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 studies the impact of graph convolutions on a binary node classification problem. In this setting, the relationships among nodes are sampled from a given stochastic block model and node features are associated with a Gaussian mixture model. Theoretical results on the performance of a two or three layer graph...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies the impact of graph convolutions on a binary node classification problem. In this setting, the relationships among nodes are sampled from a given stochastic block model and node features are associated with a Gaussian mixture model. Theoretical results on the performance of a two or three lay...
The paper proposes an exact nonnegative matrix factorization method that provides interpretable node embeddings (interpretation as the cluster/community assignment). The authors prove that under the constraint of embedding dimensionality, the embeddings can reconstruct adjacency matrix exactly. A training algorithm is ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper proposes an exact nonnegative matrix factorization method that provides interpretable node embeddings (interpretation as the cluster/community assignment). The authors prove that under the constraint of embedding dimensionality, the embeddings can reconstruct adjacency matrix exactly. A training algor...
This paper proposes a new framework for inverting backdoor triggers from backdoor classifiers. Previous approaches is only applicable to patch-based backdoor. The authors proposed an unified approaches. Experiments on various different attacks demonstrate that the proposed approach is able to generate triggers with hig...	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 proposes a new framework for inverting backdoor triggers from backdoor classifiers. Previous approaches is only applicable to patch-based backdoor. The authors proposed an unified approaches. Experiments on various different attacks demonstrate that the proposed approach is able to generate triggers ...
This paper studies the geometric convergence of NPG parametrized by log linear function. The authors assume that approximation of Q-functions by linear function, and approximation of policy with log-linear function has some error bound, and they show that employing NPG with an increasing step size result in exponential...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the geometric convergence of NPG parametrized by log linear function. The authors assume that approximation of Q-functions by linear function, and approximation of policy with log-linear function has some error bound, and they show that employing NPG with an increasing step size result in exp...
This paper proposes a novel node-level pretraining task and a graph-level contrastive learning task for molecule graph representation learning. The node-level pretraining task is to predict a discrete tokenized atom that also encodes the context information of the target masked atom. This new task is more challenging ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a novel node-level pretraining task and a graph-level contrastive learning task for molecule graph representation learning. The node-level pretraining task is to predict a discrete tokenized atom that also encodes the context information of the target masked atom. This new task is more chal...
In this work, the authors tackle the problem of long-tailed recognition. The authors highlight the issue of Class-Conditional Distribution (CCD) shift due to scarce instances. The authors propose an adaptive data augmentation method, Distributionally Robust Augmentation (DRA) to improve performance for long-tailed reco...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this work, the authors tackle the problem of long-tailed recognition. The authors highlight the issue of Class-Conditional Distribution (CCD) shift due to scarce instances. The authors propose an adaptive data augmentation method, Distributionally Robust Augmentation (DRA) to improve performance for long-tai...
In this research, for the task of tabular data learning, an algorithm that searches for features used and the base learner used jointly has been proposed. Within the research, it has been found that on a large number of different tasks, the method outperforms the existing baselines. Strengths: 1- This is a pioneering ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: In this research, for the task of tabular data learning, an algorithm that searches for features used and the base learner used jointly has been proposed. Within the research, it has been found that on a large number of different tasks, the method outperforms the existing baselines. Strengths: 1- This is a pio...
This paper proposes an imputation method for missing node features in graph neural networks. There is a node-wise diffusion as well as channel-wise diffusion that is based on the "pseudo confidence" of a missing feature node, which is inversely proportional to the distance to the closest observed feature node. Semi-sup...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes an imputation method for missing node features in graph neural networks. There is a node-wise diffusion as well as channel-wise diffusion that is based on the "pseudo confidence" of a missing feature node, which is inversely proportional to the distance to the closest observed feature node. ...
