysiology, is that useful. Why should we care about the distribution of color selectivities? Why does knowing distribution as such constitute \"understanding\"? To my mind it doesn't, at least not directly. \n\nHere's what they could have done to make a more useful investigation:\n \n (a) From a neuroscience point of view, they could have compared the properties that they measure in models to the same properties as measured in neurons the real brain. If they could show that some models are better matches on these properties to the actual neural data than others, that would be a really interesting result. That is is to say, the two isolated catalogs of selectivities (from model neurons and real neurons) alone seem pretty pointless. But if the correspondence between the two catalogs was made -- both in terms of where the model neurons and the real neurons were similar, and (especially importantly) where they were different --- that would be the beginning of nontrivial understanding. Such results would also complement a growing body of literature that attempts to link CNNs to visual brain areas. Finding good neural data is challenging, but whatever the result, the comparison would be interesting. \n\nand/or \n\n (b) From an artificial intelligence point of view, they could have shown that their metrics are *prescriptive* constraints. That is, suppose they had shown that the specific color and class selectivity indices that they compute, when imposed as a loss-function criterion on an untrained neural network, cause the network to develop useful filters and achieve significantly above-chance performance on the original task the networks were trained on. This would be a really great result, because it would not only give us a priori reason to care about the specific property metrics they chose, but it would also help contribute to efforts to find unsupervised (or semi-supervised) learning procedures, since the metrics they compute can be estimated from comparatively small numbers of stimuli and/or high-level semantic labels. To put this in perspective, imagine that they had actually tested the above hypothesis and found it to be false: that is, that their metrics, when used as loss function constraints, do not improve performance noticeably above chance performance. What would we then make of this whole investigation? It would then be reasonable to think that the measured properties were essentially epiphenomenal and didn't contribute at all to the power of neural networks in solving perceptual tasks. (The same could be said about neurophysiology experiments doing the same thing.) \n [--> NB: I've actually tried things just like this myself over the years, and have found exactly this disappointing result. Specifically, I've found a number of high-level generic statistical property of DNNs that seem like they might potentially \"interesting\", e.g. because they apparently correlate with complexity or appear to illustrate difference between low, intermediate and high layers of DNNs. Every single one of these, when imposed as optimization constraints, has basically lead nowhere on the challenging tasks (like ImageNet) that cause the DNNs to be interesting in the first place. Basically, there is to my mind no evidence at this point that highly-summarized generic statistical distributions of selectivities, like those illustrated here, place any interesting constraints on filter weights at all. Of course, I haven't tried the specific properties the authors highlight in these papers, so maybe there's something important there.]\n\nI know that both of these asks are pretty hard, but I just don't know what else to say -- this work otherwise seems like a step backwards for what the community ought to be spending its time on. \n ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Understanding trained CNNs by indexing neuron selectivity | ["Ivet Rafegas", "Maria Vanrell", "Lu\u00eds A. Alexandre"] | The impressive performance and plasticity of convolutional neural networks to solve different vision problems are shadowed by their black-box nature and its consequent lack of full understanding. To reduce this gap we propose to describe the activity of individual neurons by quantifiyng their inherent selectivity to specific properties. Our approach is based on the definition of feature selectivity indexes that allow the ranking of neurons according to specific properties. Here we report the results of exploring selectivity indexes for: (a) an image feature (color); and (b) an image label (class membership). Our contribution is a framework to seek or classify neurons by indexing on these selectivity properties. It helps to find color selective neurons, such as a red-mushroom neuron in layer conv4 or class selective neurons such as dog-face neurons in layer conv5, and establishes a methodology to derive other selectivity properties. Indexing on neuron selectivity can statistically draw how features and classes are represented through layers at a moment when the size of trained nets is growing and automatic tools to index can be helpful. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=SJBr9Mcxl | https://openreview.net/pdf?id=SJBr9Mcxl | https://openreview.net/forum?id=SJBr9Mcxl¬eId=rkn9pmD4e |
H1cc-_xSl | SJBr9Mcxl | ICLR.cc/2017/conference/-/paper177/official/review | {"title": "Solid work leading to a few interesting conclusions", "rating": "7: Good paper, accept", "review": "This paper makes three main methodological contributions:\n - definition of Neural Feature (NF) as the pixel average of the top N images that highly activation a neuron\n - ranking of neurons based on color selectivity\n - ranking of neurons based on class selectivity\n\nThe main weaknesses of the paper are that none of the methodological contributions are very significant, and no singularly significant result arises from the application of the methods.\n\nHowever, the main strengths of the paper are its assortment of moderately-sized interesting conclusions about the basic behavior of neural nets. For example, a few are:\n - \u201cIndexing on class selectivity neurons we found highly class selective neurons like digital-clock at conv2, cardoon at conv3 and ladybug at conv5, much before the fully connected layers.\u201d As far as I know, this had not been previously reported.\n - Color selective neurons are found even in higher layers. (25% color selectivity in conv5)\n - \u201cour main color axis emerge (black-white, blue-yellow, orange-cyan and cyan- magenta). Curiously, these two observations correlate with evidences in the human visual system (Shapley & Hawken (2011)).\u201d Great observation!\n\nOverall, I\u2019d recommend the paper be accepted, because although it\u2019s difficult to predict at this time, there\u2019s a fair chance that one of the \u201csmaller conclusions\u201d would turn out to be important in hindsight a few years hence.\n\n\nOther small comments:\n - The cite for \u201cLearning to generate chairs\u2026\u201d is wrong (first two authors combined resulting in a confusing cite)\n\n - What exactly is the Color Selectivity Index computing? The Opponent Color Space isn\u2019t well defined and it wasn\u2019t previously familiar to me. Intuitively it seems to be selecting for units that respond to a constant color, but the highest color selectivity NF in Fig 5 i for a unit with two colors, not one. Finally, the very last unit (lowest color selectivity) is almost the same edge pattern, but with white -> black instead of blue -> orange. Why are these considered to be so drastically different? This should probably be more clearly described.\n\n - For the sake of argument, imagine a mushroom sensitive neuron in conv5 that fires highly for mushrooms of *any* color but not for anything else. If the dataset contains only red-capped mushrooms, would the color selectivity index for this neuron be high or low? If it is high, it\u2019s somewhat misleading because the unit itself actually isn\u2019t color selective; the dataset just happens only to have red mushrooms in it. (It\u2019s a subtle point but worth considering and probably discussing in the paper)\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Understanding trained CNNs by indexing neuron selectivity | ["Ivet Rafegas", "Maria Vanrell", "Lu\u00eds A. Alexandre"] | The impressive performance and plasticity of convolutional neural networks to solve different vision problems are shadowed by their black-box nature and its consequent lack of full understanding. To reduce this gap we propose to describe the activity of individual neurons by quantifiyng their inherent selectivity to specific properties. Our approach is based on the definition of feature selectivity indexes that allow the ranking of neurons according to specific properties. Here we report the results of exploring selectivity indexes for: (a) an image feature (color); and (b) an image label (class membership). Our contribution is a framework to seek or classify neurons by indexing on these selectivity properties. It helps to find color selective neurons, such as a red-mushroom neuron in layer conv4 or class selective neurons such as dog-face neurons in layer conv5, and establishes a methodology to derive other selectivity properties. Indexing on neuron selectivity can statistically draw how features and classes are represented through layers at a moment when the size of trained nets is growing and automatic tools to index can be helpful. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=SJBr9Mcxl | https://openreview.net/pdf?id=SJBr9Mcxl | https://openreview.net/forum?id=SJBr9Mcxl¬eId=H1cc-_xSl |
SJSXiJbEe | ByG4hz5le | ICLR.cc/2017/conference/-/paper182/official/review | {"title": "Clear and efficient space+time+feature attention mechanisms for video captioning", "rating": "7: Good paper, accept", "review": "1) Summary\n\nThis paper proposes a video captioning model based on a 3D (space+time) convnet (C3D) encoder and a LSTM decoder. The authors investigate the benefits of using attention mechanisms operating both at the spatio-temporal and layer (feature abstraction) levels.\n\n2) Contributions\n\n+ Well motivated and implemented attention mechanism to handle the different shapes of C3D feature maps (along space, time, and feature dimensions).\n+ Convincing quantitative and qualitative experiments on three challenging datasets (Youtube2Text, M-VAD, MSR-VTT) showing clearly the benefit of the proposed attention mechanisms.\n+ Interesting comparison of soft vs hard attention showing a slight performance advantage for the (simpler) soft attention mechanism in this case.\n\n3) Suggestions for improvement\n\nHypercolumns comparison:\nAs mentioned during pre-review questions, it would be interesting to compare to the hypercolumns of https://arxiv.org/abs/1411.5752, as they are an alternative to the proposed attention mechanisms, with the same purpose of leveraging different feature abstraction levels.\n\nMinor clarifications in the text and figures as agreed with the authors in our pre-review discussions.\n\n4) Conclusion\n\nAlthough the novelty with existing video captioning approaches is limited, the paper is relevant to ICLR, as the proposed simple but efficient implementation and benefits of spatio-temporal + feature abstraction attention are clearly validated in this work.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Adaptive Feature Abstraction for Translating Video to Language | ["Yunchen Pu", "Martin Renqiang Min", "Zhe Gan", "Lawrence Carin"] | Previous models for video captioning often use the output from a specific layer of a Convolutional Neural Network (CNN) as video representations, preventing them from modeling rich, varying context-dependent semantics in video descriptions. In this paper, we propose a new approach to generating adaptive spatiotemporal representations of videos for a captioning task. For this purpose, novel attention mechanisms with spatiotemporal alignment is employed to adaptively and sequentially focus on different layers of CNN features (levels of feature ``abstraction''), as well as local spatiotemporal regions of the feature maps at each layer. Our approach is evaluated on three benchmark datasets: YouTube2Text, M-VAD and MSR-VTT. Along with visualizing the results and how the model works, these experiments quantitatively demonstrate the effectiveness of the proposed adaptive spatiotemporal feature abstraction for translating videos to sentences with rich semantics. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=ByG4hz5le | https://openreview.net/pdf?id=ByG4hz5le | https://openreview.net/forum?id=ByG4hz5le¬eId=SJSXiJbEe |
rJgoCe-4e | ByG4hz5le | ICLR.cc/2017/conference/-/paper182/official/review | {"title": "solid work but lack of novelty", "rating": "4: Ok but not good enough - rejection", "review": "The authors apply the image captioning architecture of Xu et al. 2015 to video captioning. The model is extended to have attention over multiple layers of the ConvNet instead of just a single layer. Experiments on YouTube2Text, M-VAD and MSR-VTT show that this works better than only using one of the layers at a time.\n\nI think this is solid work on the level of a well-executed course project or a workshop paper. The model makes sense, it is adequately described, and the experiments show that attending over multiple layers works better than attending over any one layer in isolation. Unfortunately, I don't think there is enough to get excited about here from a technical perspective and it's not clear what value the paper brings to the community. Other aspects of the paper, such as including the hard attention component, don't seem to add to the paper but take up space. \n\nIf the authors want to contribute a detailed, focused exploration of multi-level features this could become a more valuable paper, but in that case I would expect a much more thorough exploration of the choices and tradeoffs of different schemes without too many spurious aspects such as video features, hard attention, etc.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Adaptive Feature Abstraction for Translating Video to Language | ["Yunchen Pu", "Martin Renqiang Min", "Zhe Gan", "Lawrence Carin"] | Previous models for video captioning often use the output from a specific layer of a Convolutional Neural Network (CNN) as video representations, preventing them from modeling rich, varying context-dependent semantics in video descriptions. In this paper, we propose a new approach to generating adaptive spatiotemporal representations of videos for a captioning task. For this purpose, novel attention mechanisms with spatiotemporal alignment is employed to adaptively and sequentially focus on different layers of CNN features (levels of feature ``abstraction''), as well as local spatiotemporal regions of the feature maps at each layer. Our approach is evaluated on three benchmark datasets: YouTube2Text, M-VAD and MSR-VTT. Along with visualizing the results and how the model works, these experiments quantitatively demonstrate the effectiveness of the proposed adaptive spatiotemporal feature abstraction for translating videos to sentences with rich semantics. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=ByG4hz5le | https://openreview.net/pdf?id=ByG4hz5le | https://openreview.net/forum?id=ByG4hz5le¬eId=rJgoCe-4e |
r1facGf4x | ByG4hz5le | ICLR.cc/2017/conference/-/paper182/official/review | {"title": "state-of-the-art results but too incremental", "rating": "4: Ok but not good enough - rejection", "review": "This paper presents a model for video captioning with both soft and hard attention, using a C3D network for the encoder and a RNN for the decoder. Experiments are presented on YouTube2Text, M-VAD, and MSR-VTT. While the ideas of image captioning with soft and hard attention, and video captioning with soft attention, have already been demonstrated in previous work, the main contribution here is the specific architecture and attention over different layers of the CNN.\n\nThe work is well presented and the experiments clearly show the benefit of attention over multiple layers. However, in light of previous work in captioning, the contribution and resulting insights is too incremental for a conference paper at ICLR. Further experiments and analysis of the main contribution would strengthen the paper, but I would recommend resubmission to a more suitable venue.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Adaptive Feature Abstraction for Translating Video to Language | ["Yunchen Pu", "Martin Renqiang Min", "Zhe Gan", "Lawrence Carin"] | Previous models for video captioning often use the output from a specific layer of a Convolutional Neural Network (CNN) as video representations, preventing them from modeling rich, varying context-dependent semantics in video descriptions. In this paper, we propose a new approach to generating adaptive spatiotemporal representations of videos for a captioning task. For this purpose, novel attention mechanisms with spatiotemporal alignment is employed to adaptively and sequentially focus on different layers of CNN features (levels of feature ``abstraction''), as well as local spatiotemporal regions of the feature maps at each layer. Our approach is evaluated on three benchmark datasets: YouTube2Text, M-VAD and MSR-VTT. Along with visualizing the results and how the model works, these experiments quantitatively demonstrate the effectiveness of the proposed adaptive spatiotemporal feature abstraction for translating videos to sentences with rich semantics. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=ByG4hz5le | https://openreview.net/pdf?id=ByG4hz5le | https://openreview.net/forum?id=ByG4hz5le¬eId=r1facGf4x |
rJu4Ftb4x | BJjn-Yixl | ICLR.cc/2017/conference/-/paper562/official/review | {"title": "Strong experimental results, but somewhat unclear where the improvements are coming from", "rating": "5: Marginally below acceptance threshold", "review": "This paper presents an attention based recurrent approach to one-shot learning. It reports quite strong experimental results (surpassing human performance/HBPL) on the Omniglot dataset, which is somewhat surprising because it seems to make use of very standard neural network machinery. The authors also note that other have helped verify the results (did Soumith Chintala reproduce the results?) and do provide source code.\n\nAfter reading this paper, I'm left a little perplexed as to where the big performance improvements are coming from as it seems to share a lot of the same components of previous work. If the author's could report result from a broader suite of experiments like in previous work (e.g matching networks), it would much more convincing. An ablation study would also help with understanding why this model does so well.", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"} | review | 2017 | ICLR.cc/2017/conference | Attentive Recurrent Comparators | ["Pranav Shyam", "Ambedkar Dukkipati"] | Attentive Recurrent Comparators (ARCs) are a novel class of neural networks built with attention and recurrence that learn to estimate the similarity of a set of objects by cycling through them and making observations. The observations made in one object are conditioned on the observations made in all the other objects. This allows ARCs to learn to focus on the salient aspects needed to ascertain similarity. Our simplistic model that does not use any convolutions performs comparably to Deep Convolutional Siamese Networks on various visual tasks. However using ARCs and convolutional feature extractors in conjunction produces a model that is significantly better than any other method and has superior generalization capabilities. On the Omniglot dataset, ARC based models achieve an error rate of 1.5\% in the One-Shot classification task - a 2-3x reduction compared to the previous best models. This is also the first Deep Learning model to outperform humans (4.5\%) and surpass the state of the art accuracy set by the highly specialized Hierarchical Bayesian Program Learning (HBPL) system (3.3\%). | ["Deep learning", "Computer vision"] | https://openreview.net/forum?id=BJjn-Yixl | https://openreview.net/pdf?id=BJjn-Yixl | https://openreview.net/forum?id=BJjn-Yixl¬eId=rJu4Ftb4x |
