NeurIPS 2024 Papers — Page 15
Conference on Neural Information Processing Systems · 4035 papers
FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention
Yu Lu (Zhejiang University), Yi Yang (University of Technology Sydney)
GenerationData SynthesisDiffusion modelVideoTextBenchmark
🎯 What it does: We propose FreeLong, a training-free method that generates long videos (e.g., 128 frames or longer) from existing short video diffusion models (e.g., 16 frames) without the need for retraining.
FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes
Yunsong Wang (National University of Singapore), Gim Hee Lee (National University of Singapore)
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes FreeSplat, a general-purpose 3D Gaussian scattering method that can reconstruct global 3D Gaussians from indoor perspective sequences of arbitrary length and achieve free-viewpoint rendering.
FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge
Hanzhe LI, Junyu Dong (Ocean University of China)
ClassificationGenerationGenerative Adversarial NetworkImageVideo
🎯 What it does: A method for generating fake faces in the frequency domain using mixed frequency knowledge (FreqBlender) is proposed.
FreqMark: Invisible Image Watermarking via Frequency Based Optimization in Latent Space
YiYang Guo, Shangfei Wang (Tencent)
RestorationData SynthesisOptimizationAuto EncoderImage
🎯 What it does: A self-supervised image watermarking method named FreqMark is developed, which embeds secret information by optimizing perturbations in the frequency domain of the pre-trained VAE latent space.
Frequency Adaptive Normalization For Non-stationary Time Series Forecasting
Weiwei Ye (Central South University), Ning Gui (Central South University)
Time Series
🎯 What it does: This paper proposes a new instance normalization method called Frequency Adaptive Normalization (FAN) to address trend and seasonality issues in non-stationary time series.
Frequency-aware Generative Models for Multivariate Time Series Imputation
Xinyu Yang (Nankai University), Xinyang Chen (Harbin Institute of Technology)
GenerationData SynthesisAnomaly DetectionTransformerDiffusion modelTime SeriesSequentialBiomedical Data
🎯 What it does: This study focuses on the imputation of missing values in multivariate time series and proposes the FGTI model that incorporates frequency domain information into generative models.
Freya PAGE: First Optimal Time Complexity for Large-Scale Nonconvex Finite-Sum Optimization with Heterogeneous Asynchronous Computations
Alexander Tyurin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationTabular
🎯 What it does: We propose Freya PAGE, an asynchronous parallel algorithm designed for non-convex finite optimization problems, which can adaptively handle the speed heterogeneity of worker nodes and ignore computations from slow nodes.
Frieren: Efficient Video-to-Audio Generation Network with Rectified Flow Matching
Yongqi Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationData SynthesisTransformerRectified FlowVideoMultimodalityOrdinary Differential EquationAudio
🎯 What it does: This paper proposes a video-to-audio generation model called FRIEREN based on rectified flow matching, which can directly generate high-quality audio synchronized with visuals from silent videos.
From an Image to a Scene: Learning to Imagine the World from a Million 360° Videos
Matthew Wallingford (University of Washington), Ali Farhadi (University of Washington)
GenerationData SynthesisDiffusion modelOptical FlowVideo
🎯 What it does: A large-scale multi-view dataset 360-1M is constructed by collecting 1 million 360-degree videos, and ODIN based on diffusion models is trained to freely generate images of 3D scenes from a single image from any viewpoint.
From Biased to Unbiased Dynamics: An Infinitesimal Generator Approach
Timothée Devergne (Istituto Italiano di Tecnologia), Massimiliano Pontil (University College London)
Time SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: This study proposes a method to learn the characteristic functions and eigenvalues of dynamical systems from biased simulation data using an unbiased infinitesimal generator (through its inverse operator).
From Causal to Concept-Based Representation Learning
Goutham Rajendran (Carnegie Mellon University), Pradeep Kumar Ravikumar
Representation LearningTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a shift from causal representation learning to a concept-based representation learning framework, formally defining concepts as linear subspaces in latent space, and providing theoretical guarantees for identifiability under finite conditional distributions. A contrastive learning algorithm is also designed, which is validated through experiments on synthetic data, CLIP image embeddings, and large language model alignment.
From Chaos to Clarity: 3DGS in the Dark
Zhihao Li (Nanyang Technology University), Bihan Wen (Nanyang Technology University)
RestorationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: A self-supervised framework is proposed, utilizing a noise extractor and noise-robust reconstruction loss, capable of recovering high dynamic range 3D Gaussian Splatting (3DGS) with only a small number of noisy raw images.
From Dictionary to Tensor: A Scalable Multi-View Subspace Clustering Framework with Triple Information Enhancement
Zhibin Gu (Hebei Normal University), Songhe Feng (Beijing Jiaotong University)
OptimizationComputational EfficiencyImage
🎯 What it does: A scalable multi-view subspace clustering framework STONE is proposed, which utilizes enhanced anchor dictionary, hyperbolic tangent rank constraint, and anchor hypergraph Laplacian regularization to achieve triple information enhancement from dictionary to tensor.
