AAAI Conference on Artificial Intelligence Β· 1442 papers
Excluding the Impossible for Open Vocabulary Semantic Segmentation
Shiyuan Zhao (China University of Petroleum), Shuai Shao (China University of Petroleum)
CodeSegmentationContrastive LearningImage
π― What it does: Using a reverse approach of 'excluding the impossible', combined with CLIP/CLIPN, ELSE-Net is constructed to achieve open vocabulary semantic segmentation.
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Constructed the ExcluIR dataset and benchmark to evaluate the performance of retrieval models on exclusive queries, and conducted experiments on various retrieval models.
ExDDI: Explaining Drug-Drug Interaction Predictions with Natural Language
Zhaoyue Sun (University of Warwick), Yulan He (King's College London)
CodeExplainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataRetrieval-Augmented Generation
π― What it does: This paper proposes a natural language explanation task for generating drug-drug interaction (DDI) predictions and constructs a series of ExDDI models that can generate explanations for both positive and negative interactions while making predictions.
Expected Hypervolume Improvement Is a Particular Hypervolume Improvement
Jingda Deng (Xi'an Jiaotong University), Hui Li (Xi'an Jiaotong University)
CodeOptimizationGaussian SplattingPoint Cloud
π― What it does: This paper proposes to rewrite the Expected Hypervolume Improvement (EHVI) and its batch version (q EHVI) into a specific form of hypervolume improvement, providing the corresponding analytical expressions.
Expensive Multi-Objective Bayesian Optimization Based on Diffusion Models
Bingdong Li (East China Normal University), Aimin Zhou (East China Normal University)
CodeOptimizationDiffusion model
π― What it does: A Pareto set learning algorithm based on a combined diffusion model, CDM-PSL, is proposed to address expensive multi-objective Bayesian optimization problems.
CodeClassificationExplainability and InterpretabilityKnowledge DistillationTransformerVision Language ModelImageMultimodality
π― What it does: An explainable bottleneck model (XBM) is proposed, which directly generates natural language explanations through a pre-trained vision-language encoder-decoder to predict final labels, eliminating the dependence on a predefined set of concepts.
Explicit and Implicit Graduated Optimization in Deep Neural Networks
Naoki Sato (Meiji University), Hideaki Iiduka (Meiji University)
CodeOptimizationImage
π― What it does: Evaluated the performance of explicit gradient optimization algorithms on traditional benchmark functions and deep neural networks, and proposed an implicit gradient optimization algorithm using SGD with momentum, providing theoretical convergence analysis and experimental validation.
Exploit Gradient Skewness to Circumvent Byzantine Defenses for Federated Learning
Yuchen Liu (Zhejiang University), Gang Chen (Zhejiang University)
CodeFederated LearningAdversarial AttackImage
π― What it does: The study investigates the phenomenon of gradient skewness in non-IID federated learning environments and proposes a two-stage attack method based on gradient skewness called STRIKE, aimed at bypassing existing Byzantine defenses.
π― What it does: A continuous updating occupancy prediction framework (CMOP) is proposed, which dynamically iterates and updates the 3D occupancy volume by combining historical occupancy information with real-time optical flow information.
Exploiting Multimodal Spatial-temporal Patterns for Video Object Tracking
Xiantao Hu (Nanjing University), Jian Yang (Nanjing University)
CodeObject TrackingTransformerVideoMultimodality
π― What it does: A unified tracking framework STTrack based on multimodal spatiotemporal patterns is proposed, capable of continuously capturing target motion information and achieving precise localization across various modalities such as RGB, TIR, Depth, and Event.
Explore What LLM Does Not Know in Complex Question Answering
Xin Lin (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeRetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: A RAG framework KEQA based on question-answer knowledge assessment is proposed, which can accurately identify knowledge gaps in LLMs and only retrieve missing knowledge, thereby improving reasoning quality.
