review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper studied the problem of improving accuracy (performance of average subgroups) and fairness (performance in worst case subgroups). Then this paper further proposed an active learning strategy by querying the worst-subgroup to improve the accuracy-fairness boundary. Besides, this paper assumes that subgroup in... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper studied the problem of improving accuracy (performance of average subgroups) and fairness (performance in worst case subgroups). Then this paper further proposed an active learning strategy by querying the worst-subgroup to improve the accuracy-fairness boundary. Besides, this paper assumes that sub... |
This paper proposes to utilize the meta-node (average of node embeddings within a cluster) for constructing negative samples in contrastive learning. It claims that this approach is more efficient than previous sampling based methods and it leads naturally to block contrastive learning which could be beneficial. Experi... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes to utilize the meta-node (average of node embeddings within a cluster) for constructing negative samples in contrastive learning. It claims that this approach is more efficient than previous sampling based methods and it leads naturally to block contrastive learning which could be beneficial... |
This paper presents a method, dubbed Q-Pensieve, for reusing critic knowledge across training iterations in a multi-objective reinforcement learning (MORL) setting. Specifically it proposes a modification of Soft Actor Critic (SAC) in the multi-objective setting that uses snapshots of previous Q functions in the learni... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a method, dubbed Q-Pensieve, for reusing critic knowledge across training iterations in a multi-objective reinforcement learning (MORL) setting. Specifically it proposes a modification of Soft Actor Critic (SAC) in the multi-objective setting that uses snapshots of previous Q functions in th... |
This paper explains the reason behind the effectiveness of Gradient-based attacks for poisoning and evasion on semi-supervised node classification tasks on the graph.
Strength:
- It is an important and interesting problem to investigate.
- Addressing this problem from a distribution perspective is interesting
- Provi... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper explains the reason behind the effectiveness of Gradient-based attacks for poisoning and evasion on semi-supervised node classification tasks on the graph.
Strength:
- It is an important and interesting problem to investigate.
- Addressing this problem from a distribution perspective is interesting... |
The paper aims to understand language models' internal knowledge about truth values of the text that it outputs.
Given a set of yes-no questions, they train a truth classifier in an unsupervised way by first extracting the hidden states of the language model when it gives a positive or negative answer, respectively. Th... | 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 aims to understand language models' internal knowledge about truth values of the text that it outputs.
Given a set of yes-no questions, they train a truth classifier in an unsupervised way by first extracting the hidden states of the language model when it gives a positive or negative answer, respecti... |
The paper presents a copy-paste data augmentation strategy for the few-shot incremental semantic segmentation task. It starts with a base learning stage to train an initial segmentation network with fully-annotated images. In the subsequent incremental learning stages, the proposed method develops a guided sample sele... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a copy-paste data augmentation strategy for the few-shot incremental semantic segmentation task. It starts with a base learning stage to train an initial segmentation network with fully-annotated images. In the subsequent incremental learning stages, the proposed method develops a guided sam... |
This paper studies semi-supervised classification when the labels are not missing at random. In this scenario, the labeled and unlabeled data cannot be considered as from the same distribution, and the existing semi-supervised learning algorithms face issues such as over-learning on popular classes and ignoring rare cl... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper studies semi-supervised classification when the labels are not missing at random. In this scenario, the labeled and unlabeled data cannot be considered as from the same distribution, and the existing semi-supervised learning algorithms face issues such as over-learning on popular classes and ignoring... |
The paper analyzes common fairness metrics achieved by empirical risk minimization on a synthetic data generation model. The work leverages existing physics results to fully characterize solutions for two-group, binary label scenarios where the group-conditional covariate distribution follows a simple Gaussian distribu... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper analyzes common fairness metrics achieved by empirical risk minimization on a synthetic data generation model. The work leverages existing physics results to fully characterize solutions for two-group, binary label scenarios where the group-conditional covariate distribution follows a simple Gaussian ... |
The paper propose an image recognition paradigm named "classification by description". To be specifically, instead of matching an image with its categorical description, the method first queries a large language model to obtain some textual descriptions for each category and base the recognition decision on these descr... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper propose an image recognition paradigm named "classification by description". To be specifically, instead of matching an image with its categorical description, the method first queries a large language model to obtain some textual descriptions for each category and base the recognition decision on the... |
In this paper, authors challenge a usual way of designing a spectral function with a dimensional-wise mapping on Graph diffusion. Based on this, authors pay more attention on the global pattern of the specturm rather than single eigenvalue and propose a learnable set2set spectral filter. Based on the new spectral filte... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, authors challenge a usual way of designing a spectral function with a dimensional-wise mapping on Graph diffusion. Based on this, authors pay more attention on the global pattern of the specturm rather than single eigenvalue and propose a learnable set2set spectral filter. Based on the new spectr... |
This paper presents a new approach for online continual learning extending the OCM method introduced by Guo et al. (2022). The proposed approach consists in an optimisation problem amd loss function that not only enforces to learn holistic representations of the data (as in OCM) but also class-invariant features. The i... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a new approach for online continual learning extending the OCM method introduced by Guo et al. (2022). The proposed approach consists in an optimisation problem amd loss function that not only enforces to learn holistic representations of the data (as in OCM) but also class-invariant feature... |
This paper studies a novel token pruning/sparsification technique via backtracking the importance from the final layer to the first layer. In order to control the computation cost, the token importance is approximated via a forward procedure through a lightweight distilled counterpart (i.e., ApproxNet), and then a smoo... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studies a novel token pruning/sparsification technique via backtracking the importance from the final layer to the first layer. In order to control the computation cost, the token importance is approximated via a forward procedure through a lightweight distilled counterpart (i.e., ApproxNet), and the... |
The paper initiates an interesting exploration of MAE with mixture of experts (MoE). The method is quite well-motivated, with interesting and fairness-in-mind designs, and a good amount of experiments devoted to it. The final results are reported using a suite of 11 downstream classification tasks, typically used to ev... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper initiates an interesting exploration of MAE with mixture of experts (MoE). The method is quite well-motivated, with interesting and fairness-in-mind designs, and a good amount of experiments devoted to it. The final results are reported using a suite of 11 downstream classification tasks, typically us... |
The manuscript presents CRISP, an approach for hierarchical reinforcement learning from expert trajectories. Their approach relabels states from expert demonstration as subgoals to train a higher level. States are labelled as subgoals when they are the first in a subsequence that can’t be reached by the lower level fro... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The manuscript presents CRISP, an approach for hierarchical reinforcement learning from expert trajectories. Their approach relabels states from expert demonstration as subgoals to train a higher level. States are labelled as subgoals when they are the first in a subsequence that can’t be reached by the lower l... |
A novel spike encoding system for event streams is presented, that allows to reduce the amount of spikes and increase the performance of spike-learninig event-based pipelines. The method is evaluated on recognition benchmarks, as well as on Earth satellite tracking datasets.
