AAAI Conference on Artificial Intelligence Β· 1014 papers
Mitigating the Impact of False Negative in Dense Retrieval with Contrastive Confidence Regularization
Shiqi Wang (Nanjing University), Cam-Tu Nguyen (Nanjing University)
CodeRetrievalContrastive LearningText
π― What it does: This paper addresses the issue of false negative samples caused by incomplete annotations in dense retrieval, proposing a robust NCE loss based on contrastive confidence regularization and a general passage sieve filtering algorithm to filter out noisy negative samples and improve retrieval quality.
Mixup-Induced Domain Extrapolation for Domain Generalization
Meng Cao (Nanjing University of Aeronautics and Astronautics), Songcan Chen (Nanjing University of Aeronautics and Astronautics)
CodeDomain AdaptationContrastive LearningImage
π― What it does: This paper proposes an extrapolation strategy based on Mixup, called EDM, to expand the source domain space by generating extrapolated domains in domain generalization tasks, thereby enhancing the model's generalization ability.
π― What it does: A multi-modal knowledge graph fusion end-to-end neural network (MKG-FENN) is proposed for precise prediction of drug-drug interaction events.
π― What it does: This paper proposes a self-supervised point cloud representation learning framework called MM-Point based on multimodal contrastive learning, which jointly learns from multi-view 2D images and 3D point clouds of the same 3D object to better capture spatial structural information.
π― What it does: In the task of knowledge graph completion, a composite reasoning framework is introduced to unify the interpretation of various models, revealing that tensor decomposition-based models erroneously incorporate irrelevant entities into reasoning, leading to the failure of the collaborative filtering assumption. The 'combinatorial risk' is proposed and optimized to reduce erroneous reasoning, thereby enhancing model performance.
ModWaveMLP: MLP-Based Mode Decomposition and Wavelet Denoising Model to Defeat Complex Structures in Traffic Forecasting
Ke Sun (Central South University), Zhifang Liao
CodeTime Series
π― What it does: Designed and implemented the ModWaveMLP model based on MLP, using pattern decomposition and wavelet denoising for traffic prediction.
Molecular Optimization Model with Patentability Constraint
Sally Turutov (Technion Israel Institute of Technology), Kira Radinsky (Technion Israel Institute of Technology)
CodeOptimizationDrug DiscoveryAuto EncoderGraph
π― What it does: A molecular generation framework based on SMILES, called MOMP, is proposed, which jointly optimizes molecular properties and similarity to existing patented molecules (PL), achieving molecular design that improves drug properties while reducing the risk of patent infringement.
MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts
Haoqiang Guo (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextGraph
π― What it does: Combining language models with molecular pre-training models through a dual-tower architecture, utilizing text prompts to adaptively generate task-specific molecular representations.
π― What it does: A monocular hand mesh recovery method based on dual noise estimation is proposed, which gradually refines the mesh and corrects camera parameters based on benchmark parametric fitting.
Monte Carlo Tree Search in the Presence of Transition Uncertainty
Farnaz Kohankhaki (University of Alberta), Martin MΓΌller (University of Alberta)
CodeReinforcement Learning
π― What it does: The study investigates Monte Carlo Tree Search (MCTS) under model incompleteness and proposes UA-MCTS to leverage estimated uncertainty to guide the search.
π― What it does: This paper proposes MorphVAE, a soft robot morphology generation model based on variational autoencoders, which achieves the co-evolution of morphology and control through multi-task training and continuous natural selection sampling.
π― What it does: This paper proposes MotifRGC, a framework for self-supervised graph representation learning on multi-curvature Riemannian manifolds, capable of capturing motif patterns in graphs.
π― What it does: This paper proposes STCNet, which utilizes the spatial-temporal collaborative fusion of event cameras and frame images to achieve motion deblurring.
MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting
Wanlin Cai (Sichuan University), Yuankai Wu (Beijing Institute of Technology)
CodeGraph Neural NetworkTransformerTime Series
π― What it does: A multivariate time series forecasting model MSGNet based on multi-scale frequency domain analysis and adaptive graph convolution is proposed, which can capture the interrelationships of multiple series at different time scales.
Multi-Class Support Vector Machine with Maximizing Minimum Margin
Feiping Nie (Northwestern Polytechnical University), Rong Wang (Northwestern Polytechnical University)
CodeClassificationOptimizationTabular
π― What it does: A multi-class support vector machine (M3SVM) is proposed to enhance classification performance by maximizing the minimum margin between all classes.
