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 authors introduce new datasets of dynamic stability of synthetic power grids. They show that large Graph Neural Networks GNNs outperform GNNs from previous work at predicting single-node basin stability.They demonstrate that GNNs can be used to identify trouble maker-nodes in the power grids and show that GNNs tra...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors introduce new datasets of dynamic stability of synthetic power grids. They show that large Graph Neural Networks GNNs outperform GNNs from previous work at predicting single-node basin stability.They demonstrate that GNNs can be used to identify trouble maker-nodes in the power grids and show that ...
This paper introduces a theoretical framework for GANs, called MonoFlow, that challenges their usual understanding as an adversarial minimization of a distance/divergence. It interprets GAN training as the generator following a reparameterized gradient flow of the reverse KL, defined through the log density ratio estim...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This paper introduces a theoretical framework for GANs, called MonoFlow, that challenges their usual understanding as an adversarial minimization of a distance/divergence. It interprets GAN training as the generator following a reparameterized gradient flow of the reverse KL, defined through the log density rat...
The paper focuses on Unsupervised continual learning (UCL) and proposes Adaptive Update Direction Rectification (AUDR), an adaptive learning paradigm for the UCL setting. Mainly, the paper uses an Actor-critical approach, where the actor selects the best action and is updated with the predicted reward by the Critic. It...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper focuses on Unsupervised continual learning (UCL) and proposes Adaptive Update Direction Rectification (AUDR), an adaptive learning paradigm for the UCL setting. Mainly, the paper uses an Actor-critical approach, where the actor selects the best action and is updated with the predicted reward by the Cr...
The paper looks at how one could improve the performance of RL agents by leveraging planning at the level of subgoals. The authors argue that by restricting planning at a higher level of abstraction the method side-steps many of the problems associated with traditional model-based planning approaches, like inaccuracy o...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper looks at how one could improve the performance of RL agents by leveraging planning at the level of subgoals. The authors argue that by restricting planning at a higher level of abstraction the method side-steps many of the problems associated with traditional model-based planning approaches, like inac...
Traditional MAML is well used in few-shot learning domains. However, recent works show that MAML does not perform well in a fast adaptation. This paper gives a detail about feature reuse phenomenon and found the relationship between lower layer and the distribution differences in training and testing. Pros In this wor...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Traditional MAML is well used in few-shot learning domains. However, recent works show that MAML does not perform well in a fast adaptation. This paper gives a detail about feature reuse phenomenon and found the relationship between lower layer and the distribution differences in training and testing. Pros In ...
The authors propose a new intrinsic reward function for reinforcement learning, building on earlier definitions of "surprise". The proposed idea is to reward not the surprise itself but its novelty. The authors propose a specific approach to computing this intrinsic reward function and present an empirical analysis in ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose a new intrinsic reward function for reinforcement learning, building on earlier definitions of "surprise". The proposed idea is to reward not the surprise itself but its novelty. The authors propose a specific approach to computing this intrinsic reward function and present an empirical anal...
This paper presents a robust GAN-Inversion (RGI) method for mask-free image inpainting and unsupervised pixel-wise anomaly detection. Furthermore, they further propose a relaxed RGI method for achieving better results. Promising experimental results on image inpainting and anomaly detection by using the proposed method...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper presents a robust GAN-Inversion (RGI) method for mask-free image inpainting and unsupervised pixel-wise anomaly detection. Furthermore, they further propose a relaxed RGI method for achieving better results. Promising experimental results on image inpainting and anomaly detection by using the propose...
This work considers the problem of diagnosing why a policy fails when presented with an out-of-distribution task. Specifically, this work creates a framework to determine if the policy fails for a "how" reason (e.g., the policy does not know how to accomplish the provided goal) or a "what" reason (e.g., the policy does...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This work considers the problem of diagnosing why a policy fails when presented with an out-of-distribution task. Specifically, this work creates a framework to determine if the policy fails for a "how" reason (e.g., the policy does not know how to accomplish the provided goal) or a "what" reason (e.g., the pol...
