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 presents a method where out-of-domain (OOD) data are used to learning better representations via self-supervised learning (SSL) in a setting where we have unlabeled in-domain (ID) data that follow a long-tail distribution. The paper proposed to define an unsupervised "tailness" score, while an online sampling...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: The paper presents a method where out-of-domain (OOD) data are used to learning better representations via self-supervised learning (SSL) in a setting where we have unlabeled in-domain (ID) data that follow a long-tail distribution. The paper proposed to define an unsupervised "tailness" score, while an online ...
The paper investigates neural network representations under multi-task learning based on five binary classifications tasks constructed from MNIST. The authors compare using independent, individual multi-layer perceptrons (MLPs) per task with parallel MLPs that process tasks simultaneously and with task-switching MLPs ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper investigates neural network representations under multi-task learning based on five binary classifications tasks constructed from MNIST. The authors compare using independent, individual multi-layer perceptrons (MLPs) per task with parallel MLPs that process tasks simultaneously and with task-switchi...
The paper presents algorithms for computing approximately rationalizable correlated equilibria (CE) and coarse correlated equilibria (CCE) in general games. Although these bounds are not shown to be optimal (except of the case where strategies are rationalizable after a constant number of iterated elimination rounds) t...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper presents algorithms for computing approximately rationalizable correlated equilibria (CE) and coarse correlated equilibria (CCE) in general games. Although these bounds are not shown to be optimal (except of the case where strategies are rationalizable after a constant number of iterated elimination r...
This paper proposes a new initialization method for ConvNets with residual blocks and ReLU activation, called RISOTTO. It enables training deep networks with residual structures without a normalization layer. This is achieved by maintaining a dynamical isometry at the initialization time, which is implemented as initia...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new initialization method for ConvNets with residual blocks and ReLU activation, called RISOTTO. It enables training deep networks with residual structures without a normalization layer. This is achieved by maintaining a dynamical isometry at the initialization time, which is implemented a...
This paper proposes a new method for the problem of distributionally robust optimization. The learning procedure ‘Bitrate-Constrained DRO’ (BR-DRO) provides robustness to distribution shifts along groups realised by simple functions. This procedure is able to match the performance of methods that use group information ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a new method for the problem of distributionally robust optimization. The learning procedure ‘Bitrate-Constrained DRO’ (BR-DRO) provides robustness to distribution shifts along groups realised by simple functions. This procedure is able to match the performance of methods that use group info...
This paper introduces a unified information extraction (UIE) framework based on the token pair classification idea. The UIE solution allows modeling all types of IE tasks such as NER, RE and EE with just one unified model. Also the authors propose a Plusformer on top of the token-pair feature matrix for better feature ...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces a unified information extraction (UIE) framework based on the token pair classification idea. The UIE solution allows modeling all types of IE tasks such as NER, RE and EE with just one unified model. Also the authors propose a Plusformer on top of the token-pair feature matrix for better ...
This paper proposes a method for self-supervised learning based on adversarial sample generation. The proposed method is based on the probabilistic interpretation of PLS and/or CCA based on graphical models and does not require class labels or classifiers. It performs best among the domain-agnostic methods on tabular, ...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a method for self-supervised learning based on adversarial sample generation. The proposed method is based on the probabilistic interpretation of PLS and/or CCA based on graphical models and does not require class labels or classifiers. It performs best among the domain-agnostic methods on t...
The paper addresses an interesting problem for RL i.e., sample efficiency. The author proposes a Jump-Start RL (JSRL) algorithm to leverage a prior policy of any form to give a head start for exploration in RL. The proposed algorithm rolls out a pre-existing guided policy, followed by self-improving exploration policy....	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper addresses an interesting problem for RL i.e., sample efficiency. The author proposes a Jump-Start RL (JSRL) algorithm to leverage a prior policy of any form to give a head start for exploration in RL. The proposed algorithm rolls out a pre-existing guided policy, followed by self-improving exploration...
