AAAI Conference on Artificial Intelligence · 1014 papers
Data-Free Generalized Zero-Shot Learning
Bowen Tang (Beihang University), Dong Xu (The University of Hong Kong)
CodeClassificationRecognitionGenerationTransformerVision Language ModelGenerative Adversarial NetworkImage
🎯 What it does: A data-free zero-shot learning (DFZSL) framework is proposed without any real image data, achieving recognition of new categories through the recovery, alignment, and generation of the CLIP base classifier.
🎯 What it does: A data-independent model stealing attack that only queries hard labels is proposed, capable of simultaneously stealing the target model's accuracy and robustness.
🎯 What it does: A neural network predictor DCLP based on contrastive learning is designed and implemented, and the difficulty of positive samples is scheduled through curriculum learning, thereby improving prediction performance under the condition of using only a very small amount of trained network data.
De-biased Attention Supervision for Text Classification with Causality
Yiquan Wu (Zhejiang University), Kun Kuang (Zhejiang University)
CodeClassificationExplainability and InterpretabilityRecurrent Neural NetworkTransformerText
🎯 What it does: To address the label bias and word frequency bias in attention supervision (AS) for text classification tasks, a Debiased Attention Supervision (DAS) method is proposed.
Dealing with Numeric and Metric Time Constraints in PDDL3 via Compilation to Numeric Planning
Luigi Bonassi (University of Brescia), Enrico Scala (University of Brescia)
CodeTabular
🎯 What it does: This study investigates a technique for compiling PDDL3 planning problems that include mixed propositional and numerical conditions, as well as metric time constraints, into unconstrained numerical planning problems, enabling the use of existing numerical planners for solving them.
Debiasing Multimodal Sarcasm Detection with Contrastive Learning
Mengzhao Jia (Shandong University), Liqiang Jing (University of Texas at Dallas)
CodeClassificationDomain AdaptationTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: A debiased multimodal sarcasm detection framework is proposed, and an OOD evaluation task is designed to address the generalization problem caused by text bias.
Decentralized Monte Carlo Tree Search for Partially Observable Multi-Agent Pathfinding
Alexey Skrynnik (Federal Research Center for Computer Science and Control of Russian Academy of Sciences), Aleksandr Panov (Federal Research Center for Computer Science and Control of Russian Academy of Sciences)
CodeOptimizationReinforcement Learning
🎯 What it does: A decentralized lifelong multi-agent path planning method based on MCTS and a lightweight learnable policy (MATS-LP) is proposed.
Deciphering Compatibility Relationships with Textual Descriptions via Extraction and Explanation
Yu Wang (University of California San Diego), Julian McAuley (University of California San Diego)
CodeRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A PFE dataset was constructed and a two-stage pipeline model was proposed to generate compatible natural language explanations for clothing pairs.
Deciphering Raw Data in Neuro-Symbolic Learning with Provable Guarantees
Lue Tao (Nanjing University), Yuan Jiang (Nanjing University)
CodeClassificationImage
🎯 What it does: The paper studies the impact of knowledge bases on learning outcomes in neural symbolic learning, proposing a theoretical evaluation of knowledge bases through probability matrices and rank criteria, and validating their feasibility across different tasks.
Decoding Global Preferences: Temporal and Cooperative Dependency Modeling in Multi-Agent Preference-Based Reinforcement Learning
Tianchen Zhu (Beihang University), Jianxin Li (Beihang University)
CodeTransformerReinforcement LearningSequential
🎯 What it does: A Transformer-based multi-agent preference learning framework called MAPT is proposed to learn reward functions from human preferences.
Decomposing Semantic Shifts for Composed Image Retrieval
Xingyu Yang (Wuhan University), Jing Zhang (University of Sydney)
CodeRetrievalTransformerVision Language ModelImageText
🎯 What it does: This paper proposes treating text in composite image retrieval as instructions and splits the retrieval process into two steps: degradation (reference → visual prototype) and upgrading (visual prototype → target) through the Semantic Shift Network (SSN). In this process, the text is deconstructed in both forward (upgrading) and backward (degradation) directions, guiding the generation of visual prototypes and ultimately obtaining representations of target images. End-to-end training is achieved through cross-entropy retrieval loss and KL regularization.
