AAAI Conference on Artificial Intelligence Β· 1442 papers
Multi-Objective Evolution of Heuristic Using Large Language Model
Shunyu Yao (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelTabular
π― What it does: The MEoH framework is proposed, which combines large language models (LLM) with multi-objective evolutionary algorithms to automatically generate a set of non-dominated heuristic algorithms that satisfy multiple objectives (optimal gap and execution efficiency).
Multi-Objective Molecular Design Through Learning Latent Pareto Set
Yiping Liu (Hunan University), Hisao Ishibuchi (Southern University of Science and Technology)
CodeOptimizationDrug DiscoveryAuto EncoderGraph
π― What it does: A multi-objective molecular design framework named MLPS is proposed, which utilizes an encoder-decoder to map discrete chemical space to continuous latent space, and combines local Bayesian optimization with a global Pareto set learning model to achieve high-quality molecular generation for multiple objectives (such as QED, SA, GSK3Ξ², JNK3) as well as dual objectives of antifungal/toxicity.
Multi-Perspective Consolidation Enhanced Cognitive Diagnosis via Conditional Diffusion Model
Guanhao Zhao (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeTransformerDiffusion modelSequential
π― What it does: A multi-perspective integrated cognitive diagnosis framework (DMC-CDM) is proposed, which accurately captures cognitive states from a single perspective through a semantic extractor and a conditional diffusion model, and achieves information integration by maximizing mutual information across multiple perspectives.
Multi-Reference Preference Optimization for Large Language Models
Hung Le (Applied AI Institute), Svetha Venkatesh (Deakin University)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: A framework for Direct Preference Optimization using Multiple Reference Models (MRPO) is proposed to enhance the performance of large language models in various preference learning tasks.
π― What it does: A global game mechanism based on negative feedback is proposed for the multi-robot collaborative maintenance of energy supply in community task allocation problems.
Multi-Shape Matching with Cycle Consistency Basis via Functional Maps
Yifan Xia (Wuhan University), Jiayi Ma (Wuhan University)
CodeOptimizationGraph Neural NetworkMeshBenchmark
π― What it does: A multi-shape matching method based on functional mapping is proposed, achieving a balance between matching accuracy and consistency through a two-stage optimization of graph structures.
Multi-Teacher Knowledge Distillation with Reinforcement Learning for Visual Recognition
Chuanguang Yang (Institute of Computing Technology, Chinese Academy of Sciences), Yongjun Xu (Institute of Computing Technology, Chinese Academy of Sciences)
π― What it does: A multi-teacher knowledge distillation framework based on reinforcement learning, MTKD-RL, is proposed, which dynamically generates teacher weights through agents to achieve better balance and integration of multi-teacher information.
CodeRecognitionTransformerContrastive LearningMultimodalityTime Series
π― What it does: A multi-modal to single-modal emotion recognition framework (M2S) is proposed, which achieves high performance in cross-modal and single-modal tasks by pre-training on unlabeled multi-modal data and then fine-tuning on a single modality.
Multi-View Empowered Structural Graph Wordification for Language Models
Zipeng Liu (Tianjin University), Nan Feng (Tianjin University)
CodeClassificationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningGraph
π― What it does: An end-to-end framework called Dr.E is proposed, which maps graph-structured data to the vocabulary of LLM through multi-view structural enhancement and code quantization, achieving token-level alignment between graphs and LLM, thus directly outputting natural language predictions in graph node classification tasks without textual information.
Peng Su (Sichuan University), Jiancheng Lv (Sichuan University)
CodeAuto EncoderContrastive LearningImage
π― What it does: This paper proposes a multi-view particle ball contrastive clustering method called MGBCC, which captures the local structure of samples using particle ball partitioning and performs contrastive learning on cross-view particle balls in the latent space to achieve unsupervised multi-view clustering.
π― What it does: This study proposes a new multi-view pedestrian occupancy prediction task, constructs a large-scale synthetic dataset MVP-Occ, and designs a baseline model OmniOcc;
π― What it does: In the federated learning framework, the MRFF system is proposed, utilizing a lightweight Transformer as the base model and incorporating group gating networks at each layer to group users, achieving user-specific and group-level personalized joint learning, ultimately used for CTR prediction tasks.
