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 looks at the approximation capabilities of a DBN with 2 layers in which the first layer is a has binary states (both hidden and visible) and the second layer has continuous visible state. Further, it is assumed that the conditional distributions for the last layer all come from the same parental distribution...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper looks at the approximation capabilities of a DBN with 2 layers in which the first layer is a has binary states (both hidden and visible) and the second layer has continuous visible state. Further, it is assumed that the conditional distributions for the last layer all come from the same parental dist...
This submission analyzes and extends previous invariant risk minimization (IRM) formulations to a multi-objective optimization scheme, named as PAreto Invariant Risk minimization (PAIR) for finding models generalizing well and with good OOD performance. The authors motivated the work by analyzing how IRM may fail to ac...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This submission analyzes and extends previous invariant risk minimization (IRM) formulations to a multi-objective optimization scheme, named as PAreto Invariant Risk minimization (PAIR) for finding models generalizing well and with good OOD performance. The authors motivated the work by analyzing how IRM may fa...
This paper claims that it proves that fairness and trustworthy are undecidable for binary classification. The authors frame this as a first-order logic problem to guide their proof. I am not entirely sure what the point of this paper is, unfortunately. The introduction makes claims about fairness and trustworthiness no...	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 claims that it proves that fairness and trustworthy are undecidable for binary classification. The authors frame this as a first-order logic problem to guide their proof. I am not entirely sure what the point of this paper is, unfortunately. The introduction makes claims about fairness and trustworth...
This paper proposes a method to narrow the performance gap between SGD with momentum and AdamW when evaluating on the downstream tasks. The authors find that for the models where AdamW outperforms SGD, the gradients of the first layer are much larger than the gradients of the other layers. Therefore, they try to freeze...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a method to narrow the performance gap between SGD with momentum and AdamW when evaluating on the downstream tasks. The authors find that for the models where AdamW outperforms SGD, the gradients of the first layer are much larger than the gradients of the other layers. Therefore, they try t...
The paper proposes GReaT (Generation of Realistic Tabular Data), a large language model fine-tuned to generate samples for tabular datasets. First, GReaT encodes examples in a tabular dataset into sequences of text, then GReaT fine tunes a large pretrained language model using the textual representation of the tabular ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper proposes GReaT (Generation of Realistic Tabular Data), a large language model fine-tuned to generate samples for tabular datasets. First, GReaT encodes examples in a tabular dataset into sequences of text, then GReaT fine tunes a large pretrained language model using the textual representation of the ...
The paper proposes a new algorithm named FOSTER for out-of-distribution (OoD) detection in the federated setting. Since there may be no OoD sample in the federated setting, the paper proposes to use samples from the other classes that do not exist in the current client as OoD samples for local training. Since the raw d...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a new algorithm named FOSTER for out-of-distribution (OoD) detection in the federated setting. Since there may be no OoD sample in the federated setting, the paper proposes to use samples from the other classes that do not exist in the current client as OoD samples for local training. Since t...
This paper proposes a novel method for continuous unconstrained black-box optimization problems, called Optformer. It is claimed that the Optformer is inspired by the similarity between Transformer and EA. Six simple benchmark functions and a real-world problem are used to verify the performance of Optformer. In genera...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a novel method for continuous unconstrained black-box optimization problems, called Optformer. It is claimed that the Optformer is inspired by the similarity between Transformer and EA. Six simple benchmark functions and a real-world problem are used to verify the performance of Optformer. I...
This paper presents a novel method for zero-shot multi-label classification where the model is trained on seen classes and tested on unseen classes without further model tuning. The strategy uses min-max zer-sum game between the maximizer and the minimizer with minimum cost graph cuts. The proposed method has four add...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presents a novel method for zero-shot multi-label classification where the model is trained on seen classes and tested on unseen classes without further model tuning. The strategy uses min-max zer-sum game between the maximizer and the minimizer with minimum cost graph cuts. The proposed method has ...
