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 PointDP, an adversarial purification method that leverages a diffusion model as a pre-processing module to defend against 3D adversary attacks for point clouds. PointDP consists of two components (1) an off-the-shelf 3D point cloud diffusion model and (2) a classifier. Given an input point cloud, 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 proposes PointDP, an adversarial purification method that leverages a diffusion model as a pre-processing module to defend against 3D adversary attacks for point clouds. PointDP consists of two components (1) an off-the-shelf 3D point cloud diffusion model and (2) a classifier. Given an input point ...
This paper proposes MEGA, a method for combining exponential moving average (EMA) with attention. The paper evaluates MEGA on a number of benchmark datasets for sequence modeling and finds positive results. + Improvement over S4 on a number of tasks, including language modeling. + Extensive experiments and hyperparamet...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes MEGA, a method for combining exponential moving average (EMA) with attention. The paper evaluates MEGA on a number of benchmark datasets for sequence modeling and finds positive results. + Improvement over S4 on a number of tasks, including language modeling. + Extensive experiments and hype...
This paper studies safety in reinforcement learning. The goal is to maximize return while keeping the cost of visiting unsafe state-action pairs below a threshold. The paper proposes learning a classifier $F_\theta(s, a)$ that predicts whether a given state action $(s, a)$ will lead to eventually visiting an unsafe sta...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies safety in reinforcement learning. The goal is to maximize return while keeping the cost of visiting unsafe state-action pairs below a threshold. The paper proposes learning a classifier $F_\theta(s, a)$ that predicts whether a given state action $(s, a)$ will lead to eventually visiting an un...
The paper studies the problem of aggregating clusterings of a set of data points. This is a well-studied problem in we are given a collection of clusterings of a given dataset and the goal is to produce a new clustering that is "close" to the given ones. The problem is non-trivial because clusterings remain the same e...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies the problem of aggregating clusterings of a set of data points. This is a well-studied problem in we are given a collection of clusterings of a given dataset and the goal is to produce a new clustering that is "close" to the given ones. The problem is non-trivial because clusterings remain th...
This paper proposes a method that evaluates counterfactuals. The method does not focus on the algorithm (or models) for the inference of counterfactuals but on properties the inferred outputs must obey. Based on Pearl’s axiomatic definition of counterfactuals, the authors propose to check three properties: Effectivene...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes a method that evaluates counterfactuals. The method does not focus on the algorithm (or models) for the inference of counterfactuals but on properties the inferred outputs must obey. Based on Pearl’s axiomatic definition of counterfactuals, the authors propose to check three properties: Eff...
The paper proposes an optimization-based framework with influence functions for subset selection (or dataset pruning). The authors show in certain cases that the influence based method can be used to prune large datasets with only a small increase in the generalization error. From a practical point of view, this framew...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes an optimization-based framework with influence functions for subset selection (or dataset pruning). The authors show in certain cases that the influence based method can be used to prune large datasets with only a small increase in the generalization error. From a practical point of view, thi...
This paper tackles the problem of 'Unified Open-Set Recognition' (UOSR), a variant of open-set recognition (OSR) which also considers how to deal with misclassified closed-set test samples. Concretely, the closed-set test samples are split into InW and InC (wrongly and correctly classified samples respectively), and th...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper tackles the problem of 'Unified Open-Set Recognition' (UOSR), a variant of open-set recognition (OSR) which also considers how to deal with misclassified closed-set test samples. Concretely, the closed-set test samples are split into InW and InC (wrongly and correctly classified samples respectively)...
In this work, the authors take a finite-time analysis approach to quantify the impact of approximation errors on the learning performance ofWarm-Start A-C method with a given prior policy. By delving into the intricate coupling between the updates of the Actor and the Critic, the paper first provides upper bounds on th...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In this work, the authors take a finite-time analysis approach to quantify the impact of approximation errors on the learning performance ofWarm-Start A-C method with a given prior policy. By delving into the intricate coupling between the updates of the Actor and the Critic, the paper first provides upper boun...
