🎯 What it does: A knowledge distillation framework based on generative models is proposed to transfer the spatiotemporal feature semantics captured in 3D-CNNs to a lightweight student network, thereby enhancing video action recognition and detection performance.
Generative Multi-Modal Knowledge Retrieval with Large Language Models
Xinwei Long (Tsinghua University), Jie Zhou (Tencent Inc)
CodeGenerationRetrievalTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: An end-to-end generative multimodal knowledge retrieval framework, GeMKR, is proposed, which utilizes large language models to directly output knowledge cues that can be mapped to documents during the generation phase, and then retrieves the corresponding documents from the database.
🎯 What it does: A multi-modal tracking fusion mechanism GMMT based on generative models is proposed, which utilizes a generator under noisy conditions to fuse RGB-T or RGB-D features.
Generator Assisted Mixture of Experts for Feature Acquisition in Batch
Vedang Asgaonkar (Indian Institute of Technology Bombay), Abir De (Indian Institute of Technology Bombay)
CodeOptimizationMixture of ExpertsAuto EncoderGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: The study achieves optimal feature subset selection in batch feature acquisition through generator assistance, mixed experts, and random hyperplane partitioning.
🎯 What it does: This study investigates the impact of noise (Ghost Noise) caused by smaller batch sizes in Batch Normalization (BN) on model generalization and proposes Ghost Noise Injection (GNI) as a new regularization technique.
GINN-LP: A Growing Interpretable Neural Network for Discovering Multivariate Laurent Polynomial Equations
Nisal Ranasinghe (University of Melbourne), Saman Halgamuge (University of Melbourne)
CodeExplainability and InterpretabilityComputational EfficiencyTabularPhysics Related
🎯 What it does: An interpretable neural network GINN-LP is proposed, capable of automatically discovering multivariable Laurent polynomial (LP) equations from data, supporting arbitrary degrees and terms, and can automatically determine the required number of terms through a network growth strategy.
Yufei Jin (Florida Atlantic University), Xingquan Zhu (Old Dominion University)
CodeGraph Neural NetworkGraph
🎯 What it does: A Graph Label Distribution Learning (GLDL) framework is proposed to predict the label distribution of each node in a network structure.
🎯 What it does: Proposes the GLOP framework, which combines non-autoregressive (NAR) partitioning with autoregressive (AR) refinement to achieve hierarchical solving for large-scale routing problems.
Goal Alignment: Re-analyzing Value Alignment Problems Using Human-Aware AI
Malek Mechergui (Colorado State University), Sarath Sreedharan (Colorado State University)
CodeReinforcement Learning
🎯 What it does: A human cognitive model-based goal alignment framework (HAGL) is proposed, along with an interactive query algorithm designed to infer hidden true goals.
🎯 What it does: This paper presents GOALNET, a hybrid framework that alternates between predicting sub-goals using neural networks and classical planners to generate multi-stage robot execution plans from natural language instructions and world states.
🎯 What it does: Proposes two mechanisms: Gradient-Guided Modality Decoupling (GMD) and Dynamic Sharing (DS) to enhance the robustness of multimodal learning in the presence of missing modalities.
Gramformer: Learning Crowd Counting via Graph-Modulated Transformer
Hui Lin (Xi'an Jiaotong University), Deyu Meng (Harbin Institute of Technology)
CodeGraph Neural NetworkTransformerImage
🎯 What it does: This paper presents Gramformer, a model that enhances crowd counting accuracy by adjusting the attention mechanism and node features of the Transformer through a graph structure.
🎯 What it does: From a causal perspective, the GCIL method is proposed, which uses spectral graph enhancement to simulate non-causal interventions and incorporates dimension-level invariance and independence objectives to achieve causal invariant representation learning in graph contrastive learning.
Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization
Tianrui Jia (Beijing University of Posts and Telecommunications), Chuan Shi (China Mobile Information Technology Co. Ltd.)
CodeClassificationDomain AdaptationGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: This paper proposes a graph invariant learning framework based on subgraph co-mixup, combining environment mixup and invariant mixup to achieve OOD generalization for graph structures.
