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 considers community detection with unbalanced communities. Specifically, they consider the hypothesis testing problem on the degree-corrected block model, which is to distinguish whether a graph is generated from a null model with one community or an alternative model with unbalanced communities. They show...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers community detection with unbalanced communities. Specifically, they consider the hypothesis testing problem on the degree-corrected block model, which is to distinguish whether a graph is generated from a null model with one community or an alternative model with unbalanced communities. T...
This paper tackles the problem of training 3D generative models for portrait video generation with only 2D videos. In order to enforce temporal consistency and diverse motion, PV3D utilizes disentangled appearance and motion latent space. Extensive experiments demonstrate the effectiveness of the proposed approach. ## ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper tackles the problem of training 3D generative models for portrait video generation with only 2D videos. In order to enforce temporal consistency and diverse motion, PV3D utilizes disentangled appearance and motion latent space. Extensive experiments demonstrate the effectiveness of the proposed appro...
The paper at hand proposes a new inductive bias (in the form of a proxy reward) for MI-based unsupervised skill discovery. The reward is designed so that exploration is improved, in particular in environments with bottleneck states or non-trivial dynamics. This is turn motivated by the observation that MI-based methods...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper at hand proposes a new inductive bias (in the form of a proxy reward) for MI-based unsupervised skill discovery. The reward is designed so that exploration is improved, in particular in environments with bottleneck states or non-trivial dynamics. This is turn motivated by the observation that MI-based...
This paper studies the trainability of quantum neural networks (QNNs) with symmetric ansatzes and how to exploit the symmetry to improve the trainability of QNNs. More precisely, it provides a theoretical understanding of the better trainability of symmetric ansatz over asymmetric ansatz by proposing the effective quan...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper studies the trainability of quantum neural networks (QNNs) with symmetric ansatzes and how to exploit the symmetry to improve the trainability of QNNs. More precisely, it provides a theoretical understanding of the better trainability of symmetric ansatz over asymmetric ansatz by proposing the effect...
This paper empirically demonstrates that SGD learns single-hidden-layer neural networks with near-optimal sample complexity efficiently which matches the theoretical bound of an optimal learner in JV22, while according to theoretically lower bounds achieving the optimal sample complexity is intractable in the worst cas...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper empirically demonstrates that SGD learns single-hidden-layer neural networks with near-optimal sample complexity efficiently which matches the theoretical bound of an optimal learner in JV22, while according to theoretically lower bounds achieving the optimal sample complexity is intractable in the w...
The paper "WAVEFORMER: LINEAR-TIME ATTENTION WITH FORWARD AND BACKWARD WAVELET TRANSFORM" proposes integrating analysis- and synthesis-wavelet transforms into transformers. The main idea is to move attention computations into the wavelet domain, by framing the attention computation with a forward and backward wavelet ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper "WAVEFORMER: LINEAR-TIME ATTENTION WITH FORWARD AND BACKWARD WAVELET TRANSFORM" proposes integrating analysis- and synthesis-wavelet transforms into transformers. The main idea is to move attention computations into the wavelet domain, by framing the attention computation with a forward and backward ...
This paper presents an MARL algorithm to adpatively handle the entropy regularization in multi-agent RL. In this method, the level of exploration of each agent is controlled by its time-varying target entropy, which severs as a constraint in the optimization problem. To determine a proper target entropy for each agent,...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents an MARL algorithm to adpatively handle the entropy regularization in multi-agent RL. In this method, the level of exploration of each agent is controlled by its time-varying target entropy, which severs as a constraint in the optimization problem. To determine a proper target entropy for eac...
The paper presents repeated distributionally robust optimization (RDRO), a theoretical framework that extends the performative prediction with distributionally robust optimization. The authors provide a convergence analysis of RDRO and empirically demonstrate its implications for fair ML. Strengths: 1. The paper is wel...	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 presents repeated distributionally robust optimization (RDRO), a theoretical framework that extends the performative prediction with distributionally robust optimization. The authors provide a convergence analysis of RDRO and empirically demonstrate its implications for fair ML. Strengths: 1. The pape...
