NeurIPS 2024 Papers — Page 2
Conference on Neural Information Processing Systems · 4035 papers
A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis
Yue Yang (University of Pennsylvania), Mark Yatskar (University of Pennsylvania)
ClassificationDomain AdaptationTransformerLarge Language ModelImageBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This paper proposes a Knowledge-Enhanced Bottleneck Model (KnoBo), which constructs a conceptual bottleneck by retrieving medical literature and utilizing large language models, combined with structural and parameter priors, to enhance the robustness of medical imaging models under domain transfer conditions (different hospitals, races, ages, etc.).
A theoretical case-study of Scalable Oversight in Hierarchical Reinforcement Learning
Tom Yan (Carnegie Mellon University), Zachary Chase Lipton (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: This paper studies how to achieve scalable supervision under limited artificial feedback in goal-driven hierarchical reinforcement learning, proposing two hierarchical learning algorithms for metric rewards and preference feedback respectively.
A theoretical design of concept sets: improving the predictability of concept bottleneck models
Max Ruiz Luyten (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
ClassificationDomain AdaptationTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In the paper, the authors propose and validate a theoretical framework and empirical effects of concept sets in the Concept Bottleneck Model (CBM), studying the expressiveness of concept sets and the model awareness bias, and providing corresponding theoretical derivations and experimental validations.
A Theoretical Perspective for Speculative Decoding Algorithm
Ming Yin (Princeton University), Mengdi Wang (Princeton University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: To address the inference acceleration problem of Speculative Decoding, the authors established a Markov chain abstraction from a theoretical perspective, provided an exact formula for the expected number of rejections, and proved that single-core Speculative Decoding is optimal among all rejection-based algorithms; they then extended this to a batch version, deriving theoretical upper limits for acceleration and distribution bias in batch processing, and presented the Pareto optimal frontier for rejection rates and distribution bias.
A Theoretical Understanding of Self-Correction through In-context Alignment
Yifei Wang (Massachusetts Institute of Technology), Yisen Wang (Peking University)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the self-correction mechanism of LLMs and provides a theoretical analysis from the perspective of contextual alignment.
A Theory of Optimistically Universal Online Learnability for General Concept Classes
Steve Hanneke (Purdue University), Hongao Wang (Purdue University)
ClassificationOptimization
🎯 What it does: In the online learning framework, this study investigates and fully characterizes which concept classes can achieve optimistically universal online learnability under binary labels, providing the corresponding minimal data process assumptions and designing the relevant universal learning algorithms; at the same time, the theory is extended to the agnostic case, proving the equivalence of learnability in both realizable and agnostic situations.
A Topology-aware Graph Coarsening Framework for Continual Graph Learning
Xiaoxue Han (Stevens Institute of Technology), Yue Ning (Stevens Institute of Technology)
OptimizationRepresentation LearningGraph Neural NetworkGraphTime Series
🎯 What it does: A topology-aware graph coarsening framework, TA CO, is proposed to alleviate the problem of catastrophic forgetting in continual graph learning (CGL) by storing compressed graphs of previous tasks.
A Tractable Inference Perspective of Offline RL
Xuejie Liu (Peking University), Yitao Liang (Peking University)
TransformerReinforcement LearningSequential
🎯 What it does: The Trifle algorithm is proposed, utilizing a tractable probabilistic model (TPM) to make the inference process of offline RL tractable, significantly improving the performance of offline RL in action sampling and reward estimation.
A two-scale Complexity Measure for Deep Learning Models
Massimiliano Datres (University of Trento), David Sutter (IBM Research Europe)
Convolutional Neural NetworkImage
🎯 What it does: Proposes a two-scale effective dimension (2sED) based on the Fisher information matrix and its lower bound (lower 2sED) as a measure of complexity for deep learning models.
A Unified Confidence Sequence for Generalized Linear Models, with Applications to Bandits
Junghyun Lee, Kwang-Sung Jun (University of Arizona)
OptimizationReinforcement LearningTabular
🎯 What it does: A unified likelihood ratio confidence sequence (CS) framework is proposed, applicable to all self-conjugate (GLM) models, and based on this, a universal UCB algorithm OFUGLB is designed.
A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks
Hoin Jung (Purdue University), Xiaoqian Wang (Purdue University)
ClassificationGenerationRetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A unified debiasing method called SFID is proposed, which can eliminate gender bias across various modalities and tasks (zero-shot classification, text-image retrieval, image captioning, text-image generation) in visual-language models (VLMs).
A Unified Framework for 3D Scene Understanding
Wei Xu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
SegmentationKnowledge DistillationTransformerPrompt EngineeringContrastive LearningPoint Cloud
🎯 What it does: This paper presents UniSeg3D, a unified 3D scene understanding framework that can simultaneously perform six tasks: generalization, semantic segmentation, instance segmentation, interaction segmentation, reference segmentation, and open vocabulary segmentation within a single model.
