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 deal with the problem of identifying the different regions of space as they are partitioned by the parameters of a ReLU neural network, specifically identifying what combinatorial properties govern the induced partition. The authors provide a combinatorial description of the canonical polyhedral complex of...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors deal with the problem of identifying the different regions of space as they are partitioned by the parameters of a ReLU neural network, specifically identifying what combinatorial properties govern the induced partition. The authors provide a combinatorial description of the canonical polyhedral co...
This work proposes a continuous normalizing flow model, an exact likelihood generative modeling method, which is conditional and permutation invariant. The flow is driven by dynamics computed from a shared global force term and a pairwise interaction term which have shared weights across all elements thus achieving pe...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This work proposes a continuous normalizing flow model, an exact likelihood generative modeling method, which is conditional and permutation invariant. The flow is driven by dynamics computed from a shared global force term and a pairwise interaction term which have shared weights across all elements thus achi...
The paper proposes a solution to the planning as inference problem in domains where it is easy to generate samples of actions in the current state but it is expensive to simulate samples of next states. The proposed solution involves learning a parametric approximation of the soft value of each state-action pair by usi...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper proposes a solution to the planning as inference problem in domains where it is easy to generate samples of actions in the current state but it is expensive to simulate samples of next states. The proposed solution involves learning a parametric approximation of the soft value of each state-action pai...
This paper investigates the Non-IID problems of subgraph Federated Learning (FL), where the authors consider both the well-known homogeneity assumption and the under-explored heterogeneity assumption. To tackle those two problems, the authors propose the global knowledge extractor that uses global data to extract featu...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper investigates the Non-IID problems of subgraph Federated Learning (FL), where the authors consider both the well-known homogeneity assumption and the under-explored heterogeneity assumption. To tackle those two problems, the authors propose the global knowledge extractor that uses global data to extra...
This work targets to the instance-dependent noisy-label learning problem and proposes a new curriculum learning-based algorithm. In particular, the authors calculate the 'time-consistency of prediction' (TCP) scores that indicate how consistent are the predictions of samples over the course of training. Then, the propo...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This work targets to the instance-dependent noisy-label learning problem and proposes a new curriculum learning-based algorithm. In particular, the authors calculate the 'time-consistency of prediction' (TCP) scores that indicate how consistent are the predictions of samples over the course of training. Then, t...
The paper studies a setting with two environments from a specific instance of the model from Rosenfeld et al. and show that overfitting (i.e., interpolation of the training data) and invariance to the spurious features are fundamentally at odds, and that while both are achievable, they cannot be done simultaneously. Th...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies a setting with two environments from a specific instance of the model from Rosenfeld et al. and show that overfitting (i.e., interpolation of the training data) and invariance to the spurious features are fundamentally at odds, and that while both are achievable, they cannot be done simultaneo...
The paper proposes a denoising diffusion probabilistic model for generating protein backbone structures. To overcome the complication of equivariance when working with 3D structures, the approach directly models the angles that describe the orientation of residues relative to their neighbors. Results show that the mode...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes a denoising diffusion probabilistic model for generating protein backbone structures. To overcome the complication of equivariance when working with 3D structures, the approach directly models the angles that describe the orientation of residues relative to their neighbors. Results show that ...
This paper presents a new way to allow cooperation between ML agents using an approach based on a similarity measurement. Authors introduce a program called diff meta games, a novel variant of program meta games, where agents observe the similarity/dissimilarity between the policies using a single number, thus the simi...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper presents a new way to allow cooperation between ML agents using an approach based on a similarity measurement. Authors introduce a program called diff meta games, a novel variant of program meta games, where agents observe the similarity/dissimilarity between the policies using a single number, thus ...
The paper proposed a feature-based method to get OOD scores that measures the trajectory similarities between the test data and the training data. The method first gets the per-class feature prototype \mu_{l,y} at each layer based on training data, and then figures out the reference trajectory by calculating the cosine...	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 proposed a feature-based method to get OOD scores that measures the trajectory similarities between the test data and the training data. The method first gets the per-class feature prototype \mu_{l,y} at each layer based on training data, and then figures out the reference trajectory by calculating th...
In this paper, the authors addressed an open-set 3D detection problem to broaden the vocabulary of the point-cloud detection without laborious and expensive data annotation. Inspired by previous open-set works, the authors proposed OS-3DETIC consisting of two main functions as classification using image-based model and...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors addressed an open-set 3D detection problem to broaden the vocabulary of the point-cloud detection without laborious and expensive data annotation. Inspired by previous open-set works, the authors proposed OS-3DETIC consisting of two main functions as classification using image-based m...