Multi-Constellation-Inspired Single-Shot Global LiDAR Localization
Tongzhou Zhang (Jilin University), Jue Hu (Harbin Institute of Technology)
CodePose EstimationAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
π― What it does: A multi-constellation heuristic single global LiDAR positioning method is proposed: first, global descriptors are used to roughly retrieve keyframes, then several nearby observation points are selected, and a lightweight LiDAR odometer is used to estimate the distance to the observation points. Finally, the position estimate is transformed into a multi-sphere equation solution to obtain the three-dimensional pose.
Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement
Kai Shang (China University of Petroleum), Shuigen Wang (University of Technology Sydney)
CodeRestorationDiffusion modelImage
π― What it does: This paper proposes a Multi-Domain Multi-Scale Diffusion Model (MDMS) that simultaneously utilizes spatial and frequency domain information for low-light image enhancement tasks, significantly reducing checkerboard artifacts through multi-scale sampling techniques.
π― What it does: For the task of open vocabulary video visual relationship detection, a multimodal prompting method is proposed, combining spatio-temporal visual prompting with vision-guided language prompting, utilizing CLIP to achieve cross-modal knowledge transfer.
π― What it does: This paper proposes a weakly supervised 3D semantic segmentation method enhanced by multimodal similarity (MMA), which alleviates the impact of long-tail distribution by combining geometric information with RGB information and normalizing classifier weights.
Multi-Region Text-Driven Manipulation of Diffusion Imagery
Yiming Li (Shanghai Jiao Tong University), Yi Xu (China Mobile)
CodeGenerationData SynthesisVision Language ModelDiffusion modelContrastive LearningImageText
π― What it does: This paper proposes a multi-region text-driven diffusion image editing framework (MRGD) that enables the addition, deletion, and attribute modification of multiple targets in an image through given region-level text prompts.
Multi-View Dynamic Reflection Prior for Video Glass Surface Detection
Fang Liu (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
CodeObject DetectionSegmentationTransformerVideo
π― What it does: The first video glass surface detection method, VGSD-Net, is proposed, which utilizes multi-view dynamic reflection information to locate the glass surface.
π― What it does: This paper proposes MA Gumbel MuZero and MA Gumbel AlphaZero, extending MuZero to multi-agent MMDP tasks with combinatorial action spaces, and improving policy update and sampling efficiency through non-replacement sampling and Gumbel-Top-k search.
π― What it does: Proposes the AV-wav2vec2 end-to-end multi-channel multi-modal self-supervised pre-training framework, utilizing video and multi-channel audio for contrastive learning;
Multilevel Attention Network with Semi-supervised Domain Adaptation for Drug-Target Prediction
Zhousan Xie (Shanghai Jiao Tong University), Lei Xu (Guangdong Institute of Intelligence Science and Technology)
CodeDomain AdaptationDrug DiscoveryTransformerBiomedical Data
π― What it does: This paper proposes a semi-supervised domain adaptation multi-layer attention network MlanDTI for predicting drug-target interactions (DTI).
Multiscale Attention Wavelet Neural Operator for Capturing Steep Trajectories in Biochemical Systems
Jiayang Su (Guangxi Normal University), Minghan Chen (Wake Forest University)
CodeBiomedical DataOrdinary Differential Equation
π― What it does: A multi-scale attention wavelet neural operator (MAWNO) is proposed for accurately simulating the nonlinear differential equations of biochemical systems with steep trajectories, using a physics-driven training approach.
π― What it does: This paper proposes a Multi-Scale Low-Frequency Memory Network (MLFM), which maintains the original structure of CNN while inserting Low-Frequency Memory Units (LFMU) at different scales to preserve and interact with low-frequency information, thereby enhancing the performance of CNN in tasks such as image classification and semantic segmentation.
CodeGenerationData SynthesisTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: The MULTISCRIPT benchmark is proposed, defining two tasks: multimodal script generation and subsequent step prediction, and providing 6,655 task videos and text data based on WikiHow.
MusER: Musical Element-Based Regularization for Generating Symbolic Music with Emotion
Shulei Ji (Xi'an Jiaotong University), Xinyu Yang (Xi'an Jiaotong University)
CodeGenerationTransformerAuto EncoderAudio
π― What it does: This paper presents MusER, a music element decoupling and emotion generation framework based on VQ-VAE, which can separate elements such as pitch, rhythm, and dynamics in the latent space and control the emotion of generated music through element transfer.
