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 studies a method which combines MPC with DP that bypasses utility challenges of federated learning with local DP (with small datasets). authors show that the method achieves good performance by reporting their submission entry being 1st place for an iDASH 2021 challenge, at the cost of substantially more comp...	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 a method which combines MPC with DP that bypasses utility challenges of federated learning with local DP (with small datasets). authors show that the method achieves good performance by reporting their submission entry being 1st place for an iDASH 2021 challenge, at the cost of substantially m...
The paper presents a new benchmark for evaluating the robustness of models for motion forecasting of autonomous vehicles. The authors collect human-annotated labels of causal agents on the Waymo Open Motion Dataset, and perform various perturbations of the data using these labels. This data is used to evaluate the robu...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a new benchmark for evaluating the robustness of models for motion forecasting of autonomous vehicles. The authors collect human-annotated labels of causal agents on the Waymo Open Motion Dataset, and perform various perturbations of the data using these labels. This data is used to evaluate ...
This paper introduces a method for blind single image super-resolution. Contrary to the mainstream the idea of this method is to design an implicit degradation estimator that can extract information regarding the degradation without having an explicit model. This is done using ideas from Knowledge distillation where fi...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a method for blind single image super-resolution. Contrary to the mainstream the idea of this method is to design an implicit degradation estimator that can extract information regarding the degradation without having an explicit model. This is done using ideas from Knowledge distillation ...
This paper improves the efficiency of MAE training by exploiting the redundancy in reconstructed tokens. They propose to use the similarity to mean token as an emprical criterior to prune the tokens, which leads to savings in computation. They achieve 1.5x-1.9x speed up compared with MAE, while maintain similar perform...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper improves the efficiency of MAE training by exploiting the redundancy in reconstructed tokens. They propose to use the similarity to mean token as an emprical criterior to prune the tokens, which leads to savings in computation. They achieve 1.5x-1.9x speed up compared with MAE, while maintain similar...
Authors present a generalized TD learning for the MARL setting, with local updates occurring for K > 1 iterations for each instance of communication. Sample complexities at the inner loop (local TD updates) and the outer loop (consensus updates) are computed based on mild assumptions. Theoretical formulation predicts a...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Authors present a generalized TD learning for the MARL setting, with local updates occurring for K > 1 iterations for each instance of communication. Sample complexities at the inner loop (local TD updates) and the outer loop (consensus updates) are computed based on mild assumptions. Theoretical formulation pr...
The authors propose a novel multi-agent policy gradient algorithm called Advantage Constrained Proximal Policy Optimization (ACPPO). Strength: 1. The author presents a constraint coefficient to the local advantage, which is estimated by the difference between the local and fictitious joint advantage functions, to ensur...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors propose a novel multi-agent policy gradient algorithm called Advantage Constrained Proximal Policy Optimization (ACPPO). Strength: 1. The author presents a constraint coefficient to the local advantage, which is estimated by the difference between the local and fictitious joint advantage functions, ...
In this paper, the authors studied why training with the mixup loss for extra epochs often degrades test classification accuracy. They argued, through empirical observation, theoretical analysis and synthetic experiments, that the mixup loss introduces label noise, and a model may overfit to the label noise during exte...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors studied why training with the mixup loss for extra epochs often degrades test classification accuracy. They argued, through empirical observation, theoretical analysis and synthetic experiments, that the mixup loss introduces label noise, and a model may overfit to the label noise dur...
This paper presents a benchmark for partial domain adaptation (PDA) to evaluate the model selection strategies for different PDA approaches on two different real-world datasets. The experimental results show that 1) Target labels are critical for model selection strategies; 2) only one method and model selection pair p...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper presents a benchmark for partial domain adaptation (PDA) to evaluate the model selection strategies for different PDA approaches on two different real-world datasets. The experimental results show that 1) Target labels are critical for model selection strategies; 2) only one method and model selectio...