This paper attempts to learn a state abstraction based on MDP homomorphism. The goal is to reduce the size of state-action space and allow for faster RL. The proposed approach learns a backward model $(B)$ and a forward model $(F)$. The forward model $F(s, a)$ is trained to predict the next state given the current stat...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper attempts to learn a state abstraction based on MDP homomorphism. The goal is to reduce the size of state-action space and allow for faster RL. The proposed approach learns a backward model $(B)$ and a forward model $(F)$. The forward model $F(s, a)$ is trained to predict the next state given the curr...
The paper introduces a new estimator called ZGR aimed for the stable and accurate use of discrete random variables as part of approximate inference pipelines. Strengths: i) The studied problem is important for the probabilistic machine learning community ii) The paper reports comprehensive experiments that span man...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: The paper introduces a new estimator called ZGR aimed for the stable and accurate use of discrete random variables as part of approximate inference pipelines. Strengths: i) The studied problem is important for the probabilistic machine learning community ii) The paper reports comprehensive experiments that ...
The paper tackles semi-supervised learning with a focus on the most challenging setting of when there is a mismatch between the distributions of labelled and unlabelled data. The proposed method, PRG, uses the history of class transitions to sway the pseudo-labelling mechanism towards under-represented classes. This wo...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper tackles semi-supervised learning with a focus on the most challenging setting of when there is a mismatch between the distributions of labelled and unlabelled data. The proposed method, PRG, uses the history of class transitions to sway the pseudo-labelling mechanism towards under-represented classes....
The authors propose a new method for federated domain adaptation (FDA) with application to medical imaging. to improve target domain performance, authors consider a multi-task learning (MTL) framework over the source domains. Several cheap secondary tasks are considered locally (at each source client). Such local infor...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors propose a new method for federated domain adaptation (FDA) with application to medical imaging. to improve target domain performance, authors consider a multi-task learning (MTL) framework over the source domains. Several cheap secondary tasks are considered locally (at each source client). Such loc...
The authors propose to diversify the models in an ensemble forming the teachers in distillation. The specific setup the authors consider is on-line distillation (where the teachers are trained in parallel with the student) and peer-based (where the models share parts of their weights). The proposed approach to achie...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors propose to diversify the models in an ensemble forming the teachers in distillation. The specific setup the authors consider is on-line distillation (where the teachers are trained in parallel with the student) and peer-based (where the models share parts of their weights). The proposed approach ...
This paper introduces a code synthesis model that is trained to perform both autoregressive left-to-right generation as well as infilling via added mask tokens. The experimental results show improved performance on various infilling related code generation tasks. The paper is well written and easy to read. The approach...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a code synthesis model that is trained to perform both autoregressive left-to-right generation as well as infilling via added mask tokens. The experimental results show improved performance on various infilling related code generation tasks. The paper is well written and easy to read. The ...
This paper addresses the problem of learning a set of diverse policies that are simultaneously nearly-optimal for a given reward function. In their formulation, the diverse near-optimal policies are learned by maximizing an objective function related to diversity, while near-optimality is ensured through a linear const...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper addresses the problem of learning a set of diverse policies that are simultaneously nearly-optimal for a given reward function. In their formulation, the diverse near-optimal policies are learned by maximizing an objective function related to diversity, while near-optimality is ensured through a line...
This paper provides a theoretical analysis of network-based transfer learning using layer variational analysis. Their analysis is claimed to prove that the success of transfer learning is guaranteed with certain data conditions. Based on their analysis, they also propose an alternative method for network-based transfer...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper provides a theoretical analysis of network-based transfer learning using layer variational analysis. Their analysis is claimed to prove that the success of transfer learning is guaranteed with certain data conditions. Based on their analysis, they also propose an alternative method for network-based ...
This paper aims to automatically solve combinatorial optimization by leveraging unsupervised learning, learning from historical data and achieving an instance-wise good solution simultaneously. In order to do so, the authors propose a methodology for warm-starting future combinatorial optimization problem instances. T...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper aims to automatically solve combinatorial optimization by leveraging unsupervised learning, learning from historical data and achieving an instance-wise good solution simultaneously. In order to do so, the authors propose a methodology for warm-starting future combinatorial optimization problem inst...
This paper provides necessary and sufficient conditions for domain generalization. Experimental results on a few benchmark datasets are provided. Strengths: A theoretical foundation for generating representations that are domain-invariant and domain generalizable are provided. Weaknesses: Poorly written paper. Experime...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper provides necessary and sufficient conditions for domain generalization. Experimental results on a few benchmark datasets are provided. Strengths: A theoretical foundation for generating representations that are domain-invariant and domain generalizable are provided. Weaknesses: Poorly written paper. ...
