Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
Using KDE queries (as black box), the paper proposes to effciently solve various gram matrix related problems such as spectral approximation/sparsification, simulating random walks, weighted sampling, etc. In each case, a theoretical bound on the approximation/quality along with no. KDE queries is presented. Simulatio...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: Using KDE queries (as black box), the paper proposes to effciently solve various gram matrix related problems such as spectral approximation/sparsification, simulating random walks, weighted sampling, etc. In each case, a theoretical bound on the approximation/quality along with no. KDE queries is presented. S...
The authors examine how a wide variety of transformer architecture choices (and a few non-transformer architectures such as MLP-Mixers and dynamic convolutions) scale over a wide range of model sizes. Their models are implemented in a sequence-to-sequence manner, and evaluated with pretraining perplexity and SuperGlue ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors examine how a wide variety of transformer architecture choices (and a few non-transformer architectures such as MLP-Mixers and dynamic convolutions) scale over a wide range of model sizes. Their models are implemented in a sequence-to-sequence manner, and evaluated with pretraining perplexity and Su...
The paper studies pessimistic model-based offline reinforcement learning. Motivated by the fact that ensemble methods do not provide accurate uncertainty quantifiers used for constructing penalties subtracted from rewards, the authors present an entropy-regularized algorithm that learns a pessimistic model and provides...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies pessimistic model-based offline reinforcement learning. Motivated by the fact that ensemble methods do not provide accurate uncertainty quantifiers used for constructing penalties subtracted from rewards, the authors present an entropy-regularized algorithm that learns a pessimistic model and ...
The authors propose to train a policy, a state-encoder, and a dynamic model in model-based RL with the novel evidence lower bound on expected returns loss. In this way policy, encoder, and dynamic model have the same objective to maximize which differs this work from the previous ones. In model-based RL algorithms li...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose to train a policy, a state-encoder, and a dynamic model in model-based RL with the novel evidence lower bound on expected returns loss. In this way policy, encoder, and dynamic model have the same objective to maximize which differs this work from the previous ones. In model-based RL algor...
This paper proposes a novel attack, called camouflaged data poisoning attacks by leveraging a machine unlearning mechanism. This attack needs generate a pair of poisoning and camouflage datasets, which has less impact to the model on normal testing data. After triggering the machine unlearning procedure on the camoufla...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a novel attack, called camouflaged data poisoning attacks by leveraging a machine unlearning mechanism. This attack needs generate a pair of poisoning and camouflage datasets, which has less impact to the model on normal testing data. After triggering the machine unlearning procedure on the ...
The paper aims to solve the suboptimal problems caused by the pretraining models in industrial systems. And the authors analyze the problem with four aspects( the gap of uniform convergence for analyzing pretrained representations, their stochastic nature under gradient descent optimization, what model convergence mean...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper aims to solve the suboptimal problems caused by the pretraining models in industrial systems. And the authors analyze the problem with four aspects( the gap of uniform convergence for analyzing pretrained representations, their stochastic nature under gradient descent optimization, what model converge...
The authors propose to use temporal changes in object views that occur naturally in developing infants' visual environment as data augmentations for time-based self-supervised learning. Using a rendering environment they create positive and negative training pairs using 3d object manipulations such as rotations, saccad...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors propose to use temporal changes in object views that occur naturally in developing infants' visual environment as data augmentations for time-based self-supervised learning. Using a rendering environment they create positive and negative training pairs using 3d object manipulations such as rotations...
This paper proposes a visual and linguistic bounded explanation method to make part of NN models explainable by adding both attribute-wise and language-wise explanations. Specifically, they add a trainable part between the feature extractor (e.g., pre-trained Resnet) and label embedding, and the added part can be expla...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a visual and linguistic bounded explanation method to make part of NN models explainable by adding both attribute-wise and language-wise explanations. Specifically, they add a trainable part between the feature extractor (e.g., pre-trained Resnet) and label embedding, and the added part can ...
