paper_id stringlengths 9 12 | title stringlengths 10 183 | venue stringclasses 2
values | year int16 2.02k 2.03k | split stringclasses 1
value | in_eval_subset bool 1
class | abstract stringlengths 3 3.6k | num_reviews int8 1 7 | reviews listlengths 1 7 | review_template stringclasses 7
values | novelty_excerpts listlengths 0 77 | novelty_extraction_raw stringlengths 24 3.64k ⌀ | novelty_score int8 -1 2 ⌀ | novelty_score_justification stringlengths 143 1.93k ⌀ | novelty_detailed_assessment stringlengths 99 1.59k ⌀ | novelty_summary stringlengths 860 2.24k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
-0Cjhnl-dhK | Towards Uncertainties in Deep Learning that Are Accurate and Calibrated | ICLR | 2,022 | train | false | Predictive uncertainties can be characterized by two properties---calibration and sharpness. This paper introduces algorithms that ensure the calibration of any model while maintaining sharpness. They apply in both classification and regression and guarantee the strong property of distribution calibration, while being ... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors propose to decompose different common loss functions into calibration and refinement terms and show theoretically that under some (strong) assumptions, such as universal approximation, models can become perfectly calibrated without sacrificing p... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Unclear / small novelty compared to previous work",
"Performance in higher dimensions: The current method seems to only work well in very low-dimensional settings (one to two dimensions or so).",
"Empirical results: The current results are missing error bars, which makes it impossible to assess their statistic... | {
"novelty_excerpts": [
"Unclear / small novelty compared to previous work",
"Performance in higher dimensions: The current method seems to only work well in very low-dimensional settings (one to two dimensions or so).",
"Empirical results: The current results are missing error bars, which makes it imposs... | -1 | The paper introduces a recalibration algorithm that builds upon existing methods with minimal innovation. The work presents a constrained optimization framework that, while theoretically sound, lacks significant novelty in its approach. The contribution lies in its application to specific tasks, but the method's limita... | The paper proposes a recalibration algorithm that decomposes common loss functions into calibration and refinement terms, but this idea is not novel. The method works well in low-dimensional settings, but its performance in higher dimensions remains unclear. The work extends existing methods by introducing a constraine... | Novelty Score: -1
Score Justification:
The paper introduces a recalibration algorithm that builds upon existing methods with minimal innovation. The work presents a constrained optimization framework that, while theoretically sound, lacks significant novelty in its approach. The contribution lies in its application to... |
-29uFS4FiDZ | Word Sense Induction with Knowledge Distillation from BERT | ICLR | 2,022 | train | false | Pre-trained contextual language models are ubiquitously employed for language understanding tasks, but are unsuitable for resource-constrained systems. Noncontextual word embeddings are an efficient alternative in these settings. Such methods typically use one vector to encode multiple different meanings of a word, an... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a two-stage method to distill word sense knowledge from BERT into a skip-gram-like framework. The authors show that the proposed method performs well on sense selection and induction (or representation). They also verify the efficacy on t... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"One of the main weaknesses is that the results are not significant enough.",
"In Table 3, the performance of the proposed model on both criteria is worse than the existing models.",
"And the paper didn't compare with all existing sense models (see missing citations).",
"In Figure 2, most of the improvements ... | {
"novelty_excerpts": [
"One of the main weaknesses is that the results are not significant enough.",
"In Table 3, the performance of the proposed model on both criteria is worse than the existing models.",
"And the paper didn't compare with all existing sense models (see missing citations).",
"In Fig... | -1 | The paper introduces a straightforward approach to word sense induction using BERT, though the methodological novelty is low due to its similarity to existing approaches. The contributions are largely incremental, building upon existing techniques without substantial conceptual or technical innovation. | The paper presents a two-stage method for distilling word sense knowledge from BERT, which demonstrates state-of-the-art results. However, the approach is not novel, as it relies on existing techniques such as attention mechanisms and knowledge distillation. The work extends existing methods for word sense induction, b... | Novelty Score: -1
Score Justification:
The paper introduces a straightforward approach to word sense induction using BERT, though the methodological novelty is low due to its similarity to existing approaches. The contributions are largely incremental, building upon existing techniques without substantial conceptual o... |
-3yxxvDis3L | How to Improve Sample Complexity of SGD over Highly Dependent Data? | ICLR | 2,022 | train | false | Conventional machine learning applications typically assume that data samples are independently and identically distributed (i.i.d.). However, many practical scenarios naturally involve a data-generating process that produces highly dependent data samples, which are known to heavily bias the stochastic optimization pro... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis work analyzes and provides high probability convergence guarantee for periodically subsampled and minibatch SGD for optimization of convex and Lipschitz continuous function when the data stream is $\\phi$-mixing.\n\n## Main Review\nStrength:\n\nAnalyze... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"There is no novelty here.",
"almost no theoretical novelty",
"The contributions are neither significant nor novel.",
"The contributions are only marginally significant or novel."
] | {
"novelty_excerpts": [
"There is no novelty here.",
"almost no theoretical novelty",
"The contributions are neither significant nor novel.",
"The contributions are only marginally significant or novel."
]
} | 0 | The paper presents a competent analysis of SGD over dependent data, but its contributions remain largely incremental. The work builds on existing results and frameworks, providing some interesting insights but not fundamentally new ideas. The method extends existing approaches to subsampled and minibatch SGD, though th... | The paper introduces a framework for analyzing SGD over dependent data, leveraging the phi-mixing process to model data dependence. However, the analysis largely relies on existing results and frameworks, such as those presented in (Agarwal & Duchi, 2012). The work provides some interesting insights into the performanc... | Novelty Score: 0
Score Justification:
The paper presents a competent analysis of SGD over dependent data, but its contributions remain largely incremental. The work builds on existing results and frameworks, providing some interesting insights but not fundamentally new ideas. The method extends existing approaches to ... |
-6me0AsJVdu | Model Validation Using Mutated Training Labels: An Exploratory Study | ICLR | 2,022 | train | false | For out-of-sample validation, the sample set may be too small to be representative of the data distribution; the accuracy can have a large variance across different runs; excessive reuse of a fixed set of samples can lead to overfitting even if the samples are held out and not used in the training process. This paper i... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper introduces new approach for validation of ML models which is based on top of ideas imported from software development & QA (so-called mutation and metamorphic techniques). The core claimed advantage consist in the ability to avoid cross-validatio... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"A novel approach in the area of validation techniques, which are very conservative",
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prior works.",
"This work represents a first empirical study in this new direction, and a lot of ... | {
"novelty_excerpts": [
"A novel approach in the area of validation techniques, which are very conservative",
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prior works.",
"This work represents a first empirical study in t... | 1 | The paper introduces a novel approach to model validation, building on ideas from software development and QA. This approach presents a new direction in validation techniques, which are conservative in nature. However, the method's effectiveness and applicability are limited by its dependence on specific evaluation met... | The paper presents a moderately novel approach to model validation, combining mutation and metamorphic techniques from software development and QA. The work extends existing validation methods by introducing a new metric for evaluating model performance, which is based on the difference in training accuracy between mod... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to model validation, building on ideas from software development and QA. This approach presents a new direction in validation techniques, which are conservative in nature. However, the method's effectiveness and applicability are limited by it... |
-70L8lpp9DF | Hyperparameter Tuning with Renyi Differential Privacy | ICLR | 2,022 | train | false | For many differentially private algorithms, such as the prominent noisy stochastic gradient descent (DP-SGD), the analysis needed to bound the privacy leakage of a single training run is well understood. However, few studies have reasoned about the privacy leakage resulting from the multiple training runs needed to fin... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors make the following contributions regarding differentially private hyperparameter tuning :\n* As an example, the authors train an SVM with a weight penalty and show that, in presence of an outlier, a membership inference attack can be employed to... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors make the following contributions regarding differentially private hyperparameter tuning :",
"The authors propose a way to measure the utility of the algorithms by looking at the expected quantile of the output.",
"The results of their utility analysis, coupled with an experiment on MNIST, show that... | {
"novelty_excerpts": [
"The authors make the following contributions regarding differentially private hyperparameter tuning :",
"The authors propose a way to measure the utility of the algorithms by looking at the expected quantile of the output.",
"The results of their utility analysis, coupled with an experiment on ... | 1 | The paper presents moderately novel contributions to the field of differentially private hyperparameter tuning. It builds upon existing ideas and provides new insights into the problem formulation and methodological innovations. The work extends and improves upon prior approaches, offering a more comprehensive understa... | The paper introduces a novel approach to private hyperparameter tuning, which combines existing methods and provides a more general framework for analyzing the trade-offs between privacy and utility. The work extends the results of Liu and Talwar (2019) and provides a more comprehensive understanding of the problem for... | Novelty Score: 1
Score Justification:
The paper presents moderately novel contributions to the field of differentially private hyperparameter tuning. It builds upon existing ideas and provides new insights into the problem formulation and methodological innovations. The work extends and improves upon prior approaches,... |
-7UeX2KPqs | State-Action Joint Regularized Implicit Policy for Offline Reinforcement Learning | ICLR | 2,022 | train | false | Offline reinforcement learning enables learning from a fixed dataset, without further interactions with the environment. The lack of environmental interactions makes the policy training vulnerable to state-action pairs far from the training dataset and prone to missing rewarding actions. For training more effective age... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nOne of the existing Offline RL algorithms is to constrain the learned policy, such as constraining the learned policy to be consistent with the behavior policy itself or the action distribution based on state conditions, or adopting a Gaussian policy. Howev... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The technical novelty is sufficient and significant.",
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"I believe the most valuable experiments are the toy experiments to illustrate their motivations as the findings in the toy experiments may help t... | {
"novelty_excerpts": [
"The technical novelty is sufficient and significant.",
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"I believe the most valuable experiments are the toy experiments to illustrate their motivations as the findings i... | 1 | The paper introduces a novel approach to offline reinforcement learning by leveraging implicit policies and state-action visitation matching. This method extends existing work by addressing the limitations of Gaussian policies and behavior policies. However, the contributions remain largely incremental, building upon e... | The paper presents a framework that supports learning a flexible while well-regularized policy in offline RL. The method uses an implicit policy to capture the multi-modality of the action value function and controls the state-action access frequency in the data set to be consistent with the learned policy. This approa... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to offline reinforcement learning by leveraging implicit policies and state-action visitation matching. This method extends existing work by addressing the limitations of Gaussian policies and behavior policies. However, the contributions rema... |
-8sBpe7rDiV | NETWORK INSENSITIVITY TO PARAMETER NOISE VIA PARAMETER ATTACK DURING TRAINING | ICLR | 2,022 | train | false | Neuromorphic neural network processors, in the form of compute-in-memory crossbar arrays of memristors, or in the form of subthreshold analog and mixed-signal ASICs, promise enormous advantages in compute density and energy efficiency for NN-based ML tasks. However, these technologies are prone to computational non-ide... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nAuthors investigate the problem of network robustness in the presence of parameter variations. To that end, authors propose an adversarial parameter perturbation based robust training method. Proposed training approach iteratively performs adversarial attac... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"My main concern relates to the ambiguity on explicitly putting the novelty of this paper as the proposed problem is recently addressed by similar studies.",
"In fact, adversarial parameter perturbations for regularization is an approach that was demonstrated in the context of model generalization with similar re... | {
"novelty_excerpts": [
"My main concern relates to the ambiguity on explicitly putting the novelty of this paper as the proposed problem is recently addressed by similar studies.",
"In fact, adversarial parameter perturbations for regularization is an approach that was demonstrated in the context of model ge... | 1 | The paper introduces a moderately novel approach to model regularization through adversarial parameter perturbations, successfully applied to recurrent SNNs. The work presents a competent extension of existing ideas in adversarial training, but its novelty is somewhat limited by the fact that similar methods have been ... | The paper introduces a method for adversarial parameter perturbation based model regularization, which is successfully applied to recurrent SNNs. This approach builds on existing ideas in adversarial training, but its novelty is somewhat limited by the fact that similar methods have been proposed recently. The work ext... | Novelty Score: 1
Score Justification:
The paper introduces a moderately novel approach to model regularization through adversarial parameter perturbations, successfully applied to recurrent SNNs. The work presents a competent extension of existing ideas in adversarial training, but its novelty is somewhat limited by t... |
-9uy3c7b_ks | Learning Controllable Elements Oriented Representations for Reinforcement Learning | ICLR | 2,022 | train | false | Deep Reinforcement Learning (deep RL) has been successfully applied to solve various decision-making problems in recent years. However, the observations in many real-world tasks are often high dimensional and include much task-irrelevant information, limiting the applications of RL algorithms. To tackle this problem, w... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThere has been a school thought lately in RL that good representations for control should focus on abstracting out on only the task-relevant features. In high-dimensional tasks, this is especially useful in making optimal use of a limited neural network cap... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"the idea of using controllability to inform compression of state representations -- after all, what is important to the agent is what it can control about the environment.",
"Perhaps the most important situation is where there are non-controllable task-informative features.",
"The authors do point out one exam... | {
"novelty_excerpts": [
"the idea of using controllability to inform compression of state representations -- after all, what is important to the agent is what it can control about the environment.",
"Perhaps the most important situation is where there are non-controllable task-informative features.",
"The... | 0 | The paper presents a representation learning framework that builds on existing ideas in mutual information and controllability, but the contributions remain largely incremental. The approach combines and extends previous work, introducing a new objective function and experimental results that demonstrate its effectiven... | The paper introduces a novel representation learning framework, LCER, which enables the agent to extract task-relevant representations from image-based observations. The work presents a combination of mutual information and controllability to guide representation learning, extending previous approaches like CURL and PI... | Novelty Score: 0
Score Justification:
The paper presents a representation learning framework that builds on existing ideas in mutual information and controllability, but the contributions remain largely incremental. The approach combines and extends previous work, introducing a new objective function and experimental ... |
-BTmxCddppP | Revisiting Out-of-Distribution Detection: A Simple Baseline is Surprisingly Effective | ICLR | 2,022 | train | false | It is an important problem in trustworthy machine learning to recognize out-of-distribution (OOD) inputs which are inputs unrelated to the in-distribution task. Many out-of-distribution detection methods have been suggested in recent years. The goal of this paper is to recognize common objectives as well as to identify... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper analyzes different OOD detection methods and show that even if the formulation for many OOD methods were different, the binary discrimination is equivalent to those different types of methods when the rankings induced by Bayes optimal classifier a... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The paper provides insights and findings of properties of different OOD detection approaches.",
"Most of the previous work were mainly empirical and lacked mathematical foundation",
"The finding of uncertainty based methods being equivalent under the hood with binary discriminators is interesting.",
"It expl... | {
"novelty_excerpts": [
"The paper provides insights and findings of properties of different OOD detection approaches.",
"Most of the previous work were mainly empirical and lacked mathematical foundation",
"The finding of uncertainty based methods being equivalent under the hood with binary discriminator... | 1 | The paper introduces a novel connection between uncertainty-based methods and binary discriminators, providing insights into the properties of different OOD detection approaches. However, the contributions are somewhat limited by their focus on a specific set of methods and datasets, and the work builds upon existing i... | The paper presents a comprehensive analysis of existing OOD detection methods, highlighting their similarities and differences through mathematical derivations and empirical comparisons. The work extends the understanding of OOD detection by showing that uncertainty-based methods are equivalent to binary discriminators... | Novelty Score: 1
Score Justification:
The paper introduces a novel connection between uncertainty-based methods and binary discriminators, providing insights into the properties of different OOD detection approaches. However, the contributions are somewhat limited by their focus on a specific set of methods and datase... |
-Nf6TikpjQ | Multi-agent Performative Prediction: From Global Stability and Optimality to Chaos | ICLR | 2,022 | train | false | The recent framework of performative prediction is aimed at capturing settings where predictions influence the target/outcome they try to predict. In this paper, we introduce a natural multi-agent version of this framework, where multiple decision makers try to predict the same outcome. We showcase that such competiti... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studied the different behaviours of using exponentiated gradient descent (Def. 2.3) in linear regression with different learning rates. The setting is called performative prediction, which can be viewed as a special case of reinforcement learning... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"As claimed, this is the first time 'where such formal chaotic results are established in settings related to performative prediction and supervised learning more generally'.",
"However, performative prediction itself is also a special kind of game / reinforcement learning, which makes the contribution seem not v... | {
"novelty_excerpts": [
"As claimed, this is the first time 'where such formal chaotic results are established in settings related to performative prediction and supervised learning more generally'.",
"However, performative prediction itself is also a special kind of game / reinforcement learning, which makes... | 1 | The paper presents moderately novel contributions by establishing formal chaotic results in settings related to performative prediction and supervised learning. The work builds on existing techniques and observations, applying them to a new context. However, the incremental nature of the contributions is evident in the... | The paper introduces a novel application of exponentiated gradient descent in performative prediction, a special case of reinforcement learning. The work extends existing results on the chaotic behavior of gradient descent in other game settings, though the extensions remain largely incremental. The authors' use of sta... | Novelty Score: 1
Score Justification:
The paper presents moderately novel contributions by establishing formal chaotic results in settings related to performative prediction and supervised learning. The work builds on existing techniques and observations, applying them to a new context. However, the incremental nature... |
-RAFyM-YPj | Counting Substructures with Higher-Order Graph Neural Networks: Possibility and Impossibility Results | ICLR | 2,022 | train | false | While message passing Graph Neural Networks (GNNs) have become increasingly popular architectures for learning with graphs, recent works have revealed important shortcomings in their expressive power. In response, several higher-order GNNs have been proposed that substantially increase the expressive power, albeit at a... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe main contribution of this paper is to propose a novel GNN architecture called RNP-GNN. This architecture seems to be specifically designed to tackle the task of counting subgraphs. The authors prove two expressiveness results: (1) RPN-GNN succeeds in th... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The main contribution of this paper is to propose a novel GNN architecture called RNP-GNN.",
"This architecture seems to be specifically designed to tackle the task of counting subgraphs.",
"I think this paper suggests a very interesting GNN architecture, which essentially allows more expressive power than sta... | {
"novelty_excerpts": [
"The main contribution of this paper is to propose a novel GNN architecture called RNP-GNN.",
"This architecture seems to be specifically designed to tackle the task of counting subgraphs.",
"I think this paper suggests a very interesting GNN architecture, which essentially allows more expressiv... | 1 | The paper introduces a novel GNN architecture called RNP-GNN that allows more expressive power than standard MPNN without using k-th order tensors. The work presents a significant extension of existing GNN architectures, though the extensions remain largely incremental in terms of computational efficiency. The contribu... | The main novel contributions or new ideas introduced are the RNP-GNN architecture and its theoretical expressiveness results, including the universal approximation property. The paper builds on existing GNN architectures, such as MPNN and higher-order GNNs, but offers a more efficient alternative for certain tasks. How... | Novelty Score: 1
Score Justification:
The paper introduces a novel GNN architecture called RNP-GNN that allows more expressive power than standard MPNN without using k-th order tensors. The work presents a significant extension of existing GNN architectures, though the extensions remain largely incremental in terms of... |
-TSe5o7STVR | Non-Parallel Text Style Transfer with Self-Parallel Supervision | ICLR | 2,022 | train | false | The performance of existing text style transfer models is severely limited by the non-parallel datasets on which the models are trained. In non-parallel datasets, no direct mapping exists between sentences of the source and target style; the style transfer models thus only receive weak supervision of the target sentenc... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nStyle transfer is a text generation task where sentences need to be rewritten in such a way that a particular target attribute is introduced (like formality), but at the same time the content is approximately preserved. Style transfer is usually researched ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The scene graph similarity used to select pseudo parallel pairs is novel and makes sense to me.",
"The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.",
"The contributions are only marginally significant or novel."
] | {
"novelty_excerpts": [
"The scene graph similarity used to select pseudo parallel pairs is novel and makes sense to me.",
"The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.",
"The contributions are only marginally significant or novel."