The spike-based learning is a very relevant ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
A novel spike encoding system for event streams is presented, that allows to reduce the amount of spikes and increase the performance of spike-learninig event-based pipelines. The method is evaluated on recognition benchmarks, as well as on Earth satellite tracking datasets.
The spike-based learning is a very r... |
This paper studies bivariate causal direction learning from observational data. The authors focus on the general bivariate model, i.e., the post-nonlinear model, and propose a new method to learn the model. The contributions are as follows, the authors first analyze the drawbacks of the existing estimation methods, e.g... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies bivariate causal direction learning from observational data. The authors focus on the general bivariate model, i.e., the post-nonlinear model, and propose a new method to learn the model. The contributions are as follows, the authors first analyze the drawbacks of the existing estimation meth... |
The paper proposes a light-weighted plugin/algorithm to uplift the upper bound of the GNNs' expressive power, namely the 1-WL test. By constructing a multi-hop multi-color rooted subtree, the algorithm achieves or even exceeds the performance of the subgraph GNNs while using fewer computational resources.
Strength:
1. ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a light-weighted plugin/algorithm to uplift the upper bound of the GNNs' expressive power, namely the 1-WL test. By constructing a multi-hop multi-color rooted subtree, the algorithm achieves or even exceeds the performance of the subgraph GNNs while using fewer computational resources.
Stren... |
This paper analyzes the performance of the recently-introduced and influential MAUVE metric for evaluating the quality of automatically-generated text by comparing it to human-generated text at the distribution level. It begins by setting out the rationale for the MAUVE approach, starting with difficulties in applying ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper analyzes the performance of the recently-introduced and influential MAUVE metric for evaluating the quality of automatically-generated text by comparing it to human-generated text at the distribution level. It begins by setting out the rationale for the MAUVE approach, starting with difficulties in a... |
This paper proposes a simple model-agnostic framework to boost causal discovery performance by dynamically learning the adaptive weights for the score function. In particular, the proposed method leverages the bilevel optimization scheme to alternatively train a standard DAG learner first and then reweight the samples ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a simple model-agnostic framework to boost causal discovery performance by dynamically learning the adaptive weights for the score function. In particular, the proposed method leverages the bilevel optimization scheme to alternatively train a standard DAG learner first and then reweight the ... |
This paper is concerned with the problem of developing surrogates of programs. A major difficulty in this context is how train surrogates for programs with control flow. To overcome this challenge, this paper represents the program as a stratified function and uses stratified surrogates to model such functions. To ensu... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper is concerned with the problem of developing surrogates of programs. A major difficulty in this context is how train surrogates for programs with control flow. To overcome this challenge, this paper represents the program as a stratified function and uses stratified surrogates to model such functions.... |
This paper proposes a a distributed RL system built upon a state-of-the-art model-based RL
method, EfficientZero, with system support for fast distributed computation. The paper also proposes a novel technique to stabilize massively parallel model-based training. Empirical evaluations demonstrate the efficacy of the ap... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a a distributed RL system built upon a state-of-the-art model-based RL
method, EfficientZero, with system support for fast distributed computation. The paper also proposes a novel technique to stabilize massively parallel model-based training. Empirical evaluations demonstrate the efficacy o... |
The proposed work uses the hypothesis: Non-discriminative filters do not contribute a lot to the predictive performance of a network and can be removed safely. Keeping this in mind, along with a privacy based setting where the original training data and loss function are unavailable yet the a sample from the original d... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The proposed work uses the hypothesis: Non-discriminative filters do not contribute a lot to the predictive performance of a network and can be removed safely. Keeping this in mind, along with a privacy based setting where the original training data and loss function are unavailable yet the a sample from the or... |
This paper proposes a context-aware adaptive mechanism to adjust the step-size parameter rho in ADMM for solving convex quadratic programming problems, denote as CA-ADMM. It extracts the spatio-temporal context during the ADMM iterations. Numerical experiments on various type of QP sets are reported.
Strength:
i) The p... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a context-aware adaptive mechanism to adjust the step-size parameter rho in ADMM for solving convex quadratic programming problems, denote as CA-ADMM. It extracts the spatio-temporal context during the ADMM iterations. Numerical experiments on various type of QP sets are reported.
Strength:
... |
The paper introduces a new perspective for multi-task RL that to distill each control policies in each task using techniques inspired from the minimum description length principle. Specifically, the proposed algorithm learns a policy parameters distribution that both contains the control policy parameters and constrain... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper introduces a new perspective for multi-task RL that to distill each control policies in each task using techniques inspired from the minimum description length principle. Specifically, the proposed algorithm learns a policy parameters distribution that both contains the control policy parameters and c... |
This paper studies masking-based audio-visual source separation, which predicts a complex spectral mask for the audio mixture for each speaker conditioning on the mixture speech and the video of the target speaker. The authors proposed BFRNet, which is composed of a audio-visual source separation model that predicts a ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studies masking-based audio-visual source separation, which predicts a complex spectral mask for the audio mixture for each speaker conditioning on the mixture speech and the video of the target speaker. The authors proposed BFRNet, which is composed of a audio-visual source separation model that pre... |
This paper aims to learn a set of classifiers to take care of different predictive signals in the dataset. They propose that the "diversity" of the ensemble is important and the conditional independence is an effective way to realize this goal.
