AAAI 2023 Papers — Page 16
AAAI Conference on Artificial Intelligence · 1578 papers
Unfooling Perturbation-Based Post Hoc Explainers
Zachariah Carmichael (University of Notre Dame), Walter J. Scheirer (University of Notre Dame)
Anomaly DetectionExplainability and InterpretabilityAdversarial AttackTabular
🎯 What it does: This paper proposes a detection and defense framework against adversarial attacks on perturbation-based post-hoc explainers (such as LIME and SHAP), with the core being an unsupervised conditional anomaly detector KNN-CAD based on k-nearest neighbors. Based on this, we designed the attack detection algorithm CAD-Detect and the defense algorithm CAD-Defend; at the same time, new evaluation metrics (credibility of attack and defense, sparsity of explanations, infidelity, etc.) are introduced.
Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation
Yizhou Dang (Northeastern University), Hong Liu (Alibaba Group)
Recommendation SystemTransformerContrastive LearningSequential
🎯 What it does: This paper proposes a time interval-based data augmentation method for sequential data, converting non-uniform time interval sequences into uniform sequences to enhance sequence recommendation performance.
Unifying Vision-Language Representation Space with Single-Tower Transformer
Jiho Jang (Seoul National University), Nojun Kwak (Seoul National University)
RetrievalRepresentation LearningTransformerAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: A single Transformer architecture (OneR) is constructed, which encodes both images and text in a unified representation space through cross-modal mixing and context-independent contrastive learning.
UniSyn: An End-to-End Unified Model for Text-to-Speech and Singing Voice Synthesis
Yi Lei (Northwestern Polytechnical University), Dan Su (Tencent)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkAudio
🎯 What it does: This paper proposes UniSyn, an end-to-end unified model that can generate the speaking and singing voice of a target speaker using only speaking or singing data.
Universal Information Extraction as Unified Semantic Matching
Jie Lou (Baidu Inc.), Hua Wu (Baidu Inc.)
TransformerText
🎯 What it does: A unified semantic matching framework (USM) is proposed, which verbalizes the labeling scheme and splits the information extraction tasks into three types of instructions: structured (Token-Token Linking) and conceptual (Label-Token/Token-Label Linking), achieving parallel extraction of multiple extraction targets such as entities, relationships, events, and sentiments within a single model.
Universe Points Representation Learning for Partial Multi-Graph Matching
Zhakshylyk Nurlanov (Bosch Center for Artificial Intelligence), Florian Bernard (University of Bonn)
Object DetectionRepresentation LearningGraph Neural NetworkImageGraph
🎯 What it does: This paper proposes a deep learning framework based on Universe Points Representation Learning (URL) to address the problem of partial multi-graph matching;
Unlabeled Imperfect Demonstrations in Adversarial Imitation Learning
Yunke Wang (Wuhan University), Chang Xu (University of Sydney)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: A positive-negative unlabeled adversarial imitation learning framework UID is proposed to learn policies in situations where the demonstration data contains an unknown proportion of imperfect samples.
Unsupervised Cross-Domain Rumor Detection with Contrastive Learning and Cross-Attention
Hongyan Ran (Beijing Jiaotong University), Caiyan Jia (Beijing Jiaotong University)
Domain AdaptationAnomaly DetectionTransformerContrastive LearningText
🎯 What it does: This paper proposes an unsupervised cross-domain rumor detection model UCD-RD, which combines instance-level and prototype-level contrastive learning along with a cross-attention mechanism to achieve feature alignment and robustness enhancement from the source domain to the target domain.
Unsupervised Deep Embedded Fusion Representation of Single-Cell Transcriptomics
Yue Cheng (Jilin University), Xiangtao Li (City University of Hong Kong)
Representation LearningGraph Neural NetworkAuto EncoderBiomedical Data
🎯 What it does: The scDEFR model is proposed, which achieves clustering of single-cell RNA sequencing data by integrating cellular topology and transcriptomic information through deep embedding.
