AAAI 2024 Papers — Page 8
AAAI Conference on Artificial Intelligence · 2331 papers
Equity-Transformer: Solving NP-Hard Min-Max Routing Problems as Sequential Generation with Equity Context
Jiwoo Son (Korea Advanced Institute of Science and Technology), Jinkyoo Park (Korea Advanced Institute of Science and Technology)
OptimizationTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes the Equity-Transformer, which solves the min-max routing problem (multi-agent TSP and multi-agent pickup and delivery problem) using a sequential generation method, achieving efficient solutions for large-scale cities and numbers of agents.
ERL-TD: Evolutionary Reinforcement Learning Enhanced with Truncated Variance and Distillation Mutation
Qiuzhen Lin (Shenzhen University), Jianqiang Li
Knowledge DistillationReinforcement LearningSequential
🎯 What it does: A novel ERL algorithm ERL-TD is proposed, which combines multi-network Q estimation with truncated variance Bellman backup, and introduces distilled mutation to improve exploration and Q value estimation bias.
ESRL: Efficient Sampling-Based Reinforcement Learning for Sequence Generation
Chenglong Wang (Northeastern University), Jingbo Zhu (Northeastern University)
GenerationTransformerReinforcement LearningTextSequential
🎯 What it does: An efficient sampling reinforcement learning method called ESRL is proposed for training sequence generation models.
eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation
Libo Huang (Institute of Computing Technology), Yongjun Xu (Institute of Computing Technology)
ClassificationKnowledge DistillationAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A prototype-free, example-free class incremental learning framework called eTag is proposed, which utilizes embedding distillation to retain the knowledge of the feature extractor while generating task-oriented features that match the classifier to address the problem of catastrophic forgetting.
EulerMormer: Robust Eulerian Motion Magnification via Dynamic Filtering within Transformer
Fei Wang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
TransformerVideo
🎯 What it does: In this paper, the authors propose a Transformer-based EulerMormer framework that implements adaptive denoising and texture-shape separation in video motion magnification using a dynamic filtering strategy.
Evaluate Geometry of Radiance Fields with Low-Frequency Color Prior
Qihang Fang (Chinese Academy of Sciences), Liefeng Bo (Alibaba Group)
Neural Radiance FieldPoint Cloud
🎯 What it does: A new metric for evaluating the geometric quality of light radiation fields under no geometric ground truth conditions is proposed—Inverse Mean Square Residual Color (IMRC), which is theoretically and empirically validated.
EVE: Efficient Vision-Language Pre-training with Masked Prediction and Modality-Aware MoE
Junyi Chen (Sun Yat-sen University), Dongyu Zhang (Institute of Automation, Chinese Academy of Sciences)
RetrievalRepresentation LearningTransformerMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: EVE is proposed, a unified multimodal Transformer that uses only masked signals modeling (MIM+MLM) as a pre-training task, and achieves efficient visual-language pre-training through modality-aware Mixture-of-Experts (MoE).
Every Node Is Different: Dynamically Fusing Self-Supervised Tasks for Attributed Graph Clustering
Pengfei Zhu (Tianjin University), Qinghua Hu (Tianjin University)
Representation LearningGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: A dynamic fusion self-supervised learning framework DyFSS is proposed, which adaptively fuses features from multiple self-supervised tasks for each node to enhance attribute graph clustering performance.
Everything2Motion: Synchronizing Diverse Inputs via a Unified Framework for Human Motion Synthesis
Zhaoxin Fan (Psyche AI Inc), Kai Chen
GenerationData SynthesisTransformerAuto EncoderTextMultimodalityAudio
🎯 What it does: A unified framework called Everything2Motion is proposed, which can generate natural human motion sequences from multimodal inputs such as text and music.
Evidential Uncertainty-Guided Mitochondria Segmentation for 3D EM Images
Ruohua Shi (Peking University), Tingting Jiang (Peking University)
SegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a 3D mitochondrial segmentation method called EUMS-3D, which utilizes evidence uncertainty estimation guidance to simultaneously output segmentation results and corresponding uncertainties, and employs an uncertainty correction module to enhance segmentation accuracy.
Evolving Parameterized Prompt Memory for Continual Learning
Muhammad Rifki Kurniawan, Xing Wei (Xi'an Jiaotong University)
ClassificationRecognitionTransformerPrompt EngineeringImage
🎯 What it does: EvoPrompt is proposed, a continuous learning method based on learnable continuous prompts, using ViT for adaptive prompting and continuous memory fusion.
Exact Algorithms and Lowerbounds for Multiagent Path Finding: Power of Treelike Topology
Foivos Fioravantes (Czech Technical University in Prague), Michal Opler (Czech Technical University in Prague)
OptimizationGraph
🎯 What it does: This paper conducts a systematic study of the Multi-Agent Path Finding (MAPF) problem within the framework of parameterized complexity, providing feasibility and infeasibility results under various parameter combinations, and proposing corresponding FPT algorithms.