Multilingual LLMs Inherently Reward In-Language Time-Sensitive Semantic Alignment for Low-Resource Languages
Ashutosh Bajpai (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
CodeRetrievalTransformerLarge Language ModelContrastive LearningTextBenchmark
π― What it does: Improving temporal reasoning under low-resource languages, proposing a multilingual temporal reasoning dataset mTEMPREASON, and developing a cross-lingual time-sensitive semantic alignment retriever CLiTSSA;
π― What it does: This study addresses the audio-video video parsing (AVVP) task and proposes a multi-modal class-aware semantic enhancement network (MM-CSE). By decoupling mixed features into event-specific and background class features, and achieving event co-occurrence modeling and global semantic fusion at a fine-grained level, it significantly improves event localization performance.
Multiple Feature Refining Network for Visual Emotion Distribution Learning
Qinfu Xu (China University of Petroleum), Chunlei Wu (China University of Petroleum)
CodeConvolutional Neural NetworkTransformerImage
π― What it does: This paper proposes a Multi-Feature Refinement Network (MFRN) that achieves visual emotion distribution learning through spectral mixing and semantic map prompt learning.
π― What it does: This study focuses on multi-behavior recommendation and proposes the MPC model, which enhances recommendation effectiveness by utilizing multiple purchase chains and a negative transfer elimination mechanism.
π― What it does: This paper proposes an image forgery localization network called MUN, based on the M3 encoder and UN decoder, for pixel-level localization of various types of forgeries (splicing, copy-move, deletion, AI-generated) images.
π― What it does: A feature pyramid is constructed based on CLIP to obtain multi-scale video features, and the Mamba linear state space model is used to efficiently model across resolutions, thereby improving text-video retrieval performance.
Muses: 3D-Controllable Image Generation via Multi-Modal Agent Collaboration
Yanbo Ding (Shenzhen Institute of Advanced Technology), Yali Wang (Shenzhen Institute of Advanced Technology)
CodeGenerationData SynthesisLarge Language ModelSupervised Fine-TuningAgentic AIImageTextMultimodality
π― What it does: MUSES is proposed, a multimodal agent collaboration system that implements a complete process from user text to controllable 3D images.
MVREC: A General Few-shot Defect Classification Model Using Multi-View Region-Context
Shuai Lyu (Hong Kong Polytechnic University), Waikeung Wong (Hong Kong Polytechnic University)
CodeClassificationAnomaly DetectionVision Language ModelContrastive LearningImageBenchmark
π― What it does: A general few-shot defect classification framework based on multi-view region-context, MVREC, is proposed to address the issues of context information dependency and insufficient data generalization in defect classification.
CodeGenerationSafty and PrivacyDiffusion modelImage
π― What it does: A method called MYOPIA was developed, which is a no-learning sample approach that prevents personalized text-to-image models from learning real facial features by adding imperceptible perturbations to facial images, thereby protecting facial privacy.
π― What it does: This paper proposes NaviFormer, an encoder-decoder Transformer designed to aggregate spatial layout, temporal pose, and traversable frontier information for object navigation.
NeSyCoCo: A Neuro-Symbolic Concept Composer for Compositional Generalization
Danial Kamali (Michigan State University), Parisa Kordjamshidi (Michigan State University)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: A neural symbolic framework called NeSyCoCo is proposed, which generates symbolic programs using large language models and achieves combinatorial generalization for image-language reasoning tasks through soft combination and distributed word embeddings.
π― What it does: A new neural combinatorial clustering gambling algorithm, NeUClust, is proposed for the contextual combinatorial gambling problem in recommendation systems, aiming to simultaneously learn the unknown reward function and select the arm with the highest reward.
π― What it does: This paper designs Neural Conjugate Flows (NCF), which constructs a topologically conjugate flow structure through reversible Coupling Layers and affine flows, naturally satisfying physical constraints such as initial conditions and causality, and capable of approximating any ODE flow.
