NeurIPS 2024 Papers — Page 39
Conference on Neural Information Processing Systems · 4035 papers
UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation
Hanzhang Zhou (Nanyang Technological University), Kezhi Mao (Nanyang Technological University)
ClassificationGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By providing a mechanistic explanation of the FFN vectors and attention heads within large language models, this paper identifies and masks these internal components to eliminate model bias during the inference phase, thereby improving the robustness of in-context learning (ICL).
UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior
Yao Wu (Xiamen University), Yanyun Qu (Xiamen University)
SegmentationDomain AdaptationAutonomous DrivingTransformerPrompt EngineeringImagePoint Cloud
🎯 What it does: This paper proposes a unified cross-domain 3D semantic segmentation framework called UniDSeg, which utilizes the prior knowledge of visual foundation models (VFM) to enhance the model's adaptability and generalization ability across different domains.
Unified Covariate Adjustment for Causal Inference
Yonghan Jung (Purdue University), Elias Bareinboim (Columbia University)
Tabular
🎯 What it does: A Unified Covariate Adjustment (UCA) framework is proposed, which can cover various causal identification functions from backdoor, frontdoor to Tian adjustment, and based on this, a scalable and doubly robust DML-UCA estimator is designed;
Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection
Gyusam Chang (Korea University), Sangpil Kim (Korea University)
Object DetectionDomain AdaptationAutonomous DrivingTransformerSupervised Fine-TuningPoint Cloud
🎯 What it does: A unified domain generalization and adaptation framework (UDGA) is proposed to address the geometric distribution differences and label scarcity issues in multi-view 3D object detection.
Unified Generative and Discriminative Training for Multi-modal Large Language Models
Wei Chow (Zhejiang University), Qianru Sun (Singapore Management University)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A unified generative and discriminative training framework called Sugar is proposed, enabling multimodal large language models to possess both generative and retrieval capabilities.
Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement
Zhehao Huang (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
ClassificationGenerationDiffusion modelImage
🎯 What it does: This paper studies a unified gradient-based machine unlearning method that combines residual geometric enhancement;
Unified Graph Augmentations for Generalized Contrastive Learning on Graphs
Jiaming Zhuo (Hebei University of Technology), Liang Yang (Northwestern Polytechnical University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A Unified Graph Augmentation (UGA) module and a UGA-based general graph contrastive learning framework called GOUDA are proposed for flexible and low-complexity graph augmentation and contrastive learning across various graph tasks.
Unified Guidance for Geometry-Conditioned Molecular Generation
Sirine Ayadi (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
GenerationDrug DiscoveryDiffusion modelGraph
🎯 What it does: This paper proposes UniGuide, a unified geometric condition guidance framework for controlling unconditional molecular diffusion models to generate molecular geometric constraints required for structure-based, fragment-based, and ligand-based drug design tasks.
Unified Insights: Harnessing Multi-modal Data for Phenotype Imputation via View Decoupling
Qiannan Zhang (Weill Cornell Medicine), Fei Wang (Weill Cornell Medicine)
Knowledge DistillationRepresentation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningMultimodalityBiomedical DataAlzheimer's DiseaseElectronic Health Records
🎯 What it does: By decoupling the biological perspective and the phenotypic perspective within the framework of graph neural networks, this study infers missing phenotypic values in electronic medical records using multimodal biological data.
Unified Lexical Representation for Interpretable Visual-Language Alignment
Yifan Li (Fudan University), Tong He (Amazon Web Services)
RetrievalExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: The LexVLA framework is proposed, unifying vision and language into a sparse vocabulary representation, utilizing DINOv2 and Llama2 to achieve interpretable cross-modal alignment.
Unified Mechanism-Specific Amplification by Subsampling and Group Privacy Amplification
Jan Schuchardt (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
Safty and PrivacyTabular
🎯 What it does: This paper proposes a unified mechanism-specific subsampling amplification framework, utilizing conditional optimal transport theory to derive subsampling amplification and group privacy amplification under various privacy metrics such as Approximate Differential Privacy (ADP), Rényi DP, and dominating pairs, thereby obtaining tighter privacy analyses than traditional mechanism-independent upper bounds.
Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs
Alexandros Haliassos (Imperial), Maja Pantic (Meta)
RecognitionTransformerMultimodalityAudio
🎯 What it does: This paper proposes a Unified Speech Recognition (USR) framework that uses the same Transformer model to simultaneously perform auditory (ASR), visual (VSR), and audiovisual (AVSR) speech recognition tasks.
UniFL: Improve Latent Diffusion Model via Unified Feedback Learning
Jiacheng Zhang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
GenerationOptimizationComputational EfficiencyReinforcement LearningDiffusion modelImage
🎯 What it does: The UniFL framework is proposed, which uses unified feedback learning to enhance the visual quality, aesthetic appeal, and inference speed of LDM.
