AAAI 2023 Papers — Page 15
AAAI Conference on Artificial Intelligence · 1578 papers
T2-GNN: Graph Neural Networks for Graphs with Incomplete Features and Structure via Teacher-Student Distillation
Cuiying Huo (Tianjin University), Lingfei Wu (Nanjing University)
Knowledge DistillationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes the T2-GNN framework based on teacher-student distillation for handling graph neural networks with missing node features and graph structures.
T2G-FORMER: Organizing Tabular Features into Relation Graphs Promotes Heterogeneous Feature Interaction
Jiahuan Yan (Zhejiang University), Jian Wu (Zhejiang University)
ClassificationOptimizationGraph Neural NetworkTransformerTabular
🎯 What it does: A feature relationship graph (FR-Graph) based on a Graph Estimator is proposed and embedded into the Transformer structure, forming T2G-FORMER, to enhance the heterogeneous feature interaction and prediction performance of tabular data.
Tackling Data Heterogeneity in Federated Learning with Class Prototypes
Yutong Dai (Lehigh University), Ran Xu (Salesforce Research)
ClassificationFederated LearningImage
🎯 What it does: This paper proposes FedNH, an algorithm that simultaneously addresses data heterogeneity and class imbalance in federated learning, enhancing the generalization and personalization performance of local and global models by utilizing the uniformity of class prototypes and semantic information.
TaCo: Textual Attribute Recognition via Contrastive Learning
Chang Nie (Tencent), Bo Ren (Tencent)
RecognitionConvolutional Neural NetworkContrastive LearningText
🎯 What it does: This paper proposes TaCo, a contrastive learning framework for text attribute recognition.
Tagging before Alignment: Integrating Multi-Modal Tags for Video-Text Retrieval
Yizhen Chen (Tencent), Ying Shan (Tencent)
RetrievalTransformerMixture of ExpertsVision Language ModelVideoTextMultimodality
🎯 What it does: The TABLE model is proposed, which uses multimodal labels (such as objects, people, scenes, actions, audio) as anchors to explicitly fuse multimodal information from videos for more accurate video-text retrieval.
Take Your Model Further: A General Post-refinement Network for Light Field Disparity Estimation via BadPix Correction
Rongshan Chen (Beihang University), Ruixuan Cong (Beihang University)
RestorationDepth EstimationImage
🎯 What it does: Proposes the BadPix correction concept and implements a general post-processing network BpCNet to correct erroneous pixels in light field disparity maps.
Taming Continuous Posteriors for Latent Variational Dialogue Policies
Marin Vlastelica (Max Planck Institute for Intelligent Systems), Gyuri Szarvas (Amazon Development Center Germany GmbH)
Reinforcement LearningText
🎯 What it does: A new continuous posterior variational inference method TCUP is proposed for implicit action reinforcement learning in task-oriented dialogue.
Target-Aware Tracking with Long-Term Context Attention
Kaijie He (Guangxi Normal University), Zhiwen Wang (Guangxi University of Science and Technology)
Object TrackingTransformerVideo
🎯 What it does: A Long Context Attention (LCA) module based on the fusion of multi-frame target and background information is proposed, which is embedded in a Transformer to construct the TATrack target-aware tracker, along with a dynamic template update strategy based on classification confidence.
Target-Free Text-Guided Image Manipulation
Wan-Cyuan Fan (National Taiwan University), Yu-Chiang Frank Wang (NVIDIA)
Image TranslationGenerationVision Language ModelGenerative Adversarial NetworkImageText
🎯 What it does: Addressing the problem of text-guided image editing without a target image, specifically modifying a reference image based on given textual instructions without knowledge of the target image.
Task-Specific Scene Structure Representations
Jisu Shin (Gwangju Institute of Science and Technology), Hae-Gon Jeon (Gwangju Institute of Science and Technology)
Super ResolutionConvolutional Neural NetworkImage
🎯 What it does: A lightweight neural network SSGNet is proposed, which can unsupervisedly generate task-specific scene structure-guided features and can be integrated as a plug-and-play module into low-level vision tasks.
TC-DWA:Text Clustering with Dual Word-Level Augmentation
Bo Cheng (Jilin University), Yi Chang (Jilin University)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Using BERT for self-training and introducing dual word-level enhancement (anchor words and expected enhancement) to achieve unsupervised text clustering.
Teaching to Learn: Sequential Teaching of Learners with Internal States
Mustafa Mert Çelikok (Aalto University), Samuel Kaski (University of Manchester)
Meta LearningReinforcement LearningTabularSequential
🎯 What it does: The 'Teaching to Learn (TtL)' framework is proposed, which extends machine teaching to allow the internal state of learners (i.e., prior preferences) to evolve throughout the teaching process, thereby enhancing learning outcomes for future tasks.
