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
In this paper, the authors propose a new task, similar to the traditional text entity linking text where a textual entity is linked to its associated entry in Wikipedia. In this work, the authors propose to link "visual" entities to Wikipedia. Because the same entity could reasonably be linked to different Wikipedia en...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors propose a new task, similar to the traditional text entity linking text where a textual entity is linked to its associated entry in Wikipedia. In this work, the authors propose to link "visual" entities to Wikipedia. Because the same entity could reasonably be linked to different Wiki...
This paper proposes PD-MORL, which extends Envelope (Yang et al., 2019). The major difference is illustrated in Equation 5 in the paper: an extra cosine similarity term is added, and the supremum is only taken over actions, not preferences (compared to Envelope). The authors evaluated their method with Envelope on Deep...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes PD-MORL, which extends Envelope (Yang et al., 2019). The major difference is illustrated in Equation 5 in the paper: an extra cosine similarity term is added, and the supremum is only taken over actions, not preferences (compared to Envelope). The authors evaluated their method with Envelope...
This paper demonstrates the point that Self-supervised learning is immune to the choice of the probabilistic model, therefore shows robustness to model misspecification. In contrast, traditional approach of probabilistic modeling hinges on the specific predefined model. Therefore, the advantage of self-supervised learn...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper demonstrates the point that Self-supervised learning is immune to the choice of the probabilistic model, therefore shows robustness to model misspecification. In contrast, traditional approach of probabilistic modeling hinges on the specific predefined model. Therefore, the advantage of self-supervis...
In this paper, the authors propose to use three different policies to modify simultaneously the actions, morphology, and environment in a Reinforcement Learning problem. While the control policy is very standard in RL applications, this paper also considers morphology and environment modifications as Markov Decision P...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In this paper, the authors propose to use three different policies to modify simultaneously the actions, morphology, and environment in a Reinforcement Learning problem. While the control policy is very standard in RL applications, this paper also considers morphology and environment modifications as Markov De...
This paper presents In-sample Actor-Critic (IAC), an algorithm for offline RL. The policy evaluation of IAC is done only using in-distribution samples of the dataset, where each sample's (normalized) importance ratio is used for resampling probabilities. Then, TD updates are performed using the resampled samples. For p...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents In-sample Actor-Critic (IAC), an algorithm for offline RL. The policy evaluation of IAC is done only using in-distribution samples of the dataset, where each sample's (normalized) importance ratio is used for resampling probabilities. Then, TD updates are performed using the resampled sample...
The authors introduce D-BAT, a diversity-inducing regularizer for training ensembles of diverse predictors. They derive D-BAT mathematically, and evaluate it on several datasets to demonstrate that the induced diversity can help to (i) tackle shortcut learning, and (ii) improve uncertainty estimation and transferabilit...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The authors introduce D-BAT, a diversity-inducing regularizer for training ensembles of diverse predictors. They derive D-BAT mathematically, and evaluate it on several datasets to demonstrate that the induced diversity can help to (i) tackle shortcut learning, and (ii) improve uncertainty estimation and transf...
Sharpness-aware minimization (SAM) has recently been shown to improve various aspects of deep learning. In this paper, the authors show empirically that SAM also helps to improve the performance in decentralized federated learning. A main contribution of this paper is a theoretical analysis of their federated learning...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Sharpness-aware minimization (SAM) has recently been shown to improve various aspects of deep learning. In this paper, the authors show empirically that SAM also helps to improve the performance in decentralized federated learning. A main contribution of this paper is a theoretical analysis of their federated ...
This work presents two new approaches to zero-shot learning via building generative models of classifier weights. Contrastive learning is used to build a CLIP-style model which matches task description with network weights. In one approach, this CLIP-style model is used at test time to maximize match between provided t...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This work presents two new approaches to zero-shot learning via building generative models of classifier weights. Contrastive learning is used to build a CLIP-style model which matches task description with network weights. In one approach, this CLIP-style model is used at test time to maximize match between pr...
This paper challenges the common belief in literature in self supervision that augmentations need to be label preserving, in order to perform well in downstream tasks. The authors first show that when using a Viewmaker network to automatically derive good views, the augmentations derived by the model may drop significa...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper challenges the common belief in literature in self supervision that augmentations need to be label preserving, in order to perform well in downstream tasks. The authors first show that when using a Viewmaker network to automatically derive good views, the augmentations derived by the model may drop s...
This paper designs a new distance metric called Oblique(d/m, m) instead of the fashionable cosine similarity for the contrastive learning paradigm. To verify the effectiveness of the proposed simple method, the authors perform extensive experiments. ##Strength： The proposed new distance metric is neat and could be eas...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper designs a new distance metric called Oblique(d/m, m) instead of the fashionable cosine similarity for the contrastive learning paradigm. To verify the effectiveness of the proposed simple method, the authors perform extensive experiments. ##Strength： The proposed new distance metric is neat and coul...