Uniform Last-Iterate Guarantee for Bandits and Reinforcement Learning
Junyan Liu (University of Washington), Lin Yang
OptimizationReinforcement Learning
🎯 What it does: A new metric called Unified Last Iteration Guarantee (ULI) is proposed to evaluate the instantaneous and cumulative performance of reinforcement learning algorithms.
Unifying Generation and Prediction on Graphs with Latent Graph Diffusion
Cai Zhou (Massachusetts Institute of Technology), Muhan Zhang (Peking University)
ClassificationGenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelGraphStochastic Differential Equation
🎯 What it does: The Latent Graph Diffusion (LGD) framework is proposed, unifying the generation, regression, and classification tasks of graph data into a generative task; by training a diffusion model in the latent space, it achieves the simultaneous generation of nodes, edges, and graph-level features in one go; and formalizes regression/classification tasks as conditional generation problems, providing theoretical guarantees.
Unifying Homophily and Heterophily for Spectral Graph Neural Networks via Triple Filter Ensembles
Rui Duan (Guangzhou University), Haoran Yang (Tongji University)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a spectral graph convolutional neural network named TFE-GNN, which adaptively extracts homophily and heterophily information from graphs through Triple Filter Ensemble (TFE) and utilizes initial features for node classification.
UniGAD: Unifying Multi-level Graph Anomaly Detection
Yiqing Lin (Tsinghua University), Jia Li (Hong Kong University of Science and Technology)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: Proposes UniGAD, a unified framework for anomaly detection at the node, edge, and graph levels.
UniIF: Unified Molecule Inverse Folding
Zhangyang Gao (Zhejiang University), Stan Z. Li (Westlake University)
Protein Structure PredictionGraph Neural NetworkGraph
🎯 What it does: A unified molecular inverse folding model, UniIF, is proposed, capable of handling different types of molecules such as proteins, RNA, and inorganic materials within the same framework.
UniMTS: Unified Pre-training for Motion Time Series
Xiyuan Zhang (University of California San Diego), Jingbo Shang (University of California San Diego)
ClassificationPose EstimationRepresentation LearningGraph Neural NetworkLarge Language ModelContrastive LearningTextTime Series
🎯 What it does: A UniMTS model that can be uniformly pre-trained was constructed by comparing the full-joint motion sequences generated by physical simulation with text descriptions through contrastive learning.
UNION: Unsupervised 3D Object Detection using Object Appearance-based Pseudo-Classes
Ted Lentsch (Delft University of Technology), Dariu Gavrila
Object DetectionAutonomous DrivingTransformerOptical FlowPoint CloudTime Series
🎯 What it does: Proposes the UNION framework to achieve unsupervised 3D object detection: utilizing LiDAR, cameras, and temporal information to generate pseudo bounding boxes and pseudo class labels, directly training existing detectors without the need for manual annotation or multiple rounds of self-training.
Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image
Kailu Wu (Tsinghua University), Kaisheng Ma (Tsinghua University)
GenerationDiffusion modelImageMesh
🎯 What it does: A framework named Unique3D is proposed, capable of generating high-quality, high-resolution, and finely textured 3D meshes from a single real scene image within 30 seconds.
UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections
Fangjinhua Wang (ETH Zurich), Federico Tombari (Google)
RestorationGenerationNeural Radiance FieldPoint CloudMesh
🎯 What it does: Using neural implicit fields and self-learning weight fields, this paper unifies the camera view radiance field and the reflection view radiance field to achieve high-fidelity 3D reconstruction of complex scenes with reflections.
UNIT: Unifying Image and Text Recognition in One Vision Encoder
Yi Zhu (Huawei Noah's Ark Lab), Hang Xu (Huawei Noah's Ark Lab)
RecognitionTransformerVision Language ModelImageText
🎯 What it does: The UNIT framework is proposed, using a single Vision Transformer encoder to simultaneously perform image recognition and text recognition.
Unitary Convolutions for Learning on Graphs and Groups
Bobak Kiani, Melanie Weber (Harvard University)
Graph Neural NetworkGraph
🎯 What it does: Proposes unitary graph convolution to avoid over-smoothing and enhance the stability of deep networks.
United We Stand, Divided We Fall: Fingerprinting Deep Neural Networks via Adversarial Trajectories
Tianlong Xu (Huazhong University of Science and Technology), Wei Liu (Huazhong University of Science and Technology)
ClassificationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a model fingerprint extraction and verification method based on adversarial trajectories (ADV-TRA), which describes the decision boundary of the model by generating a multi-level chain of adversarial samples and constructs surface trajectories that cover multi-class decision surfaces to identify whether the model has been stolen.
