ICML 2023 Papers — Page 2
International Conference on Machine Learning · 1828 papers
Adversarial Collaborative Learning on Non-IID Features
Qinbin Li (University of California Berkeley), Dawn Song (University of California Berkeley)
Federated LearningAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: Proposes an adversarial collaborative learning framework called ADCOL in a feature non-IID environment.
Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples
Chumeng Liang (Shanghai Jiao Tong University), Haibing Guan (Shanghai Jiao Tong University)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: A method for generating adversarial examples (AdvDM) for diffusion models (LDM) is proposed and implemented to prevent the theft of artworks by AI-for-Art through techniques such as text inversion.
Adversarial Learning of Distributional Reinforcement Learning
Yang Sui (Shanghai University of Finance and Economics), Fan Zhou (University of North Carolina at Chapel Hill)
Reinforcement LearningSequential
🎯 What it does: Construct a perturbation manifold and propose the influence measure (FI) to quantify the impact of small perturbations of various components in reinforcement learning systems on performance.
Adversarial Parameter Attack on Deep Neural Networks
Lijia Yu (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Xiao-Shan Gao (Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposed and implemented an 'adversarial parameter attack' that perturbs the parameters of a trained deep neural network, significantly reducing robustness while maintaining accuracy close to the original.
Adversarial Policies Beat Superhuman Go AIs
Tony Tong Wang, Stuart Russell (University of California Berkeley)
Adversarial AttackReinforcement Learning
🎯 What it does: Trained adversarial strategies against the top Go AI KataGo, achieving win rates of 99.9% and over 97% with minimal computational resources (only 0.13% of KataGo's training) without using search or using 10^7 search nodes, respectively.
Adversarial robustness of amortized Bayesian inference
Manuel Gloeckler (University of Tübingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
Computational EfficiencyAdversarial AttackTime SeriesOrdinary Differential Equation
🎯 What it does: This study investigates the adversarial robustness when using Neural Posterior Estimation (NPE) in simulation inference and proposes a regularization method based on the Fisher information matrix to enhance robustness.
Adversarially Robust PAC Learnability of Real-Valued Functions
Idan Attias (Ben-Gurion University), Steve Hanneke (Purdue University)
OptimizationAdversarial Attack
🎯 What it does: This study investigates the PAC learning robustness of real-valued functions under adversarial attacks during testing, proposes a robust regression learning framework, and provides upper bounds on sample complexity, proving that a finite fat-shattering dimension is a sufficient condition for learnability.
Algorithmic Collective Action in Machine Learning
Moritz Hardt (Max Planck Institute for Intelligent Systems), Tijana Zrnic (University of California Berkeley)
ClassificationOptimizationTransformerLarge Language ModelText
🎯 What it does: The study investigates how a small number of participants can manipulate model learning outcomes through coordinated data modifications on a machine learning platform, proposing theoretical models and strategies in three learning scenarios: classification, risk minimization, and gradient optimization.
Algorithmic Stability of Heavy-Tailed SGD with General Loss Functions
Anant Raj (University of Illinois Urbana-Champaign), Umut Simsekli (Inria)
OptimizationStochastic Differential Equation
🎯 What it does: This paper studies the algorithmic stability of continuous-time and discrete-time stochastic differential equations with α-stable Lévy noise (corresponding to SGD with heavy-tailed noise) and translates it into an upper bound on generalization error.
Algorithms for bounding contribution for histogram estimation under user-level privacy
Yuhan Liu (Cornell University), Marco Gruteser (Google Research)
OptimizationSafty and PrivacyTabular
🎯 What it does: The study investigates how to select the optimal user contribution limit for histogram estimation under user-level differential privacy and proposes corresponding private algorithms.
Aligning Language Models with Preferences through $f$-divergence Minimization
Dongyoung Go (Naver Corporation), Marc Dymetman (Independent Researcher)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the f-DPG framework, which aligns language models with human preferences by minimizing any f-divergence to approximate any evaluable target distribution.
All in a Row: Compressed Convolution Networks for Graphs
Junshu Sun (Chinese Academy of Sciences), Qingming Huang (Chinese Academy of Sciences)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a differentiable graph regularization method and a hierarchical graph convolutional network CoCN based on diagonal convolution, aimed at achieving Euclidean convolution transfer on graph data, capable of simultaneously learning node features and structural features.
Alternately Optimized Graph Neural Networks
Haoyu Han (Michigan State University), Jiliang Tang (Michigan State University)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A training framework for graph neural networks based on single-layer optimization (ALT-OPT) has been designed and implemented, replacing the traditional end-to-end two-layer optimization with alternating optimization of pseudo-labels and MLP parameters.
Alternating Local Enumeration (TnALE): Solving Tensor Network Structure Search with Fewer Evaluations
Chao Li (RIKEN AIP), Qibin Zhao (RIKEN AIP)
CompressionOptimizationImage
🎯 What it does: A new Tensor Network structure search algorithm TnALE is proposed, which significantly reduces the number of structure evaluations by using alternating local enumeration.