The paper proposes a new divergence which is a convolution between Renyi divergence and other IPM. Strength: 1. The paper proposes a new divergence which include the Renyi divergence and other IPM. It is new and have some interesting properties. 2. The author derives some properties of the new divergence and show its...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper proposes a new divergence which is a convolution between Renyi divergence and other IPM. Strength: 1. The paper proposes a new divergence which include the Renyi divergence and other IPM. It is new and have some interesting properties. 2. The author derives some properties of the new divergence and ...
This paper presents a video self-supervised pretraining pipeline called VITO. It made several modifications over existing contrastive learning frameworks including larger crop size, improved temporal sampling scheme, and multi-scale attention feature pooling for the projector. The authors also investigated the data dom...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a video self-supervised pretraining pipeline called VITO. It made several modifications over existing contrastive learning frameworks including larger crop size, improved temporal sampling scheme, and multi-scale attention feature pooling for the projector. The authors also investigated the ...
The authors propose a deep convolution network (DECN) that mimics the operation of evolutionary search techniques, specifically recombination and selection, for black-box optimization problems. The paper first describes the general challenges in creating a convolution operator to mimic evolutionary, as well as related ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors propose a deep convolution network (DECN) that mimics the operation of evolutionary search techniques, specifically recombination and selection, for black-box optimization problems. The paper first describes the general challenges in creating a convolution operator to mimic evolutionary, as well as ...
The paper considers an autoencoder architecture where the structure in the latent space is induced by architecture of the decoder, which uses latent variables sequentially to generate samples from the target distribution. Rather than imposing a prior distribution on the latent space authors use a “hybrid sampling” appr...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper considers an autoencoder architecture where the structure in the latent space is induced by architecture of the decoder, which uses latent variables sequentially to generate samples from the target distribution. Rather than imposing a prior distribution on the latent space authors use a “hybrid sampli...
In this paper, the authors propose to use a Maximum-Entropy Rewarded Reinforcement Learning framework to select training data for NLP tasks, the goal of which is to maximize generalization. The authors experiment with A2C and SAC and experimental results show that the proposed framework could outperform several baseli...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors propose to use a Maximum-Entropy Rewarded Reinforcement Learning framework to select training data for NLP tasks, the goal of which is to maximize generalization. The authors experiment with A2C and SAC and experimental results show that the proposed framework could outperform severa...
This work has proposed a new adversarial defense method based on multi-task learning and reactive perturbation defocusing. This method has achieved very impressive restored accuracy. Strength: 1. The proposed perturbation defocusing method can restore the attacked accuracy from very low accuracy to almost the same as ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work has proposed a new adversarial defense method based on multi-task learning and reactive perturbation defocusing. This method has achieved very impressive restored accuracy. Strength: 1. The proposed perturbation defocusing method can restore the attacked accuracy from very low accuracy to almost the ...
Large language models trained on source code can generate many plausible solutions to programming problems. This work proposes to simultaneously use these models to generate tests for the same prompts, which are then used to re-rank the generated solutions, using a consensus set based algorithm. The results show that t...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Large language models trained on source code can generate many plausible solutions to programming problems. This work proposes to simultaneously use these models to generate tests for the same prompts, which are then used to re-rank the generated solutions, using a consensus set based algorithm. The results sho...
This paper proposes a method for feature distillation between a teacher model and a student model. It aims to improve the student model's accuracy by utilizing knowledge from the more capable teacher model. The method adds a linear residual module after the last convolution layer of the student model. The module firs...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a method for feature distillation between a teacher model and a student model. It aims to improve the student model's accuracy by utilizing knowledge from the more capable teacher model. The method adds a linear residual module after the last convolution layer of the student model. The mod...
In this paper, the authors proposed a novel convolutional neural network (CLEEGN) for plug-and-play automatic EEG reconstruction. CLEEGN can reconstruct subject-independent EEG without any training/calibration for a new subject. The performance of CLEEGN was validated using multiple evaluations including reconstruction...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: In this paper, the authors proposed a novel convolutional neural network (CLEEGN) for plug-and-play automatic EEG reconstruction. CLEEGN can reconstruct subject-independent EEG without any training/calibration for a new subject. The performance of CLEEGN was validated using multiple evaluations including recons...
This paper proposes a novel transformer-based model for symbolic regression (SR), which produces mathematical expression skeletons from data points. In the proposed method, a feature extractor using pointMLP is added and jointly trained using contrastive loss to realize efficient training. The experimental evaluation u...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a novel transformer-based model for symbolic regression (SR), which produces mathematical expression skeletons from data points. In the proposed method, a feature extractor using pointMLP is added and jointly trained using contrastive loss to realize efficient training. The experimental eval...
