paper_id stringlengths 19 21 | paper_title stringlengths 8 170 | paper_abstract stringlengths 8 5.01k | paper_acceptance stringclasses 18
values | meta_review stringlengths 29 10k | label stringclasses 3
values | review_ids list | review_writers list | review_contents list | review_ratings list | review_confidences list | review_reply_tos list |
|---|---|---|---|---|---|---|---|---|---|---|---|
iclr_2018_rkA1f3NpZ | Ensemble Methods as a Defense to Adversarial Perturbations Against Deep Neural Networks | Deep learning has become the state of the art approach in many machine learning problems such as classification. It has recently been shown that deep learning is highly vulnerable to adversarial perturbations. Taking the camera systems of self-driving cars as an example, small adversarial perturbations can cause the sy... | rejected-papers | The paper empirically evaluates the effectiveness of ensembles of deep networks against adversarial examples. The paper adds little to the existing literature in this area: an detailed study on "ensemble adversarial training" already exists, and the experimental evaluation in this paper is limited to MNIST and CIFAR (r... | train | [
"By2d_sBef",
"r1k7g8Oxz",
"B1fsb6bWz",
"rJB1WReNG",
"HkHuuvBMf",
"SJ1fdDHfM",
"SJ7nwDrfG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This paper describes the use of ensemble methods to improve the robustness of neural networks to adversarial examples. Adversarial examples are images that have been slightly modified (e.g. by adding some small perturbation) so that the neural network will predict a wrong class label.\n\nEnsemble methods have been... | [
7,
5,
4,
-1,
-1,
-1,
-1
] | [
3,
3,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_rkA1f3NpZ",
"iclr_2018_rkA1f3NpZ",
"iclr_2018_rkA1f3NpZ",
"iclr_2018_rkA1f3NpZ",
"By2d_sBef",
"r1k7g8Oxz",
"B1fsb6bWz"
] |
iclr_2018_HkeJVllRW | Sparse-Complementary Convolution for Efficient Model Utilization on CNNs | We introduce an efficient way to increase the accuracy of convolution neural networks (CNNs) based on high model utilization without increasing any computational complexity.
The proposed sparse-complementary convolution replaces regular convolution with sparse and complementary shapes of kernels, covering the sam... | rejected-papers | The paper studies factorizations of convolutional kernels. The proposed kernels lead to theoretical and practical efficiency improvements, but these improvements are very, very limited (for instance, Figure 5). It remains unclear how they compare to popular alternative approaches such as group convolutions (used in Res... | train | [
"BkdNAltlG",
"rkoBp1qxM",
"BJWR685ez",
"B13t7GhGf",
"SylLNMhff",
"BkQZNf3zz",
"rk6X7z2zf",
"S1XPWG2ff"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author"
] | [
"\n\nThis paper presented interesting ideas to reduce the redundancy in convolution kernels. They are very close to existing algorithms.\n\n(1)\tThe SW-SC kernel (Figure 2 (a)) is an extension of the existing shaped kernel (Figure 1 (c)).\n(2)\tThe CW-SC kernel (Figure 2 (c)) is very similar to interleaved group co... | [
6,
5,
5,
-1,
-1,
-1,
-1,
-1
] | [
5,
4,
4,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HkeJVllRW",
"iclr_2018_HkeJVllRW",
"iclr_2018_HkeJVllRW",
"iclr_2018_HkeJVllRW",
"rkoBp1qxM",
"rkoBp1qxM",
"BJWR685ez",
"BkdNAltlG"
] |
iclr_2018_ByqFhGZCW | MACHINE VS MACHINE: MINIMAX-OPTIMAL DEFENSE AGAINST ADVERSARIAL EXAMPLES | Recently, researchers have discovered that the state-of-the-art object classifiers can be fooled easily by small perturbations in the input unnoticeable to human eyes. It is known that an attacker can generate strong adversarial examples if she knows the classifier parameters. Conversely, a defender can robustify the ... | rejected-papers | The paper studies a adversarial attacks and defenses against convolutional networks based on a minimax formulation of the problem. Whilst this is an interesting direction of research, the present paper seems preliminary. In particular, compared to several other independent ICLR submissions, the empirical evaluation is ... | train | [
"B1kbvzdxz",
"B1oPNGYez",
"B1VlMxjlG",
"HkCRNN6XM",
"HJRfymTXM",
"SkfZVe6mz",
"H1tsY-hQz",
"SyJ9m4CWz",
"rymYGVR-f",
"HkLIb4Cbz",
"rygOMUuCb"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"public",
"author",
"public",
"author",
"author",
"author",
"author",
"public"
] | [
"The game-theoretic approach to attacks with / defense against adversarial examples is an important direction of the security of deep learning and I appreciate the authors to initiate this kind of study. \n\nLemma 1 summarizes properties of the solutions that are expected to have after reaching equilibria. Importan... | [
5,
6,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
3,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_ByqFhGZCW",
"iclr_2018_ByqFhGZCW",
"iclr_2018_ByqFhGZCW",
"HJRfymTXM",
"SkfZVe6mz",
"H1tsY-hQz",
"rygOMUuCb",
"B1kbvzdxz",
"B1oPNGYez",
"B1VlMxjlG",
"iclr_2018_ByqFhGZCW"
] |
iclr_2018_ryvxcPeAb | Enhancing the Transferability of Adversarial Examples with Noise Reduced Gradient | Deep neural networks provide state-of-the-art performance for many applications of interest. Unfortunately they are known to be vulnerable to adversarial examples, formed by applying small but malicious perturbations to the original inputs. Moreover, the perturbations can transfer across models: adversarial examples ge... | rejected-papers | The paper studies transferability of adversarial examples between model architectures, and proposes a method to improve this transferability. Whilst it covers an interesting and relevant line of research, the paper does not provide strong evidence for its main underling hypothesis: namely, that adversarial perturbation... | train | [
"rkzeadBxf",
"SJIOPWdgf",
"rkKt2t2xz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper postulates that an adversarial perturbation consists of a model-specific and data-specific component, and that amplification of the latter is best suited for adversarial attacks.\n\nThis paper has many grammatical errors. The article is almost always missing from nouns. Some of the sentences need changi... | [
4,
5,
5
] | [
4,
3,
4
] | [
"iclr_2018_ryvxcPeAb",
"iclr_2018_ryvxcPeAb",
"iclr_2018_ryvxcPeAb"
] |
iclr_2018_HJdXGy1RW | CrescendoNet: A Simple Deep Convolutional Neural Network with Ensemble Behavior | We introduce a new deep convolutional neural network, CrescendoNet, by stacking simple building blocks without residual connections. Each Crescendo block contains independent convolution paths with increased depths. The numbers of convolution layers and parameters are only increased linearly in Crescendo blocks. In exp... | rejected-papers | The paper proposes a new convolutional network architecture, called CrescendoNet. Whilst achieving competitive performance on CIFAR-10 and SVHN, the accuracy of the proposed model on CIFAR-100 is substantially lower than that of state-of-the-art models with fewer parameters; the paper presents no experimental results o... | train | [
"Bk1jsRM1f",
"HJwO67Klz",
"H1vOPPFez",
"r1-0rdbEz",
"SJIR8G67G",
"B1rJLMa7f",
"HJifHMaXf",
"rkpSQMrxG",
"B1ZpEWrlz",
"BJKHrDNgz",
"SJZn-KzxG",
"HynVX7Xyz",
"B17ZhCMyG",
"BJkODoG1z",
"BJSwy5MJz",
"BylimFMJf",
"Bymd5_fyG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"public",
"author",
"public",
"author",
"official_reviewer",
"author",
"official_reviewer",
"author",
"official_reviewer"
] | [
"The paper presents a new CNN architecture: CrescendoNet. It does not have skip connections yet performs quite well.\n\nOverall, I think the contributions of this paper are too marginal for acceptance in a top tier conference.\n\nThe architecture is competitive on SVHN and CIFAR 10 but not on CIFAR 100. The perform... | [
4,
5,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
5,
4,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HJdXGy1RW",
"iclr_2018_HJdXGy1RW",
"iclr_2018_HJdXGy1RW",
"SJIR8G67G",
"Bk1jsRM1f",
"HJwO67Klz",
"H1vOPPFez",
"B1ZpEWrlz",
"BJKHrDNgz",
"SJZn-KzxG",
"iclr_2018_HJdXGy1RW",
"B17ZhCMyG",
"iclr_2018_HJdXGy1RW",
"BJSwy5MJz",
"iclr_2018_HJdXGy1RW",
"Bymd5_fyG",
"iclr_2018_HJdXG... |
iclr_2018_Byj54-bAW | A Tensor Analysis on Dense Connectivity via Convolutional Arithmetic Circuits | Several state of the art convolutional networks rely on inter-connecting different layers to ease the flow of information and gradient between their input and output layers. These techniques have enabled practitioners to successfully train deep convolutional networks with hundreds of layers. Particularly, a novel way o... | rejected-papers | The paper performs a theoretical analysis of the representation power of convolutional networks with inter-layer connections. Whilst the results themselves are interesting, the current presentation of the paper stands in the way of the reading grasping and appreciating the main insights from the paper.
The authors ack... | train | [
"HywL2b9xG",
"rkoPNE9gG",
"HkXstRmZM",
"B1kgnv67f",
"r15DydamM",
"BJy-oPT7z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"SUMMARY\n\nTraditional convolutional neural networks consist of a sequence of information processing layers. However, one can relax this sequential design constraint so that higher layers receive inputs from one, some, or all preceding layers. This modification allows information to travel more freely throughout t... | [
4,
4,
5,
-1,
-1,
-1
] | [
3,
3,
3,
-1,
-1,
-1
] | [
"iclr_2018_Byj54-bAW",
"iclr_2018_Byj54-bAW",
"iclr_2018_Byj54-bAW",
"rkoPNE9gG",
"HywL2b9xG",
"HkXstRmZM"
] |
iclr_2018_B16_iGWCW | Deep Boosting of Diverse Experts | In this paper, a deep boosting algorithm is developed to
learn more discriminative ensemble classifier by seamlessly combining a set of base deep CNNs (base experts)
with diverse capabilities, e.g., these base deep CNNs are
sequentially trained to recognize a set of
object classes in an easy-to... | rejected-papers | The paper presents a boosting method and uses it to train an ensemble of convnets for image classification. The paper lacks conceptual and empirical comparisons with alternative boosting and ensembling methods. In fact, it is not even clear from the experimental results whether or not the proposed method outperforms a ... | train | [
"ry00rINxM",
"HJecicqxG",
"SJR0TtHZG",
"BJluMzjMf",
"Skm8WGozM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"public",
"public"
] | [
"This paper consider a version of boosting where in each iteration only class weights are updated rather than sample weights and apply that to a series of CNNs for object recognition tasks.\n\nWhile the paper is comprehensive in their derivations (very similar to original boosting papers and in many cases one to on... | [
2,
6,
5,
-1,
-1
] | [
5,
3,
4,
-1,
-1
] | [
"iclr_2018_B16_iGWCW",
"iclr_2018_B16_iGWCW",
"iclr_2018_B16_iGWCW",
"ry00rINxM",
"SJR0TtHZG"
] |
iclr_2018_rkmtTJZCb | Unsupervised Hierarchical Video Prediction | Much recent research has been devoted to video prediction and generation, but mostly for short-scale time horizons. The hierarchical video prediction method by Villegas et al. (2017) is an example of a state of the art method for long term video prediction. However, their method has limited applicability in practical... | rejected-papers | The paper presents a method for forward prediction in videos. The paper insufficiently motivates the proposed method and presents very limited empirical evaluations (no ablation studies, etc.) to backup its claims. This makes it difficult for the reader to put the work into the context of the broader research around l... | train | [
"BJ1X3tYgf",
"HyuKwSixM",
"B1e1owMZM",
"S1q0U7OMG",
"ByLWr7XXf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author"
] | [
"The paper treats the interesting problem of long term video prediction in complex video streams. I think the approach of adding more structure to their representation before making longer term prediction is also a reasonable one. Their approach combines an RNN that predicts an encoding of scene and then generating... | [
4,
4,
4,
-1,
-1
] | [
4,
4,
4,
-1,
-1
] | [
"iclr_2018_rkmtTJZCb",
"iclr_2018_rkmtTJZCb",
"iclr_2018_rkmtTJZCb",
"iclr_2018_rkmtTJZCb",
"S1q0U7OMG"
] |
iclr_2018_Hk-FlMbAZ | The Manifold Assumption and Defenses Against Adversarial Perturbations | In the adversarial-perturbation problem of neural networks, an adversary starts with a neural network model F and a point \bfx that F classifies correctly, and applies a \emph{small perturbation} to \bfx to produce another point \bfx′ that F classifies \emph{incorrectly}. In this paper, we propose taking into account... | rejected-papers | The original paper was sloppy in its use of mathematical constructs such as manifolds, made assumptions that are poorly motivated (see review #2 for details), and presented an empirical evaluation is preliminary. Based on the reviews, the authors have substantially revised the paper to try and address those issues by a... | train | [
"rkgspXWgf",
"HyjN-YPlz",
"S1ZQKlqxM",
"HyyW9e0ZM",
"SJ6IO2c7G",
"Sk93E_N7z",
"ryYl30GQM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"public"
] | [
"The authors argue that \"good\" classifiers naturally represent the classes in a classification as well-separated manifolds, and that adversarial examples are low-confidence examples lying near to one of these manifolds. The authors suggest \"fixing\" adversarial examples by projecting them back to the manifold, e... | [
3,
4,
5,
-1,
-1,
-1,
-1
] | [
3,
3,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_Hk-FlMbAZ",
"iclr_2018_Hk-FlMbAZ",
"iclr_2018_Hk-FlMbAZ",
"iclr_2018_Hk-FlMbAZ",
"iclr_2018_Hk-FlMbAZ",
"ryYl30GQM",
"iclr_2018_Hk-FlMbAZ"
] |
iclr_2018_rJbs5gbRW | On the Generalization Effects of DenseNet Model Structures | Modern neural network architectures take advantage of increasingly deeper layers, and various advances in their structure to achieve better performance. While traditional explicit regularization techniques like dropout, weight decay, and data augmentation are still being used in these new models, little about the regul... | rejected-papers | The paper appears unfinished in many ways: the experiments are preliminary, the paper completely ignored a large body of prior work on the subject, and the presentation needs substantial improvements. The authors did not provide a rebuttal.
I encourage the authors to refrain from submitting unfinished papers such as t... | train | [
"H1YnyaKgG",
"r1OmdAYxz",
"r1bLTRKxG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The ms analyses a number of simulations how skip connections effect the generalization of different network architectures. The experiments are somewhat interesting but they appear rather preliminary. To indeed show the claims made, error bars in the graphs would be necessary as well will more careful and more gene... | [
2,
3,
3
] | [
5,
4,
4
] | [
"iclr_2018_rJbs5gbRW",
"iclr_2018_rJbs5gbRW",
"iclr_2018_rJbs5gbRW"
] |
iclr_2018_SJCq_fZ0Z | Sparse Attentive Backtracking: Long-Range Credit Assignment in Recurrent Networks | A major drawback of backpropagation through time (BPTT) is the difficulty of learning long-term dependencies, coming from having to propagate credit information backwards through every single step of the forward computation. This makes BPTT both computationally impractical and biologically implausible. For this reason,... | rejected-papers | The authors propose to use attention over past time steps to try and solve the gradient flow problem in learning recurrent neural networks. Attention is performed over a subset of past states by a hueristic that boils to selecting best time-steps.
