AAAI 2024 Papers — Page 9
AAAI Conference on Artificial Intelligence · 2331 papers
FedLF: Layer-Wise Fair Federated Learning
Zibin Pan (Chinese University of Hong Kong), Junhua Zhao (Chinese University of Hong Kong)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes the FedLF algorithm, which utilizes hierarchical public square directions to improve fairness in federated learning.
FedLPS: Heterogeneous Federated Learning for Multiple Tasks with Local Parameter Sharing
Yongzhe Jia (Nanjing University), Wanchun Dou (Macquarie University)
ClassificationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: A heterogeneous federated learning framework named FedLPS is proposed, which utilizes shared encoders and task-specific predictors to achieve multi-task local parameter sharing, and reduces resource consumption on edge devices through channel-level pruning and heterogeneous aggregation, while maintaining or even improving model accuracy.
FedMut: Generalized Federated Learning via Stochastic Mutation
Ming Hu (Nanyang Technological University), Yang Liu (Nanyang Technological University)
Federated LearningImageText
🎯 What it does: A new federated learning method called FedMut is proposed, which generates multiple intermediate models through random mutations of the global model for local training, thereby guiding the global model to converge to a flatter optimal region and improving generalization performance.
FedNS: A Fast Sketching Newton-Type Algorithm for Federated Learning
Jian Li (Institute of Information Engineering Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering Chinese Academy of Sciences)
OptimizationFederated LearningTabular
🎯 What it does: This paper proposes Federated Newton Sketch (FedNS) and its improved version FedNDES, which utilize sketch-based square root Hessian information to achieve second-order optimization in federated learning.
FedST: Federated Style Transfer Learning for Non-IID Image Segmentation
Boyuan Ma (University of Science and Technology Beijing), Xiaojuan Ban (Liaoning Academy of Materials)
SegmentationDomain AdaptationFederated LearningConvolutional Neural NetworkDiffusion modelImageMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes a federated style transfer learning framework called FedST, which utilizes a denoising diffusion probabilistic model to achieve style decoupling and synthesis across domains, thereby alleviating the impact of non-IID in image segmentation tasks.
FedTGP: Trainable Global Prototypes with Adaptive-Margin-Enhanced Contrastive Learning for Data and Model Heterogeneity in Federated Learning
Jianqing Zhang (Shanghai Jiao Tong University), Jian Cao (Shanghai Jiao Tong University)
Federated LearningSafty and PrivacyConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A globally shareable and trainable prototype (TGP) is designed for heterogeneous federated learning, achieving greater separation between different categories through adaptive margin-enhanced contrastive learning, which guides the training of client models.
Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection
Soopil Kim (Daegu Gyeongbuk Institute of Science and Technology), Sang Hyun Park (Daegu Gyeongbuk Institute of Science and Technology)
SegmentationAnomaly DetectionImage
🎯 What it does: This paper proposes an anomaly detection framework PSAD that combines a part segmentation model based on a small number of pixel-level annotations with three types of memory banks, capable of detecting both logical and structural anomalies in industrial images.
Few-Shot Learning from Augmented Label-Uncertain Queries in Bongard-HOI
Qinqian Lei (National University of Singapore), Robby T. Tan (National University of Singapore)
ClassificationRecognitionObject DetectionMeta LearningContrastive LearningImage
🎯 What it does: A framework for few-shot human-object interaction detection is proposed, utilizing the Mean Teacher method with label-uncertain query augmentation and pseudo-label generation to enhance the model's classification performance on the Bongard-HOI and HICO-FS datasets.
Few-Shot Learning via Repurposing Ensemble of Black-Box Models
Minh Hoang (Princeton University), Trong Nghia Hoang (Washington State University)
ClassificationDomain AdaptationMeta LearningFlow-based ModelAuto EncoderImage
🎯 What it does: A black-box model reuse framework is proposed, which constructs feature integration using reversible enhancement of target inputs and black-box outputs from multiple related tasks, thereby reusing multi-model knowledge in few-shot scenarios.
Few-Shot Neural Radiance Fields under Unconstrained Illumination
SeokYeong Lee (Korea Institute of Science and Technology), Junghyun Cho (Korea Institute of Science and Technology)
GenerationData SynthesisNeural Radiance FieldImageBenchmark
🎯 What it does: This paper proposes ExtremeNeRF, which utilizes view-independent color (albedo) consistency constraints to achieve novel view synthesis in real-world environments with only a few multi-view images and inconsistent lighting.
Fewer Steps, Better Performance: Efficient Cross-Modal Clip Trimming for Video Moment Retrieval Using Language
Xiang Fang (Huazhong University of Science and Technology), Renfu Li (Huazhong University of Science and Technology)
RetrievalComputational EfficiencyKnowledge DistillationTransformerContrastive LearningVideoText
🎯 What it does: This paper proposes SpotVMR, an efficient video segment cropping method for language-driven video moment retrieval.