Temporal Knowledge Graph Reasoning with Historical Contrastive Learning
Yi Xu (Shanghai Jiao Tong University), Luoyi Fu (Shanghai Jiao Tong University)
Graph Neural NetworkContrastive LearningGraphTime Series
🎯 What it does: This paper proposes a new model for event prediction in temporal knowledge graphs, CENET, which can simultaneously utilize the dependencies of historical and non-historical events, and identify the most relevant entities through contrastive learning of queries, ultimately achieving precise inference through a masking strategy.
Temporal-Frequency Co-training for Time Series Semi-supervised Learning
Zhen Liu (South China University of Technology), Linghao Wang (South China University of Technology)
ClassificationAnomaly DetectionContrastive LearningTime Series
🎯 What it does: This paper proposes a temporal semi-supervised learning framework TS-TFC based on dual views in the time domain and frequency domain, which enhances the utilization of unlabeled data through pseudo-label propagation and collaborative training.
Tensor Compressive Sensing Fused Low-Rankness and Local-Smoothness
Xinling Liu (Southwest University), Jianjun Wang (Southwest University)
RestorationOptimizationImageVideo
🎯 What it does: A novel regularization method based on tensor compressed sensing is proposed—Tensor Correlated Total Variation (TCTV), which utilizes low-rankness and local smoothness to recover high-dimensional tensor data.
Tensorized Incomplete Multi-View Clustering with Intrinsic Graph Completion
Shuping Zhao (Shenzhen Key Laboratory of Visual Object Detection and Recognition, Harbin Institute of Technology), Bob Zhang (University of Macau)
Representation LearningMultimodality
🎯 What it does: A missing multi-view clustering method TIMVC IGC based on low-rank tensors and cross-view consistency constraints is proposed, which can simultaneously complete missing view inference, construct a complete graph, and learn co-representations.
Text to Point Cloud Localization with Relation-Enhanced Transformer
Guangzhi Wang (National University of Singapore), Mohan Kankanhalli (National University of Singapore)
Pose EstimationRetrievalAutonomous DrivingTransformerTextPoint Cloud
🎯 What it does: This study investigates the problem of locating target positions in urban-scale point clouds based on natural language instructions, proposing a two-stage framework that first retrieves matching cells and then refines the instance matching and regresses the position.
Text-DIAE: A Self-Supervised Degradation Invariant Autoencoder for Text Recognition and Document Enhancement
Mohamed Ali Souibgui (Universitat Autònoma de Barcelona), Dimosthenis Karatzas (Digital Research Center of Sfax)
RecognitionRestorationTransformerAuto EncoderContrastive LearningImageText
🎯 What it does: This paper proposes a self-supervised degradation-invariant autoencoder (Text-DIAE) that is pre-trained by applying three degradation tasks—masking, blurring, and noise—to text images, and then fine-tuned for handwritten/scenario text recognition and document image enhancement.
The Devil Is in the Frequency: Geminated Gestalt Autoencoder for Self-Supervised Visual Pre-training
Hao Liu (Tencent), Bo Ren (Tencent)
Object DetectionSegmentationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Proposes the Geminated Gestalt Autoencoder (Ge2-AE), which simultaneously uses a pixel decoder and a frequency domain decoder under the self-supervised Masked Image Modeling (MIM) framework, leveraging the complementary constraints of global semantics in the frequency domain and local details in the pixel domain for visual pre-training.
The Effect of Diversity in Meta-Learning
Ramnath Kumar (Google Research), Yoshua Bengio (Mila)
Meta LearningImage
🎯 What it does: This study investigates the impact of task diversity on model performance in meta-learning and explores the relationship between task diversity and model generalization through experimental and theoretical analysis.
The Effect of Modeling Human Rationality Level on Learning Rewards from Multiple Feedback Types
Gaurav R. Ghosal (University of California), Anca D. Dragan (University of Utah)
Autonomous DrivingReinforcement Learning from Human FeedbackTabular
🎯 What it does: This paper studies how to improve reward function learning through learning the human rationality coefficient β for each type of human feedback (demonstration, comparison, stop instructions), combining simulations and small-scale user experiments;
The Effect of Preferences in Abstract Argumentation under a Claim-Centric View
Michael Bernreiter (Institute of Logic and Computation), Stefan Woltran (Institute of Logic and Computation)
🎯 What it does: This paper studies the impact of preferences in abstract argumentation from a central perspective, exploring the effects of four common preference reductions on the semantics and complexity of claim-augmented argumentation frameworks (CAF).
The Expressive Power of Ad-Hoc Constraints for Modelling CSPs
Ruiwei Wang (National University of Singapore), Roland H.C. Yap (National University of Singapore)
🎯 What it does: This paper systematically studies the expressive power, compactness, executable operations, and queries of 14 commonly used ad-hoc constraints in modeling constraint satisfaction problems (CSP), and provides a complete comparability map.
The Implicit Regularization of Momentum Gradient Descent in Overparametrized Models
Li Wang (Northeast Normal University), Zili Yan (Beihua University)
OptimizationTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper analyzes the implicit regularization behavior of momentum gradient descent (MGD) by transforming it into a continuous-time dynamics (MGF) form in over-parameterized linear models and compares it with ridge regression. It further investigates the phenomenon where the learning rate differences of weight components in deep N-linear networks are amplified with increasing depth, revealing an implicit bias towards sparse solutions.