This paper takes a particular view on safe reinforcement learning in the form of model-predictive control. In particular, the authors present an approach, called ‘guided safe shooting (GuSS), that combines model-based planning with RL, towards a minimal violation of safety. Technically, a constrained MDP approach is re...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper takes a particular view on safe reinforcement learning in the form of model-predictive control. In particular, the authors present an approach, called ‘guided safe shooting (GuSS), that combines model-based planning with RL, towards a minimal violation of safety. Technically, a constrained MDP approa...
The authors propose an attack on federated learning where a subset f clients are compromised. The problem is formulated as a game where both players use a no-regret learning algorothm to converge to the NE. The game is simulated, as the defender does not know the true updates of compromised clients. I am surprised that...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The authors propose an attack on federated learning where a subset f clients are compromised. The problem is formulated as a game where both players use a no-regret learning algorothm to converge to the NE. The game is simulated, as the defender does not know the true updates of compromised clients. I am surpri...
This paper presents an analysis of two self-supervised training methods for lightweight vision transformers. The methods are MAE (a masking method) and MoCoV3 (a contrastive method). It shows that although MAE appears stronger when training (w/o labels) and evaluating on ImageNet, the roles reverse on downstream tasks....	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper presents an analysis of two self-supervised training methods for lightweight vision transformers. The methods are MAE (a masking method) and MoCoV3 (a contrastive method). It shows that although MAE appears stronger when training (w/o labels) and evaluating on ImageNet, the roles reverse on downstrea...
The authors proposed a new method (Rotamer Density Estimator): a flow-based generative model to estimate the probability distribution of conformation. The proposed Rotamer Density Estimator was evaluated on mutational effects for ddG (by using entropy of the probability distribution as the measure of flexibility). Str...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors proposed a new method (Rotamer Density Estimator): a flow-based generative model to estimate the probability distribution of conformation. The proposed Rotamer Density Estimator was evaluated on mutational effects for ddG (by using entropy of the probability distribution as the measure of flexibilit...
The paper studies the problem of privately releasing prefix sums in the adaptive continual release model via matrix factorization techniques. Their main results are 1.extension of prior work which was limited to single participation to general participation patterns and 2. a more efficient method based on fast Fourier ...	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 studies the problem of privately releasing prefix sums in the adaptive continual release model via matrix factorization techniques. Their main results are 1.extension of prior work which was limited to single participation to general participation patterns and 2. a more efficient method based on fast ...
This paper proposes a single-loop adaptive gradient-based algorithm for solving nonconvex-strongly-concave problem with provably efficiency. Previous works only consider directly applying adaptive update scheme to x and y separately thus break the time-scale separation rule which is vital for guaranteeing the convergen...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper proposes a single-loop adaptive gradient-based algorithm for solving nonconvex-strongly-concave problem with provably efficiency. Previous works only consider directly applying adaptive update scheme to x and y separately thus break the time-scale separation rule which is vital for guaranteeing the c...
The paper uses theoretical tools from harmonic analysis and noise stability to explain the effectiveness of data-models introduced by Ilyas et. al (2022). In addition, the same theoretical components are used to analyze group influence and identify conditions when first-order influence (linear in terms of training data...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper uses theoretical tools from harmonic analysis and noise stability to explain the effectiveness of data-models introduced by Ilyas et. al (2022). In addition, the same theoretical components are used to analyze group influence and identify conditions when first-order influence (linear in terms of train...
The paper proposes a new heuristic to estimate the accuracy of a given model for OOD test observations without labels. It uses the nuclear norm (i.e., the normalized sum of singular values) of the prediction matrix (i.e., a matrix of softmax outputs for the test data) of a given model. The measure is motivated by resul...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new heuristic to estimate the accuracy of a given model for OOD test observations without labels. It uses the nuclear norm (i.e., the normalized sum of singular values) of the prediction matrix (i.e., a matrix of softmax outputs for the test data) of a given model. The measure is motivated ...