Exploring Activation Patterns of Parameters in Language Models
Yudong Wang (Peking University), Zhifang Sui (Peking University)
CodeLarge Language ModelText
π― What it does: By evaluating parameter activation based on gradient first-order metrics, we analyze the activation distribution of different layers under same-domain and cross-domain inputs, and propose the LLMDcos metric.
π― What it does: This paper studies the ability of large language models (LLMs) to capture and update user intentions in real-time during conversations, and based on this, proposes an evaluation framework that does not use Chain-of-Thought (CoT) or prompt engineering.
Exploring Enhanced Contextual Information for Video-Level Object Tracking
Ben Kang (Dalian University of Technology), Dong Wang (Baidu Inc.)
CodeObject TrackingVideo
π― What it does: This paper proposes the MCITrack framework, which utilizes Mamba's hidden states to continuously record and transmit video-level contextual information to enhance the accuracy of visual object tracking.
Exploring More from Multiple Gait Modalities for Human Identification
Dongyang Jin (Southern University of Science and Technology), Shiqi Yu (Alibaba Group)
CodeRecognitionOptical FlowVideoMultimodality
π― What it does: This paper studies and compares the roles of three pedestrian gait modalities (shape, human segmentation, optical flow) in multimodal gait recognition and proposes a C2 Fusion method based on shared and differential features.
π― What it does: In federated learning, to address the issue of missing categories and the decline in recognition rates for minority categories caused by uneven label distribution, the FedVLS method is proposed, which simultaneously applies knowledge distillation for missing categories and logits suppression during local training.
π― What it does: This paper proposes an external reliable information-enhanced multimodal contrastive learning framework ERIC-FND, which enriches text representation with entity descriptions and improves fake news detection through multimodal contrast and cross-modal interaction.
JeongYeon Nam (NAVER Cloud AI), Taeho Kil (NAVER AI Lab)
CodeRetrievalExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Using pre-trained CLIP to detect fine-grained mismatches (such as word-level mismatches) between images and text through gradient attribution methods.
Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference
Jorge GarcΓa-Carrasco (University of Alicante), Juan Trujillo (University of Alicante)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
π― What it does: This paper proposes an automated pruning method based on Mechanism Interpretability (MI), which can extract the minimal sub-circuit from large language models (LLMs) that is only used for executing specific tasks, enabling independent inference of the sub-model without additional training or fine-tuning.
π― What it does: Designed an audio-synchronized viewing task and constructed a dataset of 20,000 fixation points from 8 participants, proposing the EyEar framework based on physical information dynamics to predict human eye gaze trajectories in visual scenes.
Yannan Bai (Duke University), Ian Zhang (Duke University)
CodeTabular
π― What it does: This paper proposes new concepts of fair shares based on individuals' utilities of other players, such as 'cake cutting shares' and 'envy-free shares', and studies the achievable approximation factors in the worst-case scenario.
Md Mahmudur Rahman (University of Maryland), Sanjay Purushotham (University of Maryland)
CodeFederated LearningBiomedical Data
π― What it does: This paper proposes FairFSA, a fair federated survival analysis framework that combines FPV and DRO to train fair and debiased survival prediction models on multi-institutional data.
π― What it does: By introducing category-agnostic all-zero inputs and incorporating uniformity loss, the fairness and robustness of classification models are enhanced.
Fairness-Accuracy Trade-Offs: A Causal Perspective
Drago Plecko, Elias Bareinboim (Columbia University)
CodeTabularFinance Related
π― What it does: This paper analyzes the trade-off between fairness and accuracy from a causal perspective, and proposes Path-Specific Excess Loss (PSEL) and the corresponding Causal Fairness Utility Ratio (CFUR) to quantify the improvement in fairness and the increase in prediction error under different causal paths.
FairTP: A Prolonged Fairness Framework for Traffic Prediction
Jiangnan Xia (Central South University), Jiannong Cao (Hong Kong Polytechnic University)
CodeGraph Neural NetworkTime Series
π― What it does: The FairTP framework is proposed, which studies long-term fairness in traffic prediction and achieves fair predictions at the regional and sensor levels through state recognition and dynamic sampling.
CodeExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: A method for edge attribution based on the computational tree perspective of self-attention message passing neural networks (Att-GNN) called GATT is proposed;
π― What it does: A cross-domain few-shot medical image segmentation model, FAMNet, is proposed to address the domain transfer issues caused by different imaging techniques.
π― What it does: The Fast and Slow Gradient Generation (FSG) framework is proposed, which approximates the gradient of the non-differentiable quantization function in binary neural network (BNN) training through a dual-branch super network, and introduces Historical Gradient Storage (HGS) and Layer Recognition Embedding (LRE) to enhance gradient quality.
Fast Computing of Dung Semantics in Acyclic Probabilistic Argumentation Frameworks
Stefano Bistarelli (University of Perugia), Carlo Taticchi (University of Perugia)
CodeComputational EfficiencyGraph
π― What it does: This paper proposes a method to quickly and accurately calculate the acceptance probabilities of arguments in acyclic (SCG, DAG) abstract argumentation frameworks using Dung semantics within the framework of probability theory (Constellation perspective).
π― What it does: This paper proposes a fast incomplete multi-view clustering method called ASCR, which unifies anchor point learning, adaptive completion and reconstruction of the anchor point-sample similarity graph, and multi-view latent embedding learning, thereby achieving efficient clustering of missing samples.
Fast Think-on-Graph: Wider, Deeper and Faster Reasoning of Large Language Model on Knowledge Graph
Xujian Liang (Beijing University Of Posts And Telecommunications), Zhaoquan Gu (Harbin Institute of Technology)
CodeTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: Proposes the FastThink-on-Graph (FastToG) framework, allowing large language models to perform multi-step reasoning on knowledge graphs at the community level.
π― What it does: A real-time object detection framework FBRT-YOLO specifically designed for aerial images is proposed, incorporating lightweight modules Feature Complementary Mapping Module (FCM) and Multi-Kernel Perception Unit (MKP) to achieve efficient small object detection.
π― What it does: A Federated Collaborative Online Monitoring framework (FCOM) is proposed, which utilizes representation learning to mine the latent structure of heterogeneous groups and implements resource allocation using an event-triggered UCB algorithm in a distributed environment.
π― What it does: Proposes the FedAA method, which utilizes reinforcement learning to dynamically adjust client aggregation weights to enhance the robustness and fairness of federated learning.
Xianjie Guo (Hefei University of Technology), Xiaoxiao Li (University of British Columbia)
CodeFederated LearningAuto EncoderTabular
π― What it does: A framework for learning causal invariant features in federated learning, called FedCIFL, is proposed, which addresses data heterogeneity (Non-IID) and out-of-distribution (OOD) issues, enabling feature selection without sharing raw data.
CodeAnomaly DetectionFederated LearningTransformerMixture of ExpertsTime Series
π― What it does: Train a temporal foundation model using the federated learning framework FFTS to avoid centralized data fusion, thereby enhancing the generalization ability of cross-domain tasks.
π― What it does: A novel Federated Unlearning framework, FedOSD, is proposed to efficiently remove specified client data from the global model without retraining, while maintaining model performance.
π― What it does: Addressed the problem of federated unsupervised domain generalization, proposing the FedGaLA method that achieves domain-invariant feature learning through local and global gradient alignment.
π― What it does: A federated weakly supervised video anomaly detection framework based on CLIP is proposed, utilizing a context-driven prompt generator for global and local contexts to achieve anomaly frame localization and detection.
FedPIA β Permuting and Integrating Adapters Leveraging Wasserstein Barycenters for Finetuning Foundation Models in Multi-Modal Federated Learning
Pramit Saha (University of Oxford), J. Alison Noble (University of Oxford)
CodeFederated LearningSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
π― What it does: This paper proposes the FedPIA framework for parameter-efficient fine-tuning of large-scale vision-language models in multimodal federated learning.