HJfCu9Amg | BJjn-Yixl | ICLR.cc/2017/conference/-/paper562/official/review | {"title": "The paper need more improvements to be accepted", "rating": "3: Clear rejection", "review": "This paper describes a method that estimates the similarity between a set of images by alternatively attend each image with a recurrent manner. The idea of the paper is interesting, which mimic the human's behavior. However, there are several cons of the paper:\n\n1. The paper is now well written. There are too many 'TODO', 'CITE' in the final version of the paper, which indicates that the paper is submitted in a rush or the authors did not take much care about the paper. I think the paper is not suitable to be published with the current version.\n\n2. The missing of the experimental results. The paper mentioned the LFW dataset. However, the paper did not provide the results on LFW dataset. (At least I did not find it in the version of Dec. 13th)\n\n3. The experiments of Omniglot dataset are not sufficient. I suggest that the paper provides some illustrations about how the model the attend two images (e.g. the trajectory of attend).", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Attentive Recurrent Comparators | ["Pranav Shyam", "Ambedkar Dukkipati"] | Attentive Recurrent Comparators (ARCs) are a novel class of neural networks built with attention and recurrence that learn to estimate the similarity of a set of objects by cycling through them and making observations. The observations made in one object are conditioned on the observations made in all the other objects. This allows ARCs to learn to focus on the salient aspects needed to ascertain similarity. Our simplistic model that does not use any convolutions performs comparably to Deep Convolutional Siamese Networks on various visual tasks. However using ARCs and convolutional feature extractors in conjunction produces a model that is significantly better than any other method and has superior generalization capabilities. On the Omniglot dataset, ARC based models achieve an error rate of 1.5\% in the One-Shot classification task - a 2-3x reduction compared to the previous best models. This is also the first Deep Learning model to outperform humans (4.5\%) and surpass the state of the art accuracy set by the highly specialized Hierarchical Bayesian Program Learning (HBPL) system (3.3\%). | ["Deep learning", "Computer vision"] | https://openreview.net/forum?id=BJjn-Yixl | https://openreview.net/pdf?id=BJjn-Yixl | https://openreview.net/forum?id=BJjn-Yixl¬eId=HJfCu9Amg |
HJWrxQM4x | BJjn-Yixl | ICLR.cc/2017/conference/-/paper562/official/review | {"title": "experimental section improved but still very weak on analysis and insight", "rating": "4: Ok but not good enough - rejection", "review": "This paper introduces an attention-based recurrent network that learns to compare images by attending iteratively back and forth between a pair of images. Experiments show state-of-the-art results on Omniglot, though a large part of the performance gain comes from when extracted convolutional features are used as input.\n\nThe paper is significantly improved from the original submission and reflects changes based on pre-review questions. However, while there was an attempt made to include more qualitative results e.g. Fig. 2, it is still relatively weak and could benefit from more examples and analysis. Also, why is the attention in Fig. 2 always attending over the full character? Although it is zooming in, shouldn\u2019t it attend to relevant parts of the character? Attending to the full character on a solid background seems a trivial solution where it is then unclear where the large performance gains are coming from.\n\nWhile the paper is much more polished now, it is still lacking in details in some respects, e.g. details of the convolutional feature extractor used that gives large performance gain.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Attentive Recurrent Comparators | ["Pranav Shyam", "Ambedkar Dukkipati"] | Attentive Recurrent Comparators (ARCs) are a novel class of neural networks built with attention and recurrence that learn to estimate the similarity of a set of objects by cycling through them and making observations. The observations made in one object are conditioned on the observations made in all the other objects. This allows ARCs to learn to focus on the salient aspects needed to ascertain similarity. Our simplistic model that does not use any convolutions performs comparably to Deep Convolutional Siamese Networks on various visual tasks. However using ARCs and convolutional feature extractors in conjunction produces a model that is significantly better than any other method and has superior generalization capabilities. On the Omniglot dataset, ARC based models achieve an error rate of 1.5\% in the One-Shot classification task - a 2-3x reduction compared to the previous best models. This is also the first Deep Learning model to outperform humans (4.5\%) and surpass the state of the art accuracy set by the highly specialized Hierarchical Bayesian Program Learning (HBPL) system (3.3\%). | ["Deep learning", "Computer vision"] | https://openreview.net/forum?id=BJjn-Yixl | https://openreview.net/pdf?id=BJjn-Yixl | https://openreview.net/forum?id=BJjn-Yixl¬eId=HJWrxQM4x |
SJZ6q-zEx | HyQWFOVge | ICLR.cc/2017/conference/-/paper23/official/review | {"title": "review", "rating": "4: Ok but not good enough - rejection", "review": "There has been substantial recent interest in representation learning, and specifically, using distance metric learning (DML) to learn representations where semantic distance between inputs can be measured. This is a topic of particular relevance / interest to ICLR. This paper poses a simple yet provocative question: can a standard SoftMax based approach learn features that match or even outperform recent state-of-the-art DML approaches? Thorough experiments seem to indicate that this is indeed the case. Comparisons are made to recent DML approaches including Magnet Loss (ICLR2016) and Lifted structure embedding (CVPR2016) and superior results are shown across a number of datasets / tasks for which the DML approaches were designed. \n\nThis main result is a bit surprising since SoftMax is a natural and trivial baseline, so it should have been properly evaluated in previous DML literature. The authors argue that previous approaches did not fully/properly tune the softmax baselines, or that comparisons were not apples-to-apples. Also, one change in the current paper is the addition of L2 normalization, which is well motivated and helps improve SoftMax feature distances. Different dimensionality reduction approaches are also tested. These changes are minor, but especially the L2 normalization proves to be a simple but effective improvement for SoftMax features.\n\nA big issue is with how pre-training is performed (in Magnet Loss the softmax baselines were pretrained for less time on ImageNet). The approach taken here is reasonable, but so is the approach in Magnet Loss (for different reasons). Ultimately, both are fine. Unfortunately, due to use of different schemes, the results are not comparable. Let me copy-paste what I wrote in an earlier comment: \n\nMy main concern about the paper is that the comparisons in Tables 1 and 2 and Figure 4 to Rippel et al. are not apples-to-apples. Basically, the papers shows that absolute results of using SoftMax w full pre-training (PT) on ImageNet is superior to any of the results in Rippel's paper (including both the Softmax and Magnet results). But as the current results show, PT appears to be critical to obtaining such good numbers - The Rippel SoftMax numbers use only 3 PT stages and are dramatically worse than the full PT on ImageNet. As it stands, I am not convinced that SoftMax is actually better than Magnet. Here is the evidence we have (I'll use Stanford Dogs as an example, but any of the datasets have the same conclusion): (1) Softmax w 3 stages of PT: 26.6% (from Rippel paper) and 32.7% (from authors' reproduction) (2) Magnet w 3 stages of PT: 24.9% (3) Softmax w full PT: 18.3% (4) Magnet w full PT: not shown From this all I see is that PT is critical for getting absolute good results. However, what about Magnet w full PT? These results are not shown either here or in the original Rippel paper (I went back and looked). As such, I do not think it is justifiable to claim superiority of Softmax to Magnet based on available evidence. (Note: I looked back carefully at Rippel's paper, and it appears that the authors use 3 PT stages as a form of \"warmup\". There is a statement that using full PT would \"defeat the purpose of pursuing DML\". I'm not sure if I agree w Rippel's statement since in the present paper there is clear evidence that full PT is hugely helpful, at least for softmax. That being said, I did not see any evidence in the Rippel paper that PT is harmful or that DML wouldn't work with full PT.)\n\nThe authors responded to my concern by claiming that \u201cfrom Rippel's results, it is no doubt that Magnet@3epochPT > Magnet@FullPT and Magnet@3epoch > Softmax@3epochPT.\u201d However, I went back to Rippel\u2019s paper, and simply the Magnet@FullPT experiment never appears. I further went and contacted Oren Rippel himself, and he verified he never ran the Magnet@FullPT experiment. I encourage the authors to contact Oren Rippel regarding this if they wish to verify (I have asked Oren Rippel to not reveal my identity). [Disclaimer: I am NOT Oren Rippel]. The authors mentioned that they are training Magnet@FullPT. If results were shown for Magnet@FullPT and also retrain the Magnet@3epochPT as a sanity check, that would help alleviate this concern. Alternatively, the language in Section 4 and the Tables could be altered to make clear that the methods use different pretraining and hence are not comparable.\n\nOverall, I am actually quite sympathetic to this work. I think it could serve as an important sanity-check paper for the community and quite relevant to ICLR. Having proper and strong SoftMax baselines should prove quite useful to the DML community and to this line of work.\n\nHowever, currently I find the main results (table 1, table 2, figure 4, etc.) to be misleading. If indeed it were the case that Magnet@3epochPT > Magnet@FullPT, then it would be fine. However, at this point as far as I know no one has actually tried Magnet@FullPT. And, given the general importance and effectiveness of pre-training, especially when transferring to small dataset, I would be hugely surprised if Magnet@FullPT was not superior by a large margin. I think either having this experiment in place or altering the writing / presentation of the results would be critical to allow for publishing.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Significance of Softmax-Based Features over Metric Learning-Based Features | ["Shota Horiguchi", "Daiki Ikami", "Kiyoharu Aizawa"] | The extraction of useful deep features is important for many computer vision tasks.
Deep features extracted from classification networks have proved to perform well in those tasks.
To obtain features of greater usefulness, end-to-end distance metric learning (DML) has been applied to train the feature extractor directly.
End-to-end DML approaches such as Magnet Loss and lifted structured feature embedding show state-of-the-art performance in several image recognition tasks.
However, in these DML studies, there were no equitable comparisons between features extracted from a DML-based network and those from a softmax-based network.
In this paper, by presenting objective comparisons between these two approaches under the same network architecture, we show that the softmax-based features are markedly better than the state-of-the-art DML features for tasks such as fine-grained recognition, attribute estimation, clustering, and retrieval. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=HyQWFOVge | https://openreview.net/pdf?id=HyQWFOVge | https://openreview.net/forum?id=HyQWFOVge¬eId=SJZ6q-zEx |
HkW_uHzVl | HyQWFOVge | ICLR.cc/2017/conference/-/paper23/official/review | {"title": "Interesting observations, but the paper seems misguided or incomplete", "rating": "7: Good paper, accept", "review": "I have a huge, big picture concern about this paper and the papers it most closely addresses (MagnetLoss and Lifted Feature Structure Embedding). I don't understand why Distance Metric Learning (DML) is being used for classification tasks (Stanford Cars 196, UCSD Birds 200, Oxford 102 flowers, Stanford Dogs, ImageNet attributes, etc). As far as I can tell, there is really only a single \"retrieval\"-like benchmark being used here - the Stanford Online Products database. All the other datasets are used in a \"classification-by-retrieval\" approach which seems contrived. While ostensibly evaluating \"retrieval\", the retrieval ground truth is totally defined by category membership so these are still classification tasks with many instances in each category. With the Online Products dataset the correspondence between queries and correct results is much more fine grained so it makes sense to think of it as a retrieval task.\n\nIt seems obvious that if your task is classification, a network trained with a classification loss will be best. Even when these datasets are used in a \"retrieval\" setting, the ground truth is still defined by category membership. It's still a classification task. \n\nI don't really see the point of using DML in these scenarios. I guess prior work claims to outperform SoftMax in these settings so this paper is fighting back against this and I should be thankful for this paper. But I think this paper's narrative is a bit off. The narrative shouldn't be \"We can get good retrieval features from softmax networks with appropriate normalization\". It should be \"It never made sense to train or evaluate these things as retrieval tasks. Direct classification is better\". For example, why are you taking the second to last layer or pool5 layer from these networks? Why aren't you taking the last layer? That should do well in these evaluations, right? Table 1 and 2 do show that using softmax probabilities directly tends to be better than doing classification-by-retrieval (works better or the same as doing retrieval with an earlier layer of features, except on Oxford flowers).\n\nGoogLeNet is quite deep and gets auxiliary supervision. By the second-to-last layer of the network, the activations could look a lot like category membership already. And category membership is all that's needed for the tasks in 4.2 and 4.3. \n \nI don't think my pre-review question was adequately addressed. I was getting at this concern by pointing out numerous scenarios where distance metric learning makes sense because you have fine-grained associations between instances at training time, NOT categorical associations -- e.g. this product photo corresponds to this photo of the object in a scene [Bell et al. 2015], this 3d model correspond to this sketch [Wang et al. 2015], this sketch corresponds to this photo [Sangkloy et al. 2016], this ground view corresponds to this aerial view [Lin et al., 2015]. DeepFace and follow-up works on LFW could also fit into this space because there are few training samples per class (few training samples per person identity). You cite DeepFace and Bell et al. 2015 but you don't compare on those benchmarks. I think those are exactly the tasks where DML makes sense.\n\nMaybe the \"retrieval on classification datasets\" would be a reasonable benchmark if the test and train classes were completely different. Then you could argue that softmax is learning a useful representation yet the last layer isn't directly useful since the categories change. But that's not the case here, is it?\n\nWith all of this said, I'm not sure whether I'm positive or negative about this paper. I think you're onto something significant -- people have been using DML where it is not appropriate -- but addressed it in the wrong way -- by using softmax for \"classification by retrieval\". But you don't need to do retrieval! Softmax is already telling you the class prediction! Why go through the extra step of finding nearest neighbors with some intermediate feature?\n\nAnonReviewer3 also raises some good points and you should be thankful that a reviewer is willing to dig so deep to help make your experiments sound! I don't think his/her concerns are disqualifying for this paper, though, as long as it is fixed.\n\nI look forward to hearing your response. I want this paper to be published, but I think it needs to be tweaked.\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Significance of Softmax-Based Features over Metric Learning-Based Features | ["Shota Horiguchi", "Daiki Ikami", "Kiyoharu Aizawa"] | The extraction of useful deep features is important for many computer vision tasks.
Deep features extracted from classification networks have proved to perform well in those tasks.
To obtain features of greater usefulness, end-to-end distance metric learning (DML) has been applied to train the feature extractor directly.
End-to-end DML approaches such as Magnet Loss and lifted structured feature embedding show state-of-the-art performance in several image recognition tasks.
However, in these DML studies, there were no equitable comparisons between features extracted from a DML-based network and those from a softmax-based network.
In this paper, by presenting objective comparisons between these two approaches under the same network architecture, we show that the softmax-based features are markedly better than the state-of-the-art DML features for tasks such as fine-grained recognition, attribute estimation, clustering, and retrieval. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=HyQWFOVge | https://openreview.net/pdf?id=HyQWFOVge | https://openreview.net/forum?id=HyQWFOVge¬eId=HkW_uHzVl |
B1Lus18Ex | HyQWFOVge | ICLR.cc/2017/conference/-/paper23/official/review | {"title": "Interesting discussion but sharing concern with reviewer 3 ", "rating": "5: Marginally below acceptance threshold", "review": "I agree with the other two reviewers that it is an interesting topic to investigate the feature learned by DML. For classification task though, I feel intuitively softmax should have advantages over distance metric learning method because the loss function is designed to assign the correct class for the given image. All the experimental results show that the softmax features work better than Rippel et al DML method. However, does it support the claim that softmax-based features work much better than DML learned features? I have doubts on this claim. \n\nAlso the experiments are a little bit misleading. What is vanilla googleNet softmax finetuned results? It seems it is not Rippel et al. (softmax prob) result. I am wondering whether the improvement comes from a) using retrieval (nearest neighbor) for classification or b) adding a new layer on top of pool5 or c) L2 normalization of the features. It is not clear to me at all. It appears to me the comparison is not apple vs apple between the proposed method and Rippel et al. \n\nIt would be great if we know adding feature reduction or adding another layer on top of pool5 can improve finetued softmax result. However, I am not sure what is the biggest contributing factor to the superior results. Before getting more clarifications from the authors, I lean toward rejection. \n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Significance of Softmax-Based Features over Metric Learning-Based Features | ["Shota Horiguchi", "Daiki Ikami", "Kiyoharu Aizawa"] | The extraction of useful deep features is important for many computer vision tasks.
Deep features extracted from classification networks have proved to perform well in those tasks.
To obtain features of greater usefulness, end-to-end distance metric learning (DML) has been applied to train the feature extractor directly.
End-to-end DML approaches such as Magnet Loss and lifted structured feature embedding show state-of-the-art performance in several image recognition tasks.
However, in these DML studies, there were no equitable comparisons between features extracted from a DML-based network and those from a softmax-based network.