This manuscript investigates the approximation error of classic deep belief networks (DBNs), in particular DBNs with two hidden layers of size m and m+1, respectively. It is demonstrated that, under both L^q-norm and Kullback-Leibler divergence, DBNs are universal approximators. Moreover, as claimed by the author, sh...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This manuscript investigates the approximation error of classic deep belief networks (DBNs), in particular DBNs with two hidden layers of size m and m+1, respectively. It is demonstrated that, under both L^q-norm and Kullback-Leibler divergence, DBNs are universal approximators. Moreover, as claimed by the au...
MULTI-LAYERED 3D GARMENTS ANIMATION This paper presents a novel deep learning-based 3D garments animation method. The method learns to model interactions between 1) cloth-body, 2) multiple cloth layers 3) wind and clothes. Specifically, the algorithm decomposes the clothes into connected patches and evaluates the defo...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: MULTI-LAYERED 3D GARMENTS ANIMATION This paper presents a novel deep learning-based 3D garments animation method. The method learns to model interactions between 1) cloth-body, 2) multiple cloth layers 3) wind and clothes. Specifically, the algorithm decomposes the clothes into connected patches and evaluates ...
This paper develops an interpretable anomaly detection framework, DIAD, by combining the NodeGAM structure with the PID objective. The NodeGAM provides an interpretable mechanism, and PID can be trained for anomaly detection. Overall, the framework makes sense to me. The paper includes detailed experimental evaluation...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper develops an interpretable anomaly detection framework, DIAD, by combining the NodeGAM structure with the PID objective. The NodeGAM provides an interpretable mechanism, and PID can be trained for anomaly detection. Overall, the framework makes sense to me. The paper includes detailed experimental ev...
This work focuses on a new problem: collaboration pure exploration with multiple tasks and general reward structures. It explores both fixed-confidence and fixed-budget settings. Nearly-matching upper and lower bounds on sampling and communication complexity are derived. Strength: 1. The paper is easy to follow and pro...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work focuses on a new problem: collaboration pure exploration with multiple tasks and general reward structures. It explores both fixed-confidence and fixed-budget settings. Nearly-matching upper and lower bounds on sampling and communication complexity are derived. Strength: 1. The paper is easy to follow...
The paper investigates robust overfitting, an important phenomenon that occurs in the adversarial training. In contrast to previous approach, the paper does not regard the neural network as a black-box model but divide a DNN into a series of layers and investigate the effect of different network layers on robust overfi...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper investigates robust overfitting, an important phenomenon that occurs in the adversarial training. In contrast to previous approach, the paper does not regard the neural network as a black-box model but divide a DNN into a series of layers and investigate the effect of different network layers on robus...
This paper presents an extension of Diffusion Probabilistic Models using Fields to model data. The basic idea is to model data as functions (fields) e.g. images are functions from $R^2$ to $R^3$, a map between pixel coordinates and color pixel value. This approach is motivated by the need to unify the generative model ...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper presents an extension of Diffusion Probabilistic Models using Fields to model data. The basic idea is to model data as functions (fields) e.g. images are functions from $R^2$ to $R^3$, a map between pixel coordinates and color pixel value. This approach is motivated by the need to unify the generativ...
In this work the authors propose an optimal transport framework algorithm called Entire Space Counter-Factual Regression (ESCFR). The method is interesting in itself, with the additions the authors propose (specially the PFOR) add some light on how neural representations could be better used for counterfactual estimati...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this work the authors propose an optimal transport framework algorithm called Entire Space Counter-Factual Regression (ESCFR). The method is interesting in itself, with the additions the authors propose (specially the PFOR) add some light on how neural representations could be better used for counterfactual ...
The paper describes a music generation approach which aims to extend the type of conditioning variables typically used in this area of research. The authors describe this as so-called description-to-sequence learning. Inspired by recent progress in text-conditioned image generation, the authors propose to derive descri...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper describes a music generation approach which aims to extend the type of conditioning variables typically used in this area of research. The authors describe this as so-called description-to-sequence learning. Inspired by recent progress in text-conditioned image generation, the authors propose to deriv...
