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 sample-efficient learning of POMDPs. The paper considers POMDPs with infinite (continuous) state and observation spaces and admitting low-rank latent transitions with features belonging to a certain feature set. The main result is the identification of a sufficient condition called past- and future-s...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies sample-efficient learning of POMDPs. The paper considers POMDPs with infinite (continuous) state and observation spaces and admitting low-rank latent transitions with features belonging to a certain feature set. The main result is the identification of a sufficient condition called past- and ...
The authors suggest viewing the problem of predicting the next event for temporal point process models as a meta-learning problem. In particular, they advocate a neural process framework, whereby windows of previous event times act as context and target input sets. A cross-attention architecture is suggested to increas...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The authors suggest viewing the problem of predicting the next event for temporal point process models as a meta-learning problem. In particular, they advocate a neural process framework, whereby windows of previous event times act as context and target input sets. A cross-attention architecture is suggested to...
This paper proposes a multi-channel self supervised learning method to overcome two issues in analyzing brain signals: (1) existing methods are often limited to a particular type of brain signal data, either the invasive data (e.g., SEEG) or non-invasive data (e.g., EEG). (2) correlations amongst different brain areas ...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a multi-channel self supervised learning method to overcome two issues in analyzing brain signals: (1) existing methods are often limited to a particular type of brain signal data, either the invasive data (e.g., SEEG) or non-invasive data (e.g., EEG). (2) correlations amongst different brai...
The authors address a challenging and important problem of feature selection. The new approach is a wrapper-type method that relies on meta-learning to select the best subset of features for supervised learning. The idea is to transform the large discrete search space into a relaxed continuous space and use gradient de...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors address a challenging and important problem of feature selection. The new approach is a wrapper-type method that relies on meta-learning to select the best subset of features for supervised learning. The idea is to transform the large discrete search space into a relaxed continuous space and use gra...
This paper studies zero-shot RL in the sense of "RL that does not require optimization of a policy when presented with a task (reward model)". It covers the literature on successor representations (SR) and forward-backward (FB) representation, proposes an unifying view and improved training losses. Then it empirically ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies zero-shot RL in the sense of "RL that does not require optimization of a policy when presented with a task (reward model)". It covers the literature on successor representations (SR) and forward-backward (FB) representation, proposes an unifying view and improved training losses. Then it empi...
This paper proposes a vision-and-language transformer that is multi-tasked to both generate an image completion, given language ("Prefix Image Modeling") and to both generate language given images ("Prefix Language Modeling"). The key claim of this paper is that by multitasking these two settings, the resulting model i...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a vision-and-language transformer that is multi-tasked to both generate an image completion, given language ("Prefix Image Modeling") and to both generate language given images ("Prefix Language Modeling"). The key claim of this paper is that by multitasking these two settings, the resulting...
This paper studies the zeroth-order optimization where the gradient information is infeasible or very expensive to access, and only function values are available. Different from most one or two-point zeroth-order estimation, this paper assumes the function f is sampled from Gaussian process as in Bayesian optimization ...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies the zeroth-order optimization where the gradient information is infeasible or very expensive to access, and only function values are available. Different from most one or two-point zeroth-order estimation, this paper assumes the function f is sampled from Gaussian process as in Bayesian optim...
This work explores the relationship between emergent communication informativeness and complexity, hyperparameters of the vector-quantized variational information bottleneck (VA-VIB) method of training EC agents, and downstream metrics like generalizability and translatability with human language. It finds that as the ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work explores the relationship between emergent communication informativeness and complexity, hyperparameters of the vector-quantized variational information bottleneck (VA-VIB) method of training EC agents, and downstream metrics like generalizability and translatability with human language. It finds that...
This paper presents a test-time adaptation (TTA) method from probabilistic perspective. Specifically, they analyze first analyze the pseudo labeling, which is a naive approach in TTA, from probabilistic point of view. Based on the analysis, they propose variational pseudo labels,　and meta-learning based algorithms. The...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper presents a test-time adaptation (TTA) method from probabilistic perspective. Specifically, they analyze first analyze the pseudo labeling, which is a naive approach in TTA, from probabilistic point of view. Based on the analysis, they propose variational pseudo labels,　and meta-learning based algorit...
The paper introduces a new framework to create recourse to algorithms with higher user satisfaction than comparable approaches, using Expected Minimum Risk and a novel optimization algorithm. Recourse here is meant to enable users to make changes in algorithmic classification, altering the result according to their pre...	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 paper introduces a new framework to create recourse to algorithms with higher user satisfaction than comparable approaches, using Expected Minimum Risk and a novel optimization algorithm. Recourse here is meant to enable users to make changes in algorithmic classification, altering the result according to t...