In this paper, by presenting objective comparisons between these two approaches under the same network architecture, we show that the softmax-based features are markedly better than the state-of-the-art DML features for tasks such as fine-grained recognition, attribute estimation, clustering, and retrieval. | ["Computer vision", "Deep learning"] | https://openreview.net/forum?id=HyQWFOVge | https://openreview.net/pdf?id=HyQWFOVge | https://openreview.net/forum?id=HyQWFOVge¬eId=B1Lus18Ex |
BkxeboW4g | SJRpRfKxx | ICLR.cc/2017/conference/-/paper83/official/review | {"title": "Interesting paper, some evaluation issues", "rating": "7: Good paper, accept", "review": "The authors formulate a recurrent deep neural network to predict human fixation locations in videos as a mixture of Gaussians. They train the model using maximum likelihood with actual fixation data. Apart from evaluating how good the model performs at predicting fixations, they combine the saliency predictions with the C3D features for action recognition.\n\nquality: I am missing a more thorough evaluation of the fixation prediction performance. The center bias performance in Table 1 differs significantly from the on in Table 2. All the state-of-the-art models reported in Table 2 have a performance worse than the center bias performance reported in Table 1. Is there really no other model better than the center bias? Additionally I am missing details on how central bias and human performance are modelled. Is human performance cross-validated?\n\nYou claim that your \"results are very close to human performance (the difference is only 3.2%). This difference is actually larger than the difference between Central Bias and your model reported in Table 1. Apart from this, it is dangerous to compare AUC performance differences due to e.g. saturation issues.\n\nclarity: the explanation for Table 3 is a bit confusing, also it is not clear why the CONV5 and the FC6 models differ in how the saliency map is used. At least one should also evaluate the CONV5 model when multiplying the input with the saliency map to see how much of the difference comes from the different ways to use the saliency map and how much from the different features.\n\nOther issues:\n\nYou cite K\u00fcmmerer et. al 2015 as a model which \"learns ... indirectly rather than from explicit information of where humans look\", however the their model has been trained on fixation data using maximum-likelihood.\n\nApart from these issues, I think the paper make a very interesting contribution to spatio-temporal fixation prediction. If the evaluation issues given above are sorted out, I will happily improve my rating.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Recurrent Mixture Density Network for Spatiotemporal Visual Attention | ["Loris Bazzani", "Hugo Larochelle", "Lorenzo Torresani"] | In many computer vision tasks, the relevant information to solve the problem at hand is mixed to irrelevant, distracting information. This has motivated researchers to design attentional models that can dynamically focus on parts of images or videos that are salient, e.g., by down-weighting irrelevant pixels. In this work, we propose a spatiotemporal attentional model that learns where to look in a video directly from human fixation data. We model visual attention with a mixture of Gaussians at each frame. This distribution is used to express the probability of saliency for each pixel. Time consistency in videos is modeled hierarchically by: 1) deep 3D convolutional features to represent spatial and short-term time relations and 2) a long short-term memory network on top that aggregates the clip-level representation of sequential clips and therefore expands the temporal domain from few frames to seconds. The parameters of the proposed model are optimized via maximum likelihood estimation using human fixations as training data, without knowledge of the action in each video. Our experiments on Hollywood2 show state-of-the-art performance on saliency prediction for video. We also show that our attentional model trained on Hollywood2 generalizes well to UCF101 and it can be leveraged to improve action classification accuracy on both datasets. | ["Computer vision", "Deep learning", "Applications"] | https://openreview.net/forum?id=SJRpRfKxx | https://openreview.net/pdf?id=SJRpRfKxx | https://openreview.net/forum?id=SJRpRfKxx¬eId=BkxeboW4g |
SyMt0S-4g | SJRpRfKxx | ICLR.cc/2017/conference/-/paper83/official/review | {"title": "Review", "rating": "6: Marginally above acceptance threshold", "review": "This work proposes to a spatiotemporal saliency network that is able to mimic human fixation patterns,\nthus helping to prune irrelevant information from the video and improve action recognition.\n\nThe work is interesting and has shown state-of-the-art results on predicting human attention on action videos.\nIt has also shown promise for helping action clip classification.\n\nThe paper would benefit from a discussion on the role of context in attention.\nFor instance, if context is important, and people give attention to context, why is it not incorporated automatically in your model?\n\nOne weak point is the action recognition section, where the comparison between the two (1)(2) and (3) seems unfair.\nThe attention weighted feature maps in fact reduce the classification performance, and only improve performance when doubling the feature and associated model complexity by concatenating the weighted maps with the original features.\n\nIs there a way to combine the context and attention without concatenation?\nThe rational for concatenating the features extracted from the original clip,\nand the features extracted from the saliency weighted clip seems to contradict the initial hypothesis that `eliminating or down-weighting pixels that are not important' will improve performance.\n\nThe authors should also mention the current state-of-the-art results in Table 4, for comparison.\n\n# Other comments:\n# Abstract\n- Typo: `mixed with irrelevant ...'\n``Time consistency in videos ... expands the temporal domain from few frames to seconds'' - These two points are not clear, probably need a re-write.\n\n# Contributions\n- 1) `The model can be trained without having to engineer spatiotemporal features' - you would need to collect training data from humans though.. \n\n# Section 3.1\nThe number of fixation points is controlled to be fixed for each frame - how is this done?\n\nIn practice we freeze the layers of the C3D network to values pretrained by Tran etal.\nWhat happens when you allow gradients to flow back to the C3D layers?\nIs it not better to allow the features to be best tuned for the final task?\n\nThe precise way in which the features are concatenated needs to be clarified in section 3.4.\n\nMinor typo:\n`we added them trained central bias'", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Recurrent Mixture Density Network for Spatiotemporal Visual Attention | ["Loris Bazzani", "Hugo Larochelle", "Lorenzo Torresani"] | In many computer vision tasks, the relevant information to solve the problem at hand is mixed to irrelevant, distracting information. This has motivated researchers to design attentional models that can dynamically focus on parts of images or videos that are salient, e.g., by down-weighting irrelevant pixels. In this work, we propose a spatiotemporal attentional model that learns where to look in a video directly from human fixation data. We model visual attention with a mixture of Gaussians at each frame. This distribution is used to express the probability of saliency for each pixel. Time consistency in videos is modeled hierarchically by: 1) deep 3D convolutional features to represent spatial and short-term time relations and 2) a long short-term memory network on top that aggregates the clip-level representation of sequential clips and therefore expands the temporal domain from few frames to seconds. The parameters of the proposed model are optimized via maximum likelihood estimation using human fixations as training data, without knowledge of the action in each video. Our experiments on Hollywood2 show state-of-the-art performance on saliency prediction for video. We also show that our attentional model trained on Hollywood2 generalizes well to UCF101 and it can be leveraged to improve action classification accuracy on both datasets. | ["Computer vision", "Deep learning", "Applications"] | https://openreview.net/forum?id=SJRpRfKxx | https://openreview.net/pdf?id=SJRpRfKxx | https://openreview.net/forum?id=SJRpRfKxx¬eId=SyMt0S-4g |
HkmLrvg4x | SJRpRfKxx | ICLR.cc/2017/conference/-/paper83/official/review | {"title": "", "rating": "6: Marginally above acceptance threshold", "review": "This paper proposes a new method for estimating visual attention in videos. The input clip is first processed by a convnet (in particular, C3D) to extract visual features. The visual features are then passed to LSTM. The hidden state at each time step in LSTM is used to generate the parameters in a Gaussian mixture model. Finally, the visual attention map is generated from the Gaussian mixture model.\n\nOverall, the idea in this paper is reasonable and the paper is well written. RNN/LSTM has been used in lots of vision problem where the outputs are discrete sequences, there has not been much work on using RNN/LSTM for problems where the output is continuous like in this paper.\n\nThe experimental results have demonstrated the effectiveness of the proposed approach. In particular, it outperforms other state-of-the-art on the saliency prediction task on the Hollywood2 datasets. It also shows improvement over baselines (e.g. C3D + SVM) on the action recognition task.\n\nMy only \"gripe\" of this paper is that this paper is missing some important baseline comparisons. In particular, it does not seem to show how the \"recurrent\" part help the overall performance. Although Table 2 shows RMDN outperforms other state-of-the-art, it might be due to the fact that it uses strong C3D features (while other methods in Table 2 use traditional handcrafted features). Since saliency prediction is essentially a dense image labeling problem (similar to semantic segmentation). For dense image labeling, there has been lots of methods proposed in the past two years, e.g. fully convolution neural network (FCN) or deconvnet. A straightforward baseline is to simply take FCN and apply it on each frame. If the proposed method still outperforms this baseline, we can know that the \"recurrent\" part really helps.\n\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Recurrent Mixture Density Network for Spatiotemporal Visual Attention | ["Loris Bazzani", "Hugo Larochelle", "Lorenzo Torresani"] | In many computer vision tasks, the relevant information to solve the problem at hand is mixed to irrelevant, distracting information. This has motivated researchers to design attentional models that can dynamically focus on parts of images or videos that are salient, e.g., by down-weighting irrelevant pixels. In this work, we propose a spatiotemporal attentional model that learns where to look in a video directly from human fixation data. We model visual attention with a mixture of Gaussians at each frame. This distribution is used to express the probability of saliency for each pixel. Time consistency in videos is modeled hierarchically by: 1) deep 3D convolutional features to represent spatial and short-term time relations and 2) a long short-term memory network on top that aggregates the clip-level representation of sequential clips and therefore expands the temporal domain from few frames to seconds. The parameters of the proposed model are optimized via maximum likelihood estimation using human fixations as training data, without knowledge of the action in each video. Our experiments on Hollywood2 show state-of-the-art performance on saliency prediction for video. We also show that our attentional model trained on Hollywood2 generalizes well to UCF101 and it can be leveraged to improve action classification accuracy on both datasets. | ["Computer vision", "Deep learning", "Applications"] | https://openreview.net/forum?id=SJRpRfKxx | https://openreview.net/pdf?id=SJRpRfKxx | https://openreview.net/forum?id=SJRpRfKxx¬eId=HkmLrvg4x |
rJ20qUlVx | ryAe2WBee | ICLR.cc/2017/conference/-/paper26/official/review | {"title": "Weak reject", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes SEM, a simple large-size multilabel learning algorithm which models the probability of each label as softmax(sigmoid(W^T X) + b), so a one-layer hidden network. This in and of itself is not novel, nor is the idea of optimizing this by adagrad. Though it's weird that the paper explicitly derives the gradient and suggests doing alternating adagrad steps instead of the more standard adagrad steps; it's unclear whether this matters at all for performance. The main trick responsible for increasing the efficiency of this model is the candidate label sampling, which is done in a relatively standard way by sampling labels proportionally to their frequency in the dataset.\n\nGiven that neither the model nor the training strategy is novel, it's surprising that the results are better than the state-of-the-art in quality and efficiency (though non-asymptotic efficiency claims are always questionable since implementation effort trades off fairly well against performance). I feel like this paper doesn't quite meet the bar.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Multi-label learning with semantic embeddings | ["Liping Jing", "MiaoMiao Cheng", "Liu Yang", "Alex Gittens", "Michael W. Mahoney"] | Multi-label learning aims to automatically assign to an instance (e.g., an image or a document) the most relevant subset of labels from a large set of possible labels. The main challenge is to maintain accurate predictions while scaling efficiently on data sets with extremely large label sets and many training data points. We propose a simple but effective neural net approach, the Semantic Embedding Model (SEM), that models the labels for an instance as draws from a multinomial distribution parametrized by nonlinear functions of the instance features. A Gauss-Siedel mini-batch adaptive gradient descent algorithm is used to fit the model. To handle extremely large label sets, we propose and experimentally validate the efficacy of fitting randomly chosen marginal label distributions. Experimental results on eight real-world data sets show that SEM garners significant performance gains over existing methods. In particular, we compare SEM to four recent state-of-the-art algorithms (NNML, BMLPL, REmbed, and SLEEC) and find that SEM uniformly outperforms these algorithms in several widely used evaluation metrics while requiring significantly less training time.
| ["Supervised Learning"] | https://openreview.net/forum?id=ryAe2WBee | https://openreview.net/pdf?id=ryAe2WBee | https://openreview.net/forum?id=ryAe2WBee¬eId=rJ20qUlVx |
HJfuJGINx | ryAe2WBee | ICLR.cc/2017/conference/-/paper26/official/review | {"title": "not very convinced", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposes a semantic embedding based approach to multilabel classification. \nConversely to previous proposals, SEM considers the underlying parameters determining the\nobserved labels are low-rank rather than that the observed label matrix is itself low-rank. \nHowever, It is not clear to what extent the difference between the two assumptions is significant\n\nSEM models the labels for an instance as draws from a multinomial distribution\nparametrized by nonlinear functions of the instance features. As such, it is a neural network.\nThe proposed training algorithm is slightly more complicated than vanilla backprop. The significance of the results compared to NNML (in particular on large datasets Delicious and EUrlex) is not very clear. \n\nThe paper is well written and the main idea is clearly presented. However, the experimental results are not significant enough to compensate the lack of conceptual novelty. \n\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Multi-label learning with semantic embeddings | ["Liping Jing", "MiaoMiao Cheng", "Liu Yang", "Alex Gittens", "Michael W. Mahoney"] | Multi-label learning aims to automatically assign to an instance (e.g., an image or a document) the most relevant subset of labels from a large set of possible labels. The main challenge is to maintain accurate predictions while scaling efficiently on data sets with extremely large label sets and many training data points. We propose a simple but effective neural net approach, the Semantic Embedding Model (SEM), that models the labels for an instance as draws from a multinomial distribution parametrized by nonlinear functions of the instance features. A Gauss-Siedel mini-batch adaptive gradient descent algorithm is used to fit the model. To handle extremely large label sets, we propose and experimentally validate the efficacy of fitting randomly chosen marginal label distributions. Experimental results on eight real-world data sets show that SEM garners significant performance gains over existing methods. In particular, we compare SEM to four recent state-of-the-art algorithms (NNML, BMLPL, REmbed, and SLEEC) and find that SEM uniformly outperforms these algorithms in several widely used evaluation metrics while requiring significantly less training time.
| ["Supervised Learning"] | https://openreview.net/forum?id=ryAe2WBee | https://openreview.net/pdf?id=ryAe2WBee | https://openreview.net/forum?id=ryAe2WBee¬eId=HJfuJGINx |
SyqLAIg4x | ryAe2WBee | ICLR.cc/2017/conference/-/paper26/official/review | {"title": "review", "rating": "4: Ok but not good enough - rejection", "review": "The paper presents the semantic embedding model for multi-label prediction.\nIn my questions, I pointed that the proposed approach assumes the number of labels to predict is known, and the authors said this was an orthogonal question, although I don't think it is!\nI was trying to understand how different is SEM from a basic MLP with softmax output which would be trained with a two step approach instead of stochastic gradient descent. It seems reasonable given their similarity to compare to this very basic baseline.\nRegarding the sampling strategy to estimate the posterior distribution, and the difference with Jean et al, I agree it is slightly different but I think you should definitely refer to it and point to the differences.\nOne last question: why is it called \"semantic\" embeddings? usually this term is used to show some semantic meaning between trained embeddings, but this doesn't seem to appear in this paper.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Multi-label learning with semantic embeddings | ["Liping Jing", "MiaoMiao Cheng", "Liu Yang", "Alex Gittens", "Michael W. Mahoney"] | Multi-label learning aims to automatically assign to an instance (e.g., an image or a document) the most relevant subset of labels from a large set of possible labels. The main challenge is to maintain accurate predictions while scaling efficiently on data sets with extremely large label sets and many training data points. We propose a simple but effective neural net approach, the Semantic Embedding Model (SEM), that models the labels for an instance as draws from a multinomial distribution parametrized by nonlinear functions of the instance features. A Gauss-Siedel mini-batch adaptive gradient descent algorithm is used to fit the model. To handle extremely large label sets, we propose and experimentally validate the efficacy of fitting randomly chosen marginal label distributions. Experimental results on eight real-world data sets show that SEM garners significant performance gains over existing methods. In particular, we compare SEM to four recent state-of-the-art algorithms (NNML, BMLPL, REmbed, and SLEEC) and find that SEM uniformly outperforms these algorithms in several widely used evaluation metrics while requiring significantly less training time.
