Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
The paper investigates the difference is speech representations towards end for streaming input vs when complete utterance input is available. Based on this investigation the work proposes future aware streaming ST which introduces few frames of future speech signals. Extensive experiments verify the improvements compa...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper investigates the difference is speech representations towards end for streaming input vs when complete utterance input is available. Based on this investigation the work proposes future aware streaming ST which introduces few frames of future speech signals. Extensive experiments verify the improvemen...
The paper propose a cross attention mechanism for classifying protein-protein interactions. The hypothesis is that information from both proteins get mixed during encoding, leading to better predictions. Paper also propose using the Wasserstein distance measure while processing the attention values. Key new ideas are (...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper propose a cross attention mechanism for classifying protein-protein interactions. The hypothesis is that information from both proteins get mixed during encoding, leading to better predictions. Paper also propose using the Wasserstein distance measure while processing the attention values. Key new ide...
This paper proposed a method called FedPA that deals with client and period drift problems. The period drift problem is caused by the asynchronized updates of each client, leading to extra bias in model aggregation. The authors proposed a learning-based aggregation strategy, that parameterizes the aggregation function ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposed a method called FedPA that deals with client and period drift problems. The period drift problem is caused by the asynchronized updates of each client, leading to extra bias in model aggregation. The authors proposed a learning-based aggregation strategy, that parameterizes the aggregation f...
This paper introduces a new kind of audio-visual navigation task which requires agent to generalize to novel audio sources in 3D indoor environments from the Matterport dataset. The proposed approach leverages multiple sources of information of which objects the source is likely coming from, where they might be found, ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a new kind of audio-visual navigation task which requires agent to generalize to novel audio sources in 3D indoor environments from the Matterport dataset. The proposed approach leverages multiple sources of information of which objects the source is likely coming from, where they might be...
This paper theoretically studies the notion of neural collapse (NC) -- the penultimate layer representations of all the examples in a specific class collapse to a single representation, and separate from other classes. Specifically, this paper studies this notion under uniform label noises, and proposed a model called ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper theoretically studies the notion of neural collapse (NC) -- the penultimate layer representations of all the examples in a specific class collapse to a single representation, and separate from other classes. Specifically, this paper studies this notion under uniform label noises, and proposed a model...
This concerns learning the hierarchical representation (e.g., a.a., backbone, all-atom) for proteins with 3D structures. A novel hierarchical graph network, termed ProNet, was developed. ProNet is flexible, efficient and effective. It is shown that ProNet representations are complete at all levels if the base 3D graph ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This concerns learning the hierarchical representation (e.g., a.a., backbone, all-atom) for proteins with 3D structures. A novel hierarchical graph network, termed ProNet, was developed. ProNet is flexible, efficient and effective. It is shown that ProNet representations are complete at all levels if the base 3...
This work studies patterns in the performance of ML models on in-distribution (ID) and out-of-distribution (OOD) data. The central claim of this work is that ID performance can be inversely correlated with OOD performance in practice, which has not been explicitly observed before in deep ML models. This work then provi...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This work studies patterns in the performance of ML models on in-distribution (ID) and out-of-distribution (OOD) data. The central claim of this work is that ID performance can be inversely correlated with OOD performance in practice, which has not been explicitly observed before in deep ML models. This work th...
The paper proposed a new RL algorithm for continuous control. A key idea is to represent the policy as a distribution of trajectories in the problem space to capture the multi-modality nature of the problem. A variational lower-bound is derived to create a practical optimization objective and with some assumptions on t...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposed a new RL algorithm for continuous control. A key idea is to represent the policy as a distribution of trajectories in the problem space to capture the multi-modality nature of the problem. A variational lower-bound is derived to create a practical optimization objective and with some assumpti...
This work studies the layer-wise representations of different VAEs. The similarity between the layer-wise representations is measured using the Centered Kernel Alignment metric and the Procrustes scores. Experiments are performed to study the effect of different hyperparameters on the same model, regularization, initia...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This work studies the layer-wise representations of different VAEs. The similarity between the layer-wise representations is measured using the Centered Kernel Alignment metric and the Procrustes scores. Experiments are performed to study the effect of different hyperparameters on the same model, regularization...
The paper studies the trade-off between the ability to achieve algorithmic resource vs. the right to be forgotten in machine learning, specifically via linear models and wide neural networks. It illustrates that these two desiderate (algorithimic resource and the right to be forgotten) are fundamentally at odds by defi...	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 studies the trade-off between the ability to achieve algorithmic resource vs. the right to be forgotten in machine learning, specifically via linear models and wide neural networks. It illustrates that these two desiderate (algorithimic resource and the right to be forgotten) are fundamentally at odds...