This paper reviews the desiderata of GNN and proposes SlenderGNN, a simple linear GNN, to meet the needs. The proposed SlenderGNN comes from the linearization framework, which can resemble various GNN designs. Towards these desiderata, the authors present several sanity check, and the proposed SlenderGNN pass all of th...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper reviews the desiderata of GNN and proposes SlenderGNN, a simple linear GNN, to meet the needs. The proposed SlenderGNN comes from the linearization framework, which can resemble various GNN designs. Towards these desiderata, the authors present several sanity check, and the proposed SlenderGNN pass a...
This paper proposes a framework to construct uncertainty sets in the online setting to control the risk of the coverage, false negative rate, or F1 score. The propose method has theoretical guarantee for risk control at the user specified level for different underlying data like distribution shifts over time in an unkn...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes a framework to construct uncertainty sets in the online setting to control the risk of the coverage, false negative rate, or F1 score. The propose method has theoretical guarantee for risk control at the user specified level for different underlying data like distribution shifts over time in...
This paper proposes an architecture and method for online training for robotic datasets Strengths: - The proposal is complete and can be trained on various sim environments - Apparently able to learn from multi robot data Weaknesses: - Tested on sim only. Online learning is too slow to be practical in real. - The pap...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes an architecture and method for online training for robotic datasets Strengths: - The proposal is complete and can be trained on various sim environments - Apparently able to learn from multi robot data Weaknesses: - Tested on sim only. Online learning is too slow to be practical in real. -...
This paper studies fast background planning, based on an abstract model over subgoals. The subgoal space is predefined, and the authors then learn expected return and discount models between subgoals, which are used in a tabular dynamic programming procedure. The resulting value function is added to the bootstrapping o...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies fast background planning, based on an abstract model over subgoals. The subgoal space is predefined, and the authors then learn expected return and discount models between subgoals, which are used in a tabular dynamic programming procedure. The resulting value function is added to the bootstr...
The paper introduces an MCMC-based framework that introduces conditional transformations that respect certain 3d structural properties of proteins with the motivation to learn conformation invariant representation. Strengths: + MCMC-based strategy for learning representations that are invariant to conditional transfo...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper introduces an MCMC-based framework that introduces conditional transformations that respect certain 3d structural properties of proteins with the motivation to learn conformation invariant representation. Strengths: + MCMC-based strategy for learning representations that are invariant to conditional...
The paper proposes to optimize a population of agents generated through self-play to cover a multi-objective Pareto front among objectives for play style and performance (skill level). Play style is quantified in terms of state changes during play against various opponents. The optimization algorithm adapts the NSGA-II...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes to optimize a population of agents generated through self-play to cover a multi-objective Pareto front among objectives for play style and performance (skill level). Play style is quantified in terms of state changes during play against various opponents. The optimization algorithm adapts the...
The authors propose an algorithm named TiAda which is a time-scale adaptive algorithm for non-convex-strongly-convex (NC-SC) minimax problems. The algorithm is a single loop and problem-specific parameter agnostic one, which are improvements over the related prior work. The authors provide insight into the design of th...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors propose an algorithm named TiAda which is a time-scale adaptive algorithm for non-convex-strongly-convex (NC-SC) minimax problems. The algorithm is a single loop and problem-specific parameter agnostic one, which are improvements over the related prior work. The authors provide insight into the desi...
This paper presents a framework for task-free continual learning. The framework consists of a method for converting any dataset to a task-free continual learning problem where information on the class is not required and is not explicit, but rather examples from different classes can be observed at any step of the lear...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper presents a framework for task-free continual learning. The framework consists of a method for converting any dataset to a task-free continual learning problem where information on the class is not required and is not explicit, but rather examples from different classes can be observed at any step of ...