This paper deals with oversquashing in GNNs, that is, when information flows between nodes (receptive field large enough), but is too compressed (squashed) into the (finite) features to be correctly exploited for prediction (because number of neighbors grows exponentially with radius of receptive field). The paper has...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper deals with oversquashing in GNNs, that is, when information flows between nodes (receptive field large enough), but is too compressed (squashed) into the (finite) features to be correctly exploited for prediction (because number of neighbors grows exponentially with radius of receptive field). The p...
This paper proposes a novel problem by combining the settings of PLL and UDA. To tackle this challenging setting, they also propose a novel approach named PAPLUDA which consists of three loss components: classification loss with soft label disambiguation; inter-domain class-prototype alignment loss, and teacher-student...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a novel problem by combining the settings of PLL and UDA. To tackle this challenging setting, they also propose a novel approach named PAPLUDA which consists of three loss components: classification loss with soft label disambiguation; inter-domain class-prototype alignment loss, and teacher...
This paper describes non-contrastive learning based self-supervised pre-training approach for the speech recognition task. As opposed to the popular methods such as Wav2vec2, Hubert, etc. which use contrastive learning for pre-training, this paper describes how non-contrastive methods such Barlow-Twins for the audio pr...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper describes non-contrastive learning based self-supervised pre-training approach for the speech recognition task. As opposed to the popular methods such as Wav2vec2, Hubert, etc. which use contrastive learning for pre-training, this paper describes how non-contrastive methods such Barlow-Twins for the ...
The paper proposes a way of deriving an encoder from a pretrained diffusion model. In particular, this encoder is stochastic and invertible, which makes it different from existing attempts. The authors show that this encoder can be deployed in an image to image translation setting, and demonstrates interesting results....	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a way of deriving an encoder from a pretrained diffusion model. In particular, this encoder is stochastic and invertible, which makes it different from existing attempts. The authors show that this encoder can be deployed in an image to image translation setting, and demonstrates interesting ...
The paper considers the impact of (pre-trained) embeddings on the performance of different GNN architectures. The datasets under considered are either images and texts with some form of "graph structures", e.g., based on “Co-viewed”, “Co-bought” and “Similar Items” (for the Amazon datasets). The authors fund that only ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper considers the impact of (pre-trained) embeddings on the performance of different GNN architectures. The datasets under considered are either images and texts with some form of "graph structures", e.g., based on “Co-viewed”, “Co-bought” and “Similar Items” (for the Amazon datasets). The authors fund th...
This paper tackles the challenge of learning an efficient representation for multi task reinforcement learning. The authors present a learning algorithm for multi-task RL with function approximation, low-rank multi task bilinear class. This algorithm incentivizes the the RL algorithm to improve sample efficiency by lev...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper tackles the challenge of learning an efficient representation for multi task reinforcement learning. The authors present a learning algorithm for multi-task RL with function approximation, low-rank multi task bilinear class. This algorithm incentivizes the the RL algorithm to improve sample efficienc...
This paper studies the impact of adaptive optimization on local deep network geometry. The main discovery is that, adaptive methods such as Adam bias the trajectories towards regions where the diagonal elements of Hessian are more uniform (compare to SGD type algorithms), and moreover such uniform diagonal Hessian tend...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies the impact of adaptive optimization on local deep network geometry. The main discovery is that, adaptive methods such as Adam bias the trajectories towards regions where the diagonal elements of Hessian are more uniform (compare to SGD type algorithms), and moreover such uniform diagonal Hess...
This paper proposes an improved pseudo-labeling method for semi-supervised learning methods. The paper first proposes a method for filtering out some unlabeled data that has high uncertainty based on Welch’s T-test and Total variance. The paper then extends the methods of SLA [1] to assign labels for unlabeled data....	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposes an improved pseudo-labeling method for semi-supervised learning methods. The paper first proposes a method for filtering out some unlabeled data that has high uncertainty based on Welch’s T-test and Total variance. The paper then extends the methods of SLA [1] to assign labels for unlabel...