FFT-Based Dynamic Token Mixer for Vision
Yuki Tatsunami (Rikkyo University), Masato Taki (Rikkyo University)
ClassificationSegmentationComputational EfficiencyTransformerImage
🎯 What it does: A dynamic filter is proposed, and two visual models, DFFormer and CDFFormer, are constructed based on the MetaFormer framework, improving the token mixer based on FFT benchmarks to achieve faster and more memory-efficient processing of high-resolution images.
FG-EmoTalk: Talking Head Video Generation with Fine-Grained Controllable Facial Expressions
Zhaoxu Sun (Xiaobing), Yang Xiang (Xiaobing)
GenerationData SynthesisGenerative Adversarial NetworkVideoAudio
🎯 What it does: We propose FG-EmoTalk, a GAN-based talking head video generation framework that enables fine-grained expression editing through the input of facial action unit (AU) intensities, supporting both audio-driven and video-driven approaches.
Finding Interpretable Class-Specific Patterns through Efficient Neural Search
Nils Philipp Walter (CISPA Helmholtz Center for Information Security), Jilles Vreeken (CISPA Helmholtz Center for Information Security)
ClassificationExplainability and InterpretabilityAuto EncoderBiomedical Data
🎯 What it does: A fully binary neural network architecture named DIFFNAPS is proposed to mine interpretable class-specific difference patterns from high-dimensional binary data, simultaneously optimizing data reconstruction and classification through multi-task objectives.
Finding Visual Saliency in Continuous Spike Stream
Lin Zhu (Beijing Institute of Technology), Hua Huang (Beijing Normal University)
Spiking Neural NetworkTransformerTime Series
🎯 What it does: This paper proposes a recursive spiking Transformer framework for full SNNs to detect visual saliency from continuous spike streams of spiking cameras.
Fine Structure-Aware Sampling: A New Sampling Training Scheme for Pixel-Aligned Implicit Models in Single-View Human Reconstruction
Kennard Yanting Chan (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
GenerationPose EstimationConvolutional Neural NetworkPoint CloudMesh
🎯 What it does: Proposes the Fine Structure-Aware Sampling (FSS) scheme and its extensions NSP and MTL for training pixel-aligned implicit models to achieve single-view human reconstruction with clothing, significantly improving the reconstruction quality of thin components and reducing noise.
Fine-Grained Distillation for Long Document Retrieval
Yucheng Zhou (University of Macau), Daxin Jiang (Microsoft Corporation)
RetrievalKnowledge DistillationTransformerText
🎯 What it does: The study proposes a fine-grained distillation framework (FGD) for long document retrieval, combining multi-granularity representation with cross-encoder distillation to maintain the efficiency of single-vector retrieval.
Fine-Grained Knowledge Selection and Restoration for Non-exemplar Class Incremental Learning
Jiang-Tian Zhai (Nankai University), Ming-Ming Cheng (Nankai University)
ClassificationKnowledge DistillationTransformerImage
🎯 What it does: A non-sample class incremental learning framework based on visual Transformer is proposed, utilizing fine-grained patch-level knowledge selection and prototype recovery to mitigate catastrophic forgetting.
Fine-Grained Multi-View Hand Reconstruction Using Inverse Rendering
Qijun Gan (Zhejiang University), Jianke Zhu (Zhejiang University)
RestorationGenerationPose EstimationGraph Neural NetworkNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper proposes a fine-grained multi-view hand mesh reconstruction method that utilizes inverse rendering to finely recover hand posture and geometric details, and achieves realistic image synthesis through a pre-trained mesh neural renderer.
Fine-Grained Prototypes Distillation for Few-Shot Object Detection
Zichen Wang (Northwestern Polytechnical University), Zhenghao Ma (Northwestern Polytechnical University)
Object DetectionKnowledge DistillationMeta LearningConvolutional Neural NetworkImage
🎯 What it does: The research focuses on few-shot object detection based on meta-learning, proposing Fine-Grained Prototype Distillation (FFA) and improving feature aggregation and fusion;
Fine-Tuning Graph Neural Networks by Preserving Graph Generative Patterns
Yifei Sun (Zhejiang University), Lei Chen (FinVolution Group)
Graph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: When fine-tuning GNNs after pre-training, the fine-tuning effect is enhanced by constructing graphon to maintain the generative patterns of downstream graphs.
Fine-Tuning Large Language Model Based Explainable Recommendation with Explainable Quality Reward
Mengyuan Yang (Zhejiang University), Jianwei Yin (Zhejiang University)
Recommendation SystemExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningGenerative Adversarial NetworkContrastive LearningText
🎯 What it does: A large language model-based explainable recommendation system LLM2ER was constructed, and unsupervised fine-tuning of explanation quality was achieved through reinforcement learning combined with two reward models (Concept Consistency Reward CCR and High-Quality Alignment Reward HQAR), resulting in LLM2ER-EQR, which can generate personalized, consistent, and high-quality text explanations.
Finite-Time Frequentist Regret Bounds of Multi-Agent Thompson Sampling on Sparse Hypergraphs
Tianyuan Jin (National University of Singapore), Pan Xu (Duke University)
OptimizationReinforcement Learning from Human FeedbackGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper studies the sparse hypergraph structure in multi-agent multi-armed bandits (MAMAB) and proposes the ε-exploration multi-agent Thompson sampling (ε-MATS) algorithm, providing its frequentist lower bound.