Summary: - Existing large language models (LLMs) require manual exemplars to acquire procedural planning knowledge in the zero-shot setting. - The paper proposed a neuro-symbolic procedural PLANner (PLAN) with commonsense-infused prompts elicited from an external knowledge base (i.e., ConceptNet) to solve the pure-la...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Summary: - Existing large language models (LLMs) require manual exemplars to acquire procedural planning knowledge in the zero-shot setting. - The paper proposed a neuro-symbolic procedural PLANner (PLAN) with commonsense-infused prompts elicited from an external knowledge base (i.e., ConceptNet) to solve the...
The authors proposed Graphsensor, a graph attention network that generates signal segments and learns their internal relationships through a multi-head approach. Their model has 56% of the model parameters compared with the best multi-head baseline while producing a 13.8% improvement in accuracy compared with the best ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors proposed Graphsensor, a graph attention network that generates signal segments and learns their internal relationships through a multi-head approach. Their model has 56% of the model parameters compared with the best multi-head baseline while producing a 13.8% improvement in accuracy compared with t...
Long-tail distribution of items is an important factor which affects the user’s experience in industrial search systems, recommendation systems, and so on. This submission tries to address the training-inference inconsistency due to the long-tail distribution. Concretely, the authors comprise a relatively large-scale o...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Long-tail distribution of items is an important factor which affects the user’s experience in industrial search systems, recommendation systems, and so on. This submission tries to address the training-inference inconsistency due to the long-tail distribution. Concretely, the authors comprise a relatively large...
In this paper, the authors investigate the problem of self-supervised pre-training on compact networks. They study the impact of view sampling strategy and design a strategy which can boost the SSL pretraining performance without knowledge distillation. The developed method consistently shows improvement on different c...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: In this paper, the authors investigate the problem of self-supervised pre-training on compact networks. They study the impact of view sampling strategy and design a strategy which can boost the SSL pretraining performance without knowledge distillation. The developed method consistently shows improvement on dif...
This paper proposes a method combining autoencoders and Bi-LSTM layers to predict the remaining useful life of physical equipments. The author argues that besides promising approaches, traditional ML methods cannot extract sufficient information from the equipment data and they need manual specialist inputs to obtain c...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes a method combining autoencoders and Bi-LSTM layers to predict the remaining useful life of physical equipments. The author argues that besides promising approaches, traditional ML methods cannot extract sufficient information from the equipment data and they need manual specialist inputs to ...
This paper aims to design a rationalization method based on causal inference. To achieve this goal, the authors first analyze the potential confounder underlined the input and output. Such latent confounder may influence the relation between the input and the rational as well as the rational and the output. To alleviat...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper aims to design a rationalization method based on causal inference. To achieve this goal, the authors first analyze the potential confounder underlined the input and output. Such latent confounder may influence the relation between the input and the rational as well as the rational and the output. To ...
This paper presents a method named Debiased Fully Test-time Adaptation (DELTA) to address the biased issue in test-time adaptation. Specifically, the authors conduct experiments to verify the claims that 1) the normalization statistics tend to fit the current test mini-batch, and 2) the test-time adaptation optimizatio...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper presents a method named Debiased Fully Test-time Adaptation (DELTA) to address the biased issue in test-time adaptation. Specifically, the authors conduct experiments to verify the claims that 1) the normalization statistics tend to fit the current test mini-batch, and 2) the test-time adaptation opt...
This paper proposes a method MaskFusion that can be incorporated into existing deep CTR models for enhancing prediction performance. This idea is to multiply input feature embeddings with mask vectors, and then concatenate them with each DNN layer in deep CTR models. The mask vectors are dependent on the input featu...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a method MaskFusion that can be incorporated into existing deep CTR models for enhancing prediction performance. This idea is to multiply input feature embeddings with mask vectors, and then concatenate them with each DNN layer in deep CTR models. The mask vectors are dependent on the inp...
