review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper follow the equivariant representation learning literature and proposed a equivariant representation learning framework for one-agent-one-object environment.
By assuming that action additively change the state of the agent and the state of the object can only be changed by agent upon impact, the proposed ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper follow the equivariant representation learning literature and proposed a equivariant representation learning framework for one-agent-one-object environment.
By assuming that action additively change the state of the agent and the state of the object can only be changed by agent upon impact, the p... |
The paper proposed a new Boundary Connectivity (BCXN) loss function for PINNs for solving PDEs with complex geometry. BCXN uses a linear interpolation to compute the points outside the domain to address the issue that the derivative stencil points are outside the domain, and then the estimated external points can be en... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposed a new Boundary Connectivity (BCXN) loss function for PINNs for solving PDEs with complex geometry. BCXN uses a linear interpolation to compute the points outside the domain to address the issue that the derivative stencil points are outside the domain, and then the estimated external points c... |
The paper explains the success of the alphahat metric [1] in predicting model generalisation: the alphahat metric consists of two components, (i) alpha, the power law exponent in the empirical spectral density, and (ii) the log spectral norm. The paper shows that these components result in a Simpson's paradox in the an... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper explains the success of the alphahat metric [1] in predicting model generalisation: the alphahat metric consists of two components, (i) alpha, the power law exponent in the empirical spectral density, and (ii) the log spectral norm. The paper shows that these components result in a Simpson's paradox i... |
This work proposes an object-centric 3D inference method that is trained in a self-supervised way using a variational auto-encoder framework and neural radience fields. While a large number of methods in the literature address object-centric self-supervised segmentation, very few also learn to infer 3D pose and shape w... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposes an object-centric 3D inference method that is trained in a self-supervised way using a variational auto-encoder framework and neural radience fields. While a large number of methods in the literature address object-centric self-supervised segmentation, very few also learn to infer 3D pose and... |
This paper introduces a new method for weakly supervised object detection with knowledge transfer, termed ProbKT. The basic idea is formulating the weakly supervised detection problem as the probabilistic logical reasoning so that the detector can be trained with arbitrary types of weak supervision. The output of the ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces a new method for weakly supervised object detection with knowledge transfer, termed ProbKT. The basic idea is formulating the weakly supervised detection problem as the probabilistic logical reasoning so that the detector can be trained with arbitrary types of weak supervision. The output... |
Given a function f : Z^d -> R, the authors construct tensor-train representations for the implicitly defined tensor. They first need to build a computational tree for the entries of the tensor that involves left and right derivative functions. From this, they show that they can construct a TT format for the tensor X_{i... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
Given a function f : Z^d -> R, the authors construct tensor-train representations for the implicitly defined tensor. They first need to build a computational tree for the entries of the tensor that involves left and right derivative functions. From this, they show that they can construct a TT format for the ten... |
Context-Dependent Gating (XDG), proposed by Masse, Grant, and Freedman, showed promise in dealing with catastrophic forgetting. In that work, gating is randomly chosen before training. This work follows up with the obvious next question: how to learn gating for different tasks. Authors show Learned Context Dependent Ga... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
Context-Dependent Gating (XDG), proposed by Masse, Grant, and Freedman, showed promise in dealing with catastrophic forgetting. In that work, gating is randomly chosen before training. This work follows up with the obvious next question: how to learn gating for different tasks. Authors show Learned Context Depe... |
The paper is in the line of works that uses reinforcement learning for combinatorial optimization (CO) problems. The main contribution is the idea of training a population of complementary agents for a given distribution of CO problems, with the objective function being optimizing the performance of the best agent inst... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper is in the line of works that uses reinforcement learning for combinatorial optimization (CO) problems. The main contribution is the idea of training a population of complementary agents for a given distribution of CO problems, with the objective function being optimizing the performance of the best ag... |
The authors study the offline RL setting with imperfect rewards. That is, a known portion of the trajectories contain possibly corrupted rewards. The approach proposed to tackle this problem is Reward Gap Minimization (RGM), which finds the reward function (correction) which, when used to solve for the optimal policy, ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors study the offline RL setting with imperfect rewards. That is, a known portion of the trajectories contain possibly corrupted rewards. The approach proposed to tackle this problem is Reward Gap Minimization (RGM), which finds the reward function (correction) which, when used to solve for the optimal ... |
This paper studies the cross-modality person re-identification problem and tries to build a new paradigm. Particularly, this paper proposes a large scale dataset called NPU-ReID that includes more identities and identities with even modality distribution. What's more, this paper presents 1) a modality augmentation meth... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studies the cross-modality person re-identification problem and tries to build a new paradigm. Particularly, this paper proposes a large scale dataset called NPU-ReID that includes more identities and identities with even modality distribution. What's more, this paper presents 1) a modality augmentat... |
This is a strong paper that proposes methods that right issues arising from concept bottleneck models. The authors show how to split concepts into explicit and implicit information, and then demonstrate how to debug concept and label errors.
Strengths
- The paper's presentation is strong. Figures are immaculate!
- 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:
This is a strong paper that proposes methods that right issues arising from concept bottleneck models. The authors show how to split concepts into explicit and implicit information, and then demonstrate how to debug concept and label errors.