The paper provides a novel methodology for identifying significant causal features in spatiotemporal data, with primary applications in biology (e.g. scRNA-Seq data). Its primary novelty lies in the judicious use of a transfer entropy characterisation of feature significance/relevance and the design of a combinatorial ...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper provides a novel methodology for identifying significant causal features in spatiotemporal data, with primary applications in biology (e.g. scRNA-Seq data). Its primary novelty lies in the judicious use of a transfer entropy characterisation of feature significance/relevance and the design of a combin...
This paper proposes a novel neural episodic-control based approach with state abstraction, named NECSA. The authors motivate their methods by utilizing latent information from historical behaviors, which has been overlooked by prior work. Specifically, this paper proposes a more comprehensive episodic memory, which con...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a novel neural episodic-control based approach with state abstraction, named NECSA. The authors motivate their methods by utilizing latent information from historical behaviors, which has been overlooked by prior work. Specifically, this paper proposes a more comprehensive episodic memory, w...
In this paper, a method called Normalization Perturbation (NP) is proposed for robust object detection under domain shift. The proposed method perturbs the channel statistics of source domain features to synthesize various latent styles. Extensive analysis and experiments verify the effectiveness of the proposed NP met...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: In this paper, a method called Normalization Perturbation (NP) is proposed for robust object detection under domain shift. The proposed method perturbs the channel statistics of source domain features to synthesize various latent styles. Extensive analysis and experiments verify the effectiveness of the propose...
In this paper, the problem of imitating an expert with access to more information about the state space in a POMDP than a learner is considered -- these are called impossibly good experts. Existing methods for imitation learning like Behavior Cloning and the DAGGER algorithm are shown to fail with concrete counterexamp...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: In this paper, the problem of imitating an expert with access to more information about the state space in a POMDP than a learner is considered -- these are called impossibly good experts. Existing methods for imitation learning like Behavior Cloning and the DAGGER algorithm are shown to fail with concrete coun...
This paper attempts to proposed a new method for so-called adversary-aware partial label learning (but the definition of this setting is not clear in the whole paper). The proposed method is an incremental improvement on a SOTA partial label learning method. Strengths: It is difficult for me to tell the strengths about...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper attempts to proposed a new method for so-called adversary-aware partial label learning (but the definition of this setting is not clear in the whole paper). The proposed method is an incremental improvement on a SOTA partial label learning method. Strengths: It is difficult for me to tell the strengt...
This paper aims to improve the communication efficiency in decentralized learning. They propose a novel algorithm, DIGEST, which uses local SGD and random walk to reduce communication costs. The random walk communication can be extended to multi-stream. They provide a standard convergence results and provide empirical...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper aims to improve the communication efficiency in decentralized learning. They propose a novel algorithm, DIGEST, which uses local SGD and random walk to reduce communication costs. The random walk communication can be extended to multi-stream. They provide a standard convergence results and provide e...
This paper tries to fix the failure of the direct combination of federated learning (FL) and adversarial training, which is indicated by the deterioration of the adversarial accuracy at the later stage of FL. The authors proposed the so-called "alpha-slack" mechanism that upweights the clients with smaller (local) adve...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper tries to fix the failure of the direct combination of federated learning (FL) and adversarial training, which is indicated by the deterioration of the adversarial accuracy at the later stage of FL. The authors proposed the so-called "alpha-slack" mechanism that upweights the clients with smaller (loc...
This paper proposes a framework for learning provably feasible approximations to optimization problems with linear constraints. This framework entails * Rewriting the original optimization problem, if it has (linear) equality constraints, in a solely inequality-constrained form, by using variable reduction techniques, ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a framework for learning provably feasible approximations to optimization problems with linear constraints. This framework entails * Rewriting the original optimization problem, if it has (linear) equality constraints, in a solely inequality-constrained form, by using variable reduction tech...
This submission studies an architecture to learn from data with missing values. The architecture is based on chaining two transformations with an attention module that enables modeling the set of observed values. The model learns by randomly masking the training data to ensure learning from different subsets of the var...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This submission studies an architecture to learn from data with missing values. The architecture is based on chaining two transformations with an attention module that enables modeling the set of observed values. The model learns by randomly masking the training data to ensure learning from different subsets of...
This paper proposes to use an adaptive induction network to rectify the prototype network for few-shot learning. The adaptive induction network is developed to address the noisy samples, while this network has degraded performance with the increase of support samples since it ignores the relative local features. Thus, ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes to use an adaptive induction network to rectify the prototype network for few-shot learning. The adaptive induction network is developed to address the noisy samples, while this network has degraded performance with the increase of support samples since it ignores the relative local features...
