Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper proposes a novel semantic image synthesis model (semantic segmentation map in, RGB image out) based on adversarial training. The main idea is to use edge maps as a structure to guide the image synthesis process. To this end a special edge-generating branch is used to generate an edge map from the semantic ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a novel semantic image synthesis model (semantic segmentation map in, RGB image out) based on adversarial training. The main idea is to use edge maps as a structure to guide the image synthesis process. To this end a special edge-generating branch is used to generate an edge map from the s...
This paper provides one benchmark named BiBench which is desifned to test the efficiency, robustness and generalization of binary algorithms. It mainly contains two parts, one for accuracy and another is for efficiency, and each part has several components. The authors build quantitative indicators to measure the binar...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper provides one benchmark named BiBench which is desifned to test the efficiency, robustness and generalization of binary algorithms. It mainly contains two parts, one for accuracy and another is for efficiency, and each part has several components. The authors build quantitative indicators to measure t...
The work introduces a novel OSLS (One-Shot Ligand Scoring) model for ligand scoring task to an unseen target based on a single context compound and its experimentally known activity to that target. This model is based on a Siamese-inspired neural network with transformer encoders. The OSLS achieves performance improvem...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The work introduces a novel OSLS (One-Shot Ligand Scoring) model for ligand scoring task to an unseen target based on a single context compound and its experimentally known activity to that target. This model is based on a Siamese-inspired neural network with transformer encoders. The OSLS achieves performance ...
The submission studies combinatorial optimization (CO) with cardinality constraints by neural networks. Different from the methods incorporating the constraints as penalties in to the objective function, the submission incorporates the constraints into the network architecture such that these constraints are guaranteed...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The submission studies combinatorial optimization (CO) with cardinality constraints by neural networks. Different from the methods incorporating the constraints as penalties in to the objective function, the submission incorporates the constraints into the network architecture such that these constraints are gu...
The paper proposed a novel non-uniform data-free quantization method. The main Idea (which is very interesting) is to find the best non-uniform quantization intervals for both weights and activations by minimizing the reconstruction error using the proposed quantization and de-quantization method. The paper provides e...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposed a novel non-uniform data-free quantization method. The main Idea (which is very interesting) is to find the best non-uniform quantization intervals for both weights and activations by minimizing the reconstruction error using the proposed quantization and de-quantization method. The paper pr...
The submission proposes a federated methodology to learn low-dimensional representations from a dataset distributed among several clients. Basically, the idea lacks novelty and is with limited contribution as it is a simple application of the MCR2 objective to a federated learning setting. The optimization is still usi...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The submission proposes a federated methodology to learn low-dimensional representations from a dataset distributed among several clients. Basically, the idea lacks novelty and is with limited contribution as it is a simple application of the MCR2 objective to a federated learning setting. The optimization is s...
This paper considers the hypothesis that the good empirical generalization of overparameterized models (neural networks in particular) may not be primarily due to an implicit bias of the optimizer (e.g. SGD), as proposed previously. Instead, it argues that the volume in parameter space of good solutions (generalizing ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper considers the hypothesis that the good empirical generalization of overparameterized models (neural networks in particular) may not be primarily due to an implicit bias of the optimizer (e.g. SGD), as proposed previously. Instead, it argues that the volume in parameter space of good solutions (gener...
This paper proposes a new graph transformer architecture based on quantum computation. The main idea is to replace the typical attention matrix by a “quantum computable” attention matrix. The main motivation for such a proposal is easily getting access to graph features (such as maximum probability assignment from an a...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new graph transformer architecture based on quantum computation. The main idea is to replace the typical attention matrix by a “quantum computable” attention matrix. The main motivation for such a proposal is easily getting access to graph features (such as maximum probability assignment f...
This paper investigates the role of position encodings in contrast to contextual encodings in BERT-based models. They illustrate two biases of these encodings: locality (a tendency to focus on nearby tokens) and symmetry (a tendency to focus at a particular distance from the target to the same degree whether the distan...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper investigates the role of position encodings in contrast to contextual encodings in BERT-based models. They illustrate two biases of these encodings: locality (a tendency to focus on nearby tokens) and symmetry (a tendency to focus at a particular distance from the target to the same degree whether th...
This paper studies fine-tuning vs freezing in the context of reinforcement learning. The authors argue that the layers that represent general-purpose features(mainly the early layers) could be frozen while those that represent task-specific features(mainly the last layers) should be fine-tuned to achieve higher perform...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies fine-tuning vs freezing in the context of reinforcement learning. The authors argue that the layers that represent general-purpose features(mainly the early layers) could be frozen while those that represent task-specific features(mainly the last layers) should be fine-tuned to achieve higher...