FLAME: A Small Language Model for Spreadsheet Formulas
Harshit Joshi (Stanford University), Gust Verbruggen (Microsoft)
RetrievalAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTabular
🎯 What it does: Trained a 60M parameter Transformer model FLAME using only Excel formulas for formula repair, completion, and similar formula retrieval.
FlexKBQA: A Flexible LLM-Powered Framework for Few-Shot Knowledge Base Question Answering
Zhenyu Li (Tsinghua University), Jianyong Wang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the FlexKBQA framework, which uses large language models (LLM) as program translators to generate executable SPARQL / S-expression and creates a large number of synthetic question-answer pairs, then trains with a lightweight model, alleviating distribution drift and semantic errors through execution-guided self-training (EGST) and intrinsic reasoning (IR);
FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework
Dongyue Guo (Sichuan University), Yi Lin (Sichuan University)
TransformerTime Series
🎯 What it does: Proposes FlightBERT++, a non-autoregressive multi-time-slot trajectory prediction framework;
Fluctuation-Based Adaptive Structured Pruning for Large Language Models
Yongqi An (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: FLAP is proposed, a training-free, structured pruning framework specifically designed for large language models (LLMs), achieving column-level pruning and compensating for losses through a volatility-based recoverability metric.
FM-OV3D: Foundation Model-Based Cross-Modal Knowledge Blending for Open-Vocabulary 3D Detection
Dongmei Zhang (Peking University), Shanghang Zhang (Peking University)
Object DetectionTransformerLarge Language ModelDiffusion modelContrastive LearningTextPoint Cloud
🎯 What it does: This paper proposes FM-OV3D, which integrates cross-modal knowledge from various pre-trained foundational models to enhance the open vocabulary detection capability of 3D point clouds.
FMRNet: Image Deraining via Frequency Mutual Revision
Kui Jiang (Harbin Institute of Technology), Xianzheng Ma (Wuhan University of Science and Technology)
Image TranslationRestorationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Proposed FMRNet frequency domain mutual correction network to achieve rain removal from rain and fog images.
FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly
Yuhan Li (Shanghai Jiao Tong University), Bingbing Ni (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: We propose FocalDreamer, a text-based 3D editing framework that allows users to generate separable, instantiable, and consistent editable components while maintaining the original base shape, by providing a rough focus area and text prompts. The components are then merged with the base through soft geometry union and rendered using dual-path rendering to synthesize a complete model, which can further generate PBR textures.
Focus Stacking with High Fidelity and Superior Visual Effects
Bo Liu (Chongqing University of Posts and Telecommunications), Bin Xiao (Chongqing University of Posts and Telecommunications)
RestorationSegmentationConvolutional Neural NetworkImageBenchmark
🎯 What it does: A foreground segmentation-diffusion elimination architecture based on optical imaging process analysis is proposed to synthesize panoramic focal images that are both high-fidelity and free of transfer region (TR) defects.
Focus-Then-Decide: Segmentation-Assisted Reinforcement Learning
Chao Chen (Nanjing University), Rui Zhao (Tencent)
SegmentationRobotic IntelligenceReinforcement LearningImage
🎯 What it does: Proposes the Focus-Then-Decide (FTD) framework, which uses an attention selector to filter the objects returned by the base segmentation model, retaining only task-relevant objects to achieve visual reinforcement learning;
Follow Your Pose: Pose-Guided Text-to-Video Generation Using Pose-Free Videos
Yue Ma (Tsinghua University), Qifeng Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisPose EstimationDiffusion modelImageVideoText
🎯 What it does: A two-stage training framework is proposed, utilizing image-pose pairs and unposed videos to generate high-quality character videos that are text-editable and pose-controllable based on Stable Diffusion.
FontDiffuser: One-Shot Font Generation via Denoising Diffusion with Multi-Scale Content Aggregation and Style Contrastive Learning
Zhenhua Yang (South China University of Technology), Lianwen Jin (South China University of Technology)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelContrastive LearningImage
🎯 What it does: Proposes the FontDiffuser method, which uses a diffusion model to achieve one-shot font generation.
Forced Exploration in Bandit Problems
Qi Han (Xi'an Jiaotong University), Fei Guo (Xi'an Jiaotong University)
OptimizationReinforcement LearningTime Series
🎯 What it does: This paper proposes a multi-armed bandit algorithm that alternates between greedy decision-making and forced exploration without requiring prior knowledge of the reward distribution parameters, and provides theoretical analysis and implementation in both steady-state and piecewise steady-state environments.
Forecasting Bimanual Object Manipulation Sequences from Unimanual Observations
Haziq Razali (Imperial), Yiannis Demiris (Imperial)
Pose EstimationRobotic IntelligenceGraph Neural NetworkAuto EncoderTime SeriesSequential
🎯 What it does: This paper proposes a method for predicting bimanual object manipulation sequences from a single-hand observation, capable of reconstructing and predicting complete bimanual actions and object movements in the absence of one hand.