Exact ASP Counting with Compact Encodings
Mohimenul Kabir (National University of Singapore), Kuldeep S. Meel
GraphBenchmark
🎯 What it does: The sharpASP framework is proposed, which achieves scalable exact answer set counting based on a new answer set definition and Copy operation.
Exact Inference for Continuous-Time Gaussian Process Dynamics
Katharina Ensinger (Bosch Center for Artificial Intelligence), Sebastian Trimpe (Institute for Data Science in Mechanical Engineering, RWTH Aachen University)
Time SeriesOrdinary Differential Equation
🎯 What it does: A framework is proposed for learning continuous-time dynamical models through precise Gaussian process inference on discrete observational data using multi-step and Taylor numerical integrators.
Exact Policy Recovery in Offline RL with Both Heavy-Tailed Rewards and Data Corruption
Yiding Chen (University of Wisconsin Madison), Xiaojin Zhu (University of Wisconsin Madison)
Reinforcement Learning
🎯 What it does: This paper studies how to achieve optimal policy recovery in an offline reinforcement learning environment where the reward distribution is heavy-tailed and the data is corrupted by attacks.
Exact, Fast and Expressive Poisson Point Processes via Squared Neural Families
Russell Tsuchida (Data61 CSIRO), Dino Sejdinovic (University of Adelaide)
OptimizationComputational EfficiencyPoint Cloud
🎯 What it does: This paper proposes the Squared Neural Poisson Point Process (SNEPPP) based on the squared norm of a two-layer neural network, providing an expressible intensity function for the Poisson point process that allows for closed-form integration, along with a convex maximum likelihood/MAP optimization scheme.
Existence Is Chaos: Enhancing 3D Human Motion Prediction with Uncertainty Consideration
Zhihao Wang (Zhejiang University), Zhao Wang (Zhejiang University)
GenerationPose EstimationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningVideoMultimodality
🎯 What it does: A 3D human motion prediction framework considering uncertainty is proposed, combining Self-Attention Graph Generation Blocks (SAGGB) with Adaptive Salient Loss to achieve dynamic weighted learning for future frames.
Expand-and-Quantize: Unsupervised Semantic Segmentation Using High-Dimensional Space and Product Quantization
Jiyoung Kim (Seoul National University), Byonghyo Shim (Seoul National University)
SegmentationCompressionImage
🎯 What it does: This paper presents EQUSS, an unsupervised semantic segmentation framework that enhances clustering through feature dimension expansion and utilizes product quantization for information compression.
ExpCLIP: Bridging Text and Facial Expressions via Semantic Alignment
Yicheng Zhong (Tencent Technology), Zhisheng Wang (Tencent Technology)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningImageVideoTextMultimodalityAudio
🎯 What it does: A method for generating emotion-driven facial animations based on natural language prompts has been developed, using ExpCLIP to achieve semantic alignment between text and facial expressions.
Expediting Contrastive Language-Image Pretraining via Self-Distilled Encoders
Bumsoo Kim (LG AI Research), Seung Hwan Kim (LG AI Research)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the ECLIPSE meta-architecture, which utilizes self-distillation encoders and token sparsification to achieve efficient contrastive visual-language pre-training.
ExpeL: LLM Agents Are Experiential Learners
Andrew Zhao (Tsinghua University), Gao Huang (Tsinghua University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the ExpeL agent, which autonomously collects successful and failed trajectories on training tasks (using Reflexion for multiple attempts) to extract natural language insights from experience. During the evaluation phase, it enhances decision-making by utilizing retrieved similar successful trajectories and distilled insights, all without the need to fine-tune the LLM parameters.
Explainable Origin-Destination Crowd Flow Interpolation via Variational Multi-Modal Recurrent Graph Auto-Encoder
Qiang Zhou (Nanjing University of Aeronautics and Astronautics), Jingbo Zhou (Dalian University of Technology)
Explainability and InterpretabilityRecurrent Neural NetworkGraph Neural NetworkAuto EncoderMultimodalityGraphTime Series
🎯 What it does: A framework named UApex was constructed, which includes a Variational Multimodal Recursive Graph Autoencoder (VMR-GAE) for interpolating OD (Origin-Destination) crowd flow, and a Shapley value-based uncertainty interpreter was developed to explain the model's prediction results.
Explaining Generalization Power of a DNN Using Interactive Concepts
Huilin Zhou (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper studies the generalization ability of deep networks from the perspective of interactive concepts, proving that lower-order interactive concepts generalize more easily, while higher-order concepts are prone to overfitting, supported by analysis and experimental validation.
Explaining Reinforcement Learning Agents through Counterfactual Action Outcomes
Yotam Amitai (Technion - Israel Institute of Technology), Ofra Amir (Technion - Israel Institute of Technology)
Autonomous DrivingExplainability and InterpretabilityReinforcement LearningVideo
🎯 What it does: This paper proposes the COViz method, which visually displays the trajectory differences between actions chosen by the RL agent in a given state and the optimal alternative actions, and combines it with the Reward Decomposition method to assess its help in user understanding of agent preferences.
