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 method for correcting bias of continual learning method using experience replay. The proposed method uses a memory from past examples to train the last layer of a neural network trained by an experience-replay-based method. This prevents the model from being biased to recently observed samples. By...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a method for correcting bias of continual learning method using experience replay. The proposed method uses a memory from past examples to train the last layer of a neural network trained by an experience-replay-based method. This prevents the model from being biased to recently observed sam...
This paper proposes the simple yet effective idea of using continuous diffusion models to generate discrete data by representing them as binary data and modeling these binary data as real numbers. The authors also propose two techniques, namely Self-Conditioning and Asymmetric Time Intervals, which lead to improvement ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper proposes the simple yet effective idea of using continuous diffusion models to generate discrete data by representing them as binary data and modeling these binary data as real numbers. The authors also propose two techniques, namely Self-Conditioning and Asymmetric Time Intervals, which lead to impr...
This paper drew a connection between soft threshold pruning and Iterative Shrinkage-Thresholding Algorithm (ISTA). Based on this interpretation of soft threshold pruning the authors derive novel threshold schedulers that appear to perform better than alternatives in terms of sparsity-accuracy trade-off. Strength: - The...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper drew a connection between soft threshold pruning and Iterative Shrinkage-Thresholding Algorithm (ISTA). Based on this interpretation of soft threshold pruning the authors derive novel threshold schedulers that appear to perform better than alternatives in terms of sparsity-accuracy trade-off. Strengt...
The paper deal with the problem of mutii-agent specialization: which is the issue related to the developing of a single general learning algorithms for a relatively diverse population. Typically each agent in a multi-agent paradigm is trained using separate behavioral policies. But this has now led way to jointly optim...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper deal with the problem of mutii-agent specialization: which is the issue related to the developing of a single general learning algorithms for a relatively diverse population. Typically each agent in a multi-agent paradigm is trained using separate behavioral policies. But this has now led way to joint...
This paper aims to tackle the combinatorial generalization problem in the offline sensorimotor object rearrangement task. By homogenizing the dynamics of different objects, the authors propose a framework named NCS that trains the correspondence matching module and plans a path from the initial state to the goal state ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper aims to tackle the combinatorial generalization problem in the offline sensorimotor object rearrangement task. By homogenizing the dynamics of different objects, the authors propose a framework named NCS that trains the correspondence matching module and plans a path from the initial state to the goa...
The paper presents an architecture that leverages the symmetry of entity-centric goal specification. The author extends the MDP framework to modeling the symmetric structure of the goal specification. The framework motivates using architectures that are agnostic to permutation of the inputs like Deep Set, Self-Attentio...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents an architecture that leverages the symmetry of entity-centric goal specification. The author extends the MDP framework to modeling the symmetric structure of the goal specification. The framework motivates using architectures that are agnostic to permutation of the inputs like Deep Set, Self-...
This paper proposes a normalizing flow-based autoencoder (AE-Flow) to address medical image anomaly detection. The AE-Flow model contains three components, an encoder to extract features of the input images, a flow model to convert extracted feature vector to the specific feature space, which follows a standard normal...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a normalizing flow-based autoencoder (AE-Flow) to address medical image anomaly detection. The AE-Flow model contains three components, an encoder to extract features of the input images, a flow model to convert extracted feature vector to the specific feature space, which follows a standar...
This paper proposes a new loss, called L* as opposed to L2 loss for training a heuristic to be used in A* search. I like the general direction of this paper. I could see how since L2 is not directly optimizing the main objective of being a good heuristic, a different objective could do better. However, I think that t...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a new loss, called L* as opposed to L2 loss for training a heuristic to be used in A* search. I like the general direction of this paper. I could see how since L2 is not directly optimizing the main objective of being a good heuristic, a different objective could do better. However, I thin...