I agree with the authors that they offer a lot of comparisons, but like... | train | [
"S1mm_s0NM",
"S1GvBVDWz",
"SySGrVOVM",
"rkL2eNONz",
"H1qBHY5eM",
"SJppHuogG",
"r1JAdgsQf",
"rJoYueomM",
"SJtYjJsmf",
"rkIdYyoXf"
] | [
"author",
"official_reviewer",
"author",
"author",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"We thank the reviewer again for reviewing our paper. We would like to ask the reviewer if there is any further questions regarding our rebuttal, especially the updated MNIST results and the comparisons with full self-attention.",
"This work proposes Sparse Attentive Backtracking, an attention-based approach to i... | [
-1,
5,
-1,
-1,
5,
8,
-1,
-1,
-1,
-1
] | [
-1,
3,
-1,
-1,
4,
4,
-1,
-1,
-1,
-1
] | [
"H1qBHY5eM",
"iclr_2018_SJCq_fZ0Z",
"H1qBHY5eM",
"S1GvBVDWz",
"iclr_2018_SJCq_fZ0Z",
"iclr_2018_SJCq_fZ0Z",
"S1GvBVDWz",
"S1GvBVDWz",
"SJppHuogG",
"H1qBHY5eM"
] |
iclr_2018_HJMN-xWC- | Learning Parsimonious Deep Feed-forward Networks | Convolutional neural networks and recurrent neural networks are designed with network structures well suited to the nature of spacial and sequential data respectively. However, the structure of standard feed-forward neural networks (FNNs) is simply a stack of fully connected layers, regardless of the feature correlatio... | rejected-papers | I am inclined to agree with R1 that there is an extensive literature on learning architectures now, and I have seen two others as part of my area chairing. This paper does not offer comparisons to existing methods for architecture learning other than very basic ones and that reduces the strength of the paper significan... | train | [
"HkQdJFuef",
"rkDPp89xz",
"BJ25hzHWf",
"S1BSve3mz",
"BJna8g27G",
"Bku44r3Mf",
"S1nXjXVbz",
"SyTaKQVZf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"official_reviewer",
"author",
"author"
] | [
"There is a vast literature on structure learning for constructing neural networks (topologies, layers, learning rates, etc.) in an automatic fashion. Your work falls under a similar category. I am a bit surprised that you have not discussed it in the paper not to mention provided a baseline to compare your method ... | [
5,
4,
5,
-1,
-1,
-1,
-1,
-1
] | [
5,
2,
2,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HJMN-xWC-",
"iclr_2018_HJMN-xWC-",
"iclr_2018_HJMN-xWC-",
"BJ25hzHWf",
"Bku44r3Mf",
"S1nXjXVbz",
"HkQdJFuef",
"rkDPp89xz"
] |
iclr_2018_Hyp3i2xRb | Overcoming the vanishing gradient problem in plain recurrent networks | Plain recurrent networks greatly suffer from the vanishing gradient problem while Gated Neural Networks (GNNs) such as Long-short Term Memory (LSTM) and Gated Recurrent Unit (GRU) deliver promising results in many sequence learning tasks through sophisticated network designs. This paper shows how we can address this pr... | rejected-papers | The authors propose to use identity + some weights in the recurrent connections to prevent vanishing gradients. The reviewers found the experiments to have weak baselines, weakening the claims of the paper. | train | [
"SyJ4HP84z",
"r19wb884z",
"B1qhp-qeG",
"rJhtcsdxf",
"Hk7vlKsxz",
"Hy7jRyD-f",
"HJcyyeDbz",
"ByAEkxD-z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"After reviewing the revised draft, I have decided to not increase the score. I think 7 is still appropriate, as I'm not too sure about the impact.",
" Thanks for your reply and clarifications.\n\nI think overall this is a very interesting direction.\nHowever, authors did not address the comparison with previou... | [
-1,
-1,
4,
2,
7,
-1,
-1,
-1
] | [
-1,
-1,
5,
4,
4,
-1,
-1,
-1
] | [
"Hy7jRyD-f",
"HJcyyeDbz",
"iclr_2018_Hyp3i2xRb",
"iclr_2018_Hyp3i2xRb",
"iclr_2018_Hyp3i2xRb",
"Hk7vlKsxz",
"B1qhp-qeG",
"rJhtcsdxf"
] |
iclr_2018_BJ78bJZCZ | Efficiently applying attention to sequential data with the Recurrent Discounted Attention unit | Recurrent Neural Networks architectures excel at processing sequences by
modelling dependencies over different timescales. The recently introduced
Recurrent Weighted Average (RWA) unit captures long term dependencies
far better than an LSTM on several challenging tasks. The RWA achieves
this by ... | rejected-papers | RDA improves on RWA, but even so, the model is inferior to the other standard RNN models. As a result R1 and R3 question the motivation for the use of this model -- something the authors should motivate. | test | [
"BkEYMCPlG",
"rkzxWW5lf",
"r1jjF59lM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The authors present RDA, the Recurrent Discounted Attention unit, that improves upon RWA, the earlier introduced Recurrent Weighted Average unit, by adding a discount factor. While the RWA was an interesting idea with bad results (far worse than the standard GRU or LSTM with standard attention except for hand-pick... | [
4,
6,
3
] | [
5,
4,
4
] | [
"iclr_2018_BJ78bJZCZ",
"iclr_2018_BJ78bJZCZ",
"iclr_2018_BJ78bJZCZ"
] |
iclr_2018_HJ8W1Q-0Z | GATED FAST WEIGHTS FOR ASSOCIATIVE RETRIEVAL | We improve previous end-to-end differentiable neural networks (NNs) with fast
weight memories. A gate mechanism updates fast weights at every time step of
a sequence through two separate outer-product-based matrices generated by slow
parts of the net. The system is trained on a complex sequence to seq... | rejected-papers | The reviewers agree that while the presented result looks interesting, it is but one result. Further, one of the reviewer finds this to be a weak comparison as well.
The novelty of the approach over the paper by Ba et. al. also is in question -- good results on multiple tasks might have made it worth exploring, but the... | train | [
"SJIM1TDez",
"BJhNZOtlz",
"Syft4Nolf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The authors present an evolution of the idea of fast weights: training a double recurrent neural network, one \"slow\" trained as usual and one \"fast\" that gets updated in every time-step based on the slow network. The authors generalize this idea in a nice way and present results on 1 experiment. On the positi... | [
3,
5,
4
] | [
5,
4,
4
] | [
"iclr_2018_HJ8W1Q-0Z",
"iclr_2018_HJ8W1Q-0Z",
"iclr_2018_HJ8W1Q-0Z"
] |
iclr_2018_rJ1RPJWAW | Learnability of Learned Neural Networks | This paper explores the simplicity of learned neural networks under various settings: learned on real vs random data, varying size/architecture and using large minibatch size vs small minibatch size. The notion of simplicity used here is that of learnability i.e., how accurately can the prediction function of a neural ... | rejected-papers | + The paper proposes an interesting empirical measure of ""learnability"" of a trained network: how well the predictive function it represents can be learned by another network. And shows it empirically seems to correlate with better generalization.
- The work is purely empirical: it features no theory relating this ... | train | [
"H17N5b5lf",
"BJ9PiZ9eG",
"rkf_j7cgG",
"rk4gLL2Qf",
"rJWuFJtGG",
"BJcNFJKMG",
"HkqzKJFff",
"HkgcOJtGG",
"r1sUdyKMM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author",
"author"
] | [
"Summary:\nThis paper presents very nice experiments comparing the complexity of various different neural networks using the notion of \"learnability\" --- the learnability of a model (N1) is defined as the \"expected agreement\" between the output of N1, and the output of another model N2 which has been trained to... | [
7,
6,
4,
-1,
-1,
-1,
-1,
-1,
-1
] | [
4,
4,
4,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_rJ1RPJWAW",
"iclr_2018_rJ1RPJWAW",
"iclr_2018_rJ1RPJWAW",
"iclr_2018_rJ1RPJWAW",
"H17N5b5lf",
"BJ9PiZ9eG",
"BJ9PiZ9eG",
"rkf_j7cgG",
"rkf_j7cgG"
] |
iclr_2018_HylgYB3pZ | Linearly Constrained Weights: Resolving the Vanishing Gradient Problem by Reducing Angle Bias | In this paper, we first identify \textit{angle bias}, a simple but remarkable phenomenon that causes the vanishing gradient problem in a multilayer perceptron (MLP) with sigmoid activation functions. We then propose \textit{linearly constrained weights (LCW)} to reduce the angle bias in a neural network, so as to train... | rejected-papers | The paper identifies an interesting problem in sigmoid deep nets, addressed diffferently by batchnorm, and proposes a different simple fix. It shows empirically that constraining neuron's weights to sum to zero improves training of a 100 layers sigmoid MLP.
The work is currenlty limited in its theoretical contribution,... | train | [
"BJr2wYHVM",
"SymE04bxf",
"H1tXgwtgG",
"r1k49d7-M",
"S1Dw9mXVf",
"Hks4KOgQG",
"S1wQ85JQf",
"HJ9xvF1XM",
"BkxFmu17z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"I appreciate the effort taken by the authors to add more experimental results and enriching the intuitions of LCW. However, as the experiments have further shown, LCW only shows consistently better training accuracy than BN on a simple dataset as CIFAR10, but not as good on the testing data (therefore poor general... | [
-1,
5,
5,
4,
-1,
-1,
-1,
-1,
-1
] | [
-1,
4,
4,
4,
-1,
-1,
-1,
-1,
-1
] | [
"S1wQ85JQf",
"iclr_2018_HylgYB3pZ",
"iclr_2018_HylgYB3pZ",
"iclr_2018_HylgYB3pZ",
"BkxFmu17z",
"iclr_2018_HylgYB3pZ",
"r1k49d7-M",
"H1tXgwtgG",
"SymE04bxf"
] |
iclr_2018_SkHkeixAW | Regularization for Deep Learning: A Taxonomy | Regularization is one of the crucial ingredients of deep learning, yet the term regularization has various definitions, and regularization methods are often studied separately from each other. In our work we present a novel, systematic, unifying taxonomy to categorize existing methods. We distinguish methods that affec... | rejected-papers | The paper is a well-written review of regularization approaches in deep learning.
It does not offer novel approaches or novel insight with empirically demonstrated usefulness
=> ICLR is not the appropriate venue for it. | test | [
"rywwiW8xz",
"HysLX85lf",
"r1Dj4EXbf",
"rkueKoimM",
"Bk1wYis7G",
"SkltKjjXG",
"Hyg2KssmM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This paper is unusual in that it is more of a review than contributing novel knowledge. It considers a taxonomy of all the ways that machine learning (mostly deep learning) methods can achieve a form of regularization. \n\nUnfortunately, it starts with a definition of regularization ('making the model generalize b... | [
5,
4,
4,
-1,
-1,
-1,
-1
] | [
5,
4,
5,
-1,
-1,
-1,
-1
] | [
"iclr_2018_SkHkeixAW",
"iclr_2018_SkHkeixAW",
"iclr_2018_SkHkeixAW",
"iclr_2018_SkHkeixAW",
"rywwiW8xz",
"HysLX85lf",
"r1Dj4EXbf"
] |
iclr_2018_r111KtCp- | Taking Apart Autoencoders: How do They Encode Geometric Shapes ? | We study the precise mechanisms which allow autoencoders to encode and decode a simple geometric shape, the disk. In this carefully controlled setting, we are able to describe the specific form of the optimal solution to the minimisation problem of the training step. We show that the autoencoder indeed approximates thi... | rejected-papers | + interesting approach for a detailed analysis of the limitations of autoencoders in solving a simple toy problem
- resulting insights somewhat trivial, not really novel, nor practically useful => lacks demonstration of a gain on non-toy task
- regularization study too limited in scope: lacking theoretical grounding... | train | [
"H1s3VoOlf",
"SJQ2Xg2eM",
"SJOrnJf-M",
"Hk4SO_-zG",
"B1edwu-zf",
"ryPkDObfM",
"BJcjHdWzG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"1. The idea is interesting, but the study is not comprehensive yet\n2. need to visualize the input data space, with the training data, test data, the 'gaps' in training data [see a recent related paper - Stoecklein et al. Deep Learning for Flow Sculpting: Insights into Efficient Learning using Scientific Simulatio... | [
4,
4,
4,
-1,
-1,
-1,
-1
] | [
5,
4,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_r111KtCp-",
"iclr_2018_r111KtCp-",
"iclr_2018_r111KtCp-",
"H1s3VoOlf",
"SJOrnJf-M",
"SJQ2Xg2eM",
"iclr_2018_r111KtCp-"
] |
iclr_2018_rkhxwltab | AANN: Absolute Artificial Neural Network | This research paper describes a simplistic architecture named as AANN: Absolute Artificial Neural Network, which can be used to create highly interpretable representations of the input data. These representations are generated by penalizing the learning of the network in such a way that those learned representations co... | rejected-papers | The paper proposes to use absolute value activations, in a joint supervised + unsupervised training (classification + deep autoencoder with tied encoder/decoder weights).
Pros:
+ simple model and approach on ideas worth revisiting
Cons:
- The paper initially approached these old ideas as novel, missing much related pr... | train | [
"SyRcMDPeG",
"S1itI_FxM",
"B1J8Hw9xz",
"H1-Y4cCmz",
"rJy6y907M"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author"
] | [
"SUMMARY \n\nThe model is an ANN whose units have the absolute value function abs as their activation function (in place of ReLU, sigmoid, etc.). The network has bi-directional connections (with equal weights) between consecutive layers, but it operates only in one direction at a time. In the forward direction, it ... | [
2,
3,
6,
-1,
-1
] | [
3,
5,
4,
-1,
-1
] | [
"iclr_2018_rkhxwltab",
"iclr_2018_rkhxwltab",
"iclr_2018_rkhxwltab",
"SyRcMDPeG",
"S1itI_FxM"
] |
iclr_2018_Hyp-JJJRW | Style Memory: Making a Classifier Network Generative | Deep networks have shown great performance in classification tasks. However, the parameters learned by the classifier networks usually discard stylistic information of the input, in favour of information strictly relevant to classification. We introduce a network that has the capacity to do both classification and reco... | rejected-papers | + Paper proposes simple joint deep autoencoder + classifier training where the hidden representation is split between (observed) class and (unobserved) style nodes.
- Empirical evaluation is very limited, focusing on only qualitative evaluation of reconstructions and interpolations (on MNIST and EMNIST).
- Unclear g... | train | [
"rkWU5vQxf",
"H109AKKlM",
"S1yZxBslG",
"ByTDsb6Qf",
"SynAU-TmG",
"ryjOU-TQM",
"Sy1frWTQM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This paper proposes to train a classifier neural network not just to classifier, but also to reconstruct a representation of its input, in order to factorize the class information from the appearance (or \"style\" as used in this paper). This is done by first using unsupervised pretraining and then fine-tuning usi... | [
3,
3,
4,
-1,
-1,
-1,
-1
] | [
5,
5,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_Hyp-JJJRW",
"iclr_2018_Hyp-JJJRW",
"iclr_2018_Hyp-JJJRW",
"rkWU5vQxf",
"H109AKKlM",
"S1yZxBslG",
"iclr_2018_Hyp-JJJRW"
] |
iclr_2018_By0ANxbRW | DNN Model Compression Under Accuracy Constraints | The growing interest to implement Deep Neural Networks (DNNs) on resource-bound hardware has motivated innovation of compression algorithms. Using these algorithms, DNN model sizes can be substantially reduced, with little to no accuracy degradation. This is achieved by either eliminating components from the model, or ... | rejected-papers | Proposed network compression method offers limited technical novelty over existing approaches, and empirical evaluations do not clearly demonstrate an advantage over current state-of-the-art.
Paper presentation quality also needs to be improved. | train | [
"rk6muOPxG",
"Hk6K4Rwlf",
"HJlSjw_ez"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"1. Summary\n\nThis paper introduced a method to learn a compressed version of a neural network such that the loss of the compressed network doesn't dramatically change.\n\n\n2. High level paper\n\n- I believe the writing is a bit sloppy. For instance equation 3 takes the minimum over all m in C but C is defined to... | [
4,
3,
3
] | [
3,
3,
5
] | [
"iclr_2018_By0ANxbRW",
"iclr_2018_By0ANxbRW",
"iclr_2018_By0ANxbRW"
] |
iclr_2018_SkiCjzNTZ | Spontaneous Symmetry Breaking in Deep Neural Networks | We propose a framework to understand the unprecedented performance and robustness of deep neural networks using field theory. Correlations between the weights within the same layer can be described by symmetries in that layer, and networks generalize better if such symmetries are broken to reduce the redundancies of th... | rejected-papers | The paper makes overly strong claims, too weakly supported by a hard to follow and insufficiently rigorous mathematical argument. Connections with a large body of relevant prior literature are missing. | train | [
"BknXbsdxG",
"By02Louez",
"Byoj85JWf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"In this paper, an number of very strong (even extraordinary) claims are made:\n\n* The abstract promises \"a framework to understand the unprecedented performance and robustness of deep neural networks using field theory.\"\n* Page 8 states that this is \"This is a first attempt to describe a neural network with a... | [
3,
3,
3
] | [
4,
3,
3
] | [
"iclr_2018_SkiCjzNTZ",
"iclr_2018_SkiCjzNTZ",
"iclr_2018_SkiCjzNTZ"
] |
iclr_2018_SJa1Nk10b | Anytime Neural Network: a Versatile Trade-off Between Computation and Accuracy | We present an approach for anytime predictions in deep neural networks (DNNs). For each test sample, an anytime predictor produces a coarse result quickly, and then continues to refine it until the test-time computational budget is depleted. Such predictors can address the growing computational problem of DNNs by autom... | rejected-papers | The paper received mixed reviews with scores of 5 (R1), 5 (R2), 7 (R3). All three reviewers raise concerns about the lack of comparisons to other methods. The rebuttal is not compelling on this point. There are quite a few methods that could be used for this application available (often with source code) and should b... | train | [
"HJMEt_FeM",
"rJeC7UixM",
"ry-xL5Dfz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper proposes an anytime neural network, which can predict anytime while training. To achieve that, the model includes auxiliary predictions which can make early predictions. Specifically, the paper presents a loss weighting scheme that considers high correlation among nearby predictions, an oscillating loss... | [
7,
5,
5
] | [
2,
3,
4
] | [
"iclr_2018_SJa1Nk10b",
"iclr_2018_SJa1Nk10b",
"iclr_2018_SJa1Nk10b"
] |
iclr_2018_B1spAqUp- | Pixel Deconvolutional Networks | Deconvolutional layers have been widely used in a variety of deep
models for up-sampling, including encoder-decoder networks for
semantic segmentation and deep generative models for unsupervised
learning. One of the key limitations of deconvolutional operations
is that they result in the so-call... | rejected-papers | The paper received borderline-negative reviews with scores of 5,5,6. A consistent issue was the weakness of the experiments: (i) lack of comparison to appropriate baselines, (ii) differences between published/reported numbers for DeepLab-ResNet (R3) and (iii) related work, e.g. Wojna paper, as raised by R1. The AC did ... | train | [
"B1YorpYxz",
"B1L5VaYgG",
"BkZQtx5lz",
"BJ3XoYe-M",
"SJnHnYg-z",
"BkqscKe-z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This paper is well written and easy to follow. The authors propose pixel deconvolutional layers for convolutional neural networks. The motivation of the proposed method, PixelDCL, is to remove the checkerboard effect of deconvolutoinal layers. \nThe method consists of adding direct dependencies among the intermedi... | [
5,
5,
6,
-1,
-1,
-1
] | [
5,
4,
4,
-1,
-1,
-1
] | [
"iclr_2018_B1spAqUp-",
"iclr_2018_B1spAqUp-",
"iclr_2018_B1spAqUp-",
"B1YorpYxz",
"B1L5VaYgG",
"BkZQtx5lz"
] |
iclr_2018_ryzm6BATZ | Image Quality Assessment Techniques Improve Training and Evaluation of Energy-Based Generative Adversarial Networks | We propose a new, multi-component energy function for energy-based Generative Adversarial Networks (GANs) based on methods from the image quality assessment literature. Our approach expands on the Boundary Equilibrium Generative Adversarial Network (BEGAN) by outlining some of the short-comings of the original energy a... | rejected-papers | The paper received borderline-negative scores (6,5,5) with R1 and R2 having significant difficulty with the clarity of the paper. Although R3 was marginally positive, they pointed out that the experiments are "extremely weak". The AC look at the paper and agrees with R3 on this point. Therefore the paper cannot be acce... | train | [
"Bk8udEEeM",
"HJZIu0Kef",
"H1NEs7Clz",
"By1yEN7bG",
"SJfxhQ7WG",
"BJq66m7Wz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"Quick summary:\nThis paper proposes an energy based formulation to the BEGAN model and modifies it to include an image quality assessment based term. The model is then trained with CelebA under different parameters settings and results are analyzed.\n\nQuality and significance:\nThis is quite a technical paper, wr... | [
5,
5,
6,
-1,
-1,
-1
] | [
3,
3,
3,
-1,
-1,
-1
] | [
"iclr_2018_ryzm6BATZ",
"iclr_2018_ryzm6BATZ",
"iclr_2018_ryzm6BATZ",
"HJZIu0Kef",
"H1NEs7Clz",
"Bk8udEEeM"
] |
iclr_2018_BJjBnN9a- | Continuous Convolutional Neural Networks for Image Classification | This paper introduces the concept of continuous convolution to neural networks and deep learning applications in general. Rather than directly using discretized information, input data is first projected into a high-dimensional Reproducing Kernel Hilbert Space (RKHS), where it can be modeled as a continuous function us... | rejected-papers | The paper received borderline negative scores: 5,6,4.