Foreseeing Reconstruction Quality of Gradient Inversion: An Optimization Perspective
Hyeong Gwon Hong (KAIST), Junmo Kim (KAIST)
OptimizationFederated LearningAdversarial AttackImage
🎯 What it does: This study investigates the impact of the loss function on the sample vulnerability ranking in gradient inversion attacks and proposes a loss-aware vulnerability proxy (LAVP) based on gradient matching loss at the Hessian's maximum/minimum eigenvalues near the true value to assess sample vulnerability.
Formal Logic Enabled Personalized Federated Learning through Property Inference
Ziyan An (Vanderbilt University), Meiyi Ma (Vanderbilt University)
Federated LearningRecurrent Neural NetworkTime Series
🎯 What it does: A personalized federated learning framework named FedSTL is proposed, which enhances the model's symbolic reasoning ability by automatically inferring the temporal logic properties of each client;
FoSp: Focus and Separation Network for Early Smoke Segmentation
Lujian Yao (East China University of Science and Technology), Kaijie Zhao (East China University of Science and Technology)
SegmentationTransformerImage
🎯 What it does: This paper proposes a network called FoSp based on Focus and Separation for fine segmentation of early smoke.
Foundations of Reactive Synthesis for Declarative Process Specifications
Luca Geatti (University of Udine), Andrey Rivkin (Technical University of Denmark)
🎯 What it does: This paper formalizes and solves the realizability and reactive synthesis problems of the DECLARE process specification for the first time, and proposes various algorithm implementations.
FoX: Formation-Aware Exploration in Multi-Agent Reinforcement Learning
Yonghyeon Jo (Ulsan National Institute of Science and Technology), Seungyul Han (Ulsan National Institute of Science and Technology)
Reinforcement LearningSequentialBenchmark
🎯 What it does: A formation-based exploration method implemented through the 'FoX' framework in multi-agent reinforcement learning is proposed.
FPRF: Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural Radiance Fields
GeonU Kim (POSTECH), Tae-Hyun Oh (Yonsei University)
Image TranslationGenerationKnowledge DistillationNeural Radiance FieldImage
🎯 What it does: We propose FPRF, a forward optical realistic style transfer method for large-scale 3D neural radiance fields, which supports multiple reference images and generates globally consistent stylized views in real-time without additional optimization.
Fractional Deep Reinforcement Learning for Age-Minimal Mobile Edge Computing
Lyudong Jin (Zhejiang University), Hao Wang (Monash University)
OptimizationReinforcement LearningTime Series
🎯 What it does: This paper proposes a scheduling framework for Mobile Edge Computing (MEC) aimed at minimizing age, jointly optimizing task update intervals and offloading strategies to minimize the time-average Age-of-Information (AoI).
Frame Semantic Role Labeling Using Arbitrary-Order Conditional Random Fields
Chaoyi Ai (ShanghaiTech University), Kewei Tu (ShanghaiTech University)
Recurrent Neural NetworkLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: Proposes a framework semantic role labeling model based on arbitrary-order conditional random fields.
FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection
Chanho Lee (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
Object DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A fully rotationally equivariant aerial image object detection framework FRED has been constructed, achieving a complete process of equivariance from input to bounding box prediction;
Frequency Shuffling and Enhancement for Open Set Recognition
Lijun Liu (Institute of Information Engineering), Chuan Wang (Institute of Information Engineering)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a frequency-domain-based open set recognition framework called FreSH, which utilizes High-Frequency Jitter (HFS) and Low-Frequency Residual (LFR) to reduce excessive focus on fine details and enhance global structure, thereby improving the model's accuracy in recognizing known classes and its ability to detect unknown classes.
Frequency Spectrum Is More Effective for Multimodal Representation and Fusion: A Multimodal Spectrum Rumor Detector
An Lao (Beijing Institute of Technology), Duoqian Miao (Tongji University)
ClassificationAnomaly DetectionRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: A model for multimodal rumor detection in the frequency domain (FSRU) is proposed.
Frequency-Adaptive Pan-Sharpening with Mixture of Experts
Xuanhua He (University of Science and Technology of China), Man Zhou (Nanyang Technological University)
Image TranslationRestorationMixture of ExpertsImage
🎯 What it does: This paper proposes a full-resolution multispectral image fusion method based on frequency adaptive mixture of experts (FAME), utilizing learnable frequency masks to process high and low-frequency information separately and fuse them dynamically;
Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning
Chuangchuang Tan (Beijing Jiaotong University), Yunchao Wei (A*STAR)
ClassificationAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A lightweight frequency domain learning network called FreqNet is proposed for general detection of unseen deepfake images with limited training data.
Frequency-Controlled Diffusion Model for Versatile Text-Guided Image-to-Image Translation
Xiang Gao (Peking University), Jiaying Liu (Peking University)
Image TranslationGenerationDiffusion modelImageText
🎯 What it does: A text-guided image-to-image translation framework called FCDiffusion based on frequency domain filtering is proposed, capable of performing multiple I2I tasks within a single model.