Shuntuo Xu (East China Normal University), Zhou Yu (East China Normal University)
CodeTabularTime Series
π― What it does: This study investigates the sufficient dimension reduction (SDR) capability of deep feedforward neural networks in regression tasks, proving that the weights of the first layer can approximate the central mean subspace, and provides statistical consistency and experimental validation.
Neural-Symbolic Collaborative Distillation: Advancing Small Language Models for Complex Reasoning Tasks
Huanxuan Liao (Chinese Academy of Sciences), Jun Zhao (Chinese Academy of Sciences)
CodeExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes Neural-Symbolic Collaborative Distillation (NesyCD), which analyzes the errors of small models (SLM) through large language models (LLM), generates and stores a symbolic knowledge base, and dynamically retrieves this knowledge during inference to enhance the performance of SLM on complex reasoning tasks.
π― What it does: A two-stage framework called NeuralFlix has been constructed to decode fMRI data and generate high-quality videos that are semantically consistent with the videos being watched.
New Compilation Languages Based on Restricted Weak Decomposability
Petr Illner (Charles University)
CodeGraph
π― What it does: This paper proposes the positive-negative weakly separable DNNF variants pwDNNF and nwDNNF, modifies the Bella compiler to generate these circuits, and designs an acceleration method based on sub-circuit replication, demonstrating that nwDNNF performs excellently in MPE computation for two-layer large domain Bayesian networks.
NightReID: A Large-Scale Nighttime Person Re-Identification Benchmark
Yuxuan Zhao (Wuhan University), Mang Ye (Wuhan University)
CodeRecognitionRetrievalTransformerImageBenchmark
π― What it does: A large-scale nighttime RGB pedestrian re-identification dataset, NightReID, is proposed, along with an EDA framework based on unsupervised image enhancement and distribution alignment, significantly improving nighttime Re-ID performance.
π― What it does: Proposes the NLGT framework for node classification, utilizing neighborhood sampling, label-enhanced feature fusion, and neighborhood mask attention.
NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning
Xin Yi (East China Normal University), Liang He (East China Normal University)
CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the Neuron-Level Safety Realignment (NLSR) framework for untrained safety recovery in the context of safety degradation issues that arise in large language models (LLMs) during fine-tuning-as-a-service scenarios.
Zhang Xiong, Beibei Yu (University of Technology Sydney)
CodeRepresentation LearningGraph Neural NetworkContrastive LearningGraphBiomedical Data
π― What it does: A model named NoiseHGNN is proposed for node representation learning in heterogeneous graphs with noisy edges, and it mitigates the impact of erroneous edges through synthetic similarity graphs and contrastive learning.
Noisy Label Calibration for Multi-View Classification
Shilin Xu (Sichuan University), Dezhong Peng (Sichuan National Innovation New Vision UHD Video Technology Co., Ltd.)
CodeClassificationTabular
π― What it does: A noise label calibration method NLC is proposed in multi-view learning, integrating cross-view maximum margin ranking, MixUp, and noise detection and calibration modules to achieve robust classification.
NOMATTERXAI: Generating βNo Matter Whatβ Alterfactual Examples for Explaining Black-Box Text Classification Models
Tuc Van Nguyen (Indiana University), Thai Le (Indiana University)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: The NOMATTERXAI algorithm is proposed, which can automatically generate 'alterfactual' examples for text classifiers to assess the model's sensitivity and fairness regarding irrelevant features.
π― What it does: For high-reflective scenes, we propose Normal-NeRF, which achieves robust estimation of surface normals and renders high-quality images through the transmission gradient estimation method and a dual activation density module.
Normalize Then Propagate: Efficient Homophilous Regularization for Few-Shot Semi-Supervised Node Classification
Baoming Zhang (Nanjing University), Chongjun Wang (Nanjing University)
CodeClassificationGraph Neural NetworkGraph
π― What it does: A graph neural network named NormProp is proposed, which first normalizes node features to the unit sphere, then propagates through low-pass filtering. It encodes category information using the direction of node vectors and aggregates consistency using the Euclidean norm, while mining supervisory signals from unlabeled nodes through homogeneity regularization to enhance generalization performance in scenarios with few labels.
