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
The paper proposes a benchmark for Federated Learning in the context of domain generalization. The authors thoroughly evaluate a number of popular DG algorithms on this benchmark. Strengths - The benchmark considers a relevant setting - The evaluations are detailed covering a variety of methods - The work highlights t...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a benchmark for Federated Learning in the context of domain generalization. The authors thoroughly evaluate a number of popular DG algorithms on this benchmark. Strengths - The benchmark considers a relevant setting - The evaluations are detailed covering a variety of methods - The work high...
This paper presents a novel method of learning robust DNNs through the use of a network containing two si models, one that learns an intermediate variable Z and then combines it with the front door criterion to learn the interventional distribution p(Y\|do(X = x)). This interventional distribution has invariance propert...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper presents a novel method of learning robust DNNs through the use of a network containing two si models, one that learns an intermediate variable Z and then combines it with the front door criterion to learn the interventional distribution p(Y\|do(X = x)). This interventional distribution has invariance...
This paper proposes a new neighborhood aggregation method based on the number of hops. With this method, a graph transformer is developed with node-wise input sequences. Moreover, a self-attention readout function is applied to integrate the representations learned from the transformer module. Strengths: 1. With node-...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a new neighborhood aggregation method based on the number of hops. With this method, a graph transformer is developed with node-wise input sequences. Moreover, a self-attention readout function is applied to integrate the representations learned from the transformer module. Strengths: 1. Wi...
This paper studies the performance fairness problem of federated learning. The authors empirically show that Q-Fair Federated Learning fails to improve performance fairness. They also show that knowledge distillation is a better way for fair federated learning without additional personalization mechanisms. Strengths: -...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies the performance fairness problem of federated learning. The authors empirically show that Q-Fair Federated Learning fails to improve performance fairness. They also show that knowledge distillation is a better way for fair federated learning without additional personalization mechanisms. Stre...
The authors provide to their knowledge the first matrix-free Newton's method in reduced subspace to obtain exact solution of cubic-regularization (CR), which is much faster than the previous CR solvers. Then, the previous ARC algorithm is expanded to ARCLQN algorithm using LQN matrices, to incorporate SGD and the propo...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors provide to their knowledge the first matrix-free Newton's method in reduced subspace to obtain exact solution of cubic-regularization (CR), which is much faster than the previous CR solvers. Then, the previous ARC algorithm is expanded to ARCLQN algorithm using LQN matrices, to incorporate SGD and t...
The paper cites the feature robustness framework of Ilyas et al. 2019 as its starting point. It claims that framework fails to explain overfitting in adversarial training (Note: that paper wasn't claiming to explain overfitting). The paper then proposes a narrower definition of feature robustness (where an attacker and...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper cites the feature robustness framework of Ilyas et al. 2019 as its starting point. It claims that framework fails to explain overfitting in adversarial training (Note: that paper wasn't claiming to explain overfitting). The paper then proposes a narrower definition of feature robustness (where an atta...
This paper considers the problem of tracking a drifting optimization target under the assumption that the loss function is a smooth function of both the optimization parameter and time. The key insight is to compute the derivative of the minimizer with respect to time by differentiating the loss function with respect t...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper considers the problem of tracking a drifting optimization target under the assumption that the loss function is a smooth function of both the optimization parameter and time. The key insight is to compute the derivative of the minimizer with respect to time by differentiating the loss function with r...
The paper proposes an Early Stopping (ES) method for Deep Image Prior (DIP) based on the running variance of the intermediate outputs. For the task of denoising, a broad set of experiments is provided to study it under different noise types and intensities and against other approaches for early stopping. Besides, some ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes an Early Stopping (ES) method for Deep Image Prior (DIP) based on the running variance of the intermediate outputs. For the task of denoising, a broad set of experiments is provided to study it under different noise types and intensities and against other approaches for early stopping. Beside...