An Adaptive Entropy-Regularization Framework for Multi-Agent Reinforcement Learning
Woojun Kim (KAIST), Youngchul Sung (KAIST)
Reinforcement Learning
🎯 What it does: Designed the ADER framework to dynamically learn the target entropy of each agent to achieve a multi-agent exploration-exploitation trade-off.
An Effective Meaningful Way to Evaluate Survival Models
Shi-ang Qi (University of Alberta), Russell Greiner (University of Alberta)
Biomedical DataElectronic Health Records
🎯 What it does: A MAE-PO metric based on pseudo-observation is proposed to evaluate survival prediction models with right censoring, and the effectiveness of this metric is assessed through the generation of realistic semi-synthetic datasets.
An Information-Theoretic Analysis of Nonstationary Bandit Learning
Seungki Min (KAIST), Daniel Russo (Columbia Business School)
🎯 What it does: This paper addresses the multi-armed bandit learning problem in non-stationary (time-varying) environments by constructing an information-theoretic framework to analyze and define the performance limits achievable by learning algorithms in such environments.
An Instrumental Variable Approach to Confounded Off-Policy Evaluation
Yang Xu (North Carolina State University), Rui Song (North Carolina State University)
Reinforcement LearningVideo
🎯 What it does: This paper proposes a framework for offline policy evaluation (OPE) using instrumental variables (IV) in the presence of unmeasured confounding in MDP (and POMDP) environments.
An Investigation into Pre-Training Object-Centric Representations for Reinforcement Learning
Jaesik Yoon (SAP), Sungjin Ahn (Korea Advanced Institute of Science and Technology)
Representation LearningRobotic IntelligenceTransformerReinforcement LearningImageBenchmark
🎯 What it does: A systematic evaluation of the effectiveness of unsupervised object-centric representation (OCR) pre-training in pixel-based reinforcement learning is conducted, and a new object-centric task benchmark is proposed.
An SDE for Modeling SAM: Theory and Insights
Enea Monzio Compagnoni (University of Basel), Aurelien Lucchi (Inria)
OptimizationTabularStochastic Differential Equation
🎯 What it does: This paper provides a rigorous derivation of the SDE continuous-time model, offering a theoretical description of SAM and its variants (USAM, DNSAM), and explains how they prefer flat minima and their behavior near saddle points.
Analysis of Error Feedback in Federated Non-Convex Optimization with Biased Compression: Fast Convergence and Partial Participation
Xiaoyun Li (LinkedIn), Ping Li (LinkedIn)
OptimizationFederated LearningImage
🎯 What it does: This paper studies the error feedback (EF) mechanism when using bias compression in federated learning, proposing the Fed-EF framework (which includes two global optimizers: SGD and AMSGrad) and providing its theoretical convergence analysis. It proves that under conditions of data heterogeneity and local steps, a linear speedup can be achieved; it also analyzes for the first time the convergence of EF under partial participation (PP), discovering an additional slow factor caused by error 'lag'.
Analyzing Convergence in Quantum Neural Networks: Deviations from Neural Tangent Kernels
Xuchen You (University of Maryland), Xiaodi Wu (University of Maryland)
TabularPhysics Related
🎯 What it does: This paper studies the convergence behavior of quantum neural networks (QNN) under gradient flow training through theoretical analysis and numerical experiments. It proves that the convergence rate of QNN under Pauli measurements is sublinear, while linear convergence can be achieved through new 'asymptotic dynamics' under over-parameterized periodic ansatz.
Analyzing Diffusion as Serial Reproduction
Raja Marjieh (Princeton University), Thomas L. Griffiths (Princeton University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper establishes a theoretical correspondence by viewing the sampling process of diffusion models as a phenomenon of serial reproduction in cognitive science, explaining its robustness to types of noise and the impact of noise scheduling on generation quality, and proposes a metric called 'inversion complexity.'
Analyzing Privacy Leakage in Machine Learning via Multiple Hypothesis Testing: A Lesson From Fano
Chuan Guo (Meta AI), Maziar Sanjabi (Meta AI)
Safty and PrivacyTabularBiomedical Data
🎯 What it does: This paper studies reconstruction attacks on discrete data under differential privacy (DP) mechanisms and provides a method for deriving an upper bound on the attack advantage using Fano's inequality, which can numerically compute the privacy guarantee level for a given ϵ.
Anchor Sampling for Federated Learning with Partial Client Participation
Feijie Wu (Purdue University), Jing Gao (Purdue University)
Federated LearningComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A federated learning framework named FedAMD is proposed, which divides some participating clients into anchor point groups and miner groups through anchor point sampling, using large-batch gradients and multi-step small-batch updates to accelerate training and improve model performance.
