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 an approach for self-supervised learning for agents to collect data, which is then relabelled with task reward and used for downstream tasks with offline RL. The approach estimates reachability of states using a contrastive loss where predicted state features and corresponding ground truth features c...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes an approach for self-supervised learning for agents to collect data, which is then relabelled with task reward and used for downstream tasks with offline RL. The approach estimates reachability of states using a contrastive loss where predicted state features and corresponding ground truth fe...
In this paper, the authors propose a method to dynamically select a subset of features in a neural network. In contrast to standard approaches, this set of features is _global_, and therefore the same subset of features will be ignored for all samples in the dataset. The proposed method consists of three ingredients: 1...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors propose a method to dynamically select a subset of features in a neural network. In contrast to standard approaches, this set of features is _global_, and therefore the same subset of features will be ignored for all samples in the dataset. The proposed method consists of three ingred...
This paper proposes a simple method called ranked policy memory (RPM) that can be plugged in any existing MARL algorithms to solve the generalization problem of MARL. The main idea of RPM is maintaining a look-up memory of the history rollout policies which are ranked by the training episode return. At each episode, th...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a simple method called ranked policy memory (RPM) that can be plugged in any existing MARL algorithms to solve the generalization problem of MARL. The main idea of RPM is maintaining a look-up memory of the history rollout policies which are ranked by the training episode return. At each epi...
The setting of this paper is few-shot class incremental learning (FSCIL). The key idea is using neural collapse to do FSCIL. This paper use a set of pre-defined classifier prototypes as ETF for the whole label space, and then aligns all classes, including base session's classes and novel classes, to the pre-defined cla...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The setting of this paper is few-shot class incremental learning (FSCIL). The key idea is using neural collapse to do FSCIL. This paper use a set of pre-defined classifier prototypes as ETF for the whole label space, and then aligns all classes, including base session's classes and novel classes, to the pre-def...
This paper tries to develop a method called GraphCG for the unsupervised discovery of steerable factors in the latent space of deep graph generative models. The authors observe that the learned representation space of some methods is not perfectly disentangled. Thus, they propose to learn the semantic-rich directions ...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper tries to develop a method called GraphCG for the unsupervised discovery of steerable factors in the latent space of deep graph generative models. The authors observe that the learned representation space of some methods is not perfectly disentangled. Thus, they propose to learn the semantic-rich dir...
The paper proposes a new method (CIDER) for representation learning and applies it for out-of-distribution detection. Given an $L_2$ normalized embedding (hyperspherical embeddings) from a deep network, the method computes prototypes, as moving averages of the embeddings corresponding to each class. The method is optim...	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 proposes a new method (CIDER) for representation learning and applies it for out-of-distribution detection. Given an $L_2$ normalized embedding (hyperspherical embeddings) from a deep network, the method computes prototypes, as moving averages of the embeddings corresponding to each class. The method ...
The paper proposes a method for generating images from text with an addition of a retrieval component. Authors use pretrained text and image encoders and train diffusion-based generative models that use an image embedding and a set of image embeddings closest to the target one to reconstruct the image. During inference...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper proposes a method for generating images from text with an addition of a retrieval component. Authors use pretrained text and image encoders and train diffusion-based generative models that use an image embedding and a set of image embeddings closest to the target one to reconstruct the image. During i...
This paper shows that self-attention mechanisms can be derived by finding the support vector expansion of a support vector regression (SVR) problem. The key contribution of this paper is the framework proposed in Section 2, which links attention mechanisms to support vector regression problems, and provides a different...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper shows that self-attention mechanisms can be derived by finding the support vector expansion of a support vector regression (SVR) problem. The key contribution of this paper is the framework proposed in Section 2, which links attention mechanisms to support vector regression problems, and provides a d...
This paper outlines some desiderata for memory tasks for RL, and proposes/releases a set of three varied benchmark tasks that meet these desiderata. The paper runs a variety of baselines on the tasks and shows that they have nice properties such as strong memory dependence and scalable difficulty. One task, searing spo...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper outlines some desiderata for memory tasks for RL, and proposes/releases a set of three varied benchmark tasks that meet these desiderata. The paper runs a variety of baselines on the tasks and shows that they have nice properties such as strong memory dependence and scalable difficulty. One task, sea...
This paper presents GC-Flow, a generative framework to generate the representation of graphs. The framework is effective for node classification and unravels the inherent structure of data for clustering. More specifically, GC-Flow takes both the advantages of normalizing flows and graph convolutions. For prediction, i...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper presents GC-Flow, a generative framework to generate the representation of graphs. The framework is effective for node classification and unravels the inherent structure of data for clustering. More specifically, GC-Flow takes both the advantages of normalizing flows and graph convolutions. For predi...