In this paper, the authors introduce TabPFN. A transformer-based neural network for classification. Unlike traditional supervised approaches, the network is pre-trained to run classification on unseen datasets. This transformer approach, runs inference on input data to produce a classification result. However, this inp...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper, the authors introduce TabPFN. A transformer-based neural network for classification. Unlike traditional supervised approaches, the network is pre-trained to run classification on unseen datasets. This transformer approach, runs inference on input data to produce a classification result. However, ...
The paper proposes a context aware variational auto encoder which modified the structure of previous C-VAE. The evaluation is on synthetic data only. Strength: - the work touches a fundamental problem. Weakness: - Only synthetic experiments are conducted. - The VAE only tested with MLP. - The data generated is ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper proposes a context aware variational auto encoder which modified the structure of previous C-VAE. The evaluation is on synthetic data only. Strength: - the work touches a fundamental problem. Weakness: - Only synthetic experiments are conducted. - The VAE only tested with MLP. - The data gener...
Combining the interests to of two separate fields of inquiry the authors of this work tackle the problem of Anytime inference for Domain Adaptation through the use of a recursive knowledge distillation. Being able to perform inference over multiple computational budgets is a beneficial strategy for deployed machine lea...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Combining the interests to of two separate fields of inquiry the authors of this work tackle the problem of Anytime inference for Domain Adaptation through the use of a recursive knowledge distillation. Being able to perform inference over multiple computational budgets is a beneficial strategy for deployed mac...
The paper studies robust fair clustering and motivates the issue of robustness under attack. The authors conduct numerical experiments to show that state-of-the-art models are highly susceptible to the attack they design. The authors then propose a Consensus Fair Clustering and numerically show its robustness. Strength...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper studies robust fair clustering and motivates the issue of robustness under attack. The authors conduct numerical experiments to show that state-of-the-art models are highly susceptible to the attack they design. The authors then propose a Consensus Fair Clustering and numerically show its robustness. ...
The paper aims to evaluate the fairness of the model without accessing sensitive group attributes. The main idea is to utilize auxiliary models of estimating group attributes by calibrating their outputs. The paper first theoretically explains why we need calibration on the auxiliary models. Then, the paper suggests a ...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper aims to evaluate the fairness of the model without accessing sensitive group attributes. The main idea is to utilize auxiliary models of estimating group attributes by calibrating their outputs. The paper first theoretically explains why we need calibration on the auxiliary models. Then, the paper sug...
The paper proposes a new model-based RL framework for continuous control in deterministic environments. The authors propose Policy Optimization with Model Planning (POMP) an algorithm that uses Differential Dynamic Programming (DDP) as the planner module. A specific instantiation of DDP is introduced which uses (D3P) w...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a new model-based RL framework for continuous control in deterministic environments. The authors propose Policy Optimization with Model Planning (POMP) an algorithm that uses Differential Dynamic Programming (DDP) as the planner module. A specific instantiation of DDP is introduced which uses...
The paper performs an analysis of fairness training methods on multiple medical imaging tasks. They conclude that no method is significantly better than basic ERM training. The paper is well written and articulates existing methods of fairness very well. Details about the experiments are not clear. Such as how the a...	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 performs an analysis of fairness training methods on multiple medical imaging tasks. They conclude that no method is significantly better than basic ERM training. The paper is well written and articulates existing methods of fairness very well. Details about the experiments are not clear. Such as h...
The paper studies the implicit bias of gradient descent on diagonal linear networks. It has been lately shown that parametrizing $x$ as $u \odot v$ in $y \approx A (u \odot v)$ induces sparsity in the gradient flow iterates on the least squares loss. In the paper a special architecture is proposed so that gradient desc...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper studies the implicit bias of gradient descent on diagonal linear networks. It has been lately shown that parametrizing $x$ as $u \odot v$ in $y \approx A (u \odot v)$ induces sparsity in the gradient flow iterates on the least squares loss. In the paper a special architecture is proposed so that gradi...