Strengths
- The paper's presentation is strong. Figures are immaculate... |
The paper presents an improvement for MAE pre-training, that instead of reconstructing all the patches, it attempts to reconstruct only a certain number of diverse/important patches. The selection for diverse/important patches is done through a momentum encoder -- it will discard patches that are closest to the mean (a... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper presents an improvement for MAE pre-training, that instead of reconstructing all the patches, it attempts to reconstruct only a certain number of diverse/important patches. The selection for diverse/important patches is done through a momentum encoder -- it will discard patches that are closest to the... |
The authors present a method that performs massive uniform exploration of simulated environments and then trains goal-conditioned SAC with an annealed success condition to achieve states discovered during exploration. They demonstrate this exploration procedure to be superior on locomotion environments to other methods... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors present a method that performs massive uniform exploration of simulated environments and then trains goal-conditioned SAC with an annealed success condition to achieve states discovered during exploration. They demonstrate this exploration procedure to be superior on locomotion environments to other... |
This paper provides some insights on the potential reason that the initial large step sizes can learn the model that can generalize well in practice. Center to the claim is an SDE with specific noise covariance structure, and the authors argue that, if we use the large step sizes at the beginning, the loss should stabi... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper provides some insights on the potential reason that the initial large step sizes can learn the model that can generalize well in practice. Center to the claim is an SDE with specific noise covariance structure, and the authors argue that, if we use the large step sizes at the beginning, the loss shou... |
This paper studies adaptive client sampling in federated optimization and formulates it as the problem of online sampling variance minimization with bandit feedback. To address this problem, the authors propose an online stochastic mirror descent (OSMD) algorithm and establish dynamic regret bound. Moreover, the conver... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies adaptive client sampling in federated optimization and formulates it as the problem of online sampling variance minimization with bandit feedback. To address this problem, the authors propose an online stochastic mirror descent (OSMD) algorithm and establish dynamic regret bound. Moreover, th... |
This paper studies the a semi-supervised community detection problem on a degree-corrected stochastic block model. Given some known labels of the nodes in the network, the authors propose a method to classified the new node based on cosine similarity with the in-sample data.
Strength
This paper is easy to read. I can... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper studies the a semi-supervised community detection problem on a degree-corrected stochastic block model. Given some known labels of the nodes in the network, the authors propose a method to classified the new node based on cosine similarity with the in-sample data.
Strength
This paper is easy to rea... |
The paper introduces two pruning techniques removing non-linearities in a network while preserving local behavior, with an eye towards producing good estimates of the true robust distance $d_{bnd}$ more quickly. The paper empirically evaluates their performance by computing a relative error between "ground truth robust... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper introduces two pruning techniques removing non-linearities in a network while preserving local behavior, with an eye towards producing good estimates of the true robust distance $d_{bnd}$ more quickly. The paper empirically evaluates their performance by computing a relative error between "ground trut... |
The paper proposes a novel, simple, scalable and efficient pre-training approach for causal, decoder-only language models. The proposed approach is called Forgetful Causal Masking (FCM) and it randomly masks out past tokens. FCM pre-training is applied to PaLM to demonstrate significant improvement in the 0-shot perfor... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a novel, simple, scalable and efficient pre-training approach for causal, decoder-only language models. The proposed approach is called Forgetful Causal Masking (FCM) and it randomly masks out past tokens. FCM pre-training is applied to PaLM to demonstrate significant improvement in the 0-sho... |
In this paper, the authors propose to add a non-convex sparsity constraint to the $l^2$ regularized discrete OT problem. In such a way, they can control the sparsity of the OT plan. Though the sparsity constraint makes the original problem nonconvex, it is still tractable through the dual and semi-dual problem. With th... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors propose to add a non-convex sparsity constraint to the $l^2$ regularized discrete OT problem. In such a way, they can control the sparsity of the OT plan. Though the sparsity constraint makes the original problem nonconvex, it is still tractable through the dual and semi-dual problem.... |
This paper aims to capture co-playr learning dynamics in MARL. The authors propose Context and History Aware Other-Shaing (CHAOS) to address this problem. The CHAOS agent is a meta-learner using RNN architecture to learns to shape its co-player. The authors conduct extensive experiments on matrix games.
One of the main... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper aims to capture co-playr learning dynamics in MARL. The authors propose Context and History Aware Other-Shaing (CHAOS) to address this problem. The CHAOS agent is a meta-learner using RNN architecture to learns to shape its co-player. The authors conduct extensive experiments on matrix games.
One of ... |
The paper introduces a new architecture that relies on polynomial expansions. The idea is to use polynomial layers, where each layer expresses a polynomial expansion of order $k$ (where $k=2$ in practice). The new architecture uses a Tucker decomposition to reduce the learnable parameters. The authors also introduce a ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper introduces a new architecture that relies on polynomial expansions. The idea is to use polynomial layers, where each layer expresses a polynomial expansion of order $k$ (where $k=2$ in practice). The new architecture uses a Tucker decomposition to reduce the learnable parameters. The authors also intr... |
They proposed a transformation to multimodal extension from the model hijacking attack by Salem et al. (2022a). This work takes a data poisoning approach while the fused dataset is used to poison a victim model. The evaluation metrics are the attack success rate and utility, where both metrics are important to attain s... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
They proposed a transformation to multimodal extension from the model hijacking attack by Salem et al. (2022a). This work takes a data poisoning approach while the fused dataset is used to poison a victim model. The evaluation metrics are the attack success rate and utility, where both metrics are important to ... |
This paper propose a masked time series autoencoders which can learn strong representations with less inductive bias or hierarchical trick. The proposed Ti-MAE bridges the connection between contrastive representation learning and generative Transformer-based methods. Ti-MAE adequately leverages all the input sequence ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper propose a masked time series autoencoders which can learn strong representations with less inductive bias or hierarchical trick. The proposed Ti-MAE bridges the connection between contrastive representation learning and generative Transformer-based methods. Ti-MAE adequately leverages all the input s... |
The paper addresses how online RL can benefit from an existing dataset and/or a policy created by offline RL. The suggested method proposes that both the offline policy (fixed) and the online policy (to be trained) provide action candidates, which are then evaluated using a Q-function (to be learned online). The evalua... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper addresses how online RL can benefit from an existing dataset and/or a policy created by offline RL. The suggested method proposes that both the offline policy (fixed) and the online policy (to be trained) provide action candidates, which are then evaluated using a Q-function (to be learned online). Th... |
This paper proposes an architectures that separately encodes sensory information and the task id in a multi-task setting. These are then concatenated to processed by a neural network that produces the action. The paper imposes constraints on the task representation (or action representation in the paper): the represent... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an architectures that separately encodes sensory information and the task id in a multi-task setting. These are then concatenated to processed by a neural network that produces the action. The paper imposes constraints on the task representation (or action representation in the paper): the r... |
This paper proposes to incorporate energy-based priors to improve the training of physics-informed neural networks (PINN), for solving partial differential equation (PDE) based inverse problems. The method is mainly validated in the application of electrical impedance tomography (EIT).
