AAAI Conference on Artificial Intelligence · 1014 papers
PG-LBO: Enhancing High-Dimensional Bayesian Optimization with Pseudo-Label and Gaussian Process Guidance
Taicai Chen (Nanjing University), Yang Gao (Nanjing University)
CodeOptimizationAuto EncoderTabular
🎯 What it does: The PG-LBO method is proposed, which improves the potential space construction of VAE-BO in high-dimensional structural optimization using pseudo-labels and GP guidance.
🎯 What it does: A one-click voice conversion method called Phoneme Hallucinator is proposed, which utilizes a small amount of target speaker audio to generate diverse speaker features, thereby achieving one-shot VC.
PICNN: A Pathway towards Interpretable Convolutional Neural Networks
Wengang Guo (Tongji University), Wei Ye (Tongji University)
CodeExplainability and InterpretabilityConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Design a Pluggable Interpretation Path (PICNN) that allows standard CNNs to group the convolutional filters in the later layers into clusters corresponding to each category, thus making it directly interpretable while maintaining recognition ability.
🎯 What it does: A pluggable extension framework PDRec is designed, which uses diffusion models as plugins to generate global user preferences, and combines historical behavior reweighting, positive sample augmentation, and noise-free negative sampling to enhance sequential recommendation performance.
Xiaopeng Li (National University of Defense Technology), Jie Yu (National University of Defense Technology)
CodeTransformerLarge Language ModelText
🎯 What it does: A precise model editing method called PMET is proposed, which utilizes the joint optimization of the multi-head self-attention (MHSA) and feed-forward network (FFN) hidden states in Transformers. It precisely updates the FFN weights using only the optimized FFN hidden states, achieving knowledge editing without changing the MHSA weights.
PoetryDiffusion: Towards Joint Semantic and Metrical Manipulation in Poetry Generation
Zhiyuan Hu (National University of Singapore), Bryan Hooi (Nanyang Technological University)
CodeGenerationDiffusion modelText
🎯 What it does: A poetry generation framework called PoetryDiffusion based on diffusion models has been constructed, which can achieve precise control over form (line count, rhyme, tonal patterns, etc.) and phonetics while maintaining semantic coherence.
Poincaré Differential Privacy for Hierarchy-Aware Graph Embedding
Yuecen Wei (Beihang University), Chunming Hu (Guangxi Normal University)
CodeSafty and PrivacyRepresentation LearningGraph Neural NetworkGaussian SplattingGraph
🎯 What it does: This paper proposes the PoinDP framework, which implements differential privacy protection for hierarchical-aware graph embedding in hyperbolic (Poincaré) space.
Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition
Hanxuan Li (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
CodeRecognitionDomain AdaptationPrompt EngineeringVision Language ModelPoint CloudMesh
🎯 What it does: By first converting point clouds into meshes, then performing realistic texture rendering and selecting the most recognizable viewpoints, the frozen CLIP model is utilized to achieve zero-shot and few-shot 3D recognition for unknown categories.
🎯 What it does: A point cloud completion framework called PointAttN based on Transformer has been designed and implemented, utilizing cross-attention (GDP unit) and self-attention (SFA unit) to achieve generation from incomplete point clouds to complete point clouds without explicit local neighborhood partitioning.
🎯 What it does: This paper proposes a multi-stage end-to-end multi-instance detection network called Polyper, based on morphological segmentation and internal multi-scale feature attention, for accurately segmenting the contours of colorectal polyps.
PORTAL: Automatic Curricula Generation for Multiagent Reinforcement Learning
Jizhou Wu (Tianjin University), Matthew E. Taylor (University of Alberta)
CodeReinforcement Learning
🎯 What it does: A multi-agent automatic curriculum learning framework named PORTAL is proposed, which utilizes dual indicators of task difficulty and task similarity to automatically select intermediate tasks, helping agents converge faster on extremely difficult collaborative tasks.
🎯 What it does: Utilizing Beltrami flow in graph neural partial differential equations to achieve dual embedding of point cloud features and positional information, and completing the correspondence at the window-patch-point three-level hierarchy through a neural ODE-driven Transformer, ultimately resulting in high-precision rigid registration outcomes.
