AAAI 2024 Papers — Page 12
AAAI Conference on Artificial Intelligence · 2331 papers
Journey to the Center of the Knowledge Neurons: Discoveries of Language-Independent Knowledge Neurons and Degenerate Knowledge Neurons
Yuheng Chen (Institute of Automation, Chinese Academy of Sciences), Jun Zhao (Institute of Automation, Chinese Academy of Sciences)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the localization and characteristics of knowledge neurons in multilingual pre-trained language models, proposing the AMIG method and discovering language-independent knowledge neurons and degenerate knowledge neurons.
KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning
Debjyoti Mondal (Samsung Research and Development Institute India), Godawari Sudhakar Rao (Samsung Research and Development Institute India)
Graph Neural NetworkTransformerLarge Language ModelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A model that combines knowledge graphs with multi-modal chain-of-thought (CoT) reasoning is proposed—KAM-CoT, which can achieve step-by-step reasoning and answer generation in multi-modal question answering tasks.
KD-Club: An Efficient Exact Algorithm with New Coloring-Based Upper Bound for the Maximum k-Defective Clique Problem
Mingming Jin (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
Graph Neural NetworkGraph
🎯 What it does: A graph coloring-based upper bound CLUB and the corresponding KD-Club algorithm are proposed for the exact solution of the maximum k-defective clique problem.
KeDuSR: Real-World Dual-Lens Super-Resolution via Kernel-Free Matching
Huanjing Yue (Tianjin University), Jingyu Yang (Tianjin University)
RestorationSuper ResolutionGenerative Adversarial NetworkOptical FlowImage
🎯 What it does: In the real dual-camera super-resolution task, the KeDuSR network is proposed, which first aligns the central region of the reference image with the low-resolution image using a global + local alignment method. Then, it performs kernel-free matching between the corners and the central region of the low-resolution image to obtain reference information for the corner areas, and finally generates high-resolution results using adaptive fusion.
Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning
Tom Nuno Wolf (Technical University Munich), Christian Wachinger (Ludwig-Maximilians-University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper studies the interpretability of ProtoPNet in case-based reasoning, proving that its commonly used pixel-level attribution maps do not satisfy interpretability axioms. Based on Shapley values, it proposes the ProtoPFaith method, providing realizable probability layer transformations and closed-form expectation/variance formulas, which can generate faithful pixel attribution maps for ProtoPNet.
Kernelized Normalizing Constant Estimation: Bridging Bayesian Quadrature and Bayesian Optimization
Xu Cai (National University of Singapore), Jonathan Scarlett (National University of Singapore)
OptimizationTabular
🎯 What it does: The study estimates the regularization constant Z = ∫ e^{-λf(x)}dx in RKHS by querying the black-box function f, exploring the difficulties when λ approaches 0, which is similar to Bayesian quadrature (BQ), and when λ approaches ∞, which is similar to Bayesian optimization (BO). It provides algorithm-independent lower bounds and upper bounds for a two-batch estimation algorithm based on Gaussian processes.
Keypoint Fusion for RGB-D Based 3D Hand Pose Estimation
Xingyu Liu (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
Pose EstimationTransformerMultimodalityPoint Cloud
🎯 What it does: Proposes the Keypoint-Fusion method, which utilizes RGB-Depth dual modalities for sparse keypoint aggregation and cross-modal fusion in 3D hand pose estimation.
KG-TREAT: Pre-training for Treatment Effect Estimation by Synergizing Patient Data with Knowledge Graphs
Ruoqi Liu (Ohio State University), Ping Zhang (Anytime.AI)
Graph Neural NetworkTransformerSupervised Fine-TuningGraphBiomedical DataElectronic Health Records
🎯 What it does: KG-TREAT proposes a pre-training and fine-tuning framework that combines large-scale observational patient data with a medical knowledge graph for treatment effect estimation.
KGDM: A Diffusion Model to Capture Multiple Relation Semantics for Knowledge Graph Embedding
Xiao Long (University of Science and Technology of China), Shafei Wang (University of Science and Technology of China)
Diffusion modelGraphBiomedical Data
🎯 What it does: A knowledge graph diffusion model (KGDM) is proposed, which transforms the entity prediction task into a conditional generation problem. It learns the probability distribution of target entities through a denoising diffusion process, thereby addressing the multi-relational semantic challenge.
KGTS: Contrastive Trajectory Similarity Learning over Prompt Knowledge Graph Embedding
Zhen Chen (University of Electronic Science and Technology of China), Shuo Shang (University of Electronic Science and Technology of China)
Recurrent Neural NetworkGraph Neural NetworkPrompt EngineeringContrastive LearningGraphTime Series
🎯 What it does: An unsupervised trajectory similarity learning framework named KGTS is proposed, which utilizes knowledge graph embedding grids, prompt-based trajectory embedding, and contrastive learning to compute trajectory similarity.
Knowledge Enhanced Representation Learning for Drug Discovery
Thanh Lam Hoang (IBM Research), Vanessa Lopez (IBM Research)
Representation LearningDrug DiscoveryGraph Neural NetworkMultimodalityGraphBiomedical Data
🎯 What it does: A multimodal knowledge graph containing seven major public data sources (approximately 30 million triples) was constructed and released, and drug and protein embeddings were obtained by pre-training a graph neural network based on this graph.
