ICML 2023 Papers — Page 19
International Conference on Machine Learning · 1828 papers
Weighted Tallying Bandits: Overcoming Intractability via Repeated Exposure Optimality
Dhruv Malik (Carnegie Mellon University), Aarti Singh (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: This paper proposes the Weighted Tallying Bandit (WTB) model, studying its complete strategy regret (CPR) under the condition of Repeated Exposure Optimality (REO), and provides an improved implementation of the Successive Elimination algorithm.
What Can Be Learnt With Wide Convolutional Neural Networks?
Francesco Cagnetta (École Polytechnique Fédérale de Lausanne), Matthieu Wyart (École Polytechnique Fédérale de Lausanne)
Convolutional Neural NetworkTabular
🎯 What it does: This paper conducts a theoretical analysis of the generalization performance of infinite-width deep convolutional neural networks (CNNs) under the kernel limit, providing spectral decompositions of their neural tangent kernel (NTK) and random feature kernel (RFK). It proves that the spectral structure corresponds to the hierarchical receptive fields of the network. Through spectral analysis and generalization bounds, it demonstrates that deep CNNs can adapt to spatial scales and significantly reduce the curse of dimensionality when learning objective functions that depend only on local input subsets; however, if the objective function is the output of a randomly initialized deep CNN, it will still encounter the curse of dimensionality. The paper also validates theoretical predictions through teacher-student experiments.
What can online reinforcement learning with function approximation benefit from general coverage conditions?
Fanghui Liu (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (Ecole Polytechnique Federale de Lausanne)
Reinforcement Learning
🎯 What it does: This paper studies the introduction and promotion of general coverage conditions in online reinforcement learning, proving that sample-efficient and computationally efficient learning algorithms can still be obtained without relying on traditional structural assumptions. By further refining the coverage conditions (such as Lp variants, density ratio realizability, partial/remnant coverage conditions) and applying them to linear MDPs, the paper provides improved regret upper bounds, achieving logarithmic regret under additional distribution conditions.
What do CNNs Learn in the First Layer and Why? A Linear Systems Perspective
Rhea Chowers (Hebrew University), Yair Weiss (Hebrew University)
Convolutional Neural NetworkImage
🎯 What it does: This paper conducts a systematic analysis of the representations in the first layer of convolutional neural networks (CNNs) and proposes the energy profile metric to quantify the sensitivity of filters to different spatial frequencies. Experiments demonstrate that CNNs with different initializations, architectures, datasets, and even training with random labels exhibit a high consistency in the energy profile of the first layer. Subsequently, an analytical formula for the energy profile is derived in a simplified linear CNN, proving that this energy profile approaches whitening during gradient descent training, and the formula is fitted to the energy profile of real nonlinear CNNs, validating the consistency between theory and practice.
What is Essential for Unseen Goal Generalization of Offline Goal-conditioned RL?
Rui Yang (Hong Kong University of Science and Technology), Tong Zhang (Hong Kong University of Science and Technology)
Robotic IntelligenceReinforcement LearningSequentialBenchmark
🎯 What it does: This study investigates the generalization problem of offline goal-conditioned reinforcement learning (offline GCRL) on unseen targets, proposing theoretical analysis and a practical algorithm (GOAT) while constructing a new evaluation benchmark.
What Makes Entities Similar? A Similarity Flooding Perspective for Multi-sourced Knowledge Graph Embeddings
Zequn Sun (Nanjing University), Wei Hu (Nanjing University)
Graph Neural NetworkGraph
🎯 What it does: This paper interprets the working principles of translation-based and aggregation-based entity alignment models by viewing the correction of entity pair similarity as a fixed point of similarity flooding, and based on this, proposes two improvement methods: similarity flooding based on entity combinations and autoregressive neighborhood aggregation.
When and How Does Known Class Help Discover Unknown Ones? Provable Understanding Through Spectral Analysis
Yiyou Sun (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
ClassificationGraph Neural NetworkContrastive LearningImage
🎯 What it does: A graph-based spectral contrastive learning framework (NSCL) is proposed to discover unknown categories with the assistance of known categories, providing a theoretical error upper bound.
