Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper proposes a self-supervised learning approach in the context of low-compute deep-learning models (e.g., DeiT-Tiny). The authors observe that a weaker self-supervised target is beneficial for small networks, and propose to match multiple views in more comparable spatial scales and contexts. Experimental result...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a self-supervised learning approach in the context of low-compute deep-learning models (e.g., DeiT-Tiny). The authors observe that a weaker self-supervised target is beneficial for small networks, and propose to match multiple views in more comparable spatial scales and contexts. Experimenta...
This paper examines whether the masking method faithfully removes the information encoded in the input variables. Then the authors propose a method to remove the effect encoded in the input variables by learning optimal baseline values for Shapley values. Experimental results demonstrate the effectiveness of the method...	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 examines whether the masking method faithfully removes the information encoded in the input variables. Then the authors propose a method to remove the effect encoded in the input variables by learning optimal baseline values for Shapley values. Experimental results demonstrate the effectiveness of th...
The paper investigates three separated techniques to improve the successful rate and perturbation efficiency of backdoor attacks. Based on the proposed attacking strategy, the authors constructed two backdoored datasets based on CIFAR-10 and CIFAR-100, in which 0.04% and 0.06% of the data are poisoned. Both of them hav...	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 investigates three separated techniques to improve the successful rate and perturbation efficiency of backdoor attacks. Based on the proposed attacking strategy, the authors constructed two backdoored datasets based on CIFAR-10 and CIFAR-100, in which 0.04% and 0.06% of the data are poisoned. Both of ...
This submission proposed an extension of NeRF called DM-NeRF, which can segment individual objects and manipulate them (translation, rotation, deformation, etc.). To supervise object segmentation, the authors use their algorithm to generate multiple 2D object code predictions and match them with ground truth object lab...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This submission proposed an extension of NeRF called DM-NeRF, which can segment individual objects and manipulate them (translation, rotation, deformation, etc.). To supervise object segmentation, the authors use their algorithm to generate multiple 2D object code predictions and match them with ground truth ob...
This paper proposes a general framework to solve the high dimensional continuum armed bandit problem and high dimensional contextual bandit problem jointly. This research problem is significant and the proposed model is novel. I agree that the proposed problem has broad scope of applications. For this problem, the pape...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a general framework to solve the high dimensional continuum armed bandit problem and high dimensional contextual bandit problem jointly. This research problem is significant and the proposed model is novel. I agree that the proposed problem has broad scope of applications. For this problem, ...
1. This paper proposed the QuasiConvex shallow neural networks (QCNN) whose optimization is equivalent to a convex feasibility problem, which can be solved efficiently with theoretical guarantees. As the building block, they first proved that the multiplication of ReLU output is quasiconvex. These building blocks can t...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: 1. This paper proposed the QuasiConvex shallow neural networks (QCNN) whose optimization is equivalent to a convex feasibility problem, which can be solved efficiently with theoretical guarantees. As the building block, they first proved that the multiplication of ReLU output is quasiconvex. These building bloc...
This paper presents a method for estimating the gauged stream flow of a watershed from the water flux daily observations at each location of a river basin. A convolutional neural network is used to perform this regression, and it is tested on real data (4 USA basin rivers). Strenghs: - the application is of importance...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents a method for estimating the gauged stream flow of a watershed from the water flux daily observations at each location of a river basin. A convolutional neural network is used to perform this regression, and it is tested on real data (4 USA basin rivers). Strenghs: - the application is of im...
This paper proposed a module inserted into a pre-trained model for parameter efficient fine-tuning. The authors empirically combine adapter, adapt pre-fix tuning, and verify the effectiveness of the design through a series of experiments. The proposed method achieves very competitive results and has the potential to ou...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a module inserted into a pre-trained model for parameter efficient fine-tuning. The authors empirically combine adapter, adapt pre-fix tuning, and verify the effectiveness of the design through a series of experiments. The proposed method achieves very competitive results and has the potenti...