| ["Supervised Learning"] | https://openreview.net/forum?id=ryAe2WBee | https://openreview.net/pdf?id=ryAe2WBee | https://openreview.net/forum?id=ryAe2WBee¬eId=SyqLAIg4x |
rJUw6gmNx | HJtN5K9gx | ICLR.cc/2017/conference/-/paper515/official/review | {"title": "Review", "rating": "6: Marginally above acceptance threshold", "review": "This paper investigates deep generative models with multiple stochastic nodes and gives them meaning by semi-supervision. From a methodological point of view, there is nothing fundamentally novel (it is very similar to the semi-supervised work of Kingma et al; although this work has sometimes more than two latent nodes, it is not a complex extension). There is a fairly classical auxiliary variable trick used to make sure the inference network for y is trained over all data points (by supposing y is in fact is a latent variable with an observation \\tilde y; the observation is y if y is observed, or uninformative for unobserved y). Alternatively, one can separate the inference used to learn the generative model (which throws out inference over y if it is observed), from an inference used to 'exercise' the model (approximate the complex p(y|x) in the model by a simpler q(y|x) - effectively inferring the target p(y|x) for the data where only x is collected). Results are strong, although on simple datasets. Overall this is a well written, interesting paper, but lacking in terms of methodological advances.\n\nMinor:\n- I feel the title is a bit too general for the content of the paper. I personally don't agree with the strong contrast made between deep generative models and graphical models (deep generative models are graphical models, but they are more typically learned and un-interpretable than classical graphical models; and having multiple stochastic variables is not exclusive to graphical models, see DRAW, Deep Kalman Filter, Recurrent VAE, etc.). The word 'structure' is a bit problematic; here, the paper seems more concerned with disentangling and semanticizing the latent representation of a generative model by supervision. It is debatable whether the models themselves have structure.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Learning Disentangled Representations in Deep Generative Models | ["N. Siddharth", "Brooks Paige", "Alban Desmaison", "Jan-Willem van de Meent", "Frank Wood", "Noah D. Goodman", "Pushmeet Kohli", "Philip H.S. Torr"] | Deep generative models provide a powerful and flexible means to learn complex distributions over data by incorporating neural networks into latent-variable models. Variational approaches to training such models introduce a probabilistic encoder that casts data, typically unsupervised, into an entangled and unstructured representation space. While unsupervised learning is often desirable, sometimes even necessary, when we lack prior knowledge about what to represent, being able to incorporate domain knowledge in characterising certain aspects of variation in the data can often help learn better disentangled representations. Here, we introduce a new formulation of semi-supervised learning in variational autoencoders that allows precisely this. It permits flexible specification of probabilistic encoders as directed graphical models via a stochastic computation graph, containing both continuous and discrete latent variables, with conditional distributions parametrised by neural networks. We demonstrate how the provision of structure, along with a few labelled examples indicating plausible values for some components of the latent space, can help quickly learn disentangled representations. We then evaluate its ability to do so, both qualitatively by exploring its generative capacity, and quantitatively by using the disentangled representation to perform classification, on a variety of models and datasets. | ["Semi-Supervised Learning", "Deep learning", "Computer vision"] | https://openreview.net/forum?id=HJtN5K9gx | https://openreview.net/pdf?id=HJtN5K9gx | https://openreview.net/forum?id=HJtN5K9gx¬eId=rJUw6gmNx |
B1FyYJmVe | HJtN5K9gx | ICLR.cc/2017/conference/-/paper515/official/review | {"title": "A variant of the semi-supervised VAE.", "rating": "6: Marginally above acceptance threshold", "review": "This paper proposed a variant of the semi-supervised VAE model which leads to a unified objective for supervised and unsupervised VAE. This variant gives software implementation of these VAE models more flexibility in specifying which variables are supervised and which are not.\n\nThis development introduces a few extra terms compared to the original semi-supervised VAE formulation proposed by Kingma et al., 2014. From the experiment results it seems that these terms do not do much as the new formulation and the performance difference between the proposed method and Kingma et al. 2014 are not very significant (Figure 5). Therefore the benefit of the new formulation is likely to be just software engineering flexibility and convenience.\n\nThis flexibility and convenience is nice to have, but it is better to demonstrate a few situations where the proposed method can be applied while for other previous methods it is non-trivial to do.\n\nThe paper's title and the way it is written make me expect a lot more than what is currently in the paper. I was expecting to see, for example, structured hidden variable model for the posterior (page 4, top), or really \"structured interpretation\" of the generative model (title), but I didn't see any of these. The main contribution of this paper (a variant of the semi-supervised VAE model) is quite far from these.\n\nAside from these, the plug-in estimation for discrete variables only works when the function h(x,y) is a continuous function of y. If however, h(x, y) is not continuous in y, for example h takes one form when y=1 and another form when y=2, then the approach of using Expectation[y] to replace y will not work. Therefore the \"plug-in\" estimation has its limitations.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Learning Disentangled Representations in Deep Generative Models | ["N. Siddharth", "Brooks Paige", "Alban Desmaison", "Jan-Willem van de Meent", "Frank Wood", "Noah D. Goodman", "Pushmeet Kohli", "Philip H.S. Torr"] | Deep generative models provide a powerful and flexible means to learn complex distributions over data by incorporating neural networks into latent-variable models. Variational approaches to training such models introduce a probabilistic encoder that casts data, typically unsupervised, into an entangled and unstructured representation space. While unsupervised learning is often desirable, sometimes even necessary, when we lack prior knowledge about what to represent, being able to incorporate domain knowledge in characterising certain aspects of variation in the data can often help learn better disentangled representations. Here, we introduce a new formulation of semi-supervised learning in variational autoencoders that allows precisely this. It permits flexible specification of probabilistic encoders as directed graphical models via a stochastic computation graph, containing both continuous and discrete latent variables, with conditional distributions parametrised by neural networks. We demonstrate how the provision of structure, along with a few labelled examples indicating plausible values for some components of the latent space, can help quickly learn disentangled representations. We then evaluate its ability to do so, both qualitatively by exploring its generative capacity, and quantitatively by using the disentangled representation to perform classification, on a variety of models and datasets. | ["Semi-Supervised Learning", "Deep learning", "Computer vision"] | https://openreview.net/forum?id=HJtN5K9gx | https://openreview.net/pdf?id=HJtN5K9gx | https://openreview.net/forum?id=HJtN5K9gx¬eId=B1FyYJmVe |
HyW_r-fEl | HJtN5K9gx | ICLR.cc/2017/conference/-/paper515/official/review | {"title": "", "rating": "5: Marginally below acceptance threshold", "review": "This paper introduces a variant of the semi-supervised variational auto-encoder (VAE) framework. The authors present a way of introducing structure (observed variables) inside the recognition network.\n\nI find that the presentation of the inference with auxiliary variables could be avoided, as it actually makes the presentation unnecessarily complicated. Specifically, the expressions with auxiliary variables are helpful for devising a unified implementation, but modeling-wise one can get the same model without these auxiliary variables and recover a minimal extension of VAE where part of the generating space is actually observed. The observed variables mean that the posterior needs to also condition on those, so as to incorporate the information they convey. The way this is done in this paper is actually not very different from Kingma et al. 2014, and I am surprised that the experiments show a large deviation in these two methods' results. Given the similarity of the models, it'd be useful if the authors could give a possible explanation on the superiority of their method compared to Kingma et al. 2014. By the way, I was wondering if the experimental setup is the same as in Kingma et al. 2014 for the results of Fig. 5 (bottom) - the authors mention that they use CNNs for feature extraction but from the paper it's not clear if Kingma et al. do the same. \n\nOn a related note, I was wondering the same for the comparison with Jampani et al. 2015. In particular, is that model also using the same rate of supervision for a fair comparison?\n\nThe experiment in section 4.3 is interesting and demonstrates a useful property of the approach.\n\nThe discussion of the supervision rate (and the pre-review answer) is helpful in giving some insight about what is a successful training protocol to use in semi-supervised learning.\n\nOverall, the paper is interesting but the title and introduction made me expect something more from it. From the title I expected a method for interpreting general deep generative models, instead the described approach was about a semi-supervised variant of VAE - naturally including labelled examples disentangles the latent space, but this is a general property of any semi-supervised probabilistic model and not unique to the approach described here. Moreover, from the intro I expected to see a more general approximation scheme for the variational posterior (similar to Ranganath et al. 2015 which trully allows very flexible distributions), however this is not the case here.\n\nGiven the above, the contributions of this paper are in defining a slight variant of the semi-supervised VAE, and (perhaps more importantly) formulating it in a way that is amendable to easier automation in terms of software. But methodologically there is not much contribution to the current literature. The authors mention that they plan to extend the framework in the probabilistic programming setting. It seems indeed that this would be a very promising and useful extension. \n\nMinor note: three of Kingma's papers are all cited in the main text as Kingma et al. 2014, causing confusion. I suggest using Kingma et al. 2014a etc.\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Learning Disentangled Representations in Deep Generative Models | ["N. Siddharth", "Brooks Paige", "Alban Desmaison", "Jan-Willem van de Meent", "Frank Wood", "Noah D. Goodman", "Pushmeet Kohli", "Philip H.S. Torr"] | Deep generative models provide a powerful and flexible means to learn complex distributions over data by incorporating neural networks into latent-variable models. Variational approaches to training such models introduce a probabilistic encoder that casts data, typically unsupervised, into an entangled and unstructured representation space. While unsupervised learning is often desirable, sometimes even necessary, when we lack prior knowledge about what to represent, being able to incorporate domain knowledge in characterising certain aspects of variation in the data can often help learn better disentangled representations. Here, we introduce a new formulation of semi-supervised learning in variational autoencoders that allows precisely this. It permits flexible specification of probabilistic encoders as directed graphical models via a stochastic computation graph, containing both continuous and discrete latent variables, with conditional distributions parametrised by neural networks. We demonstrate how the provision of structure, along with a few labelled examples indicating plausible values for some components of the latent space, can help quickly learn disentangled representations. We then evaluate its ability to do so, both qualitatively by exploring its generative capacity, and quantitatively by using the disentangled representation to perform classification, on a variety of models and datasets. | ["Semi-Supervised Learning", "Deep learning", "Computer vision"] | https://openreview.net/forum?id=HJtN5K9gx | https://openreview.net/pdf?id=HJtN5K9gx | https://openreview.net/forum?id=HJtN5K9gx¬eId=HyW_r-fEl |
rJygw6GVg | SkJeEtclx | ICLR.cc/2017/conference/-/paper474/official/review | {"title": "Review", "rating": "4: Ok but not good enough - rejection", "review": "The authors propose a \"hierarchical\" attention model for video captioning. They introduce a model composed of three parts: the temporal modeler (TEM) that takes as input the video sequence and outputs a sequential representation of the video to the HAM; the hierarchical attention/memory mechanism (HAM) implements a soft-attention mechanism over the sequential video representation; and finally a decoder that generates a caption. \n\nRelated to the second series of questions above, it seems as though the authors have chosen to refer to their use of an LSTM (or equivalent RNN) as the output of the Bahdanau et al (2015) attention mechanism as a hierarchical memory mechanism. I am actually sympathetic to this terminology in the sense that the recent popularity of memory-based models seems to neglect the memory implicit in the LSTM state vector, but that said, this seems to seriously misrepresent the significance fo the contribution of this paper. \n\nI appreciate the ablation study presented in Table 1. Not enough researchers bother with this kind of analysis. But it does show that the value of the contributions is not actually clear. In particular the case for the TEM is quite weak.\n\nRegarding the quantitative evaluation presented in Table 2, the authors are carving out a fairly specific set of features to describe the set of \"fair\" comparators from the literature. Given the variability of the models and alternate training datasets that are in use, I would find it more compelling if the authors just set about trying to achieve the best results they can, if that includes the fine-tuning of the frame model, so be it. The value of this work is as an application paper, so the discovery and incorporation of elements that can significantly improve performance would seems warranted. \n\nOverall, at this point, I do not see a sufficient contribution to warrant publication in ICLR.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Memory-augmented Attention Modelling for Videos | ["Rasool Fakoor", "Abdel-rahman Mohamed", "Margaret Mitchell", "Sing Bing Kang", "Pushmeet Kohli"] | Recent works on neural architectures have demonstrated the utility of attention mechanisms for a wide variety of tasks. Attention models used for problems such as image captioning typically depend on the image under consideration, as well as the previous sequence of words that come before the word currently being generated. While these types of models have produced impressive results, they are not able to model the higher-order interactions involved in problems such as video description/captioning, where the relationship between parts of the video and the concepts being depicted is complex. Motivated by these observations, we propose a novel memory-based attention model for video description. Our model utilizes memories of past attention when reasoning about where to attend to in the current time step, similar to the central executive system proposed in human cognition. This allows the model to not only reason about local attention more effectively, it allows it to consider the entire sequence of video frames while generating each word. Evaluation on the challenging and popular MSVD and Charades datasets show that the proposed architecture outperforms all previously proposed methods and leads to a new state of the art results in the video description. | ["Deep learning", "Multi-modal learning", "Computer vision"] | https://openreview.net/forum?id=SkJeEtclx | https://openreview.net/pdf?id=SkJeEtclx | https://openreview.net/forum?id=SkJeEtclx¬eId=rJygw6GVg |
Hk8khIG4l | SkJeEtclx | ICLR.cc/2017/conference/-/paper474/official/review | {"title": "Review: Novelty and performance claims do not hold, missing clarity, no significant improvements over ablations.", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposes an attention-based approach for video description. The approach uses three LSTMs and two attention mechanisms to sequentially predict words from a sequence of frames.\nIn the LSTM-encoder of the frames (TEM), the first attention approach predicts a spatial attention per frame, and computes the weighted average. The second LSTM (HAM) predicts an attention over the hidden states of the encoder LSTM.\nThe third LSTM which run temporally in parallel to the second LSTM generates the sentence, one word at a time.\n\n\nStrength:\n===============\n\n-\tThe paper works on a relevant and interesting problem.\n-\tUsing 2 layers of attention in the proposed way have to my knowledge not been used before for video description. The exact architecture is thus novel (but the work claims much more without sufficient attribution, see blow)\n-\tThe experiments are evaluated on two datasets, MSVD and Charades, showing performance on the level of related work for MSVD and improvements for Charades.\n\nWeaknesses:\n===============\n\n1.\tClaims about the contribution/novelty of the model seem not to hold: \n1.1.\tOne of the main contributions is the Hierarchical Attention/Memory (HAM):\n1.1.1.\tIt is not clear to me how the presented model (Eq 6-8), are significantly different from the presented model in Xu et al / Yao et al. While Xu et al. attends over spatial image locations and Yao et al. attend over frames, this model attends over encoded video representations h_v^i. A slight difference might be that Xu et al. use the same LSTM to generate, while this model uses an additional LSTM for the decoding.\n1.1.2.\tThe paper states in section 3.2 \u201cwe propose f_m to memorize the previous attention\u201d, however H_m^{t\u2019-1} only consist of the last hidden state. Furthermore, the model f_m does not have access to the \u201cattention\u201d \\alpha. This was also discussed in comments by others, but remains unclear.\n1.1.3.\tIn the discussion of comments the authors claim that \u201cattention not only is a function a current time step but also a function of all previous attentions and network states.\u201d: While it is true that there is a dependency but that is true also for any LSTM, however the model does not have access to the previous network states as H_g^{t\u2019-1} only consist of the last hidden state, as well as H_m^{t\u2019-1} [at least that is what the formulas say and what Figure 1 suggests]. \n1.1.4.\tThe authors claim to have multi-layer attention in HAM, however it remains unclear where the multi-layer comes from.\n1.2.\tThe paper states that in section 3.1. \u201c[CNN] features tend to discard the low level information useful in modeling the motion in the video (Ballas et al., 2016).\u201d This suggests that the approach which follows attacks this problem. However, it cannot model motion as attention \\rho between frames is not available when predicting the next frame. Also, it is not clear how the model can capture anything \u201clow level\u201d as it operates on rather high level VGG conv 5 features.\n\n2.\tRelated work: The difference of HAM to Yao et al. and Xu et al. should be made more clear / or these papers should be cited in the HAM section.\n\n3.\tConceptual Limitation of the model: The model has two independent attention mechanisms, a spatial one, and a temporal one. The spatial (within a frame) is independent of the sentence generation. It thus cannot attend to different aspects of the frames for different words which would make sense. E.g. if the sentence is \u201cthe dog jumps on the trampoline\u201d, the model should focus on the dog when saying \u201cdog\u201d and on the trampoline when saying \u201ctrampoline\u201d, however, as the spatial attention is fixed this is difficult. Also, the encoder model does not have an explicitly way to look at different aspects in the frame during the encoding so might likely get stuck and always predict the same spatial attention for all frames (or it might e.g. always attend to the dog which moves around, but never on the scene).\n\n4.\tEq 11 contradicts Fig 1: How is the model exactly receiving the previous word as input. Eq. 11 suggests it is the softmax. If this is the case, the authors should emphasize this in the text as this is unusual. More common would be to use the ground truth previous word during training (which Fig 11 suggests) and the \u201chardmax\u201d, i.e. the highest predicted previous word encoded as one-hot vector at test time.\n\n5.\tClarity:\n5.1.\tIt would be helpful if the same notation would be used in Eq 2-5 and 6-9. Why is a different notation required?\n5.2.\tIt would be helpful if Fig 1 could contain more details or additional figures for the corresponding parts would be added. If space is a problem, e.g. the well-known equations for LSTM, softmax (Eq 2), and log likelihood loss (Eq 12) could be omitted or inlined.\n\n6.\tEvaluation:\n6.1.\tThe paper claims that the \u201cthe proposed architecture outperforms all previously proposed methods and leads to a new state of the art results\u201d.\n6.1.1.\tFor the MSVD dataset this clearly is wrong, even given the same feature representation. Pan et al. (2016 a) in Table 2 achieve higher METEOR (33.10).\n6.1.2.\tFor this strong claim, I would also expect that it outperforms all previous results independent of the features used, which is also wrong again, Yu et al achieve higher performance in all compared metrics.\n6.1.3.\tFor Charades dataset, this claim is also too bold as hardly any methods have been evaluated on this dataset, so at least all the ablations reported in Table 1 should also be reported for the Charades dataset, to make for this dataset any stronger claims.\n6.2.\tMissing qualitative results of attention: The authors should show qualitative results of the attention, for both attention mechanisms to understand if anything sensible is happening there. How diverse is the spatial and the temporal attention? Is it peaky or rather uniform?\n6.3.\tPerformance improvement is not significant over model ablations: The improvements over Att+No TEM is only 0.5 Meteor, 0.7 Blue@4 and the performance drops for CIDEr by 1.7.\n6.4.\tMissing human evaluation: I disagree with the authors that a human evaluation is not feasible. 1. An evaluation on a subset of the test data is not so difficult. 2. Even if other authors do not provide their code/model [and some do], they are typically happy to share the predicted sentences which is sufficient and even better for human evaluation [if not I would explicitly mention that some authors did not share sentences, as this seems clearly wrong]. 3. For model ablations the sentences are available to the authors.\n\n7.\tSeveral of the comments raised by reviewers/others have not yet been incorporated in a revised version of the paper and/or are still not clear from the explanations given. E.g. including SPICE evaluation and making fixes seems trivial.\t\n\n8.\tHyperparameters are inconsistent: Why are the hyperparemters inconsistent between the ablation analysis (40 frames are sampled) and the performance comparison (8 frames)? Should this not be selected on the validation set? What is the performance of all the ablations with 8 frames?\n\nOther (minor/discussion points)\n-\tEquation 10: what happens with h_m, and h_g, the LSTM formulas provided only handle two inputs. Are h_m and h_g concatenated.\n-\tThere is a section 4.1 but no 4.2.\n-\tThe paper states in section 4.1 \u201cour proposed architecture can alone not only learn a representation for video that can model the temporal structure of a video sequence, but also a representation that can effectively map visual space to the language space.\u201d However, this seems to be true also for many/most other approaches, e.g. [Venugopalan et al. 2015 ICCV]\n\nSummary:\n===============\n\nWhile the paper makes strong claims w.r.t. to the approach and results, the approach lacks novelty and the results are not convincing over related work and ablations. Furthermore, improved clarity and visualizations of the model and attention results would benefit the paper.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Memory-augmented Attention Modelling for Videos | ["Rasool Fakoor", "Abdel-rahman Mohamed", "Margaret Mitchell", "Sing Bing Kang", "Pushmeet Kohli"] | Recent works on neural architectures have demonstrated the utility of attention mechanisms for a wide variety of tasks. Attention models used for problems such as image captioning typically depend on the image under consideration, as well as the previous sequence of words that come before the word currently being generated. While these types of models have produced impressive results, they are not able to model the higher-order interactions involved in problems such as video description/captioning, where the relationship between parts of the video and the concepts being depicted is complex. Motivated by these observations, we propose a novel memory-based attention model for video description. Our model utilizes memories of past attention when reasoning about where to attend to in the current time step, similar to the central executive system proposed in human cognition. This allows the model to not only reason about local attention more effectively, it allows it to consider the entire sequence of video frames while generating each word. Evaluation on the challenging and popular MSVD and Charades datasets show that the proposed architecture outperforms all previously proposed methods and leads to a new state of the art results in the video description. | ["Deep learning", "Multi-modal learning", "Computer vision"] | https://openreview.net/forum?id=SkJeEtclx | https://openreview.net/pdf?id=SkJeEtclx | https://openreview.net/forum?id=SkJeEtclx¬eId=Hk8khIG4l |