This paper introduces a scalable and privacy-enhanced graph generative model to learn and reproduce the distribution of real-world graphs with node attributes/labels. The proposed model satisfies benchmark effectiveness, scalability and privacy guarantee. Strength - The proposed graph generative model fills the g...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces a scalable and privacy-enhanced graph generative model to learn and reproduce the distribution of real-world graphs with node attributes/labels. The proposed model satisfies benchmark effectiveness, scalability and privacy guarantee. Strength - The proposed graph generative model fil...
This paper proposes a benchmarking framework for class-out-of-distribution detection with various levels of detection difficulty. This work benchmarks this technique's application to ImageNet with 525 publicly available pre-trained ImageNet-1K classifiers. Based on this benchmark, the authors identify several trends in...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a benchmarking framework for class-out-of-distribution detection with various levels of detection difficulty. This work benchmarks this technique's application to ImageNet with 525 publicly available pre-trained ImageNet-1K classifiers. Based on this benchmark, the authors identify several t...
In this work, the authors propose to forecast irregular time series using state-space model whose dynamics is specified by a linear ODE. The proposed method meets several desirable properties to time series forecast such as self-consistency, forward stability and allowing for missing values. The experiments show the pr...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work, the authors propose to forecast irregular time series using state-space model whose dynamics is specified by a linear ODE. The proposed method meets several desirable properties to time series forecast such as self-consistency, forward stability and allowing for missing values. The experiments sho...
This paper studies constrained decentralized bilevel optimization where the inner-loop problem is finite-sum, smooth and strongly convex, and the outer-loop problem is finite-sum, smooth and nonconvex. By combing (a) gradient tracking, (b) variance reduction and (c) a pseudo-gradient trick based on proximal operator, ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies constrained decentralized bilevel optimization where the inner-loop problem is finite-sum, smooth and strongly convex, and the outer-loop problem is finite-sum, smooth and nonconvex. By combing (a) gradient tracking, (b) variance reduction and (c) a pseudo-gradient trick based on proximal op...
This paper proposes a method to learn predictors that are invariant under counterfactual changes of certain covariates. The proposed method adds a model-agnostic regularization term based on the notion of conditional independence and kernel mean embeddings, to enforce the so called "counterfactual invariance" during tr...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a method to learn predictors that are invariant under counterfactual changes of certain covariates. The proposed method adds a model-agnostic regularization term based on the notion of conditional independence and kernel mean embeddings, to enforce the so called "counterfactual invariance" d...
This paper considers the problem of fair classification by post-processing the decisions of an existing (possibly unfair) classifier. The authors focus on two metrics — equality of opportunity and equalized odds. The main idea is to set the decision threshold appropriately based on the protected group and true outcome...	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 considers the problem of fair classification by post-processing the decisions of an existing (possibly unfair) classifier. The authors focus on two metrics — equality of opportunity and equalized odds. The main idea is to set the decision threshold appropriately based on the protected group and true...
This paper extends existing transformer-based autoregressive indoor scene synthesis method by introducing a new conditioning mechanism that allow the input sequence to contain, at any location, an arbitrary number of masked tokens, whose values are to be predicted. Three set of positional encodings are introduced to en...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper extends existing transformer-based autoregressive indoor scene synthesis method by introducing a new conditioning mechanism that allow the input sequence to contain, at any location, an arbitrary number of masked tokens, whose values are to be predicted. Three set of positional encodings are introduc...
This paper proposed a novel hybrid design where an edge-device selectively queries the cloud only on those hard instances that the cloud can classify correctly to optimize accuracy under latency and edge-device constraints. An end-to-end method to train neural architectures, base predictors, and routing models is also ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a novel hybrid design where an edge-device selectively queries the cloud only on those hard instances that the cloud can classify correctly to optimize accuracy under latency and edge-device constraints. An end-to-end method to train neural architectures, base predictors, and routing models ...