Strength:
- The paper is well wri... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes to incorporate energy-based priors to improve the training of physics-informed neural networks (PINN), for solving partial differential equation (PDE) based inverse problems. The method is mainly validated in the application of electrical impedance tomography (EIT).
Strength:
- The paper is ... |
In this work, the authors propose a hierarchy (analogous to the WL hierarchy for graphs) $-$ for graphs with geometric features (Typically node features with coordinate information, velocity, etc) which can then be used to characterize the expressive power of equivariant and invariant geometric graph neural network lay... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this work, the authors propose a hierarchy (analogous to the WL hierarchy for graphs) $-$ for graphs with geometric features (Typically node features with coordinate information, velocity, etc) which can then be used to characterize the expressive power of equivariant and invariant geometric graph neural net... |
This paper proposed a data augmentation method PASTA for syn2real generalization. Based on the finding that real images exhibit larger high frequency variances than synthetic ones, PASTA perturbs the high frequency bands of the amplitude spectrum of the synthetic images. In doing so, the high frequency variance is larg... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a data augmentation method PASTA for syn2real generalization. Based on the finding that real images exhibit larger high frequency variances than synthetic ones, PASTA perturbs the high frequency bands of the amplitude spectrum of the synthetic images. In doing so, the high frequency variance... |
The paper proposes a greedy strategy to feature selection based on information maximization criterion.
Strengths:
--- Clear and easy to read
--- Empirical evaluations are interesting and useful.
Weaknesses:
--- The paper lacks solid theoretical justification. They cite Das/Kempe 2011 who analyze linear regression fu... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a greedy strategy to feature selection based on information maximization criterion.
Strengths:
--- Clear and easy to read
--- Empirical evaluations are interesting and useful.
Weaknesses:
--- The paper lacks solid theoretical justification. They cite Das/Kempe 2011 who analyze linear regre... |
The authors propose an approach to model the time dependencies, cross-channel dependencies, and stochasticity for multivariate time series (MTS) modeling. Specifically, an adaptive graph Transformer is first designed to learn the cross-channel relationships in MTS. To model the stochasticity, a channel embedding guided... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors propose an approach to model the time dependencies, cross-channel dependencies, and stochasticity for multivariate time series (MTS) modeling. Specifically, an adaptive graph Transformer is first designed to learn the cross-channel relationships in MTS. To model the stochasticity, a channel embeddin... |
This paper discussed the transfer learning in relatively relaxed conditions. Particularly, the paper assumes (1) no label overlapping space between source and target tasks, (2) no source dataset access when training the target task, and (3) the inconsistent model architectures.
To tackle the transfer learning problem ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper discussed the transfer learning in relatively relaxed conditions. Particularly, the paper assumes (1) no label overlapping space between source and target tasks, (2) no source dataset access when training the target task, and (3) the inconsistent model architectures.
To tackle the transfer learning ... |
The paper presents an approach for learning POMDP models from data based on the idea of factoring the transition function of the POMDP into a product of two mappings, going trough a low-dimensional bottleneck layer that will effectively represent the hidden state. Although this is a sound idea, there is no empirical ve... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents an approach for learning POMDP models from data based on the idea of factoring the transition function of the POMDP into a product of two mappings, going trough a low-dimensional bottleneck layer that will effectively represent the hidden state. Although this is a sound idea, there is no empi... |
This paper illustrates a new generative foundation model for multi-property optimization in molecular de novo design. The authors leverages transformer based self-supervised learning to pre-train the model with carefully designed input sequence. The input token sequence contains multiple prompts which are human interpr... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper illustrates a new generative foundation model for multi-property optimization in molecular de novo design. The authors leverages transformer based self-supervised learning to pre-train the model with carefully designed input sequence. The input token sequence contains multiple prompts which are human... |
This work proposes a variant of BERT-like pretraining method for computer vision (the so-called Masked Image Modeling). The method can generalize well to both vision transformers and CNNs, thus is described as architecture agnostic.
The main contributions of this work are:
- using the mean color values to fill the cor... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This work proposes a variant of BERT-like pretraining method for computer vision (the so-called Masked Image Modeling). The method can generalize well to both vision transformers and CNNs, thus is described as architecture agnostic.
The main contributions of this work are:
- using the mean color values to fill... |
This paper targets to solve real-world robotic manipulation tasks by leveraging a large demonstration dataset and a small amount of target task demonstrations.
The proposed method first pre-train a policy on the large dataset with multi-task offline RL, then fine-tune on the small demonstration data again with same obj... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper targets to solve real-world robotic manipulation tasks by leveraging a large demonstration dataset and a small amount of target task demonstrations.
The proposed method first pre-train a policy on the large dataset with multi-task offline RL, then fine-tune on the small demonstration data again with ... |
The paper visually analyzes the behavior of ViT and CLIP models. The paper also shows the difference between ViT and CNNs and provides many observations.
Strength
---
- To my knowledge, this is the first work that visualizes the features of transformers.
Weakness
---
- The paper is missing references that study 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 visually analyzes the behavior of ViT and CLIP models. The paper also shows the difference between ViT and CNNs and provides many observations.
Strength
---
- To my knowledge, this is the first work that visualizes the features of transformers.
Weakness
---
- The paper is missing references that st... |
This paper proposes a neural network-based equalizer that mimics the forward-backward algorithm for wireline communication. The authors also tried pruning the neural network.
Strength:
The proposed solution has lower computational complexity than the forward-backward (FB) algorithm, while the BER performance is better ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a neural network-based equalizer that mimics the forward-backward algorithm for wireline communication. The authors also tried pruning the neural network.
Strength:
The proposed solution has lower computational complexity than the forward-backward (FB) algorithm, while the BER performance is... |
This paper adresses data-free model stealing, an ambitious framework where one attempts to build a model that imitates a target one without accessing the target parameter or any training data point (true or substitute ones). In this field, SOTA approaches use a student network whose outputs are meant to minimize a dist... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper adresses data-free model stealing, an ambitious framework where one attempts to build a model that imitates a target one without accessing the target parameter or any training data point (true or substitute ones). In this field, SOTA approaches use a student network whose outputs are meant to minimiz... |
The paper proposes a method of transfer based on the assumption of common dynamics and variable reward signals. The method leverages artificial pseudo-rewards to span a space of probably $Q$-functions related to the down-stream tasks.