This paper examines the perspective of interpreting masked language models (MLMs) as a generative model, where the unmasking procedure corresponds to predicting unary conditionals. The paper notes that the different unmasked tokens are predicted in a conditionally independent way, and focuses previous attempts that ins...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper examines the perspective of interpreting masked language models (MLMs) as a generative model, where the unmasking procedure corresponds to predicting unary conditionals. The paper notes that the different unmasked tokens are predicted in a conditionally independent way, and focuses previous attempts ...
- This works address two majors challenges in continual learning: catastrophic forgetting and knowledge transfer (KT) across tasks applied to NLP tasks. - In doing so, the approach is based on sub-networks, and computes task similaity via gradients and mask importance where catastrophic forgetting is controlled using ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: - This works address two majors challenges in continual learning: catastrophic forgetting and knowledge transfer (KT) across tasks applied to NLP tasks. - In doing so, the approach is based on sub-networks, and computes task similaity via gradients and mask importance where catastrophic forgetting is controlle...
Update after rebuttal The authors added an important control experiment (fluid in-polarity) which shows that the main improvements result from transferring a favorable initial set of weights, whereas actually freezing the polarity (which is the main NI inspiration) has a a marginal effect (and it is conceivable t...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: Update after rebuttal The authors added an important control experiment (fluid in-polarity) which shows that the main improvements result from transferring a favorable initial set of weights, whereas actually freezing the polarity (which is the main NI inspiration) has a a marginal effect (and it is conce...
This paper proposes a new keypoint representation learning method through novel losses inspired by information theory. One of the losses encourages the keypoints to represent the scene without information loss, and the other guides the keypoints to represent well the dynamics of the video through conditional entropy an...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a new keypoint representation learning method through novel losses inspired by information theory. One of the losses encourages the keypoints to represent the scene without information loss, and the other guides the keypoints to represent well the dynamics of the video through conditional en...
This paper addresses the generalization ability of language-instructed agents of gSCAN. Based on the Meta-Sequence-to-Sequence learning approach and meta-seq2seq architecture of Lake 2019, they issue the statistical action-bias of gSCAN split-H and extend the Meta-Sequence-to-Sequence learning approach to the specially...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper addresses the generalization ability of language-instructed agents of gSCAN. Based on the Meta-Sequence-to-Sequence learning approach and meta-seq2seq architecture of Lake 2019, they issue the statistical action-bias of gSCAN split-H and extend the Meta-Sequence-to-Sequence learning approach to the s...
In this paper, the authors present an energy saving approach that replaces dot product operations with an equivalent add-shift-add operation. In the context of neural networks with integer weights, the authors demonstrate how such a technique can save significant amounts of energy while controlling for the accuracy deg...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors present an energy saving approach that replaces dot product operations with an equivalent add-shift-add operation. In the context of neural networks with integer weights, the authors demonstrate how such a technique can save significant amounts of energy while controlling for the accu...
The authors introduce a new class of stochastic processes that they think of as a generalization of diffusion models to infinite-dimensional inputs. - Strength: Any attempt to mary the pragmatism, effectiveness and scalability of deep learning with the "principledness" of Bayesian nonparametrics is always laudabl...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The authors introduce a new class of stochastic processes that they think of as a generalization of diffusion models to infinite-dimensional inputs. - Strength: Any attempt to mary the pragmatism, effectiveness and scalability of deep learning with the "principledness" of Bayesian nonparametrics is always...
Existing diffusion-based cross-modal generation methods mainly establish the cross-modal relationships by incorporating the cross-modal prior model into the variational lower bound of the diffusion model. However, the authors claim that this method may lead to the loss of the cross-modal correspondence in the denoising...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Existing diffusion-based cross-modal generation methods mainly establish the cross-modal relationships by incorporating the cross-modal prior model into the variational lower bound of the diffusion model. However, the authors claim that this method may lead to the loss of the cross-modal correspondence in the d...