The authors study the problem of federated recommendation with implicit feedback, and propose a new framework with personalized score function and personalized item embedding (without user embedding), which is illustrated in Figure 1(c). Strength: 1 The idea of using personalized item embedding in federated recommendat...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The authors study the problem of federated recommendation with implicit feedback, and propose a new framework with personalized score function and personalized item embedding (without user embedding), which is illustrated in Figure 1(c). Strength: 1 The idea of using personalized item embedding in federated rec...
The author adopts knowledge transfer to solve the inhabited stability and generalization issue of differential NAS. The method first trains the weights of the main model with alpha fixed, then transfer the knowledge from main to auxiliary by quadruple relative similarities, and finally optimizes the alpha of the main m...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The author adopts knowledge transfer to solve the inhabited stability and generalization issue of differential NAS. The method first trains the weights of the main model with alpha fixed, then transfer the knowledge from main to auxiliary by quadruple relative similarities, and finally optimizes the alpha of th...
This paper focused on the research problem of link prediction with GNNs. The authors comprehensively analyzed why the subgraph-based GNNs performs good on link prediction, as well as their efficiency problems. The authors then proposed to estimate the key structural information with subgraph sketching. Doing so combine...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper focused on the research problem of link prediction with GNNs. The authors comprehensively analyzed why the subgraph-based GNNs performs good on link prediction, as well as their efficiency problems. The authors then proposed to estimate the key structural information with subgraph sketching. Doing so...
Summary: This paper proposed to care about article generation with image information where the image info is used in an entity-aware way instead of directly using the visual features. That is, the embedded image in the associated article is firstly transformed into the textual space in the form of textual captions and...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Summary: This paper proposed to care about article generation with image information where the image info is used in an entity-aware way instead of directly using the visual features. That is, the embedded image in the associated article is firstly transformed into the textual space in the form of textual capt...
The manuscript proposes to tackle the continual domain shift learning issue, in which the model is trained on a source domain and serveral unlabeled target domains. The authors highlight the pain points of the issue are concluded with three: 1. model generalization on 'before and during' training doamins(TDG); 2. bette...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The manuscript proposes to tackle the continual domain shift learning issue, in which the model is trained on a source domain and serveral unlabeled target domains. The authors highlight the pain points of the issue are concluded with three: 1. model generalization on 'before and during' training doamins(TDG); ...
This paper proposes a technique for removing data points for standard supervised learning problems. Their best performing method partially drops data from a class with a sufficient F1 score, and shows that they can train on ~6M total less data points while only reducing ImageNet accuracy by 1 percentage point. Strengt...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes a technique for removing data points for standard supervised learning problems. Their best performing method partially drops data from a class with a sufficient F1 score, and shows that they can train on ~6M total less data points while only reducing ImageNet accuracy by 1 percentage point. ...
The paper extends hierarchical transformers (neural network models for synthesising circuits from linear time specifications) to build models for the repair of circuits so that they satisfy a certain linear time property. The models, which are also used in an iterative procedure to do reactive synthesis, are trained ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper extends hierarchical transformers (neural network models for synthesising circuits from linear time specifications) to build models for the repair of circuits so that they satisfy a certain linear time property. The models, which are also used in an iterative procedure to do reactive synthesis, are ...
The paper proposes the $\sigma$-reparameterised transformer. The authors note a phenomenon in during training of transformers where the 'attention entropy' (essentially, how sharply peaked the attention scores are) decreases and then increases again. In the cases where the transformer fails to learn, this attention ent...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper proposes the $\sigma$-reparameterised transformer. The authors note a phenomenon in during training of transformers where the 'attention entropy' (essentially, how sharply peaked the attention scores are) decreases and then increases again. In the cases where the transformer fails to learn, this atten...