The Influence of Dimensions on the Complexity of Computing Decision Trees
Stephen G. Kobourov (University of Arizona), Jules Wulms (TU Wien)
🎯 What it does: This paper studies the computational complexity of learning the minimum decision tree as the dimension d of the feature space changes, and presents various algorithms and lower bounds.
The Linear Distance Traveling Tournament Problem Allows an EPTAS
Jingyang Zhao (University of Electronic Science and Technology of China), Mingyu Xiao (University of Electronic Science and Technology of China)
Optimization
🎯 What it does: An effective polynomial-time approximation scheme (EPTAS) is proposed for the linear distance traveling tournament problem (LDTTPk), which can achieve a total travel distance not exceeding (1+ε) times the optimal solution for any given error ε.
The Multi-Agent Transportation Problem
Pascal Bachor (Albert-Ludwigs-Universitat Freiburg), Bernhard Nebel (Albert-Ludwigs-Universitat Freiburg)
OptimizationAgentic AIGraph
🎯 What it does: A multi-agent transportation problem (MAT) is proposed, along with a complete algorithm based on SAT solving;
The Parameterized Complexity of Network Microaggregation
Václav Blažej (Czech Technical University in Prague), Kirill Simonov (Hasso Plattner Institute)
🎯 What it does: This paper conducts a systematic parameterized complexity analysis of the Network Microaggregation problem and its connected variants, proposing various fixed-parameter algorithms based on tree width and vertex cover number, along with corresponding lower bounds.
The Perils of Trial-and-Error Reward Design: Misdesign through Overfitting and Invalid Task Specifications
Serena Booth (Bosch), Alessandro Allievi (Bosch)
Reinforcement Learning
🎯 What it does: This study investigates the issues of overfitting of the reward function and ineffective task specification in reinforcement learning, validated through large-scale experiments and expert user studies.
The Role of Heuristics and Biases during Complex Choices with an AI Teammate
Nikolos Gurney (Institute for Creative Technologies), David V. Pynadath (Institute for Creative Technologies)
Tabular
🎯 What it does: This study investigates the anchoring and framing effects in complex decision-making tasks when humans collaborate with AI assistants during a search process. It utilizes a 'rugged terrain' search experiment based on a rolling disk to compare behaviors and outcomes between independent decision-making and collaboration with AI.
The Sufficiency of Off-Policyness and Soft Clipping: PPO Is Still Insufficient according to an Off-Policy Measure
Xing Chen (Jilin University), Yi Chang (Jilin University)
Reinforcement LearningSequential
🎯 What it does: An improved PPO algorithm called P3O is proposed, which utilizes the importance sampling ratio with sigmoid preprocessing to achieve broader policy space exploration.
The Unreasonable Effectiveness of Deep Evidential Regression
Nis Meinert (Pasteur Labs), Alexander Lavin (Pasteur Labs)
Depth EstimationImage
🎯 What it does: This paper analyzes and evaluates the effectiveness of Deep Evidential Regression (DER) in uncertainty estimation, revealing its theoretical flaws and providing suggestions for improvement.
The Value of AI Guidance in Human Examination of Synthetically-Generated Faces
Aidan Boyd (University of Notre Dame), Adam Czajka (University of Notre Dame)
RecognitionGenerationExplainability and InterpretabilityConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper evaluates the effectiveness of AI-assisted human recognition of synthetic faces through large-scale experiments and compares the impact of different AI prompting methods and model training approaches on non-expert human detection of synthetic faces.
Tight Inapproximability for Graphical Games
Argyrios Deligkas (Royal Holloway), Themistoklis Melissourgos (University of Essex)
Graph
🎯 What it does: The computational complexity of approximate Nash equilibria in two-action graphical games is studied.
Tight Performance Guarantees of Imitator Policies with Continuous Actions
Davide Maran (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
Reinforcement Learning
🎯 What it does: This paper addresses the problem of behavior cloning (BC) in continuous action spaces, proposing a performance upper bound based on Wasserstein distance. It proves that under Lipschitz MDPs and policies, the value function is always Hölder continuous, and further obtains stronger theoretical guarantees through noise injection techniques, while also conducting empirical validation in OpenAI Gym continuous control environments.
Tighter Robust Upper Bounds for Options via No-Regret Learning
Shan Xue (Southwestern University of Finance and Economics), Liang Xu (Southwestern University of Finance and Economics)
OptimizationReinforcement LearningFinance Related
🎯 What it does: A new robust upper price for European options and average price Asian options is constructed through the η-momentum learning strategy.
Time Series Contrastive Learning with Information-Aware Augmentations
Dongsheng Luo (Florida International University), Xiang Zhang (Pennsylvania State University)
Representation LearningMeta LearningContrastive LearningTime Series
🎯 What it does: This paper studies a contrastive learning method for time series called InfoTS, which achieves better representation learning through information theory-guided adaptive data augmentation.