SyTtIyf4x | SkJeEtclx | ICLR.cc/2017/conference/-/paper474/official/review | {"title": "lacks polish and clarity, moderately original but weak results", "rating": "4: Ok but not good enough - rejection", "review": "\nThis paper addresses video captioning with a TEM-HAM architecture, where a HAM module attends over attended outputs of the TEM module when generating the description. This gives a kind of 2-level attention. The model is evaluated on the Charades and MSVD datasets.\n\n1. Quality/Clarify: I found this paper to be poorly written and relatively hard to understand. As far as I can tell the TEM module of Section 3.1 is a straight-forward attention frame encoder of Bahdanau et al. 2015 or Xu et al. 2015. The decoder of Section 3.3 is a standard LSTM with log likelihood. The HAM module of Section 3.2 is the novel module but is not very well described. It looks to be an attention LSTM where the attention is over the TEM LSTM outputs, but the attention weights are additionally conditioned on the decoder state. There are a lot of small problems with the description, such as notational discrepancy in using \\textbf in equations and then not using it in the text. Also, I spent a long time trying to understand what f_m is. The authors say: \n\n\"In order to let the network remember what has been attended before and the temporal\nstructure of a video, we propose f_m to memorize the previous attention and encoded version of an\ninput video with language model. Using f_m not only enables the network to memorize previous\nattention and frames, but also to learn multi-layer attention over an input video and corresponding\nlanguage.\"\n\nWhere one f_m is bold and the other f_m is not. Due to words such as \"we propose f_m\" assumed this was some kind of a novel technical contribution I couldn't find any details about but it is specified later in Section 3.3 at the end that f_m is in fact just an LSTM. It's not clear why this piece of information is in Section 3.3, which discusses the decoder. The paper is sloppy in other parts. For example in Table 1 some numbers have 1 significant digit and some have 2. The semantics of the horizontal line in Table 2 are not explained in text. \n\n2. Experimental results: The ablation study shows mixed results when adding TEM and HAM to the model. Looking at METEOR which was shown to have the highest correlation to humans in the COCO paper compared to the other evaluation criteria, adding TEM+HAM improves the model from 31.20 to 31.70. It is not clear how significant this improvement is, especially given that the test set is only 670 videos. I have doubts over this result. In Table 2, the METEOR score of Pan et al. 2016a is higher [33.10 vs. 31.80], but this discrepancy is not addressed in text. This is surprising because the authors explicitly claim \"state of the art results\".\n\n3. Originality/Significance: The paper introduces an additional layer of attention over a more standard sequence to sequence setup, which is argued to alleviate the burden on the LSTM's memory. This is moderately novel but I don't believe that the experimental results make it sufficiently clear that it is also worth doing. If the paper made the standard model somehow simpler instead of more complex I would be more inclined to judge it favorably.\n\n\nMinor:\nIn response to the author's comment \"not sure what causes to think of RBM. We don't model any part of our architecture using RBM. We'd be appreciated if you please elaborate more about your confusion about figure 1 so we can address it accordingly.\", I created a diagram to hopefully make this more clear: https://imgur.com/a/4MJaG . It is very common to use 2 rows of circles with lines between them pairwise to denote an RBM. The lines indicate undirected edges of the graphical model. A typical RBM diagram can have, for example, 3 circles on top and 4 circles on the bottom joined with edges. Your diagram has 4 circles on the top and 5 circles on the bottom joined with edges. However, not all of your circles from the top and bottom are connected, which further adds to the difficulty. In particular, your last circle on the bottom is only connected to 2 circles on the top. If the authors are trying to denote a neural network I would advise using arrows instead of lines. However, since the HAM module uses an LSTM just like the encoder and the decoder it is unclear why this module has a visually distinct appearance at all."} | review | 2017 | ICLR.cc/2017/conference | Memory-augmented Attention Modelling for Videos | ["Rasool Fakoor", "Abdel-rahman Mohamed", "Margaret Mitchell", "Sing Bing Kang", "Pushmeet Kohli"] | Recent works on neural architectures have demonstrated the utility of attention mechanisms for a wide variety of tasks. Attention models used for problems such as image captioning typically depend on the image under consideration, as well as the previous sequence of words that come before the word currently being generated. While these types of models have produced impressive results, they are not able to model the higher-order interactions involved in problems such as video description/captioning, where the relationship between parts of the video and the concepts being depicted is complex. Motivated by these observations, we propose a novel memory-based attention model for video description. Our model utilizes memories of past attention when reasoning about where to attend to in the current time step, similar to the central executive system proposed in human cognition. This allows the model to not only reason about local attention more effectively, it allows it to consider the entire sequence of video frames while generating each word. Evaluation on the challenging and popular MSVD and Charades datasets show that the proposed architecture outperforms all previously proposed methods and leads to a new state of the art results in the video description. | ["Deep learning", "Multi-modal learning", "Computer vision"] | https://openreview.net/forum?id=SkJeEtclx | https://openreview.net/pdf?id=SkJeEtclx | https://openreview.net/forum?id=SkJeEtclx¬eId=SyTtIyf4x |
rkAXNP7Ex | r1R5Z19le | ICLR.cc/2017/conference/-/paper158/official/review | {"title": "Not clearly enough related to previous work.", "rating": "4: Ok but not good enough - rejection", "review": "The authors propose a semi-supervised technique for neural networks which includes two objectives: (1) the neural net embeddings of two samples with identical labels is constrained to be closer than the embeddings of samples with different labels (2) the embedding of an unlabeled example is constrained to be close to the embeddings of the closest labeled sample (and far away from the other ones).\n\nWhile the authors list a number of previous works, they do not relate them very well with their approach, and actual differences appear unclear. In particular, the approach seems rather incremental with respect to (Hadsell et al. 2006), the way the neighbors are chosen being the main difference; in that respect, (Weston et al, 2008 or 2012) (which could be viewed as an application of DrLim from Hadsell et al, applied to semi-supervised learning) is even closer to the approach proposed here.\n\nThere is also no balancing constraint in the proposed approach, which was known to be crucial in all these models from the 2000s.\n\nConcerning the experimental results, reported performance are very good; given the approach is very close to previous work, it is hard to know if good performance come from better neural net architectures or something else. This should be clarified.\n\nIn summary, the novelty is not clearly defined in this paper; differences with existing literature should be highlighted. Experimental results are very good, but it is hard to know where comes the difference in performance with previous (very related) work.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Semi-supervised deep learning by metric embedding | ["Elad Hoffer", "Nir Ailon"] | Deep networks are successfully used as classification models yielding state-of-the-art results when trained on a large number of labeled samples. These models, however, are usually much less suited for semi-supervised problems because of their tendency to overfit easily when trained on small amounts of data. In this work we will explore a new training objective that is targeting a semi-supervised regime with only a small subset of labeled data. This criterion is based on a deep metric embedding over distance relations within the set of labeled samples, together with constraints over the embeddings of the unlabeled set. The final learned representations are discriminative in euclidean space, and hence can be used with subsequent nearest-neighbor classification using the labeled samples. | ["Deep learning", "Semi-Supervised Learning"] | https://openreview.net/forum?id=r1R5Z19le | https://openreview.net/pdf?id=r1R5Z19le | https://openreview.net/forum?id=r1R5Z19le¬eId=rkAXNP7Ex |
S1ZMS80Qx | r1R5Z19le | ICLR.cc/2017/conference/-/paper158/official/review | {"title": "", "rating": "6: Marginally above acceptance threshold", "review": "This work presents an embedding approach for semi-supervised learning with neural nets, in the presence of little labeled data. The intuition is to learn a metric embedding that forms \u201cclusters\u201d with the following desiderata: two labeled examples from the class should have a smaller distance in this embedding compared to any example from another class & a given unlabeled example embedding will be closer to all of the embeddings of *some* label (i.e. that a given unlabeled example will be \u201cmatched\u201d to one cluster). The paper formulates these intuitions as two differentiable losses and does gradient descent on their sum.\n\nIt\u2019s unclear to me how different is this work from the sum of Hoffer & Ailon (2015) (which is eq. 3) and Grandvalet & Bengio (2004) (seems to be related to eq. 4). Would be nice if the authors not only cited the previous work but summarized the actual differences.\nIn Section 5.1, the authors say that Szegedy et al. (2015) use random noise in the targets -- is that actually true? I think only soft targets are used (which are not noisy).\n\nDoes the choice of \\lambda_{1,2} make a difference?\n\nHow is k for k-NN actually chosen? Is there a validation set?\n\nFigure 1 would benefit from showing where the labeled examples were at the beginning of training (relative to each other / rest of the data).\n\nThe submission seems overall OK, but somewhat light on actual data-driven or theoretical insights. I would\u2019ve liked experiments showing the influence of data set sizes at the very least, and ablation experiments that showed the influence of each of the corresponding losses.\n\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Semi-supervised deep learning by metric embedding | ["Elad Hoffer", "Nir Ailon"] | Deep networks are successfully used as classification models yielding state-of-the-art results when trained on a large number of labeled samples. These models, however, are usually much less suited for semi-supervised problems because of their tendency to overfit easily when trained on small amounts of data. In this work we will explore a new training objective that is targeting a semi-supervised regime with only a small subset of labeled data. This criterion is based on a deep metric embedding over distance relations within the set of labeled samples, together with constraints over the embeddings of the unlabeled set. The final learned representations are discriminative in euclidean space, and hence can be used with subsequent nearest-neighbor classification using the labeled samples. | ["Deep learning", "Semi-Supervised Learning"] | https://openreview.net/forum?id=r1R5Z19le | https://openreview.net/pdf?id=r1R5Z19le | https://openreview.net/forum?id=r1R5Z19le¬eId=S1ZMS80Qx |
r1pQ4-zNe | S1J0E-71l | ICLR.cc/2017/conference/-/paper2/official/review | {"title": "Misleading", "rating": "3: Clear rejection", "review": "This paper proposes to use previous error signal of the output layer as an additional input to recurrent update function in order to enhance the modelling power of a dynamic system such as RNNs. \n\n-This paper makes an erroneous assumption: test label information is not given in most of the real world applications, except few applications. This means that the language modelling task, which is the only experiment of this paper, may not be the right task to test this approach. Also, comparing against the models that do not use test error signal at inference time is unfair. We cannot just say that the test label information is being observed, this only holds in online-prediction problems.\n\n-The experiment is only conducted on one dataset, reporting state-of-the-art result, but unfortunately this is not true. There are already more than four papers reporting better numbers than the one reported in this task, however the author did not cite them. I understand that this paper came before the other papers, but the manuscript should be updated before the final decision.\n\n-The model size is still missing and without this information, it is hard to judge the contribution of the proposed trick.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Surprisal-Driven Feedback in Recurrent Networks | ["Kamil Rocki"] | Recurrent neural nets are widely used for predicting temporal data. Their inherent deep feedforward structure allows learning complex sequential patterns. It is believed that top-down feedback might be an important missing ingredient which in theory could help disambiguate similar patterns depending on broader context. In this paper, we introduce surprisal-driven recurrent networks, which take into account past error information when making new predictions. This is achieved by continuously monitoring the discrepancy between most recent predictions and the actual observations. Furthermore, we show that it outperforms other stochastic and fully deterministic approaches on enwik8 character level prediction task achieving 1.37 BPC. | ["Unsupervised Learning", "Applications", "Deep learning"] | https://openreview.net/forum?id=S1J0E-71l | https://openreview.net/pdf?id=S1J0E-71l | https://openreview.net/forum?id=S1J0E-71l¬eId=r1pQ4-zNe |
SyWZXPLEg | S1J0E-71l | ICLR.cc/2017/conference/-/paper2/official/review | {"title": "Badly Written", "rating": "4: Ok but not good enough - rejection", "review": "Summary:\nThis paper proposes to use surprisal-driven feedback for training recurrent neural networks where they feedback the next-step prediction error of the network as an input to the network. Authors have shown a result on language modeling tasks.\n\nContributions:\nThe introduction of surprisal-driven feedback, which is just the feedback from the errors of the model from the previous time-steps.\n\nQuestions:\nA point which is not fully clear from the paper is whether if you have used the ground-truth labels on the test set for the surprisal feedback part of the model? I assume that authors do that since they claim that they use the misprediction error as additional input.\n\nCriticisms:\nThe paper is really badly written, authors should rethink the organization of the paper.\nMost of the equations presented in the paper, about BPTT are not necessary for the main-text and could be moved to Appendix. \nThe justification is not convincing enough.\nExperimental results are lacking, only results on a single dataset are provided.\nAlthough the authors claim that they got SOTA on enwiki8, there are other papers such as the HyperNetworks that got better results (1.34) than the result they achieve. This claim is wrong.\nThe model requires the ground-truth labels for the test-set, however, this assumption really limits the application of this technique to a very limited set of applications(more or less rules out most conditional language modeling tasks).\n\nHigh-level Review:\n Pros: \n - A simple modification of the model that seems to improve the results and it is an interesting modification.\n\n Cons:\n - The authors need to use test-set labels.\n - Writing of the paper is bad.\n - The authors assume that they have access to the ground-truth labels during the test-set.\n - Experimental results are lacking", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Surprisal-Driven Feedback in Recurrent Networks | ["Kamil Rocki"] | Recurrent neural nets are widely used for predicting temporal data. Their inherent deep feedforward structure allows learning complex sequential patterns. It is believed that top-down feedback might be an important missing ingredient which in theory could help disambiguate similar patterns depending on broader context. In this paper, we introduce surprisal-driven recurrent networks, which take into account past error information when making new predictions. This is achieved by continuously monitoring the discrepancy between most recent predictions and the actual observations. Furthermore, we show that it outperforms other stochastic and fully deterministic approaches on enwik8 character level prediction task achieving 1.37 BPC. | ["Unsupervised Learning", "Applications", "Deep learning"] | https://openreview.net/forum?id=S1J0E-71l | https://openreview.net/pdf?id=S1J0E-71l | https://openreview.net/forum?id=S1J0E-71l¬eId=SyWZXPLEg |
SJvqO1GEl | S1J0E-71l | ICLR.cc/2017/conference/-/paper2/official/review | {"title": "Need some revisions", "rating": "3: Clear rejection", "review": "\nThis paper proposes to leverage \"surprisal\" as top-down signal in RNN. More specifically author uses the error corresponding to the previous prediction as an extra input at the current timestep in a LSTM.\n\nThe general idea of suprising-driven feedback is interesting for online prediction task. It is a simple enough idea that seems to bring some significant improvements. However, the paper in its current form has some important flaws.\n\n- Overall, the paper writing could be improved. In particular, section 2.4 and 2.5 is composed mostly by the equations of the forward and backward propagation of feedback RNN and feedback LSTM. However, author provides no analysis along with those equations. It is therefore not clear what insight the author tries to express in those sections. In addition, feedback RNN is not evaluated in the experimental section, so it is not clear why feedback RNN is described.\n\n- The experimental evaluation is limited. Only one dataset enwik8 is explored. I think it is necessary to try the idea on different datasets to see if feedback LSTM sees some consistent improvements.\nAlso, author claims state-of-art on enwik8, but hypernetwork, already cited in the paper, achieves better results (1.34 BPC, table 4 in the hypernetworks paper).\n\n- Author only compares to methods that do not use last prediction error as extra signal. I would argue that a comparison with dynamic evaluation would be more fair. \n Feedback LSTM uses prediction error as extra input in the forward prop, while dynamic evaluation backprop it through the network and change the weight accordingly. Also they don't propagate the prediction error in the same way, they both leverage \"extra\" supervised information through the prediction errors.\n\n\nIn summary:\nPros: \n- Interesting idea\n- Seems to improve performances\n\nCons:\n- Paper writing\n- Weak evaluation (only one dataset)\n- Compare only with approaches that does not use the last-timestep error signal", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Surprisal-Driven Feedback in Recurrent Networks | ["Kamil Rocki"] | Recurrent neural nets are widely used for predicting temporal data. Their inherent deep feedforward structure allows learning complex sequential patterns. It is believed that top-down feedback might be an important missing ingredient which in theory could help disambiguate similar patterns depending on broader context. In this paper, we introduce surprisal-driven recurrent networks, which take into account past error information when making new predictions. This is achieved by continuously monitoring the discrepancy between most recent predictions and the actual observations. Furthermore, we show that it outperforms other stochastic and fully deterministic approaches on enwik8 character level prediction task achieving 1.37 BPC. | ["Unsupervised Learning", "Applications", "Deep learning"] | https://openreview.net/forum?id=S1J0E-71l | https://openreview.net/pdf?id=S1J0E-71l | https://openreview.net/forum?id=S1J0E-71l¬eId=SJvqO1GEl |
SJ3MEJPEg | Skq89Scxx | ICLR.cc/2017/conference/-/paper263/official/review | {"title": "Great trick worth publishing, but is there enough material for a full paper?", "rating": "7: Good paper, accept", "review": "This heuristic to improve gradient descent in image classification is simple and effective, but this looks to me more like a workshop track paper. Demonstration of the algorithm is limited to one task (CIFAR) and there is no theory to support it, so we do not know how it will generalize on other tasks\n\nWorking on DNNs for NLP, I find some observations in the paper opposite to my own experience. In particular, with architectures that combine a wide variety of layer types (embedding, RNN, CNN, gating), I found that ADAM-type techniques far outperform simple SGD with momentum, as they save searching for the right learning rate for each type of layer. But ADAM only works well combined with Poliak averaging, as it fluctuates a lot from one batch to another.\n\nRevision:\n- the authors substantially improved the contents of the paper, including experiments on another set than Cifar\n- the workshop track has been modified to breakthrough work, so my recommendation for it is not longer appropriate\nI have therefore improved my rating", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | SGDR: Stochastic Gradient Descent with Warm Restarts | ["Ilya Loshchilov", "Frank Hutter"] | Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets,
where we demonstrate new state-of-the-art results at 3.14\% and 16.21\%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at \\ \url{https://github.com/loshchil/SGDR} | ["Deep learning", "Optimization"] | https://openreview.net/forum?id=Skq89Scxx | https://openreview.net/pdf?id=Skq89Scxx | https://openreview.net/forum?id=Skq89Scxx¬eId=SJ3MEJPEg |
SkMtuVxEx | Skq89Scxx | ICLR.cc/2017/conference/-/paper263/official/review | {"title": "An effective method to improve convergence of neural network training", "rating": "7: Good paper, accept", "review": "This paper describes a way to speed up convergence through sudden increases of otherwise monotonically decreasing learning rates. Several techniques are presented in a clear way and parameterized method is proposed and evaluated on the CIFAR task. The concept is easy to understand and the authors chose state-of-the-art models to show the performance of their algorithm. The relevance of these results goes beyond image classification.\n\n\nPros:\n\n- Simple and effective method to improve convergence\n- Good evaluation on well known database\n\n\nCons:\n\n- Connection of introduction and topic of the paper is a bit unclear\n- Fig 2, 4 and 5 are hard to read. Lines are out of bounds and maybe only the best setting for T_0 and T_mult would be clearer. The baseline also doesn't seem to converge\n\nRemarks:\nAn loss surface for T_0 against T_mult would be very helpful. Also understanding the relationship of network depth and the performance of this method would add value to this analysis.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | SGDR: Stochastic Gradient Descent with Warm Restarts | ["Ilya Loshchilov", "Frank Hutter"] | Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets,
where we demonstrate new state-of-the-art results at 3.14\% and 16.21\%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at \\ \url{https://github.com/loshchil/SGDR} | ["Deep learning", "Optimization"] | https://openreview.net/forum?id=Skq89Scxx | https://openreview.net/pdf?id=Skq89Scxx | https://openreview.net/forum?id=Skq89Scxx¬eId=SkMtuVxEx |
By3oLYeNe | Skq89Scxx | ICLR.cc/2017/conference/-/paper263/official/review | {"title": "", "rating": "7: Good paper, accept", "review": "This an interesting investigation into learning rate schedules, bringing in the idea of restarts, often overlooked in deep learning. The paper does a thorough study on non-trivial datasets, and while the outcomes are not fully conclusive, the results are very good and the approach is novel enough to warrant publication. \n\nI thank the authors for revising the paper based on my concerns.\n\nTypos:\n- \u201cflesh\u201d -> \u201cflush\u201d", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | SGDR: Stochastic Gradient Descent with Warm Restarts | ["Ilya Loshchilov", "Frank Hutter"] | Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets,
where we demonstrate new state-of-the-art results at 3.14\% and 16.21\%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at \\ \url{https://github.com/loshchil/SGDR} | ["Deep learning", "Optimization"] | https://openreview.net/forum?id=Skq89Scxx | https://openreview.net/pdf?id=Skq89Scxx | https://openreview.net/forum?id=Skq89Scxx¬eId=By3oLYeNe |
rJXJeHYEg | Sk2iistgg | ICLR.cc/2017/conference/-/paper128/official/review | {"title": "This paper presents an approach to non-linear kernel dimensionality reduction with a trace norm regularizer in the feature space. The authors proposed an iterative minimization approach in order to obtain a local optimum of a relaxed problem. The paper contains errors and the experimental evaluation is not convincing. Only old techniques are compared against in very toy datasets. ", "rating": "4: Ok but not good enough - rejection", "review": "This paper presents an approach to non-linear kernel dimensionality reduction with a trace norm regularizer in the feature space. The authors proposed an iterative minimization approach in order to obtain a local optimum of a relaxed problem. \nThe paper contains errors and the experimental evaluation is not convincing. Only old techniques are compared against in very toy datasets. \n\nThe authors claim state-of-the-art, however, the oil dataset is not a real benchmark, and the comparisons are to very old approaches. \nThe experimental evaluation should demonstrate robustness to more complex noise and outliers, as this was one of the motivations in the introduction.\n\nThe authors do not address the out-of-sample problem. This is a problem of kernel-based methods vs LVMs, and thus should be address here.\n\n\nThe paper contains errors:\n\n- The last paragraph of section 1 says that this paper proposes a closed form solution to robust KPCA. This is simply wrong, as the proposed approach consists of iteratively solving iterativey a set of closed form updates and Levenberg-Marquard optimizationd. This is not any more closed form!\n\n- In the same paragraph (and later in the text) the authors claim that the proposed approach can be trivially generalized to incorporate other cost functions. This is not true, as in general there will be no more inner loop closed form updates and the authors will need to solve a much more complex optimization problem. \n\n- The third paragraph of section 2 claims that this paper presents a novel energy minimization framework to solve problems of the general form of eq. (2). However, this is not what the authors solve at the end. They solve a different problem that has been subject to at least two relaxations. It is not clear how solving for a local optima of this double relaxed problem is related to the original problem they want to solve. \n\n- The paper says that Geiger et al defined non linearities on a latent space of pre-defined dimensionality. This is wrong. This paper discovers the dimensionality of the latent space by means of a regularizer that encourages the singular values to be sparse. Thus, it does not have a fixed dimensionality, the latent space is just bounded to be smaller or equal than the dimensionality of the original space. \n\n\nIt is not clear to me why the author say for LVMs such as GPLVM that \"the latent space is learned a priority with clean training data\". One can use different noise models within the GP framework. Furthermore, the proposed approach assumes Gaussian noise (see eq. 6), which is also the trivial case for GP-based LVMs. \n\n\nIt is not clear what the authors mean in the paper by \"pre-training\" or saying that techniques do not have a training phase. KPCA is trained via a closed-form update, but there is still training. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Non-linear Dimensionality Regularizer for Solving Inverse Problems | ["Ravi Garg", "Anders Eriksson", "Ian Reid"] | Consider an ill-posed inverse problem of estimating causal factors from observations, one of which is known to lie near some (unknown) low-dimensional, non-linear manifold expressed by a predefined Mercer-kernel. Solving this problem requires simultaneous estimation of these factors and learning the low-dimensional representation for them. In this work, we introduce a novel non-linear dimensionality regularization technique for solving such problems without pre-training.