]
} | 1 | The paper introduces novel ideas in style transfer, particularly in the use of scene graphs for semantic preservation and the application of imitation learning to refine pseudo-parallel data. However, the contributions remain somewhat incremental, building upon existing methods in machine translation and pretraining. T... | The paper introduces a novel approach to style transfer by combining scene graphs and imitation learning. This method extends existing work in machine translation and pretraining, but its application to style transfer is a new direction. The use of scene graphs for semantic preservation is a novel aspect, as it has not... | Novelty Score: 1
Score Justification:
The paper introduces novel ideas in style transfer, particularly in the use of scene graphs for semantic preservation and the application of imitation learning to refine pseudo-parallel data. However, the contributions remain somewhat incremental, building upon existing methods in... |
-Txy_1wHJ4f | Safe Deep RL in 3D Environments using Human Feedback | ICLR | 2,022 | train | false | Agents should avoid unsafe behaviour during both training and deployment. This typically requires a simulator and a procedural specification of unsafe behaviour. Unfortunately, a simulator is not always available, and procedurally specifying constraints can be difficult or impossible for many real-world tasks. A recent... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes an extension to ReQueST, which learn learns a neural simulator of the environment from safe human trajectories and then learns a reward model from human feedback. This work extends ReQueST by dense reward sketches on imagines trajectories... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors clearly state their contributions and delineate it from previous work.",
"The use of real human data and reward sketches is laudable, and an important contribution.",
"This only seems to be a very incremental work over the existing algorithm.",
"Additionally benchmarking ReQueST on 3D environment... | {
"novelty_excerpts": [
"The authors clearly state their contributions and delineate it from previous work.",
"The use of real human data and reward sketches is laudable, and an important contribution.",
"This only seems to be a very incremental work over the existing algorithm.",
"Additionally benchm... | -1 | The paper introduces a minor extension to the ReQueST algorithm by incorporating dense reward sketches on imagined trajectories, but the contribution remains largely incremental. The work builds upon existing ideas without significantly advancing the field. The approach fails to demonstrate substantial novelty in its p... | The paper extends the ReQueST algorithm by incorporating dense reward sketches on imagined trajectories, though the extensions remain largely incremental. The work presents a competent application of known methods without substantial conceptual or technical innovation. The contribution lies in the use of real human dat... | Novelty Score: -1
Score Justification:
The paper introduces a minor extension to the ReQueST algorithm by incorporating dense reward sketches on imagined trajectories, but the contribution remains largely incremental. The work builds upon existing ideas without significantly advancing the field. The approach fails to ... |
-dzXGe2FyW6 | Equalized Robustness: Towards Sustainable Fairness Under Distributional Shifts | ICLR | 2,022 | train | false | Increasing concerns have been raised on deep learning fairness in recent years. Existing fairness metrics and algorithms mainly focus on the discrimination of model performance across different groups on in-distribution data. It remains unclear whether the fairness achieved on in-distribution data can be generalized to... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studies fairness under distribution shifts. The model claims to be fair in robustness against distribution shifts with several experiments being conducted. \n\n## Main Review\nPoints in favor\n+ Fairness is a hot and timely topic.\n+ Selection of... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors overstate several of their contributions while the actual contribution is unknown.",
"Rather it extends equalized model performance to distribution shift settings.",
"Similarity, the 'new' learning algorithm is basically a combination of existing techniques such as Hessian matrix and adversarial le... | {
"novelty_excerpts": [
"The authors overstate several of their contributions while the actual contribution is unknown.",
"Rather it extends equalized model performance to distribution shift settings.",
"Similarity, the 'new' learning algorithm is basically a combination of existing techniques such as Hes... | 0 | The paper extends existing fairness goals and metrics to distribution shift settings, though the actual contribution remains largely incremental. The proposed method combines existing techniques, such as Hessian matrix and adversarial learning frameworks, without introducing fundamentally new ideas. The work presents m... | The paper introduces a method that extends equalized model performance to distribution shift settings, building on existing fairness goals and metrics. The approach combines existing techniques, such as Hessian matrix and adversarial learning frameworks, to achieve robust fairness. However, the method does not propose ... | Novelty Score: 0
Score Justification:
The paper extends existing fairness goals and metrics to distribution shift settings, though the actual contribution remains largely incremental. The proposed method combines existing techniques, such as Hessian matrix and adversarial learning frameworks, without introducing funda... |
-e7awdzWsOc | Towards Structured Dynamic Sparse Pre-Training of BERT | ICLR | 2,022 | train | false | Identifying algorithms for computational efficient unsupervised training of large language models is an important and active area of research.
In this work, we develop and study a straightforward, dynamic always-sparse pre-training approach for BERT language modeling, which leverages periodic compression steps based o... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper studies the feasibility to use the dynamic sparsity (DynSparse) method of continuously pruning and re-allocating network weights during training for large unsupervised (self-supervised) language models, namely BERT. Towards this end, the authors p... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"However, the presented paper does not provide much novelty or method advancement.",
"The authors simply apply an already existing method to one type of network.",
"More importantly, no generalized “lessons-learned” from these studies are drawn, and the translation towards other network architectures is not dis... | {
"novelty_excerpts": [
"However, the presented paper does not provide much novelty or method advancement.",
"The authors simply apply an already existing method to one type of network.",
"More importantly, no generalized “lessons-learned” from these studies are drawn, and the translation towards other ne... | 0 | The paper presents a straightforward application of an existing method to a specific network architecture, lacking significant scientific impact. The work builds on existing ideas, but the contributions are only marginally significant or novel. The method extends existing approaches to sparse training, though the exten... | The paper introduces a dynamic sparse pre-training method for BERT, which is an application of an existing approach to a specific network architecture. The work presents a block-sparse DynSparse scheme for higher effective utilization of FLOPS on real hardware accelerators, but this is not a fundamentally new idea. The... | Novelty Score: 0
Score Justification:
The paper presents a straightforward application of an existing method to a specific network architecture, lacking significant scientific impact. The work builds on existing ideas, but the contributions are only marginally significant or novel. The method extends existing approach... |
-geBFMKGlkq | Density-based Clustering with Kernel Diffusion | ICLR | 2,022 | train | false | Finding a suitable density function is essential for density-based clustering algorithms such as DBSCAN and DPC. A naive density corresponding to the indicator function of a unit $d$-dimensional Euclidean ball is commonly used in these algorithms. Such density suffers from capturing local features in complex datasets. ... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper presents a methodology for defining the density reference function required in density-based clustering methods such as DBSCAN and DPC. The density function is obtained as the limiting probability density of a diffusion process and it takes into a... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The proposed idea is interesting and novel.",
"However, there are several concerns to be addressed:",
"A major drawback of the method is that it includes additional hyperparameters compared to the naïve approach.",
"As a general remark, density-based methods are not considered effective in the case of small ... | {
"novelty_excerpts": [
"The proposed idea is interesting and novel.",
"However, there are several concerns to be addressed:",
"A major drawback of the method is that it includes additional hyperparameters compared to the naïve approach.",
"As a general remark, density-based methods are not considered... | 1 | The paper introduces a new approach to density-based clustering by constructing a Markov process and utilizing its stationary distribution as a density function. However, the contributions remain largely incremental, building on existing ideas in spectral clustering and density-based clustering methods. The work presen... | The main novel contribution of the paper lies in its application of the stationary distribution of a Markov process as a density function for density-based clustering methods. This approach builds on existing ideas in spectral clustering and density-based clustering, but the paper presents a competent extension of thes... | Novelty Score: 1
Score Justification:
The paper introduces a new approach to density-based clustering by constructing a Markov process and utilizing its stationary distribution as a density function. However, the contributions remain largely incremental, building on existing ideas in spectral clustering and density-ba... |
-h5rboREox7 | Double Descent in Adversarial Training: An Implicit Label Noise Perspective | ICLR | 2,022 | train | false | Here, we show that the robust overfitting shall be viewed as the early part of an epoch-wise double descent --- the robust test error will start to decrease again after training the model for a considerable number of epochs. Inspired by our observations, we further advance the analyses of double descent to understand r... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper demonstrates that \"robust overfitting\" during adversarial training is an early part of an epoch-wise double descent phenomenon for relatively large models. The authors also demonstrate that the overfitting behavior is observed as we select larg... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This work provides a novel and interesting analysis for explaining the robust overfitting problem of adversarial training.",
"I believe the novel analysis will provide a new perspective for the community to explain some other phenomenons.",
"The contributions are significant, and do not exist in prior works.",... | {
"novelty_excerpts": [
"This work provides a novel and interesting analysis for explaining the robust overfitting problem of adversarial training.",
"I believe the novel analysis will provide a new perspective for the community to explain some other phenomenons.",
"The contributions are significant, and ... | 1 | The paper introduces a novel connection between robust overfitting and the epoch-wise double descent phenomenon, though the extensions remain largely incremental. The work presents a new perspective on robust overfitting, but its novelty is somewhat tempered by the existence of similar observations in prior literature.... | The paper introduces a novel analysis for explaining robust overfitting in adversarial training by connecting it to the epoch-wise double descent phenomenon. This connection is relatively new, but the analysis itself is somewhat standard, building on existing ideas in the field. The work presents a new perspective on r... | Novelty Score: 1
Score Justification:
The paper introduces a novel connection between robust overfitting and the epoch-wise double descent phenomenon, though the extensions remain largely incremental. The work presents a new perspective on robust overfitting, but its novelty is somewhat tempered by the existence of si... |
-llS6TiOew | Fairness in Representation for Multilingual NLP: Insights from Controlled Experiments on Conditional Language Modeling | ICLR | 2,022 | train | false | We perform systematically and fairly controlled experiments with the 6-layer Transformer to investigate the hardness in conditional-language-modeling languages which have been traditionally considered morphologically rich (AR and RU) and poor (ZH). We evaluate through statistical comparisons across 30 possible languag... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nAt its core, this paper focuses on testing whether languages with different morphological complexity equally challenging to language-model in a bilingual / MT-like setting using a 6 layer Transformer architecture. Specifically, it tests the assumption whet... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The claim that this paper'renders a decade-long debate on morphology irrelevant' is made multiple times in this paper, and in my opinion, is too strong a claim to be made on the basis of experiments on a single task (conditional LM), on a handful of languages (however typologically diverse they may be), and using ... | {
"novelty_excerpts": [
"The claim that this paper'renders a decade-long debate on morphology irrelevant' is made multiple times in this paper, and in my opinion, is too strong a claim to be made on the basis of experiments on a single task (conditional LM), on a handful of languages (however typologically divers... | 1 | The paper presents moderately novel contributions by highlighting the importance of input representation in language-modeling, particularly in a multilingual setting. However, the findings are somewhat limited by the use of a single task, a handful of languages, and a specific architecture, which restricts the generali... | The paper introduces a novel perspective on the role of input representation in language-modeling, which is a significant contribution to the field. The work extends existing research on character-level neural models and input representation, but its novelty is somewhat tempered by the use of a single task and a limite... | Novelty Score: 1
Score Justification:
The paper presents moderately novel contributions by highlighting the importance of input representation in language-modeling, particularly in a multilingual setting. However, the findings are somewhat limited by the use of a single task, a handful of languages, and a specific arc... |
-ngwPqanCEZ | Representation-Agnostic Shape Fields | ICLR | 2,022 | train | false | 3D shape analysis has been widely explored in the era of deep learning. Numerous models have been developed for various 3D data representation formats, e.g., MeshCNN for meshes, PointNet for point clouds and VoxNet for voxels. In this study, we present Representation-Agnostic Shape Fields (RASF), a generalizable and co... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper presents a method for 3D shape analysis by means of local shape embedding. The key idea is to use a learnable multi-channel 3D grid to embed local shapes in the input 3D object, similar to word embedding in NLP; hence, the method can be pre-trai... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The paper claims that the proposed method is generalizable (i.e., could be used in different 3D representations, backbones and downstream tasks) and computation-efficient shape embedding layer for 3D deep learning.",
"Though the paper claims 'generalizable', the method simply re-samples points for mesh and voxel... | {
"novelty_excerpts": [
"The paper claims that the proposed method is generalizable (i.e., could be used in different 3D representations, backbones and downstream tasks) and computation-efficient shape embedding layer for 3D deep learning.",
"Though the paper claims 'generalizable', the method simply re-sampl... | 1 | The paper introduces a generalizable shape embedding layer for 3D deep learning, building on the idea of word embeddings in NLP. However, the method's originality is somewhat limited by its reliance on existing techniques, such as re-sampling points for mesh and voxel inputs. The contributions are significant and somew... | The main novel contribution lies in the application of a learnable multi-channel 3D grid to embed local shapes in the input 3D object, enabling pre-training and various downstream tasks. The approach combines existing ideas, such as local shape extraction and embedding, with a novel indexing mechanism using a 3D grid. ... | Novelty Score: 1
Score Justification:
The paper introduces a generalizable shape embedding layer for 3D deep learning, building on the idea of word embeddings in NLP. However, the method's originality is somewhat limited by its reliance on existing techniques, such as re-sampling points for mesh and voxel inputs. The ... |
-spj8FZD4y2 | Contextual Multi-Armed Bandit with Communication Constraints | ICLR | 2,022 | train | false | We consider a remote Contextual Multi-Armed Bandit (CMAB) problem, in which the decision-maker observes the context and the reward, but must communicate the actions to be taken by the agents over a rate-limited communication channel. This can model, for example, a personalized ad placement application, where the conten... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper focus on the contextual multi-armed bandit with communication constraints. It provides some discussions and numerical results.\n\n## Main Review\nMajor comments:\n1. Most importantly, there is no complete theoretical analysis (e.g. the upper boun... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The contributions are only marginally significant or novel.",
"Hence, the contribution of the work is not clear."
] | {
"novelty_excerpts": [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The contributions are only marginally significant or novel.",
"Hence, the contribution of the work is not clear."
]
} | 0 | The paper presents a work that shows minor variations or applications of known methods without substantial conceptual or technical innovation. The contributions are somewhat standard and lack a clear, distinct departure from existing approaches. | The paper introduces a contextual multi-armed bandit with communication constraints, though this setup is not clearly distinct from existing papers. The work extends existing bandit algorithms, though the extensions remain largely incremental. The approach combines elements of existing methods without significantly adv... | Novelty Score: 0
Score Justification:
The paper presents a work that shows minor variations or applications of known methods without substantial conceptual or technical innovation. The contributions are somewhat standard and lack a clear, distinct departure from existing approaches.
Detailed Assessment:
The paper int... |
-u8EliRNW8k | Speech-MLP: a simple MLP architecture for speech processing | ICLR | 2,022 | train | false | Overparameterized transformer-based architectures have shown remarkable performance in recent years, achieving state-of-the-art results in speech processing tasks such as speech recognition, speech synthesis, keyword spotting, and speech enhancement et al. The main assumption is that with the underlying self-attention ... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes a new architecture for speech processing, dubbed \"speech-MLP\". Besides input layers and output task-specific layers, speech-MLP consists of linear layers, residual connections, instance or layer normalization, GELU activations, and a la... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"One main contribution of the paper, the \"split and glue\" layer, seems like a groupwise convolution layer.",
"An unfold followed by a linear layer *is* a convolution. This significantly reduces the novelty and undermines the claim that the \"architecture involves simple linear transformations only\".",
"Prior... | {
"novelty_excerpts": [
"One main contribution of the paper, the \"split and glue\" layer, seems like a groupwise convolution layer.",
"An unfold followed by a linear layer *is* a convolution. This significantly reduces the novelty and undermines the claim that the \"architecture involves simple linear transf... | 0 | The paper presents a variation of existing architectures with a new layer, but its novelty is undermined by similarities to groupwise convolution layers and the lack of clear advances over prior methods. The work builds on existing approaches, primarily the Mixer MLP architecture, and its contributions are largely incr... | The paper introduces a new architecture, Speech-MLP, which combines elements of Mixer MLP and other existing approaches. The split-and-glue layer, a key component of Speech-MLP, is similar to groupwise convolution layers, reducing the novelty of the architecture. The work extends existing methods in speech processing, ... | Novelty Score: 0
Score Justification:
The paper presents a variation of existing architectures with a new layer, but its novelty is undermined by similarities to groupwise convolution layers and the lack of clear advances over prior methods. The work builds on existing approaches, primarily the Mixer MLP architecture,... |
-uPIaaZdMLF | Attentional meta-learners for few-shot polythetic classification | ICLR | 2,022 | train | false | Polythetic classifications, based on shared patterns of features that need neither be universal nor constant among members of a class, are common in the natural world and greatly outnumber monothetic classifications over a set of features. We show that threshold meta-learners, such as Prototypical Networks, require an ... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper first considers the limitations of threshold and attentional classifiers. They proposed an attention-based method for feature selection to address the problems of threshold classifiers and attentional classifiers. The experiments on several syn... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Lack of technological innovation. The proposed feature-selection mechanism is too simple and the only technological innovation.",
"How to choose repetitions R? It is a hyper-parameter, is the bigger the R the better? Does it have no upper bound?",
"Self-attention also has some parameters for the transformatio... | {
"novelty_excerpts": [
"Lack of technological innovation. The proposed feature-selection mechanism is too simple and the only technological innovation.",
"How to choose repetitions R? It is a hyper-parameter, is the bigger the R the better? Does it have no upper bound?",
"Self-attention also has some pa... | 1 | The paper presents a moderately novel approach to few-shot learning by introducing the concept of polythetic classification and proposing an attention-based feature refinement technique. While the method builds upon existing ideas, it offers a new perspective on the problem and demonstrates improvements over baseline a... | The paper introduces a new problem formulation for few-shot learning, focusing on polythetic classification, where class membership is determined by combinations of present and absent features. The work presents an attention-based feature refinement technique that outperforms baseline approaches on polythetic few-shot ... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to few-shot learning by introducing the concept of polythetic classification and proposing an attention-based feature refinement technique. While the method builds upon existing ideas, it offers a new perspective on the problem and de... |
-w2oomO6qgc | GeneDisco: A Benchmark for Experimental Design in Drug Discovery | ICLR | 2,022 | train | false | In vitro cellular experimentation with genetic interventions, using for example CRISPR technologies, is an essential step in early-stage drug discovery and target validation that serves to assess initial hypotheses about causal associations between biological mechanisms and disease pathologies. With billions of potenti... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper presents a benchmark for experimental design in drug discovery based on previously public data sets. The benchmark includes three standardized feature sets describing possible interventions and four different genome-wide CRISPR experiment results ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"- I feel it lacks the kind of machine learning novelty and significance that would be expected from an ICLR publication, and hence some other forum could be more suitable",
"- A straightforward paper that presents a useful-looking benchmark dataset for drug discovery. Machine learning novelty and significance is... | {
"novelty_excerpts": [
"- I feel it lacks the kind of machine learning novelty and significance that would be expected from an ICLR publication, and hence some other forum could be more suitable",
"- A straightforward paper that presents a useful-looking benchmark dataset for drug discovery. Machine learning... | 1 | The paper presents a moderately novel contribution to the field of experimental design in drug discovery by introducing a benchmarking suite for evaluating active learning strategies. The work builds on existing ideas but offers a new experimental design benchmark suite with unique challenges and an evaluation methodol... | The paper introduces a novel benchmarking suite for experimental design in drug discovery, which includes a new experimental design benchmark suite of genetic experiments for CRISPR screens. The work extends existing approaches by providing a cost-efficient method for evaluating active learning strategies. However, the... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel contribution to the field of experimental design in drug discovery by introducing a benchmarking suite for evaluating active learning strategies. The work builds on existing ideas but offers a new experimental design benchmark suite with uniqu... |
-xhk0O7iAc0 | A Topological View of Rule Learning in Knowledge Graphs | ICLR | 2,022 | train | false | Inductive relation prediction is an important learning task for knowledge graph completion. One can use the existence of rules, namely a sequence of relations, to predict the relation between two entities. Previous works view rules as paths and primarily focus on the searching of paths between entities. The space of pa... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes an approach to predict the confidence of missing edges in a knowledge graph in the purely inductive setting, in which the test knowledge graph does not share any entities with the training graph. The key idea is to use cycles (or more spe... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are only marginally significant or novel.",
"The contributions are only marginally significant or novel."
] | {
"novelty_excerpts": [
"The contributions are only marginally significant or novel.",
"The contributions are only marginally significant or novel."
]
} | -1 | The paper introduces a convoluted approach to inductive relation prediction, which is neither well-argued for nor convincing. The contributions are only marginally significant or novel, and the method is unnecessarily complicated. The work presents a limited extension of existing ideas, with significant overlap with pr... | The paper builds on existing methods for inductive relation prediction, but its approach is overly complex and lacks clear justification. The use of cycle representation is not distinct from path representation in multi-hop reasoning methods, and the method's scalability is unclear. The paper's claims about the importa... | Novelty Score: -1
Score Justification:
The paper introduces a convoluted approach to inductive relation prediction, which is neither well-argued for nor convincing. The contributions are only marginally significant or novel, and the method is unnecessarily complicated. The work presents a limited extension of existing... |
01AMRlen9wJ | Online Hyperparameter Meta-Learning with Hypergradient Distillation | ICLR | 2,022 | train | false | Many gradient-based meta-learning methods assume a set of parameters that do not participate in inner-optimization, which can be considered as hyperparameters. Although such hyperparameters can be optimized using the existing gradient-based hyperparameter optimization (HO) methods, they suffer from the following issues... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors propose a hyperparameter optimization algorithm in meta-learning, where parameters w/o being involved in the inner loop optimization are treated as hyperparameters. The proposed algorithms approximate the second-order hypergradients via knowledg... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The contributions are significant and somewhat new."
] | {
"novelty_excerpts": [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The contributions are significant and somewhat new."
]
} | 1 | The paper introduces a new approach to hyperparameter optimization in meta-learning, approximating second-order hypergradients via knowledge distillation. However, the contribution lies in its competent extension and combination of existing ideas rather than fundamentally new concepts. The work presents a somewhat new ... | The method builds on existing knowledge distillation techniques and applies them to the problem of hyperparameter optimization in meta-learning. The approach combines existing ideas in a novel way, but the extensions remain largely incremental. The paper evaluates the effectiveness of the proposed method on image class... | Novelty Score: 1
Score Justification:
The paper introduces a new approach to hyperparameter optimization in meta-learning, approximating second-order hypergradients via knowledge distillation. However, the contribution lies in its competent extension and combination of existing ideas rather than fundamentally new conc... |
03RLpj-tc_ | Crystal Diffusion Variational Autoencoder for Periodic Material Generation | ICLR | 2,022 | train | false | Generating the periodic structure of stable materials is a long-standing challenge for the material design community. This task is difficult because stable materials only exist in a low-dimensional subspace of all possible periodic arrangements of atoms: 1) the coordinates must lie in the local energy minimum defined b... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes an approach to generate stable crystal structures using variational autoencoders without having to train on out-of-equilibrium data, which is scarce for materials. Beyond that, benchmark datasets and tasks for materials generation are pro... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The proposed score-based generative model for crystal structure generation, i.e., CDVAE, is novel.",
"Although similar techniques have already been explored in other structure generation problems [1,2], I still believe the proposed model is quite novel as it addresses several important challenges in material str... | {
"novelty_excerpts": [
"The proposed score-based generative model for crystal structure generation, i.e., CDVAE, is novel.",
"Although similar techniques have already been explored in other structure generation problems [1,2], I still believe the proposed model is quite novel as it addresses several importan... | 1 | The paper introduces a novel approach to generating stable crystal structures using variational autoencoders without requiring out-of-equilibrium data. The method presents a moderately novel contribution by addressing several important challenges in material structure generation, such as periodic boundary conditions an... | The paper introduces a score-based Diffusion Variational Autoencoder (CDVAE) for crystal structure generation, which is a novel approach to tackling the challenges of material structure generation. The method incorporates periodic boundary conditions and composition prediction, addressing significant challenges in the ... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to generating stable crystal structures using variational autoencoders without requiring out-of-equilibrium data. The method presents a moderately novel contribution by addressing several important challenges in material structure generation, ... |
04pGUg0-pdZ | Finite-Time Convergence and Sample Complexity of Multi-Agent Actor-Critic Reinforcement Learning with Average Reward | ICLR | 2,022 | train | false | In this paper, we establish the first finite-time convergence result of the actor-critic algorithm for fully decentralized multi-agent reinforcement learning (MARL) problems with average reward.