This paper studied the differential privacy preservation for the generative model and proposed incorporating the DP-SGD strategy into Diffusion Model training. Strength: 1. This paper is well-organized and easy to follow. 2. The experimental results on MNIST, Fashion-MNIST, and CelebA datasets validate the effectivene...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper studied the differential privacy preservation for the generative model and proposed incorporating the DP-SGD strategy into Diffusion Model training. Strength: 1. This paper is well-organized and easy to follow. 2. The experimental results on MNIST, Fashion-MNIST, and CelebA datasets validate the eff...
The polysemy of data introduces a big challenge for the labeling process. This paper proposes an implicit distribution representation method based on label distribution learning, which also considers the uncertainty of label values, in which the Gaussian prior and self-attention-based methods are also adopted for learn...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The polysemy of data introduces a big challenge for the labeling process. This paper proposes an implicit distribution representation method based on label distribution learning, which also considers the uncertainty of label values, in which the Gaussian prior and self-attention-based methods are also adopted f...
In this paper the authors propose the Composite Slice Transformer (CST) which consists of a composition of attentions applied to a stacked, slice representation of the input sequence at different scales, coupled with a multi-scale volatile instant positional embedding. To do this, first the input sequence of length N ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper the authors propose the Composite Slice Transformer (CST) which consists of a composition of attentions applied to a stacked, slice representation of the input sequence at different scales, coupled with a multi-scale volatile instant positional embedding. To do this, first the input sequence of l...
This paper tackles the problem of learning representation robust to biased data, i.e. data containing easy-to-learn but unrelated features. It first derives \epsilon-SupInfoNCE, which adds control of the minimal distance between positive and negative samples to the traditional contrastive losses. Then it proposes FairK...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper tackles the problem of learning representation robust to biased data, i.e. data containing easy-to-learn but unrelated features. It first derives \epsilon-SupInfoNCE, which adds control of the minimal distance between positive and negative samples to the traditional contrastive losses. Then it propos...
This paper develops a new differentially private (DP) framework to tackle the problem of decentralised consensus learning. This framework is based on the stochastic Decentralised Krasnosel'skii-Mann (D-KM) iteration (its privatised version referred to as "DP-KM" in this paper). It provides both theoretical and empirica...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper develops a new differentially private (DP) framework to tackle the problem of decentralised consensus learning. This framework is based on the stochastic Decentralised Krasnosel'skii-Mann (D-KM) iteration (its privatised version referred to as "DP-KM" in this paper). It provides both theoretical and ...
In this paper authors propose a mechanism to prune a convolutional neural network model in a relatively data-free manner i.e., they do not utilize training data or loss function for retraining the pruned model. However unlike the actual data-free pruning techniques they assume the availability of moments of class-condi...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper authors propose a mechanism to prune a convolutional neural network model in a relatively data-free manner i.e., they do not utilize training data or loss function for retraining the pruned model. However unlike the actual data-free pruning techniques they assume the availability of moments of cla...
This paper attempts to use diffusion models for action generation in imitation learning set ups. The paper contributes : 1. modifications to diffusion models to adapt them to action generation 2. discussion on effectiveness of CFG 3. sampling techniques for action generation The paper tests on a few datasets including ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper attempts to use diffusion models for action generation in imitation learning set ups. The paper contributes : 1. modifications to diffusion models to adapt them to action generation 2. discussion on effectiveness of CFG 3. sampling techniques for action generation The paper tests on a few datasets in...
The work proposes a theoretical approach for choosing hyper-parameters in unsupervised domain adaptation. The main strategy is to compute an aggregation of models with target error bound, which theoretically relies on the extension of importance weighted least squares to linear aggregation of vector-valued functions. T...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The work proposes a theoretical approach for choosing hyper-parameters in unsupervised domain adaptation. The main strategy is to compute an aggregation of models with target error bound, which theoretically relies on the extension of importance weighted least squares to linear aggregation of vector-valued func...
