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 authors focus on the choice of task head in fine tuning, and how the task head controls feature adaptation and downstream model quality. The gradient was decomposed as the product between a direction and an energy, and analysis was done on the effect of the energy vs feature adaptation after FT step. The paper's...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors focus on the choice of task head in fine tuning, and how the task head controls feature adaptation and downstream model quality. The gradient was decomposed as the product between a direction and an energy, and analysis was done on the effect of the energy vs feature adaptation after FT step. The...
This paper is about model distillation for dense prediction tasks, specifically for object detection and semantic segmentation. Prior works on this topic identify per-region distillation as an important ingredient to a successful knowledge transfer. This paper proposes a novel way to identify what regions should be dis...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper is about model distillation for dense prediction tasks, specifically for object detection and semantic segmentation. Prior works on this topic identify per-region distillation as an important ingredient to a successful knowledge transfer. This paper proposes a novel way to identify what regions shoul...
Recently, it has been observed that although neural networks achieve very good accuracy on average over the dataset but they can have very low accuracy on certain subgroups of the distribution. There have been many works on trying to fix this gap. This paper proposes an algorithm to improve the accuracy on different su...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Recently, it has been observed that although neural networks achieve very good accuracy on average over the dataset but they can have very low accuracy on certain subgroups of the distribution. There have been many works on trying to fix this gap. This paper proposes an algorithm to improve the accuracy on diff...
This paper present a DNN model similarity measure using input gradient transferability. The basic hypothesis is that if two neural networks are similar, adversarial attack transferability will be high. Additionally two topics are investigated: (1) Which network component contributes to the model diversity? (2) impact o...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: This paper present a DNN model similarity measure using input gradient transferability. The basic hypothesis is that if two neural networks are similar, adversarial attack transferability will be high. Additionally two topics are investigated: (1) Which network component contributes to the model diversity? (2) ...
The paper uses the information generated by the symbol system to help the neural system model to better reason and test CoT reasoning ability, which creates a Dataset PRONTOQA which aim at testing the reasoning ability of big model, and proposes a new method to measure the correctness of reasoning steps rather than the...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper uses the information generated by the symbol system to help the neural system model to better reason and test CoT reasoning ability, which creates a Dataset PRONTOQA which aim at testing the reasoning ability of big model, and proposes a new method to measure the correctness of reasoning steps rather ...
This paper proposes a multi-modal Masked Auto-Encoder for audio-visual tasks. More specifically, the proposed method combines contrastive learning and masked data modeling for better joint and coordinated audio-visual representation. The model based on self-supervised training achieves a new SOTA on VGGSound and great ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a multi-modal Masked Auto-Encoder for audio-visual tasks. More specifically, the proposed method combines contrastive learning and masked data modeling for better joint and coordinated audio-visual representation. The model based on self-supervised training achieves a new SOTA on VGGSound an...
This paper introduces a theoretical analysis to understand why mask reconstruction pre-training (MRP) works well in downstream tasks. It's found MRP can capture all discriminative semantics of each class in the pre-training dataset and thus utilize them to help downstream tasks. It provides solid proof, straightforward...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper introduces a theoretical analysis to understand why mask reconstruction pre-training (MRP) works well in downstream tasks. It's found MRP can capture all discriminative semantics of each class in the pre-training dataset and thus utilize them to help downstream tasks. It provides solid proof, straigh...
Paper studying social and environmental impact of recent developments in ML (with a focus on DL) on biology and chemistry research. Strengths: Lots of information and very interesting topic Weaknesses: Minimally coherent structure. At some points it almost seems like a stream of thoughts dump. Mixed topics Unorthodox i...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Paper studying social and environmental impact of recent developments in ML (with a focus on DL) on biology and chemistry research. Strengths: Lots of information and very interesting topic Weaknesses: Minimally coherent structure. At some points it almost seems like a stream of thoughts dump. Mixed topics Unor...