In this problem, a set of $N$ agents work cooperatively to maximize the global average reward through interacting with their... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper considers cooperative multi-agent reinforcement learning (MARL) for average reward MDPs with fully decentralized actor-critic methods. In particular, the authors make some progress on top of existing works in this direction, and in particular (Zh... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This paper makes some relatively non-trivial progress on top of (Zhang et al., 2018).",
"In particular, while (Zhang et al., 2018) requires that each agent observes the joint actions of all agents, this paper only requires each agent to observe their own private action.",
"This paper also derives the first non... | {
"novelty_excerpts": [
"This paper makes some relatively non-trivial progress on top of (Zhang et al., 2018).",
"In particular, while (Zhang et al., 2018) requires that each agent observes the joint actions of all agents, this paper only requires each agent to observe their own private action.",
"This paper also deriv... | 1 | The paper introduces a novel extension to existing cooperative multi-agent reinforcement learning methods by removing the assumption of joint actions being observable to all agents. This modification leads to improved theoretical guarantees and numerical results. However, the contributions remain largely incremental, b... | The paper presents a moderately novel approach by modifying the actor updates with mini-batch TD sharing to accommodate private action observations. This extension allows for improved convergence rates and sample complexities in fully decentralized MARL. The work builds on existing methods, such as (Zhang et al., 2018)... | Novelty Score: 1
Score Justification:
The paper introduces a novel extension to existing cooperative multi-agent reinforcement learning methods by removing the assumption of joint actions being observable to all agents. This modification leads to improved theoretical guarantees and numerical results. However, the cont... |
057dxuWpfx | Shaped Rewards Bias Emergent Language | ICLR | 2,022 | train | false | One of the primary characteristics of emergent phenomena is that they are determined by the basic properties of the system whence they emerge as opposed to explicitly designed constraints. Reinforcement learning is often used to elicit such phenomena which specifically arise from the pressure to maximize reward. We dis... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper examines the effect of reward shaping in the context of emergent communication, and specifically with respect to the Shannon entropy of the emerged language. The paper showed that shaping the rewards can lead to higher entropy than training with j... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"use of PPO in emergent communication is new",
"empirical exploration of effect of reward shaping on the entropy of the emergent language is new",
"theoretical analysis of why the reward shaping affects the entropy of the emergent language is new",
"The contributions are significant, and do not exist in prior... | {
"novelty_excerpts": [
"use of PPO in emergent communication is new",
"empirical exploration of effect of reward shaping on the entropy of the emergent language is new",
"theoretical analysis of why the reward shaping affects the entropy of the emergent language is new",
"The contributions are signif... | 1 | The paper introduces novel approaches to emergent communication, including the use of PPO and empirical exploration of reward shaping's effect on language entropy. However, the theoretical analysis of reward shaping's impact on entropy is somewhat limited in its scope and application. The work presents moderately novel... | The paper introduces the use of PPO in emergent communication, which is a new approach in this field. The empirical exploration of reward shaping's effect on language entropy is also a novel aspect of the work. However, the theoretical analysis of reward shaping's impact on entropy is somewhat limited in its scope and ... | Novelty Score: 1
Score Justification:
The paper introduces novel approaches to emergent communication, including the use of PPO and empirical exploration of reward shaping's effect on language entropy. However, the theoretical analysis of reward shaping's impact on entropy is somewhat limited in its scope and applicat... |
06fUz_bJStS | Differentially Private SGD with Sparse Gradients | ICLR | 2,022 | train | false | A large number of recent studies reveal that networks and their optimization updates contain information about potentially private training data. To protect sensitive training data, differential privacy has been adopted in deep learning to provide rigorously defined and measurable privacy. However, differentially priva... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper aims to improve the effectiveness of deep learning under differential privacy. It proposes a simple dimension reduction method that could potentially improve model accuracy as well as mitigate the computational and memory overhead of some previous... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The problem the paper aims to address is an important one.",
"Machine learning with differential privacy is a very active research area, and this paper is a valuable addition to the literature in this field.",
"The proposed method is simple to understand and implement.",
"It is also general and applicable to... | {
"novelty_excerpts": [
"The problem the paper aims to address is an important one.",
"Machine learning with differential privacy is a very active research area, and this paper is a valuable addition to the literature in this field.",
"The proposed method is simple to understand and implement.",
"It i... | 0 | The paper presents a dimension reduction technique called random freeze, which is a relatively simple approach to improve the effectiveness of deep learning under differential privacy. However, the method is not fully novel, as it has appeared in other papers, and its algorithmic contribution is limited. The work exten... | The paper introduces a dimension reduction technique called random freeze, which randomly selects a subset of model parameters and zeros out corresponding gradients during the training process. The approach builds on existing methods for private machine learning, but its novelty lies in its simplicity and ease of imple... | Novelty Score: 0
Score Justification:
The paper presents a dimension reduction technique called random freeze, which is a relatively simple approach to improve the effectiveness of deep learning under differential privacy. However, the method is not fully novel, as it has appeared in other papers, and its algorithmic ... |
085y6YPaYjP | Zero-Shot Self-Supervised Learning for MRI Reconstruction | ICLR | 2,022 | train | false | Deep learning (DL) has emerged as a powerful tool for accelerated MRI reconstruction, but often necessitates a database of fully-sampled measurements for training. Recent self-supervised and unsupervised learning approaches enable training without fully-sampled data. However, a database of undersampled measurements may... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis submission deals with MR images reconstruction in a context where the raw data is under-sampled. This data (which is in the Fourier domain) can be under-sampled to accelerate the imaging exam and thus improve clinical workflow and it is thus a very rel... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This data (which is in the Fourier domain) can be under-sampled to accelerate the imaging exam and thus improve clinical workflow and it is thus a very relevant research topic.",
"The submission relies on recent deep learning models that explicitly incorporate some physical aspects of MR image acquisition (multi... | {
"novelty_excerpts": [
"This data (which is in the Fourier domain) can be under-sampled to accelerate the imaging exam and thus improve clinical workflow and it is thus a very relevant research topic.",
"The submission relies on recent deep learning models that explicitly incorporate some physical aspects of... | 0 | The paper presents a competent extension of existing ideas in MRI reconstruction, combining zero-shot learning, plug-and-play architectures, and transfer learning. However, the contributions remain largely incremental, building upon prior work without introducing fundamentally new concepts. The method extends existing ... | The paper introduces a zero-shot learning approach for MRI reconstruction, leveraging plug-and-play architectures and transfer learning to achieve good results. The method combines existing ideas, but the contributions are not novel in a fundamental sense. The work presents a competent application of existing technique... | Novelty Score: 0
Score Justification:
The paper presents a competent extension of existing ideas in MRI reconstruction, combining zero-shot learning, plug-and-play architectures, and transfer learning. However, the contributions remain largely incremental, building upon prior work without introducing fundamentally new... |
0DLwqQLmqV | NAS-Bench-Suite: NAS Evaluation is (Now) Surprisingly Easy | ICLR | 2,022 | train | false | The release of tabular benchmarks, such as NAS-Bench-101 and NAS-Bench-201, has significantly lowered the computational overhead for conducting scientific research in neural architecture search (NAS). Although they have been widely adopted and used to tune real-world NAS algorithms, these benchmarks are limited to smal... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a unified interface for access to a collection of Neural Architecture Search (NAS) benchmarks. With experiments performed on across multiple search spaces and datasets, the authors show that some conclusions drawn from a small subset of ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"the work lacks deeper insights from the experiments performed that could guide the community towards better methods.",
"In my view the two main contributions of this work are (1) showcasing the limitations of drawing conclusions from single benchmarks; (2) providing the community with a larger analysis tool.",
... | {
"novelty_excerpts": [
"the work lacks deeper insights from the experiments performed that could guide the community towards better methods.",
"In my view the two main contributions of this work are (1) showcasing the limitations of drawing conclusions from single benchmarks; (2) providing the community with... | 1 | The paper presents a moderately novel contribution by introducing a unified interface for access to a collection of Neural Architecture Search (NAS) benchmarks, highlighting the limitations of drawing conclusions from single benchmarks, and providing a larger analysis tool for the community. The work builds upon existi... | The paper introduces a unified interface for NAS benchmarks, which is a moderately novel contribution. The work presents a comprehensive analysis of existing NAS benchmarks, highlighting the limitations of drawing conclusions from single benchmarks. The approach combines existing search spaces and architecture datasets... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel contribution by introducing a unified interface for access to a collection of Neural Architecture Search (NAS) benchmarks, highlighting the limitations of drawing conclusions from single benchmarks, and providing a larger analysis tool for the... |
0DecTiJFbm | A New Perspective on Fluid Simulation: An Image-to-Image Translation Task via Neural Networks | ICLR | 2,022 | train | false | Standard numerical methods for creating simulation models in the field of fluid dynamics are designed to be close to perfection, which results in high computational effort and high computation times in many cases. Unfortunately, there is no mathematical way to decrease this correctness in cases where only approximate p... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper uses a cGAN to model the fluid flow of a Karman vortex street in image-space. This is intended as a more efficient method for creating 'visual' simulations compared to classical numerical simulation.\n\n## Main Review\nThe paper uses the cGAN meth... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"There is no novelty in the method (this is 1:1 reimplementation of Isola et al), nor the task (Karman vortex street is a standard benchmark task for both classical, and ML-based physics prediction papers).",
"It's unclear to me what benefits an 'image-based' approach would bring.",
"The only difference in the ... | {"novelty_excerpts": [
"There is no novelty in the method (this is 1:1 reimplementation of Isola et al), nor the task (Karman vortex street is a standard benchmark task for both classical, and ML-based physics prediction papers).",
"It's unclear to me what benefits an 'image-based' approach would bring.",
"The only dif... | -1 | The paper introduces a standard implementation of a cGAN on a standard benchmark task for physics prediction without significant innovation, comparison to other ML approaches, or interesting analysis and discussion. The work does not bring any new insights to the community, and its contributions are neither significant... | The paper uses a cGAN method in [Isola et al 2017] on image sequences generated by a fluid solver, which is a 1:1 reimplementation of Isola et al without any novelty in the method or task. The paper does not contain any baseline comparisons to other methods or further analysis, and its contributions are limited to a st... | Novelty Score: -1
Score Justification:
The paper introduces a standard implementation of a cGAN on a standard benchmark task for physics prediction without significant innovation, comparison to other ML approaches, or interesting analysis and discussion. The work does not bring any new insights to the community, and i... |
0EL4vLgYKRW | Plan Better Amid Conservatism: Offline Multi-Agent Reinforcement Learning with Actor Rectification | ICLR | 2,022 | train | false | The idea of conservatism has led to significant progress in offline reinforcement learning (RL) where an agent learns from pre-collected datasets. However, it is still an open question to resolve offline RL in the more practical multi-agent setting as many real-world scenarios involve interaction among multiple agents.... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a simple yet effective multi-agent offline reinforcement learning algorithm, called Offline Multi-Agent RL with Actor Rectification, which combines CQL (first-order optimization) and evolutionary algorithms (zero-order optimization) to e... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The local optimum problem is successfully solved using zero-order optimization for offline training in the multi-agent scenario for the first time.",
"Although the approach proposed in this paper achieves significant performance, the following issues are not explained or investigated very clearly.",
"So it is ... | {
"novelty_excerpts": [
"The local optimum problem is successfully solved using zero-order optimization for offline training in the multi-agent scenario for the first time.",
"Although the approach proposed in this paper achieves significant performance, the following issues are not explained or investigated ... | 1 | The paper introduces a novel combination of CQL and evolutionary algorithms for offline multi-agent reinforcement learning, which presents a moderately novel approach to addressing the local optimum problem in multi-agent tasks. The work builds on existing methods but contributes a new perspective on their applicabilit... | The method extends the applicability of CQL to multi-agent tasks by incorporating evolutionary algorithms, though the extensions remain largely incremental. The paper successfully solves the local optimum problem using zero-order optimization for offline training in the multi-agent scenario, which is a novel applicatio... | Novelty Score: 1
Score Justification:
The paper introduces a novel combination of CQL and evolutionary algorithms for offline multi-agent reinforcement learning, which presents a moderately novel approach to addressing the local optimum problem in multi-agent tasks. The work builds on existing methods but contributes ... |
0EXmFzUn5I | Pyraformer: Low-Complexity Pyramidal Attention for Long-Range Time Series Modeling and Forecasting | ICLR | 2,022 | train | false | Accurate prediction of the future given the past based on time series data is of paramount importance, since it opens the door for decision making and risk management ahead of time. In practice, the challenge is to build a flexible but parsimonious model that can capture a wide range of temporal dependencies. In this p... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors propose a new architecture to tackle the problem of long sequence temporal forecasting (LSTF) – which looks at capturing long-range dependencies in time-series data by providing more direct paths between the output and distant history. \n\nThe P... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors propose a new architecture to tackle the problem of long sequence temporal forecasting (LSTF) – which looks at capturing long-range dependencies in time-series data by providing more direct paths between the output and distant history.",
"The paper makes a strong contribution to long sequence tempora... | {
"novelty_excerpts": [
"The authors propose a new architecture to tackle the problem of long sequence temporal forecasting (LSTF) – which looks at capturing long-range dependencies in time-series data by providing more direct paths between the output and distant history.",
"The paper makes a strong contribut... | 1 | null | null | Novelty Score: 1
The paper introduces a moderately novel approach to long sequence temporal forecasting, building on existing transformer architectures with a pyramidal attention module that captures multi-resolution temporal dependencies. The work presents a competent extension and combination of existing ideas, with... |
0HkFxvSRDSW | Role Diversity Matters: A Study of Cooperative Training Strategies for Multi-Agent RL | ICLR | 2,022 | train | false | Cooperative multi-agent reinforcement learning (MARL) is making rapid progress for solving tasks in a grid world and real-world scenarios, in which agents are given different attributes and goals. For example, in Starcraft II battle tasks, agents are initialized with the various move, defense, and attack abilities acco... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nTo address the problem of algorithm choosing in different MARL tasks, this paper proposes to use role diversity as a metric to describe MARL tasks. For choosing parameter sharing, communication mechanism, and credit assignment strategy, this paper defines t... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The contributions are significant, and do not exist in prior works."
] | {
"novelty_excerpts": [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The contributions are significant, and do not exist in prior works."
]
} | 1 | The paper introduces a novel framework for describing MARL tasks using role diversity metrics, which is a significant contribution to the field. However, the approach builds on existing ideas and does not present fundamentally new concepts or problem formulations. The work presents a competent extension and combination... | The paper introduces a novel framework for describing MARL tasks using role diversity metrics, which is a significant contribution to the field. The approach combines policy-based, trajectory-based, and contribution-based metrics to describe the diversity of roles in MARL tasks. The work presents a competent extension ... | Novelty Score: 1
Score Justification:
The paper introduces a novel framework for describing MARL tasks using role diversity metrics, which is a significant contribution to the field. However, the approach builds on existing ideas and does not present fundamentally new concepts or problem formulations. The work present... |
0J98XyjlQ1 | D$^2$-GCN: Data-Dependent GCNs for Boosting Both Efficiency and Scalability | ICLR | 2,022 | train | false | Graph Convolutional Networks (GCNs) have gained an increasing attention thanks to their state-of-the-art (SOTA) performance in graph-based learning tasks. However, their sheer number of node features and large adjacency matrix limit their deployment into real-world applications, as they impose the following challenges:... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper attempts to address the training efficiency and scalability of GCNs. To be specific, a so-called Data-Dependent dynamic GCN framework is proposed, in which node-wise skepping, edge-wise skipping, an bit-wise skipping are integrated via gate funct... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Integrating node-wise skepping, edge-wise skipping, an bit-wise skipping to reduce the nodes, prune the edges, and compress the bit-precision is interesting and somewhat novel.",
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prio... | {
"novelty_excerpts": [
"Integrating node-wise skepping, edge-wise skipping, an bit-wise skipping to reduce the nodes, prune the edges, and compress the bit-precision is interesting and somewhat novel.",
"The contributions are significant, and do not exist in prior works.",
"The contributions are signifi... | 1 | The paper introduces a data-dependent dynamic GCN framework that integrates node-wise, edge-wise, and bit-wise skipping techniques to improve training efficiency and scalability. The approach builds on existing GCN methods but presents a novel combination of techniques to address the over-smoothing issue. However, the ... | The main novel contributions or new ideas introduced are the integration of node-wise, edge-wise, and bit-wise skipping techniques to reduce the nodes, prune the edges, and compress the bit-precision, respectively. The work presents a novel combination of existing methods to address the over-smoothing issue in GCNs. Ho... | Novelty Score: 1
Score Justification:
The paper introduces a data-dependent dynamic GCN framework that integrates node-wise, edge-wise, and bit-wise skipping techniques to improve training efficiency and scalability. The approach builds on existing GCN methods but presents a novel combination of techniques to address ... |
0JzqUlIVVDd | KL Guided Domain Adaptation | ICLR | 2,022 | train | false | Domain adaptation is an important problem and often needed for real-world applications. In this problem, instead of i.i.d. training and testing datapoints, we assume that the source (training) data and the target (testing) data have different distributions. With that setting, the empirical risk minimization training pr... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes a domain adaptation method that is guided by the reverse Kullback-Leibler (KL) divergence between the target and source domains in the representation space. To this end, it is assumed that the representation z is sufficient and that p_S(y... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"I find the paper of a great potential contribution to the field."
] | {
"novelty_excerpts": [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"I find the paper of a great potential contribution to the field."
]
} | 0 | The paper presents a domain adaptation method that builds on existing ideas and approaches, introducing a new bound on the test loss based on the training loss and a quantity based on the divergence of the joint distribution between the target and source domain. However, the contributions are somewhat standard and show... | The paper introduces a domain adaptation method that combines existing ideas and approaches, deriving a bound on the test loss based on the training loss and a quantity based on the divergence of the joint distribution between the target and source domain. The method extends existing domain adaptation methods, but the ... | Novelty Score: 0
Score Justification:
The paper presents a domain adaptation method that builds on existing ideas and approaches, introducing a new bound on the test loss based on the training loss and a quantity based on the divergence of the joint distribution between the target and source domain. However, the contr... |
0Kj5mhn6sw | Gesture2Vec: Clustering Gestures using Representation Learning Methods for Co-speech Gesture Generation | ICLR | 2,022 | train | false | Co-speech gestures are a principal component in conveying messages and enhancing interaction experiences between humans. Similarly, the co-speech gesture is a key ingredient in human-agent interaction including both virtual agents and robots. Existing machine learning approaches have yielded only marginal success in le... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper tackles with co-speech gesture generation. First a denoising autoencoder is applied to the body pose detected at each frame. Then, a Vector Quantized Denoising Variational Autoencoder learns to create a motion vocabulary by parameterizing the pos... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors present a method to generate co-speech gestures by learning a vector-quantized representation space of gestures coupled with a machine translation from natural language sentences to that representation space.",
"They design a denoising variational autoencoder (DVAE) to learn the encoding space corres... | {
"novelty_excerpts": [
"The authors present a method to generate co-speech gestures by learning a vector-quantized representation space of gestures coupled with a machine translation from natural language sentences to that representation space.",
"They design a denoising variational autoencoder (DVAE) to lea... | -1 | The paper introduces a method for co-speech gesture generation that builds on existing approaches, primarily variational autoencoders and vector quantized variational autoencoders, without significantly advancing the field. The work presents incremental contributions, such as adding denoising properties to the existing... | The paper extends existing work on variational autoencoders and vector quantized variational autoencoders for co-speech gesture generation, but the contributions are largely incremental. The method's novelty lies in its ability to generate plausible gesture sequences, but this is achieved through a combination of exist... | Novelty Score: -1
Score Justification:
The paper introduces a method for co-speech gesture generation that builds on existing approaches, primarily variational autoencoders and vector quantized variational autoencoders, without significantly advancing the field. The work presents incremental contributions, such as add... |
0Q6BzWbvg0P | Less is More: Dimension Reduction Finds On-Manifold Adversarial Examples in Hard-Label Attacks | ICLR | 2,022 | train | false | Designing deep networks robust to adversarial examples remains an open problem. Likewise, recent zeroth-order hard-label attacks on image classification models have shown comparable performance to their first-order, gradient-level alternatives. It was recently shown in the gradient-level setting that regular adversaria... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studies zeroth-order hard-label adversarial attacks. In particular, the authors explore the connection between how the gradient can reveal information about the data manifold, and its dependence on data dimensionality. They also empirically consi... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"That being said, the paper is poorly written: without a clear hypothesis, or convincing experimental or theoretical results.",
"It is not clear why the LPIPS distance or the gradient deviation are quantities we should care about in practice.",
"Unfortunately, my concerns with the paper regarding the presentati... | {
"novelty_excerpts": [
"That being said, the paper is poorly written: without a clear hypothesis, or convincing experimental or theoretical results.",
"It is not clear why the LPIPS distance or the gradient deviation are quantities we should care about in practice.",
"Unfortunately, my concerns with the ... | -1 | The paper introduces a method for zeroth-order hard-label adversarial attacks, but it fails to provide a clear hypothesis, convincing experimental or theoretical results, and sufficient justification for its assumptions. The method builds on existing ideas without significant innovation, and its contributions are not n... | The paper presents a method for black-box attacks against publicly available deep learning systems, but its approach remains largely incremental, extending existing ideas without substantial conceptual or technical innovation. The method explores the connection between gradient information and data manifold, but its pr... | Novelty Score: -1
Score Justification:
The paper introduces a method for zeroth-order hard-label adversarial attacks, but it fails to provide a clear hypothesis, convincing experimental or theoretical results, and sufficient justification for its assumptions. The method builds on existing ideas without significant inn... |
0RDcd5Axok | Towards a Unified View of Parameter-Efficient Transfer Learning | ICLR | 2,022 | train | false | Fine-tuning large pretrained language models on downstream tasks has become the de-facto learning paradigm in NLP. However, conventional approaches fine-tune all the parameters of the pretrained model, which becomes prohibitive as the model size and the number of tasks grow. Recent work has proposed a variety of parame... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper formulates different approaches for parameter-efficient transfer learning (such as adapters, prefix-tuning, LoRA) under a common framework. Approaches vary alongside different design dimensions: functional form, insertion form (sequential or para... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prior works."