Explicit Visual Prompts for Visual Object Tracking
Liangtao Shi (Guangxi Normal University), Xianxian Li (Harbin Institute of Technology)
Object TrackingTransformerVideo
🎯 What it does: A visual tracking framework called EVPTrack is proposed, which is based on explicit visual prompts (spatio-temporal and multi-scale). It utilizes token propagation to convey spatio-temporal information and directly fuses prompts with image tokens through a transformer encoder, thereby avoiding the template update problem.
Explicitly Perceiving and Preserving the Local Geometric Structures for 3D Point Cloud Attack
Daizong Liu (Peking University), Wei Hu (Peking University)
Adversarial AttackGraph Neural NetworkDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a multi-scale multi-layer local spectral filtering geometric perception attack (MGA) that generates more covert adversarial samples by explicitly capturing different local geometric structures.
Exploiting Auxiliary Caption for Video Grounding
Hongxiang Li (Peking University), Yuexian Zou (Peking University)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes an Auxiliary Caption Network (ACNet) that enhances video semantic localization performance by filtering reliable auxiliary captions from dense video subtitles and using these captions to provide prior knowledge for video localization tasks.
Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection
Xiaoyuan Guan (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes using the element-mean difference of the CNN's terminal feature vector as auxiliary information for OOD detection, combined with energy scores to construct a new OOD score.
Exploiting Geometry for Treatment Effect Estimation via Optimal Transport
Yuguang Yan (Guangdong University of Technology), Michael Kwok-Po Ng (Hong Kong Baptist University)
OptimizationTabular
🎯 What it does: This paper proposes an optimal transport-based reweighting method (OTCI) that utilizes cross-group and within-group geometric information to balance the covariate distribution in the observed data, thereby more accurately estimating the average treatment effect.
Exploiting Label Skews in Federated Learning with Model Concatenation
Yiqun Diao (National University of Singapore), Bingsheng He (National University of Singapore)
Federated LearningImage
🎯 What it does: Proposes FedConcat to address the label skew problem in FL through model concatenation and clustering.
Exploiting Polarized Material Cues for Robust Car Detection
Wen Dong (Dalian University of Technology), Xin Yang (Dalian University of Technology)
Object DetectionAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: A multi-modal vehicle detection network PCDNet is proposed, which integrates RGB and linear polarization information to achieve robust vehicle detection in extreme lighting, rain, fog, and dense vehicle scenarios.
Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training
Xi Chen (University of Zurich and ETH Zurich), Tobi Delbruck (University of Zurich and ETH Zurich)
Computational EfficiencyRecurrent Neural NetworkAudio
🎯 What it does: A method is proposed that utilizes the temporal sparsity of Delta networks to implement sparse BPTT in the backpropagation of RNNs, significantly reducing training computation and memory access.
Exploiting the Social-Like Prior in Transformer for Visual Reasoning
Yudong Han (Shandong University), Liqiang Nie (Harbin Institute of Technology)
RecognitionRepresentation LearningTransformerImage
🎯 What it does: This paper proposes a Social-Like Transformer module based on social network theory, which alleviates rank collapse caused by deep layers and enhances representation discrimination by incorporating structured interactions and pillar validation into self-attention, thereby improving performance on visual reasoning tasks.
Explore 3D Dance Generation via Reward Model from Automatically-Ranked Demonstrations
Zilin Wang (Shenzhen International Graduate School Tsinghua University), Zhiyong Wu (Shenzhen International Graduate School Tsinghua University)
GenerationTransformerReinforcement LearningVideo
🎯 What it does: The E3D2 framework is proposed, which utilizes automatically ranked dance demonstrations to train a reward model and explores generating richer and more human-preferred music-driven 3D dances through reinforcement learning.
Exploring Base-Class Suppression with Prior Guidance for Bias-Free One-Shot Object Detection
Wenwen Zhang (Zhejiang University), Eryun Liu (Zhejiang University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Base Class Suppression and Prior Guided Network (BSPG) that eliminates base class interference through a base class predictor in one-shot object detection and uses a non-parametric prior map to guide detection, enhancing unbiased detection performance.
Exploring Channel-Aware Typical Features for Out-of-Distribution Detection
Rundong He (Shandong University), Yongshun Gong (Shandong University)
Anomaly DetectionImage
🎯 What it does: This paper proposes a new method for outlier detection called LAPS, which utilizes channel-aware typical features and improves the limitations of traditional fixed-ratio typical sets.
Exploring Diverse Representations for Open Set Recognition
Yu Wang (Tianjin University), Qinghua Hu (Tianjin University)
RecognitionRepresentation LearningMixture of ExpertsImage
🎯 What it does: Proposes a multi-expert attention fusion model MEDAF for open set recognition.
Exploring Domain Incremental Video Highlights Detection with the LiveFood Benchmark
Sen Pei (ByteDance Inc.), Xiaojie Jin (Institute of Automation, Chinese Academy of Sciences)
ClassificationRecognitionDomain AdaptationConvolutional Neural NetworkTransformerVideoBenchmark
🎯 What it does: This paper proposes an incremental video highlight detection task and implements a no-retraining incremental learning approach in new domains based on the Global Prototype Encoding (GPE) model.