Novelty-Guided Data Reuse for Efficient and Diversified Multi-Agent Reinforcement Learning
Yangkun Chen (Shenzhen International Graduate School Tsinghua University), Jiafei Lyu (Shenzhen International Graduate School Tsinghua University)
CodeReinforcement LearningTabular
π― What it does: To address the issues of low sample utilization and insufficient agent diversity in multi-agent reinforcement learning, a sample reuse method based on observation novelty called MANGER is proposed. It dynamically adjusts the policy update frequency for each agent and introduces shared and independent layers in the critic network to enhance agent diversity.
π― What it does: This study investigates the self-consistency of fine-grained mediating variables in causal reasoning by large language models (causal cognitive consistency) and proposes evaluation metrics.
CodeOptimizationComputational EfficiencyPhysics Related
π― What it does: A new training method called Good Lattice Training (GLT) is proposed to accelerate the learning process of Physics-Informed Neural Networks (PINNs), particularly in solving partial differential equations (PDEs).
NumbOD: A Spatial-Frequency Fusion Attack Against Object Detectors
Ziqi Zhou (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
CodeObject DetectionAdversarial AttackImage
π― What it does: A model-agnostic spatial-frequency fusion adversarial attack, NumbOD, is designed to achieve imperceptible and global failure against object detectors.
π― What it does: A deepfake detection framework suitable for open-world social networks, ODDN, is proposed to address the issue of missing paired data between compressed images and original images.
Offline Multi-Agent Reinforcement Learning via In-Sample Sequential Policy Optimization
Zongkai Liu (Sun Yat-sen University), Xuetao Ding (Meituan)
CodeReinforcement LearningTabularSequential
π― What it does: This paper proposes an offline multi-agent reinforcement learning algorithm named InSPO, which utilizes sequential in-sample policy optimization to learn a joint policy for multiple agents, avoiding the out-of-distribution (OOD) joint action and local optimal convergence issues in offline scenarios.
π― What it does: Proposes the OmniQuery framework, which combines active learning with source-free domain adaptation to achieve cross-modal 3D semantic segmentation.
π― What it does: A lightweight method called OmniMark is proposed for the Latent Diffusion Model (LDM), which allows each generated image to carry an invisible fingerprint without affecting the generation quality, enabling model attribution and accountability tracking.
Huiwen Dong (Beijing Normal University), Ninh Pham (University of Auckland)
CodeOptimizationComputational EfficiencyTime Series
π― What it does: By employing random sampling kNN computation on high-dimensional datasets, SamHub efficiently identifies hub points with a time complexity of O(sn).
π― What it does: This paper studies and proposes a differential equation-based graph neural network model called SWAN, which introduces antisymmetric parameterization in both spatial and weight domains to achieve global and local non-dissipative characteristics, thereby addressing the oversquashing problem in traditional message-passing networks.
CodeSafty and PrivacyComputational EfficiencyConvolutional Neural NetworkImage
π― What it does: This paper systematically analyzes the errors of probabilistic truncation in Privacy-Preserving Machine Learning (PPML) and proposes methods such as deterministic truncation protocols, ReLU channel scaling techniques, and zero-keeping random mappings, significantly reducing communication overhead and improving inference speed.
Randy Lefebvre (Laval University), Audrey Durand (Laval University)
CodeReinforcement Learning
π― What it does: This paper discusses the impact of shallow planning (i.e., shorter discount factors) on the bias-variance trade-off in reinforcement learning within partially observable environments, and provides a new planning loss bound based on MDP structural parameters.
On the Expressiveness and Length Generalization of Selective State Space Models on Regular Languages
Aleksandar Terzic (IBM Research), Abbas Rahimi (IBM Research)
CodeRecurrent Neural NetworkTransformerSequential
π― What it does: This paper proposes a new Selective Dense State Space Model (SD-SSM) for perfectly learning and generalizing regular language (finite state machine) tasks within a single-layer model.