Graph Learning in 4D: A Quaternion-Valued Laplacian to Enhance Spectral GCNs
Stefano Fiorini (Italian Institute of Technology), Enza Messina (University of Milano-Bicocca)
CodeGraph Neural NetworkGraph
🎯 What it does: This paper proposes a quaternion Laplacian matrix and the corresponding spectral graph convolutional network QuaterGCN, designed to handle directed graphs with arbitrary weights, arbitrary signs, and bidirectional edges (digons).
🎯 What it does: This paper proposes a new semi-supervised node classification graph neural network GNN-SATA, which utilizes a soft association mechanism to separate attribute and topology representations, and dynamically adjusts adjacency relationships through graph pruning (GPM) and graph augmentation (GAM) modules.
Yijun Tian (University of Notre Dame), Panpan Xu (Amazon)
CodeGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraphBiomedical Data
🎯 What it does: This paper proposes a prompting method based on Graph Neural Networks—Graph Neural Prompting (GNP)—to inject structured knowledge from Knowledge Graphs (KG) into pre-trained large language models (LLM) through soft prompts, thereby enhancing their performance in tasks such as question answering.
Graph of Thoughts: Solving Elaborate Problems with Large Language Models
Maciej Besta (ETH Zurich), Torsten Hoefler (Cledar)
CodeLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The Graph of Thoughts (GoT) framework is proposed, modeling the reasoning process of LLMs as an arbitrary graph structure, allowing thoughts (LLM outputs) to combine, aggregate, and iterate based on graph relationships, thereby enhancing multi-step reasoning capabilities.
🎯 What it does: This paper proposes the Graph Reasoning Transformers (GRT) model, which enhances knowledge-augmented question answering performance through a triplet-level graph encoder, cross-modal representation alignment pre-training, and attention bias in cross-modal fusion.
Graph-Aware Contrasting for Multivariate Time-Series Classification
Yucheng Wang (Institute for Infocomm Research, A*STAR), Zhenghua Chen (Nanyang Technological University)
CodeClassificationGraph Neural NetworkTransformerContrastive LearningTime Series
🎯 What it does: A graph contrastive learning framework for multivariate time series (TS-GAC) is proposed, achieving spatial consistency and temporal consistency through graph structures.
🎯 What it does: Proposes GridFormer, which utilizes a point-grid transformer to bridge point clouds and space through attention between point and grid features, generating a continuous occupancy field;
CodeRecognitionObject DetectionTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes GroundVLP, a zero-shot visual localization method that utilizes existing visual language pre-trained models (VLP) and open vocabulary object detectors (OVD) to accomplish referential expression recognition and phrase localization tasks without the need for additional annotated data.
GSO-Net: Grid Surface Optimization via Learning Geometric Constraints
Chaoyun Wang (Xi'an Jiaotong University), Caigui Jiang (Xi'an Jiaotong University)
CodeOptimizationConvolutional Neural NetworkMesh
🎯 What it does: Proposes encoding mesh surfaces as geometric images and completing developability, unfolding, and denoising optimization through a self-supervised learning network called GSO-Net.
Guiding a Harsh-Environments Robust Detector via RAW Data Characteristic Mining
Hongyang Chen (Xi'an Jiaotong University), Kaisheng Ma (Tsinghua University)
CodeObject DetectionImage
🎯 What it does: This study investigates the feasibility of using RAW images for object detection in harsh environments and proposes a complete RAW detection pipeline (PRD), a raw noise degradation benchmark (RCB), and a nonlinear regularization method (FR).
GxVAEs: Two Joint VAEs Generate Hit Molecules from Gene Expression Profiles
Chen Li (Nagoya University), Yoshihiro Yamanishi (Nagoya University)
CodeGenerationData SynthesisDrug DiscoveryRecurrent Neural NetworkAuto EncoderBiomedical Data
🎯 What it does: This study proposes a dual VAE model (ProfileVAE + MolVAE) to generate 'hit' molecules with targeted activity and drug similarity from gene expression profiles, utilizing latent features of gene expression to guide molecular generation.
