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 using placebos (chosen from a free image stream such as Google Images) in class incremental learning. It formulates the policy training process as an online Markov Decision Process, and achieve improvement over prior works. Strength: 1. The paper is written well. 2. Clear improvement is achieved ove...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes using placebos (chosen from a free image stream such as Google Images) in class incremental learning. It formulates the policy training process as an online Markov Decision Process, and achieve improvement over prior works. Strength: 1. The paper is written well. 2. Clear improvement is achi...
This paper proposes a novel contrastive learning algorithm to estimate the occupancy measure of future states in the offline RL setting. It tackles the well-known issue of extrapolation error in existing algorithms for offline RL, in which rewards outside of the underlying data distribution is not well captured by the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a novel contrastive learning algorithm to estimate the occupancy measure of future states in the offline RL setting. It tackles the well-known issue of extrapolation error in existing algorithms for offline RL, in which rewards outside of the underlying data distribution is not well captured...
This paper suggests an interesting idea of real-world actions often being grounded by observations and the shared action-observation features naturally emerging for coordination between agents. This paper hypothesizes that the policy architecture is crucial for emerging the shared action-observation features. The exper...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper suggests an interesting idea of real-world actions often being grounded by observations and the shared action-observation features naturally emerging for coordination between agents. This paper hypothesizes that the policy architecture is crucial for emerging the shared action-observation features. T...
Type inference is a challenging task for the Python language given its dynamic nature. Scholars have recently proposed machine learning (ML)-based techniques to infer types for Python. Previous techniques infer types based on seen examples during training, i.e, they cannot synthesize types, which hinders their ability ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Type inference is a challenging task for the Python language given its dynamic nature. Scholars have recently proposed machine learning (ML)-based techniques to infer types for Python. Previous techniques infer types based on seen examples during training, i.e, they cannot synthesize types, which hinders their ...
This paper proposes to model the latent state representation in RL with hyperbolic space. They find that naively applying existing hyperbolic deep learning methods is not helpful, and introduce two techniques (spectral normalization and feature rescaling) to address the issues. The integrated method (S-RYM) is evaluate...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes to model the latent state representation in RL with hyperbolic space. They find that naively applying existing hyperbolic deep learning methods is not helpful, and introduce two techniques (spectral normalization and feature rescaling) to address the issues. The integrated method (S-RYM) is ...
This work proposes to unify generalization methods by accounting for potential data-generating processes. The paper is presented to unify potential choice of generalization methods under a canonical causal graph which can account for i) label indepedent attributes, ii) label dependent attributes and iii) environment va...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes to unify generalization methods by accounting for potential data-generating processes. The paper is presented to unify potential choice of generalization methods under a canonical causal graph which can account for i) label indepedent attributes, ii) label dependent attributes and iii) enviro...
The paper proposes two main contributions to the direction of HTML modeling/understanding: 1. A new task and dataset, Description Generation, derived from CommonCrawl with 85K (HTML, element, description) tuples. 2. Experiments across a suite of three HTML tasks (Description Generation, semantic classification f...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes two main contributions to the direction of HTML modeling/understanding: 1. A new task and dataset, Description Generation, derived from CommonCrawl with 85K (HTML, element, description) tuples. 2. Experiments across a suite of three HTML tasks (Description Generation, semantic classifi...
This paper proposed a learnable visual word method to interpret the model prediction behaviors. The authors proposed 2 new modules: semantic visual word learning and dual fidelity preservation. The semantic visual word removes the separation loss from the protopnet, and the dual fidelity preservation adds attention ali...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a learnable visual word method to interpret the model prediction behaviors. The authors proposed 2 new modules: semantic visual word learning and dual fidelity preservation. The semantic visual word removes the separation loss from the protopnet, and the dual fidelity preservation adds atten...
This paper proposes to solve a dynamical optimal transport problem where: given a sequence of probability measures sampled in time, the goal is to recover an accurate trajectory of the measures for all intermediate times. Such a scenario is highlighted over two practical applications - single-cell RNA sequencing and di...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes to solve a dynamical optimal transport problem where: given a sequence of probability measures sampled in time, the goal is to recover an accurate trajectory of the measures for all intermediate times. Such a scenario is highlighted over two practical applications - single-cell RNA sequencin...
The paper investigates the impact of regularization in solving extensive-form games (EFGs). It shows that after adding the regularization terms to different pay-off functions, several convergence results could be achieved with either improved convergence rate or weaker assumptions. More specifically, for the dilated o...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper investigates the impact of regularization in solving extensive-form games (EFGs). It shows that after adding the regularization terms to different pay-off functions, several convergence results could be achieved with either improved convergence rate or weaker assumptions. More specifically, for the d...