A Unified Principle of Pessimism for Offline Reinforcement Learning under Model Mismatch
Yue Wang (University of Central Florida), Shaofeng Zou (Arizona State University)
OptimizationReinforcement Learning
🎯 What it does: This paper addresses the challenges in offline reinforcement learning (RL) under model mismatch, proposing a unified pessimistic principle aimed at optimizing performance through distributionally robust Markov decision processes (MDP).
A Unifying Normative Framework of Decision Confidence
Amelia Johnson, Koosha Khalvati (Allen Institute)
OptimizationReinforcement LearningTabular
🎯 What it does: A unified normative framework is proposed to measure decision confidence using probabilistic models and map it to planning as inference (maximum entropy reinforcement learning).
A Unifying Post-Processing Framework for Multi-Objective Learn-to-Defer Problems
Mohammad-Amin Charusaie (Max Planck Institute for Intelligent Systems), Samira Samadi (Max Planck Institute for Intelligent Systems)
Anomaly DetectionOptimizationTabular
🎯 What it does: A post-processing framework based on the d-dimensional generalized Neyman-Pearson rule is proposed to simultaneously satisfy accuracy and various constraints (fairness, expert intervention budget, anomaly detection, etc.) in the multi-objective learn-to-defer (L2D) problem.
A Universal Growth Rate for Learning with Smooth Surrogate Losses
Anqi Mao (Courant Institute), Yutao Zhong (Courant Institute)
Optimization
🎯 What it does: This study investigates the H-consistency upper bounds of smooth surrogate losses and proves that they exhibit a square root growth rate in both binary and multi-class classification.
A Versatile Diffusion Transformer with Mixture of Noise Levels for Audiovisual Generation
Gwanghyun Kim (Seoul National University), Krishna Somandepalli (Google DeepMind)
GenerationData SynthesisTransformerDiffusion modelVideoMultimodalityAudio
🎯 What it does: A variable noise level diffusion framework (MoNL) is proposed, and based on this, an audio-video latent diffusion transformer (AVDiT) is constructed to accomplish various audio-video generation tasks, such as cross-modal generation, interpolation, and continuation, within a single model.
A versatile informative diffusion model for single-cell ATAC-seq data generation and analysis
Lei huang, Manolis Kellis (Massachusetts Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderBiomedical Data
🎯 What it does: Proposes the ATAC-Diff framework, which utilizes conditional diffusion models and an auxiliary GMM+MI module to generate and analyze single-cell ATAC-seq data.
A Walsh Hadamard Derived Linear Vector Symbolic Architecture
Mohammad Mahmudul Alam (University of Maryland, Baltimore County), James Holt (Laboratory for Physical Sciences)
ClassificationRecognitionContrastive LearningImage
🎯 What it does: This paper proposes a Hadamard transform-based linear vector symbolic architecture (HLB) and provides implementations for binding, unbinding, initialization, and projection.
A-FedPD: Aligning Dual-Drift is All Federated Primal-Dual Learning Needs
Yan Sun (University of Sydney), Dacheng Tao (Nanyang Technological University)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes A-FedPD, which utilizes virtual dual variable updates to eliminate the 'dual drift' problem that arises in partial participation federated learning, thereby improving training stability and convergence speed.
A2PO: Towards Effective Offline Reinforcement Learning from an Advantage-aware Perspective
Yunpeng Qing (Zhejiang University), Mingli Song (Zhejiang University)
Reinforcement LearningAuto EncoderTabular
🎯 What it does: A new offline reinforcement learning method A2PO is proposed, which addresses constraint conflicts in mixed-quality datasets through advantage-aware policy optimization.
Abductive Reasoning in Logical Credal Networks
Radu Marinescu (IBM Research), Alexander G. Gray
🎯 What it does: Define and solve the Maximum A Posteriori (MAP) and Marginal MAP (MMAP) inference tasks in Logic Credible Networks (LCNs), proposing two types of solution approaches: exact search and approximate message passing.
Abrupt Learning in Transformers: A Case Study on Matrix Completion
Pulkit Gopalani (University of Michigan), Wei Hu (University of Michigan)
TransformerLarge Language ModelTabular
🎯 What it does: This study explores the training dynamics of the Transformer model in the low-rank matrix completion task, finding that the training loss experiences a plateau in the early stages of training, followed by a sudden drop to near-optimal values.
Absorb & Escape: Overcoming Single Model Limitations in Generating Heterogeneous Genomic Sequences
Zehui Li (Imperial College London), Yiren Zhao (Imperial College London)
GenerationData SynthesisOptimizationDiffusion modelBiomedical Data
🎯 What it does: This paper proposes an Absorb & Escape (A&E) framework and a fast implementation called Fast A&E, which utilizes a combination of pre-trained autoregressive models and diffusion models to generate high-quality genomic sequences.