π― What it does: A multi-source watermark tracking method MuST is proposed to identify and extract copyright information from multi-source synthesized images.
π― What it does: Using only 2D bounding box annotations, we jointly train an instance segmentation model for images and LiDAR, and enhance segmentation quality through a multi-modal pseudo-label generation and correction module.
Narrowing the Gap between Supervised and Unsupervised Sentence Representation Learning with Large Language Model
Mingxin Li (Beihang University), Yongyi Mao (University of Ottawa)
CodeRepresentation LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: This paper conducts detailed experiments on the training process of supervised and unsupervised contrastive sentence embedding (CSE) and finds that the complexity of similarity patterns determines the performance gap; subsequently, it utilizes the contextual learning ability of large language models to generate data with complex similarity patterns and introduces hierarchical triplet loss to fully leverage the hierarchical scoring of STS data, ultimately significantly narrowing the performance gap between supervised and unsupervised CSE.
NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models
Gengze Zhou (University of Adelaide), Qi Wu (University of Adelaide)
CodeTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityChain-of-Thought
π― What it does: Proposes NavGPT, a zero-shot visual-language navigation agent based on large language models, which completes instruction following by transforming visual perception into natural language and performing explicit reasoning with LLM.
π― What it does: A benchmark for Open Set Skeleton Action Recognition (OS-SAR) has been established to evaluate existing open set methods and propose the CrossMax method to improve the open set recognition performance of skeleton actions.
Near-Optimal Resilient Aggregation Rules for Distributed Learning Using 1-Center and 1-Mean Clustering with Outliers
Yuhao Yi (Sichuan University), Jiancheng Lv (Sichuan University)
CodeAnomaly DetectionFederated LearningImage
π― What it does: A robust aggregation rule based on a 1-center/1-mean clustering (including outliers) approximation algorithm is proposed, and a two-phase voting framework (2PRASHB) is designed to enhance the robustness of distributed learning under Byzantine attacks.
π― What it does: This paper proposes the NeBLa model, which can reconstruct the 3D structure of the oral cavity using only a single panoramic X-ray image.
π― What it does: A Negative Pre-aware Cross-modal Matching (NPC) method is proposed, which improves the noise robustness of cross-modal matching by pre-evaluating the negative impact of each sample and assigning adaptive confidence weights to the samples, and fine-tuning on CLIP ViTβB/32.
Neighborhood-Enhanced 3D Human Pose Estimation with Monocular LiDAR in Long-Range Outdoor Scenes
Jingyi Zhang (Xiamen University), Cheng Wang (Xiamen University)
CodePose EstimationTransformerPoint Cloud
π― What it does: In large-scale outdoor scenes, 3D human pose estimation has been improved by utilizing point clouds collected from monocular LiDAR, combined with spatial and structural consistency information from background neighborhood (3BN) and scanning neighborhood (3SN);
π― What it does: This paper proposes NeRF-LiDAR, which utilizes NeRF for implicit reconstruction of real driving scenes to generate realistic LiDAR point clouds and semantic labels.
NestE: Modeling Nested Relational Structures for Knowledge Graph Reasoning
Bo Xiong (University of Stuttgart), Steffen Staab (Griffith University)
CodeGraph
π― What it does: This paper proposes a new knowledge graph embedding method called NestE, which is designed to simultaneously capture the semantic relationships between atomic facts (triples) and nested facts.
Kevin Xia (Columbia University), Elias Bareinboim (Columbia University)
CodeData SynthesisExplainability and InterpretabilityRepresentation LearningAuto EncoderGenerative Adversarial NetworkImageTabular
π― What it does: This paper proposes a neural causal abstraction framework based on 'abstract functions (Ο)', utilizing interpretable variable clustering (inter-/intravariable) to construct high-level causal models, and provides theoretical tools such as hierarchical consistency (Q_Ο-consistency) and abstract conditions (AIC) for inference.