On the Relationship Between Monotone and Squared Probabilistic Circuits
Benjie Wang (University of California, Los Angeles), Guy Van den Broeck (University of California, Los Angeles)
CodeImageTabular
π― What it does: Proposes Inception PCs (Deep Sum-Square-Sum Structure), unifying and extending the expressive power of monotonic PCs and square PCs;
One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation
Huishi Luo (Beihang University), Deqing Wang (Beihang University)
CodeRecommendation SystemMixture of ExpertsTabular
π― What it does: This study proposes a multi-domain recommendation framework called AREAD, which can handle dozens of domains and enhances recommendation performance through hierarchical expert fusion, expert mask pruning, and popularity-based counterfactual augmentation.
One Node One Model: Featuring the Missing-Half for Graph Clustering
Xuanting Xie (University of Electronic Science and Technology of China), Wenyu Chen (University of Electronic Science and Technology of China)
CodeGraph Neural NetworkContrastive LearningGraph
π― What it does: A graph clustering framework based on 'one model per node' (FPGC) is proposed, which selects clustering-related features through the squeeze-excitation mechanism and enhances low-order feature interactions using feature crossing.
One Token Can Help! Learning Scalable and Pluggable Virtual Tokens for Retrieval-Augmented Large Language Models
Yutao Zhu (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: A pluggable and extensible virtual token method called SPRING is designed to enhance the performance of LLM in retrieval-augmented generation (RAG) scenarios while maintaining the general generation capability of the original model.
π― What it does: A OneBatchPAM algorithm is proposed to quickly estimate the k-medoids objective using a single batch (size O(log n)), significantly reducing computational and storage costs.
Online Guidance Graph Optimization for Lifelong Multi-Agent Path Finding
Hongzhi Zang (Tsinghua University), Jiaoyang Li (Monash University)
CodeOptimizationConvolutional Neural NetworkGraph
π― What it does: An online guidance strategy is proposed, which dynamically updates the guidance map using real-time traffic information to enhance the throughput of lifelong multi-agent path planning.
π― What it does: A virtual try-on method OOTDiffusion based on latent diffusion models is proposed, which can generate realistic try-on images while maintaining details.
Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models
Lucio La Cava (University of Calabria), Andrea Tagarelli (University of Calabria)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: This study investigates the inherent personalities of 12 open-source LLMs in MBTI and BFI personality assessments, as well as their imitation abilities under conditional prompts and role prompts.
π― What it does: This paper proposes a framework called UAGA for open cross-network node classification, aimed at completing node classification and unknown detection when the target network contains unknown categories not present in the source network.
Open-world Radio Frequency Fingerprint Identification via Augmented Semi-supervised Learning
Zehua Han (Beihang University), Wenrui Ding (Beihang University)
CodeClassificationRecognitionTransformerContrastive LearningTime Series
π― What it does: Under open world conditions, a self-supervised pre-training based radio frequency fingerprint identification framework called OpenRFI is proposed, along with a dedicated Roinformer feature extraction model.
Shide Du (Fuzhou University), Wenzhong Guo (Fuzhou University)
CodeExplainability and InterpretabilityRepresentation LearningImage
π― What it does: A multi-view learning framework for open environments, OpenViewer, is proposed to address interpretability and generalization issues.
CodeObject DetectionObject TrackingSegmentationRecurrent Neural NetworkTransformerVision Language ModelContrastive LearningVideo
π― What it does: The InstFormer framework is proposed to achieve Open Vocabulary Video Instance Segmentation (OpenVIS), capable of simultaneously detecting, segmenting, and tracking instances of any category in videos.
π― What it does: This paper proposes a dynamic programming and branch-and-bound algorithm called ConTree, which constructs optimal classification trees on continuous feature data within a given size constraint.
π― What it does: An optimal control method based on the perspective of Instance-Solution Control Operator is proposed, and a new neural operator architecture, Neural Adaptive Spectral Method (NASM), is implemented for direct one-time solving of OCP.
Optimized Gradient Clipping for Noisy Label Learning
Xichen Ye (Shanghai University), Cheng Jin (Fudan University)
CodeOptimizationImage
π― What it does: A dynamic threshold-based gradient clipping method called OGC is proposed to suppress the negative impact of noisy labels on gradient distribution during the training process.