Recently chain-of-thought prompting ("lets think step by step") has demonstrated big language models can perform several reasoning tasks well, e.g. match word problem. However, it is known to not perform well when the problem to solve is harder than the few-shot examples in the prompt. To address the problem, this pape...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Recently chain-of-thought prompting ("lets think step by step") has demonstrated big language models can perform several reasoning tasks well, e.g. match word problem. However, it is known to not perform well when the problem to solve is harder than the few-shot examples in the prompt. To address the problem, t...
[response to author's rebuttal: see https://openreview.net/forum?id=mCmerkTCG2S&noteId=mpYhm0LBTQD] The paper "Brain-like representational straightening of natural movies in robust feedforward neural networks" investigates to which degree standard ImageNet-trained models and "robust" networks (in terms of adversarial ...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: [response to author's rebuttal: see https://openreview.net/forum?id=mCmerkTCG2S&noteId=mpYhm0LBTQD] The paper "Brain-like representational straightening of natural movies in robust feedforward neural networks" investigates to which degree standard ImageNet-trained models and "robust" networks (in terms of adve...
The paper proposes a Sentiment-oriented Transformer-based Variational Autoencoder (SO-TVAE) model for the automatic live video commenting task. The SO-TVAE takes sentiment clues of the surrounding comments to generate new automatic live video comments. The model also adopts the VAE to achieve commenting sentences with ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a Sentiment-oriented Transformer-based Variational Autoencoder (SO-TVAE) model for the automatic live video commenting task. The SO-TVAE takes sentiment clues of the surrounding comments to generate new automatic live video comments. The model also adopts the VAE to achieve commenting sentenc...
This paper proposes a deterministic VAE for unsupervised disentanglement learning with Euler encodings. By introducing an architectural inductive bias with the Eular layer, this approach can learn disentanglement without using independence regularization terms or additional supervisions. In addition, the authors presen...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a deterministic VAE for unsupervised disentanglement learning with Euler encodings. By introducing an architectural inductive bias with the Eular layer, this approach can learn disentanglement without using independence regularization terms or additional supervisions. In addition, the author...
In this paper, the authors proposed an interesting approach based on what I can tell to be a novel connection between vision transformers and evolutionary algorithms. The paper then modifies the transformer architecture to implement evolutionary-specific operations like crossover, mutation and selection. The authors th...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors proposed an interesting approach based on what I can tell to be a novel connection between vision transformers and evolutionary algorithms. The paper then modifies the transformer architecture to implement evolutionary-specific operations like crossover, mutation and selection. The au...
In this work the authors investigate protocols for utilizing pre-trained models for downstream tasks. Motivated by the observation that different protocols perform differently according to several metrics under varying levels of distribution shift, the authors propose to analyse such protocols under the light of simpli...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this work the authors investigate protocols for utilizing pre-trained models for downstream tasks. Motivated by the observation that different protocols perform differently according to several metrics under varying levels of distribution shift, the authors propose to analyse such protocols under the light o...
The paper proposes a new algorithm for reward-free offline reinforcement learning when a few expert demonstrations are available. The core idea is to assign rewards to the unlabelled trajectories based on their distance to the expert demonstrations such that the closer the trajectories they are to the expert demonstrat...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a new algorithm for reward-free offline reinforcement learning when a few expert demonstrations are available. The core idea is to assign rewards to the unlabelled trajectories based on their distance to the expert demonstrations such that the closer the trajectories they are to the expert de...
This paper proposed a graph anomaly detection method through mutual information maximization. The key contributions are proposing (i) an orthogonal projection layer for the decision boundary correction and (ii) a two co-centered hyperspheres structure for estimating the normal distribution. Experimental results on mult...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposed a graph anomaly detection method through mutual information maximization. The key contributions are proposing (i) an orthogonal projection layer for the decision boundary correction and (ii) a two co-centered hyperspheres structure for estimating the normal distribution. Experimental results...