In this paper, the author provides a comprehensive formulation of the attention flows for different architectures of transformers (encoder, decoder, encoder-decoder). In order to account for the auto-regressive decoder structure, the author adjusted the attention flow to ensure the positional independence of the comput...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: In this paper, the author provides a comprehensive formulation of the attention flows for different architectures of transformers (encoder, decoder, encoder-decoder). In order to account for the auto-regressive decoder structure, the author adjusted the attention flow to ensure the positional independence of th...
This paper proposed a gradient-guided importance sampling method for learning binary energy based models. The idea is to combine ratio matching with scalable gradient of the energy function for more efficient computation. The paper is clearly written. Numerical experiments show the advantage of the proposed method over...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper proposed a gradient-guided importance sampling method for learning binary energy based models. The idea is to combine ratio matching with scalable gradient of the energy function for more efficient computation. The paper is clearly written. Numerical experiments show the advantage of the proposed met...
This paper proposes to use implicit neural representations to model and efficiently solve PDE dynamics. This allows to solve dynamics in the latent space instead of the whole parameter space (lower dimension), without discretizing it. The latent space and network are trained on a range of sample PDEs for a given task, ...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes to use implicit neural representations to model and efficiently solve PDE dynamics. This allows to solve dynamics in the latent space instead of the whole parameter space (lower dimension), without discretizing it. The latent space and network are trained on a range of sample PDEs for a give...
The paper analyzes the reactive defense against model stealing called dataset inference that was proposed at ICLR 2021. It is shown in this submission that dataset inference suffers from false positives (FP) and false negatives (FN). For FPs - it is presented that dataset inference can incorrectly resolve the model own...	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 analyzes the reactive defense against model stealing called dataset inference that was proposed at ICLR 2021. It is shown in this submission that dataset inference suffers from false positives (FP) and false negatives (FN). For FPs - it is presented that dataset inference can incorrectly resolve the m...
This paper studies the behavior of the sim-to-real gap for partially observed linear-Gaussian systems with quadratic cost. The authors formalize the sim-to-real gap as the finite-horizon minimax regret of a policy across a set of plausible simulators. The main argument is that the finite-horizon minimax regret can be b...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the behavior of the sim-to-real gap for partially observed linear-Gaussian systems with quadratic cost. The authors formalize the sim-to-real gap as the finite-horizon minimax regret of a policy across a set of plausible simulators. The main argument is that the finite-horizon minimax regret ...
In this paper the author address the partial-label learning problem, a typical weakly supervised learning problem, where each training instance is equipped with a set of candidate labels among which only one is the true label. In this scenario, the main challenge lies in candidate label disambiguation. The authors rais...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: In this paper the author address the partial-label learning problem, a typical weakly supervised learning problem, where each training instance is equipped with a set of candidate labels among which only one is the true label. In this scenario, the main challenge lies in candidate label disambiguation. The auth...
This paper proposes a privacy-preserving framework (SNFL) in federated learning by incorporating the conversion from artificial neural networks to spiking neural networks. SNFL is validated to robust to gradient inversion attack and backdoor attack and it guarantees privacy while improving accuracy. Strength: This pap...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a privacy-preserving framework (SNFL) in federated learning by incorporating the conversion from artificial neural networks to spiking neural networks. SNFL is validated to robust to gradient inversion attack and backdoor attack and it guarantees privacy while improving accuracy. Strength: ...
This paper tackles a challenging problem of translating sign language to natural language. The problem's challenging nature arises from the data scarcity and the modality gap between video and text. The authors approach the problem from the perspective of learning a joint latent space for video and text, guided by mult...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper tackles a challenging problem of translating sign language to natural language. The problem's challenging nature arises from the data scarcity and the modality gap between video and text. The authors approach the problem from the perspective of learning a joint latent space for video and text, guided...
The manuscript proposes to bridge the gap between post-hoc explanation methods and interpretable-by-design methods. This is achieved by a pipeline where given a black-box model, an interpretable component iteratively distills parts of the representation from the blackbox with uncovered (undistilled) parts forwarded to...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The manuscript proposes to bridge the gap between post-hoc explanation methods and interpretable-by-design methods. This is achieved by a pipeline where given a black-box model, an interpretable component iteratively distills parts of the representation from the blackbox with uncovered (undistilled) parts forw...
This paper proposes a probabilistic model to learn disentangled object representation in an unsupervised fashion. The generative model is expressed through an observation module and a dynamic module, where the representations of entities are factorized into interaction-relevant relational features and interaction-irrel...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper proposes a probabilistic model to learn disentangled object representation in an unsupervised fashion. The generative model is expressed through an observation module and a dynamic module, where the representations of entities are factorized into interaction-relevant relational features and interacti...