PPIDSG: A Privacy-Preserving Image Distribution Sharing Scheme with GAN in Federated Learning
Yuting Ma (University of Science and Technology of China), Xiaohua Xu (University of Science and Technology of China)
CodeGenerationFederated LearningSafty and PrivacyAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A privacy protection scheme PPIDSG is proposed in federated learning that does not upload classifier parameters. It utilizes block scrambling encryption, GAN image distribution sharing, and local classification training to defend against image reconstruction, label inference, and membership inference attacks.
PPO-Clip Attains Global Optimality: Towards Deeper Understandings of Clipping
Nai-Chieh Huang (National Yang Ming Chiao Tung University), I-Chen Wu (National Yang Ming Chiao Tung University)
CodeOptimizationReinforcement Learning
🎯 What it does: This paper studies the global convergence of the PPO-Clip algorithm, establishing for the first time the global convergence results of the PPO-Clip variant in both tabular and neural network function approximation settings.
Predicting Real-World Penny Auction Durations by Integrating Game Theory and Machine Learning
Yujia Wang (Beijing Institute of Technology), Haoran Yu (Beijing Institute of Technology)
CodeMixture of ExpertsTabular
🎯 What it does: A three-stage framework (ADAPT) is proposed, which first predicts auction duration using game theory, then utilizes a multi-branch mixture density network combined with product description embeddings and game theory predictions, and finally outputs the actual auction duration distribution.
🎯 What it does: This paper studies the unlabeled video face pre-training framework PrefAce, which utilizes multi-scale keypoint-guided self-distillation and instance-level updates of the FaceFeat Cache to learn general and transferable face representations.
PREFER: Prompt Ensemble Learning via Feedback-Reflect-Refine
Chenrui Zhang (Meituan Inc), Mingchen Cai (Meituan Inc)
CodeClassificationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: An automatic and jointly optimized Prompt integration framework called PREFER is proposed, which generates and optimizes Prompts through an iterative feedback-reflection-refinement loop, and enhances stability by combining forward and backward bidirectional bagging.
🎯 What it does: The PAD-CLRec model is proposed, which utilizes contrastive learning to align user preferences, content features, and collaborative features, addressing the cold start recommendation problem.
🎯 What it does: This paper studies a two-stage pre-routing timing prediction model, PreRoutGNN, which can accurately estimate timing metrics (slack, slope, net delay, cell delay) before routing, thereby accelerating the IC design process.
Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring
Peng Tang (Technical University of Munich), Tobias Lasser (Technical University of Munich)
CodeRestorationTransformerImage
🎯 What it does: This paper proposes a single image defocus deblurring method called P2IKT, which utilizes prior knowledge and predicted inverse kernel transformation to achieve a divide-and-conquer deblurring approach.
🎯 What it does: A probabilistic neural circuit (PNC) is proposed, which can be interpreted as a mixture of deep Bayesian networks, along with its hierarchical construction and executable queryable form.
Probabilistic Offline Policy Ranking with Approximate Bayesian Computation
Longchao Da (Arizona State University), Hua Wei (Arizona State University)
CodeReinforcement LearningAgentic AITabular
🎯 What it does: The Probabilistic Offline Policy Ranking (POPR) framework is proposed, which constructs the posterior performance distribution of candidate policies using expert data to achieve offline policy ranking; at the same time, the POPR-EABC (Energy-based Approximate Bayesian Computation) method is introduced to estimate the posterior distribution.
Progressive Distillation Based on Masked Generation Feature Method for Knowledge Graph Completion
Cunhang Fan (Anhui University), Zhao Lv (Tsinghua University)
CodeCompressionKnowledge DistillationTransformerLarge Language ModelGraph
🎯 What it does: A progressive distillation method based on mask-generated features (PMD) is proposed to compress the parameter size of pre-trained language models in knowledge graph completion tasks while maintaining or even improving performance.
🎯 What it does: A backdoor defense framework called PIPD is proposed, which utilizes progressive data poisoning isolation and selective training to build a clean model.
Prompt to Transfer: Sim-to-Real Transfer for Traffic Signal Control with Prompt Learning
Longchao Da (Arizona State University), Hua Wei (Arizona State University)
CodeAutonomous DrivingOptimizationGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTabularTime SeriesChain-of-Thought
🎯 What it does: This paper proposes PromptGAT, a grounded action transformation method that utilizes large language models (LLM) for sim-to-real transfer in traffic signal control (TSC) tasks, enabling better modeling of road dynamics without the need to collect large amounts of real data.