Pros:
* The proposed fast adaption method for ensemble learning can address... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper aims to learn a set of classifiers to take care of different predictive signals in the dataset. They propose that the "diversity" of the ensemble is important and the conditional independence is an effective way to realize this goal.
Pros:
* The proposed fast adaption method for ensemble learning can... |
The task is automatic speech recognition (ASR). The paper proposed PATCorrect, a new correction model which operates on the output of another speech recognition model and tries to improve the output by correcting errors.
The proposed model works in a non-autoregressive way and thus can run in parallel, efficiently on ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The task is automatic speech recognition (ASR). The paper proposed PATCorrect, a new correction model which operates on the output of another speech recognition model and tries to improve the output by correcting errors.
The proposed model works in a non-autoregressive way and thus can run in parallel, efficie... |
The paper investigates the tension between consistency and simplicity in interpretability. The goal is to find a model from a restricted function class (e.g. of simple functions) that matches the network on as much of the input space as possible. The paper focuses on Boolean functions, takes the function space to be fu... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper investigates the tension between consistency and simplicity in interpretability. The goal is to find a model from a restricted function class (e.g. of simple functions) that matches the network on as much of the input space as possible. The paper focuses on Boolean functions, takes the function space ... |
This paper addresses the problem of continual learning under a novel Continual Image-Text Modeling setting where the model learns a sequence of multiple image-text matching tasks. The authors claim that transferring the old knowledge from the historical model parameters and periodically updating the current model to th... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper addresses the problem of continual learning under a novel Continual Image-Text Modeling setting where the model learns a sequence of multiple image-text matching tasks. The authors claim that transferring the old knowledge from the historical model parameters and periodically updating the current mod... |
This paper studies the formal output reachability verification of message passing neural networks (MPNN). The paper claims that the verification is impossible if for general graphs (unbounded size, nontrivial degree and node labels). But the verification is possible if there is a bound on the degree of the graph.
St... | 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 studies the formal output reachability verification of message passing neural networks (MPNN). The paper claims that the verification is impossible if for general graphs (unbounded size, nontrivial degree and node labels). But the verification is possible if there is a bound on the degree of the grap... |
This paper studies the computation of Nash Equilibrium for two-team zero-sum games, and
1. proves that computing an approximate (possibly mixed) NE in two-team zero-sum games is CLS-hard.
2. gives a simple family of two-team zero-sum games where some common online, first-order min-max optimization methods (GDA, OGDA,... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the computation of Nash Equilibrium for two-team zero-sum games, and
1. proves that computing an approximate (possibly mixed) NE in two-team zero-sum games is CLS-hard.
2. gives a simple family of two-team zero-sum games where some common online, first-order min-max optimization methods (GD... |
This paper introduced Trajectory Translation (TR2) framework that seeks to train low-level policies by translating an abstract trajectory into executable actions. TR2 is designed to solve tasks in three steps: 1) build a paired abstract environment by simplifying geometry and physics, generating abstract trajectories... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduced Trajectory Translation (TR2) framework that seeks to train low-level policies by translating an abstract trajectory into executable actions. TR2 is designed to solve tasks in three steps: 1) build a paired abstract environment by simplifying geometry and physics, generating abstract traj... |
The paper extends the previous work DINO by interpreting it a new way, as a von Mises-Fisher mixture model, and in turn proposes DINO-vMF model and achieves very good performance on numerous tasks.
Strength: The interpretation of vMF is appealing, especially leading to removing the L2 regularization to enable more fl... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper extends the previous work DINO by interpreting it a new way, as a von Mises-Fisher mixture model, and in turn proposes DINO-vMF model and achieves very good performance on numerous tasks.
Strength: The interpretation of vMF is appealing, especially leading to removing the L2 regularization to enable... |
The proposed method is as simple as aligning representations in the latent space to learn a domain-invariant policy. Conceptually that makes sense as it can generalise better in the presence of small perturbations and changes in an environment. This allows to do some domain adaptation with zero-shot learning whereby ev... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The proposed method is as simple as aligning representations in the latent space to learn a domain-invariant policy. Conceptually that makes sense as it can generalise better in the presence of small perturbations and changes in an environment. This allows to do some domain adaptation with zero-shot learning wh... |
In this paper a transformer is trained to infer from a set of scalar time series observations the dynamical law behind these observations, in symbolic form. A large data set of equations and trajectories generated by these is generated, from which various test set scenarios are created. Performance is evaluated both in... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
In this paper a transformer is trained to infer from a set of scalar time series observations the dynamical law behind these observations, in symbolic form. A large data set of equations and trajectories generated by these is generated, from which various test set scenarios are created. Performance is evaluated... |
Unlike other previous works, the work first studies the fairness issues for face recognition in terms of the neural architecture and hyperparameter search. The authors run a large-scale study of 29 architectures from ViT to Xception for 355 models in total with different hyperparameters and spend 88,493 GPU hours. They... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Unlike other previous works, the work first studies the fairness issues for face recognition in terms of the neural architecture and hyperparameter search. The authors run a large-scale study of 29 architectures from ViT to Xception for 355 models in total with different hyperparameters and spend 88,493 GPU hou... |
The paper proposes a framework for global counterfactual explanations while considering the bias that can arise across different subgroups and the need to allow for different translations. The paper also provides an analysis of categorical feature translations. Emprirical results support the theoretical analysis.
Stren... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a framework for global counterfactual explanations while considering the bias that can arise across different subgroups and the need to allow for different translations. The paper also provides an analysis of categorical feature translations. Emprirical results support the theoretical analysi... |
The paper adapts adversarial attack strategies to multivariate probabilistic forecasting models, focusing on attacks that might be difficult to detect. They also adapt defense mechanisms to adversarial attacks to the time-series forecasting setup and demonstrate how multivariate problems differ from univariate ones.