This paper targets the problem that the edges in a graph may have different levels of difficulty to learn and proposes a curriculum learning-based model to gradually incorporate edges during learning according to their difficulties. The difficulty level is obtained by self-supervised learning. Experiments are conducted...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper targets the problem that the edges in a graph may have different levels of difficulty to learn and proposes a curriculum learning-based model to gradually incorporate edges during learning according to their difficulties. The difficulty level is obtained by self-supervised learning. Experiments are c...
This paper focuses on multi-morphology and multi-task generalization in robotics. There are three main contributions to this paper. First, a comprehensive benchmark called MxT-Bench is developed for training and evaluating morphology-task generalization. This benchmark also supports the scalable procedural generation o...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper focuses on multi-morphology and multi-task generalization in robotics. There are three main contributions to this paper. First, a comprehensive benchmark called MxT-Bench is developed for training and evaluating morphology-task generalization. This benchmark also supports the scalable procedural gene...
In this paper authors study wide stochastic networks, where the weights are sampled from a Gaussian with diagonal covariance. Authors aim to give a convergence analysis when the means and variances are obtained via gradient descent trained on the PAC-Bayesian bound objective. Authors obtain these results by introduc...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: In this paper authors study wide stochastic networks, where the weights are sampled from a Gaussian with diagonal covariance. Authors aim to give a convergence analysis when the means and variances are obtained via gradient descent trained on the PAC-Bayesian bound objective. Authors obtain these results by ...
The authors propose a neural network pruning method which perserves trainability after pruning. The authors argue that the proposed approach will make the network easier to train after pruning therefore lead to a better finetuned model. First, the proposed method use L1 norm to select important and unimportant filters....	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose a neural network pruning method which perserves trainability after pruning. The authors argue that the proposed approach will make the network easier to train after pruning therefore lead to a better finetuned model. First, the proposed method use L1 norm to select important and unimportant ...
This paper proposed the H3 architecture to narrow the gap between state space models and attention models on language modeling. Basically, H3 uses a shift SSM and a diagonal SSM to replace the non-linear function in linear attention. The shift SSM is designed to remember tokens in the input sequence while the multiplic...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed the H3 architecture to narrow the gap between state space models and attention models on language modeling. Basically, H3 uses a shift SSM and a diagonal SSM to replace the non-linear function in linear attention. The shift SSM is designed to remember tokens in the input sequence while the m...
the paper proposes the modification of AdaSAM to SAM (sharpness-aware minimization) that records running averages of first and second moment (based on the "SAM gradient"), and adjusts the parameter updated based on these estimates a la Adam and AdaGrad. authors show improved results with AdaSAM on GLUE benchmarks. str...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: the paper proposes the modification of AdaSAM to SAM (sharpness-aware minimization) that records running averages of first and second moment (based on the "SAM gradient"), and adjusts the parameter updated based on these estimates a la Adam and AdaGrad. authors show improved results with AdaSAM on GLUE benchmar...
This paper presents a simple and effective Self-attentive Rationale guided Graph Contrastive Learning for graph network learning. The code idea is to learn both node- and edge-wise rationale-aware views. Extensive experimental results show the effectiveness of the proposed method. Strength: 1. The writing is good. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper presents a simple and effective Self-attentive Rationale guided Graph Contrastive Learning for graph network learning. The code idea is to learn both node- and edge-wise rationale-aware views. Extensive experimental results show the effectiveness of the proposed method. Strength: 1. The writing is...
This paper proposes to directly model feature embeddings from a line graph. The line graph is comprised of nodes as the original temporal edges, and new set of edges between the line graph nodes are captured by interactions between the original temporal edges. The paper addresses critical ML application in the context...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes to directly model feature embeddings from a line graph. The line graph is comprised of nodes as the original temporal edges, and new set of edges between the line graph nodes are captured by interactions between the original temporal edges. The paper addresses critical ML application in the...
This paper proposes two methods for equivariance-aware neural architectural search (NAS). The motivation for this work stems from the fact that symmetries present in a data set might often be imperfect or non-explicitly known. The first method is based on the proposed equivariance relaxation morphism, a procedure that ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes two methods for equivariance-aware neural architectural search (NAS). The motivation for this work stems from the fact that symmetries present in a data set might often be imperfect or non-explicitly known. The first method is based on the proposed equivariance relaxation morphism, a procedu...
This paper introduces a new measure of conditional independence for multivariate continuous variables, based on the equations of Daudin, 1980. Although previous tests have also used other equivalent forms of equations from (Daudin, 1980), the proposed CI measure and statistic seem to enjoy the advantage of avoiding m...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper introduces a new measure of conditional independence for multivariate continuous variables, based on the equations of Daudin, 1980. Although previous tests have also used other equivalent forms of equations from (Daudin, 1980), the proposed CI measure and statistic seem to enjoy the advantage of av...