Abstract Reward Processes: Leveraging State Abstraction for Consistent Off-Policy Evaluation
Shreyas Chaudhari (University of Massachusetts), Philip S. Thomas (University of Massachusetts)
Reinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: A STAR framework based on state abstraction is proposed for consistent and low-variance off-policy evaluation.
Abstracted Shapes as Tokens - A Generalizable and Interpretable Model for Time-series Classification
Yunshi Wen (Rensselaer Polytechnic Institute), Anak Agung Julius (Rensselaer Polytechnic Institute)
ClassificationExplainability and InterpretabilityTransformerTime Series
🎯 What it does: A self-supervised pre-training model VQShape is proposed, using interpretable abstract shapes and attributes as discrete symbols for time series data;
Accelerated Regularized Learning in Finite N-Person Games
Kyriakos Lotidis (Stanford University), Nicholas Bambos (Stanford University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: An accelerated regularization learning method FT-XL is proposed, improving traditional exponential/multiplicative weight learning, achieving faster convergence in finite player games;
Accelerating Augmentation Invariance Pretraining
Jinhong Lin (University of Wisconsin Madison), Pedro Morgado (University of Wisconsin Madison)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A self-supervised pre-training acceleration framework based on Vision Transformer is proposed, which reduces gradient estimation costs through two sequence compression strategies: random token dropout and variable patch size, thereby accelerating model training.
Accelerating Blockwise Parallel Language Models with Draft Refinement
Taehyeon Kim (Korea Advanced Institute of Science and Technology), Adrian Benton (Google Research)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper analyzes the distribution of block parallel decoding (BPD) drafts and proposes two lightweight re-scoring methods to enhance the generation speed of BPD.
Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity
Haoxuan Chen (Stanford University), Grant M. Rotskoff (Stanford University)
GenerationComputational EfficiencyDiffusion modelStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A parallel inference algorithm for diffusion models, PIADM-SDE and PIADM-ODE, is proposed, which utilizes Picard iteration and exponential integrators to parallelize the sampling process in blocks, achieving a reduction in time complexity from linear to polynomial logarithmic.
Accelerating ERM for data-driven algorithm design using output-sensitive techniques
Maria Florina Balcan, Dravyansh Sharma (Carnegie Mellon University)
OptimizationComputational EfficiencyTabular
🎯 What it does: An empirical risk minimization (ERM) solving framework aimed at data-driven algorithm design is proposed, significantly accelerating the learning of parameterized algorithm families using an output-sensitive multi-dimensional hyperplane enumeration method.
Accelerating Greedy Coordinate Gradient and General Prompt Optimization via Probe Sampling
Yiran Zhao (National University of Singapore), Michael Shieh (National University of Singapore)
OptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the Probe Sampling algorithm, which uses a small draft model to filter candidate prompts in GCG search, significantly accelerating the adversarial prompt optimization for LLMs.
Accelerating Matroid Optimization through Fast Imprecise Oracles
Franziska Eberle (Technical University of Berlin), Jens Schlöter (Centrum Wiskunde en Informatica)
Optimization
🎯 What it does: This paper proposes a solution to the maximum weight problem of bases under two types of oracles (exact and fast rough), designing a series of algorithms based on error metrics and providing matching lower bounds.
Accelerating Nash Equilibrium Convergence in Monte Carlo Settings Through Counterfactual Value Based Fictitious Play
Qi Ju (Huazhong University of Science and Technology), YunFeng Luo (Huazhong University of Science and Technology)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackReinforcement LearningAgentic AI
🎯 What it does: A Monte Carlo sampling-based method for solving adversarial games is proposed—Monte Carlo Counterfactual Value-Based Fictitious Play (MCCFVFP).
Accelerating Non-Maximum Suppression: A Graph Theory Perspective
King-Siong Si (Xi'an Jiaotong University), Hao Sun (Xi'an Jiaotong University)
Object DetectionComputational EfficiencyImageBenchmark
🎯 What it does: A systematic analysis of the internal structure of Non-Maximum Suppression (NMS) from the perspective of graph theory is presented, proposing two efficient implementations (QSI-NMS, eQSI-NMS, and BOE-NMS), and constructing the NMS-Bench benchmark for quick evaluation and comparison of different NMS algorithms.
Accelerating Pre-training of Multimodal LLMs via Chain-of-Sight
Ziyuan Huang (Ant Group), Ming Yang (Ant Group)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Reduce the number of visual tokens during the pre-training phase through the Chain-of-Sight module, thereby accelerating the pre-training of multimodal LLMs.