The authors consider the problem of learning DAG from heterogeneous datasets that share the same DAG. This problem statement is quite restricted. From Eq(4) we can see that the authors not only encourage the DAG to have the same structure, but also to have the same parameter. While in Figure 1, the authors assume that ...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The authors consider the problem of learning DAG from heterogeneous datasets that share the same DAG. This problem statement is quite restricted. From Eq(4) we can see that the authors not only encourage the DAG to have the same structure, but also to have the same parameter. While in Figure 1, the authors assu...
This paper proposes a workflow for using neural networks to solve optimisation problems while guaranteeing the feasibility of the solution. Strength: 1. The paper tackles an important problem of guaranteeing correctness of DNN-based solution (i.e., does not violate constraints) for the optimization problem, which is ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a workflow for using neural networks to solve optimisation problems while guaranteeing the feasibility of the solution. Strength: 1. The paper tackles an important problem of guaranteeing correctness of DNN-based solution (i.e., does not violate constraints) for the optimization problem, w...
The paper presents two different ways of training a decision tree with fairness criteria and then providing a zero-knowledge proof certifying the fairness without revealing any further information. I think this is in a great application of zero-knowledge proofs. Interesting and well-executed application of zero-knowl...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper presents two different ways of training a decision tree with fairness criteria and then providing a zero-knowledge proof certifying the fairness without revealing any further information. I think this is in a great application of zero-knowledge proofs. Interesting and well-executed application of ze...
This paper studies reward-free learning in linear MDPs. It proposes a computationally efficient algorithm LSVI-RFE and improves the upper bound to O(H^4d^2/eps^2). An Omega(H^3d^2/eps^2) lower bound is also provided. Strengths: 1. The LSVI-RFE algorithm is computationally efficient and improves the current state of th...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies reward-free learning in linear MDPs. It proposes a computationally efficient algorithm LSVI-RFE and improves the upper bound to O(H^4d^2/eps^2). An Omega(H^3d^2/eps^2) lower bound is also provided. Strengths: 1. The LSVI-RFE algorithm is computationally efficient and improves the current sta...
This study gives the solution structure of the label smoothing (LS) loss with positivity constraint, which is an orthogonal variant of the original neural collapse configurations. The paper further explores why LS loss has a better generalization than CE loss when label corruption emerges based on the memorization-dila...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This study gives the solution structure of the label smoothing (LS) loss with positivity constraint, which is an orthogonal variant of the original neural collapse configurations. The paper further explores why LS loss has a better generalization than CE loss when label corruption emerges based on the memorizat...
This paper introduces a new algorithm, Prometheus, for the composite stochastic bilevel optimization problem in the decentralized setting. In particular, the paper considers the setting where we have a distributed optimization problem on $m$ agents, each of which has a loss function of the form $l(x) + h(x)$ where $l(x...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper introduces a new algorithm, Prometheus, for the composite stochastic bilevel optimization problem in the decentralized setting. In particular, the paper considers the setting where we have a distributed optimization problem on $m$ agents, each of which has a loss function of the form $l(x) + h(x)$ wh...
The paper proposes a new dimensionality reduction method, named RTD-AE, that attempts to preserve the topological structure of the data when they are compressed to a lower-dimensional space. This is done in a standard autoencoder fashion where the loss function consists of a reconstruction loss and a topology-preservin...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes a new dimensionality reduction method, named RTD-AE, that attempts to preserve the topological structure of the data when they are compressed to a lower-dimensional space. This is done in a standard autoencoder fashion where the loss function consists of a reconstruction loss and a topology-p...
The paper presents a new fairness metric that can measure fairness for multiple sensitive attributes of any type. The author also developed a learning algorithm by using their fairness criteria as a regularization term. The paper aims to solve an important problem which is ensuring fairness for any number and any type ...	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 presents a new fairness metric that can measure fairness for multiple sensitive attributes of any type. The author also developed a learning algorithm by using their fairness criteria as a regularization term. The paper aims to solve an important problem which is ensuring fairness for any number and a...