The paper proposes Unleashing Mask (UM) with the goal of maintaining a model's capability to distinguish between out-of-distribution (OOD) and in-distribution (ID) training examples. The paper posits that as classification models train and the training loss decreases, there is a point where simple measures such as OD...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes Unleashing Mask (UM) with the goal of maintaining a model's capability to distinguish between out-of-distribution (OOD) and in-distribution (ID) training examples. The paper posits that as classification models train and the training loss decreases, there is a point where simple measures su...
This paper proposes a diffusion model MeshDiffusion for 3D meshes generation. The meshes are embedded in a deformable tetrahedral grid, and then employ a score based method to train a diffusion model on this direct parameterization. Besides using in the 3D space, the network can be used when only 2.5D information is a...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposes a diffusion model MeshDiffusion for 3D meshes generation. The meshes are embedded in a deformable tetrahedral grid, and then employ a score based method to train a diffusion model on this direct parameterization. Besides using in the 3D space, the network can be used when only 2.5D informat...
This paper studies parametric offline rl with a class of general differentiable function class approximation (DFA). The differentiable function class they consider is the set of functions $f(\theta, \phi)$ whose argument can be decomposed into two parts: $\theta$ is the parameter vector and $\phi$ is the feature map. T...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies parametric offline rl with a class of general differentiable function class approximation (DFA). The differentiable function class they consider is the set of functions $f(\theta, \phi)$ whose argument can be decomposed into two parts: $\theta$ is the parameter vector and $\phi$ is the featur...
This paper follows a line of papers who question the importance of massive corpora for pre-training MLMs. The authors show that by pre-training MLMs on downstream corpora (which are far smaller than standard pre-training corpora), they reach similar performance on the corresponding downstream task to that of off-the-sh...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper follows a line of papers who question the importance of massive corpora for pre-training MLMs. The authors show that by pre-training MLMs on downstream corpora (which are far smaller than standard pre-training corpora), they reach similar performance on the corresponding downstream task to that of of...
This paper proposes the idea of using time-based data augmentations to aid general purpose machine vision systems. Authors provide a thorough and very detailed psychological motivation for why this may work, and later use simulated agents's view points to simulate the time-based components of learning. Authors find tha...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper proposes the idea of using time-based data augmentations to aid general purpose machine vision systems. Authors provide a thorough and very detailed psychological motivation for why this may work, and later use simulated agents's view points to simulate the time-based components of learning. Authors ...
This paper proposes the continual federated learning problem which has properties of both federated learning and continual learning, and proposes modifications such as model consolidation and enforced consistency to stabilize ACGAN to solve the proposed problem. The proposed approach achieves the best performance in th...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes the continual federated learning problem which has properties of both federated learning and continual learning, and proposes modifications such as model consolidation and enforced consistency to stabilize ACGAN to solve the proposed problem. The proposed approach achieves the best performan...
This contribution tackles the domain generalization problem by assuming that every real-world distribution is composed of a mixture of elementary distributions, which remain invariant across different domains. This assumption was named I.E.D. (Invariant Elementary Distribution). The authors presented a lemma to support...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This contribution tackles the domain generalization problem by assuming that every real-world distribution is composed of a mixture of elementary distributions, which remain invariant across different domains. This assumption was named I.E.D. (Invariant Elementary Distribution). The authors presented a lemma to...
This paper proves that conditional ReLU networks can memorize any dataset of size n, in such a way that performing inference with this network requires only $O(\log{n})$ operations per input. It also proves that this is the best possible result, assuming mild conditions on the dataset. The best known results for uncond...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves that conditional ReLU networks can memorize any dataset of size n, in such a way that performing inference with this network requires only $O(\log{n})$ operations per input. It also proves that this is the best possible result, assuming mild conditions on the dataset. The best known results fo...
This paper proposes a mutual learning framework called ML-PLL for partial label learning under competitive label noise. The proposed ML-PLL method disambiguates the irrelevant label noise with noise label correction and learns a couple classifiers (a prediction network based classifier and a class-prototype based class...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper proposes a mutual learning framework called ML-PLL for partial label learning under competitive label noise. The proposed ML-PLL method disambiguates the irrelevant label noise with noise label correction and learns a couple classifiers (a prediction network based classifier and a class-prototype bas...
This paper proposes a variational autoencoder strategy to offer robustness to hospital-specific X-ray characteristics. The methodology aims to disentangle these centre-specific measurement characteristics from the features predictive of the condition. The paper presents theoretical intuition and evidence of the stronge...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a variational autoencoder strategy to offer robustness to hospital-specific X-ray characteristics. The methodology aims to disentangle these centre-specific measurement characteristics from the features predictive of the condition. The paper presents theoretical intuition and evidence of the...
