Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The paper (FedACG) proposes the following additive regularization term to the learning objective function for federated learning clients: $$ \quad\quad \frac{\beta}{2}\|\|\theta - (\theta^{t-1} + \lambda m^{t-1}) \|\|^2 $$ where $\beta$ and $\lambda$ are tuning parameters. $\theta$ is the model parameters. The $m$ momen...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper (FedACG) proposes the following additive regularization term to the learning objective function for federated learning clients: $$ \quad\quad \frac{\beta}{2}\|\|\theta - (\theta^{t-1} + \lambda m^{t-1}) \|\|^2 $$ where $\beta$ and $\lambda$ are tuning parameters. $\theta$ is the model parameters. The $...
This paper analyzes the maximum-margin bias of gradient flow on quasi-homogeneous networks under data separation. In contrast to homogeneous networks, different parameters in a quasi-homogeneous network are allowed to have different homogeneity factors; this model can represent bias terms, skip connections, normalizati...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper analyzes the maximum-margin bias of gradient flow on quasi-homogeneous networks under data separation. In contrast to homogeneous networks, different parameters in a quasi-homogeneous network are allowed to have different homogeneity factors; this model can represent bias terms, skip connections, nor...
This paper provides a sample complexity bound for Transformers, as well as the required number of SGD steps. These bounds are for the model to achieve 0 generalization loss (using the hinge loss). The task is binary classification, where the label depends only on the "label-relevant patterns" in the input. The bound ...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper provides a sample complexity bound for Transformers, as well as the required number of SGD steps. These bounds are for the model to achieve 0 generalization loss (using the hinge loss). The task is binary classification, where the label depends only on the "label-relevant patterns" in the input. Th...
In this work the authors take advantage of the analogy between ResNets and (neural) ODEs, and try to use analysis tools from ODEs to analyze blocks in a ResNet. In particular, they focus on the notion of "stiffness" - a set of properties which make ODEs difficult to integrate, and prone to numerical instability. They c...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this work the authors take advantage of the analogy between ResNets and (neural) ODEs, and try to use analysis tools from ODEs to analyze blocks in a ResNet. In particular, they focus on the notion of "stiffness" - a set of properties which make ODEs difficult to integrate, and prone to numerical instability...
Prior work in continual learning has focused on addressing the catastrophic forgetting (CF) problem. However, this paper notes that "knowledge transfer" (KT) between tasks is also important, and proposes a new method which "addresses both issues." To address CF, each task is associated with a learned mask identifying...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Prior work in continual learning has focused on addressing the catastrophic forgetting (CF) problem. However, this paper notes that "knowledge transfer" (KT) between tasks is also important, and proposes a new method which "addresses both issues." To address CF, each task is associated with a learned mask ide...
For unsupervised reinforcement learning, this paper proposes (1) a set of metrics to evaluate the exploration and skill learning of the agent without specific downstream tasks; (2) an automatic curriculum to train the agent with different intrinsic rewards in different stages, formulated by multi-armed bandit. Strength...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: For unsupervised reinforcement learning, this paper proposes (1) a set of metrics to evaluate the exploration and skill learning of the agent without specific downstream tasks; (2) an automatic curriculum to train the agent with different intrinsic rewards in different stages, formulated by multi-armed bandit. ...
This paper focuses on accelerating guided diffusion sampling, where the backward ODE to be solved consists of two parts: the first one from the diffusion contribution and the second one from the classifier contribution. It is empirically found that non-splitting high-order ODE solvers does not work well when the number...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper focuses on accelerating guided diffusion sampling, where the backward ODE to be solved consists of two parts: the first one from the diffusion contribution and the second one from the classifier contribution. It is empirically found that non-splitting high-order ODE solvers does not work well when th...
- The paper addresses the problem of handling domain-shifts that arises in generative learnt channel models in E2E communication systems in a few-shot setting. - The proposed domain adaptation approach is tailored around a Mixture Density Network (MDN) representing the channel model. In here, the approach: - learns...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: - The paper addresses the problem of handling domain-shifts that arises in generative learnt channel models in E2E communication systems in a few-shot setting. - The proposed domain adaptation approach is tailored around a Mixture Density Network (MDN) representing the channel model. In here, the approach: ...