Answering Complex Logical Queries on Knowledge Graphs via Query Computation Tree Optimization
Yushi Bai (Tsinghua University), Lei Hou (Tsinghua University)
OptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: The QTO (Query Computation Tree Optimization) method is proposed, which utilizes pre-trained knowledge graph embeddings (KGE) to score first-order subqueries and accurately solves the optimal entity allocation on the query computation tree through forward-backward recursive propagation, ensuring both the optimality of the answers and the interpretability of intermediate variables.
Anti-Exploration by Random Network Distillation
Alexander Nikulin (Tinkoff), Sergey Kolesnikov (Tinkoff)
Knowledge DistillationReinforcement LearningTabularBenchmark
🎯 What it does: A new offline RL method called SAC-RND is developed, which utilizes random network distillation for uncertainty estimation and suppresses exploration in offline RL.
Applied Online Algorithms with Heterogeneous Predictors
Jessica Maghakian (Stony Brook University), Zhenhua Liu
OptimizationReinforcement LearningTime Series
🎯 What it does: This paper studies the combination of online algorithms and diverse predictors, proposing an algorithm for the rental/purchase problem using parameter prediction and input prediction, and validating it on the task of minimizing bandwidth costs in distributed systems.
Approximate Causal Effect Identification under Weak Confounding
Ziwei Jiang (Purdue University), Murat Kocaoglu (Purdue University)
Tabular
🎯 What it does: This paper addresses the problem of causal effect identification under weak confounding (low-entropy unobserved confounding variables) and proposes a linear programming method with entropy constraints to provide upper and lower bounds for causal effects.
Approximate Stein Classes for Truncated Density Estimation
Daniel James Williams, Song Liu (University of Bristol)
Score-based Model
🎯 What it does: An approximate Stein class and Truncated Kernelized Stein Divergence (TKSD) are proposed to estimate truncated density models using only boundary point samples when the unknown truncation boundary is present.
Approximately Optimal Core Shapes for Tensor Decompositions
Mehrdad Ghadiri (Georgia Tech), Vahab Mirrokni (Google Research)
OptimizationBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A combinatorial optimization method is proposed to find the optimal core tensor shape (multilinear rank) in size-constrained Tucker decomposition;
Approximation Algorithms for Fair Range Clustering
Sedjro Salomon Hotegni (African Institute for Mathematical Sciences), Ali Vakilian (Toyota Technological Institute)
Optimization
🎯 What it does: This paper studies the problem of fair range clustering, aiming to select k centers from different population groups to minimize clustering costs while ensuring that each group has at least minimal representation in the center set and that no group dominates the center set.
Approximation and Estimation Ability of Transformers for Sequence-to-Sequence Functions with Infinite Dimensional Input
Shokichi Takakura (University of Tokyo), Taiji Suzuki (University of Tokyo)
TransformerImage
🎯 What it does: This paper theoretically analyzes the approximation and estimation capabilities of Transformers when handling infinite-dimensional sequence-to-sequence functions. It proves that under mixed/aniso-smooth and variable-smooth target functions, Transformers can avoid the curse of dimensionality and achieve near-optimal error rates.
Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN
Siyuan Li (Westlake University), Stan Z. Li (Westlake University)
Object DetectionSegmentationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes a unified mask image modeling framework A MIM for Transformer and CNN, aimed at enhancing general visual representation through mask learning of mid-level interactions.
Are Diffusion Models Vulnerable to Membership Inference Attacks?
Jinhao Duan (Drexel University), Kaidi Xu (Drexel University)
Adversarial AttackDiffusion modelImage
🎯 What it does: This paper studies the vulnerability of diffusion models under membership inference attacks and proposes an attack method called SecMI based on stepwise posterior error comparison.
Are Equivariant Equilibrium Approximators Beneficial?
Zhijian Duan (Peking University), Xiaotie Deng (Peking University)
🎯 What it does: This study investigates the theoretical properties of equivariant balance approximators, providing analyses of generalization bounds, sample complexity, and approximation performance, and proving their advantages and limitations.
Are Gaussian Data All You Need? The Extents and Limits of Universality in High-Dimensional Generalized Linear Estimation
Luca Pesce (Ecole Polytechnique Federale de Lausanne), Ludovic Stephan (Ecole Polytechnique Federale de Lausanne)
ClassificationOptimizationImage
🎯 What it does: This paper studies the generalized linear estimation (GLM) of Gaussian mixture data in high dimensions, providing precise asymptotic expressions for training and testing errors, and exploring when a single Gaussian model is sufficient to describe learning errors.
Are labels informative in semi-supervised learning? Estimating and leveraging the missing-data mechanism.
Aude Sportisse (Cote d'Azur University), Pierre-Alexandre Mattei (Cote d'Azur University)
ClassificationData-Centric LearningSupervised Fine-TuningImage
🎯 What it does: This study addresses the problem of non-random missing labels (MNAR) in semi-supervised learning, proposing to estimate the missing mechanism and debias any SSL algorithm through inverse propensity weighting.