Th... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper adapts adversarial attack strategies to multivariate probabilistic forecasting models, focusing on attacks that might be difficult to detect. They also adapt defense mechanisms to adversarial attacks to the time-series forecasting setup and demonstrate how multivariate problems differ from univariate ... |
The paper makes two contributions for single-output deep learning models with smooth activations:
1) Upper bounds the spectral norm of the model's Hessian over a layer-wise spectral norm, which is larger than the euclidean ball that was studied in previous work. Hence, the result holds with relaxed conditions on how cl... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper makes two contributions for single-output deep learning models with smooth activations:
1) Upper bounds the spectral norm of the model's Hessian over a layer-wise spectral norm, which is larger than the euclidean ball that was studied in previous work. Hence, the result holds with relaxed conditions o... |
The paper improves the deep generative models used for constraining solution spaces in inverse problems by making distributions of latent variables obey a Gaussian. Deep generative models such as StyleGAN2 and Glow are widely employed for the regularizes in inverse problems. Given latent variables that obey a Gaussian,... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper improves the deep generative models used for constraining solution spaces in inverse problems by making distributions of latent variables obey a Gaussian. Deep generative models such as StyleGAN2 and Glow are widely employed for the regularizes in inverse problems. Given latent variables that obey a G... |
The authors propose a method for generating news articles conditioned on entities to appear in an article. The authors claim that existing methods for generating news articles only model text and ignore visual content and secondly that these methods do not explicitly account for named entities to be mentioned in genera... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a method for generating news articles conditioned on entities to appear in an article. The authors claim that existing methods for generating news articles only model text and ignore visual content and secondly that these methods do not explicitly account for named entities to be mentioned i... |
1. This paper interpreted self-supervised contrastive learning (SSCL) as a type of stochastic neighbor embedding (SNE) methods that preserve the pairwise similarities specified by the data augmentations.
2. Based on the connection with SNE, this paper provided theoretical insights for domain-agnostic data augmentation... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
1. This paper interpreted self-supervised contrastive learning (SSCL) as a type of stochastic neighbor embedding (SNE) methods that preserve the pairwise similarities specified by the data augmentations.
2. Based on the connection with SNE, this paper provided theoretical insights for domain-agnostic data augm... |
This paper proposes a novel approach for text-to-3D generation based on pre-trained text-conditioned diffusion models (Imagen in this paper). More specifically, the authors learn a neural radiance field by rendering at random views and use the pretrained diffusion model as the objective function to back-prop gradients ... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper proposes a novel approach for text-to-3D generation based on pre-trained text-conditioned diffusion models (Imagen in this paper). More specifically, the authors learn a neural radiance field by rendering at random views and use the pretrained diffusion model as the objective function to back-prop gr... |
The authors proposed a layer-wise learnable threshold approach for SNNs. The proposed method design two separate surrogate gradient paths for the gradient of membrane potential and learnable threshold. Different classification and object detection tasks reflect the effectivenesses of the proposed LTSNN training scheme.... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors proposed a layer-wise learnable threshold approach for SNNs. The proposed method design two separate surrogate gradient paths for the gradient of membrane potential and learnable threshold. Different classification and object detection tasks reflect the effectivenesses of the proposed LTSNN training... |
This work introduces FedTHE(+), a way to allow for test-time personalization of federated learning models. FedTHE works by introducing and training a scalar that interpolates global and local classifier predictions. This scalar is tuned at test time by minimising the entropy of the interpolated logits while being regul... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work introduces FedTHE(+), a way to allow for test-time personalization of federated learning models. FedTHE works by introducing and training a scalar that interpolates global and local classifier predictions. This scalar is tuned at test time by minimising the entropy of the interpolated logits while bei... |
The paper proposes an intuitive prompt-to-prompt editing framework, where the edits are controlled by text only. While large-scale language-image (LLI) models are collecting attention to generate an image from text, they have limitation to control by only text. The paper propose to control the attention maps of the ed... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper proposes an intuitive prompt-to-prompt editing framework, where the edits are controlled by text only. While large-scale language-image (LLI) models are collecting attention to generate an image from text, they have limitation to control by only text. The paper propose to control the attention maps o... |
This paper proposed a new building block, which combined MBConv and Attention, for neural network design and the authors show its efficiency and effectiveness on many visual tasks.
Strength
1. The authors demonstrate the efficiency and effectiveness on the proposed MOAT module.
2. The ablation studies are comprehensive... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed a new building block, which combined MBConv and Attention, for neural network design and the authors show its efficiency and effectiveness on many visual tasks.
Strength
1. The authors demonstrate the efficiency and effectiveness on the proposed MOAT module.
2. The ablation studies are compr... |
The paper undertakes a minimax analysis of a formulation of the domain generalisation problem. When considering an optimistic case where one has access to infinite data from a finite number of domains, it is determined that the number of required domains to obtain a solution within performance of the Bayes optimal clas... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper undertakes a minimax analysis of a formulation of the domain generalisation problem. When considering an optimistic case where one has access to infinite data from a finite number of domains, it is determined that the number of required domains to obtain a solution within performance of the Bayes opti... |
This paper proposes a new approach to meta-learning that addresses vulnerability to pseudo-labeling errors in state of the art meta-learning methods by improving pseudo-labeling using a momentum network and a queue of previous batches. The main insight in this work comes from the recognition that pseudo-labels are usua... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a new approach to meta-learning that addresses vulnerability to pseudo-labeling errors in state of the art meta-learning methods by improving pseudo-labeling using a momentum network and a queue of previous batches. The main insight in this work comes from the recognition that pseudo-labels ... |
This work proposes a SE3-equivariant EBM with high sample efficiency and a competent generalization capability.
Strength
* The proposed method is well-motivated theoretically, and its practical instantiation incorporatoins several components to accomendate the innate instability and intractability in the theoretical f... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes a SE3-equivariant EBM with high sample efficiency and a competent generalization capability.
Strength
* The proposed method is well-motivated theoretically, and its practical instantiation incorporatoins several components to accomendate the innate instability and intractability in the theor... |
The authors present a complex methodology of how to generate compact MLP-models that would highly accurately and smoothly represent mathematical functions.