From Instance Training to Instruction Learning: Task Adapters Generation from Instructions
Huanxuan Liao (Chinese Academy of Sciences), Jun Zhao (Chinese Academy of Sciences)
Knowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes a task-specific adapter generation through instruction learning (TAGI), enabling the construction of task-specific models without instance training, simulating how humans learn skills through understanding instructions.
From Linear to Linearizable Optimization: A Novel Framework with Applications to Stationary and Non-stationary DR-submodular Optimization
Mohammad Pedramfar (McGill University and Mila), Vaneet Aggarwal (Purdue University)
Optimization
🎯 What it does: This paper introduces the concept of upper linearizable/quadratic functions and constructs a general meta-algorithm to transform linear/quadratic optimization algorithms into optimization algorithms for upper linearizable functions; it is applied to online and offline optimization of DR-submodular and upper convex functions.
From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection
Xinlei Wang (University of Sydney), Junhua Zhao (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTime SeriesFinance RelatedChain-of-Thought
🎯 What it does: A framework is proposed that combines large language models with reasoning agents to predict news events by integrating them with time series data;
From Similarity to Superiority: Channel Clustering for Time Series Forecasting
Jialin Chen (Yale University), Rex Ying (Yale University)
Time SeriesFinance Related
🎯 What it does: Designed and implemented the Channel Clustering Module (CCM), which manages channel clustering for multi-channel time series prediction by clustering channels and learning prototypes.
From Text to Trajectory: Exploring Complex Constraint Representation and Decomposition in Safe Reinforcement Learning
Pusen Dong (Beihang University), Jianxin Li (Beihang University)
TransformerReinforcement LearningContrastive LearningTextMultimodality
🎯 What it does: The Trajectory-level Textual Constraints Translator (TTCT) method is proposed, which maps trajectory-level textual constraints to cost functions, addressing the issues of constraint identification and cost sparsity in safe reinforcement learning.
From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $\alpha$-NeuS
Haoran Zhang (Institute of Software, Chinese Academy of Sciences), Ying He (Nanyang Technological University)
GenerationData SynthesisNeural Radiance FieldPoint CloudMeshBenchmark
🎯 What it does: A new α-NeuS method is proposed, achieving simultaneous reconstruction of transparent and opaque objects.
From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models
Zhuoshi Pan (Tsinghua University), Sijia Liu (Michigan State University)
GenerationAdversarial AttackData-Centric LearningDiffusion modelImage
🎯 What it does: This study investigates the use of BadNets-style data poisoning attacks in diffusion models, revealing the bidirectional poisoning effects (trigger amplification and phase shift) and their impact on generation quality and classification robustness.
From Unstructured Data to In-Context Learning: Exploring What Tasks Can Be Learned and When
Kevin Christian Wibisono (University of Michigan), Yixin Wang (University of Michigan)
TransformerLarge Language ModelText
🎯 What it does: This study investigates how LLM achieves in-context learning (ICL) without structured pre-training data and reveals the underlying mechanisms.
Frozen-DETR: Enhancing DETR with Image Understanding from Frozen Foundation Models
Shenghao Fu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: This paper proposes using frozen visual foundation models (such as CLIP) as feature enhancers to improve query-based object detectors (DETR series) by using class tokens as image queries and patch tokens for encoder feature fusion, thereby enhancing detection performance.
Frustratingly Easy Test-Time Adaptation of Vision-Language Models
Matteo Farina (University of Trento), Elisa Ricci (Fondazione Bruno Kessler)
ClassificationDomain AdaptationTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a zero-temperature testing adaptation method called ZERO, which achieves instant enhancement of visual-language models through single forward propagation.
FSP-Laplace: Function-Space Priors for the Laplace Approximation in Bayesian Deep Learning
Tristan Cinquin (University of Tübingen), Robert Bamler (University of Tübingen)
OptimizationImageTime Series
🎯 What it does: A deep Bayesian neural network method using Gaussian process priors in function space and performing Laplace approximation (FSP-LAPLACE) is proposed.
FUG: Feature-Universal Graph Contrastive Pre-training for Graphs with Diverse Node Features
Jitao Zhao (Tianjin University), Zhiyong Feng (National University of Singapore)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A Feature-Universal Graph (FUG) contrastive pre-training strategy is proposed, enabling graph neural networks to directly transfer on graph data with different node feature shapes without the need to rebuild models or preprocess features.
FUGAL: Feature-fortified Unrestricted Graph Alignment
Aditya Bommakanti (Indian Institute of Technology Delhi), Panagiotis Karras (University of Copenhagen)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes FUGAL, a mediator-free graph alignment method that operates directly on the adjacency matrix. It relaxes the permutation matrix into a doubly stochastic matrix, uses Frank-Wolfe and Sinkhorn-Knopp iterations for solving, incorporates a linear alignment regularization based on structural features into the objective, and finally obtains the final matching through the Hungarian algorithm.