The paper claims to extend the previously proposed CTRL framework for representation learning to unsupervised setting. Paper proposes self-consistency and augmented sample compression based constraints to define an objective function for representation learning. Several experiments evaluate the representations learned ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper claims to extend the previously proposed CTRL framework for representation learning to unsupervised setting. Paper proposes self-consistency and augmented sample compression based constraints to define an objective function for representation learning. Several experiments evaluate the representations ...
This paper attempts to introduce robust algorithms against label noise into backdoor defense. In particular, the authors propose a meta-algorithm that can transform an existing noisy label defense into one that defends against backdoor attacks. The experimental results demonstrate the effectiveness of the proposed meth...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper attempts to introduce robust algorithms against label noise into backdoor defense. In particular, the authors propose a meta-algorithm that can transform an existing noisy label defense into one that defends against backdoor attacks. The experimental results demonstrate the effectiveness of the propo...
This paper proposes a "hyper label model" for weak supervision that can infer ground-truth labels for any weak supervision dataset without retraining. The authors evaluate their hyper label model in accuracy and efficiency on 14 benchmark datasets. The hyper label model is more efficient and on average slightly more ac...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a "hyper label model" for weak supervision that can infer ground-truth labels for any weak supervision dataset without retraining. The authors evaluate their hyper label model in accuracy and efficiency on 14 benchmark datasets. The hyper label model is more efficient and on average slightly...
This paper proposes a hierarchical topic modeling approach based on conditional transport, combining previous research on hierarchical topic models with recent research on conditional transport. The idea is that documents are represented by discrete distributions over word embeddings, topics at each level of the hierar...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes a hierarchical topic modeling approach based on conditional transport, combining previous research on hierarchical topic models with recent research on conditional transport. The idea is that documents are represented by discrete distributions over word embeddings, topics at each level of th...
The paper proposes to apply a subcomponent-wise softmax to SSL representations (z -> partition -> softmax -> concatenate -> new z), dubbed SEM. They empirically show that this simple transformation combined with a large increase in the dimensionality of the representation yields significant improvements in standard SS...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper proposes to apply a subcomponent-wise softmax to SSL representations (z -> partition -> softmax -> concatenate -> new z), dubbed SEM. They empirically show that this simple transformation combined with a large increase in the dimensionality of the representation yields significant improvements in sta...
This paper is concerned with the implicit bias of gradient flow and gradient descent in two-layer fully connected neural networks with leaky ReLU activations when the training data are nearly-orthogonal. For gradient flow, this paper leverages the implicit bias for homogeneous neural networks to show that gradient flow...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper is concerned with the implicit bias of gradient flow and gradient descent in two-layer fully connected neural networks with leaky ReLU activations when the training data are nearly-orthogonal. For gradient flow, this paper leverages the implicit bias for homogeneous neural networks to show that gradi...
The paper proposes Stay-On-the-Ridge $(STON'R)$ algorithm, which according to the authors it is the first method that is guaranteed to converge to a local min-max equilibrium for smooth nonconvex-nonconcave objectives. The proposed method is a second-order algorithm that provably escapes limit cycles as long as it is i...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper proposes Stay-On-the-Ridge $(STON'R)$ algorithm, which according to the authors it is the first method that is guaranteed to converge to a local min-max equilibrium for smooth nonconvex-nonconcave objectives. The proposed method is a second-order algorithm that provably escapes limit cycles as long as...
In this work, the authors propose a new feature extraction framework called MuHS to improve few-shot semantic segmentation using multi-coarse heterogeneity-aware optimization. The MuHS is optimized with three key components: cross-sample attention, cross-region interaction, and masked image segmentation. The empirical ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work, the authors propose a new feature extraction framework called MuHS to improve few-shot semantic segmentation using multi-coarse heterogeneity-aware optimization. The MuHS is optimized with three key components: cross-sample attention, cross-region interaction, and masked image segmentation. The em...