We re-formulate Kernel-PCA as an energy minimization problem in which low dimensionality constraints are introduced as regularization terms in the energy.
To the best of our knowledge, ours is the first attempt to create a dimensionality regularizer in the KPCA framework. Our approach relies on robustly penalizing the rank of the recovered factors directly in the implicit feature space to create
their low-dimensional approximations in closed form. Our approach performs robust KPCA in the presence of missing data and noise.
We demonstrate state-of-the-art results on predicting missing entries in the standard oil flow dataset. Additionally, we evaluate our method on the challenging problem of Non-Rigid Structure from Motion and our approach delivers promising results on CMU mocap dataset despite the presence of significant occlusions and noise. | ["Computer vision", "Optimization", "Structured prediction"] | https://openreview.net/forum?id=Sk2iistgg | https://openreview.net/pdf?id=Sk2iistgg | https://openreview.net/forum?id=Sk2iistgg¬eId=rJXJeHYEg |
ryNp9bzre | Sk2iistgg | ICLR.cc/2017/conference/-/paper128/official/review | {"title": "Not clear", "rating": "3: Clear rejection", "review": "This paper considers an alternate formulation of Kernel PCA with rank constraints incorporated as a regularization term in the objective. The writing is not clear. The focus keeps shifting from estimating \u201ccausal factors\u201d, to nonlinear dimensionality reduction to Kernel PCA to ill-posed inverse problems. The problem reformulation of Kernel PCA uses somewhat standard tricks and it is not clear what are the advantages of the proposed approach over the existing methods as there is no theoretical analysis of the overall approach or empirical comparison with existing state-of-the-art. \n\n- Not sure what the authors mean by \u201ccausal factors\u201d. There is a reference to it in Abstract and in Problem formulation on page 3 without any definition/discussion.\n\n- In KPCA, I am not sure why one is interested in step (iii) outlined on page 2 of finding a pre-image for each\n\n- Authors outline two key disadvantages of the existing KPCA approach. The first one, that of low-dimensional manifold assumption not holding exactly, has received lots of attention in the machine learning literature. It is common to assume that the data lies near a low-dimensional manifold rather than on a low-dimensional manifold. Second disadvantage is somewhat unclear as finding \u201ca data point (pre-image) corresponding to each projection in the input space\u201d is not a standard step in KPCA. \n\n- On page 3, you never define $\\mathcal{X} \\times N$, $\\mathcal{Y} \\times N$, $\\mathcal{H} \\times N$. Clearly, they cannot be cartesian products. I have to assume that notation somehow implies N-tuples. \n\n- On page 3, Section 2, $\\mathcal{X}$ and $\\mathcal{Y}$ are sets. What do you mean by $\\mathcal{Y} \\ll \\mathcal{X}$\n\n- On page 5, $\\mathcal{S}^n$ is never defined. \n\n- Experiments: None of the standard algorithms for matrix completion such as OptSpace or SVT were considered \n\n- Experiments: There is no comparison with alternate existing approaches for Non-rigid structure from motion. \n\n- Proof of the main result Theorem 3.1: To get from (16) to (17) using the Holder inequality (as stated) one would end up with a term that involves sum of fourth powers of weights w_{ij}. Why would they equal to one using the orthonormal constraints? It would be useful to give more details here, as I don\u2019t see how the argument goes through at this point. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Non-linear Dimensionality Regularizer for Solving Inverse Problems | ["Ravi Garg", "Anders Eriksson", "Ian Reid"] | Consider an ill-posed inverse problem of estimating causal factors from observations, one of which is known to lie near some (unknown) low-dimensional, non-linear manifold expressed by a predefined Mercer-kernel. Solving this problem requires simultaneous estimation of these factors and learning the low-dimensional representation for them. In this work, we introduce a novel non-linear dimensionality regularization technique for solving such problems without pre-training.
We re-formulate Kernel-PCA as an energy minimization problem in which low dimensionality constraints are introduced as regularization terms in the energy.
To the best of our knowledge, ours is the first attempt to create a dimensionality regularizer in the KPCA framework. Our approach relies on robustly penalizing the rank of the recovered factors directly in the implicit feature space to create
their low-dimensional approximations in closed form. Our approach performs robust KPCA in the presence of missing data and noise.
We demonstrate state-of-the-art results on predicting missing entries in the standard oil flow dataset. Additionally, we evaluate our method on the challenging problem of Non-Rigid Structure from Motion and our approach delivers promising results on CMU mocap dataset despite the presence of significant occlusions and noise. | ["Computer vision", "Optimization", "Structured prediction"] | https://openreview.net/forum?id=Sk2iistgg | https://openreview.net/pdf?id=Sk2iistgg | https://openreview.net/forum?id=Sk2iistgg¬eId=ryNp9bzre |
B1yevitBg | Sk2iistgg | ICLR.cc/2017/conference/-/paper128/official/review | {"title": "Lacking in several aspects; limited novelty", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposes a nonlinear regularizer for solving ill-posed inverse problems. The latent variables (or causal factors) corresponding to the observed data are assumed to lie near a low dimensional subspace in an RKHS induced by a predetermined kernel. The proposed regularizer can be seen as an extension of the linear low-rank assumption on the latent factors. A nuclear norm penalty on the Cholesky factor of the kernel matrix is used as a relaxation for the dimensionality of the subspace. Empirical results are reported on two tasks involving linear inverse problems -- missing feature imputation, and estimating non-rigid 3D structures from a sequence of 2D orthographic projections -- and the proposed method is shown to outperform linear low-rank regularizer. \n\nThe clarity of the paper has scope for improvement (particularly, Introduction) - the back and forth b/w dimensionality reduction techniques and inverse problems is confusing at times. Clearly defining the ill-posed inverse problem first and then motivating the need for a regularizer (which brings dimensionality reduction techniques into the picture) may be a more clear flow in my opinion. \n\nThe motivation behind relaxation of rank() in Eq 1 to nuclear-norm in Eq 2 is not clear to me in this setting. The relaxation does not yield a convex problem over S,C (Eq 5) and also increases the computations (Algo 2 needs to do full SVD of K(S) every time). The authors should discuss pros/cons over the alternate approach that fixes the rank of C (which can be selected using cross-validation, in the same way as $\\tau$ is selected), leaving just the first two terms in Eq 5. For this simpler objective, an interesting question to ask would be -- are there kernel functions for which it can solved in a scalable manner? \n\nThe proposed alternating optimization approach in the current form is computationally intensive and seems hard to scale to even moderate sized data -- in every iteration one needs to compute the kernel matrix over S and perform full SVD over the kernel matrix (Algo 2). Empirical evaluations are also not extensive -- (i) the dataset used for feature imputation is old and non-standard, (ii) for structure estimation from motion on CMU dataset, the paper only compares with linear low-rank regularization, (iii) there is no comment/study on the convergence of the alternating procedure (Algo 1). \n\n\n\n\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Non-linear Dimensionality Regularizer for Solving Inverse Problems | ["Ravi Garg", "Anders Eriksson", "Ian Reid"] | Consider an ill-posed inverse problem of estimating causal factors from observations, one of which is known to lie near some (unknown) low-dimensional, non-linear manifold expressed by a predefined Mercer-kernel. Solving this problem requires simultaneous estimation of these factors and learning the low-dimensional representation for them. In this work, we introduce a novel non-linear dimensionality regularization technique for solving such problems without pre-training.
We re-formulate Kernel-PCA as an energy minimization problem in which low dimensionality constraints are introduced as regularization terms in the energy.
To the best of our knowledge, ours is the first attempt to create a dimensionality regularizer in the KPCA framework. Our approach relies on robustly penalizing the rank of the recovered factors directly in the implicit feature space to create
their low-dimensional approximations in closed form. Our approach performs robust KPCA in the presence of missing data and noise.