This work analyzes the benign overfitting when the logistic regression is mild over-parameterization, in the setting of label noise. The core theory results are: 1) excess risk can be lower bounded at infinite SGD iteration; 2) the risk with early stopping can be upper bounded by 1/n (the number of samples. The authors...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work analyzes the benign overfitting when the logistic regression is mild over-parameterization, in the setting of label noise. The core theory results are: 1) excess risk can be lower bounded at infinite SGD iteration; 2) the risk with early stopping can be upper bounded by 1/n (the number of samples. The...
This paper proposes a new SNN model, named d-block model. By adding stochastic absolute refractory periods and recurrent conductance latencies, a d-block SNN can reduce the number of sequential computations using fast vectorized operations. Strength: The proposed model achieves fewer sequential operations and lower e...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a new SNN model, named d-block model. By adding stochastic absolute refractory periods and recurrent conductance latencies, a d-block SNN can reduce the number of sequential computations using fast vectorized operations. Strength: The proposed model achieves fewer sequential operations and...
The author(s) studied the multi-objective optmization problem and proposed a new algorithm called MoCo that provably converges to a Pareto stationary point in the stochastic gradient setting. Experiments on both toy and real data are conducted to support the proposed method. Strengths: - The paper is well-motivated and...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: The author(s) studied the multi-objective optmization problem and proposed a new algorithm called MoCo that provably converges to a Pareto stationary point in the stochastic gradient setting. Experiments on both toy and real data are conducted to support the proposed method. Strengths: - The paper is well-motiv...
This paper studies Federated stochastic optimization with client sampling. Under the assumption of second-order similarity condition, the authors propose a variance-reduction based proximal point algorithm which enjoy a better convergence rate wrt the number of clients. They also proposed an accelerated version which e...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies Federated stochastic optimization with client sampling. Under the assumption of second-order similarity condition, the authors propose a variance-reduction based proximal point algorithm which enjoy a better convergence rate wrt the number of clients. They also proposed an accelerated version...
The authors propose an approach to distillation where the teacher provides both a predictive target and scaffolds the student's prediction by censoring hard-to-learn examples. In the case where the student has far fewer parameters than the teacher this scaffolding leads to a smoother loss landscape for the student and ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose an approach to distillation where the teacher provides both a predictive target and scaffolds the student's prediction by censoring hard-to-learn examples. In the case where the student has far fewer parameters than the teacher this scaffolding leads to a smoother loss landscape for the stud...
This paper aims to learn a generative model of 3D scenes from large unaligned datasets. The paper argues that previous work only produced good results for well-aligned datasets (e.g. human and animal faces), and that directly applying such methods to unaligned data did not work well. In order to address this problem, t...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper aims to learn a generative model of 3D scenes from large unaligned datasets. The paper argues that previous work only produced good results for well-aligned datasets (e.g. human and animal faces), and that directly applying such methods to unaligned data did not work well. In order to address this pr...
This paper extends the Deep Eikonal Solver of Liechtenstein et al. It achieves higher empirical accuracy by allowing the local solver to see a 3-ring rather than 2-ring neighborhood, and by training against ground truth distances computed on a finer mesh. The results display generalization from polynomial surfaces to s...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper extends the Deep Eikonal Solver of Liechtenstein et al. It achieves higher empirical accuracy by allowing the local solver to see a 3-ring rather than 2-ring neighborhood, and by training against ground truth distances computed on a finer mesh. The results display generalization from polynomial surfa...
This paper studies imbalanced semi-supervised learning. SOTA imbalanced semi-supervised learning methods are often based on pseudo-labeling and consistency regularization, which still relies on confidence thresholding. In this paper, the authors formulate the pseudo-labeling problem as a classification problem, i.e., u...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper studies imbalanced semi-supervised learning. SOTA imbalanced semi-supervised learning methods are often based on pseudo-labeling and consistency regularization, which still relies on confidence thresholding. In this paper, the authors formulate the pseudo-labeling problem as a classification problem,...
The paper introduces a new family of probability distributions on SO(3) called Rotation Laplace distributions. The literature has established how a matrix Fisher distribution on SO(3) approximates a zero-mean multivariate Gaussian distribution on the tangent space $T_{\mathbf{R}_0}SO(3)$, where $\mathbf{R}_0$ is the mo...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper introduces a new family of probability distributions on SO(3) called Rotation Laplace distributions. The literature has established how a matrix Fisher distribution on SO(3) approximates a zero-mean multivariate Gaussian distribution on the tangent space $T_{\mathbf{R}_0}SO(3)$, where $\mathbf{R}_0$ i...
