AAAI 2024 Papers — Page 13
AAAI Conference on Artificial Intelligence · 2331 papers
Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation
Tianhao Huang (Nankai University), Xiaojie Yuan (Nankai University)
Recommendation SystemTransformerSequential
🎯 What it does: This paper proposes a Mobility Tree based on a hierarchical tree structure and designs the MTNet model, which utilizes multi-granularity time slot nodes to learn users' next location preferences at different time periods.
Learning to Approximate Adaptive Kernel Convolution on Graphs
Jaeyoon Sim (Pohang University of Science and Technology), Won Hwa Kim (Pohang University of Science and Technology)
ClassificationGraph Neural NetworkGraphAlzheimer's Disease
🎯 What it does: A learnable adaptive diffusion kernel is introduced in graph neural networks, utilizing polynomial approximations (Chebyshev, Hermite, Laguerre) to implement heat kernel convolution. Based on this, the diffusion scale for each node is learned end-to-end, significantly improving the performance of node classification and graph classification tasks.
Learning to Learn Better Visual Prompts
Fengxiang Wang (National University of Defense Technology), Long Lan (National University of Defense Technology)
ClassificationMeta LearningTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: Based on VLM pre-training models (such as CLIP), a meta-learning-based visual prompt tuning framework called LoL (Learning to Learn) is proposed. It significantly enhances the model's generalization ability to new classes by first training learnable prompts using CoOp, and then performing N-way K-shot meta-learning (episodic) fine-tuning on the base classes.
Learning to Learn in Interactive Constraint Acquisition
Dimosthenis Tsouros (KU Leuven), Tias Guns (KU Leuven)
OptimizationMeta LearningTabularBenchmark
🎯 What it does: This study investigates the introduction of statistical machine learning methods into Constraint Acquisition (CA) to reduce the number of required queries.
Learning to Manipulate Artistic Images
Wei Guo (Zhejiang University), Qian Zheng (Zhejiang University)
Image TranslationGenerationImage
🎯 What it does: A mask-based zero-shot artistic image manipulation network SIM-Net is proposed, capable of modifying artistic images in any style without using semantic labels or training data.
Learning to Optimize Permutation Flow Shop Scheduling via Graph-Based Imitation Learning
Longkang Li (Chinese University of Hong Kong), Baoyuan Wu (National University of Singapore)
OptimizationGraph Neural NetworkReinforcement LearningGraphTabular
🎯 What it does: Use a graph-structured imitation learning method to solve the permutation flow shop scheduling (PFSS) problem, outputting the job arrangement that minimizes makespan.
Learning to Pivot as a Smart Expert
Tianhao Liu (Shanghai University of Finance and Economics), Yinyu Ye (Stanford University)
OptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: Two types of intelligent pivot experts based on global optimal benchmarks and local information have been designed and implemented, and the behavior of the experts has been approximated through imitation learning using Graph Convolutional Neural Networks (GCNN), thereby improving the pivot selection of the simplex method;
Learning to Prompt Knowledge Transfer for Open-World Continual Learning
Yujie Li (Southwestern University of Finance and Economics), Tianrui Li (Southwest Jiaotong University)
ClassificationKnowledge DistillationPrompt EngineeringImage
🎯 What it does: The Pro-KT model is proposed, which utilizes a Prompt Bank to encode task-general and task-specific knowledge and achieves unknown detection and incremental learning in open-ended continual learning.
Learning to Rank in Generative Retrieval
Yongqi Li (Hong Kong Polytechnic University), Wenjie Li (Microsoft)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the LTRGR framework, which adds a learning-to-rank phase on top of the generative retrieval model, allowing the model to directly optimize the ranking of retrieval results.
Learning to Reweight for Generalizable Graph Neural Network
Zhengyu Chen (Zhejiang University), Fei Wu (Alibaba Group)
ClassificationOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This study investigates the generalization problem of graph neural networks (GNNs) in out-of-distribution (OOD) scenarios and proposes the L2R-GNN framework, which utilizes nonlinear graph decorrelation and bi-level optimization to achieve graph sample reweighting, enhancing the model's predictive performance on graphs with unknown distributions.
Learning to Stop Cut Generation for Efficient Mixed-Integer Linear Programming
Haotian Ling (University of Science and Technology of China), Jie Wang (University of Science and Technology of China)
OptimizationComputational EfficiencyGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This study proposes a reinforcement learning-based method for learning cut generation stopping strategies, aimed at improving the efficiency of Mixed Integer Linear Programming (MILP) solving.
Learning to Unlearn: Instance-Wise Unlearning for Pre-trained Classifiers
Sungmin Cha (New York University), Moontae Lee (LG AI Research)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A framework for instance-level forgetting is proposed, which only uses pre-trained models and samples to be deleted, to remove sensitive information through misclassification or re-labeling while maintaining the predictive performance of the remaining data.