🎯 What it does: This paper proposes DIVIDE, a robust multi-view clustering method that utilizes random walks to identify high-order neighbors and simultaneously corrects false negatives and false positives through decoupled contrastive learning.
Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees
Weijia Zhang (University of Newcastle), Xuanhui Zhang (Nanjing University)
CodeTabularTime SeriesElectronic Health Records
🎯 What it does: A deep Copula-based survival analysis method is proposed, which can automatically learn the dependency structure from data without pre-specifying the Copula in the presence of dependent censoring;
Deep Homography Estimation for Visual Place Recognition
Feng Lu (Tsinghua University), Chun Yuan (Tsinghua University)
CodeRecognitionRetrievalTransformerImage
🎯 What it does: A deep homography estimation network (DHE) based on Transformer is proposed for two-stage retrieval in visual location recognition, replacing RANSAC with differentiable homography fitting for geometric verification, and designing an unsupervised REI loss for end-to-end joint training.
🎯 What it does: A two-stage JARNet network is proposed, which combines optical flow correction and spatial frequency domain residual modules to recover distortions and blurriness in line-scan camera images.
Deep Semantic Graph Transformer for Multi-View 3D Human Pose Estimation
Lijun Zhang (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences), Yu Shi (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences)
🎯 What it does: A multi-view 3D human pose estimation framework based on a deep semantic graph Transformer is proposed, which integrates multi-view semantic features and gradually fuses spatial-temporal information.
Deep Unfolded Network with Intrinsic Supervision for Pan-Sharpening
Hebaixu Wang (Wuhan University), Jiayi Ma (Wuhan University)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: An interpretable deep unfolding network is proposed, which achieves the fusion of PAN and MS images (pansharpening) through spatial consistency and spectral projection priors.
CodeAutonomous DrivingTransformerContrastive LearningSimultaneous Localization and MappingPoint Cloud
🎯 What it does: DeepPointMap proposes a unified neural descriptor framework that achieves high-precision LiDAR SLAM and lightweight map construction using sparse neural features;
Delegation-Relegation for Boolean Matrix Factorization
Florent Avellaneda (University of Quebec at Montreal), Roger Villemaire (University of Quebec at Montreal)
CodeOptimizationTabular
🎯 What it does: This paper proposes two operators, delegation and relegation, to simplify Boolean matrices, thereby reducing the number of 1 elements that need to be factored while maintaining or guaranteeing the minimum rank, significantly lowering the solving time for constraint-based BMF.
Yacine Izza (National University of Singapore), Joao Marques-Silva (IRIT CNRS)
CodeExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: This paper proposes and implements 'inflated explanations' by expanding the range (or set) of feature values in traditional suspicious explanations to provide more informative explanations.
Designing Biological Sequences without Prior Knowledge Using Evolutionary Reinforcement Learning
Xi Zeng (Northwestern Polytechnical University), Jiajie Peng (Northwestern Polytechnical University)
CodeOptimizationConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningBiomedical Data
🎯 What it does: A novel evolutionary reinforcement learning framework named ERLBioSeq is proposed for designing biological sequences such as DNA, RNA, and proteins under conditions of no prior knowledge.
Detecting and Preventing Hallucinations in Large Vision Language Models
Anisha Gunjal (University of Texas), Erhan Bas (Scale AI)
CodeOptimizationTransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: This paper constructs the M-HalDetect multimodal hallucination detection dataset and trains reward models and direct optimization models based on this dataset to reduce the hallucination rate of large visual language models.
🎯 What it does: This study investigates the detectability of unlearnable examples and defense methods, proving their linear separability and proposing two detection algorithms; subsequently, a defense scheme based on strong data augmentation and adversarial noise generated by simple networks is designed; and a theoretical boundary between toxicity budget and adversarial training budget is provided.