Are Large Kernels Better Teachers than Transformers for ConvNets?
Tianjin Huang (Eindhoven University of Technology), Shiwei Liu (Eindhoven University of Technology)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates the use of large-kernel convolutional networks as teachers to enhance the performance of small-kernel networks through knowledge distillation, and conducts a systematic comparison with Vision Transformer teachers.
Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations
Yongyi Yang (University of Michigan), Wei Hu (University of Michigan)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper explores whether the final layer representation retains the fine-grained structure of the input distribution after the convergence of neural networks by refining the training labels and training on networks such as ResNet-18, challenging the completeness of the traditional Neural Collapse theory.
Are Random Decompositions all we need in High Dimensional Bayesian Optimisation?
Juliusz Krzysztof Ziomek (Huawei Noah's Ark Lab), Haitham Bou Ammar (Huawei Noah's Ark Lab)
OptimizationTabular
🎯 What it does: This paper proposes the use of random tree decomposition in high-dimensional Bayesian optimization as a replacement for traditional data-based decomposition learning, constructing the Random Decomposition Upper Confidence Bound (RDUCB) algorithm, and proving its theoretical guarantee of optimal error and information gain balance in expectation.
Arithmetic Sampling: Parallel Diverse Decoding for Large Language Models
Luke Vilnis (Google Research), Sumit Sanghai (Google Research)
GenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: A parallel diversified decoding method based on arithmetic coding is proposed, which uses an implicit arithmetic codebook to map sampling points to sequences, achieving unbiased and diverse sampling.
Atari-5: Distilling the Arcade Learning Environment down to Five Games
Matthew Aitchison (Australian National University), Marcus Hutter (DeepMind)
Knowledge DistillationReinforcement LearningVideoBenchmark
🎯 What it does: This paper proposes a systematic method based on linear regression to select a representative subset from a large number of Atari games and approximate the median score of the complete ALE benchmark using this subset.
Attention-Based Recurrence for Multi-Agent Reinforcement Learning under Stochastic Partial Observability
Thomy Phan (University of Southern California), Claudia Linnhoff-Popien (LMU Munich)
Recurrent Neural NetworkTransformerReinforcement LearningSequentialBenchmark
🎯 What it does: The AERIAL method is proposed, which approximates the multi-agent recursive value function using the memory states of attention-aggregated agents, thereby achieving more robust distributed decision-making in randomly partially observable Dec-POMDP environments, and introduces the MessySMAC benchmark based on SMAC.
Attribute-Efficient PAC Learning of Low-Degree Polynomial Threshold Functions with Nasty Noise
Shiwei Zeng (Stevens Institute of Technology), Jie Shen (Stevens Institute of Technology)
ClassificationOptimization
🎯 What it does: This paper studies the properties of low-degree polynomial threshold functions (PTFs) for efficient PAC learning, particularly K-sparse PTFs, and proposes a new algorithm that can PAC learn this class of functions under Gaussian marginal distributions with a sample complexity of O(K^4 d ε^(-2d) log^5(d) n), even in the presence of adversarial noise.
Attributing Image Generative Models using Latent Fingerprints
Guangyu Nie (Arizona State University), Yi Ren (Arizona State University)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Proposes embedding subtle semantic variations as fingerprints in the latent space of generative models to achieve traceability of the generative models;
AudioLDM: Text-to-Audio Generation with Latent Diffusion Models
Haohe Liu (University of Surrey), Mark D Plumbley
GenerationData SynthesisSuper ResolutionDiffusion modelMultimodalityAudio
🎯 What it does: This study proposes a continuous latent diffusion model based on CLAP embeddings for text-to-audio (TTA) generation.
Auto-Differentiation of Relational Computations for Very Large Scale Machine Learning
Yuxin Tang (Rice University), Chris Jermaine (Rice University)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes an automatic differentiation framework based on relational algebra, which can automatically convert relational queries into gradient calculations and directly train large-scale machine learning models in relational databases.
AutoCoreset: An Automatic Practical Coreset Construction Framework
Alaa Maalouf (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
ClassificationOptimizationTabularFinance Related
🎯 What it does: AutoCoreset is an automated core set construction framework where users only need to provide input data and the corresponding loss function. The system can generate a core set that meets the specified error without any problem-specific sensitivity calculations or manual operations.
Automated Search for Conjectures on Mathematical Constants using Analysis of Integer Sequences
Ofir Razon (Technion Israel Institute of Technology), Ido Kaminer (Technion Israel Institute of Technology)
SequentialPhysics Related
🎯 What it does: The ESMA (Enumerated Signed-continued-fraction Massey-Approve) algorithm is proposed, which utilizes the Berlekamp-Massey algorithm to identify linear recurrence patterns in integer sequences of constants (through signed numerator extended continued fractions), thereby automatically generating new conjectural formulas about mathematical constants (such as e, e², tan(1), Bessel function values).
