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 deals with multi-modal joint training methods on deep neural networks. Related to this topic, several works have reported that the best uni-modal networks outperform the multi-modal networks even though the multi-modal networks receive more information. This work is focused on this topic, and the main contri...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper deals with multi-modal joint training methods on deep neural networks. Related to this topic, several works have reported that the best uni-modal networks outperform the multi-modal networks even though the multi-modal networks receive more information. This work is focused on this topic, and the mai...
The paper proposes an analysis of a solution for achieving compositional generalization on the grounded SCAN dataset, in particular on its compositional split "H", i.e. the novel adverb-verb combination setting. The model is required to generalize to creating commands for instructions such as `pull X while spinning`, b...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes an analysis of a solution for achieving compositional generalization on the grounded SCAN dataset, in particular on its compositional split "H", i.e. the novel adverb-verb combination setting. The model is required to generalize to creating commands for instructions such as `pull X while spin...
This paper studied an interesting problem where the time series forecasting model may fail to benefit from the extra information in a multivariate setting. Strengths 1. A novel problem is studied 2. The experiment design is comprehensive Weaknesses: 1. Some model settings issues are not clearly explained in the paper...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studied an interesting problem where the time series forecasting model may fail to benefit from the extra information in a multivariate setting. Strengths 1. A novel problem is studied 2. The experiment design is comprehensive Weaknesses: 1. Some model settings issues are not clearly explained in t...
The paper proposes a method for prompt learning in vision and language models that aims at producing a variety of templates that can be used by the text encoder to provide better classification weights in a few-shot scenario. In particular, the paper observes that learning a single prompt is not sufficient to represe...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes a method for prompt learning in vision and language models that aims at producing a variety of templates that can be used by the text encoder to provide better classification weights in a few-shot scenario. In particular, the paper observes that learning a single prompt is not sufficient to...
The paper proposes a neural vocoder based on improved GAN architecture and the use of a bigger training dataset. The system aims at a "universal" vocoding function that works on various unseen categories of signals, including different acoustic signatures and languages. The main claims are the proposed GAN-based appro...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a neural vocoder based on improved GAN architecture and the use of a bigger training dataset. The system aims at a "universal" vocoding function that works on various unseen categories of signals, including different acoustic signatures and languages. The main claims are the proposed GAN-bas...
This paper proposes an approach for topology-preserving representation learning (dimensionality reduction). The topological similarity between data points in original and latent spaces was obtained by minimizing the Representation Topology Divergence (RTD) between original data and latent representations. They demonstr...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes an approach for topology-preserving representation learning (dimensionality reduction). The topological similarity between data points in original and latent spaces was obtained by minimizing the Representation Topology Divergence (RTD) between original data and latent representations. They ...
The paper proposes Make-A-Video, an approach for generating short video clips conditioned on a given open domain text. The model does not need any text-video pairs and instead relies on learning dynamics from unlabeled videos and applying them to existing text to image systems. The approach also applies multiple levels...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: The paper proposes Make-A-Video, an approach for generating short video clips conditioned on a given open domain text. The model does not need any text-video pairs and instead relies on learning dynamics from unlabeled videos and applying them to existing text to image systems. The approach also applies multipl...
This paper presents a new method for learning deep closed-loop control policies from training on data generated by open-loop optimal control solvers. The paper proposes a resampling procedure that iteratively retrains the controller in states that it may not have seen in the initial training data. The approach was test...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents a new method for learning deep closed-loop control policies from training on data generated by open-loop optimal control solvers. The paper proposes a resampling procedure that iteratively retrains the controller in states that it may not have seen in the initial training data. The approach ...
This paper proposes a novel and efficient data augmentation method for knowledge distillation, by building on top of representation interpolation augmentation methods. Existing data augmentation methods can be divided as three types and have their own limitations as follows: 1) representation interpolation - the newly ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a novel and efficient data augmentation method for knowledge distillation, by building on top of representation interpolation augmentation methods. Existing data augmentation methods can be divided as three types and have their own limitations as follows: 1) representation interpolation - th...
The paper tackles the problem of representation learning with GANs. To do so, it introduces a novel adversarial objective that is composed of a distribution-matching objective and a clustering objective. In addition, they propose a novel regularization algorithm to ensure the smoothness of the discriminator. Compared t...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper tackles the problem of representation learning with GANs. To do so, it introduces a novel adversarial objective that is composed of a distribution-matching objective and a clustering objective. In addition, they propose a novel regularization algorithm to ensure the smoothness of the discriminator. Co...