The paper proposes a method to handle the species distribution modeling (SDM) problem at a global scale, which is challenging and important. Traditional models need presence-absence data for training and are unable to take advantage of large-scale data. In contrast, with several scalable loss functions, the method in t...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes a method to handle the species distribution modeling (SDM) problem at a global scale, which is challenging and important. Traditional models need presence-absence data for training and are unable to take advantage of large-scale data. In contrast, with several scalable loss functions, the met...
This work introduces model-based reinforcement learning to the domain of spiking recurrent neural networks. To this end, the authors used two subnetworks – an “agent” one for computing policy and a “model” one for predicting future rewards and states of the environment. The authors then formulated local learning rules ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work introduces model-based reinforcement learning to the domain of spiking recurrent neural networks. To this end, the authors used two subnetworks – an “agent” one for computing policy and a “model” one for predicting future rewards and states of the environment. The authors then formulated local learnin...
This paper studies machine learning in settings where where classes have rare subcategories (potentially as small as one example in the training set). The paper creates a data model, and shows that while a model that properly learns features can achieve very high performance, models that do not learn these features ha...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies machine learning in settings where where classes have rare subcategories (potentially as small as one example in the training set). The paper creates a data model, and shows that while a model that properly learns features can achieve very high performance, models that do not learn these fea...
The paper proposes a method to cast fine-grained visual parsing into analogical inference instead of previous methods that mapping input scenes to part labels. The main idea is to retrieve the new input point cloud feature embedding from memories of labelled feature embeddings. Then the several results with the highest...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a method to cast fine-grained visual parsing into analogical inference instead of previous methods that mapping input scenes to part labels. The main idea is to retrieve the new input point cloud feature embedding from memories of labelled feature embeddings. Then the several results with the...
The paper proposed a solution to the practical problem of contextual speech synthesis to address TTS model limitation for highly phonetically contextual languages such as Chinese. The paper employs the use of conformer architecture baseline and introduces 3 main changes to enable paragraph-level expressiveness - segmen...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposed a solution to the practical problem of contextual speech synthesis to address TTS model limitation for highly phonetically contextual languages such as Chinese. The paper employs the use of conformer architecture baseline and introduces 3 main changes to enable paragraph-level expressiveness ...
This paper studies the landscape of a 2-layer neural network with $r$ hidden neurons. It adopts a teacher-student network setting where the training data is generated by a 2-layer network (teacher network) with $k$ hidden neurons. This paper establishes the following results: 1) For $r=1$, prove that if the teacher n...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the landscape of a 2-layer neural network with $r$ hidden neurons. It adopts a teacher-student network setting where the training data is generated by a 2-layer network (teacher network) with $k$ hidden neurons. This paper establishes the following results: 1) For $r=1$, prove that if the t...
The paper draws attention to the societal and environmental impacts of machine learning research in chemistry and biology. The paper specifically makes a contribution in the following components such as: environmental considerations, citation inequality, academic brain-drain, and scientific considerations: The study m...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper draws attention to the societal and environmental impacts of machine learning research in chemistry and biology. The paper specifically makes a contribution in the following components such as: environmental considerations, citation inequality, academic brain-drain, and scientific considerations: The...
Paper tries to answer: what do skip connections and normalisation layers do? Can we train without them? The authors attempt to analyze how our dependency on normalization and skip connections can be removed in deep NNs (specifically, in the self-attention operation of transformers). They do this by porting the “Deep K...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: Paper tries to answer: what do skip connections and normalisation layers do? Can we train without them? The authors attempt to analyze how our dependency on normalization and skip connections can be removed in deep NNs (specifically, in the self-attention operation of transformers). They do this by porting the...
This paper presents a drop-in differentiable module (DiffRes) that automatically adjusts temporal resolutions of spectrogram input for audio classification. The DiffRes module computes frame-level importance score and calculate a warp matrix to dynamically scale temporal resolution. The guide loss encourages empty fra...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a drop-in differentiable module (DiffRes) that automatically adjusts temporal resolutions of spectrogram input for audio classification. The DiffRes module computes frame-level importance score and calculate a warp matrix to dynamically scale temporal resolution. The guide loss encourages e...
This paper proposes a NAS method to find robust neural network that can defend adversarial attracks. It combine knowledge distillation with NAS to pursue better performacne. The authors conduct experiments on multiple tiny datasets, such as CIFAR-10, ImageNet-100. Strength: -It is interesing to find robust architectur...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a NAS method to find robust neural network that can defend adversarial attracks. It combine knowledge distillation with NAS to pursue better performacne. The authors conduct experiments on multiple tiny datasets, such as CIFAR-10, ImageNet-100. Strength: -It is interesing to find robust arc...