The paper analyzes the gradient flow learning of a particular mean-field (MF) network with more than 2 layers and shows that it can learn a radial target function (and a specific data distribution) with width polynomial in the data dimension $d$. This target function was previously shown to be inapproximable by 2-layer...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper analyzes the gradient flow learning of a particular mean-field (MF) network with more than 2 layers and shows that it can learn a radial target function (and a specific data distribution) with width polynomial in the data dimension $d$. This target function was previously shown to be inapproximable by...
The paper studies the effect of overparameterization under the existence of minority groups. The paper provides theoretical characterizations based on existing work, and ran experiments on the fixed point equations to show that the overparamterization does reflect recent empirical findings in this high dimensional setu...	Recommendation: 3: reject, not good enough	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies the effect of overparameterization under the existence of minority groups. The paper provides theoretical characterizations based on existing work, and ran experiments on the fixed point equations to show that the overparamterization does reflect recent empirical findings in this high dimensio...
This paper proposes a new method for out-of-distribution (OOD) generalization on graphs. The authors claim that the existing works mainly focus on the OOD issue of correlation shift, while another type, covariate shift, remains largely unexplored. Then, a graph augmentation strategy called AdvCA is proposed to handle t...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes a new method for out-of-distribution (OOD) generalization on graphs. The authors claim that the existing works mainly focus on the OOD issue of correlation shift, while another type, covariate shift, remains largely unexplored. Then, a graph augmentation strategy called AdvCA is proposed to ...
The paper focuses on learning graph neural networks (GNN) under the federated (FL) scheme. As for FL under vertically distributed data, "each client holds a subgraph of the global graph, part of the features for nodes in this subgraph, and part of the whole model; all clients collaboratively predict node properties." T...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper focuses on learning graph neural networks (GNN) under the federated (FL) scheme. As for FL under vertically distributed data, "each client holds a subgraph of the global graph, part of the features for nodes in this subgraph, and part of the whole model; all clients collaboratively predict node proper...
This manuscript studies the instance-level label restoration in mini-batch training under the federated learning structure. The authors proposed a strong method that is capable of restoring labels via gradient inversion attack even when batch size is as large as 4096, which was usually considered a challenging task. Th...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This manuscript studies the instance-level label restoration in mini-batch training under the federated learning structure. The authors proposed a strong method that is capable of restoring labels via gradient inversion attack even when batch size is as large as 4096, which was usually considered a challenging ...
The paper presents a method for learning object-centric representations from single 2D images without supervision. Unlike previous work that learns by reconstructing the current frame, the proposed method learns by predicting object movement in future frames. Specifically, the model is trained from triplets of 2D image...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper presents a method for learning object-centric representations from single 2D images without supervision. Unlike previous work that learns by reconstructing the current frame, the proposed method learns by predicting object movement in future frames. Specifically, the model is trained from triplets of ...
This paper considers the adversarial linear adversarial MDP and proposes the first efficient algorithm that achieves $\sqrt{K}$ bound. Moreover, they also give a lower bound. Their techniques are based on the occupancy measurements which is initially proposed by Jin et al. (2020b). Compared to previous techniques, this...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper considers the adversarial linear adversarial MDP and proposes the first efficient algorithm that achieves $\sqrt{K}$ bound. Moreover, they also give a lower bound. Their techniques are based on the occupancy measurements which is initially proposed by Jin et al. (2020b). Compared to previous techniqu...
This work proposes a 3D-aware generative model to disentangle 3D shapes, camera poses, object appearance, and background appearance when synthesizing high-quality images, which could realize non-rigid shape variation in an object category exclusively from 2D image supervision. To achieve a clear disentanglement, it pr...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This work proposes a 3D-aware generative model to disentangle 3D shapes, camera poses, object appearance, and background appearance when synthesizing high-quality images, which could realize non-rigid shape variation in an object category exclusively from 2D image supervision. To achieve a clear disentanglemen...