Learning Ultrametric Trees for Optimal Transport Regression
Samantha Chen (University of California), Yusu Wang (University of California)
OptimizationGraphTabular
🎯 What it does: A method for learning ultra-metric trees based on projected gradient descent is proposed to approximate the 1-Wasserstein distance in discrete metric spaces.
Learning Uncertainty-Aware Temporally-Extended Actions
Joongkyu Lee (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement LearningSequential
🎯 What it does: A time extension algorithm UTE based on uncertainty perception is proposed, which learns to dynamically select the extension length during repeated action execution and adjusts the exploration and exploitation strategy based on the uncertainty of the Q-value.
Learning Visual Abstract Reasoning through Dual-Stream Networks
Kai Zhao (Beijing Normal University), Bailu Si (Beijing Normal University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a Dual-stream Reasoning Network (DRNet) that extracts image features by simultaneously using two branches: CNN (for processing local/object information) and ViT (for processing global/spatial information), and learns abstract rules in a rule extractor to solve visual abstract reasoning tasks of Raven’s Progressive Matrices (RPM).
Learning with Noisy Labels Using Hyperspherical Margin Weighting
Shuo Zhang (Southeast University), Chengyu Liu (Southeast University)
ClassificationData-Centric LearningImage
🎯 What it does: The IAM metric and HMW weighting method are proposed to improve the robustness of learning on datasets with noisy labels.
Learning-Augmented Online Algorithm for Two-Level Ski-Rental Problem
Keyuan Zhang (Virginia Tech), Bo Ji (Virginia Tech)
OptimizationTabularTime Series
🎯 What it does: A robust online algorithm RDTSR and a learning-enhanced algorithm LADTSR based on machine learning predictions are proposed and implemented for the two-tier ski rental problem, addressing optimal cost decisions under various payment options (rental, single purchase, combination purchase).
Leaving the Nest: Going beyond Local Loss Functions for Predict-Then-Optimize
Sanket Shah (Harvard University), Milind Tambe (Harvard University)
OptimizationTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a new framework called 'Efficient Global Losses' (EGL) for learning task-specific loss functions in Predict-then-Optimize (PtO) to improve decision quality.
LERE: Learning-Based Low-Rank Matrix Recovery with Rank Estimation
Zhengqin Xu (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationOptimizationVideo
🎯 What it does: A learning-driven low-rank matrix recovery method called LERE is proposed, which first estimates the matrix rank using a new R_GDE rule, and then iteratively solves on randomly sampled submatrices using a 17-step FRMNN deep network, recovering the complete low-rank matrix through the correlation of row/column submatrices; this process does not require prior rank information and can handle sparse noise matrices with large condition numbers.
Less Is More: Label Recommendation for Weakly Supervised Point Cloud Semantic Segmentation
Zhiyi Pan (Peking University), Ge Li (Tencent)
SegmentationRecommendation SystemPoint Cloud
🎯 What it does: A weakly supervised point cloud semantic segmentation label recommendation framework is proposed, which includes pre-indication task learning, cross-scene dual-layer clustering for recommending sparse labels, and a location-based LabelAttention module.
Let All Be Whitened: Multi-Teacher Distillation for Efficient Visual Retrieval
Zhe Ma (Zhejiang University), Lei Yang (Ant Group)
RetrievalComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: Utilizing a multi-teacher distillation framework (Whiten-MTD) to transfer knowledge from various large pre-trained retrieval models to a lightweight student model, achieving efficient visual retrieval.
Let There Be Sound: Reconstructing High Quality Speech from Silent Videos
Ji-Hoon Kim (Korea Advanced Institute of Science and Technology), Joon Son Chung (Korea Advanced Institute of Science and Technology)
RecognitionGenerationData SynthesisConvolutional Neural NetworkTransformerFlow-based ModelGenerative Adversarial NetworkVideoAudio
🎯 What it does: This paper proposes a system for reconstructing high-quality speech from silent videos (Lip-to-Speech) by mapping lip movements to speech.
Levenshtein Distance Embedding with Poisson Regression for DNA Storage
Xiang Wei (Tianjin University), Wei Yu (China Mobile Research Institute)
Convolutional Neural NetworkRecurrent Neural NetworkSequentialBiomedical Data
🎯 What it does: A Poisson regression-based Levenshtein distance embedding method is proposed, and the relationship between embedding dimension and approximation accuracy is provided through theoretical analysis.
Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision
Nicholas Lui (Stanford University), Douwe Kiela (Meta AI)
ClassificationGenerationVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Utilize diffusion models to generate facial images of different races and construct a fairness evaluation dataset for the task of occupational classification.
Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection
Hongzhi Gao (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
Object DetectionAutonomous DrivingTransformerContrastive LearningImagePoint Cloud
🎯 What it does: A weakly semi-supervised 3D object detection method based on a teacher-student framework, called Point-DETR3D, is proposed. It utilizes single-point annotations as weak supervision information and enhances detection accuracy through explicit location query initialization, cross-modal deformable RoI fusion, and point-guided self-supervised learning.