Unsupervised Deep Learning for Phase Retrieval via Teacher-Student Distillation
Yuhui Quan (South China University of Technology), Hui Ji (National University of Singapore)
RestorationKnowledge DistillationRecurrent Neural NetworkImage
🎯 What it does: A completely unsupervised deep learning framework is proposed for phase recovery through teacher-student distillation.
Unsupervised Deep Video Denoising with Untrained Network
Huan Zheng (National University of Singapore), Hui Ji (National University of Singapore)
RestorationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: A completely unsupervised deep video denoising method that does not rely on clean video samples is proposed.
Unsupervised Domain Adaptation for Medical Image Segmentation by Selective Entropy Constraints and Adaptive Semantic Alignment
Wei Feng (Monash University), Zongyuan Ge (Monash University)
SegmentationDomain AdaptationGenerative Adversarial NetworkMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A framework for unsupervised domain adaptation based on selective entropy constraints and adaptive semantic alignment is proposed for cross-modal medical image segmentation.
Unsupervised Explanation Generation via Correct Instantiations
Sijie Cheng (Fudan University), Lingpeng Kong (Shanghai Artificial Intelligence Laboratory)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A two-stage unsupervised explanation generation framework called NEON is designed and implemented, which first automatically generates correct instances that are similar to the commonality of erroneous statements, and then uses these instances to induce large pre-trained language models to implicitly infer conflict points and generate natural language explanations.
Unsupervised Hierarchical Domain Adaptation for Adverse Weather Optical Flow
Hanyu Zhou (Huazhong University of Science and Technology), Luxin Yan (Sun Yat-sen University)
Domain AdaptationGenerative Adversarial NetworkContrastive LearningOptical FlowImageVideo
🎯 What it does: This paper proposes an unsupervised hierarchical motion-boundary adaptation framework that transfers optical flow knowledge from clean domains to adverse weather domains, enabling optical flow estimation in adverse weather conditions.
Unsupervised Legal Evidence Retrieval via Contrastive Learning with Approximate Aggregated Positive
Feng Yao (Tsinghua University), Weixing Shen (Tsinghua University)
RetrievalTransformerContrastive LearningText
🎯 What it does: By constructing a dense retrieval model, this paper automatically retrieves oral evidence related to facts in criminal cases, helping judges quickly verify facts.
Unsupervised Multi-Exposure Image Fusion Breaking Exposure Limits via Contrastive Learning
Han Xu (Wuhan University), Jiayi Ma (Wuhan University)
RestorationContrastive LearningImage
🎯 What it does: A novel unsupervised multi-exposure image fusion method MEF-CL is proposed, which utilizes contrastive learning to achieve exposure limitation breakthrough;
Unsupervised Paraphrasing under Syntax Knowledge
Tianyuan Liu (Shandong University), Bin Gong (Shandong University)
GenerationRecurrent Neural NetworkContrastive LearningText
🎯 What it does: This paper proposes an unsupervised grammar-controlled paraphrasing method called CKPara, which utilizes dependency syntax trees as grammatical guidance and combines pre-trained compositional word knowledge to achieve consistency constraints at both the word and sentence levels.
Untangled: A Complete Dynamic Topological Logic
David Fernández-Duque (Czech Academy of Sciences), Yoàv Montacute (University of Cambridge)
🎯 What it does: This paper presents a complete and finitely axiomatizable framework of Dynamic Topological Logic (DGL) defined on scatter spaces, addressing the previously unsolved DGL problem that was only complete under extended languages.
Untargeted Attack against Federated Recommendation Systems via Poisonous Item Embeddings and the Defense
Yang Yu (University of Science and Technology of China), Zaixi Zhang (University of Science and Technology of China)
Recommendation SystemFederated LearningAdversarial AttackContrastive LearningTabular
🎯 What it does: For federated recommendation systems, we propose ClusterAttack, which uploads toxic gradient aggregates of item embeddings, leading to the failure of recommendation ranking. At the same time, we design the UNION defense mechanism, which uses contrastive learning to make item embeddings tend towards a uniform distribution, thereby detecting and filtering malicious gradients.