Time-Aware Random Walk Diffusion to Improve Dynamic Graph Learning
Jong-whi Lee (Jeonbuk National University), Jinhong Jung (Jeonbuk National University)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A time-aware random walk diffusion method named TIARA is proposed to enhance the spatial and temporal locality of dynamic graphs, thereby improving the learning effectiveness of dynamic graph neural networks.
TinyNeRF: Towards 100 x Compression of Voxel Radiance Fields
Tianli Zhao (University of Chinese Academy of Sciences), Jian Cheng (Institute of Automation Chinese Academy of Sciences)
CompressionNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes TinyNeRF, which achieves a high compression rate for NeRF voxel models through three steps: frequency domain transformation, pruning, and quantization.
Token Mixing: Parameter-Efficient Transfer Learning from Image-Language to Video-Language
Yuqi Liu (Renmin University of China), Qin Jin (Renmin University of China)
GenerationRetrievalTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: A Token Mixing strategy is proposed to achieve parameter-efficient transfer of image-language models to video-language tasks without adding extra modules.
TopicFM: Robust and Interpretable Topic-Assisted Feature Matching
Khang Truong Giang (Korea Advanced Institute of Science and Technology), Sungho Jo (Korea Advanced Institute of Science and Technology)
RecognitionImage TranslationRetrievalExplainability and InterpretabilityTransformerImage
🎯 What it does: Proposes a TopicFM method based on topic modeling, utilizing Transformer to infer high-level semantic topics in images and enhance features for robust image matching.
Topological Distance Games
Martin Bullinger (Technical University of Munich), Warut Suksompong (National University of Singapore)
🎯 What it does: A novel vertex allocation game called Topological Distance Games (TDGs) is proposed, which studies the allocation of agents on a topological graph where agents' intrinsic preferences for others and their distances jointly determine utility. The existence, complexity, and dynamic properties of jump-stable solutions are analyzed.
Topological Pooling on Graphs
Yuzhou Chen (Temple University), Yulia R. Gel (University of Texas at Dallas)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a differentiable graph pooling layer called Wit-TopoPool, which can simultaneously capture both local and global topological features of graphs to enhance graph classification performance.
TOT：Topology-Aware Optimal Transport for Multimodal Hate Detection
Linhao Zhang (Aerospace Information Research Institute, Chinese Academy of Sciences), Shiyao Yan (Aerospace Information Research Institute, Chinese Academy of Sciences)
ClassificationRecognitionOptimizationGraph Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a topology-aware optimal transport framework (TOT) to identify implicitly aligned hate memes, addressing the semantic gap and alignment challenges between visual and textual modalities.
Tournament Fixing Parameterized by Feedback Vertex Set Number Is FPT
Meirav Zehavi (Ben-Gurion University of the Negev)
OptimizationGraph
🎯 What it does: This paper studies the problem of 'adjusting the competition system' in single-elimination tournaments and presents an FPT algorithm parameterized by the feedback vertex set size k, with a running time of 2^{O(k log k)}·n^{O(1)}.
Toward a Perspectivist Turn in Ground Truthing for Predictive Computing
Federico Cabitza (University of Milano-Bicocca), Valerio Basile (University of Turin)
ImageBiomedical DataReview/Survey Paper
🎯 What it does: This paper proposes and argues for the adoption of a 'Perspectivism' paradigm in the process of machine learning data labeling and Ground Truthing, which retains and utilizes the differences and diversity among labelers, rather than unifying them into a single 'gold standard';
Toward Robust Diagnosis: A Contour Attention Preserving Adversarial Defense for COVID-19 Detection
Kun Xiang (Sun Yat-sen University), Shancheng Jiang (Sun Yat-sen University)
ClassificationDomain AdaptationAdversarial AttackTransformerContrastive LearningImageComputed Tomography
🎯 What it does: This paper proposes a novel adversarial defense method called CAP, based on maintaining attention to lung contours, aimed at enhancing the robustness of COVID-19 CT image classification.
Towards a Holistic Understanding of Mathematical Questions with Contrastive Pre-training
Yuting Ning (University of Science and Technology of China), Shijin Wang (iFLYTEK Co., Ltd.)
TransformerContrastive LearningText
🎯 What it does: QuesCo is proposed, a comparative pre-training model for mathematical problems that enhances the generation of surface-diverse but semantically similar samples through a dual-layer content and structure approach, and learns overall semantic representations through knowledge hierarchy-based ranking and ranking contrastive loss.
Towards Automated Modeling Assistance: An Efficient Approach for Repairing Flawed Planning Domains
Songtuan Lin (Australian National University), Pascal Bercher (Australian National University)
OptimizationComputational EfficiencyTabularBenchmark
🎯 What it does: An automatic domain repair method based on conflict detection and minimal hitting set is proposed, which can quickly locate and minimally repair the planning domain under the premise of a given erroneous plan, making the plan executable.