🎯 What it does: Proposes H2GFormer, which utilizes RGB images to achieve 3D semantic scene completion through a Transformer, focusing on handling horizontal differences and the importance of object edges;
🎯 What it does: Proposed the HAGO-Graph hierarchical geometric graph and HAGO-Net model for complete and hierarchical representation of molecular geometry and learning molecular representations;
🎯 What it does: Improved the motion refinement method for 3D hand-object interaction, refining rough gesture sequences affected by noise and jitter to make interactions more natural and coherent.
🎯 What it does: This paper proposes an online deep clustering method that does not require data augmentation. It prevents clustering collapse by applying Bayesian regularization to hard assignments within batches during training, achieving high-quality representation learning.
Hardness of Random Reordered Encodings of Parity for Resolution and CDCL
Leroy Chew (TU Wien), Stefan Szeider (TU Wien)
CodeGraph
🎯 What it does: This paper proves that viewing randomly permuted Parity encoding as a Tseitin formula leads to exponential length/time requirements in Resolution and CDCL solving due to the linear tree width of its graph. It further introduces a random Parity Addition formula, demonstrating the same exponential difficulty; at the same time, it provides O(n log n) DRAT short proofs for these formulas.
HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors
Xiao Wang (Anhui University), Yonghong Tian (Peking University)
CodeRecognitionTransformerVideo
🎯 What it does: A large-scale real event camera dataset HARDVS (300 classes, 100K event sequences) is proposed, and a Transformer-based ESTF model is designed for human action recognition in event streams.
🎯 What it does: A deep active learning framework based on Beta scores, called BESRA, is proposed to address the label imbalance problem in multi-label text classification.
HDMixer: Hierarchical Dependency with Extendable Patch for Multivariate Time Series Forecasting
Qihe Huang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeTime Series
🎯 What it does: A pure MLP structure called HDMixer is proposed, which implements multivariate time series forecasting using scalable chunking (LEP) and hierarchical dependency detectors (HDE).
Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables
Haisong Gong (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
CodeGraph Neural NetworkLarge Language ModelTextTabular
🎯 What it does: This paper proposes a fact-checking model based on a word-level heterogeneous graph, HeterFC, which performs authenticity reasoning on text and table evidence.
🎯 What it does: A Heterogeneous Testing Time Training (HTT) framework is proposed, which utilizes the Cross-Identity Cross-Modal Distance Loss (CIM) and Multi-Modal Testing Time Training (MTT) self-supervised strategy to significantly enhance the generalization performance of multi-modal person re-identification in unknown domains.
HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces
Jiaxin Pan (University of Stuttgart), Steffen Staab (University of Stuttgart)
CodeGraph Neural NetworkGraphTime Series
🎯 What it does: A Temporal Knowledge Graph Embedding model HGE based on heterogeneous geometric subspaces is proposed, utilizing the multiplication of complex numbers, split-complex numbers, and dual numbers, along with two attention mechanisms for integration.
Hierarchical and Incremental Structural Entropy Minimization for Unsupervised Social Event Detection
Yuwei Cao (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
CodeGraph Neural NetworkTextGraph
🎯 What it does: This paper proposes and implements HISEvent, an unsupervised social event detection framework that first minimizes the incremental one-dimensional structural entropy (1D SE) to supplement semantically similar edges in the message graph, and then utilizes hierarchical two-dimensional structural entropy (2D SE) to minimize the segmentation of the graph, thereby obtaining social event clusters.
🎯 What it does: In unsupervised graph representation learning, a method called Hierarchical Topological Isomorphism Knowledge Embedding for Graph Contrastive Learning (HTML) is proposed. This method allows the GNN encoder to learn isomorphic information at the graph level and subgraph level through knowledge distillation, adding isomorphic loss in addition to the original contrastive loss.
🎯 What it does: A multi-objective stochastic linear bandit (MOSLB) algorithm is proposed for mixed Pareto-lexicographic order, along with theoretical analysis and experimental validation of two mixed orders (MPL-PC and MPL-PL).