When do Minimax-fair Learning and Empirical Risk Minimization Coincide?
Harvineet Singh (New York University), Chris Russell (Amazon Web Services)
Tabular
🎯 What it does: This paper explores the conditions under which conventional empirical risk minimization (ERM) and minimax-fair learning perform similarly in terms of worst-group error, providing theoretical proofs and large-scale empirical validation.
When does Privileged information Explain Away Label Noise?
Guillermo Ortiz-Jimenez (Ecole Polytechnique Federale de Lausanne), Effrosyni Kokiopoulou (Google Research)
ClassificationRecognitionKnowledge DistillationData-Centric LearningImage
🎯 What it does: A large-scale experiment was conducted to evaluate the effectiveness of using privileged information (PI) to mitigate label noise, analyzing different characteristics of PI and proposing improvement methods.
When is Realizability Sufficient for Off-Policy Reinforcement Learning?
Andrea Zanette (University of California)
Reinforcement Learning
🎯 What it does: Analyzed the statistical complexity of offline reinforcement learning under the condition of realizability without requiring Bellman completeness, and proposed a new concept of local inherent Bellman error.
When Personalization Harms Performance: Reconsidering the Use of Group Attributes in Prediction
Vinith Menon Suriyakumar, Berk Ustun (University of California San Diego)
Biomedical DataElectronic Health Records
🎯 What it does: The study investigates the phenomenon where the use of group attributes in personalized prediction models may lead to poorer predictions for certain groups, and proposes the 'fair use' criteria to assess and address this issue.
When Sparsity Meets Contrastive Models: Less Graph Data Can Bring Better Class-Balanced Representations
Chunhui Zhang (Brandeis University), Chuxu Zhang (Brandeis University)
ClassificationRepresentation LearningData-Centric LearningGraph Neural NetworkContrastive LearningImageGraph
🎯 What it does: Proposes the DataDec framework, which dynamically sparsifies graph data and models, utilizing contrastive learning to filter important samples based on gradients, addressing the issues of imbalance in graph data and computational burden.
Which Features are Learnt by Contrastive Learning? On the Role of Simplicity Bias in Class Collapse and Feature Suppression
Yihao Xue (University of California), Baharan Mirzasoleiman (University of California)
OptimizationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper systematically elaborates on two failure modes in contrastive learning from a theoretical perspective—class collapse and feature suppression—and proves that the simplicity bias of gradient descent is the fundamental cause of these two issues.
Which Invariance Should We Transfer? A Causal Minimax Learning Approach
Mingzhou Liu (Peking University), Yizhou Wang (Peking University)
Biomedical DataAlzheimer's Disease
🎯 What it does: A framework based on causal minimization learning is proposed to select the optimal transferable invariant subset from the training environment to enhance the model's robustness in unseen environments.
Which is Better for Learning with Noisy Labels: The Semi-supervised Method or Modeling Label Noise?
Yu Yao (Mohamed bin Zayed University of Artificial Intelligence), Tongliang Liu (University of Sydney)
ClassificationData-Centric LearningContrastive LearningTabular
🎯 What it does: The study investigates the performance differences between semi-supervised learning (SSL) methods and noise modeling methods in the presence of label noise, explaining their applicable scenarios from a causal generative process perspective. It also proposes the CDNL estimator for automatically detecting the causal/anti-causal structure of data.
Which Tricks are Important for Learning to Rank?
Ivan Lyzhin (Yandex), Liudmila Prokhorenkova (Yandex)
Recommendation SystemOptimizationTabularBenchmark
🎯 What it does: This paper provides a unified analysis of existing learning-to-rank methods based on gradient boosting decision trees and proposes an improved algorithm, YetiLoss, which can be optimized for any ranking metric.
Who Needs to Know? Minimal Knowledge for Optimal Coordination
Niklas Lauffer (University of California), Stuart Russell (University of California)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper studies the minimum necessary knowledge of partners in cooperative games and proposes a Strategic Equivalence Relation (SER) framework to distinguish between strategy-relevant and irrelevant information, thereby enabling efficient computation of optimal responses.