The authors response to R1 question about the motivations was "...thus can achieve similar classification results with much smaller network sizes. This translates into smaller memory requirements, faster computational speeds and higher expressivity." If this is rea... | train | [
"rklk7M9ez",
"S1_ETHcef",
"rknFtmdff",
"H1N0Bs5Xz",
"Sk_lrgkmz",
"SkC64gJ7M",
"ByIjVlJ7z",
"HkS27lkmz",
"Hys4mg1mz",
"ByTIk_5pZ"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author",
"author",
"public"
] | [
"The paper introduces the notion of continuous convolutional neural networks. \nThe main idea of the paper is to project examples into an RK Hilbert space\nand performs convolution and filtering into that space. Interestingly, the\nfilters defined in the Hilbert space have parameters that are learnable.\n\nWhile t... | [
5,
6,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
2,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_BJjBnN9a-",
"iclr_2018_BJjBnN9a-",
"iclr_2018_BJjBnN9a-",
"ByTIk_5pZ",
"iclr_2018_BJjBnN9a-",
"rklk7M9ez",
"rklk7M9ez",
"S1_ETHcef",
"rknFtmdff",
"iclr_2018_BJjBnN9a-"
] |
iclr_2018_rkQu4Wb0Z | DNN Representations as Codewords: Manipulating Statistical Properties via Penalty Regularization | Performance of Deep Neural Network (DNN) heavily depends on the characteristics of hidden layer representations. Unlike the codewords of channel coding, however, the representations of learning cannot be directly designed or controlled. Therefore, we develop a family of penalty regularizers where each one aims to affec... | rejected-papers | The paper received scores of 5,5,5, with the reviewers agreeing the paper was marginally below the acceptance threshold. The main issue, raised by both R2 and R3 was that connection between representation learning in deep nets and coding theory was not fully justified/made. With no reviewer advocating acceptance, it i... | test | [
"B1H8H5YxG",
"HkSWIWqez",
"BkuLn46ez",
"r18LCpbQG",
"H1uR66W7G",
"SkTKap-Qf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This paper presents a set of regularizers which aims for manipulating the statistical properties like sparsity, variance and covariance. While some of the proposed regularizers are applied to weights, most are applied to hidden representations of neural networks. Class-wise regularizations are also investigated fo... | [
5,
5,
5,
-1,
-1,
-1
] | [
5,
3,
4,
-1,
-1,
-1
] | [
"iclr_2018_rkQu4Wb0Z",
"iclr_2018_rkQu4Wb0Z",
"iclr_2018_rkQu4Wb0Z",
"B1H8H5YxG",
"HkSWIWqez",
"BkuLn46ez"
] |
iclr_2018_HkMhoDITb | Reinforcement Learning via Replica Stacking of Quantum Measurements for the Training of Quantum Boltzmann Machines | Recent theoretical and experimental results suggest the possibility of using current and near-future quantum hardware in challenging sampling tasks. In this paper, we introduce free-energy-based reinforcement learning (FERL) as an application of quantum hardware. We propose a method for processing a quantum annealer’s ... | rejected-papers | All three reviewers agreed that the paper was an interesting, giving a demonstration of what quantum computer could achieve. However, they all also felt that the topic was outside the main interests of the conference and better suited to other venues, e.g. a quantum computation workshop. The AC agrees with them. Thus u... | train | [
"BysYuqXxz",
"rkW8HjOlz",
"SJ22faFez",
"BJsN2Ia7G",
"rylNA8a7z",
"r1Hv6IamG",
"H12QoLpQG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"Summary: The paper demonstrates the use of a quantum annealing machine to solve a free-energy based reinforcement learning problem. Experimental results are demonstrated on a toy gridworld task, where if I understand correctly it does better than a DQN and a method based on RBM-free-energy approximation (Sallans a... | [
4,
6,
4,
-1,
-1,
-1,
-1
] | [
3,
4,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HkMhoDITb",
"iclr_2018_HkMhoDITb",
"iclr_2018_HkMhoDITb",
"SJ22faFez",
"BysYuqXxz",
"rkW8HjOlz",
"iclr_2018_HkMhoDITb"
] |
iclr_2018_SksY3deAW | Learning Deep ResNet Blocks Sequentially using Boosting Theory | We prove a multiclass boosting theory for the ResNet architectures which simultaneously creates a new technique for multiclass boosting and provides a new algorithm for ResNet-style architectures. Our proposed training algorithm, BoostResNet, is particularly suitable in non-differentiable architectures. Our method on... | rejected-papers | All three reviewers felt that the paper was just below the acceptance threshold, with scores of 5,4,5. R1 felt there were problems in the proofs, but the authors rebuttal satisfactorily addressed this. R3 and the authors had an extended discussion with the authors, but did not revise their score from its initial value ... | train | [
"BkUBAzFlG",
"SJMDsFilM",
"HJQdsudQM",
"S1wa3wp7f",
"HJTLF8nmf",
"BJChJEi7M",
"SyHTUGwfM",
"SyUUmkDGG",
"By4KyasWf",
"rJOTfTjZf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"official_reviewer",
"author",
"author",
"official_reviewer",
"author",
"author"
] | [
"Disclaimer: I reviewed this paper for NIPS as well and many of comments made by reviewers at that time still apply to this version of the paper as well, although presentation has overall improved.\n\nThe paper presents a boosting-style algorithm for training deep residual networks. Convergence analysis for trainin... | [
5,
4,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
4,
4,
3,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_SksY3deAW",
"iclr_2018_SksY3deAW",
"iclr_2018_SksY3deAW",
"HJTLF8nmf",
"SyHTUGwfM",
"HJQdsudQM",
"SyUUmkDGG",
"rJOTfTjZf",
"SJMDsFilM",
"BkUBAzFlG"
] |
iclr_2018_BJgPCveAW | Characterizing Sparse Connectivity Patterns in Neural Networks | We propose a novel way of reducing the number of parameters in the storage-hungry fully connected layers of a neural network by using pre-defined sparsity, where the majority of connections are absent prior to starting training. Our results indicate that convolutional neural networks can operate without any loss of acc... | rejected-papers | The paper received weak scores: 4,4,5. R2 complained about clarity. R3's point about the lack of fully connected layers in current SOA deepnets is very valid and the authors response far from convincing. Unfortunately the major revision provided by the authors was not commented on by the reviewers, but many of the majo... | train | [
"S1951kYef",
"Bk-lFRWWz",
"HyYy75NMG",
"HyBhKWsXf",
"HJhs-1o7G",
"Syblgksmz",
"r1wbCAcmM",
"r12EC0cQG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author"
] | [
"This paper examines sparse connection patterns in upper layers of convolutional image classification networks. Networks with very few connections in the upper layers are experimentally determined to perform almost as well as those with full connection masks. Heuristics for distributing connections among windows/... | [
4,
5,
4,
-1,
-1,
-1,
-1,
-1
] | [
3,
3,
3,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_BJgPCveAW",
"iclr_2018_BJgPCveAW",
"iclr_2018_BJgPCveAW",
"iclr_2018_BJgPCveAW",
"HyYy75NMG",
"Bk-lFRWWz",
"S1951kYef",
"r1wbCAcmM"
] |
iclr_2018_S1fHmlbCW | Neural Networks for irregularly observed continuous-time Stochastic Processes | Designing neural networks for continuous-time stochastic processes is challenging, especially when observations are made irregularly. In this article, we analyze neural networks from a frame theoretic perspective to identify the sufficient conditions that enable smoothly recoverable representations of signals in L^2(R)... | rejected-papers | The scores were not favorable: 5,5,2. R2 felt the motivation of the paper was inadequate. R3 raised numerous technical points, some of which were addressed in the rebuttal, but not all. R3 continues to have issue with some of the results. The AC agrees with R3's concerns and feels that the paper cannot be accepted in i... | train | [
"r1wssyDlf",
"ByeAk4weM",
"HJZM5e9eM",
"ryPpT3xEM",
"S1RntVn7z",
"BkO8q42mG",
"SJvEcE2mG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"The authors proved that convolutional neural networks with Leaky ReLU activation function are nonlinear frames, and similar results hold for non-uniformly sampled time-series as well. My main concern on this part is that theory is too rough and its link to the later part of the paper is weak. Although frames are s... | [
5,
5,
2,
-1,
-1,
-1,
-1
] | [
5,
4,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_S1fHmlbCW",
"iclr_2018_S1fHmlbCW",
"iclr_2018_S1fHmlbCW",
"S1RntVn7z",
"HJZM5e9eM",
"r1wssyDlf",
"ByeAk4weM"
] |
iclr_2018_SkrHeXbCW | Learning Representations for Faster Similarity Search | In high dimensions, the performance of nearest neighbor algorithms depends crucially on structure in the data.
While traditional nearest neighbor datasets consisted mostly of hand-crafted feature vectors, an increasing number of datasets comes from representations learned with neural networks.
We study the ... | rejected-papers | The paper received three good quality reviews which were in agreement that the paper was below the acceptance threshold. The authors are encouraged to follow the suggestions from the reviews to revise the paper and resubmit to another venue. | val | [
"ry39-uBxG",
"BkcUX-5eG",
"rJIi7bcgM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"\nThe context is indexing images with descriptor vectors obtained from a DNN. This paper studies the impact of changing the classification part on top of the DNN on the ability to index the descriptors with a LSH or a kd-tree algorithm. The modifications include: applying or not a ReLU and batch normalization (and... | [
4,
4,
4
] | [
5,
5,
4
] | [
"iclr_2018_SkrHeXbCW",
"iclr_2018_SkrHeXbCW",
"iclr_2018_SkrHeXbCW"
] |
iclr_2018_B1mAkPxCZ | VOCABULARY-INFORMED VISUAL FEATURE AUGMENTATION FOR ONE-SHOT LEARNING | A natural solution for one-shot learning is to augment training data to handle the data deficiency problem. However, directly augmenting in the image domain may not necessarily generate training data that sufficiently explore the intra-class space for one-shot classification. Inspired by the recent vocabulary-informed ... | rejected-papers | Two reviewers recommended rejection, and one was on the edge. There was no rebuttal to address the concerns and questions posed by the reviewers. | train | [
"Sk5zOVceG",
"SyBcch5lM",
"H1Wo7H2Zf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper proposes a feature augmentation method for one-shot learning. The proposed approach is very interesting. However, the method needs to be further clarified and the experiments need to be improved. \n\nDetails:\n1. The citation format used in the paper is not appropriate, which makes the paper, especiall... | [
4,
6,
5
] | [
3,
4,
4
] | [
"iclr_2018_B1mAkPxCZ",
"iclr_2018_B1mAkPxCZ",
"iclr_2018_B1mAkPxCZ"
] |
iclr_2018_HymYLebCb | Network Signatures from Image Representation of Adjacency Matrices: Deep/Transfer Learning for Subgraph Classification | We propose a novel subgraph image representation for classification of network fragments with the target being their parent networks. The graph image representation is based on 2D image embeddings of adjacency matrices. We use this image representation in two modes. First, as the input to a machine learning algorithm. ... | rejected-papers | The main idea of the paper is to transform graph classification into image representation (via adjacency matrices). Two reviewers are positive, while one is negative. The concerns are novelty (as mentioned by R2), while the last reviewer thinks the method is too simple and unprincipled (here the AC agrees with authors ... | train | [
"SJBw7As1f",
"r1jbeZ9xM",
"ry9izvnlf",
"HJVVl5q7M",
"HJgG5RbXM",
"SyLJxCbXG",
"HyhR2pZmf",
"rJyCr_Pef",
"ry6_ZhLgM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"public",
"public",
"public",
"public",
"official_reviewer"
] | [
"The paper proposes to use 2-d image representation techniques as a means of learning representations of graphs via their adjacency matrices. The adjacency matrix (or a subgraph of it) is first re-ordered to produce some canonical ordering which can then be fed into an image representation method. This can then be ... | [
3,
6,
6,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
3,
3,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HymYLebCb",
"iclr_2018_HymYLebCb",
"iclr_2018_HymYLebCb",
"iclr_2018_HymYLebCb",
"r1jbeZ9xM",
"SJBw7As1f",
"ry9izvnlf",
"ry6_ZhLgM",
"iclr_2018_HymYLebCb"
] |
iclr_2018_Skvd-myR- | Learning Non-Metric Visual Similarity for Image Retrieval | Measuring visual (dis)similarity between two or more instances within a data distribution is a fundamental task in many applications, specially in image retrieval. Theoretically, non-metric distances are able to generate a more complex and accurate similarity model than metric distances, provided that the non-linear da... | rejected-papers | Two reviewers recommended rejection, and the last reviewer votes for acceptance. The authors provided a rebuttal, including the end-to-end experiment (although the AC agrees with the authors that this experiment is not crucial to the paper). The AC read the paper and the reviews. While there are clearly interesting asp... | train | [
"By32fJqlG",
"SkT3Sw9lG",
"SySI56ubG",
"BJ5wfNKfG",
"H1WLMsUzM",
"Sksd0LB-G",
"HyG22UHbG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"The authors of this work propose learning a similarity measure for visual similarity and obtain, by doing that, an improvement in the very well-known datasets of Oxford and Paris for image retrieval. The work takes high-level image representations generated with an existing architecture (R-MAC), and train on top a... | [
7,
4,
3,
-1,
-1,
-1,
-1
] | [
5,
4,
5,
-1,
-1,
-1,
-1
] | [
"iclr_2018_Skvd-myR-",
"iclr_2018_Skvd-myR-",
"iclr_2018_Skvd-myR-",
"iclr_2018_Skvd-myR-",
"SySI56ubG",
"SkT3Sw9lG",
"By32fJqlG"
] |
iclr_2018_SJw03ceRW | GENERATIVE LOW-SHOT NETWORK EXPANSION | Conventional deep learning classifiers are static in the sense that they are trained on
a predefined set of classes and learning to classify a novel class typically requires
re-training. In this work, we address the problem of Low-shot network-expansion
learning. We introduce a learning framework whic... | rejected-papers | Two reviewers recommended rejection, and one is slightly more positive. The main concern is that the experiments are not convincing (ie, the number of base and added classes is very small). Furthermore, while the paper introduces several interesting ideas, the AC agrees with the second reviewer that each of these could... | train | [
"r1R0L9Def",
"r1W-h8Flz",
"BJK7re9ez",
"B1Fj2sjmf",
"BJWd2ismz",
"SyTJhosXM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"On few-shot learning problem, this paper presents a simple yet powerful distillation method where the base network is augmented with additional weights to classify the novel classes, while keeping the weights of the base network unchanged. Thus the so-called hard distillation is proposed. This paper is well-writte... | [
6,
4,
4,
-1,
-1,
-1
] | [
4,
3,
4,
-1,
-1,
-1
] | [
"iclr_2018_SJw03ceRW",
"iclr_2018_SJw03ceRW",
"iclr_2018_SJw03ceRW",
"r1R0L9Def",
"r1W-h8Flz",
"BJK7re9ez"
] |
iclr_2018_BJluxbWC- | Unseen Class Discovery in Open-world Classification | This paper concerns open-world classification, where the classifier not only needs to classify test examples into seen classes that have appeared in training but also reject examples from unseen or novel classes that have not appeared in training. Specifically, this paper focuses on discovering the hidden unseen classe... | rejected-papers | Three reviewers recommended rejection and there was no rebuttal to overturn their recommendation. | train | [
"HkWZFMVxf",
"ry2G0fvxM",
"HkFsIQclG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The main goal of this paper is to cluster images from classes unseen during training.\nThis is an interesting extension of the open-world paradigm, where at test time, the classifier has to identify images beloning to the C seen classes during training, but also identify (reject) images which were previously unsee... | [
5,
5,
4
] | [
4,
5,
4
] | [
"iclr_2018_BJluxbWC-",
"iclr_2018_BJluxbWC-",
"iclr_2018_BJluxbWC-"
] |
iclr_2018_HJr4QJ26W | Improving image generative models with human interactions | GANs provide a framework for training generative models which mimic a data distribution. However, in many cases we wish to train a generative model to optimize some auxiliary objective function within the data it generates, such as making more aesthetically pleasing images. In some cases, these objective functions are ... | rejected-papers | The reviewers agree that the idea of incorporating humans in the training of generative adversarial networks is interesting and worthwhile exploring. However, they felt that the paper fell short in providing strong support for their approach. The AC agrees. The authors are encouraged to strengthen their work and resubm... | train | [
"SyRiTtIgG",
"r1P_bO5eG",
"SyuHdLJ-z",
"SJjcxSFQG",
"SJ_XxStXM",
"B1NQTNYXz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"+ Quality:\nThe paper discusses an interesting direction of incorporating humans in the training of a generative adversarial networks in the hope of improving generated samples. I personally find this exciting/refreshing and will be useful in the future of machine learning.\n\nHowever, the paper shows only prelimi... | [
4,
5,
4,
-1,
-1,
-1
] | [
5,
3,
4,
-1,
-1,
-1
] | [
"iclr_2018_HJr4QJ26W",
"iclr_2018_HJr4QJ26W",
"iclr_2018_HJr4QJ26W",
"SyRiTtIgG",
"r1P_bO5eG",
"SyuHdLJ-z"
] |
iclr_2018_SJVHY9lCb | Learning to Select: Problem, Solution, and Applications | We propose a "Learning to Select" problem that selects the best among the flexible size candidates. This makes decisions based not only on the properties of the candidate, but also on the environment in which they belong to. For example, job dispatching in the manufacturing factory is a typical "Learning to Select" pro... | rejected-papers | Three reviewers recommended rejection and there was no rebuttal. | val | [
"r1uDK-Yez",
"rJW87QaxM",
"SyAd7DuZf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The paper proposed a new framework called `Learning to select’, in which a best candidate needs to be identified in the decision making process such as job dispatching. A CNN architecture is designed, called `Variable-Length CNN’, to solve this problem.\n\nMy major concern is on the definition of the proposed conc... | [
4,
4,
4
] | [
4,
4,
5
] | [
"iclr_2018_SJVHY9lCb",
"iclr_2018_SJVHY9lCb",
"iclr_2018_SJVHY9lCb"
] |
iclr_2018_HJDUjKeA- | Learning objects from pixels | We show how discrete objects can be learnt in an unsupervised fashion from pixels, and how to perform reinforcement learning using this object representation.