High-Dimensional Analysis for Generalized Nonlinear Regression: From Asymptotics to Algorithm
Jian Li (Institute of Information Engineering), Weiping Wang (Institute of Information Engineering)
CodeImage
🎯 What it does: This paper proposes a general framework for high-dimensional nonlinear regression analysis, covering various feature mappings such as random features, neural networks, projections, and subsampling. Under this framework, asymptotic risk expressions for ridge regression and ridgeless regression in under/over-parameterized regimes are provided, revealing the relationship between risk and effective dimension. Based on this theory, a trainable random feature regression model (RFRed) is designed, which minimizes effective dimension by simultaneously optimizing feature mapping, regularization parameters, and subsampling matrices, thereby enhancing generalization ability. Experiments on datasets such as MNIST validate the impact of double descent curves, subsampling, and regularization on generalization, and comparisons are made with traditional ridge regression and random projection methods.
🎯 What it does: High-fidelity 3D head avatar reconstruction is achieved by introducing Spatially-Varying Expression (SVE) into NeRF, which provides detailed expression and geometric details for NeRF; a coarse-to-fine training strategy (geometric initialization and adaptive importance sampling) is proposed to enhance geometric and rendering quality.
🎯 What it does: This paper proposes a two-stage gradient inversion attack, which first uses evolutionary algorithms and feature mutation coefficients to accurately recover the labels of batch training samples, and then achieves high-fidelity batch image recovery through stepwise gradient matching, gradient dropping, and image prior scheduling.
Higher-Order Graph Convolutional Network with Flower-Petals Laplacians on Simplicial Complexes
Yiming Huang (University of Electronic Science and Technology of China), Linyuan Lü
CodeClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a high-order graph convolutional network (HiGCN) based on the flower petal-flower petal (FP) model, which models and quantifies high-order interactions through the FP Laplacian and learnable polynomial filters.
History Matters: Temporal Knowledge Editing in Large Language Model
Xunjian Yin (Peking University), Xiaojun Wan (Peking University)
CodeTransformerLarge Language ModelPrompt EngineeringTextTime SeriesBenchmark
🎯 What it does: This paper proposes and evaluates the Temporal Knowledge Editing (TKE) task, constructs the ATOKE benchmark dataset, and explores and improves large model knowledge editing methods to retain historical knowledge.
Hot or Cold? Adaptive Temperature Sampling for Code Generation with Large Language Models
Yuqi Zhu (Academy of Military Sciences), Hong Mei (Peking University)
CodeGenerationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Researches decoding strategies in code generation and proposes an adaptive temperature sampling method that dynamically adjusts the temperature based on the difficulty of token predictions;
How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection
Yiyang Yao (Northwestern Polytechnical University), Qing Wang (Northwestern Polytechnical University)
CodeObject DetectionTransformerVision Language ModelImageMultimodalityBenchmark
🎯 What it does: A new open vocabulary detection (OVD) evaluation benchmark, OVDEval, is proposed, covering 9 sub-tasks, with the design of fine-grained hard negative samples and the introduction of the NMS-AP evaluation metric.
🎯 What it does: A new framework called Subspace Nearest Neighbor (SNN) is proposed for detecting out-of-distribution (OOD) samples, aimed at overcoming the curse of dimensionality in high-dimensional feature spaces.
How to Protect Copyright Data in Optimization of Large Language Models?
Timothy Chu (Google), Chiwun Yang (Sun Yat-sen University)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A training method called Copyright Regression is proposed to prevent Transformers/LLMs from outputting content similar to copyrighted data in the training set during optimization.
🎯 What it does: The HR-Pro framework is proposed, which enhances temporal action localization performance by leveraging the reliability of point-level annotations through dual propagation at the segment and instance levels.
Hybrid-SORT: Weak Cues Matter for Online Multi-Object Tracking
Mingzhan Yang (Dalian University of Technology), Dong Wang (Dalian University of Technology)
CodeObject TrackingVideoBenchmark
🎯 What it does: This paper proposes Hybrid-SORT, which improves the association process of multi-object tracking by incorporating weak cues such as confidence, height state, and velocity direction based on SORT.