Prompt-Based Distribution Alignment for Unsupervised Domain Adaptation
Shuanghao Bai (Xi'an Jiaotong University), Badong Chen (Xi'an Jiaotong University)
CodeDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A Prompt-based Distribution Alignment (PDA) method is proposed, which utilizes two branches (the base branch and the alignment branch) for prompt learning in visual-language models, achieving unsupervised domain adaptation.
🎯 What it does: By constructing Semantic-Aware Audio Prompts (SAP) and Correlation Adapters (ColA), audio information is injected as prompts into the visual foundation model SAM, enabling audio-driven semantic segmentation.
🎯 What it does: An adaptive graph contrastive learning framework RAGCL is proposed, targeting the wide structural characteristics of rumor propagation trees, utilizing node centrality to guide data augmentation to highlight subtrees with high information content.
Prot2Text: Multimodal Protein’s Function Generation with GNNs and Transformers
Hadi Abdine (Ecole Polytechnique), Michalis Vazirgiannis (Ecole Polytechnique)
CodeGenerationProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelMultimodality
🎯 What it does: The Prot2Text model is proposed, which transforms the protein function prediction task into free text generation, utilizing multimodal information (protein sequences, 3D structures, text annotations) to generate detailed functional descriptions.
ProxyDet: Synthesizing Proxy Novel Classes via Classwise Mixup for Open-Vocabulary Object Detection
Joonhyun Jeong (NAVER Cloud), Heesu Kim (Korea Advanced Institute of Science and Technology)
CodeObject DetectionVision Language ModelImage
🎯 What it does: By using linear combinations of base classes to generate proxy new categories during the training phase, and training the object detector with proxy loss, the generalization ability of open vocabulary object detection is enhanced.
PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction
Lirong Wu (Westlake University), Stan Z. Li (Westlake University)
CodeDrug DiscoveryRecurrent Neural NetworkGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: This paper proposes a multi-scale protein sequence-structure contrastive learning framework (PSC-CPI) for predicting the interaction patterns and affinities between compounds and proteins, addressing the issues of modality missing and domain transfer.
🎯 What it does: This paper proposes a post-training multi-bit quantization framework (PTMQ) that enables real-time switching between different bit widths with just one calibration, supporting both uniform and mixed precision.
🎯 What it does: This paper proposes a method called PTUS for 'talking upper-body synthesis' that can simultaneously generate realistic upper body and facial animations from source images and driving videos.
QAGait: Revisit Gait Recognition from a Quality Perspective
Zengbin Wang (Beijing University of Posts and Telecommunications), Shibiao Xu (Beijing University of Posts and Telecommunications)
CodeRecognitionImageVideo
🎯 What it does: This paper proposes the QAGait framework, which first conducts quality assessment of gait profiles (maximum connected area, template matching), followed by tilt-aware alignment and data augmentation, and finally introduces quality-aware QACE and QATriplet losses during the feature learning phase to achieve full-process control of gait recognition quality.
🎯 What it does: This paper upgrades the existing quantized compressed sensing algorithm QCS-SGM to QCS-SGM+, achieving effective recovery of general (non-orthogonal) measurement matrices.
QI-IRA: Quantum-Inspired Interactive Ranking Aggregation for Person Re-identification
Chunyu Hu (Wuhan University), Hao Huang (Wuhan University)
CodeRecognitionRetrievalImageVideo
🎯 What it does: A quantum-inspired interactive ranking aggregation (QI-IRA) method is proposed for multi-method ranking aggregation in person re-identification tasks.
🎯 What it does: A generative sampling framework QDGS based on quality diversity optimization is proposed, which uses language prompts to guide the generator to uniformly cover the multi-dimensional attribute space in the latent space, generating balanced synthetic data.
🎯 What it does: Created the QuerySum multi-document query-focused summarization dataset, which includes 27,041 non-factual (What/How/Why) question-answer samples, and provides similar query clusters and corresponding summaries for each query.
🎯 What it does: This paper proposes a Transformer network that utilizes millimeter-wave radar point clouds to simultaneously achieve moving object segmentation (MOS) and ego-vehicle velocity estimation (EVE).
