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
This paper proposed an unsupervised learning framework for denoising, without assuming noise priors. Three stages are designed, i.e., generating noise, psedo supervised learning and recorrupted2recorrupted. The results on synthetic Gaussian noise and real noise datasets show the proposed method is comparable with exist...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed an unsupervised learning framework for denoising, without assuming noise priors. Three stages are designed, i.e., generating noise, psedo supervised learning and recorrupted2recorrupted. The results on synthetic Gaussian noise and real noise datasets show the proposed method is comparable wi...
Background: DP-SGD gives guarantees for the whole history of gradients. It very difficult to obtain guarantees that would hold only for the final model, so then it is natural to think, how could we use the history of gradients since due to post-processing that will not have any additional privacy cost. The paper propo...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: Background: DP-SGD gives guarantees for the whole history of gradients. It very difficult to obtain guarantees that would hold only for the final model, so then it is natural to think, how could we use the history of gradients since due to post-processing that will not have any additional privacy cost. The pap...
This paper proposes a novel supervised learning model in the scenarios of crowdsourcing based on AUM, which is a confidence indicator for each task. The proposed method generalizes the existing AUM indicator to the crowdsourcing setting as WAUM (the Weighted Area Under the Margin), and the WAUM identifies harmful data ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a novel supervised learning model in the scenarios of crowdsourcing based on AUM, which is a confidence indicator for each task. The proposed method generalizes the existing AUM indicator to the crowdsourcing setting as WAUM (the Weighted Area Under the Margin), and the WAUM identifies harmf...
The paper formulates a third-order accurate fast-marching style algorithm for finding geodesic distances on surfaces embedded in 3D. The local solver is a trained neural network which operates on the 3-ring of the point in question. The algorithm has quasilinear complexity ultimately. For training, a multi-resolution a...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper formulates a third-order accurate fast-marching style algorithm for finding geodesic distances on surfaces embedded in 3D. The local solver is a trained neural network which operates on the 3-ring of the point in question. The algorithm has quasilinear complexity ultimately. For training, a multi-reso...
The paper studies the problem of solving the alleged inconsistencies between the policy evaluation and policy improvement steps of the “policy iteration” algorithm. The paper proposes a method called CASA as a solution which suggests some modifications on the gradient calculations of standard algorithms such as PPO. *S...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies the problem of solving the alleged inconsistencies between the policy evaluation and policy improvement steps of the “policy iteration” algorithm. The paper proposes a method called CASA as a solution which suggests some modifications on the gradient calculations of standard algorithms such as...
The authors propose a novel variational causal inference model for constructing the gene expression counts in cells after counterfactual perturbations that relies on information from two sources: individual features embedded in the outcome Y, and response distributions of similar individuals that did indeed receive the...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors propose a novel variational causal inference model for constructing the gene expression counts in cells after counterfactual perturbations that relies on information from two sources: individual features embedded in the outcome Y, and response distributions of similar individuals that did indeed rec...
This paper proposes a way to define a metric (called GFMMD) for non-negative graph signals in a way to account for geometric structure. It comes from by mimicing the so-called Integral Probability Metrics, particularly 1-Wasserstein distance for discrete distributions. GFMMD can be computed rapidly (relying on approxim...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a way to define a metric (called GFMMD) for non-negative graph signals in a way to account for geometric structure. It comes from by mimicing the so-called Integral Probability Metrics, particularly 1-Wasserstein distance for discrete distributions. GFMMD can be computed rapidly (relying on ...
This paper proposes RTC, an imitation learning approach for stochastic environments that aims to learn policies that avoid mode collapse and cover a wide range of behaviors shown in demonstrations. The paper argues that current methods relying on inferring agent types to explain multi-modal behaviors in stochastic envi...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes RTC, an imitation learning approach for stochastic environments that aims to learn policies that avoid mode collapse and cover a wide range of behaviors shown in demonstrations. The paper argues that current methods relying on inferring agent types to explain multi-modal behaviors in stochas...
This paper proposes OTGNet to tackle the problem of continual learning/class-incremental learning on graphs. OTGNet consists of two components. The first component is an open/close triad selection and replay method to avoid catastrophic forgetting of existing knowledge. The second component is an information bottleneck...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes OTGNet to tackle the problem of continual learning/class-incremental learning on graphs. OTGNet consists of two components. The first component is an open/close triad selection and replay method to avoid catastrophic forgetting of existing knowledge. The second component is an information bo...
This paper studies an interesting layer-grafting technique to promote the combination of CL and MIM for representation learning. Specially, the authors find that directly combining such two paradigms shows the negative influence due to the gradient conflicts, and discovers that the sequential cascade in a certain order...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper studies an interesting layer-grafting technique to promote the combination of CL and MIM for representation learning. Specially, the authors find that directly combining such two paradigms shows the negative influence due to the gradient conflicts, and discovers that the sequential cascade in a certa...