The paper considers sim-to-real transfer and approaches this problem from a theoretical perspective. It uses LQG to model the system dynamics and establishes the upper and lower bounds on the sim-to-real gap. A new algorithm is also developed for transfer learning. Strengths: 1. Sim-to-real transfer is an important p...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper considers sim-to-real transfer and approaches this problem from a theoretical perspective. It uses LQG to model the system dynamics and establishes the upper and lower bounds on the sim-to-real gap. A new algorithm is also developed for transfer learning. Strengths: 1. Sim-to-real transfer is an imp...
The paper studies how to visualize high sensitivity directions in reinforcement learning. Given a set of states, authors proposed a RADEN algorithm to find the leading eigenvector of certain covariance matrix relevant to the perturbation analysis. The obtained leading eigenvector indicates the directions of vulnerabili...	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 studies how to visualize high sensitivity directions in reinforcement learning. Given a set of states, authors proposed a RADEN algorithm to find the leading eigenvector of certain covariance matrix relevant to the perturbation analysis. The obtained leading eigenvector indicates the directions of vul...
In this paper, the authors propose a transformer-based method for the temporal action proposal generation task. Compared with existing methods, the main difference lies in the choice of visual features, which are obtained from the Laban Movement Analysis (LMA) method that are originally designed for action analysis. Th...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors propose a transformer-based method for the temporal action proposal generation task. Compared with existing methods, the main difference lies in the choice of visual features, which are obtained from the Laban Movement Analysis (LMA) method that are originally designed for action anal...
The paper proposes MPCFormer, a novel two-stage approach for private transformer inference with MPC. The first stage replaces bottleneck functions in the pre-trained transformer (ie, functions with high computational/communication complexity under MPC), with MPC-friendly approximations. The second stage uses knowledge ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The paper proposes MPCFormer, a novel two-stage approach for private transformer inference with MPC. The first stage replaces bottleneck functions in the pre-trained transformer (ie, functions with high computational/communication complexity under MPC), with MPC-friendly approximations. The second stage uses kn...
This paper attempted to explain the superiority of row-normalized adjacency matrices as the graph convolution operator for node classification in GNNs from the viewpoint of graph neural tangent kernels (GNTK). This paper derived GNTKs for GNNs with finite and infinite layers (and with an infinite number of units) for v...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper attempted to explain the superiority of row-normalized adjacency matrices as the graph convolution operator for node classification in GNNs from the viewpoint of graph neural tangent kernels (GNTK). This paper derived GNTKs for GNNs with finite and infinite layers (and with an infinite number of unit...
This paper extends the 3D GAN to more diverse and un-aligned datasets (eg., ImageNet). This problem setting change is significant and very challenging, as we have learned from 2D fields (StyleGAN3--> StyleGAN-XL). This paper introduces several new modules to adapt an existing model EpiGRAF to this new setting including...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper extends the 3D GAN to more diverse and un-aligned datasets (eg., ImageNet). This problem setting change is significant and very challenging, as we have learned from 2D fields (StyleGAN3--> StyleGAN-XL). This paper introduces several new modules to adapt an existing model EpiGRAF to this new setting i...
Based on heavy-tailed Levy dynamics can produce both large jumps and small roaming to explore the sampling space, resulting in better sampling results than Langevin dynamics with a lacking of large jumps, the authors explore a new class of score-based generative models (SGMs) with sampling based on the Levy dynamics. H...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: Based on heavy-tailed Levy dynamics can produce both large jumps and small roaming to explore the sampling space, resulting in better sampling results than Langevin dynamics with a lacking of large jumps, the authors explore a new class of score-based generative models (SGMs) with sampling based on the Levy dyn...
This paper proposes to discovery causal graphs in term of conditional summary graphs with conditional stationary time series data. Authors first present the framework for learning conditional time series data and then presented a VAE framework to instantiate the framework. Empirical study shows the superior performance...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper proposes to discovery causal graphs in term of conditional summary graphs with conditional stationary time series data. Authors first present the framework for learning conditional time series data and then presented a VAE framework to instantiate the framework. Empirical study shows the superior per...