] | {
"novelty_excerpts": [
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prior works."
]
} | 1 | The paper introduces a unified framework for parameter-efficient transfer learning approaches, which enables the derivation of new methods and their comparison to existing ones. The work presents a set of design dimensions and shows how changing these dimensions leads to new approaches, including parallel adapters and ... | The main novel contributions or new ideas introduced are the unified framework and the derivation of new methods, such as parallel adapters and scaled PA, through the application of design dimensions. The work extends existing approaches by providing a common framework and comparing their performance. However, the exte... | Novelty Score: 1
Score Justification:
The paper introduces a unified framework for parameter-efficient transfer learning approaches, which enables the derivation of new methods and their comparison to existing ones. The work presents a set of design dimensions and shows how changing these dimensions leads to new appro... |
0RqDp8FCW5Z | W-CTC: a Connectionist Temporal Classification Loss with Wild Cards | ICLR | 2,022 | train | false | Connectionist Temporal Classification (CTC) loss is commonly used in sequence learning applications. For example, in Automatic Speech Recognition (ASR) task, the training data consists of pairs of audio (input sequence) and text (output label),without temporal alignment information. Standard CTC computes a loss by aggr... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a method to utilize partially labeled data for the CTC loss. In order to handle untrascribed lables, it introduces a wild-card for the beginning of the sequence and uses unconstrained endpoints for the ending of the sequence. It also dis... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contribution of the paper is the formal introduction of the approach to CTC and to give experimental results to confirm the effectiveness.",
"Although it is valuable to have a paper discussing this topic, it is not very clear the actual benefit in real scenarios. Given that the technique itself it not very n... | {"novelty_excerpts": [
"The contribution of the paper is the formal introduction of the approach to CTC and to give experimental results to confirm the effectiveness.",
"Although it is valuable to have a paper discussing this topic, it is not very clear the actual benefit in real scenarios. Given that the technique its... | 1 | The paper presents a moderately novel approach to handling incomplete labels in CTC training, extending the existing dynamic time warping algorithm to CTC. The contribution lies in its application to CTC, demonstrating improved results on incomplete labels. However, the technique itself is not entirely new, and the wor... | The paper introduces a method to handle incomplete labels in CTC training by using wild-cards for missing beginnings and unconstrained endpoints for missing ends. The approach builds on existing dynamic time warping algorithms, such as SPRING, and applies them to CTC, demonstrating improved results on incomplete labels... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to handling incomplete labels in CTC training, extending the existing dynamic time warping algorithm to CTC. The contribution lies in its application to CTC, demonstrating improved results on incomplete labels. However, the technique ... |
0Tnl8uBHfQw | Deep Classifiers with Label Noise Modeling and Distance Awareness | ICLR | 2,022 | train | false | Uncertainty estimation in deep learning has recently emerged as a crucial area of interest to advance reliability and robustness in safety-critical applications. While there have been many proposed methods that either focus on distance-aware model uncertainties for out-of-distribution detection or on input-dependent la... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studies the combination between model uncertainty and data uncertainty based on spectral-normalized Gaussian process. Theoretical results show that heteroscedastic SNGP allows for joint modeling of model and data uncertainties. The paper then pro... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The first contribution, that jointly model two different types of uncertainty, as far as I know, make the first step for noisy label.",
"However，the contribution of the whole paper can not meet the requirement of ICRL as (1) limited novelty.",
"The proposed method is a limited improvement of SNGP.",
"The co... | {
"novelty_excerpts": [
"The first contribution, that jointly model two different types of uncertainty, as far as I know, make the first step for noisy label.",
"However，the contribution of the whole paper can not meet the requirement of ICRL as (1) limited novelty.",
"The proposed method is a limited im... | 0 | The paper introduces a method that jointly models model and data uncertainties, which is a somewhat novel approach. However, the contribution lies in its incremental improvement of the Spectral-Normalised Gaussian Process (SNGP) method, rather than presenting fundamentally new ideas. | The paper presents a method that combines heteroscedasticity and SNGP for uncertainty estimation, which is a moderately novel approach. The work extends existing methods for uncertainty estimation, but the extensions remain largely incremental. The method's ability to jointly model model and data uncertainties is a som... | Novelty Score: 0
Score Justification:
The paper introduces a method that jointly models model and data uncertainties, which is a somewhat novel approach. However, the contribution lies in its incremental improvement of the Spectral-Normalised Gaussian Process (SNGP) method, rather than presenting fundamentally new ide... |
0UXT6PpRpW | Large-Scale Representation Learning on Graphs via Bootstrapping | ICLR | 2,022 | train | false | Self-supervised learning provides a promising path towards eliminating the need for costly label information in representation learning on graphs. However, to achieve state-of-the-art performance, methods often need large numbers of negative examples and rely on complex augmentations. This can be prohibitively expens... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper introduces Bootstrapped Graph Latents (BGRL) to learn graph representation by predicting alternative augmentations of the input. The experiment results on a wide range of datasets exhibit appealing performance.\n\n## Main Review\nThe paper can b... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The novelty of the method and the conceptual contribution is extremely limited since the proposed model is a pretty straightforward application of the BYOL [1] model on graph data.",
"The difference of not having a projection head cannot be deemed as a contribution or novelty, since the reason is the low dimensi... | {
"novelty_excerpts": [
"The novelty of the method and the conceptual contribution is extremely limited since the proposed model is a pretty straightforward application of the BYOL [1] model on graph data.",
"The difference of not having a projection head cannot be deemed as a contribution or novelty, since t... | 0 | The paper introduces a straightforward application of the BYOL model on graph data, lacking substantial conceptual or technical innovation. The work presents a competent extension of existing ideas, but its contributions remain largely incremental. The method builds on the BYOL scheme, adapting it to graph representati... | The paper extends the BYOL model to graph representation learning tasks, leveraging its advantages in boosting task performance and enhancing model scalability. However, the technical contribution remains incremental, as the method largely replicates existing approaches with minimal innovation. The work does not introd... | Novelty Score: 0
Score Justification:
The paper introduces a straightforward application of the BYOL model on graph data, lacking substantial conceptual or technical innovation. The work presents a competent extension of existing ideas, but its contributions remain largely incremental. The method builds on the BYOL sc... |
0bXmbOt1oq | Towards Learning to Speak and Hear Through Multi-Agent Communication over a Continuous Acoustic Channel | ICLR | 2,022 | train | false | While multi-agent reinforcement learning has been used as an effective means to study emergent communication between agents, existing work has focused almost exclusively on communication with discrete symbols. Human communication often takes place (and emerged) over a continuous acoustic channel; human infants acquire ... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes to move in the direction of emerging oral communication, instead of written communication. They create an architecture where the sender agent first outputs a string of symbols, which map statically to a fixed table of phonems. The sender ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Unfortunately, there has been no comparisons made with this work that seems to be the most relevant to the proposed method.",
"The idea of using signalling to make one agent learn about the underlying concept and active listening to help the listener solve the given downstream task is already explored in Eccles ... | {
"novelty_excerpts": [
"Unfortunately, there has been no comparisons made with this work that seems to be the most relevant to the proposed method.",
"The idea of using signalling to make one agent learn about the underlying concept and active listening to help the listener solve the given downstream task is... | -1 | The paper presents a largely incremental approach to emergent communication, building on existing methods with minimal innovation. The work introduces a continuous communication channel, but the experiments and analysis fail to highlight the novelty of this environment. The contribution lies in exploring a new directio... | The paper introduces a continuous communication channel for emergent language learning, which is a new direction in the field. However, the experiments and analysis are largely standard, showing minor variations or applications of known methods without substantial conceptual or technical innovation. The work builds on ... | Novelty Score: -1
Score Justification:
The paper presents a largely incremental approach to emergent communication, building on existing methods with minimal innovation. The work introduces a continuous communication channel, but the experiments and analysis fail to highlight the novelty of this environment. The contr... |
0d1mLPC2q2 | Understanding the Success of Knowledge Distillation -- A Data Augmentation Perspective | ICLR | 2,022 | train | false | Knowledge distillation (KD) is a general neural network training approach that uses a teacher model to guide a student model. Many works have explored the rationale for its success. However, its interplay with data augmentation (DA) has not been well understood so far. In this paper, we are motivated by an interesting ... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper presents a generic framework to explain an observation that by using more training iterations, knowledge distillation (KD) losses benefit more from data augmentation (DA) than cross entropy (CE) losses do. The paper proposes an efficient DA appro... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"- This paper points out the fact that in-depth analysis of data augmentation (DA) in the field of knowledge distillation is rarely performed.",
"- The proposed augmentation techniques effectively converge with existing knowledge distillation algorithms to show that they can lead to further performance improvemen... | {
"novelty_excerpts": [
"- This paper points out the fact that in-depth analysis of data augmentation (DA) in the field of knowledge distillation is rarely performed.",
"- The proposed augmentation techniques effectively converge with existing knowledge distillation algorithms to show that they can lead to fu... | 0 | The paper presents a generic framework to explain the success of knowledge distillation (KD) approaches, but its contributions are largely incremental and build upon existing ideas. The work extends prior research on data augmentation (DA) in KD by proposing a new DA approach and demonstrating its effectiveness, but th... | The paper introduces a new DA approach that combines existing methods to improve the performance of KD algorithms, but this approach is not a groundbreaking innovation. The work presents a generic framework to explain the success of KD approaches, but the explanation is not mathematically rigorous and lacks a clear und... | Novelty Score: 0
Score Justification:
The paper presents a generic framework to explain the success of knowledge distillation (KD) approaches, but its contributions are largely incremental and build upon existing ideas. The work extends prior research on data augmentation (DA) in KD by proposing a new DA approach and ... |
0kNbTghw7q | Improving Generative Adversarial Networks via Adversarial Learning in Latent Space | ICLR | 2,022 | train | false | Generative Adversarial Networks (GANs) have been widely studied as generative models, which map a latent distribution to the target distribution. Although many efforts have been made in terms of backbone architecture design, loss function, and training techniques, few results have been obtained on how the sampling in l... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors demonstrate that the generator in a GAN is a continuous function two latent codes that are close in the latent space are mapped to two images that are close in the pixel space. However, the quality of the generated images is not preserved as qua... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"However, the quality of the generated images is not preserved as quality is not a continuous function in pixel space.",
"To address this issue, the authors propose to transform the original latent codes and demonstrate that it results in better generation quality and diversity.",
"Another limitation of the wor... | {
"novelty_excerpts": [
"However, the quality of the generated images is not preserved as quality is not a continuous function in pixel space.",
"To address this issue, the authors propose to transform the original latent codes and demonstrate that it results in better generation quality and diversity.",
... | 0 | The paper presents a competent extension and combination of existing ideas in generative adversarial networks (GANs), but it does not introduce fundamentally new concepts or approaches. The work builds upon existing techniques and methods, such as adversarial attacks and MSGAN, to improve the quality and diversity of g... | The paper introduces a method that combines two existing techniques to improve the quality and diversity of GAN-generated samples. The approach applies I-FGSM to shift latent vectors and MSGAN to mine hard latent vectors for the discriminator. However, the method remains largely incremental, as it does not significantl... | Novelty Score: 0
Score Justification:
The paper presents a competent extension and combination of existing ideas in generative adversarial networks (GANs), but it does not introduce fundamentally new concepts or approaches. The work builds upon existing techniques and methods, such as adversarial attacks and MSGAN, to... |
0kPL3xO4R5 | Fast topological clustering with Wasserstein distance | ICLR | 2,022 | train | false | The topological patterns exhibited by many real-world networks motivate the development of topology-based methods for assessing the similarity of networks. However, extracting topological structure is difficult, especially for large and dense networks whose node degrees range over multiple orders of magnitude. In this ... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nGiven a weighted graph, authors construct a filtration on top of it that is associated with a persistence diagram. This persistence diagram summarizes the topology of the weighted graph. Authors propose then to cluster different graphs using either the dist... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"To overcome the main computational challenges that arise when having to compute distances/barycenters of PDs, authors propose to consider simpler persistence diagrams that are one-dimensional (that is all points in the PDs have coordinates of the form $(-\\infty,d_i)$ or $(b_i,+\\infty)$).",
"It is well-known th... | {
"novelty_excerpts": [
"To overcome the main computational challenges that arise when having to compute distances/barycenters of PDs, authors propose to consider simpler persistence diagrams that are one-dimensional (that is all points in the PDs have coordinates of the form $(-\infty,d_i)$ or $(b_i,+\infty)$)."... | 1 | The paper presents a moderately novel approach to clustering graphs using persistence diagrams, combining topological and geometrical features. While the method leverages existing concepts, it introduces a novel algorithm that efficiently reduces the cost of matching topological features between data sets. However, the... | The method builds on existing work in persistent homology and optimal transport, but it extends the idea of using persistence diagrams for graph clustering by incorporating geometrical features and a regularization term. The approach is novel in its ability to efficiently compute distances and barycenters of persistenc... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to clustering graphs using persistence diagrams, combining topological and geometrical features. While the method leverages existing concepts, it introduces a novel algorithm that efficiently reduces the cost of matching topological f... |
0m4c9ZfDrDt | Generalizing Successor Features to continuous domains for Multi-task Learning | ICLR | 2,022 | train | false | The deep reinforcement learning (RL) framework has shown great promise to tackle sequential decision-making problems, where the agent learns to behave optimally through interactions with the environment and receiving rewards.
The ability of an RL agent to learn different reward functions concurrently has many benefits,... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis work looks at the use of successor features for solving simple continuous control tasks (in particular, reaching to different locations and door closing). The two contributions they enumerate are a ``practical implementation of SF framework for continu... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"I have some concerns about the novelty of this work.",
"Hunt et al. (2018) applied successor features to the max-ent RL framework.",
"Kulkarni et al. (2016) learned jointly φ and w.",
"Goal conditioned policies are an old idea.",
"However, the empirical results here are not particularly challenging tasks,"... | {
"novelty_excerpts": [
"I have some concerns about the novelty of this work.",
"Hunt et al. (2018) applied successor features to the max-ent RL framework.",
"Kulkarni et al. (2016) learned jointly φ and w.",
"Goal conditioned policies are an old idea.",
"However, the empirical results here are no... | -1 | The paper introduces a method that combines existing ideas in reinforcement learning, but it does not bring fundamentally new concepts or approaches. The work presents a practical implementation of successor features for continuous state and action domains, but this contribution lies in its application rather than its ... | The paper extends the use of successor features to continuous action spaces, though the extensions remain largely incremental. The method builds on existing ideas in goal-conditioned policies and actor-critic algorithms, and it does not significantly advance the field. The work presents a novel decomposition of success... | Novelty Score: -1
Score Justification:
The paper introduces a method that combines existing ideas in reinforcement learning, but it does not bring fundamentally new concepts or approaches. The work presents a practical implementation of successor features for continuous state and action domains, but this contribution ... |
0n1UvVzW99x | Synthetic Reduced Nearest Neighbor Model for Regression | ICLR | 2,022 | train | false | Nearest neighbor models are among the most established and accurate approaches to machine learning. In this paper, we investigate Synthetic Reduced Nearest Neighbor (SRNN) as a novel approach to regression tasks. Existing prototype nearest neighbor models are initialized by training a k-means model over each class. How... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe main contributions of this submission is to adapt Synthetic Reduced Nearest Neighbor (SRNN) for regression tasks. Technical details and comparative studies are provided accordingly.\n\n## Main Review\n[COMMENTS AFTER REBUTTAL]\n\nI have read the rebutt... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The model presented in this paper is new in the sense that it has never been proposed previously.",
"In my opinion, the main contribution of this submission is to adapt Synthetic Reduced Nearest Neighbor (SRNN) for regression tasks,",
"However, as I mentioned above, SRNN aims at decreasing memory and computati... | {
"novelty_excerpts": [
"The model presented in this paper is new in the sense that it has never been proposed previously.",
"In my opinion, the main contribution of this submission is to adapt Synthetic Reduced Nearest Neighbor (SRNN) for regression tasks,",
"However, as I mentioned above, SRNN aims at d... | 0 | The paper presents a competent adaptation of Synthetic Reduced Nearest Neighbor (SRNN) for regression tasks, though the contributions remain largely incremental. The work builds on existing ideas without introducing fundamentally new concepts or problem formulations. The method extends SRNN, but the extensions remain l... | The paper introduces a novel application of SRNN for regression tasks, leveraging the traditional method's ability to decrease memory and computation requirements. However, the motivation behind this adaptation is unclear, and the experimental design is criticized for being incomplete and poorly described. The work pre... | Novelty Score: 0
Score Justification:
The paper presents a competent adaptation of Synthetic Reduced Nearest Neighbor (SRNN) for regression tasks, though the contributions remain largely incremental. The work builds on existing ideas without introducing fundamentally new concepts or problem formulations. The method ex... |
0no8Motr-zO | An Experimental Design Perspective on Model-Based Reinforcement Learning | ICLR | 2,022 | train | false | In many practical applications of RL, it is expensive to observe state transitions from the environment. For example, in the problem of plasma control for nuclear fusion, computing the next state for a given state-action pair requires querying an expensive transition function which can lead to many hours of computer si... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper presents an active learning algorithm for model based reinforcement learning. The algorithm is based on a new acquisition function that finds the most informative transition to query an oracle model in order to achieve better performance with res... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Most model based RL approaches are based on data efficiency, so clearly the idea of applying active learning is interesting.",
"If we already have a transition oracle that we can query, what is the point of learning a transition model?",
"There is not a practical implementation in an RL setting.",
"The state... | {
"novelty_excerpts": [
"Most model based RL approaches are based on data efficiency, so clearly the idea of applying active learning is interesting.",
"If we already have a transition oracle that we can query, what is the point of learning a transition model?",
"There is not a practical implementation in... | 1 | The paper presents a moderately novel approach to model-based reinforcement learning, introducing an acquisition function based on expected information gain to efficiently query an oracle model. The method builds on existing ideas in active learning and model-based RL, but its combination and application to the specifi... | The paper introduces an acquisition function that approximates the optimal policy and computes information gain on the sequence of states visited by the optimal policy, which is a novel and clever approach to active learning in model-based RL. The method combines existing ideas in active learning and model-based RL, bu... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to model-based reinforcement learning, introducing an acquisition function based on expected information gain to efficiently query an oracle model. The method builds on existing ideas in active learning and model-based RL, but its com... |
0oSM3TC9Z5a | Learning to Persuade | ICLR | 2,022 | train | false | In the standard Bayesian persuasion model, an informed sender looks to design a signaling scheme to partially reveal the information to an uninformed receiver, so as to influence the behavior of the receiver. This kind of strategic interaction abounds in the real world. However, the standard model relies crucially on s... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper studies how to relax certain stringent assumptions made in a Bayesian persuasion framework. In particular, the paper looks to relax assumptions where the sender knows the receiver’s utility function and the receiver’s behavior is completely ration... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"However, to my best knowledge, simply relax either one of these assumptions is far from trivial.",
"I’d suggest the authors to focus on relaxing one of them, instead trying to relax them in a single framework.",
"The contributions are only marginally significant or novel.",
"The contributions are only margin... | {
"novelty_excerpts": [
"However, to my best knowledge, simply relax either one of these assumptions is far from trivial.",
"I’d suggest the authors to focus on relaxing one of them, instead trying to relax them in a single framework.",
"The contributions are only marginally significant or novel.",
"T... | -1 | The paper's approach to relaxing assumptions in Bayesian persuasion is largely incremental, building on existing methods without substantial conceptual or technical innovation. The use of machine learning to supplant unknown elements is not novel, as previous work has explored similar approaches. The contributions are ... | The paper introduces a framework for relaxing assumptions in Bayesian persuasion, but the approach is largely based on existing methods and does not significantly advance the field. The use of neural networks to simulate the necessary optimisation patterns is not novel, as previous work has explored similar approaches.... | Novelty Score: -1
Score Justification:
The paper's approach to relaxing assumptions in Bayesian persuasion is largely incremental, building on existing methods without substantial conceptual or technical innovation. The use of machine learning to supplant unknown elements is not novel, as previous work has explored si... |
0q0REJNgtg | Retrieval-Augmented Reinforcement Learning | ICLR | 2,022 | train | false | Most deep reinforcement learning (RL) algorithms distill experience into parametric behavior policies or value functions via gradient updates. While effective, this approach has several disadvantages: (1) it is computationally expensive, (2) it can take many updates to integrate experiences into the parametric model, (... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposed a state-augmentation strategy based on a fully differentiable model for online and offline reinforcement learning.\nThe authors evaluate the proposal on two algorithms, DQN and R2D2.\nThe augmentation through retrieval approach is decompo... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contribution is marginal, state augmentation is a reasonably classic topic and the end-2-end retrieval process could have been replaced by an attention model over the dataset of trajectories.",
"We would have hoped to see this baseline studied deeply to justify the approach.",
"In the introduction, the aut... | {
"novelty_excerpts": [
"The contribution is marginal, state augmentation is a reasonably classic topic and the end-2-end retrieval process could have been replaced by an attention model over the dataset of trajectories.",
"We would have hoped to see this baseline studied deeply to justify the approach.",
... | 0 | The paper presents a retrieval-based state augmentation approach for online and offline reinforcement learning, though the method builds on existing ideas and techniques. The contribution lies in its application and combination of known methods, rather than introducing fundamentally new concepts or approaches. The work... | The paper introduces a retrieval-based state augmentation approach that combines existing methods to improve reinforcement learning performance. However, the approach is largely an extension of existing techniques, and its novelty is somewhat limited by its reliance on established methods. The work presents a competent... | Novelty Score: 0
Score Justification:
The paper presents a retrieval-based state augmentation approach for online and offline reinforcement learning, though the method builds on existing ideas and techniques. The contribution lies in its application and combination of known methods, rather than introducing fundamental... |
0rcbOaoBXbg | Neural Spectral Marked Point Processes | ICLR | 2,022 | train | false | Self- and mutually-exciting point processes are popular models in machine learning and statistics for dependent discrete event data. To date, most existing models assume stationary kernels (including the classical Hawkes processes) and simple parametric models. Modern applications with complex event data require more g... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nIn this paper, the authors introduce a neural network framework modulating the excitation functions within the self and mutually exciting count processes.\n\n## Main Review\n\nI find that this paper has interesting ideas, and it addresses an important quest... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are only marginally significant or novel.",
"The contributions are only marginally significant or novel.",
"I find that this paper has interesting ideas, and it addresses an important question in learning self-exciting processes in the presence of non-stationarity."