Exploring Equation as a Better Intermediate Meaning Representation for Numerical Reasoning of Large Language Models
Dingzirui Wang (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the use of equations as an intermediate semantic representation to solve numerical reasoning tasks, and enhances the equation generation capability of LLMs through the BRIDGE method.
Exploring Gradient Explosion in Generative Adversarial Imitation Learning: A Probabilistic Perspective
Wanying Wang (Shanghai University), Yangchun Zhang (Shanghai University)
GenerationReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: The study investigates the gradient explosion phenomenon in Generative Adversarial Imitation Learning (GAIL), proposes a probabilistic theoretical analysis, and designs a reward clipping method called CREDO to alleviate the instability of DE-GAIL training.
Exploring Large Language Model for Graph Data Understanding in Online Job Recommendations
Likang Wu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Recommendation SystemGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraph
🎯 What it does: A recruitment recommendation framework GLRec based on large language models is proposed, utilizing meta-paths of behavior graphs to construct prompts, enabling LLMs to understand user behavior graphs and generate personalized recommendation results.
Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models
Mingzhao Yang (Fudan University), Xiangyang Xue (Fudan University)
Data SynthesisFederated LearningDiffusion modelImage
🎯 What it does: A semi-supervised federated learning method called FedDISC is proposed, which utilizes a pre-trained diffusion model to generate synthetic data that conforms to the distribution of each client for training the global model in a single round of communication without client-side training.
Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation
Zhewei Yao (Microsoft), Yuxiong He (Microsoft)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A systematic evaluation of the post-training quantization (PTQ) effects of large language models (OPT, BLOOM) under different sizes (125M–176B) and various quantization schemes (weights, activations, joint) is conducted, and a low-rank compensation (LoRC) method is proposed to enhance quantization quality.
Exploring Self- and Cross-Triplet Correlations for Human-Object Interaction Detection
Weibo Jiang (Harbin Institute of Technology), Honghai Liu (Harbin Institute of Technology)
Object DetectionKnowledge DistillationGraph Neural NetworkTransformerContrastive LearningMultimodality
🎯 What it does: A framework for HOI detection that combines self-Triplet aggregation and cross-Triplet dependencies (SCTC) is proposed, which models the relationships within each candidate triplet and between triplets using graph neural networks, and utilizes CLIP for knowledge distillation.
Exploring Sparse Visual Prompt for Domain Adaptive Dense Prediction
Senqiao Yang (Peking University), Shanghang Zhang (Peking University)
SegmentationDepth EstimationDomain AdaptationAutonomous DrivingTransformerPrompt EngineeringImage
🎯 What it does: Proposes Sparse Visual Domain Prompt (SVDP) and applies it to dense prediction tasks in test-time domain adaptation.
Exploring Temporal Feature Correlation for Efficient and Stable Video Semantic Segmentation
Matthieu Lin (Tsinghua University), Yong-Jin Liu (Tsinghua University)
SegmentationComputational EfficiencyVideo
🎯 What it does: This paper proposes an efficient and stable video semantic segmentation method, addressing the issues of feature misalignment and prediction inconsistency present in the keyframe paradigm.
Exploring Transformer Extrapolation
Zhen Qin (Open NLPLab), Hui Deng (Northwestern Polytechnical University)
TransformerText
🎯 What it does: The study investigates whether Transformer can support longer sequences during inference (length extrapolation), proposes and verifies the convergence conditions of RPE (Relative Position Encoding), derives the theoretical receptive field (TRF), and conducts experimental validation using two new RPEs.
Exponential Hardness of Optimization from the Locality in Quantum Neural Networks
Hao-Kai Zhang (Institute for Advanced Study, Tsinghua University), Xin Wang (Institute for Quantum Computing, Baidu Research)
OptimizationTabularSequentialPhysics Related
🎯 What it does: This study investigates the adjustable range of local controlled units in random quantum neural networks and demonstrates that their maximum impact on the loss function decays exponentially with the number of qubits, leading to an exponential increase in optimization difficulty.
Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection
Zhongjie Ba (Zhejiang University), Kui Ren (Zhejiang University)
ClassificationRecognitionAnomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: A deepfake detection framework based on the information bottleneck is proposed, which utilizes local information blocks to extract separated local features and fuse them into global semantic features;
Expressive Forecasting of 3D Whole-Body Human Motions
Pengxiang Ding (Westlake University), Donglin Wang
Pose EstimationGraph Neural NetworkVideo
🎯 What it does: This paper proposes a framework based on Encoding-Alignment-Interaction (EAI) that can simultaneously predict the 3D full-body motion of major human joints and hand details.