Detection-Based Intermediate Supervision for Visual Question Answering
Yuhang Liu (ByteDance Inc.), Dangyang Chen (Ping An Property & Casualty Insurance Company of China)
CodeRecognitionObject DetectionExplainability and InterpretabilityTransformerVision Language ModelMultimodality
🎯 What it does: A detection-based intermediate supervision (DIS) method is proposed, which transforms intermediate inference results into serializable detection boxes, ground truths, and answer sequences. It utilizes an autoregressive decoder to supervise the inference state of the visual question answering model, improving answer accuracy and reasoning consistency.
DGA-GNN: Dynamic Grouping Aggregation GNN for Fraud Detection
Mingjiang Duan (Zhejiang University), Xinyu Wang (Zhejiang University)
CodeAnomaly DetectionGraph Neural NetworkGraphTabularFinance Related
🎯 What it does: A dynamic grouping aggregation graph neural network (DGA-GNN) is proposed, which processes non-additive attributes through decision tree binning encoding and uses feedback-based dynamic grouping to bipartition neighboring nodes, ultimately achieving more discriminative fraud detection.
DGCLUSTER: A Neural Framework for Attributed Graph Clustering via Modularity Maximization
Aritra Bhowmick (New York University), Sourav Medya (University of Illinois)
CodeGraph Neural NetworkGraph
🎯 What it does: Proposes the DGCLUSTER framework, which utilizes GNN to learn node embeddings and achieves modular maximization of graph clustering without the need to preset the number of clusters based on embedding similarity.
🎯 What it does: For the text-video retrieval task, a Dynamic Global-Local Prompt Tuning (DGL) framework is proposed, which utilizes a shared latent space to generate cross-modal prompts and incorporates global-local attention in the visual encoder to capture both overall and frame-level information of the video.
🎯 What it does: This paper proposes the DI-V2X model, which utilizes a teacher-student distillation framework to learn domain-invariant 3D object detection representations in vehicle-infrastructure collaborative perception.
🎯 What it does: This paper proposes the 'Partial Label' problem in cross-domain image retrieval (PCIR) and designs a new method called DiDA, which utilizes Prototype Score Unit Learning (PSUL) and Prototype Domain Alignment (PBDA) to achieve label disambiguation and cross-domain feature alignment, thereby improving retrieval performance.
DifAttack: Query-Efficient Black-Box Adversarial Attack via Disentangled Feature Space
Jun Liu (University of Macau), Jinyu Tian (Beijing Normal University)
CodeAdversarial AttackAuto EncoderImage
🎯 What it does: This paper proposes a black-box adversarial attack method called DifAttack based on feature space decoupling. It first uses an autoencoder to split the latent features of images into adversarial features and visual features. During the attack phase, only the adversarial features are optimized to generate adversarial samples while keeping the visual features unchanged, and finally controls the perturbation magnitude through projection.
🎯 What it does: This paper proposes DiffSED, a sound event detection framework based on a denoising diffusion model, which utilizes noise latent queries to progressively denoise in the Transformer decoder, generating event time boundaries and category labels.
Diffusion Language-Shapelets for Semi-supervised Time-Series Classification
Zhen Liu (South China University of Technology), Qianli Ma (South China University of Technology)
CodeClassificationExplainability and InterpretabilityDiffusion modelContrastive LearningTime Series
🎯 What it does: The DiffShape model is proposed and implemented, which combines self-supervised diffusion learning with language contrastive learning to generate interpretable shape subsequences (shapelets) and is trained on semi-supervised time series classification tasks.
🎯 What it does: A diffusion probability model-based edge detector called DiffusionEdge is proposed, which can generate accurate and sharp edge maps directly at the original resolution without the need for post-processing.
🎯 What it does: This paper proposes DiffusionTrack, a multi-object tracking framework that unifies object detection and association into a denoising diffusion process for tracking.
CodeAnomaly DetectionGraph Neural NetworkReinforcement LearningContrastive LearningGraphFinance Related
🎯 What it does: This paper proposes DiG-In-GNN, which generates distinguishable guiding nodes through multi-scale contrastive learning and uses reinforcement learning for fine-grained neighbor selection, enhancing the effectiveness of multi-relational graph fraud detection.