] | {
"novelty_excerpts": [
"The contributions are only marginally significant or novel.",
"The contributions are only marginally significant or novel.",
"I find that this paper has interesting ideas, and it addresses an important question in learning self-exciting processes in the presence of non-stationarit... | 1 | The paper presents moderately novel ideas by introducing a neural network framework modulating the excitation functions within self and mutually exciting count processes. However, the contributions are largely incremental, building upon existing approaches with some minor variations. The work extends existing methods, ... | The method extends the self-exciting point process framework, though the extensions remain largely incremental. The paper introduces a neural network framework modulating the excitation functions, which is a moderately novel contribution. However, the approach builds upon existing ideas, such as the modulation of the e... | Novelty Score: 1
Score Justification:
The paper presents moderately novel ideas by introducing a neural network framework modulating the excitation functions within self and mutually exciting count processes. However, the contributions are largely incremental, building upon existing approaches with some minor variatio... |
0rjx6jy25R4 | Classify and Generate Reciprocally: Simultaneous Positive-Unlabelled Learning and Conditional Generation with Extra Data | ICLR | 2,022 | train | false | The scarcity of class-labeled data is a ubiquitous bottleneck in a wide range of machine learning problems. While abundant unlabeled data normally exist and provide a potential solution, it is extremely challenging to exploit them. In this paper, we address this problem by leveraging Positive-Unlabeled~(PU) classificat... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes to leverage Positive-Unlabeled (PU) learning and the conditional generation with extra unlabeled data to simultaneously improve classification and generation performance. The authors show that their method is especially effective for han... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"However, it seems to be the fact that utilizing PU and conditional generation is the unique contribution of this work.",
"Such a new perspective does provide new insights to the community and can motivate future research in the field of semi-supervised learning.",
"To this end, the authors designed the model o... | {
"novelty_excerpts": [
"However, it seems to be the fact that utilizing PU and conditional generation is the unique contribution of this work.",
"Such a new perspective does provide new insights to the community and can motivate future research in the field of semi-supervised learning.",
"To this end, th... | 0 | The paper presents a combination of Positive-Unlabeled learning and conditional generation to improve classification and generation performance, though this integration remains largely incremental. The work builds on existing approaches in semi-supervised learning and GANs, introducing a new type of GAN network, but th... | The paper introduces a novel framework for semi-supervised learning that combines Positive-Unlabeled classification and conditional generation, leveraging extra unlabeled data to improve performance. However, this integration is largely an extension of existing approaches in semi-supervised learning and GANs, and the m... | Novelty Score: 0
Score Justification:
The paper presents a combination of Positive-Unlabeled learning and conditional generation to improve classification and generation performance, though this integration remains largely incremental. The work builds on existing approaches in semi-supervised learning and GANs, introd... |
0sEIBFb4cs | Practical Adversarial Attacks on Brain--Computer Interfaces | ICLR | 2,022 | train | false | Deep learning has been widely employed in brain--computer interfaces (BCIs) to decode a subject's intentions based on recorded brain activities enabling direct interaction with computers and machines. BCI systems play a crucial role in motor rehabilitation and have recently experienced a significant market boost as con... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nAuthors study the vulnerability of brain-computer interface (BCI) systems from a security perspective, and present a physiologically plausible adversarial attack approach on BCI systems that can induce denial-of-service after being transmitted by external d... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"In this context, it is not really easy for me to understand the motivation of investigating adversarial attacks to induce DoS for a MI-BCI system in experiments, or at least it is not clearly put in the paper.",
"Going further, how to possibly overcome this vulnerability that is proposed in the paper is also not... | {
"novelty_excerpts": [
"In this context, it is not really easy for me to understand the motivation of investigating adversarial attacks to induce DoS for a MI-BCI system in experiments, or at least it is not clearly put in the paper.",
"Going further, how to possibly overcome this vulnerability that is propo... | 0 | The paper presents a study on adversarial attacks on brain-computer interface (BCI) systems, but its contributions are somewhat standard, showing minor variations or applications of known methods without substantial conceptual or technical innovation. | The paper introduces a method to generate more realistic, smoother representations of adversarial attacks on BCI systems, but this approach builds on existing methods for generating smooth adversarial examples for different electrical body signals. The work presents a study on the vulnerability of BCI systems to advers... | Novelty Score: 0
Score Justification:
The paper presents a study on adversarial attacks on brain-computer interface (BCI) systems, but its contributions are somewhat standard, showing minor variations or applications of known methods without substantial conceptual or technical innovation.
Detailed Assessment:
The pap... |
0sgntlpKDOz | Learning Graphon Mean Field Games and Approximate Nash Equilibria | ICLR | 2,022 | train | false | Recent advances at the intersection of dense large graph limits and mean field games have begun to enable the scalable analysis of a broad class of dynamical sequential games with large numbers of agents. So far, results have been largely limited to graphon mean field systems with continuous-time diffusive or jump dyna... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studies discrete time dense graph mean field games (GMFG). It discusses the settings of both finite-agent graph game and GMFG and shows that under mild Lipschitz continuous assumptions, the Nash equilibrium of the GMFG exists. It establishes that... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors should give some discussions on when the assumption holds or how to verify this assumption, especially the existence of the Lipschitz mapping ηΔ, which is an element chosen from a set mapping (e.g., how it is chosen).",
"More formal justifications of this algorithm is needed to understand how it’s de... | {
"novelty_excerpts": [
"The authors should give some discussions on when the assumption holds or how to verify this assumption, especially the existence of the Lipschitz mapping \u03b7\u0394, which is an element chosen from a set mapping (e.g., how it is chosen).",
"More formal justifications of this algorit... | 1 | The paper introduces a novel framework for discrete-time graphon mean-field games, extending existing work on mean-field games and graphon-based models. However, the contributions remain largely incremental, building upon existing approaches without substantial conceptual or technical innovation. | The paper presents a discrete-time graphon mean-field game (GMFG) framework as a limiting model to approximate large population dense graph games, establishing existence of graphon mean-field equilibrium (GMFE) and a graphon mean-field approximation result. The proposed learning algorithms and numerical experiments add... | Novelty Score: 1
Score Justification:
The paper introduces a novel framework for discrete-time graphon mean-field games, extending existing work on mean-field games and graphon-based models. However, the contributions remain largely incremental, building upon existing approaches without substantial conceptual or techn... |
0xiJLKH-ufZ | Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models | ICLR | 2,022 | train | false | Diffusion probabilistic models (DPMs) represent a class of powerful generative models. Despite their success, the inference of DPMs is expensive since it generally needs to iterate over thousands of timesteps. A key problem in the inference is to estimate the variance in each timestep of the reverse process. In this wo... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nIn Sec. 2, the authors review the general framework of Diffusion Probabilistic Models~(DPMs) with non-Markovian forward processes from [1] which has DDPMs [2] and DDIMs [1] as special cases. They emphasize that the variance of the reverse process, such as i... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The paper provides an important insight in DPMs and shows improved results for pretrained models by a simple post-processing technique.",
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.... | {
"novelty_excerpts": [
"The paper provides an important insight in DPMs and shows improved results for pretrained models by a simple post-processing technique.",
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant and somewhat new. Aspects of the contributions exis... | 1 | The paper introduces a novel post-processing technique for diffusion probabilistic models, which improves results for pretrained models by optimizing the variance of the reverse process. The work presents a new approach to estimating the optimal variance, which is a significant contribution to the field. However, the m... | The main novel contribution lies in the derivation of the optimal variance of the reverse process and its application to improve the performance of diffusion probabilistic models. The work extends existing methods by providing a more principled approach to estimating the variance, which is a crucial component of the re... | Novelty Score: 1
Score Justification:
The paper introduces a novel post-processing technique for diffusion probabilistic models, which improves results for pretrained models by optimizing the variance of the reverse process. The work presents a new approach to estimating the optimal variance, which is a significant co... |
0ze7XgWcYNV | Learning When and What to Ask: a Hierarchical Reinforcement Learning Framework | ICLR | 2,022 | train | false | Reliable AI agents should be mindful of the limits of their knowledge and consult humans when sensing that they do not have sufficient knowledge to make sound decisions. We formulate a hierarchical reinforcement learning framework for learning to decide when to request additional information from humans and what type o... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper presents a reinforcement learning framework where the agent is able to \"ask for help\" to receive from a human additional information to more easily solve the task. The paper is kinda Robotic Navigation-oriented by the way tasks and \"additional ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"it's uncertain if the paper presents any significant contribution due to major omissions on the related works list",
"there has been a large amount of works with the same high-level idea in the last years",
"most of the related works are simply neglected and not even mentioned in the manuscript",
"works from... | {
"novelty_excerpts": [
"it's uncertain if the paper presents any significant contribution due to major omissions on the related works list",
"there has been a large amount of works with the same high-level idea in the last years",
"most of the related works are simply neglected and not even mentioned in ... | 0 | The paper presents a competent extension of existing ideas in reinforcement learning, particularly in the area of transfer learning and action advising. However, the work builds upon a well-established framework without introducing fundamentally new concepts or approaches. The method extends existing methods, though th... | The paper introduces a hierarchical reinforcement learning framework that enables agents to seek help from assistants in complex tasks. The approach combines existing ideas in transfer learning and action advising, but does not significantly advance the field. The work presents a framework that can be extended further,... | Novelty Score: 0
Score Justification:
The paper presents a competent extension of existing ideas in reinforcement learning, particularly in the area of transfer learning and action advising. However, the work builds upon a well-established framework without introducing fundamentally new concepts or approaches. The met... |
1-YP2squpa7 | Deep learning via message passing algorithms based on belief propagation | ICLR | 2,022 | train | false | Message-passing algorithms based on the Belief Propagation (BP) equations constitute a well-known distributed computational scheme. It is exact on tree-like graphical models and has also proven to be effective in many problems defined on graphs with loops (from inference to optimization, from signal processing to clust... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\n[Summary] \n\nThis paper develops a class of fBP-based message-passing algorithms by adding a \n“reinforcement “ term to the BP equations and shows equivalent performance to the binary networks in experiments.\n\n## Main Review\n[Main Strengths] \n\nThis pa... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This paper develops a class of fBP-based message-passing algorithms by adding a'reinforcement'term to the BP equations and shows equivalent performance to the binary networks in experiments.",
"The fundamental shortcoming of this study is that, while the basic concept is intriguing, it does not appear to make a ... | {
"novelty_excerpts": [
"This paper develops a class of fBP-based message-passing algorithms by adding a'reinforcement'term to the BP equations and shows equivalent performance to the binary networks in experiments.",
"The fundamental shortcoming of this study is that, while the basic concept is intriguing, i... | 0 | The paper presents a method that builds on existing approaches to neural network training, but the contributions remain largely incremental. The work does not introduce fundamentally new ideas or approaches that significantly advance the field. | The paper introduces a class of fBP-based message-passing algorithms that show equivalent performance to binary networks in experiments. However, the method does not make a significant impact, much like various flavors of stochastic gradient descent (SGD). The work extends existing approaches to neural network training... | Novelty Score: 0
Score Justification:
The paper presents a method that builds on existing approaches to neural network training, but the contributions remain largely incremental. The work does not introduce fundamentally new ideas or approaches that significantly advance the field.
Detailed Assessment:
The paper intr... |
14F3fI6MGxX | A Generalized Weighted Optimization Method for Computational Learning and Inversion | ICLR | 2,022 | train | false | The generalization capacity of various machine learning models exhibits different phenomena in the under- and over-parameterized regimes. In this paper, we focus on regression models such as feature regression and kernel regression and analyze a generalized weighted least-squares optimization method for computational l... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper starts from the random Fourier model, generalizes the recent results on weighting the coefficients to adopt weighted least squares loss. Under different settings of noise levels and over-/under- parameterized regimes, error analysis has been prov... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prior works."
] | {
"novelty_excerpts": [
"The contributions are significant, and do not exist in prior works.",
"The contributions are significant, and do not exist in prior works."
]
} | 1 | The paper introduces a generalized weighted least-squares optimization method that allows weighting in both the parameter and data spaces, extending previous work on the random Fourier model. This approach presents a moderately novel contribution by combining existing ideas in a new way, but its significance is somewha... | The paper builds on the random Fourier model setting and generalizes the weighted least-squares loss to adopt weighted least squares loss, providing a solid analysis under different settings of noise levels and over-/under-parameterized regimes. The work presents a substantial extension of the work by Xie et al. (2020)... | Novelty Score: 1
Score Justification:
The paper introduces a generalized weighted least-squares optimization method that allows weighting in both the parameter and data spaces, extending previous work on the random Fourier model. This approach presents a moderately novel contribution by combining existing ideas in a n... |
14kbUbOaZUc | Metric Learning on Temporal Graphs via Few-Shot Examples | ICLR | 2,022 | train | false | Graph metric learning methods aim to learn the distance metric over graphs such that similar graphs are closer and dissimilar graphs are farther apart. This is of critical importance in many graph classification applications such as drug discovery and epidemics categorization. In many real-world applications, the graph... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis authors present a novel method for learning representations for time-varying graphs which allows for incorporating information at different time-scales using their streaming-snapshot model. The streaming-snapshot model has the following parts:\n* Each ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors present a novel method for learning representations for time-varying graphs which allows for incorporating information at different time-scales using their streaming-snapshot model.",
"This authors present a novel method for learning representations for time-varying graphs",
"First attempt to provi... | {"novelty_excerpts": [
"The authors present a novel method for learning representations for time-varying graphs which allows for incorporating information at different time-scales using their streaming-snapshot model.",
"This authors present a novel method for learning representations for time-varying graphs",
"First a... | 0 | The paper presents a competent extension of existing ideas in graph representation learning, combining established methodologies like GCNs and ProtoNets with novel time-aware node representations and attention mechanisms. However, the contributions remain largely incremental, building upon existing approaches without i... | The method extends existing graph representation learning techniques by incorporating time-aware node representations and attention mechanisms, allowing for the learning of dynamic graph structures. However, the approach remains largely based on established methodologies, with the primary novelty residing in its applic... | Novelty Score: 0
Score Justification:
The paper presents a competent extension of existing ideas in graph representation learning, combining established methodologies like GCNs and ProtoNets with novel time-aware node representations and attention mechanisms. However, the contributions remain largely incremental, buil... |
18Ys0-PzyPI | Online Ad Hoc Teamwork under Partial Observability | ICLR | 2,022 | train | false | Autonomous agents often need to work together as a team to accomplish complex cooperative tasks. Due to privacy and other realistic constraints, agents might need to collaborate with previously unknown teammates on the fly. This problem is known as ad hoc teamwork, which remains a core research challenge. Prior works u... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper addresses the problem of ad hoc teamwork, where an agent learns how to coordinate with a team of unknown teammates. Previous works typically assume that the teammates belong to one of a finite set of possible (known) types; at test time, the ad ho... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The proposed approach strikes me as novel in the context of ad hoc teamwork",
"It is also clearly different from the'main trend' on ad hoc teamwork, which mostly disregards partial observability.",
"My main objection is the assumption that---during training---the agent has full access to the other agent's and ... | {
"novelty_excerpts": [
"The proposed approach strikes me as novel in the context of ad hoc teamwork",
"It is also clearly different from the'main trend' on ad hoc teamwork, which mostly disregards partial observability.",
"My main objection is the assumption that---during training---the agent has full ac... | 1 | The paper introduces a novel approach to ad hoc teamwork by learning a latent representation of teammate behavior, which is a moderately novel contribution. The method builds on existing ideas in latent-variable modeling and CTDE MARL, but presents a new way of utilizing these concepts in the context of ad hoc teamwork... | The paper presents a moderately novel approach to ad hoc teamwork by learning a latent representation of teammate behavior. The method extends existing ideas in latent-variable modeling and CTDE MARL, but presents a new way of utilizing these concepts in the context of ad hoc teamwork. The approach is somewhat new, but... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to ad hoc teamwork by learning a latent representation of teammate behavior, which is a moderately novel contribution. The method builds on existing ideas in latent-variable modeling and CTDE MARL, but presents a new way of utilizing these con... |
1DUwCRNAbA | An Investigation into the Role of Author Demographics in ICLR Participation and Review | ICLR | 2,022 | train | false | As machine learning conferences grow rapidly, many are concerned that individuals will be left behind on the basis of traits such as gender and geography. We leverage historic ICLR submissions from 2017 to 2021 to investigate the impact of gender and country of origin both on representation and paper review outcomes a... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper conducts an observational study of disparities in the ICLR 2017-2021 reviewing process. The work centers around the analysis of disparities in reviewing scores and acceptance rates across gender, number of authors, countries, papers topics, and in... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are neither significant nor novel.",
"The paper is not well-motivated and does not seem to fit within the general scope of ICLR",
"A notable example is the work by Blau and Kahn: Blau, Francine D., and Lawrence M. Kahn. 2017. \"The Gender Wage Gap: Extent, Trends, and Explanations.\" Journal ... | {
"novelty_excerpts": [
"The contributions are neither significant nor novel.",
"The paper is not well-motivated and does not seem to fit within the general scope of ICLR",
"A notable example is the work by Blau and Kahn: Blau, Francine D., and Lawrence M. Kahn. 2017. \"The Gender Wage Gap: Extent, Trends... | -1 | The paper presents an analysis of disparities in the ICLR reviewing process, but its findings largely overlap with those of a similar study by Tran et al. (2020). The work does not introduce fundamentally new ideas or approaches, instead relying on exploratory data analysis and regression results that uncover associati... | The paper introduces an observational study of disparities in the ICLR reviewing process, but its contributions are largely incremental, building on existing work by Tran et al. (2020). The analysis focuses on univariate patterns and does not account for potential confounding variables, which limits its ability to esta... | Novelty Score: -1
Score Justification:
The paper presents an analysis of disparities in the ICLR reviewing process, but its findings largely overlap with those of a similar study by Tran et al. (2020). The work does not introduce fundamentally new ideas or approaches, instead relying on exploratory data analysis and r... |
1HxTO6CTkz | Unifying Likelihood-free Inference with Black-box Optimization and Beyond | ICLR | 2,022 | train | false | Black-box optimization formulations for biological sequence design have drawn recent attention due to their promising potential impact on the pharmaceutical industry. In this work, we propose to unify two seemingly distinct worlds: likelihood-free inference and black-box optimization, under one probabilistic framework.... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper relates likelihood free inference to methods for biological sequences design and proposes new sequence design methods based on this insight.\n\n## Main Review\n# Major comments\n1) The paper is missing a related works section with an overview of e... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The outlined relation between likelihood free inference and sequence design (blackbox optimization) is interesting; I am not aware of any existing papers with this insight.",
"It is not described clearly enough how the proposed methods differ from existing methods for sequence design such as Bayesian model-based... | {
"novelty_excerpts": [
"The outlined relation between likelihood free inference and sequence design (blackbox optimization) is interesting; I am not aware of any existing papers with this insight.",
"It is not described clearly enough how the proposed methods differ from existing methods for sequence design such as Ba... | 1 | The paper introduces a novel connection between likelihood-free inference and black-box sequence design, which is moderately novel as it builds on existing ideas but presents a fresh perspective. The work presents several new methods that combine insights from these two fields, though the extensions remain largely incr... | The paper's main novel contribution is the establishment of a connection between likelihood-free inference and black-box sequence design, which allows for the adaptation of existing methods from one field to the other. This connection is moderately novel, as it builds on existing ideas but presents a fresh perspective ... | Novelty Score: 1
The paper introduces a novel connection between likelihood-free inference and black-box sequence design, which is moderately novel as it builds on existing ideas but presents a fresh perspective. The work presents several new methods that combine insights from these two fields, though the extensions r... |
1NvflqAdoom | Neural Networks as Kernel Learners: The Silent Alignment Effect | ICLR | 2,022 | train | false | Neural networks in the lazy training regime converge to kernel machines. Can neural networks in the rich feature learning regime learn a kernel machine with a data-dependent kernel? We demonstrate that this can indeed happen due to a phenomenon we term silent alignment, which requires that the tangent kernel of a netwo... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nIn this paper the authors show that in some cases, a neural network, trained in the rich regime, can behave like a kernel predictor where the kernel is the neural tangent kernel at the end of training. Specifically, this happens when the kernel changes befo... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The authors identify an interesting silent alignment effect, where the trained network acts as kernel predictor, and provide a thorough analysis for 2-layers linear networks and deep linear networks.",
"I think the authors should contrast their results with recent results on implicit bias of linear networks with... | {
"novelty_excerpts": [
"The authors identify an interesting silent alignment effect, where the trained network acts as kernel predictor, and provide a thorough analysis for 2-layers linear networks and deep linear networks.",
"I think the authors should contrast their results with recent results on implicit ... | 1 | The paper introduces a novel phenomenon called the "silent alignment effect," where neural networks behave like kernel predictors under specific conditions. This contribution is moderately novel, as it builds upon existing work in neural tangent kernels but provides a new perspective on the learning dynamics of neural ... | The main novel contribution lies in the identification and characterization of the silent alignment effect, which reveals a new aspect of neural network behavior. The work presents a thorough analysis of this phenomenon in linear networks and demonstrates its existence in non-linear networks as well. However, the analy... | Novelty Score: 1
Score Justification:
The paper introduces a novel phenomenon called the "silent alignment effect," where neural networks behave like kernel predictors under specific conditions. This contribution is moderately novel, as it builds upon existing work in neural tangent kernels but provides a new perspect... |
1O5UK-zoK8g | Adaptive Generalization for Semantic Segmentation | ICLR | 2,022 | train | false | Out-of-distribution robustness remains a salient weakness of current state-of-the-art models for semantic segmentation. Until recently, research on generalization followed a restrictive assumption that the model parameters remain fixed after the training process. In this work, we empirically study an adaptive inference... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studies an existing problem - domain generalization for semantic segmentation of urban scenes. The technique keys of this work has two main points: instance-adaptive batch normalization and testing-time training using pseudo labels. Instance-adap... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The novelty of this work is the main drawback.",
"The first point instance-adaptive normalization is a trivial solution, which is the tradeoff of previous t-BN and p-BN.",
"Second, there are a lot of work applying pseudo labels in semantic segmentation, such as Naive-students [ECCV2020].",
"Even though this ... | {
"novelty_excerpts": [
"The novelty of this work is the main drawback.",
"The first point instance-adaptive normalization is a trivial solution, which is the tradeoff of previous t-BN and p-BN.",
"Second, there are a lot of work applying pseudo labels in semantic segmentation, such as Naive-students [ECC... | 0 | The paper presents a tradeoff between existing instance normalization techniques, and its method for generating pseudo labels for test-time training is not novel. The work shows better results than previous methods but lacks significant originality. | The paper introduces instance-adaptive batch normalization, which is a tradeoff between previous t-BN and p-BN techniques. The method for generating pseudo labels for test-time training is not novel, as it has been applied in other works such as Naive-students [ECCV2020]. The paper's main contribution lies in its abili... | Novelty Score: 0
Score Justification:
The paper presents a tradeoff between existing instance normalization techniques, and its method for generating pseudo labels for test-time training is not novel. The work shows better results than previous methods but lacks significant originality.