This paper first compares the performance of contrastive and non-contrastive methods on the graph link prediction task. Second, a novel model i.e. Triplet-BGRL (T-BGRL) is proposed based on Bootstrapped Graph Latents (BGRL). The proposed T-BGRL is categorised as a non-contrastive method and it relies on an efficient co...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper first compares the performance of contrastive and non-contrastive methods on the graph link prediction task. Second, a novel model i.e. Triplet-BGRL (T-BGRL) is proposed based on Bootstrapped Graph Latents (BGRL). The proposed T-BGRL is categorised as a non-contrastive method and it relies on an effi...
The paper proposes a transformer based method to solve offline meta reinforcement learning and offline reinforcement learning tasks. An auto-regressive transformer is learned on the existing trajectory, and prompt tuning is used to guide the prediction towards policy improvement. There is an additional contrastive loss...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes a transformer based method to solve offline meta reinforcement learning and offline reinforcement learning tasks. An auto-regressive transformer is learned on the existing trajectory, and prompt tuning is used to guide the prediction towards policy improvement. There is an additional contrast...
This paper improves the previous path integration based model interpretation works by overcoming the limit of absolute based attribute scores. The method is mainly consisting of a) building a better "reference" than the manual way previously; b) identifying the interpolation point in the non-linear path through the rel...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper improves the previous path integration based model interpretation works by overcoming the limit of absolute based attribute scores. The method is mainly consisting of a) building a better "reference" than the manual way previously; b) identifying the interpolation point in the non-linear path through...
This paper considers using autoregressive models to predict the different state dimensions when learning a transition model. The authors argue that autoregressive models are able to capture correlations between state dimensions and therefore work better than standard neural nets in the context of model-based offline RL...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper considers using autoregressive models to predict the different state dimensions when learning a transition model. The authors argue that autoregressive models are able to capture correlations between state dimensions and therefore work better than standard neural nets in the context of model-based of...
Summary Based on former findings, this paper first proposes a new definition of robust features with taking dynamics into consideration. Focused on the phenomenon that robust overfitting usually starts after learning rate decay, the authors provide an explanation of robust overfitting based on several experiments. New...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: Summary Based on former findings, this paper first proposes a new definition of robust features with taking dynamics into consideration. Focused on the phenomenon that robust overfitting usually starts after learning rate decay, the authors provide an explanation of robust overfitting based on several experime...
The paper studies the linear mode connectivity problem empirically and uncovers a new phenomenon called variance collapse (where activation variances of interpolated networks between two parent networks reduces significantly). Test accuracy barrier reduction is shown successfully for the interpolated networks (for CIFA...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper studies the linear mode connectivity problem empirically and uncovers a new phenomenon called variance collapse (where activation variances of interpolated networks between two parent networks reduces significantly). Test accuracy barrier reduction is shown successfully for the interpolated networks (...
This paper proposes a novel algorithm for offline RL that works by dividing the state-action space into discrete latent regions, learning a deterministic policy for each, and acting with a mixture of these policies. The authors demonstrate the effectiveness of their method on a series of experiments from the D4RL bench...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes a novel algorithm for offline RL that works by dividing the state-action space into discrete latent regions, learning a deterministic policy for each, and acting with a mixture of these policies. The authors demonstrate the effectiveness of their method on a series of experiments from the D4...
This paper presents a new approach for object depth estimation by using audio and video propagation times. Strengths I really like this paper. Although the idea of using time of arrival differences is simple, it hasn't been explored so thoroughly before. The execution is nice. One main advantage is that this method wor...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a new approach for object depth estimation by using audio and video propagation times. Strengths I really like this paper. Although the idea of using time of arrival differences is simple, it hasn't been explored so thoroughly before. The execution is nice. One main advantage is that this me...
The paper identifies that with Tensor Programming Optimisation in general the dataset collected is significantly imbalanced. Using this they develop an active learning approach which favours the selection of new samples as those which are hoped to be in areas with less prior knowledge. The authors show that this allows...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper identifies that with Tensor Programming Optimisation in general the dataset collected is significantly imbalanced. Using this they develop an active learning approach which favours the selection of new samples as those which are hoped to be in areas with less prior knowledge. The authors show that thi...
This paper proposes a method for deepfake detection against adversarial attacks. It proposes to use an ensemble of models whose inputs are subsets of frequency spectrum. It also theoretically proves that the ensemble of models reduce the dimension of the adversarial subspace, which could increase the robustness. Streng...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a method for deepfake detection against adversarial attacks. It proposes to use an ensemble of models whose inputs are subsets of frequency spectrum. It also theoretically proves that the ensemble of models reduce the dimension of the adversarial subspace, which could increase the robustness...