Leveraging Local Variance for Pseudo-Label Selection in Semi-supervised Learning
Zeping Min (Peking University), Chengfei Li (TAL Education Group)
ClassificationRecognitionImageAudio
🎯 What it does: In semi-supervised learning, the Local Variance Match (LVM) method is proposed to enhance model performance by incorporating local variance filtering of pseudo-labels.
Leveraging Normalization Layer in Adapters with Progressive Learning and Adaptive Distillation for Cross-Domain Few-Shot Learning
YongJin Yang, Se-Young Yun (KAIST)
Domain AdaptationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a cross-domain few-shot learning framework called ProLAD, which combines an Adapter with normalization layers, progressive learning, and adaptive distillation to adaptively select suitable Adapters for fine-tuning across different domains.
Leveraging Partial Symmetry for Multi-Agent Reinforcement Learning
Xin Yu (Beihang University), Wenjun Wu (Beihang University)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningSequential
🎯 What it does: A theoretical framework for partially symmetric Markov games is proposed, and based on this, an Adaptive Symmetric Exploitation (PSE) framework is designed to enhance the sample efficiency and overall performance of multi-agent reinforcement learning.
LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition
Youbing Hu (Harbin Institute of Technology), Zhijun Li (Harbin Institute of Technology)
ClassificationRecognitionComputational EfficiencyTransformerImage
🎯 What it does: A two-stage Vision Transformer framework called LF-ViT is proposed, which first performs localization on downsampled images. If the confidence is insufficient, it uses Neighborhood Global Class Attention (NGCA) to locate class-discriminative regions in the full-resolution image, and then continues inference only in that region, achieving efficient image classification.
LGMRec: Local and Global Graph Learning for Multimodal Recommendation
Zhiqiang Guo (Huazhong University of Science and Technology), Bin Ruan (Wuhan Digital Engineering Institute)
Recommendation SystemGraph Neural NetworkContrastive LearningMultimodality
🎯 What it does: This paper proposes a new multimodal recommendation framework LGMRec, which can simultaneously capture users' local interests and global interests.
Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing
Han Liu (Dalian University of Technology), Xianchao Zhang (Dalian University of Technology)
ClassificationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A method is proposed for zero-shot and few-shot text classification that generates class anchors through a pre-trained language model without the need for known class samples, thereby reconstructing multi-class tasks into binary classification tasks.
Lifting by Image – Leveraging Image Cues for Accurate 3D Human Pose Estimation
Feng Zhou (Beijing University of Posts and Telecommunications), Peiyang Li (Beijing University of Posts and Telecommunications)
Pose EstimationTransformerImage
🎯 What it does: This paper proposes a two-stage 3D human pose estimation framework that enhances the 'lifting from 2D pose' method using image features;
LimeAttack: Local Explainable Method for Textual Hard-Label Adversarial Attack
Hai Zhu (University of Science and Technology of China), Kai Liu (Lazada)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkRecurrent Neural NetworkTransformerText
🎯 What it does: A text adversarial attack method based on hard labels, LimeAttack, is proposed. This method estimates the importance of words using the local interpretable model-agnostic explanations (LIME), and combines beam search and semantic similarity sampling rules to generate high-quality adversarial samples under a very limited query budget.
Limitations of Face Image Generation
Harrison Rosenberg (University of Wisconsin Madison), Ramya Korlakai Vinayak (University of Wisconsin Madison)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: This study proposes an end-to-end framework that utilizes text-to-image diffusion models (Stable Diffusion v2.1 and Realistic Vision) combined with SEGA to generate a diverse facial dataset with fine-grained social attributes (race, gender, age, beard, glasses, smile), and conducts quantitative evaluation, facial recognition validation, and user experience research on the generated results.
Limited Memory Online Gradient Descent for Kernelized Pairwise Learning with Dynamic Averaging
Hilal AlQuabeh (Mohamed bin Zayed University of Artificial Intelligence), Bin Gu (Jilin University)
OptimizationTabular
🎯 What it does: A lightweight online kernelized pairwise learning algorithm is proposed, combining moving average gradients and random samples, using random Fourier features to approximate the kernel function, achieving O(T) complexity.
Limited Query Graph Connectivity Test
Mingyu Guo (University of Adelaide), Hung Nguyen (University of Adelaide)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A Limited Query Graph Connectivity Test model is proposed, along with a scalable exact algorithm and a heuristic method based on query limits that can be solved on large-scale graphs.
Limited-Supervised Multi-Label Learning with Dependency Noise
Yejiang Wang (Northeastern University), Xingwei Wang (Singapore Institute of Technology)
ClassificationOptimizationSupervised Fine-TuningTabular
🎯 What it does: This paper proposes a finite supervised multi-label learning method MLDN that simultaneously identifies instance-dependent and label-dependent noise, and achieves noise recovery through noise matrix decomposition and manifold constraints.