Authors build upon the mitigation of back-propagation and aim at reaching higher accuracies. They first highlight these limits of BP. They then make very strong claims about soft Hebb, an algorithm capable of notably learn in a self-supervised manner or without feedback signals. They evaluate their algorithms or standa...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: Authors build upon the mitigation of back-propagation and aim at reaching higher accuracies. They first highlight these limits of BP. They then make very strong claims about soft Hebb, an algorithm capable of notably learn in a self-supervised manner or without feedback signals. They evaluate their algorithms o...
This paper proposes a straightforward hyper-parameter-free approach to embed discrete solvers as differentiable layers into neural networks. In detail, during the backward pass, the Jacobian $\frac{\partial y(\omega)}{\partial \omega}$ is simply treated as a negative identity matrix and a theoretical justification is p...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a straightforward hyper-parameter-free approach to embed discrete solvers as differentiable layers into neural networks. In detail, during the backward pass, the Jacobian $\frac{\partial y(\omega)}{\partial \omega}$ is simply treated as a negative identity matrix and a theoretical justificat...
This paper introduces a method (B-DISTILL) to distill an (expensive) trained model into an ensemble of weak learners. These learners are meant to provide efficient inference, with progressively better results as the ensemble size increases: this allows for a straightforward way to trade off accuracy and compute at infe...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces a method (B-DISTILL) to distill an (expensive) trained model into an ensemble of weak learners. These learners are meant to provide efficient inference, with progressively better results as the ensemble size increases: this allows for a straightforward way to trade off accuracy and compute...
This paper proposes an animal-pose estimation method which can be applied to multiple animal species. Firstly the proposed method divides a set of animal keypoints into several groups. The keypoints in one group are supposed to have close relation, i.e. intra-group keypoints may provide a localization cue to each other...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an animal-pose estimation method which can be applied to multiple animal species. Firstly the proposed method divides a set of animal keypoints into several groups. The keypoints in one group are supposed to have close relation, i.e. intra-group keypoints may provide a localization cue to ea...
This paper provides a representation view of latent variable models in linear MDPs, and further propose a computationally efficient algorithm to implement it for both online and offline RL. The authors also theoretically and empirically demonstrate the proposed approach. Strength: + The paper is well organised and...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper provides a representation view of latent variable models in linear MDPs, and further propose a computationally efficient algorithm to implement it for both online and offline RL. The authors also theoretically and empirically demonstrate the proposed approach. Strength: + The paper is well organ...
This paper targets the problem of Class-Incremental-Learning with two major contributions. First, the authors propose to evaluate previous methods with aligned memory budgets to fairly compare exampler-based and model-based approaches. Second, the authors conduct empirical observations on the effect of different layers...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper targets the problem of Class-Incremental-Learning with two major contributions. First, the authors propose to evaluate previous methods with aligned memory budgets to fairly compare exampler-based and model-based approaches. Second, the authors conduct empirical observations on the effect of differen...
The authors propose a more realistic multi-agent communication setup and provide a corresponding experimental environment. The authors also propose a corresponding learning framework based on MI to solve the corresponding communication problem, and the experiments show that their approach has significant advantages. St...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose a more realistic multi-agent communication setup and provide a corresponding experimental environment. The authors also propose a corresponding learning framework based on MI to solve the corresponding communication problem, and the experiments show that their approach has significant advant...
This work proposes an interactive segmentation framework that leverages a pre-trained ViT model to perform semi-supervised segmentation tasks on otherwise challenging images. Due to the nature of the proposed framework and ambiguity in evaluation for interactive segmentation framework, no baseline is provided or compar...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This work proposes an interactive segmentation framework that leverages a pre-trained ViT model to perform semi-supervised segmentation tasks on otherwise challenging images. Due to the nature of the proposed framework and ambiguity in evaluation for interactive segmentation framework, no baseline is provided o...