Expressive Multi-Agent Communication via Identity-Aware Learning
Wei Du (China University of Mining and Technology), Ling Ding (Tianjin University)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningAgentic AIGraph
🎯 What it does: The IDEAL protocol is proposed, which enhances cooperation performance by incorporating identity coloring and heterogeneous parameters into the message passing of graph neural networks, allowing multiple agents to obtain distinguishable features under the same observation due to different neighborhoods.
F³-Pruning: A Training-Free and Generalized Pruning Strategy towards Faster and Finer Text-to-Video Synthesis
Sitong Su (University of Electronic Science and Technology of China), Jingkuan Song (University of Electronic Science and Technology of China)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelVideoText
🎯 What it does: This paper proposes F³-Pruning, a training-free, general pruning strategy that accelerates inference and improves video quality by removing redundant temporal attention weights in text-to-video models.
FaceCoresetNet: Differentiable Coresets for Face Set Recognition
Gil Shapira (Bar-Ilan University), Yosi Keller (Bar-Ilan University)
RecognitionTransformerContrastive LearningImageVideo
🎯 What it does: This paper proposes a facial set recognition method called FaceCoresetNet based on differentiable core set selection. It first selects a core subset from an unbounded image collection that balances quality and diversity, then enhances core features using self-attention and cross-attention, and finally obtains a fixed-length template representation for face recognition.
FaceRSA: RSA-Aware Facial Identity Cryptography Framework
Zhongyi Zhang (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
RecognitionGenerationData SynthesisSafty and PrivacyGenerative Adversarial NetworkImage
🎯 What it does: A facial identity encryption and decryption framework called FaceRSA has been designed and implemented, achieving reversible facial anonymization and recovery by mapping passwords to identity-related layers in the StyleGAN latent space.
FacetCRS: Multi-Faceted Preference Learning for Pricking Filter Bubbles in Conversational Recommender System
Yongsen Zheng (Sun Yat-sen University), Liang Lin (Guangdong University of Technology)
Recommendation SystemGraph Neural NetworkTransformerText
🎯 What it does: This paper proposes FacetCRS, a dialog-based recommendation system based on multi-faceted preference learning, which uses natural language interaction to timely 'burst' filter bubbles, enhancing the diversity and quality of recommendations.
FACL-Attack: Frequency-Aware Contrastive Learning for Transferable Adversarial Attacks
Hunmin Yang (Visual Intelligence Lab, KAIST), Kuk-Jin Yoon (Visual Intelligence Lab, KAIST)
Adversarial AttackGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A frequency-domain-based transferable adversarial attack framework called FACL-Attack is proposed, which enhances the cross-domain and cross-model transferability of black-box attacks through frequency domain randomization and contrastive learning.
Fact-Driven Logical Reasoning for Machine Reading Comprehension
Siru Ouyang (University of Illinois), Hai Zhao (Shanghai Jiao Tong University)
Graph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: This paper proposes a hierarchical graph model called Focal Reasoner based on factual units, which constructs factual units by extracting the subject-verb-object structure from sentences and simultaneously models sentence-level and entity-level interactions on a hypergraph to achieve logical reasoning.
Factored Online Planning in Many-Agent POMDPs
Maris F. L. Galesloot, Nils Jansen (Ruhr-University Bochum)
Reinforcement LearningMixture of ExpertsGraphBenchmark
🎯 What it does: A multi-agent POMDP online planning framework is proposed that combines weighted particle filtering and coordination graph value decomposition to achieve scalability for a large number of agents.
Factorized Diffusion Autoencoder for Unsupervised Disentangled Representation Learning
Ancong Wu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
GenerationRepresentation LearningDiffusion modelAuto EncoderImage
🎯 What it does: An unsupervised decoupled representation learning framework is constructed, which decomposes images into multiple 'content + mask' pairs and reconstructs images through a conditional diffusion model, thereby learning interpretable visual concepts.
Factorized Explainer for Graph Neural Networks
Rundong Huang (Technical University of Munich), Dongsheng Luo (Florida International University)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: For post-hoc instance explanation of graph neural networks, the authors analyze and improve the information bottleneck (GIB) objective, propose an improved version of GIB, and design K‑FactExplainer to achieve more accurate subgraph explanations.
Fair Allocation of Items in Multiple Regions
Houyu Zhou (City University of Hong Kong), Minming Li (City University of Hong Kong)
🎯 What it does: This paper proposes a multi-region fair allocation model, focusing on the fair allocation problem in scenarios where divisible and indivisible items are considered, and each agent can only obtain items from a single region;
Fair Lotteries for Participatory Budgeting
Haris Aziz (University of New South Wales), Toby Walsh (University of New South Wales)
🎯 What it does: This paper proposes the use of randomization methods in participatory budgeting (PB) to achieve budget allocation rules that satisfy both ex-ante fairness and ex-post fairness.
Fair Participation via Sequential Policies
Reilly Raab (University of California), Yang Liu (University of Washington)
Recommendation SystemOptimizationTabularSequential
🎯 What it does: This paper proposes a long-term fairness optimization framework that considers distribution shifts caused by policies, and provides a corresponding executable policy update method.