Automatic Data Augmentation via Invariance-Constrained Learning
Ignacio Hounie (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
OptimizationData-Centric LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes an automatic data augmentation framework based on constraint learning, treating data augmentation as an invariance constraint problem, and adaptively generating augmented distributions through MCMC sampling.
Automatic Intrinsic Reward Shaping for Exploration in Deep Reinforcement Learning
Mingqi Yuan (Hong Kong Polytechnic University), Wenjun Zeng (Eastern Institute for Advanced Study)
Reinforcement Learning
🎯 What it does: Enhancing exploration capabilities in reinforcement learning by automatically selecting and using different intrinsic reward functions.
Automatically Auditing Large Language Models via Discrete Optimization
Erik Jones (University of California, Berkeley), Jacob Steinhardt (University of California, Berkeley)
OptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a framework that views model auditing as a discrete optimization problem and automatically searches for input-output pairs that can trigger target behaviors (such as toxic outputs, language switching, specific suffixes, etc.) using the Autoregressive Coordinate Ascent algorithm (ARCA).
Automatically marginalized MCMC in probabilistic programming
Jinlin Lai (University of Massachusetts Amherst), Daniel Sheldon (University of Massachusetts Amherst)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes an automated method for edge reversal and variable marginalization of directed graphical models generated by probabilistic programming languages, significantly reducing the variable dimension of HMC sampling and improving sampling efficiency.
Autoregressive Diffusion Model for Graph Generation
Lingkai Kong (Georgia Institute of Technology), Chao Zhang (Georgia Institute of Technology)
GenerationData SynthesisOptimizationGraph Neural NetworkReinforcement LearningDiffusion modelGraph
🎯 What it does: A self-regressive diffusion model called GRAPHARM is proposed, which generates graph structures in discrete graph space through node absorption diffusion.
Auxiliary Learning as an Asymmetric Bargaining Game
Aviv Shamsian (Bar-Ilan University), Ethan Fetaya
ImageAudio
🎯 What it does: The AuxiNash method is proposed, modeling auxiliary learning as an asymmetric game, enhancing the performance of the main task through dynamic learning task weights.
Auxiliary Modality Learning with Generalized Curriculum Distillation
Yu Shen (University of Maryland), Ming Lin
Autonomous DrivingKnowledge DistillationMultimodalityPoint Cloud
🎯 What it does: This paper proposes the Auxiliary Modality Learning (AML) framework, systematically defines and classifies auxiliary modalities and network architectures, and introduces the Smart Auxiliary Modality Distillation (SAMD) method to automatically select the optimal auxiliary modality and enhance the performance of the main modality through curriculum distillation under the 'super model' condition.
Averaged Method of Multipliers for Bi-Level Optimization without Lower-Level Strong Convexity
Risheng Liu (Dalian University of Technology), Jin Zhang (Southern University of Science and Technology)
OptimizationImage
🎯 What it does: A single-loop bi-level average multiplier method (sl-BAMM) is proposed to solve bi-level optimization problems without requiring strong convexity conditions in the lower level.
B-Learner: Quasi-Oracle Bounds on Heterogeneous Causal Effects Under Hidden Confounding
Miruna Oprescu (Cornell University), Uri Shalit (Technion Israel Institute of Technology)
TabularFinance Related
🎯 What it does: Proposes the B-Learner method for estimating the upper and lower bounds of conditional average treatment effects (CATE) in the presence of hidden confounding.
Bag of Tricks for Training Data Extraction from Language Models
Weichen Yu (Institute of Automation, Chinese Academy of Sciences), Shuicheng YAN
Hyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In response to the issue of training data leakage in language models, this paper systematically evaluates and improves various techniques in the two main stages of training data extraction: generation and ranking, proposing a complete extraction process based on GPT-Neo.
Bandit Multi-linear DR-Submodular Maximization and Its Applications on Adversarial Submodular Bandits
Zongqi Wan (Institute of Computing Technology, Chinese Academy of Sciences), Zhijie Zhang (Fujian Province Center for Applied Mathematics, Fuzhou University)
Optimization
🎯 What it does: This paper proposes a bandit maximization algorithm for online polynomial DR-submodular functions with diffuse feedback, called BanditMLSM, and extends it to general DR-submodular functions (BanditDRSM). By constructing a continuous polynomial extension, the two types of combinatorial diffuse submodular optimization (submodular maximization under partition matroid constraints and sequential submodular maximization) are transformed into the aforementioned problem, resulting in new sub-linear regret rate upper bounds.
Bandit Online Linear Optimization with Hints and Queries
Aditya Bhaskara (University of Utah), Manish Purohit (Google Research)
Optimization
🎯 What it does: This paper studies the online linear optimization problem with hints or query capabilities (Bandit OLO), analyzing that it is still impossible to break the standard O(√T) upper bound with only confidence hints. It proposes a model that allows for actively querying the cost function values, achieving a logarithmic expected regret upper bound under this model. Subsequently, in a stricter model with limited response feedback, it provides a robustness guarantee of logarithmic order plus a √B term by combining self-concordant regularization and exploration strategies.
