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 proposes a theoretical analysis to understand the intrinsic dimension for deterministic continuous-time State and action space problems. This method is motivated by understanding the representations required to train optimal policies in these types of spaces will help us better understand the structure of ne...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a theoretical analysis to understand the intrinsic dimension for deterministic continuous-time State and action space problems. This method is motivated by understanding the representations required to train optimal policies in these types of spaces will help us better understand the structu...
The model consists of three parts: encoder, decod er, and RFTM, The encoder makes a nonlinear transform to the input data x in the latent space. The latent data goes through a re-representation (in the same dimensionality) by the RFTM. Then the data is reconstructed by the decoder. The anomaly of the test sample is me...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The model consists of three parts: encoder, decod er, and RFTM, The encoder makes a nonlinear transform to the input data x in the latent space. The latent data goes through a re-representation (in the same dimensionality) by the RFTM. Then the data is reconstructed by the decoder. The anomaly of the test samp...
The paper presents a hierarchical algorithm based on the locality of data points. Entities are represented in the D dimensional space as D-dimensional hypercubes and are assigned colors based on whether they are completely/partially contained within another entity. The algorithm first inserts entities into a tree, colo...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper presents a hierarchical algorithm based on the locality of data points. Entities are represented in the D dimensional space as D-dimensional hypercubes and are assigned colors based on whether they are completely/partially contained within another entity. The algorithm first inserts entities into a tr...
The paper studies the effect of using pre-trained neural network weights on federated optimization. Several combinations of server and client optimizers are considered in combination with 2 different models for vision and language tasks. It was shown that pre-trained initialization reduces the negative effects of data ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper studies the effect of using pre-trained neural network weights on federated optimization. Several combinations of server and client optimizers are considered in combination with 2 different models for vision and language tasks. It was shown that pre-trained initialization reduces the negative effects ...
This work propose a Deep Power Law (DPL) which uses the scaling law property of learning curves for hyperparameter optimization, by improving fidelity of HOP for deep learning. DPL ensembles predicts the performance of hyperparameter configurations in low-budget regimes as a probabilistic surrogator for Bayesian optimi...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work propose a Deep Power Law (DPL) which uses the scaling law property of learning curves for hyperparameter optimization, by improving fidelity of HOP for deep learning. DPL ensembles predicts the performance of hyperparameter configurations in low-budget regimes as a probabilistic surrogator for Bayesia...
This paper proposed a federated framework for correlation tests. And demonstrates an empirical evaluation of chi-square test and G-test under this framework. Strength: - The problem considered is important. - The proposed solution is intuitive and not too complex. It's also backed by theoretical analysis and security p...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposed a federated framework for correlation tests. And demonstrates an empirical evaluation of chi-square test and G-test under this framework. Strength: - The problem considered is important. - The proposed solution is intuitive and not too complex. It's also backed by theoretical analysis and se...
This work examines the generalization error and model error of single-step MAML in the mixed linear regression task. Unlike prior works, this work explicitly considers optimization in the over-parameterized regime, and derives model-error bounds in the non-asymptotic regime. Through theoretical derivations, the authors...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work examines the generalization error and model error of single-step MAML in the mixed linear regression task. Unlike prior works, this work explicitly considers optimization in the over-parameterized regime, and derives model-error bounds in the non-asymptotic regime. Through theoretical derivations, the...
The authors introduce a new architecture for forecasting non-stationary time series, the ‘Koopman Neural Forecaster (KNF)’. KNF combines local and global Koopman matrix operators assembled from a set of basis functions, and also includes a feedback operator based on local prediction errors. Parameters of the Koopman op...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors introduce a new architecture for forecasting non-stationary time series, the ‘Koopman Neural Forecaster (KNF)’. KNF combines local and global Koopman matrix operators assembled from a set of basis functions, and also includes a feedback operator based on local prediction errors. Parameters of the Ko...
This paper suggests a Machine Learning (ML) driven method to support solving the Navier-Stokes equation of incompressible flows, a classical problem in Computational Fluid Dynamics (CFD). The submitted approach learns to infer the stencils of advection-diffusion fluids in a Finite Volume Method framework. The authors...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper suggests a Machine Learning (ML) driven method to support solving the Navier-Stokes equation of incompressible flows, a classical problem in Computational Fluid Dynamics (CFD). The submitted approach learns to infer the stencils of advection-diffusion fluids in a Finite Volume Method framework. The...
In this paper the authors study the problem of learning a fair representation given a dataset. In order to learn fair representations, previous works have balanced the reconstruction loss with an adversarial loss using a weighting term. This weighting term is a central problem identified by the authors as blind selecti...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper the authors study the problem of learning a fair representation given a dataset. In order to learn fair representations, previous works have balanced the reconstruction loss with an adversarial loss using a weighting term. This weighting term is a central problem identified by the authors as blind...