π― What it does: In Physics-Informed Machine Learning (PIML), by combining neural oscillators (CoRNN, LEM) with Physics-Informed Neural Networks (PINN), the PINN is first trained within the training domain, and its output is then used as a time series input to the oscillator, achieving accurate extrapolation to unseen time domains (or parameter domains);
NILUT: Conditional Neural Implicit 3D Lookup Tables for Image Enhancement
Marcos V. Conde (University of Wurzburg), Radu Timofte (York University)
CodeComputational EfficiencyImage
π― What it does: An implicit 3D LUT (NILUT) that can run efficiently on mobile devices and a condition version (CNILUT) with multiple styles are proposed for image enhancement and color adjustment.
NN-Steiner: A Mixed Neural-Algorithmic Approach for the Rectilinear Steiner Minimum Tree Problem
Andrew B. Kahng (University of California San Diego), Chien-Yi Yang (University of California San Diego)
CodeOptimizationPoint Cloud
π― What it does: A new hybrid neural algorithm framework called NN-Steiner is proposed for solving the Rectilinear Steiner Minimum Tree (RSMT) problem, which combines Arora's PTAS algorithm.
No Prejudice! Fair Federated Graph Neural Networks for Personalized Recommendation
Nimesh Agrawal (Indian Institute of Technology), Jayadeva (Indian Institute of Technology)
CodeRecommendation SystemFederated LearningSafty and PrivacyGraph Neural NetworkGraph
π― What it does: A fair federated graph neural network framework F2PGNN is proposed, achieving fair personalized recommendations for user groups in a federated learning environment while protecting the privacy of users' sensitive attributes.
π― What it does: A framework called No Prior Mask (NPM) is proposed, which can filter redundant actions in the action space without prior knowledge. It dynamically generates state-related action masks by learning the similarity factors between actions, reducing redundant exploration in reinforcement learning (RL).
NodeMixup: Tackling Under-Reaching for Graph Neural Networks
Weigang Lu (Xidian University), Long Jin (Xidian University)
CodeGraph Neural NetworkGraph
π― What it does: The NodeMixup method is proposed, which enhances the reachability of GNNs and alleviates the under-reaching problem by performing mixup at the node level between labeled nodes and pseudo-labeled unlabeled nodes.
π― What it does: A fine-grained word-level denoising framework for noisy image-text pairs, NIC (Noise-aware Image Captioning), is proposed. It adaptively identifies and corrects erroneous words, making the image captioning model more robust on noisy data.
π― What it does: A sample-free domain incremental object detection method is proposed, utilizing a frozen baseline model to achieve continuous learning and generalization by learning the domain bias of each new domain.
π― What it does: A new multimodal novel category discovery method based on paired chest X-ray images and text is proposed, utilizing two networks to encode images and text respectively, and evaluating the quality of pseudo-labels through internal consistency, thereby generating weighted synthetic pseudo-labels for joint training.
π― What it does: In the multi-contrast MRI reconstruction task, a deep unfolding network (Range-Null Empowered Unfolding Network, RNU) is proposed that utilizes Range-Null Decomposition and Correlation Decoupling Learning Block. It achieves data consistency by performing proximal mapping only in the null space and separates low-frequency global (isotropy) and high-frequency local (anisotropy) features in multi-modal fusion, ultimately achieving higher quality MRI reconstruction.
CodeAutonomous DrivingRecurrent Neural NetworkVision Language ModelImageTextMultimodalityPoint CloudBenchmark
π― What it does: This study investigates the visual question answering (VQA) task in autonomous driving scenarios, constructing a large-scale dataset called NuScenes-QA that includes images, LiDAR, and multi-frame information, and proposes a baseline model for evaluation.
O^2-Recon: Completing 3D Reconstruction of Occluded Objects in the Scene with a Pre-trained 2D Diffusion Model
Yubin Hu (Tsinghua University), Yong-Jin Liu (Beijing Jiaotong University)
CodeRestorationGenerationDiffusion modelImage
π― What it does: A framework named O-Recon is proposed for image inpainting of occluded regions using a pre-trained 2D diffusion model, and complete 3D object reconstruction is achieved based on the repaired images and depth information utilizing neural implicit surfaces.
π― What it does: This paper proposes an audio-visual question answering model based on fine-grained visual objects, and enhances the understanding and utilization of cross-modal information through an adaptive positive sample learning strategy.
π― What it does: A single-source domain generalization object detection method OA-DG is proposed, which includes OA-Mix data augmentation and OA-Loss contrastive learning.