🎯 What it does: A memory-oriented continual learning framework is proposed, which uses a working memory network to learn new tasks and recalls old knowledge at the parameter level through a Generative Adversarial Meta Model (GAMM), addressing the stability-plasticity dilemma.
🎯 What it does: A theoretical analysis of motion blur caused by ultra-high-speed motion is proposed, and a robust peak representation method is designed based on spatiotemporal context learning, along with the construction of the first real ultra-high-speed peak recognition dataset UHSR.
Reconciling Predictive and Statistical Parity: A Causal Approach
Drago Plecko (Columbia University), Elias Bareinboim (Columbia University)
CodeTabular
🎯 What it does: This paper unifies the conflict between Statistical Parity (SP) and Predictive Parity (PP) through causal decomposition methods and proposes a spectrum of 'Business Necessity' (BN), providing methods for judgment and implementation.
Sofia Lemons (University of New Hampshire), Carlos Linares Lopez (Universidad Carlos III de Madrid)
CodeOptimization
🎯 What it does: A new anytime heuristic search algorithm called Rectangle Search is proposed, which is based on beam search with increasing width/depth exploration, aimed at solving deep local minimum problems.
🎯 What it does: This study addresses the issue of missing modalities and the imbalance of different missing rates in multimodal learning, proposing the RedCore model to achieve cross-modal representation learning and adaptively adjust the supervisory weights of each modality.
🎯 What it does: A diffusion model distillation framework called SFERD is proposed, which is based on attention guidance and semantic gradient prediction, to reduce the spatial fitting error between the teacher and student models, thereby achieving high-quality image generation in very few steps (even a single step).
🎯 What it does: This paper proposes MUTR, a unified multimodal temporal Transformer for video object segmentation based on language or audio instructions.
🎯 What it does: This paper studies the conflict between the graph neural network (GNN) encoder and the InfoNCE loss in graph contrastive learning (GCL), and proposes the ReGCL framework to alleviate this conflict through gradient-guided structural learning and gradient-weighted InfoNCE, thereby improving the quality of node representations.
🎯 What it does: Designed the Region-aware Exposure Correction Network (RECNet), which simultaneously corrects mixed exposure images through a region-aware de-exposure module and a multi-scale recovery unit.
Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation
Debaditya Shome (Queen's University), Ali Etemad (Queen's University)
CodeGenerationData SynthesisDiffusion modelTime SeriesBiomedical DataElectronic Health RecordsElectrocardiogramBenchmark
🎯 What it does: A regional decoupled diffusion model (RDDM) is proposed to convert photoplethysmography (PPG) signals into high-fidelity electrocardiogram (ECG) signals.
REGLO: Provable Neural Network Repair for Global Robustness Properties
Feisi Fu (Boston University), Wenchao Li (Boston University)
CodeOptimizationAdversarial AttackTabularFinance Related
🎯 What it does: A neural network repair framework called REGLO is proposed, which is based on global robustness properties. It allows for post-hoc weight adjustments of pre-trained models to meet given global robustness constraints (including individual fairness).
Fengpeng Li (University of Macau), Jiantao Zhou (University of Macau)
CodeClassificationImage
🎯 What it does: This paper proposes the Regroup Median Loss (RML) method for robustly estimating sample loss under label noise conditions, and further constructs a semi-supervised learning framework based on RML to enhance the model's generalization performance on noisy data.
Ziyang Li (University of Pennsylvania), Mayur Naik (University of Pennsylvania)
CodeRetrievalRecommendation SystemTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper presents VIEIRA, a declarative framework based on relational programming, designed to unify the invocation and combination of multimodal foundational models (such as GPT, CLIP, SAM, etc.), enabling seamless interaction and reasoning from text, images to structured databases.
Relax Image-Specific Prompt Requirement in SAM: A Single Generic Prompt for Segmenting Camouflaged Objects
Jian Hu (Queen Mary University of London), Weitong Cai (Queen Mary University of London)
CodeObject DetectionSegmentationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: An unsupervised CAMO detection method called GenSAM is proposed, which utilizes general task descriptions to automatically generate image-specific visual prompts for the segmentation of concealed objects.
Cai Xu (Xidian University), Xiyue Gao (Xidian University)
CodeClassificationRecognitionImageVideo
🎯 What it does: This paper proposes the Reliable Conflict Multi-Perspective Learning (RCML) problem and designs an evidence theory-based conflict opinion aggregation method, ECML, to provide reliable decisions and their confidence in the presence of conflicting perspectives in the data.
CodeGenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Proposes the Self-RDGS method, which utilizes LLM to automatically generate sentences and selects reliable training data through ranking to construct a high-quality training set for low-resource relation extraction tasks.
Reproduce, Replicate, Reevaluate. The Long but Safe Way to Extend Machine Learning Methods
Luisa Werner (Univ. Grenoble Alpes), Damien Graux (Trinity College Dublin)
CodeGraph Neural NetworkGraph
🎯 What it does: Using a three-step process of reproduction-reimplementation-reevaluation, we systematically reproduce, reimplement, and cross-evaluate the Knowledge Enhanced Neural Network (KENN) experiments on Citation graphs, summarizing experiences and obstacles.
Resisting Backdoor Attacks in Federated Learning via Bidirectional Elections and Individual Perspective
Zhen Qin (Zhejiang University), Shuiguang Deng (Zhejiang University)
CodeFederated LearningAuto EncoderImage
🎯 What it does: A dual-election-based backdoor attack framework called Snowball is proposed, which utilizes individual perspectives to vote on model updates and perform incremental filtering, and employs VAE to learn model differences to further exclude malicious updates.
🎯 What it does: A residual attention learning framework ResMatch and its sparse version sResMatch are designed to improve the accuracy of image feature matching.
Responding to the Call: Exploring Automatic Music Composition Using a Knowledge-Enhanced Model
Zhejing Hu (Hong Kong Polytechnic University), Qianwen Luo (Shenzhen University)
CodeGenerationTransformerAudio
🎯 What it does: This paper studies the application of call-and-response in automatic music creation, constructing the first CRD dataset containing 19,155 call-and-response pairs, and proposes a knowledge-enhanced Seq2Seq generator (CRG) to produce creatively responsive music.
🎯 What it does: This study proposes a Dimensional Rationale framework in graph contrastive learning, utilizing a learnable dimension weight network and redundancy reduction regularization, and employs meta-learning for backdoor adjustment to eliminate irrelevant information, enhancing the discriminability and transferability of graph representations.
🎯 What it does: This paper establishes a connection between Graph Masked Autoencoders (GraphMAE) and Graph Contrastive Learning (GCL) through theoretical analysis, pointing out their limitations in alignment and uniformity, and subsequently proposes the AUG-MAE model to overcome these issues.
🎯 What it does: The researchers introduced momentum (ordinary momentum and positive-negative momentum) into the diffusion model to escape local minima, thereby reducing the probability of generating singular images.
Rethinking Propagation for Unsupervised Graph Domain Adaptation
Meihan Liu (Zhejiang University), Jiajun Bu (Tsinghua University)
CodeDomain AdaptationGraph Neural NetworkGraph
🎯 What it does: A heterogeneous A2GNN model is proposed, which re-evaluates the role of the propagation layer in GNNs for unsupervised graph domain adaptation, providing both theoretical and experimental support.
Sandesh Kamath (Indian Institute of Technology), Vineeth N Balasubramanian (Microsoft Research)
CodeExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper addresses the robustness issue of model interpretation (feature importance mapping) by proposing more reasonable evaluation metrics and improvement methods, and conducting experimental validation on different datasets and models.
🎯 What it does: A point-based two-stage referential expression understanding framework is proposed, transforming the localization and understanding process into point-level cross-modal understanding and point-level localization.
Reverse Multi-Choice Dialogue Commonsense Inference with Graph-of-Thought
Li Zheng (Wuhan University), Chong Teng (Wuhan University)
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A three-step reverse exclusion graph thinking framework (ReX-GoT) is proposed for dialogue commonsense multiple-choice question answering.
🎯 What it does: A method to address the issue of minority class forgetting in domain adaptation black-box predictors, called RFC, is proposed, which includes two modules: Selective Training (ST) and Neighborhood Clustering (NC).
🎯 What it does: This study investigates the necessity of causal disentangled representations for downstream performance in abstract visual reasoning tasks.
Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction
Monika Jain (Indraprastha Institute of Information Technology), Kuldeep Singh (Cerence)
CodeExplainability and InterpretabilityKnowledge DistillationGraph Neural NetworkText
🎯 What it does: This paper reframes the document-level relation extraction task as knowledge graph link prediction and improves relation prediction performance by integrating document reasoning with external knowledge (Wikidata, WordNet).
Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum
Fan Xu (Beijing Jiaotong University), Hai Wan (Tsinghua University)
CodeAnomaly DetectionOptimizationGraph Neural NetworkContrastive LearningGraphFinance Related
🎯 What it does: A semi-supervised graph neural network named SEC-GFD is proposed to address the highly heterogeneous and class-imbalanced fraud detection task.
🎯 What it does: This paper proposes an unconstrained reinforcement learning framework that incorporates cumulative costs in the state space and applies reward penalties to trajectories that violate constraints, effectively addressing safety reinforcement learning problems under various constraints (expected cost, Value-at-Risk, Conditional Value-at-Risk, worst-case). This framework is embedded in DQN and SAC, resulting in two algorithms: Safe DQN and Safe SAC.
🎯 What it does: A GAN training method called RG-GAN is proposed, which enhances generation quality and reduces model parameters through dynamic weight pruning and regeneration (recursive regeneration and incremental regeneration).
🎯 What it does: An algorithm named RLfOLD is proposed, which trains driving strategies in the CARLA urban driving environment using online demonstrations combined with reinforcement learning.
Robust Evaluation Measures for Evaluating Social Biases in Masked Language Models
Yang Liu (Tianjin University)
CodeTransformerLarge Language ModelText
🎯 What it does: This paper proposes the use of Gaussian distribution to represent the pseudo-log-likelihood (PLL) score of PLM, and constructs two new bias assessment metrics, KLS and JSS, based on KL and JS divergence, aimed at measuring the social bias of masked language models (MLM).
🎯 What it does: A robust two-stage few-shot named entity recognition method based on boundary discrimination and correlation purification, called BDCP, is proposed to resist text adversarial attacks.
Robust Loss Functions for Training Decision Trees with Noisy Labels
Jonathan Wilton (University of Queensland), Nan Ye (University of Queensland)
CodeClassificationOptimizationImageTabular
🎯 What it does: This study proposes a robust loss function for training decision trees with noisy labeled data, introducing a conservative loss and distribution loss framework, and designing a negative exponential (NE) loss based on this, achieving an analyzable impurity calculation.
🎯 What it does: This paper proposes a robust visual recognition method in an environment with class imbalance and the coexistence of open-world noise, primarily addressing the issues of label noise correction and sample selection.
Robustly Train Normalizing Flows via KL Divergence Regularization
Kun Song (Mohamed bin Zayed University of Artificial Intelligence), Fakhri Karray (Mohamed bin Zayed University of Artificial Intelligence)
CodeAnomaly DetectionFlow-based ModelImage
🎯 What it does: This paper proposes a training method based on KL divergence regularization, enhancing the robustness of regularization, especially for datasets with insufficient samples or the presence of outliers.
Root Cause Analysis in Microservice Using Neural Granger Causal Discovery
Cheng-Ming Lin (National Yang Ming Chiao Tung University), Wen-Chih Peng (National Yang Ming Chiao Tung University)
CodeAnomaly DetectionContrastive LearningTime Series
🎯 What it does: A root cause analysis framework RUN based on neural Granger causality discovery and contrastive learning is proposed for root cause localization of anomalies in microservice systems.
s-ID: Causal Effect Identification in a Sub-population
Amir Mohammad Abouei (École Polytechnique Fédérale de Lausanne), Negar Kiyavash (École Polytechnique Fédérale de Lausanne)
Code
🎯 What it does: A new problem of identifying causal effects in subpopulations (subsamples affected by selection bias) is proposed (S-ID), and its necessary and sufficient conditions are provided.
🎯 What it does: A spatial-spectral bidirectional cyclic diffusion framework S CycleDiff is proposed for high-resolution hyperspectral image reconstruction.
S2WAT: Image Style Transfer via Hierarchical Vision Transformer Using Strips Window Attention
Chiyu Zhang (Nanjing University of Aeronautics and Astronautics), Jun Yang (Sichuan Normal University)
CodeImage TranslationTransformerImage
🎯 What it does: This paper proposes an image style transfer framework S2WAT based on hierarchical visual Transformer, utilizing Strips Window Attention (SpW Attention) combined with multi-scale window attention and dynamically merging through Attn Merge, successfully addressing the localization (grid-like) problem caused by traditional window attention, achieving style transfer that captures both local and global features simultaneously.