Unveiling the Black Box of PLMs with Semantic Anchors: Towards Interpretable Neural Semantic Parsing
Lunyiu Nie (Tsinghua University), Jidong Zhai (Tsinghua University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A pre-trained language model fine-tuning framework based on a hierarchical decoder is proposed, incorporating an intermediate supervision task for semantic anchor extraction and alignment, which enhances semantic parsing quality and explicitly reveals internal representations.
USDNL: Uncertainty-Based Single Dropout in Noisy Label Learning
Yuanzhuo Xu (Wuhan University), Ruizhi Chen (Wuhan University)
ClassificationComputational EfficiencyData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A noise label robust training method based on single dropout, USDNL, is proposed, which utilizes prior knowledge from the early training stage to estimate the model's determinacy (epistemic uncertainty) through single dropout, and combines cross-entropy loss to filter clean samples for training.
User-Controllable Arbitrary Style Transfer via Entropy Regularization
Jiaxin Cheng (University of Southern California), Prem Natarajan (Amazon)
Image TranslationDomain AdaptationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A user-controllable arbitrary style transfer (ε-Assign-and-Mix) framework based on entropy regularization is proposed, which can produce diverse style transfer results while maintaining speed and quality.
USER: Unsupervised Structural Entropy-Based Robust Graph Neural Network
Yifei Wang (University of Auckland), Jiamou Liu (University of Auckland)
Anomaly DetectionRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes an unsupervised robust graph neural network framework called USER, which constructs a harmless graph using a learnable adjacency matrix and unlabeled data, thereby offsetting random perturbations in the graph and learning more robust node representations.
Utility Maximizer or Value Maximizer: Mechanism Design for Mixed Bidders in Online Advertising
Hongtao Lv (Shandong University), Fan Wu (Shanghai Jiao Tong University)
Optimization
🎯 What it does: A sincere (IC, IR) bidding mechanism is proposed for hybrid advertisers (utility maximizers and value maximizers), with two algorithms, MPU and MPR, designed for public and private categories respectively.
Utilizing Prior Solutions for Reward Shaping and Composition in Entropy-Regularized Reinforcement Learning
Jacob Adamczyk (University of Massachusetts Boston), Rahul V. Kulkarni (San Jose State University)
Reinforcement Learning
🎯 What it does: This paper proposes a general framework for utilizing solutions from solved tasks for reward shaping and task composition in entropy-regularized reinforcement learning, thereby achieving rapid transfer learning.
Value-Consistent Representation Learning for Data-Efficient Reinforcement Learning
Yang Yue (Tsinghua University), Shuicheng Yan (Sea AI Lab)
Representation LearningReinforcement LearningContrastive LearningVideo
🎯 What it does: This paper proposes a Value Consistent Representation Learning (VCR) method that enhances the sample efficiency of RL by predicting the value of imagined future states and aligning them with real states.
Variable-Based Calibration for Machine Learning Classifiers
Markelle Kelly (University of California, Irvine), Padhraic Smyth (University of California, Irvine)
ClassificationImageTextTabular
🎯 What it does: This paper proposes variable-based calibration (VCE, VECE) metrics and visualization methods, and calibrates specific variables to reveal systematic errors hidden by traditional ECE.
Variational Wasserstein Barycenters with C-cyclical Monotonicity Regularization
Jinjin Chi (Jilin University), Renchu Guan (Jilin University)
OptimizationTabular
🎯 What it does: A variational Wasserstein barycenter method based on c-cyclic monotonicity regularization is proposed, which directly estimates the barycenter of continuous distributions using samples.