- The paper formulates instance segmentation as a RL problem based on a stateless actor-critic setup, achieving end-to-end learning while exploiting prior knowledge on instance appearance instead of using ground truth instance supervision. - The paper introduces a strategy for spatial decomposition of rewards based on ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: - The paper formulates instance segmentation as a RL problem based on a stateless actor-critic setup, achieving end-to-end learning while exploiting prior knowledge on instance appearance instead of using ground truth instance supervision. - The paper introduces a strategy for spatial decomposition of rewards b...
This paper proposes a new basic data generation framework where each data point is associated with a function and the `noises' are assumed to be from the variation of its function rather than its output used in classical ERM. Two corresponding models are proposed for unsupervised learning (generative model) and superv...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new basic data generation framework where each data point is associated with a function and the `noises' are assumed to be from the variation of its function rather than its output used in classical ERM. Two corresponding models are proposed for unsupervised learning (generative model) an...
The paper studies the failure cases in FSL and proposes an efficient algorithm to extract the difficult tasks from large-scale datasets. Based on the proposed algorithm, the paper builds a new test-only few-shot classification benchmark named HARD-META-DATASET++. Strength: 1) The paper is writing clearly and easy to re...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies the failure cases in FSL and proposes an efficient algorithm to extract the difficult tasks from large-scale datasets. Based on the proposed algorithm, the paper builds a new test-only few-shot classification benchmark named HARD-META-DATASET++. Strength: 1) The paper is writing clearly and ea...
This paper studies if Vision Transformers (ViTs) could outperform CNNs in vision-based RL tasks. The authors compared existing self-supervised strategies and proposed a new self-supervised approach designed for the sequential observations in RL. The newly proposed method combines previous ideas from VICReg (Bardes et a...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies if Vision Transformers (ViTs) could outperform CNNs in vision-based RL tasks. The authors compared existing self-supervised strategies and proposed a new self-supervised approach designed for the sequential observations in RL. The newly proposed method combines previous ideas from VICReg (Bar...
This work examines the uncertainty calibration of various image classifiers. A fairly comprehensive list of classification models, calibration techniques, and metrics are examined experimentally. From experimental results, this work draws several observations and recommendations in both model type and training techniqu...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work examines the uncertainty calibration of various image classifiers. A fairly comprehensive list of classification models, calibration techniques, and metrics are examined experimentally. From experimental results, this work draws several observations and recommendations in both model type and training ...
The idea is to define an appropriate notion of W2 projection of P_0 onto given measures P_1,..,P_J . The proposed definition avoids the standard bi-level optimization based one and hence is efficient to compute. Simulations on a problem of estimating causal effect are presented. Strength: 1. Paper is well written with ...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The idea is to define an appropriate notion of W2 projection of P_0 onto given measures P_1,..,P_J . The proposed definition avoids the standard bi-level optimization based one and hence is efficient to compute. Simulations on a problem of estimating causal effect are presented. Strength: 1. Paper is well writt...
This paper proposes a new pre-training task tailored for end-to-end self-driving. The idea is to take an unsupervised monocular depth model (for instance [1]) and distill the component that takes as input 2 images and predicts the pose between them into a single-image model. The resultant model therefore learns a repre...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new pre-training task tailored for end-to-end self-driving. The idea is to take an unsupervised monocular depth model (for instance [1]) and distill the component that takes as input 2 images and predicts the pose between them into a single-image model. The resultant model therefore learns...
The paper proposes SlenderGNN, a new GNN method that is accurate, robust, and interpretable. Authors report experimental results comparing SlenderGNN's performance with a set of alternative methods. They also conduct an ablated study to assess the importance of individual modules of the proposed method in its performan...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposes SlenderGNN, a new GNN method that is accurate, robust, and interpretable. Authors report experimental results comparing SlenderGNN's performance with a set of alternative methods. They also conduct an ablated study to assess the importance of individual modules of the proposed method in its p...