Accelerating Relative Entropy Coding with Space Partitioning
Jiajun He (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
CompressionComputational EfficiencyImage
🎯 What it does: A relative entropy coding (REC) method utilizing spatial partitioning is proposed, significantly reducing encoding time while maintaining encoding length.
Accelerating Transformers with Spectrum-Preserving Token Merging
Hoai-Chau Tran (German Research Center for Artificial Intelligence), Mathias Niepert (Max Planck Research School for Intelligent Systems)
RetrievalComputational EfficiencyTransformerImageText
🎯 What it does: A Token merging algorithm named PITOME is proposed, which uses energy scores to protect important tokens and significantly reduces the computational load of Transformers.
Acceleration Exists! Optimization Problems When Oracle Can Only Compare Objective Function Values
Aleksandr Lobanov (Moscow Institute of Physics and Technology), Andrey Krasnov
OptimizationTabular
🎯 What it does: This paper proposes a new method for designing optimization algorithms that utilizes Order Oracle to solve black-box optimization problems, particularly demonstrating the existence of accelerated optimization algorithms in non-convex, convex, and strongly convex settings.
Accuracy is Not All You Need
Abhinav Dutta (Microsoft Research), Ramachandran Ramjee (Microsoft Research)
GenerationCompressionTransformerLarge Language ModelText
🎯 What it does: This paper reveals that even when accuracy remains unchanged, there can be a significant number of 'flips' (correct↔incorrect) in model responses through experiments on various quantization, pruning, and sparsification compression techniques applied to large language models, and demonstrates that this phenomenon is related to the differences in compression methods.
Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation
Yinghao Wu (Xiamen University), Yipeng Qin (Cardiff University)
Pose EstimationTransformerSequential
🎯 What it does: A sequence structure module (SSM) is added to the Transformer model to utilize the spatial and temporal structural information of fixed-length sequences in the task of attitude estimation with sparse inertial sensors, improving the accuracy and smoothness of attitude prediction.
ACES: Generating a Diversity of Challenging Programming Puzzles with Autotelic Generative Models
Julien Pourcel (Inria), Laetitia Teodorescu (Inria)
GenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a method for automatically generating diverse and controllable difficulty Python programming puzzles.
ACFun: Abstract-Concrete Fusion Facial Stylization
Jiapeng Ji (Xidian University), Cheng Deng (Xidian University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper proposes an ACFun method for face stylization that requires only a pair of images, utilizing the Abstract Fusion Module (AFun) and the Concrete Fusion Module (CFun) to extract abstract and concrete features of style and face, respectively, and achieves the fusion of style and face information through image-oriented alignment loss.
Achievable distributional robustness when the robust risk is only partially identified
Julia Kostin (ETH Zurich), Fanny Yang (ETH Zurich)
TabularBiomedical Data
🎯 What it does: Proposes the worst-case robust risk under a partially identifiable robustness framework, and derives its lower bound and optimal predictor in linear models.
Achievable Fairness on Your Data With Utility Guarantees
Muhammad Faaiz Taufiq (ByteDance Research), Yang Liu (University of California Santa Cruz)
OptimizationSupervised Fine-TuningImageTextTabular
🎯 What it does: This paper proposes a method that can obtain a complete and accurate accuracy-fairness trade-off curve after a single training session, and provides an algorithm with statistical confidence intervals to help evaluate the upper and lower limits of model fairness across different datasets.
Achieving $\tilde{O}(1/\epsilon)$ Sample Complexity for Constrained Markov Decision Process
Jiashuo Jiang (Hong Kong University of Science and Technology), Yinyu Ye (Stanford University)
OptimizationReinforcement LearningTabular
🎯 What it does: A new online solving algorithm based on linear programming is proposed for solving constrained Markov decision processes (CMDP) under unknown transition probabilities, and an upper bound on its sample complexity is provided.
Achieving Constant Regret in Linear Markov Decision Processes
Weitong Zhang (University of North Carolina at Chapel Hill), Quanquan Gu (University of California Los Angeles)
Reinforcement Learning
🎯 What it does: A new reinforcement learning algorithm Cert‑LSVI‑UCB is proposed, achieving a constant cumulative return error independent of the number of trials under linear MDPs (including cases with error constraints).
Achieving Domain-Independent Certified Robustness via Knowledge Continuity
Alan Sun (Carnegie Mellon University), Soroush Vosoughi (Dartmouth College)
ClassificationAdversarial AttackText
🎯 What it does: A new definition of robustness called Knowledge Continuity is proposed, along with its theoretical proof and practical applications.
Achieving Linear Convergence with Parameter-Free Algorithms in Decentralized Optimization
Ilya Kuruzov (Innopolis University), Alexander Gasnikov (Innopolis University)
OptimizationTabular
🎯 What it does: A parameter-independent algorithm is proposed for decentralized networks that does not require prior knowledge of gradient smoothness or network connectivity parameters, and achieves step size adaptation through local backtracking line search.