Knowledge Graph Error Detection with Contrastive Confidence Adaption
Xiangyu Liu (Nanjing University), Wei Hu (Nanjing University)
Anomaly DetectionTransformerContrastive LearningGraph
🎯 What it does: A knowledge graph error detection model named CCA is proposed, which reconstructs triples by jointly using text (BERT) and graph structure (Transformer), and combines interactive contrastive learning with adaptive confidence dynamic constraint training.
Knowledge Graph Prompting for Multi-Document Question Answering
Yu Wang (Vanderbilt University), Tyler Derr (Vanderbilt University)
TransformerLarge Language ModelPrompt EngineeringTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper addresses the multi-document question answering (MD-QA) task by proposing the use of knowledge graphs (KG) to construct a graph traversal agent driven by LLMs for context retrieval and reasoning, thereby improving answer quality.
Knowledge Guided Semi-supervised Learning for Quality Assessment of User Generated Videos
Shankhanil Mitra (Indian Institute of Science), Rajiv Soundararajan (Indian Institute of Science)
Representation LearningTransformerContrastive LearningVideo
🎯 What it does: This paper designs a self-supervised spatiotemporal video quality representation learning framework ST-VQRL and a semi-supervised video quality assessment method SSL-VQA based on knowledge transfer, which can train high-performance models using only a small amount of labeled videos and a large number of unlabeled videos.
Knowledge-Aware Explainable Reciprocal Recommendation
Kai-Huang Lai (Sun Yat-sen University), Min Chen (South China University of Technology)
Recommendation SystemExplainability and InterpretabilityRecurrent Neural NetworkGraph Neural NetworkTabular
🎯 What it does: A knowledge graph-driven, interpretable bidirectional recommendation system called KAERR is proposed to address the sparsity problem in bidirectional recommendations, encoding and integrating from the perspectives of candidates and positions using meta-paths.
Knowledge-Aware Neuron Interpretation for Scene Classification
Yong Guan (Tsinghua University), Jeff Z. Pan (University of Edinburgh)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A neuron explanation framework based on knowledge graphs is proposed, utilizing three techniques: core concepts, concept filtering, and model tuning, to explain the predictions of image scene classification models.
Knowledge-Aware Parameter Coaching for Personalized Federated Learning
Mingjian Zhi (Northeastern University), Tianao Xiang (Huazhong University of Science and Technology)
Federated LearningImage
🎯 What it does: A knowledge-aware parameter coaching framework is designed and implemented, achieving fine-grained sharing of hierarchical knowledge among clients through a relational cube in federated learning to enhance the performance of personalized models.
Knowledge-Enhanced Historical Document Segmentation and Recognition
En-Hao Gao (Nanjing University), Wang-Zhou Dai (Nanjing University)
RecognitionSegmentationConvolutional Neural NetworkImage
🎯 What it does: In this work, the authors propose the KESAR method, which utilizes an inductive learning framework to embed knowledge into segmentation and recognition models for character segmentation and recognition of historical document images.
KPA-Tracker: Towards Robust and Real-Time Category-Level Articulated Object 6D Pose Tracking
Liu Liu (Hefei University of Technology), Meng Wang (University of Science and Technology of China)
Object TrackingPose EstimationVideoPoint Cloud
🎯 What it does: This paper proposes KPA-Tracker, a category-level joint object 6D position tracking framework based on 3D keypoints, capable of tracking the pose of each rigid part in online real-time scenes.
Kumaraswamy Wavelet for Heterophilic Scene Graph Generation
Lianggangxu Chen (East China Normal University), Gaoqi He (East China Normal University)
GenerationGraph Neural NetworkGraph
🎯 What it does: A Kumaraswamy wavelet-based graph neural network (KWGNN) is proposed to address the heterophily problem in scene graph generation.
Label Attentive Distillation for GNN-Based Graph Classification
Xiaobin Hong (Nanjing University), Sanglu Lu (Chinese University of Hong Kong Shenzhen)
ClassificationKnowledge DistillationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes LAD-GNN, which is based on label attention distillation. It utilizes a teacher-student framework to generate ideal embeddings through a label attention encoder, addressing the misalignment issue between GNN node embeddings and graph labels, thereby enhancing graph classification performance.
Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training
Jianwu Li (Beijing Institute of Technology), Tianfei Zhou (Nanjing University of Science and Technology)
SegmentationMeta LearningTransformerImage
🎯 What it does: Utilize self-supervised pixel embeddings for pixel clustering under unsupervised conditions, automatically construct pseudo meta-tasks, and train few-shot semantic segmentation models;
Labels Need Prompts Too: Mask Matching for Natural Language Understanding Tasks
Bo Li (Peking University), Shikun Zhang (Boston University)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A 'Mask Matching' method is proposed for natural language understanding tasks, where both the input text and label names are wrapped with prompts to obtain mask representations on both sides, and classification predictions are made by matching these two types of mask vectors.