The authors proposed a RNN with latent embedding that uses optimization at inference time to generate internal contextual signals allowing the agent to parse its temporal experience into discrete events and organize learning about them. They showed that the model trained on tasks sequentially using weight updates with ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The authors proposed a RNN with latent embedding that uses optimization at inference time to generate internal contextual signals allowing the agent to parse its temporal experience into discrete events and organize learning about them. They showed that the model trained on tasks sequentially using weight updat...
This paper proposes a new task of video highlight detection (VHD) in a domain-incremental setting. First, the authors define the task of incremental video highlights detection and introduce a gourmet dataset named LiveFood that they've carefully collected to facilitate research in this new task. Second, they argue that...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new task of video highlight detection (VHD) in a domain-incremental setting. First, the authors define the task of incremental video highlights detection and introduce a gourmet dataset named LiveFood that they've carefully collected to facilitate research in this new task. Second, they ar...
The authors propose the use of score-based generative modeling (i.e., diffusion models) for general tabular synthesis, and demonstrate the effectiveness this approach across a large number of tabular datasets and synthesizers (e.g., CTGAN, VEEGAN, TableGAN). Building on this approach, the authors introduce a novel sel...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The authors propose the use of score-based generative modeling (i.e., diffusion models) for general tabular synthesis, and demonstrate the effectiveness this approach across a large number of tabular datasets and synthesizers (e.g., CTGAN, VEEGAN, TableGAN). Building on this approach, the authors introduce a n...
The paper takles the problem of overfitting in the RL setting, where early stopping is not directly applicable due to the evolution of the dataset. The proposed method adjusts the update to data (UTD) ratio during training which trades off between underfitting and overfitting. The experiments are conducted using a mode...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper takles the problem of overfitting in the RL setting, where early stopping is not directly applicable due to the evolution of the dataset. The proposed method adjusts the update to data (UTD) ratio during training which trades off between underfitting and overfitting. The experiments are conducted usin...
This paper tackles the subgraph-level Federated Learning (FL), where each client has the individual subgraph of the larger global graph. Then, to tackle this task, the authors propose to reconstruct the neighborhood information of the subgraph based on the rooted tree structure. In particular, the rooted tree for the p...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper tackles the subgraph-level Federated Learning (FL), where each client has the individual subgraph of the larger global graph. Then, to tackle this task, the authors propose to reconstruct the neighborhood information of the subgraph based on the rooted tree structure. In particular, the rooted tree f...
Main concern. How to accommodate GNNs to deal naturally with heterophily and provide a theoretical analysis (from signal processing perspective). The core idea is to transform/combine low-pass filter into/with high-pass, for instance. There are two learnable parameters: w and \eta whose values lead to low-pass or high-...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Main concern. How to accommodate GNNs to deal naturally with heterophily and provide a theoretical analysis (from signal processing perspective). The core idea is to transform/combine low-pass filter into/with high-pass, for instance. There are two learnable parameters: w and \eta whose values lead to low-pass ...
In this work, the authors propose a new framework called Diamante that is aimed at improving the performance of open domain chatbots by incorporating human feedback (both explicit and implicit preferences). The proposed framework has a generation-evaluation training paradigm that is aimed to optimized response generat...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this work, the authors propose a new framework called Diamante that is aimed at improving the performance of open domain chatbots by incorporating human feedback (both explicit and implicit preferences). The proposed framework has a generation-evaluation training paradigm that is aimed to optimized response...
A new region based explanation method is proposed that computes attribution scores for features as "escape distances" from a region/polytope defined by the user as input. strengths - The proposed technical idea of region based explanations is quite novel and useful. The idea of using escape distances is good. - Alth...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: A new region based explanation method is proposed that computes attribution scores for features as "escape distances" from a region/polytope defined by the user as input. strengths - The proposed technical idea of region based explanations is quite novel and useful. The idea of using escape distances is good....
This paper analyzes the generalization contribution of local training in Federated learning. Their results show the key to promote generalization (conditional Wasserstein distance) and therefore this paper proposes decoupling the deep models for harnessing client drift and protecting privacy. Experimental results show ...	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 analyzes the generalization contribution of local training in Federated learning. Their results show the key to promote generalization (conditional Wasserstein distance) and therefore this paper proposes decoupling the deep models for harnessing client drift and protecting privacy. Experimental resul...
This paper proposes a novel sampling algorithm for distributions learned by neural language models. The authors start by highlighting a pitfall of “naively” sampling from these models, where high likelihood tokens being picked leads to their future probability increasing and to repetitive loops in the samples. To incr...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a novel sampling algorithm for distributions learned by neural language models. The authors start by highlighting a pitfall of “naively” sampling from these models, where high likelihood tokens being picked leads to their future probability increasing and to repetitive loops in the samples. ...