FedSum: Data-Efficient Federated Learning Under Data Scarcity Scenario for Text Summarization
Zhiyong Ma (South China University of Technology), Jian Chen (Washington State University)
CodeFederated LearningTransformerSupervised Fine-TuningTextBiomedical Data
π― What it does: In the task of text summarization with data scarcity under federated learning, the FedSum framework is proposed, which combines depth (sample level) and breadth (knowledge level) expansion to enhance model performance.
FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts
Yichen Gong (Tsinghua University), Xiaoyun Wang (Tsinghua University)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: This paper studies a black-box jailbreak method called FigStep that converts text instructions into formatted images, successfully bypassing the safety alignment of multimodal language models.
Filling Memory Gaps: Enhancing Continual Semantic Parsing via SQL Syntax Variance-Guided LLMs Without Real Data Replay
Ruiheng Liu (Harbin Institute of Technology), Bailong Yang (Xi'an Research Institute of High-Tech)
CodeKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This work proposes a continuous semantic parsing framework LECSP that does not rely on historical data replay or ideal scene constraints.
π― What it does: A new anomaly detection method called FiCo is proposed, which addresses the performance degradation caused by distribution shifts through two steps: distribution-specific compensation and distribution-invariant filtering.
FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting
Yulong Wang (Nankai University), Kai Wang (Nankai University)
CodeTime Series
π― What it does: The FilterTS model is designed to extract stable and varying frequency components from multivariate time series using frequency domain filtering techniques, improving prediction accuracy.
FIND: A Framework for Discovering Formulas in Data
Tingxiong Xiao (Tsinghua University), Jinli Suo (Tsinghua University)
CodeTabularPhysics Related
π― What it does: Proposed the FIND framework, which utilizes hidden layers to generate dimensionless latent variables and directly discovers physical formulas from observational data through polynomial expression layers;
Fine-grained Adaptive Visual Prompt for Generative Medical Visual Question Answering
Ting Yu (Hangzhou Normal University), Ke Zhang (Hangzhou Dianzi University)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
π― What it does: Proposes the FAVP framework, utilizing fine-grained adaptive visual prompts (AVPC) and a hierarchical answer generator (HAG) to enhance the precise localization and answer quality of generative medical visual question answering.
Fine-Grained Graph Representation Learning for Heterogeneous Mobile Networks with Attentive Fusion and Contrastive Learning
Shengheng Liu (National Mobile Communications Research Laboratory), Yongming Huang (National Mobile Communications Research Laboratory)
CodeAutonomous DrivingRepresentation LearningGraph Neural NetworkContrastive LearningGraphTime Series
π― What it does: An unsupervised data and model-driven graph structure learning framework DMGSL is proposed for the automated construction and fine-grained optimization of the network topology of Wireless Data Knowledge Graphs (WDKG), and its effectiveness is validated in node classification tasks.
Fine-Grained Perception in Panoramic Scenes: A Novel Task, Dataset, and Method for Object Importance Ranking
Jia Song (China University of Petroleum), Shanchen Pang (China University of Petroleum)
CodeObject DetectionSegmentationImage
π― What it does: The FOIR-360 task is proposed, which predicts the fine-grained importance ranking of all instances in a 360-degree image, and the first 360Rank dataset is constructed.
π― What it does: A first-layer network called the Anti-Interference Noise Filter (ANF) is proposed, which achieves inherent robustness against adversarial attacks through large convolutional kernels, more filters, and max pooling.
Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Models
Weihao Ye (Xiamen University), Yiyi Zhou (Xiamen University)
CodeOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
π― What it does: Proposes FitPrune, a training-free visual token pruning method that utilizes attention distribution to quickly generate pruning schemes;
CodeData SynthesisAutonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality
π― What it does: Designed and trained a city visual-language navigation agent FLAME based on a multimodal LLM (Flamingo), employing a three-stage fine-tuning process and automatic data synthesis to achieve efficient street scene understanding and path planning.