This paper shows several methods on how to use temporal information in stream-based active learning. It shows why classical pool-based AL methods cannot be used in this domain. It also shows that some methods can even outperform pool-based methods in terms of data savings. This paper closes the gap between pool and str...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper shows several methods on how to use temporal information in stream-based active learning. It shows why classical pool-based AL methods cannot be used in this domain. It also shows that some methods can even outperform pool-based methods in terms of data savings. This paper closes the gap between pool...
This paper proposes relational transformers, extending the transformers architecture to general graph-structured data. Specifically, the proposed relational transformers is a graph-to-graph model with feature updates of nodes, edges, and the global graph vector at each layer. Relational transformers achieve remarkable ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes relational transformers, extending the transformers architecture to general graph-structured data. Specifically, the proposed relational transformers is a graph-to-graph model with feature updates of nodes, edges, and the global graph vector at each layer. Relational transformers achieve rem...
This paper uses the energy loss for training normalizing flows. This produces a consistent estimator that doesn't require computing determinants of the Jacobian of the transformation, making it more flexible in terms of the choice of architecture. The paper also introduces a few tricks to further improve scaling and/or...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This paper uses the energy loss for training normalizing flows. This produces a consistent estimator that doesn't require computing determinants of the Jacobian of the transformation, making it more flexible in terms of the choice of architecture. The paper also introduces a few tricks to further improve scalin...
This paper tackles the challenging goal: developing a general GNN framework with provably expressive power while maintaining the scalability of the message-passing scheme. The authors first model the edges among neighbors as a multiset of multisets and formulate the NC-1-WL graph isomorphism. They show that the express...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper tackles the challenging goal: developing a general GNN framework with provably expressive power while maintaining the scalability of the message-passing scheme. The authors first model the edges among neighbors as a multiset of multisets and formulate the NC-1-WL graph isomorphism. They show that the...
In this paper, the authors develop notions of comprehensibility and transparency as indicator of explainability, and develop formal notions to introduce these as constraints during model training. The authors also propose a new technique for generating saliency maps. Subsequently, they show that inclusion of these cons...	Recommendation: 8: accept, good paper	Area: Optimization (eg, convex and non-convex optimization)	Review: In this paper, the authors develop notions of comprehensibility and transparency as indicator of explainability, and develop formal notions to introduce these as constraints during model training. The authors also propose a new technique for generating saliency maps. Subsequently, they show that inclusion of th...
This paper proposes a new generative model by using inverse heat dissipation. The proposed model is motivated by diffusion models and the empirical success of coarse-to-fine modelling. Specifically, the authors propose to synthesize images by iteratively inverting the heat equation, which is basically a PDE that locall...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper proposes a new generative model by using inverse heat dissipation. The proposed model is motivated by diffusion models and the empirical success of coarse-to-fine modelling. Specifically, the authors propose to synthesize images by iteratively inverting the heat equation, which is basically a PDE tha...
This paper presents a Momentum Stiefel Optimizer with several pleasant properties. The paper is well structured and well written; Since I was not familiar with this field and not able to confirm every single proof/algorithm, I am unable to assess this paper. The paper is well structured and well written; Since I was n...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper presents a Momentum Stiefel Optimizer with several pleasant properties. The paper is well structured and well written; Since I was not familiar with this field and not able to confirm every single proof/algorithm, I am unable to assess this paper. The paper is well structured and well written; Since...
Test-time adaptation updates a model to reduce generalization on shifted data. Such adaptation methods need to choose a loss for adaptation and parameters to update, and this work's main contribution is to introduce visual prompts as a parameterization. The proposed Data-efficient Prompt Tuning (DePT) method includes a...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: Test-time adaptation updates a model to reduce generalization on shifted data. Such adaptation methods need to choose a loss for adaptation and parameters to update, and this work's main contribution is to introduce visual prompts as a parameterization. The proposed Data-efficient Prompt Tuning (DePT) method in...
This work draws connections between the DCI metrics, commonly used in the deep learning literature on disentanglement, and notions of representation identifiability, commonly used in the literature on independent component analysis (ICA). More precisely, Corollary 3.5 provides conditions under which D=C=1 and K=L impli...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work draws connections between the DCI metrics, commonly used in the deep learning literature on disentanglement, and notions of representation identifiability, commonly used in the literature on independent component analysis (ICA). More precisely, Corollary 3.5 provides conditions under which D=C=1 and K...