This work reviewed the existing solution of Dual BN for adversarial robustness by focusing on the affine parameters in BN. It is argued that the stronger adversarial robustness in Dual BN is mainly attributed to additional affine parameters. Extensive ablation studies are presented to validate the argument. Moreover, i...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work reviewed the existing solution of Dual BN for adversarial robustness by focusing on the affine parameters in BN. It is argued that the stronger adversarial robustness in Dual BN is mainly attributed to additional affine parameters. Extensive ablation studies are presented to validate the argument. Mor...
This paper proposes O3F, which is an action selection algorithm when doing an offline-to-online fine tuning. Since value estimates and policies we had was trained under a conservative objective (in many offline Rl algorithms), it may be too pessimistic in online fine-tuning, and may not improve much. To alleviate this ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes O3F, which is an action selection algorithm when doing an offline-to-online fine tuning. Since value estimates and policies we had was trained under a conservative objective (in many offline Rl algorithms), it may be too pessimistic in online fine-tuning, and may not improve much. To allevia...
This paper focuses on unlearning federated clustering problems with k-means++ initialization and secure compressed multiset aggregation. The idea of unlearning is to perform k-means++ initialization at each local client. By exploiting the nature of k-means++ approximation, the centroid set C may remain unchanged if rem...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper focuses on unlearning federated clustering problems with k-means++ initialization and secure compressed multiset aggregation. The idea of unlearning is to perform k-means++ initialization at each local client. By exploiting the nature of k-means++ approximation, the centroid set C may remain unchange...
This paper proposes a new communication compression method for federated learning with heterogeneous data. In particular, the compression scheme is personalized to each client. The authors show that several new techniques, such as downstream compression personalization, and element-wise aggregation, are critical to ach...	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 paper proposes a new communication compression method for federated learning with heterogeneous data. In particular, the compression scheme is personalized to each client. The authors show that several new techniques, such as downstream compression personalization, and element-wise aggregation, are critica...
The paper aims to provide an explanation of double descent using classical VC theory. The primary argument is that since the test error can be bounded by a function involving both the training error and the VC dimension, one may observe double descent if one minimizes the training error and then “minimizes the VC dimen...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper aims to provide an explanation of double descent using classical VC theory. The primary argument is that since the test error can be bounded by a function involving both the training error and the VC dimension, one may observe double descent if one minimizes the training error and then “minimizes the ...
This paper theoretically analyzed the trade-off between universality (measured by the average performance of multiple tasks) and label efficiency (measured by the amount of labeled data needed for a downstream task) in constrastive learning. Both empirical evidence and theoretical guarantee were given. (I'm...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper theoretically analyzed the trade-off between universality (measured by the average performance of multiple tasks) and label efficiency (measured by the amount of labeled data needed for a downstream task) in constrastive learning. Both empirical evidence and theoretical guarantee were giv...
The paper claims that humans do not always favor high-likelihood texts; decoding using high likelihood leads to repetition and boredom. So, the paper proposes a sampling method. It rescales the high-likelihood probabilities in a distribution (determined by quartiles and interquartile range aka IQR, as shown in Equation...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper claims that humans do not always favor high-likelihood texts; decoding using high likelihood leads to repetition and boredom. So, the paper proposes a sampling method. It rescales the high-likelihood probabilities in a distribution (determined by quartiles and interquartile range aka IQR, as shown in ...
This paper describes an approach for learning a reward function in preference-based reinforcement learning based on transformers. This provides the advantages that the model allows credit assignment within the behavior trajectory to correctly weight significant state/actions, and that less feedback samples that has ty...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper describes an approach for learning a reward function in preference-based reinforcement learning based on transformers. This provides the advantages that the model allows credit assignment within the behavior trajectory to correctly weight significant state/actions, and that less feedback samples tha...
In this paper, the authors propose to regularize the value function approximation in MBRL in order to learn policies without having to use an ensemble of models (as is done in MBPO bu Janner). After the authors present evidence that the ensemble of models is indeed regularizing the value function by using the value-aw...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In this paper, the authors propose to regularize the value function approximation in MBRL in order to learn policies without having to use an ensemble of models (as is done in MBPO bu Janner). After the authors present evidence that the ensemble of models is indeed regularizing the value function by using the ...
The paper presents BigVGAN, a GAN based vocoder that generalizes well for OOD scenarios without additional finetuning. They seem to achieve state-of-the-art performance for audio synthesis in a variety of novel scenarios like new speakers, unseen recording environments etc. The proposed model is able to reproduce a lot...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents BigVGAN, a GAN based vocoder that generalizes well for OOD scenarios without additional finetuning. They seem to achieve state-of-the-art performance for audio synthesis in a variety of novel scenarios like new speakers, unseen recording environments etc. The proposed model is able to reprodu...