In this paper, the authors propose the Deep Graph Neural Diffusion method, which ensures the over-smoothing issue and guarantees stability of the model in theoretical. The experimental results also show the improvement of the proposed method compared with the baseline methods. Strength: 1. Convincible theoretical anal...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors propose the Deep Graph Neural Diffusion method, which ensures the over-smoothing issue and guarantees stability of the model in theoretical. The experimental results also show the improvement of the proposed method compared with the baseline methods. Strength: 1. Convincible theoreti...
This paper presents an unsupervised GNN pre-training technique for brain networks. The authors first designed a two-level contrastive learning method and then propose a data-driven atlas mapping technique for mapping different ROI systems. The method was evaluated in three brain network datasets. The results showed the...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents an unsupervised GNN pre-training technique for brain networks. The authors first designed a two-level contrastive learning method and then propose a data-driven atlas mapping technique for mapping different ROI systems. The method was evaluated in three brain network datasets. The results sh...
The paper introduces a novel method to generate triangular meshes for the the input 3D point cloud. The authors introduce multiple anchors to divide the neighborhood space of each input 3D point. A deep neural network is then used to predict the presence and the locations of cirumcenters in the point local neighborhoo...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper introduces a novel method to generate triangular meshes for the the input 3D point cloud. The authors introduce multiple anchors to divide the neighborhood space of each input 3D point. A deep neural network is then used to predict the presence and the locations of cirumcenters in the point local nei...
The authors pose and attempt to answer a new question if the perceptual gradients of the model imply the robustness of the model to adversarial attacks. To answer that, they proposed a new method that trains the model to have input gradients that align with ground-truth input gradients. They introduced several methods ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors pose and attempt to answer a new question if the perceptual gradients of the model imply the robustness of the model to adversarial attacks. To answer that, they proposed a new method that trains the model to have input gradients that align with ground-truth input gradients. They introduced several ...
This work presents a graph structure refinement layer named GLAM based on a neighborhood attention schema slightly different from the Softmax attention, which disentangles the optimization of node embeddings and graph structures. The proposed GLAM layer learns more sparse subgraphs than prior arts by giving the known a...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work presents a graph structure refinement layer named GLAM based on a neighborhood attention schema slightly different from the Softmax attention, which disentangles the optimization of node embeddings and graph structures. The proposed GLAM layer learns more sparse subgraphs than prior arts by giving the...
The paper presents a model of hippocampal / entorhinal representation that handles both physical and conceptual space with a common mathematical formulation, drawing on successor representations from RL, and building on dimensionality reduction accounts of grid cells such as the work of Dordek et al. The authors propos...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The paper presents a model of hippocampal / entorhinal representation that handles both physical and conceptual space with a common mathematical formulation, drawing on successor representations from RL, and building on dimensionality reduction accounts of grid cells such as the work of Dordek et al. The author...
This mostly theoretical work presents a framework of GNNs in the context of Dirichlet energy minimization/maximization. GNNs are framed as estimators for gradients of energy functionals that point to either of those two directions. It shows that for certain classes of graphs and GNNs, attractive or repulsive forces bet...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This mostly theoretical work presents a framework of GNNs in the context of Dirichlet energy minimization/maximization. GNNs are framed as estimators for gradients of energy functionals that point to either of those two directions. It shows that for certain classes of graphs and GNNs, attractive or repulsive fo...
The paper proposed to decompose video data (3D, space 2D + time 1D) into 3 data planes. Each of the data plane is only 2D. This allows the T2D model to create tokens in 3 2D planes separately. ViT 3D self-attention is decomposed along each plane. Compared with a single 3D self-attention, using the three 2D self-atten...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposed to decompose video data (3D, space 2D + time 1D) into 3 data planes. Each of the data plane is only 2D. This allows the T2D model to create tokens in 3 2D planes separately. ViT 3D self-attention is decomposed along each plane. Compared with a single 3D self-attention, using the three 2D se...