This paper studies a score-based generative model (SGM) for tabular data. They found there's a difficulty in adopting SGM to tabular modeling that the training process is unstable. Hence they introduced a self-paced learning algorithm, a variation of curriculum learning, to train SGM. In experiments, the proposed metho...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper studies a score-based generative model (SGM) for tabular data. They found there's a difficulty in adopting SGM to tabular modeling that the training process is unstable. Hence they introduced a self-paced learning algorithm, a variation of curriculum learning, to train SGM. In experiments, the propos...
This paper proposes a method to perform activation compressed training (ACT) for pointwise non-linear activation functions. Instead of storing a quantized version of the input x, the proposed approach stores a quantized version of the activation function's gradient f'(x), which is multiplied with the gradient in the ba...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a method to perform activation compressed training (ACT) for pointwise non-linear activation functions. Instead of storing a quantized version of the input x, the proposed approach stores a quantized version of the activation function's gradient f'(x), which is multiplied with the gradient i...
The paper reveals an interesting and important failure pattern of the pre-trained vision-language models (VLMs): they are insensitive to object attributes, relations, and even word orders. They created several tests from Visual Genome, COCO caption, and Flickr30K, and show several representative VLMs (BLIP, CLIP, etc.)...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper reveals an interesting and important failure pattern of the pre-trained vision-language models (VLMs): they are insensitive to object attributes, relations, and even word orders. They created several tests from Visual Genome, COCO caption, and Flickr30K, and show several representative VLMs (BLIP, CLI...
The paper mainly investigates the principle differences between contrastive learning (CL) and masked image modeling (MIM). That’s the CL is shape-biased, which learns low-frequencies, and the MIM is texture-biased, which exploits high-frequencies. The Conclusion is interesting and the analysis is thorough. Strength: ...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The paper mainly investigates the principle differences between contrastive learning (CL) and masked image modeling (MIM). That’s the CL is shape-biased, which learns low-frequencies, and the MIM is texture-biased, which exploits high-frequencies. The Conclusion is interesting and the analysis is thorough. Stre...
This paper study the problem of multiple objective reinforcement learning. As previous methods that search for Pareto front are not sample efficient. The paper propose the Q-Pensieve method, which updates the policy with learned Q-networks from past iterations. The paper provide some theoretical analysis. Finally, exp...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper study the problem of multiple objective reinforcement learning. As previous methods that search for Pareto front are not sample efficient. The paper propose the Q-Pensieve method, which updates the policy with learned Q-networks from past iterations. The paper provide some theoretical analysis. Fina...
The paper investigates the capability of transformer-based in-context learners (ICL) to implicitly implement learning algorithms. The authors consider a study-case of linear regression models to validate their hypothesis that ICL encode context-specific parametric models: 1) ICL implement linear models by encoding grad...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper investigates the capability of transformer-based in-context learners (ICL) to implicitly implement learning algorithms. The authors consider a study-case of linear regression models to validate their hypothesis that ICL encode context-specific parametric models: 1) ICL implement linear models by encod...
The paper proposed dense correlation fields for action recognition. Based on CorrNet, the paper proposed to model temporal long-term correlation by computing correlation for frames that are far apart. The paper also designs a spatial hierarchical architecture to aggregate the information at different granularities. The...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposed dense correlation fields for action recognition. Based on CorrNet, the paper proposed to model temporal long-term correlation by computing correlation for frames that are far apart. The paper also designs a spatial hierarchical architecture to aggregate the information at different granularit...
This paper proposes a generative modeling setup for weak supervision in images. The claim is that by doing so, there are improvements both to generative modeling and to the pseudolabeling qualities of the model, and they work in concert. They do this through proofs and experimental evaluations. The setup is a GAN arc...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a generative modeling setup for weak supervision in images. The claim is that by doing so, there are improvements both to generative modeling and to the pseudolabeling qualities of the model, and they work in concert. They do this through proofs and experimental evaluations. The setup is a...