Linear-Time Algorithms for Front-Door Adjustment in Causal Graphs
Marcel Wienöbst, Maciej Liśkiewicz (University of Luebeck)
Graph
🎯 What it does: Developed a linear time algorithm for finding front-door adjustment sets and their minimized versions, and implemented multi-language code;
Linear-Time Verification of Data-Aware Processes Modulo Theories via Covers and Automata
Alessandro Gianola (Instituto Superior Técnico), Sarah Winkler (Free University of Bozen-Bolzano)
TabularBenchmark
🎯 What it does: A linear temporal logic (LTLf) verification method for Data-Aware Processes Modulo Theories (DMT) is proposed, constructing a semi-decision procedure and providing several decidable subclasses, unifying and generalizing previous results for specific data types or safety properties.
LINGO-Space: Language-Conditioned Incremental Grounding for Space
Dohyun Kim (Korea Advanced Institute of Science and Technology), Daehyung Park (Korea Advanced Institute of Science and Technology)
Robotic IntelligenceGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: A language condition incremental spatial localization method for composite instructions, LINGO-Space, is proposed, which predicts spatial position distribution using polar coordinate distribution and updates it gradually.
Link Prediction in Multilayer Networks via Cross-Network Embedding
Guojing Ren (Anhui University), Hai-Feng Zhang (Beijing Normal University)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the link prediction problem in multilayer networks and proposes a cross-network embedding model.
LION: Implicit Vision Prompt Tuning
Haixin Wang (Peking University), Qi Tian (Huawei)
ClassificationConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: A lightweight visual prompt tuning framework called LION is proposed, which utilizes two layers of implicit balancing layers to achieve task-specific tuning of pre-trained visual models while keeping the model weights frozen.
LISR: Learning Linear 3D Implicit Surface Representation Using Compactly Supported Radial Basis Functions
Atharva Pandey (Indian Institute of Technology Jodhpur), Santanu Chaudhury (Indian Institute of Technology Jodhpur)
RestorationRepresentation LearningPoint Cloud
🎯 What it does: A linear implicit surface representation (LISR) is proposed using locally supported radial basis functions (CSRBF) to reconstruct three-dimensional shapes from incomplete and noisy 3D point clouds.
Live and Learn: Continual Action Clustering with Incremental Views
Xiaoqiang Yan (Zhengzhou University), Hui Yu (University of Portsmouth)
ClassificationRecognitionObject DetectionOptimizationSupervised Fine-TuningVideo
🎯 What it does: A continuous learning multi-view action clustering method (CAC) is proposed, which can continuously update clustering results and retain historical knowledge as new camera views are gradually added.
LLM vs Small Model? Large Language Model Based Text Augmentation Enhanced Personality Detection Model
Linmei Hu (Beijing Institute of Technology), Liqiang Nie (Beijing University of Posts and Telecommunications)
ClassificationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a method that combines the incremental text generated by large language models (LLMs) from semantic, emotional, and linguistic perspectives with contrastive learning, using LLM-generated label explanations to enhance the performance of smaller models in the task of personality detection in social media posts.
LLMEval: A Preliminary Study on How to Evaluate Large Language Models
Yue Zhang (Fudan University), Xuanjing Huang (Fudan University)
TransformerLarge Language ModelText
🎯 What it does: Constructed the LLMEval dataset to systematically evaluate 20 types of LLMs, comparing manual and automatic evaluations, different scoring methods, evaluator types, and ranking systems.
LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs
Yan Wang (Ant Group), Sheng Li (University of Virginia)
Recommendation SystemGraph Neural NetworkLarge Language ModelPrompt EngineeringSequential
🎯 What it does: Utilize large language models to generate personalized inference graphs and embed them into traditional sequential recommendation models to enhance recommendation performance.
LMD: Faster Image Reconstruction with Latent Masking Diffusion
Zhiyuan Ma (Tsinghua University), Bowen Zhou (Tsinghua University)
RestorationGenerationCompressionTransformerDiffusion modelAuto EncoderImage
🎯 What it does: A framework for image reconstruction based on latent space projection and progressive masking diffusion (LMD) is proposed, integrating VQ-VAE latent projection, MAE encoder-decoder, and a diffusion scheduler with adjustable masking ratios.
Local-Global Multi-Modal Distillation for Weakly-Supervised Temporal Video Grounding
Peijun Bao (Nanyang Technological University), Alex C. Kot (Peng Cheng Laboratory)
RecognitionKnowledge DistillationContrastive LearningOptical FlowVideoMultimodality
🎯 What it does: A local-global multimodal distillation framework is proposed, using a multimodal teacher (RGB + optical flow) to guide a unimodal student to achieve weakly supervised temporal video localization.
Locality Preserving Refinement for Shape Matching with Functional Maps
Yifan Xia (Wuhan University), Jiayi Ma (Wuhan University)
Mesh
🎯 What it does: A two-stage local consistency point-to-point mapping refinement method called LOPR is proposed to eliminate outliers and improve correspondence accuracy in non-rigid shape matching.