UniTS: A Unified Multi-Task Time Series Model
Shanghua Gao (Harvard University), Marinka Zitnik (Harvard University)
ClassificationAnomaly DetectionTransformerPrompt EngineeringTime SeriesBiomedical DataFinance Related
🎯 What it does: A unified multi-task time series model called UNITS is proposed, capable of simultaneously handling various tasks such as forecasting, classification, anomaly detection, and imputation.
Unity by Diversity: Improved Representation Learning for Multimodal VAEs
Thomas M. Sutter (ETH Zurich), Stephan Mandt (University of California Irvine)
GenerationRepresentation LearningMixture of ExpertsAuto EncoderMultimodalityBiomedical Data
🎯 What it does: A multi-modal variational autoencoder (MMVM VAE) based on a mixed expert prior is proposed, achieving soft sharing of latent variables across different modalities.
Universal Exact Compression of Differentially Private Mechanisms
Yanxiao Liu (Chinese University of Hong Kong), Cheuk Ting Li (Chinese University of Hong Kong)
CompressionSafty and PrivacyTabular
🎯 What it does: A new scheme for compressing different privacy mechanisms is proposed - Poisson Private Representation (PPR), which can accurately simulate any local or global differential privacy mechanism and compress its communication volume.
Universal In-Context Approximation By Prompting Fully Recurrent Models
Aleksandar Petrov (University of Oxford), Adel Bibi (University of Oxford)
Recurrent Neural NetworkPrompt Engineering
🎯 What it does: This study constructs a programming language called LSRL and uses it to prove that various recursive neural networks (RNN, LSTM, GRU, linear RNN, and their gated variants) can approximate any function through in-context prompting, demonstrating that they are universal in-context approximators.
Universal Neural Functionals
Allan Zhou (Stanford University), James Harrison (Google DeepMind)
OptimizationRecurrent Neural NetworkTransformerTextSequential
🎯 What it does: A universal neural functional body (UNF) is proposed that can automatically construct permutation equivariant networks in arbitrary weight spaces, and it is used for tasks such as predicting model generalization and training learning optimizers.
Universal Online Convex Optimization with $1$ Projection per Round
Wenhao Yang (Nanjing University), Lijun Zhang (Nanjing University)
Optimization
🎯 What it does: An online convex optimization algorithm is proposed, which requires only one projection per round to achieve optimal tuning performance for general convex functions, exponentially concave functions, and strongly convex functions.
Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators
Benedikt Alkin (John Kepler University Linz), Johannes Brandstetter (John Kepler University Linz)
TransformerPoint CloudPhysics Related
🎯 What it does: A unified Transformer-based neural operator framework called UPT has been developed for efficient scaling to spatiotemporal problems across different grids, particles, and scales, enabling dynamic propagation in a compressed latent space and arbitrary spatiotemporal queries.
Universal Rates for Active Learning
Steve Hanneke (Purdue University), Grigoris Velegkas (Yale University)
ClassificationOptimization
🎯 What it does: This study investigates the optimal learning curve of binary classifiers in active learning and provides a complete characterization of achievable learning rates under different distributions.
Universal Rates of Empirical Risk Minimization
Steve Hanneke (Purdue University), Mingyue Xu (Purdue University)
🎯 What it does: This paper studies the universal learning rates of the empirical risk minimization (ERM) algorithm under realizable settings and proves that it satisfies four categories (exponential, linear, logarithmic/linear, and arbitrarily slow) of complete partition.
Universal Sample Coding
Szymon Kobus (Imperial College London), Deniz Gunduz
CompressionFederated LearningText
🎯 What it does: A method of 'universal sample coding' is proposed to generate multiple independent and identically distributed samples from an unknown distribution at the receiver with the minimum number of bits.
Universality in Transfer Learning for Linear Models
Reza Ghane (California Institute of Technology), Babak Hassibi (California Institute of Technology)
Domain AdaptationOptimization
🎯 What it does: This study investigates the transfer learning and fine-tuning issues under linear models, providing a rigorous analysis of the generalization error and training error for regression and binary classification in the limit of large models;
Universality of AdaGrad Stepsizes for Stochastic Optimization: Inexact Oracle, Acceleration and Variance Reduction
Anton Rodomanov (CISPA Helmholtz Center for Information Security), Sebastian U Stich
Optimization
🎯 What it does: A unified AdaGrad step size adaptive gradient method (basic version UniSgd and accelerated version UniFastSgd) is proposed, suitable for convex composite optimization problems where the main function is approximately smooth and can only be accessed through (possibly biased) stochastic gradient oracles. Convergence analysis is provided under various scenarios (uniform variance, implicit variance reduction, SVRG variance reduction).