π― What it does: An implicit neural representation based on octrees, called OTIAS, is proposed for the fusion of multispectral and hyperspectral images.
π― What it does: A low-rank t-SVD generative model based on learnable orthogonal transformations is proposed for the recovery of multi-dimensional inverse problems (such as tensor completion, CASSI reconstruction, and MSI denoising).
π― What it does: A dynamic facial expression recognition method OUS is proposed, which integrates scene context with facial features to address cognitive biases caused by traditional methods neglecting the scene.
π― What it does: This work proposes a long text recognition method based on substring matching called SMTR, and constructs the first long text benchmark LTB to address the issue of recognizing long texts when trained solely on short text datasets.
π― What it does: Proposes the GPro model, which learns causal invariant substructures in graph structures through multi-step evolutionary reasoning to achieve OOD generalization;
OV-DQUO: Open-Vocabulary DETR with Denoising Text Query Training and Open-World Unknown Objects Supervision
Junjie Wang (Harbin Institute of Technology), Yong Xu (Chongqing Research Institute of HIT)
CodeObject DetectionTransformerVision Language ModelContrastive LearningImage
π― What it does: This paper proposes the OV-DQUO framework, which effectively alleviates the confidence bias in open vocabulary detection through open-world pseudo-labeling, wildcard matching, and denoised text query training.
Overcoming Heterogeneous Data in Federated Medical Vision-Language Pre-training: A Triple-Embedding Model Selector Approach
Aowen Wang (Zhejiang University), Shiting Wen (NingboTech University)
CodeFederated LearningSafty and PrivacyTransformerVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
π― What it does: The PMS-FM framework is proposed to address data heterogeneity and privacy issues in medical audiovisual language pre-training, achieving cross-institutional federated pre-training through a personalized model selector and triple embedding.
Anand Krishna (National University of Singapore), Vincent Y. F. Tan (National University of Singapore)
CodeOptimizationReinforcement Learning
π― What it does: This paper proposes a new performance metric for Markov Bandits (MAB) called p-mean regret, which can balance fairness and efficiency.
PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation
Sarthak Kumar Maharana (University of Texas at Dallas), Yunhui Guo (University of Texas at Dallas)
CodeDomain AdaptationOptimizationImage
π― What it does: A continuous testing time adaptive learning rate (CTTA) method named PALM is proposed, which can automatically select the layers that need fine-tuning and dynamically adjust their learning rates without relying on pseudo-labels, thereby enhancing the model's robustness in domain shift environments.
π― What it does: A two-stage parameter-efficient fine-tuning framework called PanAdapter is proposed, which utilizes the spatial-spectral prior information from a pre-trained image restoration model to achieve high-quality image fusion.
π― What it does: This paper proposes Pareto Continual Learning (ParetoCL), which treats the stability-plasticity trade-off in continual learning as a multi-objective optimization problem, using a preference-conditioned model to learn within a single network and dynamically adapt to different trade-offs during inference.
ParGo: Bridging Vision-Language with Partial and Global Views
An-Lan Wang (ByteDance China), Wei-Shi Zheng (Sun Yat-sen University)
CodeGenerationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: The ParGo project is proposed, which combines visual features through partial views and global views to project them into the language model space, better bridging the visual encoder and large language model.
π― What it does: This paper proposes a single-step class-level unlearning method (Partially-Blinded Unlearning, PBU) that utilizes only the data of the class to be forgotten to eliminate the information of that class from a pre-trained network while maintaining the performance of other classes.
PAT: Pruning-Aware Tuning for Large Language Models
Yijiang Liu (Nanjing University), Li Du (Nanjing University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A Pruning-Aware Tuning (PAT) framework is proposed, which combines structured pruning and fine-tuning by inserting Hybrid Sparsification Modules (HSM) between the attention and feedforward networks, and implementing a unified sparse mask to achieve dynamic pruning of the hidden dimensions of large language models.
π― What it does: This paper proposes a LiDAR semantic segmentation framework called PC-BEV, which fuses polar and Cartesian coordinate segmentation only within the BEV space, and achieves efficient dense feature fusion through a pre-computed remap method.