The submission aims to distinguish informative data and uninformative data under a designed probabilistic setting. Under this setting, the true labels of informative v.s. uninformative are missing. To deal with this issue, the authors propose a learning method simultaneously learning the predictor and the selector, wit...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The submission aims to distinguish informative data and uninformative data under a designed probabilistic setting. Under this setting, the true labels of informative v.s. uninformative are missing. To deal with this issue, the authors propose a learning method simultaneously learning the predictor and the selec...
This paper presents Masked Vector Quantization (MVQ), which is an extension of hierarchical Vector Quantized Variational AutoEncoder (VQ-VAE [a]). MVQ introduces mask configuration on the secondary code vector and demonstrates its effectiveness on image generation with shorter sequences (for example, less than 20) per ...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This paper presents Masked Vector Quantization (MVQ), which is an extension of hierarchical Vector Quantized Variational AutoEncoder (VQ-VAE [a]). MVQ introduces mask configuration on the secondary code vector and demonstrates its effectiveness on image generation with shorter sequences (for example, less than ...
This paper studies in-depth the implicit denoising effect in graph neural networks, which remains an open issue from the theoretical perspective. Rigorous theoretical analysis is provided to uncover the underlying philosophy of GNNs for graph signal denoising. The theoretical results and discussion suggest that the imp...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies in-depth the implicit denoising effect in graph neural networks, which remains an open issue from the theoretical perspective. Rigorous theoretical analysis is provided to uncover the underlying philosophy of GNNs for graph signal denoising. The theoretical results and discussion suggest that...
This paper considers the problem of Elicitation inference optimization where the goal is to get the perspectives of N participants on K different perspectives by minimizing the total number of elicitations. Elicitation could be direct: does participant i agree with perspective j or comparative, does participant i agree...	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 considers the problem of Elicitation inference optimization where the goal is to get the perspectives of N participants on K different perspectives by minimizing the total number of elicitations. Elicitation could be direct: does participant i agree with perspective j or comparative, does participant...
The authors propose 3DiM, a diffusion model for novel view synthesis. The input to the model is a set of images with poses representing different views of the same object plus a target pose; the output of the model is a new image of the same object from the target pose. By applying the model repeatedly, the authors can...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The authors propose 3DiM, a diffusion model for novel view synthesis. The input to the model is a set of images with poses representing different views of the same object plus a target pose; the output of the model is a new image of the same object from the target pose. By applying the model repeatedly, the aut...
This paper proposes a variant of the Mixup method that considers the label mismatch problem by decoupling (i.e., removing) the competitor's class in the softmax which can potentially cause mismatching between the mixed label and sample. The authors formulate the decoupled mixup cross-entropy loss ($L_{DM(CE)}$) that be...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a variant of the Mixup method that considers the label mismatch problem by decoupling (i.e., removing) the competitor's class in the softmax which can potentially cause mismatching between the mixed label and sample. The authors formulate the decoupled mixup cross-entropy loss ($L_{DM(CE)}$)...
The paper focuses on inverse protein folding, a fundamental and challenging problem in protein science. The authors propose a new regularization term by including the trained AlphaFold model and its predicted structural confidence metric. To improve the efficiency of the model optimization loop, they distill the AlphaF...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper focuses on inverse protein folding, a fundamental and challenging problem in protein science. The authors propose a new regularization term by including the trained AlphaFold model and its predicted structural confidence metric. To improve the efficiency of the model optimization loop, they distill th...
This paper presents a novel object-centric representation called Block-Slot Representation, which, unlike the conventional slot representation, provides concept-level disentanglement within a slot, such as color, texture, and position. In comparison with the previous methods, the approach demonstrated significantly bet...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper presents a novel object-centric representation called Block-Slot Representation, which, unlike the conventional slot representation, provides concept-level disentanglement within a slot, such as color, texture, and position. In comparison with the previous methods, the approach demonstrated significa...