DINGO: Towards Diverse and Fine-Grained Instruction-Following Evaluation
Zihui Gu (Renmin University of China), Ju Fan (Renmin University of China)
CodeLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A multi-level, fine-grained evaluation dataset DINGO has been constructed, covering diverse user instructions, and various LLMs have been evaluated using LLM-as-a-judge.
🎯 What it does: A neural network estimator for directly calculating the likelihood ratio between parameter pairs (DNRE) is proposed, and based on this, a Monte-Carlo posterior approximation and numerically stable gradient estimation are derived to compare Hamiltonian Monte Carlo (HMC) and random walk Metropolis-Hastings (MH) in simulating likelihood-free inference, with the method applied to the quadrotor design problem.
🎯 What it does: A Dirichlet-based prediction calibration method (DPC) is proposed, which reduces the overconfidence problem in learning from noisy labels and improves example selection effectiveness by modifying softmax and training with the Dirichlet distribution.
🎯 What it does: Proposes the DiSCO framework, which utilizes the diffusion Schrödinger bridge for post-processing optimization of existing molecular conformations, making the generated 3D conformations closer to the true energy distribution.
Discrepancy and Uncertainty Aware Denoising Knowledge Distillation for Zero-Shot Cross-Lingual Named Entity Recognition
Ling Ge (Beihang University), Hong Zhang (National Computer Network Emergency Response Technical Team)
CodeRecognitionKnowledge DistillationText
🎯 What it does: The DenKD model is proposed, which uses denoising knowledge distillation to address the issue of pseudo-label noise in zero-shot cross-lingual named entity recognition.
🎯 What it does: This paper proposes the Discriminative Forest GAN (ForestGAN), which constructs a forest of multiple independent discriminators using a bootstrap method to enhance the diversity of GAN generation.
🎯 What it does: Transforming a Transformer-based autoregressive (AR) model into a non-autoregressive (NAR) model through knowledge distillation for fast inference in vehicle routing problems (VRP).
🎯 What it does: A self-distillation-based framework DIRK is proposed for instance-dependent partial label learning (IDPLL), and a representation refinement module DIRK-REF is added to enhance feature representation and classification performance.
🎯 What it does: This paper proposes DistilVPR—a cross-modal knowledge distillation pipeline that enhances the performance of unimodal students in visual place recognition by utilizing multi-agent and multi-manifold relationships.
🎯 What it does: The study proposes a Variational Inference with Variational Adversarial Networks (VIV) to automatically generate instrumental variables that satisfy the conditions of relevance, exclusivity, and exogeneity in the absence of effective candidate instrumental variables, to support causal inference.
🎯 What it does: A method for unbiased distribution estimation for slide recommendation (SUnO) is proposed, which can estimate the reward distribution of the target policy based on offline data.
🎯 What it does: This paper proposes the DiG-SST framework, which combines a prefix-enhanced end-to-end speech translation model with a dynamic read-write strategy based on normalized cosine divergence, achieving low-latency and high-quality synchronous speech translation.
🎯 What it does: In the context of federated learning, the DACS method is proposed for the domain generalization task of person re-identification (re-ID), utilizing a style transfer model to generate diverse and realistic synthetic samples, thereby enhancing the generalization ability of local models.
🎯 What it does: A hybrid pre-training framework for person search is proposed, which utilizes data from detection and re-identification sub-tasks for joint learning and reduces domain differences through a domain alignment module.
DME: Unveiling the Bias for Better Generalized Monocular Depth Estimation
Songsong Yu (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
CodeDepth EstimationConvolutional Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: Analyzes the long-tail distribution and its correlation with simulation in monocular depth estimation, and proposes a Distance-based Multi-Expert (DME) network that achieves depth prediction fusion across different distance segments through pixel-level routing; simultaneously designs a two-stage training strategy and experimentally validates its performance across multiple datasets.
DMMR: Cross-Subject Domain Generalization for EEG-Based Emotion Recognition via Denoising Mixed Mutual Reconstruction
Yiming Wang (Xi'an Jiaotong University), Yujiao Tang (Xi'an Jiaotong University)
CodeRecognitionDomain AdaptationAuto EncoderTime SeriesBiomedical Data
🎯 What it does: A self-supervised mixed mutual reconstruction (DMMR) model is proposed for cross-subject domain generalization in EEG emotion recognition.