Full-Atom Peptide Design with Geometric Latent Diffusion
Xiangzhe Kong (Tsinghua University), Yang Liu (Tsinghua University)
Drug DiscoveryProtein Structure PredictionDiffusion modelAuto EncoderBiomedical Data
🎯 What it does: This paper proposes a geometric latent diffusion model called PepGLAD for full-atom peptide design given binding sites.
Full-Distance Evasion of Pedestrian Detectors in the Physical World
Zhi Cheng (Tsinghua University), Xiaolin Hu (Tsinghua University)
Object DetectionAdversarial AttackConvolutional Neural NetworkImagePhysics Related
🎯 What it does: A full-distance attack (FDA) method is proposed to generate effective adversarial patterns against pedestrian detectors in the physical world.
Fully Distributed, Flexible Compositional Visual Representations via Soft Tensor Products
Bethia Sun (University of New South Wales Sydney), Yang Song (University of New South Wales Sydney)
Representation LearningAuto EncoderImage
🎯 What it does: This paper proposes a fully distributed and flexible Soft TPR (Soft Tensor Product Representation) and the corresponding Soft TPR Autoencoder for learning distributed representations of decomposable components in visual data.
Fully Explicit Dynamic Gaussian Splatting
Junoh Lee (Gwangju Institute of Science and Technology), Hae-Gon Jeon (Gwangju Institute of Science and Technology)
GenerationData SynthesisOptimizationGaussian SplattingVideo
🎯 What it does: This paper proposes Explicit 4D Gaussian Splatting (Ex4DGS), which explicitly models dynamic 3D Gaussians through keyframe interpolation and separates static and dynamic Gaussians during training, achieving high-quality, real-time dynamic viewpoint synthesis.
Fully Unconstrained Online Learning
Ashok Cutkosky (Boston University), Zakaria Mhammedi (Google Research)
Optimization
🎯 What it does: This paper proposes a completely unconstrained, parameter-free online convex optimization algorithm that achieves nearly optimal sublinear cumulative loss upper bounds without knowing the gradient upper bound G or the norm of the comparison point ‖w*‖.
Functional Bilevel Optimization for Machine Learning
Ieva Petrulionytė, Michael Arbel (Univ Grenoble Alpes)
OptimizationReinforcement LearningImageTabular
🎯 What it does: A new paradigm for bi-level optimization is proposed in function space rather than parameter space, along with the corresponding function implicit differentiation theory and an implementable algorithm (FuncID).
Functional Gradient Flows for Constrained Sampling
Shiyue Zhang (Peking University), Cheng Zhang (Peking University)
OptimizationTabularStochastic Differential Equation
🎯 What it does: A constrained particle variational inference method based on neural networks for piecewise velocity fields (CFG) is proposed, which can sample within any shape of constrained domains.
Functionally Constrained Algorithm Solves Convex Simple Bilevel Problem
Huaqing Zhang (Tsinghua University), Jingzhao Zhang (Tsinghua University)
Optimization
🎯 What it does: This paper studies convex bilevel optimization problems, proving that any zero-respect first-order method cannot achieve an absolute optimal solution, and proposes a new algorithm called FC-BiO, which can achieve near-optimal iteration complexity under weak optimality error.
Fundamental Convergence Analysis of Sharpness-Aware Minimization
Pham Duy Khanh (Ho Chi Minh City University of Education), Dat Ba Tran (Wayne State University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the basic convergence properties of Sharpness-Aware Minimization (SAM) and its normalized variants in convex and non-convex optimization, and provides a comprehensive convergence analysis for the unnormalized USAM; it then validates its effectiveness in practical deep learning through experiments.
Fundamental Limits of Prompt Compression: A Rate-Distortion Framework for Black-Box Language Models
Alliot Nagle (University of Texas at Austin), Hyeji Kim (University of Texas at Austin)
CompressionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A theoretical framework for prompt compression of black-box large language models is proposed, deriving its rate-distortion function and experimentally validating the gap with theoretical limits; at the same time, an Adaptive QuerySelect algorithm that is query-aware and variable-rate is introduced.
FUSE: Fast Unified Simulation and Estimation for PDEs
Levi E. Lingsch (ETH Zurich), Georgios Kissas (ETH Zurich)
Flow-based ModelTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A unified framework FUSE is proposed, which integrates forward simulation of PDEs and backward parameter estimation, allowing for simultaneous simulation and parameter inference after a one-time training.
FuseAnyPart: Diffusion-Driven Facial Parts Swapping via Multiple Reference Images
Zheng Yu (Shanghai Jiao Tong University), Senzhang Wang (Central South University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: FuseAnyPart is a facial component swapping method based on diffusion models, capable of seamlessly blending components such as eyes, nose, and mouth from multiple source images onto a target face to generate high-quality, natural synthetic faces.