HybridGait: A Benchmark for Spatial-Temporal Cloth-Changing Gait Recognition with Hybrid Explorations
Yilan Dong (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
CodeRecognitionTransformerVideoMeshBenchmark
🎯 What it does: A gait recognition benchmark CCGait is proposed, covering time and spatial clothing changes in the wild, and robust recognition against clothing and viewpoint variations is achieved based on a three-branch hybrid framework called HybridGait.
🎯 What it does: This paper proposes the Hyperbolic Graph Diffusion Model (HGDM), a two-stage graph generation method: first, it encodes nodes into hyperbolic space using a Hyperbolic Variational Autoencoder (HVAE), and then learns the joint distribution of nodes and the adjacency matrix in that space using a score-based diffusion model, thereby generating graphs with hierarchical structures.
🎯 What it does: A hyper-correlated evolution (HCE) framework based on hierarchical aggregation and correlation refinement is proposed for class-incremental learning in video classification.
HyperFast: Instant Classification for Tabular Data
David Bonet (Universitat Polit'cnica de Catalunya), Alexander G. Ioannidis (Stanford University)
CodeClassificationMeta LearningTabular
🎯 What it does: We propose HyperFast, a pre-trained hypernetwork that can generate lightweight neural network weights for new tasks based on a given support set in a single forward pass, enabling instant classification.
Hyperspectral Image Reconstruction via Combinatorial Embedding of Cross-Channel Spatio-Spectral Clues
Xingxing Yang (Hong Kong Baptist University), Zaifeng Yang (Agency for Science Technology and Research)
CodeRestorationTransformerImage
🎯 What it does: For the reconstruction of hyperspectral images from RGB images, a CESST framework is proposed, which independently extracts the spatial-spectral features of each RGB channel in a high-dimensional embedding space, and then achieves cross-channel information fusion through a combinatorial attention module.
i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance
Haoyang Chen (Southeast University), Yan Lyu (Nanjing University of Science and Technology)
CodeRecommendation SystemRecurrent Neural NetworkReinforcement LearningTabularTime Series
🎯 What it does: This paper proposes i-Rebalance, a personalized vehicle repositioning method based on deep reinforcement learning, which considers drivers' cruising preferences and makes decisions based on whether drivers accept recommendations, aiming to achieve supply-demand balance and enhance driver income.
Identification of Necessary Semantic Undertakers in the Causal View for Image-Text Matching
Huatian Zhang (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
CodeRetrievalRecurrent Neural NetworkVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a causal perspective-based image-text matching method, which first theoretically derives the probability of the necessity of semantic sharing degree for segment pairs (PN_f), and implements a Necessary Undertaker Identification Framework (NUIF). It filters the most critical visual/language segments for matching through adaptive representation and two quantification methods of PN_f (PN_f‑d and PN_f‑r).
IGAMT: Privacy-Preserving Electronic Health Record Synthesization with Heterogeneity and Irregularity
Wenjie Wang (ShanghaiTech University), Li Xiong (Emory University)
CodeData SynthesisSafty and PrivacyTransformerGenerative Adversarial NetworkBiomedical DataElectronic Health Records
🎯 What it does: A framework named IGAMT is proposed for synthesizing electronic health record data with temporal heterogeneity, missing values, and heterogeneous features while ensuring differential privacy.
Image as a Language: Revisiting Scene Text Recognition via Balanced, Unified and Synchronized Vision-Language Reasoning Network
Jiajun Wei (East China Normal University), Umapada Pal (Indian Statistical Institute)
CodeRecognitionTransformerVision Language ModelImageText
🎯 What it does: This paper proposes BUSNet, a balanced unified synchronous visual-language reasoning network that treats images as noisy text for length dimension concatenation, improving scene text recognition.