Towards Better Visualizing the Decision Basis of Networks via Unfold and Conquer Attribution Guidance
Jung-Ho Hong (Korea University), Seong-Whan Lee (Korea University)
Explainability and InterpretabilityImage
🎯 What it does: A post-hoc framework called UCAG is proposed, which achieves finer-grained and more interpretable visualization results by spatially splitting the input image and aggregating local attributions.
Towards Complex Scenarios: Building End-to-End Task-Oriented Dialogue System across Multiple Knowledge Bases
Libo Qin (Central South University), Wanxiang Che (Harbin Institute of Technology)
Recurrent Neural NetworkGraph Neural NetworkText
🎯 What it does: This paper proposes a knowledge graph attention network named KoK-HAN to address the end-to-end task-oriented dialogue system problem in the context of multiple knowledge bases.
Towards Credible Human Evaluation of Open-Domain Dialog Systems Using Interactive Setup
Sijia Liu (Amazon), Dilek Hakkani-Tur (Amazon)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper conducts round-by-round evaluation of open-domain dialogue models through an interactive evaluation framework and extends it to multi-model comparisons, providing three new evaluation settings: SOBA, SATA-Random, and SATA-User, systematically assessing their sensitivity and consistency.
Towards Decision-Friendly AUC: Learning Multi-Classifier with AUCµ
Peifeng Gao (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
ClassificationImage
🎯 What it does: A multi-class classifier training framework based on AUC µ is proposed, divided into two stages: first, maximize AUC µ through a custom differentiable surrogate loss, and then learn thresholds to improve F1.
Towards Diverse, Relevant and Coherent Open-Domain Dialogue Generation via Hybrid Latent Variables
Bin Sun (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
GenerationTransformerSupervised Fine-TuningText
🎯 What it does: A hybrid latent variable (Hybrid Latent Variable, HLV) and conditional hybrid variational Transformer (CHVT) are proposed for open-domain dialogue generation, balancing diversity, relevance, and coherence.
Towards Efficient and Domain-Agnostic Evasion Attack with High-Dimensional Categorical Inputs
Hongyan Bao (King Abdullah University of Science and Technology), Xiangliang Zhang (University of Notre Dame)
Computational EfficiencyAdversarial AttackTextTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a domain-agnostic attack method called FEAT for high-dimensional categorical inputs, aimed at efficiently finding feasible adversarial perturbations.
Towards Fine-Grained Explainability for Heterogeneous Graph Neural Network
Tong Li (Shanghai Jiao Tong University), Caleb Chen Cao (Huawei Research Hong Kong)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper proposes the xPath framework to provide fine-grained interpretability for the node classification task of black-box heterogeneous graph neural networks (HGN), specifically by giving explanations of causal nodes and their influence paths.
Towards Global Video Scene Segmentation with Context-Aware Transformer
Yang Yang (Nanjing University of Science and Technology), Dingyin Xia (HUAWEI CBG Edu AI Lab)
SegmentationTransformerVideo
🎯 What it does: The Context-Aware Transformer (CAT) model is proposed, which learns high-quality shot representations through self-supervised pre-training tasks, ultimately used for video scene segmentation.
Towards Good Practices for Missing Modality Robust Action Recognition
Sangmin Woo (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
RecognitionTransformerAuto EncoderVideoMultimodality
🎯 What it does: This paper proposes a training and inference scheme for multimodal action recognition in the case of missing modalities, which mainly includes data augmentation, selection of fusion methods, and an autoencoder (ActionMAE) to achieve reconstruction and prediction of the missing modalities.
Towards In-Distribution Compatible Out-of-Distribution Detection
Boxi Wu (Zhejiang University), Wei Liu (Tencent Data Platform)
Anomaly DetectionImage
🎯 What it does: This paper proposes an Outlier Exposure method (ICE) that maintains compatibility with the original distribution while improving OOD detection and IN classification accuracy.
Towards Inference Efficient Deep Ensemble Learning
Ziyue Li (Microsoft Research), Dongsheng Li (Microsoft Research)
ClassificationComputational EfficiencyRecurrent Neural NetworkMixture of ExpertsImage
🎯 What it does: A learnable sequential multi-model ensemble framework IRENE is designed to reduce inference costs through dynamic stopping decisions.
Towards Interpreting and Utilizing Symmetry Property in Adversarial Examples
Shibin Mei (Shanghai Jiao Tong University), Shengchao Yuan (Shanghai Jiao Tong University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Attack Proportion metric, systematically studies the misclassification distribution of adversarial examples across different categories, and reveals the symmetry property of adversarial examples and the maximum symmetry pair. Based on this symmetry, a symmetry-aware regularization loss is designed, incorporating inter-class and intra-class constraints in adversarial training to enhance model robustness.
Towards More Robust Interpretation via Local Gradient Alignment
Sunghwan Joo (Sungkyunkwan University), Taesup Moon (Seoul National University)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a gradient alignment regularization method that combines ℓ2 and cosine distance to enhance the robustness of deep network feature attribution.