π― What it does: An efficient, fast, and general diffusion model distillation method named Flash Diffusion is proposed for generating high-quality images with only two steps (or fewer) of sampling.
π― What it does: The FlexDataset framework is proposed, achieving composition-to-data (C2D) generation based on scene composition, capable of synthesizing multiple instances and multiple categories (MIMC) scenes at once and automatically generating pixel-level annotations for tasks such as salient object detection, depth estimation, and semantic/instance segmentation.
π― What it does: A framework based on Hierarchical Flexible Sharpness-Aware Minimization (FedFSA) is designed in personalized federated learning, applying larger perturbations to the layers with the highest sharpness during local training to enhance the model's generalization performance.
FlexiTex: Enhancing Texture Generation via Visual Guidance
Dadong Jiang (Tianjin University), Zhihui Ke (Tianjin University)
CodeGenerationData SynthesisDiffusion modelMesh
π― What it does: We propose FlexiTex, a multi-view texture generation framework that supports text and image conditions, achieving high-quality, globally consistent textures through visual guidance enhancement and direction-aware adaptation.
π― What it does: This paper proposes FlowPolicy, a framework for generating single-step real-time robot policies using 3D point clouds as conditions, achieved through consistency flow matching.
π― What it does: A dual-stage dynamic fusion network FMPM-DNet is proposed for the fusion of high-resolution PAN images and low-resolution hyperspectral images;
FNIN: A Fourier Neural Operator-based Numerical Integration Network for Surface-from-gradients
Jiaqi Leng (Ocean University of China), Hao Fan (Ocean University of China)
CodeRestorationOptimizationImage
π― What it does: A two-stage numerical integration network (FNIN) based on the Fourier Neural Operator is proposed for recovering three-dimensional surfaces from gradient fields.
Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition
Changwei Wang (Qilu University of Technology), Shibiao Xu (Beijing University of Posts and Telecommunications)
CodeRecognitionRetrievalTransformerImage
π― What it does: This paper proposes a two-stage visual place recognition framework called FoL, which explicitly models reliable discrimination of local regions and combines pseudo-correspondence weakly supervised training to achieve improvements in the overall retrieval and re-ranking performance.
π― What it does: A unified foreground segmentation framework called FOCUS is proposed, which can simultaneously handle various foreground segmentation tasks (SOD, COD, SD, DBD, FD) and generate boundary-aware masks.
Maximilian SchΓ€ffeler, Mohammad Abdulaziz (King's College London)
CodeOptimizationReinforcement Learning
π― What it does: This paper achieves the formal verification of Approximate Policy Iteration (API) for factored Markov Decision Processes (MDP) through interactive theorem proving, and provides an executable, proven correct implementation.
π― What it does: A continuous segmentation framework FR 2 Seg based on Fourier transform is designed, utilizing low-frequency information to achieve domain-specific and domain-invariant knowledge learning;
π― What it does: This paper proposes FDS-GS, which achieves frequency-aware density control in 3D Gaussian Splatting by reparameterizing scale and associating it with density.
Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning
En Fu (University of Science and Technology Beijing), Yanyan Hu (University of Science and Technology Beijing)
CodeClassificationRepresentation LearningContrastive LearningTime Series
π― What it does: A novel non-contrastive learning framework called Frequency-masked Embedding Inference (FEI) is proposed, which performs embedding inference through frequency masking prompts, eliminating the need for positive and negative sample pairs.
π― What it does: This paper studies multimodal multi-party conversation (MMC) understanding, constructs the Friends-MMC dataset, and proposes two fundamental tasks: speaker identification and response prediction.
π― What it does: This study proposes the StoryMind framework, which utilizes multi-agent LLMs to automatically generate and review a large-scale deep video understanding dataset called FriendsQA, and evaluates 10 top VideoQA models on it.