More precisely, we construct a differentiable mapping from an image to a discrete tabular list of objects, where each object consists of a different... | rejected-papers | All three reviewers recommended rejection and there was no rebuttal. | train | [
"Sk9f1tBlz",
"BJTS11qlz",
"SJ6V5roeG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper learns to construct masks and feature representations from an input image, in order to represent objects. This is applied to the relatively simple domain of Atari games video input (compared to natural images). The paper is completely inadequate in respect to related work; it re-invents known techniques... | [
3,
4,
4
] | [
4,
3,
4
] | [
"iclr_2018_HJDUjKeA-",
"iclr_2018_HJDUjKeA-",
"iclr_2018_HJDUjKeA-"
] |
iclr_2018_SkAK2jg0b | An Out-of-the-box Full-network Embedding for Convolutional Neural Networks | Transfer learning for feature extraction can be used to exploit deep representations in contexts where there is very few training data, where there are limited computational resources, or when tuning the hyper-parameters needed for training is not an option. While previous contributions to feature extraction propose em... | rejected-papers | Three reviewers recommended rejection, and there was no rebuttal. | train | [
"BkNxXL5ef",
"H1n46uTxf",
"By2YMc3WM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper proposes an out-of-the-box embedding for image classification task. Instead of taking one single layer output from pre-trained network as the feature vector for new dataset, the method first extracts the activations from all the layers, then runs spatial average pooling on all convolutional layers, then... | [
3,
4,
4
] | [
4,
5,
5
] | [
"iclr_2018_SkAK2jg0b",
"iclr_2018_SkAK2jg0b",
"iclr_2018_SkAK2jg0b"
] |
iclr_2018_BkoCeqgR- | On the Construction and Evaluation of Color Invariant Networks | This is an empirical paper which constructs color invariant networks and evaluates their performances on a realistic data set. The paper studies the simplest possible case of color invariance: invariance under pixel-wise permutation of the color channels. Thus the network is aware not of the specific color object, but ... | rejected-papers | Three reviewers recommend rejection and there is no rebuttal. | train | [
"HJzYXIOxG",
"B1eq0Hqlz",
"rJrWUW0gM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The paper proposes and evaluates a method to make neural networks for image recognition color invariant.\n\nThe contribution of the paper is: \n - some proposed methods to extract a color-invariant representation\n - an experimental evaluation of the methods on the cifar 10 dataset\n - a new dataset \"crashed cars... | [
4,
3,
3
] | [
4,
4,
4
] | [
"iclr_2018_BkoCeqgR-",
"iclr_2018_BkoCeqgR-",
"iclr_2018_BkoCeqgR-"
] |
iclr_2018_Bym0cU1CZ | Towards Interpretable Chit-chat: Open Domain Dialogue Generation with Dialogue Acts | Conventional methods model open domain dialogue generation as a black box through end-to-end learning from large scale conversation data. In this work, we make the first step to open the black box by introducing dialogue acts into open domain dialogue generation. The dialogue acts are generally designed and reveal how ... | rejected-papers | This work takes dialogue acts into account to generate responses in a human-machine conversation. However, incorporating dialogue acts into open-domain dialogue was already the focus of Zhao et al's ACL 2017 paper, Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders, a... | val | [
"rkBh-NNSf",
"rknNBk4BM",
"SJ7bhA2fM",
"rJiRgVkSz",
"rJNoU0uEf",
"BkY9B2d4G",
"ry2CUfcxz",
"HklWe9qxz",
"ryiSW8nef",
"BknHTJTfz",
"SydD_y6fG",
"r1aPFC2zz"
] | [
"author",
"public",
"author",
"author",
"author",
"public",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"Thank for reminding us the paper.\n\nThis is a system description paper in which the system is the one built by MILA for the 1st Alexa Prize competition. In the system, dialogue acts are used as one of the 1458 features for learning a response selection model and a special feature in the following abstract disco... | [
-1,
-1,
-1,
-1,
-1,
-1,
7,
4,
7,
-1,
-1,
-1
] | [
-1,
-1,
-1,
-1,
-1,
-1,
3,
5,
4,
-1,
-1,
-1
] | [
"rknNBk4BM",
"iclr_2018_Bym0cU1CZ",
"HklWe9qxz",
"iclr_2018_Bym0cU1CZ",
"BkY9B2d4G",
"iclr_2018_Bym0cU1CZ",
"iclr_2018_Bym0cU1CZ",
"iclr_2018_Bym0cU1CZ",
"iclr_2018_Bym0cU1CZ",
"iclr_2018_Bym0cU1CZ",
"ry2CUfcxz",
"ryiSW8nef"
] |
iclr_2018_Bki1Ct1AW | Baseline-corrected space-by-time non-negative matrix factorization for decoding single trial population spike trains | Activity of populations of sensory neurons carries stimulus information in both the temporal and the spatial dimensions. This poses the question of how to compactly represent all the information that the population codes carry across all these dimensions. Here, we developed an analytical method to factorize a large num... | rejected-papers | This work is incremental compared to previous work, solving very specific challenges, and would probably appeal to only a very limited fraction of ICLR's audience. | train | [
"SJz_9SFgz",
"ByuRMz5eG",
"S1Ib7Lcxf",
"B1eWVftmf",
"BkMdCktmM",
"ry7fAJKXf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This study proposes the use of non-negative matrix factorization accounting for baseline by subtracting the pre-stimulus baseline from each trial and subsequently decompose the data using a 3-way factorization thereby identifying spatial and temporal modules as well as their signed activation. The method is used o... | [
4,
6,
6,
-1,
-1,
-1
] | [
4,
3,
3,
-1,
-1,
-1
] | [
"iclr_2018_Bki1Ct1AW",
"iclr_2018_Bki1Ct1AW",
"iclr_2018_Bki1Ct1AW",
"SJz_9SFgz",
"ByuRMz5eG",
"S1Ib7Lcxf"
] |
iclr_2018_S1GUgxgCW | Latent Topic Conversational Models | Despite much success in many large-scale language tasks, sequence-to-sequence (seq2seq) models have not been an ideal choice for conversational modeling as they tend to generate generic and repetitive responses. In this paper, we propose a Latent Topic Conversational Model (LTCM) that augments the seq2seq model with a ... | rejected-papers | This paper combines existing models to detect topics and generate responses, and the resulting model is shown to be slightly preferred by human evaluators over baselines. This is quite incremental and the results are not impressive enough to stand on their own merit. | val | [
"S1B9wqGrz",
"HyY3SM9NM",
"H1E8RNcxz",
"rypOZF5eG",
"r1ahhhJWM",
"HJeDIJaXG",
"B1nrBm57f",
"rJnFHQ5XM",
"HkES4X57M",
"SJG0X7cQf"
] | [
"author",
"public",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"public",
"public",
"public",
"public"
] | [
"The main contribution of the paper is three-fold as mentioned in this post (– General comments on Contributions–): \n1) We were first to be able to jointly learn the neural topic and seq2seq models.\n2) The paper offers a better understanding/training of latent models for languages.\n3) Both an extensive evaluatio... | [
-1,
-1,
4,
5,
6,
-1,
-1,
-1,
-1,
-1
] | [
-1,
-1,
3,
4,
4,
-1,
-1,
-1,
-1,
-1
] | [
"HyY3SM9NM",
"iclr_2018_S1GUgxgCW",
"iclr_2018_S1GUgxgCW",
"iclr_2018_S1GUgxgCW",
"iclr_2018_S1GUgxgCW",
"iclr_2018_S1GUgxgCW",
"rypOZF5eG",
"H1E8RNcxz",
"r1ahhhJWM",
"iclr_2018_S1GUgxgCW"
] |
iclr_2018_SkBHr1WRW | Ego-CNN: An Ego Network-based Representation of Graphs Detecting Critical Structures | While existing graph embedding models can generate useful embedding vectors that perform well on graph-related tasks, what valuable information can be jointly learned by a graph embedding model is less discussed. In this paper, we consider the possibility of detecting critical structures by a graph embedding model. We ... | rejected-papers | This paper deals with the important topic of learning better graph representations and shows promise in helping to detect critical substructures of graph that would help with the interpretability of representations. Unfortunately, this work fails to accurately portray how it relates to previous work (in particular, Nie... | val | [
"Hk3rCW5ef",
"rk7Oq1oxG",
"H1FOVLn-G",
"rkp69mLfM",
"r1lYsm8Mz",
"Sk5zj78zz",
"S1cjuXIMf",
"rk9fnGXRZ"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author"
] | [
"The authors proposed a convolutional framework based on merging ego-networks. It combines graph embedding layers with task driven output layers, producing interpretable results for critical structure detection. While based on existing embedding methods such as Patchy-San, the contribution of ego-centric convolutio... | [
7,
4,
4,
-1,
-1,
-1,
-1,
-1
] | [
3,
4,
4,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_SkBHr1WRW",
"iclr_2018_SkBHr1WRW",
"iclr_2018_SkBHr1WRW",
"rk7Oq1oxG",
"iclr_2018_SkBHr1WRW",
"H1FOVLn-G",
"Hk3rCW5ef",
"iclr_2018_SkBHr1WRW"
] |
iclr_2018_BJ6anzb0Z | Multimodal Sentiment Analysis To Explore the Structure of Emotions | We propose a novel approach to multimodal sentiment analysis using deep neural
networks combining visual recognition and natural language processing. Our
goal is different than the standard sentiment analysis goal of predicting whether
a sentence expresses positive or negative sentiment; instead, we a... | rejected-papers | This work combines words and images from Tumblr to provide more fine-grained sentiment analysis than just positive-negative. The contribution is too slight, as a straightforward combination of existing architectures applied on an emotion classification task with conclusions that aren't well motivated and are not provid... | train | [
"rJA29bLxf",
"BJ2J7pFgf",
"HJcw0y5eM",
"B1AeovaXM",
"Ske1iPpQG",
"HkZ_Yw6mM",
"S1U4Kv67f"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"\nThe authors present a study that aims at inferring the \"emotional\" tags provided by Thumblr users starting from images and texts in the captions. For text processing the authors use a standard LSTM taking as input GLOVE vectors of words in a sentence. For visual information, authors use a pretrained CNN (with ... | [
6,
4,
5,
-1,
-1,
-1,
-1
] | [
5,
5,
5,
-1,
-1,
-1,
-1
] | [
"iclr_2018_BJ6anzb0Z",
"iclr_2018_BJ6anzb0Z",
"iclr_2018_BJ6anzb0Z",
"BJ2J7pFgf",
"BJ2J7pFgf",
"rJA29bLxf",
"HJcw0y5eM"
] |
iclr_2018_rJ7yZ2P6- | Enhance Word Representation for Out-of-Vocabulary on Ubuntu Dialogue Corpus | Ubuntu dialogue corpus is the largest public available dialogue corpus to make it feasible to build end-to-end
deep neural network models directly from the conversation data. One challenge of Ubuntu dialogue corpus is
the large number of out-of-vocabulary words. In this paper we proposed an algorithm which... | rejected-papers | This paper's idea is to augment pre-trained word embeddings on a large corpus with embeddings learned on the data of interest. This is shown to yield better results than the pre-trained word embeddings alone. This contribution is too limited to justify publication at iclr. | train | [
"rJdjmmLez",
"BkomChuxf",
"H1RMVeqgz",
"r1RKt4ffM",
"BJzBG4Gfz",
"ryX6HZzzM",
"r1lgG9n-z",
"HJAQJq2Wf",
"rJswuIPbG",
"HJAg9rDZM",
"SJEsevIZf",
"B1BaLmQWz",
"SyDbmz6gf",
"BkpgMaheG",
"B15ZQj2gf",
"rk9R5BjgM",
"r1UNAmogz",
"HkFtNMclM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"public",
"author",
"author",
"author",
"public",
"author",
"author",
"public",
"author",
"public",
"author",
"author",
"public"
] | [
"Summary:\nThis paper proposes an approach to improve the out-of-vocabulary embedding prediction for the task of modeling dialogue conversations. The proposed approach uses generic embeddings and combines them with the embeddings trained on the training dataset in a straightforward string-matching algorithm. In add... | [
6,
3,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
5,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_rJ7yZ2P6-",
"iclr_2018_rJ7yZ2P6-",
"iclr_2018_rJ7yZ2P6-",
"ryX6HZzzM",
"ryX6HZzzM",
"iclr_2018_rJ7yZ2P6-",
"BkomChuxf",
"SJEsevIZf",
"HJAg9rDZM",
"BkpgMaheG",
"B1BaLmQWz",
"H1RMVeqgz",
"BkpgMaheG",
"B15ZQj2gf",
"r1UNAmogz",
"HkFtNMclM",
"HkFtNMclM",
"iclr_2018_rJ7yZ2P6-"... |
iclr_2018_By5SY2gA- | Towards Building Affect sensitive Word Distributions | Learning word representations from large available corpora relies on the distributional hypothesis that words present in similar contexts tend to have similar meanings. Recent work has shown that word representations learnt in this manner lack sentiment information which, fortunately, can be leveraged using external kn... | rejected-papers | This work attempts to incorporate affect information from additional resources into word embeddings. This is a valuable goal, but the methods used are very similar to existing ones, and the experimental results are not quite convincing enough to make a strong enough case for accepting the paper. | train | [
"HyaDR19xG",
"ByVH7MslM",
"H1YXWe6gM",
"HJqOCra7M",
"r1t5jrTmG",
"HyW-sH6XM",
"H17McHTmf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This paper proposed to use affect lexica to improve word embeddings. They extended the training objective functions of Word2vec and Glove with the affect information. The resulting embeddings were evaluated not only on word similarity tasks but also on a bunch of downstream applications such as sentiment analysis.... | [
6,
4,
4,
-1,
-1,
-1,
-1
] | [
3,
5,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_By5SY2gA-",
"iclr_2018_By5SY2gA-",
"iclr_2018_By5SY2gA-",
"HyaDR19xG",
"ByVH7MslM",
"H1YXWe6gM",
"iclr_2018_By5SY2gA-"
] |
iclr_2018_r17lFgZ0Z | Relevance of Unsupervised Metrics in Task-Oriented Dialogue for Evaluating Natural Language Generation | Automated metrics such as BLEU are widely used in the machine translation literature. They have also been used recently in the dialogue community for evaluating dialogue response generation. However, previous work in dialogue response generation has shown that these metrics do not correlate strongly with human judgment... | rejected-papers | This paper tackles a very important problem: evaluating natural language generation. The paper presents an overview of existing unsupervised metrics, and looks at how they correlate with human evaluation scores. This is important work and the empirical conclusions are useful to the community, but the datasets used are ... | train | [
"Hy0xrQegf",
"SJWidMceG",
"rkEMEscxf",
"HyBXcDamM",
"Bk66KPaQz",
"ByiJYDTXM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This paper's main thesis is that automatic metrics like BLEU, ROUGE, or METEOR is suitable for task-oriented natural language generation (NLG). In particular, the paper presents a counterargument to \"How NOT To Evaluate Your Dialogue System...\" where Wei et al argue that automatic metrics are not correlated or o... | [
4,
5,
5,
-1,
-1,
-1
] | [
4,
4,
3,
-1,
-1,
-1
] | [
"iclr_2018_r17lFgZ0Z",
"iclr_2018_r17lFgZ0Z",
"iclr_2018_r17lFgZ0Z",
"Hy0xrQegf",
"SJWidMceG",
"rkEMEscxf"
] |
iclr_2018_ByhthReRb | A Neural Method for Goal-Oriented Dialog Systems to interact with Named Entities | Many goal-oriented dialog tasks, especially ones in which the dialog system has to interact with external knowledge sources such as databases, have to handle a large number of Named Entities (NEs). There are at least two challenges in handling NEs using neural methods in such settings: individual NEs may occur only rar... | rejected-papers | This work deals with the important task of capturing named entities in a goal-directed setting. The description of the work and the experiments are not ready for publication; for example, it is unclear whether the proposed method would have an advantage over existing methods such as the match type features that are onl... | train | [
"SJP-xVtlf",
"B1hNwrFeM",
"S18iNb5lG",
"r1qKHZ77f",
"SyAtG-7mM",
"Skp_ybmXz",
"BJpwneQQM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"The paper proposes to generate embedding of named-entities on the fly during dialogue sessions. If the text is from the user, a named entity recognizer is used. If it is from the bot response, then it is known which words are named entities therefore embedding can be constructed directly. The idea has some novelty... | [
6,
4,
3,
-1,
-1,
-1,
-1
] | [
3,
3,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_ByhthReRb",
"iclr_2018_ByhthReRb",
"iclr_2018_ByhthReRb",
"SJP-xVtlf",
"B1hNwrFeM",
"S18iNb5lG",
"iclr_2018_ByhthReRb"
] |
iclr_2018_HJXyS7bRb | A Goal-oriented Neural Conversation Model by Self-Play | Building chatbots that can accomplish goals such as booking a flight ticket is an unsolved problem in natural language understanding. Much progress has been made to build conversation models using techniques such as sequence2sequence modeling. One challenge in applying such techniques to building goal-oriented conversa... | rejected-papers | While using self-play for training a goal-oriented dialogue system makes sense, the contribution of this paper compared to previous work (that the paper itself cites) seems too minor, and the limitations of using toy synthetic data further weaken the work.
