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 introduces M&Ms framework that aims to generate controllable scene generation with given visual descriptions. The proposed framework takes the background scene and set of objects along with the object location and size as conditions and synthesis the photorealistic collages. In addition to realistic image ge...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper introduces M&Ms framework that aims to generate controllable scene generation with given visual descriptions. The proposed framework takes the background scene and set of objects along with the object location and size as conditions and synthesis the photorealistic collages. In addition to realistic ...
The paper proposes neural constrain inference (NCI), a dynamics model which predicts future trajectories by inferring interactions between particles represented as sets of energy constraints, and sampling from the resulting energy-based model (EBM) using Langevin dynamics. It is demonstrated on various dynamics task th...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper proposes neural constrain inference (NCI), a dynamics model which predicts future trajectories by inferring interactions between particles represented as sets of energy constraints, and sampling from the resulting energy-based model (EBM) using Langevin dynamics. It is demonstrated on various dynamics...
This paper proposes a conditional selective inference (SI) approach for obtaining p-values for saliency maps. By leveraging the piecewise linearity of ReLU networks, they develop a method for performing the necessary hypothesis testing in an automated fashion across arbitrary network architectures. In experiments, thei...	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 proposes a conditional selective inference (SI) approach for obtaining p-values for saliency maps. By leveraging the piecewise linearity of ReLU networks, they develop a method for performing the necessary hypothesis testing in an automated fashion across arbitrary network architectures. In experimen...
This paper proposes a CroMA framework to incorporate multiple sensor modalities and close the domain gap between training and deployment for self-driving. They utilize LiDAR sensor during the training phase with knowledge distillation paradigm to enhance the camera-only model. Adversarial learning is adopted to address...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a CroMA framework to incorporate multiple sensor modalities and close the domain gap between training and deployment for self-driving. They utilize LiDAR sensor during the training phase with knowledge distillation paradigm to enhance the camera-only model. Adversarial learning is adopted to...
This paper presents a run-time channel squeezing technique. The proposed technique mainly works on the convolutional block and can serve as a plug and play component to many CNNs. The proposed methods does not keeps all possible parameters but can reduce the number of FLOPs to enable fast inference on low-power devices...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a run-time channel squeezing technique. The proposed technique mainly works on the convolutional block and can serve as a plug and play component to many CNNs. The proposed methods does not keeps all possible parameters but can reduce the number of FLOPs to enable fast inference on low-power...
This manuscript presents a new stochastic embedding method named Maximum Entropy Information Bottleneck (MEIB). Authors show that the well-known VIB is an upper bound of MEIB. Experimental results show that MEIB performs more robustly against adversarial attacks or other types of perturbation than VIB. The proposed met...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This manuscript presents a new stochastic embedding method named Maximum Entropy Information Bottleneck (MEIB). Authors show that the well-known VIB is an upper bound of MEIB. Experimental results show that MEIB performs more robustly against adversarial attacks or other types of perturbation than VIB. The prop...
The authors of this paper introduce a new approach to safe (or constrained) reinforcement learning (RL) that leverages a concept called scenario-based programming. In essence, this programming paradigm is a means to inject domain knowledge to the RL loop, by running such a (set of) program(s) in parallel with the RL ag...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors of this paper introduce a new approach to safe (or constrained) reinforcement learning (RL) that leverages a concept called scenario-based programming. In essence, this programming paradigm is a means to inject domain knowledge to the RL loop, by running such a (set of) program(s) in parallel with t...
This paper proposes Physics Augmented Continuum Neural Radiance Fields (PAC-NeRF), to estimate both the unknown geometry and physical parameters of highly dynamic objects from multi-view videos. The core idea of this paper is a hybrid Eulerian-Lagrangian representation of the neural radiance field. The proposed method ...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes Physics Augmented Continuum Neural Radiance Fields (PAC-NeRF), to estimate both the unknown geometry and physical parameters of highly dynamic objects from multi-view videos. The core idea of this paper is a hybrid Eulerian-Lagrangian representation of the neural radiance field. The proposed...