Whose Opinions Do Language Models Reflect?
Shibani Santurkar (Stanford University), Tatsunori Hashimoto (Stanford University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the construction of the 'OpinionQA' dataset using public opinion surveys and quantitatively evaluates the alignment of language models (LM) with opinions on subjective questions across 60 social groups in the United States, revealing significant biases in existing LMs regarding representativeness, adjustability, and consistency.
Why do Nearest Neighbor Language Models Work?
Frank F. Xu (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
RetrievalOptimizationTextRetrieval-Augmented Generation
🎯 What it does: A deep analysis of the k-nearest-neighbor language model (k-NN-LM) is conducted to explore why it can significantly reduce perplexity even when using the same retrieval corpus as the original training set. A general formula is proposed, and systematic ablation experiments are performed.
Why does Throwing Away Data Improve Worst-Group Error?
Kamalika Chaudhuri (Meta AI), David Lopez-Paz (Meta AI)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: Analyzes how discarding excess data improves the worst group error in the case of sample imbalance (categories or groups), and explains the geometric shift of the maximum margin linear classifier using extreme value theory.
Why Is Public Pretraining Necessary for Private Model Training?
Arun Ganesh (Google), Lun Wang (University of Washington)
OptimizationSafty and PrivacyTextAudio
🎯 What it does: This study investigates why public pre-training is necessary in the training of models under differential privacy, providing both theoretical proof and experimental validation.
Why Random Pruning Is All We Need to Start Sparse
Advait Harshal Gadhikar (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
ClassificationOptimizationGraph Neural NetworkImageTabular
🎯 What it does: The effectiveness of random pruning in training sparse networks is studied, and it is proven that random sparse networks can approximate any target network when the width is appropriately increased;
Why Target Networks Stabilise Temporal Difference Methods
Mattie Fellows (University of Oxford), Shimon Whiteson (University of Oxford)
Reinforcement LearningTabularSequential
🎯 What it does: The study proves that the target network can stabilize traditional TD learning through the Partial Fitting Policy Evaluation (PFPE) framework.
Width and Depth Limits Commute in Residual Networks
Soufiane Hayou (National University of Singapore), Greg Yang (Microsoft Research AI)
Stochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies the initialization behavior of residual networks (ResNet) as the width and depth approach infinity, proving that the limits of width and depth can be exchanged, and providing explicit limits for the pre-activation distribution and neural covariance;
WL meet VC
Christopher Morris (Aachen University), Martin Grohe (Aachen University)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper studies the generalization ability of Graph Neural Networks (GNN) in graph-level prediction tasks through VC dimension theory, and relates it to the expressiveness of the 1-dimensional Weisfeiler-Leman (1-WL) algorithm. It provides upper and lower bounds for the VC dimension under two scenarios: with and without a size limit on the graphs, and validates the theoretical predictions through experiments.
Wrapped Cauchy Distributed Angular Softmax for Long-Tailed Visual Recognition
Boran Han (Amazon Web Services)
ClassificationRecognitionImage
🎯 What it does: In the long-tail visual recognition task, the Wrapped Cauchy Distributed Angular Softmax (WCDAS) method is proposed to address the overfitting problem caused by sample imbalance and noise.
X-Paste: Revisiting Scalable Copy-Paste for Instance Segmentation using CLIP and StableDiffusion
Hanqing Zhao (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
Object DetectionSegmentationData SynthesisConvolutional Neural NetworkDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes X-Paste, a scalable Copy-Paste enhancement framework that utilizes CLIP and StableDiffusion to automatically collect and filter large-scale instances, generating high-quality masks that are then collaged onto background images for training.
XTab: Cross-table Pretraining for Tabular Transformers
Bingzhao Zhu (Cornell University), Mahsa Shoaran (EPFL)
Federated LearningRepresentation LearningTransformerContrastive LearningTabularBenchmark
🎯 What it does: Using the Transformer model for cross-table pre-training on tabular data, we propose the XTab framework, which trains a shared Transformer backbone using a federated learning approach.