The paper proposes a novel algorithm to learn from imperfect teacher's policy and achieves super-teacher performance. It provides theoretical justification, performance bounds and experimental results with comparison against strong baselines. Strengths: - The paper addresses a key problem of learning from an imperfect ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a novel algorithm to learn from imperfect teacher's policy and achieves super-teacher performance. It provides theoretical justification, performance bounds and experimental results with comparison against strong baselines. Strengths: - The paper addresses a key problem of learning from an im...
This paper introduces RAIN, a method to learn relational graphs between different nodes ("agents") of a system from time-series data describing the trajectories followed by said nodes. RAIN is composed of a) an encoder computing an embedding of the trajectories seen so far, b) a graph extractor that infers the intera...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces RAIN, a method to learn relational graphs between different nodes ("agents") of a system from time-series data describing the trajectories followed by said nodes. RAIN is composed of a) an encoder computing an embedding of the trajectories seen so far, b) a graph extractor that infers th...
Authors explore the problem of robustifying a video transformer using a pre-trained image transformer as a surrogate model. Authors argue that using adversarial patterns optimal for large image models (such as Deit, DINO and CLIP) improves adversarial robustness of other image models, but is not suitable out-of-the-box...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Authors explore the problem of robustifying a video transformer using a pre-trained image transformer as a surrogate model. Authors argue that using adversarial patterns optimal for large image models (such as Deit, DINO and CLIP) improves adversarial robustness of other image models, but is not suitable out-of...
This paper introduces a numerically stable and parallelizable version of the entropy semi-ring for ASR training. For numerical stability, the paper introduces log entropy semiring with $<\log p(e), \log(-p(e)\log p(e))> $. For the knowledge distillation, which is based on the hard and soft labels, the paper proposes on...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a numerically stable and parallelizable version of the entropy semi-ring for ASR training. For numerical stability, the paper introduces log entropy semiring with $<\log p(e), \log(-p(e)\log p(e))> $. For the knowledge distillation, which is based on the hard and soft labels, the paper pro...
This paper proposes to use cycle consistency to help learn domain-invariant features for unsupervised domain generalization. Pro: 1. Previous method DiVAE only considers the reconstruction loss \|\|y_m-x_m\|\|. The proposed loss considers how to utilize the information of y_n and use cycle consistency to further help the...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes to use cycle consistency to help learn domain-invariant features for unsupervised domain generalization. Pro: 1. Previous method DiVAE only considers the reconstruction loss \|\|y_m-x_m\|\|. The proposed loss considers how to utilize the information of y_n and use cycle consistency to further ...
The authors, presents PINTO a two stage pipeline to improve rationale-based language reasoning. The first step is to prompt a language model (20B in this case) to generate rationale given a question, while the second step is to fine tune a smaller (<1B) LM with the pair (Question, generated rationale) -> Answer. The a...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors, presents PINTO a two stage pipeline to improve rationale-based language reasoning. The first step is to prompt a language model (20B in this case) to generate rationale given a question, while the second step is to fine tune a smaller (<1B) LM with the pair (Question, generated rationale) -> Answer...
The paper introduces a DRL based approach to perform Loop Vectorization in LLVM IR. They represent LLVM IR as a graph and then learn an embedding in an unsupervised manner using a GNN, which they later refine using a FCNN to decide the VF and IF in the loop vectorizer. Strengths: * The paper improves on structural repr...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper introduces a DRL based approach to perform Loop Vectorization in LLVM IR. They represent LLVM IR as a graph and then learn an embedding in an unsupervised manner using a GNN, which they later refine using a FCNN to decide the VF and IF in the loop vectorizer. Strengths: * The paper improves on structu...
This paper proposes a new graph Transformer architecture that contains a sequence of tokens obtained by aggregating neighbors' features. It shows improved expressive power with information from multi-hop neighborhoods. So, the model treats a node as a sequence of tokens. The authors claim that the model has better scal...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new graph Transformer architecture that contains a sequence of tokens obtained by aggregating neighbors' features. It shows improved expressive power with information from multi-hop neighborhoods. So, the model treats a node as a sequence of tokens. The authors claim that the model has bet...