Strengths:
- well-motivated and important problem
- sound method
Weaknesses:
- sca... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a method of transfer based on the assumption of common dynamics and variable reward signals. The method leverages artificial pseudo-rewards to span a space of probably $Q$-functions related to the down-stream tasks.
Strengths:
- well-motivated and important problem
- sound method
Weaknesse... |
This paper introduces a diffusion adversarial representation learning (DARL) model, for self-supervised vessel segmentation. The major novelty appears to be the usage of a diffusion module for estimating latent features, which are then used by a generation model to estimate both vessel segmentation masks and synthetic ... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper introduces a diffusion adversarial representation learning (DARL) model, for self-supervised vessel segmentation. The major novelty appears to be the usage of a diffusion module for estimating latent features, which are then used by a generation model to estimate both vessel segmentation masks and sy... |
Using Othello game as the testbed for their research, the paper addresses a challenging question of whether transformers learn reasonable world-state representations, or simply exploit surface-level statistics of the data. First, the authors identify that the board state is recoverable from the network's representation... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
Using Othello game as the testbed for their research, the paper addresses a challenging question of whether transformers learn reasonable world-state representations, or simply exploit surface-level statistics of the data. First, the authors identify that the board state is recoverable from the network's repres... |
This paper proposes a Distributional return Meta-Gradient algorithm (DrMG) algorithm, which combines distributional RL (like C51) and meta-gradient RL. Specifically, the authors formulate a distributional return in a meta-gradient RL setup and derive a meta-update rule to learn the adaptive distributional return with m... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a Distributional return Meta-Gradient algorithm (DrMG) algorithm, which combines distributional RL (like C51) and meta-gradient RL. Specifically, the authors formulate a distributional return in a meta-gradient RL setup and derive a meta-update rule to learn the adaptive distributional retur... |
This paper proposed a novel method for knowledge distillation in multi-label learning scenarios. Multi-label learning aims to solve the problem where multiple positive labels exists in a given single sample. However, general knowledge distillation methods mainly focus on single-label learning scenario while ignoring th... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed a novel method for knowledge distillation in multi-label learning scenarios. Multi-label learning aims to solve the problem where multiple positive labels exists in a given single sample. However, general knowledge distillation methods mainly focus on single-label learning scenario while ign... |
The paper presents a method to take advantage of the weaknesses of the Go program KataGo. While the method is fairly elaborate using MCTS and having gray access to KataGo, the winning strategy is not based on superior play, but on the exploitation of a specific rule regarding the termination/scoring of the game.
Find... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents a method to take advantage of the weaknesses of the Go program KataGo. While the method is fairly elaborate using MCTS and having gray access to KataGo, the winning strategy is not based on superior play, but on the exploitation of a specific rule regarding the termination/scoring of the game... |
The work couples an RNN with a latent space representation that can infer the current task in a continual learning setup. The latent space is inspired by the role of thalamus, and enables the RNN to adapt to new tasks by changing the latent representation rather than the internal dynamics of the RNN.
### Strengths
A ... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The work couples an RNN with a latent space representation that can infer the current task in a continual learning setup. The latent space is inspired by the role of thalamus, and enables the RNN to adapt to new tasks by changing the latent representation rather than the internal dynamics of the RNN.
### Stren... |
The paper studies the problem of reward free learning, where the setup involves a pre-training stage (when an intrinsic reward is used based on different prior methods) and a fine-tuning stage (when a particular extrinsic reward is provided). The claim is that model-based methods perform better than model-free methods ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies the problem of reward free learning, where the setup involves a pre-training stage (when an intrinsic reward is used based on different prior methods) and a fine-tuning stage (when a particular extrinsic reward is provided). The claim is that model-based methods perform better than model-free ... |
This proposes the unsupervised 2D-to-3D lifting approach by exprimenting with 5 network structures: 1) Popularly used 2D-to-3D lifting model, 2) extracting features from body and legs and estimate 3D full pose by concatenating 2 features, 3) Separately estimating poses for body and legs, 4) Separately estimate 5 part f... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This proposes the unsupervised 2D-to-3D lifting approach by exprimenting with 5 network structures: 1) Popularly used 2D-to-3D lifting model, 2) extracting features from body and legs and estimate 3D full pose by concatenating 2 features, 3) Separately estimating poses for body and legs, 4) Separately estimate ... |
This paper proposed to use a pre-trained diffusion model as a generative prior for BFR. With a diffused estimator, the LQ images with different degradation can be transformed into the latent space of the diffusion model. A weighted L2 loss has been proposed to train the diffused estimator. Generally, the paper is well-... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed to use a pre-trained diffusion model as a generative prior for BFR. With a diffused estimator, the LQ images with different degradation can be transformed into the latent space of the diffusion model. A weighted L2 loss has been proposed to train the diffused estimator. Generally, the paper ... |
This paper proposes task vectors, which are subtraction of pre-trained weights from the weights fine-tuned on downstream tasks,
and perform add or subtract the task vector to the pre-trained weight to steer pre-trained models. With the task vector, we can remove undesirable behavior of pre-trained model (e.g. generati... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes task vectors, which are subtraction of pre-trained weights from the weights fine-tuned on downstream tasks,
and perform add or subtract the task vector to the pre-trained weight to steer pre-trained models. With the task vector, we can remove undesirable behavior of pre-trained model (e.g. ... |
This work studies how and to what degree deep networks perform memorization of samples in the clean label setting. The authors build upon previous work and employ the stability-based memorization score developped in [1] which measures how much the prediction of a
model on a given sample changes if the sample is present... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work studies how and to what degree deep networks perform memorization of samples in the clean label setting. The authors build upon previous work and employ the stability-based memorization score developped in [1] which measures how much the prediction of a
model on a given sample changes if the sample is... |
This paper addresses a generalization problem in RL, in which the agent interacts with a set of MDPs taken from an unknown task distribution during training, and subsequently minimizes the (expectation) of the regret in a test MDP taken from the same distribution. The paper first studies an asymptotic setting in which ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper addresses a generalization problem in RL, in which the agent interacts with a set of MDPs taken from an unknown task distribution during training, and subsequently minimizes the (expectation) of the regret in a test MDP taken from the same distribution. The paper first studies an asymptotic setting i... |