Detailed Assessment:
The paper... |
1OHZX4YDqhT | FedNAS: Federated Deep Learning via Neural Architecture Search | ICLR | 2,022 | train | false | Federated Learning (FL) is an effective learning framework used when data cannotbe centralized due to privacy, communication costs, and regulatory restrictions.While there have been many algorithmic advances in FL, significantly less effort hasbeen made on model development, and most works in FL employ predefined model... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a method for automatic model design in the context of federated learning and introduces an auto-ml system under the framework of federated learning.\n\n\n\n## Main Review\nStrength:\n1. The paper is generally clearly written.\n2. The cod... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The motivation is not clear. The abstract says 'We propose Federated Neural Architecture Search (FedNAS) for automating the model design process in FL.' and the introduction says 'We aim to address data heterogeneity in FL via... NAS.' For me, it is not clear whether the paper aims to automate model design in FL o... | {"novelty_excerpts": [
"The motivation is not clear. The abstract says 'We propose Federated Neural Architecture Search (FedNAS) for automating the model design process in FL.' and the introduction says 'We aim to address data heterogeneity in FL via... NAS.' For me, it is not clear whether the paper aims to automate m... | 0 | The paper presents a method that combines existing neural architecture search (NAS) and federated learning (FL) approaches, though the combination remains largely incremental. The work extends existing NAS methods to the FL setting, introducing some personalization but without significant conceptual or technical innova... | The paper introduces a method that combines NAS and FL to address data heterogeneity in FL, though this combination is not novel in itself. The approach builds on existing NAS methods, such as MiLeNAS, and extends them to the FL setting, introducing some personalization. However, the method's novelty is limited by its ... | Novelty Score: 0
Score Justification:
The paper presents a method that combines existing neural architecture search (NAS) and federated learning (FL) approaches, though the combination remains largely incremental. The work extends existing NAS methods to the FL setting, introducing some personalization but without sig... |
1QxveKM654 | Genome Sequence Reconstruction Using Gated Graph Convolutional Network | ICLR | 2,022 | train | false | A quest to determine the human DNA sequence from telomere to telomere started three decades ago and was finally finished in 2021. This accomplishment was a result of a tremendous effort of numerous experts with an abundance of data, various tools, and often included manual inspection during genome reconstruction. There... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors propose a graph neural network approach for genome assembly. In the framework of overlap-layout-consensus, a graph structure is formed in which reads are nodes and overlaps between them are edges. In theory, one desires a Hamiltonian path that v... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This paper has nontrivial contributions on genome assembly area, and I think this work is promising.",
"it would be great if the authors address issues as follows:",
"I think this paper is one of the applications or the extensions of existing work [1], [2], [3], [4].",
"the authors should supplement other ac... | {
"novelty_excerpts": [
"This paper has nontrivial contributions on genome assembly area, and I think this work is promising.",
"it would be great if the authors address issues as follows:",
"I think this paper is one of the applications or the extensions of existing work [1], [2], [3], [4].",
"the au... | 0 | The paper presents a competent extension of existing ideas in the field of genome assembly, leveraging graph neural networks to improve the process. However, the contributions remain largely incremental, building upon established approaches without introducing fundamentally new concepts. The work's novelty is further d... | The paper introduces a novel application of graph neural networks to the genome assembly problem, extending the use of neural networks in this area. However, the method's reliance on simulated data and the lack of comprehensive results limit its novelty. The work builds on existing approaches, such as the Overlap-Layou... | Novelty Score: 0
Score Justification:
The paper presents a competent extension of existing ideas in the field of genome assembly, leveraging graph neural networks to improve the process. However, the contributions remain largely incremental, building upon established approaches without introducing fundamentally new co... |
1R_PRbQK2eu | Dual Training of Energy-Based Models with Overparametrized Shallow Neural Networks | ICLR | 2,022 | train | false | Energy-based models (EBMs) are generative models that are usually trained via maximum likelihood estimation. This approach becomes challenging in generic situations where the trained energy is nonconvex, due to the need to sample the Gibbs distribution associated with this energy. Using general Fenchel duality results,... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper studies the training of Energy Based Models (EBM). The “standard” technique is maximum likelihood maximization of the observed dataset. While theoretically sound, there are practical difficulties in simulating the MCMC dynamics due to large basin... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The connection on a continuum of maximum likelihood and score based training is very interesting from a theoretical point of view, and I think this is a novel and valuable contribution.",
"what I really miss is an implementation of the idea outlined at the end of section 4, which is what the authors refer to as ... | {
"novelty_excerpts": [
"The connection on a continuum of maximum likelihood and score based training is very interesting from a theoretical point of view, and I think this is a novel and valuable contribution.",
"what I really miss is an implementation of the idea outlined at the end of section 4, which is w... | 1 | The paper introduces a novel connection on a continuum between maximum likelihood and score-based training, which is a valuable contribution to the field. However, the work builds upon existing ideas and concepts, and its originality is somewhat limited by its reliance on prior research. The contribution lies in its th... | The paper presents a theoretical framework that connects maximum likelihood and score-based training, leveraging known results about Fenchel dualities. This connection is a significant advancement, as it provides a new perspective on the relationship between these two training modalities. However, the work's originalit... | Novelty Score: 1
Score Justification:
The paper introduces a novel connection on a continuum between maximum likelihood and score-based training, which is a valuable contribution to the field. However, the work builds upon existing ideas and concepts, and its originality is somewhat limited by its reliance on prior re... |
1T5FmILBsq2 | SGORNN: Combining Scalar Gates and Orthogonal Constraints in Recurrent Networks | ICLR | 2,022 | train | false | Recurrent Neural Network (RNN) models have been applied in different domains, producing high accuracies on time-dependent data. However, RNNs have long suffered from exploding gradients during training, mainly due to their recurrent process. In this context, we propose a variant of the scalar gated FastRNN architecture... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis work proposes to approach and partially solve (under certain constraints) the exploding gradient problem of RNN. In practice, the authors combined two existing methods (residual connection + unitary RNNs) in a single novel model. Demonstrations are giv... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"In practice, the authors combined two existing methods (residual connection + unitary RNNs) in a single novel model.",
"SGORNN outperform models that are based either on residual connection or unitary recurrent connections only.",
"Indeed, the implementation of this solution (such as given in supplementary mat... | {
"novelty_excerpts": [
"In practice, the authors combined two existing methods (residual connection + unitary RNNs) in a single novel model.",
"SGORNN outperform models that are based either on residual connection or unitary recurrent connections only.",
"Indeed, the implementation of this solution (such... | 0 | null | null | Novelty Score: 0
The paper presents a limited novelty as it combines existing methods (residual connection + unitary RNNs) in a single model, which is a somewhat standard approach in the field. The work extends existing ideas, but the extensions remain largely incremental. The contribution lies in its application and ... |
1W0z96MFEoH | Benchmarking the Spectrum of Agent Capabilities | ICLR | 2,022 | train | false | Evaluating the general abilities of intelligent agents requires complex simulation environments. Existing benchmarks typically evaluate only one narrow task per environment, requiring researchers to perform expensive training runs on many different environments. We introduce Crafter, an open world survival game with vi... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe work introduces a gridworld-style benchmark inspired by Minecraft. It is similar in style to MineRL, but is significantly more computationally efficient, and easier to use (can be run directly without requiring e.g. an external Java application). Despit... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contribution seems useful.",
"Having an environment designed specifically for RL in mind is also useful in terms of permitting modifications in the future, and enabling debugging by directly accessing internal state.",
"I expect many researchers to continue to prefer MineRL.",
"It's significantly less co... | {
"novelty_excerpts": [
"The contribution seems useful.",
"Having an environment designed specifically for RL in mind is also useful in terms of permitting modifications in the future, and enabling debugging by directly accessing internal state.",
"I expect many researchers to continue to prefer MineRL.",... | 1 | The paper introduces a novel contribution, a gridworld-style benchmark inspired by Minecraft, which is computationally efficient and easier to use than MineRL. However, the environment design is not entirely novel, as it is similar to prior work, particularly Mine RL. The paper also presents a new RL environment, Craft... | The main novel contributions or new ideas introduced are the gridworld-style benchmark inspired by Minecraft and the Crafter environment, which have some interesting aspects, such as the need to find food, water, shelter, and sleep. The Crafter environment is designed to support learning of general policies, and its pr... | Novelty Score: 1
Score Justification:
The paper introduces a novel contribution, a gridworld-style benchmark inspired by Minecraft, which is computationally efficient and easier to use than MineRL. However, the environment design is not entirely novel, as it is similar to prior work, particularly Mine RL. The paper al... |
1Z5P--ntu8 | On the Global Convergence of Gradient Descent for multi-layer ResNets in the mean-field regime | ICLR | 2,022 | train | false | Finding the optimal configuration of parameters in ResNet is a nonconvex minimization problem, but first-order methods nevertheless find the global optimum in the overparameterized regime. We study this phenomenon with mean-field analysis, by translating the training process of ResNet to a gradient-flow partial differe... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proves a global convergence result for GD in resnets, whose residual blocks are mean field two-layer neural nets. This is taken with a double limit, where the depth limit is the neural ODE and the width limit is the two-layer MF limit. To prove gl... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"As said in my review, there are reasons to believe why the contribution of this paper is a clearly meaningful one on top of prior works in similar resnet mean field setups, namely the removal of the full support assumption on $\\rho_{t=\\infty}$.",
"This is a technical paper that removes a strong assumption from... | {
"novelty_excerpts": [
"As said in my review, there are reasons to believe why the contribution of this paper is a clearly meaningful one on top of prior works in similar resnet mean field setups, namely the removal of the full support assumption on $\rho_{t=\infty}$.",
"This is a technical paper that removes a strong... | 0 | The paper presents a technically sound contribution that removes a strong assumption from the analysis of Lu et al 2020, though the techniques are not surprising. The work builds on existing approaches and extends them to a specific problem setup, but the novelty lies in the technical execution rather than the conceptu... | The paper introduces a rigorous proof for the global convergence of gradient descent in ResNets, removing the full support assumption on $\rho_{t=\infty}$. The work extends existing results in the mean-field regime to multilayer networks, but the extensions remain largely incremental. The approach combines existing tec... | Novelty Score: 0
Score Justification:
The paper presents a technically sound contribution that removes a strong assumption from the analysis of Lu et al 2020, though the techniques are not surprising. The work builds on existing approaches and extends them to a specific problem setup, but the novelty lies in the techn... |
1Zxv7TdLquI | YOUR AUTOREGRESSIVE GENERATIVE MODEL CAN BE BETTER IF YOU TREAT IT AS AN ENERGY-BASED ONE | ICLR | 2,022 | train | false | Autoregressive generative models are commonly used, especially for those tasks involving sequential data. They have, however, been plagued by a slew of inherent flaws due to the intrinsic characteristics of chain-style conditional modeling (e.g., exposure bias or lack of long-range coherence), severely limiting their a... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe authors propose an approach to integrate energy based models (EBM) into into general auto-regressive generative models ARGM. They propose to combine the negative log-likelihood loss of auto-regressive models with a second energy based loss, which serves... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The direction of joint training of AR model and EBM is promising."
] | {
"novelty_excerpts": [
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior work.",
"The direction of joint training of AR model and EBM is promising."