| train | [
"SkIuskXef",
"B1uZU0Kgf",
"SyZknb5ez"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"I like the idea of coupling the language and the conversation model. This is in line with the latest trends of constructing end-to-end NN models that deal with the conversation in a holistic manner. The idea of enforcing information isolation is brilliant. Creating hidden information and allowing the two-party mod... | [
6,
4,
3
] | [
3,
3,
4
] | [
"iclr_2018_HJXyS7bRb",
"iclr_2018_HJXyS7bRb",
"iclr_2018_HJXyS7bRb"
] |
iclr_2018_Syl3_2JCZ | A Self-Organizing Memory Network | Working memory requires information about external stimuli to be represented in the brain even after those stimuli go away. This information is encoded in the activities of neurons, and neural activities change over timescales of tens of milliseconds. Information in working memory, however, is retained for tens of seco... | rejected-papers | This work extends Druckmann and Chklowskii, 2012 and demonstrates some interesting properties of the new model. This would be of interest to a neuroscience audience, but the focus is off for ICLR. | test | [
"S1VMieDlG",
"SytEarFxz",
"S1Vm7Z7GG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This is a great discussion on an interesting problem in computational neuroscience, that of holding an attractor memory stable even though individual neurons fluctuate. The previously published idea is that this is possible when the sum of all these memory neurons remain constant for the specific readout network, ... | [
3,
4,
4
] | [
2,
4,
4
] | [
"iclr_2018_Syl3_2JCZ",
"iclr_2018_Syl3_2JCZ",
"iclr_2018_Syl3_2JCZ"
] |
iclr_2018_HJXOfZ-AZ | When and where do feed-forward neural networks learn localist representations? | According to parallel distributed processing (PDP) theory in psychology, neural networks (NN) learn distributed rather than interpretable localist representations. This view has been held so strongly that few researchers have analysed single units to determine if this assumption is correct. However, recent results from... | rejected-papers | This work looks at what factors can lead to the emergence of selectivity (to certain categories) in units of a neural network. While this is an intriguing area to explore, this work uses settings that are quite toy-ish, making it a very hard to see how the observations could generalize to more realistic architectures o... | train | [
"ryBhOOXlM",
"ryuoSrKxM",
"Bko4LP9eM",
"Hkpg-16mG",
"BkhtWLYzG",
"Bk65YSKff",
"rJbgBrYGf",
"SJoDNBYfz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"public",
"public",
"public",
"public",
"public"
] | [
"The authors ask when the hidden layer units of a multi-layer feed-forward neural network will display selectivity to object categories. They train 3-layer ANNs to categorize binary patterns, and find that typically at least some of the hidden layer units are category selective. The number of category selective (\"... | [
3,
3,
5,
-1,
-1,
-1,
-1,
-1
] | [
3,
5,
4,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HJXOfZ-AZ",
"iclr_2018_HJXOfZ-AZ",
"iclr_2018_HJXOfZ-AZ",
"iclr_2018_HJXOfZ-AZ",
"ryBhOOXlM",
"ryuoSrKxM",
"Bko4LP9eM",
"Bko4LP9eM"
] |
iclr_2018_SJiHOSeR- | Contextual memory bandit for pro-active dialog engagement | An objective of pro-activity in dialog systems is to enhance the usability of conversational
agents by enabling them to initiate conversation on their own. While
dialog systems have become increasingly popular during the last couple of years,
current task oriented dialog systems are still mainly react... | rejected-papers | This paper is lacking in terms of clarity and experimentation, and would require a lot of additional work to bring it to the standards of any high quality venue. | train | [
"ByJXMOblf",
"H10BwZ9xM",
"Byjc8N5lz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This article propose to combine a form of contextual Thompson sampling policy with memory networks to handle dialog engagement in mobility interfaces.\n\nThe idea of using contextual bandits (especially Thompson sampling) instead of state of the art approximate RL algorithms (like DQN, AC3, or fittedQ) is surprisi... | [
2,
3,
3
] | [
5,
4,
4
] | [
"iclr_2018_SJiHOSeR-",
"iclr_2018_SJiHOSeR-",
"iclr_2018_SJiHOSeR-"
] |
iclr_2018_BkpXqwUTZ | Iterative temporal differencing with fixed random feedback alignment support spike-time dependent plasticity in vanilla backpropagation for deep learning | In vanilla backpropagation (VBP), activation function matters considerably in terms of non-linearity and differentiability.
Vanishing gradient has been an important problem related to the bad choice of activation function in deep learning (DL).
This work shows that a differentiable activation function is no... | rejected-papers | This paper is nowhere near standards for publication anywhere. | train | [
"rJUnrQDlG",
"rkFKYjdlz",
"r18pMHFgM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"- This paper is not well written and incomplete. There is no clear explanation of what exactly the authors want to achieve in the paper, what exactly is their approach/contribution, experimental setup, and analysis of their results. \n\n- The paper is hard to read due to many abbreviations, e.g., the last paragrap... | [
3,
2,
2
] | [
4,
5,
5
] | [
"iclr_2018_BkpXqwUTZ",
"iclr_2018_BkpXqwUTZ",
"iclr_2018_BkpXqwUTZ"
] |
iclr_2018_BJ_QxP1AZ | Unleashing the Potential of CNNs for Interpretable Few-Shot Learning | Convolutional neural networks (CNNs) have been generally acknowledged as one of the driving forces for the advancement of computer vision. Despite their promising performances on many tasks, CNNs still face major obstacles on the road to achieving ideal machine intelligence. One is that CNNs are complex and hard to int... | rejected-papers | The paper builds on earlier work by Wang et al (2015) on Visual Concepts (VCs) and explores the use of VCs for few-shot learning setting for novel classes.
The work, as pointed out by two reviewers is somewhat incremental in nature, with main novelty being the demonstration of utilities of VCs for few shot learning. T... | train | [
"By1Lg8Ygz",
"BJaIDpBEM",
"rJxG4bqlG",
"rkj8j16lM",
"S1SBDs27z",
"SJ_tUo2QM",
"HJ55NohXz",
"S1WHVi3mG",
"rktB7jnXf",
"H1AGk5_Gz",
"ry84z1fff"
] | [
"official_reviewer",
"author",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author",
"author",
"public"
] | [
"My main concern for this paper is that the description of the Visual Concepts is completely unclear for me. At some point I thought I did understand it, but then the next equation didnt make sense anymore... If I understand correctly, f_p is a representation of *all images* of a specific layer *k* at/around pixel ... | [
5,
-1,
7,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
4,
-1,
4,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_BJ_QxP1AZ",
"rktB7jnXf",
"iclr_2018_BJ_QxP1AZ",
"iclr_2018_BJ_QxP1AZ",
"SJ_tUo2QM",
"By1Lg8Ygz",
"rJxG4bqlG",
"rkj8j16lM",
"iclr_2018_BJ_QxP1AZ",
"ry84z1fff",
"iclr_2018_BJ_QxP1AZ"
] |
iclr_2018_B1i7ezW0- | Semi-Supervised Learning via New Deep Network Inversion | We exploit a recently derived inversion scheme for arbitrary deep neural networks to develop a new semi-supervised learning framework that applies to a wide range of systems and problems.
The approach reaches current state-of-the-art methods on MNIST and provides reasonable performances on SVHN and CIFAR10. Thr... | rejected-papers | The paper proposes a novel approach for DNN inversion mainly targeted towards semi-supervised learning. However the semi-supervised learning results are not competitive enough. Although the authors mention in the author-response that semi-supervised learning is not the main goal of the paper, the experiments and claims... | val | [
"BJDhjYXlz",
"HkBSVnBxG",
"HyvZDmueM",
"BJ2lgU6Xf",
"H1qHBZOQM",
"HkqVJzQff",
"rJHWkfXzM",
"r1wA0Zmff"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"official_reviewer",
"author",
"author",
"author"
] | [
"After reading the revision:\n\nThe authors addressed my detailed questions on experiments. It appears sometimes the entropy loss (which is not the main contribution of the paper) is essential to improve performance; this obscures the main contribution.\n\nOn the other hand, the theoretical part of the paper is not... | [
5,
4,
7,
-1,
-1,
-1,
-1,
-1
] | [
4,
5,
2,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_B1i7ezW0-",
"iclr_2018_B1i7ezW0-",
"iclr_2018_B1i7ezW0-",
"H1qHBZOQM",
"rJHWkfXzM",
"BJDhjYXlz",
"HkBSVnBxG",
"HyvZDmueM"
] |
iclr_2018_ByzvHagA- | Disentangled activations in deep networks | Deep neural networks have been tremendously successful in a number of tasks.
One of the main reasons for this is their capability to automatically
learn representations of data in levels of abstraction,
increasingly disentangling the data as the internal transformations are applied.
In this pape... | rejected-papers | The novelty of the paper is limited and it lacks on comparisons with relevant baselines, as pointed out by the reviewers. | train | [
"H1dMyvFgG",
"HJbgHsYlM",
"rkhaGO9gG",
"Hy0nCO6XG",
"BkbMROTQG",
"Sy5z-uTXf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"The authors propose a penalization term that enforces decorrelation between the dimensions of the representation \nThey show that it can be included as additional term in cost functions to train generic models.\nThe idea is simple and it seems to work for the presented examples.\n\nHowever, they talk about gradien... | [
6,
5,
4,
-1,
-1,
-1
] | [
3,
4,
3,
-1,
-1,
-1
] | [
"iclr_2018_ByzvHagA-",
"iclr_2018_ByzvHagA-",
"iclr_2018_ByzvHagA-",
"H1dMyvFgG",
"HJbgHsYlM",
"rkhaGO9gG"
] |
iclr_2018_SySpa-Z0Z | From Information Bottleneck To Activation Norm Penalty | Many regularization methods have been proposed to prevent overfitting in neural networks. Recently, a regularization method has been proposed to optimize the variational lower bound of the Information Bottleneck Lagrangian. However, this method cannot be generalized to regular neural network architectures. We present t... | rejected-papers | All reviewers have acknowledged that the proposed regularization is novel and also results in some empirical improvements on the reported language modeling and image classification tasks. However there are serious concerns on writing and rigor (reviewers Anon1 and Anon3) of the paper. The authors have not uploaded any ... | train | [
"rJgVu2HgM",
"Hks-5ZwlG",
"rJhMyTnlf",
"r1w72HamG",
"rkxNtHpXG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author"
] | [
"The paper puts forward Activation Norm Penalty (\"ANR\", an L_2 type regularization on the activations), deriving it from the Information Bottleneck principle. As usual with Information Bottleneck style constructions, the loss takes on a variational form.\n\nThe experiments demonstrate small but consistent gains w... | [
7,
4,
4,
-1,
-1
] | [
3,
3,
4,
-1,
-1
] | [
"iclr_2018_SySpa-Z0Z",
"iclr_2018_SySpa-Z0Z",
"iclr_2018_SySpa-Z0Z",
"rJgVu2HgM",
"Hks-5ZwlG"
] |
iclr_2018_ryykVe-0W | Learning Independent Features with Adversarial Nets for Non-linear ICA | Reliable measures of statistical dependence could potentially be useful tools for learning independent features and performing tasks like source separation using Independent Component Analysis (ICA). Unfortunately, many of such measures, like the mutual information, are hard to estimate and optimize directly. We prop... | rejected-papers | The paper proposes the use of GANs to match the joint distribution of features to the product of their marginals for ICA. The approach is totally plausible but reviewers have complaints about lack of rigor and analysis in terms of (i) mixing conditions under which the proposed GAN based approach will work, given that I... | train | [
"SybdzO8Nf",
"H1hlWndxM",
"HyoEDdvxG",
"ry2lpp_ez",
"SytMSVTmf",
"Sk1uKXTXf",
"BkIBV767G"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"Thank you for your response.\n\nMaximizing independence under general mixing conditions does not necessarily lead to the recovery of the underlying independent sources (even up to the standard ambiguities); this is one of the major motivations why the linear and post-nonlinear ICA (PNL-ICA) tasks have been conside... | [
-1,
5,
3,
6,
-1,
-1,
-1
] | [
-1,
5,
5,
3,
-1,
-1,
-1
] | [
"SytMSVTmf",
"iclr_2018_ryykVe-0W",
"iclr_2018_ryykVe-0W",
"iclr_2018_ryykVe-0W",
"HyoEDdvxG",
"H1hlWndxM",
"ry2lpp_ez"
] |
iclr_2018_ry0WOxbRZ | IVE-GAN: Invariant Encoding Generative Adversarial Networks | Generative adversarial networks (GANs) are a powerful framework for generative tasks. However, they are difficult to train and tend to miss modes of the true data generation process. Although GANs can learn a rich representation of the covered modes of the data in their latent space, the framework misses an inverse map... | rejected-papers | Reviewers recognize that the method proposed is somewhat novel but have strong reservations on the experimental evaluation. Discussion of some relevant papers is also missing (eg, Li et al, 2017 : ALICE). The authors have not responded to the many concerns expressed by the reviewers. | train | [
"Sk2jcxFlM",
"SJqfr3qlf",
"rJgavPneM",
"B1eAjclgG",
"HyPVk6p1G"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"public"
] | [
"The paper proposes a modified GAN objective, summarized in Eq.(3). It consists of two parts:\n(A) Classic GAN term: \\E_{x ~ P_{data} } \\log D'(x) + \\E_{z ~ P_{Z}, z' ~ P_{Z'} } \\log D'( G(z',E(x)) )\n(B) Invariant Encoding term: \\E_{x ~ P_{data} } [ \\log D(T(x),x) + \\E_{z' ~ P_{Z'} } \\log D( G(z',E(x)... | [
5,
4,
5,
-1,
-1
] | [
4,
5,
4,
-1,
-1
] | [
"iclr_2018_ry0WOxbRZ",
"iclr_2018_ry0WOxbRZ",
"iclr_2018_ry0WOxbRZ",
"HyPVk6p1G",
"iclr_2018_ry0WOxbRZ"
] |
iclr_2018_S17mtzbRb | Forced Apart: Discovering Disentangled Representations Without Exhaustive Labels | Learning a better representation with neural networks is a challenging problem, which has been tackled from different perspectives in the past few years. In this work, we focus on learning a representation that would be useful in a clustering task. We introduce two novel loss components that substantially improve the q... | rejected-papers | The paper proposes two regularizers for encouraging "clustered feature embeddings" (use of "disentangled" in the title is misleading). Reviewers have raised points about the lack of proper motivation and justification of the regularizers. There are also concerns on the experiments conducted to evaluate the method, incl... | train | [
"BkH22ZFxG",
"HyDv0CYgz",
"r1qbYHogz",
"BkUT6TW7G",
"B1etRT-7f",
"HyYLRp-7G",
"BytXAaW7f"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This paper proposes two regularization terms to encourage learning disentangled representations. One term is applied to weight parameters of a layer just like weight decay. The other is applied to the activations of the target layer (e.g., the penultimate layer). The core part of both regularization terms is a com... | [
5,
5,
4,
-1,
-1,
-1,
-1
] | [
5,
4,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_S17mtzbRb",
"iclr_2018_S17mtzbRb",
"iclr_2018_S17mtzbRb",
"iclr_2018_S17mtzbRb",
"BkH22ZFxG",
"HyDv0CYgz",
"r1qbYHogz"
] |
iclr_2018_SyYYPdg0- | Counterfactual Image Networks | We capitalize on the natural compositional structure of images in order to learn object segmentation with weakly labeled images. The intuition behind our approach is that removing objects from images will yield natural images, however removing random patches will yield unnatural images. We leverage this signal to devel... | rejected-papers | All reviewers acknowledge that the idea of the paper is interesting but have expressed serious concerns on empirical evaluations. The paper is not suitable for publication in its current form. | train | [
"HJSdXVqxG",
"SyqB7xHlz",
"rkFHVe5lf",
"ry6MYLa7z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author"
] | [
"This paper creates a layered representation in order to better learn segmentation from unlabeled images. It is well motivated, as Fig. 1 clearly shows the idea that if the segmentation was removed properly, the result would still be a natural image. However, the method itself as described in the paper leaves many ... | [
4,
5,
4,
-1
] | [
4,
4,
4,
-1
] | [
"iclr_2018_SyYYPdg0-",
"iclr_2018_SyYYPdg0-",
"iclr_2018_SyYYPdg0-",
"iclr_2018_SyYYPdg0-"
] |
iclr_2018_rJTGkKxAZ | Learning Generative Models with Locally Disentangled Latent Factors | One of the most successful techniques in generative models has been decomposing a complicated generation task into a series of simpler generation tasks. For example, generating an image at a low resolution and then learning to refine that into a high resolution image often improves results substantially. Here we expl... | rejected-papers | Reviewers recognize the proposed method of hierarchical extension to ALI to be potentially novel and interesting but have expressed strong concerns on the experiments section. The paper also needs to have comparisons with relevant hierarchical generative model baselines. Not suitable for publication in its current form... | val | [
"H13up9Klf",
"ByKf-8jlG",
"HypMNiy-G",
"Sy_g2vaQG",
"HkTqLPaQz",
"r1F-NvpQM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This paper proposed a method called Locally Disentangled Factors for hierarchical latent variable generative model, which can be seen as a hierarchical variant of Adversarially Learned Inference (Dumoulin el atl. 2017). The idea seems to be a valid variant, however, the quality of the paper is not good. The introd... | [
4,
6,
3,
-1,
-1,
-1
] | [
4,
3,
4,
-1,
-1,
-1
] | [
"iclr_2018_rJTGkKxAZ",
"iclr_2018_rJTGkKxAZ",
"iclr_2018_rJTGkKxAZ",
"H13up9Klf",
"ByKf-8jlG",
"HypMNiy-G"
] |
iclr_2018_rJL6pz-CZ | Transfer Learning on Manifolds via Learned Transport Operators | Within-class variation in a high-dimensional dataset can be modeled as being on a low-dimensional manifold due to the constraints of the physical processes producing that variation (e.g., translation, illumination, etc.). We desire a method for learning a representation of the manifolds induced by identity-preserving t... | rejected-papers | Learning identity-preserving transformations from unlabeled data is definitely an important and useful direction. However the paper does not have convincing experiments to establish the effectiveness of the proposed method on real datasets which is a crucial limitation in my view, given that the paper is largely based ... | train | [
"rypjqWBlf",
"ryjTQZ9xz",
"HklkGPeeG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper propose to learn manifold transport operators via a dictionary learning framework that alternatively optimize a dictionary of transformations and coefficients defining the transformation between random pairs of data points. Experiments on the swiss roll and synthetic rotated images on USPS digits show t... | [
5,
4,
4
] | [
4,
4,
4
] | [
"iclr_2018_rJL6pz-CZ",
"iclr_2018_rJL6pz-CZ",
"iclr_2018_rJL6pz-CZ"
] |
iclr_2018_HJaDJZ-0W | Block-Sparse Recurrent Neural Networks | Recurrent Neural Networks (RNNs) are used in state-of-the-art models in domains such as speech recognition, machine translation, and language modelling. Sparsity is a technique to reduce compute and memory requirements of deep learning models. Sparse RNNs are easier to deploy on devices and high-end server processors. ... | rejected-papers | Pros
-- Interesting approach to induce sparsity, trains faster than alternative approaches
Cons
-- Fairly complex set of heuristics for pruning weights
-- Han et al. works well, although the authors claim it takes more time to train, which may not not hold for all training sets and doesn’t seem like a strong enough rea... | train | [
"ByH5FWfgf",
"rJWbr5zxM",
"Hk8Ah7_eG",
"Hkid-pzNf",
"HJCr-6z4z",
"rkyLQgkNM",
"HktlUv6XM",
"B19rTJ8GM",
"S1QJz0Wfz",
"rkegZRbMz",
"r18hlCZzG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"official_reviewer",
"author",
"official_reviewer",
"author",
"author",
"author"
] | [
"The authors propose a block sparsity pruning approach to compress RNNs. There are several ways. One is using group LASSO to promote sparsity. The other is to prune, but with a very specialized schedule as to the pruning and pruning weight, motivated by the work of Narang et al 2017 for non-group sparsity. The bl... | [
5,
7,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
4,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HJaDJZ-0W",
"iclr_2018_HJaDJZ-0W",
"iclr_2018_HJaDJZ-0W",
"rkyLQgkNM",
"rkyLQgkNM",
"B19rTJ8GM",
"iclr_2018_HJaDJZ-0W",
"S1QJz0Wfz",
"ByH5FWfgf",
"rJWbr5zxM",
"Hk8Ah7_eG"
] |
iclr_2018_B1NGT8xCZ | Principled Hybrids of Generative and Discriminative Domain Adaptation | We propose a probabilistic framework for domain adaptation that blends both generative and discriminative modeling in a principled way. Under this framework, generative and discriminative models correspond to specific choices of the prior over parameters. This provides us a very general way to interpolate between gener... | rejected-papers | Pros
-- Nice way to formulate domain adaptation in a Bayesian framework that explains why autoencoder and domain difference losses are useful.