FuseFL: One-Shot Federated Learning through the Lens of Causality with Progressive Model Fusion
Zhenheng Tang (Hong Kong Baptist University), Xiaowen Chu (Hong Kong University of Science and Technology)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: A one-round communication federated learning method called FuseFL is proposed, which reduces communication costs and enhances model generalization ability by layer-wise fusion of intermediate features from client models.
FuseMoE: Mixture-of-Experts Transformers for Fleximodal Fusion
Xing Han (Johns Hopkins University), Suchi Saria (Johns Hopkins University)
TransformerMixture of ExpertsMultimodalityBiomedical DataElectronic Health Records
🎯 What it does: FuseMoE is a sparse expert mixture Transformer model designed for FlexiModal data, capable of efficient fusion and prediction in the presence of missing modalities and irregular temporal sampling.
G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering
Xiaoxin He (National University of Singapore), Bryan Hooi (National University of Singapore)
GenerationRetrievalExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: A question-answering benchmark for real text graphs has been constructed, and the G-Retriever framework has been proposed to support dialogue with graphs.
G2D: From Global to Dense Radiography Representation Learning via Vision-Language Pre-training
Che Liu (Imperial College London), Rossella Arcucci (Imperial College London)
Object DetectionSegmentationRepresentation LearningConvolutional Neural NetworkVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: G2D proposes a medical vision-language pre-training framework that can learn both global and pixel-level visual representations simultaneously.
G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models
Pengyue Jia (City University of Hong Kong), Dawei Yin (Baidu Inc.)
RetrievalTransformerVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a global image geolocation framework named G3, which utilizes a retrieval-enhanced generation method to achieve more accurate coordinate predictions.
GACL: Exemplar-Free Generalized Analytic Continual Learning
Huiping Zhuang (South China University of Technology), Cen Chen (Shenzhen Institute, Hunan University)
ClassificationRecognitionTransformerImage
🎯 What it does: A novel example-free generalized analysis continuous learning method (GACL) is proposed, capable of completing the GCIL task without storing historical samples.
GAMap: Zero-Shot Object Goal Navigation with Multi-Scale Geometric-Affordance Guidance
Shuaihang Yuan (New York University Abu Dhabi), Yi Fang (New York University Abu Dhabi)
Robotic IntelligenceLarge Language ModelVision Language ModelPoint CloudBenchmark
🎯 What it does: This paper proposes a zero-shot target navigation method called GAMap, which constructs a map using geometric components and available attributes, and guides the robot's exploration through multi-scale similarity scoring.
GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation
Haoran Lu (Peking University), Hao Dong (Peking University)
Robotic IntelligenceReinforcement LearningMeshBenchmark
🎯 What it does: A unified GarmentLab environment and benchmark have been established for realistic and diverse experiments in garment and flexible object manipulation.
Gated Inference Network: Inference and Learning State-Space Models
Hamidreza Hashempoor (Seoul National University), Wan Choi (Seoul National University)
Autonomous DrivingOptimizationRecurrent Neural NetworkAuto EncoderVideo
🎯 What it does: This paper proposes the Gated Inference Network (GIN), a scalable Bayesian filtering-smoothing algorithm based on GRU, for approximate inference and learning of state space models in high-dimensional noisy videos.
Gated Slot Attention for Efficient Linear-Time Sequence Modeling
Yu Zhang (Soochow University), Guohong Fu (Soochow University)
Recurrent Neural NetworkTransformerSupervised Fine-TuningTextSequential
🎯 What it does: A Gated Slot Attention (GSA) mechanism is proposed, incorporating gating based on ABC to balance memory forgetting and recent bias, improving the performance of linear attention models on long sequences and memory-intensive tasks while maintaining a small state size.
Gaussian Approximation and Multiplier Bootstrap for Polyak-Ruppert Averaged Linear Stochastic Approximation with Applications to TD Learning
Sergey Samsonov (Higher School of Economics), Alexey Naumov (Higher School of Economics)
Reinforcement LearningTime Series
🎯 What it does: This paper studies the Berry-Esseen bound of the Polyak-Ruppert average iteration of the Linear Stochastic Approximation (LSA) algorithm and proves the non-asymptotic validity of the confidence interval for LSA parameter estimation based on the multiplier bootstrap method.
Gaussian Graph Network: Learning Efficient and Generalizable Gaussian Representations from Multi-view Images
Shengjun Zhang (Tsinghua University), Yueqi Duan (Tsinghua University)
RestorationGenerationData SynthesisGraph Neural NetworkGaussian SplattingImage
🎯 What it does: This paper proposes the Gaussian Graph Network (GGN), which generates efficient and generalizable Gaussian representations by constructing Gaussian graphs and implementing graph networks at the Gaussian level, thereby improving the quality of multi-view novel view synthesis.