LAFA: Multimodal Knowledge Graph Completion with Link Aware Fusion and Aggregation
Bin Shang (Xi'an Jiaotong University), Di Wang (Xidian University)
Knowledge DistillationRepresentation LearningGraph Neural NetworkTransformerVision Language ModelMultimodalityGraph
🎯 What it does: This paper proposes a multimodal knowledge graph completion model named LAFA, which generates entity embeddings using a link-aware fusion and aggregation mechanism, and employs decoders such as ConvE for link prediction.
LAMM: Label Alignment for Multi-Modal Prompt Learning
Jingsheng Gao (Shanghai Jiao Tong University), Yuzhuo Fu (Southeast University)
ClassificationDomain AdaptationKnowledge DistillationPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a label alignment method called LAMM, which dynamically learns the embeddings of categories in downstream tasks, significantly improving the performance of multimodal prompt learning in few-shot, cross-domain, and incremental learning scenarios.
LAMPAT: Low-Rank Adaption for Multilingual Paraphrasing Using Adversarial Training
Khoi M. Le (VinAI Research), Anh Tuan Luu (Nanyang Technological University)
GenerationData SynthesisOptimizationTransformerLarge Language ModelGenerative Adversarial NetworkText
🎯 What it does: An unsupervised multilingual paraphrase generation model LAMPAT is proposed, which achieves high-quality and diverse paraphrase generation using low-rank adaptation and virtual adversarial training.
LaneGraph2Seq: Lane Topology Extraction with Language Model via Vertex-Edge Encoding and Connectivity Enhancement
Renyuan Peng (Fudan University), Li Zhang (Fudan University)
Autonomous DrivingTransformerLarge Language ModelGraph
🎯 What it does: Proposes LaneGraph2Seq, a method for sequential prediction of lane graphs using a Transformer language model.
Language-Guided Transformer for Federated Multi-Label Classification
I-Jieh Liu (National Taiwan University), Yu-Chiang Frank Wang (NVIDIA)
ClassificationFederated LearningTransformerImage
🎯 What it does: A federated multi-label classification framework FedLGT based on language-guided Transformer is proposed, which can utilize the label association information of clients for local training in multi-label tasks and guide local learning through global model knowledge.
Large Language Models Are Clinical Reasoners: Reasoning-Aware Diagnosis Framework with Prompt-Generated Rationales
Taeyoon Kwon (Yonsei University), Jinyoung Yeo (Yonsei University)
ClassificationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBiomedical DataMagnetic Resonance ImagingAlzheimer's DiseaseChain-of-Thought
🎯 What it does: A framework for 'reasoning-aware diagnosis' based on large language models is proposed and validated, utilizing prompt-based LLMs to automatically generate clinical Chain-of-Thought explanations (Clinical CoT) for disease diagnosis, followed by knowledge distillation to transfer diagnostic capabilities to smaller unimodal and multimodal models.
Large Language Models Are Neurosymbolic Reasoners
Meng Fang (University of Liverpool), Jun Wang (University College London)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Exploring the application of LLM as a neural symbolic reasoner in text games, designing LLM agents to interact with external symbolic modules to complete symbolic tasks.
Large Occluded Human Image Completion via Image-Prior Cooperating
Hengrun Zhao (Dalian University of Technology), Lijun Wang (Dalian University of Technology)
RestorationSegmentationTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a two-stage 'Image-Prior Collaborative Completion' framework to simultaneously recover images and human segmentation priors in human images with large occlusions, significantly improving the reconstruction quality of occluded human areas.
Large-Scale Multi-Robot Coverage Path Planning via Local Search
Jingtao Tang (Simon Fraser University), Hang Ma (Simon Fraser University)
OptimizationRobotic IntelligenceSimultaneous Localization and MappingGraph
🎯 What it does: A new local search-based multi-robot coverage path planning framework LS-MCPP is proposed, and an Extended-STC (ESTC) algorithm that can directly search on the decomposed graph is designed.
Large-Scale Non-convex Stochastic Constrained Distributionally Robust Optimization
Qi Zhang (University at Buffalo), Shaofeng Zou (University at Buffalo)
OptimizationImage
🎯 What it does: This paper studies a stochastic algorithm for large-scale non-convex constrained distributionally robust optimization (DRO) and provides theoretical convergence and complexity analysis.
Latent Diffusion Transformer for Probabilistic Time Series Forecasting
Shibo Feng (Nanyang Technological University), Peilin Zhao (Tencent)
GenerationOptimizationTransformerDiffusion modelAuto EncoderTime Series
🎯 What it does: A multivariate time series probabilistic forecasting framework based on the Latent Diffusion Transformer (LDT) is proposed, which first compresses the time series into a low-dimensional latent space using a statistical adaptive autoencoder, then generates future latent representations in that space using a self-conditioned guided non-autoregressive diffusion generator, and finally decodes to obtain the predicted values.
Latent Space Editing in Transformer-Based Flow Matching
Vincent Tao Hu (University of Amsterdam), Cees Snoek
Image TranslationGenerationTransformerFlow-based ModelAuto EncoderImageMultimodalityOrdinary Differential Equation
🎯 What it does: This paper proposes and implements an image editing method based on the Flow Matching and Transformer (U-ViT) framework, utilizing the u-space of the initial layer in U-ViT for controllable and composable semantic editing. It also develops a semantic direction interpolation technique that works under an adaptive ODE solver and a local prompt re-weighting method, enabling the addition, deletion, replacement, and fine-tuning of image attributes.