🎯 What it does: A unified document image enhancement framework, DocNLC, is proposed, which achieves unified processing of various degradation types through normalization and latent contrastive learning.
🎯 What it does: This paper proposes a Disentangled Object-Centric Transformer (DOCTR) that simultaneously performs instance segmentation, pose estimation, and mesh reconstruction of point cloud scenes through a single network, addressing the challenges of modeling and optimizing inter-object relationships in traditional multi-stage pipelines.
🎯 What it does: This paper studies backdoor attacks in few-shot learning (FSL) and proposes an FLBA attack method based on maximizing embedding bias triggers and hidden perturbations.
🎯 What it does: A differentiable algorithm that combines graph neural networks and Lagrangian decomposition is proposed for efficiently solving the LP relaxation of 0-1 integer linear programming.
🎯 What it does: This paper proposes Vital Phase Augmentation (VIPAug), which enhances the model's robustness to input distortion and phase fluctuations by performing limited transformations on the image phase in the frequency domain and combining it with amplitude enhancement.
🎯 What it does: This paper proposes a domain-invariant learning algorithm for Gaussian processes, DIL-GP, which enhances its generalization ability to out-of-domain data by adaptively partitioning data and performing a min-max optimization of the IRM penalty, and extends it to Bayesian optimization.
🎯 What it does: By introducing a specific domain foundation model (LSDM) and conducting domain-controlled prompt learning on the visual and language branches, the zero-shot recognition performance of CLIP on remote sensing and medical images is improved.
Double Buffers CEM-TD3: More Efficient Evolution and Richer Exploration
Sheng Zhu (Jilin University), Daolong An (Jilin University)
CodeReinforcement Learning
🎯 What it does: This paper proposes Double Buffers CEM-TD3, which improves the evolutionary and gradient learning process of the original CEM-TD3 using a double buffering mechanism, enhancing evolutionary efficiency and population diversity.
DP-AdamBC: Your DP-Adam Is Actually DP-SGD (Unless You Apply Bias Correction)
Qiaoyue Tang (University of British Columbia), Mathias Lécuyer (University of British Columbia)
CodeOptimizationSafty and PrivacyGraph Neural NetworkImageTextGraph
🎯 What it does: An optimization algorithm called DP-AdamBC is proposed to correct the bias of DP Adam, restoring the original behavior of Adam under privacy training;
DPA-P2PNet: Deformable Proposal-Aware P2PNet for Accurate Point-Based Cell Detection
Zhongyi Shui (Zhejiang University), Lin Yang (Westlake University)
CodeObject DetectionSegmentationConvolutional Neural NetworkContrastive LearningImageBiomedical Data
🎯 What it does: An end-to-end point-based cell detection model DPA-P2PNet is proposed, supporting multi-scale decoding, deformable point candidates, and multi-field input, and for the first time, self-supervised pre-training is conducted on a large-scale immunohistochemical image dataset.
DR-Label: Label Deconstruction and Reconstruction of GNN Models for Catalysis Systems
Bowen Wang (Chinese University of Hong Kong), Pheng Ann Heng
CodeGraph Neural NetworkGraph
🎯 What it does: This paper studies a graph neural network supervision and prediction strategy named DR-Label, aimed at reducing the multiplicity of edge representations and sensitivity to changes in graph structure in catalyst-adsorbate systems.
🎯 What it does: A dual-stream analytical learning (DS-AL) framework is proposed for exemplar-free class incremental learning (CIL), achieving continuous learning without catastrophic forgetting through the mainstream C-RLS closed-form solution and the compensation stream DAC module.
DSD²: Can We Dodge Sparse Double Descent and Compress the Neural Network Worry-Free?
Victor Quétu (Telecom Paris Institute Polytechnique de Paris), Enzo Tartaglione (Telecom Paris Institute Polytechnique de Paris)
CodeCompressionKnowledge DistillationImage
🎯 What it does: This study proposes a learning framework that avoids sparse double descent through knowledge distillation, enhancing the performance of sparse networks without sacrificing accuracy.