The paper studies when Random Fourier Features (RFF) is able to create a Johnson-Lindenstrauss-like result for preserving the distances between the feature mappings of vectors given a shift-invariant kernel. Prior work emphasized different metrics, like getting an additive error guarantee on the pairwise distances, ins...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies when Random Fourier Features (RFF) is able to create a Johnson-Lindenstrauss-like result for preserving the distances between the feature mappings of vectors given a shift-invariant kernel. Prior work emphasized different metrics, like getting an additive error guarantee on the pairwise distan...
The paper proposes a new way to estimate Shapely values for Vision Transformers. It is required to evaluate the model on partial inputs for estimating Shapely values. For CNNs, this is difficult (e.g. masking areas with gray produces out-of-distribution samples). However, the inputs of vision transformer can be more ea...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes a new way to estimate Shapely values for Vision Transformers. It is required to evaluate the model on partial inputs for estimating Shapely values. For CNNs, this is difficult (e.g. masking areas with gray produces out-of-distribution samples). However, the inputs of vision transformer can be...
1. The paper argues in favor of random sampling (termed as soft sampling) over other methods - specifically coreset based. The argument is well understood and to summarize a. The advantage of coresets in accuracy over random sampling is limited in some (maybe authors want to argue over many / most? But the evidenc...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: 1. The paper argues in favor of random sampling (termed as soft sampling) over other methods - specifically coreset based. The argument is well understood and to summarize a. The advantage of coresets in accuracy over random sampling is limited in some (maybe authors want to argue over many / most? But the...
The paper focuses on learning GNNs on open temporal graph where new nodes with novel classes are also added to the graph. The proposed method solves two major issues that existing temporal GNN methods have under this setting: learning with added heterophily and catastrophic forgetting. To solve the catastrophic forgett...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper focuses on learning GNNs on open temporal graph where new nodes with novel classes are also added to the graph. The proposed method solves two major issues that existing temporal GNN methods have under this setting: learning with added heterophily and catastrophic forgetting. To solve the catastrophic...
The paper proposes a method called Goodness of Fit Autoencoder (GoFAE) which uses goodness of fit (GoF) to regularize the posterior at the mini-batch level and selects a regularization coefficient at the global level (a few mini-batches) by making the distribution of p-values from each mini-batch uniform. Empirical res...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper proposes a method called Goodness of Fit Autoencoder (GoFAE) which uses goodness of fit (GoF) to regularize the posterior at the mini-batch level and selects a regularization coefficient at the global level (a few mini-batches) by making the distribution of p-values from each mini-batch uniform. Empir...
This paper aims at accelerating translation inference without quality degeneration. The authors extend the idea of "Predict-Verify" to "Draft-Verify" and propose SpecDec which leverages a non-autoregressive model to produce the next k tokens followed by an autoregressive model verifying these predictions in parallel. O...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper aims at accelerating translation inference without quality degeneration. The authors extend the idea of "Predict-Verify" to "Draft-Verify" and propose SpecDec which leverages a non-autoregressive model to produce the next k tokens followed by an autoregressive model verifying these predictions in par...
This papar focuses on the model-based off-policy evaluation(OPE) method, where there exist both limited coverage of state-action space and sensitive initialization during training the transition model. To address these, this paper presents the variational latent branching model(VLBM) to obtain as much information from ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This papar focuses on the model-based off-policy evaluation(OPE) method, where there exist both limited coverage of state-action space and sensitive initialization during training the transition model. To address these, this paper presents the variational latent branching model(VLBM) to obtain as much informati...
The paper aims to establish a dataset / benchmark for gene regulatory network inference from single-cell data. The paper is hard to follow for me. From the abstract and introduction, it seems it aims at providing a dataset / benchmark for GRN inference that would be very easy to understand and use by researchers from o...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper aims to establish a dataset / benchmark for gene regulatory network inference from single-cell data. The paper is hard to follow for me. From the abstract and introduction, it seems it aims at providing a dataset / benchmark for GRN inference that would be very easy to understand and use by researcher...