Cons
-- Model closely follows the framework, but the overall strategy is similar to previous models (but with improved rationale).
-- Experimental section can be improved. It w... | test | [
"SyeuDWqef",
"HJ6gA_L4G",
"HkTrtoKlf",
"SySd0xngf",
"B14UUUxzM",
"SJ9QI8ezf",
"BJne88xff",
"rkAhH8eff"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This is a very well-written paper that shows how to successfully use (generative) autoencoders together with the (discriminative) domain adversarial neural network (DANN) of Ganin et al.\nThe construction is simple but nicely backed by a probabilistic analysis of the domain adaptation problem.\n\nThe only criticis... | [
6,
-1,
5,
5,
-1,
-1,
-1,
-1
] | [
3,
-1,
4,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_B1NGT8xCZ",
"SJ9QI8ezf",
"iclr_2018_B1NGT8xCZ",
"iclr_2018_B1NGT8xCZ",
"HkTrtoKlf",
"SyeuDWqef",
"SySd0xngf",
"iclr_2018_B1NGT8xCZ"
] |
iclr_2018_B1tC-LT6W | Trace norm regularization and faster inference for embedded speech recognition RNNs | We propose and evaluate new techniques for compressing and speeding up dense matrix multiplications as found in the fully connected and recurrent layers of neural networks for embedded large vocabulary continuous speech recognition (LVCSR). For compression, we introduce and study a trace norm regularization technique f... | rejected-papers | Pros
-- Shows alternative strategies to train low-rank factored weight matrices for recurrent nets.
Cons
-- Minor modifications (and gains) over other forms of regularization like L2.
-- Results are only on an ASR task, so it’s not entirely clear how they’ll work on other tasks.
As pointed out by the reviewers, unles... | train | [
"Hk1q_liEG",
"Bk-k0Ctgz",
"BJgpEMDEM",
"HJcmDCFgz",
"By29r_AeG",
"HkhMCaWQM",
"BJ_PnTW7M",
"S1UpspbmG"
] | [
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"Indeed, at this point, it seems hard to escape the conclusion that trace norm regularization is not substantially or at all superior to L2 regularization with respect to the number of parameters versus CER trade-off. In retrospect, we should have written the paper quite differently to highlight the comparison of r... | [
-1,
4,
-1,
5,
5,
-1,
-1,
-1
] | [
-1,
3,
-1,
5,
3,
-1,
-1,
-1
] | [
"BJgpEMDEM",
"iclr_2018_B1tC-LT6W",
"BJ_PnTW7M",
"iclr_2018_B1tC-LT6W",
"iclr_2018_B1tC-LT6W",
"HJcmDCFgz",
"Bk-k0Ctgz",
"By29r_AeG"
] |
iclr_2018_SkJd_y-Cb | Word2net: Deep Representations of Language | Word embeddings extract semantic features of words from large datasets of text.
Most embedding methods rely on a log-bilinear model to predict the occurrence
of a word in a context of other words. Here we propose word2net, a method that
replaces their linear parametrization with neural networks. For e... | rejected-papers | Pros
-- Extends embeddings to use a richer representation; simple yet interesting improvement on Mikolov et al. work.
Cons
-- All of the reviewers pointed out that the experimental evaluations needs improvement. The authors should find better ways to improve both quantitative (e.g., accuracy in analogies as in Mikolov ... | train | [
"rkSCZ0uxG",
"SkH7fQYez",
"rJADQOqlG",
"H1a4LD6mz",
"BkmVBP6mM",
"H1z6NDTmz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"The paper extends SGNS as follows. In SGNS, each word x is associated with vectors a_x and r_x. Given a set of context words C, the model calculates the probability that the target word is x by a dot product between a_x and the average of {r_c: c in C}. The paper generalizes this computation to an arbitrary netwo... | [
5,
4,
4,
-1,
-1,
-1
] | [
5,
4,
4,
-1,
-1,
-1
] | [
"iclr_2018_SkJd_y-Cb",
"iclr_2018_SkJd_y-Cb",
"iclr_2018_SkJd_y-Cb",
"rkSCZ0uxG",
"SkH7fQYez",
"rJADQOqlG"
] |
iclr_2018_Hyig0zb0Z | Gated ConvNets for Letter-Based ASR | In this paper we introduce a new speech recognition system, leveraging a simple letter-based ConvNet acoustic model. The acoustic model requires only audio transcription for training -- no alignment annotations, nor any forced alignment step is needed. At inference, our decoder takes only a word list and a language mod... | rejected-papers | Pros
-- Competitive results on LibriSpeech.
Cons
-- Limited novelty, and lacks enough comparisons.
-- Comparison with other end-to-end approaches, and on other commonly used datasets, like WSJ, missing.
-- Gated convnets have already been proposed.
-- Letter based systems have been shown to be competitive to phone base... | train | [
"SkEUwDtkG",
"Hyp5jKfxf",
"Byw1O6Fgz",
"Byv3IsSzG",
"Bkdm2tXMG",
"S1R0OMGGG",
"Sk7YcE1fG",
"rkU2e9DJM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"public",
"official_reviewer",
"author",
"public"
] | [
"This paper applies gated convolutional neural networks [1] to speech recognition, using the training criterion ASG [2]. It is fair to say that this paper contains almost no novelty.\n\nThis paper starts by bashing the complexity of conventional HMM systems, and states the benefits of their approach. However, all o... | [
3,
6,
4,
-1,
-1,
-1,
-1,
-1
] | [
5,
5,
4,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_Hyig0zb0Z",
"iclr_2018_Hyig0zb0Z",
"iclr_2018_Hyig0zb0Z",
"S1R0OMGGG",
"iclr_2018_Hyig0zb0Z",
"Sk7YcE1fG",
"iclr_2018_Hyig0zb0Z",
"iclr_2018_Hyig0zb0Z"
] |
iclr_2018_S1Ow_e-Rb | How do deep convolutional neural networks learn from raw audio waveforms? | Prior work on speech and audio processing has demonstrated the ability to obtain excellent performance when learning directly from raw audio waveforms using convolutional neural networks (CNNs). However, the exact inner workings of a CNN remain unclear, which hinders further developments and improvements into this dire... | rejected-papers | The reviewers rightly point out that presented analysis is limiting and that the experimental results are not extensive enough. Moreover, several existing work that use raw waveforms have interesting analysis of what the network is trying to learn. Given these comments, the AC recommends that the paper be rejected.
| train | [
"B1DXXLUNG",
"r19T4gcgM",
"HydgKG5ez",
"rJ37LJolM",
"SkMmiXhXG",
"HyycjgXzz",
"B1wpueQGf",
"rJc5ue7fG",
"ByLRPe7fz",
"SJZsPxmGM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author",
"author"
] | [
"0. Noted, thanks for clarifying this.\n\n1. I'm still not convinced. You say that \"the processing of raw waveforms will also go through a similar time-feature representation\", but how do you know where that is? At which point in the network do you call something a \"time-feature representation\"? Isn't every lay... | [
-1,
3,
3,
2,
-1,
-1,
-1,
-1,
-1,
-1
] | [
-1,
4,
5,
5,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"SJZsPxmGM",
"iclr_2018_S1Ow_e-Rb",
"iclr_2018_S1Ow_e-Rb",
"iclr_2018_S1Ow_e-Rb",
"iclr_2018_S1Ow_e-Rb",
"r19T4gcgM",
"HydgKG5ez",
"HydgKG5ez",
"rJ37LJolM",
"rJ37LJolM"
] |
iclr_2018_BybQ7zWCb | “Style” Transfer for Musical Audio Using Multiple Time-Frequency Representations | Neural Style Transfer has become a popular technique for
generating images of distinct artistic styles using convolutional neural networks. This
recent success in image style transfer has raised the question of
whether similar methods can be leveraged to alter the “style” of musical
audio. In th... | rejected-papers | The paper extends an existing work with three different frequency representations of audios and necessary network structure modifications for music style transfer.
It is an interesting study but does not provide "sufficiently novel or justified contributions compared to the baseline approach of Ulyanov and Lebedev". Al... | train | [
"rJAGTitgG",
"B1Lbmvmef",
"rJ978GcgG",
"BJPoXdaXf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author"
] | [
"This paper describes improvements to a system described in a blog post for musical style transfer. Such a system is difficult to evaluate, but examples are presented where the style of one song is applied to the content of another. These audio examples show that the system produces somewhat reasonable mixtures o... | [
6,
4,
7,
-1
] | [
4,
4,
3,
-1
] | [
"iclr_2018_BybQ7zWCb",
"iclr_2018_BybQ7zWCb",
"iclr_2018_BybQ7zWCb",
"iclr_2018_BybQ7zWCb"
] |
iclr_2018_SJ60SbW0b | Modeling Latent Attention Within Neural Networks | Deep neural networks are able to solve tasks across a variety of domains and modalities of data. Despite many empirical successes, we lack the ability to clearly understand and interpret the learned mechanisms that contribute to such effective behaviors and more critically, failure modes. In this work, we present a gen... | rejected-papers | The proposed LAN provides a visualization of the selectivity of networks to its inputs. It takes a trained network as golden target and estimates an LAN to predict masks that can be applied on inputs to generate the same outputs.
But the significance of the proposed method is unclear, "what is the potential usage of th... | train | [
"Hy3t5U5gz",
"Sk3pyHoez",
"HkkF8qngG",
"r1nN7a4Qz",
"ByjI56eXG",
"rJakzaEmM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"The main contribution of the paper is to propose to learn a Latent Attention Network (LAN) that can help to visualize the inner structure of a deep neural network. To this end, the paper propose a novel training objective that can learn to tell the importance of each dimension of input. It is very interesting. How... | [
4,
5,
7,
-1,
-1,
-1
] | [
4,
4,
4,
-1,
-1,
-1
] | [
"iclr_2018_SJ60SbW0b",
"iclr_2018_SJ60SbW0b",
"iclr_2018_SJ60SbW0b",
"Sk3pyHoez",
"Hy3t5U5gz",
"HkkF8qngG"
] |
iclr_2018_rkmoiMbCb | Tandem Blocks in Deep Convolutional Neural Networks | Due to the success of residual networks (resnets) and related architectures, shortcut connections have quickly become standard tools for building convolutional neural networks. The explanations in the literature for the apparent effectiveness of shortcuts are varied and often contradictory. We hypothesize that shortcut... | rejected-papers | The paper presents a good analysis on the use of different linear maps instead of identity shortcuts for resnet.
It is interesting to the community but the experimental justification is insufficient.