DTF-AT: Decoupled Time-Frequency Audio Transformer for Event Classification
Tony Alex (Surrey Institute for People Centred AI), Philip JB Jackson (Surrey Institute for People Centred AI)
CodeClassificationTransformerAudio
🎯 What it does: Proposes the DTF-AT audio Transformer, which achieves audio event classification through a time-frequency decoupled convolutional branch and a multi-scale MaxViT structure;
🎯 What it does: This paper proposes the DTMFormer dynamic Token merging Transformer block, which addresses the attention collapse problem in medical image segmentation and improves convergence and performance.
Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks
Xiaofan Que (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
CodeMeta LearningImage
🎯 What it does: A dual-layer curriculum meta-learning framework (DCML) is proposed, which constructs curricula at both the category level and the sample level to improve robustness against noisy few-shot learning.
🎯 What it does: This paper proposes a Dual Prior Enhanced Decoding Network (DPADN), which splits the human-object interaction detection task into two sub-tasks: human-object pair detection and interaction recognition. It utilizes prior information provided by external object classifiers and verb discriminators to assist decoding, thereby alleviating the recognition difficulties caused by long-tail distributions.
Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging
Fulin Luo (Chongqing University), Tan Guo (Chongqing University)
CodeRestorationCompressionTransformerImage
🎯 What it does: This paper proposes a Dual-Window Multi-Scale Transformer (DWMT) based on Transformer for the reconstruction of hyperspectral snapshot compressive imaging, constructing a two-stage U-Net network and integrating a dual-branch encoder.
🎯 What it does: A multi-view action recognition framework DVANet based on a Transformer decoder is proposed, which separates action features from view features using learnable queries and achieves view-invariant action representation through supervised contrastive learning.
🎯 What it does: This paper proposes a Dynamic Reactive Spiking Graph Neural Network (DRSGNN) that enhances the model's expressive power and energy efficiency through optimizable threshold spiking neurons and learnable graph positional information.
Dynamic Semantic-Based Spatial Graph Convolution Network for Skeleton-Based Human Action Recognition
Jianyang Xie (University of Liverpool), Yalin Zheng (University of Liverpool)
CodeRecognitionGraph Neural NetworkVideo
🎯 What it does: A dynamic semantic-driven spatial graph convolutional network (DS-GCN) is proposed, which achieves more accurate modeling of skeletal actions by dynamically encoding the types of joints and edges in graph convolution.
🎯 What it does: In semi-supervised continual learning, the issue of catastrophic forgetting with unlabeled data is studied, and a Dynamic Subgraph Distillation (DSGD) method is proposed;
🎯 What it does: A Dynamic Weighted Combiner (DWC) is proposed for Mixed Modal Image Retrieval (MMIR), addressing issues of modality importance imbalance, label noise, and modality gap through feature editing, dynamic soft labels, and multimodal contrastive learning.
E2E-AT: A Unified Framework for Tackling Uncertainty in Task-Aware End-to-End Learning
Wangkun Xu (Imperial College London), Fei Teng (Imperial College London)
CodeOptimizationAdversarial AttackTime Series
🎯 What it does: This paper proposes a unified robust framework E2E-AT, which systematically models and trains two types of uncertainties in end-to-end learning: input features and unpredictable constraint optimization parameters.
Earthfarsser: Versatile Spatio-Temporal Dynamical Systems Modeling in One Model
Hao Wu (University of Science and Technology of China), Kun Wang (University of Science and Technology of China)
CodeConvolutional Neural NetworkTransformerVideoTime SeriesPhysics Related
🎯 What it does: This paper presents EarthFarseer, a unified spatiotemporal modeling framework that balances local convolution with global Fourier-Transformer, and introduces a continuous-time Fourier-temporal transform to capture long-term temporal dependencies.
CodeClassificationAnomaly DetectionConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: To address the problem of external sample detection under long-tail distribution, the EAT framework is proposed, which combines virtual labels, multiple rejection categories, tail category enhancement, model ensemble, and fine-tuning.
🎯 What it does: This paper proposes a heterogeneous dynamic graph model ECHO-GL based on earnings call semantics to predict stock price movements over various time windows.