The paper proposes a novel differentiable dynamic time-warping algorithm. The approach is advantageous over existing variants as it outputs a learnable warping path between the time series representations. The authors test their approach on two distinct tasks: audio-score alignment and visual place recognition, and ac...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a novel differentiable dynamic time-warping algorithm. The approach is advantageous over existing variants as it outputs a learnable warping path between the time series representations. The authors test their approach on two distinct tasks: audio-score alignment and visual place recognition...
This submission contributes a masking based reconstruction approach with a transformer architecture for self-supervised learning based on quadratic reconstruction error with adverserial perturbation of the input. The method is benchmarked on a set of classic datasets where it is shown to predict better than standard m...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This submission contributes a masking based reconstruction approach with a transformer architecture for self-supervised learning based on quadratic reconstruction error with adverserial perturbation of the input. The method is benchmarked on a set of classic datasets where it is shown to predict better than st...
This paper studies quantum kernel learning. Their observation seems that the dimension of the feature maps associated with quantum kernels is exponential but can be processed by quantum boxes. The problem with these kernel is that their leading eigenvalues are too small so that these models do not generalize well. The ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper studies quantum kernel learning. Their observation seems that the dimension of the feature maps associated with quantum kernels is exponential but can be processed by quantum boxes. The problem with these kernel is that their leading eigenvalues are too small so that these models do not generalize we...
This paper presents a model for selecting test-time data augmentation to boost a classification model. The proposed method is based on RNN to gradually output multiple augmentations, selected from a predefined augmentation set. The idea is based on loss prediction for finding suitable augmentations. In each step of RNN...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a model for selecting test-time data augmentation to boost a classification model. The proposed method is based on RNN to gradually output multiple augmentations, selected from a predefined augmentation set. The idea is based on loss prediction for finding suitable augmentations. In each ste...
The authors consider the general problem of multi-solution optimization (MSO), where one is interested in obtaining a sampling (perhaps all) of the minimizers of each of a family of (nonconvex) functions $\{ f_{\tau} : \tau \in \mathcal{T} \}$ defined on a low-dimensional domain $\mathcal{X}$ (here, $\mathcal{T}$ repre...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: The authors consider the general problem of multi-solution optimization (MSO), where one is interested in obtaining a sampling (perhaps all) of the minimizers of each of a family of (nonconvex) functions $\{ f_{\tau} : \tau \in \mathcal{T} \}$ defined on a low-dimensional domain $\mathcal{X}$ (here, $\mathcal{T...
This paper investigates the question: how can we train agents to behave 'morally' in text-based games, while achieving a high task reward? The paper proposes a two stage learning method, which is iterated over time. In my understanding, this involves training both a task-specific Q function ('task policy'), which only...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper investigates the question: how can we train agents to behave 'morally' in text-based games, while achieving a high task reward? The paper proposes a two stage learning method, which is iterated over time. In my understanding, this involves training both a task-specific Q function ('task policy'), wh...
In this paper, the authors consider the problem of computing subgradients of non-smooth non-convex functions that are given as programs that are based on the small set of elementary functions. Their main contributions are the following: 1. They analyze the computational complexity of the algorithm suggested by Bolt...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: In this paper, the authors consider the problem of computing subgradients of non-smooth non-convex functions that are given as programs that are based on the small set of elementary functions. Their main contributions are the following: 1. They analyze the computational complexity of the algorithm suggested...
The paper introduces an approach for capturing long-term dependencies of complex time series. The paper incorporates some ideas from the classical dynamical system modeling literature into machine learning. Strengths: i) The proposed method is very comprehensive. It is tailored in fine details including computational f...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper introduces an approach for capturing long-term dependencies of complex time series. The paper incorporates some ideas from the classical dynamical system modeling literature into machine learning. Strengths: i) The proposed method is very comprehensive. It is tailored in fine details including computa...