]
} | 1 | The paper introduces a novel approach to integrating energy-based models (EBM) into general auto-regressive generative models (ARGM), which mitigates exposure bias. However, the method builds on existing ideas and does not demonstrate significant superiority over other methods. The contribution lies in its competent ex... | The paper presents a moderately novel approach to addressing exposure bias in ARGM by combining them with EBM. The method extends the capabilities of ARGM by incorporating EBM, though the extensions remain largely incremental. The work does not introduce fundamentally new ideas or problem formulations but rather applie... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to integrating energy-based models (EBM) into general auto-regressive generative models (ARGM), which mitigates exposure bias. However, the method builds on existing ideas and does not demonstrate significant superiority over other methods. Th... |
1bEaEzGwfhP | Learning to Model Editing Processes | ICLR | 2,022 | train | false | Most existing sequence generation models produce outputs in one pass, usually left-to-right. However, this is in contrast with a more natural approach that humans use in generating content; iterative refinement and editing. Recent work has introduced edit-based models for various tasks (such as neural machine translati... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nExisting edit-based models condition on a single sentence or document. This paper argues that such frameworks do not represent processes well in which the same document undergoes multiple iterations of revisions, for example the edit histories of Wikipedia ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Previous approaches only consider predicting editing based on the current text (therefore just one revision history). This paper considers more revision to predict and it show more revisions (up to three) are beneficial.",
"One of the metrics proposed in the paper is slightly strange. Why to measure the perplexi... | {
"novelty_excerpts": [
"Previous approaches only consider predicting editing based on the current text (therefore just one revision history). This paper considers more revision to predict and it show more revisions (up to three) are beneficial.",
"One of the metrics proposed in the paper is slightly strange.... | 1 | The paper introduces a novel approach to modeling editing processes by conditioning on multiple revisions, which has the potential to improve the accuracy of editing process modeling. However, the evaluation does not convincingly demonstrate this novelty, and the contributions remain somewhat incremental. | The paper presents a technically sound and novel method for modeling iterative editing processes, which is a significant departure from previous single-step editing approaches. However, the method's novelty is somewhat diminished by its reliance on existing metrics and the lack of comparisons with single-step edit mode... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to modeling editing processes by conditioning on multiple revisions, which has the potential to improve the accuracy of editing process modeling. However, the evaluation does not convincingly demonstrate this novelty, and the contributions rem... |
1gEb_H1DEqZ | Logic Pre-Training of Language Models | ICLR | 2,022 | train | false | Pre-trained language models (PrLMs) have been shown useful for enhancing a broad range of natural language understanding (NLU) tasks. However, the capacity for capturing logic relations in challenging NLU still remains a bottleneck even for state-of-the-art PrLM enhancement, which greatly stalled their reasoning abilit... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes to add the dependency parsing information of sentences as part of the pre-training objectives to the pre-training of BERT. In particular, it designs three pre-training tasks: logical connective masking, logical structure completion, and ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The main drawback of this paper is that it does not convince me that dependency parsing is a very important additional information for pre-training language models.",
"According to previous studies [1], the pre-training of BERT actually already contains morphology, syntax, and semantics information.",
"And the... | {
"novelty_excerpts": [
"The main drawback of this paper is that it does not convince me that dependency parsing is a very important additional information for pre-training language models.",
"According to previous studies [1], the pre-training of BERT actually already contains morphology, syntax, and semanti... | 0 | The paper presents a method that combines existing ideas and techniques to enhance pre-trained language models with logical reasoning capabilities. The approach builds on prior work in syntactic parsing and pre-training objectives, but does not introduce fundamentally new ideas or significantly advance the field. | The paper introduces a new pre-training objective that incorporates logical structure completion, logical connective masking, and logical path prediction. However, these objectives are largely based on existing methods and do not significantly extend the state-of-the-art. The work presents a competent extension of exis... | Novelty Score: 0
Score Justification:
The paper presents a method that combines existing ideas and techniques to enhance pre-trained language models with logical reasoning capabilities. The approach builds on prior work in syntactic parsing and pre-training objectives, but does not introduce fundamentally new ideas or... |
1kqWZlj4QYJ | Learning Two-Step Hybrid Policy for Graph-Based Interpretable Reinforcement Learning | ICLR | 2,022 | train | false | We present a two-step hybrid reinforcement learning (RL) policy that is designed to generate interpretable and robust hierarchical policies
on the RL problem with graph-based input. Unlike prior deep reinforcement learning policies parameterized by an end-to-end black-box graph neural network, our approach disentangles... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper introduces a two-step method for RL problems with graph-based inputs. This method trains a teacher policy which it then uses to construct a policy \"pruner\" which predicts the kind of action needed. It then uses a ideas from retrieval to select ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"I am worried that this framework is not very generalizable outside of this specific problem.",
"This method would not work at all in any other kind of environment, such as one with visual perception instead of symbolic input, in continuous/physics based environments, or even games such as Atari where there aren'... | {
"novelty_excerpts": [
"I am worried that this framework is not very generalizable outside of this specific problem.",
"This method would not work at all in any other kind of environment, such as one with visual perception instead of symbolic input, in continuous/physics based environments, or even games suc... | 1 | The paper presents a moderately novel approach to improving the interpretability of agents trained on reinforcement learning problems with graph-based inputs. The method introduces a two-step hybrid model, combining action pruning and selection, which shows promise in generalizing to unseen environments and being more ... | The main novel contributions of the paper lie in its two-step hybrid model, which reduces the search space by breaking down potential actions into groups sharing similar semantics. This approach shows some originality in its application to graph-based reinforcement learning problems, particularly in the context of text... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to improving the interpretability of agents trained on reinforcement learning problems with graph-based inputs. The method introduces a two-step hybrid model, combining action pruning and selection, which shows promise in generalizing... |
1uf_kj0GUF- | Nonparametric Learning of Two-Layer ReLU Residual Units | ICLR | 2,022 | train | false | We describe an algorithm that learns two-layer residual units using rectified linear unit (ReLU) activation: suppose the input $\mathbf{x}$ is from a distribution with support space $\mathbb{R}^d$ and the ground-truth generative model is a residual unit of this type, given by $\mathbf{y} = \boldsymbol{B}^\ast\left[\lef... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\n\nThis paper considers the special case of two-layer ReLU residual units and introduces a nonparametric convex approach to optimize the training objective. The experiments also highlight the effectiveness of the proposed approach with respect to SGD applied... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"I think that the main contribution of this paper is incremental due to the following reasons.",
"First of all, the model seems to be so simple that it even doesn't allow permutation and/or scaling invariance.",
"Moreover, the paper lacks satisfactory comparative experimental analysis.",
"The contributions ar... | {
"novelty_excerpts": [
"I think that the main contribution of this paper is incremental due to the following reasons.",
"First of all, the model seems to be so simple that it even doesn't allow permutation and/or scaling invariance.",
"Moreover, the paper lacks satisfactory comparative experimental analy... | 1 | The paper presents a moderately novel approach to learning two-layer ReLU residual networks, introducing a nonparametric convex optimization method that outperforms SGD. While the method builds on existing ideas, it offers a new perspective on the problem and demonstrates its effectiveness through extensive experimenta... | The main novel contribution of the paper lies in its nonparametric convex optimization approach, which allows for the learning of two-layer ReLU residual networks without making strong parametric assumptions. This method extends existing work on convex optimization for neural networks, though the extensions remain larg... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to learning two-layer ReLU residual networks, introducing a nonparametric convex optimization method that outperforms SGD. While the method builds on existing ideas, it offers a new perspective on the problem and demonstrates its effe... |
1ugNpm7W6E | Cold Brew: Distilling Graph Node Representations with Incomplete or Missing Neighborhoods | ICLR | 2,022 | train | false | Graph Neural Networks (GNNs) have achieved state-of-the-art performance in node classification, regression, and recommendation tasks. GNNs work well when rich and high-quality connections are available. However, their effectiveness is often jeopardized in many real-world graphs in which node degrees have power-law dist... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe work focuses on a very practical problem of strict cold start(SCS) recommendations which is a highly prevalent and relevant problem. The author’s main contribution to address the SCS is to use GNN with knowledge distillation – this proposed solution doe... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The proposed solution is technically sound and is well supported by the extensive evaluation presented.",
"The proposed method to adopt GNN with knowledge distillation to solve SCS problem relies on well-known concepts and previous works, but the learnable Structural Embedding and model selection methodology is ... | {
"novelty_excerpts": [
"The proposed solution is technically sound and is well supported by the extensive evaluation presented.",
"The proposed method to adopt GNN with knowledge distillation to solve SCS problem relies on well-known concepts and previous works, but the learnable Structural Embedding and mod... | 1 | The paper introduces a novel approach to learning cold-start node embeddings in graphs using a knowledge-distillation framework, which combines neighbor aggregator and structure embedding to learn node embeddings. However, the method builds upon existing techniques and does not significantly advance the field. The cont... | The paper presents a technically sound method for addressing the cold-start problem in graph-based recommendations. The learnable Structural Embedding and model selection methodology is a novel contribution, but its practical adoption is largely incremental. The approach combines existing techniques, such as teacher-st... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to learning cold-start node embeddings in graphs using a knowledge-distillation framework, which combines neighbor aggregator and structure embedding to learn node embeddings. However, the method builds upon existing techniques and does not si... |
1v1N7Zhmgcx | Maximum Likelihood Training of Parametrized Diffusion Model | ICLR | 2,022 | train | false | Whereas the diverse variations of the diffusion model exist in image synthesis, the previous variations have not innovated the diffusing mechanism by maintaining the static linear diffusion. Meanwhile, it is intuitive that there would be more promising diffusion pattern adapted to the data distribution. This paper intr... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe author proposed a framework that combines normalizing flow with diffusion models, allowing non-linear diffusions to be used for inference and generation. Specifically, the author first discussed the similarity between VAE and diffusion models to motivat... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The only difference is the replacement of the encoder and decoder with normalizing flows.",
"What are the actual empirical advantages?",
"Does PDM reduce the required number of generation steps like LSGM compared to typical linear diffusions?",
"It seems that LSGM produces lower FID scores compared to PDM.",... | {
"novelty_excerpts": [
"The only difference is the replacement of the encoder and decoder with normalizing flows.",
"What are the actual empirical advantages?",
"Does PDM reduce the required number of generation steps like LSGM compared to typical linear diffusions?",
"It seems that LSGM produces lower FID scores comp... | 1 | The paper introduces a novel approach that combines normalizing flow with diffusion models, allowing non-linear diffusions to be used for inference and generation. This combination presents a moderately novel contribution, as it builds upon existing ideas in diffusion models and normalizing flows. However, the work als... | The method extends the concept of diffusion models by incorporating normalizing flows, which enables non-linear diffusions. This approach presents a moderately novel contribution, as it combines existing ideas in a new way. The work also shows some limitations in originality, as it is largely an extension of existing a... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach that combines normalizing flow with diffusion models, allowing non-linear diffusions to be used for inference and generation. This combination presents a moderately novel contribution, as it builds upon existing ideas in diffusion models and n... |
1wVvweK3oIb | Simple GNN Regularisation for 3D Molecular Property Prediction and Beyond | ICLR | 2,022 | train | false | In this paper we show that simple noisy regularisation can be an effective way to address oversmoothing. We first argue that regularisers ad-dressing oversmoothing should both penalise node latent similarity and encourage meaningful node representations. From this observation we derive “Noisy Nodes”,a simple technique ... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a noise-injected training mechanism for graph neural network. The idea is to perturb the node attributes and add an auxiliary loss for the noise-correction task. The authors. Implementing this strategy for the GNS model And observed impr... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The method is pretty simple, but the results are very positive on several benchmarks, which has achieved improvements over several baselines.",
"The proposed idea is intuitive and sound, and the results are positive.",
"However, important baselines and detailed experiments needs to be added to support the cla... | {
"novelty_excerpts": [
"The method is pretty simple, but the results are very positive on several benchmarks, which has achieved improvements over several baselines.",
"The proposed idea is intuitive and sound, and the results are positive.",
"However, important baselines and detailed experiments needs ... | 1 | The paper introduces a simple regularization method for Graph Neural Networks, building on existing ideas from other domains, but its application to GNNs is novel. The method shows improved performance on various tasks and datasets, though its extensions remain largely incremental. The contribution lies in its practica... | The paper presents a novel approach to reducing oversmoothing in Graph Neural Networks through the introduction of a denoising mechanism, called Noisy Nodes (NN). This method adds a small noise to the input node representation and an auxiliary loss for predicting the uncorrupted node representation. The paper demonstra... | Novelty Score: 1
Score Justification:
The paper introduces a simple regularization method for Graph Neural Networks, building on existing ideas from other domains, but its application to GNNs is novel. The method shows improved performance on various tasks and datasets, though its extensions remain largely incremental... |
1xXvPrAshao | Learning Multimodal VAEs through Mutual Supervision | ICLR | 2,022 | train | false | Multimodal VAEs seek to model the joint distribution over heterogeneous data (e.g.\ vision, language), whilst also capturing a shared representation across such modalities. Prior work has typically combined information from the modalities by reconciling idiosyncratic representations directly in the recognition model th... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper presents an approach to model the joint distribution over data from different modalities such as vision and language. It seeks to address a limitation that is common to prior work on multimodal VAEs, which have typically combined information acro... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Experimental results show that the proposed method outperforms the conventional methods, MVAE and MMVAE, in both partial and complete settings.",
"However, why does MEME, which is just an extension of semi-supervised VAEs, perform better than MVAE and MMVAE even in the complete setting?",
"What parts of MEME a... | {
"novelty_excerpts": [
"Experimental results show that the proposed method outperforms the conventional methods, MVAE and MMVAE, in both partial and complete settings.",
"However, why does MEME, which is just an extension of semi-supervised VAEs, perform better than MVAE and MMVAE even in the complete settin... | 1 | The paper introduces a novel formulation of multimodal VAEs that circumvents the need for explicit combinations of modalities. However, the contributions are somewhat new but also build upon existing ideas, showing significant overlap with prior work. The approach presents a competent extension of existing methods, wit... | The paper introduces a novel formulation of multimodal VAEs that exploits mutual supervision between different modalities. This approach allows for the joint distribution to be modeled under partially observed settings, which is a significant contribution. However, the method extends existing semi-supervised VAEs and m... | Novelty Score: 1
Score Justification:
The paper introduces a novel formulation of multimodal VAEs that circumvents the need for explicit combinations of modalities. However, the contributions are somewhat new but also build upon existing ideas, showing significant overlap with prior work. The approach presents a compe... |
2-mkiUs9Jx7 | Stein Latent Optimization for Generative Adversarial Networks | ICLR | 2,022 | train | false | Generative adversarial networks (GANs) with clustered latent spaces can perform conditional generation in a completely unsupervised manner. In the real world, the salient attributes of unlabeled data can be imbalanced. However, most of existing unsupervised conditional GANs cannot cluster attributes of these data in th... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper uses Gaussian mixture as a prior to address the scenario of imbalanced attributes in GANs latent space clustering. The study derives Stein latent optimization that provides reparameterizable gradient estimations when assuming a Gaussian mixture p... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The idea of using Gaussian mixture as a prior for imbalanced attributes is intuitive and technically sound.",
"In the theoretical aspect, this paper derives Stein latent optimization based on Stein’s lemma to estimate the gradient of parameters.",
"Compared with ClusterGAN, this paper address the importance of... | {"novelty_excerpts": [
"The idea of using Gaussian mixture as a prior for imbalanced attributes is intuitive and technically sound.",
"In the theoretical aspect, this paper derives Stein latent optimization based on Stein’s lemma to estimate the gradient of parameters.",
"Compared with ClusterGAN, this paper address th... | 1 | The paper introduces a novel approach to addressing imbalanced attributes in GANs latent space clustering by using a Gaussian mixture as a prior and deriving Stein latent optimization. However, the contributions are somewhat limited in their originality as they build upon existing ideas and methods. The work presents a... | The paper introduces a novel approach to addressing imbalanced attributes in GANs latent space clustering by using a Gaussian mixture as a prior and deriving Stein latent optimization. This method extends existing ideas in GANs and clustering, but the novelty lies in its application to unsupervised conditional GANs. Th... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to addressing imbalanced attributes in GANs latent space clustering by using a Gaussian mixture as a prior and deriving Stein latent optimization. However, the contributions are somewhat limited in their originality as they build upon existing... |
25kzAhUB1lz | Direct then Diffuse: Incremental Unsupervised Skill Discovery for State Covering and Goal Reaching | ICLR | 2,022 | train | false | Learning meaningful behaviors in the absence of reward is a difficult problem in reinforcement learning. A desirable and challenging unsupervised objective is to learn a set of diverse skills that provide a thorough coverage of the state space while being directed, i.e., reliably reaching distinct regions of the enviro... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes a novel algorithm for learning unsupervised skills based on empowerment. Specifically, direct-and-diffuse policies are proposed that first optimize empowerment and then after a fixed number of episode steps optimize action entropy. This y... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The paper proposes a novel algorithm for learning unsupervised skills based on empowerment.",
"This yields policies that go to a certain point in a directed manner and then move around that point at random.",
"Further, the UPSIDE algorithm is based on tracking the policies with high discriminability (i.e. low ... | {
"novelty_excerpts": [
"The paper proposes a novel algorithm for learning unsupervised skills based on empowerment.",
"This yields policies that go to a certain point in a directed manner and then move around that point at random.",
"Further, the UPSIDE algorithm is based on tracking the policies with hi... | 1 | The paper introduces a novel approach to unsupervised skill discovery by combining directed and diffusive policies, which significantly outperforms prior work and provides important insights into what properties are important for a successful skill discovery algorithm. The work presents a new method for stacking skills... | The paper introduces a novel algorithm for learning unsupervised skills based on empowerment, which combines direct-and-diffuse policies to achieve good performance on pointmass, cheetah, walker, and ant tasks. The work presents a new method for stacking skills sequentially for better exploration, which is a significan... | Novelty Score: 1
Score Justification:
The paper introduces a novel approach to unsupervised skill discovery by combining directed and diffusive policies, which significantly outperforms prior work and provides important insights into what properties are important for a successful skill discovery algorithm. The work pr... |
28ib9tf6zhr | Patch-Fool: Are Vision Transformers Always Robust Against Adversarial Perturbations? | ICLR | 2,022 | train | false | Vision transformers (ViTs) have recently set off a new wave in neural architecture design thanks to their record-breaking performance in various vision tasks. In parallel, to fulfill the goal of deploying ViTs into real-world vision applications, their robustness against potential malicious attacks has gained increasin... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a Patch-Fool attack which fools the self-attention mechanism of ViTs to evaluate the robustness of ViTs and CNNs-based models. Prior works investigated the robustness of ViTs and CNNs under the adversarial attacks designed mainly for CNN... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This paper provides a new perspective for evaluating the robustness of ViTs which is novel and insightful.",
"From my point of view, the conclusion that the perturbation density is the key factor that influences the robustness ranking between ViTs and CNNs is significant to this field and can provide a better un... | {
"novelty_excerpts": [
"This paper provides a new perspective for evaluating the robustness of ViTs which is novel and insightful.",
"From my point of view, the conclusion that the perturbation density is the key factor that influences the robustness ranking between ViTs and CNNs is significant to this field... | 1 | The paper introduces a novel attack framework, Patch-Fool, which effectively evaluates the robustness of Vision Transformers (ViTs) by targeting their self-attention mechanism. The work presents a new perspective on ViT robustness, revealing that ViTs are weaker learners than CNNs under specific perturbation conditions... | The main novel contributions or new ideas introduced are the Patch-Fool attack framework and the Spare Patch-Fool extension, which demonstrate the vulnerability of ViTs to attention-aware attacks. The paper's investigation into the impact of perturbation density on robustness ranking between ViTs and CNNs provides insi... | Novelty Score: 1
Score Justification:
The paper introduces a novel attack framework, Patch-Fool, which effectively evaluates the robustness of Vision Transformers (ViTs) by targeting their self-attention mechanism. The work presents a new perspective on ViT robustness, revealing that ViTs are weaker learners than CNNs... |
2DJn3E7lXu | What to expect of hardware metric predictors in NAS | ICLR | 2,022 | train | false | Modern Neural Architecture Search (NAS) focuses on finding the best performing architectures in hardware-aware settings; e.g., those with an optimal tradeoff of accuracy and latency. Due to many advantages of prediction models over live measurements, the search process is often guided by estimates of how well each cons... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nIn neural architecture search (NAS), performance predictors are an important tool, because predicting attributes such as accuracy can save costly measurements. Hardware metric predictors predict metrics such as latency in addition to accuracy. This paper gi... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The major concern of this paper is the technical novelty since no new technique is proposed and the provided insights are relatively intuitive.",
"More novel insights are expected.",
"For example, as demonstrated in HW-NAS-Bench, for some search spaces (i.e., a sequential space like FBNet space) and many devic... | {
"novelty_excerpts": [
"The major concern of this paper is the technical novelty since no new technique is proposed and the provided insights are relatively intuitive.",
"More novel insights are expected.",
"For example, as demonstrated in HW-NAS-Bench, for some search spaces (i.e., a sequential space li... | -1 | The paper presents a comprehensive evaluation of hardware performance predictors, but it does not introduce any fundamentally new ideas or approaches. The work primarily focuses on extending existing methods to a new domain, hardware-aware neural architecture search. | The paper introduces a large-scale study of hardware metric predictors, comparing their performance across different networks, devices, and tasks. However, the methodological innovations and experimental insights presented are largely incremental, building upon existing work in the field of neural architecture search. ... | Novelty Score: -1
Score Justification:
The paper presents a comprehensive evaluation of hardware performance predictors, but it does not introduce any fundamentally new ideas or approaches. The work primarily focuses on extending existing methods to a new domain, hardware-aware neural architecture search.