VASR: Visual Analogies of Situation Recognition
Yonatan Bitton (Hebrew University of Jerusalem), Gabriel Stanovsky (Hebrew University of Jerusalem)
RecognitionTransformerContrastive LearningImageTextBenchmark
🎯 What it does: This paper constructs a visual analogy task—Visual Analogical Situations Recognition (VASR), which automatically generates approximately 500,000 scene-level analogy instances using situation recognition (imSitu) annotations and the CLIP model, and obtains 3,820 gold standard samples through crowdsourcing validation, releasing the dataset publicly.
VBLC: Visibility Boosting and Logit-Constraint Learning for Domain Adaptive Semantic Segmentation under Adverse Conditions
Mingjia Li (Beijing Institute of Technology), Xinjing Cheng (Tsinghua University)
Object DetectionSegmentationDomain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes an unsupervised domain adaptation framework VBLC that does not rely on normal image pairs, utilizing a visibility enhancement module and logarithmic it constraint learning to improve the robustness of semantic segmentation models under adverse weather conditions.
Vector Causal Inference between Two Groups of Variables
Jonas Wahl (Technische Universitat Berlin), Jakob Runge (Technische Universitat Berlin)
Time Series
🎯 What it does: This paper proposes a new non-parametric constraint method for inferring the causal relationship between two vector random variables from observational data without dimensionality reduction.
Very Fast, Approximate Counterfactual Explanations for Decision Forests
Miguel Á. Carreira-Perpinan (University of California), Suryabhan Singh Hada (University of California)
Explainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: The study proposes a fast approximate adversarial explanation method (LIRE) for decision forests (such as random forests) to find the minimal modifications that make the model output meet the target prediction for input instances.
Video Compression Artifact Reduction by Fusing Motion Compensation and Global Context in a Swin-CNN Based Parallel Architecture
Xinjian Zhang (Fudan University), Weishan Zhang (China University of Petroleum)
RestorationCompressionConvolutional Neural NetworkTransformerVideo
🎯 What it does: A novel spatiotemporal compensation fusion framework (STCF) is proposed, which removes video compression artifacts by integrating motion compensation and global context through parallel Ada-CNN and Swin self-attention.
Video Event Extraction via Tracking Visual States of Arguments
Guang Yang (Tsinghua University), Shih-Fu Chang (Columbia University)
RecognitionObject DetectionObject TrackingConvolutional Neural NetworkRecurrent Neural NetworkTransformerVideo
🎯 What it does: A framework for event extraction based on the visual state changes of argumentative entities in videos is proposed, which identifies video events and their semantic roles by tracking pixel changes, positional displacements, and multi-object interactions.
Video Object of Interest Segmentation
Siyuan Zhou (Shanghai Jiao Tong University), Li Niu (Alibaba Group)
Object TrackingSegmentationTransformerImageVideoBenchmark
🎯 What it does: This paper proposes the Video Object Interest Segmentation (VOIS) task, which aims to simultaneously segment and track all relevant objects in a video based on a given target image.
Video-Text Pre-training with Learned Regions for Retrieval
Rui Yan (Nanjing University of Science and Technology), Jinhui Tang (Tongji University)
RetrievalTransformerContrastive LearningVideoText
🎯 What it does: In video-text pre-training, RegionLearner is proposed, which can explicitly capture the semantics of objects in videos by clustering patch features of video frames, learning maskable regions, and constructing region-level spatiotemporal graphs without using any positional supervision. This module is embedded in an end-to-end pre-training framework.
VideoDubber: Machine Translation with Speech-Aware Length Control for Video Dubbing
Yihan Wu (Renmin University of China), Jiang Bian (Microsoft)
TransformerVideoTextAudio
🎯 What it does: The VideoDubber system was designed and implemented for video dubbing, ensuring that the synthesized voice is synchronized with the video by directly controlling the duration matching between the translated text and the original speech during the machine translation phase.
VIDM: Video Implicit Diffusion Models
Kangfu Mei (Johns Hopkins University), Vishal Patel (Johns Hopkins University)
GenerationData SynthesisDiffusion modelOptical FlowVideo
🎯 What it does: This paper proposes a Video Implicit Diffusion Model (VIDM), which models content and motion separately using diffusion models, and achieves high-quality video synthesis by employing implicit conditioning and positional group normalization during the generation process.