A new vector quantization method is proposed, named Vector Quantized Wasserstein Auto-Encoder (VQ-WAE). The presented VQ-WAE employs the Wasserstein (WS) distance in both the observation x space and the latent z space to encourage matching, mimicking the existing Wasserstein Auto-Encoder. Experiments on MNIST, CIFAR10,...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: A new vector quantization method is proposed, named Vector Quantized Wasserstein Auto-Encoder (VQ-WAE). The presented VQ-WAE employs the Wasserstein (WS) distance in both the observation x space and the latent z space to encourage matching, mimicking the existing Wasserstein Auto-Encoder. Experiments on MNIST, ...
This paper proposed a specialized optimizer for optimizing the DNN to have good compression ability, i.e. the optimized model can be pruned/quantized to smaller models which have good accuracy. The paper is based on a formulation related to the sharpness aware minimization, by adding a compression related constraint p...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a specialized optimizer for optimizing the DNN to have good compression ability, i.e. the optimized model can be pruned/quantized to smaller models which have good accuracy. The paper is based on a formulation related to the sharpness aware minimization, by adding a compression related cons...
Several variants of GFlowNets are proposed in the context of multi-agent systems with centralized training and decentralized execution. The challenge considered in this paper is that we want to resulting policy to factorize across the agents, so that each agent can independently sample its next action given only a loca...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: Several variants of GFlowNets are proposed in the context of multi-agent systems with centralized training and decentralized execution. The challenge considered in this paper is that we want to resulting policy to factorize across the agents, so that each agent can independently sample its next action given onl...
The paper proposes approaches to identify the optimal frame rate for audio classification tasks. Specifically, they propose a series of strategies to identify the optimal rate by inserting the frame-rate identification module between the traditional feature extraction step and the final classifier and casting the probl...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes approaches to identify the optimal frame rate for audio classification tasks. Specifically, they propose a series of strategies to identify the optimal rate by inserting the frame-rate identification module between the traditional feature extraction step and the final classifier and casting t...
The paper introduces a new gradient inversion method in Federated Learning. The proposed approach called "Learning To Invert" (LTI) directly attempts to learn the mapping between the gradient of a sample and the corresponding input sample using an auxiliary dataset. A simple multi-layer perceptron (MLP) is used to lea...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper introduces a new gradient inversion method in Federated Learning. The proposed approach called "Learning To Invert" (LTI) directly attempts to learn the mapping between the gradient of a sample and the corresponding input sample using an auxiliary dataset. A simple multi-layer perceptron (MLP) is use...
Conditional language models often perform poorly on out-of-domain (OOD) data. This might be a reason why they shouldn't be tasked with generating anything for OOD examples at all: generations might be of poor quality and even more unpredictable than normal outputs. In order to predict which examples a conditional langu...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Conditional language models often perform poorly on out-of-domain (OOD) data. This might be a reason why they shouldn't be tasked with generating anything for OOD examples at all: generations might be of poor quality and even more unpredictable than normal outputs. In order to predict which examples a condition...
The paper proposes an algorithm (pre-cluster and merge) to better disambiguate heterogenous treatment effects in non-targeted clinical trials (where there is a hidden confounder of the patient of whether the patient was sick/healthy/other confounder variable-- with same observable patient properties/covariates otherwis...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes an algorithm (pre-cluster and merge) to better disambiguate heterogenous treatment effects in non-targeted clinical trials (where there is a hidden confounder of the patient of whether the patient was sick/healthy/other confounder variable-- with same observable patient properties/covariates ...
The paper aims at learning the causal representation from time-series data with the access of interventions. The formulation is a state-space model, which the causal process is unobserved and the observation is a funtion of the unobserved causal process. It generalizes CITRIS (Lippe et al., 2022b) to include instantane...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper aims at learning the causal representation from time-series data with the access of interventions. The formulation is a state-space model, which the causal process is unobserved and the observation is a funtion of the unobserved causal process. It generalizes CITRIS (Lippe et al., 2022b) to include in...