Summary:
This paper examines multi-class classification in settings with out-of-distribution (OOD) data. Here, this refers to test data where the labels are not available at training time. The proposed solution is hierarchical classification, with the leaf nodes of the tree being the classes seen in training. The tree... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Summary:
This paper examines multi-class classification in settings with out-of-distribution (OOD) data. Here, this refers to test data where the labels are not available at training time. The proposed solution is hierarchical classification, with the leaf nodes of the tree being the classes seen in training. ... |
This paper proposes a method for improving robustness to common corruptions via L2-norm bounded data augmentations of varying sizes. The authors show theoretically that data augmentation of this form results in models with flatter minima in parameter space, which has been shown to result in better generalization. This ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method for improving robustness to common corruptions via L2-norm bounded data augmentations of varying sizes. The authors show theoretically that data augmentation of this form results in models with flatter minima in parameter space, which has been shown to result in better generalizatio... |
This paper studies how to address the issue of hyper-parameter selection in unsupervised domain adaptation. Specifically, the authors propose a method that subsequently computes a linear aggregation of the models. In addition, theories for bounding the target error are added to make the method convincing. Experiments s... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies how to address the issue of hyper-parameter selection in unsupervised domain adaptation. Specifically, the authors propose a method that subsequently computes a linear aggregation of the models. In addition, theories for bounding the target error are added to make the method convincing. Exper... |
This paper presents near-optimal coresets for robust Euclidean clustering. Given k, z, m, and a set of n points P in R^d, we look for a k-clustering of P excluding a set of outliers L of size at most m, minimizing the sum of (power z of) distances of points to their centers. This gives k-median and k-means objectives f... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper presents near-optimal coresets for robust Euclidean clustering. Given k, z, m, and a set of n points P in R^d, we look for a k-clustering of P excluding a set of outliers L of size at most m, minimizing the sum of (power z of) distances of points to their centers. This gives k-median and k-means obje... |
This paper focuses on helping deep learning models learn fundamental geometric transformations efficiently. Specifically, LatFormer is proposed to incorporate lattice symmetry biases into attention mechanisms by modulating the attention weights using learned soft masks. Experiments are conducted on both synthetic tasks... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper focuses on helping deep learning models learn fundamental geometric transformations efficiently. Specifically, LatFormer is proposed to incorporate lattice symmetry biases into attention mechanisms by modulating the attention weights using learned soft masks. Experiments are conducted on both synthet... |
This works studies the relation between policy evaluation and policy improvement in the context where the value-function(s) and the policy share the same parameterization, and suggests a way to consolidates the two type of updates.
The basic ideas underlying the paper are interesting and could potentially be a valid c... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This works studies the relation between policy evaluation and policy improvement in the context where the value-function(s) and the policy share the same parameterization, and suggests a way to consolidates the two type of updates.
The basic ideas underlying the paper are interesting and could potentially be a... |
The paper addresses the limits of GAN based vocoders to generalize to new/different voices not seen in the training set.
It does this by scaling up the generator block of the model an using periodic activations. Specifically the paper call for using the SNAKE activation function, in combination with 2x upsampling and d... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper addresses the limits of GAN based vocoders to generalize to new/different voices not seen in the training set.
It does this by scaling up the generator block of the model an using periodic activations. Specifically the paper call for using the SNAKE activation function, in combination with 2x upsampli... |
The authors propose a novel concept-based explanation method: Concept Gradient (CG). In contrast to previous work, they use non-linear concept functions and show how the standard Concept Activation Vector (CAV) approach can be generalized to this setting.
Experimental results show superior performance of CG over CAV.
#... | 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 propose a novel concept-based explanation method: Concept Gradient (CG). In contrast to previous work, they use non-linear concept functions and show how the standard Concept Activation Vector (CAV) approach can be generalized to this setting.
Experimental results show superior performance of CG ove... |
This paper focus on the novel view synthesis with only sparse inputs. The authors propose a method called GeCo-NeRF that enforces geometric consistency to regularize the training of NeRF. Specifically, input images are warped to unseen viewpoints for supervision at feature level. They also filter out erroneous warped p... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper focus on the novel view synthesis with only sparse inputs. The authors propose a method called GeCo-NeRF that enforces geometric consistency to regularize the training of NeRF. Specifically, input images are warped to unseen viewpoints for supervision at feature level. They also filter out erroneous ... |
The paper proposes a solution for federated semi-supervised learning. the proposal lies averaging the manifold in the decision space over multiple clients. The sharing is done using two cryptographical protocols. The paper evaluates on three datasets
Pros:
- The idea to use manifold for pseudo-labeling has not been ... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a solution for federated semi-supervised learning. the proposal lies averaging the manifold in the decision space over multiple clients. The sharing is done using two cryptographical protocols. The paper evaluates on three datasets
Pros:
- The idea to use manifold for pseudo-labeling has n... |
This paper studies popular domain adaptation methods under scenarios where both label distribution and conditionals may shift. Specifically, the authors introduce a large-scale benchmark for relaxed label shift settings which consists of 11 vision datasets spanning more than 200 distribution shift pairs. 12 popular dom... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper studies popular domain adaptation methods under scenarios where both label distribution and conditionals may shift. Specifically, the authors introduce a large-scale benchmark for relaxed label shift settings which consists of 11 vision datasets spanning more than 200 distribution shift pairs. 12 pop... |
This work suggests that the attention matrix, a key module in Transformer but computationally inefficient, can be approximated by low-rank matrix multiplications. Given the arguments, the authors proposed a new efficient Transformer variant using random projections. Experiments, including popularly used LRA benchmarks,... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work suggests that the attention matrix, a key module in Transformer but computationally inefficient, can be approximated by low-rank matrix multiplications. Given the arguments, the authors proposed a new efficient Transformer variant using random projections. Experiments, including popularly used LRA ben... |
The paper presents an idea to select important examples to be buffered in continual learning. The proposed Consistency Aware Sampling (CAWS) considers examples that are easy to learn and representative of previous tasks. Empirical investigation has been conducted to evaluate the effectiveness of the proposed method.