The authors introduced a novel vision-language multi modality model (Spotlight) for mobile UI understanding tasks. Specifically this model is designed for tasks for mapping visual components in the UI to natural language text. In the context of UI understanding models, this paper presents 2 major contributions. 1. The...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors introduced a novel vision-language multi modality model (Spotlight) for mobile UI understanding tasks. Specifically this model is designed for tasks for mapping visual components in the UI to natural language text. In the context of UI understanding models, this paper presents 2 major contributions....
This paper proposes a quantitative similarity score between different neural architectures based on the adversarial attack transferability. This smiliarity helps to understand the component-level architecture design, and leads to better understanding of the relationship between model similarity of model ensemble perfor...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a quantitative similarity score between different neural architectures based on the adversarial attack transferability. This smiliarity helps to understand the component-level architecture design, and leads to better understanding of the relationship between model similarity of model ensembl...
The authors provide a new pooling strategy for GNN on learning physical systems on unstructured meshes along with a down sampling and up sampling method to reduce overhead between levels. Current GNN method to infer physical states suffers from two limitations. The computation complexity is high and over-smoothing prob...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors provide a new pooling strategy for GNN on learning physical systems on unstructured meshes along with a down sampling and up sampling method to reduce overhead between levels. Current GNN method to infer physical states suffers from two limitations. The computation complexity is high and over-smooth...
The paper explores the idea of addressing catastrophic forgetting through rehearsal using a memory buffer. It studies strategies to select the most important samples for storage in a memory buffer. The core idea is to design a consistency score that ranks samples for storage according to how easy they are to learn and ...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper explores the idea of addressing catastrophic forgetting through rehearsal using a memory buffer. It studies strategies to select the most important samples for storage in a memory buffer. The core idea is to design a consistency score that ranks samples for storage according to how easy they are to le...
The paper proposes "Concept-Monitor", a method for analyzing the training process of a given neural network architecture from the perspective of interpretability. More specifically, the proposed method applies concept detectors at every step (i.e. iteration or epoch) of the training process. The detected concepts are ...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes "Concept-Monitor", a method for analyzing the training process of a given neural network architecture from the perspective of interpretability. More specifically, the proposed method applies concept detectors at every step (i.e. iteration or epoch) of the training process. The detected conce...
This paper is motivated by the limited availability of actions in datasets when training RL agents. To better leverage the large amount of data not paired with actions, this paper proposes to learn a model to annotate actions for the dataset. More concretely, this paper proposes Action Limited PreTraining (ALPT), which...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper is motivated by the limited availability of actions in datasets when training RL agents. To better leverage the large amount of data not paired with actions, this paper proposes to learn a model to annotate actions for the dataset. More concretely, this paper proposes Action Limited PreTraining (ALPT...
The manuscript studies the estimation error of fitted Q-evaluation with convolutional neural networks under the assumption of a low dimensional state-action space. The main contribution is that the decay rate of the estimation error only depends on the dimension of the state-action space rather than the ambient dimens...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The manuscript studies the estimation error of fitted Q-evaluation with convolutional neural networks under the assumption of a low dimensional state-action space. The main contribution is that the decay rate of the estimation error only depends on the dimension of the state-action space rather than the ambien...
The paper proposes a combination of three hardware-accelerated techniques (Gecko, Quantum Mantissa, and BitChop) for optimizing the memory footprint of model training through low-precision floating point tensors. The approach is inspired by the observation that in some scenarios, the value distribution of training stat...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a combination of three hardware-accelerated techniques (Gecko, Quantum Mantissa, and BitChop) for optimizing the memory footprint of model training through low-precision floating point tensors. The approach is inspired by the observation that in some scenarios, the value distribution of train...
The paper proposes a two-stage curriculum learning strategy -- TuneUp -- to improve GNN's performance on tail nodes of small node degrees. The paper is motivated by the observation that GNNs perform worse on nodes with smaller degrees. In this first stage, TuneUp conducts standard training without augmentations. In ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes a two-stage curriculum learning strategy -- TuneUp -- to improve GNN's performance on tail nodes of small node degrees. The paper is motivated by the observation that GNNs perform worse on nodes with smaller degrees. In this first stage, TuneUp conducts standard training without augmentati...
This paper presents a method to generate counterfactual explanations based on two components: (i) a feature selector and (ii) and an end-to-end network. While (i) makes sure that relevant features are selected, (ii) enables the model to enforce additional constraints (e.g., dealing with categorical variables). Strength...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper presents a method to generate counterfactual explanations based on two components: (i) a feature selector and (ii) and an end-to-end network. While (i) makes sure that relevant features are selected, (ii) enables the model to enforce additional constraints (e.g., dealing with categorical variables). ...