This paper studies federated learning where the local data on each client are non-iid in the sense that they may encounter openset noisy labels. To solve the problem, they extend the so-called peer loss which is studied in centralized label noise learning to the federated learning scenario and call the proposed method ...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper studies federated learning where the local data on each client are non-iid in the sense that they may encounter openset noisy labels. To solve the problem, they extend the so-called peer loss which is studied in centralized label noise learning to the federated learning scenario and call the proposed...
This paper proposes to use a queue of embeddings of samples from the unlabeled dataset to improve the performance of self-supervised learning. It presents that the proposed method have achieved success on the tabular datasets. Strength: This paper has conducted some interesting experiments. Weaknesses: 1. The techni...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper proposes to use a queue of embeddings of samples from the unlabeled dataset to improve the performance of self-supervised learning. It presents that the proposed method have achieved success on the tabular datasets. Strength: This paper has conducted some interesting experiments. Weaknesses: 1. Th...
This paper aims to transfer the point cloud knowledge from a LiDAR sensor during the training phase to the camera-only testing scenario. The authors apply knowledge distillation with a LiDAR teacher and a camera student. They also use the multi-level adversarial learning mechanism to adapt the features learned from dif...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper aims to transfer the point cloud knowledge from a LiDAR sensor during the training phase to the camera-only testing scenario. The authors apply knowledge distillation with a LiDAR teacher and a camera student. They also use the multi-level adversarial learning mechanism to adapt the features learned ...
This paper finds that non-smooth ReLU activation function weaken the adversarial robustness and smooth activation functions such as SiLU can improve the adversarial robustness. The experiments are based on large-scale ImageNet dataset and show the state-of-the-art performance. Strength - This paper focuses on ImageNet...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper finds that non-smooth ReLU activation function weaken the adversarial robustness and smooth activation functions such as SiLU can improve the adversarial robustness. The experiments are based on large-scale ImageNet dataset and show the state-of-the-art performance. Strength - This paper focuses on ...
The paper demonstrates that large language models (LLMs) can be self-improved by generating high-confidence rationale augmented answers for unlabelled questions using an LLM and use that data to fine-tune the same LLM to improve itself. The paper empirically shows that this approach improves the reasoning abilities of ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper demonstrates that large language models (LLMs) can be self-improved by generating high-confidence rationale augmented answers for unlabelled questions using an LLM and use that data to fine-tune the same LLM to improve itself. The paper empirically shows that this approach improves the reasoning abili...
This paper introduces S-Mixup. It is a graph Mixup method, that can better utilize the node-level information between the pair-wise graph. ## Strengths 1. The problem is well-motivated. Indeed, graph data is highly structured, and the standard augmentation methods can easily destroy the key substructures, which can le...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces S-Mixup. It is a graph Mixup method, that can better utilize the node-level information between the pair-wise graph. ## Strengths 1. The problem is well-motivated. Indeed, graph data is highly structured, and the standard augmentation methods can easily destroy the key substructures, whic...
In this article, the authors propose to study and improve fairness in graph neural networks. More precisely, they provide conditions under which the graph aggregation steps increase prediction biases, and they propose an optimization problem that allows to find an optimal graph topology that reduces biases while preser...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this article, the authors propose to study and improve fairness in graph neural networks. More precisely, they provide conditions under which the graph aggregation steps increase prediction biases, and they propose an optimization problem that allows to find an optimal graph topology that reduces biases whil...
This work attempts to deal with node heterogeneity for graph learning task. The proposed method dissects the node representations into feature and structure views. Finally a prototype-based node representation method is designed by separately modeling feature view with prototype-based MLP network and encoding graph str...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This work attempts to deal with node heterogeneity for graph learning task. The proposed method dissects the node representations into feature and structure views. Finally a prototype-based node representation method is designed by separately modeling feature view with prototype-based MLP network and encoding g...