Gaussian Process Bandits for Top-k Recommendations
Mohit Yadav (University of Massachusetts Amherst), Daniel Sheldon (University of Massachusetts Amherst)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: This paper proposes a context Bandit algorithm GP-TopK based on Gaussian processes and the Kendall kernel for optimal ranking of Top-k recommendations in a full feedback environment.
GaussianCube: A Structured and Explicit Radiance Representation for 3D Generative Modeling
Bowen Zhang (University of Science and Technology of China), Baining Guo (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: We propose GaussianCube, a structured explicit radiance representation that fits and maps a fixed number of high-quality Gaussians to a voxel grid through density constraints for 3D generation tasks.
GaussianCut: Interactive segmentation via graph cut for 3D Gaussian Splatting
Umangi Jain (University of Toronto), Igor Gilitschenski (University of Toronto)
SegmentationGaussian SplattingVideoPoint Cloud
🎯 What it does: Implement interactive multi-view segmentation on existing 3D Gaussian Splatting, allowing users to quickly specify target objects through points, doodles, or text from any view and obtain high-quality 3D segmentation of foreground/background;
GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting
Xiufeng Huang (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
Data SynthesisSafty and PrivacyKnowledge DistillationGaussian SplattingPoint Cloud
🎯 What it does: A covert digital watermarking method for the 3D Gaussian Splatting model is proposed, which embeds copyright information in the 3D Gaussian parameters while maintaining rendering quality, and can recover this information from the parameters or rendered images.
GAVEL: Generating Games via Evolution and Language Models
Graham Todd (New York University), Julian Togelius (New York University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Automatically generate novel playable games using large language models and evolutionary algorithms on the Ludii game description language.
GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning
Guibin Zhang (Tongji University), Kun Wang (NTU)
Graph Neural NetworkGraph
🎯 What it does: A soft data pruning method GDeR based on prototype learning is proposed, which dynamically maintains the training subset to improve the training efficiency, balance, and robustness of graph neural networks.
GenArtist: Multimodal LLM as an Agent for Unified Image Generation and Editing
Zhenyu Wang (Tsinghua University), Xihui Liu (University of Hong Kong)
Image TranslationGenerationTransformerLarge Language ModelAgentic AIVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: We propose GenArtist, a unified image generation and editing system that utilizes a multimodal large language model (MLLM) agent to decompose user instructions, plan tree-structured execution, and perform self-checking and correction, calling various generation and editing tools to complete tasks.
Gene-Gene Relationship Modeling Based on Genetic Evidence for Single-Cell RNA-Seq Data Imputation
Daeho Um (Samsung Advanced Institute of Technology), Yunha Yeo (Korea University)
OptimizationData-Centric LearningBiomedical Data
🎯 What it does: A framework named scCR is proposed for imputing missing values in scRNA-seq data, utilizing the associations and ablation relationships between genes to complete the cell-gene matrix.
General Articulated Objects Manipulation in Real Images via Part-Aware Diffusion Process
Zhou FANG, Cewu Lu (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The PA-Diffusion model is proposed for editing joint objects in real images based on text or human interaction.
General bounds on the quality of Bayesian coresets
Trevor Campbell (University of British Columbia)
🎯 What it does: This paper studies the theoretical error of Bayesian coreset approximating posterior distributions, providing general upper and lower bounds for Kullback-Leibler divergence, and applying these results to two commonly used coreset construction methods.
General Detection-based Text Line Recognition
Raphael Baena (Ecole des Ponts), Mathieu Aubry (Ecole des Ponts)
RecognitionTransformerText
🎯 What it does: This paper proposes a detection-based text line recognition method (DTLR), which achieves universal line text recognition for different writing systems (Latin letters, Chinese characters, and ciphers) through large-scale synthetic data pre-training and line-level fine-tuning on a Transformer detector.
Generalizable and Animatable Gaussian Head Avatar
Xuangeng Chu (University of Tokyo), Tatsuya Harada (University of Tokyo)
RestorationGenerationGaussian SplattingImageVideo
🎯 What it does: A real-time replayable full-body 3D Gaussian model based on a single image is proposed, achieving unidirectional forward head reconstruction and expression control.
Generalizable Implicit Motion Modeling for Video Frame Interpolation
Zujin Guo (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
Image TranslationRestorationOptical FlowVideo
🎯 What it does: A general implicit motion modeling framework GIMM is proposed, which utilizes a pre-trained optical flow estimator's bidirectional optical flow and coordinate network to predict bidirectional optical flow at any time step, thereby achieving video frame interpolation.
Generalizable Person Re-identification via Balancing Alignment and Uniformity
Yoonki Cho (Korea Advanced Institute of Science and Technology), Sung-eui Yoon
RecognitionRetrievalDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This study addresses and resolves the bias effect caused by data augmentation in domain generalization for person re-identification, proposing the BAU framework that maintains alignment and uniformity in the representation space.