The paper introduced a new concept extensibility, to explore the openness of CLIP on visual recognition task. Two new evaluation metrics Acc-E and Acc-S were proposed to quantify CLIP's extensibility and stability. The paper argued the confusion between text embeddings of different classes made the CLIP unstable in ext...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper introduced a new concept extensibility, to explore the openness of CLIP on visual recognition task. Two new evaluation metrics Acc-E and Acc-S were proposed to quantify CLIP's extensibility and stability. The paper argued the confusion between text embeddings of different classes made the CLIP unstabl...
This paper proposed a novel generative model GA-NTK that mitigates the drawbacks of traditional GANs trained with the alternative SGD method. This is achieved by using a closed-form discriminator based on a neural tangent kernel (NTK) instead of a neural network. The author proved the convergence of the proposed model ...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This paper proposed a novel generative model GA-NTK that mitigates the drawbacks of traditional GANs trained with the alternative SGD method. This is achieved by using a closed-form discriminator based on a neural tangent kernel (NTK) instead of a neural network. The author proved the convergence of the propose...
This paper uses a deterministic policy gradient for a deep contextual bandit with additive noises for exploration, and through extensive experiments discusses that this approach is easier for continuous and multi-dimensional action spaces, compared to others. Strengths: Extending a recent state-of-the-art method from a...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper uses a deterministic policy gradient for a deep contextual bandit with additive noises for exploration, and through extensive experiments discusses that this approach is easier for continuous and multi-dimensional action spaces, compared to others. Strengths: Extending a recent state-of-the-art metho...
This work theoretically proves the training dynamics of deep and wide probabilistic neural networks and generalization guarantees described by the minimum eigenvalue of the probabilistic NTK. The PAC-Bayesian framework follows the previous classical NTK based, lazy-training results on DNNs, e.g., using the squared loss...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work theoretically proves the training dynamics of deep and wide probabilistic neural networks and generalization guarantees described by the minimum eigenvalue of the probabilistic NTK. The PAC-Bayesian framework follows the previous classical NTK based, lazy-training results on DNNs, e.g., using the squa...
In this paper, the authors presented Compressed Number Line (CNL), a deep neural network model that can extract time-dependent latent variables from sensory inputs, and use them to make predictions through Laplace transform-based timeline representation. This method essentially generalizes the timeline model to number ...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: In this paper, the authors presented Compressed Number Line (CNL), a deep neural network model that can extract time-dependent latent variables from sensory inputs, and use them to make predictions through Laplace transform-based timeline representation. This method essentially generalizes the timeline model to...
This paper presents an approach to leveraging a language modeling objective to train an energy function which can be used together with a sampling procedure to generate trajectories for discrete-control tasks. The approach uses a masked language modeling loss to train a bi-directional language model to fit expert demon...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents an approach to leveraging a language modeling objective to train an energy function which can be used together with a sampling procedure to generate trajectories for discrete-control tasks. The approach uses a masked language modeling loss to train a bi-directional language model to fit expe...
This paper considers an important problem of label noise in the training data. Specifically, it studies the effect of label error on a model's group-based disparity metrics, with more focus on smaller groups in the data. Then, the authors of the paper take a step further by considering a method based on influence funct...	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 considers an important problem of label noise in the training data. Specifically, it studies the effect of label error on a model's group-based disparity metrics, with more focus on smaller groups in the data. Then, the authors of the paper take a step further by considering a method based on influen...
The paper proposes a node assignmet based technique for graph data mixup, as mixup is a standard, popular and effective way of data augmentation for images, and its effectiveness is also recently explored in the graph area. The experimental results are convincing. Pros: 1) The paper is clearly written and the idea is ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a node assignmet based technique for graph data mixup, as mixup is a standard, popular and effective way of data augmentation for images, and its effectiveness is also recently explored in the graph area. The experimental results are convincing. Pros: 1) The paper is clearly written and the ...
The paper studies the risk (generalization error) of the PCA least squares estimator and shows that dimension reduction can avoid the peaking in the risk curve. The analysis is divided into two parts: (i) precise bias-variance decomposition in the proportional limit for isotropic data, and (ii) non-asymptotic bound on ...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies the risk (generalization error) of the PCA least squares estimator and shows that dimension reduction can avoid the peaking in the risk curve. The analysis is divided into two parts: (i) precise bias-variance decomposition in the proportional limit for isotropic data, and (ii) non-asymptotic b...
The paper studies the first price auction in a repeated setting and with budget constraints. The authors propose some algorithms showing sublinear regret under different levels of knowledge of the learner in the setting. The authors also provide some synthetically generated experiments to check the performances of the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies the first price auction in a repeated setting and with budget constraints. The authors propose some algorithms showing sublinear regret under different levels of knowledge of the learner in the setting. The authors also provide some synthetically generated experiments to check the performances...