Fairness under Covariate Shift: Improving Fairness-Accuracy Tradeoff with Few Unlabeled Test Samples
Shreyas Havaldar (Google Research India), Aravindan Raghuveer (Google Research India)
Domain AdaptationOptimizationRepresentation LearningTabularBenchmark
🎯 What it does: A joint weighted entropy objective and representation matching loss function is proposed to achieve a trade-off between fairness and accuracy in the presence of a small number of unlabeled test samples and covariate shift.
Fairness without Demographics through Shared Latent Space-Based Debiasing
Rashidul Islam (Visa Research), Yiwei Cai (Visa Research)
Domain AdaptationGenerative Adversarial NetworkContrastive LearningTabular
🎯 What it does: This paper proposes a shared latent space-based debiasing method (SLSD) and its relaxed version (R-SLSD), which maps data from the source domain (with protected attributes) and the target domain (without protected attributes) to a shared representation using deep CCA. It then employs consistency training and adversarial debiasing to achieve fair learning in the target domain.
FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization
Cheng Yang (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
Graph Neural NetworkGraphFinance Related
🎯 What it does: A fair graph neural network framework called FairSIN is proposed, which eliminates sensitive bias in node representations by introducing fairness-promoting features (F3) before information propagation.
FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning
Maryam Badar (L3S Research Center, Leibniz University), Marco Fisichella (L3S Research Center, Leibniz University)
OptimizationFederated LearningTabular
🎯 What it does: Proposes the FairTrade framework to achieve a Pareto optimal trade-off between accuracy and fairness in federated learning;
FairWASP: Fast and Optimal Fair Wasserstein Pre-processing
Zikai Xiong (Massachusetts Institute of Technology), Manuela Veloso (J.P. Morgan AI Research)
OptimizationTabular
🎯 What it does: A preprocessing method named FairWASP is proposed, which reconstructs the dataset by learning integer sample weights (equivalent to duplicating or deleting samples) to minimize the Wasserstein distance between the reconstructed distribution and the original distribution while satisfying demographic parity.
Faithful Model Explanations through Energy-Constrained Conformal Counterfactuals
Patrick Altmeyer (Delft University of Technology), Cynthia C. S. Liem (Delft University of Technology)
Anomaly DetectionOptimizationExplainability and InterpretabilityData-Centric LearningImageTabularFinance RelatedStochastic Differential Equation
🎯 What it does: This paper proposes an energy-constrained and confidence-prediction-based counterfactual generation method called ECCCo, which aims to generate explanations that are both trustworthy and consistent with the data distribution in black-box models.
Far3D: Expanding the Horizon for Surround-View 3D Object Detection
Xiaohui Jiang (Beijing Institute of Technology), Xiangyu Zhang (MEGVII Technology)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes a sparse query framework called Far3D, which generates 3D adaptive queries using high-quality 2D detection priors and achieves long-range 3D object detection at 150 m through viewpoint-aware aggregation and range modulation denoising.
FashionERN: Enhance-and-Refine Network for Composed Fashion Image Retrieval
Yanzhe Chen (Peking University), Lele Cheng (Kuaishou Technology)
RetrievalTransformerContrastive LearningImageText
🎯 What it does: This paper proposes FashionERN (Fashion Enhance-and-Refine Network) for the task of composite fashion image retrieval. It first enhances the text encoder through a three-branch Modifier Enhancement (TME) model, and then refines the visual features using a Dual-guided Vision Refinement (DVR) model, thereby improving retrieval accuracy.
Fast and Controllable Post-training Sparsity: Learning Optimal Sparsity Allocation with Global Constraint in Minutes
Ruihao Gong (Beihang University), Xianglong Liu (Beihang University)
OptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: A fast and controllable post-training sparsification framework called FCPTS has been developed, which utilizes a differentiable bridge function to learn global sparsity rate allocation, generating high-accuracy sparse models within minutes.
Fast Inter-frame Motion Prediction for Compressed Dynamic Point Cloud Attribute Enhancement
Wang Liu (Peking University), Xingming Mu (Peking University)
RestorationCompressionConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: Two end-to-end compressed dynamic point cloud attribute enhancement models are proposed: DAE-S (single frame) and DAE-MP (multi-frame), along with a fast adjacent frame motion prediction module (IFMP) to achieve feature alignment and fusion.
Fast Machine Unlearning without Retraining through Selective Synaptic Dampening
Jack Foster (University of Cambridge), Alexandra Brintrup (University of Cambridge)
RecognitionComputational EfficiencyData-Centric LearningImage
🎯 What it does: A machine forgetting method called Selective Synaptic Dampening (SSD) is proposed, which does not require retraining and is post-hoc. It achieves rapid forgetting and maintains performance on remaining data by selecting and moderately dampening model parameters based on the Fisher information matrix.
Faster Stochastic Variance Reduction Methods for Compositional MiniMax Optimization
Jin Liu (Central South University), Zhe Qu (Central South University)
OptimizationReinforcement LearningImage
🎯 What it does: Two methods, NSTORM and ADA-NSTORM, are proposed for combinatorial minimax optimization problems, achieving a theoretical sample complexity of O(κ/ε^{3/2}) without requiring large batches.