We demonstrate state-of-the-art results on predicting missing entries in the standard oil flow dataset. Additionally, we evaluate our method on the challenging problem of Non-Rigid Structure from Motion and our approach delivers promising results on CMU mocap dataset despite the presence of significant occlusions and noise. | ["Computer vision", "Optimization", "Structured prediction"] | https://openreview.net/forum?id=Sk2iistgg | https://openreview.net/pdf?id=Sk2iistgg | https://openreview.net/forum?id=Sk2iistgg¬eId=B1yevitBg |
BkLVK-NEx | HyET6tYex | ICLR.cc/2017/conference/-/paper114/official/review | {"title": "Hard to see universality yet.", "rating": "5: Marginally below acceptance threshold", "review": "The idea of universality that is independent of input distribution and dimension, depending only on the algorithm is an appealing one. However, as an empirical study, this paper comes up somewhat short:\n\n1. Exactly one algorithm is shown for the deep learning example. It would have been more convincing to compare distributions with one or more algorithms. \n\n2. The definition (1), and much of the work of Section 2.1 seems to have already been covered in Deift (2014), Section 1.3. In that paper, a number of different algorithms for the solution of linear systems are considered, and then the concept of universality becomes more plausible. I do not see enough of such algorithmic comparisons in this paper (same problem setup, different algorithms).\n\n3. It seems to me that what practitioners might care about in practice are both the mean and variance in running times; these quantities are buried in (1). So I question how useful the distribution itself might be for algorithm tuning. \n\nAt the least, many more empirical comparisons should be provided to convince me that the universality holds across a broad range of algorithms.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Universality in halting time | ["Levent Sagun", "Thomas Trogdon", "Yann LeCun"] | The authors present empirical distributions for the halting time (measured by the number of iterations to reach a given accuracy) of optimization algorithms applied to two random systems: spin glasses and deep learning. Given an algorithm, which we take to be both the optimization routine and the form of the random landscape, the fluctuations of the halting time follow a distribution that remains unchanged even when the input is changed drastically. We observe two main classes, a Gumbel-like distribution that appears in Google searches, human decision times, QR factorization and spin glasses, and a Gaussian-like distribution that appears in conjugate gradient method, deep network with MNIST input data and deep network with random input data. This empirical evidence suggests presence of a class of distributions for which the halting time is independent of the underlying distribution under some conditions. | ["Optimization"] | https://openreview.net/forum?id=HyET6tYex | https://openreview.net/pdf?id=HyET6tYex | https://openreview.net/forum?id=HyET6tYex¬eId=BkLVK-NEx |
H1Wt5Cd4x | HyET6tYex | ICLR.cc/2017/conference/-/paper114/official/review | {"title": "Unusual research. Unsound methodology. conclusions are not supported by the content.", "rating": "2: Strong rejection", "review": "Summary\n\n\nFor several algorithms, previous research has shown that the halting time follows a two-parameter distribution (the so-called universal property investigated by the authors). In this work, the authors extend the investigation to new algorithms (spin-glass, gradient descent in deep learning).\n\nAn algorithm is considered to satisfy the universality property when the centered/scaled halting time fluctuations (empirical distribution of halting times) depend on the algorithm but do not depend on the target accuracy epsilon, an intrinsic measure of dimension N, the probability distribution/random ensemble. (This is clear from Eq 1 where on the left the empirical halting time distribution depends on epsilon, N, A, E and on the right, the approximation only depends on the algorithm)\n\nThe authors argue that empirically, the universal property is observed when both algorithms (spin glass and deep learning) perform well and that it is not observed when they do not perform well.\n\nA moment-based indicator is introduced to assess whether universality is observed.\n\n\nReview\n\n\nThis paper presents several problems.\n\n\n\npage 2: \u201c[\u2026] for sufficiently large N and eps = eps(N)\u201d\n\nThe dependence of epsilon on N is troubling.\n\n\n\npage 3: \u201cUniversality is a measure of stability in an algorithm [\u2026] For example [\u2026] halting time for the power method [\u2026] has infinite expectation and hence this type of universality is *not* present. One could use this to conclude that the power method is naive. Therefore the presence of universality is a desirable feature of a numerical method\u201d\n\nNo. An algorithm is naive if there are better ways to answer the problem. One could not conclude from a halting time with infinite expectation (e.g. solving a problem extremely quickly 99% of the time, and looping forever in 1% of cases) or infinite variance, that the algorithm is naive.\u2028\nMoreover the universal property is more restrictive than having a finite halting time expectation. Even if in many specific cases, having a finite halting time expectation is a desirable property, showing that the presence of universality is desirable would require a demonstration that the other more restrictive aspects are also desirable.\u2028\nAlso, the paragraph only concerns one algorithm. why would the conclusions generalise to all numerical methods ?\u2028\nEven if the universality property is arguably desirable (i.e. event if the conclusion of this paragraph is assumed correct), the paragraph does not support the given conclusion.\n\n\n\nComparing Eq 1 and figures 2,3,4,5\u2028\nFrom Eq 1, universality means that the centered/scaled halting time fluctuations (which depend on A, epsilon, N, E) can be approximated by a distribution that only depends on A (not on epsilon, N, E) but in the experiments only E varies (figures 2,3,4,5). The validity of the approximation with varying epsilon or N is never tested\n\n\n\nThe ensembles/distributions parameter E (on which halting fluctuations should not depend) and the algorithm A (on which halting fluctuations are allowed to depend) are not well defined, especially w.r.t. the common use of the words. In the optimisation setting we are told that the functional form of the landscape function is part of A (in answer to the question of a reviewer) but what is part of the functional form ? what about computations where the landscape has no known functional form (black box) ?\n\n\n\nThe conclusion claims that the paper \u201cattempts to exhibit cases\u201d where one can answer 5 questions in a robust and quantitative way.\n\nQuestion 1: \u201cWhat are the conditions on the ensembles and the model that lead to such universality ?\u201d\u2028The only quantitative way would be to use the moments based indicator however there is only one example of universality not being observed which concerns only one algorithm (conjugate gradient) and one type of failure (when M = N). This does not demonstrate robustness of the method.\n\nQuestion 2: \u201cWhat constitutes a good set of hyper parameters for a given algorithm ?\u201d\nThe proposed way to choose would be to test whether universality is observed. If it is then the hyper parameters are good, if not the hyper parameters are bad. The correspondance between bad hyper-parameters and observing no universality concerns only one algorithm and one type of failure. Other algorithms may fail in the universal regime or perform well in the non universal regime. The paper does not show how to answer this question in a robust way.\n\nQuestion 3: \"How can we go beyond inspection when tuning a system ?\u2028\"\nThe question is too vague and general and there is probably no robust and quantitative way to answer it at all.\n\nQuestion 4: \"How can we infer if an algorithm is a good match to the system at hand ?\u2028\"\nThe paper fails to demonstrate convincingly that universality is either a good or robust way to approach the very few studied algorithms. The suggested generalisation to all systems and algorithms is extremely far fetched.\n\nQuestion 5: \"What is the connection between the universal regime and the structure of the landscape ?\"\n\u2028Same as before, the question is extremely vague and cannot be answered in a robust or quantitative way at all. The fact that what corresponds to A and what corresponds to E is not clear does not help.\n\n\nIn the conclusion it is written that the paper validates the claim that universality is present in all or nearly all sensible computation. It does not. The paper does not properly test whether universality is present (only 1 parameter in 3 that should not vary is tested). The paper does not properly test whether universality is lost when the computation is no longer sensible (only one failure case tested). Finally the experiments do not apply to all or nearly all computations but only to very few specific algorithms.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Universality in halting time | ["Levent Sagun", "Thomas Trogdon", "Yann LeCun"] | The authors present empirical distributions for the halting time (measured by the number of iterations to reach a given accuracy) of optimization algorithms applied to two random systems: spin glasses and deep learning. Given an algorithm, which we take to be both the optimization routine and the form of the random landscape, the fluctuations of the halting time follow a distribution that remains unchanged even when the input is changed drastically. We observe two main classes, a Gumbel-like distribution that appears in Google searches, human decision times, QR factorization and spin glasses, and a Gaussian-like distribution that appears in conjugate gradient method, deep network with MNIST input data and deep network with random input data. This empirical evidence suggests presence of a class of distributions for which the halting time is independent of the underlying distribution under some conditions. | ["Optimization"] | https://openreview.net/forum?id=HyET6tYex | https://openreview.net/pdf?id=HyET6tYex | https://openreview.net/forum?id=HyET6tYex¬eId=H1Wt5Cd4x |
r1uorDYBe | HyET6tYex | ICLR.cc/2017/conference/-/paper114/official/review | {"title": "interesting idea, shortcomings in evaluation and clarity", "rating": "5: Marginally below acceptance threshold", "review": "The authors explore whether the halting time distributions for various algorithms in various settings exhibit \"universality\", i.e. after rescaling to zero mean and unit variance, the distribution does not depend on stopping parameter, dimensionality and ensemble.\n\nThe idea of the described universality is very interesting. However I see several shortcomings in the paper:\n\nIn order to be of practical relevance, the actual stopping time might be more relevant than the scaled one. The discussion of exponential tailed halting time distributions is a good start, but I am not sure how often this might be actually helpful. Still, the findings in the paper might be interesting from a theoretical point of view.\n\nEspecially for ICLR, I think it would have been more interesting to look into comparisons between stochastic gradient descent, momentum, ADAM etc on different deep learning architectures. Over which of those parameters does universality hold?. How can different initializations influence the halting time distribution? I would expect a sensible initialization to cut of part of the right tail of the distribution.\n\nAdditionally, I found the paper quite hard to read. Here are some clarity issues:\n\n- abstract: \"even when the input is changed drastically\": From the abstract I'm not sure what \"input\" refers to, here\n- I. Introduction: \"where the stopping condition is, essentially, the time to find the minimum\": this doesn't seem to make sense, a condition is not a time. I guess the authors wanted to say that the stopping condition is that the minimum has been reached?\n- I.1 the notions of dimension N, epsilon and ensemble E are introduced without any clarification what they are. From the later parts of the paper I got some ideas and examples, but here it is very hard to understand what these parameters should be (just some examples would be already helpful)\n- I.3 \"We use x^\\ell for \\ell \\in Z=\\{1, \\dots, S\\} where Z is a random sample from of training samples\" This formulation doesn't make sense. Either Z is a random sample, or Z={1, ..., S}.\n- II.1 it took me a long time to find the meaning of M. As this parameter seems to be crucial for universality in this case, it would be very helpful to point out more explicitly what it refers to.\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Universality in halting time | ["Levent Sagun", "Thomas Trogdon", "Yann LeCun"] | The authors present empirical distributions for the halting time (measured by the number of iterations to reach a given accuracy) of optimization algorithms applied to two random systems: spin glasses and deep learning. Given an algorithm, which we take to be both the optimization routine and the form of the random landscape, the fluctuations of the halting time follow a distribution that remains unchanged even when the input is changed drastically. We observe two main classes, a Gumbel-like distribution that appears in Google searches, human decision times, QR factorization and spin glasses, and a Gaussian-like distribution that appears in conjugate gradient method, deep network with MNIST input data and deep network with random input data. This empirical evidence suggests presence of a class of distributions for which the halting time is independent of the underlying distribution under some conditions. | ["Optimization"] | https://openreview.net/forum?id=HyET6tYex | https://openreview.net/pdf?id=HyET6tYex | https://openreview.net/forum?id=HyET6tYex¬eId=r1uorDYBe |
BJwnboxEl | HkvS3Mqxe | ICLR.cc/2017/conference/-/paper183/official/review | {"title": "Useful topic, but not very solid techinique", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes two pruning methods to reduce the computation of deep neural network. In particular, whole feature maps and the kernel connections can be removed with not much decrease of classification accuracy. \nHowever, this paper also has the following problems. \n1)\tThe method is somehow trivial, since the pruning masks are mainly chosen by simple random sampling. The novelty and scalability are both limited. \n2)\tExperiment results are mainly focused on the classification rate and the ideal complexity. As a paper on improving computation efficiency, it should include results on practical time consumption. It is very common that reducing numbers of operations may not lead to reduced computational time on a highly parallel platform (e.g., GPU). \n3)\tIt is more important to improve the computational efficiency on large-scale models (e.g., ImageNet classification network) than on small models (e.g., MNIST, CIFAR network). However, results on large-scale network is missing.\n4)\t(*Logical validity of the proposed method*) For feature map pruning, what if just to train reduced-size network is trained from scratch without transfer any knowledge from the pretrained large network? Is it possible to get the same accuracy? If so, it will simply indicate the hyper-parameter is not optimal for the original network. Experimental results are necessary to clarify the necessity of feature map pruning. \nNote that I agree with that a smaller network may be more generalizable than a larger network. \n\n----------------------------------------------\n\nComments to the authors's response:\n\nThanks for replying to my comments. \n\n1) I still believe that the proposed methods are trivial.\n2) It is nice to show GPU implementation. Compared to existing toolboxes (e.g., Torch, Caffe, TensorFlow), is the implementation of convolution efficient enough?\n3) Experiments on Cifar-100 are helpful (better than cifar-10), but it is not really large-scale, where speed-up is not so critical. ImageNet and Places datasets are examples of large-scale datasets.\n4) The author did not reply to the question wrt the validity of the proposed methods. This question is critical. \n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Coarse Pruning of Convolutional Neural Networks with Random Masks | ["Sajid Anwar", "Wonyong Sung"] | The learning capability of a neural network improves with increasing depth at
higher computational costs. Wider layers with dense kernel connectivity patterns
further increase this cost and may hinder real-time inference. We propose feature
map and kernel level pruning for reducing the computational complexity of
a deep convolutional neural network. Pruning feature maps reduces the width
of a layer and hence does not need any sparse representation. Further, kernel
pruning changes the dense connectivity pattern into a sparse one. Due to coarse
nature, these pruning granularities can be exploited by GPUs and VLSI based
implementations. We propose a simple strategy to choose the least adversarial
pruning masks. The proposed approach is generic and can select good pruning
masks for feature map, kernel and intra-kernel pruning. The pruning masks are
generated randomly, and the best performing one is selected using the evaluation
set. The sufficient number of random pruning masks to try depends on the pruning
ratio, and is around 100 when 40% complexity reduction is needed. The pruned
network is retrained to compensate for the loss in accuracy. We have extensively
evaluated the proposed approach with the CIFAR-10, SVHN and MNIST datasets.
Experiments with the CIFAR-10 dataset show that more than 85% sparsity can be
induced in the convolution layers with less than 1% increase in the misclassification
rate of the baseline network. | ["convolutional neural networks", "increase", "feature map", "coarse pruning", "random masks", "capability", "neural network", "depth", "higher computational costs"] | https://openreview.net/forum?id=HkvS3Mqxe | https://openreview.net/pdf?id=HkvS3Mqxe | https://openreview.net/forum?id=HkvS3Mqxe¬eId=BJwnboxEl |
ryuDdOkEx | HkvS3Mqxe | ICLR.cc/2017/conference/-/paper183/official/review | {"title": "N random trails to get best pruning?", "rating": "6: Marginally above acceptance threshold", "review": "Summary: There are many different pruning techniques to reduce memory footprint of CNN models, and those techniques have different granularities (layer, maps, kernel or intra kernel), pruning ratio and sparsity of representation. The work proposes a method to choose the best pruning masks out to many trials. Tested on CIFAR-10, SVHN and MNIST.\n\nPros:\nProposes a method to choose pruning mask out of N trials. \nAnalysis on different pruning methods.\n\nCons & Questions:\n\u201cThe proposed strategy selects the best pruned network through N random pruning trials. This approach enables one to select pruning mask in one shot and is simpler than the multi-step technique.\u201d How can one get the best pruning mask in one shot if you ran N random pruning trials? (answered)\nMissing tests of the approach with bigger CNN: like AlexNet, VGG, GoogLeNet or ResNet. (extended to VGG ok)\nSince reducing model size for embedded systems is the final goal, then showing how much memory space in MB is saved with the proposed technique compared with other approaches like Han et al. (2015) would be good.\n\nMisc:\nTypo in figure 6 a) caption: \u201cFeatuer\u201d (corrected)\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Coarse Pruning of Convolutional Neural Networks with Random Masks | ["Sajid Anwar", "Wonyong Sung"] | The learning capability of a neural network improves with increasing depth at
higher computational costs. Wider layers with dense kernel connectivity patterns
further increase this cost and may hinder real-time inference. We propose feature
map and kernel level pruning for reducing the computational complexity of
a deep convolutional neural network. Pruning feature maps reduces the width
of a layer and hence does not need any sparse representation. Further, kernel
pruning changes the dense connectivity pattern into a sparse one. Due to coarse
nature, these pruning granularities can be exploited by GPUs and VLSI based
implementations. We propose a simple strategy to choose the least adversarial
pruning masks. The proposed approach is generic and can select good pruning
masks for feature map, kernel and intra-kernel pruning. The pruning masks are
generated randomly, and the best performing one is selected using the evaluation
set. The sufficient number of random pruning masks to try depends on the pruning
ratio, and is around 100 when 40% complexity reduction is needed. The pruned
network is retrained to compensate for the loss in accuracy. We have extensively
evaluated the proposed approach with the CIFAR-10, SVHN and MNIST datasets.
Experiments with the CIFAR-10 dataset show that more than 85% sparsity can be
induced in the convolution layers with less than 1% increase in the misclassification
rate of the baseline network. | ["convolutional neural networks", "increase", "feature map", "coarse pruning", "random masks", "capability", "neural network", "depth", "higher computational costs"] | https://openreview.net/forum?id=HkvS3Mqxe | https://openreview.net/pdf?id=HkvS3Mqxe | https://openreview.net/forum?id=HkvS3Mqxe¬eId=ryuDdOkEx |
rJxBszzNg | HkvS3Mqxe | ICLR.cc/2017/conference/-/paper183/official/review | {"title": "Paper review", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes a simple randomized algorithm for selecting which weights in a ConvNet to prune in order to reduce theoretical FLOPs when evaluating a deep neural network. The paper provides a nice taxonomy or pruning granularity from coarse (layer-wise) to fine (intra-kernel). The pruning strategy is empirically driven and uses a validation set to select the best model from N randomly pruned models. Makes claims in the intro about this being \"one shot\" and \"near optimal\" that cannot be supported: it is \"N-shot\" in the sense that N networks are generated and tested and there is no evidence or theory that the found solution is \"near optimal.\"\n\nPros:\n- Nice taxonomy of pruning levels\n- Comparison to the recent weight-sum pruning method\n\nCons:\n- Experimental evaluation does not touch upon recent models (ResNets) and large scale datasets (ImageNet)\n- Paper is somewhat hard to follow\n- Feature map pruning can obviously accelerate computation without specialized sparse implementations of convolution, but this is not the case for finer grained sparsity; since this paper considers fine-grained sparsity it should provide some evidence that introducing that sparsity can yield performance improvements\n\nAnother experimental downside is that the paper does not evaluate the impact of filter pruning on transfer learning. For example, there is not much direct interest in the tasks of MNIST, CIFAR10, or even ImageNet. Instead, a main interest in both academia and industry is the value of the learned representation for transferring to other tasks. One might expect pruning to harm transfer learning. It's possible that the while the main task has about the same performance, transfer learning is strongly hurt. This paper has missed an opportunity to explore that direction.\n\nIn summary, the proposed method is simple, which is good, but the experimental evaluation is somewhat incomplete and does not cover recent models and larger scale datasets.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Coarse Pruning of Convolutional Neural Networks with Random Masks | ["Sajid Anwar", "Wonyong Sung"] | The learning capability of a neural network improves with increasing depth at
higher computational costs. Wider layers with dense kernel connectivity patterns
further increase this cost and may hinder real-time inference. We propose feature
map and kernel level pruning for reducing the computational complexity of
a deep convolutional neural network. Pruning feature maps reduces the width
of a layer and hence does not need any sparse representation. Further, kernel
pruning changes the dense connectivity pattern into a sparse one. Due to coarse
nature, these pruning granularities can be exploited by GPUs and VLSI based
implementations. We propose a simple strategy to choose the least adversarial
pruning masks. The proposed approach is generic and can select good pruning
masks for feature map, kernel and intra-kernel pruning. The pruning masks are
generated randomly, and the best performing one is selected using the evaluation
set. The sufficient number of random pruning masks to try depends on the pruning
ratio, and is around 100 when 40% complexity reduction is needed. The pruned
network is retrained to compensate for the loss in accuracy. We have extensively
evaluated the proposed approach with the CIFAR-10, SVHN and MNIST datasets.