Generalizablity of Memorization Neural Network
Lijia Yu (Chinese Academy of Sciences), Yibo Miao (Chinese Academy of Sciences)
Tabular
🎯 What it does: This paper provides the first theoretical analysis of the generalization ability of memory neural networks, exploring the expressive power of neural networks on interpolating finite datasets.
Generalization Analysis for Label-Specific Representation Learning
Yifan Zhang, Min-Ling Zhang (Southeast University)
Representation Learning
🎯 What it does: This paper proposes a theoretical generalization analysis framework for label-specific representation learning (LSRL) and provides the corresponding risk upper bound.
Generalization Bound and Learning Methods for Data-Driven Projections in Linear Programming
Shinsaku Sakaue (University of Tokyo), Taihei Oki (Hokkaido University)
OptimizationTabular
🎯 What it does: This paper studies a data-driven projection method that uses a trained projection matrix to reduce the dimensionality of high-dimensional linear programming for solving, and then maps the solution back to the original space, improving the efficiency and quality of the solution.
Generalization Bounds via Conditional $f$-Information
Ziqiao Wang (Tongji University), Yongyi Mao (University of Ottawa)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A general upper bound on the generalized error is provided under the conditional f-information framework.
Generalization Error Bounds for Two-stage Recommender Systems with Tree Structure
Jin Zhang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Recommendation SystemScore-based ModelTabular
🎯 What it does: This paper studies the upper bound of generalization error for tree-structured two-stage recommendation systems.
Generalization of Hamiltonian algorithms
Andreas Maurer (Italian Institute of Technology)
🎯 What it does: This paper proposes a general method for deriving upper bounds on the generalization error of stochastic learning algorithms, particularly those in Hamiltonian form, and provides specific applications in the Gibbs algorithm, randomized stability algorithms, and PAC-Bayes theory with data-dependent priors.
Generalize or Detect? Towards Robust Semantic Segmentation Under Multiple Distribution Shifts
Zhitong Gao (ShanghaiTech University), Xuming He (ShanghaiTech University)
SegmentationDomain AdaptationAnomaly DetectionImage
🎯 What it does: By jointly generating data with semantic shift and domain shift, and employing a two-stage noise adaptive training, the detection and generalization capabilities of semantic segmentation under multiple distribution shifts are enhanced.
Generalized Eigenvalue Problems with Generative Priors
Zhaoqiang Liu (University of Electronic Science and Technology of China), Junren Chen (University of Hong Kong)
OptimizationImage
🎯 What it does: This paper proposes a method to solve the Generalized Generalized Eigenvalue Problem (GGEP) under the prior of generative models, providing the statistical convergence rate of the global optimal solution and an iterative algorithm called PRFM.
Generalized Fast Exact Conformalization
Diyang Li (Cornell University)
TabularOrdinary Differential Equation
🎯 What it does: A general fast and accurate regularization method is proposed, utilizing the piecewise smooth structure of the solution path to obtain a complete prediction interval by solving first-order ordinary differential equations, significantly reducing computational costs.
Generalized Linear Bandits with Limited Adaptivity
Ayush Sawarni (Stanford University), Gaurav Sinha (Microsoft Research India)
Tabular
🎯 What it does: Under finite adaptive constraints, the bandit problem in a generalized linear context is studied, and two algorithms, B-GLinCB (M1 setting) and RS-GLinCB (M2 setting), are proposed;
Generalized Protein Pocket Generation with Prior-Informed Flow Matching
ZAIXI ZHANG, Qi Liu (University of Science and Technology of China)
GenerationDrug DiscoveryProtein Structure PredictionTransformerFlow-based ModelMultimodalityBiomedical Data
🎯 What it does: A protein-ligand pocket generation model called PocketFlow is proposed and implemented, which can generate structurally reasonable and high-affinity protein pockets under the conditions of a given protein framework and different types of ligands (small molecules, peptides, RNA).
Generalized Tensor Decomposition for Understanding Multi-Output Regression under Combinatorial Shifts
Andong Wang (RIKEN Advanced Intelligence Project), Qibin Zhao (RIKEN Advanced Intelligence Project)
Domain AdaptationOptimizationTabular
🎯 What it does: This study investigates the issue of combination distribution shift (CDS) in multi-output regression caused by the inability of the training distribution to cover all input feature combinations. It proposes the Functional t-SVD (Ft-SVD) theoretical framework and, based on this, models multi-output regression as a missing non-random (MNAR) tensor completion problem. Furthermore, it designs a dual-stage empirical risk minimization (ERM-DS) algorithm and provides theoretical generalization guarantees.
Generalizing CNNs to graphs with learnable neighborhood quantization
Isaac Osafo Nkansah (Weill Cornell Medicine), Logan Grosenick (Weill Cornell Medicine)
Graph Neural NetworkImageGraph
🎯 What it does: Proposed Quantized Graph Convolution Networks (QGCN) and Quantized Graph Residual Networks (QGRN), which strictly generalize CNN convolution layers to graph data and achieve generalization through learnable neighborhood quantization.