Bandits with Knapsacks: Advice on Time-Varying Demands
Lixing Lyu (National University of Singapore), Wang Chi Cheung (National University of Singapore)
Recommendation SystemOptimizationReinforcement LearningTime SeriesSequential
🎯 What it does: A framework is proposed to improve decision-making in the non-stationary weighted knapsack problem by utilizing online predictions of total demand.
Banker Online Mirror Descent: A Universal Approach for Delayed Online Bandit Learning
Jiatai Huang (Institute for Interdisciplinary Information Sciences), Longbo Huang (Institute for Interdisciplinary Information Sciences)
OptimizationFinance Related
🎯 What it does: Proposes the Banker-OMD framework, decoupling Online Mirror Descent (OMD) from the delayed feedback mechanism, achieving a unified algorithm design for various delayed online Bandit learning tasks.
Bayes-optimal Learning of Deep Random Networks of Extensive-width
Hugo Cui (École Polytechnique Fédérale de Lausanne), Lenka Zdeborova (École Polytechnique Fédérale de Lausanne)
ClassificationOptimizationTabular
🎯 What it does: This paper studies the problem of learning the objective function of deep and wide random neural networks, proposing a closed-form expression for the Bayesian optimal test error, applicable to both regression and classification tasks.
Bayesian Design Principles for Frequentist Sequential Learning
Yunbei Xu (Columbia University), assaf zeevi
Reinforcement LearningSequential
🎯 What it does: A unified Bayesian design principle is proposed for constructing adaptive algorithms in frequency learning, utilizing the algorithm information ratio to balance exploration and exploitation.
Bayesian Estimation of Differential Privacy
Santiago Zanella-Beguelin (Microsoft), Daniel Jones
Safty and PrivacyImageText
🎯 What it does: A differential privacy budget (ɓε) estimation method based on Bayesian posterior distribution is proposed, and an end-to-end privacy assessment system is implemented.
Bayesian Neural Networks Avoid Encoding Complex and Perturbation-Sensitive Concepts
Qihan Ren (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the representational capacity of mean-field variational Bayesian neural networks (BNN), proving that BNNs are more difficult to learn high-order (complex) interaction concepts compared to ordinary DNNs, and experimentally validating this conclusion.
Bayesian online change point detection with Hilbert space approximate Student-t process
Jeremy Sellier (University College London), Petros Dellaportas (University College London)
Anomaly DetectionComputational EfficiencyTime Series
🎯 What it does: A Bayesian online change point detection framework based on the Student-t process and low-rank kernel approximation in Hilbert space is proposed.
Bayesian Progressive Deep Topic Model with Knowledge Informed Textual Data Coarsening Process
Zhibin Duan (Xidian University), Mingyuan Zhou (University of Texas at Austin)
Graph Neural NetworkText
🎯 What it does: A knowledge-driven text data coarsening process and the corresponding advanced generative model ProGBN are proposed, which enhance the deep representation and quality of topics in topic models through a coarse-to-fine generation process.
Bayesian Reparameterization of Reward-Conditioned Reinforcement Learning with Energy-based Models
Wenhao Ding (Carnegie Mellon University), Marco Pavone (Stanford University)
TransformerReinforcement LearningDiffusion modelContrastive LearningTabularSequential
🎯 What it does: This paper proposes Bayesian Reparameterized Reward-Conditioned RL (BR-RCRL), which enhances the generalization and robustness to out-of-distribution (OOD) conditions of RCRL through Bayesian reparameterization and energy models.
Beam Tree Recursive Cells
Jishnu Ray Chowdhury (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
Recurrent Neural NetworkText
🎯 What it does: This paper proposes the Beam Tree Recursive Cell (BT-Cell), a recursive neural network framework based on Beam Search, designed to automatically induce the hierarchical structure of text without relying on golden trees and to support gradient propagation.
BEATs: Audio Pre-Training with Acoustic Tokenizers
Sanyuan Chen (Harbin Institute of Technology), Furu Wei (Microsoft Research Asia)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningAudio
🎯 What it does: Designed and implemented the BEATS framework, using discrete label prediction to replace reconstruction loss for audio self-supervised pre-training.
Behavior Contrastive Learning for Unsupervised Skill Discovery
Rushuai Yang (Shanghai Artificial Intelligence Laboratory), Xuelong Li (Northwestern Polytechnical University)
Robotic IntelligenceReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper proposes an unsupervised skill discovery method based on behavior contrastive learning (BeCL), which maximizes the mutual information between different trajectory states under the same skill, achieving both diverse skill learning and improved state coverage.