Image Safeguarding: Reasoning with Conditional Vision Language Model and Obfuscating Unsafe Content Counterfactually
Mazal Bethany (University of Texas at San Antonio), Peyman Najafirad
CodeSegmentationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: A Conditional Visual-Language Model (ConditionalVLM) has been developed to generate accurate justifications for unsafe images, and a sub-object segmentation algorithm based on adversarial explanations (CSE) has been proposed to minimize the masking of unsafe areas while keeping safe areas unchanged.
🎯 What it does: This paper proposes a safety reinforcement learning framework that does not require estimating cost constraints, using an incremental good trajectory imitation and bad trajectory avoidance method to directly improve the policy;
Imitation of Life: A Search Engine for Biologically Inspired Design
Hen Emuna (Hebrew University of Jerusalem), Dafna Shahaf (Hebrew University of Jerusalem)
CodeRetrievalTransformerLarge Language ModelText
🎯 What it does: An automated search engine named BARCODE has been developed to mine bio-inspired solutions corresponding to engineering challenges from large-scale natural language texts such as Wikipedia.
🎯 What it does: This paper proposes an unbiased adversarial distillation method (IPAD) for the scenario where the teacher model in data-free knowledge distillation (DFKD) comes from an imbalanced dataset, addressing the issues of the generator's excessive focus on minority classes and the collapse of the main class patterns.
Zain Alabedeen Ali (Moscow Institute of Physics and Technology), Konstantin Yakovlev (Federal Research Center for Computer Science and Control of the Russian Academy of Sciences)
🎯 What it does: A bulk search algorithm (Bulk Search) for anonymous multi-agent path planning based on flow networks is proposed, achieving faster path solving by compressing the search state.
🎯 What it does: This paper proposes a short text classification model named GIFT, which combines heterogeneous graph learning and contrastive learning, and introduces SVD dimensionality reduction to generate augmented views and seed K-means based weak label assignment to enhance representation learning.
🎯 What it does: A multi-layer perceptron network called HPENet based on high-dimensional position encoding (HPE) is proposed for 3D classification, segmentation, and detection tasks of point clouds.
Improving Factual Error Correction by Learning to Inject Factual Errors
Xingwei He (University of Hong Kong), Siu Ming Yiu (University of Hong Kong)
CodeGenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: A method named LIFE for unsupervised factual error correction is proposed, which automatically generates error-correct text pairs and performs corrections using a 'mask-then-corrupt-then-correct' process.
Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis
James R. Kirk (Integrated Cognition), John E. Laird (Integrated Cognition)
CodeRobotic IntelligenceTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: A method called STARS is proposed, allowing embedded agents to utilize large language models (LLM) to obtain executable task goal descriptions in a single learning instance, and to evaluate, repair, and select feasible answers through the agent's own reasoning mechanism.
🎯 What it does: Proposes a multi-source multi-lingual knowledge-driven dialogue generation task (MMKRG) and its dataset MMK-DailyDialog, and introduces the MMK-BART model to address cross-language conflicts and repetition issues.
🎯 What it does: A one-stage Panoptic Narrative Grounding method, XPNG, is proposed, which utilizes semantic and visual relationships to achieve more accurate instance and region localization.
Improving PTM Site Prediction by Coupling of Multi-Granularity Structure and Multi-Scale Sequence Representation
Zhengyi Li (Huazhong Agricultural University), Wen Zhang (Huazhong Agricultural University)
CodeProtein Structure PredictionConvolutional Neural NetworkRecurrent Neural NetworkTransformerContrastive LearningBiomedical Data
🎯 What it does: A PTM site prediction method called PTM-CMGMS is proposed, which combines multi-granularity structural information (atom, amino acid, whole protein) with multi-scale sequence representation;
🎯 What it does: A neural field representation based on decomposed low-rank tensors is proposed, which jointly optimizes camera pose and scene geometry under the condition of having only 2D image supervision.
Improving the Robustness of Knowledge-Grounded Dialogue via Contrastive Learning
Jiaan Wang (Soochow University), An Liu (Hong Kong Polytechnic University)
CodeTransformerContrastive LearningText
🎯 What it does: This paper proposes an entity-driven contrastive learning framework called EnCo, aimed at enhancing the robustness of knowledge-driven dialogue systems when faced with real noise (such as typos, incomplete or incorrect entities).