Core of the methodology (contributions):
1. a novel hierarchical MLP-model called WGN (weight generating network) - one part of the model ("weight generator") gene... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors present a complex methodology of how to generate compact MLP-models that would highly accurately and smoothly represent mathematical functions.
Core of the methodology (contributions):
1. a novel hierarchical MLP-model called WGN (weight generating network) - one part of the model ("weight generato... |
This paper proposes an open world instance segmentation framework, which contains three stages: 1) it proposes an auxiliary task of predicting total foreground on the basis of existing query-based mask2former model. 2) it includes a cross-task consistency loss and improve results by correcting outputs between two tasks... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes an open world instance segmentation framework, which contains three stages: 1) it proposes an auxiliary task of predicting total foreground on the basis of existing query-based mask2former model. 2) it includes a cross-task consistency loss and improve results by correcting outputs between t... |
This paper presents three techniques on masked autoencoder (MAE)-based pre-training for downstream tasks.
The first is to apply dropout to the attention modules of ViT.
The second is a modification of the image normalization method from using per-patch statistics to using per-dataset statistics.
The third is to adjust ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents three techniques on masked autoencoder (MAE)-based pre-training for downstream tasks.
The first is to apply dropout to the attention modules of ViT.
The second is a modification of the image normalization method from using per-patch statistics to using per-dataset statistics.
The third is to... |
This paper proposes an approximation of the neural tangent kernel (NTK) for models with a large output dimension. Specifically, considering a model with output dimension $O$, for a specific pair of data examples, the standard definition of NTK gives an $O\times O$ matrix. This paper tries to approximate the standard NT... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an approximation of the neural tangent kernel (NTK) for models with a large output dimension. Specifically, considering a model with output dimension $O$, for a specific pair of data examples, the standard definition of NTK gives an $O\times O$ matrix. This paper tries to approximate the sta... |
Thank you for your submission. I mentioned the main points in each section, and I elaborated on the points in the Detailed Comments.
## Summary
The paper proposes TED, an auxiliary task for deep RL agents that encourages learning disentangled representations. TED is shown to improve generalization across color variat... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Thank you for your submission. I mentioned the main points in each section, and I elaborated on the points in the Detailed Comments.
## Summary
The paper proposes TED, an auxiliary task for deep RL agents that encourages learning disentangled representations. TED is shown to improve generalization across colo... |
The paper proposes a volumetric motion estimation framework specifically in the context of fluids. The framework can be trained end-to-end from only single-view image sequences. Experiments show that the proposed framework can successfully generalize across different inputs and outperforms single-scene optimization-bas... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a volumetric motion estimation framework specifically in the context of fluids. The framework can be trained end-to-end from only single-view image sequences. Experiments show that the proposed framework can successfully generalize across different inputs and outperforms single-scene optimiza... |
The paper presents a generative model that draws on the diffusion-based and score-based models that generate samples through a sequence of intermediate distributions, but proposes different parameterisation, training criteria and sampling procedure. The model is motivated largely by the fact that prior models treat tim... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper presents a generative model that draws on the diffusion-based and score-based models that generate samples through a sequence of intermediate distributions, but proposes different parameterisation, training criteria and sampling procedure. The model is motivated largely by the fact that prior models t... |
This paper provides a post-processing to the LINKX with the low-rank constraint on the adjacency matrix. Then, it employs softimpute alternating least square to reduce the complexity. Finally, it bridges the connection between the proposed methods and subspace clustering based on low-rank recovery. Experimental evaluat... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper provides a post-processing to the LINKX with the low-rank constraint on the adjacency matrix. Then, it employs softimpute alternating least square to reduce the complexity. Finally, it bridges the connection between the proposed methods and subspace clustering based on low-rank recovery. Experimental... |
This paper introduces the Short-Term Memory Convolution (STMC) layers, which is designed to be a faster version of the convolution layers (in terms of computation time) and also achieve lower latency for real-time application. The paper first presents the approach, based on shift registers. The proposed layer is then e... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces the Short-Term Memory Convolution (STMC) layers, which is designed to be a faster version of the convolution layers (in terms of computation time) and also achieve lower latency for real-time application. The paper first presents the approach, based on shift registers. The proposed layer i... |
This paper is the first provable policy optimization algorithm for robust MDP in online RL with a $\ell_1$ uncertainty set, along with a finite-sample complexity bound. The main contribution is that compared to the previous sample complexity of robust MDP, under an online RL setting, it gives the first provable optimiz... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper is the first provable policy optimization algorithm for robust MDP in online RL with a $\ell_1$ uncertainty set, along with a finite-sample complexity bound. The main contribution is that compared to the previous sample complexity of robust MDP, under an online RL setting, it gives the first provable... |
This paper proposes an approach to achieve low-rank training for CNNs. Concretely, the authors consider the form of tucker-2 decomposition to build the convolutional kernel, while during training the orthogonality regularization is imposed on the non-core matrices. The authors experimentally demonstrate the effectivene... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an approach to achieve low-rank training for CNNs. Concretely, the authors consider the form of tucker-2 decomposition to build the convolutional kernel, while during training the orthogonality regularization is imposed on the non-core matrices. The authors experimentally demonstrate the eff... |
This paper is on safe reinforcement learning (RL) with unknown stochastic environments. Under the safety definition in terms of reachability, the authors aim to enforce hard constraints rather than soft constraints that are often the main target of constrained Markov decision processes (CMDP) literature. In this paper,... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper is on safe reinforcement learning (RL) with unknown stochastic environments. Under the safety definition in terms of reachability, the authors aim to enforce hard constraints rather than soft constraints that are often the main target of constrained Markov decision processes (CMDP) literature. In thi... |
This work proposed a method for generating online handwriting data based on diffusion models. On several datasets, the proposed method outperforms previous approaches.
Strength:
- To my best knowledge, this is the first work that explores diffusion models to generate temporal sequences.
- The proposed method overcomes ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This work proposed a method for generating online handwriting data based on diffusion models. On several datasets, the proposed method outperforms previous approaches.
Strength:
- To my best knowledge, this is the first work that explores diffusion models to generate temporal sequences.