This paper proposes a test-time adaptation method based visual prompt tuning. After training the deep model along with a set of prompts to convergence, the proposed DePT fixes the backbone parameters and only fine tunes the prompts on the target domain. Based on this learning scheme, DePT integrates multiple self-train...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a test-time adaptation method based visual prompt tuning. After training the deep model along with a set of prompts to convergence, the proposed DePT fixes the backbone parameters and only fine tunes the prompts on the target domain. Based on this learning scheme, DePT integrates multiple se...
Learning decision trees from data is a simple and rich problem. The classic decision tree learning algorithms use a greedy strategy to select the best feature to evenly split the dataset. In this work, the authors look at k features and select an attribute that achieves maximal accuracy. This recovers the classic strat...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: Learning decision trees from data is a simple and rich problem. The classic decision tree learning algorithms use a greedy strategy to select the best feature to evenly split the dataset. In this work, the authors look at k features and select an attribute that achieves maximal accuracy. This recovers the class...
This paper firstly analyzes the properties of mean teacher in a masked data modeling task with a simplified linear model. The mean teacher contributes to the training by gradient correction, and it is similar to the memory queue. Several empirical findings about the mean teacher are introduced based on the experimental...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper firstly analyzes the properties of mean teacher in a masked data modeling task with a simplified linear model. The mean teacher contributes to the training by gradient correction, and it is similar to the memory queue. Several empirical findings about the mean teacher are introduced based on the expe...
The paper tackles the global class imbalance problem in Federated Learning and proposes a novel class distribution estimation method. The estimation method is motivated by the observation that the averaged output probability vector from a classifier at the beginning of training is numerically close to the class ratio. ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper tackles the global class imbalance problem in Federated Learning and proposes a novel class distribution estimation method. The estimation method is motivated by the observation that the averaged output probability vector from a classifier at the beginning of training is numerically close to the class...
The paper presents autograph, a GNN-based approach to automatic loop vectorization. Autograph represents the program as a graph, computes an embedding for each node using a standard GNN forward pass, then trains a RL agent to predict vectorization factors based on node embeddings. The paper evaluates on a set of standa...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents autograph, a GNN-based approach to automatic loop vectorization. Autograph represents the program as a graph, computes an embedding for each node using a standard GNN forward pass, then trains a RL agent to predict vectorization factors based on node embeddings. The paper evaluates on a set o...
The paper tackles the problem of detection of adversarial examples targeting (deep) neural networks. Specifically, the paper models the underlying operations of neural networks as a “dynamic process”, wherein information is propagated “over time” (i.e., goes through layers of neurons). By acting upon such intuition, th...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper tackles the problem of detection of adversarial examples targeting (deep) neural networks. Specifically, the paper models the underlying operations of neural networks as a “dynamic process”, wherein information is propagated “over time” (i.e., goes through layers of neurons). By acting upon such intui...
This paper focuses on recognizing actions based on the motion information in the videos. The proposed idea (Dense Correlation Fields, DCF) which can be applied to different CNN backbones, computes frame to frame similarity to find spatial-temporal correlation fields. - DCF is a simple and novel idea that outperforms S...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper focuses on recognizing actions based on the motion information in the videos. The proposed idea (Dense Correlation Fields, DCF) which can be applied to different CNN backbones, computes frame to frame similarity to find spatial-temporal correlation fields. - DCF is a simple and novel idea that outpe...
This paper proposes a hyperparameter optimization algorithm that leverages the power law phenomenon of the learning curves. In particular, the proposed method DPL assumes the surrogate $\hat f(\lambda)$ of the true learning curve $f(\lambda)$ obeys a power law function, which can be parameterized by a neural network. ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a hyperparameter optimization algorithm that leverages the power law phenomenon of the learning curves. In particular, the proposed method DPL assumes the surrogate $\hat f(\lambda)$ of the true learning curve $f(\lambda)$ obeys a power law function, which can be parameterized by a neural n...