LatestEval: Addressing Data Contamination in Language Model Evaluation through Dynamic and Time-Sensitive Test Construction
Yucheng Li (University of Surrey), Chenghua Lin (University of Manchester)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper presents LatestEval, an automatic benchmark for reading comprehension evaluation that constructs the latest text while avoiding data contamination.
LaViP: Language-Grounded Visual Prompting
Nilakshan Kunananthaseelan (Monash University), Mehrtash Harandi (Monash University)
ClassificationRecognitionDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a language-guided visual prompting method called LaViP, which generates language-based input perception visual prompts at the input end of the visual encoder, enabling downstream task transfer of VLM without modifying model parameters.
Layer Collaboration in the Forward-Forward Algorithm
Guy Lorberbom (University of Illinois at Urbana-Champaign), Tamir Hazan (Hebrew University of Jerusalem)
OptimizationComputational EfficiencyImage
🎯 What it does: This study investigates the hierarchical collaboration problem of the forward-forward algorithm and proposes a collaborative forward-forward method that improves the information flow between layers.
Layer Compression of Deep Networks with Straight Flows
Chengyue Gong (University of Texas at Austin), Qiang Liu (Meta)
SegmentationCompressionKnowledge DistillationConvolutional Neural NetworkTransformerImageOrdinary Differential Equation
🎯 What it does: Maps the original deep network to a continuous-time neural flow model, uses the ReFlow operation to straighten the trajectory, and then obtains a first-order flow model through knowledge distillation, thereby achieving model compression and acceleration.
Layer-Wise Representation Fusion for Compositional Generalization
Yafang Zheng (Xiamen University), Xiaodong Shi (Xiamen University)
Representation LearningTransformerText
🎯 What it does: This paper proposes a Layer-wise Representation Fusion (LRF) framework, which introduces a fuse-attention module in each layer of the encoder and decoder of the Transformer, gradually fusing the syntactic and semantic information from the previous layers to address the representation entanglement (RE) problem that arises in the top layer of the model, thereby enhancing the combinatorial generalization ability of sequence-to-sequence models.
LDMVFI: Video Frame Interpolation with Latent Diffusion Models
Duolikun Danier (University of Bristol), David Bull (University of Bristol)
RestorationGenerationData SynthesisDiffusion modelAuto EncoderVideo
🎯 What it does: A method for video frame interpolation using latent diffusion models is proposed, treating VFI as a conditional generation task.
LDS2AE: Local Diffusion Shared-Specific Autoencoder for Multimodal Remote Sensing Image Classification with Arbitrary Missing Modalities
Jiahui Qu (Xidian University), Yufei Yang (Xidian University)
ClassificationTransformerDiffusion modelAuto EncoderImageMultimodality
🎯 What it does: A Local Diffusion Shared-Specific Autoencoder (LDS AE) has been designed and implemented, capable of completing multi-modal remote sensing image classification tasks with any missing modality using only one model.
Learn How to See: Collaborative Embodied Learning for Object Detection and Camera Adjusting
Lingdong Shen (University of Chinese Academy of Sciences), Zichen Wang (University of Chinese Academy of Sciences)
Object DetectionTransformerReinforcement LearningImage
🎯 What it does: Proposes the Student-Teacher Framework (STF), which utilizes replay buffer, teacher rewards, and GPT student networks to achieve collaborative learning of camera actions, thereby enhancing active object detection performance.
Learn the Force We Can: Enabling Sparse Motion Control in Multi-Object Video Generation
Aram Davtyan (University of Bern), Paolo Favaro (University of Bern)
GenerationData SynthesisGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: This paper proposes an unsupervised and controllable multi-object video generation model called YODA, which utilizes sparse motion input to achieve motion control of objects in the video.
Learn to Follow: Decentralized Lifelong Multi-Agent Pathfinding via Planning and Learning
Alexey Skrynnik (AIRI), Aleksandr Panov (AIRI)
OptimizationRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper studies decentralized lifelong multi-agent path planning (LMAPF) and proposes the FOLLOWER method, which combines heuristic planning with reinforcement learning.
Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation
Weiming Liu (Zhejiang University), Yew Soon Ong
Recommendation SystemAuto EncoderContrastive LearningTextOrdinary Differential Equation
🎯 What it does: A cross-domain recommendation model JPEDET is proposed for non-overlapping users/items, which combines user ratings and review information, utilizing dynamic embedding transfer to achieve bidirectional knowledge transfer.
Learning Bayesian Network Classifiers to Minimize the Class Variable Parameters
Shouta Sugahara (University of Electro-Communications), Maomi Ueno (University of Electro-Communications)
ClassificationOptimizationTabular
🎯 What it does: A new method for learning Bayesian network classifiers is proposed, which can find I-map structures with the least class variable parameters in all parentless variable structures and ensure asymptotic classification equivalence.