This paper proposes a measure of conditional independence called CIRCE that can be used as a regularizer in a large-scale mini-batch training. The key idea of CIRCE is that the conditional independence holds if and only if Eq. (4) holds, so we can measure the conditional dependence by how much the equation is violated....	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a measure of conditional independence called CIRCE that can be used as a regularizer in a large-scale mini-batch training. The key idea of CIRCE is that the conditional independence holds if and only if Eq. (4) holds, so we can measure the conditional dependence by how much the equation is v...
This paper proposes a meta-learning approach based on generation of new tasks from the columns of the training data. The authors's technique targets tabular data through modification of the task generation using randomization of the columns as well as an unsupervised validation technique. Through generation of such new...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a meta-learning approach based on generation of new tasks from the columns of the training data. The authors's technique targets tabular data through modification of the task generation using randomization of the columns as well as an unsupervised validation technique. Through generation of ...
This paper proposed a transformer-based architecture to discover diverse behaviors from the offline dataset. Strength: 1. This paper is well-writing. 2. The discovery of diverse behaviors is an interesting problem. Weakness: 1. The experiment results are not good enough. The proposed method is much worse than baseline...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposed a transformer-based architecture to discover diverse behaviors from the offline dataset. Strength: 1. This paper is well-writing. 2. The discovery of diverse behaviors is an interesting problem. Weakness: 1. The experiment results are not good enough. The proposed method is much worse than ...
The paper presents a way of performing knowledge distillation using contrastive learning. The authors point out that existing distillation methods that use contrastive learning have two main issues: (i) memory intensive - as it needs to store the representations of the whole dataset to construct negatives, (ii) the neg...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a way of performing knowledge distillation using contrastive learning. The authors point out that existing distillation methods that use contrastive learning have two main issues: (i) memory intensive - as it needs to store the representations of the whole dataset to construct negatives, (ii)...
The paper describes a proposal to enhance real-valued neural autoregressive density estimator (RNADE) by utilizing RNN-based mechanisms to be applicable over a concept-drifting data stream for density estimation and classification tasks. Using RNNs for estimating mean and variance, and constraining the number of parame...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper describes a proposal to enhance real-valued neural autoregressive density estimator (RNADE) by utilizing RNN-based mechanisms to be applicable over a concept-drifting data stream for density estimation and classification tasks. Using RNNs for estimating mean and variance, and constraining the number o...
This work presents a new probabilistic algorithm ("DiscoBAX") for subset selection that aims to approximately optimize phenotype movement in genomic intervention and can be useful in drug discovery tasks according to the authors. The method identifies a set of interventions whose elements will trigger maximum expected ...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This work presents a new probabilistic algorithm ("DiscoBAX") for subset selection that aims to approximately optimize phenotype movement in genomic intervention and can be useful in drug discovery tasks according to the authors. The method identifies a set of interventions whose elements will trigger maximum e...
This manuscript proposes a method, PA-LoFTE, for solving feature matching between images which generally follows a detection-description-matching three-stage pipeline. In order to achieve that, the authors utilize depth information from a depth predictor to generate 3D position embedding, then combine visual features, ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This manuscript proposes a method, PA-LoFTE, for solving feature matching between images which generally follows a detection-description-matching three-stage pipeline. In order to achieve that, the authors utilize depth information from a depth predictor to generate 3D position embedding, then combine visual fe...
In this paper, the authors transplant the idea of spiking neurons to MLP. Strength: + The authors ensure multiplication-free inference. + This work may be the first attempt for combining LIF and MLP. Weakness: - The motivation for this work is not convincing for me. In the abstract, the authors claimed that the ...	Recommendation: 3: reject, not good enough	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: In this paper, the authors transplant the idea of spiking neurons to MLP. Strength: + The authors ensure multiplication-free inference. + This work may be the first attempt for combining LIF and MLP. Weakness: - The motivation for this work is not convincing for me. In the abstract, the authors claimed t...
This paper proposed a new finding that between the image content and the additive degradation, deep networks tend to learn the less complex element in the separation task. To solve the generalization problem in image deraining, it tried to train the deraining networks with the fewer and less complex background images. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed a new finding that between the image content and the additive degradation, deep networks tend to learn the less complex element in the separation task. To solve the generalization problem in image deraining, it tried to train the deraining networks with the fewer and less complex background ...
In this paper, the authors try to address the multi-instance interactive segmentation task instead of salient interactive segmentation via the Label Propagation algorithm with patch feature representation from Transformers. In particular, the authors utilize keys from the last ViT-B layer for feature representation wh...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: In this paper, the authors try to address the multi-instance interactive segmentation task instead of salient interactive segmentation via the Label Propagation algorithm with patch feature representation from Transformers. In particular, the authors utilize keys from the last ViT-B layer for feature represent...