- The proposed method ov... |
Paper presents a post-hoc method to produce interpretable global rules to explain NER classifiers. Rules are extracted with a data mining approach that gathers labeling rule patterns by FP-growth algorithms, prunes the rules by removing soft-duplicated rules, and selects rules that maximizes the F1 score. Selected rule... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Paper presents a post-hoc method to produce interpretable global rules to explain NER classifiers. Rules are extracted with a data mining approach that gathers labeling rule patterns by FP-growth algorithms, prunes the rules by removing soft-duplicated rules, and selects rules that maximizes the F1 score. Selec... |
This paper proposes a method "interventional rationalization (Inter-RAT)" to discover the causal rationales. Inter-RAT tries to detect spurious correlations between the input and rationales, and between rationales and results, respectively, by analyzing the causalities among them and identifying the confounder in the c... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a method "interventional rationalization (Inter-RAT)" to discover the causal rationales. Inter-RAT tries to detect spurious correlations between the input and rationales, and between rationales and results, respectively, by analyzing the causalities among them and identifying the confounder ... |
This work propose a method to tackle the train-test inconsistent problem, e.g. distribution/domain/task shift exists at test time. The method tries to find a way to linearly combine multi models (obtained from some samples interpolated via train and test data) so that it can adapt to seen tasks as well as unseen interp... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work propose a method to tackle the train-test inconsistent problem, e.g. distribution/domain/task shift exists at test time. The method tries to find a way to linearly combine multi models (obtained from some samples interpolated via train and test data) so that it can adapt to seen tasks as well as unsee... |
The paper investigates the effect of task-dependent context in multi-task learning for the task-switching networks. Particularly, the authors show the task-switching networks operate in a regime between two extreme cases of individual and parallel networks. The authors also study the shared representations, the role of... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper investigates the effect of task-dependent context in multi-task learning for the task-switching networks. Particularly, the authors show the task-switching networks operate in a regime between two extreme cases of individual and parallel networks. The authors also study the shared representations, the... |
This paper proposes to train noise-conditional autoregressive models across a continuum of noise levels, analogously to diffusion models. The authors argue that such models improve density estimation performance under a least-significant-bit perturbation scheme, and also improve sample generation.
# Strengths
The pape... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper proposes to train noise-conditional autoregressive models across a continuum of noise levels, analogously to diffusion models. The authors argue that such models improve density estimation performance under a least-significant-bit perturbation scheme, and also improve sample generation.
# Strengths
... |
This paper tackles the data heterogeneity issue in Federated Learning (FL) by letting local clients leverage model parameters from nodes with similar data distributions. Each client approximates the data similarity of its neighbors through a weighing parameter $w_{ij}$ which is learned by EM optimization. In order to a... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper tackles the data heterogeneity issue in Federated Learning (FL) by letting local clients leverage model parameters from nodes with similar data distributions. Each client approximates the data similarity of its neighbors through a weighing parameter $w_{ij}$ which is learned by EM optimization. In or... |
This paper proposed to abandon the samples near the end of the trajectory due to an estimation error in the GAE estimator. Experiments show the effectiveness of the method.
Strength:
The proposed method is very simple and easy to implement.
It's also interesting to see that abandoning parts of the data near the end... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposed to abandon the samples near the end of the trajectory due to an estimation error in the GAE estimator. Experiments show the effectiveness of the method.
Strength:
The proposed method is very simple and easy to implement.
It's also interesting to see that abandoning parts of the data near... |
The authors proposed a novel subgraph-wise sampling method, Local Message Compensation(LMC), to accelerate the training of GNNs on large-scale graphs. They mainly focus on solving the neighbor explosion problem. Their main concern is finding a solution to the neighbor explosion issue. I think their research has a signi... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors proposed a novel subgraph-wise sampling method, Local Message Compensation(LMC), to accelerate the training of GNNs on large-scale graphs. They mainly focus on solving the neighbor explosion problem. Their main concern is finding a solution to the neighbor explosion issue. I think their research has... |
The paper presents a new HCNN implementation to perform private inference (PI) called HyPHEN. It proposes a replication-based convolution method (RAConv), which is innovatively alternated with the channel-aligned convolution method (CAConv). It also proposes a hybrid packing method, which is a combination of two existi... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper presents a new HCNN implementation to perform private inference (PI) called HyPHEN. It proposes a replication-based convolution method (RAConv), which is innovatively alternated with the channel-aligned convolution method (CAConv). It also proposes a hybrid packing method, which is a combination of tw... |
This work presents a “conversion +fine-tuning” two-step method for training SNNs for text classification. Meanwhile, a new encoding method for converting pre-trained word embeddings to the spiking version is proposed. As results, the converted SNNs achieved comparable results compared with DNNs on text classification b... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This work presents a “conversion +fine-tuning” two-step method for training SNNs for text classification. Meanwhile, a new encoding method for converting pre-trained word embeddings to the spiking version is proposed. As results, the converted SNNs achieved comparable results compared with DNNs on text classifi... |
This paper presents OTOv2, an improved version of Only-Train-Once (OTOv1) framework. It introduces two major improvements over OTOv1, including automated zero-invariant groups (ZIGs) partition and a new optimizer called Dual Half-Space Projected Gradient
(DHSPG). Experiments are conducted on several small image classif... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents OTOv2, an improved version of Only-Train-Once (OTOv1) framework. It introduces two major improvements over OTOv1, including automated zero-invariant groups (ZIGs) partition and a new optimizer called Dual Half-Space Projected Gradient
(DHSPG). Experiments are conducted on several small image... |
This work aims to utilize the encoded vision-and-language correlation in the trained CLIP embedding space to propose a DeCap method to perform zero-shot captioning. In implementations, this work first constructs an auto-encoder language model to learn to generate the given sentence based on the text embedding encoded b... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work aims to utilize the encoded vision-and-language correlation in the trained CLIP embedding space to propose a DeCap method to perform zero-shot captioning. In implementations, this work first constructs an auto-encoder language model to learn to generate the given sentence based on the text embedding e... |
The authors of this paper present a training strategy to improve performance in ML MD. In this strategy, first bias of inaccurate data is calculated using the model trained on accurate, then the bias-aware inaccurate data is used to train an ML MD model from scratch, and finally, the model is fine-tuned using accurate ... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors of this paper present a training strategy to improve performance in ML MD. In this strategy, first bias of inaccurate data is calculated using the model trained on accurate, then the bias-aware inaccurate data is used to train an ML MD model from scratch, and finally, the model is fine-tuned using a... |
This paper explores the how two types of curiosity, lifelong and episodic, contribute to performance and exploration when training an agent using PPO within three different hard exploration tasks from the MiniGrid evironment: KeyCorridor, MultiRoom, and ObstructedMaze. The authors claim that previous work in this setti... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper explores the how two types of curiosity, lifelong and episodic, contribute to performance and exploration when training an agent using PPO within three different hard exploration tasks from the MiniGrid evironment: KeyCorridor, MultiRoom, and ObstructedMaze. The authors claim that previous work in th... |
This paper studies an important problem, i.e., given a pre-trained large teacher model, how to transfer its knowledge to a compact low-compute smaller model. The proposed method is developed on top of the standard knowledge distillation approach. The authors assume that there is an available encoder and decoder, such t... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies an important problem, i.e., given a pre-trained large teacher model, how to transfer its knowledge to a compact low-compute smaller model. The proposed method is developed on top of the standard knowledge distillation approach. The authors assume that there is an available encoder and decoder... |
This work proposes an appearance contrastive loss to apply the self-supervised learned 2D visual feature for 3D representations.