This work proposes a graph gating method for advancing the capacity of overcoming over-smoothing problem of GNN models. Theoretical proof and analysis are provided to demonstrate its superiority on dealing with over-smoothing issue with the help of ordinal differential equation (ODE). Experimental results look impressi...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes a graph gating method for advancing the capacity of overcoming over-smoothing problem of GNN models. Theoretical proof and analysis are provided to demonstrate its superiority on dealing with over-smoothing issue with the help of ordinal differential equation (ODE). Experimental results look ...
This paper proposes a new attack method. Unlike previous methods that perturbs the training data, the proposed method only perturb the label of a selected subset of training data. This essentially finds a small population of data and train the model to misclassify them consistently. The benefit is that at inference sta...	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 proposes a new attack method. Unlike previous methods that perturbs the training data, the proposed method only perturb the label of a selected subset of training data. This essentially finds a small population of data and train the model to misclassify them consistently. The benefit is that at infer...
This submission considers neural networks trained with completely random labels with or without data augmentations. The main observation is that when data augmentations are enabled, neural networks learn meaningful representation of data. This is supported by fitting a k-NN classifier on learned representations using g...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This submission considers neural networks trained with completely random labels with or without data augmentations. The main observation is that when data augmentations are enabled, neural networks learn meaningful representation of data. This is supported by fitting a k-NN classifier on learned representations...
This paper studies well-known Adam algorithm. Previous work shows that Adam can converge to a neighborhood of stationary point if the gradients are bounded (functions are smooth). In this paper, the authors show that, Adam can guarantees a even better results under a milder condition, i.e., L_0,L_1 smoothness, under wh...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies well-known Adam algorithm. Previous work shows that Adam can converge to a neighborhood of stationary point if the gradients are bounded (functions are smooth). In this paper, the authors show that, Adam can guarantees a even better results under a milder condition, i.e., L_0,L_1 smoothness, ...
The paper provides a differentiable method for learning temporal knowledge graphs in KGs. While differentiable rule-learning in the context of KG’s already exists, and rule-learning in the context of temporal KG’s already exists, this paper is the first to provide a differentiable rule learning method for temporal KGs....	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper provides a differentiable method for learning temporal knowledge graphs in KGs. While differentiable rule-learning in the context of KG’s already exists, and rule-learning in the context of temporal KG’s already exists, this paper is the first to provide a differentiable rule learning method for tempo...
This paper presents a novel way to evaluate the quality of a pre-trained representation. It requires no real data and downstream tasks for the evaluation. In the evaluation, it utilizes synthetic data generated from the mixture of two Gaussian distributions. The authors prove the soundness of the method using the assum...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a novel way to evaluate the quality of a pre-trained representation. It requires no real data and downstream tasks for the evaluation. In the evaluation, it utilizes synthetic data generated from the mixture of two Gaussian distributions. The authors prove the soundness of the method using t...
In this paper, a new quantum enhanced GNN is proposed from the perspective of quantum mechanics. The paper regards the graph as a quantum system, transforms the graph topology into the interaction of a group of qubits, and uses the long-range correlation of the quantum system to calculate the relationship between diffe...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this paper, a new quantum enhanced GNN is proposed from the perspective of quantum mechanics. The paper regards the graph as a quantum system, transforms the graph topology into the interaction of a group of qubits, and uses the long-range correlation of the quantum system to calculate the relationship betwe...
This paper focuses on the problem of training dual encoding models on decentralized datasets, which has few existing works. Moreover, the authors consider a challenging scenario that each client possesses a small and non-IID dataset, where directly utilizing the existing centralized methods decreases efficacy. Hence, t...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper focuses on the problem of training dual encoding models on decentralized datasets, which has few existing works. Moreover, the authors consider a challenging scenario that each client possesses a small and non-IID dataset, where directly utilizing the existing centralized methods decreases efficacy. ...