Generalizing Consistency Policy to Visual RL with Prioritized Proximal Experience Regularization
Haoran Li (Institute of Automation Chinese Academy of Sciences University of Chinese Academy of Sciences), Dongbin Zhao (Institute of Automation Chinese Academy of Sciences University of Chinese Academy of Sciences)
Reinforcement LearningImage
🎯 What it does: In visual reinforcement learning, a consistency model-based strategy with prioritized proximal experience regularization (CP3ER) is proposed to address the policy collapse issue caused by traditional Q-loss.
Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling
Wanghan Xu (Shanghai Jiao Tong University), LEI BAI
TransformerTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A hybrid model called WeatherGFT, which integrates physical equations and neural networks, has been developed to achieve finer-grained (30-minute) forecasts on hourly data.
Generate Universal Adversarial Perturbations for Few-Shot Learning
Yiman Hu (Huazhong University of Science and Technology), Yuhua Li (Huazhong University of Science and Technology)
ClassificationAdversarial AttackMeta LearningConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: In the context of few-shot learning (FSL), this paper proposes a unified attack framework called FSAFW, which generates universal adversarial perturbations (UAP) that can generalize across various FSL training paradigms (fine-tuning, meta-learning, metric learning).
Generated and Pseudo Content guided Prototype Refinement for Few-shot Point Cloud Segmentation
Lili Wei (Beijing Jiaotong University), Jun Liu (Lancaster University)
SegmentationKnowledge DistillationLarge Language ModelContrastive LearningPoint Cloud
🎯 What it does: A prototype-based few-shot 3D point cloud semantic segmentation framework GPCPR is proposed, which enhances prototype quality using text content generated by LLM and pseudo-query context;
Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search
Nicola Dainese (Aalto University), Pekka Marttinen (Aalto University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the use of large language models (LLM) to generate world models in the form of Python code (Code World Models, CWM), and implements self-debugging and iterative generation through the GIF-MCTS method;
Generating compositional scenes via Text-to-image RGBA Instance Generation
Alessandro Fontanella (University of Edinburgh), Sarah Parisot (Microsoft Research)
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: A multi-layer RGBA object generation and noise mixing synthesis method is proposed, achieving fine-grained control over object attributes and layout in text-generated images, as well as interactive scene stitching.
Generating Highly Designable Proteins with Geometric Algebra Flow Matching
Simon Wagner (Heidelberg Institute for Theoretical Studies), Jan Stuehmer
Protein Structure PredictionFlow-based ModelBiomedical Data
🎯 What it does: A protein backbone generation model based on geometric algebra, GAFL, is proposed, which combines the Flow Matching framework and improves AlphaFold2's Invariant Point Attention (IPA) to Clifford Frame Attention (CFA), achieving more geometrically expressive message passing.
Generating Origin-Destination Matrices in Neural Spatial Interaction Models
Ioannis Zachos (Cambridge University), Theodoros Damoulas (University of Warwick)
GenerationOptimizationTabularStochastic Differential Equation
🎯 What it does: The GeNSIT framework is proposed, which utilizes neural networks and the Harris-Winston SDE to simultaneously learn spatial interaction models at both continuous and discrete levels, directly sampling and recovering the real OD matrix in the discrete OD matrix space.
Generative Adversarial Model-Based Optimization via Source Critic Regularization
Michael S Yao, Osbert Bastani (University of Pennsylvania)
OptimizationDrug DiscoveryAuto EncoderGenerative Adversarial NetworkTabular
🎯 What it does: A generative adversarial optimization framework using adaptive source critic regularization (aSCR) is proposed for offline model optimization, preventing out-of-domain predictions caused by model overfitting in the optimization trajectory.
Generative Forests
Richard Nock (Google Research), Mathieu Guillame-Bert (Google)
GenerationData SynthesisTabular
🎯 What it does: A generative model based on a decision tree ensemble is proposed—Generative Forests (GF), along with a simple training algorithm GF.BOOST that has convergence guarantees.
Generative Fractional Diffusion Models
Gabriel Nobis (Fraunhofer HHI), Wojciech Samek (Fraunhofer HHI)
GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A continuous-time fractional generative model (GFDM) driven by Markovian approximated fractional Brownian motion is proposed, and its reverse-time process is derived.
Generative Hierarchical Materials Search
Sherry Yang (Google DeepMind), Ekin Dogus Cubuk (Google DeepMind)
GenerationOptimizationGraph Neural NetworkTransformerLarge Language ModelDiffusion modelTextRetrieval-Augmented Generation
🎯 What it does: A layered generative system (GenMS) has been designed to convert natural language instructions into crystal structures that meet multiple objectives (synthesis feasibility, structural legality, low energy).