Detailed As... |
2DJwuD-elOt | Hybrid Cloud-Edge Networks for Efficient Inference | ICLR | 2,022 | train | false | Although deep neural networks (DNNs) achieve state-of-the-art accuracy on large-scale and fine-grained prediction tasks, they are high capacity models and often cannot be deployed on edge devices. As such, two distinct paradigms have emerged in parallel: 1) edge device inference for low-level tasks, 2) cloud-based infe... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes to combine the inference on edge and cloud together, taking advantage of the communication-free inference on edge and the high-accuracy model on cloud. It utilizes the OFA to obtain models with different resource requirement at low cost. ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"It is interesting to have models with different MACs on cloud and edge to improve the inference accuracy with minimal latency increase.",
"The idea of introducing a routing model to route more challenging inputs to the cloud is quite novel.",
"The paper proposes to provide a better tradeoff between the latency... | {
"novelty_excerpts": [
"It is interesting to have models with different MACs on cloud and edge to improve the inference accuracy with minimal latency increase.",
"The idea of introducing a routing model to route more challenging inputs to the cloud is quite novel.",
"The paper proposes to provide a bette... | 1 | The paper introduces a moderately novel approach to combining edge and cloud inference, leveraging a routing model to route challenging inputs to the cloud. This idea builds upon existing concepts but presents a competent extension and combination of existing ideas. The work's originality lies in its ability to balance... | The paper presents a principled solution to the problem of improving edge device inference accuracy by employing a hybrid approach with a low-capacity model on the edge and a high-capacity model on the cloud. The query routing model running on the edge device decides which query needs to be handled by the cloud model, ... | Novelty Score: 1
Score Justification:
The paper introduces a moderately novel approach to combining edge and cloud inference, leveraging a routing model to route challenging inputs to the cloud. This idea builds upon existing concepts but presents a competent extension and combination of existing ideas. The work's ori... |
2I1wy0y6xo | Stability analysis of SGD through the normalized loss function | ICLR | 2,022 | train | false | We prove new generalization bounds for stochastic gradient descent for both the convex and non-convex cases. Our analysis is based on the stability framework. We analyze stability with respect to the normalized version of the loss function used for training. This leads to investigating a form of angle-wise stability in... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper considers the problem of understanding the generalization of SGD using the stability framework. The well-known result in this line of work is the paper by Hardt'16. In Hardt'16, the stability is measured using the difference between the \"actual\... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The main observation by the authors in this paper is that in many cases, the loss function is invariant to the scaling of weights.",
"Their main results are the new stability analysis for this new notion for convex and non-convex settings.",
"For the non-convex the authors define a new measure for generalizati... | {
"novelty_excerpts": [
"The main observation by the authors in this paper is that in many cases, the loss function is invariant to the scaling of weights.",
"Their main results are the new stability analysis for this new notion for convex and non-convex settings.",
"For the non-convex the authors define ... | 0 | The paper presents a reformulation of the stability analysis using the normalized loss function, which is a minor variation of existing approaches. The work builds upon the stability framework introduced in Hardt'16, but the extensions remain largely incremental. The paper's contributions are somewhat standard, showing... | The paper introduces a new stability analysis for the normalized loss function, which is a reformulation of the existing stability framework. The work extends the analysis to convex and non-convex settings, but the results are not fundamentally new. The approach combines existing ideas from the stability framework and ... | Novelty Score: 0
Score Justification:
The paper presents a reformulation of the stability analysis using the normalized loss function, which is a minor variation of existing approaches. The work builds upon the stability framework introduced in Hardt'16, but the extensions remain largely incremental. The paper's contr... |
2M0WXSP6Qi | Information-theoretic stochastic contrastive conditional GAN: InfoSCC-GAN | ICLR | 2,022 | train | false | Conditional generation is a subclass of generative problems when the output of generation is conditioned by a class attributes’ information. In this paper, we present a new stochastic contrastive conditional generative adversarial network (InfoSCC-GAN) with explorable latent space. The InfoSCC-GAN architecture is based... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a GAN for conditional generation. The authors combine an unsupervised contrastive encoder, stochastic EigenGAN generator, and a classifier. InfoSCC-GAN can perform image generation conditioned on external attributes by maximizing mutual ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The technical novelty of the paper is limited.",
"InfoSCC-GAN introduces a classifier that minimizes cross-entropy, which is the same as AC-GAN [1] or ClusterGAN [2].",
"A method for selecting inner attributes using K-means clustering on features already exists [3].",
"The novelty here is that InfoSCC-GAN pr... | {
"novelty_excerpts": [
"The technical novelty of the paper is limited.",
"InfoSCC-GAN introduces a classifier that minimizes cross-entropy, which is the same as AC-GAN [1] or ClusterGAN [2].",
"A method for selecting inner attributes using K-means clustering on features already exists [3].",
"The nov... | 0 | The paper presents a combination of existing ideas, building on contrastive learning, stochastic EigenGAN generators, and attribute clustering. While it introduces a novel application of these concepts, the technical novelty is limited, and the contributions are somewhat standard. The work extends existing approaches b... | The paper introduces a stochastic contrastive conditional GAN that combines an encoder, an attributes' classifier, and a stochastic EigenGAN generator. The method extends existing approaches in conditional image generation, but the novelty lies in its application of contrastive learning and attribute clustering. The wo... | Novelty Score: 0
Score Justification:
The paper presents a combination of existing ideas, building on contrastive learning, stochastic EigenGAN generators, and attribute clustering. While it introduces a novel application of these concepts, the technical novelty is limited, and the contributions are somewhat standard.... |
2NqIV8dzR7N | Automatic Termination for Hyperparameter Optimization | ICLR | 2,022 | train | false | Bayesian optimization (BO) is a widely popular approach for the hyperparameter optimization (HPO) of machine learning algorithms. At its core, BO iteratively evaluates promising configurations until a user-defined budget, such as wall-clock time or number of iterations, is exhausted. While the final performance after t... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes an automatic termination criterion for Bayesian optimization (BO) by using the upper bound of the simple reget. Various experiments are conducted to demonstrate that with the utilization of the termination criterion, computation and ener... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The major contribution of this paper lies in the two propositions.",
"Plenty of related works are even not mentioned in the paper, e.g. Freeze-thaw BO [1], Multi-fidelity BO [2], Hyperband BO [3], and BOS-BO [4], these works all discussed how to achieve efficiency and saving budget in BO.",
"A detailed discus... | {
"novelty_excerpts": [
"The major contribution of this paper lies in the two propositions.",
"Plenty of related works are even not mentioned in the paper, e.g. Freeze-thaw BO [1], Multi-fidelity BO [2], Hyperband BO [3], and BOS-BO [4], these works all discussed how to achieve efficiency and saving budget i... | 0 | The paper presents a competent extension of existing ideas in Bayesian optimization, though the contributions remain largely incremental. The work builds on prior research in the field, introducing a new approach for automatic termination of hyperparameter optimization. However, the novelty of the approach is limited b... | The paper introduces a new termination criterion for Bayesian optimization, which combines elements of existing approaches. The work extends the concept of regret in Bayesian optimization, proposing a new bound for the optimization error. However, the bound lacks theoretical guarantee, and the method relies heavily on ... | Novelty Score: 0
Score Justification:
The paper presents a competent extension of existing ideas in Bayesian optimization, though the contributions remain largely incremental. The work builds on prior research in the field, introducing a new approach for automatic termination of hyperparameter optimization. However, t... |
2RNpZ8S4alJ | KINet: Keypoint Interaction Networks for Unsupervised Forward Modeling | ICLR | 2,022 | train | false | Object-centric representation is an essential abstraction for physical reasoning and forward prediction. Most existing approaches learn this representation through extensive supervision (e.g, object class and bounding box) although such ground-truth information is not readily accessible in reality. To address this, we ... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper introduces KINet (Keypoint Interaction Network)---an end-to-end unsupervised framework to reason about object interactions in complex systems based on a keypoint representation. There are three major steps in this model: extracting keypoint coord... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"In general, this work is a patchwork of a condition-based generative model and a raw image preprocessing method at the input and output ends, leading to insufficient novelty.",
"For key point detection, Jakab et al. (2018) thoughts are followed.",
"The initial graph Gt with its nodes and edges is built through... | {
"novelty_excerpts": [
"In general, this work is a patchwork of a condition-based generative model and a raw image preprocessing method at the input and output ends, leading to insufficient novelty.",
"For key point detection, Jakab et al. (2018) thoughts are followed.",
"The initial graph Gt with its no... | 0 | The paper presents a combination of existing ideas, including unsupervised keypoint extraction and forward modeling, without introducing fundamentally new concepts or significant methodological innovations. The approach builds on established techniques, such as graph-based representations and contrastive loss, but does... | The paper introduces a framework for keypoint interaction networks (KINet) that combines unsupervised keypoint extraction and forward modeling. However, the method largely relies on existing techniques, including the use of graph-based representations and contrastive loss. The work presents a competent extension of exi... | Novelty Score: 0
Score Justification:
The paper presents a combination of existing ideas, including unsupervised keypoint extraction and forward modeling, without introducing fundamentally new concepts or significant methodological innovations. The approach builds on established techniques, such as graph-based represe... |
2RYOwBOFesi | An Empirical Study of Pre-trained Models on Out-of-distribution Generalization | ICLR | 2,022 | train | false | Generalizing to out-of-distribution (OOD) data -- that is, data from domains unseen during training -- is a key challenge in modern machine learning, which has only recently received much attention. Some existing approaches propose leveraging larger models and pre-training on larger datasets. In this paper, we provide ... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper empirically studies pre-trained model on OOD generalization. Specifically, the authors consider four datasets in this paper: PACS, Office-Home, VLCS and TerraIncognita. Based on the experimental results, the authors concludes three main observati... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"However, I have found some of the results unclear or incorrect.",
"It seems to me that the authors want to claim that the smaller lr, the better performance. This result is not well-supported.",
"More experiments with smaller learning rates such as $10^{-5}$, $10^{-6}$ and $10^{-7}$ are needed.",
"We can see... | {
"novelty_excerpts": [
"However, I have found some of the results unclear or incorrect.",
"It seems to me that the authors want to claim that the smaller lr, the better performance. This result is not well-supported.",
"More experiments with smaller learning rates such as $10^{-5}$, $10^{-6}$ and $10^{-7... | -1 | The paper introduces an incremental approach to improving out-of-distribution (OOD) generalization using pre-training models, building upon existing ideas without significant innovation. The work presents a straightforward extension of prior research, leveraging larger models and datasets to enhance OOD performance. Th... | The paper extends the idea of using pre-training models to improve OOD generalization, as previously explored in ICML 2021 and ICLR 2021 papers. The method combines existing techniques, such as adversarial training and pre-training, to achieve better OOD performance. The work presents a comprehensive experimental analy... | Novelty Score: -1
Score Justification:
The paper introduces an incremental approach to improving out-of-distribution (OOD) generalization using pre-training models, building upon existing ideas without significant innovation. The work presents a straightforward extension of prior research, leveraging larger models and... |
2_dQlkDHnvN | Defending Backdoor Data Poisoning Attacks by Using Noisy Label Defense Algorithm | ICLR | 2,022 | train | false | Training deep neural networks with data corruption is a challenging problem. One example of such corruption is the backdoor data poisoning attack, in which an adversary strategically injects a backdoor trigger to a small fraction of the training data to subtly compromise the training process. Consequently, the trained ... | 3 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper aims to apply noisy-label defense algorithms to defend general backdoor attacks. The authors propose a meta-algorithm by adding noisy labels to the training set and optimizing outer minimization and inner maximization of the model by the noisy-la... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The method proposed is simple yet effective.",
"The authors theoretically proved that noisy-label defense algorithms can be leveraged to against backdoor attacks, but the method seems not very innovative,",
"The contributions are significant and somewhat new.",
"Aspects of the contributions exist in prior wo... | {
"novelty_excerpts": [
"The method proposed is simple yet effective.",
"The authors theoretically proved that noisy-label defense algorithms can be leveraged to against backdoor attacks, but the method seems not very innovative,",
"The contributions are significant and somewhat new.",
"Aspects of the... | 1 | The paper introduces a meta-algorithm that leverages noisy-label defense algorithms to defend against backdoor attacks, though the approach builds on existing techniques and lacks significant innovation. The work presents a novel application of noisy-label algorithms to a new problem domain, but the methodological cont... | The paper introduces a meta-algorithm that combines noisy-label defense algorithms with adversarial training to defend against backdoor attacks. The approach extends existing noisy-label algorithms to a new problem domain, but the methodological contributions remain largely incremental, and the work does not present fu... | Novelty Score: 1
Score Justification:
The paper introduces a meta-algorithm that leverages noisy-label defense algorithms to defend against backdoor attacks, though the approach builds on existing techniques and lacks significant innovation. The work presents a novel application of noisy-label algorithms to a new prob... |
2_vhkAMARk | Escaping limit cycles: Global convergence for constrained nonconvex-nonconcave minimax problems | ICLR | 2,022 | train | false | This paper introduces a new extragradient-type algorithm for a class of nonconvex-nonconcave minimax problems. It is well-known that finding a local solution for general minimax problems is computationally intractable. This observation has recently motivated the study of structures sufficient for convergence of first o... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper constructs methods named CurvatureEG+ (and Adaptive EG+ and CEG+), built upon a recently proposed EG+ [Diakonikolas et al., 2021], a variant of EG, that works under the weak MVI condition. Most importantly, the CurvatureEG+ (and Adaptive EG+/CEG+... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"A larger range of ρ of the weak MVI: The Adaptive EG+ and (a line-search-version) CurvatureEG+ work for a wider range of ρ, ρ>−12 or ρ>−γk2 respectively, compared to that of EG+, ρ>−18.",
"One might be view this improvement negligible (which I did when I glanced through the paper), but the wider region includes,... | {
"novelty_excerpts": [
"A larger range of ρ of the weak MVI: The Adaptive EG+ and (a line-search-version) CurvatureEG+ work for a wider range of ρ, ρ>−12 or ρ>−γk2 respectively, compared to that of EG+, ρ>−18.",
"One might be view this improvement negligible (which I did when I glanced through the paper), bu... | 1 | The paper introduces a unified framework of EG using the projection onto a hyperplane, which extends the applicability of EG to a wider range of settings. The proposed method, CurvatureEG+, has a wider range of applicability compared to EG+, though the improvement in convergence guarantee is not significant. The contri... | The main novel contribution of the paper is the unified framework of EG, which includes a non-adaptive variant, CEG+, covering both EG+ and FBF. The approach builds on existing methods, such as EG+ and FBF, and extends their applicability to a wider range of settings. The paper presents a stochastic variant of the meth... | Novelty Score: 1
Score Justification:
The paper introduces a unified framework of EG using the projection onto a hyperplane, which extends the applicability of EG to a wider range of settings. The proposed method, CurvatureEG+, has a wider range of applicability compared to EG+, though the improvement in convergence g... |
2bO2x8NAIMB | Should We Be Pre-training? An Argument for End-task Aware Training as an Alternative | ICLR | 2,022 | train | false | In most settings of practical concern, machine learning practitioners know in advance what end-task they wish to boost with auxiliary tasks. However, widely used methods for leveraging auxiliary data like pre-training and its continued-pretraining variant are end-task agnostic: they rarely, if ever, exploit knowledge o... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper makes the argument that generic pre-training (on auxiliary tasks) is inferior to task-specific pre-training. The authors argue that the final task should be learned together with the auxiliary tasks in a multi-task setup, which they call MT-TARTAN... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"Their results should be of interest to the community.",
"Although they raise the question whether pre-training is necessary, they then don't actually compare against a model that is not pre-trained.",
"Rather, they show that after pre-training it might not be necessary to further pre-train on large amounts of ... | {
"novelty_excerpts": [
"Their results should be of interest to the community.",
"Although they raise the question whether pre-training is necessary, they then don't actually compare against a model that is not pre-trained.",
"Rather, they show that after pre-training it might not be necessary to further ... | 1 | The paper presents a moderately novel approach to pre-training and fine-tuning tasks, combining them in a multi-task setup. The work extends existing ideas on pre-training and fine-tuning, but the contributions remain largely incremental. The method introduces a new algorithm, META-TARTAN, which uses meta-learning to m... | The paper introduces a multi-task setup, MT-TARTAN, which combines pre-training and fine-tuning tasks. The work extends existing ideas on pre-training and fine-tuning, but the contributions remain largely incremental. The method presents a new algorithm, META-TARTAN, which uses meta-learning to mitigate the impact of a... | Novelty Score: 1
Score Justification:
The paper presents a moderately novel approach to pre-training and fine-tuning tasks, combining them in a multi-task setup. The work extends existing ideas on pre-training and fine-tuning, but the contributions remain largely incremental. The method introduces a new algorithm, MET... |
2d4riGOpmU8 | Sequential Covariate Shift Detection Using Classifier Two-Sample Tests | ICLR | 2,022 | train | false | A standard assumption in supervised learning is that the training data and test data are from the same distribution. However, this assumption often fails to hold in practice, which can cause the learned model to perform poorly. We consider the problem of detecting covariate shift, where the covariate distribution shift... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper proposes a new method to sequentially detect covariate shift which is the problem encountered when training and test covariates are not equally distributed. This methods builds on existing work that uses classifiers to distinguish between the two... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The contribution of this paper is proposing a sequential version, where the classifier is trained in an online manner and the accuracy computed on a sliding window (of training and test samples) is used for each test set observation.",
"It is argued (sections 1 and 2) that sequential tests are not appropriate i... | {"novelty_excerpts": [
"The contribution of this paper is proposing a sequential version, where the classifier is trained in an online manner and the accuracy computed on a sliding window (of training and test samples) is used for each test set observation.",
"It is argued (sections 1 and 2) that sequential tests are ... | 1 | The paper introduces a new method for sequentially detecting covariate shift, which is a practical and significant contribution to the field. However, the approach builds on existing work and does not fundamentally change the underlying problem formulation or methodology. The method's novelty lies in its ability to be ... | The paper presents a novel sequential classifier two-sample test for dynamically detecting distribution shifts, which is a significant contribution to the field of AI reliability and trustworthy ML. The approach is moderately novel, as it extends existing two-sample tests to an online setting, but the underlying proble... | Novelty Score: 1
Score Justification:
The paper introduces a new method for sequentially detecting covariate shift, which is a practical and significant contribution to the field. However, the approach builds on existing work and does not fundamentally change the underlying problem formulation or methodology. The meth... |
2eXhNpHeW6E | R5: Rule Discovery with Reinforced and Recurrent Relational Reasoning | ICLR | 2,022 | train | false | Systematicity, i.e., the ability to recombine known parts and rules to form new sequences while reasoning over relational data, is critical to machine intelligence. A model with strong systematicity is able to train on small-scale tasks and generalize to large-scale tasks. In this paper, we propose R5, a relational rea... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThis paper presents a novel method for rule induction. The main idea is to apply\nreinforcement learning in the task of relational pathfinding within a finite\nHerbrand base. The reinforcement learner uses MCTS to find the best routes to\nestablish the path... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This paper presents a novel method for rule induction.",
"The main contribution of this work is applying reinforcement learning in the relational pathfinding problem.",
"The proposed approach works surprisingly well in the experiments, and the combination of neural-guided heuristic learning for MCTS and the sy... | {
"novelty_excerpts": [
"This paper presents a novel method for rule induction.",
"The main contribution of this work is applying reinforcement learning in the relational pathfinding problem.",
"The proposed approach works surprisingly well in the experiments, and the combination of neural-guided heuristi... | 1 | The paper introduces a novel method for rule induction by applying reinforcement learning to relational pathfinding. The approach presents a significant contribution to the field, though aspects of the work exist in prior research. The method's originality lies in its combination of neural-guided heuristic learning and... | The paper presents a novel method for rule induction by applying reinforcement learning to relational pathfinding. This approach combines neural-guided heuristic learning for MCTS and symbolic rewarding, which performs well in the relational pathfinding problem. The work extends previous research in this direction by i... | Novelty Score: 1
Score Justification:
The paper introduces a novel method for rule induction by applying reinforcement learning to relational pathfinding. The approach presents a significant contribution to the field, though aspects of the work exist in prior research. The method's originality lies in its combination ... |
2f1z55GVQN | Critical Points in Quantum Generative Models | ICLR | 2,022 | train | false | One of the most important properties of neural networks is the clustering of local minima of the loss function near the global minimum, enabling efficient training. Though generative models implemented on quantum computers are known to be more expressive than their traditional counterparts, it has empirically been obse... | 1 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe work attempts to connect the optimization of quantum generative models with random fields on manifolds. Some theoretical and numerical analyses are provided. \n\nThe major idea is to apply random matrix theory for variational quantum circuits. When the ... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The work attempts to connect the optimization of quantum generative models with random fields on manifolds.",
"The major idea is to apply random matrix theory for variational quantum circuits.",
"The work leverages the concepts of random matrix theory to analyze the optimization performance of quantum generati... | {
"novelty_excerpts": [
"The work attempts to connect the optimization of quantum generative models with random fields on manifolds.",
"The major idea is to apply random matrix theory for variational quantum circuits.",
"The work leverages the concepts of random matrix theory to analyze the optimization p... | 1 | The paper introduces a new approach by leveraging random matrix theory for variational quantum circuit training, though the extensions remain largely incremental. The work presents a combination of existing ideas, building upon prior research in the field. The contribution lies in its application of random matrix theor... | The paper builds on existing research in variational quantum circuit training and random matrix theory, combining these ideas to analyze the optimization performance of quantum generative models. The work extends the application of random matrix theory to a new area, but the novelty is somewhat limited by the lack of s... | Novelty Score: 1
Score Justification:
The paper introduces a new approach by leveraging random matrix theory for variational quantum circuit training, though the extensions remain largely incremental. The work presents a combination of existing ideas, building upon prior research in the field. The contribution lies in... |
2ggNjUisGyr | Partial Wasserstein Adversarial Network for Non-rigid Point Set Registration | ICLR | 2,022 | train | false | Given two point sets, the problem of registration is to recover a transformation that matches one set to the other. This task is challenging due to the presence of large number of outliers, the unknown non-rigid deformations and the large sizes of point sets. To obtain strong robustness against outliers, we formulate t... | 2 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\nThe paper proposes a method for registration of two point sets by formulating the problem as partial distribution matching problem. The authors utilise the partial Wasserstein (PW) discrepancy, and show how its gradient can be computed. They further propose... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"the authors are able to make a novel contribution to a classic problem.",
"It is pleasing that the point sets are modelled as discrete distributions without the need of making smoothing like GMM based methods for instance do.",
"clearly a strength of the paper even though it is a harder part of the paper to re... | {
"novelty_excerpts": [
"the authors are able to make a novel contribution to a classic problem.",
"It is pleasing that the point sets are modelled as discrete distributions without the need of making smoothing like GMM based methods for instance do.",
"clearly a strength of the paper even though it is a ... | 1 | null | null | Novelty Score: 1
The paper introduces moderately novel contributions to the classic problem of point set registration. The work presents a novel formulation of the registration problem as a partial distribution matching problem and derives the Kantorovich–Rubinstein duality for the partial Wasserstein discrepancy. The... |
2hMEdc35xZ6 | Defect Transfer GAN: Diverse Defect Synthesis for Data Augmentation | ICLR | 2,022 | train | false | Large amounts of data are a common requirement for many deep learning approaches. However, data is not always equally available at large scale for all classes. For example, on highly optimized production lines, defective samples are hardly acquired while non-defective samples come almost for free. The defects however o... | 5 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\n\nThe paper prposes a StarGAN based model to distangle the defect foreground, background, transfer the style of foreground and then synthesize the defected image for different products. The quality of the synthesized defect images are evaluated using FID, K... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"The novelty and contribution of methodology is limited, particularly for ICLR 2022.",
"the novelty and contribution of methodology is limited",
"Much more defects and products need to be tested and evaluated.",
"the number of categories for defects are too limited and it's not convincing to support the claim... | {
"novelty_excerpts": [
"The novelty and contribution of methodology is limited, particularly for ICLR 2022.",
"the novelty and contribution of methodology is limited",
"Much more defects and products need to be tested and evaluated.",
"the number of categories for defects are too limited and it's not... | -1 | The paper introduces a StarGAN-based approach for generating synthetic images of various defects, but the technical novelty is limited, and the contributions are only marginally significant or novel. The method heavily relies on existing architectures, such as StarGAN v2, with minimal innovation in terms of methodology... | The paper builds on existing GAN formulations, primarily StarGAN v2, with additional components that are largely incremental. The approach extends the capability of GANs for data augmentation in the context of defect classification, but the novelty lies in the application rather than the underlying methodology. The met... | Novelty Score: -1
Score Justification:
The paper introduces a StarGAN-based approach for generating synthetic images of various defects, but the technical novelty is limited, and the contributions are only marginally significant or novel. The method heavily relies on existing architectures, such as StarGAN v2, with mi... |
2jYxq9_TkpG | Network Pruning Optimization by Simulated Annealing Algorithm | ICLR | 2,022 | train | false | One critical problem of large neural networks is over-parameterization with a large number of weight parameters. This becomes an obstacle to implement networks in edge devices as well as limiting the development of industrial applications by engineers for machine learning problems. Plenty of papers have shown that the ... | 4 | [
{
"review_text": "# Review\n\n## Summary Of The Paper\n This paper presents a novel simulated annealing method for pruning neural networks. Both one-shot pruning and gradual pruning strategies are discussed in this study. Experiments are conducted based on a simple neural network structure with only two hidden... | ## Summary Of The Paper
Provide a brief summary of the paper and its contributions.
## Main Review
Please list both the strengths and weaknesses of the paper. When discussing weaknesses, please provide concrete, actionable feedback on the paper.
## Summary Of The Review
Please provide a short summary justifying your ... | [
"This paper presents a novel simulated annealing method for pruning neural networks.",
"Although the idea seems interesting,",
"The contributions are only marginally significant or novel.",
"The contributions are only marginally significant or novel."
] | {
"novelty_excerpts": [
"This paper presents a novel simulated annealing method for pruning neural networks.",
"Although the idea seems interesting,",
"The contributions are only marginally significant or novel.",
"The contributions are only marginally significant or novel."
]
} | -1 | The paper introduces a simulated annealing method for pruning neural networks, but the contributions are largely incremental and build upon existing pruning methods. The work presents a novel application of simulated annealing, but the novelty is not apparent compared to existing pruning methods. The approach combines ... | The paper presents a simulated annealing method for pruning neural networks, which is an interesting but not novel application of the technique. The method is an extension of existing pruning methods, and the novelty is not apparent compared to other black-box optimization methods. The work does not provide a fundament... | Novelty Score: -1
Score Justification:
The paper introduces a simulated annealing method for pruning neural networks, but the contributions are largely incremental and build upon existing pruning methods. The work presents a novel application of simulated annealing, but the novelty is not apparent compared to existing... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.