The paper proposes to apply an ensemble of siamese CNNs to detect and describe keypoints. The parameters of these CNNS are not trained but set randomly. The keypoint descriptors produced by each siamese CNN are matched using a classical nearest neighbor with ratio test followed by a mutual nearest neighbor test. All th...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes to apply an ensemble of siamese CNNs to detect and describe keypoints. The parameters of these CNNS are not trained but set randomly. The keypoint descriptors produced by each siamese CNN are matched using a classical nearest neighbor with ratio test followed by a mutual nearest neighbor test...
This manuscript provides an information theoretical perspective for multiview self-supervised learning (SSL). In specific, the authors show that a lower bound of mutual information is a useful objective function to learn informative representations. Additionally, optimizing such an objective function could be regarded ...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This manuscript provides an information theoretical perspective for multiview self-supervised learning (SSL). In specific, the authors show that a lower bound of mutual information is a useful objective function to learn informative representations. Additionally, optimizing such an objective function could be r...
The authors address the problem of outlier-robust group inference via gradient space clustering. They extract features first and then calculate the gradient of the extracted features. Finally, they cluster in gradient space. Strength: The paper is well written at the beginning and validate features in gradient space is...	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 authors address the problem of outlier-robust group inference via gradient space clustering. They extract features first and then calculate the gradient of the extracted features. Finally, they cluster in gradient space. Strength: The paper is well written at the beginning and validate features in gradient ...
The paper discusses the issue of adversarial training over-fitting, pointing out that a pre-specified perturbation budget is not optimal as the training progresses the perturbation budget should be adjusted accordingly. Based on this intuition, the author proposes a new adversarial training method that generates advers...	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 discusses the issue of adversarial training over-fitting, pointing out that a pre-specified perturbation budget is not optimal as the training progresses the perturbation budget should be adjusted accordingly. Based on this intuition, the author proposes a new adversarial training method that generate...
The paper presents Q-Pensieve, a method using versions of past Q functions to update a new Q function. This method enables information sharing across policies. The authors instantiate the idea in a soft actor-critic algorithm. Experiment results on DST, LunarLander, and several MuJoCo environments are provided where Q-...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper presents Q-Pensieve, a method using versions of past Q functions to update a new Q function. This method enables information sharing across policies. The authors instantiate the idea in a soft actor-critic algorithm. Experiment results on DST, LunarLander, and several MuJoCo environments are provided ...
This paper studies learning composable task codes for soft-prompting of language models. The authors propose an approach similar to VQ-VAE where each task is associated with a set of discrete codes and embedding of these codes from a codebook lookup table is stacked as a soft-prompt for a language model. First, a 2D em...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies learning composable task codes for soft-prompting of language models. The authors propose an approach similar to VQ-VAE where each task is associated with a set of discrete codes and embedding of these codes from a codebook lookup table is stacked as a soft-prompt for a language model. First,...
This paper considers the problem of unlabelled sparse recovery under multiple measurements. More formally, Given $(Y, X)$ such that $Y = \Pi^* X B^+ W$, where $W \in \mathbb R^{n \times m}$, $W_{i, j} \sim N(0, \sigma^2)$, $X \in \mathbb R^{n \times p}$, $X_{i, j} \sim N(0, 1)$ and each column of $B^ \in \mathbb R^{p...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper considers the problem of unlabelled sparse recovery under multiple measurements. More formally, Given $(Y, X)$ such that $Y = \Pi^* X B^+ W$, where $W \in \mathbb R^{n \times m}$, $W_{i, j} \sim N(0, \sigma^2)$, $X \in \mathbb R^{n \times p}$, $X_{i, j} \sim N(0, 1)$ and each column of $B^ \in \mat...