Learning Broadcast Protocols
Dana Fisman (Ben Gurion University), Swen Jacobs (CISPA Helmholtz Center for Information Security)
🎯 What it does: A learning method for a broadcast protocol (Fine BP) for learning an arbitrary number of processes is proposed.
Learning by Erasing: Conditional Entropy Based Transferable Out-of-Distribution Detection
Meng Xing (Tianjin University), Changjae Oh (Queen Mary University of London)
Anomaly DetectionAuto EncoderImage
🎯 What it does: A transferable OOD detection method based on conditional entropy, CETOOD, is constructed. It captures the differences in conditional entropy distributions of different ID datasets through an image erasure strategy and uncertainty estimation network, enabling cross-dataset OOD detection without retraining.
Learning Cluster-Wise Anchors for Multi-View Clustering
Chao Zhang (Nanjing University), Huaxiong Li (Nanjing University)
OptimizationImage
🎯 What it does: A clustering-based anchor point learning method called CAMVC is proposed to improve the representation of anchor points and subspaces in multi-view clustering.
Learning Coalition Structures with Games
Yixuan Even Xu (Tsinghua University), Fei Fang (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: An important but previously overlooked problem—Coalition Structure Learning (CSL)—is proposed and studied by designing a series of games to infer the coalition structure among agents.
Learning Content-Enhanced Mask Transformer for Domain Generalized Urban-Scene Segmentation
Qi Bi (University of Amsterdam), Theo Gevers (University of Amsterdam)
SegmentationDomain AdaptationAutonomous DrivingTransformerImage
🎯 What it does: This paper proposes CMFormer, a content-enhanced Mask Transformer for domain generalization in urban scene semantic segmentation.
Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling
Shujuan Li (Tsinghua University), Zhizhong Han (Wayne State University)
GenerationData SynthesisOptimizationNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes a point cloud upsampling method based on unsigned distance fields and local distance indicators (LDI) that can achieve point cloud densification at different scales without the need for retraining.
Learning Deformable Hypothesis Sampling for Accurate PatchMatch Multi-View Stereo
Hongjie Li (Wuhan University), Hanjiang Xiong (Wuhan University)
Depth EstimationImage
🎯 What it does: Proposes a learnable Deformable Hypothesis Sampler embedded in the PatchMatch MVS framework to enhance depth estimation accuracy.
Learning Dense Correspondence for NeRF-Based Face Reenactment
Songlin Yang (University of Chinese Academy of Sciences), Jing Dong (Chinese Academy of Sciences)
GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImageVideo
🎯 What it does: A one-shot multi-view facial reenactment framework is proposed, utilizing a tri-plane NeRF representation to learn the dense correspondence between different facial tri-planes without relying on 3DMM priors.
Learning Diffusions under Uncertainty
Hao Huang (Wuhan University), Chuanhui Yang (Ant Group)
OptimizationGraph Neural NetworkGraph
🎯 What it does: The study infers the structure of diffusion networks given only the infection probabilities of nodes and proposes an iterative maximization algorithm based on nonlinear regression called PIND.
Learning Discrete-Time Major-Minor Mean Field Games
Kai Cui (Technische Universitat Darmstadt), Heinz Koeppl (Technische Universitat Darmstadt)
Reinforcement LearningSequential
🎯 What it does: A discrete-time multi-leader-follower equilibrium game (M3FG) framework is proposed for scalable analysis of multi-player non-cooperative games that include dominant players and public noise.
Learning Discriminative Noise Guidance for Image Forgery Detection and Localization
Jiaying Zhu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
Object DetectionAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a two-step noise-guided scheme for image forgery detection and localization using noise inconsistency.
Learning Diverse Risk Preferences in Population-Based Self-Play
Yuhua Jiang (Tsinghua University), Qianchuan Zhao (Tsinghua University)
Reinforcement Learning
🎯 What it does: A self-play method based on risk preference learning is proposed—Risk-Sensitive PPO (RPPO), which is embedded into a Population-Based Self-Play framework to obtain RPBT, aimed at generating diverse and robust playing strategies.
Learning Domain-Independent Heuristics for Grounded and Lifted Planning
Dillon Z. Chen (Australian National University), Felipe Trevizan (Australian National University)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes three new graph representations (SLG, FLG, LLG) for learning domain-independent heuristic estimates in classical planning tasks through graph neural networks, and implements a search framework GOOSE based on these graphs. Through theoretical analysis and experimental validation, it demonstrates that these graphs have greater expressive power than the existing STRIPS-HGN and can achieve higher search coverage and better plan quality in multi-domain tasks.
Learning Efficient and Robust Multi-Agent Communication via Graph Information Bottleneck
Shifei Ding (China University of Mining and Technology), Jian Zhang (China University of Mining and Technology)
Adversarial AttackRobotic IntelligenceGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A multi-agent communication learning mechanism called MAGI based on Graph Information Bottleneck (GIB) is proposed, which can reduce the sensitivity of communication messages to neighbor features while maintaining action coordination, achieving robust and efficient communication learning.