This paper proposed a multi-agent interactive game that provides formal interpretability for classification problem. In particular, the interactive game is inspired by interactive proof systems. In addition, the authors connected information measures such as conditional entropy and average precision to completeness and...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposed a multi-agent interactive game that provides formal interpretability for classification problem. In particular, the interactive game is inspired by interactive proof systems. In addition, the authors connected information measures such as conditional entropy and average precision to complete...
For the possible repetition in Continual Learning, the paper makes contributes in the following three areas: (1) The author proposes two stochastic scenario generators that produce a wide range of CIR scenarios starting from a single dataset and a few control parameters. (2) The author conducts a comprehensive evaluati...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: For the possible repetition in Continual Learning, the paper makes contributes in the following three areas: (1) The author proposes two stochastic scenario generators that produce a wide range of CIR scenarios starting from a single dataset and a few control parameters. (2) The author conducts a comprehensive ...
This paper proposes a new deep learning model for long-term time series forecasting with distribution shifts based on Koopman operator theory. In comparison to existing work, the authors utilize measurement functions (e.g. sine functions) in latent space, and they formulate forecasting using global and temporally local...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a new deep learning model for long-term time series forecasting with distribution shifts based on Koopman operator theory. In comparison to existing work, the authors utilize measurement functions (e.g. sine functions) in latent space, and they formulate forecasting using global and temporal...
In this paper the authors propose an empirical analysis of neural kernel bandits, proposed as a UCB based alternative to Bayesian neural networks. Bayesian NNs have a high computational requirement as often an ensemble must be maintained in order to bootstrap a sample from the posterior, as in the case of Thompson samp...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: In this paper the authors propose an empirical analysis of neural kernel bandits, proposed as a UCB based alternative to Bayesian neural networks. Bayesian NNs have a high computational requirement as often an ensemble must be maintained in order to bootstrap a sample from the posterior, as in the case of Thomp...
The authors propose a contrastive learning scheme where not the feature representations of individual samples, but an aggregated feature representation of multiple samples (a set) is considered. The authors claim improvements over a SimCLR baseline on CIFAR, STL, ImageNet-100 and full ImageNet. Major Weaknesses -...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors propose a contrastive learning scheme where not the feature representations of individual samples, but an aggregated feature representation of multiple samples (a set) is considered. The authors claim improvements over a SimCLR baseline on CIFAR, STL, ImageNet-100 and full ImageNet. **Major Weaknes...
This paper introduces two variants of covariance-robust minimax probability machines (MPM) which the authors term as Bures MPM and Cramer-Rao MPM based on different statistical divergences between Gaussian distributions. The authors discuss how to estimate the parameter for these new variants of MPM and discuss the pot...	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 introduces two variants of covariance-robust minimax probability machines (MPM) which the authors term as Bures MPM and Cramer-Rao MPM based on different statistical divergences between Gaussian distributions. The authors discuss how to estimate the parameter for these new variants of MPM and discuss...
Motivated by membership inference attacks against differentially private stochastic gradient descent (DP-SGD), the paper bounds the total variation (TV) distance between two runs of DP-SGD on two neighboring data sets. DP-SGD amounts to subsampling and adding Gaussian noise to adaptively chosen vectors. The main result...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Motivated by membership inference attacks against differentially private stochastic gradient descent (DP-SGD), the paper bounds the total variation (TV) distance between two runs of DP-SGD on two neighboring data sets. DP-SGD amounts to subsampling and adding Gaussian noise to adaptively chosen vectors. The mai...
This work studies the how the accuracy of a fixed point changes when multiple models with increasing sizes are used for predictions which it defines as the learning profile of a point. They show that there can be points which are positively correlated with the average accuracy of models and there are certain points whi...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This work studies the how the accuracy of a fixed point changes when multiple models with increasing sizes are used for predictions which it defines as the learning profile of a point. They show that there can be points which are positively correlated with the average accuracy of models and there are certain po...