OCEAN-MBRL: Offline Conservative Exploration for Model-Based Offline Reinforcement Learning
Fan Wu (Institute of Software, Chinese Academy of Sciences), Ling Li (Institute of Software, Chinese Academy of Sciences)
CodeReinforcement LearningTabular
π― What it does: Developed the OCEAN plugin, which addresses the issue of biased exploration in model-based offline RL by incorporating a conservative exploration mechanism during the rollout phase.
ODTrack: Online Dense Temporal Token Learning for Visual Tracking
Yaozong Zheng (Guangxi Normal University), Xianxian Li (Wuzhou University)
CodeObject TrackingTransformerVideo
π― What it does: A video-level object tracking framework called ODTrack is proposed, which achieves dense associations between consecutive frames in video streams using an online token propagation method.
π― What it does: This paper proposes a framework that treats offline black-box optimization as offline reinforcement learning, where the learning policy determines the step size for gradient search, guided by a surrogate model.
Yuning Cui (Shenzhen Campus of Sun Yat-sen University), Alois Knoll (Technical University of Munich)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: An efficient convolutional network named OKNet is proposed for image restoration tasks such as dehazing, snow removal, and defocus correction.
Omnidirectional Image Super-resolution via Bi-projection Fusion
Jiangang Wang (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)
CodeRestorationSuper ResolutionTransformerImage
π― What it does: This paper proposes a dual-projection panoramic image super-resolution network (BPOSR), which utilizes both equirectangular projection (ERP) and cubic mapping projection (CMP) to enhance the super-resolution quality of 360Β° images.
On Estimating the Gradient of the Expected Information Gain in Bayesian Experimental Design
Ziqiao Ao (University of Birmingham), Jinglai Li
CodeOptimizationReinforcement LearningTabular
π― What it does: This paper proposes two gradient-based methods for estimating expected information gain (EIG) (UEEG-MCMC and BEEG-AP) for gradient optimization in Bayesian experimental design.
π― What it does: A new sequence-aware loss function is proposed to reduce the estimation gap between predicted trajectories and actual trajectories in diffusion models, thereby improving image generation quality.
On Partial Optimal Transport: Revising the Infeasibility of Sinkhorn and Efficient Gradient Methods
Anh Duc Nguyen (National University of Singapore), Kim-Chuan Toh (National University of Singapore)
CodeOptimizationPoint Cloud
π― What it does: The research addresses and solves the infeasibility problem of Partial Optimal Transport (POT) in large-scale issues, proposing efficient algorithms such as Sinkhorn, APDAGD, and Dual Extrapolation.
π― What it does: A hierarchical topic model called TraCo is proposed, which is based on optimal transport plans and context-aware decoupling, to generate topic hierarchies with high affinity, coherence, and diversity.
One Step Closer to Unbiased Aleatoric Uncertainty Estimation
Wang Zhang (Massachusetts Institute of Technology), Lam M. Nguyen (IBM Research)
CodeImageTabularOrdinary Differential Equation
π― What it does: A new unbiased data uncertainty (aleatoric) estimation method called Denoising Variance Attenuation (DVA) is proposed and validated, which approaches the true noise level through active denoising and normalized gradient descent on noise.
π― What it does: A new fine-tuning method called OLOR is proposed, which combines weight rollback and hierarchical penalties to alleviate knowledge forgetting and improve performance on downstream visual tasks.
π― What it does: A one-forward and one-backward tracking quantization method (BLAQ) is proposed, which addresses the gradient 'sawtooth' problem and improves convergence speed by first performing a forward probing gradient at each step, followed by a backward update of the quantized gradient, thus solving issues present in traditional loss-aware quantization methods.
Online Conversion Rate Prediction via Multi-Interval Screening and Synthesizing under Delayed Feedback
Qiming Liu (Chinese Academy of Sciences), Qing He (Chinese Academy of Sciences)
CodeRecommendation SystemOptimizationTabularFinance Related
π― What it does: The MISS model is proposed, which achieves online CVR prediction through multi-interval selection and a lightweight synthesis framework, addressing the label bias problem caused by delayed feedback.
Online Submodular Maximization via Online Convex Optimization
Tareq Si Salem (Northeastern University), Stratis Ioannidis (Northeastern University)
CodeOptimizationTabular
π― What it does: This paper proposes a method to transform Online Submodular Optimization (OSM) into Online Convex Optimization (OCO), and achieves online submodular maximization problems based on WTP functions (including various applications such as influence maximization and facility location) under arbitrary base matrix constraints through randomized rounding and convex relaxation.