Achieving Near-Optimal Convergence for Distributed Minimax Optimization with Adaptive Stepsizes
Yan Huang (Zhejiang University), Jinming Xu (Zhejiang University)
OptimizationGenerative Adversarial NetworkTabular
🎯 What it does: A distributed adaptive minimax optimization method called D-AdaST is proposed, which achieves global step size consistency through a step size tracking protocol and obtains near-optimal convergence in non-convex-strongly convex problems.
Achieving Optimal Clustering in Gaussian Mixture Models with Anisotropic Covariance Structures
Xin Chen (Princeton University), Anderson Ye Zhang (University of Pennsylvania)
OptimizationImage
🎯 What it does: The study addresses the clustering problem under Gaussian Mixture Models (GMM) with unknown and potentially different covariance matrices, providing lower bounds on optimal clustering error in two scenarios (homogeneous and heterogeneous), and proposes an improved Lloyd algorithm to achieve extremum optimal clustering error.
Achieving Tractable Minimax Optimal Regret in Average Reward MDPs
Victor Boone (University of Grenoble Alpes), Zihan Zhang (Princeton University)
Reinforcement Learning
🎯 What it does: A new algorithm called PMEVI-DT is proposed, which can achieve manageable minimax optimal regret in average reward Markov Decision Processes (MDP) without requiring prior knowledge.
Acoustic Volume Rendering for Neural Impulse Response Fields
Zitong Lan (University of Pennsylvania), Mingmin Zhao (University of Pennsylvania)
MeshPhysics RelatedAudio
🎯 What it does: A neural acoustic response field based on frequency domain volume rendering and spherical surface integration (AVR) is proposed, which can synthesize accurate impulse responses under any orientation of the sound source and listener.
ActAnywhere: Subject-Aware Video Background Generation
Boxiao Pan (Stanford University), Jimei Yang (Runway)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: This paper proposes a method for automatically generating video backgrounds that match the motion of foreground subjects, using a single frame background image to create a complete video with realistic interactions with the subject.
ActFusion: a Unified Diffusion Model for Action Segmentation and Anticipation
Dayoung Gong (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
RecognitionSegmentationTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a unified diffusion model (ActFusion) that simultaneously performs temporal action segmentation and long-term action prediction in videos, utilizing learnable masking tokens to achieve segmentation of visible parts and prediction of invisible parts.
Action Gaps and Advantages in Continuous-Time Distributional Reinforcement Learning
Harley Wiltzer (Mila-Quebec AI Institute McGill University), Yash Jhaveri (Rutgers University Newark)
Reinforcement LearningTime SeriesFinance RelatedStochastic Differential Equation
🎯 What it does: This study investigates continuous-time distributed reinforcement learning, where action values collapse as decision frequency increases. It proposes the concept of distributed superiority and designs an algorithm called DSUP based on this concept to address performance instability under high-frequency decision-making.
Action Imitation in Common Action Space for Customized Action Image Synthesis
Wang Lin (Zhejiang University), Hanwang Zhang (Nanyang Technological University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The TwinAct method is proposed to achieve customized action image generation with decoupled actions and actors under few-shot conditions.
Activating Self-Attention for Multi-Scene Absolute Pose Regression
Miso Lee (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
Pose EstimationTransformerImage
🎯 What it does: A solution is proposed to activate the self-attention of the Transformer encoder for multi-scene absolute pose regression.
Activation Map Compression through Tensor Decomposition for Deep Learning
Le-Trung Nguyen (Telecom Paris), Van-Tam Nguyen (Telecom Paris)
ClassificationSegmentationCompressionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: When training deep networks on edge devices, this paper compresses storage by using low-rank tensor decomposition (SVD and HOSVD) on activation maps, significantly reducing the memory requirements for backpropagation.
Active Classification with Few Queries under Misspecification
Vasilis Kontonis (University of Texas at Austin), Christos Tzamos (University of Athens)
ClassificationOptimization
🎯 What it does: This paper proposes a new query language called 'Threshold Statistical Query' (TSQ) and demonstrates its polynomial logarithmic level query complexity under Massart and random classification noise in half-space learning; it also shows that under adversarial noise, the query efficiency of TSQ cannot exceed the linear lower bound.
Active Learning for Derivative-Based Global Sensitivity Analysis with Gaussian Processes
Syrine Belakaria (Stanford University), Eytan Bakshy (Meta)
Tabular
🎯 What it does: An active learning method for derivative-based global sensitivity analysis (DGSM) is proposed, which directly targets the quantification metrics of gradient, absolute gradient, and squared gradient for sample-efficient collection.