Friendly Attacks to Improve Channel Coding Reliability
Anastasiia Kurmukova (Imperial College London), Deniz Gunduz (Imperial College London)
OptimizationAdversarial Attack
🎯 What it does: This paper proposes a 'friendly attack' strategy that enhances the decoding reliability of error correction codes by adding small perturbations to the modulation codewords without altering the receiver.
FRIH: Fine-Grained Region-Aware Image Harmonization
Jinlong Peng (Tencent), Boshen Zhang (Tencent)
Image HarmonizationRestorationConvolutional Neural NetworkImage
🎯 What it does: A two-stage Fine-grained Region-Aware Image Harmonization (FRIH) framework is proposed, where a global U-Net is first used for coarse adjustment of the foreground, followed by adaptive clustering to obtain sub-masks, which are refined in a lightweight cascading module, and finally, a fusion prediction is utilized to obtain the final synthesized image.
From Coarse to Fine: A Distillation Method for Fine-Grained Emotion-Causal Span Pair Extraction in Conversation
Xinhao Chen (East China Normal University), Aimin Zhou (East China Normal University)
RecognitionKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the extraction of fine-grained emotions and causal spans, proposing the KD-ECS method based on knowledge distillation, and constructs the FG-RECCON fine-grained emotional causal dataset.
From GARCH to Neural Network for Volatility Forecast
Pengfei Zhao (BNU-HKBU United International College), Dik Lun Lee (Hong Kong University of Science and Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkTransformerTime SeriesSequentialFinance Related
🎯 What it does: This paper studies the equivalence relationship between GARCH series models and neural networks, and based on this, proposes the GARCH-NN framework, constructing interpretable volatility prediction models such as GARCH-LSTM.
From Past to Future: Rethinking Eligibility Traces
Dhawal Gupta (University of Massachusetts), Bruno Castro da Silva (Amazon)
Reinforcement LearningSequential
🎯 What it does: This paper explores the credit allocation problem using eligibility traces under nonlinear function approximation, proposing a bidirectional value function that combines forward and backward value functions, and provides online updates and convergence proofs.
From Retrieval to Generation: A Simple and Unified Generative Model for End-to-End Task-Oriented Dialogue
Zeyuan Ding (Dalian University of Technology), Hongfei Lin (Dalian University of Technology)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A unified generative model called Uni-ToD is proposed, which simultaneously accomplishes knowledge retrieval and dialogue generation using a single language model.
From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks
Xianda Zhang (Peking University), Xiaoying Bai (University of Glasgow)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A dual-stage framework called FTT is proposed, which combines self-attention distillation and semi-supervised learning to eliminate backdoors in deep networks and improve accuracy.
Frozen CLIP Transformer Is an Efficient Point Cloud Encoder
Xiaoshui Huang (Shanghai AI Laboratory), Wanli Ouyang (Shanghai AI Laboratory)
ClassificationObject DetectionSegmentationTransformerSupervised Fine-TuningContrastive LearningPoint Cloud
🎯 What it does: Utilizing a frozen CLIP Transformer as a point cloud encoder, combined with a point cloud tokenizer and task tokens, to directly fine-tune on various point cloud tasks.
Frugal LMs Trained to Invoke Symbolic Solvers Achieve Parameter-Efficient Arithmetic Reasoning
Subhabrata Dutta (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Bombay)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper presents the SYRELM system, which utilizes small-scale language models to translate natural language arithmetic problems into formal language (FL) expressions and calls a symbolic solver to complete the reasoning.
FT-GAN: Fine-Grained Tune Modeling for Chinese Opera Synthesis
Meizhen Zheng (Xiamen University), Yiting Yan (Xiamen University)
GenerationData SynthesisGenerative Adversarial NetworkAudio
🎯 What it does: This paper constructs the first high-quality audio-text alignment dataset for Gezi opera and proposes a fine-grained tuning model FT-GAN specifically for Chinese opera, along with a speech pre-training strategy, achieving high-quality singing voice synthesis for Gezi opera and other operas such as Peking opera.
Full Bayesian Significance Testing for Neural Networks
Zehua Liu (Beihang University), Yue He (Tsinghua University)
ImageTabular
🎯 What it does: A full Bayesian significance test based on Bayesian neural networks (n FBST) is proposed, which can perform significance testing of feature importance at global, local, and instance levels.
Full-Body Motion Reconstruction with Sparse Sensing from Graph Perspective
Feiyu Yao (Huawei Technologies), Li Yi (Tsinghua University)
Pose EstimationGraph Neural NetworkGraph
🎯 What it does: This paper proposes the use of Body Pose Graph to reconstruct full-body motion from sparse sensor data obtained from VR systems (head-mounted displays and two controllers).
Fully Data-Driven Pseudo Label Estimation for Pointly-Supervised Panoptic Segmentation
Jing Li (University of Chinese Academy of Sciences), Zhaoxiang Zhang
Object DetectionSegmentationTransformerImage
🎯 What it does: A completely data-driven pseudo-label branch is proposed, which utilizes sparse point labels to generate full-pixel pseudo-labels for training a panorama segmentation model focused on point annotations.
Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data
Yucheng Wang (Institute for Infocomm Research, Astar), Zhenghua Chen (Nanyang Technological University)
Graph Neural NetworkTime Series
🎯 What it does: A fully connected spatio-temporal graph neural network (FC-STGNN) is proposed to learn the features of multivariate time series.
Fusing Conditional Submodular GAN and Programmatic Weak Supervision
Kumar Shubham (Indian Institute of Science), Prathosh AP (Indian Institute of Science)
ClassificationGenerationGenerative Adversarial NetworkImage
🎯 What it does: A joint training framework guided by a noise-aware classifier is constructed by combining conditional GANs with programmatic weak supervision, and representative samples are selected through submodular optimal subset selection to improve label quality and image generation effects.
Fusion-Vital: Video-RF Fusion Transformer for Advanced Remote Physiological Measurement
Jae-Ho Choi (Stanford University), Kyung-Tae Kim (Pohang University of Science and Technology)
TransformerVideoMultimodality
🎯 What it does: By integrating RGB video and RF sensor signals, a multi-level time-difference feature fusion is achieved using Transformer, enabling remote monitoring of breathing and heart rate.
FusionFormer: A Concise Unified Feature Fusion Transformer for 3D Pose Estimation
Yanlu Cai (Fudan University), Cheng Jin (Fudan University)
Pose EstimationTransformerVideo
🎯 What it does: A unified feature fusion transformer, FusionFormer, is designed to integrate multi-view and multi-frame information for camera-parameter-free 3D human pose estimation.
G-Adapter: Towards Structure-Aware Parameter-Efficient Transfer Learning for Graph Transformer Networks
Anchun Gui (Xiamen University), Han Xiao (Xiamen University)
Graph Neural NetworkTransformerGraph
🎯 What it does: The study transfers parameter-efficient fine-tuning techniques to graph Transformer networks and proposes G-Adapter to address the issue of feature distribution shift.
G-NAS: Generalizable Neural Architecture Search for Single Domain Generalization Object Detection
Fan Wu (Shanghai Jiao Tong University), Nanyang Ye (Huawei Noah's Ark Lab)
Object DetectionDomain AdaptationNeural Architecture SearchImage
🎯 What it does: This paper proposes a single-domain generalization object detection framework G-NAS based on differentiable neural architecture search, addressing the generalization problem of models trained on a single source domain to multiple target domains.
G^2SAM: Graph-Based Global Semantic Awareness Method for Multimodal Sarcasm Detection
Yiwei Wei (Tianjin University), Meng Chen (JD AI Research)
ClassificationGraph Neural NetworkContrastive LearningImageTextMultimodality
🎯 What it does: A global semantic awareness method based on graphs, G SAM, is proposed, which generates multimodal graph representations using a fine-grained graph alignment model, and determines sarcasm during the inference phase through kNN global semantic similarity voting. Additionally, label-aware graph contrastive learning is introduced to enhance semantic consistency.
G2L-CariGAN: Caricature Generation from Global Structure to Local Features
Xin Huang (Tongji University), Jinyuan Jia (Tongji University)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A GAN-based G2L-CariGAN model is designed to achieve exaggerated facial generation of characters using the global structure and local features of reference comics while maintaining character identity recognition.
G2P-DDM: Generating Sign Pose Sequence from Gloss Sequence with Discrete Diffusion Model
Pan Xie (Beihang University), Zexian Li (Beihang University)
GenerationPose EstimationTransformerDiffusion modelAuto EncoderVideoSequential
🎯 What it does: This paper proposes a new framework for generating sign language pose sequences from sign language vocabulary sequences (G2P‑DDM), primarily by discretizing the continuous pose space and using a discrete diffusion model for conditional generation.
GAD-PVI: A General Accelerated Dynamic-Weight Particle-Based Variational Inference Framework
Fangyikang Wang (Zhejiang University), Hui Qian (Zhejiang University)
Tabular
🎯 What it does: This paper proposes a particle variational inference framework GAD-PVI that simultaneously accelerates position updates and adjusts dynamic weights.
GAMC: An Unsupervised Method for Fake News Detection Using Graph Autoencoder with Masking
Shu Yin (Northwestern Polytechnical University), Zhen Wang (Northwestern Polytechnical University)
ClassificationAnomaly DetectionGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: An unsupervised fake news detection method called GAMC is proposed, which utilizes graph autoencoders, feature masking, and contrastive learning to learn latent features from news propagation graphs.
Gated Attention Coding for Training High-Performance and Efficient Spiking Neural Networks
Xuerui Qiu (Institute of Automation, Chinese Academy of Sciences), Guoqi Li (Institute of Automation, Chinese Academy of Sciences)
ClassificationSpiking Neural NetworkImage
🎯 What it does: A Gated Attention Coding (GAC) pre-coding module is proposed in SNN, which efficiently encodes static images into powerful and spatiotemporally dynamic spike sequences using multidimensional gated attention units while maintaining spike-driven characteristics.