This paper proposes a framework based on some existing works to incorporate gene regulatory networks (Garcia-Alonso et al., 2019), protein-protein interaction (Chereda et al., 2021) networks, protein-pathway networks (Elmarakeby et al., 2021), and relations among genes, proteins, and pathways for biological discoveries...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a framework based on some existing works to incorporate gene regulatory networks (Garcia-Alonso et al., 2019), protein-protein interaction (Chereda et al., 2021) networks, protein-pathway networks (Elmarakeby et al., 2021), and relations among genes, proteins, and pathways for biological dis...
In this paper, the authors proposed to use relaxed attention in both self attention and cross attention part of transformer models. In details, the authors proposed to add a smoothing term on the cross attention module with a fuzzy realaxation coefficient drawn from a Gaussian distribution. Extensive experiments were d...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors proposed to use relaxed attention in both self attention and cross attention part of transformer models. In details, the authors proposed to add a smoothing term on the cross attention module with a fuzzy realaxation coefficient drawn from a Gaussian distribution. Extensive experiment...
This paper raises a gap in current OOD detection approaches that rely too heavily on the training examples, and falsely mark images out of the train set but with similar semantic meaning. The authors coin the term "intended distribution" to capture both the training set, but also the rest of the "semantically similar...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper raises a gap in current OOD detection approaches that rely too heavily on the training examples, and falsely mark images out of the train set but with similar semantic meaning. The authors coin the term "intended distribution" to capture both the training set, but also the rest of the "semantically...
Studies link prediction for hyper-relational knowledge graphs (in which edges are associated with have attribute/entity pairs). Transforms the hyper-relational KG into a standard KG using reification, then runs a relational GNN for link prediction. Strengths: S1. Simple approach (but see W1). S2. Good experimental r...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Studies link prediction for hyper-relational knowledge graphs (in which edges are associated with have attribute/entity pairs). Transforms the hyper-relational KG into a standard KG using reification, then runs a relational GNN for link prediction. Strengths: S1. Simple approach (but see W1). S2. Good experi...
This paper studies the problem of how overparametrization affects the ERM's performance on minority groups theoretically. To be more specific, this paper theoretically shows that under the overparametrazation condition and ERM algorithm, the minority groups will tend to have worse performance on the regression tasks, a...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the problem of how overparametrization affects the ERM's performance on minority groups theoretically. To be more specific, this paper theoretically shows that under the overparametrazation condition and ERM algorithm, the minority groups will tend to have worse performance on the regression ...
This paper proposes to model the evolution of latent variables in continuous time models as the solution of first-order partial differential equations. The authors builds upon the method of characteristics for solving PDEs and show that their approach provides both theoretical and practical advantages. In particular, t...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to model the evolution of latent variables in continuous time models as the solution of first-order partial differential equations. The authors builds upon the method of characteristics for solving PDEs and show that their approach provides both theoretical and practical advantages. In parti...
This paper proposes instance-reweighted adversarial training (IRAT) which weighs the instances based on the suggested standard in the adversarial loss term. This paper suggests the margin-based standard which weighs more when the instance is near the decision boundary. While the previous margin-based instance-reweighte...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes instance-reweighted adversarial training (IRAT) which weighs the instances based on the suggested standard in the adversarial loss term. This paper suggests the margin-based standard which weighs more when the instance is near the decision boundary. While the previous margin-based instance-r...
In this paper, the authors propose a new approach for solving PDEs using non-equispaced Fourier Neural Operators. The authors draw some analogs between vision mixer models and FNO and then build over it to obtain the non-equispaced FNO. For the non-equispaced FNO, the authors use a non-equispaced interpolation function...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: In this paper, the authors propose a new approach for solving PDEs using non-equispaced Fourier Neural Operators. The authors draw some analogs between vision mixer models and FNO and then build over it to obtain the non-equispaced FNO. For the non-equispaced FNO, the authors use a non-equispaced interpolation ...