Learning Encodings for Constructive Neural Combinatorial Optimization Needs to Regret
Rui Sun (Southern University of Science and Technology), Zhenkun Wang (Southern University of Science and Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: A general plugin called LCH-Regret is proposed, which introduces a regret mechanism into Learning Constructive Heuristics (LCH);
Learning Explicit Contact for Implicit Reconstruction of Hand-Held Objects from Monocular Images
Junxing Hu (University of Chinese Academy of Sciences), Zhenan Sun
Object DetectionSegmentationPose EstimationGraph Neural NetworkTransformerImage
🎯 What it does: This study investigates how to predict the contact points between a hand and a handheld object using monocular RGB images, and utilizes this explicit contact information to improve the implicit 3D reconstruction of the object.
Learning from Ambiguous Demonstrations with Self-Explanation Guided Reinforcement Learning
Yantian Zha (Arizona State University), Subbarao Kambhampati (Arizona State University)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: A self-explanatory guided RL learning framework SERLfD is proposed, which alleviates the negative impact of ambiguous demonstrations on RL learning by identifying task-related relationships through a self-explanatory network.
Learning from Failure: Improving Meeting Summarization without Good Samples
Ke Wang (Huawei IT Innovation and Research Center), Wei Peng (Huawei IT Innovation and Research Center)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A cold start alignment framework called Score Tuning is proposed, which utilizes asynchronous numerical human feedback to improve meeting summary generation.
Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration
Gang Wu (Harbin Institute of Technology), Xianming Liu (Harbin Institute of Technology)
RestorationSuper ResolutionContrastive LearningImage
🎯 What it does: Utilize the target model itself to generate negative samples for retraining existing image restoration models through model contrastive learning;
Learning from Polar Representation: An Extreme-Adaptive Model for Long-Term Time Series Forecasting
Yanhong Li (Santa Clara University), David Anastasiu
Convolutional Neural NetworkRecurrent Neural NetworkTime Series
🎯 What it does: An extreme adaptability model DAN is proposed, which uses polar coordinate representation learning and distance-weighted multi-loss techniques to predict long-period water flow and handle extreme events.
Learning GAI-Decomposable Utility Models for Multiattribute Decision Making
Margot Herin (Sorbonne University), Nataliya Sokolovska (Sorbonne University)
Tabular
🎯 What it does: This paper proposes a method for learning a Generalized Additive Independent (GAI) utility model in multi-attribute decision-making, which can automatically identify attribute interaction factors and learn the corresponding sub-utility functions, applicable to both continuous and discrete attributes.
Learning Generalized Medical Image Segmentation from Decoupled Feature Queries
Qi Bi (Wuhan University), Yefeng Zheng (Tencent)
SegmentationDomain AdaptationTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A domain generalization medical image segmentation method based on Decoupled Feature Query (DFQ) is proposed, aiming to address the issues of feature redundancy and alignment in cross-domain images.
Learning Generalized Segmentation for Foggy-Scenes by Bi-directional Wavelet Guidance
Qi Bi (University of Amsterdam), Theo Gevers (University of Amsterdam)
SegmentationDomain AdaptationAutonomous DrivingTransformerImage
🎯 What it does: A foggy scene semantic segmentation method is proposed under the domain generalization setting, utilizing a bidirectional wavelet-guided self-attention mechanism (BWG) to separate low-frequency content from high-frequency urban style and fog style, and training only on clear images to achieve good generalization capability for any unknown foggy scenes.
Learning Hierarchical Prompt with Structured Linguistic Knowledge for Vision-Language Models
Yubin Wang (Tongji University), Cairong Zhao (Tongji University)
ClassificationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a Hierarchical Prompt Tuning (HPT) method that enhances the performance of visual-language models in few-shot and cross-domain classification tasks by utilizing category descriptions and their structured relationships (entities, attributes, and their associations) generated by large language models, while also incorporating traditional text prompts.
Learning Hybrid Dynamics Models with Simulator-Informed Latent States
Katharina Ensinger (Bosch Center for Artificial Intelligence), Sebastian Trimpe (Institute for Data Science in Mechanical Engineering, RWTH Aachen University)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: A hybrid dynamic model learning method based on the KKL observer is proposed, which corrects the latent states under the guidance of the black-box simulator output, thereby reducing prediction errors.
Learning Image Demoiréing from Unpaired Real Data
Yunshan Zhong (Xiamen University), Rongrong Ji (Xiamen University)
RestorationGenerationGenerative Adversarial NetworkImage
🎯 What it does: Using unpaired real moiré patterns and non-pattern images, we first generate pseudo moiré images through grouping + GAN, and then train existing de-moiré networks with these pseudo-aligned data.
Learning in Online Principal-Agent Interactions: The Power of Menus
Minbiao Han (University of Chicago), Haifeng Xu (University of Virginia)
Optimization
🎯 What it does: This paper studies the problem of using menus to learn agent privacy types in online master-slave interactions, proposing several algorithms and providing upper bounds on sample complexity, covering various instances such as Stackelberg games, contract design, and security games.
Learning Invariant Inter-pixel Correlations for Superpixel Generation
Sen Xu (Beijing Jiaotong University), Lixin Liao (DaoAI Robotics Inc.)
SegmentationGenerationConvolutional Neural NetworkImage
🎯 What it does: The Content Disentangle Superpixel (CDS) algorithm is proposed, which utilizes auxiliary modalities to separate the style noise of training data, thereby learning pixel correlations that are independent of image content and have better generalization capabilities, used for generating high-quality superpixels.