This paper concerns the approximation power of ReLU FNNs for Sobolev (or actually Hölder) smooth functions defined on spheres. This paper constructs ReLU FNNs with certain architectures to obtain their approximation results. The main idea is to leverage approximation capability of the homogeneous harmonic polynomials o...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper concerns the approximation power of ReLU FNNs for Sobolev (or actually Hölder) smooth functions defined on spheres. This paper constructs ReLU FNNs with certain architectures to obtain their approximation results. The main idea is to leverage approximation capability of the homogeneous harmonic polyn...
The authors studied a method that can detect the presence of unstable states. They achieve that by measuring the gap between the perturbed cost and the first order approximation. I this gap is large (one the direction is chosen in a worst case fashion) then the second order curvature is large as well (in a worst case s...	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 authors studied a method that can detect the presence of unstable states. They achieve that by measuring the gap between the perturbed cost and the first order approximation. I this gap is large (one the direction is chosen in a worst case fashion) then the second order curvature is large as well (in a wors...
This paper presents a novel targeted preference-based adversarial attack against deep reinforcement learning (DRL) agents, so that the DRL agents would show extreme behaviors desired by adversaries. In particular, the proposed PALM adopts human preference, an intention policy and a weighting function to guide the adver...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents a novel targeted preference-based adversarial attack against deep reinforcement learning (DRL) agents, so that the DRL agents would show extreme behaviors desired by adversaries. In particular, the proposed PALM adopts human preference, an intention policy and a weighting function to guide t...
This paper proposes a new method for ranking of offline RL policies with off-policy evaluation (OPE). The ranking is produced with a model that 1) learns a pairwise policy representation with a transformer architecture, 2) uses a crowd layer to aggregate OPE scores of other methods. In the experimental results the auth...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new method for ranking of offline RL policies with off-policy evaluation (OPE). The ranking is produced with a model that 1) learns a pairwise policy representation with a transformer architecture, 2) uses a crowd layer to aggregate OPE scores of other methods. In the experimental results ...
This paper aims to investigate the secrets of why contrastive learning for sentence representation learning (SRL) works from the perspective of alignment and uniformity (Wang & Isola, 2020). While alignment & uniformity are usually utilized in the literature as a tool for demonstrating the effectiveness of some SRL ap...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper aims to investigate the secrets of why contrastive learning for sentence representation learning (SRL) works from the perspective of alignment and uniformity (Wang & Isola, 2020). While alignment & uniformity are usually utilized in the literature as a tool for demonstrating the effectiveness of som...
This paper proposes a GNN based approach for vulnerability detection. The control flow graph (CFG) of a program is augmented with dataflow information and graph learning is applied to predict vulnerability of the program. This work proposes an “abstract dataflow embedding” to represent important dataflow properties tha...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a GNN based approach for vulnerability detection. The control flow graph (CFG) of a program is augmented with dataflow information and graph learning is applied to predict vulnerability of the program. This work proposes an “abstract dataflow embedding” to represent important dataflow proper...
The paper proposes a strategy to find a parameter-efficient tuneable architecture from a given pre-trained backbone neural network. In particular, the strategy consits of four phases exploring 1) how to group layers in the backbone neural network; 2) how to allocate the tuneable parameters within each group; 3) how to ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a strategy to find a parameter-efficient tuneable architecture from a given pre-trained backbone neural network. In particular, the strategy consits of four phases exploring 1) how to group layers in the backbone neural network; 2) how to allocate the tuneable parameters within each group; 3)...
This paper presents an interactive algorithm for multi-class/multi-instance segmentation based on the vision features obtained from self-supervised pre-training. Previously, such vision features, e.g. DINO features, have been used for unsupervised segmentation. This work extends previous work to the interactive segment...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper presents an interactive algorithm for multi-class/multi-instance segmentation based on the vision features obtained from self-supervised pre-training. Previously, such vision features, e.g. DINO features, have been used for unsupervised segmentation. This work extends previous work to the interactive...
The paper provides a new regulariser term for training energy-based models, which promotes feature diversity. A theoretical analysis on the generalisation performance of energy-based models using the PAC-learning framework gives a solid motivating evidence on the need of the regularizer. Specifically, the authors consi...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper provides a new regulariser term for training energy-based models, which promotes feature diversity. A theoretical analysis on the generalisation performance of energy-based models using the PAC-learning framework gives a solid motivating evidence on the need of the regularizer. Specifically, the autho...
The paper proposes an unsupervised learning rule, NeAW, combining Hebbian and anti-Hebbian learning and applies it to 3D point cloud classification. The learning rule is motivated by the observation that Hebbian and anti-Hebbian learning don't produce diverse representations across neurons (there is low variance in neu...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper proposes an unsupervised learning rule, NeAW, combining Hebbian and anti-Hebbian learning and applies it to 3D point cloud classification. The learning rule is motivated by the observation that Hebbian and anti-Hebbian learning don't produce diverse representations across neurons (there is low varianc...