1) As pointed out by the reviewer that this work shows "that on small size networks Tandem Block outperforms Residual Bl... | train | [
"BkWQMuIBz",
"rJyDBCYgG",
"ryUzS184G",
"Hk1CjNZyM",
"ry1gCAFlM",
"HkL83kfVz",
"HJ3dHdW4z",
"BJVy_t14M",
"HkLjGd6mM",
"B1m4fdpmz",
"Sy_eGOTXf",
"Byzn-uamG",
"BJKUbSHJf",
"BkYT2CzkM"
] | [
"author",
"official_reviewer",
"author",
"official_reviewer",
"official_reviewer",
"author",
"official_reviewer",
"public",
"author",
"author",
"author",
"author",
"author",
"official_reviewer"
] | [
"We appreciate that all of our reviewers responded to our updated paper and we are pleased to see that we managed to address nearly all of their questions and concerns. The only remaining criticisms regard the size of the networks in our experiments. While we were careful to recreate the meta-architecture of Zagoru... | [
-1,
5,
-1,
7,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
-1,
4,
-1,
4,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_rkmoiMbCb",
"iclr_2018_rkmoiMbCb",
"BJVy_t14M",
"iclr_2018_rkmoiMbCb",
"iclr_2018_rkmoiMbCb",
"BJVy_t14M",
"B1m4fdpmz",
"iclr_2018_rkmoiMbCb",
"iclr_2018_rkmoiMbCb",
"Hk1CjNZyM",
"rJyDBCYgG",
"ry1gCAFlM",
"BkYT2CzkM",
"iclr_2018_rkmoiMbCb"
] |
iclr_2018_S1PWi_lC- | Multi-task Learning on MNIST Image Datasets | We apply multi-task learning to image classification tasks on MNIST-like datasets. MNIST dataset has been referred to as the {\em drosophila} of machine learning and has been the testbed of many learning theories. The NotMNIST dataset and the FashionMNIST dataset have been created with the MNIST dataset as reference. I... | rejected-papers | the paper validates the benefit of multi-task learning on MNIST datasets, which is not sufficient for ICLR publication | train | [
"Sk-OB6wlM",
"rJPeuDOxM",
"B19pUEjlz",
"rkmQzUoQG",
"rkUS0Hj7f",
"Sy09XLsmf",
"rJgnUCmGM",
"rJVCZ7zzz",
"By4d_zGff",
"Hyo0zyzGz",
"HJdSRRnWz",
"ByLrDaeGf",
"HJLn9FM-z",
"S1za-QyZf",
"BJnd4ZTlf",
"SyveRxjlM",
"BJ-kwu9eM",
"SJV6JGqgM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"public",
"public",
"public",
"public",
"author",
"author",
"public",
"author",
"public",
"author",
"public"
] | [
"This paper presents a multi-task neural network for classification on MNIST-like datasets.\n\nThe main concern is that the technical innovation is limited. It is well known that multi-task learning can lead to performance improvement on similar tasks/datasets. This does not need to be verified in MNIST-like datase... | [
5,
4,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
4,
5,
3,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_S1PWi_lC-",
"iclr_2018_S1PWi_lC-",
"iclr_2018_S1PWi_lC-",
"Sk-OB6wlM",
"rJPeuDOxM",
"B19pUEjlz",
"Hyo0zyzGz",
"iclr_2018_S1PWi_lC-",
"iclr_2018_S1PWi_lC-",
"iclr_2018_S1PWi_lC-",
"iclr_2018_S1PWi_lC-",
"HJdSRRnWz",
"S1za-QyZf",
"BJ-kwu9eM",
"SyveRxjlM",
"BJ-kwu9eM",
"SJV6J... |
iclr_2018_B1p461b0W | Deep Learning is Robust to Massive Label Noise | Deep neural networks trained on large supervised datasets have led to impressive results in recent years. However, since well-annotated datasets can be prohibitively expensive and time-consuming to collect, recent work has explored the use of larger but noisy datasets that can be more easily obtained. In this paper, we... | rejected-papers | The paper studies the robustness of deep learning against label noise on MNIST, CIFAR-10 and ImageNet. But the generalization of the claim "deep learning is robust to massive label noise" is still questionable due to the limited noise types investigated.
The paper presents some tricks to improve learning with high labe... | test | [
"B1STvPVyG",
"HJTr9mqgM",
"BkC-glk-M",
"BktM0Ppmf",
"rJqJbCqXf",
"BkKLl097G",
"rkfFCmgzz",
"ByRhoHM1G",
"H1pqnCHCZ"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"public",
"author",
"public"
] | [
"The paper makes a bold claim, that deep neural networks are robust to arbitrary level of noise. It also implies that this would be true for any type of noise, and support this later claim using experiments on CIFAR and MNIST with three noise types: (1) uniform label noise (2) non-uniform but image-independent labe... | [
5,
4,
5,
-1,
-1,
-1,
-1,
-1,
-1
] | [
4,
5,
5,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_B1p461b0W",
"iclr_2018_B1p461b0W",
"iclr_2018_B1p461b0W",
"BkC-glk-M",
"B1STvPVyG",
"HJTr9mqgM",
"iclr_2018_B1p461b0W",
"H1pqnCHCZ",
"iclr_2018_B1p461b0W"
] |
iclr_2018_H1O0KGC6b | Post-training for Deep Learning | One of the main challenges of deep learning methods is the choice of an appropriate training strategy. In particular, additional steps, such as unsupervised pre-training, have been shown to greatly improve the performances of deep structures. In this article, we propose an extra training step, called post-training, whi... | rejected-papers | * the proposed fine-tuning of only the last layer is not novel enough
* experiments are not sufficient to isolate the differences to support the benefit of post-training
| train | [
"HyyQKdklf",
"ry63clwxz",
"H1zHTO2ef",
"H1orCErGz",
"HJL3TESff",
"HkrvrXMGz",
"r1QnTOhJf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"public",
"public"
] | [
"Summary: \nBased on ideas within the context of kernel theory, the authors consider post-training of NNs as an extra training step, which only optimizes the last layer of the network.\nThis additional step makes sure that the embedding, or representation, of the data is used in the best possible way for the consid... | [
5,
3,
4,
-1,
-1,
-1,
-1
] | [
4,
5,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_H1O0KGC6b",
"iclr_2018_H1O0KGC6b",
"iclr_2018_H1O0KGC6b",
"HkrvrXMGz",
"iclr_2018_H1O0KGC6b",
"iclr_2018_H1O0KGC6b",
"iclr_2018_H1O0KGC6b"
] |
iclr_2018_H1-oTz-Cb | Parametrizing filters of a CNN with a GAN | It is commonly agreed that the use of relevant invariances as a good statistical bias is important in machine-learning. However, most approaches that explicitely incorporate invariances into a model architecture only make use of very simple transformations, such as translations and rotations. Hence, there is a need for... | rejected-papers | The experiments are not sufficient to support the claim. The authors plan to improve it for future publication. | val | [
"ByXiDMPlG",
"Bk7nEndeG",
"HJoUOC5eM",
"rJsyhEaXM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author"
] | [
"The paper proposes an approach to learning a distribution over filters of a CNN. The method is based on a adversarial training: the generator produces filters, and the discriminator aims to distinguish the activation maps produced by real filters from those produced by the generated ones. \n\nPros:\n1) The general... | [
2,
4,
4,
-1
] | [
4,
4,
5,
-1
] | [
"iclr_2018_H1-oTz-Cb",
"iclr_2018_H1-oTz-Cb",
"iclr_2018_H1-oTz-Cb",
"iclr_2018_H1-oTz-Cb"
] |
iclr_2018_HkwrqtlR- | WHAT ARE GANS USEFUL FOR? | GANs have shown how deep neural networks can be used for generative modeling, aiming at achieving the same impact that they brought for discriminative modeling. The first results were impressive, GANs were shown to be able to generate samples in high dimensional structured spaces, like images and text, that were no cop... | rejected-papers | As the reviewers said, it is unclear what the main contribution of the paper is. | train | [
"rymXy2ggz",
"HkqCiwdlf",
"rJE3PW5gM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"The main take-away messages of this paper seem to be:\n\n1. GANs don't really match the target distribution. Some previous theory supports this, and some experiments are provided here demonstrating that the failure seems to be largely in under-sampling the tails, and sometimes perhaps in introducing spurious modes... | [
3,
3,
3
] | [
5,
4,
5
] | [
"iclr_2018_HkwrqtlR-",
"iclr_2018_HkwrqtlR-",
"iclr_2018_HkwrqtlR-"
] |
iclr_2018_BJQPG5lR- | Avoiding degradation in deep feed-forward networks by phasing out skip-connections | A widely observed phenomenon in deep learning is the degradation problem: increasing
the depth of a network leads to a decrease in performance on both test and training data. Novel architectures such as ResNets and Highway networks have addressed this issue by introducing various flavors of skip-connections or ga... | rejected-papers | Pros:
+ Interesting perspective on training deep networks
Cons:
- Not a lot of practical significance: why would one want to use this algorithm over standard methods like ResNets or highway networks given that the proposed algorithm is more complex than established methods?
| test | [
"B1LHtBsEM",
"ByISCAelf",
"HkX303_ez",
"SyaalQ9lM",
"HJOgTRn7f",
"HJe2nZnmf",
"S1Hrk2cQf",
"Bka78GFXG",
"r1-JIMK7z",
"BJERu9XXz",
"HyFpv57XG",
"H1C8Pcmmf",
"HyqZ1Io1z",
"HJaKBU9kG",
"HJV2WiFyf",
"BJiI_YZkG"
] | [
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"official_reviewer",
"author",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"official_reviewer",
"author",
"official_reviewer"
] | [
"Thank you for your feedback as well as for increasing the score of the paper. We respond to the remaining comments below.\n\nSimilarity to Savarese et al., (2016)\nWe thank the reviewer for suggesting this paper, which we will reference in an updated manuscript. As noted by the reviewer, the similarity between our... | [
-1,
6,
6,
6,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
-1,
4,
4,
5,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"SyaalQ9lM",
"iclr_2018_BJQPG5lR-",
"iclr_2018_BJQPG5lR-",
"iclr_2018_BJQPG5lR-",
"iclr_2018_BJQPG5lR-",
"S1Hrk2cQf",
"r1-JIMK7z",
"BJERu9XXz",
"BJERu9XXz",
"ByISCAelf",
"HkX303_ez",
"SyaalQ9lM",
"HJaKBU9kG",
"iclr_2018_BJQPG5lR-",
"BJiI_YZkG",
"iclr_2018_BJQPG5lR-"
] |
iclr_2018_B1D6ty-A- | Training Autoencoders by Alternating Minimization | We present DANTE, a novel method for training neural networks, in particular autoencoders, using the alternating minimization principle. DANTE provides a distinct perspective in lieu of traditional gradient-based backpropagation techniques commonly used to train deep networks. It utilizes an adaptation of quasi-convex ... | rejected-papers | Pros:
+ Interesting alternative algorithm for training autoencoders
Cons:
- Not a lot of practical value because DANTE does not outperform SGD in terms of time or classification performance using autoencoder features.
This is an interesting and well-written paper that doesn't quite meet the threshold for ICLR accepta... | train | [
"HJTgwzPVM",
"H1Q0pKwlf",
"HJJaJoveM",
"Syrb4g5gz",
"SyFN3vmNf",
"ry_SthoGG"
] | [
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author"
] | [
"We thank you for going through our revised draft, and sharing your concern. \n\nIn our setup for Theorem A.3, we have a single *multi-output* layer, so the label set given x is given by (y being a vector):\n\n y = [ max{0, cx}, max{0,dx} ]\n\nAssuming the setup mentioned in the comment with a,b,c and d as s... | [
-1,
6,
4,
7,
-1,
-1
] | [
-1,
4,
4,
5,
-1,
-1
] | [
"SyFN3vmNf",
"iclr_2018_B1D6ty-A-",
"iclr_2018_B1D6ty-A-",
"iclr_2018_B1D6ty-A-",
"ry_SthoGG",
"iclr_2018_B1D6ty-A-"
] |
iclr_2018_SyL9u-WA- | Stabilizing Gradients for Deep Neural Networks via Efficient SVD Parameterization | Vanishing and exploding gradients are two of the main obstacles in training deep neural networks, especially in capturing long range dependencies in recurrent neural networks (RNNs). In this paper, we present an efficient parametrization of the transition matrix of an RNN that allows us to stabilize the gradients that ... | rejected-papers | Pros:
+ Clearly written paper.
+ Good theoretical analysis of the expressivity of the proposed model.
+ Efficient model update is appealing.
+ Reviewers appreciated the addition of results on the copy and adding tasks in Appendix C.
Cons:
- Evaluation was on less-standard RNN tasks. A language modeling task should ha... | val | [
"Syi9ojdgf",
"Bkyxj89lM",
"H1Z5gRJZf",
"rk61wXVQM",
"BJttB7EXf",
"HJtJL7Emz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This paper proposed a new parametrization scheme for weight matrices in neural network based on the Householder reflectors to solve the gradient vanishing and exploding problems in training. The proposed method improved two previous papers:\n1) stronger expressive power than Mahammedi et al. (2017),\n2) faster gr... | [
7,
5,
5,
-1,
-1,
-1
] | [
4,
4,
3,
-1,
-1,
-1
] | [
"iclr_2018_SyL9u-WA-",
"iclr_2018_SyL9u-WA-",
"iclr_2018_SyL9u-WA-",
"H1Z5gRJZf",
"Bkyxj89lM",
"Syi9ojdgf"
] |
iclr_2018_B1twdMCab | Dynamic Integration of Background Knowledge in Neural NLU Systems | Common-sense or background knowledge is required to understand natural language, but in most neural natural language understanding (NLU) systems, the requisite background knowledge is indirectly acquired from static corpora. We develop a new reading architecture for the dynamic integration of explicit background knowle... | rejected-papers | Pros:
+ The paper is very clearly written.
+ The proposed re-embedding approach is easily implemented and can be integrated into fancier architectures.
Cons:
- A lot of the gains reported come from lemmatization, and the gains from background knowledge become marginal when used on a stronger baseline (e.g., ESIM with ... | train | [
"SkdyfcOlf",
"BJdxRnOlz",
"ByyFhLYgM",
"Bk8Ppr4Wf",
"Skxg6NE-M",
"HJzqfHVZz",
"B1DVGLEbG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"This paper proposes a model for adding background knowledge to natural language understanding tasks. The model reads the relevant text and then more assertions gathered from background knowledge before determining the final prediction. The authors show this leads to some improvement on multiple tasks like question... | [
5,
5,
6,
-1,
-1,
-1,
-1
] | [
3,
4,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_B1twdMCab",
"iclr_2018_B1twdMCab",
"iclr_2018_B1twdMCab",
"BJdxRnOlz",
"ByyFhLYgM",
"SkdyfcOlf",
"BJdxRnOlz"
] |
iclr_2018_SJ19eUg0- | BLOCK-DIAGONAL HESSIAN-FREE OPTIMIZATION FOR TRAINING NEURAL NETWORKS | Second-order methods for neural network optimization have several advantages over methods based on first-order gradient descent, including better scaling to large mini-batch sizes and fewer updates needed for convergence. But they are rarely applied to deep learning in practice because of high computational cost and th... | rejected-papers | Pros:
+ Clearly written paper.
Cons:
- Limited empirical evaluation: paper should compare to first-order methods with well-tuned hyperparameters, since the block Hessian-free hyperparameters likely were well tuned, and plots of convergence as a function of time need to be included.
- Somewhat limited novelty in that b... | train | [
"SyJaBw1eG",
"BkCQ4Ywgz",
"SJ1GzQKxz",
"ry6E_jXVz",
"BkI3GiqXz",
"H11iXicmz",
"rJ-_ZscQz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"public",
"author",
"author",
"author"
] | [
"Summary: \nThe paper considers second-order optimization methods for training of neural networks.\nIn particular, the contribution of the paper is a Hessian-free method that works on blocks of parameters (this is a user defined splitting of the parameters in blocks, e.g., parameters of each layer is one block, or ... | [
6,
4,
6,
-1,
-1,
-1,
-1
] | [
4,
4,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_SJ19eUg0-",
"iclr_2018_SJ19eUg0-",
"iclr_2018_SJ19eUg0-",
"iclr_2018_SJ19eUg0-",
"BkCQ4Ywgz",
"SyJaBw1eG",
"SJ1GzQKxz"
] |
iclr_2018_H1bhRHeA- | Unbiased scalable softmax optimization | Recent neural network and language models have begun to rely on softmax distributions with an extremely large number of categories. In this context calculating the softmax normalizing constant is prohibitively expensive. This has spurred a growing literature of efficiently computable but biased estimates of the softmax... | rejected-papers | The key concern from the reviewers that was not addressed is that none of the experimental results illustrate convergence vs. time instead of convergence vs. number of iterations. While the authors point out that their method is O(ND) instead of O(KND), the reviewers really wanted to see graphs demonstrating this, giv... | train | [
"BJ_UlF_lM",
"HJBXAQFeM",
"HJIPOSAbf",
"H1-65w37f",
"BkuezBxfz",
"SJCdbBezM",
"HygceHgMG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"The paper presents interesting algorithms for minimizing softmax with many classes. The objective function is a multi-class classification problem (using softmax loss) and with linear model. The main idea is to rewrite the obj as double-sum using the dual formulation and then apply SGD to solve it. At each iterati... | [
5,
5,
5,
-1,
-1,
-1,
-1
] | [
4,
3,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_H1bhRHeA-",
"iclr_2018_H1bhRHeA-",
"iclr_2018_H1bhRHeA-",
"iclr_2018_H1bhRHeA-",
"BJ_UlF_lM",
"HJBXAQFeM",
"HJIPOSAbf"
] |
iclr_2018_HyKZyYlRZ | Large Scale Multi-Domain Multi-Task Learning with MultiModel | Deep learning yields great results across many fields,
from speech recognition, image classification, to translation.
But for each problem, getting a deep model to work well involves
research into the architecture and a long period of tuning.
We present a single model that yields good res... | rejected-papers |
Pros:
+ Interesting and promising approach to multi-domain, multi-task learning.
+ Paper is clearly written.
Cons:
- Reads more like a technical report than a research paper: more space should be devoted to explaining the design decisions behind the model and the challenges involved, as this will help others tackle s... | train | [
"rJ6r_eqlf",
"rJEgR88Nf",
"Sy4tKnHNM",
"S1VMUmbeM",
"rke2Ms7ez",
"Syf0DbVNM",
"r1iTXuT7G",
"ry8Hwv67z",
"HyKMBPp7z"
] | [
"official_reviewer",
"official_reviewer",
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"The paper presents a multi-task, multi-domain model based on deep neural networks. The proposed model is able to take inputs from various domains (image, text, speech) and solves multiple tasks, such as image captioning, machine translation or speech recognition. The proposed model is composed of several features ... | [
6,
-1,
-1,
3,
6,
-1,
-1,
-1,
-1
] | [
3,
-1,
-1,
5,
4,
-1,
-1,
-1,
-1
] | [
"iclr_2018_HyKZyYlRZ",
"Sy4tKnHNM",
"Syf0DbVNM",
"iclr_2018_HyKZyYlRZ",
"iclr_2018_HyKZyYlRZ",
"r1iTXuT7G",
"S1VMUmbeM",
"rke2Ms7ez",
"rJ6r_eqlf"
] |
iclr_2018_rJ4uaX2aW | Large Batch Training of Convolutional Networks with Layer-wise Adaptive Rate Scaling | A common way to speed up training of large convolutional networks is to add computational units. Training is then performed using data-parallel synchronous Stochastic Gradient Descent (SGD) with a mini-batch divided between computational units. With an increase in the number of nodes, the batch size grows. However, t... | rejected-papers | Pros:
+ The proposed large-batch, synchronous SGD method is able to generalize at larger batch sizes than previous approaches (e.g., Goyal et al., 2017).