Gaussian Mixture Proposals with Pull-Push Learning Scheme to Capture Diverse Events for Weakly Supervised Temporal Video Grounding
Sunoh Kim (Seoul National University), Jin Young Choi (NAVER CLOVA)
Convolutional Neural NetworkTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A weakly supervised temporal video localization framework is proposed, utilizing learnable Gaussian Mixture Proposals (GMP) and a Pull-Push Scheme to capture diverse events.
Gaussian Process Neural Additive Models
Wei Zhang (Columbia University), John Paisley (Columbia University)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a Gaussian Process-based Neural Additive Model (GP-NAM), which models the shape function of each feature as a one-dimensional Gaussian process by approximating the RBF kernel with Random Fourier Features (RFF), resulting in an interpretable, parameter-efficient, and convex additive model.
Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process
Mingjie Xu (Beihang University), Feng Lu (Peng Cheng Laboratory)
Pose EstimationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A gaze normalization-based auxiliary learning framework GFAL is proposed to enhance cross-domain gaze estimation performance.
Gaze Target Detection by Merging Human Attention and Activity Cues
Yaokun Yang (Beihang University), Feng Lu (Beihang University)
Object DetectionPose EstimationConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes a gaze target detection method that integrates human attention and activity cues, achieving precise detection of gaze targets in complex backgrounds through a soft gaze attention module and a body part-object interaction attention module.
GCNext: Towards the Unity of Graph Convolutions for Human Motion Prediction
Xinshun Wang (Sun Yat-sen University), Mengyuan Liu (University of Central Florida)
Pose EstimationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A universal graph convolution (UniGC) and dynamic graph convolution network framework GCNext are proposed, which can dynamically select the most suitable type of graph convolution for each layer and each sample, and can be used for training from scratch or refining existing GCNs.
Generalisation through Negation and Predicate Invention
David M. Cerna (Czech Academy of Sciences Institute of Computer Science), Andrew Cropper (University of Oxford)
GraphTabular
🎯 What it does: A method for inductive logic programming learning that combines negation and predicate invention is proposed, capable of generating polar programs and achieving recursive and optimized normal logic program learning.
Generalising Planning Environment Redesign
Alberto Pozanco (J.P. Morgan), Daniel Borrajo (J.P. Morgan)
OptimizationSafty and PrivacyTabular
🎯 What it does: In the task of redesigning planning environments, a general, metric-independent method is proposed that can minimize action deletions to redesign the environment based on the goals and metrics of any interested party; a series of new privacy and distance-related metrics are also introduced.
Generalizable Fourier Augmentation for Unsupervised Video Object Segmentation
Huihui Song (Nanjing University of Information Science and Technology), Dong Liu (Walmart Global Tech)
SegmentationDomain AdaptationTransformerOptical FlowVideo
🎯 What it does: Proposes a Generalizable Fourier Augmentation (GFA) framework that enhances the generalization ability of unsupervised video object segmentation by applying FFT to intermediate features in the Transformer, performing Gaussian sampling on the magnitude, and using EMA on the phase.
Generalizable Sleep Staging via Multi-Level Domain Alignment
Jiquan Wang (Zhejiang University), Gang Pan (Zhejiang University)
ClassificationDomain AdaptationConvolutional Neural NetworkTransformerAuto EncoderTime SeriesBiomedical Data
🎯 What it does: A domain generalization framework for sleep staging, SleepDG, is proposed, which can achieve high-accuracy staging on unseen datasets.
Generalizable Task Representation Learning for Offline Meta-Reinforcement Learning with Data Limitations
Renzhe Zhou (Nanjing University), Yang Yu (Nanjing University)
Representation LearningMeta LearningReinforcement LearningAuto Encoder
🎯 What it does: The GENTLE framework is proposed, which learns general task representations through Task Auto-Encoder (TAE) in offline meta reinforcement learning, and constructs pseudo transitions using action, dynamics, and reward relabeling to overcome the issues of insufficient task quantity and behavioral diversity.
Generalization Analysis of Machine Learning Algorithms via the Worst-Case Data-Generating Probability Measure
Xinying Zou (INRIA), Eitan Altman (Université d'Avignon)
🎯 What it does: Proposes the 'worst-case data generation probability measure' and proves it to be a Gibbs measure, thereby obtaining closed-form expressions for key indicators such as expected loss, empirical risk, and generalization error.
Generalize for Future: Slow and Fast Trajectory Learning for CTR Prediction
Jian Zhu (JD.COM), Jingping Shao (JD.COM)
Recommendation SystemTabular
🎯 What it does: A Slow and Fast Trajectory Learning (SFTL) framework is proposed to address temporal distribution drift in CTR prediction.
Generalized Bradley-Terry Models for Score Estimation from Paired Comparisons
Julien Fageot (Tournesol Association), Oscar Villemaud (Tournesol Association)
🎯 What it does: This paper proposes a general Bradley-Terry model (GBT) framework for transforming pairwise comparisons into object scores, providing theoretical guarantees such as strict convexity, monotonicity, and Lipschitz robustness within this framework.