This paper adaptively learns a supra-graph, representing non-static correlations between any two variables at any two timestamps, to capture high-resolution spatial-temporal dependencies, and define FGSO that has the capacity of scale-free learning parameters in the Fourier space. Accordingly, authors construct a compl...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper adaptively learns a supra-graph, representing non-static correlations between any two variables at any two timestamps, to capture high-resolution spatial-temporal dependencies, and define FGSO that has the capacity of scale-free learning parameters in the Fourier space. Accordingly, authors construct...
The paper proposes a new automatic metric called SMART for evaluating model-generated text. Here, sentences are used as basic units of matching instead of tokens, in order to support long and multi-sentence texts. For this, the authors use a sentence-level soft matching function, and experiment with both string-based a...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a new automatic metric called SMART for evaluating model-generated text. Here, sentences are used as basic units of matching instead of tokens, in order to support long and multi-sentence texts. For this, the authors use a sentence-level soft matching function, and experiment with both string...
The paper provides an analysis of the denoising effect of GNNs and reveals the impact of graph size, connectivity, and GNN architectures. It also proposes a robust Neumann graph convolution (RNGC) model based on the defined adversarial graph signal denoising problem. # Strength: 1. The paper is clearly written and ea...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper provides an analysis of the denoising effect of GNNs and reveals the impact of graph size, connectivity, and GNN architectures. It also proposes a robust Neumann graph convolution (RNGC) model based on the defined adversarial graph signal denoising problem. # Strength: 1. The paper is clearly writte...
The paper provides an algorithm to improve zero-shot open-vocabulary classifiers by better prompting them. The authors propose to (1) create sub-classes for each parent classification category using existing (human-created) or inferred (GPT-3 generated) label hierarchies, (2) perform standard zero-shot classification o...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper provides an algorithm to improve zero-shot open-vocabulary classifiers by better prompting them. The authors propose to (1) create sub-classes for each parent classification category using existing (human-created) or inferred (GPT-3 generated) label hierarchies, (2) perform standard zero-shot classifi...
This work introduces a new class of Q-learning-based algorithm for both online and offline RL for continuous control tasks based on extreme value theory. By using the insight that the maximum of i.i.d. random variables with exponential tails has a Gumbel distribution, the authors derived a new set of update rules which...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work introduces a new class of Q-learning-based algorithm for both online and offline RL for continuous control tasks based on extreme value theory. By using the insight that the maximum of i.i.d. random variables with exponential tails has a Gumbel distribution, the authors derived a new set of update rul...
This paper formulates the problem of training neural networks with threshold activation functions to a convex optimization problem. As a result, global minima can be obtained with standard convex optimizers. Strengths: 1. The paper is well-written and easy to follow. The notations are clear. 2. Equivalenting neural ne...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper formulates the problem of training neural networks with threshold activation functions to a convex optimization problem. As a result, global minima can be obtained with standard convex optimizers. Strengths: 1. The paper is well-written and easy to follow. The notations are clear. 2. Equivalenting n...
There are two successful directions in DL nowadays to build models based on both labeled and large amount of unlabeled data: pseudo-labeling (semi-supervised) and SSL (self-supervised learning) with further fine-tuning. There are few papers which try to properly combine both SSL and supervised loss right from the begin...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: There are two successful directions in DL nowadays to build models based on both labeled and large amount of unlabeled data: pseudo-labeling (semi-supervised) and SSL (self-supervised learning) with further fine-tuning. There are few papers which try to properly combine both SSL and supervised loss right from t...
This paper studied the convergence of gradient descent algorithm in training the DeepONet in the overparameterized regime. Specifically, the authors proved a spectral norm bound for the Hessian of DeepONets with smooth activation functions; this bound implies convergence of GD using the RSC property. For ReLU activate...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studied the convergence of gradient descent algorithm in training the DeepONet in the overparameterized regime. Specifically, the authors proved a spectral norm bound for the Hessian of DeepONets with smooth activation functions; this bound implies convergence of GD using the RSC property. For ReLU ...