Generative Modeling of Molecular Dynamics Trajectories
Bowen Jing (Massachusetts Institute of Technology), Bonnie Berger (Massachusetts Institute of Technology)
GenerationData SynthesisDrug DiscoveryTransformerTime SeriesSequentialStochastic Differential Equation
🎯 What it does: MDGEN is proposed, a molecular dynamics trajectory simulation framework based on generative models, supporting forward simulation, interpolation, upsampling, and filling incomplete trajectories;
Generative Modelling of Structurally Constrained Graphs
Manuel Madeira (École Polytechnique Fédérale de Lausanne), Pascal Frossard (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: Proposes the ConStruct framework, which utilizes a graph discrete diffusion model to enforce hard structural constraints (such as planar graphs, acyclic graphs, etc.) during the generation process while maintaining consistency between training and sampling distributions.
Generative Retrieval Meets Multi-Graded Relevance
Yubao Tang (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
RetrievalTransformerAuto EncoderContrastive LearningText
🎯 What it does: A generative retrieval framework GR 2 is proposed for multi-level relevance, including document ID generation and multi-level constraint contrastive learning.
Generative Semi-supervised Graph Anomaly Detection
Hezhe Qiao (Singapore Management University), Guansong Pang (Singapore Management University)
Anomaly DetectionGraph Neural NetworkGraphFinance Related
🎯 What it does: In the context of semi-supervised graph anomaly detection, a learning framework called GGAD is proposed, which is based on generating pseudo-anomalous nodes to enhance anomaly detection performance when only a portion of normal nodes is provided.
Genetic-guided GFlowNets for Sample Efficient Molecular Optimization
Hyeonah Kim (Korea Advanced Institute of Science and Technology), Jinkyoo Park (Korea Advanced Institute of Science and Technology)
OptimizationDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A molecular optimization method combining domain-specific genetic algorithms with GFlowNet (Genetic GFN) is proposed, which generates high-reward molecules through genetic search and conducts unbiased sampling training using GFlowNet to enhance sample efficiency.
GENOT: Entropic (Gromov) Wasserstein Flow Matching with Applications to Single-Cell Genomics
Dominik Klein (Helmholtz Munich), marco cuturi
OptimizationDrug DiscoveryFlow-based ModelGenerative Adversarial NetworkMultimodalityBiomedical Data
🎯 What it does: This paper presents GENOT, a flow-matching-based neural optimization transport framework that can learn entropy-regularized optimal transport (OT) coupling under arbitrary cost functions (linear, quadratic, fused, and unbalanced) and achieve cell trajectory inference, drug response prediction, and cross-modal translation in single-cell genomics.
GenRec: Unifying Video Generation and Recognition with Diffusion Models
Zejia Weng (Fudan University), Yu-Gang Jiang (Fudan University)
RecognitionGenerationDiffusion modelVideo
🎯 What it does: This paper presents GenRec, a unified video diffusion model capable of achieving high-quality video generation and video recognition within the same network, enhancing robustness in frame-missing scenarios through random frame conditioning and masking mechanisms.
GenRL: Multimodal-foundation world models for generalization in embodied agents
Pietro Mazzaglia (University of Ghent), Sai Rajeswar
Robotic IntelligenceReinforcement LearningVision Language ModelWorld ModelImageVideoMultimodality
🎯 What it does: The GenRL framework is proposed, aligning visual-language foundational models with generative world models in reinforcement learning, enabling agents to learn multi-task behaviors through visual or language prompts alone, and to complete learning in imagination within the model without additional data.
GenWarp: Single Image to Novel Views with Semantic-Preserving Generative Warping
Junyoung Seo (Sony AI), Yuki Mitsufuji (Sony Group Corporation)
Image TranslationGenerationData SynthesisDiffusion modelImage
🎯 What it does: The GenWarp framework is proposed to achieve generative perspective transformation of a single image, generating high-quality new perspective images through a semantic-preserving generative distortion method.
Geodesic Optimization for Predictive Shift Adaptation on EEG data
Apolline Mellot (Inria), Denis Alexander Engemann
Domain AdaptationOptimizationBiomedical Data
🎯 What it does: A multi-source domain adaptation method named GOPSA is proposed, which can simultaneously handle the joint distribution shift of features X and target y on the SPD matrix manifold. The method achieves adaptation to new domains during testing by learning domain-specific geodesic centering parameters and a global regression model.
GeoLRM: Geometry-Aware Large Reconstruction Model for High-Quality 3D Gaussian Generation
Chubin Zhang (Tsinghua University), Yansong Tang (Tsinghua University)
GenerationData SynthesisTransformerGenerative Adversarial NetworkPoint CloudMesh
🎯 What it does: GeoLRM is proposed, a geometry-aware large reconstruction model that can generate high-quality 512k 3D Gaussian models using 21 input views with only 11 GB of GPU memory.