Locally Rainbow Paths
Till Fluschnik (TU Clausthal), Malte Renken (Technische Universität Berlin)
🎯 What it does: This paper proposes and studies the algorithmic problems of finding Locally Rainbow Paths and Locally Rainbow Walks in vertex-colored directed graphs, analyzes their computational complexity, and provides various parameterized algorithms and hardness lower bounds.
LogFormer: A Pre-train and Tuning Pipeline for Log Anomaly Detection
Hongcheng Guo (Cloudwise Research), Qi Tian (Huawei)
Anomaly DetectionTransformerTextChain-of-Thought
🎯 What it does: Proposes the LogFormer framework, which includes the Log-Attention module and a two-stage process of pre-training + adapter fine-tuning for cross-domain log anomaly detection.
LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer
Yuxin Cao (Tsinghua University), Jin Lu (CSIRO Data61)
RecognitionAdversarial AttackRecurrent Neural NetworkReinforcement LearningVideo
🎯 What it does: This study proposes a black-box video attack framework called LogoStyleFool, which deceives video recognition models by perturbing stylized logos overlaid in the corners of videos.
Long-Tailed Learning as Multi-Objective Optimization
Weiqi Li (Tianjin University), Wei Feng
OptimizationImageBenchmark
🎯 What it does: This paper proposes a method to transform long-tail classification tasks into multi-objective optimization problems, and achieves collaborative updates of gradients for head and tail categories through Gradient Balancing Grouping (GBG).
Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation
Yuheng Jia (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A long-tail biased label learning method is proposed, which constructs a head classifier and a tail classifier to work collaboratively, utilizing soft pseudo-labels and class distribution estimation to achieve high-quality label disambiguation.
Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains
Yimu Wang (University of Waterloo), Hongyang Zhang (University of Maryland)
Domain AdaptationImageBenchmark
🎯 What it does: This paper studies the impact of the number of training domains on the small population error in the testing domain and presents a theoretical analysis of the loss of domain generalization due to the lack of training domains.
Low Category Uncertainty and High Training Potential Instance Learning for Unsupervised Domain Adaptation
Xinyu Zhang (Jilin University), Shuai Lü
Domain AdaptationContrastive LearningImage
🎯 What it does: An unsupervised domain adaptation method based on low category uncertainty and high training potential instance learning (LUHP) is proposed.
Low-Distortion Clustering with Ordinal and Limited Cardinal Information
Jakob Burkhardt (Aarhus University), Sudarshan Shyam (Aarhus University)
Optimization
🎯 What it does: This paper studies the quantification of distortion in clustering problems given only ranking information and proposes a series of algorithms: providing 2-distortion and 4-distortion for k-center, a constant distortion low-query and zero-query dual-criteria algorithm for (k,z)-clustering, and a constant distortion query algorithm for facility location.
Low-Latency Space-Time Supersampling for Real-Time Rendering
Ruian He (Fudan University), Bo Yan (Fudan University)
Super ResolutionComputational EfficiencyConvolutional Neural NetworkImageVideo
🎯 What it does: A unified spatiotemporal super sampling (STSS) framework is proposed, which can simultaneously address the issues of low resolution and low frame rate in real-time rendering while maintaining low latency.
Low-Light Face Super-resolution via Illumination, Structure, and Texture Associated Representation
Chenyang Wang (Harbin Institute of Technology), Xianming Liu (Harbin Institute of Technology)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: For the super-resolution task of low-resolution facial images under low light conditions, a joint low-light compensation and facial structure recovery framework (IC-FSRNet) is proposed, and based on this, a detail enhancement network (DENet) using a diffusion model is further employed to improve image quality.
Low-Rank Kernel Tensor Learning for Incomplete Multi-View Clustering
Tingting Wu (Beijing Jiaotong University), Jiazheng Yuan (Beijing Open University)
OptimizationTabular
🎯 What it does: A low-rank kernel tensor learning method LRKT-IMVC is proposed, which combines the imputation of missing kernel matrices with multi-view clustering.
LRANet: Towards Accurate and Efficient Scene Text Detection with Low-Rank Approximation Network
Yuchen Su (Fudan University), Yu-Gang Jiang (Fudan University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes LRANet, an efficient arbitrary shape text detection network utilizing low-rank approximation and dual allocation;
LRS: Enhancing Adversarial Transferability through Lipschitz Regularized Surrogate
Tao Wu (Missouri University of Science and Technology), Donald C. Wunsch II (Missouri University of Science and Technology)
Adversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes the Lipschitz Regularized Surrogate (LRS) method, which fine-tunes a pre-trained surrogate model by applying first or second-order Lipschitz regularization to enhance the transferability of adversarial examples, thereby significantly improving the attack success rate without altering existing transfer-based black-box attacks.