This paper presents a new algorithm, referred to as the LAVA method, for evaluating the minimal weight arborescence of a directed graph which each of the edges carries a weight. It was shown in the literature that this question can be reduced to the Multi-Pair Shortest Path (MPSP) problem. The present paper investigate...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper presents a new algorithm, referred to as the LAVA method, for evaluating the minimal weight arborescence of a directed graph which each of the edges carries a weight. It was shown in the literature that this question can be reduced to the Multi-Pair Shortest Path (MPSP) problem. The present paper inv...
This paper proposes a Gaussian mixture model (GMM) for binary classification problems with imbalanced groups. The labels are generated according to a ground truth neural network. Authors mention a few observations according to the theoretical analysis and provided empirical evidence using CelebA and CIFAR datasets. Str...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proposes a Gaussian mixture model (GMM) for binary classification problems with imbalanced groups. The labels are generated according to a ground truth neural network. Authors mention a few observations according to the theoretical analysis and provided empirical evidence using CelebA and CIFAR datas...
This work proposes TransFool for generating non-targeted adversarial examples against neural machine translation models. One core idea is to utilize an autoregressive language model (GPT-2) to add a language model loss term, which helps generate fluent adversarial examples. They also add a similarity loss, which constr...	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 work proposes TransFool for generating non-targeted adversarial examples against neural machine translation models. One core idea is to utilize an autoregressive language model (GPT-2) to add a language model loss term, which helps generate fluent adversarial examples. They also add a similarity loss, whic...
The authors propose Pruning with Output Error Minimization (POEM) technique to improve model accuracy after structured pruning. In order to include activation function in the objective function to be optimized, the authors suggest the weighted least squares method while the authors claim the previous methods minimize t...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose Pruning with Output Error Minimization (POEM) technique to improve model accuracy after structured pruning. In order to include activation function in the objective function to be optimized, the authors suggest the weighted least squares method while the authors claim the previous methods mi...
The paper proposes a novel method for adversarial imitation learning that restricts the family of reward functions to a specific function that is monotonically decreasing in the reconstruction error of an auto-encoder (I think that "reconstruction-error" and "auto-encoder" are misnomers, as the encoder-decoder network ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a novel method for adversarial imitation learning that restricts the family of reward functions to a specific function that is monotonically decreasing in the reconstruction error of an auto-encoder (I think that "reconstruction-error" and "auto-encoder" are misnomers, as the encoder-decoder ...
This paper suggests an algorithm for offline reinforcement learning based on first applying a noise contrastive objective to estimate a quantity proportional to $p(s_{t+\Delta t}\|s_t,a_t)/p(s_{t+\Delta t})$, where $\Delta t$ belongs to a geometric distribution of future times. That is, the ratio of the discounted futur...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper suggests an algorithm for offline reinforcement learning based on first applying a noise contrastive objective to estimate a quantity proportional to $p(s_{t+\Delta t}\|s_t,a_t)/p(s_{t+\Delta t})$, where $\Delta t$ belongs to a geometric distribution of future times. That is, the ratio of the discount...
The paper presents a new approach for learning QUBO formulation for problems to be solved on quantum annealers using machine learning in comparison to previous work on quantum computer vision that relied on hand-crafted QUBO formulations. The paper proposes learning to regress a coefficient matrix for QUBO using contra...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper presents a new approach for learning QUBO formulation for problems to be solved on quantum annealers using machine learning in comparison to previous work on quantum computer vision that relied on hand-crafted QUBO formulations. The paper proposes learning to regress a coefficient matrix for QUBO usin...
The paper proposes Equivariant Diffusion-Hypergraph Neural Network (ED-HNN) for hypergraph structured data. ED-HNN is implemented as a message-passing neural network on the star expansion of the hypergraph and can provably represent any continuous hypergraph diffusion operator. ED-HNN competes with recent HNNs on hyp...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes Equivariant Diffusion-Hypergraph Neural Network (ED-HNN) for hypergraph structured data. ED-HNN is implemented as a message-passing neural network on the star expansion of the hypergraph and can provably represent any continuous hypergraph diffusion operator. ED-HNN competes with recent HNN...