The paper collected a small-scale exercise video dataset with fine-grained labeling. Some baselines such as I3D, SI-EN, and GCN-based models (e.g., ST-GCN and MS-G3D) are provided for evaluation in the settings of few-shot learning. The baselines do not represent state-of-the-art models in their corresponding modalitie...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper collected a small-scale exercise video dataset with fine-grained labeling. Some baselines such as I3D, SI-EN, and GCN-based models (e.g., ST-GCN and MS-G3D) are provided for evaluation in the settings of few-shot learning. The baselines do not represent state-of-the-art models in their corresponding m...
Authors present a provably convergent MARL algorithmic template HAML, from which two existing MARL algorithms are derived as demonstration. In addition, further two MARL algorithms are cast as HAML instances to rid of their shortcoming defined as 'traps' in the paper. SMAC and MAMuJoCo evaluation shows better performan...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Authors present a provably convergent MARL algorithmic template HAML, from which two existing MARL algorithms are derived as demonstration. In addition, further two MARL algorithms are cast as HAML instances to rid of their shortcoming defined as 'traps' in the paper. SMAC and MAMuJoCo evaluation shows better p...
This paper focuses on the fair learning problem under the covariate shift. The main contribution is the weighted entropy loss component of the regularized objective function. In the experiments, the proposed method achieves a better tradeoff between accuracy and fairness. Strength Fairness under the covariate shift i...	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 paper focuses on the fair learning problem under the covariate shift. The main contribution is the weighted entropy loss component of the regularized objective function. In the experiments, the proposed method achieves a better tradeoff between accuracy and fairness. Strength Fairness under the covariate...
This work considers the problem of neural network inversion, which lies at the core of many downstream tasks incorporating neural networks, such as inverse problems. Naively inverting a network via gradient descent is challenging, due to the induced non-convex landscape with potentially many local minima. The authors p...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work considers the problem of neural network inversion, which lies at the core of many downstream tasks incorporating neural networks, such as inverse problems. Naively inverting a network via gradient descent is challenging, due to the induced non-convex landscape with potentially many local minima. The a...
The work studies the generalization of fairness constraints in Machine Learning systems. The authors argue that while current algorithms are optimized for generalizing on the whole dataset, their efficacy on minority classes for imbalanced datasets have not been extensively covered They introduce their Flexible Imbal...	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 work studies the generalization of fairness constraints in Machine Learning systems. The authors argue that while current algorithms are optimized for generalizing on the whole dataset, their efficacy on minority classes for imbalanced datasets have not been extensively covered They introduce their Flexib...
The paper addresses the problem of reward extrapolation in inverse reinforcement learning (IRL). The authors propose a modification of the MaxEntIRL to properly address the problem by leveraging an additional dataset collected by a behavioral policy and properly adjusting the loss function, introducing additional terms...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper addresses the problem of reward extrapolation in inverse reinforcement learning (IRL). The authors propose a modification of the MaxEntIRL to properly address the problem by leveraging an additional dataset collected by a behavioral policy and properly adjusting the loss function, introducing addition...
This paper focuses on domain generalization for object detection, especially for the application of autonomous driving. Authors claim that features from images in different domains have different statistics, thus propose to perturb the feature statistics during the training. Experiments are conducted on several autonom...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper focuses on domain generalization for object detection, especially for the application of autonomous driving. Authors claim that features from images in different domains have different statistics, thus propose to perturb the feature statistics during the training. Experiments are conducted on several...
The authors propose two new synthetic likelihood methods for simulation based inference using energy based models, UNLE and AUNLE. They introduces a tilting trick and an amortized sequential model to perform posterior inference with intractable likelihoods and improve both modeling quality and inference budget. Stregt...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors propose two new synthetic likelihood methods for simulation based inference using energy based models, UNLE and AUNLE. They introduces a tilting trick and an amortized sequential model to perform posterior inference with intractable likelihoods and improve both modeling quality and inference budget...