LSTKC: Long Short-Term Knowledge Consolidation for Lifelong Person Re-identification
Kunlun Xu (Peking University), Jiahuan Zhou (Peking University)
RecognitionRetrievalKnowledge DistillationConvolutional Neural NetworkImageBenchmark
🎯 What it does: This paper proposes a sample-free lifelong person re-identification method called LSTKC, which addresses the problem of catastrophic forgetting through a short-term knowledge correction module (R-STKT) and a long-term knowledge integration module (E-LTKC).
Lyapunov-Stable Deep Equilibrium Models
Haoyu Chu (Beijing Jiaotong University), Yuto Miyatake (Osaka University)
ClassificationAdversarial AttackConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: A deep equilibrium (DEQ) model called LyaDEQ based on Lyapunov stability has been constructed to enhance robustness against adversarial attacks.
M-BEV: Masked BEV Perception for Robust Autonomous Driving
Siran Chen (Shenzhen Institute of Advanced Technology), Yali Wang (Shenzhen Institute of Advanced Technology)
Object DetectionAutonomous DrivingTransformerImage
🎯 What it does: A Mask BEV (M-BEV) framework is proposed, utilizing random masking and reconstructing camera views to enhance the robustness of multi-camera BEV perception in scenarios of camera failure.
M2Doc: A Multi-Modal Fusion Approach for Document Layout Analysis
Ning Zhang (South China University of Technology), Lianwen Jin (Platform of AI)
Object DetectionSegmentationConvolutional Neural NetworkLarge Language ModelImageTextMultimodalityPhysics Related
🎯 What it does: This paper proposes a pluggable multimodal fusion method M2Doc, which transforms existing unimodal document layout analysis detectors into multimodal detectors, significantly improving layout detection accuracy.
M2SD:Multiple Mixing Self-Distillation for Few-Shot Class-Incremental Learning
Jinhao Lin (South China University of Technology), RongHua Luo (South China University of Technology)
ClassificationKnowledge DistillationImage
🎯 What it does: This paper proposes the M2SD (Multiple Mixing Self-Distillation) method to address the problem of catastrophic forgetting in few-shot incremental learning. It expands the feature space and enhances classification performance through dual-branch virtual class mixing distillation and self-distillation with attention enhancement.
M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy
Hansong Zhang (Institute of Information Engineering, Chinese Academy of Sciences), Shiming Ge (Shanghai University)
ClassificationData SynthesisOptimizationGenerative Adversarial NetworkImage
🎯 What it does: This paper studies the task of dataset condensation and proposes a new method called M3D based on Maximum Mean Discrepancy (MMD), which achieves higher-order moment alignment in Reproducing Kernel Hilbert Space (RKHS) to generate more informative synthetic samples.
M3SOT: Multi-Frame, Multi-Field, Multi-Space 3D Single Object Tracking
Jiaming Liu (Xidian University), Can Qin (Northeastern University)
Object TrackingAutonomous DrivingTransformerPoint Cloud
🎯 What it does: A 3D single-object tracking framework named M3SOT is proposed, which utilizes multi-frame, different receptive fields, and multi-task space to fully exploit the spatiotemporal information of point clouds, enhancing tracking accuracy and robustness in sparse point clouds.
MA-Net: Rethinking Neural Unit in the Light of Astrocytes
Mengqiao Han (Beijing Institute of Technology), Xiabi Liu (Beijing Institute of Technology)
ClassificationObject DetectionSegmentationRecurrent Neural NetworkTransformerImage
🎯 What it does: A multi-astrocyte-neuron (MA-N) model is proposed, and based on this, MA-Net is constructed to achieve dynamically adjustable connections during training, enhancing multi-task performance.
Machine Learning-Powered Combinatorial Clock Auction
Ermis Nikiforos Soumalias (University of Zurich), Sven Seuken (University of Zurich)
OptimizationTabularBenchmark
🎯 What it does: In iterative combinatorial auctions, machine learning models are used to approximate bidders' value functions, and information is obtained through demand queries rather than value queries.
Machine-Created Universal Language for Cross-Lingual Transfer
Yaobo Liang (Microsoft Research Asia), Nan Duan (Microsoft Research Asia)
TransformerContrastive LearningText
🎯 What it does: A general-purpose language MUL was designed and implemented, which is automatically generated by machines and used as a cross-language intermediate language for cross-language transfer;
MagiCapture: High-Resolution Multi-Concept Portrait Customization
Junha Hyung (KAIST), Jaegul Choo (KAIST)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Generate high-resolution portraits with multiple concepts personalized from a small number of reference images.
Make Lossy Compression Meaningful for Low-Light Images
Shilv Cai (Huazhong University of Science and Technology), Xu Zou (Peking University)
RestorationCompressionTransformerAuto EncoderImage
🎯 What it does: An end-to-end joint low-light image compression and enhancement model is proposed, which completes low-light enhancement directly during the compression process, rather than the traditional methods of compressing first and then enhancing or enhancing first and then compressing.