Learning MDL Logic Programs from Noisy Data
Céline Hocquette (University of Oxford), Andrew Cropper (University of Oxford)
Drug DiscoveryBiomedical DataAlzheimer's Disease
🎯 What it does: An ILP method for learning Minimum Description Length (MDL) logic programs (including recursion and predicate invention) under noisy data is proposed.
Learning Multi-Modal Cross-Scale Deformable Transformer Network for Unregistered Hyperspectral Image Super-resolution
Wenqian Dong (Xidian University), Shaoxiong Hou (Xidian University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: A multi-modal cross-scale deformable transformation network (M2DTN) is proposed to achieve super-resolution of unaligned hyperspectral images.
Learning Multi-Object Positional Relationships via Emergent Communication
Yicheng Feng (Peking University), Zongqing Lu (Peking University)
Convolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningContrastive LearningImage
🎯 What it does: This paper trains agents in a reference game to express and convey spatial position information between two geometric shapes through discrete symbol sequences, and verifies the generalization ability of this language in new multi-step MDP tasks.
Learning Multi-Scale Video-Text Correspondence for Weakly Supervised Temporal Article Gronding
Wenjia Geng (Tsinghua University), Yansong Tang (Tsinghua University)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A multi-scale video-text correspondence learning framework (MVTCL) is proposed, specifically designed to address the weakly supervised multi-scale article video localization (WSAG) task;
Learning Multi-Task Sparse Representation Based on Fisher Information
Yayu Zhang (Shanxi University), Qingfu Zhang (City University of Hong Kong)
SegmentationDepth EstimationRepresentation LearningImage
🎯 What it does: A sparse multi-task learning method based on Fisher information (FSMTL) is proposed, which achieves knowledge transfer by learning a sparse shared subspace for each task.
Learning Multimodal Volumetric Features for Large-Scale Neuron Tracing
Qihua Chen (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
SegmentationRepresentation LearningConvolutional Neural NetworkContrastive LearningImagePoint Cloud
🎯 What it does: A large-scale FlyTracing dataset is constructed, utilizing connection-aware contrastive learning to learn dense embeddings of EM images, and integrating these embeddings with 3D morphological representations (point clouds/voxels) to predict connections between over-segmented neuronal fragments, thereby reducing manual correction work.
Learning Not to Regret
David Sychrovský (Charles University), Martin Schmid (EquiLibre Technologies)
OptimizationMeta LearningReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningTabularSequential
🎯 What it does: This paper proposes a 'learning to avoid regret' framework for games sampled from a distribution, utilizing meta-learning methods to automatically generate regret-avoiding algorithms for that distribution, achieving rapid convergence in matrix games and river card poker distributions.
Learning Only When It Matters: Cost-Aware Long-Tailed Classification
Yu-Cheng He (Nanjing University), Zhi-Hua Zhou (Nanjing University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: Proposes a cost-aware long-tail classification problem and designs the AugARP framework to focus on learning high-cost tail classes through adaptive region partitioning and instance augmentation.
Learning Optimal Advantage from Preferences and Mistaking It for Reward
W. Bradley Knox (University of Texas at Austin), Scott Niekum (University of Massachusetts Amherst)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningSequential
🎯 What it does: Analyzes and experiments on the scenario where preferences are mistakenly viewed as based on partial rewards but are actually driven by regret, showing that the learning algorithm actually obtains the optimal advantage function (rather than the reward function), and explores the impact of this misunderstanding on policy learning and RLHF fine-tuning.
Learning Performance Maximizing Ensembles with Explainability Guarantees
Vincent Pisztora (Pennsylvania State University), Jia Li (Pennsylvania State University)
Explainability and InterpretabilityTabularBenchmark
🎯 What it does: This paper proposes an EEG method that optimally allocates between interpretable (glass box) and black box models based on the 'explainability level' of observation points, maximizing explainability while maintaining performance.
Learning Persistent Community Structures in Dynamic Networks via Topological Data Analysis
Dexu Kong, Yang Li (Shenzhen International Graduate School Tsinghua University)
Representation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: An end-to-end dynamic community detection framework is proposed, combining a Matrix Factorization Clustering module (MFC) and a Topological Regularization module (TopoReg), to enhance temporal consistency and clustering accuracy by maintaining the persistent topology of community networks.
Learning Planning Domains from Non-redundant Fully-Observed Traces: Theoretical Foundations and Complexity Analysis
Pascal Bachor (Albert-Ludwigs-Universitat Freiburg), Gregor Behnke (Universiteit van Amsterdam)
🎯 What it does: In a fully observable classical planning environment, the theoretical foundation and complexity analysis of learning action models (preconditions and effects) from non-redundant observation trajectories are studied. The concept of the most restrictive domain is proposed, and it is proven that determining whether a trajectory is sufficiently or perfectly proven in this domain can be completed in polynomial time, with perfect proof being coNP-complete. After introducing expert knowledge constraints, the complexity increases to NP-complete or Σ₂ⁿ-complete.