The paper proposes a new framework to achieve fair outcomes in settings where access to sensitive attributes is not available. The method is called Unsupervised Locality-base Proxy Label assignment (ULPL) and is based on assigning proxy labels according to model predictions for poor vs well-modeled instances. The autho...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes a new framework to achieve fair outcomes in settings where access to sensitive attributes is not available. The method is called Unsupervised Locality-base Proxy Label assignment (ULPL) and is based on assigning proxy labels according to model predictions for poor vs well-modeled instances. T...
This paper proposes a novel framework to conduct evaluation-free model selection for graph learning models, i.e., without having to train/evaluate any model on the new graph. The framework learns latent embeddings for observed models and the corresponding performance on observed graphs. Moreover, meta-graph features ar...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a novel framework to conduct evaluation-free model selection for graph learning models, i.e., without having to train/evaluate any model on the new graph. The framework learns latent embeddings for observed models and the corresponding performance on observed graphs. Moreover, meta-graph fea...
This work derived a new deep adversary CCA method, which is called adCCA, from DCCA framework. The authors frame CCA problem under the assumption that the representations of different modalities obey an identical distribution in a common latent subspace. Compared to other DNN-based CCA methods, the authors claim 1) the...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work derived a new deep adversary CCA method, which is called adCCA, from DCCA framework. The authors frame CCA problem under the assumption that the representations of different modalities obey an identical distribution in a common latent subspace. Compared to other DNN-based CCA methods, the authors clai...
For the mobile UI understanding tasks, the authors achieve advanced performance on multi-task learning and few-shot learning by pretraining the vision-language model on the proposed 2.69M dataset. This pretraining-finetune framework is easily scalable to other UI modeling tasks and not needs the view hierarchy as auxil...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: For the mobile UI understanding tasks, the authors achieve advanced performance on multi-task learning and few-shot learning by pretraining the vision-language model on the proposed 2.69M dataset. This pretraining-finetune framework is easily scalable to other UI modeling tasks and not needs the view hierarchy ...
The paper proposed a communication-efficient federated learning algorithm, FedACG to accelerate the convergence of training. Specifically, by adding a momentum to the aggregated weight and using this term as a regularizer, the authors show that the model can converge faster with higher accuracy theoretically and empiri...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper proposed a communication-efficient federated learning algorithm, FedACG to accelerate the convergence of training. Specifically, by adding a momentum to the aggregated weight and using this term as a regularizer, the authors show that the model can converge faster with higher accuracy theoretically an...
This paper proves an isoperimetric theorem for the image space. Specifically, it shows, in the pixel space of images, for every class of size less than half of the space, most of the points in the class are located on the boundary of the class (hence a small adversarial perturbation will move the point out of the class...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves an isoperimetric theorem for the image space. Specifically, it shows, in the pixel space of images, for every class of size less than half of the space, most of the points in the class are located on the boundary of the class (hence a small adversarial perturbation will move the point out of t...
This paper proposes an interpretable transport map based method for explaining the distribution shift. It further proposes methods to calculate and obtain k-spare and k-cluster transport, based on pre-defined hyperparameters. Extensive evaluation has been performed to demonstrate the promising of the proposed methods. ...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes an interpretable transport map based method for explaining the distribution shift. It further proposes methods to calculate and obtain k-spare and k-cluster transport, based on pre-defined hyperparameters. Extensive evaluation has been performed to demonstrate the promising of the proposed m...
This paper proposes a method that integrates a rule-based control policy with online/offline reinforcement learning. The proposed method, RUBICON, is an extension of TD3+BC, which is one of the state-of-the-art offline RL algorithms. The basic idea is to switch the actor's loss function based on the state-action value ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a method that integrates a rule-based control policy with online/offline reinforcement learning. The proposed method, RUBICON, is an extension of TD3+BC, which is one of the state-of-the-art offline RL algorithms. The basic idea is to switch the actor's loss function based on the state-actio...