🎯 What it does: A cross-domain trajectory prediction framework based on Continuous Stochastic Differential Equations (NSDE) is proposed, which can simultaneously handle different time step configurations and trajectory noise generated by the data collection process, and achieve adaptive robustness to trajectory errors through dataset-specific diffusion networks.
🎯 What it does: An end-to-end missing multi-view clustering method called ICMVC is proposed, which can simultaneously handle missing value processing, representation learning, and clustering assignment.
🎯 What it does: This paper proposes a deep kernel Gaussian process model named ICDKGP, designed to handle irregular, sparse, and potentially abrupt longitudinal data, by modeling continuity and abrupt changes through the combination of zero-mean GP and a deterministic mean function.
Inducing Point Operator Transformer: A Flexible and Scalable Architecture for Solving PDEs
Seungjun Lee (Alsemy), TaeiL Oh
CodeTransformerMeshPhysics Related
🎯 What it does: This paper proposes the Inducing Point Operator Transformer (IPOT), a PDE operator learning framework that can handle arbitrary input-output discretization and is scalable.
Inference and Learning in Dynamic Decision Networks Using Knowledge Compilation
Gabriele Venturato (KU Leuven), Luc De Raedt (KU Leuven)
CodeReinforcement LearningTabular
🎯 What it does: A dynamic decision circuit (DDC) based on knowledge compilation and a corresponding value iteration algorithm mapl-cirup are proposed to perform Bellman updates in dynamic decision networks (DDN), completing MDP planning and offline reinforcement learning.
Information Design for Congestion Games with Unknown Demand
Svenja M. Griesbach (Technische Universität Berlin), Tim Koglin (Goethe Universität Frankfurt)
CodeOptimizationGraph
🎯 What it does: This paper studies how information design can influence players' equilibrium behavior through public signals in non-autonomous network congestion games with unknown demand, and proposes an algorithm for optimizing signal schemes to minimize total expected costs. For single-source single-target networks with linear costs, it provides a Fully Polynomial-Time Approximation Scheme (FPTAS) for two states, a graph structure determination for complete information disclosure optimality, and a linear programming (LP) solving framework based on support sets; experiments are conducted on real traffic network instances for validation.
INFORMEDQX: Informed Conflict Detection for Over-Constrained Problems
Viet-Man Le (Graz University of Technology), Mathias Uta (Siemens Energy AG)
CodeOptimizationTabular
🎯 What it does: This paper proposes the INFORMEDQX algorithm, which improves QUICKXPLAIN to utilize historical conflict knowledge for conflict detection.
Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision
Chase Walker (University of Central Florida), Rickard Ewetz (Lockheed Martin)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A new path integral attribution method is proposed - Integrated Decision Gradients (IDG), which addresses the saturation problem of Integrated Gradients by weighting the gradients according to the output logit change rate.
🎯 What it does: This study investigates predicting human intentions in untrimmed videos and proposes the ICCA loss based on instance confidence and category accuracy, as well as an intention evolution learning method.
Interactive Hyperparameter Optimization in Multi-Objective Problems via Preference Learning
Joseph Giovanelli (Alma Mater Studiorum University of Bologna), Marius Lindauer (Institute of Artificial Intelligence L3S Research Center Leibniz University Hannover)
CodeOptimizationHyperparameter SearchTabular
🎯 What it does: An interactive multi-objective machine learning hyperparameter optimization method is proposed, which allows users to compare different Pareto front binary examples to learn a personalized quality metric, and then uses this metric as a loss function to perform hyperparameter search in SMAC.
InterpretARA: Enhancing Hybrid Automatic Readability Assessment with Linguistic Feature Interpreter and Contrastive Learning
Jinshan Zeng (Jiangxi Normal University), Qing Huang (Jiangxi University of Science and Technology)
CodeClassificationExplainability and InterpretabilityRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Proposes the InterpretARA mixed readability assessment model, which combines a language feature interpreter with contrastive learning to extract and integrate deep representations at the document and paragraph levels.