Towards Optimal Randomized Strategies in Adversarial Example Game
Jiahao Xie (Zhejiang University), Hui Qian (Zhejiang University)
OptimizationAdversarial AttackImageStochastic Differential Equation
🎯 What it does: This paper proposes the FRAT algorithm, which can solve the mixed Nash equilibrium of adversarial example games in continuous parameter spaces, thereby achieving fully randomized robust training.
Towards Real-Time Panoptic Narrative Grounding by an End-to-End Grounding Network
Haowei Wang (Xiamen University), Xiaoshuai Sun (Tencent)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: An end-to-end single-stage panoptic narrative grounding network (EPNG) is proposed, which can generate pixel masks corresponding to text descriptions in real-time, eliminating the candidate mask generation process of traditional two-stage methods.
Towards Real-Time Segmentation on the Edge
Yanyu Li (Northeastern University), Yanzhi Wang (Northeastern University)
SegmentationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a real-time semantic segmentation framework for edge devices, combining self-attention modules and lightweight convolutions, and incorporating latency constraints during the search phase.
Towards Reliable Item Sampling for Recommendation Evaluation
Dong Li (Kent State University), Zhi Liu (iLambda)
Recommendation SystemTabular
🎯 What it does: This paper proposes an improved project sampling evaluation method for more accurately estimating the Top-K metrics of recommendation systems.
Towards Reliable Neural Machine Translation with Consistency-Aware Meta-Learning
Rongxiang Weng (Soochow University), Min Zhang (Soochow University)
Meta LearningTransformerText
🎯 What it does: A Consistency-Aware Meta-Learning (CAML) framework is proposed, which utilizes a combination of Transformer and MAML to learn semantically consistent representations in the outer loop and the mapping from these representations to target sentences in the inner loop, thereby enhancing the robustness and reliability of NMT to source diversity.
Towards Robust Metrics for Concept Representation Evaluation
Mateo Espinosa Zarlenga (University of Cambridge), Mateja Jamnik (University of Cambridge)
Representation LearningTabular
🎯 What it does: Two new metrics (Oracle Impurity Score OIS and Niche Impurity Score NIS) are proposed to measure the purity of concept representations in concept learning models.
Towards Voice Reconstruction from EEG during Imagined Speech
Young-Eun Lee (Korea University), Seong-Whan Lee (Korea University)
GenerationDomain AdaptationRecurrent Neural NetworkGenerative Adversarial NetworkTime SeriesAudio
🎯 What it does: This study proposes a generative framework named NeuroTalk, which utilizes non-invasive EEG (imagined speech) to reconstruct the user's own voice.
Tracking and Reconstructing Hand Object Interactions from Point Cloud Sequences in the Wild
Jiayi Chen (Peking University), He Wang (Stanford University)
Object TrackingPose EstimationVideoPoint Cloud
🎯 What it does: Using deep point cloud sequences in a wild environment to track and reconstruct the pose and shape of hands and objects in real-time.
Trafformer: Unify Time and Space in Traffic Prediction
Di Jin (Tianjin University), Yu-Bin Yang (Nanjing University)
TransformerTime Series
🎯 What it does: Designed and proposed the Trafformer model, which predicts multi-step traffic speed in one go using unified spatiotemporal encoding and fully connected self-attention.
Training Meta-Surrogate Model for Transferable Adversarial Attack
Yunxiao Qin (Communication University of China), Cho-Jui Hsieh (University of California)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Meta-Transfer Attack (MTA) framework, which trains a Meta-Surrogate Model (MSM) so that attacks on the MSM can better transfer to the target model;
Training-Time Attacks against K-nearest Neighbors
Ara Vartanian (University of Wisconsin Madison), Scott Alfeld (Amherst College)
ClassificationAdversarial AttackImage
🎯 What it does: The research focuses on poisoning attacks during the training phase of the k-NN classifier, proving that the optimal attack is NP-Hard, and proposes online algorithms CHOPPA and GIT-ACHOPPA2 for constructing single-point and multi-point attacks, providing approximate performance guarantees.
Transfer Learning Enhanced DeepONet for Long-Time Prediction of Evolution Equations
Wuzhe Xu (University of Minnesota), Li Wang (University of Massachusetts Amherst)
Time SeriesPhysics Related
🎯 What it does: Using a physics-informed DeepONet combined with transfer learning, we construct an operator network that evolves over time to predict the long-term solutions of evolution equations.
Transferable Post-hoc Calibration on Pretrained Transformers in Noisy Text Classification
Jun Zhang (Tsinghua University), Ling Feng (National Innovation Institute of Defense Technology, Chinese Academy of Military Science)
ClassificationTransformerTextBenchmark
🎯 What it does: This study investigates the post-calibration problem of pre-trained transformers in text classification under noisy environments, proposing a transferable temperature scaling method based on distribution shift metrics (entropy and MMD);
Transformation-Equivariant 3D Object Detection for Autonomous Driving
Hai Wu (Xiamen University), Cheng Wang (Xiamen University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkPoint Cloud
🎯 What it does: An efficient Transformation-Equivariant 3D object detection framework TED is proposed, which utilizes sparse convolution to extract multi-transform equivariant features, and compresses these features into lightweight representations through TeBEV pooling and TiVoxel pooling, achieving real-time and high-precision object detection.