This paper tackles the question of policy diversity in RL, with the specific goal of optimizing the diversity of obtained policies _within_ different quality levels (as one would, for example, want different AI difficulty levels within video games). To do so, the authors first propose a framework that describes this ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper tackles the question of policy diversity in RL, with the specific goal of optimizing the diversity of obtained policies _within_ different quality levels (as one would, for example, want different AI difficulty levels within video games). To do so, the authors first propose a framework that describ...
The authors use interval-bound propagation to work with few-shot learning problems. Interval arithmetic was used to model the manifold of data devoted to task interpolation. Finally, the authors apply interval architecture to model-agnostic meta-learning and prototype-based metric-learning paradigms. 1. Related work se...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors use interval-bound propagation to work with few-shot learning problems. Interval arithmetic was used to model the manifold of data devoted to task interpolation. Finally, the authors apply interval architecture to model-agnostic meta-learning and prototype-based metric-learning paradigms. 1. Related...
The paper studies reinforcement learning (RL) in high-dimensional continuous action spaces. It introduces Variational Reparametrized Policy (VRP), which formulates the policy as a generative model for optimal trajectories, optimized using a variational approach. The paper claims that this allows modeling complex and mu...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies reinforcement learning (RL) in high-dimensional continuous action spaces. It introduces Variational Reparametrized Policy (VRP), which formulates the policy as a generative model for optimal trajectories, optimized using a variational approach. The paper claims that this allows modeling comple...
This work proposed a deep probabilistic framework for modeling the distributions of each time-series together using a soft distributional coherency regularization (SDCR) over forecast distributions. It has the advantages of 1) adapting to both strong and weak hierarchical consistent datasets and 2) producing well calib...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work proposed a deep probabilistic framework for modeling the distributions of each time-series together using a soft distributional coherency regularization (SDCR) over forecast distributions. It has the advantages of 1) adapting to both strong and weak hierarchical consistent datasets and 2) producing we...
The authors present a novel and simple approach for contrastive learning by leveraging LDA-based data augmentation at document level; and StanfordNLP-based antonym substitution at sentence level for dissimilar sentence pairs. Existing approaches like SimCSE (dropout-only), SSMBA and MASKER (masking-based), DeCLUTR ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors present a novel and simple approach for contrastive learning by leveraging LDA-based data augmentation at document level; and StanfordNLP-based antonym substitution at sentence level for dissimilar sentence pairs. Existing approaches like SimCSE (dropout-only), SSMBA and MASKER (masking-based), ...
The paper proposes a semantic manipulation framework, combining low-level and high-level conditions, for object removal, object replacement, semantic relationship change and object addition. Quantitative and qualitative experiments validate the effectiveness of this design. ### Strengths 1. Each of the four mentioned p...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a semantic manipulation framework, combining low-level and high-level conditions, for object removal, object replacement, semantic relationship change and object addition. Quantitative and qualitative experiments validate the effectiveness of this design. ### Strengths 1. Each of the four men...
In this paper, an interesting finds are demonstrated: instead of always finetuning the last layer of the model, tuning different layers works best for different types of distribution shifts. 3 types of shifts are discussed: (1) Input-level shift, first-layer finetuning is better; (2) Feature-level shift, mid-later bloc...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper, an interesting finds are demonstrated: instead of always finetuning the last layer of the model, tuning different layers works best for different types of distribution shifts. 3 types of shifts are discussed: (1) Input-level shift, first-layer finetuning is better; (2) Feature-level shift, mid-la...
The authors study the problem of spurious correlations during continual learning. They divide the problem into two cases: the typical spurious correlation setting, and a local spurious correlation setting where the spuriousness is due to limited data within each task. They conduct experiments for each of the two cases,...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors study the problem of spurious correlations during continual learning. They divide the problem into two cases: the typical spurious correlation setting, and a local spurious correlation setting where the spuriousness is due to limited data within each task. They conduct experiments for each of the tw...