This paper considers distributionally robust probabilistic supervised learning. The ambiguity set is constructed to include distributions that share the same marginal with the empirical distribution and are no more than ε away from the empirical in terms of first-order feature moment divergence. The strong duality is s...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper considers distributionally robust probabilistic supervised learning. The ambiguity set is constructed to include distributions that share the same marginal with the empirical distribution and are no more than ε away from the empirical in terms of first-order feature moment divergence. The strong dual...
This work proposes a new unrolled network architecture for inverse problems. It's main contribution is to replace the large feed-forward architecture (e.g., DnCNN) used in typical unrolled architectures with a small recurrent architecture, known as a Deep Equilibrium (DEQ) model. The DEQ model iterates until it converg...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work proposes a new unrolled network architecture for inverse problems. It's main contribution is to replace the large feed-forward architecture (e.g., DnCNN) used in typical unrolled architectures with a small recurrent architecture, known as a Deep Equilibrium (DEQ) model. The DEQ model iterates until it...
This paper studies how fixed-length Transformer networks can approximate anisotropic Besov and mixed smooth Besov spaces. They show that the expected approximation error of fixed-length Transformer networks is mildly dependent on feature dimension $d$ and input length $l$, and is independent of $d$ and $l$ for anisotr...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies how fixed-length Transformer networks can approximate anisotropic Besov and mixed smooth Besov spaces. They show that the expected approximation error of fixed-length Transformer networks is mildly dependent on feature dimension $d$ and input length $l$, and is independent of $d$ and $l$ for...
This paper proposes a variational expectation maximization framework to jointly train the language model and the graph neural network for representation learning for text-attributed graphs. The approach is clearly motivated, mathematically principled, and empirically effective and efficient. I believe this paper makes ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a variational expectation maximization framework to jointly train the language model and the graph neural network for representation learning for text-attributed graphs. The approach is clearly motivated, mathematically principled, and empirically effective and efficient. I believe this pape...
This paper proposes a greedy-based approach for online feature selection. This paper defines greedy online feature selection and also provides an iterative procedure to implement the greedy approach with a deep learning approach. Strengths: + The problem is interesting and meaningful. Feature selection, especially o...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a greedy-based approach for online feature selection. This paper defines greedy online feature selection and also provides an iterative procedure to implement the greedy approach with a deep learning approach. Strengths: + The problem is interesting and meaningful. Feature selection, espe...
* The paper adds better inductive bias to Transformer attention by using a linear recurrence in the form of exponential moving average to contextualize the queries and keys. * The paper improves simple damped EMA into a multidimensional damped EMA turning it effectively in a simplified form of S4 where a diagonal bou...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: * The paper adds better inductive bias to Transformer attention by using a linear recurrence in the form of exponential moving average to contextualize the queries and keys. * The paper improves simple damped EMA into a multidimensional damped EMA turning it effectively in a simplified form of S4 where a diag...
This submission is considering if multiple federated learning tasks on a single device can be coordinated and consolidated to ease resource constrain and energy consumption. A multi-tenant FL system called MuFL is proposed that 1) Initially merges "similar" training activities into one activity with a multi-task archit...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This submission is considering if multiple federated learning tasks on a single device can be coordinated and consolidated to ease resource constrain and energy consumption. A multi-tenant FL system called MuFL is proposed that 1) Initially merges "similar" training activities into one activity with a multi-tas...
This paper provides a new regularization perspective on analyzing why the distribution RL methods perform better than the expectation-based RL methods empirically. By leveraging the decomposition assumption of the action-value distributions, this paper proposes to express the objective function of NeuralFZI as the expe...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper provides a new regularization perspective on analyzing why the distribution RL methods perform better than the expectation-based RL methods empirically. By leveraging the decomposition assumption of the action-value distributions, this paper proposes to express the objective function of NeuralFZI as ...
In this paper, the authors propose an accelerated first-order method for the optimization of the Hadamard manifold. They theoretically show that the proposed algorithm enjoys the same convergence rate as Nesterov's accelerated gradient descent method in Euclidian space, under the assumption that the objective function ...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: In this paper, the authors propose an accelerated first-order method for the optimization of the Hadamard manifold. They theoretically show that the proposed algorithm enjoys the same convergence rate as Nesterov's accelerated gradient descent method in Euclidian space, under the assumption that the objective f...
This paper proposes a solution to “embodied reference understanding”, which aims to locate so-called “referents” which are objects referred to by text and gestures. The core observation of this work is that referents do not lie on the elbow-wrist line but rather on the eye-wrist line, which they call the “virtual touch...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a solution to “embodied reference understanding”, which aims to locate so-called “referents” which are objects referred to by text and gestures. The core observation of this work is that referents do not lie on the elbow-wrist line but rather on the eye-wrist line, which they call the “virtu...