This paper is an empirical study of the generalization abilities of the deep RL algorithm IMPALA pre-trained on multiple tasks. Pre-training is done on variants of a single game from the Atari benchmark and fine-tuning/testing is done on held-out variants. The authors show that by limiting to variants of a single game,...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper is an empirical study of the generalization abilities of the deep RL algorithm IMPALA pre-trained on multiple tasks. Pre-training is done on variants of a single game from the Atari benchmark and fine-tuning/testing is done on held-out variants. The authors show that by limiting to variants of a sing...
This paper proposes a novel specialized attention mechanism for CLIP models to induce hierarchical feature discovery in vision and language each modality. The authors successfully validate their claims with various experimental results, and they demonstrate the significant performance improvement of CLIP (Contrastive ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a novel specialized attention mechanism for CLIP models to induce hierarchical feature discovery in vision and language each modality. The authors successfully validate their claims with various experimental results, and they demonstrate the significant performance improvement of CLIP (Cont...
This work focuses on improving the use of knowledge in knowledge-augmented Pre-Trained Language Models (PTLM). Authors hypothesize that using knowledge for all instances can backfire and lead the model towards incorrect predictions, but simultaneously for some instances, external knowledge is required. To combat this i...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work focuses on improving the use of knowledge in knowledge-augmented Pre-Trained Language Models (PTLM). Authors hypothesize that using knowledge for all instances can backfire and lead the model towards incorrect predictions, but simultaneously for some instances, external knowledge is required. To comba...
The authors proposed SMART, a new automatic evaluation metric text generation models. - Proposal 1: To deal with multiple sentences, they proposed SMART-N and SMART-L, in which the calculation unit of ROUGE-N and ROUGE-L is changed from words to sentences. In SMART, several metrics such as BLEURT and ANLI are used to s...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors proposed SMART, a new automatic evaluation metric text generation models. - Proposal 1: To deal with multiple sentences, they proposed SMART-N and SMART-L, in which the calculation unit of ROUGE-N and ROUGE-L is changed from words to sentences. In SMART, several metrics such as BLEURT and ANLI are u...
Mode collapse is a problem in imitating multi-modal trajectories. Hierarchical method use latent code to represent the agent type or context variable which is inferred during the training and later used at test time, and the policy should be conditioned on the latent code. Context variable include internal (intrinsic s...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: Mode collapse is a problem in imitating multi-modal trajectories. Hierarchical method use latent code to represent the agent type or context variable which is inferred during the training and later used at test time, and the policy should be conditioned on the latent code. Context variable include internal (int...
This paper proposes an out-of-distribution (OOD) detection method for text generation based on the concept of information measure. More specifically, the author considers two measurements, the Renyi divergence, and the Fisher-Rao distance under both the no-reference and the reference scenarios. The empirical results sh...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes an out-of-distribution (OOD) detection method for text generation based on the concept of information measure. More specifically, the author considers two measurements, the Renyi divergence, and the Fisher-Rao distance under both the no-reference and the reference scenarios. The empirical re...
The paper proposes a novel benchmark, called CAB, to compare the efficiency of different attention mechanisms in transformers. The current widely used benchmark LRA focusses only on noncausal self-attention. CAB seeks to evaluate four types of attention mechanisms: the combinations of causal/ non-causal and self/ cros...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper proposes a novel benchmark, called CAB, to compare the efficiency of different attention mechanisms in transformers. The current widely used benchmark LRA focusses only on noncausal self-attention. CAB seeks to evaluate four types of attention mechanisms: the combinations of causal/ non-causal and se...
This paper investigates systematic generalization of multi-label classifications in settings where the input space is shared between the different labels. The authors suggest that deep learning models are biased towards feature reuse, which conflicts with systematic generalization to new combinations of known classes. ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper investigates systematic generalization of multi-label classifications in settings where the input space is shared between the different labels. The authors suggest that deep learning models are biased towards feature reuse, which conflicts with systematic generalization to new combinations of known c...