Visually Grounded Commonsense Knowledge Acquisition
Yuan Yao (Tsinghua University), Maosong Sun
Knowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes CLEVER, a framework for unsupervised multi-instance learning that utilizes images for visual anchoring of common sense knowledge extraction.
VLTinT: Visual-Linguistic Transformer-in-Transformer for Coherent Video Paragraph Captioning
Kashu Yamazaki (University of Arkansas), Ngan Le (University of Arkansas)
Object DetectionGenerationTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes the Visual-Linguistic Transformer-in-Transformer (VLTinT) model for generating coherent video paragraph descriptions.
Voting with Preference Intensities
Anson Kahng (University of Rochester), Nisarg Shah (University of Toronto)
🎯 What it does: This paper studies the use of intensity markers (double arrows) to represent preference intensity in voting and designs approximately optimal voting rules within the 'distortion' framework.
Warm-Starting Nested Rollout Policy Adaptation with Optimal Stopping
Chen Dang (Orange Labs), Pierre-Henri Wuillemin (Sorbonne Université)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes the Meta-NRPA algorithm, which combines optimal stopping theory with Nested Rollout Policy Adaptation (NRPA) for warm starting, and introduces exploration techniques such as Force Exploration, ε-greedy, and dynamic learning rates, significantly improving the performance of NRPA on three types of problems, and providing a new lower bound for the Snake-in-the-Box problem.
Was Fixing This Really That Hard? On the Complexity of Correcting HTN Domains
Songtuan Lin (Australian National University), Pascal Bercher (Australian National University)
🎯 What it does: This paper studies the computational complexity of model correction in hierarchical task network (HTN) planning domains under given whitelist and blacklist schemes.
Wasserstein Actor-Critic: Directed Exploration via Optimism for Continuous-Actions Control
Amarildo Likmeta (University of Bologna), Marcello Restelli (Politecnico di Milano)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper studies a Wasserstein distance-based Actor-Critic algorithm (WAC) that achieves directed exploration in continuous action spaces.
Wasserstein Graph Distance Based on L1–Approximated Tree Edit Distance between Weisfeiler–Lehman Subtrees
Zhongxi Fang (Waseda University), Hiroyuki Kasai (Waseda University)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: Proposes the Wasserstein distance based on WL subtree (WWLS) for L1-approximate tree edit distance, capturing fine-grained differences in graph structures and providing a new graph similarity measure.
WaveForM: Graph Enhanced Wavelet Learning for Long Sequence Forecasting of Multivariate Time Series
Fuhao Yang (Beijing Institute of Technology), Mingzhong Wang (The University of the Sunshine Coast)
Recurrent Neural NetworkGraph Neural NetworkTime Series
🎯 What it does: This paper proposes WaveForM, an end-to-end multivariate time series forecasting framework based on discrete wavelet transform and graph convolution.
Weakly Supervised 3D Multi-Person Pose Estimation for Large-Scale Scenes Based on Monocular Camera and Single LiDAR
Peishan Cong (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
Pose EstimationRecurrent Neural NetworkImagePoint Cloud
🎯 What it does: A weakly supervised 3D multi-person pose estimation method called FusionPose is proposed, which can perform pose regression in large-scale scenes without 3D annotations using a single camera and a single LiDAR.
Weakly Supervised 3D Segmentation via Receptive-Driven Pseudo Label Consistency and Structural Consistency
Yuxiang Lan (Xiamen University), Zongze Wu (Shenzhen University)
SegmentationPoint Cloud
🎯 What it does: This paper proposes a weakly supervised point cloud semantic segmentation framework called RPSC, which enhances segmentation performance under sparse annotations by utilizing receptive field-driven pseudo-label consistency and structural consistency.