Make Prompts Adaptable: Bayesian Modeling for Vision-Language Prompt Learning with Data-Dependent Prior
Youngjae Cho (Korea Advanced Institute of Science and Technology), Il-Chul Moon (Korea Advanced Institute of Science and Technology)
ClassificationDomain AdaptationPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes an Adaptive Particle Prompt Learning (APP) based on a Bayesian framework for prompt learning in visual-language pre-training models, addressing overfitting and distribution shift issues in few-shot learning.
Make RepVGG Greater Again: A Quantization-Aware Approach
Xiangxiang Chu (Meituan), Bo Zhang (Meituan)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A quantization-friendly improved structure for RepVGG, named QARepVGG, is proposed and implemented, maintaining accuracy close to FP32 under INT8 inference.
Manifold Constraints for Imperceptible Adversarial Attacks on Point Clouds
Keke Tang (Guangzhou University), Zhihong Tian (Guangzhou University)
Adversarial AttackAuto EncoderPoint Cloud
🎯 What it does: A dual manifold mapping based on invertible autoencoders is proposed, which maps 3D point clouds to simple parametric shapes, adding distance and angle constraints in the parametric space to enhance the effectiveness of imperceptible adversarial attacks on point clouds.
Manifold-Based Verbalizer Space Re-embedding for Tuning-Free Prompt-Based Classification
Haochun Wang (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
ClassificationTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: A completely parameter-free prompt-based classification method is proposed, which utilizes Local Linear Embedding with Intra-class Neighborhood Constraints (LLE-INC) to re-embed the [MASK] word vector space, and on this basis, optional contrastive learning can be used to enhance performance.
Manipulation-Robust Selection of Citizens’ Assemblies
Bailey Flanigan (Carnegie Mellon University), Sven Wang (Massachusetts Institute of Technology)
🎯 What it does: This paper studies the manipulability of the citizens' assembly participant selection algorithm when faced with self-reporting features, demonstrating that existing mainstream fair algorithms, Leximin and Nash, are easily manipulable. It proposes the use of the glyph-p norm with a strong convex objective as a new selection algorithm, theoretically providing its rate of decrease in manipulability and experimentally validating its superiority.
MAPTree: Beating “Optimal” Decision Trees with Bayesian Decision Trees
Colin Sullivan (Stanford University), Sebastian Thrun (Stanford University)
ClassificationOptimizationTabular
🎯 What it does: The MAPTree algorithm is proposed, which uses the AND/OR graph search method to find the maximum a posteriori decision tree on the BCART posterior distribution.
Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context
Haochong Xia (Nanyang Technological University), Bo An (Nanyang Technological University)
GenerationData SynthesisRecommendation SystemAnomaly DetectionOptimizationAuto EncoderGenerative Adversarial NetworkTabularTime SeriesFinance Related
🎯 What it does: A financial market data generation framework called Market-GAN based on semantic context is proposed, and a Contextual Market Dataset is constructed.
Mask-Homo: Pseudo Plane Mask-Guided Unsupervised Multi-Homography Estimation
Yasi Wang (Samsung Research China), Qiang Wang (Samsung Research China)
TransformerOptical FlowImage
🎯 What it does: Proposes the Mask-Homo framework, achieving pseudo-plane mask-guided multi-Homography estimation to address alignment issues caused by depth differences among multiple planes.
MaskDiff: Modeling Mask Distribution with Diffusion Probabilistic Model for Few-Shot Instance Segmentation
Minh-Quan Le (University of Science), Minh-Triet Tran (University of Science)
SegmentationDomain AdaptationDiffusion modelImage
🎯 What it does: Proposes the MaskDiff method, which utilizes a conditional diffusion probability model to model the distribution of binary masks in few-shot instance segmentation, achieving more robust mask generation.
MASTER: Market-Guided Stock Transformer for Stock Price Forecasting
Tong Li (Shanghai Jiao Tong University), Sen Huang (Alibaba Group)
TransformerTabularTime SeriesFinance Related
🎯 What it does: A market-guided stock transformer named MASTER is proposed, which captures instantaneous stock price correlations through multi-step aggregated cross moments and achieves automatic feature selection.
Mastering Context-to-Label Representation Transformation for Event Causality Identification with Diffusion Models
Hieu Man (University of Oregon), Thien Huu Nguyen (University of Oregon)
RecognitionTransformerDiffusion modelText
🎯 What it does: A method for event causal identification (ECI) based on diffusion models is proposed, achieving gradual denoising and information extraction in the process of converting context to label representation.
MatchDet: A Collaborative Framework for Image Matching and Object Detection
Jinxiang Lai (Tencent Youtu Lab), Chengjie Wang (Tencent Youtu Lab)
RecognitionObject DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: A collaborative framework called MatchDet is proposed, which jointly learns image matching and object detection to achieve mutual enhancement.
MathAttack: Attacking Large Language Models towards Math Solving Ability
Zihao Zhou (Xi'an Jiaotong-Liverpool University), Kaizhu Huang (ShanghaiTech University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: The MathAttack model is proposed, which utilizes logical entity recognition and word-level attack methods to assess the robustness of large language models' mathematical problem-solving abilities.