Active Learning of General Halfspaces: Label Queries vs Membership Queries
Ilias Diakonikolas (University of Wisconsin Madison), Mingchen Ma (University of Wisconsin Madison)
ClassificationOptimizationTabular
🎯 What it does: The study investigates the problem of learning general half-spaces under Gaussian distribution, particularly in the cases of label queries and membership queries.
Active learning of neural population dynamics using two-photon holographic optogenetics
Andrew Wagenmaker (University of California), Kevin Jamieson (University of Washington)
Reinforcement LearningTime SeriesBiomedical Data
🎯 What it does: Utilizing two-photon holographic optogenetics to design optical stimulation patterns, actively learning the dynamics of neural populations and causal connectivity matrices.
Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios
Nicolás Astorga (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
OptimizationData-Centric LearningTransformerLarge Language ModelTabular
🎯 What it does: A proactive learning framework named POCA is proposed, specifically designed to address the issues of partial observation and cost-constrained data collection.
Active Perception for Grasp Detection via Neural Graspness Field
Haoxiang Ma (Beihang University), Di Huang (Beihang University)
Object DetectionRobotic IntelligenceNeural Radiance FieldPoint Cloud
🎯 What it does: A proactive perception framework based on Neural Grasp Field (NGF) is proposed for efficiently detecting grasp poses in cluttered environments.
Active preference learning for ordering items in- and out-of-sample
Herman Bergström (Chalmers University of Technology and University of Gothenburg), Fredrik D. Johansson (Chalmers University of Technology and University of Gothenburg)
Recommendation SystemOptimizationImageTabular
🎯 What it does: Learn complete rankings on item sets with contextual attributes through Active Preference Learning, and provide an upper bound on ranking error along with an implemented active sampling strategy.
Active Sequential Posterior Estimation for Sample-Efficient Simulation-Based Inference
Sam Griesemer (University of Southern California), Yan Liu (University of Southern California)
Flow-based ModelTabular
🎯 What it does: The Active Sequential Neural Posterior Estimation (ASNPE) method is proposed for efficient inference of posterior distributions in expensive simulation models.
Active Set Ordering
Quoc Phong Nguyen (Deakin University), Patrick Jaillet (Massachusetts Institute of Technology)
OptimizationTabularTime Series
🎯 What it does: This paper proposes the active set ranking problem and presents the MP algorithm based on Gaussian process mean predictions, which estimates the top-k set of a black-box function with a limited budget, providing theoretical regret guarantees and experimental validation.
Active, anytime-valid risk controlling prediction sets
Ziyu Xu (Carnegie Mellon University), Paul Mineiro (Microsoft)
OptimizationImage
🎯 What it does: A method for anytime-valid risk control prediction sets (RCPS) is proposed and extended to the active labeling scenario to achieve label efficiency and safety guarantees.
ActSort: An active-learning accelerated cell sorting algorithm for large-scale calcium imaging datasets
Yiqi Jiang (Stanford University), Mark Schnitzer
ClassificationComputational EfficiencyBiomedical DataBenchmark
🎯 What it does: This paper presents ActSort, an active learning-based cell sorting algorithm that can quickly perform quality control of cell candidates in large calcium imaging datasets.
Ad Auctions for LLMs via Retrieval Augmented Generation
MohammadTaghi Hajiaghayi (University of Maryland), Suho Shin (University of Maryland)
GenerationRecommendation SystemTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a 'paragraph auction' mechanism based on Retrieval-Augmented Generation (RAG), aimed at reasonably embedding advertisements into the generated text of large language models (LLMs) and providing bidding and display opportunities for advertisers through an auction process, maintaining a balance between advertising revenue and text quality.
Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting
Zongjiang Shang (Zhejiang University), Dongliang Cui (Zhejiang University)
Graph Neural NetworkTransformerTime Series
🎯 What it does: Proposes Ada-MSHyper, an adaptive multi-scale hypergraph transformer model aimed at improving time series forecasting.
AdaFlow: Imitation Learning with Variance-Adaptive Flow-Based Policies
Xixi Hu (University of Texas at Austin), Bo Liu (University of Texas at Austin)
Robotic IntelligenceFlow-based ModelRectified FlowSequentialOrdinary Differential Equation
🎯 What it does: This paper proposes AdaFlow, an adaptive modal behavior cloning method based on flow models, which can achieve near real-time action generation while ensuring multimodal decision-making.
Adam on Local Time: Addressing Nonstationarity in RL with Relative Adam Timesteps
Benjamin Ellis (University of Oxford), Jakob Nicolaus Foerster
OptimizationReinforcement LearningSequential
🎯 What it does: In response to the impact of non-stationarity in RL on the Adam optimizer, Adam-Rel is proposed to control the update size by resetting the time step rather than the momentum.