This paper studies the learning dynamics of deep neural networks using the quadratic models. Specifically, the "catapult phase" is identified under this model when the learning rate is high, while it is not for linear models. Numerical simulations are provided to support the theoretical analysis. Strenghths: 1. ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies the learning dynamics of deep neural networks using the quadratic models. Specifically, the "catapult phase" is identified under this model when the learning rate is high, while it is not for linear models. Numerical simulations are provided to support the theoretical analysis. Strenght...
This paper proposes a way to make an existing dataset (MBPP), an execution-based code completion benchmark, multilingual: while the original MBPP only contained python code, the new dataset (MBXP) contains a total of 10 programming languages. Problems in function completion datasets consists of prompt, *test statemen...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper proposes a way to make an existing dataset (MBPP), an execution-based code completion benchmark, multilingual: while the original MBPP only contained python code, the new dataset (MBXP) contains a total of 10 programming languages. Problems in function completion datasets consists of prompt, *test ...
This paper introduces the method of learning interpretable trees for representing reward functions by following a previously-proposed reward tree model. The paper proposes several updates to the reward tree inference algorithm, including an NLL-based objective for estimating trajectory-level returns, a new criterion fa...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper introduces the method of learning interpretable trees for representing reward functions by following a previously-proposed reward tree model. The paper proposes several updates to the reward tree inference algorithm, including an NLL-based objective for estimating trajectory-level returns, a new crit...
This paper proposes to leverage the injetced prompt token to capture the coarse class feature to benefit the visual recognition. With the additive corase class labels the proposed approach can outperform the baseline with less than 2% additional parameters. Strength: 1. The idea of using additive tokens to extract coa...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes to leverage the injetced prompt token to capture the coarse class feature to benefit the visual recognition. With the additive corase class labels the proposed approach can outperform the baseline with less than 2% additional parameters. Strength: 1. The idea of using additive tokens to ext...
The paper introduces a risk-averse mean-variance equilibrium to multi-agent games, and the authors also showed the existence of such a risk-averse equilibrium. A fictitious play type of learning algorithm is proposed to solve for the equilibrium and a corresponding RL-based approximation algorithm enjoys good empirical...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces a risk-averse mean-variance equilibrium to multi-agent games, and the authors also showed the existence of such a risk-averse equilibrium. A fictitious play type of learning algorithm is proposed to solve for the equilibrium and a corresponding RL-based approximation algorithm enjoys good e...
The paper proposed a method for Blind Face Restoration (BFR) using pre-trained diffusion models for faces. The main novelty is a design of a transition model from a low quality image that estimates an intermediate state of the diffusion process of a high quality image. This allows the method to use a pre-trained diffus...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposed a method for Blind Face Restoration (BFR) using pre-trained diffusion models for faces. The main novelty is a design of a transition model from a low quality image that estimates an intermediate state of the diffusion process of a high quality image. This allows the method to use a pre-traine...
This paper proposes an adversarial example detection mechanism for residual networks, as well as a regularization method to improve the detection performance. The detection mechanism is based upon viewing the deep network as a dynamical system. The regularization method is an existing method based on optimal transport ...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes an adversarial example detection mechanism for residual networks, as well as a regularization method to improve the detection performance. The detection mechanism is based upon viewing the deep network as a dynamical system. The regularization method is an existing method based on optimal tr...
The authors propose DyDecNet, a new network architecture for counting sounds in highly polyphonic environments. The method contains a series of T-F band-pass filters which double in each layer to provide a learnable alternative to other frequency representation methods. It then contains a backbone and regresses the fin...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose DyDecNet, a new network architecture for counting sounds in highly polyphonic environments. The method contains a series of T-F band-pass filters which double in each layer to provide a learnable alternative to other frequency representation methods. It then contains a backbone and regresses...