The authors propose a transformer-based pre-training and fine-tuning framework for TEE from observational data. The proposed method is pre-trained to learn representative contextual patient representations. Then, fine-tuned on labeled patient data for TEE. The proposed method is evaluated on 4 downstream TEE tasks. S...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose a transformer-based pre-training and fine-tuning framework for TEE from observational data. The proposed method is pre-trained to learn representative contextual patient representations. Then, fine-tuned on labeled patient data for TEE. The proposed method is evaluated on 4 downstream TEE ta...
This paper analyzes the convergence of gradient descent on overparameterized neural networks with smooth activation functions. The authors show that the Hessian of the neural network has a small spectral norm within a special ball of the initialization, and this ball is allowed to be significantly larger than those all...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper analyzes the convergence of gradient descent on overparameterized neural networks with smooth activation functions. The authors show that the Hessian of the neural network has a small spectral norm within a special ball of the initialization, and this ball is allowed to be significantly larger than t...
This paper presents a multivariate time-series forecasting framework that combines GAT with a forecasting framework. The paper provides detail about the GAT framework. The way they integrate it seems to have been inspired by Wu et al. IJCAI 19 but differs in the choice of the specific GAT parameterization. The second h...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a multivariate time-series forecasting framework that combines GAT with a forecasting framework. The paper provides detail about the GAT framework. The way they integrate it seems to have been inspired by Wu et al. IJCAI 19 but differs in the choice of the specific GAT parameterization. The ...
This paper proposed to unify relative positional encoding methods by formulating them as a quadratic form and introduced a class of decomposable relative positional encoding that is equipable by linear attention. Strength: - The unified form of relative positional encoding is meaningful in the sense that this formulati...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed to unify relative positional encoding methods by formulating them as a quadratic form and introduced a class of decomposable relative positional encoding that is equipable by linear attention. Strength: - The unified form of relative positional encoding is meaningful in the sense that this f...
This paper studies reward-free RL with safe exploration. The authors consider the scenario where a safe baseline policy is given beforehand, propose a unified algorithmic framework called SWEET, and instantiate the SWEET framework to the tabular and low-rank MDP settings. Their algorithms utilize the concavity and cont...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper studies reward-free RL with safe exploration. The authors consider the scenario where a safe baseline policy is given beforehand, propose a unified algorithmic framework called SWEET, and instantiate the SWEET framework to the tabular and low-rank MDP settings. Their algorithms utilize the concavity ...
The paper proposes a distillation framework that makes a network learn with crops and augmentations from a single image. The learning target comes from a pre-trained teacher's outputs. The results obtained is of decent level (69% ImageNet top1). Various design options and internal properties of the resulting network ar...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a distillation framework that makes a network learn with crops and augmentations from a single image. The learning target comes from a pre-trained teacher's outputs. The results obtained is of decent level (69% ImageNet top1). Various design options and internal properties of the resulting ne...
The authors propose non-monotonic self-terminating (NMST) decoding algorithm, improving upon [Welleck,et. al. 2020] ST algorithm. The paper first show that a language model is non-monotonic if it is trained on dataset that has the same prefix with two different lengths samples. It also points out that the monotonicity ...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose non-monotonic self-terminating (NMST) decoding algorithm, improving upon [Welleck,et. al. 2020] ST algorithm. The paper first show that a language model is non-monotonic if it is trained on dataset that has the same prefix with two different lengths samples. It also points out that the monot...
This work focuses on the many-to-one matching problem in the bandits, and the uniqueness condition is fundamental in their problem setup. The author(s) designs the MO-UCB-D4 algorithm and derives an upper bound. The paper also presents experimental results to convince the efficacy of the proposed algorithm. Strength: ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This work focuses on the many-to-one matching problem in the bandits, and the uniqueness condition is fundamental in their problem setup. The author(s) designs the MO-UCB-D4 algorithm and derives an upper bound. The paper also presents experimental results to convince the efficacy of the proposed algorithm. St...