OntoFact: Unveiling Fantastic Fact-Skeleton of LLMs via Ontology-Driven Reinforcement Learning
Ziyu Shang (Southeast University), Ke Ji (Southeast University)
CodeOptimizationReinforcement LearningGraph
π― What it does: The OntoFact framework is proposed, which automatically generates high-error-rate test cases based on ontology reinforcement learning to evaluate the factual omissions of LLMs.
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelVideoText
π― What it does: This paper proposes the Open Vocabulary Video Relation Extraction task (OVRE) and constructs the Moments-OVRE dataset, which contains 180K videos, aimed at generating all relation triples related to video actions.
Optimal Makespan in a Minute Timespan! A Scalable Multi-Robot Goal Assignment Algorithm for Minimizing Mission Time
Aakash (Indian Institute of Technology Kanpur), Indranil Saha (Indian Institute of Technology Kanpur)
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingTabular
π― What it does: This paper proposes a scalable multi-robot task allocation algorithm, OM, which minimizes the makespan by solving only the necessary robot-target pairs through heuristic estimation and lazy path computation.
Optimal Survival Trees: A Dynamic Programming Approach
Tim Huisman (Delft University of Technology), Emir DemiroviΔ (Delft University of Technology)
CodeOptimizationTabular
π― What it does: An optimal survival tree algorithm called SurTree based on dynamic programming is proposed, which can obtain a global optimal solution under the constraints of given tree depth and number of branching nodes, and significantly improves scalability through a specialized depth-two precomputation algorithm.
Optimizing the Optimization of Planning Domains by Automatic Action Schema Splitting
Mojtaba Elahi (Aalto University), Jussi Rintanen (Aalto University)
CodeOptimization
π― What it does: An improved action splitting method in the planning domain automatically decomposes macro actions into smaller micro actions during the preprocessing phase by utilizing domain invariants, priority orders, and other information, significantly reducing the number of generated ground actions and alleviating preprocessing bottlenecks.
π― What it does: A novel Orthogonal Dictionary Guided Shape Completion Network (ODGNet) is proposed, achieving high-quality reconstruction of missing areas in point clouds through an improved Seed Generation U-Net and a learnable orthogonal shape dictionary.
π― What it does: A method for OOD detection called COCL is proposed and implemented for long-tail recognition scenarios, addressing the confusion between head class and tail class samples with OOD samples.
OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models
Changhun Lee (POSTECH), Eunhyeok Park (POSTECH)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A mixed-precision weight quantization method based on activation outlier awareness, called OWQ, is proposed, along with a fine-tuning technique called WCT that only updates high-precision 'weak columns' for efficient inference and task adaptation in large language models.
p-Laplacian Adaptation for Generative Pre-trained Vision-Language Models
Haoyuan Wu (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
CodeGraph Neural NetworkTransformerVision Language ModelImageMultimodality
π― What it does: This paper proposes a p-Laplacian-based adapter (p-adapter) that maps adapter tuning to graph information transmission on attention maps to address the issue of heterogeneous attention maps.
π― What it does: A new method called Pano-NeRF is proposed, which synthesizes high dynamic range new views from sparse low dynamic range panoramic images, utilizing geometric information for reconstruction.
Pantypes: Diverse Representatives for Self-Explainable Models
Rune Kjærsgaard (Technical University of Denmark), Line Clemmensen (Physikalisch Technische Bundesanstalt)
CodeExplainability and InterpretabilityAuto EncoderImage
π― What it does: This paper proposes Pantypes, which learns sparse diverse prototypes in the ProtoVAE framework using volume loss based on deterministic point processes, and implements dynamic pruning of prototypes.
Parallel Beam Search Algorithms for Domain-Independent Dynamic Programming
Ryo Kuroiwa (University of Toronto), J. Christopher Beck (University of Toronto)
CodeOptimizationTabular
π― What it does: Three parallel beam search algorithms (shared beam search, hash distributed beam search 1 and 2) are proposed for domain-independent dynamic programming (DIDP), and a multi-threaded CABS solver is implemented, significantly improving the solving speed.
Johannes K. Fichte (LinkΓΆping University), Matthias SchlΓΆgel (TU Wien)
CodeOptimizationTabular
π― What it does: This paper proposes a parallel experimental method based on memory hierarchy and cache partitioning to improve the reproducibility and stability of combinatorial optimization experiments.