Generalized Planning for the Abstraction and Reasoning Corpus
Chao Lei (University of Melbourne), Krista A. Ehinger (University of Melbourne)
🎯 What it does: A generalized planning-based ARC solver GPAR is proposed, which maps each ARC task to a PDDL planning problem and implements object-level conditions and loops through external functions and pointer references;
Generalized Planning in PDDL Domains with Pretrained Large Language Models
Tom Silver (Massachusetts Institute of Technology), Michael Katz (Massachusetts Institute of Technology)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study utilizes the pre-trained GPT-4 model, combined with chain reasoning and automatic debugging techniques, to generate a general-purpose program capable of handling a range of PDDL planning tasks, referred to as the universal planner.
Generalized Variational Inference via Optimal Transport
Jinjin Chi (Jilin University), Hongbin Pei (Xi'an Jiaotong University)
OptimizationAuto EncoderImageTabular
🎯 What it does: This paper proposes a variational inference method called VOT that uses optimal transport distance (OT) to measure the difference between the prior and the variational posterior. By introducing a λ hyperparameter, OT constraints are gradually incorporated, forming a gradient-compatible black-box optimization framework.
Generalizing across Temporal Domains with Koopman Operators
Qiuhao Zeng (University of Western Ontario), Boyu Wang (Beihang University)
Domain AdaptationAuto EncoderTime SeriesSequential
🎯 What it does: This paper proposes and implements a temporal domain generalization model based on the Koopman operator (Temporal Koopman Networks, TKNets), aimed at predicting the distribution of the target domain and performing classification by utilizing the temporal evolution patterns between source domains.
Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection
Songmin Dai (Shanghai University), Xiangyang Xue (Fudan University)
Anomaly DetectionConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes a prior-free unsupervised anomaly detection framework called GRAD, which utilizes the diffusion model PatchDiff to generate multi-level dense contrast patterns, and achieves efficient detection and localization through a self-supervised reweighting mechanism and a lightweight FCN detector.
Generating Images of Rare Concepts Using Pre-trained Diffusion Models
Dvir Samuel (Bar-Ilan University), Gal Chechik (OriginAI)
GenerationData SynthesisOptimizationDiffusion modelContrastive LearningImage
🎯 What it does: The SeedSelect method is proposed, which optimizes seeds in the noise space and generates high-quality images of rare concepts by utilizing a small number of reference images in conjunction with a pre-trained diffusion model.
Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback
Shreyas Bhat Brahmavar (Birla Institute of Technology and Science Pilani), Raviprasad Aduri (TCS Innovation Labs)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringBiomedical Data
🎯 What it does: The study utilizes large language models (LLMs) and a logical feedback mechanism to propose an iterative LMLF method that automatically refines logical constraints and generates drug lead molecules.
Generating Universal Adversarial Perturbations for Quantum Classifiers
Gautham Anil (Indian Institute of Technology Madras), Apurva Narayan (University of Western Ontario)
ClassificationAdversarial AttackAuto EncoderGenerative Adversarial NetworkImagePhysics Related
🎯 What it does: This study proposes two methods for generating universal adversarial perturbations (UAP) for quantum classifiers: additive UAP based on classical generative models (QuGAP-A) and unitary UAP based on quantum generative models (QuGAP-U), and verifies their attack effects on classifiers based on parameterized quantum circuits (PQC).
Generative Calibration of Inaccurate Annotation for Label Distribution Learning
Liang He (Nanjing University of Science and Technology), Xiuyi Jia (Nanjing University of Science and Technology)
Data-Centric LearningAuto EncoderGenerative Adversarial NetworkTabular
🎯 What it does: This study investigates how to correct inaccurate label distribution annotations to improve the performance of label distribution learning (LDL) models.
Generative Model Perception Rectification Algorithm for Trade-Off between Diversity and Quality
Guipeng Lan (Tianjin University), Jiabao Wen (Tianjin University)
GenerationData SynthesisDiffusion modelFlow-based ModelGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A dynamic model-aware correction algorithm (DMPRA) is designed and validated to balance the diversity and quality of generative models.
Generative Model-Based Feature Knowledge Distillation for Action Recognition
Guiqin Wang (Xi'an Jiaotong University), Shusen Yang (Xi'an Jiaotong University)
RecognitionCompressionKnowledge DistillationConvolutional Neural NetworkAuto EncoderVideo
🎯 What it does: A knowledge distillation framework based on generative models is proposed to transfer the spatiotemporal feature semantics captured in 3D-CNNs to a lightweight student network, thereby enhancing video action recognition and detection performance.
Generative Multi-Modal Knowledge Retrieval with Large Language Models
Xinwei Long (Tsinghua University), Jie Zhou (Tencent Inc)
GenerationRetrievalTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: An end-to-end generative multimodal knowledge retrieval framework, GeMKR, is proposed, which utilizes large language models to directly output knowledge cues that can be mapped to documents during the generation phase, and then retrieves the corresponding documents from the database.