Maxileximin Envy Allocations and Connected Goods
Gianluigi Greco (University of Calabria), Francesco Scarcello (University of Calabria)
Graph
🎯 What it does: This paper proposes and studies a new fair allocation scheme—maxileximin allocation, which considers both the minimization of envy (in lexicographic order) and the maximization of social welfare, and systematically analyzes its computational complexity in the context of fair allocation with connectivity constraints (Connected Fair Division);
Maximizing Nash Social Welfare under Two-Sided Preferences
Pallavi Jain (Indian Institute of Technology Jodhpur), Rohit Vaish (Indian Institute of Technology Delhi)
Optimization
🎯 What it does: This paper studies the matching problem of maximizing Nash social welfare under bilateral preferences (workers and firms), proposing various algorithms and complexity conclusions.
Maximizing the Success Probability of Policy Allocations in Online Systems
Artem Betlei (Criteo AI Lab), Benjamin Heymann (Criteo AI Lab)
OptimizationTabular
🎯 What it does: A framework for bid strategy allocation at the user timeline level is proposed, with an optimization objective focused on 'success probability' rather than expected revenue, introducing the SuccessProbaMax algorithm.
MCA: Moment Channel Attention Networks
Yangbo Jiang (Zhejiang University), Nenggan Zheng (Zhejiang University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A channel attention mechanism based on higher-order moment aggregation, called MCA (Moment Channel Attention), is proposed, which enhances the network's expressive capability by utilizing statistical moment information.
MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning
Ying Mo (Beihang University), Zhoujun Li (Meituan)
RecognitionContrastive LearningText
🎯 What it does: A multi-perspective contrastive learning framework MCL-NER is proposed, which transforms the cross-lingual NER task into token pair relationship classification and achieves cross-lingual alignment through contrastive learning of semantics and token-to-token relationships.
MCSSME: Multi-Task Contrastive Learning for Semi-supervised Singing Melody Extraction from Polyphonic Music
Shuai Yu (Donghua University)
Contrastive LearningAudio
🎯 What it does: A semi-supervised singing melody extraction framework called MCSSME based on multi-task contrastive learning is proposed.
MDFL: Multi-Domain Diffusion-Driven Feature Learning
Daixun Li (Xidian University), Yunsong Li (Xidian University)
ClassificationDiffusion modelImageMultimodality
🎯 What it does: This paper proposes a multi-domain diffusion-driven feature learning framework (MDFL) for joint feature extraction and classification of high-dimensional remote sensing images.
MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction
Hao Qian (Ant Group), Jun Zhou (Ant Group)
Recommendation SystemOptimizationGraph Neural NetworkTransformerGraphTime SeriesFinance Related
🎯 What it does: This paper proposes a Multi-Relational Dynamic Graph Neural Network (MDGNN) that constructs daily multi-relational graphs (including stock-stock, stock-bank-stock, stock-industry-industry-stock meta-paths, etc.) to capture the multidimensional relationships and temporal evolution between stocks using hierarchical graph attention embedding and Transformer, aimed at predicting the intraday price fluctuation probabilities of stocks.
Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework
Weixi Weng (Tsinghua University), Chun Yuan (Tsinghua University)
Object DetectionDomain AdaptationTransformerGenerative Adversarial NetworkImage
🎯 What it does: A two-stage domain adaptation object detection framework MTM based on Mean Teacher DETR is proposed, which utilizes target-similar domain images for pre-training, and then uses pseudo-labels and introduces OQKT during the self-training phase; at the same time, two masking feature alignment strategies, Masked Domain Query-based Feature Alignment (MDQFA) and Masked Token-Wise Feature Alignment (MTWFA), are designed to alleviate domain shift;
Measuring Self-Supervised Representation Quality for Downstream Classification Using Discriminative Features
Neha Kalibhat (University of Maryland), Soheil Feizi (Meta AI)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Analyzed the representation space of self-supervised learning models, discovered unsupervised discriminative features, and proposed Q-Score to evaluate and enhance representation quality.
Measuring Task Similarity and Its Implication in Fine-Tuning Graph Neural Networks
Renhong Huang (Zhejiang University), Yang Yang (FinVolution Group)
Graph Neural NetworkContrastive LearningGraph
🎯 What it does: This study investigates the similarity between graph pre-training models and downstream tasks, proposing a task consistency metric and a Bridge-Tune fine-tuning strategy based on this metric.
MedBench: A Large-Scale Chinese Benchmark for Evaluating Medical Large Language Models
Yan Cai (East China Normal University), Liang He (East China Normal University)
TransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsBenchmarkChain-of-Thought
🎯 What it does: MedBench is proposed — a large Chinese medical benchmark containing 40,041 questions, covering three stages of medical examinations (National Physician Qualification, Resident Physician Standardized Training, Physician Practice Qualification) as well as over 2,000 real electronic health record question-answer pairs, aimed at uniformly assessing the knowledge mastery and reasoning ability of large medical language models.