Learning Real-World Image De-weathering with Imperfect Supervision
Xiaohui Liu (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
RestorationConvolutional Neural NetworkTransformerContrastive LearningImageVideo
🎯 What it does: A unified method is proposed, which first uses a multi-frame Consistent Label Constructor to generate pseudo-labels that are consistent with the input images and remove most weather damage, and then combines the pseudo-labels with the original imperfect labels through an Information Allocation Strategy to guide the training of the de-weathering model.
Learning Reduced Fluid Dynamics
Zherong Pan (Lightspeed Studios), Kui Wu (Lightspeed Studios)
OptimizationComputational EfficiencyTime SeriesPhysics Related
🎯 What it does: This paper proposes a model-based learning method to construct low-dimensional, strictly time-reversible, energy-preserving fluid dynamics systems, and finely tunes the basis space through gradient optimization.
Learning Representations on the Unit Sphere: Investigating Angular Gaussian and Von Mises-Fisher Distributions for Online Continual Learning
Nicolas Michel (Gustave Eiffel University), Jean-François Bercher (Gustave Eiffel University)
Representation LearningConvolutional Neural NetworkImage
🎯 What it does: A representation learning method based on maximum a posteriori estimation is proposed, using angular Gaussian distribution and von Mises-Fisher distribution on the unit sphere to learn fixed-direction representations and achieve online continual learning with memory replay.
Learning Robust Rationales for Model Explainability: A Guidance-Based Approach
Shuaibo Hu (Hefei University of Technology), Kui Yu (Hefei University of Technology)
Explainability and InterpretabilityText
🎯 What it does: A selection-prediction framework G-RAT based on a guiding module is proposed to improve the text selective rationalization model, addressing its degradation and failure issues.
Learning Safe Action Models with Partial Observability
Hai S. Le (Washington University in St. Louis), Roni Stern (Ben Gurion University of the Negev)
OptimizationSafty and PrivacyRobotic IntelligenceSequentialBenchmark
🎯 What it does: This paper proposes two algorithms, PI-SAM and EPI-SAM, for safely learning PDDL action models from partially observable trajectories, ensuring that the generated plans are executable and can achieve the goals.
Learning Small Decision Trees for Data of Low Rank-Width
Konrad K. Dabrowski (Newcastle University), Stefan Szeider (TU Wien)
ClassificationOptimization
🎯 What it does: This paper proposes a minimum decision tree learning method for low-rank wide data.
Learning Small Decision Trees with Few Outliers: A Parameterized Perspective
Harmender Gahlawat (Ben Gurion University of the Negev), Meirav Zehavi (Ben Gurion University of the Negev)
Anomaly DetectionOptimization
🎯 What it does: This study investigates the parameterized and kernelization complexity of learning small decision trees in the presence of a limited number of anomalous samples. It proves the W[1]-hardness, FPT nature, and compression impossibility of related problems, and provides a minimal polynomial kernel.
Learning Spatially Collaged Fourier Bases for Implicit Neural Representation
Jason Chun Lok Li (University of Hong Kong), Ngai Wong (University of Hong Kong)
RestorationGenerationImageVideoPoint Cloud
🎯 What it does: A new implicit neural representation framework SCONE is proposed, which utilizes learnable spatial masks to stitch Fourier bases of different frequencies to their respective suitable image/video/3D spatial regions, achieving finer local reconstruction.
Learning Subject-Aware Cropping by Outpainting Professional Photos
James Hong (Stanford University), Kayvon Fatahalian
Object DetectionGenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes a weakly supervised subject-aware cropping method called GenCrop, which utilizes image extension techniques to automatically generate cropping pairs from professional stock photos and train a cropping model.
Learning Task-Aware Language-Image Representation for Class-Incremental Object Detection
Hongquan Zhang (East China Normal University), Yuan Xie (Xiamen University)
Object DetectionTransformerVision Language ModelImageText
🎯 What it does: This paper proposes a task-aware language-image representation method to address the problem of catastrophic forgetting in class-incremental object detection.
Learning Temporal Resolution in Spectrogram for Audio Classification
Haohe Liu (University of Surrey), Mark D. Plumbley (The Chinese University of Hong Kong)
ClassificationComputational EfficiencyConvolutional Neural NetworkAudio
🎯 What it does: This paper proposes a differentiable time resolution learning module, DiffRes, which adaptively merges unimportant time frames in audio spectrograms, thereby reducing the time dimension and computational cost while maintaining or improving classification performance.
Learning the Causal Structure of Networked Dynamical Systems under Latent Nodes and Structured Noise
Augusto Santos (Instituto de Telecomunicacoes), José M. F. Moura (Carnegie Mellon University)
GraphTime Series
🎯 What it does: Learning the hidden causal network structure in linear network dynamical systems with only partial node observations and the presence of colored noise.
Learning the Topology and Behavior of Discrete Dynamical Systems
Zirou Qiu (University of Virginia), Anil Vullikanti (University of Virginia)
Graph
🎯 What it does: This paper studies how to learn both unknown topology and unknown threshold interaction functions from observed system dynamics within the PAC framework, that is, simultaneously inferring the network structure and node behavior of discrete dynamical systems.