Aiming at the problems of the aggregations of target deviated neighbors on dynamic graphs (the authors call it topology-task discordance), this paper revisits node-wise relationships based on a dynamic homophily theory from the spatial-temporal perspectives, and proposes a model GReTo consisting of two stages, signed t...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Aiming at the problems of the aggregations of target deviated neighbors on dynamic graphs (the authors call it topology-task discordance), this paper revisits node-wise relationships based on a dynamic homophily theory from the spatial-temporal perspectives, and proposes a model GReTo consisting of two stages, ...
When the distribution gap between the source domain and the target domain is very large, some intermediate domains are usually used to make the model gradually adapt to the target domain. The author proposes the GOTA method to generate intermediate domains based on Wasserstein geodesic. Empirically, this GOAT framework...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: When the distribution gap between the source domain and the target domain is very large, some intermediate domains are usually used to make the model gradually adapt to the target domain. The author proposes the GOTA method to generate intermediate domains based on Wasserstein geodesic. Empirically, this GOAT f...
This work propose a simple and novel self-supervised learning framework for time series representation learning. Instead of using contrastive representation learning, this work directly follow the masked data modeling and optimize the reconstruction loss of the randomly masked data points. Empirical studies on real-wor...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work propose a simple and novel self-supervised learning framework for time series representation learning. Instead of using contrastive representation learning, this work directly follow the masked data modeling and optimize the reconstruction loss of the randomly masked data points. Empirical studies on ...
The paper proposes a method to model dynamical systems in a latent space in a smooth way. To go back and forth from the latent space, a decoder (called "ansatz") and an encoder network are used, that are task dependent. One first trick proposed in the paper is to propose some cyclic consistency for this encoder decoder...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper proposes a method to model dynamical systems in a latent space in a smooth way. To go back and forth from the latent space, a decoder (called "ansatz") and an encoder network are used, that are task dependent. One first trick proposed in the paper is to propose some cyclic consistency for this encoder...
This paper proposed two improvements over the baseline model PESNet. The first one, PlaNet, use the local energy computed in each PESNET training step to train a surrogate model, which takes in the geometry of nucleus and output final energy, namely no dependence on electrons. At inference time, only the surrogate mode...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposed two improvements over the baseline model PESNet. The first one, PlaNet, use the local energy computed in each PESNET training step to train a surrogate model, which takes in the geometry of nucleus and output final energy, namely no dependence on electrons. At inference time, only the surrog...
The paper studies transfer learning in multi-agent setting, where existing works based on graph neural networks or attention mechanisms achieves generalization across tasks implicitly through the generalization of neural network function approximations. The authors propose to explicitly model the task relationships by ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies transfer learning in multi-agent setting, where existing works based on graph neural networks or attention mechanisms achieves generalization across tasks implicitly through the generalization of neural network function approximations. The authors propose to explicitly model the task relations...
This work proves regret bounds for multi-arm bandit under distributed pure DP and RDP only using discrete noise. The theoretical results are verified by experiments. Strength - This work fills in the gap for distributed pure DP in previous works and the result is tight (matches the central DP lower bound) - The RDP upp...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work proves regret bounds for multi-arm bandit under distributed pure DP and RDP only using discrete noise. The theoretical results are verified by experiments. Strength - This work fills in the gap for distributed pure DP in previous works and the result is tight (matches the central DP lower bound) - The...
The authors study how to learn an effective global model on private and decentralized datasets. Toward this end, the authors propose to use consensus graph to fuse feature representations for a final prediction. The experimental results on four real-life datasets are used to validate the effectiveness of the proposed f...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The authors study how to learn an effective global model on private and decentralized datasets. Toward this end, the authors propose to use consensus graph to fuse feature representations for a final prediction. The experimental results on four real-life datasets are used to validate the effectiveness of the pr...