FAVOR: Full-Body AR-Driven Virtual Object Rearrangement Guided by Instruction Text
Kailin Li (Shanghai Jiao Tong University), Cewu Lu (South China University of Technology)
GenerationRobotic IntelligenceLarge Language ModelVision Language ModelVideoText
🎯 What it does: A dataset named FAVOR was constructed, consisting of 3,000 full-body human-robot interaction rearrangement sequences collected using motion capture and AR glasses. Based on this dataset, the FAVORITE pipeline was proposed, enabling the generation of complete grasping and placing actions from text instructions to 3D scene localization.
FD3D: Exploiting Foreground Depth Map for Feature-Supervised Monocular 3D Object Detection
Zizhang Wu (Fudan University), Jian Pu (Fudan University)
Object DetectionAutonomous DrivingGenerative Adversarial NetworkImage
🎯 What it does: Proposes the FD3D framework, which utilizes the annotated foreground depth map (AFOD) to provide feature supervision for monocular 3D object detection models, thereby improving detection accuracy;
Feature Distribution Matching by Optimal Transport for Effective and Robust Coreset Selection
Weiwei Xiao (Harbin Institute of Technology), Jingyong Su (Harbin Institute of Technology)
ClassificationSupervised Fine-TuningImage
🎯 What it does: A coreset selection method based on feature distribution matching, FDMat, is proposed, aiming to reduce the distribution bias between the subset and the original dataset through optimal transport.
Feature Fusion from Head to Tail for Long-Tailed Visual Recognition
Mengke Li (Guangdong Laboratory of Artificial Intelligence and Digital Economy), Hui Huang (Hong Kong Baptist University)
ClassificationRecognitionImage
🎯 What it does: A head-to-tail feature fusion method (H2T) is proposed for long-tail visual recognition, which enhances the semantic diversity of tail classes by randomly replacing some channels in the tail class feature map with head class features.
Feature Transportation Improves Graph Neural Networks
Moshe Eliasof (University of Cambridge), Eran Treister (Ben-Gurion University of the Negev)
ClassificationGraph Neural NetworkGraphTime SeriesOrdinary Differential Equation
🎯 What it does: A graph neural network ADR-GNN based on advection-diffusion-reaction (ADR) PDE is proposed, utilizing learnable transport, diffusion, and reaction operators to address the limitations of traditional GNNs such as node feature transfer and heterogeneity.
FeatWalk: Enhancing Few-Shot Classification through Local View Leveraging
Dalong Chen (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
ClassificationMeta LearningConvolutional Neural NetworkImage
🎯 What it does: A pluggable FeatWalk module is proposed to enhance the classification performance of few-shot learning by fusing randomly sampled local views with the original global features.
Fed-QSSL: A Framework for Personalized Federated Learning under Bitwidth and Data Heterogeneity
Yiyue Chen (University of Texas at Austin), Chianing Wang (Toyota InfoTech Lab)
Federated LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A Fed-QSSL framework is proposed to address the issues of data distribution heterogeneity and client bit-width heterogeneity in federated learning.
FedA3I: Annotation Quality-Aware Aggregation for Federated Medical Image Segmentation against Heterogeneous Annotation Noise
Nannan Wu (Huazhong University of Science and Technology), Li Yu (Huazhong University of Science and Technology)
SegmentationFederated LearningConvolutional Neural NetworkImageBiomedical DataUltrasound
🎯 What it does: A non-IID annotation noise modeling and robust training method for multi-source data is proposed in federated medical image segmentation;
FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update
Ji Liu (Hithink RoyalFlush Information Network Co., Ltd.), Dejing Dou (Boston Consulting Group)
Federated LearningConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningImageText
🎯 What it does: An asynchronous federated learning framework FedASMU is proposed, combining dynamic delay-aware model aggregation and local adaptive model adjustment to alleviate system and statistical heterogeneity.
FedCD: Federated Semi-Supervised Learning with Class Awareness Balance via Dual Teachers
Yuzhi Liu (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
Federated LearningKnowledge DistillationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Proposes the FedCD framework, which utilizes dual teacher distillation and class-aware balancing to address knowledge drift and class imbalance issues in federated semi-supervised learning.
FedCompetitors: Harmonious Collaboration in Federated Learning with Competing Participants
Shanli Tan (National Engineering Research Center of Mobile Network Technologies, Beijing University of Posts and Telecommunications), Xiaofeng Tao (State Key Laboratory for Novel Software Technology, Nanjing University)
Federated LearningGraph Neural NetworkImageBiomedical DataElectronic Health Records
🎯 What it does: The FedCompetitors framework is proposed to handle competitive participants in the federated learning ecosystem, ensuring complementary data while avoiding conflicts of interest, ultimately building a scalable collaborative network.