Benign Overfitting in Deep Neural Networks under Lazy Training
Zhenyu Zhu (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (Ecole Polytechnique Federale de Lausanne)
ClassificationOptimization
🎯 What it does: The study investigates the phenomenon of 'benign overfitting' in shallow ReLU networks and extends it to deep networks, proving that under lazy training conditions, deep networks can achieve zero training error on noisy label data and approach the Bayes optimal test error.
Benign Overfitting in Two-layer ReLU Convolutional Neural Networks
Yiwen Kou (University of California), Quanquan Gu (University of California)
Convolutional Neural Network
🎯 What it does: This study investigates the benign overfitting phenomenon of two-layer ReLU convolutional neural networks in the presence of label-flipping noise, providing convergence bounds for training error and risk limits for testing error.
Best Arm Identification in Multi-Agent Multi-Armed Bandits
Filippo Vannella (KTH Royal Institute of Technology), Jaeseong Jeong (Ericsson)
OptimizationGraph
🎯 What it does: This paper studies the optimal arm identification problem in multi-agent multi-armed bandits (MAMAB) with a factor graph as the reward structure, deriving instance-specific lower bounds and providing a realizable algorithm MF-TaS;
Best of Both Worlds Policy Optimization
Christoph Dann (Google Research), Julian Zimmert
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a new strategy optimization algorithm aimed at achieving optimal outcomes in tabular Markov Decision Processes (MDP), capable of obtaining good regret bounds under both adversarial and stochastic losses.
Better Diffusion Models Further Improve Adversarial Training
Zekai Wang (Wuhan University), Shuicheng YAN
GenerationAdversarial AttackConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: By using more advanced diffusion models (EDM) to generate high-quality conditional images, combined with adversarial training (TRADES) to enhance the model's robustness against adversarial attacks.
Better Training of GFlowNets with Local Credit and Incomplete Trajectories
Ling Pan (Mila Quebec AI Institute), Yoshua Bengio (Mila Quebec AI Institute)
GenerationData SynthesisGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes the Forward-Looking GFlowNet (FL-GFN) method, which utilizes energy/reward information from intermediate states to achieve more efficient credit assignment and supports training using only incomplete trajectories.
Beyond Exponentially Fast Mixing in Average-Reward Reinforcement Learning via Multi-Level Monte Carlo Actor-Critic
Wesley Suttle, Dinesh Manocha (University of Maryland)
Reinforcement Learning
🎯 What it does: This paper proposes a multi-layer Monte Carlo based Average Reward Actor-Critic algorithm (MAC) for training policies in Markov Decision Processes (MDPs) with slow mixing speeds.
Beyond Homophily: Reconstructing Structure for Graph-agnostic Clustering
Erlin Pan (University of Electronic Science and Technology of China), zhao kang
Representation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A graph-independent node clustering method is proposed, achieving unsupervised high-quality clustering through graph reconstruction, mixed filters, and a dual graph clustering network.
Beyond In-Domain Scenarios: Robust Density-Aware Calibration
Christian Tomani (Technical University of Munich), Daniel Cremers (Technical University of Munich)
Domain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: A density-aware calibration method based on KNN (DAC) is proposed, which performs sample-dependent temperature scaling on the logits of a trained network by estimating and weighting the density of hidden layer features, in order to improve calibration performance under domain shift and OOD conditions.
Beyond Lipschitz Smoothness: A Tighter Analysis for Nonconvex Optimization
Zhengmian Hu (University of Maryland), Heng Huang (University of Pittsburgh)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a more detailed smoothness assumption by separating positive and negative curvature, and provides a more compact convergence analysis for methods such as Lookahead and Local SGD.
Beyond Reward: Offline Preference-guided Policy Optimization
Yachen Kang (Zhejiang University), Donglin Wang (Westlake University)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningSequential
🎯 What it does: A one-time offline preference-guided policy optimization (OPPO) framework is proposed, which directly learns policies in high-dimensional context space using preference labels without the need to first learn a scalar reward function.
Beyond the Edge of Stability via Two-step Gradient Updates
Lei Chen (New York University), Joan Bruna (New York University)
OptimizationTabular
🎯 What it does: This paper proposes a theory of the existence of stable periodic 2 orbits under high-order derivative conditions by analyzing two-step updates of gradient descent (GD) outside the 'stability boundary', and proves the convergence properties of GD at high learning rates in representative models such as 1D, 2D, single-layer ReLU neurons, and matrix factorization.
Beyond the Universal Law of Robustness: Sharper Laws for Random Features and Neural Tangent Kernels
Simone Bombari (Institute of Science and Technology), Marco Mondelli (Institute of Science and Technology)
OptimizationAdversarial AttackImage
🎯 What it does: This paper provides a theoretical analysis of the empirical risk minimization (ERM) solutions for two types of models: Random Features (RF) and Neural Tangent Kernel (NTK), and presents precise theorems regarding their adversarial robustness.