This work proposes to use reinforcement learning to design graph augmentation mechanism used in graph contrastive learning. The benefit of using RL is that the way of graph augmentation can be designed via a markov decision process where a long-term reward and the dependence between different parts of the graph are bet...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This work proposes to use reinforcement learning to design graph augmentation mechanism used in graph contrastive learning. The benefit of using RL is that the way of graph augmentation can be designed via a markov decision process where a long-term reward and the dependence between different parts of the graph...
The paper tries to detect and prevent overfitting by using the training loss and validation loss history. For detecting overfitting, the paper propose a strategy to collect a labelled dataset with binary labels: overfit and no overfit. The paper shows that a time-series classifier can be trained, and empirically showed...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper tries to detect and prevent overfitting by using the training loss and validation loss history. For detecting overfitting, the paper propose a strategy to collect a labelled dataset with binary labels: overfit and no overfit. The paper shows that a time-series classifier can be trained, and empiricall...
The paper proposes a joint (mixed) FL strategy that jointly trains using a combination of centralized and federated models without transferring any data across server or client and vice versa. # Pros: - The paper is well-written, easy to understand and follow along, however, I must admit that I did not pay an in dep...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes a joint (mixed) FL strategy that jointly trains using a combination of centralized and federated models without transferring any data across server or client and vice versa. # Pros: - The paper is well-written, easy to understand and follow along, however, I must admit that I did not pay a...
This paper provides certification guarantees on robustness of gradient based explainers by providing upper bounds on the largest adversarial change that can be made to the explanation from these explainers given bounded manipulation of either the input features or model parameters. Additionally, these bounds are differ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides certification guarantees on robustness of gradient based explainers by providing upper bounds on the largest adversarial change that can be made to the explanation from these explainers given bounded manipulation of either the input features or model parameters. Additionally, these bounds ar...
The authors propose an approach to the problem of transferring learned behaviors between tasks in a sequential setting, called Attentive Priors for Expressive and Transferable Skills (APES). They first present two theorems: a) with more input variables, the shift of the input distributions in terms of KL divergence get...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose an approach to the problem of transferring learned behaviors between tasks in a sequential setting, called Attentive Priors for Expressive and Transferable Skills (APES). They first present two theorems: a) with more input variables, the shift of the input distributions in terms of KL diverg...
This work introduces a new model for agents to perform vision and language task. The method, named InstructRL relies on a vision and language transformer (ViLT) trained on passive datasets with language and images aligned with description. This ViLT is used to convert observations and descriptions into tokens, which ar...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work introduces a new model for agents to perform vision and language task. The method, named InstructRL relies on a vision and language transformer (ViLT) trained on passive datasets with language and images aligned with description. This ViLT is used to convert observations and descriptions into tokens, ...
The authors tackle the problem of model uncertainty estimation as the acquisition function in Active Learning for object detection. To this end, they first adapt Evidential Deep Learning (EDL) to predict the epistemic uncertainty estimate for each bounding box separately. They modify EDL in three ways: (a) replace the...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The authors tackle the problem of model uncertainty estimation as the acquisition function in Active Learning for object detection. To this end, they first adapt Evidential Deep Learning (EDL) to predict the epistemic uncertainty estimate for each bounding box separately. They modify EDL in three ways: (a) rep...
This paper presents Hetero-SSFL, a framework for SSL in FL that enables diverse clients (heterogeneous in terms of their compute capabilities and/or data distribution) to collaboratively learn a good SSL model. At the end of N rounds, the SSL model gets personalised (via linear evaluation protocol) to each client's dat...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper presents Hetero-SSFL, a framework for SSL in FL that enables diverse clients (heterogeneous in terms of their compute capabilities and/or data distribution) to collaboratively learn a good SSL model. At the end of N rounds, the SSL model gets personalised (via linear evaluation protocol) to each clie...
The authors of this submission proposed a deep non-stationary kernel (DNSK) modeling Spatio-Temporal Point Processes (STPP) for potentially non-stationary events in continuous time and space. The authors focused on Hawkes process assuming that the influences from the past events are linearly additive and in turn mode...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The authors of this submission proposed a deep non-stationary kernel (DNSK) modeling Spatio-Temporal Point Processes (STPP) for potentially non-stationary events in continuous time and space. The authors focused on Hawkes process assuming that the influences from the past events are linearly additive and in t...
To answer a pre-posed question about the contrastive views of graph representation learning, this paper studies task-oriented contrastive learning from a counterfactual perspective in node property prediction tasks. A model-agnostic framework, G-Censor was proposed to generate both positive and negative views in graph ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: To answer a pre-posed question about the contrastive views of graph representation learning, this paper studies task-oriented contrastive learning from a counterfactual perspective in node property prediction tasks. A model-agnostic framework, G-Censor was proposed to generate both positive and negative views i...