The paper studies learning with graphs under the covariate shift assumption. The predictor is decomposed into two functions; one is the feature extractor of the graph, f, and the other is the discriminator, g. The generalization performance of the predictor h(G)=g(f(G)) for a graph G is governed by the Lipschitz consta...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies learning with graphs under the covariate shift assumption. The predictor is decomposed into two functions; one is the feature extractor of the graph, f, and the other is the discriminator, g. The generalization performance of the predictor h(G)=g(f(G)) for a graph G is governed by the Lipschit...
The paper considers classification problems with small tabular datasets (a few thousand samples, and up to ten classes and 100 features). It presents a pre-trained model that solves such problems within a second, outperforming tree-ensembles and AutoML methods that run for an hour. The pre-trained model is a 12-layer t...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper considers classification problems with small tabular datasets (a few thousand samples, and up to ten classes and 100 features). It presents a pre-trained model that solves such problems within a second, outperforming tree-ensembles and AutoML methods that run for an hour. The pre-trained model is a 12...
The authors propose a way to manipulate intermediate latent features of pre-trained generative models to edit local parts of the image. Specifically, they decompose a tensor of sampled latent features into two tensors, in which one could be interpreted as the set of appearance features and the other as the set of salie...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The authors propose a way to manipulate intermediate latent features of pre-trained generative models to edit local parts of the image. Specifically, they decompose a tensor of sampled latent features into two tensors, in which one could be interpreted as the set of appearance features and the other as the set ...
The paper defines a notion of risk-averse equilibrium, develops some theory surrounding the notion, and runs some experiments illustrating that their notion achieves a better trade-off between utility and variance than other notions. Broadly speaking, risk aversion is usually implemented by passing the utility function...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper defines a notion of risk-averse equilibrium, develops some theory surrounding the notion, and runs some experiments illustrating that their notion achieves a better trade-off between utility and variance than other notions. Broadly speaking, risk aversion is usually implemented by passing the utility ...
The paper proposed a new hyper parameter optimization method using tensor completion algorithm. The new algorithm consider a N dimensional tensor with its dimension representing tunable parameters and values representing feedbacks (e.g. validation loss). A Tucker factorization is used to for tensor completion in this p...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposed a new hyper parameter optimization method using tensor completion algorithm. The new algorithm consider a N dimensional tensor with its dimension representing tunable parameters and values representing feedbacks (e.g. validation loss). A Tucker factorization is used to for tensor completion i...
I have reviewed this paper in another conference. This paper introduces a technique called Recursive Interpolation Method (RIM) for time series augmentation. The augmented samples preserve the original inherent time-series dynamics. Theoretical analysis is given to guarantee the test performance. Experiments are conduc...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: I have reviewed this paper in another conference. This paper introduces a technique called Recursive Interpolation Method (RIM) for time series augmentation. The augmented samples preserve the original inherent time-series dynamics. Theoretical analysis is given to guarantee the test performance. Experiments ar...
The paper addresses the problem of modelling a PDE in the means of neural networks. The main motivation stems from the fact that Clifford algebras is a very popular tool in modelling physical relationships. As a consequence, the paper proposes to use the product of the algebra (instead of the products of a scalar field...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper addresses the problem of modelling a PDE in the means of neural networks. The main motivation stems from the fact that Clifford algebras is a very popular tool in modelling physical relationships. As a consequence, the paper proposes to use the product of the algebra (instead of the products of a scal...
This paper proposes an efficient and general component GrafT for Transformer-based vision architectures. The GrafT can introduce multi-scale features for early-stage layers, thus achieving better results. The authors demonstrate the effectiveness of the proposed GrafT on several tasks and datasets. Strengths * The des...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an efficient and general component GrafT for Transformer-based vision architectures. The GrafT can introduce multi-scale features for early-stage layers, thus achieving better results. The authors demonstrate the effectiveness of the proposed GrafT on several tasks and datasets. Strengths *...