Adam with model exponential moving average is effective for nonconvex optimization
Kwangjun Ahn (Microsoft Research), Ashok Cutkosky (Boston University)
Optimization
🎯 What it does: Analyzes the convergence of Adam and EMA in non-convex optimization and proves that it can achieve optimal rates.
AdanCA: Neural Cellular Automata As Adaptors For More Robust Vision Transformer
Yitao Xu (École polytechnique fédérale de Lausanne), Sabine Susstrunk
Domain AdaptationComputational EfficiencyAdversarial AttackTransformerImage
🎯 What it does: Incorporating a pluggable Neural Cellular Automaton (AdaNCA) into the visual Transformer to enhance the model's robustness against adversarial attacks and out-of-distribution samples.
AdaNeg: Adaptive Negative Proxy Guided OOD Detection with Vision-Language Models
Yabin Zhang (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Anomaly DetectionTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: The AdaNeg method is proposed, which dynamically generates adaptive negative proxies through a feature memory bank during testing to better align with the OOD distribution.
AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases
Jun Xia (Westlake University), Stan Z. Li (Westlake University)
RecognitionData-Centric LearningTransformerBiomedical Data
🎯 What it does: This paper proposes the AdaNovo framework, which utilizes Conditional Mutual Information (CMI) to re-weight training loss in order to improve the performance of de-biasing in de novo peptide sequencing, particularly for the identification of PTMs.
AdaPKC: PeakConv with Adaptive Peak Receptive Field for Radar Semantic Segmentation
Teng Li (Shenzhen International Graduate School Tsinghua University), Zhe Ma (Intelligent Science and Technology Academy of CASIC)
SegmentationConvolutional Neural NetworkPoint Cloud
🎯 What it does: For the radar semantic segmentation task, Adaptive Peak-aware Convolution (AdaPKC) is proposed with two implementations (metric-based AdaPKC ξ and learning-based AdaPKC θ). Additionally, a threshold online switching fine-tuning strategy (FiTOS) is introduced to further enhance performance.
Adaptable Logical Control for Large Language Models
Honghua Zhang (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
GenerationKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: The Ctrl-G framework is proposed, using hidden Markov models (HMM) as an interpretable approximation of large language models, and employing deterministic finite automata (DFA) for logical constraint control, thereby achieving inferable and satisfiable generation.
Adapting Diffusion Models for Improved Prompt Compliance and Controllable Image Synthesis
Deepak Sridhar (University of California), Nuno Vasconcelos (University of California)
GenerationData SynthesisPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes a factor graph-based diffusion model (FG-DM) that models the joint distribution of images and various visual conditions (semantic segmentation, depth, normals, sketches, etc.) to achieve higher prompt compliance, controllable image generation, and fine-grained editing.
Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models
Gen Li (Chinese University of Hong Kong), Yuling Yan (University of Wisconsin Madison)
Diffusion modelScore-based ModelTabular
🎯 What it does: This paper conducts a theoretical analysis of the coefficient design for the DDPM sampler in the case where the target distribution is near low-dimensional manifolds in high-dimensional space, and provides a coefficient scheme that achieves dimension-independent convergence rates.
Adaptive $Q$-Aid for Conditional Supervised Learning in Offline Reinforcement Learning
Jeonghye Kim (Korea Advanced Institute of Science and Technology), Youngchul Sung (Korea Advanced Institute of Science and Technology)
Reinforcement LearningTabular
🎯 What it does: This paper proposes an algorithm called QCS, which combines return-conditioned supervised learning (RCSL) with the concatenation capability of the Q-function;
Adaptive and Optimal Second-order Optimistic Methods for Minimax Optimization
Ruichen Jiang (University of Texas at Austin), Aryan Mokhtari (University of Texas at Austin)
Optimization
🎯 What it does: An adaptive, line search-free second-order optimistic method is proposed to solve convex-concave minimax optimization problems, along with a parameter-free version.
Adaptive Depth Networks with Skippable Sub-Paths
Woochul Kang (Incheon National University), Hyungseop Lee (Incheon National University)
Computational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A network architecture with predictable and adjustable depth is designed, dividing the residual blocks into necessary and skipable sub-paths, and using self-distillation to refine features in the latter, allowing multiple depth sub-networks to be obtained from a single training.
Adaptive Domain Learning for Cross-domain Image Denoising
Zian Qian (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
RestorationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: An Adaptive Domain Learning (ADL) framework is proposed, which trains a cross-domain RAW image denoising model using a small amount of target sensor data combined with existing source domain data.
Adaptive Experimentation When You Can't Experiment
Yao Zhao (University of Arizona), Lalit K Jain
🎯 What it does: This paper proposes a framework for adaptive experiments using incentive design and linear structural equation models (including covariates) in situations where direct random experiments are not feasible, referred to as the 'Constrained Pure Exploration Transductive Linear Bandit' (CPET-LB) problem, and provides corresponding algorithms.