The paper proposed a method for generating stable and safe policies for stochastic delay-differential equations. Specifically, the paper achieves this by learning a LaSalle's type stability certificate as an MLP and extending it to a stochastic control barrier function. # Strengths: - The paper relies on minimal assump...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper proposed a method for generating stable and safe policies for stochastic delay-differential equations. Specifically, the paper achieves this by learning a LaSalle's type stability certificate as an MLP and extending it to a stochastic control barrier function. # Strengths: - The paper relies on minima...
The paper studies 'probabilistic dynamics model ensemble' method in reinforcement learning. It proposes that (1) an important factor to the convergence of RL is the Lipschitz condition of the value function and (2) model ensemble helps to regularize the Lipschitz condition in the training. The paper provides both theor...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper studies 'probabilistic dynamics model ensemble' method in reinforcement learning. It proposes that (1) an important factor to the convergence of RL is the Lipschitz condition of the value function and (2) model ensemble helps to regularize the Lipschitz condition in the training. The paper provides bo...
The paper proposes a novel communication scheme, Sequential Communication (SeqComm), for cooperative multi-agent reinforcement learning (MARL). In communication for cooperation in MARL, circular dependencies can sometimes occur. This is caused by synchronization in communication. The proposed model assumes asynchronous...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a novel communication scheme, Sequential Communication (SeqComm), for cooperative multi-agent reinforcement learning (MARL). In communication for cooperation in MARL, circular dependencies can sometimes occur. This is caused by synchronization in communication. The proposed model assumes asyn...
In the existing meta-gradient studies, the adaptive return is simply formulated in the form of expected cumulative rewards. In this paper, the authors address the meta learning problem of the distributional actor-critic method with quantile distributions by considering the distributional $\lambda$ return and apply the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In the existing meta-gradient studies, the adaptive return is simply formulated in the form of expected cumulative rewards. In this paper, the authors address the meta learning problem of the distributional actor-critic method with quantile distributions by considering the distributional $\lambda$ return and ap...
The paper proposes a new adaptive quantization strategy (and aggregation criteria), AQUILA, by adaptively determining the quantization bits per round per client in lazily aggreagated federated learning. The quantization level $b^*$ is chosen based on the $l_\infty$ and $l_2$ norm of the difference between the current g...	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 proposes a new adaptive quantization strategy (and aggregation criteria), AQUILA, by adaptively determining the quantization bits per round per client in lazily aggreagated federated learning. The quantization level $b^*$ is chosen based on the $l_\infty$ and $l_2$ norm of the difference between the c...
The paper proposes to bring MIMO for CNNs to vision transformers. Specifically, it designs a source attributing module at last layer to separate the input source and perform classification for different input. Experiments are mainly conducted on small-scale datasets, CIFAR, Imagenet-100 to show its effectiveness. St...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes to bring MIMO for CNNs to vision transformers. Specifically, it designs a source attributing module at last layer to separate the input source and perform classification for different input. Experiments are mainly conducted on small-scale datasets, CIFAR, Imagenet-100 to show its effectivenes...
This paper proposes to predict solutions of PDE locally and unite them into a global solution. The proposed neural operator scales to large resolutions by leveraging local and global structures by decomposing both the input domain and the operator's parameter space. The proposed method can be trained with much fewer sa...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to predict solutions of PDE locally and unite them into a global solution. The proposed neural operator scales to large resolutions by leveraging local and global structures by decomposing both the input domain and the operator's parameter space. The proposed method can be trained with much ...
This paper proposes a new target value estimate approach, Graph backup, based on tree-backup. The main idea is to leverage the crossover on state space to accelerate value propagation using a graph. In graph backup, a state-transition graph is built and the target value for each value backup is estimated by looking at ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a new target value estimate approach, Graph backup, based on tree-backup. The main idea is to leverage the crossover on state space to accelerate value propagation using a graph. In graph backup, a state-transition graph is built and the target value for each value backup is estimated by loo...