Avyukta Manjunatha Vummintala (International Institute of Information Technology), Sujit Gujar (International Institute of Information Technology)
CodeClassificationOptimizationTabular
π― What it does: A post-processing method called FROC is proposed to convert the scoring function of a trained binary classifier into a probabilistic classifier that satisfies the 'Ξ΅-Equalized ROC' fairness constraint without retraining the model, while ensuring minimal AUC loss.
From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach
Xilin Wang (Beihang University), Zihan Zhou (Manycore Tech Inc.)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: This paper proposes CAD2PROGRAM, a method for recovering 3D parametric models from 2D CAD drawings based on a vision-language model.
π― What it does: This paper proposes two methods, DEFT and ADA-DEFT, which induce activation sparsity in the MLP of Transformers through a differentiable sparse loss within the framework of parameter-efficient fine-tuning (PEFT), thereby improving inference efficiency while maintaining downstream task performance.
π― What it does: A testing-time augmentation-based FSL Rectifier framework is proposed, which corrects the outlier effects of test samples by generating synthetic samples.
FSTA-SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks
Kairong Yu (Zhejiang University), Qi Xu (Dalian University of Technology)
CodeSpiking Neural NetworkImage
π― What it does: This paper explores the learning preferences of intermediate output spikes in SNNs through frequency domain analysis and proposes a Frequency-based Space-Time Attention (FSTA) module that effectively suppresses redundant spikes and enhances feature learning.
CodeExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkImage
π― What it does: This study investigates a method that combines functional connectivity groups of neural networks with those of humans, characterizing the functional topology of large neural networks through persistent graph homology and Wasserstein statistics, and validating its interpretability using unsupervised clustering.
Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free Backdoor Mitigation
Weilin Lin (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
CodeOptimizationAdversarial AttackImage
π― What it does: This paper proposes a completely data-independent post-training anti-poisoning method called OTBR. It first prunes the model using randomly generated untrained neuron weight changes (NWC), and then aligns and merges the pruned model with the original model through optimal transport (OT) to restore performance and suppress backdoors.
G2LDetect: A Global-to-Local Approach for Hallucination Detection
Xiaoxia Cheng (Zhejiang University), Weiming Lu (Zhejiang University)
CodeAnomaly DetectionTransformerLarge Language ModelTextBenchmark
π― What it does: A global-to-local hallucination detection framework G2LDetect is proposed, which first constructs a hierarchical tree structure for global representation of text, and then extracts local details for detection and aggregates results along paths from the tree.
Jiarong Yang (South China University of Technology), Yuan Liu (South China University of Technology)
CodeFederated LearningImage
π― What it does: An asynchronous split federated learning framework GAS is proposed, which utilizes activation buffers, model buffers, and generative activation techniques to mitigate model update bias caused by slow devices and improve convergence speed.
Gaussian Graphical Modelling Without Independence Assumptions for Uncentered Data
Bailey Andrew (University of Leeds), Luisa Cutillo (University of Leeds)
CodeOptimizationGraph Neural NetworkVideoGraphBiomedical Data
π― What it does: This paper proposes the incorporation of a mean estimation wrapper into the multi-axis Gaussian graphical model with a Kronecker-sum structure to correct the bias caused by the traditional zero-mean assumption.
GCD-Sampling: A General Cross-scale Decoupled Sampling for Point Cloud
Tao Dai (Shenzhen University), Zexuan Zhu (Tsinghua University)
CodeClassificationSegmentationPoint Cloud
π― What it does: A general cross-scale decoupled point cloud sampling method called GCD-sampling is proposed, which can adaptively sample point clouds without changing the target network structure, and achieve fine control over the sampling point positions through cross-scale feature fusion and convex combination learning.
GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation
Xu Wang (University of Science and Technology of China), Zihan Lin (University of Science and Technology of China)
CodeObject DetectionKnowledge DistillationVision Language ModelImage
π― What it does: The GCD method is proposed for incremental object detection tasks, aiming to prevent catastrophic forgetting in visual-language detectors through global semantic alignment and semantic correspondence distillation.