This paper designs an entity-aware and image-information-integrated article generation method. The entities in the article are recognized with the help of image information and labeled with entity types. The paper claims that by annotating these entities on a target sequence, it enables the model to perceive entity inf...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper designs an entity-aware and image-information-integrated article generation method. The entities in the article are recognized with the help of image information and labeled with entity types. The paper claims that by annotating these entities on a target sequence, it enables the model to perceive en...
The paper aims to discover generalizable multi-agent coordination skills from multi-task offline data (multi-task marl with offline data). The insight is that we can extract universal skills for coordination from offline multi-agent multi-task data. The authors propose the ODIS algorithm which consists of two stages. F...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper aims to discover generalizable multi-agent coordination skills from multi-task offline data (multi-task marl with offline data). The insight is that we can extract universal skills for coordination from offline multi-agent multi-task data. The authors propose the ODIS algorithm which consists of two s...
This paper proposes a model and pre-training method for estimating conditional expected outcomes as part of causal estimation and evaluates this approach using real and semi-synthetic data. I am of two minds on this paper. On the one hand, I do not think the proposed method is particularly novel and I think the paper...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a model and pre-training method for estimating conditional expected outcomes as part of causal estimation and evaluates this approach using real and semi-synthetic data. I am of two minds on this paper. On the one hand, I do not think the proposed method is particularly novel and I think t...
The paper studies RL on language models, and makes 3 contributions: 1) RL4LMs, a modular library for optimizing language generators with RL. 2) GRUE, a benchmark of 6 language generation tasks with reward functions. 3) NLPO, a new algorithm which improves on PPO for RL on LMs. Strengths: + The lack of open-source benc...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper studies RL on language models, and makes 3 contributions: 1) RL4LMs, a modular library for optimizing language generators with RL. 2) GRUE, a benchmark of 6 language generation tasks with reward functions. 3) NLPO, a new algorithm which improves on PPO for RL on LMs. Strengths: + The lack of open-sou...
This paper extended Q-Learning to continuous action space by the action discretization with value decomposition in MARL. The experiment results show that the proposed method is competitive with the state-of-art continuous actor-critic methods. Strength: The method of the paper is very clear. The method of appl...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper extended Q-Learning to continuous action space by the action discretization with value decomposition in MARL. The experiment results show that the proposed method is competitive with the state-of-art continuous actor-critic methods. Strength: The method of the paper is very clear. The method...
This paper is inspired by cognitive science that young children actively acquire language through interactions with their surrounding environment and caretakers. Specifically one critical mechanism of language learning is the ability to infer the mental states of other agents in social environments, referred to as Theo...	Recommendation: 5: marginally below the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper is inspired by cognitive science that young children actively acquire language through interactions with their surrounding environment and caretakers. Specifically one critical mechanism of language learning is the ability to infer the mental states of other agents in social environments, referred to...
This work addresses the problem of errors in structure learning algorithms. The authors propose to reweight poorly fit samples in order to improve the efficacy of the underlying algorithm. A proof is shown under linear models for specific scoring rules in the asymptotic regime. Empirical results show the proposed meth...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This work addresses the problem of errors in structure learning algorithms. The authors propose to reweight poorly fit samples in order to improve the efficacy of the underlying algorithm. A proof is shown under linear models for specific scoring rules in the asymptotic regime. Empirical results show the propo...
Suppose a set of historical solutions of a DNN model are obtained by applying SGD over a sequence of epochs, the paper considers finding the optimal combination of the historical solutions by minimizing the training loss subject to an equality constraint. The proposed optimisation process is referred to as trainable we...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: Suppose a set of historical solutions of a DNN model are obtained by applying SGD over a sequence of epochs, the paper considers finding the optimal combination of the historical solutions by minimizing the training loss subject to an equality constraint. The proposed optimisation process is referred to as trai...
This paper presents a novel framework, called Flash-to-Bang Depth (FBDepth), for passive sound-source depth estimation . The authors use audio-visual correspondence and optical flow manipulation to get decimeter-level depth accuracy. The proposed audio-visual depth estimation system uses video, audio and optical flow t...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a novel framework, called Flash-to-Bang Depth (FBDepth), for passive sound-source depth estimation . The authors use audio-visual correspondence and optical flow manipulation to get decimeter-level depth accuracy. The proposed audio-visual depth estimation system uses video, audio and optica...