Cons:
- Evaluation on more than one task would make the paper more convincing.
- The addition of more hyperparameters makes the proposed algorithm less appealing.
- ... | train | [
"ry33G5FxM",
"Sk-rQ7qxf",
"S1b88I5xM",
"BJLfcihbz",
"r1rHHihWG",
"S1J7jF2-z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"This paper provides an optimization approach for large batch training of CNN with layer-wise adaptive learning rates. \nIt starts from the observation that the ratio between the L2-norm of parameters and that of gradients on parameters varies\nsignificantly in the optimization, and then introduce a local learning... | [
5,
4,
5,
-1,
-1,
-1
] | [
3,
4,
5,
-1,
-1,
-1
] | [
"iclr_2018_rJ4uaX2aW",
"iclr_2018_rJ4uaX2aW",
"iclr_2018_rJ4uaX2aW",
"ry33G5FxM",
"Sk-rQ7qxf",
"S1b88I5xM"
] |
iclr_2018_HJSA_e1AW | Normalized Direction-preserving Adam | Optimization algorithms for training deep models not only affects the convergence rate and stability of the training process, but are also highly related to the generalization performance of trained models. While adaptive algorithms, such as Adam and RMSprop, have shown better optimization performance than stochastic g... | rejected-papers | The paper proposes a modification to Adam which is intended to ensure that the direction of weight update lies in the span of the historical gradients and to ensure that the effective learning rate does not decrease as the magnitudes of the weights increase. The reviewers wanted a clearer justification of the changes ... | test | [
"BktDDeo4z",
"Bk_mQkcgM",
"S1-Kfe5lM",
"SJ_kd7ixf",
"HJq1ofsmf",
"HJy8rIyGM",
"ryVoBIJMG",
"Hk71lIyMG",
"SkCcbLkGf",
"rkEceLJzG",
"HkcU1LJfM",
"Bks3kZqgM",
"SJ9yMkqlf",
"r1864VYlG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author",
"author",
"author",
"author",
"author",
"public"
] | [
"Thank the authors for their response. I am disappointed the latest revised paper does not provide any further insight into the ND-Adam updates. It is with much regret that my score remains the same. ",
"Method:\n\nThe paper is missing analysis of some important related works such as\n\n\"Beyond convexity: Stocha... | [
-1,
5,
5,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
-1,
4,
5,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"HkcU1LJfM",
"iclr_2018_HJSA_e1AW",
"iclr_2018_HJSA_e1AW",
"iclr_2018_HJSA_e1AW",
"iclr_2018_HJSA_e1AW",
"Bk_mQkcgM",
"Bk_mQkcgM",
"S1-Kfe5lM",
"Bk_mQkcgM",
"Bk_mQkcgM",
"SJ_kd7ixf",
"iclr_2018_HJSA_e1AW",
"r1864VYlG",
"iclr_2018_HJSA_e1AW"
] |
iclr_2018_SybqeKgA- | On Batch Adaptive Training for Deep Learning: Lower Loss and Larger Step Size | Mini-batch gradient descent and its variants are commonly used in deep learning. The principle of mini-batch gradient descent is to use noisy gradient calculated on a batch to estimate the real gradient, thus balancing the computation cost per iteration and the uncertainty of noisy gradient. However, its batch size is ... | rejected-papers | The reviewers generally thought the proposed algorithm was a straightforward extension of Yin et al., 2017, and not enough for a new paper. They also objected to a lack of test results (to show generalization), but the authors did provide these in their revision.
Pros:
+ Adaptive batch sizing is useful, especially if... | train | [
"HkvBT03Jf",
"HktgWy7xM",
"r1b1AtYlG",
"SyW3N3jQf",
"rJkjw8sMz",
"Byg8vUiMM",
"H1ilDIiMG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author"
] | [
"The authors propose extending the recently-proposed adaptive batch-size approach of Yin et al. to an update that includes momentum, and perform more comprehensive experiments than in the Yin et al. paper validating their approach.\n\nThe basic idea makes a great deal of intuitive sense: inaccurate gradient estimat... | [
5,
5,
4,
-1,
-1,
-1,
-1
] | [
3,
4,
3,
-1,
-1,
-1,
-1
] | [
"iclr_2018_SybqeKgA-",
"iclr_2018_SybqeKgA-",
"iclr_2018_SybqeKgA-",
"iclr_2018_SybqeKgA-",
"HkvBT03Jf",
"HktgWy7xM",
"r1b1AtYlG"
] |
iclr_2018_H1A5ztj3b | Super-Convergence: Very Fast Training of Residual Networks Using Large Learning Rates | In this paper, we show a phenomenon, which we named ``super-convergence'', where residual networks can be trained using an order of magnitude fewer iterations than is used with standard training methods. The existence of super-convergence is relevant to understanding why deep networks generalize well. One of the key... | rejected-papers | The paper reports unusally rapid convergence of the ResNet-56 model on CIFAR-10 when a single cycle of a cyclic learning rate schedule is used. The effect is analyzed from several different perspectives. However, the reviewers were not convinced because the effect is only observed for one task, so they question the si... | train | [
"rJfAp3Zef",
"H1yQ04YxG",
"Hyn-NPJbz",
"ByWgFx2Mf",
"HkvgXwlfM",
"SyXNU83Wf",
"r1J5783bz",
"rJJuxtxeG",
"H1Df4xU0Z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"public",
"public",
"author",
"author",
"author",
"public"
] | [
"The paper discusses a phenomenon where neural network training in very specific settings can profit much from a schedule including large learning rates. Unfortunately, this paper feels to be hastily written and can only be read when accompanied with several references as key parts (CLR) are not described and thus ... | [
4,
4,
4,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
4,
3,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_H1A5ztj3b",
"iclr_2018_H1A5ztj3b",
"iclr_2018_H1A5ztj3b",
"H1yQ04YxG",
"iclr_2018_H1A5ztj3b",
"iclr_2018_H1A5ztj3b",
"iclr_2018_H1A5ztj3b",
"iclr_2018_H1A5ztj3b",
"iclr_2018_H1A5ztj3b"
] |
iclr_2018_ByJ7obb0b | Understanding and Exploiting the Low-Rank Structure of Deep Networks | Training methods for deep networks are primarily variants on stochastic gradient descent. Techniques that use (approximate) second-order information are rarely used because of the computational cost and noise associated with those approaches in deep learning contexts. However, in this paper, we show how feedforward d... | rejected-papers | The reviewers thought that idea of trying to exploit low-rank structure in the loss gradients of a feedforward network to improve training was interesting; however they expressed many concerns about the clarity of the presentation, quality of the empirical evaluation, and significance of the result (since the tests wer... | train | [
"H1b4hwrxf",
"HyyKwSvlG",
"Hk22fOybz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"[Main comments]\n\n* The authors made a really odd choice of notation, which made the equations hard to follow.\nApparently, that notation is used in differential geometry, but I have never seen it used in\nan ML paper. If you talk about outer product structure, show some outer products!\n\n* The function f that t... | [
4,
5,
2
] | [
4,
4,
4
] | [
"iclr_2018_ByJ7obb0b",
"iclr_2018_ByJ7obb0b",
"iclr_2018_ByJ7obb0b"
] |
iclr_2018_HJg1NTGZRZ | Bit-Regularized Optimization of Neural Nets | We present a novel regularization strategy for training neural networks which we call ``BitNet''. The parameters of neural networks are usually unconstrained and have a dynamic range dispersed over a real valued range. Our key idea is to control the expressive power of the network by dynamically quantizing the range an... | rejected-papers | Pros:
+ The idea of end-to-end training that simultaneously learns the weights and appropriate precision for those weights is very appealing.
Cons:
- Experimental results are far from the state-of-the-art, which makes the empirical evaluation unconvincing.
- More justification is needed for the update of the number of... | train | [
"B1qKE_Tmz",
"HkrrVdTQG",
"B1S1NdT7G",
"HkGFQdT7z",
"Syi6-muxf",
"Sy4v3Bolz",
"Bk1-V2plG",
"B1jyA87EG",
"B1IKaxXbf",
"H12bJi4eM",
"BkchXu6yG"
] | [
"author",
"author",
"author",
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"public"
] | [
"1- Only small networks on relatively small datasets are tested. \n>The results on VGG networks (larger networks) is being computed and will be included in the camera ready submission. \n\n2-The results on MNIST and CIFAR-10 are not good enough...\n>We found that our low performance on MNIST was caused by using 4X4... | [
-1,
-1,
-1,
-1,
4,
3,
4,
-1,
-1,
-1,
-1
] | [
-1,
-1,
-1,
-1,
5,
4,
4,
-1,
-1,
-1,
-1
] | [
"Syi6-muxf",
"Sy4v3Bolz",
"Bk1-V2plG",
"iclr_2018_HJg1NTGZRZ",
"iclr_2018_HJg1NTGZRZ",
"iclr_2018_HJg1NTGZRZ",
"iclr_2018_HJg1NTGZRZ",
"Syi6-muxf",
"Sy4v3Bolz",
"BkchXu6yG",
"iclr_2018_HJg1NTGZRZ"
] |
iclr_2018_r1Kr3TyAb | ANALYSIS ON GRADIENT PROPAGATION IN BATCH NORMALIZED RESIDUAL NETWORKS | We conduct a mathematical analysis on the Batch normalization (BN) effect on gradient backpropagation in residual network training in this work, which is believed to play a critical role in addressing the gradient vanishing/explosion problem. Specifically, by analyzing the mean and variance behavior of the input and th... | rejected-papers | Two of the reviewers liked the intent of the paper -- to analyze gradient flow in residual networks and understand the tradeoffs between width and depth in such networks. However, all reviewers flagged a number of problems in the paper, and the authors did not participate in the discussion period.
Pros:
+ Interesting... | test | [
"rkZAtAaxM",
"B1BeKAgyz",
"H1oD2H5xG"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This manuscript is fairly well-written, and discusses how the batch normalization step helps to stabilize the scale of the gradients. Intriguingly, the analysis suggests that using a shallower but wider resnet should provide competitive performance, which is supported by empirical evidence. This work should help... | [
4,
1,
4
] | [
4,
5,
5
] | [
"iclr_2018_r1Kr3TyAb",
"iclr_2018_r1Kr3TyAb",
"iclr_2018_r1Kr3TyAb"
] |
iclr_2018_rJUBryZ0W | Lifelong Learning by Adjusting Priors | In representational lifelong learning an agent aims to continually learn to solve novel tasks while updating its representation in light of previous tasks. Under the assumption that future tasks are related to previous tasks, representations should be learned in such a way that they capture the common structure across ... | rejected-papers | The author's revisions addressed clarity issues and some experimental issues (e.g., including MAML results in the comparison). The work takes an original path to an important problem (transfer learning, essentially). There is a question of significance, and this is due to the fact that the empirical comparisons are sti... | train | [
"SJAvHP8BM",
"rJX9THLHf",
"HyJlQlQWf",
"HyUIRnVSG",
"r1IHf4oVf",
"H1qBI-5gG",
"B1l2Xr8Ef",
"HyY2Sogef",
"H1LX97Czz",
"H1zptXAzM",
"r1-wtXRzz"
] | [
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author"
] | [
"We thank the area chair for the helpful comment.\nIndeed there was a problem with the constant, please see our response to AnonReviewer1.\n \nP.S.\nSince the submission of the revised paper we added more experiments that demonstrate the meta-learning performance in varied task-environments and with different numb... | [
-1,
-1,
6,
-1,
-1,
6,
-1,
6,
-1,
-1,
-1
] | [
-1,
-1,
4,
-1,
-1,
4,
-1,
4,
-1,
-1,
-1
] | [
"rJX9THLHf",
"iclr_2018_rJUBryZ0W",
"iclr_2018_rJUBryZ0W",
"H1LX97Czz",
"B1l2Xr8Ef",
"iclr_2018_rJUBryZ0W",
"H1qBI-5gG",
"iclr_2018_rJUBryZ0W",
"HyJlQlQWf",
"H1qBI-5gG",
"HyY2Sogef"
] |
iclr_2018_SkPoRg10b | Rethinking generalization requires revisiting old ideas: statistical mechanics approaches and complex learning behavior | We describe an approach to understand the peculiar and counterintuitive generalization properties of deep neural networks. The approach involves going beyond worst-case theoretical capacity control frameworks that have been popular in machine learning in recent years to revisit old ideas in the statistical mechanics o... | rejected-papers | The concerns raised by AnonReviewer3 point out that, despite the effort of the authors to bridge the SM / ML divide, there is still some work to be done. The gulf between thermodynamic limits and finite effects is oft-cited in the author response. This seems to be a catch all. This gap needs to be addressed early. The ... | train | [
"HJdaJu51z",
"S1IjqJvxf",
"SyBLHu_gM",
"r1bcRWlNf",
"BJM0h8dZz",
"SJ_fu6dZz",
"S16pvadWM",
"ryU4P6_-M",
"ByHfw6dbM",
"S1pJwadbG",
"SkXpUTuZM",
"ryQF86Obz"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author",
"author",
"author",
"author",
"author"
] | [
"The authors suggest that ideas from statistical mechanics will help to understand the \"peculiar and counterintuitive generalization properties of deep neural networks.\" The paper's key claim (from the abstract) is that their approach \"provides a strong qualitative description of recently-observed empirical resu... | [
3,
6,
7,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
3,
5,
4,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_SkPoRg10b",
"iclr_2018_SkPoRg10b",
"iclr_2018_SkPoRg10b",
"HJdaJu51z",
"iclr_2018_SkPoRg10b",
"HJdaJu51z",
"HJdaJu51z",
"HJdaJu51z",
"HJdaJu51z",
"HJdaJu51z",
"HJdaJu51z",
"HJdaJu51z"
] |
iclr_2018_ByJWeR1AW | Data augmentation instead of explicit regularization | Modern deep artificial neural networks have achieved impressive results through models with very large capacity---compared to the number of training examples---that control overfitting with the help of different forms of regularization. Regularization can be implicit, as is the case of stochastic gradient descent or pa... | rejected-papers | The reviewers agree that the authors have made an interesting contribution studying the effect of data augmentation, but they also agree that the claims made by the paper require a broader empirical study beyond the limited number of tasks surveyed in the current revision. I urge the authors to follow this advice and ... | train | [
"S1KIF7olf",
"BJHwtGogM",
"rJrLHq3yz",
"ryojlc5Xz",
"SkzKEY9QM",
"B1FxNt97G",
"S1U2QK5mG",
"BJvBW-rXM"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer",
"author",
"author",
"author",
"author",
"author"
] | [
"This paper presents an empirical study of whether data augmentation can be a substitute for explicit regularization of weight decay and dropout. It is a well written and well organized paper. However, overall I do not find the authors’ premises and conclusions to be well supported by the results and would sugges... | [
5,
5,
5,
-1,
-1,
-1,
-1,
-1
] | [
4,
4,
4,
-1,
-1,
-1,
-1,
-1
] | [
"iclr_2018_ByJWeR1AW",
"iclr_2018_ByJWeR1AW",
"iclr_2018_ByJWeR1AW",
"iclr_2018_ByJWeR1AW",
"rJrLHq3yz",
"BJHwtGogM",
"BJHwtGogM",
"S1KIF7olf"
] |
iclr_2018_ByED-X-0W | Parametric Information Bottleneck to Optimize Stochastic Neural Networks | In this paper, we present a layer-wise learning of stochastic neural networks (SNNs) in an information-theoretic perspective. In each layer of an SNN, the compression and the relevance are defined to quantify the amount of information that the layer contains about the input space and the target space, respectively. We ... | rejected-papers | The reviewers are in agreement that while the paper is interesting, both the clarity of presentation and experimental rigor could be improved. The committee feels this paper is not ready for publication at ICLR 2018 inits current form. | train | [
"r1D1-BKgz",
"rJYxl_YxG",
"SJcOWb5gf"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"This paper proposes a learning method (PIB) based on the information bottleneck framework.\nPIB pursues the very natural intuition outlined in the information bottleneck literature: hidden layers of deep nets compress the input X while maintaining sufficient information to predict the output Y.\nIt should be noted... | [
4,
6,
4
] | [
4,
4,
4
] | [
"iclr_2018_ByED-X-0W",
"iclr_2018_ByED-X-0W",
"iclr_2018_ByED-X-0W"
] |
iclr_2018_HJZiRkZC- | Byte-Level Recursive Convolutional Auto-Encoder for Text | This article proposes to auto-encode text at byte-level using convolutional networks with a recursive architecture. The motivation is to explore whether it is possible to have scalable and homogeneous text generation at byte-level in a non-sequential fashion through the simple task of auto-encoding. We show that non-se... | rejected-papers | This paper presents a method for using byte level convolutional networks for building text-based autoencoders. They show that these models do well compared to RNN-based methods which model text in a sequence. Evaluation is solely based on byte level prediction error. The committee feels that the paper would have be... | train | [
"BJuY5u9gf",
"BkLqBW6lG",
"ryQsImE-z"
] | [
"official_reviewer",
"official_reviewer",
"official_reviewer"
] | [
"\nThis paper presents a convolutional auto-encoder architecture for text encoding and generation. It works on the character level and contains a recursive structure which scales with the length of the input text. Building on the recent state-of-the-art in terms of architectural components, the paper shows the feas... | [
7,
5,
5
] | [
4,
3,
5
] | [
"iclr_2018_HJZiRkZC-",
"iclr_2018_HJZiRkZC-",
"iclr_2018_HJZiRkZC-"
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.