PDBs Go Numeric: Pattern-Database Heuristics for Simple Numeric Planning
Daniel Gnad (Linkoping University), Alexander Shleyfman (Bar-Ilan University)
CodeOptimizationReinforcement LearningBenchmark
π― What it does: A heuristic for simple numerical planning (SNP) with integer variables is proposed, along with various methods to handle the infinite state space of numerical variables.
π― What it does: A framework for RGB semantic segmentation based on pseudo depth is proposed, utilizing a multi-source pseudo depth aggregation module (PDAM) and a diffusion model (PDDM) to achieve high-precision segmentation of indoor scenes.
π― What it does: This paper proposes an input-agnostic Prompt mechanism called PEARL for class-incremental learning, addressing the knowledge confusion problem caused by traditional Query-Select mechanisms.
Pedestrian Attribute Recognition: A New Benchmark Dataset and a Large Language Model Augmented Framework
Jiandong Jin (Anhui University), Chenglong Li (Anhui University)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningImageBenchmark
π― What it does: A new large-scale cross-domain pedestrian attribute recognition dataset MSP60K is proposed, and based on this dataset, the LLM-PAR framework is developed, which combines visual Transformer, MEQ-Former, and large language models to achieve joint learning of attribute classification and text description; a systematic benchmark of 17 existing PAR methods is conducted under random and cross-domain splits; the performance improvement of LLM-PAR is validated on MSP60K and other public datasets.
PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing
Huaizhuo Liu (Beihang University), Yurui Liu (Beihang University)
CodeRestorationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImagePhysics Related
π― What it does: This paper proposes a physics-based embedded illumination estimation framework (PEIE) for adaptive dehazing in real-world scenarios.
π― What it does: This work proposes PerReactor, an offline personalized multi-appropriate facial response generation (PMAFRG) framework that can generate diverse and personalized facial responses for a given speaker's behavior.
π― What it does: This paper proposes a personalized clustering framework called PCL, which utilizes active querying of the most uncertain and difficult negative samples to learn user preference-oriented feature representations, achieving clustering results that align with user needs.
Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach
Zhiwei Li (University of Technology Sydney), Chengqi Zhang (Hong Kong Polytechnic University)
CodeRecommendation SystemFederated LearningSafty and PrivacyAuto EncoderTabular
π― What it does: A federated collaborative filtering framework named FedDAE is proposed, which utilizes Variational AutoEncoder (VAE) combined with dual encoders and a gating network to achieve personalized recommendations without uploading raw data.
Yangxuan Zhou (Zhejiang University), Gang Pan (Zhejiang University)
CodeDomain AdaptationContrastive LearningTime SeriesBiomedical Data
π― What it does: A source-free unsupervised individual domain adaptation (SF-UIDA) framework is proposed for personalized rapid adaptation of sleep staging models for each new subject without accessing source data.
π― What it does: The pFedES framework is proposed, which adds a shared proxy homogeneous feature extractor in front of each heterogeneous client model in federated learning, and achieves bidirectional transfer of global knowledge and local personalized knowledge through alternating training and aggregation.
π― What it does: This paper proposes the PhyCamo framework for multi-view physical adversarial attacks, utilizing diffusion models for data enhancement, contrastive learning to improve robustness, and achieving efficient attacks on encoders.
π― What it does: A frequency domain data augmentation method called PhysAug is proposed, based on an atmospheric optical physical model, for single-source generalized object detection.
PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement
Wei Qian (Hefei University of Technology), Meng Wang (Zhejiang University)
CodeTransformerDiffusion modelBiomedical Data
π― What it does: A diffusion model combining physiological dynamics separation (PhysDiff) is proposed for remote photoplethysmography (rPPG) signal estimation.
π― What it does: This paper proposes two techniques: Pinwheel-shaped Convolution (PConv) and Scale-Based Dynamic Loss (SD Loss) to enhance feature extraction and regression performance for infrared small target detection, and constructs a large-scale single-frame infrared small target dataset SIRST-UAVB.