Unlearnable 3D Point Clouds: Class-wise Transformation Is All You Need
Xianlong Wang (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
Anomaly DetectionData-Centric LearningGraph Neural NetworkPoint Cloud
🎯 What it does: A 'non-learnable' framework for 3D point clouds is proposed, utilizing category-adaptive multi-transformations (rotation, scaling, distortion, shearing) to generate data that cannot be learned by unauthorized models, and a scheme is provided for recovery of learning through inverse transformations.
Unleashing Multispectral Video's Potential in Semantic Segmentation: A Semi-supervised Viewpoint and New UAV-View Benchmark
Wei Ji (University of Alberta), Hongxia Jin (Samsung AI Center-Mountain View)
SegmentationConvolutional Neural NetworkVideoMultimodalityBenchmark
🎯 What it does: This paper first constructs a multispectral video semantic segmentation dataset MVUAV based on the UAV bird's-eye view and proposes a semi-supervised multispectral video segmentation framework SemiMV to fully utilize sparse annotations and massive unlabeled videos.
Unleashing Region Understanding in Intermediate Layers for MLLM-based Referring Expression Generation
Yaoyuan Liang (Tsinghua University), Shao-Lun Huang (Tsinghua University)
GenerationTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: This paper proposes a method for generating referential expressions based on a multimodal large language model, utilizing intermediate layer region information for decoding, and filtering high-quality descriptions through cyclic consistency discrimination to reduce object hallucination.
Unleashing the Denoising Capability of Diffusion Prior for Solving Inverse Problems
Jiawei Zhang (Tsinghua University), Yuantao Gu (Tsinghua University)
RestorationOptimizationDiffusion modelImageAudio
🎯 What it does: A framework for solving inverse problems based on diffusion models is proposed—ProjDiff, which constructs auxiliary variables using the diffusion process to transform noisy observations into equivalent noise samples, and solves a two-variable constrained optimization problem through projected gradient descent.
Unleashing the Potential of the Diffusion Model in Few-shot Semantic Segmentation
Muzhi Zhu (Zhejiang University), Chunhua Shen (Zhejiang University)
SegmentationDiffusion modelImage
🎯 What it does: A framework called DiffewS is proposed, which utilizes the Stable Diffusion model (Stable Diffusion 2.1) for few-shot semantic segmentation, generating target segmentation masks directly in the latent space.
Unlock the Intermittent Control Ability of Model Free Reinforcement Learning
Jiashun Liu (Tianjin University), Tangjie Lv (NetEase)
Robotic IntelligenceReinforcement LearningAuto EncoderSequential
🎯 What it does: The MARS method is proposed to map multi-step action sequences to a low-dimensional latent space to address the intermittent control problem in reinforcement learning.
Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance
Jiwan Hur (KAIST), Junmo Kim (KAIST)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A self-guidance sampling method is proposed, improving the image synthesis quality and diversity of Masked Generative Models (MGMs);
Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought
Qiguang Chen (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
OptimizationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed and validated the 'Reasoning Boundary Framework (RBF)' to quantitatively assess and optimize the chain of thought (CoT) capabilities of large language models.
Unlocking the Potential of Global Human Expertise
Elliot Meyerson (Cognizant AI Labs), Risto Miikkulainen (University of Texas at Austin)
OptimizationKnowledge DistillationTime Series
🎯 What it does: The RHEA framework was designed and implemented, first distilling diverse policy models submitted by global experts into neural networks through behavioral cloning, and then using evolutionary search to crossover and mutate these networks, ultimately obtaining a multi-objective policy set that better balances case numbers and intervention costs.
Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models
Sanae Lotfi (New York University), Andrew Gordon Wilson (New York University)
GenerationCompressionTransformerLarge Language ModelText
🎯 What it does: This paper proposes a token-level generalization bound based on Markov properties, which provides a non-empty generalization bound for the next token prediction task of LLMs.
Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback
Hamish Ivison (University of Washington), Hannaneh Hajishirzi (University of Washington)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: This paper explores and compares four core elements of preference-based learning: preference data, learning algorithms, reward models, and strategy training prompts. It also proposes best practices for using PPO with large-scale reward models under the condition of synthesizing high-quality data.
Unraveling the Gradient Descent Dynamics of Transformers
Bingqing Song (University of Minnesota), Mingyi Hong (University of Minnesota)
ClassificationOptimizationTransformerImageText
🎯 What it does: This study investigates the gradient descent convergence behavior of a single-layer Transformer under different attention kernels (Softmax and Gaussian), providing theoretical convergence conditions and experimental validation on text classification and image path tasks.
Unravelling in Collaborative Learning
Aymeric Capitaine (Centre de Mathématiques Appliquées CNRS École Polytechnique), Alain Oliviero Durmus
🎯 What it does: This paper explores the phenomenon of 'disintegration' caused by information asymmetry in data quality during strategic learner collaboration, and designs a non-transferable probability verification mechanism to restore a complete alliance.