One existing problem in information extraction is that it consists of a wide range of tasks. Previous attempts to unify all IE tasks with one architecture are either complex or time-consuming to decode. To this end, this paper proposes a simple yet effective paradigm for unified IE by interpreting all IE tasks as span ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: One existing problem in information extraction is that it consists of a wide range of tasks. Previous attempts to unify all IE tasks with one architecture are either complex or time-consuming to decode. To this end, this paper proposes a simple yet effective paradigm for unified IE by interpreting all IE tasks ...
This paper proposes a very neat idea of embedding contrastive learning into energy-based model learning. On top of the normal latent-variable energy-based models, the authors introduce augmentation and contrastive representation learning to regularize the latent space. The general idea is to still keep the empirical KL...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposes a very neat idea of embedding contrastive learning into energy-based model learning. On top of the normal latent-variable energy-based models, the authors introduce augmentation and contrastive representation learning to regularize the latent space. The general idea is to still keep the empi...
The paper aims to solve both graph data imbalance and unnecessary model-level computation burden in a unified framework. Specifically, the authors first examine the challenges from theoretical and empirical perspectives. Then, the authors propose GraphDec, a novel data-model dynamic sparsity framework to address the ch...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper aims to solve both graph data imbalance and unnecessary model-level computation burden in a unified framework. Specifically, the authors first examine the challenges from theoretical and empirical perspectives. Then, the authors propose GraphDec, a novel data-model dynamic sparsity framework to addres...
The paper presents an algorithm to estimate conditional quantiles for time-series predictions, and proposes a new parameterized Elliot activation function in LSTM gates. The conditional quantiles in predictions are used to identify anomalies. 1. There should be a set of experiments that compare the anomaly detection pe...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper presents an algorithm to estimate conditional quantiles for time-series predictions, and proposes a new parameterized Elliot activation function in LSTM gates. The conditional quantiles in predictions are used to identify anomalies. 1. There should be a set of experiments that compare the anomaly dete...
The paper introduces an agent architecture intended to facilitate the reuse of existing knowledge, called ‘knowledge grounded RL’. The knowledge grounding consists of a dictionary with randomly initialized learnable keys paired with hard-coded policies (these could possibly also be pre-trained, in principle) which is a...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces an agent architecture intended to facilitate the reuse of existing knowledge, called ‘knowledge grounded RL’. The knowledge grounding consists of a dictionary with randomly initialized learnable keys paired with hard-coded policies (these could possibly also be pre-trained, in principle) wh...
This paper studies computing Nash equilibria in two-player zero-sum Markov games, by proposing a single-loop and symmetric algorithm with last-iterate convergence guarantee. The derived iteration complexity improves over previous works by a factor of $H$. # Strength - The iteration complexity improves over the best pr...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies computing Nash equilibria in two-player zero-sum Markov games, by proposing a single-loop and symmetric algorithm with last-iterate convergence guarantee. The derived iteration complexity improves over previous works by a factor of $H$. # Strength - The iteration complexity improves over the...
This work reconsiders the problem of finding stationary points of finite-sum of $n$ smooth functions, where the optimal complexity for stochastic first-order methods is already known to be $\mathcal{O}(n + n^{1/2} \epsilon^{-1})$. The authors proposed to do a refined analysis since the $\mathcal{O}$ could hide dependen...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This work reconsiders the problem of finding stationary points of finite-sum of $n$ smooth functions, where the optimal complexity for stochastic first-order methods is already known to be $\mathcal{O}(n + n^{1/2} \epsilon^{-1})$. The authors proposed to do a refined analysis since the $\mathcal{O}$ could hide ...
The paper proposes a mechanism to keep long-tail events in episodic memory to improve RL performance in Zipfian-like environments. Through contrastive learning, observations' rarity score is estimated and stored in a familiarity buffer. After certain training epochs, only the top rarest events from the buffer transit t...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a mechanism to keep long-tail events in episodic memory to improve RL performance in Zipfian-like environments. Through contrastive learning, observations' rarity score is estimated and stored in a familiarity buffer. After certain training epochs, only the top rarest events from the buffer t...