This work discusses how regularizing the representations of deep neural networks can help improve model generalization when the dataset is noisy. In particular, the authors first discuss how limiting the capacity of a neural network can help improve generalization in noisy data settings. This is achieved via the regula...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This work discusses how regularizing the representations of deep neural networks can help improve model generalization when the dataset is noisy. In particular, the authors first discuss how limiting the capacity of a neural network can help improve generalization in noisy data settings. This is achieved via th...
The submission proposed the hidden Markov based GPFR mixture model (HM-GPFR) by describing the time series data from the perspective of both fine and coarse level. The time series data at fine level is characterized by Gaussian process model and and that at coarse level is characterized by a hidden Markov process. To a...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The submission proposed the hidden Markov based GPFR mixture model (HM-GPFR) by describing the time series data from the perspective of both fine and coarse level. The time series data at fine level is characterized by Gaussian process model and and that at coarse level is characterized by a hidden Markov proce...
This work conducts an analysis of expressive power of geometric graph neural networks. Analogous to WL test for generic GNNs, this work proposes Geometric WL (GWL) test and its constraint version IGWL for geometric graphs. It characterizes an expressive-power gap between GWL and IGWL, which is then accordingly extended...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work conducts an analysis of expressive power of geometric graph neural networks. Analogous to WL test for generic GNNs, this work proposes Geometric WL (GWL) test and its constraint version IGWL for geometric graphs. It characterizes an expressive-power gap between GWL and IGWL, which is then accordingly ...
1. The paper tackles the problem of client drift (locally) in federated learning (FL) due to heterogeneous client data distributions. Besides, they target the period drift (globally), which is the inter-communication heterogeneity of data distributions. 2. They propose a learning based parameterized aggregator called F...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: 1. The paper tackles the problem of client drift (locally) in federated learning (FL) due to heterogeneous client data distributions. Besides, they target the period drift (globally), which is the inter-communication heterogeneity of data distributions. 2. They propose a learning based parameterized aggregator ...
In this paper authors proposed a variational information pursuit for interpretable classification model / estimation scheme. Authors idea is motivated by generative variant that they denote G-IP. Authors propose a complete framework with model defintions explanation of the training scheme and proof that loss does what ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper authors proposed a variational information pursuit for interpretable classification model / estimation scheme. Authors idea is motivated by generative variant that they denote G-IP. Authors propose a complete framework with model defintions explanation of the training scheme and proof that loss do...
This paper studies the problem of the heterogeneity of local data distributions in federated learning. The authors argued that the distribution overlaps are not consistent but scattered in local clients. They proposed to infer the local data manifolds to learn from the informative overlaps and estimate the data density...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies the problem of the heterogeneity of local data distributions in federated learning. The authors argued that the distribution overlaps are not consistent but scattered in local clients. They proposed to infer the local data manifolds to learn from the informative overlaps and estimate the data...
The paper proposes a new routing strategy, Routing Entropy Minimization (REM), for capsule networks. The core idea of REM is to minimize the entropy of capsule parse trees by pruning so that the activated connections between capsules are stronger and fewer, which will improve the interpretability of the capsule network...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes a new routing strategy, Routing Entropy Minimization (REM), for capsule networks. The core idea of REM is to minimize the entropy of capsule parse trees by pruning so that the activated connections between capsules are stronger and fewer, which will improve the interpretability of the capsule...
The paper proposes speculative decoding, a decoding scheme for conditional NLG that aims to achieve the quality of traditional autoregressive approaches while retaining the speedup improvements of non-autoregressive approaches. The proposed framework combines a non-autoregressive transformer (NAT) to generate "draft" s...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes speculative decoding, a decoding scheme for conditional NLG that aims to achieve the quality of traditional autoregressive approaches while retaining the speedup improvements of non-autoregressive approaches. The proposed framework combines a non-autoregressive transformer (NAT) to generate "...