Experiments with the CIFAR-10 dataset show that more than 85% sparsity can be
induced in the convolution layers with less than 1% increase in the misclassification
rate of the baseline network. | ["convolutional neural networks", "increase", "feature map", "coarse pruning", "random masks", "capability", "neural network", "depth", "higher computational costs"] | https://openreview.net/forum?id=HkvS3Mqxe | https://openreview.net/pdf?id=HkvS3Mqxe | https://openreview.net/forum?id=HkvS3Mqxe¬eId=rJxBszzNg |
HyqxOUfVl | rJe-Pr9le | ICLR.cc/2017/conference/-/paper253/official/review | {"title": "Interesting ideas but investigated too superficially", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes a model-based reinforcement learning approach focusing on predicting future rewards given a current state and future actions. This is achieved with a \"residual recurrent neural network\", that outputs the expected reward increase at various time steps in the future. To demonstrate the usefulness of this approach, experiments are conducted on Atari games, with a simple playing strategy that consists in evaluating random sequences of moves and picking the one with highest expected reward (and low enough chance of dying). Interestingly, out of the 3 games tested, one of them exhibits better performance when the agent is trained in a multitask setting (i.e. learning all games simultaneously), hinting that transfer learning is occurring.\n\nThis submission is easy enough to read, and the reward prediction architecture looks like an original and sound idea. There are however several points that I believe prevent this work from reaching the ICLR bar, as detailed below.\n\nThe first issue is the discrepancy between the algorithm proposed in Section 3 vs its actual implementation in Section 4 (experiments): in Section 3 the output is supposed to be the expected accumulated reward in future time steps (as a single scalar), while in experiments it is instead two numbers, one which is the probability of dying and another one which is the probability of having a higher score without dying. This might work better, but it also means the idea as presented in the main body of the paper is not actually evaluated (and I guess it would not work well, as otherwise why implement it differently?)\n\nIn addition, the experimental results are quite limited: only on 3 games that were hand-picked to be easy enough, and no comparison to other RL techniques (DQN & friends). I realize that the main focus of the paper is not about exhibiting state-of-the-art results, since the policy being used is only a simple heuristic to show that the model predictions can ne used to drive decisions. That being said, I think experiments should have tried to demonstrate how to use this model to obtain better reinforcement learning algorithms: there is actually no reinforcement learning done here, since the model is a supervised algorithm, used in a manually-defined hardcoded policy. Another question that could have been addressed (but was not) in the experiments is how good these predictions are (e.g. classification error on dying probability, MSE on future rewards, ...), compared to simpler baselines.\n\nFinally, the paper's \"previous work\" section is too limited, focusing only on DQN and in particular saying very little on the topic of model-based RL. I think a paper like for instance \"Action-Conditional Video Prediction using Deep Networks in Atari Games\" should have been an obvious \"must cite\".\n\nMinor comments:\n- Notations are unusual, with \"a\" denoting a state rather than an action, this is potentially confusing and I see no reason to stray away from standard RL notations\n- Using a dot for tensor concatenation is not a great choice either, since the dot usually indicates a dot product\n- The r_i in 3.2.2 is a residual that has nothing to do with r_i the reward\n- c_i is defined as \"The control that was performed at time i\", but instead it seems to be the control performed at time i-1\n- There is a recurrent confusion between mean and median in 3.2.2\n- x should not be used in Observation 1 since the x from Fig. 3 does not go through layer normalization\n- The inequality in Observation 1 should be about |x_i|, not x_i\n- Observation 1 (with its proof) takes too much space for such a simple result\n- In 3.2.3 the first r_j should be r_i\n- The probability of dying comes out of nowhere in 3.3, since we do not know yet it will be an output of the model\n- \"Our approach is not able to learn from good strategies\" => did you mean \"*only* from good strategies\"?\n- Please say that in Fig. 4 \"fc\" means \"fully connected\"\n- It would be nice also to say how the architecture of Fig. 4 differs from the classical DQN architecture from Mnih et al (2015)\n- Please clarify r_j2 as per your answer in OpenReview comments\n- Table 3 says \"After one iteration\" but has \"PRL Iteration 2\" in it, which is confusing\n- \"Figure 5 shows that not only there is no degradation in Pong and Demon Attack\"=> to me it seems to be a bit worse, actually\n- \"A model that has learned only from random play is able to play at least 7 times better.\" => not clear where this 7 comes from\n- \"Demon Attack's plot in Figure 5c shows a potential problem we mentioned earlier\" => where was it mentioned?\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Multi-task learning with deep model based reinforcement learning | ["Asier Mujika"] | In recent years, model-free methods that use deep learning have achieved great success in many different reinforcement learning environments. Most successful approaches focus on solving a single task, while multi-task reinforcement learning remains an open problem. In this paper, we present a model based approach to deep reinforcement learning which we use to solve different tasks simultaneously. We show that our approach not only does not degrade but actually benefits from learning multiple tasks. For our model, we also present a new kind of recurrent neural network inspired by residual networks that decouples memory from computation allowing to model complex environments that do not require lots of memory. The code will be released before ICLR 2017. | ["Reinforcement Learning", "Deep learning", "Games", "Transfer Learning"] | https://openreview.net/forum?id=rJe-Pr9le | https://openreview.net/pdf?id=rJe-Pr9le | https://openreview.net/forum?id=rJe-Pr9le¬eId=HyqxOUfVl |
rJIwQv-Ve | rJe-Pr9le | ICLR.cc/2017/conference/-/paper253/official/review | {"title": "Final Review", "rating": "2: Strong rejection", "review": "The term strategy is a bit ambiguous. Could you please explain more in formal terms what is strategy?\nIs r the discounted Return at time t, or the reward at time t?\nCould the author compare the method to TD learning?\nThe paper is vague and using many RL terms with different meanings without clarifying those diversions.\n\"So, the output for a given state-actions pair is always same\". Q function by definition is the value of (state, action). So as long as the policy is deterministic the output would be always same too. How's this different from Q learning?\nThe model description doesn't specify what is the policy, and it's only being mentioned in data generation part.\nWhy is it a model based approach?\nThe learning curves are only for 19 iterations, which does not give any useful information. The final results are clearly nothing comparable to previous works. The model is only being tested on three games.\n\nThe paper is vague and using informal language or sometimes misusing the common RL terms. The experiments are very small scale and even in that scenario performing very bad. It's not clear, why it's a model-based approach. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Multi-task learning with deep model based reinforcement learning | ["Asier Mujika"] | In recent years, model-free methods that use deep learning have achieved great success in many different reinforcement learning environments. Most successful approaches focus on solving a single task, while multi-task reinforcement learning remains an open problem. In this paper, we present a model based approach to deep reinforcement learning which we use to solve different tasks simultaneously. We show that our approach not only does not degrade but actually benefits from learning multiple tasks. For our model, we also present a new kind of recurrent neural network inspired by residual networks that decouples memory from computation allowing to model complex environments that do not require lots of memory. The code will be released before ICLR 2017. | ["Reinforcement Learning", "Deep learning", "Games", "Transfer Learning"] | https://openreview.net/forum?id=rJe-Pr9le | https://openreview.net/pdf?id=rJe-Pr9le | https://openreview.net/forum?id=rJe-Pr9le¬eId=rJIwQv-Ve |
BkZFQVxVx | rJe-Pr9le | ICLR.cc/2017/conference/-/paper253/official/review | {"title": "Encouraging results but the approach is too ad-hoc", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes a new approach to model based reinforcement learning and\nevaluates it on 3 ATARI games. The approach involves training a model that\npredicts a sequence of rewards and probabilities of losing a life given a\ncontext of frames and a sequence of actions. The controller samples random\nsequences of actions and executes the one that balances the probabilities of\nearning a point and losing a life given some thresholds. The proposed system\nlearns to play 3 Atari games both individually and when trained on all 3 in a\nmulti-task setup at super-human level.\n\nThe results presented in the paper are very encouraging but there are many\nad-hoc design choices in the design of the system. The paper also provides\nlittle insight into the importance of the different components of the system.\n\nMain concerns:\n- The way predicted rewards and life loss probabilities are combined is very ad-hoc.\n The natural way to do this would be by learning a Q-value, instead different\n rules are devised for different games.\n- Is a model actually being learned and improved? It would be good to see\n predictions for several actions sequences from some carefully chosen start\n states. This would be good to see both on a game where the approach works and\n on a game where it fails. The learning progress could also be measured by\n plotting the training loss on a fixed holdout set of sequences.\n- How important is the proposed RRNN architecture? Would it still work without\n the residual connections? Would a standard LSTM also work?\n\nMinor points:\n- Intro, paragraph 2 - There is a lot of much earlier work on using models in\n RL. For example, see Dyna and \"Memory approaches to reinforcement learning in\n non-Markovian domains\" by Lin and Mitchell to name just two.\n- Section 3.1 - Minor point, but using a_i to represent the observation is\n unusual. Why not use o_i for observations and a_i for actions?\n- Section 3.2.2 - Notation again, r_i was used earlier to represent the\n reward at time i but it is being used again for something else.\n- Observation 1 seems somewhat out of place. Citing the layer normalization\n paper for the motivation is enough.\n- Section 3.2.2, second last paragraph - How is memory decoupled from\n computation here? Models like neural turning machines accomplish this by using\n an external memory, but this looks like an RNN with skip connections.\n- Section 3.3, second paragraph - Whether the model overfits or not depends on\n the data. The approach doesn't work with demonstrations precisely because it\n would overfit.\n- Figure 4 - The reference for Batch Normalization should be Ioffe and Szegedy\n instead of Morimoto et al.\n\nOverall I think the paper has some really promising ideas and encouraging\nresults but is missing a few exploratory/ablation experiments and some polish.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Multi-task learning with deep model based reinforcement learning | ["Asier Mujika"] | In recent years, model-free methods that use deep learning have achieved great success in many different reinforcement learning environments. Most successful approaches focus on solving a single task, while multi-task reinforcement learning remains an open problem. In this paper, we present a model based approach to deep reinforcement learning which we use to solve different tasks simultaneously. We show that our approach not only does not degrade but actually benefits from learning multiple tasks. For our model, we also present a new kind of recurrent neural network inspired by residual networks that decouples memory from computation allowing to model complex environments that do not require lots of memory. The code will be released before ICLR 2017. | ["Reinforcement Learning", "Deep learning", "Games", "Transfer Learning"] | https://openreview.net/forum?id=rJe-Pr9le | https://openreview.net/pdf?id=rJe-Pr9le | https://openreview.net/forum?id=rJe-Pr9le¬eId=BkZFQVxVx |
BJqeUcbVe | ryrGawqex | ICLR.cc/2017/conference/-/paper440/official/review | {"title": "", "rating": "8: Top 50% of accepted papers, clear accept", "review": "The paper presents a novel strategy to deal with dynamic computation graphs. They arise, when the computation is dynamically influenced by the input data, such as in LSTMs. The authors propose an `unrolling' strategy over the operations done at every step, which allows a new kind of batching of inputs.\n\nThe presented idea is novel and the results clearly indicate the potential of the approach. For the sake of clarity of the presentation I would drop parts of Section 3 (\"A combinator library for neural networks\") which presents technical details that are in general interesting, but do not help the understanding of the core idea of the paper. The presented experimental results on the \"Stanford Sentiment Treebank\" are in my opinion not supporting the claim of the paper, which is towards speed, than a little bit confusing. It is important to point out that even though the presented ensemble \"[...] variant sets a new state-of-the-art on both subtasks\" [p. 8], this is not due to the framework, not even due to the model (comp. lines 4 and 2 of Tab. 2), but probably, and this can only be speculated about, due to the ensemble averaging. I would appreciate a clearer argumentation in this respect.\n\nUpdate on Jan. 17th:\nafter the authors update for their newest revision, I increase my rating to 8 due to the again improved, now very clear argumentation.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Learning with Dynamic Computation Graphs | ["Moshe Looks", "Marcello Herreshoff", "DeLesley Hutchins", "Peter Norvig"] | Neural networks that compute over graph structures are a natural fit for problems in a variety of domains, including natural language (parse trees) and cheminformatics (molecular graphs). However, since the computation graph has a different shape and size for every input, such networks do not directly support batched training or inference. They are also difficult to implement in popular deep learning libraries, which are based on static data-flow graphs. We introduce a technique called dynamic batching, which not only batches together operations between different input graphs of dissimilar shape, but also between different nodes within a single input graph. The technique allows us to create static graphs, using popular libraries, that emulate dynamic computation graphs of arbitrary shape and size. We further present a high-level library of compositional blocks that simplifies the creation of dynamic graph models. Using the library, we demonstrate concise and batch-wise parallel implementations for a variety of models from the literature. | ["Deep learning"] | https://openreview.net/forum?id=ryrGawqex | https://openreview.net/pdf?id=ryrGawqex | https://openreview.net/forum?id=ryrGawqex¬eId=BJqeUcbVe |
SyOL-J8Nx | ryrGawqex | ICLR.cc/2017/conference/-/paper440/official/review | {"title": "Description of a promising software package.", "rating": "7: Good paper, accept", "review": "Authors describe implementation of TensorFlow Fold which allows one to run various computations without modifying computation graph. They achieve this by creating a generic scheduler as a TensorFlow computation graph, which can accept graph description as input and execute it.\n\nThey show clear benefits to this approach for tasks where computation changes for each datapoint, such as the case with TreeRNN.\n\nIn the experiments, they compare against having static batch (same graph structure repeated many times) and batch size 1.\n\nThe reason my score is 7 and not higher is because they do not provide comparison to the main alternative of their method -- someone could create a new TensorFlow graph for each dynamic batch. In other words, instead of using their graph as the scheduling algorithm, one could explicitly create each non-uniform batch as a TensorFlow graph, and run that using standard TensorFlow.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"} | review | 2017 | ICLR.cc/2017/conference | Deep Learning with Dynamic Computation Graphs | ["Moshe Looks", "Marcello Herreshoff", "DeLesley Hutchins", "Peter Norvig"] | Neural networks that compute over graph structures are a natural fit for problems in a variety of domains, including natural language (parse trees) and cheminformatics (molecular graphs). However, since the computation graph has a different shape and size for every input, such networks do not directly support batched training or inference. They are also difficult to implement in popular deep learning libraries, which are based on static data-flow graphs. We introduce a technique called dynamic batching, which not only batches together operations between different input graphs of dissimilar shape, but also between different nodes within a single input graph. The technique allows us to create static graphs, using popular libraries, that emulate dynamic computation graphs of arbitrary shape and size. We further present a high-level library of compositional blocks that simplifies the creation of dynamic graph models. Using the library, we demonstrate concise and batch-wise parallel implementations for a variety of models from the literature. | ["Deep learning"] | https://openreview.net/forum?id=ryrGawqex | https://openreview.net/pdf?id=ryrGawqex | https://openreview.net/forum?id=ryrGawqex¬eId=SyOL-J8Nx |
SkRHZS-Ee | ryrGawqex | ICLR.cc/2017/conference/-/paper440/official/review | {"title": "A new method to optimize computation graphs", "rating": "8: Top 50% of accepted papers, clear accept", "review": "The paper describes a novel technique to improve the efficiency of computation graphs in deep learning frameworks. An impressive speedup can be observed in their implementation within TensorFlow. The content is presented with sufficient clarity, although some more graphical illustrations could be useful. This work is relevant in order to achieve highest performance in neural network training.\n\n\nPros:\n\n- significant speed improvements through dynamic batching\n- source code provided\n\n\nCons:\n\n- the effect on a large real-world (ASR, SMT) would allow the reader to put the improvements better into context\n- presentation/vizualisation can be improved ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | Deep Learning with Dynamic Computation Graphs | ["Moshe Looks", "Marcello Herreshoff", "DeLesley Hutchins", "Peter Norvig"] | Neural networks that compute over graph structures are a natural fit for problems in a variety of domains, including natural language (parse trees) and cheminformatics (molecular graphs). However, since the computation graph has a different shape and size for every input, such networks do not directly support batched training or inference. They are also difficult to implement in popular deep learning libraries, which are based on static data-flow graphs. We introduce a technique called dynamic batching, which not only batches together operations between different input graphs of dissimilar shape, but also between different nodes within a single input graph. The technique allows us to create static graphs, using popular libraries, that emulate dynamic computation graphs of arbitrary shape and size. We further present a high-level library of compositional blocks that simplifies the creation of dynamic graph models. Using the library, we demonstrate concise and batch-wise parallel implementations for a variety of models from the literature. | ["Deep learning"] | https://openreview.net/forum?id=ryrGawqex | https://openreview.net/pdf?id=ryrGawqex | https://openreview.net/forum?id=ryrGawqex¬eId=SkRHZS-Ee |
B1O_F2eVl | SkwSJ99ex | ICLR.cc/2017/conference/-/paper527/official/review | {"title": "", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes to reduce model size and evaluation time of deep CNN models on mobile devices by converting multiple layers into single layer and then retraining the converted model. The paper showed that the computation time can be reduced by 3x to 5x with only 0.4% accuracy loss on a specific model.\n\nReducing model sizes and speeding up model evaluation are important in many applications. I have several concerns:\n\n1. There are many techniques that can reduce model sizes. For example, it has been shown by several groups that using the teacher-student approach, people can achieve the same and sometimes even better accuracy than the teacher (big model) using a much smaller model. However, this paper does not compare any one of them.\n2. The technique proposed in this paper is limited in its applicability since it's designed specifically for the models discussed in the paper. \n3. Replacing several layers with single layer is a relatively standard procedure. For example, the mean variance normalization layer and batch normalization layer can all be absorbed without retraining or losing accuracy.\n\nBTW, the DNN low-rank approximation technique was first proposed in speech recognition. e.g., \n\nXue, J., Li, J. and Gong, Y., 2013, August. Restructuring of deep neural network acoustic models with singular value decomposition. In INTERSPEECH (pp. 2365-2369).\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"} | review | 2017 | ICLR.cc/2017/conference | DeepRebirth: A General Approach for Accelerating Deep Neural Network Execution on Mobile Devices | ["Dawei Li", "Xiaolong Wang", "Deguang Kong", "Mooi Choo Chuah"] | Deploying deep neural networks on mobile devices is a challenging task due to computation complexity and memory intensity. Existing works solve this problem by reducing model size using weight approximation methods based on dimension reduction (i.e., SVD, Tucker decomposition and Quantization). However, the execution speed of these compressed models are still far below the real-time processing requirement of mobile services. To address this limitation, we propose a novel acceleration framework: DeepRebirth by exploring the deep learning model parameter sparsity through merging the parameter-free layers with their neighbor convolution layers to a single dense layer. The design of DeepRebirth is motivated by the key observation: some layers (i.e., normalization and pooling) in deep learning models actually consume a large portion of computational time even few learned parameters are involved, and acceleration of these layers has the potential to improve the processing speed significantly. Essentially, the functionality of several merged layers is replaced by the new dense layer – rebirth layer in DeepRebirth. In order to preserve the same functionality, the rebirth layer model parameters are re-trained to be functionality equivalent to the original several merged layers. The extensive experiments performed on ImageNet using several popular mobile devices demonstrate that DeepRebirth is not only providing huge speed-up in model deployment and significant memory saving but also maintaining the model accuracy, i.e., 3x-5x speed-up and energy saving on GoogLeNet with only 0.4% accuracy drop on top-5 categorization in ImageNet. Further, by combining with other model compression techniques, DeepRebirth offers an average of 65ms model forwarding time on each image using Samsung Galaxy S6 with only 2.4% accuracy drop. In addition, 2.5x run-time memory saving is achieved with rebirth layers. | ["deeprebirth", "layers", "general", "mobile devices", "functionality", "imagenet", "accuracy drop", "mobile devices deeprebirth", "deep neural networks"] | https://openreview.net/forum?id=SkwSJ99ex | https://openreview.net/pdf?id=SkwSJ99ex | https://openreview.net/forum?id=SkwSJ99ex¬eId=B1O_F2eVl |