Unrolled denoising networks provably learn to perform optimal Bayesian inference
Aayush Karan (Harvard University), Yonina C. Eldar (Weizmann Institute of Science)
RestorationOptimizationTabular
🎯 What it does: The research investigates whether networks trained using gradient descent through unrolled denoising networks can achieve Bayesian optimal inference performance in inverse problems such as compressed sensing and rank-one matrix estimation.
Unscrambling disease progression at scale: fast inference of event permutations with optimal transport
Peter A. Wijeratne (University of Sussex), Daniel C. Alexander (University College London)
OptimizationExplainability and InterpretabilityComputational EfficiencyBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A variational event model based on optimal transport (vEBM) is proposed, which can achieve fast and interpretable disease progression sequence inference for thousands of high-dimensional features without the need for dimensionality reduction, and for the first time realizes pixel-level event sequences of brain images and retinal images.
UnSeg: One Universal Unlearnable Example Generator is Enough against All Image Segmentation
Ye Sun (Fudan University), Yu-Gang Jiang (Fudan University)
SegmentationOptimizationImage
🎯 What it does: A framework for generating unlearned samples (UnSeg) aimed at image segmentation tasks is proposed, utilizing the pre-trained Segment Anything Model to learn a general noise generator through a dual-layer optimal optimization, allowing any image with a given mask to generate unlearnable noise in a single forward pass, thereby rendering the model ineffective during training.
Unsupervised Anomaly Detection in The Presence of Missing Values
Feng Xiao (Chinese University of Hong Kong), Jicong Fan (Shenzhen Research Institute of Big Data)
Anomaly DetectionTabular
🎯 What it does: An end-to-end unsupervised anomaly detection method called ImAD is proposed, which can be directly trained and inferred on data with missing values.
Unsupervised Discovery of Formulas for Mathematical Constants
Michael Shalyt (Technion Israel Institute of Technology), Ido Kaminer (Technion Israel Institute of Technology)
🎯 What it does: This paper proposes an unsupervised clustering method based on the dynamical characteristics of PCF (Polynomial Continued Fractions) and uses this method to automatically discover and verify hundreds of thousands of formulas, resulting in the introduction of hundreds of new mathematical constant formulas.
Unsupervised Homography Estimation on Multimodal Image Pair via Alternating Optimization
Sanghyeob Song (Seoul National University), Sungroh Yoon (Seoul National University)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: An unsupervised multimodal image planar transformation estimation framework called AltO is proposed, which utilizes alternating optimization to handle geometric and modal discrepancies separately.
Unsupervised Modality Adaptation with Text-to-Image Diffusion Models for Semantic Segmentation
Ruihao Xia (East China University of Science and Technology), Pan Zhou (Singapore Management University)
SegmentationDomain AdaptationDiffusion modelImageMultimodality
🎯 What it does: Proposes the MADM method, extending unsupervised domain adaptation to multi-modal semantic segmentation; utilizes a pre-trained text-image diffusion model for cross-modal feature extraction and pseudo-label generation;
Unsupervised Object Detection with Theoretical Guarantees
Marian Longa (University of Oxford), Joao F. Henriques
Object DetectionConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A fully equivariant autoencoder architecture is proposed to achieve unsupervised object detection, and a theoretical upper bound on the error between latent variables and the true object positions is provided.
Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms
Mengyu Zhao (Rutgers University), Shirin Jalali (Rutgers University)
RestorationCompressionVideo
🎯 What it does: Using untrained neural networks (UNN) and deep image priors, a theoretical analysis framework is proposed, and the SCI-BDVP (Bagged-DVP) algorithm is designed to achieve high-quality video/spectral reconstruction from single-frame compressed imaging.
Unveil Benign Overfitting for Transformer in Vision: Training Dynamics, Convergence, and Generalization
Jiarui Jiang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
TransformerImage
🎯 What it does: This paper conducts a theoretical study on the training dynamics, convergence, and generalization performance of Vision Transformers (ViT), revealing the benign and harmful overfitting phenomena when the dataset is overfitted, and provides precise threshold conditions to determine the two situations.
Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness
Weilin Lin (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
ClassificationAnomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: For backdoor defense in the post-training phase, a two-stage method is proposed: first, the neuron weight changes (NWC) are obtained through 'clean unlearning' of the backdoored model, and these changes are used to zero-reset the high-weight-changing sub-weights to eliminate the backdoor; subsequently, an activation-aware fine-tuning (with gradient norm constraints) is performed on the reset model to restore clean accuracy.
Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?