Th... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper presents an idea to select important examples to be buffered in continual learning. The proposed Consistency Aware Sampling (CAWS) considers examples that are easy to learn and representative of previous tasks. Empirical investigation has been conducted to evaluate the effectiveness of the proposed me... |
This paper explores further scaling up preconditioned conjugate gradient (PC-CG) based GP regression by parallelizing the computation of the Gram matrix as a one-off and storing it to disk in "sub-blocks" to be loaded in the RAM in a multi-threaded fashion as/when needed.
This contrasts with standard PC-CG where the ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper explores further scaling up preconditioned conjugate gradient (PC-CG) based GP regression by parallelizing the computation of the Gram matrix as a one-off and storing it to disk in "sub-blocks" to be loaded in the RAM in a multi-threaded fashion as/when needed.
This contrasts with standard PC-CG wh... |
This paper proposes a hybrid framework for modeling N-body systems. It uses a classical solver for the non-thermal particle distribution. The neural networks are used in two components: (1) in modeling the thermal part using a continuum representation, (2) in coupling the continuum model and the particle model. The pro... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a hybrid framework for modeling N-body systems. It uses a classical solver for the non-thermal particle distribution. The neural networks are used in two components: (1) in modeling the thermal part using a continuum representation, (2) in coupling the continuum model and the particle model.... |
This paper focused on the message-passing schemes of a GNN model. It integrates the anisotropic state based on Cartesian multiples to the message-passing models to compensate for the hidden features. Such operation enhances the interactions, i.e., (1) anisotropic long-range, (2) the surrounding fields and particles not... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper focused on the message-passing schemes of a GNN model. It integrates the anisotropic state based on Cartesian multiples to the message-passing models to compensate for the hidden features. Such operation enhances the interactions, i.e., (1) anisotropic long-range, (2) the surrounding fields and parti... |
The paper proposes the primal-dual framework to derive popular self-attention mechanisms. The framework bases on a support vector expansion by solving a support vector regression problem. The paper shows that several popular self-attention mechanisms can be derived under the framework. It also invents two new self-atte... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes the primal-dual framework to derive popular self-attention mechanisms. The framework bases on a support vector expansion by solving a support vector regression problem. The paper shows that several popular self-attention mechanisms can be derived under the framework. It also invents two new s... |
This work extensively investigates continuous learning based methods for prequential minimum description length.
More precisely, the authors train MDL models using online learning with rehearsal, and propose two techniques for improving the results.
- Forward-calibration, to optimize a calibration parameter $\beta$ ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work extensively investigates continuous learning based methods for prequential minimum description length.
More precisely, the authors train MDL models using online learning with rehearsal, and propose two techniques for improving the results.
- Forward-calibration, to optimize a calibration parameter ... |
The paper proposes a new method for solving L1 regularized problems using redundant parametrization. Theoretical results show that the regularization on the redundant parameters is equivalent to L1 (overall, and group-wise). These results are then applied towards various use-cases ranging from simple models (Lasso) to ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new method for solving L1 regularized problems using redundant parametrization. Theoretical results show that the regularization on the redundant parameters is equivalent to L1 (overall, and group-wise). These results are then applied towards various use-cases ranging from simple models (La... |
I thank the authors for providing a thoughtful reply addressing most of my concerns. Together with the clarifications I am now convinced that this work is a good contribution. I have adjusted my score accordingly.
The paper proposes a biologically plausible source separation method with local updates and batched onlin... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
I thank the authors for providing a thoughtful reply addressing most of my concerns. Together with the clarifications I am now convinced that this work is a good contribution. I have adjusted my score accordingly.
The paper proposes a biologically plausible source separation method with local updates and batch... |
This paper investigated an important research problem in point process modeling and aimed to model multi-class event sequence in a network. As the paper pointed out, one of the recent work Omi et al. (2019) is known to be computationally efficient and flexible in representing complex triggering effects between events, ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigated an important research problem in point process modeling and aimed to model multi-class event sequence in a network. As the paper pointed out, one of the recent work Omi et al. (2019) is known to be computationally efficient and flexible in representing complex triggering effects between ... |
This paper analyzes the accuracy of PrivBayes by giving upper bounds on the distances between real and synthetic data and on utility errors of synthetic data from downstream supervised learning tasks. The authors also proved the lower bound for total variation distance for \epsilon-DP synthetic data generator.
The aut... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper analyzes the accuracy of PrivBayes by giving upper bounds on the distances between real and synthetic data and on utility errors of synthetic data from downstream supervised learning tasks. The authors also proved the lower bound for total variation distance for \epsilon-DP synthetic data generator.... |
This paper proposed a spatio-temporal entropy score (STEntr-Score) with a refinement factor to handle the discrepancy of visual information in spatial and temporal dimensions, through dynamically leveraging the correlation between the feature map size and kernel size depth-wisely. Then the proposed entropy-based 3D CNN... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a spatio-temporal entropy score (STEntr-Score) with a refinement factor to handle the discrepancy of visual information in spatial and temporal dimensions, through dynamically leveraging the correlation between the feature map size and kernel size depth-wisely. Then the proposed entropy-base... |
This paper studies convex formulations of the problem of training parallel ReLU networks.
These are models of the form \sum_{i=1}^K f_i(x; \theta^i), where each f_i is a ReLU network.
The problem gives algorithms for exactly/approximately learning these networks with path regularization.
The result is by a convex fo... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies convex formulations of the problem of training parallel ReLU networks.
These are models of the form \sum_{i=1}^K f_i(x; \theta^i), where each f_i is a ReLU network.
The problem gives algorithms for exactly/approximately learning these networks with path regularization.