S3A: Towards Realistic Zero-Shot Classification via Self Structural Semantic Alignment
Sheng Zhang (Mohamed bin Zayed University of Artificial Intelligence), Fahad Shahbaz Khan (Linköping University)
CodeClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: The Self Structural Semantic Alignment (S3A) framework is proposed, utilizing unsupervised image clustering, voting, LLM prompting, and realignment self-learning to address the real zero-shot classification problem under large vocabularies.
🎯 What it does: A spatially aligned and adaptive visual prompt model SA VP 2 is proposed, which aligns a two-dimensional prompt map with the image feature map to improve visual prompt tuning.
🎯 What it does: A sample reconstruction-based model extraction attack detection method (SAME) is proposed, which can identify malicious queries without relying on additional OOD data, user query records, or access to white-box models, and can be used in conjunction with active defense strategies.
Sample Efficient Reinforcement Learning with Partial Dynamics Knowledge
Meshal Alharbi (Massachusetts Institute of Technology), Munther Dahleh (Massachusetts Institute of Technology)
CodeOptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the sample complexity problem in online reinforcement learning with known or efficiently learnable partial dynamics knowledge (in the form of additive disturbance). It proposes an optimization Q-learning algorithm based on UCB (UCB-f) that achieves a √T regret independent of S and A when the dynamics are fully known; when only noise estimates are available, the sample complexity remains independent of the state and action dimensions and is related to the approximation error of f and the Lipschitz constant of the value function.
🎯 What it does: A method for incomplete multi-view clustering based on sample-level cross-view similarity learning (SCSL) is proposed, which can construct a complete similarity matrix between all sample pairs and perform spectral clustering.
SAT-Based Algorithms for Regular Graph Pattern Matching
Miguel Terra-Neves (OutSystems), Antonio Alegria (Zharta)
CodeGraph Neural NetworkGraph
🎯 What it does: A SAT-based Regular Graph Pattern Matching (ReGaP) algorithm is proposed, which can check complex structural properties in graphs through declarative specifications, extending the traditional graph isomorphism problem.
🎯 What it does: A SAVSR network is designed to achieve video super-resolution at arbitrary scales (non-integer and non-symmetric ratios) under a single model.
🎯 What it does: A framework called SC-NeuS is proposed, which can jointly learn neural implicit surfaces and camera poses from sparse and noisy camera poses.
🎯 What it does: This paper studies a Boolean network trap space enumeration method based on answer set programming (tsconj), achieving efficient computation of both minimum and maximum trap spaces.
🎯 What it does: A scalable geometric fracture assembly framework is proposed, utilizing a collaborative creation space to allow multiple assemblers to gradually complete the assembly, and introducing geometric conflict loss to avoid local optima and collision issues.
🎯 What it does: An evolutionary reinforcement learning agent (EVORL) combined with a coarse-fine detection framework is used to adaptively optimize the target scale in drone aerial images to improve detection accuracy.
Scaling and Masking: A New Paradigm of Data Sampling for Image and Video Quality Assessment
Yongxu Liu (Xidian University), Jinjian Wu (Xidian University)
CodeTransformerImageVideo
🎯 What it does: This paper proposes a sampling method based on scaling and masking, called SAMA, which captures both local details and global semantics of images/videos while maintaining a single-branch model.
ScanERU: Interactive 3D Visual Grounding Based on Embodied Reference Understanding
Ziyang Lu (University of Electronic Science and Technology of China), Heng Tao Shen (University of Science and Technology of China)
CodeRecognitionObject DetectionTransformerVision Language ModelTextPoint Cloud
🎯 What it does: The task of 'Embodied Reference Understanding (ERU)' is proposed, which utilizes natural language and human gestures to jointly locate target objects in 3D point cloud scenes, and based on this, a complete system from data collection, synthesis to model training is designed.
🎯 What it does: A skeleton action recognition framework SCD-Net based on self-supervised contrastive learning is proposed, which utilizes a dual-path decoupled encoder to extract pure spatial and temporal features from skeleton sequences and achieves cross-domain comparison through global anchors.