This paper proposes a new video Transformer architecture, termed UniformerV2, that extends the pre-trained ViTs (on images) to the video action recognition tasks. Specifically, UniformerV2 introduces local temporal aggregation, global temporal modeling and multi-stage fusion to enhance the original image-based ViTs to ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new video Transformer architecture, termed UniformerV2, that extends the pre-trained ViTs (on images) to the video action recognition tasks. Specifically, UniformerV2 introduces local temporal aggregation, global temporal modeling and multi-stage fusion to enhance the original image-based ...
The authors present a neural network method BFReg-NN for several tasks involving RNA-seq data. This method defines its architecture using biological databases that annotate gene regulatory networks, protein-protein interactions and pathways, represented as graphs/hypergraphs. The neural network uses graph attention to ...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors present a neural network method BFReg-NN for several tasks involving RNA-seq data. This method defines its architecture using biological databases that annotate gene regulatory networks, protein-protein interactions and pathways, represented as graphs/hypergraphs. The neural network uses graph atten...
This paper proposes a bi-level knowledge integration strategy that incorporates the prior knowledge from CLIP for weakly-supervised HOI detection. This paper also exploits CLIP textual embeddings of HOI labels as a relational knowledge bank, which is adopted to enhance the HOI representation with an image-wise HOI rec...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a bi-level knowledge integration strategy that incorporates the prior knowledge from CLIP for weakly-supervised HOI detection. This paper also exploits CLIP textual embeddings of HOI labels as a relational knowledge bank, which is adopted to enhance the HOI representation with an image-wise...
The paper aims to connect an effective regularization technique used in deep learning for visions with a special kind of measuring independence technique through kernel method. This connection would show that the regularization's goal is to suppress correlation of learned representations. Then the authors proceed to va...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper aims to connect an effective regularization technique used in deep learning for visions with a special kind of measuring independence technique through kernel method. This connection would show that the regularization's goal is to suppress correlation of learned representations. Then the authors proce...
The authors proposed an algorithm for planning that uses learned components for generating subgoals (states up to k steps away), generating one low-level action between two states (conditional low-level policy, CLLP), verifying whether CLLP can find an action between two states, and predicting the distance between curr...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors proposed an algorithm for planning that uses learned components for generating subgoals (states up to k steps away), generating one low-level action between two states (conditional low-level policy, CLLP), verifying whether CLLP can find an action between two states, and predicting the distance betw...
The paper proposes a variant of an already existing algorithm CMA-ME for Qualty-Diversity (QD) problems by introducing a learning rate based annealing function. Strength: 1) The paper presents a systematic study with adequate empirical experiments. Weakness: 1) Significance and relevance of QD problem in machine lea...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: The paper proposes a variant of an already existing algorithm CMA-ME for Qualty-Diversity (QD) problems by introducing a learning rate based annealing function. Strength: 1) The paper presents a systematic study with adequate empirical experiments. Weakness: 1) Significance and relevance of QD problem in mac...
The paper describes INSPIRE, a methodology to provide recourse sets to users from which they can select a counterfactual instance which can overturn a potentially bad decision taken by a machine learning model. The authors show a method to efficiently sample counterfactuals tailored to the personal user’s cost function...	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 describes INSPIRE, a methodology to provide recourse sets to users from which they can select a counterfactual instance which can overturn a potentially bad decision taken by a machine learning model. The authors show a method to efficiently sample counterfactuals tailored to the personal user’s cost ...
This paper introduces a feasible way to perform curriculum learning in multi-agent reinforcement learning (MARL). It describes well why curriculum learning cannot be applied to MARL seamlessly and proposes sound solutions to these problems. The experimental evaluation is conducted in a satisfactory way, although the re...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces a feasible way to perform curriculum learning in multi-agent reinforcement learning (MARL). It describes well why curriculum learning cannot be applied to MARL seamlessly and proposes sound solutions to these problems. The experimental evaluation is conducted in a satisfactory way, althoug...
This paper studies the problem of alleviating convergence issues in FL in the face of non-iid data. In particular, the authors propose methods which align the gradient and the direction of update with the past update at the server, and align the server update with the aggregated client model update. The alignment is do...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper studies the problem of alleviating convergence issues in FL in the face of non-iid data. In particular, the authors propose methods which align the gradient and the direction of update with the past update at the server, and align the server update with the aggregated client model update. The alignme...
This paper introduces DINo, a continuous time and space solver for partial differential equations. The authors proposed to model the solution of the PDE as an implicit neural representation conditioned on the initial condition through an hyper-network. The dynamics is modeled as a learned latent ordinary differential e...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper introduces DINo, a continuous time and space solver for partial differential equations. The authors proposed to model the solution of the PDE as an implicit neural representation conditioned on the initial condition through an hyper-network. The dynamics is modeled as a learned latent ordinary differ...