Invariant Random Forest: Tree-Based Model Solution for OOD Generalization
Yufan Liao (Renmin University of China), Xing Yan (City University of Hong Kong)
CodeClassificationDomain AdaptationOptimizationExplainability and InterpretabilityTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a decision tree and random forest model based on theoretical invariance (Invariant Decision Tree and Invariant Random Forest). By adding a cross-environment invariance penalty term to the tree splitting criterion, it suppresses the use of features that change with the environment, enhancing out-of-distribution (OOD) generalization ability.
🎯 What it does: A data-independent IP removal attack called IPRemover is proposed, which utilizes model inversion to generate training data and evades DNN fingerprint and watermark detection through virtual ensemble knowledge distillation.
🎯 What it does: This paper proposes a soft channel pruning method called IRPruneDet based on wavelet structural regularization, aimed at achieving lightweight and efficient infrared small target detection models.
Is a Large Language Model a Good Annotator for Event Extraction?
Ruirui Chen (Institute of High Performance Computing), Dongkyu Choi (Institute of High Performance Computing)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Using large language models (LLMs) as expert annotators to generate additional annotated samples that conform to the original data distribution for the event extraction task, thereby alleviating data scarcity and long-tail distribution issues, and improving model performance.
IS-DARTS: Stabilizing DARTS through Precise Measurement on Candidate Importance
Hongyi He (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
CodeOptimizationNeural Architecture SearchImage
🎯 What it does: The IS-DARTS method is proposed, which evaluates the importance of candidate operations by measuring their Fisher information and combines a multi-step shrinking strategy to improve and stabilize DARTS.
🎯 What it does: This paper proposes ISP-Teacher, which utilizes self-supervised ISP degradation learning and a discretization regularization Teacher-Student framework to achieve target detection in unlabeled low-light environments.
🎯 What it does: A model called IVP-VAE based on continuous-time variational autoencoders has been designed for modeling, predicting, and classifying unordered and sparse electronic health record (EHR) time series.
🎯 What it does: A joint denoising and demosaicing (JDD) network is proposed for spike cameras with color filter arrays, capable of recovering high-quality color images from the binarized Bayer-pattern spike stream.
Joint Learning Neuronal Skeleton and Brain Circuit Topology with Permutation Invariant Encoders for Neuron Classification
Minghui Liao (Wuhan University), Bo Du (Wuhan University)
CodeClassificationGraph Neural NetworkGraph
🎯 What it does: A neuron classification framework called NeuNet was designed and implemented, which simultaneously utilizes neuron skeleton morphology and brain circuit topology information, and two whole-brain reconstruction datasets were made public.
Jointly Improving the Sample and Communication Complexities in Decentralized Stochastic Minimax Optimization
Xuan Zhang (Pennsylvania State University), Yangyang Xu (Rensselaer Polytechnic Institute)
CodeOptimizationTabular
🎯 What it does: A single-cycle distributed gradient descent ascent algorithm DGDA-VR is proposed to solve non-convex-strongly concave minimax problems with multiple agents, distributed data, and access only to unbiased stochastic gradients.
Jointly Modeling Spatio-Temporal Features of Tactile Signals for Action Classification
Jimmy Lin (Institute for AI Industry Research), Yang Liu (Institute for AI Industry Research)
CodeClassificationTransformerTime Series
🎯 What it does: A Transformer model that can simultaneously capture spatial and temporal features of tactile signals (STAT) is proposed, enhancing action classification performance through spatial embedding, temporal embedding, and a temporal pre-training task.
Journey to the Center of the Knowledge Neurons: Discoveries of Language-Independent Knowledge Neurons and Degenerate Knowledge Neurons
Yuheng Chen (Institute of Automation, Chinese Academy of Sciences), Jun Zhao (Institute of Automation, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelText
🎯 What it does: This study investigates the localization and characteristics of knowledge neurons in multilingual pre-trained language models, proposing the AMIG method and discovering language-independent knowledge neurons and degenerate knowledge neurons.