Predicting protein-molecule interaction by combining a pre-trained protein language model and a GNN. The method was evaluated empirically in the context of odor perception. Strengths: + Protein-molecule interaction is generally a problem of interest to many communities. + The proposed method is an interesting way to co...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: Predicting protein-molecule interaction by combining a pre-trained protein language model and a GNN. The method was evaluated empirically in the context of odor perception. Strengths: + Protein-molecule interaction is generally a problem of interest to many communities. + The proposed method is an interesting w...
This paper presents SE(3) equivariant energy-based models for sample-efficient visual robotic manipulation from point clouds. It also provides theoretical conditions for bi-equivariant energy-based models, along with sampling strategies over the SE(3) manifold. Experiments on simple simulated pick-place and stick-in-tr...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents SE(3) equivariant energy-based models for sample-efficient visual robotic manipulation from point clouds. It also provides theoretical conditions for bi-equivariant energy-based models, along with sampling strategies over the SE(3) manifold. Experiments on simple simulated pick-place and sti...
The paper proposes a pre-training strategy for graph neural networks based on line graphs. The method is then benchmarks on several molecular property prediction tasks. Strengths: - The work approaches a relevant problem, how to pre-train molecular encoders - Clearly written and structured paper Weaknesses: - The lack...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper proposes a pre-training strategy for graph neural networks based on line graphs. The method is then benchmarks on several molecular property prediction tasks. Strengths: - The work approaches a relevant problem, how to pre-train molecular encoders - Clearly written and structured paper Weaknesses: - ...
Summary: This paper explores solving the masking augmentation problem in convolutional neural networks. This topic is very important. The authors' method is also very clever. However, unfortunately, its performance is not satisfactory. Moreover, it does not solve the more important problem of masking image modeling. T...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: Summary: This paper explores solving the masking augmentation problem in convolutional neural networks. This topic is very important. The authors' method is also very clever. However, unfortunately, its performance is not satisfactory. Moreover, it does not solve the more important problem of masking image mod...
The paper motivates the interest for making bidirectional language models better few-shot/zero-shot reasoners from natural language prompts/instructions. The paper argues that bidirectional models can potentially have richer representations but the unidirectional training in the causal decoder-only models typically hap...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper motivates the interest for making bidirectional language models better few-shot/zero-shot reasoners from natural language prompts/instructions. The paper argues that bidirectional models can potentially have richer representations but the unidirectional training in the causal decoder-only models typic...
This paper considers the problem of off-policy evaluation using deep convolutional neural network. The main contribution of this paper is a theoretical justification of the ability that a deep neural network captures low-rank structures in high-dimensional datasets in off-policy RL settings. Although preliminary result...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper considers the problem of off-policy evaluation using deep convolutional neural network. The main contribution of this paper is a theoretical justification of the ability that a deep neural network captures low-rank structures in high-dimensional datasets in off-policy RL settings. Although preliminar...
This paper presents an approach for combining rule based policies and RL methods, such as TD3+BC and TD3. The proposed approach performs well (though I have no clear calibration of the tasks being benchmarked upon). The resulting policy attains good performance and can improve upon heuristics obtained through the rule ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper presents an approach for combining rule based policies and RL methods, such as TD3+BC and TD3. The proposed approach performs well (though I have no clear calibration of the tasks being benchmarked upon). The resulting policy attains good performance and can improve upon heuristics obtained through t...
The paper proposes a new generative model for tabular data, as a VAE whose decoder structure is built around GraphNN layers. The rationale for this structure is given in terms of inductive biases: on structured data like images or text, the successful models are built around architectures that take advantage of good i...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper proposes a new generative model for tabular data, as a VAE whose decoder structure is built around GraphNN layers. The rationale for this structure is given in terms of inductive biases: on structured data like images or text, the successful models are built around architectures that take advantage o...