This paper analyses representations learned by neural network. The analysis is done via isometric mapping of the probabilistic model into a lower dimensional maniforld such that it preserves the pairwise distance between models. They propose a distance metric to compare different learning trajectories. They also use I...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper analyses representations learned by neural network. The analysis is done via isometric mapping of the probabilistic model into a lower dimensional maniforld such that it preserves the pairwise distance between models. They propose a distance metric to compare different learning trajectories. They al...
In (Safran et al. 2019), the authors describe a function which is $ReLU(1-\lVert x\rVert)$, which can be well approximated by a depth 3-network, while a depth 2-network requires $\Omega(\exp(d\log d))$ width for the same. In this paper the authors show that, this separation is also, algorithmic, in the sense that one c...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In (Safran et al. 2019), the authors describe a function which is $ReLU(1-\lVert x\rVert)$, which can be well approximated by a depth 3-network, while a depth 2-network requires $\Omega(\exp(d\log d))$ width for the same. In this paper the authors show that, this separation is also, algorithmic, in the sense th...
The paper hypothesise that function sharing is one of the reasons why deep learning models can’t perform systematic generalization. The paper demonstrate that as the degree of parameter sharing increases, the systematic generalization drops. From the practical stand point, the papers argues for sparsity in models that ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper hypothesise that function sharing is one of the reasons why deep learning models can’t perform systematic generalization. The paper demonstrate that as the degree of parameter sharing increases, the systematic generalization drops. From the practical stand point, the papers argues for sparsity in mode...
In this work, the authors propose a new method for federated semi-supervised learning relying on a coarse grained regulator and fine grained regulator. This work handles scenarios where the labeled data and unlabeled data share different distribution. Empirical results show that the proposed method is able to outperfor...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: In this work, the authors propose a new method for federated semi-supervised learning relying on a coarse grained regulator and fine grained regulator. This work handles scenarios where the labeled data and unlabeled data share different distribution. Empirical results show that the proposed method is able to o...
This paper puts forth a fully-cooperative multi-agent/task algorithm, MATTAR, that can be used to transfer knowledge learned from source tasks to improve learning unseen tasks at test time. MATTAR learns a high-level mapping from task representations to state, observation, and reward function parameters and combines th...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper puts forth a fully-cooperative multi-agent/task algorithm, MATTAR, that can be used to transfer knowledge learned from source tasks to improve learning unseen tasks at test time. MATTAR learns a high-level mapping from task representations to state, observation, and reward function parameters and com...
This paper considers the problem of OOD generalization in L2O, i.e. meta-learning an optimizer that is effective on objectives not seen during training. It proposes an algorithm based on MAML (Finn et al. 2017) and analyzes its generalization ability, measured by the loss at testing time. Unsurprisingly, generalization...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper considers the problem of OOD generalization in L2O, i.e. meta-learning an optimizer that is effective on objectives not seen during training. It proposes an algorithm based on MAML (Finn et al. 2017) and analyzes its generalization ability, measured by the loss at testing time. Unsurprisingly, genera...
This paper studies Federated Long-tailed Data, which is a highly challenging task. To address this task, this paper proposes to use a privacy-preserving proxy to guide model training and develops a label-distribution-agnostic ensemble learning framework. Both theoretical analysis and empirical verification are provided...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies Federated Long-tailed Data, which is a highly challenging task. To address this task, this paper proposes to use a privacy-preserving proxy to guide model training and develops a label-distribution-agnostic ensemble learning framework. Both theoretical analysis and empirical verification are ...
This paper focuses on the edge noisy scenario in the link prediction task and proposes a method named Robust Graph Information Bottleneck. The self-supervised learning technique and data reparametrization mechanism are utilized to instantiate RGIB. Strength: 1. The coupled edge noise in the link prediction task is a g...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focuses on the edge noisy scenario in the link prediction task and proposes a method named Robust Graph Information Bottleneck. The self-supervised learning technique and data reparametrization mechanism are utilized to instantiate RGIB. Strength: 1. The coupled edge noise in the link prediction tas...
This work revisits the GRBM to improve its learning and sampling process. One key innovation is introducing Langevin sampling to GRBMs. Two variants are proposed: sampling the visible marginal potential, and hybrid Gibbs-Langevin sampling that trades off between Gibbs updates of hidden units and Langevin updates of vis...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This work revisits the GRBM to improve its learning and sampling process. One key innovation is introducing Langevin sampling to GRBMs. Two variants are proposed: sampling the visible marginal potential, and hybrid Gibbs-Langevin sampling that trades off between Gibbs updates of hidden units and Langevin update...