This paper proposes to use parametric transforms such as exponential function or stick-breaking transformation on the outputs of Bayesian neural network (BNN) and turn them into positive random vectors or simplex vectors. This new transformed BNN module can be plugged into probabilistic graphic models as a replacement ...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: This paper proposes to use parametric transforms such as exponential function or stick-breaking transformation on the outputs of Bayesian neural network (BNN) and turn them into positive random vectors or simplex vectors. This new transformed BNN module can be plugged into probabilistic graphic models as a repl...
This paper proposes a new topic-aware transformer that alleviates the gap between the pretrained language model and the target domain. The paper introduces an additional topic latent variable $z$ as an additional condition when generating new tokens. The new topic latent variables come from the topic steering layer, wh...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper proposes a new topic-aware transformer that alleviates the gap between the pretrained language model and the target domain. The paper introduces an additional topic latent variable $z$ as an additional condition when generating new tokens. The new topic latent variables come from the topic steering l...
This paper studies reducing selection bias and confounders bias in estimating average causal effects (ATE) in observational data. This paper provides the SC-CFR algorithm for simultaneously addressing confounder and collider bias to achieve unbiased ATE estimation. Specifically, SC-CFR first computes the magnitude of t...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper studies reducing selection bias and confounders bias in estimating average causal effects (ATE) in observational data. This paper provides the SC-CFR algorithm for simultaneously addressing confounder and collider bias to achieve unbiased ATE estimation. Specifically, SC-CFR first computes the magnit...
The paper proposes a data valuation method that is oblivious to the downstream learning algorithm. The main idea is to evaluate the training data by a class-wise Wasserstein distance between the training and the validation set. They prove that the class-wise Wasserstein distance approximates the performance for any giv...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The paper proposes a data valuation method that is oblivious to the downstream learning algorithm. The main idea is to evaluate the training data by a class-wise Wasserstein distance between the training and the validation set. They prove that the class-wise Wasserstein distance approximates the performance for...
The paper presents a semi-supervised learning method for semantic segmentation, which learns a multi-prototype representation for each semantic class and a feature refinement step based on those prototypes. To achieve this, it develops a multi-prototype contrastive loss function and an attention-based prototype-to-feat...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper presents a semi-supervised learning method for semantic segmentation, which learns a multi-prototype representation for each semantic class and a feature refinement step based on those prototypes. To achieve this, it develops a multi-prototype contrastive loss function and an attention-based prototype...
The paper develops techniques for distributed matrix multiplication with adversarially corrupt nodes and stragglers. A error correction coding scheme is developed that simultaneously detects corrupt computations via random-key-based verification, and can decode in the presence of stragglers. Strengths: + The simultan...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper develops techniques for distributed matrix multiplication with adversarially corrupt nodes and stragglers. A error correction coding scheme is developed that simultaneously detects corrupt computations via random-key-based verification, and can decode in the presence of stragglers. Strengths: + The ...
The paper aims to make CNN suitable for data of arbitrary resolution, dimensionality and length without any structural changes. A Continuous CNN is proposed by introducing continuous convolutional kernels which is a data independent parameterization for convolutional weight. The proposed CCNN can work on on sequential ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper aims to make CNN suitable for data of arbitrary resolution, dimensionality and length without any structural changes. A Continuous CNN is proposed by introducing continuous convolutional kernels which is a data independent parameterization for convolutional weight. The proposed CCNN can work on on seq...
This paper considers a new type of adversary that chooses to only perturb a subset of data. The authors propose a defensive strategy that is formulated as a minimax problem that involves a worst-case subset selection procedure. This paper uses a greedy algorithm to realize the worst-case subset selection. The empirical...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper considers a new type of adversary that chooses to only perturb a subset of data. The authors propose a defensive strategy that is formulated as a minimax problem that involves a worst-case subset selection procedure. This paper uses a greedy algorithm to realize the worst-case subset selection. The e...