Haoang Chi (National University of Defense Technology), Bo Han (Hong Kong Baptist University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper explores the causal reasoning ability of large language models, demonstrating that they only possess shallow (level-1) causal reasoning based on learned knowledge, rather than deep (level-2) true causal reasoning.
Unveiling Encoder-Free Vision-Language Models
Haiwen Diao (Dalian University of Technology), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A vision-language model EVE was proposed and trained, which is a decoder-only model without a visual encoder.
Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers
Siyu Chen (Yale University), Zhuoran Yang (Yale University)
TransformerSequential
🎯 What it does: This study investigates the training dynamics of a two-layer Transformer on n-gram Markov chain data and proves that the gradient flow converges to a limit model that implements the generalized induction head (GIH) mechanism, thereby achieving in-context learning.
Unveiling LoRA Intrinsic Ranks via Salience Analysis
Wenjun Ke (Southeast University), Yining Li (Southeast University)
Supervised Fine-TuningTextTime Series
🎯 What it does: An adaptive method for dynamically allocating the internal rank of LoRA through significance analysis of intra-time series rank is proposed, called SalientLoRA.
Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models
Ziyi Zhou (Southern University of Science and Technology), Xuetao Wei (Lingnan University)
Large Language ModelText
🎯 What it does: This paper evaluates the symmetric moral consistency of large language models in moral contexts and reveals the impact of position bias and choice bias on consistency assessment. It proposes the tSMC framework to correct biases and calculate the corrected consistency scores.
Unveiling the Hidden Structure of Self-Attention via Kernel Principal Component Analysis
Rachel Teo, Tan Minh Nguyen
OptimizationAdversarial AttackTransformerImageText
🎯 What it does: This paper derives the self-attention mechanism of the Transformer from the perspective of Kernel Principal Component Analysis (Kernel PCA) and proposes a robust self-attention mechanism called RPC-Attention, which further enhances the model's robustness against interference and attacks.
Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation
Nayeon Kim (Samsung Advanced Institute of Technology), Sujin Jang (Samsung Advanced Institute of Technology)
Object DetectionAutonomous DrivingTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A clip-level based online vectorized HD map construction framework called MapUnveiler is proposed.
Unveiling The Matthew Effect Across Channels: Assessing Layer Width Sufficiency via Weight Norm Variance
Yiting Chen (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerImageTextGraph
🎯 What it does: Dynamic analysis of the weight norm variance across different channels in the same layer reveals two distinct variance evolution patterns (IS and DS) for wide and narrow layers during training. A width adaptation strategy is designed based on this pattern, adjusting the width layer by layer in CNN architectures such as VGG and ResNet, which reduces parameters while enhancing performance.
Unveiling the Potential of Robustness in Selecting Conditional Average Treatment Effect Estimators
Yiyan HUANG, Qi WU
OptimizationReinforcement LearningTabular
🎯 What it does: A distributionally robust metric (DRM) is proposed to select CATE estimators, avoiding the need to fit irrelevant parameters and focusing on robustness under distributional uncertainty.
Unveiling the Tapestry of Consistency in Large Vision-Language Models
Yuan Zhang (Peking University), Haoyuan Guo (ByteDance Inc)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes the ConBench benchmark to evaluate the consistency of large visual language models (LVLM) in different solution spaces, and systematically analyzes their performance through experiments.
Unveiling User Satisfaction and Creator Productivity Trade-Offs in Recommendation Platforms
Fan Yao (University of Virginia), Hongning Wang (University of Virginia)
Recommendation SystemOptimizationTabular
🎯 What it does: A Cournot Content Creation Competition (C4) model is constructed to study how the exploration intensity of recommendation algorithms affects the trade-off between the production quantity of content creators and user satisfaction, and provides the corresponding optimal exploration strategy.
Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation
Xingru Huang (Hangzhou Dianzi University), Yaoqi Sun (Hangzhou Dianzi University)
SegmentationConvolutional Neural NetworkBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A jump connection structure uC based on 2D U-Net is proposed for 3D medical image segmentation, and uC 3DU-Net is constructed accordingly.
UPS: Unified Projection Sharing for Lightweight Single-Image Super-resolution and Beyond
Kun Zhou (Chinese University of Hong Kong Shenzhen), Jiangbo Lu (SmartMore Corporation)
RestorationSuper ResolutionTransformerImage
🎯 What it does: A Unified Projection Sharing (UPS) algorithm is proposed for lightweight single-image super-resolution and other image restoration tasks.
UQ-Guided Hyperparameter Optimization for Iterative Learners
Jiesong Liu (North Carolina State University), Xipeng Shen (Renmin University of China)
OptimizationHyperparameter SearchTabularBenchmark
🎯 What it does: The UQ-guided scheme is proposed, incorporating uncertainty during the model training process into the hyperparameter optimization (HPO) workflow;
UQE: A Query Engine for Unstructured Databases
Hanjun Dai (Google DeepMind), Dale Schuurmans (University of Alberta)
Large Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper presents UQE (Unstructured Query Engine), an engine capable of executing analytical queries on databases containing unstructured data using UQL (Universal Query Language).