The paper studies the problem of agnostically learning a ReLU function via gradient descent using samples drawn from Gaussian (or even a more general class of) distributions. The main result states that under a suitable random initialization, if gradient descent is run for sufficiently many iterations on the populatio...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies the problem of agnostically learning a ReLU function via gradient descent using samples drawn from Gaussian (or even a more general class of) distributions. The main result states that under a suitable random initialization, if gradient descent is run for sufficiently many iterations on the p...
The paper proposes to treat clean and adversarial samples as different domains and co-train them by separate classification tokens of a VIT, which produces adversarial model soup to trade off between clean and robust accuracy by simple interpolation. The authors present a good alternative for advprop with fewer paramet...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes to treat clean and adversarial samples as different domains and co-train them by separate classification tokens of a VIT, which produces adversarial model soup to trade off between clean and robust accuracy by simple interpolation. The authors present a good alternative for advprop with fewer...
Overall, the paper provides a novel and intuitive understanding about adversarial trained models. The paper identifies a new phenomenon and points out its unrealized threats which will inspire new related research. Corresponding index (Gini coeffiencient) is designed to illustrate such phenomenon. Strength: 1. The str...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Overall, the paper provides a novel and intuitive understanding about adversarial trained models. The paper identifies a new phenomenon and points out its unrealized threats which will inspire new related research. Corresponding index (Gini coeffiencient) is designed to illustrate such phenomenon. Strength: 1....
This paper proves under some particular situation, e.g. the input vector is the concatenation of a 1-sparse feature vector and a noise vector, under certain assumptions, the two layer CNN (whose second layer is assumed to be fixed as all 1’s), converges to “bad” local optima with high probability when trained using ADA...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves under some particular situation, e.g. the input vector is the concatenation of a 1-sparse feature vector and a noise vector, under certain assumptions, the two layer CNN (whose second layer is assumed to be fixed as all 1’s), converges to “bad” local optima with high probability when trained u...
This work introduces a framework for the acoustic modeling of musical instruments. The authors present an open-source and open-access framework for the generation of numerical muscial acoustics. I am not a person in acoustic areas. Thus, everything in this area should be new and exciting to me. However, from this desc...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work introduces a framework for the acoustic modeling of musical instruments. The authors present an open-source and open-access framework for the generation of numerical muscial acoustics. I am not a person in acoustic areas. Thus, everything in this area should be new and exciting to me. However, from t...
This paper focuses on continual image-text embedding, reducing the computation and storage resources using all data in the training phase. For the first time, the paper identifies the important role of direct parameter transfer (between the historical and main models) in continual learning and the proposed DHA outperfo...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper focuses on continual image-text embedding, reducing the computation and storage resources using all data in the training phase. For the first time, the paper identifies the important role of direct parameter transfer (between the historical and main models) in continual learning and the proposed DHA ...
The authors empirically study how combining replay ratio increases with network parameter resets can greatly improve sample efficiency of off-policy methods. Their experiments demonstrate superior performance on existing benchmarks with simple modifications to existing methods. ## Strengths Clarity: The paper is w...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors empirically study how combining replay ratio increases with network parameter resets can greatly improve sample efficiency of off-policy methods. Their experiments demonstrate superior performance on existing benchmarks with simple modifications to existing methods. ## Strengths Clarity: The pa...
This paper studies the question of whether pre-trained model-based RL can be effectively fine-tuned to solve downstream tasks. To answer this question, during pre-training, this paper proposed to deploy a teacher-student framework: learn several sing-task teacher models via offline RL and distill these single-task poli...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the question of whether pre-trained model-based RL can be effectively fine-tuned to solve downstream tasks. To answer this question, during pre-training, this paper proposed to deploy a teacher-student framework: learn several sing-task teacher models via offline RL and distill these single-t...
This paper proposes to combine graph structure with transformer to model multivariate time series, and use probabilistic modeling method and variational inference to remove random noise in time series. The experiments on the task of anomaly detection and time series forecasting show the effectiveness of the proposed me...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes to combine graph structure with transformer to model multivariate time series, and use probabilistic modeling method and variational inference to remove random noise in time series. The experiments on the task of anomaly detection and time series forecasting show the effectiveness of the pro...
This paper proposes to regard the system identification problem as a reinforcement learning problem. Compared with traditional supervised solutions, the proposed reinforcement method has advantages over delayed effects, high non-linearity, non-stationarity, partial observability, and error accumulation when using boots...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes to regard the system identification problem as a reinforcement learning problem. Compared with traditional supervised solutions, the proposed reinforcement method has advantages over delayed effects, high non-linearity, non-stationarity, partial observability, and error accumulation when usi...