Beyond Uniform Lipschitz Condition in Differentially Private Optimization
Rudrajit Das (University of Texas Austin), sujay sanghavi
OptimizationSafty and PrivacyReinforcement LearningImage
🎯 What it does: This paper proposes a non-uniform Lipschitz assumption for the DP-SGD gradient and provides theoretical guidance on the clipping threshold in the case of approximate perfect fitting, along with corresponding convergence analysis.
Bi-directional Masks for Efficient N:M Sparse Training
Yuxin Zhang (Xiamen University), Rongrong Ji (Xiamen University)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Proposes the Bi-directional Masks (Bi-Mask) method for efficient training of N:M fine-grained sparse networks, achieving training acceleration through forward and backward separation of sparse masks and row permutation.
Biases in Evaluation of Molecular Optimization Methods and Bias Reduction Strategies
Hiroshi Kajino (IBM Research), Takayuki Osogami (IBM Research)
OptimizationDrug DiscoveryReinforcement LearningTabular
🎯 What it does: This study evaluates molecular optimization methods, theoretically analyzes and quantifies two types of biases in plugin performance estimators (model misjudgment and sample reuse), and proposes various bias correction and mitigation strategies, including bootstrapping, covariate shift adaptation, double robust estimation, and generator constraints.
BiBench: Benchmarking and Analyzing Network Binarization
Haotong Qin (Beihang University), Xianglong Liu (Beihang University)
Convolutional Neural NetworkTransformerImageMultimodalityBenchmark
🎯 What it does: This paper proposes and implements BiBench, a comprehensive evaluation benchmark covering 8 types of binarization algorithms, 9 datasets, 13 network architectures, 2 deployment libraries, and 14 hardware chips.
Bidirectional Adaptation for Robust Semi-Supervised Learning with Inconsistent Data Distributions
Lin-Han Jia (Nanjing University), Yu-Feng Li (Huawei Technologies)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A theoretical framework is proposed for the inconsistency between label distribution and feature distribution in semi-supervised learning, along with a corresponding bidirectional adaptation method;
Bidirectional Learning for Offline Model-based Biological Sequence Design
Can Chen, Mark Coates (McGill University)
OptimizationDrug DiscoveryTransformerLarge Language ModelBiomedical Data
🎯 What it does: A model-driven biological sequence design method BIB is proposed, which constructs a proxy model using a pre-trained language model and a linear head, and achieves closed-form loss for bidirectional learning through deep linearization.
Bidirectional Looking with A Novel Double Exponential Moving Average to Adaptive and Non-adaptive Momentum Optimizers
Yineng Chen (Wuhan University), hai zhao
ClassificationOptimizationImageTextAudio
🎯 What it does: Proposed the ADMeta bidirectional observation optimizer framework, combining double exponential moving average with dynamic Lookahead;
Bigger, Better, Faster: Human-level Atari with human-level efficiency
Max Schwarzer (Google DeepMind), Pablo Samuel Castro
Convolutional Neural NetworkReinforcement LearningVideoBenchmark
🎯 What it does: The BBF (Bigger, Better, Faster) algorithm is proposed and implemented, achieving superhuman performance within the Atari 100K benchmark (only 100K steps) by utilizing large-scale networks and a series of design improvements.
Bilevel Optimization with Coupled Decision-Dependent Distributions
Songtao Lu (IBM Research)
OptimizationTabular
🎯 What it does: A bi-level optimization model that includes decision-related data distribution is proposed, and two solving algorithms, Bi-RRM and Bi-SGD, are designed;
BiRT: Bio-inspired Replay in Vision Transformers for Continual Learning
Kishaan Jeeveswaran (NavInfo Europe), Elahe Arani (NavInfo Europe)
TransformerImage
🎯 What it does: This paper proposes BiRT, a continual learning method based on visual Transformers, which enhances generalization and suppresses catastrophic forgetting during representation replay using controllable noise.
Bit Allocation using Optimization
Tongda Xu (Institute for AI Industry Research Tsinghua University), Ya-Qin Zhang (Tsinghua University)
CompressionOptimizationVideo
🎯 What it does: This paper proposes an optimized bit allocation method that directly solves the optimal bit allocation at the pixel level by equating the GoP-level likelihood in neural video compression with semi-adaptive variational inference (SAVI).
Blackout Diffusion: Generative Diffusion Models in Discrete-State Spaces
Javier E. Santos (Los Alamos National Laboratory), Yen Ting Lin (Los Alamos National Laboratory)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A diffusion model theory for discrete state spaces is proposed, and a 'Blackout Diffusion' model based on pure death processes is implemented.
BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
Junnan Li (Salesforce Research), Steven Hoi (Salesforce Research)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: BLIP-2 pretrains a vision-language model by inserting a lightweight query Transformer (Q-Former) between a frozen image encoder and a large language model, completing zero-shot generation and retrieval tasks from images to text.