A new geometry contrastive loss is proposed for object segmentation to involve geometric information for segmentation clustering.
The segmentation of the proposed method has a more refined re... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This work proposes an appearance contrastive loss to apply the self-supervised learned 2D visual feature for 3D representations.
A new geometry contrastive loss is proposed for object segmentation to involve geometric information for segmentation clustering.
The segmentation of the proposed method has a more re... |
This paper proposes a set of new methods to deal with the "catastrophic forgetting" issue for federated learning over heterogeneous data. The main idea of this method is to project the newly sampled gradients into the set with acute angle with previous gradients which is kind of like efficient regularization that has... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a set of new methods to deal with the "catastrophic forgetting" issue for federated learning over heterogeneous data. The main idea of this method is to project the newly sampled gradients into the set with acute angle with previous gradients which is kind of like efficient regularization ... |
This work explores the high dimensional limit of a permuted linear regression problem with Gaussian sensing matrix and additive Gaussian noise. Thanks to the introduction of a tailored graphical model, the authors build a Message Passing (MP) algorithm to reconstruct the permutation matrix $\Pi$. The theoretical analy... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This work explores the high dimensional limit of a permuted linear regression problem with Gaussian sensing matrix and additive Gaussian noise. Thanks to the introduction of a tailored graphical model, the authors build a Message Passing (MP) algorithm to reconstruct the permutation matrix $\Pi$. The theoretic... |
This paper evaluates whether or not large language models understand implicature. They create (automatically) a dataset that allows them to evaluate pre-trained models’ understanding of this phenomenon. They evaluate LLMs pre-trained on large text corpora and also ones that are fine-tuned with RL feedback to see the di... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper evaluates whether or not large language models understand implicature. They create (automatically) a dataset that allows them to evaluate pre-trained models’ understanding of this phenomenon. They evaluate LLMs pre-trained on large text corpora and also ones that are fine-tuned with RL feedback to se... |
This paper proposes an unsupervised and parameter-free method that functionally projects transformer models into the space of all tree-structured models. By using this method, they show that transformers for three different tasks become more tree-like over the course of training, in some cases unsupervisedly recovering... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes an unsupervised and parameter-free method that functionally projects transformer models into the space of all tree-structured models. By using this method, they show that transformers for three different tasks become more tree-like over the course of training, in some cases unsupervisedly re... |
This paper proposes a new algorithm called Policy Optimization with Model Planning (POMP), for continuous control problems that incorporates a model-based planner derived from differential dynamic programming (DDP). DDP cannot be practically applied directly as a planner since it has a high computational cost and requi... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a new algorithm called Policy Optimization with Model Planning (POMP), for continuous control problems that incorporates a model-based planner derived from differential dynamic programming (DDP). DDP cannot be practically applied directly as a planner since it has a high computational cost a... |
The paper proposes a Sparse Gaussian Process Attention (SGPA) for the Transformer architecture. The SGPA is used instead of the classical scaled dot-product attention and directly performs an approximation of the Bayesian inference in the attention blocks in Transformers. In addition, the authors provide decoupled SGPA... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a Sparse Gaussian Process Attention (SGPA) for the Transformer architecture. The SGPA is used instead of the classical scaled dot-product attention and directly performs an approximation of the Bayesian inference in the attention blocks in Transformers. In addition, the authors provide decoup... |
The work addresses the issue of oversquashing in GNN in proposing a new approach of rewiring graphs so as to improve the spectral gap of the graph (as it has been studied that spectral gap is a key feature in preventing oversquashing). The proposed rewiring algorithm comes from a study of perturbation of the Fiedler ve... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The work addresses the issue of oversquashing in GNN in proposing a new approach of rewiring graphs so as to improve the spectral gap of the graph (as it has been studied that spectral gap is a key feature in preventing oversquashing). The proposed rewiring algorithm comes from a study of perturbation of the Fi... |
This paper proposes to use heat dissipation to define a forward process (blurring) in the diffusion model. It uses eigendecomposition to decompose the Laplace operation and uses a variational model, similar to DDPM to learn the generative process. Gaussian noise is added to ensure the diversity of the generated images.... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper proposes to use heat dissipation to define a forward process (blurring) in the diffusion model. It uses eigendecomposition to decompose the Laplace operation and uses a variational model, similar to DDPM to learn the generative process. Gaussian noise is added to ensure the diversity of the generated... |
The authors propose DiffDock, an approach to global molecular docking as a generative modeling problem that aims to map the non-Euclidean manifold of ligand poses to the degrees of freedom involved in docking sampling by developing a diffusion process on this space. They propose an objective formalism that aims to maxi... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose DiffDock, an approach to global molecular docking as a generative modeling problem that aims to map the non-Euclidean manifold of ligand poses to the degrees of freedom involved in docking sampling by developing a diffusion process on this space. They propose an objective formalism that aims... |
This paper proposes to learn object-centric representation with an energy-based model. This energy model takes a visual scene and a set of object-centric latent variables as input. Latent variables are inferred from visual observations through gradient-based MCMC sampling, where the gradient is derived from the energy ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to learn object-centric representation with an energy-based model. This energy model takes a visual scene and a set of object-centric latent variables as input. Latent variables are inferred from visual observations through gradient-based MCMC sampling, where the gradient is derived from the... |