TransLO: A Window-Based Masked Point Transformer Framework for Large-Scale LiDAR Odometry
Jiuming Liu (Shanghai Jiao Tong University), Hesheng Wang (China University of Mining and Technology)
Pose EstimationAutonomous DrivingTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A window attention-based point cloud Transformer network (TransLO) is proposed, which projects LiDAR point clouds into 2D pseudo-images and uses window mask self-attention and cross-frame attention to achieve end-to-end LiDAR odometry estimation.
TransPath: Learning Heuristics for Grid-Based Pathfinding via Transformers
Daniil Kirilenko (Federal Research Center for Computer Science and Control of Russian Academy of Sciences), Konstantin Yakovlev (Federal Research Center for Computer Science and Control of Russian Academy of Sciences)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: The study predicts correction factors and path probabilities for grid maps by combining CNN and Transformer, and applies them to weighted A* and Focal Search to enhance the efficiency of grid path searching.
TransVCL: Attention-Enhanced Video Copy Localization Network with Flexible Supervision
Sifeng He (Ant Group), Jiandong Zhang (Ant Group)
RecognitionObject DetectionTransformerVideo
🎯 What it does: An end-to-end TransVCL network is proposed, which enhances frame features using Transformers to generate a differentiable similarity matrix, and achieves video segment localization through object detection.
Tree Learning: Optimal Sample Complexity and Algorithms
Dmitrii Avdiukhin (Indiana University), Faraz Mirza (Thomas Jefferson High School for Science and Technology)
Tabular
🎯 What it does: The study learns hierarchical tree representations from labeled tuples (such as triplets) and provides upper bounds on the optimal sample complexity in PAC learning and online learning scenarios; it also proposes a near-linear time tree construction algorithm.
Tree-Structured Trajectory Encoding for Vision-and-Language Navigation
Xinzhe Zhou (Peking University), Yadong Mu (Peking University)
TransformerReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a tree-structured trajectory encoding method that utilizes Tree-Transformer for fine-grained modeling of navigation trajectories, and designs a tree-based action space and Tree-nDTW reward, significantly enhancing the error correction and path recognition capabilities of visual-language navigation (VLN) models.
TrEP: Transformer-Based Evidential Prediction for Pedestrian Intention with Uncertainty
Zhengming Zhang (Purdue University), Zhengming Ding (Tulane University)
Autonomous DrivingTransformerTime Series
🎯 What it does: A Transformer-based evidence learning model, TrEP, has been designed and implemented for predicting pedestrian crossing intentions from the perspective of an autonomous vehicle, along with providing uncertainty estimates.
Tricking the Hashing Trick: A Tight Lower Bound on the Robustness of CountSketch to Adaptive Inputs
Edith Cohen (Google Research), Uri Stemmer (Google Research)
OptimizationAdversarial Attack
🎯 What it does: This paper studies the robustness of CountSketch and feature hashing under adaptive inputs, proposing an attack method that generates a highly biased adversarial input vector after O(ℓ²) queries, revealing the inherent vulnerability of this fundamental method.
TrOCR: Transformer-Based Optical Character Recognition with Pre-trained Models
Minghao Li (Beihang University), Furu Wei (Microsoft Corporation)
RecognitionTransformerImageText
🎯 What it does: We propose TrOCR, an end-to-end Transformer OCR model that directly recognizes text line images using pre-trained image and text Transformers, generating results with wordpiece units.
Truncate-Split-Contrast: A Framework for Learning from Mislabeled Videos
Zixiao Wang (Tsinghua University), Jue Wang (Tencent AI Lab)
ClassificationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: A framework named NEAT is proposed to handle noisy labels in video classification tasks.
Trusted Fine-Grained Image Classification through Hierarchical Evidence Fusion
Zhikang Xu (Shanghai University), Zihao Li (Shanghai University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A trustworthy fine-grained image classification method based on hierarchical evidence fusion is proposed;
Truthful Mechanisms for Steiner Tree Problems
Jinshan Zhang (Zhejiang University), Jianwei Yin (Zhejiang University)
OptimizationGraph
🎯 What it does: A sincere expectation mechanism based on LP iterative relaxation and decomposition techniques is designed to construct Steiner trees in self-interested edge-weighted graphs.
Two Heads Are Better than One: Image-Point Cloud Network for Depth-Based 3D Hand Pose Estimation
Pengfei Ren (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
Pose EstimationDepth EstimationConvolutional Neural NetworkGraph Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes an IPNet network that simultaneously utilizes depth images and point clouds to achieve 3D hand pose estimation.