The paper proposes a method that learns deep prototype representations, by revisiting the nearest centroids clustering in the setting of deep network training. The authors construct a framework to combine the gradient descent kind of learning of neural networks with the distance-based learning of cluster centroids. In ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a method that learns deep prototype representations, by revisiting the nearest centroids clustering in the setting of deep network training. The authors construct a framework to combine the gradient descent kind of learning of neural networks with the distance-based learning of cluster centro...
This paper proposes a transformer model for generating repair for software vulnerabilities. It trains a separate token-level classifier model to generate masks called vulnerability queries and incorporates these masks in both the encoder and decoder parts. Through experimental evaluation on Big-Vul and CVEFixes, it sho...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a transformer model for generating repair for software vulnerabilities. It trains a separate token-level classifier model to generate masks called vulnerability queries and incorporates these masks in both the encoder and decoder parts. Through experimental evaluation on Big-Vul and CVEFixes...
In this paper, the authors consider the problem of finding dynamics that not only converge to Nash Equilibria but also converge to "good" equilibria according to some objective function, e.g., the social-welfare. The main results are the following: 1. The social welfare of the 0-replicator dynamics is higher than the...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this paper, the authors consider the problem of finding dynamics that not only converge to Nash Equilibria but also converge to "good" equilibria according to some objective function, e.g., the social-welfare. The main results are the following: 1. The social welfare of the 0-replicator dynamics is higher ...
This paper tries to tackle the problem of data heterogeneity in federated learning with differential privacy by attempting to divide the features into private and generalizable features. They ask the question of what is necessary to share to learn global models and can the remaining data stay on the client as local mod...	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 tries to tackle the problem of data heterogeneity in federated learning with differential privacy by attempting to divide the features into private and generalizable features. They ask the question of what is necessary to share to learn global models and can the remaining data stay on the client as l...
This paper addresses the class-imbalance problem in graph learning. The authors propose Graph Decantation (GraphDec) framework for learning balanced graph representation in a self-supervised manner. The key procedure of GraphDec is to select some informative samples from unbalanced node data or graph data during traini...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper addresses the class-imbalance problem in graph learning. The authors propose Graph Decantation (GraphDec) framework for learning balanced graph representation in a self-supervised manner. The key procedure of GraphDec is to select some informative samples from unbalanced node data or graph data durin...
The paper explores methods for discovering diverse policies for RL problems, specifically focusing on MARL. This is well-motivated, as these types of methods have become increasingly prominent in recent years and played a part of large scale successes such as AlphaStar. The paper then disappointingly focuses on a new a...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper explores methods for discovering diverse policies for RL problems, specifically focusing on MARL. This is well-motivated, as these types of methods have become increasingly prominent in recent years and played a part of large scale successes such as AlphaStar. The paper then disappointingly focuses on...
This paper proposed a new method for model editing in pre-trained language models, which has received growing attention from the deep learning community in the past few years. The authors point out that the setting of sequential model editing (in which the model must be edited with a stream of edits, rather than a sm...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a new method for model editing in pre-trained language models, which has received growing attention from the deep learning community in the past few years. The authors point out that the setting of sequential model editing (in which the model must be edited with a stream of edits, rather t...
The paper studies the learning-based low-rank approximation problem and proposes a new algorithm, which is based on the tucker tensor decomposition. The new optimization problem based on tensor decomposition can be seen as a relaxation of the original LRA problem. The experimental result suggests that the proposed algo...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper studies the learning-based low-rank approximation problem and proposes a new algorithm, which is based on the tucker tensor decomposition. The new optimization problem based on tensor decomposition can be seen as a relaxation of the original LRA problem. The experimental result suggests that the propo...
The overall goal of the submission is a learning formulation that simultaneously models inter-class discrimination while promoting intra class variation in classification. To this end, it considers a linear subspace approach that is scaleless and thus in theory more suitable for long tail classification than the vector...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The overall goal of the submission is a learning formulation that simultaneously models inter-class discrimination while promoting intra class variation in classification. To this end, it considers a linear subspace approach that is scaleless and thus in theory more suitable for long tail classification than th...
This work focuses on graph model selection without training by matching the meta-feature and model representations. Extensive experiments are conducted to demonstrate the effective and efficiency of the proposed method. Basically, the overall idea has close relation to pioneering work MetaOD. By comparison, this work a...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work focuses on graph model selection without training by matching the meta-feature and model representations. Extensive experiments are conducted to demonstrate the effective and efficiency of the proposed method. Basically, the overall idea has close relation to pioneering work MetaOD. By comparison, thi...