This paper studied the problem of learning agent polices with datasets that have limited labeled actions. Action Limited PreTraining (ALPT) is presented, which pretrain an inverse dynamics model (IDM) on multiple environments to provide accurate action labels for decision transformer (DT) agent on an action limited tar...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studied the problem of learning agent polices with datasets that have limited labeled actions. Action Limited PreTraining (ALPT) is presented, which pretrain an inverse dynamics model (IDM) on multiple environments to provide accurate action labels for decision transformer (DT) agent on an action lim...
The paper proposes an instance-dependent PLL approach by decomposing the generation process of candidate labels into two processes and explicitly modeling the processes using different probability distributions, which the risk estimator is built upon. The MAP technique is employed to create the final empirical risk est...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes an instance-dependent PLL approach by decomposing the generation process of candidate labels into two processes and explicitly modeling the processes using different probability distributions, which the risk estimator is built upon. The MAP technique is employed to create the final empirical ...
This work proposes a novel data representation framework, namely, the spacetime graph. Both theoretical and experimental analyses are provided to show the effectiveness and superiority of the proposed framework. - Strengths: 1. The authors present the definition of basic concepts in spacetime differential geometry ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This work proposes a novel data representation framework, namely, the spacetime graph. Both theoretical and experimental analyses are provided to show the effectiveness and superiority of the proposed framework. - Strengths: 1. The authors present the definition of basic concepts in spacetime differential g...
The paper studies the template inversion attack against face recognition (FR) systems. It uses results in the literature to train an encoder that generates the latent representation of StyleGAN and, using StyleGAN decoder, produces a realistic face, that is subsequently used to attack a FR system. Strength. The paper...	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 studies the template inversion attack against face recognition (FR) systems. It uses results in the literature to train an encoder that generates the latent representation of StyleGAN and, using StyleGAN decoder, produces a realistic face, that is subsequently used to attack a FR system. Strength. T...
The paper considers the problem of provably efficient reinforcement learning on predictive state representations (PSRs) with function approximation. The main contribution of the paper is a PAC algorithm for PSRs with polynomial sample complexity when competing with the globally optimal policy. Although sharing similar ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper considers the problem of provably efficient reinforcement learning on predictive state representations (PSRs) with function approximation. The main contribution of the paper is a PAC algorithm for PSRs with polynomial sample complexity when competing with the globally optimal policy. Although sharing ...
This paper addressed test-time adaptation by incorporating uncertainty. This is inspired by the fact that pseudo labels could be confidently wrong, thus pseudo labels are treated as a distribution and test-time training is carried out on the pseudo labels sampled from the distribution. A meta-learning approach is furth...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper addressed test-time adaptation by incorporating uncertainty. This is inspired by the fact that pseudo labels could be confidently wrong, thus pseudo labels are treated as a distribution and test-time training is carried out on the pseudo labels sampled from the distribution. A meta-learning approach ...
The paper proposes an automatic method to look for hard subpopulations and spurious correlations in datasets, by training SVMs in CLIP space to predict whether the original model will predict the data point correctly. Experiments show their method identified hard examples. Strengths: The paper proposes an automatic m...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes an automatic method to look for hard subpopulations and spurious correlations in datasets, by training SVMs in CLIP space to predict whether the original model will predict the data point correctly. Experiments show their method identified hard examples. Strengths: The paper proposes an aut...
The paper proposes a deep learning model for time series forecasting that reflects the irregularity in time series. Irregularity is based on Fourier series and and the authors employ it to design the Irregularity Representation Block that captures, preserves, and learns the irregularity representation of time series d...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a deep learning model for time series forecasting that reflects the irregularity in time series. Irregularity is based on Fourier series and and the authors employ it to design the Irregularity Representation Block that captures, preserves, and learns the irregularity representation of time ...
This work presents a straightforward, but impressive extension of prior work on “Behavior Transformers” for learning policies from offline, reward-free data, by introducing the ability to learn goal-conditioned policies from undirected play data. Crucially, the play data used to learn policies in this work in inherentl...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This work presents a straightforward, but impressive extension of prior work on “Behavior Transformers” for learning policies from offline, reward-free data, by introducing the ability to learn goal-conditioned policies from undirected play data. Crucially, the play data used to learn policies in this work in i...
This paper provides a theoretical analysis of contrastive learning in the setting where the hypothesis class has significantly smaller dimension than the data distribution (as measured by the number of clusters). They instantiate their framework on a few example settings and show that their results provide theoretical ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper provides a theoretical analysis of contrastive learning in the setting where the hypothesis class has significantly smaller dimension than the data distribution (as measured by the number of clusters). They instantiate their framework on a few example settings and show that their results provide theo...