The result is by a c... |
This paper analyzes offline policy learning in a confounded setting for contextual bandits and sequential decision-making. The paper introduces the notion of super policy learning in the setup where expert-recommended actions are available at each step in addition to the past historical data. Under this setup, the auth... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper analyzes offline policy learning in a confounded setting for contextual bandits and sequential decision-making. The paper introduces the notion of super policy learning in the setup where expert-recommended actions are available at each step in addition to the past historical data. Under this setup, ... |
The paper considers using VAEs for out of distribution (OOD) detection . The authors argue that the standard normal prior for latent variables is too concentrated to allow a proper representation of "in distribution" data and propose to use instead a distribution that is concentrated around a sphere in the latent space... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper considers using VAEs for out of distribution (OOD) detection . The authors argue that the standard normal prior for latent variables is too concentrated to allow a proper representation of "in distribution" data and propose to use instead a distribution that is concentrated around a sphere in the late... |
This work aims to provide a new perspective to explain the robust overfitting and improve the adversarial robustness built upon the proposed viewpoint. To do so, the authors revisit and generalize the static framework of feature robustness. Built upon the proposed framework, the authors claim that the balance between t... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work aims to provide a new perspective to explain the robust overfitting and improve the adversarial robustness built upon the proposed viewpoint. To do so, the authors revisit and generalize the static framework of feature robustness. Built upon the proposed framework, the authors claim that the balance b... |
This work proposes a technique called AudioPure where pre-trained diffusion-based speech denoisers are used for adversarial defense. The diffusion model is intended to be used in forward and backward steps to arrive from adversarial signal to purified audio. The paper proposes to use DiffWave and DiffSpec to be able to... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work proposes a technique called AudioPure where pre-trained diffusion-based speech denoisers are used for adversarial defense. The diffusion model is intended to be used in forward and backward steps to arrive from adversarial signal to purified audio. The paper proposes to use DiffWave and DiffSpec to be... |
Recently, FGSM-based adversarial training faced an emergency problem of catastrophic overfitting. To mitigate this, the authors first observed that the catastrophic overfitting is related to some abnormal examples, whose loss decreases under the FGSM attack. Based on this observation, in this paper they proposed a new ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
Recently, FGSM-based adversarial training faced an emergency problem of catastrophic overfitting. To mitigate this, the authors first observed that the catastrophic overfitting is related to some abnormal examples, whose loss decreases under the FGSM attack. Based on this observation, in this paper they propose... |
The paper proposes a heuristic pipeline to improve existing active learning methods using controllable augmentation.
Strength:
1. The proposed method applies to a wide range of augmentation methods and acquisition methods.
2. The authors clearly explain the heuristics in the proposed algorithm.
Weakness:
1. My primar... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper proposes a heuristic pipeline to improve existing active learning methods using controllable augmentation.
Strength:
1. The proposed method applies to a wide range of augmentation methods and acquisition methods.
2. The authors clearly explain the heuristics in the proposed algorithm.
Weakness:
1. M... |
This paper presents the "Planning Exploratory Goals" algorithm, which identifies goal states that are likely to result in many novel observations when used to initialize an subsequent explore-phase upon reaching a termination condition en route to the aforementioned goal. The algorithm overcomes prior issues, such as w... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents the "Planning Exploratory Goals" algorithm, which identifies goal states that are likely to result in many novel observations when used to initialize an subsequent explore-phase upon reaching a termination condition en route to the aforementioned goal. The algorithm overcomes prior issues, s... |
This paper aims to perform parameter-efficient adaption for vision tasks, where only a small number of parameters are trained to adapt to new tasks (image classification tasks in this paper). To address this problem, the authors propose to exploit the estimated sensitivity to decide which parameters to be tuned, and th... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper aims to perform parameter-efficient adaption for vision tasks, where only a small number of parameters are trained to adapt to new tasks (image classification tasks in this paper). To address this problem, the authors propose to exploit the estimated sensitivity to decide which parameters to be tuned... |
The paper studies molecule generation, proposing a new model using a retrieval mechanism, which has recently shown to be successful in NLP.
The method is evaluated on several established benchmarks and reaches performance comparable to state of the art
strengths
- comprehensive evaluation
- new, reasonably motivated m... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper studies molecule generation, proposing a new model using a retrieval mechanism, which has recently shown to be successful in NLP.
The method is evaluated on several established benchmarks and reaches performance comparable to state of the art
strengths
- comprehensive evaluation
- new, reasonably mot... |
The paper proposes to meta-learn an instance-level weight for real labels and pseudo-labels when learning with noisy labels.
Pros
1. The paper is well-written and easy to follow.
2. Theoretical guarantee of the convergence of the proposed method is provided.
Cons
My major concern is that the proposed method is n... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes to meta-learn an instance-level weight for real labels and pseudo-labels when learning with noisy labels.
Pros
1. The paper is well-written and easy to follow.
2. Theoretical guarantee of the convergence of the proposed method is provided.
Cons
My major concern is that the proposed met... |
In this paper, the authors propose to learn the radius in ball query with gumble-sigmoid trick. Experiments on both KITTI and Waymo show the algorithm is more efficient and effective than previous state-of-the-art point-based 3d object detectors.
However, the core contribution of this paper, the dynamic ball query, is ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose to learn the radius in ball query with gumble-sigmoid trick. Experiments on both KITTI and Waymo show the algorithm is more efficient and effective than previous state-of-the-art point-based 3d object detectors.