This paper introduces deep metric learning for generalized zero-shot learning, and proposes a novel framework named MLIS to avoid overfitting to seen classes. MLIS disentangles the effect of semantics on feature learning and classification, making the training of the two tasks more effective. Moreover, this paper conca...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper introduces deep metric learning for generalized zero-shot learning, and proposes a novel framework named MLIS to avoid overfitting to seen classes. MLIS disentangles the effect of semantics on feature learning and classification, making the training of the two tasks more effective. Moreover, this pap...
This paper proposes a dynamic scheduling strategy for the value of the weight decay hyperparameter. During the training procedure, the authors keep the ratio of the value of weight decay value to the gradient of cross-entropy loss (dubbed as DoG) at a constant. Empirically, this adaptive weight decay strategy can help ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a dynamic scheduling strategy for the value of the weight decay hyperparameter. During the training procedure, the authors keep the ratio of the value of weight decay value to the gradient of cross-entropy loss (dubbed as DoG) at a constant. Empirically, this adaptive weight decay strategy c...
The paper studies representation learning for Markov games with low-rank structures. Under the low-rank assumption, the authors develop two representation learning methods to learn the underlying transitions: model-based and model-free representation learning and add the UCB bonus into the estimated Q-value functions f...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies representation learning for Markov games with low-rank structures. Under the low-rank assumption, the authors develop two representation learning methods to learn the underlying transitions: model-based and model-free representation learning and add the UCB bonus into the estimated Q-value fun...
The proposed method is a offline language-based goal-conditioned reinforcement learning. It needs human annotator for labeling basic tasks with languages. It implements large languages models for relabeling and cross-trajectory chaining, in order to increase the diversity of the task set. Experiments are conducted to s...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The proposed method is a offline language-based goal-conditioned reinforcement learning. It needs human annotator for labeling basic tasks with languages. It implements large languages models for relabeling and cross-trajectory chaining, in order to increase the diversity of the task set. Experiments are conduc...
This work addresses the problem of using large language models for understanding HTML. Unlike prior work which attempt to solve this problem using dedicated architectures and training procedures and/or large HTML corpora, this work employs large language models pretrained on natural language text and evaluates their pe...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work addresses the problem of using large language models for understanding HTML. Unlike prior work which attempt to solve this problem using dedicated architectures and training procedures and/or large HTML corpora, this work employs large language models pretrained on natural language text and evaluates ...
This paper proposes a curriculum learning approach for reinforcement learning (RL) that changes the morphology of the agent as well as the environment for training agents that are robust to test-time environments. The authors provide thorough experimental results and ablation studies in three domains and demonstrate th...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a curriculum learning approach for reinforcement learning (RL) that changes the morphology of the agent as well as the environment for training agents that are robust to test-time environments. The authors provide thorough experimental results and ablation studies in three domains and demons...
This paper focuses on the setting of unsupervised domain generalization, where the multiple source domains do not have label supervision and the target domain is unseen. Based on the previous work (i.e., DiMAE) that applies the masked-reconstruction-modeling as a self-supervised task, this paper further proposes two mo...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper focuses on the setting of unsupervised domain generalization, where the multiple source domains do not have label supervision and the target domain is unseen. Based on the previous work (i.e., DiMAE) that applies the masked-reconstruction-modeling as a self-supervised task, this paper further propose...
The authors have introduced the concept of coverability, and related it to the sample efficiency in the online reinforcement learning. This idea is very interesting and the paper is well written with detailed description of everything. Strengths - The paper is well written. - The authors have addressed a crucial ch...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors have introduced the concept of coverability, and related it to the sample efficiency in the online reinforcement learning. This idea is very interesting and the paper is well written with detailed description of everything. Strengths - The paper is well written. - The authors have addressed a cr...