π― What it does: A backward synthesis prediction framework GDiffRetro is proposed, which is based on dual graph enhanced molecular representation and 3D conditional diffusion. It first identifies the reaction center to generate synthetic fragments, and then uses a diffusion model to generate complete reactants from the synthetic fragments.
π― What it does: A general contrastive clustering framework GeCC is proposed, which utilizes a clustering-guided domain transformation modeling module and attention weights to enhance unsupervised clustering performance.
Rui Lv (University of Science and Technology of China), Linbo Zhu (University of Science and Technology of China)
CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
π― What it does: A generative intelligent agent GenAL based on large language models has been designed for adaptive learning path recommendation, which includes a global thinking agent and a local teaching agent, capable of dynamically recommending exercises based on learners' historical records and text semantics.
Generalized Convergence Analysis of Tsetlin Automaton Based Algorithms: A Probabilistic Approach to Concept Learning
Mohamed-Bachir Belaid (NILU Climate and Environmental Research Institute), Anis Yazidi (Oslo Metropolitan University)
CodeClassificationOptimizationTabular
π― What it does: Introduces the Probabilistic Concept Learning (PCL) framework and proves its almost certain convergence to any literal conjunction under noise-free samples, while validating its effectiveness through experiments.
Generalized Implicit Neural Representations for Dynamic Molecular Surface Modeling
Fang Wu (Stanford University), Stan Z. Li (Westlake University)
CodeDrug DiscoveryProtein Structure PredictionMixture of ExpertsPoint Cloud
π― What it does: This paper studies dynamic protein surface modeling and proposes a dynamic surface representation framework MoE-DSR based on Mixture-of-Experts.
Generalizing Constraint Models in Constraint Acquisition
Dimos Tsouros (KU Leuven), Tias Guns (University College Cork)
CodeClassificationOptimizationExplainability and InterpretabilityTabularBenchmark
π― What it does: This paper proposes a constraint acquisition method called GENCON, which utilizes constraint-level classification to learn parameterized constraint models, thereby achieving model generalization across instances.
Generating Counterfactual Explanations Under Temporal Constraints
Andrei Buliga (Fondazione Bruno Kessler), Massimiliano Ronzani (Fondazione Bruno Kessler)
CodeOptimizationExplainability and InterpretabilityTime SeriesSequential
π― What it does: This paper proposes a counterfactual explanation method for generating process trajectories under time constraints, which ensures that the generated trajectories always satisfy the business process background knowledge expressed by Linear Temporal Logic (LTLp).
π― What it does: This paper proposes Generative Medical Segmentation (GMS), which utilizes a pre-trained vision foundation model (Stable Diffusion VAE) to extract latent representations of images and masks, trains a lightweight latent mapping network to achieve mapping from image latent to mask latent, and then decodes to obtain segmentation results.
GenPlan: Generative Sequence Models as Adaptive Planners
Akash Karthikeyan (University of Waterloo), Yash Vardhan Pant (University of Waterloo)
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerFlow-based ModelSequential
π― What it does: A generative planning method called GenPlan based on a discrete flow model is proposed, utilizing an energy-guided denoising process to achieve multi-task adaptive planning.
π― What it does: This paper proposes GeoBEV, a 3D object detection framework based on multi-view images, achieving finer geometric recovery through high-resolution BEV representation.
π― What it does: Aiming at semi-supervised learning for mixed-type tabular data, the GFTab framework is proposed to learn robust representations and perform classification from a limited number of labeled samples and a large number of unlabeled samples.
π― What it does: This paper proposes a post-processing algorithm called GHOST, which models each class feature using a multivariate Gaussian (diagonal covariance) and performs Z-score normalization on logits to distinguish between known and unknown samples, thereby enhancing open-set recognition and OOD detection.
π― What it does: The GLAM (Global-Local variation Awareness Mamba-based world model) framework is proposed, utilizing two parallel Mamba modules, GMamba and LMamba, to perceive state changes at both global and local levels, thereby achieving high-quality world model inference and sample-efficient model-based reinforcement learning.