This paper proposes an adaptation of the popular MaskedAE model for timeseries data. S - solid experimental work in terms of datasets and models compared - good results with multiple ablation tests W - lack of connection between motivation (disentanglement) and the actual proposed method - non-significant result c...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an adaptation of the popular MaskedAE model for timeseries data. S - solid experimental work in terms of datasets and models compared - good results with multiple ablation tests W - lack of connection between motivation (disentanglement) and the actual proposed method - non-significant ...
The authors argue that contributions of knowledge from the teacher to the student network should be layer dependent. Adaptive block-wise learning automatically balances the contribution of knowledge between the student and teacher for each block. The paper talks in generalities and employs a general abstract notation...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors argue that contributions of knowledge from the teacher to the student network should be layer dependent. Adaptive block-wise learning automatically balances the contribution of knowledge between the student and teacher for each block. The paper talks in generalities and employs a general abstract ...
This work shows that after finetuning the pretrained language models on a book corpus, the models align better with brains though their language ability is not better. This better alignment seems general to all brain regions and is also shown in two different metrics. After further evaluating the alignment increase on ...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This work shows that after finetuning the pretrained language models on a book corpus, the models align better with brains though their language ability is not better. This better alignment seems general to all brain regions and is also shown in two different metrics. After further evaluating the alignment incr...
In this paper, the authors proposed a joint Gaussian mixture model (JGMM)-based post-hoc explanation method, which applies inter-layer deep features in a probabilistic model. The JGMM can explain deep features and inter-layer deep feature relationship on the latent component variables in its GMMs. The JGMM is applicabl...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper, the authors proposed a joint Gaussian mixture model (JGMM)-based post-hoc explanation method, which applies inter-layer deep features in a probabilistic model. The JGMM can explain deep features and inter-layer deep feature relationship on the latent component variables in its GMMs. The JGMM is a...
In the existing application of language model (LM) in RL-based dialog management (DM) models, the agent is trained at the word-level and thus results in a complex action space. To address this issue, this paper proposes a novel mixture of expert language models (MoE-LM) and performs the planning at the utterance level ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In the existing application of language model (LM) in RL-based dialog management (DM) models, the agent is trained at the word-level and thus results in a complex action space. To address this issue, this paper proposes a novel mixture of expert language models (MoE-LM) and performs the planning at the utteranc...
In this paper, authors proposed a new methodology by rendering text as images, making it possible to transfer representations across languages based on orthographic similarity or the co-activation of pixels. The pre-trained step is learning reconstructing the pixels of masked patches instead of predicting a distributio...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, authors proposed a new methodology by rendering text as images, making it possible to transfer representations across languages based on orthographic similarity or the co-activation of pixels. The pre-trained step is learning reconstructing the pixels of masked patches instead of predicting a dis...
The authors consider FL problem in the scenario where some nodes do not participate in the learning and provide a theory that analysis this scenario. Moreover, the authors consider SA-FL: the server takes part in the optimization process to mitigate the corruption from partial participation of the nodes. Before I start...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors consider FL problem in the scenario where some nodes do not participate in the learning and provide a theory that analysis this scenario. Moreover, the authors consider SA-FL: the server takes part in the optimization process to mitigate the corruption from partial participation of the nodes. Before...
This paper proposes using LLMs to understand underlying reward functions given a few language examples, then demonstrate that RL agents trained with this LLM-provided sparse reward function nearly match those trained with true underlying reward functions on some simple tasks. ## Strengths *****************************...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes using LLMs to understand underlying reward functions given a few language examples, then demonstrate that RL agents trained with this LLM-provided sparse reward function nearly match those trained with true underlying reward functions on some simple tasks. ## Strengths *********************...
	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Recommendation: 8
The paper considers gradient compression in federated/distributed learning with two additional requirements: robustness against Byzantine clients and differential privacy (DP). The paper proposes a sign-based gradient compression method, called $\beta$-stochastic signSGD, which applies clipping and a stochastic sign op...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper considers gradient compression in federated/distributed learning with two additional requirements: robustness against Byzantine clients and differential privacy (DP). The paper proposes a sign-based gradient compression method, called $\beta$-stochastic signSGD, which applies clipping and a stochastic...
This paper addresses molecular property prediction in the few-shot setting by suggesting a novel Modern Hopfield Network architecture and testing it on the molecule-specific few-shot dataset FS-Mol. The novel architecture seeks to make use of the context available in the wider molecular training set through means other...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper addresses molecular property prediction in the few-shot setting by suggesting a novel Modern Hopfield Network architecture and testing it on the molecule-specific few-shot dataset FS-Mol. The novel architecture seeks to make use of the context available in the wider molecular training set through mea...