This paper introduces GRAPHSENSOR architecture to capture the internal relationships in time-series data. GRAPHSENSOR is composed of single segment representation module using convolutional encoders and relationship learning module using graph-based self multi-head attention. Applying GRAPHSENSOR to two datasets of Sl...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper introduces GRAPHSENSOR architecture to capture the internal relationships in time-series data. GRAPHSENSOR is composed of single segment representation module using convolutional encoders and relationship learning module using graph-based self multi-head attention. Applying GRAPHSENSOR to two datase...
This work adopts the feature learning framework by Allen-Zhu and Li (2021), and further introduces latent orthogonal feature views. The authors prove that under certain assumptions on features, a two-layer smooth-ReLU network can learn all features with midpoint mixup, whereas ERM can only learn one feature. Strength: ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This work adopts the feature learning framework by Allen-Zhu and Li (2021), and further introduces latent orthogonal feature views. The authors prove that under certain assumptions on features, a two-layer smooth-ReLU network can learn all features with midpoint mixup, whereas ERM can only learn one feature. St...
This paper proposes a method that learns to map input sequences into log-compressed number lines. The proposed method is evaluated on event prediction using synthetic data in 3 different settings and shown to outperform RNNs, LSTMs and several other baselines (despite having fewer parameters). Strengths: This p...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper proposes a method that learns to map input sequences into log-compressed number lines. The proposed method is evaluated on event prediction using synthetic data in 3 different settings and shown to outperform RNNs, LSTMs and several other baselines (despite having fewer parameters). *Strengths:...
This paper proposes to use dynamic prior knowledge (DPK) that uses the teacher’s feature as inputs for the feature distillation by mixing the student features and using an encoder-decoder framework that learns the teacher’s feature as the target. Using DPK has an advantage in that the performance of the student network...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes to use dynamic prior knowledge (DPK) that uses the teacher’s feature as inputs for the feature distillation by mixing the student features and using an encoder-decoder framework that learns the teacher’s feature as the target. Using DPK has an advantage in that the performance of the student...
This paper uses Lipschitz regularization (in the context of the dual for f-divergences) to construct gradient flows for sampling from target distributions. These distributions can be empirical (i.e. discrete), and the construction can be applied to build generative models. Experiments show improved performance for gene...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper uses Lipschitz regularization (in the context of the dual for f-divergences) to construct gradient flows for sampling from target distributions. These distributions can be empirical (i.e. discrete), and the construction can be applied to build generative models. Experiments show improved performance ...
This work focused on reward-free RL with low-rank MDP. The author proposed a new model-based reward-free (RAFFLE) algorithm and provided a theoretical guarantee with polynomial sample complexity. In addition, the author offers a lower sample complexity bound for any algorithm within the low-rank MDP. Finally, the autho...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work focused on reward-free RL with low-rank MDP. The author proposed a new model-based reward-free (RAFFLE) algorithm and provided a theoretical guarantee with polynomial sample complexity. In addition, the author offers a lower sample complexity bound for any algorithm within the low-rank MDP. Finally, t...
This paper addresses the limitation of previous methods for Bregman divergence learning and proposes a solution named Neural Bregman Divergences (NBD) using Input Convex Neural Network (ICNN), which can be computed efficiently and gives finer resolution to the generating function. Experimental results verify the perfor...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper addresses the limitation of previous methods for Bregman divergence learning and proposes a solution named Neural Bregman Divergences (NBD) using Input Convex Neural Network (ICNN), which can be computed efficiently and gives finer resolution to the generating function. Experimental results verify th...
This paper introduces some new analysis technique of the convergence of AdaGrad on the unbounded domain, in the smooth convex optimization setting. The authors prove that several variants of AdaGrad can converge (in both average and the last literature sense) to the optimum without the need of additional assumptions. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper introduces some new analysis technique of the convergence of AdaGrad on the unbounded domain, in the smooth convex optimization setting. The authors prove that several variants of AdaGrad can converge (in both average and the last literature sense) to the optimum without the need of additional assump...