This paper presents Decision Diffuser, a conditional generative model for sequential decision making. It frames offline sequential decision making as conditional generative modeling by considering two other variables: constraints and skills. Conditioning on a single constraint or skill during training leads to behavio... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents Decision Diffuser, a conditional generative model for sequential decision making. It frames offline sequential decision making as conditional generative modeling by considering two other variables: constraints and skills. Conditioning on a single constraint or skill during training leads to... |
The paper addresses the problem of efficiently understanding the relationship between the input-output mapping of a neural network over a large and representative set of input spaces. To do so, the authors draw an analogy between a neural network model and physics's density of states problem. In doing so, they can desi... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper addresses the problem of efficiently understanding the relationship between the input-output mapping of a neural network over a large and representative set of input spaces. To do so, the authors draw an analogy between a neural network model and physics's density of states problem. In doing so, they ... |
This paper first proposes a continuous-time framework for quantifying the speed of decision boundary changes and empirically shows conflicting dynamics in adversarial training: the decision boundary moves closer to some examples even if the model just learns these examples. This paper proposes a novel adversarial train... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper first proposes a continuous-time framework for quantifying the speed of decision boundary changes and empirically shows conflicting dynamics in adversarial training: the decision boundary moves closer to some examples even if the model just learns these examples. This paper proposes a novel adversari... |
The paper touches on an important problem in embodied AI, object affordance understanding. The authors formulate the object affordance in a grasp pose generation paradigm. The grasp pose is predicted by a three-stage pipeline: A contact map is first predicted by a cVAE and then is input to a GraspNet to obtain the hand... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper touches on an important problem in embodied AI, object affordance understanding. The authors formulate the object affordance in a grasp pose generation paradigm. The grasp pose is predicted by a three-stage pipeline: A contact map is first predicted by a cVAE and then is input to a GraspNet to obtain ... |
This paper proposes a method to compute temporal saliency in probabilistic forecasting. The saliency of each timestamp in time-series data is measured based on its gradient w.r.t. the output entropy. Authors construct a synthesized dataset to verify the effectiveness of the proposed method.
[Strength]
1. This paper fo... | 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 proposes a method to compute temporal saliency in probabilistic forecasting. The saliency of each timestamp in time-series data is measured based on its gradient w.r.t. the output entropy. Authors construct a synthesized dataset to verify the effectiveness of the proposed method.
[Strength]
1. This ... |
The paper studies a very interesting and novel problem setting: how to split the data in a way that when a model is trained on one partition cannot generalize to the second partition: while this task does not make realistic sense, careful studying of it will surely reveal many other properties that are important in the... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper studies a very interesting and novel problem setting: how to split the data in a way that when a model is trained on one partition cannot generalize to the second partition: while this task does not make realistic sense, careful studying of it will surely reveal many other properties that are importan... |
The paper proposes a new optimization algorithm for deep neural networks that incorporates aspects of curvature information via meta-learning. In particular, authors model the inverse of the Fisher information matrix that appears in natural gradient descent via a set of additional parameters that are optimized jointly ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes a new optimization algorithm for deep neural networks that incorporates aspects of curvature information via meta-learning. In particular, authors model the inverse of the Fisher information matrix that appears in natural gradient descent via a set of additional parameters that are optimized ... |
The authors proposed a method to obtain a good and fast explanation at the testing time. In particular, the authors focused on the approach that learns an explainer model to mimic the “accurate explanation” at test time (e.g., predict the Shapley value).
The authors pointed out that a major issue of existing methods is... | 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 authors proposed a method to obtain a good and fast explanation at the testing time. In particular, the authors focused on the approach that learns an explainer model to mimic the “accurate explanation” at test time (e.g., predict the Shapley value).
The authors pointed out that a major issue of existing me... |
This paper proposes “context distillation”, a method to “internalize” the information provided by abstract instructions, concrete demonstration examples, or scratchpad (model intermediate output that helps reasoning). They first use a language model to sample some input (e.g., movie reviews), and then they use the teac... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes “context distillation”, a method to “internalize” the information provided by abstract instructions, concrete demonstration examples, or scratchpad (model intermediate output that helps reasoning). They first use a language model to sample some input (e.g., movie reviews), and then they use ... |
This paper tackles the emerging topic of graph contrastive learning in recommender system, by proposing a lightweight and effective method. Particularly, the augmented view is automatically generated based on singular value decomposition technique, which is interesting and novel to me. With the augmented user-item grap... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper tackles the emerging topic of graph contrastive learning in recommender system, by proposing a lightweight and effective method. Particularly, the augmented view is automatically generated based on singular value decomposition technique, which is interesting and novel to me. With the augmented user-i... |
This paper provides a theoretical analysis of Inverse soft Q-Learn (IQ-Learn) and proposes a novel algorithm named Least Squares Inverse Q-learning (LS-IQ), which outperforms state-of-the-art algorithms such as GAIL, VAIL, IQ, and SQIL. At first, the authors show that the maximum entropy IR objective with the regulariz... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper provides a theoretical analysis of Inverse soft Q-Learn (IQ-Learn) and proposes a novel algorithm named Least Squares Inverse Q-learning (LS-IQ), which outperforms state-of-the-art algorithms such as GAIL, VAIL, IQ, and SQIL. At first, the authors show that the maximum entropy IR objective with the r... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.