This paper conducts a systematic study of scaling behavior of different model architectures. Strength: Extensive experiments have been done in this paper. Weaknesses: (1). There is a lack of innovation. All of the conclusions drawn from experiments are pretty under expectation and there are no surprises. Major...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper conducts a systematic study of scaling behavior of different model architectures. Strength: Extensive experiments have been done in this paper. Weaknesses: (1). There is a lack of innovation. All of the conclusions drawn from experiments are pretty under expectation and there are no surprise...
This paper studies data selection to construct a subset of full data, which is an important research topic in data-efficient deep learning. The authors piont out that prior works on data selection are always specially designed for certain cases, which makes it hard to apply them in practice, since realistic scenes are ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper studies data selection to construct a subset of full data, which is an important research topic in data-efficient deep learning. The authors piont out that prior works on data selection are always specially designed for certain cases, which makes it hard to apply them in practice, since realistic sce...
This paper analyzes the stability issues in Q learning and SARSA with epsilon greedy exploration. The authors use Differential Inclusion (DI) theory to analyze the behavior of Q learning and SARSA. The authors use numerical examples to illustrate their theory. Strength: The stability issue of Q learning and SARSA is v...	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper analyzes the stability issues in Q learning and SARSA with epsilon greedy exploration. The authors use Differential Inclusion (DI) theory to analyze the behavior of Q learning and SARSA. The authors use numerical examples to illustrate their theory. Strength: The stability issue of Q learning and SA...
This paper proposes to use a score-based model for the task of removing structured noise. Specifically, the method uses a pretrained (unconditional) NCSNv2 model as the signal prior and uses modified update rule during sampling that incorporates the score of a separate noise model. Experimental results show that score...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper proposes to use a score-based model for the task of removing structured noise. Specifically, the method uses a pretrained (unconditional) NCSNv2 model as the signal prior and uses modified update rule during sampling that incorporates the score of a separate noise model. Experimental results show th...
This paper proposes InLay, a new layer module which claims to capture internal relationships between input objects. Ultimately, for each input sequence $X \in \mathbb R^{k \times n}$ (interpreted as $k$ elements with $n$ features), the output of an InLay layer is given by $\tanh\left(\frac{XQ(XK)^\mathrm T}{\sqrt{4n}}\...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes InLay, a new layer module which claims to capture internal relationships between input objects. Ultimately, for each input sequence $X \in \mathbb R^{k \times n}$ (interpreted as $k$ elements with $n$ features), the output of an InLay layer is given by $\tanh\left(\frac{XQ(XK)^\mathrm T}{\sq...
This work proposed a fast sampling method, termed as diffusion exponential integrator (DEIS), for diffusion models with a new discretization of the reverse process. In particular, this work first investigates the existing ODE solvers for diffusion models and found that reducing the discretization error is crucial for f...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This work proposed a fast sampling method, termed as diffusion exponential integrator (DEIS), for diffusion models with a new discretization of the reverse process. In particular, this work first investigates the existing ODE solvers for diffusion models and found that reducing the discretization error is cruci...
This paper proposes a pretrained image-to-text model for visual language understanding, which has a wide range of application in sources such as textbooks with diagrams, webpages with images and tables, and mobile apps with buttons and forms. The model has a simple and general architecture which only takes an image as ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a pretrained image-to-text model for visual language understanding, which has a wide range of application in sources such as textbooks with diagrams, webpages with images and tables, and mobile apps with buttons and forms. The model has a simple and general architecture which only takes an i...
The authors propose JSRL to explore and learn by continuously using existing polices to reach a state before adopting a learnable policy. The authors emphasize that when using existing policy to select the arrival state, one can consider gradually decreasing the time step of their decisions, thus providing a more stabl...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose JSRL to explore and learn by continuously using existing polices to reach a state before adopting a learnable policy. The authors emphasize that when using existing policy to select the arrival state, one can consider gradually decreasing the time step of their decisions, thus providing a mo...