This paper seeks to define appropriate representations of the coviariates for treatment effect estimation in the setting of multiple treatments with continuous dosage parameters. Previous work has looked at this problem in the binary treatment setting. As in previous work, the paper defines a standard ML loss and a cou...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper seeks to define appropriate representations of the coviariates for treatment effect estimation in the setting of multiple treatments with continuous dosage parameters. Previous work has looked at this problem in the binary treatment setting. As in previous work, the paper defines a standard ML loss a...
In this paper, the authors propose Jump-Start Reinforcement Learning (JSRL) which utilizes a pre-trained guide policy to form a curriculum of starting states for a different exploration policy. Theoretical analysis shows that with a properly chosen training and evaluation algorithm, JSRL achieves a polynomial sample co...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In this paper, the authors propose Jump-Start Reinforcement Learning (JSRL) which utilizes a pre-trained guide policy to form a curriculum of starting states for a different exploration policy. Theoretical analysis shows that with a properly chosen training and evaluation algorithm, JSRL achieves a polynomial s...
This paper considers the robustness of reinforcement learning with respect to observational disturbances. They proposed to use a recently proposed lexicographic optimization framework. Experiments on 3 grid worlds type environment and compare against an adversarial training RL baseline and one naive baseline without co...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper considers the robustness of reinforcement learning with respect to observational disturbances. They proposed to use a recently proposed lexicographic optimization framework. Experiments on 3 grid worlds type environment and compare against an adversarial training RL baseline and one naive baseline wi...
The paper focuses on online bias correction during task-free continual learning. The paper first shows, both theoretically and empirically, why simple experience replay biases on the results of the recent stream observation. Second, the paper introduces the metric to quantify prediction biases. Using the observation fr...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The paper focuses on online bias correction during task-free continual learning. The paper first shows, both theoretically and empirically, why simple experience replay biases on the results of the recent stream observation. Second, the paper introduces the metric to quantify prediction biases. Using the observ...
This paper studies whether prompts that effectively extract information from a language model can also be used to probe other language models for the same information. The authors show that this is indeed not the case. They propose an approach to induce prompts by mixing language models at training time and show that i...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper studies whether prompts that effectively extract information from a language model can also be used to probe other language models for the same information. The authors show that this is indeed not the case. They propose an approach to induce prompts by mixing language models at training time and sho...
This paper presents a fundamental impossibility result associated with the neural network function class. It is proven that given a specific network architecture (e.g., depth and width) with ReLU activation function and norm-bounded parameters, there exists a neural network function u such that u cannot be determined *...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper presents a fundamental impossibility result associated with the neural network function class. It is proven that given a specific network architecture (e.g., depth and width) with ReLU activation function and norm-bounded parameters, there exists a neural network function u such that u cannot be dete...
This paper studies how well-correlated human judgments and LLM predictions are for causal and moral reasoning tasks. They collect a dataset of human judgments for a number of text problems used in cognitive science literature for probing human judgments, including expert judgments for relevant factors in the judgment p...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper studies how well-correlated human judgments and LLM predictions are for causal and moral reasoning tasks. They collect a dataset of human judgments for a number of text problems used in cognitive science literature for probing human judgments, including expert judgments for relevant factors in the ju...
This paper tackles an important problem facing autonomous vehicle perception - how to distinguish between a 3d object and a 2d representation of a 3d object? A useful insight of this paper is that having information of the target object over time leads to important features that help solve this problem. The paper pro...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper tackles an important problem facing autonomous vehicle perception - how to distinguish between a 3d object and a 2d representation of a 3d object? A useful insight of this paper is that having information of the target object over time leads to important features that help solve this problem. The p...
In this work the authors tackle the question of the infinite-width limit of neural networks trained using adaptive optimizers (like ADAM). They show that if the step taken in the optimizer depends on a non-linear, scale-invariant function of the previous step, then the NTK can be computed by applying said function to $...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this work the authors tackle the question of the infinite-width limit of neural networks trained using adaptive optimizers (like ADAM). They show that if the step taken in the optimizer depends on a non-linear, scale-invariant function of the previous step, then the NTK can be computed by applying said funct...
The paper describes a mean to lower the high sensibility of decision trees using randomization of algorithm construction. Some theorems are given and experiments are done showing that in general, trees obtained by the randomization process are less sensitive. +: an interesting problem treated in an original way, that ...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper describes a mean to lower the high sensibility of decision trees using randomization of algorithm construction. Some theorems are given and experiments are done showing that in general, trees obtained by the randomization process are less sensitive. +: an interesting problem treated in an original wa...