The authors propose a risk-aware bayesian reinforcement learning method to tackle the problem of safe exploration. Specifically, the authors assume that the agent maintains a Dirichlet-Categorical model of the MDP, and propose a method to derive an approximate bound on the confidence that the risk is below a certain le...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose a risk-aware bayesian reinforcement learning method to tackle the problem of safe exploration. Specifically, the authors assume that the agent maintains a Dirichlet-Categorical model of the MDP, and propose a method to derive an approximate bound on the confidence that the risk is below a ce...
The paper makes the argument that current OOD detectors are faulty since they are tied to the bias in the data collection process. The authors show examples where several existing OOD detectors would have classified an auxiliary test data set as OOD, but the classifier performs well on those examples. The authors argu...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper makes the argument that current OOD detectors are faulty since they are tied to the bias in the data collection process. The authors show examples where several existing OOD detectors would have classified an auxiliary test data set as OOD, but the classifier performs well on those examples. The auth...
- This work extends the study of BEiT, i.e. BERT-style pre-training of image transformers. Specifically, this work proposes to adopt a semantic-rich visual tokenizer distilled from the semantically-rich CLIP model, promoting the MIM process to focus more on semantic-level. - In order to learn better and compact code...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: - This work extends the study of BEiT, i.e. BERT-style pre-training of image transformers. Specifically, this work proposes to adopt a semantic-rich visual tokenizer distilled from the semantically-rich CLIP model, promoting the MIM process to focus more on semantic-level. - In order to learn better and comp...
This paper investigate the structrual similarity between winning tickets. In particular, they argue that the signs in weight connection play a critical role. Strength: * The topic of whether lottery tickets (LTs) is unique or share some similarity is of great interests to the communicty. Weakness: * The paper does n...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper investigate the structrual similarity between winning tickets. In particular, they argue that the signs in weight connection play a critical role. Strength: * The topic of whether lottery tickets (LTs) is unique or share some similarity is of great interests to the communicty. Weakness: * The pape...
The paper considers video-based object-centric learning tasks. The proposed model achieves object-centric learning by predicting per-object motion and future frame observation. Experiments show that the proposed method outperforms a collection of single-frame baselines. - Strength - This work exploits the ...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper considers video-based object-centric learning tasks. The proposed model achieves object-centric learning by predicting per-object motion and future frame observation. Experiments show that the proposed method outperforms a collection of single-frame baselines. - Strength - This work explo...
In this paper, the authors propose theoretical analyses for the generalization of GNNs. For the in-distribution case, the authors propose an improved bound regarding graph classification compared to the existing PAC-Bayes results. For the out-of-distribution generalization, the authors propose an analysis for node clas...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this paper, the authors propose theoretical analyses for the generalization of GNNs. For the in-distribution case, the authors propose an improved bound regarding graph classification compared to the existing PAC-Bayes results. For the out-of-distribution generalization, the authors propose an analysis for n...
This paper presents training strategy that allows to training optimal transformers under user-defined budgets. The budget training strategy is based on investigating the redundancy in attention heads, hidden dimensions in MLP, and visual tokens. The training strategy could adjust the activation rate of the model along...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents training strategy that allows to training optimal transformers under user-defined budgets. The budget training strategy is based on investigating the redundancy in attention heads, hidden dimensions in MLP, and visual tokens. The training strategy could adjust the activation rate of the mod...
Overall Summary This paper presents an approach aims to selecting a subset of training data that (1) makes training more computationally efficient while (2) incurring little-to-no loss in accuracy. The proposed approach is analyzed in terms of its generalization performance in addition to computational requirements...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Overall Summary This paper presents an approach aims to selecting a subset of training data that (1) makes training more computationally efficient while (2) incurring little-to-no loss in accuracy. The proposed approach is analyzed in terms of its generalization performance in addition to computational requ...