UrbanKGent: A Unified Large Language Model Agent Framework for Urban Knowledge Graph Construction
Yansong Ning (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextGraphChain-of-Thought
🎯 What it does: This study proposes the UrbanKGent framework, which utilizes large language models (LLMs) along with a custom instruction set, tool calls, and iterative trajectory optimization to achieve automated and low-cost construction of urban knowledge graphs (UrbanKG).
User-Creator Feature Polarization in Recommender Systems with Dual Influence
Tao Lin (Harvard University), Yang Liu (University of California)
Recommendation SystemTabular
🎯 What it does: This paper proposes a user-creator characteristic dynamics model, proving that under dual influences, the system inevitably tends toward polarization, and explores the impact of various designs on polarization and diversity.
User-item fairness tradeoffs in recommendations
Sophie Greenwood (Cornell Tech), Nikhil Garg (Cornell Tech)
Recommendation SystemOptimizationText
🎯 What it does: This paper studies the trade-off between user fairness and item fairness in recommendation systems, proposing the concept of 'fair price'. It analyzes the loss incurred when item fairness constraints are added while maximizing user fairness, and further explores how fairness constraints can amplify estimation costs in the case of preference misestimation. The authors provide two key phenomena through theoretical derivation and construct an empirical system on the arXiv paper recommendation task to validate the theoretical conclusions.
Using Noise to Infer Aspects of Simplicity Without Learning
Zachery Boner (Duke University), Cynthia Rudin (Duke University)
ClassificationRecommendation SystemAnomaly DetectionOptimizationData-Centric LearningTabularFinance Related
🎯 What it does: This study investigates the impact of noise in the data generation process, proving that noise is equivalent to implicit regularization. It quantitatively analyzes the relationship between noise levels and model simplicity (number of leaves, weight norm) as well as the simplification of the Rashomon set. Through theoretical proof and empirical validation, it demonstrates that simpler models with comparable accuracy can be obtained in the presence of noise.
Using Surrogates in Covariate-adjusted Response-adaptive Randomization Experiments with Delayed Outcomes
Lei Shi (University of California), Jingshen Wang (University of California)
Biomedical DataElectronic Health Records
🎯 What it does: This paper proposes a new design method that combines covariate-adjusted response-adaptive randomization (CARA) in situations where the primary outcome is delayed and observable surrogate outcomes are available.
Using Time-Aware Graph Neural Networks to Predict Temporal Centralities in Dynamic Graphs
Franziska Heeg (Julius-Maximilans-Universität Würzburg), Ingo Scholtes (Julius-Maximilans-Universität Würzburg)
Graph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes the use of a time-aware graph neural network (DBGNN) to predict the temporal betweenness and closeness of nodes in temporal graphs.
Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary
Zhuoyan Li (Purdue University), Ming Yin (Purdue University)
Recommendation SystemExplainability and InterpretabilityTextTabular
🎯 What it does: Construct a human behavior model to quantify how humans integrate AI recommendations and explanations in AI-assisted decision-making, and manipulate AI explanations using this model to achieve benevolent or malevolent decision guidance.
Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series
Ilan Naiman (Ben Gurion University of Negev), Omri Azencot (Ben Gurion University of Negev)
GenerationData SynthesisAnomaly DetectionDiffusion modelTime SeriesSequentialFinance RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a unified generative temporal model: first, temporal signals are mapped to two-dimensional images through reversible transformations (delay embedding or short-time Fourier transform), then a pre-existing visual diffusion model (EDM) is used to generate images, and finally, the inverse transformation is applied to recover the temporal signals, thus achieving generation, interpolation, and extrapolation of short, long, and ultra-long sequences.
UV-free Texture Generation with Denoising and Geodesic Heat Diffusion
Simone Foti (Imperial College London), Tolga Birdal (Imperial College London)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: A denoising diffusion model (UV3-TeD) has been developed to directly generate point cloud textures on 3D mesh surfaces.
Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack
Tiansheng Huang (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes an alignment defense method called Vaccine against LLM fine-tuning-as-a-service attacks, addressing the security failure issue of alignment models caused by a small amount of malicious data during user fine-tuning.
Validating Climate Models with Spherical Convolutional Wasserstein Distance
Robert C. Garrett (University of Illinois Urbana-Champaign), Bo Li (University of Illinois Urbana-Champaign)
Time Series
🎯 What it does: This paper proposes and implements the Spherical Convolution Wasserstein Distance (SCWD) to compare the distribution differences between climate models and observational/reanalysis data on the sphere, and uses this metric to assess the accuracy of CMIP5/6 historical simulations.