This paper is about improving the expressivity of large scale probabilistic circuits (PCs). Finding a good starting point for EM based learning of these large latent variable models is problematic and the authors propose one such solution to this problem. The main idea is to obtain semantic-aware assignments (called s...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper is about improving the expressivity of large scale probabilistic circuits (PCs). Finding a good starting point for EM based learning of these large latent variable models is problematic and the authors propose one such solution to this problem. The main idea is to obtain semantic-aware assignments (...
In order to train GNN efficiently on large graphs with a single machine and external storage, this paper proposes a locality-aware training scheme to reduce the data movement time. First, it partitions the graph using METIS. The main idea is to save a mega-batch, which is randomly sampled from partitions in the main me...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In order to train GNN efficiently on large graphs with a single machine and external storage, this paper proposes a locality-aware training scheme to reduce the data movement time. First, it partitions the graph using METIS. The main idea is to save a mega-batch, which is randomly sampled from partitions in the...
The paper proposes MultiMix which is a variant of the Mixup data augmentation technique. While Mixup creates augmented data by only considering interpolation between two data points, MultiMix considers the interpolation between $m$ data samples. In greater detail, MultiMix creates $n$ new augmented samples from $m$ dat...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes MultiMix which is a variant of the Mixup data augmentation technique. While Mixup creates augmented data by only considering interpolation between two data points, MultiMix considers the interpolation between $m$ data samples. In greater detail, MultiMix creates $n$ new augmented samples from...
The paper start from an interesting empirical finding that a transformer trained with ERM outperforms CNN trained with domain generalization (DG) algorithms on DG task. They show some experiments and math to prove that. They propose a new GMoE and show its superiority on DG. The paper focus on an interesting aspect an...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper start from an interesting empirical finding that a transformer trained with ERM outperforms CNN trained with domain generalization (DG) algorithms on DG task. They show some experiments and math to prove that. They propose a new GMoE and show its superiority on DG. The paper focus on an interesting a...
The paper introduces an asymmetric loss called Softmax MSE to address the overestimation issue in value-based off-policy learning. Instead of taking a minimum of an ensemble of Q-functions, as was proposed in TD3, the authors propose penalizing overestimation in TD learning. # Strength 1. The idea is exciting and novel...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces an asymmetric loss called Softmax MSE to address the overestimation issue in value-based off-policy learning. Instead of taking a minimum of an ensemble of Q-functions, as was proposed in TD3, the authors propose penalizing overestimation in TD learning. # Strength 1. The idea is exciting a...
The paper introduces Implicit Value Regularization (IVR) framework, provides some interesting theoretical results that might help better understand offline methods such as CQL and IQL, and then under this framework, a new offline RL method called Sparse Q-learning (SQL) is proposed. Empirical results from D4RL benchm...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces Implicit Value Regularization (IVR) framework, provides some interesting theoretical results that might help better understand offline methods such as CQL and IQL, and then under this framework, a new offline RL method called Sparse Q-learning (SQL) is proposed. Empirical results from D4R...
This paper may most accurately be summarized as giving polynomial-time algorithms that obtain sublinear regret in linear contextual MDPs when the transition and Q-functions are linear in the state-action features and the context, and that only need to re-solve the MDP logarithmically many times w.r.t. the number of tas...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper may most accurately be summarized as giving polynomial-time algorithms that obtain sublinear regret in linear contextual MDPs when the transition and Q-functions are linear in the state-action features and the context, and that only need to re-solve the MDP logarithmically many times w.r.t. the numbe...
Test-time model adaptation updates model parameters during inference in order to reduce generalization error on shifted data. Continual test-time adaptation, the setting of this work, does so for varying shifts without knowledge of when the shift itself changes over time. The purpose of this work is to improve the effi...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: Test-time model adaptation updates model parameters during inference in order to reduce generalization error on shifted data. Continual test-time adaptation, the setting of this work, does so for varying shifts without knowledge of when the shift itself changes over time. The purpose of this work is to improve ...