This paper considers new drug recommendations as a few-shot learning problem. The proposed model can quickly adapt to the recommendation for a new drug with limited prescription data from a few support patients by addressing complex relations among diseases and drugs and numerous false-negative patients. Experiments o...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper considers new drug recommendations as a few-shot learning problem. The proposed model can quickly adapt to the recommendation for a new drug with limited prescription data from a few support patients by addressing complex relations among diseases and drugs and numerous false-negative patients. Exper...
The authors present a transformer-encoder framework for multipleThe authors present a encoder/transformer method for learning representations of EHR data, for multiple prediction tasks. They demonstrate that their method out-performs other DL methods on one prediction task using MIMIC-III data. Strengths: - This pap...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors present a transformer-encoder framework for multipleThe authors present a encoder/transformer method for learning representations of EHR data, for multiple prediction tasks. They demonstrate that their method out-performs other DL methods on one prediction task using MIMIC-III data. Strengths: - ...
This paper focused on learning a structured node representation on top of a GNN model. It follows the idea of FlowGMM and applies normalizing flow to the GNN model. The proposed GC-Flow model leads to significantly better clustering results and improves node classification. Strengths: 1. The proposed GC-Flow model is ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focused on learning a structured node representation on top of a GNN model. It follows the idea of FlowGMM and applies normalizing flow to the GNN model. The proposed GC-Flow model leads to significantly better clustering results and improves node classification. Strengths: 1. The proposed GC-Flow m...
This paper studies how to reduce ReLU operation more efficiently to reduce the communication and latency overhead of privacy inference. They demonstrate the relation between a layer’s sensitivity towards pruning and its associated ReLU sensitivity and introduce an automated layer-wise ReLU sensitivity evaluation strate...	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 studies how to reduce ReLU operation more efficiently to reduce the communication and latency overhead of privacy inference. They demonstrate the relation between a layer’s sensitivity towards pruning and its associated ReLU sensitivity and introduce an automated layer-wise ReLU sensitivity evaluatio...
Note: Throughout the review, I will make a distinction between dense mini-batch SGD, i.e., mini-batch SGD with batch size $B$ and large mini-batch SGD with batch size $B\cdot H$, where $H$ is the number of local steps, for the local-SGD algorithm with the exact computation and communication cost, and $B$ is the bas...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: Note: Throughout the review, I will make a distinction between dense mini-batch SGD, i.e., mini-batch SGD with batch size $B$ and large mini-batch SGD with batch size $B\cdot H$, where $H$ is the number of local steps, for the local-SGD algorithm with the exact computation and communication cost, and $B$ is...
This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning.Some carefully designed experiments have verified theoretical results. Strength: Thi...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning.Some carefully designed experiments have verified theoretical results. Stren...
This paper proposed a continual learning method by constructing low-coherence subspace projector for each new task. Given a new task, the projector matrix is constructed by minimizing the coherence to the previous task projectors and within the current projector to be optimized, in the oblique manifold. The proposed ap...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a continual learning method by constructing low-coherence subspace projector for each new task. Given a new task, the projector matrix is constructed by minimizing the coherence to the previous task projectors and within the current projector to be optimized, in the oblique manifold. The pro...
The authors propose a variant of relational networks to reason about interactions between objects and between objects and an agent in a manipulation task. They train policies with this architecture on a few simulated manipulation tasks, and demonstrate out-of-distribution generalization to variants of the task not seen...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose a variant of relational networks to reason about interactions between objects and between objects and an agent in a manipulation task. They train policies with this architecture on a few simulated manipulation tasks, and demonstrate out-of-distribution generalization to variants of the task ...
This paper proposes a new distillation technique for learning Information Retrieval (IR) models whereby not only the scores of the documents are aligned with the teacher, but also the representations. The paper exhibits two propositions showing (1) a relationship involving the empirical distillation loss and the repres...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new distillation technique for learning Information Retrieval (IR) models whereby not only the scores of the documents are aligned with the teacher, but also the representations. The paper exhibits two propositions showing (1) a relationship involving the empirical distillation loss and th...
This paper presents an approach to use diffusion models as a manner to parameterize a meta-learning algorithm, where the network perbutation prediction by the diffusion process adapts a base model to a separate task. # Strengths Novelty. The problem studied by the paper is interesting and is new to my knowledge #...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presents an approach to use diffusion models as a manner to parameterize a meta-learning algorithm, where the network perbutation prediction by the diffusion process adapts a base model to a separate task. # Strengths Novelty. The problem studied by the paper is interesting and is new to my know...
This paper proposes the opinion that the relative representation, the representation of data described by a fixed set of anchor representations, is invariant among different randomized training factors like seeds, optimization strategies, training steps, and even architectures. To prove their opinion, the authors then ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes the opinion that the relative representation, the representation of data described by a fixed set of anchor representations, is invariant among different randomized training factors like seeds, optimization strategies, training steps, and even architectures. To prove their opinion, the autho...