Value-Based Deep Multi-Agent Reinforcement Learning with Dynamic Sparse Training
Pihe Hu (Tsinghua University), Longbo Huang (Tsinghua University)
Computational EfficiencyReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes a multi-agent sparse training framework called MAST, which addresses the training issues of value-based deep multi-agent reinforcement learning (MARL) under sparse networks, significantly reducing the computational load and model size for training and inference.
Variance estimation in compound decision theory under boundedness
Subhodh Kotekal (University of Chicago)
🎯 What it does: Under the theory of composite decision-making, the estimation of Gaussian noise variance under mean constraints is studied.
Variational Delayed Policy Optimization
Qingyuan Wu (University of Southampton), Chao Huang (University of Southampton)
OptimizationTransformerReinforcement LearningAgentic AITabular
🎯 What it does: A Variational Delayed Policy Optimization (VDPO) is proposed to address the reinforcement learning problem under observation delays.
Variational Distillation of Diffusion Policies into Mixture of Experts
Hongyi Zhou (Karlsruhe Institute of Technology), Rudolf Lioutikov (Karlsruhe Institute of Technology)
Knowledge DistillationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerMixture of ExpertsDiffusion modelScore-based ModelMultimodality
🎯 What it does: Proposes Variational Diffusion Distillation (VDD), which compresses the pre-trained diffusion strategy into a Mixture of Experts (MoE) form, preserving the multimodal representation of the diffusion model while achieving faster inference and computable likelihood;
Variational Flow Matching for Graph Generation
Floor Eijkelboom (University of Amsterdam), Jan-Willem van de Meent (University of Amsterdam)
GenerationData SynthesisGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: A Variational Flow Matching (VFM) framework is proposed, and based on this, a CatFlow method for discrete graph data is designed, which can train continuous regularized flows without explicitly solving ODEs.
Variational Multi-scale Representation for Estimating Uncertainty in 3D Gaussian Splatting
Ruiqi Li (Hong Kong Baptist University), Yiu-ming Cheung (Hong Kong Baptist University)
OptimizationComputational EfficiencyRepresentation LearningGaussian SplattingPoint Cloud
🎯 What it does: A 3D Gaussian Spray (3DGS) uncertainty estimation method based on variational multi-scale representation is proposed, which can provide predictive confidence during rendering and automatically eliminate noise Gaussians.
VASA-1: Lifelike Audio-Driven Talking Faces Generated in Real Time
Sicheng Xu (Microsoft Research Asia), Baining Guo (Microsoft Research Asia)
GenerationData SynthesisTransformerDiffusion modelVideoMultimodalityAudio
🎯 What it does: The VASA-1 framework is proposed to generate high-quality, real-time talking head videos with synchronized lip movements and rich, natural head motions from a single portrait and audio speech.
VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks
Yang Li (Georgia State University), Shihao Ji (University of Connecticut)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes VB-LoRA, an extremely parameter-efficient fine-tuning method implemented through a vector pool, balancing storage and performance.
VCR-GauS: View Consistent Depth-Normal Regularizer for Gaussian Surface Reconstruction
Hanlin Chen (National University of Singapore), Gim Hee Lee (National University of Singapore)
SegmentationGenerationDepth EstimationGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes VCR-GauS, which achieves high-quality surface reconstruction of 3D Gaussian Splatting using perspective-consistent depth-normal regularization.
Vector Quantization Prompting for Continual Learning
Li Jiao (Communication University of China), Qiang Xu (Chinese University of Hong Kong)
ClassificationRecognitionTransformerPrompt EngineeringImage
🎯 What it does: In the context of class-incremental continual learning, the VQ-Prompt method is proposed, which achieves the discretization and selection of task knowledge through a discrete prompt pool on a fixed pre-trained ViT model.
VeLoRA: Memory Efficient Training using Rank-1 Sub-Token Projections
Roy Miles (Huawei), Jiankang Deng
CompressionComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: We propose VeLoRA, which utilizes fixed one-dimensional subspace for grouping projection compression of forward activations and roughly reconstructs them during backpropagation, significantly reducing the storage requirements for intermediate activations during large model training.
Verifiably Robust Conformal Prediction
Linus Jeary (King's College London), Nicola Paoletti (King's College London)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A verifiable robust conformal prediction framework (VRCP) based on neural network verification is proposed, which can maintain coverage guarantees of the prediction set under adversarial perturbations.
Verified Code Transpilation with LLMs
Sahil Bhatia (University of California Berkeley), Alvin Cheung (University of California Berkeley)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A verification enhancement framework based on large language models, LLMLIFT, has been developed, which can automatically convert source language programs into target DSLs and generate formal equivalence proofs.