Adaptive Exploration for Data-Efficient General Value Function Evaluations
Arushi Jain (McGill University), Doina Precup (McGill University)
Reinforcement LearningTabular
🎯 What it does: This paper proposes and implements GVFExplorer, which can efficiently evaluate multiple Generalized Value Functions (GVFs) in parallel by adaptively learning a single behavior policy, and achieves minimum variance updates on offline data using TD variance estimation.
Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare
Hanwei Zhu (City University of Hong Kong), Shiqi Wang (City University of Hong Kong)
OptimizationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A large model-based NR-IQA model called Compare2Score is proposed, which learns image quality comparisons and converts them into continuous scores.
Adaptive Important Region Selection with Reinforced Hierarchical Search for Dense Object Detection
Dingrong Wang (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
Object DetectionReinforcement LearningImage
🎯 What it does: Proposes the Adaptive Important Region Selection (AIRS) framework, which utilizes reinforcement learning for adaptive hierarchical search on feature pyramids to reduce false positives in dense detection.
Adaptive Labeling for Efficient Out-of-distribution Model Evaluation
Daksh Mittal (Columbia University), Hongseok Namkoong (Columbia University)
Domain AdaptationReinforcement LearningBiomedical Data
🎯 What it does: An adaptive label sampling framework is proposed for efficiently evaluating the performance of machine learning models under severe distribution shifts.
Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment
Wei Li (University of Birmingham), Shengjie Sun (AISpeech Co., Ltd.)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: Proposes the Adaptive Layer Sparsity (ALS) method, which compresses large language models by adaptively allocating layer-wise sparsity rates while maintaining or improving inference performance.
Adaptive Passive-Aggressive Framework for Online Regression with Side Information
Runhao Shi (Hong Kong University of Science and Technology), Daniel P. Palomar (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencySupervised Fine-TuningTabularTime SeriesFinance Related
🎯 What it does: An adaptive passive-attack framework (APAS) is proposed for online regression problems, particularly in handling complex scenarios with additional information.
Adaptive Preference Scaling for Reinforcement Learning with Human Feedback
Ilgee Hong (Georgia Institute of Technology), Tuo Zhao (Georgia Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes an Adaptive Preference Scaling (APS) loss for reward function learning in Reinforcement Learning from Human Feedback (RLHF);
Adaptive Proximal Gradient Method for Convex Optimization
Yura Malitsky (University of Vienna), Konstantin Mishchenko (Samsung AI Center)
Optimization
🎯 What it does: An adaptive gradient descent and adaptive proximal gradient descent algorithm (AdGD/AdProxGD) based on local curvature information is proposed, which does not require a global Lipschitz constant and can automatically adjust the step size, theoretically converging to the optimal solution of convex problems.
Adaptive Randomized Smoothing: Certified Adversarial Robustness for Multi-Step Defences
Saiyue Lyu (University of British Columbia), Mathias Lécuyer (Google DeepMind)
OptimizationAdversarial AttackGaussian SplattingImageBenchmark
🎯 What it does: Proposes Adaptive Randomized Smoothing (ARS), which adaptively processes inputs during testing to enhance robustness.
Adaptive Sampling for Efficient Softmax Approximation
Tavor Baharav, Mert Pilanci (Stanford University)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextSequential
🎯 What it does: Proposes the AdaptiveSoftmax algorithm, which uses multi-armed bandit adaptive sampling to quickly approximate the top k outputs of softmax;
Adaptive Variance Reduction for Stochastic Optimization under Weaker Assumptions
Wei Jiang (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationImageText
🎯 What it does: This paper proposes Ada-STORM, an adaptive variance reduction algorithm that achieves optimal convergence rates in scenarios such as non-convex optimization, component optimization, and finite summation.
Adaptive Visual Scene Understanding: Incremental Scene Graph Generation
Naitik Khandelwal (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
GenerationData SynthesisSafty and PrivacyTransformerDiffusion modelImageBenchmark
🎯 What it does: A continuous scene graph generation (CSEGG) framework suitable for visual scenes is proposed, and three benchmarks for learning scenes are constructed;
AdaptiveISP: Learning an Adaptive Image Signal Processor for Object Detection
Yujin Wang (Shanghai AI Laboratory), Jinwei Gu (Chinese University of Hong Kong)
Object DetectionReinforcement LearningImage
🎯 What it does: Designed and implemented an adaptive image signal processor (AdaptiveISP) that utilizes deep reinforcement learning to dynamically generate ISP pipelines and parameters for object detection tasks during inference, adjusting the processing flow in real-time to maximize detection performance.