EcomGPT: Instruction-Tuning Large Language Models with Chain-of-Task Tasks for E-commerce
Yangning Li (Tsinghua University), Yong Jiang (Alibaba Group)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: A dataset for instruction tuning aimed at e-commerce, EcomInstruct (approximately 2.5 million instructions, 134 tasks), and a specialized LLM EcomGPT trained on this dataset are proposed.
EDA: Evolving and Distinct Anchors for Multimodal Motion Prediction
Longzhong Lin (Zhejiang University), Yue Wang (Zhejiang University)
CodeAutonomous DrivingTransformerMultimodality
🎯 What it does: Proposed and implemented an Evolvable and Unique Anchor (EDA) method, which significantly enhances regression and scoring capabilities in multimodal motion prediction by gradually updating anchors through a multi-layer decoder and using NMS before matching.
CodeRecurrent Neural NetworkTransformerLarge Language ModelGraph
🎯 What it does: This paper proposes knowledge graph embedding (KGE) based on language models for fast and data-efficient editing tasks (EDIT/ADD), and constructs the corresponding datasets.
Efficient Axiomatization of OWL 2 EL Ontologies from Data by Means of Formal Concept Analysis
Francesco Kriegel (Technische Universitat Dresden)
CodeGraph
🎯 What it does: This paper proposes an algorithm based on Formal Concept Analysis for efficiently constructing a complete OWL 2 EL TBox (concept inclusion, range restrictions, and role inclusion) from graph data, and provides an implementable Scala implementation.
🎯 What it does: This paper proposes an efficient image super-resolution method called ECDP, which utilizes continuous-time conditional diffusion models and probabilistic flow sampling.
Efficient Nonparametric Tensor Decomposition for Binary and Count Data
Zerui Tao (Tokyo University of Agriculture and Technology), Qibin Zhao (RIKEN Center for Advanced Intelligence Project)
CodeTabularStochastic Differential Equation
🎯 What it does: This paper proposes an efficient non-parametric tensor decomposition model, ENTED, for handling binary and count tensor data, aiming to achieve higher quality tensor completion.
EG-NAS: Neural Architecture Search with Fast Evolutionary Exploration
Zicheng Cai (Guangdong University of Technology), Yutao Lai (Guangdong University of Technology)
CodeNeural Architecture SearchImage
🎯 What it does: This paper proposes EG-NAS, a neural architecture search framework that combines gradient descent with an improved evolutionary strategy, aiming to reduce search costs and avoid getting trapped in local optima.
Anne-Marie George (University of Oslo), Christos Dimitrakakis (University of Neuchatel)
Code
🎯 What it does: The paper models the problem of collecting voter preferences as a dueling bandits problem and proposes a method for estimating Kemeny rankings under PAC (Probably Approximately Correct) conditions.
Embracing Language Inclusivity and Diversity in CLIP through Continual Language Learning
Bang Yang (Peking University), Yuexian Zou (Peking University)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a continuous language learning framework based on CLIP, called CLL-CLIP, which focuses on training only the scalable token embedding layer. It combines cross-modal InfoNCE and cross-language MSE objectives to achieve alignment between images and multilingual text, and employs the TEIR method for unified distribution initialization and frequency regularization of token embeddings to mitigate catastrophic forgetting.
🎯 What it does: This paper proposes an Early Mixed GAN Attack (EMGAN) for Split Federated Learning, which effectively extracts the server-side model through three modules: early learning, multiple GANs, and ProperMix.
🎯 What it does: This paper proposes an emotional dialogue speech synthesis model (ECSS) that can accurately capture and generate emotional speech within the context of dialogue.
Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations
Changqing Qiu (Beijing Institute of Technology), Yining Zhang (Peking University)
CodeExplainability and InterpretabilityConvolutional Neural NetworkScore-based ModelImage
🎯 What it does: A novel explanation method FG-CAM is studied, which gradually enhances the resolution of explanations by utilizing the relationships between feature maps of adjacent layers, thereby generating fine-grained and highly credible interpretable results at shallow and input layers.
🎯 What it does: A dual-layer graph self-supervised pre-training framework DGPM is proposed, which automatically discovers subgraph patterns and achieves node-subgraph interactive learning.