FedCSL: A Scalable and Accurate Approach to Federated Causal Structure Learning
Xianjie Guo (Hefei University of Technology), Jiuyong Li (University of South Australia)
Federated LearningSafty and PrivacyComputational EfficiencyTabular
🎯 What it does: A scalable and accurate federated causal structure learning method, FedCSL, is proposed to address the scalability and accuracy issues of existing methods under high-dimensional data and sample imbalance.
FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning
Haokun Chen (Siemens AG), Volker Tresp (LMU Munich)
Federated LearningKnowledge DistillationTransformerVision Language ModelMultimodality
🎯 What it does: Proposes the FedDAT framework for parameter-efficient fine-tuning of large foundational models in a multimodal heterogeneous federated learning environment.
FedDiv: Collaborative Noise Filtering for Federated Learning with Noisy Labels
Jichang Li (Sun Yat-sen University), Yizhou Yu (University of Hong Kong)
Federated LearningImage
🎯 What it does: Designed and implemented the FedDiv framework to handle noisy labeled data in federated learning, achieving noise detection and sample re-labeling through a global noise filter and a prediction consistency sampler.
Federated Adaptive Prompt Tuning for Multi-Domain Collaborative Learning
Shangchao Su (Fudan University), Xiangyang Xue (Fudan University)
ClassificationDomain AdaptationFederated LearningPrompt EngineeringContrastive LearningImage
🎯 What it does: Proposes the FedAPT cross-domain federated learning framework, which utilizes prompt tuning of the CLIP pre-trained model to achieve multi-domain image classification.
Federated Causality Learning with Explainable Adaptive Optimization
Dezhi Yang (Shandong University), Jinglin Zhang (Shandong University)
OptimizationFederated LearningExplainability and InterpretabilityGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Proposes FedCausal to learn unified causal graphs in a federated environment.
Federated Contextual Cascading Bandits with Asynchronous Communication and Heterogeneous Users
Hantao Yang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Recommendation SystemFederated LearningTabular
🎯 What it does: Proposed FedC3UCB-H, which addresses the issues of asynchronous communication and user heterogeneity in federated context-aware cascading bandits.
Federated Graph Learning under Domain Shift with Generalizable Prototypes
Guancheng Wan (Wuhan University), Mang Ye (Wuhan University)
Domain AdaptationFederated LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A federated graph learning framework named FGGP is proposed, which utilizes generalizable prototypes to achieve shared model training of multi-domain graph data, addressing the global model generalization problem under domain shift.
Federated Label-Noise Learning with Local Diversity Product Regularization
Xiaochen Zhou (Shanghai Jiao Tong University), Xudong Wang (Shanghai Jiao Tong University)
Federated LearningImage
🎯 What it does: A learning framework called FedLNL is proposed for label noise in a federated learning environment, which alternately updates the noise transformation matrix (NTM) and the classifier on local devices, using Bayesian inference and local diversity product regularization to achieve denoising and model training.
Federated Learning with Extremely Noisy Clients via Negative Distillation
Yang Lu (Xiamen University), Hanzi Wang (Xiamen University)
Federated LearningKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: In federated learning, the FedNed method is proposed to address the issue of extreme noisy clients (noise rate > 90%): it first identifies extreme noisy clients through model uncertainty, and then employs negative distillation and pseudo-label training to fully utilize information from all clients and enhance the performance of the global model.
Federated Modality-Specific Encoders and Multimodal Anchors for Personalized Brain Tumor Segmentation
Qian Dai (Xiamen University), Yefeng Zheng (Tencent)
SegmentationFederated LearningTransformerMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the FedMEMA framework, which employs a federated modality-specific encoder and a server-side multimodal fusion decoder, and optimizes both the full-modal server model and the personalized model of the client with missing modalities simultaneously through multi-anchor cross-attention calibration.
Federated Partial Label Learning with Local-Adaptive Augmentation and Regularization
Yan Yan (Carleton University), Yuhong Guo (Carleton University)
ClassificationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper studies the problem of partial label learning in a federated learning environment (FedPLL) and proposes the FedPLL-LAAR method.
Federated X-armed Bandit
Wenjie Li (Purdue University), Guang Lin (Purdue University)
OptimizationFederated LearningTabular
🎯 What it does: A federated X-armed bandit framework is proposed, and the Fed-PNE algorithm is designed to achieve global optimal search for multiple clients through hierarchical partitioning and node elimination.
FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning
Xinyuan Ji (Xi'an Jiaotong University), Yang Liu (University of California Santa Cruz)
Federated LearningImage
🎯 What it does: We propose FedFixer, which utilizes a dual model structure of global and personalized models to effectively denoise and train models in federated learning against heterogeneous label noise.
FedGCR: Achieving Performance and Fairness for Federated Learning with Distinct Client Types via Group Customization and Reweighting
Shu-Ling Cheng (National Taiwan University), Ming-Syan Chen (National Taiwan University)
Federated LearningTransformerContrastive LearningImage
🎯 What it does: The paper proposes FedGCR, which aims to improve performance and fairness in federated learning with different client types.