The paper attempts to investigate the problem of why object detection models fail with adversarial attacks. The authors tried to look into the classification and localization branch independently to locate the issue. Also, different types of object detection networks were investigated for this issue. According to exper...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: The paper attempts to investigate the problem of why object detection models fail with adversarial attacks. The authors tried to look into the classification and localization branch independently to locate the issue. Also, different types of object detection networks were investigated for this issue. According ...
Authors propose a method for learning with high label noise that is based on learning a predictor and a selector function. They introduce a loss function that weights the abstained points by the predictor’s correctness. They present a theoretical analysis in the noisy setting for picking the best selector function giv...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: Authors propose a method for learning with high label noise that is based on learning a predictor and a selector function. They introduce a loss function that weights the abstained points by the predictor’s correctness. They present a theoretical analysis in the noisy setting for picking the best selector func...
This paper extends MAE from single modality to audio and visual modalities by combining contrastive learning and masked prediction pretext tasks. Experiments show the improvements on VGGSound dataset and better retrieval score. Strength: With the advent of MAE and audioMAE, it is natural to extend such framework to m...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper extends MAE from single modality to audio and visual modalities by combining contrastive learning and masked prediction pretext tasks. Experiments show the improvements on VGGSound dataset and better retrieval score. Strength: With the advent of MAE and audioMAE, it is natural to extend such framew...
The paper proposes a solution for selecting models for time-series anomaly detection. Starting from the assumption that no single model performs best across all datasets, the paper argues for the importance of AutoML solutions in that area. As a result, the paper first studies surrogate measures that correlate with the...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposes a solution for selecting models for time-series anomaly detection. Starting from the assumption that no single model performs best across all datasets, the paper argues for the importance of AutoML solutions in that area. As a result, the paper first studies surrogate measures that correlate ...
The authors present a method for learning a model that outputs probabilities wherein the output probability is the output of an optimization over an uncertainty set rather than, say the result of a softmax. They show how to learn a model using this approach and compare the method to simple alternatives on several datas...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The authors present a method for learning a model that outputs probabilities wherein the output probability is the output of an optimization over an uncertainty set rather than, say the result of a softmax. They show how to learn a model using this approach and compare the method to simple alternatives on sever...
The paper presents an overall pretraining approach for object detection with a fixed pretrained backbone. A transformer-based detector is used. The main contribution of the paper is the approach to selecting positive examples in a contrastive learning method. Strength: - The overall paper is well-written and defined. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper presents an overall pretraining approach for object detection with a fixed pretrained backbone. A transformer-based detector is used. The main contribution of the paper is the approach to selecting positive examples in a contrastive learning method. Strength: - The overall paper is well-written and d...
The paper proposes a general framework which unifies problem complexity measures with function approximation in the model-based and model-free lens for reinforcement learning. In particular, the authors propose Admissible Bellman Characterization (ABC) class that subsumes many prior models for problem complexity in th...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper proposes a general framework which unifies problem complexity measures with function approximation in the model-based and model-free lens for reinforcement learning. In particular, the authors propose Admissible Bellman Characterization (ABC) class that subsumes many prior models for problem complexi...
This paper introduces an approach to train a neural network for a multiclass image classification task. The main contribution is to add an auxiliary task, where Binary Labels are used as supervision. The Binary Labels can be carefully constructed or generated randomly based on the number of classes. The proposed approa...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper introduces an approach to train a neural network for a multiclass image classification task. The main contribution is to add an auxiliary task, where Binary Labels are used as supervision. The Binary Labels can be carefully constructed or generated randomly based on the number of classes. The propose...
The paper studies identifiability of label noise transition matrix T(X), which plays a crucial role in learning with noisy labels. Most of the works assume access to it or rely on some methods to estimate it. Understanding when such a noise matrix is identifiable is an important aspect of the problem. The paper conside...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper studies identifiability of label noise transition matrix T(X), which plays a crucial role in learning with noisy labels. Most of the works assume access to it or rely on some methods to estimate it. Understanding when such a noise matrix is identifiable is an important aspect of the problem. The paper...
The paper studies the problem of task continual learning wherein new tasks are continuously learned in an incremental fashion. Specifically, the paper aims to achieve two key objectives for a successful continual learning system: preventing catastrophic forgetting of the previous tasks and enabling knowledge transfer t...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper studies the problem of task continual learning wherein new tasks are continuously learned in an incremental fashion. Specifically, the paper aims to achieve two key objectives for a successful continual learning system: preventing catastrophic forgetting of the previous tasks and enabling knowledge tr...
Through multiple iteration of rebuttal, the authors have provided empirical evidence to address my concern. I am willing to raise my score to 6, conditioned on polishing the method section. The following are my original review: This paper proposes to sample rows and columns of key and query matrices to approximate att...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Through multiple iteration of rebuttal, the authors have provided empirical evidence to address my concern. I am willing to raise my score to 6, conditioned on polishing the method section. The following are my original review: This paper proposes to sample rows and columns of key and query matrices to approxi...