In this work, the authors reformulate the problem of intent detection as a question-answering task by treating utterances and labels as questions and answers. A two stage training schema is employed by utilizing question-answering retrieval architecture and batch contrastive loss. The first of the training stages is fo...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work, the authors reformulate the problem of intent detection as a question-answering task by treating utterances and labels as questions and answers. A two stage training schema is employed by utilizing question-answering retrieval architecture and batch contrastive loss. The first of the training stag...
This paper proposes the single image depth estimation method with variational constraints via first-order difference. This paper proposes V-layer which computes the depth difference map and the weight map, which will be utilized to obtain the initial depth map. The initial depth map will be upsampled and refined throug...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes the single image depth estimation method with variational constraints via first-order difference. This paper proposes V-layer which computes the depth difference map and the weight map, which will be utilized to obtain the initial depth map. The initial depth map will be upsampled and refine...
This paper proposes a model OBPose for unsupervised 3D object segmentation from RGB-D images/videos. OBPose first represents each object as disentangled location and appearance information, then re-renders the scene with a NERF decoder. Experimental results prove the effectiveness of the proposed object representation....	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a model OBPose for unsupervised 3D object segmentation from RGB-D images/videos. OBPose first represents each object as disentangled location and appearance information, then re-renders the scene with a NERF decoder. Experimental results prove the effectiveness of the proposed object represe...
The paper proposes to learn isometric representations in neural networks to preserve the structure in the output end. Then locally isometric layer is proposed to learn the required output representation, which is shown to be robust on MNIST compared with the cross-entropy training baseline. Strength: 1. The Locally Is...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes to learn isometric representations in neural networks to preserve the structure in the output end. Then locally isometric layer is proposed to learn the required output representation, which is shown to be robust on MNIST compared with the cross-entropy training baseline. Strength: 1. The Lo...
The paper proposes a self-supervised method for Category-level 6D Object Pose Estimation. Specifically, the paper proposes an architecture and a number of geometry-based consistency losses allowing simultaneous estimation of the shape of specific instance and its pose. The experimental sections, demonstrates improvemen...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper proposes a self-supervised method for Category-level 6D Object Pose Estimation. Specifically, the paper proposes an architecture and a number of geometry-based consistency losses allowing simultaneous estimation of the shape of specific instance and its pose. The experimental sections, demonstrates im...
This paper reports the performance of SFDA methods under the new setting, called double-transfer, via a large-scale empirical evaluation. The double-transfer represents two-stage transfer learning setting in which the pre-trained model is firstly fine-tuned with a source domain dataset and is then adapted to the target...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper reports the performance of SFDA methods under the new setting, called double-transfer, via a large-scale empirical evaluation. The double-transfer represents two-stage transfer learning setting in which the pre-trained model is firstly fine-tuned with a source domain dataset and is then adapted to th...
Detailed annotations are often not available in practice, and only high-level image information is available on the target dataset; this work presents a fine-tuning method using ProbKT to update object detection models from a pre-trained architecture. Nevertheless, the presented knowledge transfer method can be built u...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Detailed annotations are often not available in practice, and only high-level image information is available on the target dataset; this work presents a fine-tuning method using ProbKT to update object detection models from a pre-trained architecture. Nevertheless, the presented knowledge transfer method can be...
The paper focuses on the task of textual inversion which in essence tries to capture a concept in a set of images (either style, abstract, object, or relations) as a single new "word" which can then we used to guide the generation of the generative models based on this new "word". An example would be to extract out abs...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper focuses on the task of textual inversion which in essence tries to capture a concept in a set of images (either style, abstract, object, or relations) as a single new "word" which can then we used to guide the generation of the generative models based on this new "word". An example would be to extract...