The paper proposes a new notion of fairness called Equal Improvability (EI) that requires a binary classifier to be equally difficult to change predictions for each sensitive attribution group. To achieve this property, three different penalizations are discussed: covariance-based EI penalty, KDE-based EI penalty, and ...	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 notion of fairness called Equal Improvability (EI) that requires a binary classifier to be equally difficult to change predictions for each sensitive attribution group. To achieve this property, three different penalizations are discussed: covariance-based EI penalty, KDE-based EI penal...
The authors propose a new sampling strategy for the construction of batches containing hard-negatives (avoiding false negatives in the batch) for contrastive learning methods. According to previous studies, it is legitimate to assume that hard negative pairs contribute the most in training the network. The method is ba...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors propose a new sampling strategy for the construction of batches containing hard-negatives (avoiding false negatives in the batch) for contrastive learning methods. According to previous studies, it is legitimate to assume that hard negative pairs contribute the most in training the network. The meth...
Authors propose prompt generator based on the whole code repository for the code completion tasks. Authors show a significant improvement over baseline Codex results by using their prompt generator. Strengths: - Addresses an important and novel area of good prompt generation for LLM tasks - Approach does not require ac...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Authors propose prompt generator based on the whole code repository for the code completion tasks. Authors show a significant improvement over baseline Codex results by using their prompt generator. Strengths: - Addresses an important and novel area of good prompt generation for LLM tasks - Approach does not re...
This work proposes an early stopping mechanism for DIP based on the windowed moving variance of the output images. An early stop is detected when the variance does not decrease after a certain number of training steps. The proposed method is observed to have small PSNR and SSIM gaps compared to other early stopping met...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work proposes an early stopping mechanism for DIP based on the windowed moving variance of the output images. An early stop is detected when the variance does not decrease after a certain number of training steps. The proposed method is observed to have small PSNR and SSIM gaps compared to other early stop...
This paper solves the problem that existing learned index methods adopt a fixed ε for all the learned segments, which neglecting the diverse characteristics of different data localities and propose a mathematically-grounded learned index framework with dynamic ε, which is efficient and pluggable to existing learned ind...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper solves the problem that existing learned index methods adopt a fixed ε for all the learned segments, which neglecting the diverse characteristics of different data localities and propose a mathematically-grounded learned index framework with dynamic ε, which is efficient and pluggable to existing lea...
This work is concerned with improving classification of brain activity (MEG). Typically, this would be done by training a model for each subject. But this work presents a strategy for pooling data across multiple subjects (n=15) to produce a "group-level" model. In the naive approach, this can be done by combining data...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This work is concerned with improving classification of brain activity (MEG). Typically, this would be done by training a model for each subject. But this work presents a strategy for pooling data across multiple subjects (n=15) to produce a "group-level" model. In the naive approach, this can be done by combin...
This paper works on expanding the exact training dynamics analysis of Saxe et al 2019, which considers interrelated semantic groupings, to also cover systematic behavior. For this analysis, they consider the effect of combinatorial behavior in the dataset on the resulting gradient dynamics. In particular, they consider...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper works on expanding the exact training dynamics analysis of Saxe et al 2019, which considers interrelated semantic groupings, to also cover systematic behavior. For this analysis, they consider the effect of combinatorial behavior in the dataset on the resulting gradient dynamics. In particular, they ...
The paper investigates an efficient Transformer based method for high-quality image restoration. Specifically, the proposed Attention Retractable Transformer (ART) incorporates sparse and dense attention to sense larger receptive field. The alternative manner of those two attention blocks help capture global and local ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper investigates an efficient Transformer based method for high-quality image restoration. Specifically, the proposed Attention Retractable Transformer (ART) incorporates sparse and dense attention to sense larger receptive field. The alternative manner of those two attention blocks help capture global an...
This paper observes that the dynamics of the ensemble model provide a distinction between clean and noisy labeled data in the training phase. The diversity of noisy labeled data is much larger than that of clean labeled data. Based on this observation, author proposes DisagreeNet which composes of the three steps : (1)...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper observes that the dynamics of the ensemble model provide a distinction between clean and noisy labeled data in the training phase. The diversity of noisy labeled data is much larger than that of clean labeled data. Based on this observation, author proposes DisagreeNet which composes of the three ste...
This paper proposes the Dual Graph Lottery Ticket framework to obtain the triple-win graph lottery ticket. Regularization-based network pruning and hierarchical graph sparsification are designed to jointly get the sparsified graph and subnetwork. What’s more, the graph information theory guarantees the explainability o...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes the Dual Graph Lottery Ticket framework to obtain the triple-win graph lottery ticket. Regularization-based network pruning and hierarchical graph sparsification are designed to jointly get the sparsified graph and subnetwork. What’s more, the graph information theory guarantees the explaina...