π― What it does: Proposed and learned the Projected Bellman Operator (PBO), replacing the traditional empirical Bellman operator to eliminate the projection step and reduce dependence on sampling; based on this, implemented two algorithms: ProFQI and ProDQN.
Alaleh Ahmadianshalchi (Washington State University), Janardhan Rao Doppa (Washington State University)
CodeOptimizationTabular
π― What it does: This paper proposes a new batch multi-objective Bayesian optimization method (PDBO), which adaptively selects acquisition functions and utilizes Determinantal Point Process (DPP) to achieve diversified search of the Pareto front.
π― What it does: The PARSAC method is proposed, which uses deep networks to predict multiple sets of sample weights and inlier weights in one go, achieving multi-model geometric fitting within a parallel RANSAC framework.
patchDPCC: A Patchwise Deep Compression Framework for Dynamic Point Clouds
Zirui Pan (Shandong University), Yao Liu (Rutgers University)
CodeCompressionAuto EncoderPoint Cloud
π― What it does: For dynamic point cloud compression, a patchDPCC framework is proposed, which first divides the point cloud frames into patch groups with a fixed number of points, and then compresses them using a point-based deep network.
PathAsst: A Generative Foundation AI Assistant towards Artificial General Intelligence of Pathology
Yuxuan Sun (Zhejiang University), Lin Yang (Westlake University)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityBiomedical Data
π― What it does: PathAsst has been developed, a multimodal generative AI assistant for pathology, which integrates a dedicated visual encoder PathCLIP, the Vicuna-13B language model, as well as 8 specialized sub-models and a literature retrieval tool;
PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering
Bingheng Li (University of Electronic Science and Technology of China), Zhao Kang (University of Electronic Science and Technology of China)
CodeClassificationGraph Neural NetworkGraph
π― What it does: A two-layer filtering mechanism is proposed, combining heterogeneous graph heat kernels and low-pass filters to construct PC-Conv convolution and PCNet networks, achieving simultaneous processing of node classification for homogeneous and heterogeneous graphs.
π― What it does: A generalized framework PDE+ based on partial differential equations (PDE) is proposed, which directly enhances the smoothness of neural network functions by introducing adaptive distribution diffusion (ADD) into the transport equation, thereby improving generalization ability to unseen distributions.
CodeRobotic IntelligenceReinforcement LearningAgentic AI
π― What it does: A 'Peer Learning' framework is proposed, allowing a group of reinforcement learning agents to communicate with each other through action suggestions in independent environments, learning complex strategies from scratch.
π― What it does: Designed and implemented PerFedRLNAS, a framework that automatically searches for personalized model structures and weights for each client in a federated learning environment through reinforcement learning.
Performative Federated Learning: A Solution to Model-Dependent and Heterogeneous Distribution Shifts
Kun Jin (University of Michigan), Mingyan Liu (University of Michigan)
CodeOptimizationFederated LearningImageTabular
π― What it does: Proposes an executable federated learning framework, studies the impact of model-driven distribution migration in federated learning, and provides corresponding PS (stable) and PO (optimal) solutions.
Personalized LoRA for Human-Centered Text Understanding
You Zhang (Yunnan University), Xuejie Zhang (Yunnan University)
CodeClassificationRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: For the Human-Centric Text Understanding (HCTU) task, a Pluggable Personalized Low-Rank Adapter (PLoRA) is proposed, which achieves lightweight personalized fine-tuning of the model by combining User Knowledge Injection (PKI) with LoRA, and addresses the cold start (zero-shot/few-shot) problem using techniques such as PDropout and MIM.
π― What it does: This paper studies how to achieve personalized decision-making under a limited strategy budget in areas with high regulatory costs, proposing a representative MDP (r-MDP) framework and designing two deep reinforcement learning algorithms.
Perturbation-Invariant Adversarial Training for Neural Ranking Models: Improving the Effectiveness-Robustness Trade-Off
Yu-An Liu (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
CodeRetrievalAdversarial AttackTransformerText
π― What it does: This paper proposes a perturbation-invariant adversarial training (PIAT) for neural retrieval models, which enhances the balance between model performance and robustness on both natural and adversarial samples by simultaneously optimizing natural ranking loss and boundary ranking loss during training.