This paper provides an analytic overview of the ``sharp Sinkhorn loss'' and its derivatives with respect both the input weights and the cost matrix (that typically depends on the input locations). Formally, given two discretes probability measures $\mu = \sum_i a_i \delta_{x_i}$ and $\nu = \sum_j b_j \delta_{y_j}$, and...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper provides an analytic overview of the ``sharp Sinkhorn loss'' and its derivatives with respect both the input weights and the cost matrix (that typically depends on the input locations). Formally, given two discretes probability measures $\mu = \sum_i a_i \delta_{x_i}$ and $\nu = \sum_j b_j \delta_{y_...
This paper studies byzantine robust decentralized learning. In this setting, a fraction of workers is byzantine. First, the paper argues why the consensus is vulnerable to byzantine attacks, then they propose DISSENSUS, a decentralized attack to steer away from the true consensus. Finally, they come up with a robust ca...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies byzantine robust decentralized learning. In this setting, a fraction of workers is byzantine. First, the paper argues why the consensus is vulnerable to byzantine attacks, then they propose DISSENSUS, a decentralized attack to steer away from the true consensus. Finally, they come up with a r...
This paper explains that the exact solution can be numerically found in DDIM with fast sampling. Extending this explanation, this paper proposes "generalized DDIM" that modifies parameterization of the score networks of existing diffusion models to achieve 20x acceleration and comparable or much better FID even on non-...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper explains that the exact solution can be numerically found in DDIM with fast sampling. Extending this explanation, this paper proposes "generalized DDIM" that modifies parameterization of the score networks of existing diffusion models to achieve 20x acceleration and comparable or much better FID even...
The paper builds on Split learning [Thapa et al. 2022], where the first few layers of a neural network model are stored and shared across clients and the server, while the remaining layers are only stored and trained on the server. Since the communication cost is high to "make the connection" between the split layers, ...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper builds on Split learning [Thapa et al. 2022], where the first few layers of a neural network model are stored and shared across clients and the server, while the remaining layers are only stored and trained on the server. Since the communication cost is high to "make the connection" between the split ...
The paper studies the negative transfer phenomenon of MAE pre-trained models. Authors first experimentally show the existence of negative transfer and then show that natively using MoE can not solve the problem. Then, a clustering-based method is proposed to solve the problem. Generally, the paper is well-organized and...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper studies the negative transfer phenomenon of MAE pre-trained models. Authors first experimentally show the existence of negative transfer and then show that natively using MoE can not solve the problem. Then, a clustering-based method is proposed to solve the problem. Generally, the paper is well-organ...
This work utilizes (1) a Long Short-Term Memory (LSTM) autoencoder to learn unsupervised latent representations of multivariate time series data, and (2) a supervised Bidirectional LSTM (Bi-LSTM) that is initialized with the representations generated in step (1) and tasked to predict a continuous-valued output. The res...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This work utilizes (1) a Long Short-Term Memory (LSTM) autoencoder to learn unsupervised latent representations of multivariate time series data, and (2) a supervised Bidirectional LSTM (Bi-LSTM) that is initialized with the representations generated in step (1) and tasked to predict a continuous-valued output....
This paper considers vertical federated learning settings. It addresses the homogeneity and synchronicity problems by proposing a federated feature fusion (F^3) framework and learning a consensus graph among local owners. Experiments on four real-world time series datasets demonstrate the effectiveness of F^3. Strength...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper considers vertical federated learning settings. It addresses the homogeneity and synchronicity problems by proposing a federated feature fusion (F^3) framework and learning a consensus graph among local owners. Experiments on four real-world time series datasets demonstrate the effectiveness of F^3. ...
This paper provides a practical algorithm for differentially private linear regression. Algorithmically, this proposed method is a combination of Theil-Sen estimator for linear regression and private mean estimator using Tukey median appeared in (Brown at el 2021) and (Liu at el 2021). Prior works have provable guarant...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides a practical algorithm for differentially private linear regression. Algorithmically, this proposed method is a combination of Theil-Sen estimator for linear regression and private mean estimator using Tukey median appeared in (Brown at el 2021) and (Liu at el 2021). Prior works have provable...