This work proposes a transformer-based architecture with a masked modeling task to operate on intra-cranial recordings directly. Some of the important components of this model include a content-aware loss to capture the burst-like nature of neural activity, a superlet representation of the raw neural data and a masking...	Recommendation: 8: accept, good paper	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This work proposes a transformer-based architecture with a masked modeling task to operate on intra-cranial recordings directly. Some of the important components of this model include a content-aware loss to capture the burst-like nature of neural activity, a superlet representation of the raw neural data and a...
The authors describe a classification framework that consists of a verifier (Arthur) and two provers (one collaborative [Merlin] and one adversarial [Morgana]). The provers aim to convince/trick the verifier in making a correct/incorrect classification. Theoretical results: * **Mutual information bound at equilib...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors describe a classification framework that consists of a verifier (Arthur) and two provers (one collaborative [Merlin] and one adversarial [Morgana]). The provers aim to convince/trick the verifier in making a correct/incorrect classification. Theoretical results: * **Mutual information bound at...
This paper proposes a method for accounting for the geometry of the data in devising methods for enforcing pessimism in offline RL. To do so, the paper ends up training a state-conditioned distance function between actions, and enforces a policy constraint using the learned distance function. The hope is that this is u...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a method for accounting for the geometry of the data in devising methods for enforcing pessimism in offline RL. To do so, the paper ends up training a state-conditioned distance function between actions, and enforces a policy constraint using the learned distance function. The hope is that t...
1. The authors try to infer the underlying dynamics of some unknown dynamical system based on sparsely sampled observations and external inputs. 2. The authors build a framework to infer such a dynamical system using 1) neural networks as function approximators for a generic continuous-time dynamical system, and 2) gra...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: 1. The authors try to infer the underlying dynamics of some unknown dynamical system based on sparsely sampled observations and external inputs. 2. The authors build a framework to infer such a dynamical system using 1) neural networks as function approximators for a generic continuous-time dynamical system, an...
This work proposes a proportional sampling method to estimate the ratios between Shapley values by introducing effective coalitions. Then, it combines this method and the existing Integrated Gradients to propose the Shapley Integrated Gradients that can explain a deep neural network (DNN) prediction by attributing the ...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This work proposes a proportional sampling method to estimate the ratios between Shapley values by introducing effective coalitions. Then, it combines this method and the existing Integrated Gradients to propose the Shapley Integrated Gradients that can explain a deep neural network (DNN) prediction by attribut...
Image retrieval model can be updated when there is more training data or improved architecture, but it is not straightforward since old and new feature spaces are not compatible, leading to poor retrieval results between new query and old gallery embeddings. A naive solution is to replace the old features by new featur...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Image retrieval model can be updated when there is more training data or improved architecture, but it is not straightforward since old and new feature spaces are not compatible, leading to poor retrieval results between new query and old gallery embeddings. A naive solution is to replace the old features by ne...
This paper proposes a knowledge distillation (KD) framework that leverages supervision complexity to measure the alignment between teacher-provided supervision and the student’s neural tangent kernel. It first shows how supervision complexity is defined and measured for kernel-based classifiers, which is then extended ...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper proposes a knowledge distillation (KD) framework that leverages supervision complexity to measure the alignment between teacher-provided supervision and the student’s neural tangent kernel. It first shows how supervision complexity is defined and measured for kernel-based classifiers, which is then e...
This paper investigates tasks in which the "reward hypothesis" is not the most appropriate framework. The authors demonstrate three types of objectives: MORL, risk sensitive RL, and Modal objectives, that cannot mathematically be reduced to a singular MDP formulation. The mathematical demonstrations are themselves int...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper investigates tasks in which the "reward hypothesis" is not the most appropriate framework. The authors demonstrate three types of objectives: MORL, risk sensitive RL, and Modal objectives, that cannot mathematically be reduced to a singular MDP formulation. The mathematical demonstrations are themse...