The paper proposes an application of SoftHebb to deep convolutional neural networks. SoftHebb has been proposed before (supposedly by the same authors?), it is a soft winner-take-all version of the Hebbian learning rule implemented using the softmax function. The novelty of the paper lies in making SoftHebb scale to mu...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: The paper proposes an application of SoftHebb to deep convolutional neural networks. SoftHebb has been proposed before (supposedly by the same authors?), it is a soft winner-take-all version of the Hebbian learning rule implemented using the softmax function. The novelty of the paper lies in making SoftHebb sca...
The authors evaluate the relation between episodic and global novelty bonus with regard to contextual MDPs. They show two simplified examples for demonstrating the different advantages of the two novelty bonus types. Additionally, they explore 6 different combinations of the episodic E3B bonus with global novelty bonus...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The authors evaluate the relation between episodic and global novelty bonus with regard to contextual MDPs. They show two simplified examples for demonstrating the different advantages of the two novelty bonus types. Additionally, they explore 6 different combinations of the episodic E3B bonus with global novel...
The paper proposes a new method based on the CLIP model for continual vision-language representation learning. The authors find that simple continual training in the CLIP model degenerates the performance on multimodal retrieval tasks, named Cognitive Disorder (CD). To tackle CD, they suggest Mod-X that selectively ali...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes a new method based on the CLIP model for continual vision-language representation learning. The authors find that simple continual training in the CLIP model degenerates the performance on multimodal retrieval tasks, named Cognitive Disorder (CD). To tackle CD, they suggest Mod-X that selecti...
This paper proposes to combine meta learning with few-shot adaptation of language models, by essentially building upon methods such as Frozen (Tsimpoukelli et al., 2021) which maps image representations into language model token space, but adding a meta-mapper network, acting as a meta-learner, to more efficiently perf...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes to combine meta learning with few-shot adaptation of language models, by essentially building upon methods such as Frozen (Tsimpoukelli et al., 2021) which maps image representations into language model token space, but adding a meta-mapper network, acting as a meta-learner, to more efficien...
The authors present a novel application of differential geometry and Hamiltonian flows to graph neural networks and demonstrate its applicability to the common GNN problem of node classification. The theory gives rise to a GNN model which is then shown to perform relatively well on standard datasets and somewhat to res...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The authors present a novel application of differential geometry and Hamiltonian flows to graph neural networks and demonstrate its applicability to the common GNN problem of node classification. The theory gives rise to a GNN model which is then shown to perform relatively well on standard datasets and somewha...
This paper proposes, based on SIREN, a new interpretable INR architecture for time series called iSIREN. Building up on that, the paper proposes to use the iSIREN set-up in the context of hypernetworks for time series generating. Both methods differ from its plain counter version (SIREN and SIREN-based hypernetwork) in...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes, based on SIREN, a new interpretable INR architecture for time series called iSIREN. Building up on that, the paper proposes to use the iSIREN set-up in the context of hypernetworks for time series generating. Both methods differ from its plain counter version (SIREN and SIREN-based hypernet...
This paper proposes a new hierarchy of expressive isomorphism test algorithm and GNN model, with combining k-WL tuple idea and rooted subgraph idea from subgraph GNNs. More specifically, the author still work on subgraph GNNs with each node representing the encoding of d-hop rooted subgraph around the node. Unliking 1-...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new hierarchy of expressive isomorphism test algorithm and GNN model, with combining k-WL tuple idea and rooted subgraph idea from subgraph GNNs. More specifically, the author still work on subgraph GNNs with each node representing the encoding of d-hop rooted subgraph around the node. Unl...
This paper proposes a metrics called "susceptibility" that measure the model's resistance to memorization by using randomly labeled data. The authors provide both theoretical and empirical observations that the models which are resistant to memorization will have high test accuracies. Therefore, the extent of memorizat...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a metrics called "susceptibility" that measure the model's resistance to memorization by using randomly labeled data. The authors provide both theoretical and empirical observations that the models which are resistant to memorization will have high test accuracies. Therefore, the extent of m...