However, the core contribution of this paper, the dynamic ball qu... |
A new method based on Gaussian distribution with AINet is proposed to address the problem of noisy samples in few-shot learning. More Specifically, a novel AINet is developed to focus on query-related support samples to decrease the weight of less informative samples. In addition, Yeo-Johnson transformation is further ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
A new method based on Gaussian distribution with AINet is proposed to address the problem of noisy samples in few-shot learning. More Specifically, a novel AINet is developed to focus on query-related support samples to decrease the weight of less informative samples. In addition, Yeo-Johnson transformation is ... |
The paper considers infinite-horizon discounted Markov decision process with finite state and action space. By re-formulating the natural policy gradient (NPG) and the Q-NPG methods as approximate versions of the policy mirror descent (PMD) method, the paper shows that both methods enjoy linear convergence rates and \t... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper considers infinite-horizon discounted Markov decision process with finite state and action space. By re-formulating the natural policy gradient (NPG) and the Q-NPG methods as approximate versions of the policy mirror descent (PMD) method, the paper shows that both methods enjoy linear convergence rate... |
This paper proposes the method of few-shot prompting for internet-augmented language models consisting of three steps: 1) retrieval, which searches top 50 paragraphs using Google search, 2) prompting, where four prompts are applied on each of top 50 retrieved paragraphs to generate 200 candidate answers in total, 3) re... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes the method of few-shot prompting for internet-augmented language models consisting of three steps: 1) retrieval, which searches top 50 paragraphs using Google search, 2) prompting, where four prompts are applied on each of top 50 retrieved paragraphs to generate 200 candidate answers in tota... |
This paper discusses an improved training procedure to solve GFlowNet problems. The paper begins by outlining GFlowNets, a DAG with exactly one root node where "actions" are decisions about which edge to follow in the DAG. When the agent reaches a terminal node (a node with no edges leaving that node), the "episode" en... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper discusses an improved training procedure to solve GFlowNet problems. The paper begins by outlining GFlowNets, a DAG with exactly one root node where "actions" are decisions about which edge to follow in the DAG. When the agent reaches a terminal node (a node with no edges leaving that node), the "epi... |
The paper proposes an objective for domain generalization. The main findings involve two types of representations: domain invariant - the distribution of the target conditioned on such a representation is the same across training environments; and domain domain-general - the representation is domain invariant across al... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes an objective for domain generalization. The main findings involve two types of representations: domain invariant - the distribution of the target conditioned on such a representation is the same across training environments; and domain domain-general - the representation is domain invariant a... |
The authors provide a quantitative measure to check the correspondence of feature importance as detected by a model with human perception of a feature's importance, which is incorporated in a framework to quantify model trustworthiness.
+ The paper is well motivated and well written
+ Recent work seems to have been di... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors provide a quantitative measure to check the correspondence of feature importance as detected by a model with human perception of a feature's importance, which is incorporated in a framework to quantify model trustworthiness.
+ The paper is well motivated and well written
+ Recent work seems to have... |
This paper proposes AANG to automatically generate a suite of auxiliary objectives for end tasks. AANG builds a unified framework/pipeline that consists of several parts: input data, input transformation, model representation, and output. AANG uses algorithmic stability to search for the best combination in each part. ... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes AANG to automatically generate a suite of auxiliary objectives for end tasks. AANG builds a unified framework/pipeline that consists of several parts: input data, input transformation, model representation, and output. AANG uses algorithmic stability to search for the best combination in eac... |
In this paper, the authors propose a new, data-driven method to solve optimal control problems that is significantly faster than existing methods at execution time and produces similar suboptimality of solution when compared to other approaches. The method is explained through the lens of learning an operator that maps... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this paper, the authors propose a new, data-driven method to solve optimal control problems that is significantly faster than existing methods at execution time and produces similar suboptimality of solution when compared to other approaches. The method is explained through the lens of learning an operator t... |
The authors use the Bhattacharyya distance to define notions like the distance between learning trajectories on classification tasks. They also define a simple notion of transfer learning, and use that to compute learning trajectory distances across different transfer tasks and
Using the Bhattacharyya distance gives a... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors use the Bhattacharyya distance to define notions like the distance between learning trajectories on classification tasks. They also define a simple notion of transfer learning, and use that to compute learning trajectory distances across different transfer tasks and
Using the Bhattacharyya distance... |
This paper proposes a hyper-prompted training mechanism for large-scale text retrieval across tasks of different domains.
Strength:
1. The problem of cross-domain text retrieval that this paper studies is important.
2. The paper is well written with clear motivation, method description, and experiment setup.
3. The e... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a hyper-prompted training mechanism for large-scale text retrieval across tasks of different domains.
Strength:
1. The problem of cross-domain text retrieval that this paper studies is important.
2. The paper is well written with clear motivation, method description, and experiment setup.
... |
This paper studies the effect of adversarial training on the algorithmic recourse by deriving both cost bounds and validity bounds. The correctness of the proposed bounds are verified through numerical simulation.
Strength:
S1) Effects of adversarial training on algorithmic resource are analyzed by establishing bou... | 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 effect of adversarial training on the algorithmic recourse by deriving both cost bounds and validity bounds. The correctness of the proposed bounds are verified through numerical simulation.
Strength:
S1) Effects of adversarial training on algorithmic resource are analyzed by establis... |
This paper provides a provable reinforcement learning algorithm for low-rank MDPs. The key difference from past work on this literature is that the transition model $T(s' \mid s, a)$ has a decomposition $<\mu(s'), \phi(s, a)>$ which can be an infinite-dimensional inner product, however, it is required that $\phi$ and $... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper provides a provable reinforcement learning algorithm for low-rank MDPs. The key difference from past work on this literature is that the transition model $T(s' \mid s, a)$ has a decomposition $<\mu(s'), \phi(s, a)>$ which can be an infinite-dimensional inner product, however, it is required that $\ph... |
The paper presents a method to learn from “play data” in a robot manipulation domain. The “play data” is first popularized by Lynch et al., 2018 and is defined as robot experience data collected through human teleoperation where the human teleoperator controls the robot to achieve a variety of goals in a domain. The di... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper presents a method to learn from “play data” in a robot manipulation domain. The “play data” is first popularized by Lynch et al., 2018 and is defined as robot experience data collected through human teleoperation where the human teleoperator controls the robot to achieve a variety of goals in a domain... |
This paper considers an adversarial setting for contrastive learning: the attacker is allowed to add imperceptible perturbation to all training samples of CL, which will reduce CL's effectiveness as a feature extractor. Experiments show that downstream task's performance will be lowered by the poisoning.
**Strength**
... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper considers an adversarial setting for contrastive learning: the attacker is allowed to add imperceptible perturbation to all training samples of CL, which will reduce CL's effectiveness as a feature extractor. Experiments show that downstream task's performance will be lowered by the poisoning.
**Stre... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.