The paper is a contribution to the understanding the capabilities of transformers to model languages. The first part of the paper provides results on the capabilities of (shallow) transformers to simulate computations of semiautomata at length T. The second part of the paper is an experimental study on whether learning...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper is a contribution to the understanding the capabilities of transformers to model languages. The first part of the paper provides results on the capabilities of (shallow) transformers to simulate computations of semiautomata at length T. The second part of the paper is an experimental study on whether ...
The paper provides a primal-dual formulation to derive self-attention as a solution to the support vector regression problem having the primal formulation in the form of a neural network layer. The formulation allows the authors to derive various types of attention: linear, softmax, sparse, and multi-head attention usi...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper provides a primal-dual formulation to derive self-attention as a solution to the support vector regression problem having the primal formulation in the form of a neural network layer. The formulation allows the authors to derive various types of attention: linear, softmax, sparse, and multi-head atten...
The paper proposes to deal with the data heterogeneity in federated learning by local-global feature alignment and weighted combination of local classifiers. The whole model is decomposed into two parts, a feature extractor $f$ and a linear classifier $g$ built on $f$. The feature alignment is achieved by extracting fe...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: The paper proposes to deal with the data heterogeneity in federated learning by local-global feature alignment and weighted combination of local classifiers. The whole model is decomposed into two parts, a feature extractor $f$ and a linear classifier $g$ built on $f$. The feature alignment is achieved by extra...
This work proposes an HPO method that learns to select the best HP from the rankings of HPs. The main insight comes from the authors that the optimal strategy for training surrogates is to preserve the ranks of the performances of HPO as an L2R problem. More specifically, the method meta-learns neural network surrogate...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes an HPO method that learns to select the best HP from the rankings of HPs. The main insight comes from the authors that the optimal strategy for training surrogates is to preserve the ranks of the performances of HPO as an L2R problem. More specifically, the method meta-learns neural network s...
This paper builds on the PESNet model of NN-based wave functions for small polyatomic molecules [this reviewers recommends avoiding re-using names for model architectures to avoid compromising anonymity). The works proposes two innovations. One is learning a surrogate model of the VMC energies to avoid O(N^4) scaling i...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper builds on the PESNet model of NN-based wave functions for small polyatomic molecules [this reviewers recommends avoiding re-using names for model architectures to avoid compromising anonymity). The works proposes two innovations. One is learning a surrogate model of the VMC energies to avoid O(N^4) s...
This paper proposes a debiasing framework using low-rank regularization and self-supervised learning techinque. Specifically, Authors find that rank regularization may force deep networks to focus on spurious attributes and the biased models with strong regularization can effectively probe out the bias-conflicting samp...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a debiasing framework using low-rank regularization and self-supervised learning techinque. Specifically, Authors find that rank regularization may force deep networks to focus on spurious attributes and the biased models with strong regularization can effectively probe out the bias-conflict...
This paper proposes to combine the slot-based model with the gaussian mixture model (GMM) to improve the object-centric model. It explicitly represents the slot as the clustering center and uses the distance between slots to learn the mixture model. The experiments show certain improvements compared with some previous ...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper proposes to combine the slot-based model with the gaussian mixture model (GMM) to improve the object-centric model. It explicitly represents the slot as the clustering center and uses the distance between slots to learn the mixture model. The experiments show certain improvements compared with some p...
This paper proposes a new method of multiple-instance learning from high resolution images, which uses a patch selection scheme that processes available patches in mini-batches, scores them, then aggregates the top-M highest-scoring patches’ embeddings to reach a bag/image-level prediction. The proposed IPS Transformer...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new method of multiple-instance learning from high resolution images, which uses a patch selection scheme that processes available patches in mini-batches, scores them, then aggregates the top-M highest-scoring patches’ embeddings to reach a bag/image-level prediction. The proposed IPS Tra...