This paper studies offline multi-objective reinforcement learning (offline MORL). In the first half of the paper, the authors introduce a new benchmarking dataset. For 6 tasks it contains trajectories with different preferences, where for each preference we have data from an expert policy (with similar preference profi...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies offline multi-objective reinforcement learning (offline MORL). In the first half of the paper, the authors introduce a new benchmarking dataset. For 6 tasks it contains trajectories with different preferences, where for each preference we have data from an expert policy (with similar preferen...
Sparse Mixture-of-Experts (SMoE) have been widely used to increase the representation capacity of transformers with similar computational cost as dense transformers. SMoE consists of many experts, only few of which are activate for any given input example. This increased representational power comes with two issues: (...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: Sparse Mixture-of-Experts (SMoE) have been widely used to increase the representation capacity of transformers with similar computational cost as dense transformers. SMoE consists of many experts, only few of which are activate for any given input example. This increased representational power comes with two is...
This paper studies a new performance measuring method that measures the performance of a collection of models when evaluated on a single input point. The authors compare the models' average performance on the test distribution and their pointwise performance on an individual point. The empirical results show that for s...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper studies a new performance measuring method that measures the performance of a collection of models when evaluated on a single input point. The authors compare the models' average performance on the test distribution and their pointwise performance on an individual point. The empirical results show th...
The paper introduces a state abstraction method for neural episodic control. States or state-action pairs are mapped to grid coordinates and dimensionality is reduced further by random projections. The method is suitable for state spaces with a limited number of independent dimensions. The method provides better result...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces a state abstraction method for neural episodic control. States or state-action pairs are mapped to grid coordinates and dimensionality is reduced further by random projections. The method is suitable for state spaces with a limited number of independent dimensions. The method provides bette...
This paper uses the transformer to leverage a sizeable online resume dataset by pretraining and then fine-tuning it on the small and carefully constructed longitudinal survey datasets. According to the results based on their experiments, their approach shows a significant improvement compared to the current state of th...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper uses the transformer to leverage a sizeable online resume dataset by pretraining and then fine-tuning it on the small and carefully constructed longitudinal survey datasets. According to the results based on their experiments, their approach shows a significant improvement compared to the current sta...
The paper considers the case of population-based approaches for exploration in reinforcement learning, in particular the Quality-Diversity approach, an evolutionary algorithm to create collections of high performing solutions. The authors propose two algorithm variations of QD which take into consideration the learned ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper considers the case of population-based approaches for exploration in reinforcement learning, in particular the Quality-Diversity approach, an evolutionary algorithm to create collections of high performing solutions. The authors propose two algorithm variations of QD which take into consideration the ...
This paper introduces a parameterized similarity measure that shows significant improvement over existing closed-form similarity measures. Strength: - The paper is well-written and easy to follow. - The method is simple yet effective. - The empirical analysis is comprehensive. - The provided benchmark can be...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a parameterized similarity measure that shows significant improvement over existing closed-form similarity measures. Strength: - The paper is well-written and easy to follow. - The method is simple yet effective. - The empirical analysis is comprehensive. - The provided benchmar...
The authors propose CO3, Cooperative Contrastive Learning, and Contextual Shape Prediction for unsupervised 3D representation learning in outdoor scenes. By using DAIR-V2X, a vehicle-infrastructure-cooperation dataset, the data mitigates challenges due to the moving objects that would make it hard to find correct corre...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The authors propose CO3, Cooperative Contrastive Learning, and Contextual Shape Prediction for unsupervised 3D representation learning in outdoor scenes. By using DAIR-V2X, a vehicle-infrastructure-cooperation dataset, the data mitigates challenges due to the moving objects that would make it hard to find corre...
This paper proposes an adversarial training framework for resource-constrained federated learning. The proposed framework decouples the entire model into small modules to fit into the edge device memory. Adversarial training is only performed on a single module in each communication round. Strengths: 1. The paper aims ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an adversarial training framework for resource-constrained federated learning. The proposed framework decouples the entire model into small modules to fit into the edge device memory. Adversarial training is only performed on a single module in each communication round. Strengths: 1. The pap...
This paper designs a circuit GNN (CktGNN) to facilitate the learning of circuit graphs and electronic design automation. CktGNN is a generation of the nested GNN structure to DAGs. The paper also presents a circuit benchmark dataset OCB with open-source code for automated circuit design methods evaluation. Strength: 1....	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper designs a circuit GNN (CktGNN) to facilitate the learning of circuit graphs and electronic design automation. CktGNN is a generation of the nested GNN structure to DAGs. The paper also presents a circuit benchmark dataset OCB with open-source code for automated circuit design methods evaluation. Stre...