This paper presented a novel fragment (motif) based generative model for molecules. The motif vocabulary is mined from the molecule dataset by iterative merging small motifs into larger ones and keeping the most frequent ones from each iteration. Then a generative model is learned to construct the motifs into molecules...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presented a novel fragment (motif) based generative model for molecules. The motif vocabulary is mined from the molecule dataset by iterative merging small motifs into larger ones and keeping the most frequent ones from each iteration. Then a generative model is learned to construct the motifs into m...
The paper proposes a unified self-supervised learning framework for both representation learning and generation. The proposed approach first discretizes an input image into a sequence of tokens using VQGAN and concatenate them with text tokens. It then trains a bidirectional Transformer encoder using the masked token p...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a unified self-supervised learning framework for both representation learning and generation. The proposed approach first discretizes an input image into a sequence of tokens using VQGAN and concatenate them with text tokens. It then trains a bidirectional Transformer encoder using the masked...
This paper proposed an unlearnable strategy, TUE. It aims to simultaneously enjoy training-wise transferability and data-wise transferability by enhancing linear separability. Experimental results show that the proposed method shows advantages in both transferability settings. Strength: 1. This paper proposed a simple...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposed an unlearnable strategy, TUE. It aims to simultaneously enjoy training-wise transferability and data-wise transferability by enhancing linear separability. Experimental results show that the proposed method shows advantages in both transferability settings. Strength: 1. This paper proposed ...
This paper proves the exponential convergence of gradient flow on multi-layer linear models in which the loss function f satisfies the gradient dominance property. It also provides a lower bound on the convergence rate that depends on the imbalance matrices and the least singular value of the weight product. Strength:...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proves the exponential convergence of gradient flow on multi-layer linear models in which the loss function f satisfies the gradient dominance property. It also provides a lower bound on the convergence rate that depends on the imbalance matrices and the least singular value of the weight product. S...
The paper empirically evaluates how machine learning models forget the training samples during the training process. The paper empirically demonstrates that machine learning models do memorize the training samples once it sees them, but tend to forget them gradually once the training proceeds without seeing them again....	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper empirically evaluates how machine learning models forget the training samples during the training process. The paper empirically demonstrates that machine learning models do memorize the training samples once it sees them, but tend to forget them gradually once the training proceeds without seeing the...
In this work, the authors propose an efficient attention EVA, where the work reals that exact softmax attention can be recovered from RFA by manipulating each control variate. The mathematic analysis is provided to prove the dissecting RFA with control variates. And the authors also implemented their EVA and applied it...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this work, the authors propose an efficient attention EVA, where the work reals that exact softmax attention can be recovered from RFA by manipulating each control variate. The mathematic analysis is provided to prove the dissecting RFA with control variates. And the authors also implemented their EVA and ap...
The paper explores the topics of crossmodal knowledge distillation (KD), to transfer knowledge across modalities. To facilitate better understanding of crossmodal KD, the paper proposed a hypothesis that modality-general decisive features are the crucial factor that determines the efficacy of crossmodal KD. Strength: (...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper explores the topics of crossmodal knowledge distillation (KD), to transfer knowledge across modalities. To facilitate better understanding of crossmodal KD, the paper proposed a hypothesis that modality-general decisive features are the crucial factor that determines the efficacy of crossmodal KD. Str...
The authors of this paper investigate a unified model for multiple tasks in molecular science. The model named MolJET is based on Transformer and inspired by unification models in NLP. MolJET takes a prompt sequence and is able to conduct multiple tasks. Experimental results on several tasks show better or competitive ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The authors of this paper investigate a unified model for multiple tasks in molecular science. The model named MolJET is based on Transformer and inspired by unification models in NLP. MolJET takes a prompt sequence and is able to conduct multiple tasks. Experimental results on several tasks show better or comp...
This paper considers offline RL with return conditioning, similar to the recent line of works on decision transformers. It considers the problem of return conditioning especially when rewards can be arbitrary depending on the offline data distribution, even if considering expert trajectories. It considers two key meth...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper considers offline RL with return conditioning, similar to the recent line of works on decision transformers. It considers the problem of return conditioning especially when rewards can be arbitrary depending on the offline data distribution, even if considering expert trajectories. It considers two ...
This paper discussed the robust transfer learning through min-max principle. Specifically, this paper adopted chi^2/Hellinger distance to measure the joint distribution. Through min-max formula, this paper focused on the task reweighting approach such that to learn the source target weight, embedding and its downstream...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper discussed the robust transfer learning through min-max principle. Specifically, this paper adopted chi^2/Hellinger distance to measure the joint distribution. Through min-max formula, this paper focused on the task reweighting approach such that to learn the source target weight, embedding and its do...