Two Views of Constrained Differential Privacy: Belief Revision and Update
Likang Liu (Renmin University of China), Yuan Feng (University of Technology Sydney)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper provides two perspectives on constrained differential privacy mechanisms: belief revision and belief updating. Through these two perspectives, it explores how to combine randomized differential privacy mechanisms with deterministic constraints to maintain the standard trade-off between privacy protection and data utility.
UCoL: Unsupervised Learning of Discriminative Facial Representations via Uncertainty-Aware Contrast
Hao Wang (ByteDance Inc), Liying Chi (The University of Sydney)
RecognitionRepresentation LearningContrastive LearningImage
🎯 What it does: A completely unsupervised facial representation learning framework UCoL is proposed, combining a dual-path momentum contrast network and an uncertainty-aware kNN self-labeling mechanism to achieve high-quality construction and contrastive learning of positive and negative pairs.
UEQMS: UMAP Embedded Quick Mean Shift Algorithm for High Dimensional Clustering
Abhishek Kumar (TCG Creast), Rammohan Mallipeddi (Indian Statistical Institute)
OptimizationImageTabular
🎯 What it does: The convergence of MeanShift++ has been studied and improved, and Quick Mean Shift (QMS) has been proposed to eliminate redundant calculations and further integrate UMAP dimensionality reduction, forming the UEQMS algorithm for high-dimensional clustering.
Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method
Tao Wang (Nanjing University), Tong Lu (Rakuten Institute of Technology)
RestorationTransformerImageBenchmark
🎯 What it does: A novel dataset called UHD-LOL for ultra-high resolution (4K/8K) low-light image enhancement is proposed, and based on this dataset, the LLFormer Transformer model is introduced to achieve UHD-LLIE.
Ultrafast Euclidean Shortest Path Computation Using Hub Labeling
Jinchun Du (Monash University), Muhammad Aamir Cheema (Monash University)
OptimizationComputational EfficiencySimultaneous Localization and MappingGraphBenchmark
🎯 What it does: A new Euclidean shortest path calculation method, called Euclidean Hub Labeling (EHL), is proposed for computing the shortest path in a Euclidean plane with polygonal obstacles.
Unbalanced CO-optimal Transport
Quang Huy Tran (Université Bretagne Sud), Ritambhara Singh (Brown University)
Domain AdaptationAnomaly DetectionOptimizationImageMultimodalityBiomedical Data
🎯 What it does: Proposed and theoretically proved the COOT extended UCOOT that can handle imbalanced samples and features, and verified its robustness against noise/outliers.
Unbiased Heterogeneous Scene Graph Generation with Relation-Aware Message Passing Neural Network
Kanghoon Yoon (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)
Object DetectionGenerationGraph Neural NetworkGraph
🎯 What it does: An unbiased heterogeneous scene graph generation framework HetSGG is proposed, and a relation-aware message passing network RMP is designed to better capture the contextual information of object relationships.
Uncertainty-Aware Image Captioning
Zhengcong Fei (Meituan), Xiaolin Wei (Meituan)
GenerationTransformerImageText
🎯 What it does: A framework for image caption generation based on uncertainty awareness, called UAIC, is proposed, which generates captions in parallel from easy to difficult using an insertion Transformer.
Uncertainty-Aware Self-Training for Low-Resource Neural Sequence Labeling
Jianing Wang (East China Normal University), Aoying Zhou (East China Normal University)
ClassificationRecognitionTransformerSupervised Fine-TuningTextSequential
🎯 What it does: This paper proposes SeqUST, a framework that utilizes uncertainty-aware self-training to address the low-resource neural sequence labeling problem.
Understanding Representation Learnability of Nonlinear Self-Supervised Learning
Ruofeng Yang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Representation LearningContrastive LearningTabular
🎯 What it does: This study investigates the learning outcomes of a single-layer nonlinear self-supervised learning (SSL) model, demonstrating that it can converge to a local optimum after gradient descent training and accurately describes the feature representation corresponding to that optimum solution.
Understanding the Generalization Performance of Spectral Clustering Algorithms
Shaojie Li (Renmin University of China), Yong Liu (Renmin University of China)
Tabular
🎯 What it does: This paper studies the generalization performance of commonly used spectral clustering algorithms (Relaxed RatioCut and Relaxed NCut), derives the excess risk bounds for continuous and discrete solutions, and proposes the GPOD algorithm, which penalizes key quantities and supports efficient clustering of new samples without the need for re-feature decomposition.
Underwater Ranker: Learn Which Is Better and How to Be Better
Chunle Guo (Nankai University), Chongyi Li (Nanyang Technological University)
RestorationTransformerImage
🎯 What it does: This paper proposes a ranking-based underwater image quality assessment method called URanker based on Transformer, and constructs a dedicated underwater image ranking dataset called URankerSet. It also demonstrates that URanker can serve as additional supervision to enhance the performance of underwater image enhancement networks, and introduces a Normalization Tail to further improve enhancement quality.