Weakly-Guided Self-Supervised Pretraining for Temporal Activity Detection
Kumara Kahatapitiya (Stony Brook University), Gang Hua (Wormpex AI Research)
ClassificationRecognitionObject DetectionConvolutional Neural NetworkSupervised Fine-TuningVideo
🎯 What it does: Using weak labels to generate frame-level pseudo labels and employing video augmentation methods such as volume freezing, MixUp, and CutMix, self-supervised pre-training is conducted on large classification datasets to enhance temporal activity detection performance.
Weakly-Supervised Camouflaged Object Detection with Scribble Annotations
Ruozhen He (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
Object DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Developed the first weakly supervised camouflage object detection framework based on line annotations and created the S-COD dataset.
Weakly-Supervised Semantic Segmentation for Histopathology Images Based on Dataset Synthesis and Feature Consistency Constraint
Zijie Fang (Tsinghua University), Yongbing Zhang (Harbin Institute of Technology)
SegmentationData SynthesisConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a weakly supervised semantic segmentation framework called PistoSeg, which generates a pixel-level annotated dataset by performing Mosaic synthesis on images of a single tissue category. It first trains a preliminary segmentation network on this synthetic dataset, and then refines the CAM and the pseudo-mask obtained from the preliminary segmentation using attention feature consistency, ultimately producing high-quality pseudo-masks for training a fine segmentation model.
Weight Predictor Network with Feature Selection for Small Sample Tabular Biomedical Data
Andrei Margeloiu (University of Cambridge), Mateja Jamnik (University of Cambridge)
ClassificationOptimizationComputational EfficiencyTabularBiomedical Data
🎯 What it does: A framework called WPFS, based on weight prediction networks and feature selection, has been designed and implemented for classification on high-dimensional, extremely small sample biomedical tabular data, significantly reducing learnable parameters and achieving global feature selection.
Weighted Policy Constraints for Offline Reinforcement Learning
Zhiyong Peng (National University of Defense Technology), Zongtan Zhou (National University of Defense Technology)
Reinforcement LearningTabular
🎯 What it does: A Weighted Policy Constraints (wPC) offline reinforcement learning algorithm is proposed and implemented, which applies constraints only to the desirable state-action pairs in the dataset during training, thereby alleviating distribution shift and improving policy performance.
What Do You MEME? Generating Explanations for Visual Semantic Role Labelling in Memes
Shivam Sharma (Indraprastha Institute of Information Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the EXCLAIM task, which involves generating natural language explanations for entities (heroes, villains, victims) in memes, and constructs the corresponding dataset ExHVV.
What Does Your Face Sound Like? 3D Face Shape towards Voice
Zhihan Yang (Tsinghua University), Jia Jia (Tsinghua University)
GenerationData SynthesisTransformerAudio
🎯 What it does: A speech generation framework based on 3D facial shapes is proposed, which maps the 3D facial model, texture, facial attributes, and demographic information to speaker embeddings to synthesize speech.
When Congestion Games Meet Mobile Crowdsourcing: Selective Information Disclosure
Hongbo Li (Singapore University of Technology and Design), Lingjie Duan (Singapore University of Technology and Design)
OptimizationReinforcement LearningGraph
🎯 What it does: This paper studies how to incentivize self-interested users to achieve the optimal exploration-exploitation trade-off in a dynamic congestion game through information mechanisms in a mobile crowdsourcing environment, proposing a Selective Information Disclosure (SID) mechanism.
When Neural Networks Fail to Generalize? A Model Sensitivity Perspective
Jiajin Zhang (Rensselaer Polytechnic Institute), Pingkun Yan (Rensselaer Polytechnic Institute)
Domain AdaptationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a framework for adversarial data augmentation in the frequency domain called SADA, aimed at enhancing single-source domain generalization by suppressing the model's spectral sensitivity.
When Online Learning Meets ODE: Learning without Forgetting on Variable Feature Space
Diyang Li (Nanjing University of Information Science and Technology), Bin Gu (MBZUAI)
TabularOrdinary Differential Equation
🎯 What it does: This paper proposes a dynamic feature learning system (DFLS) based on ordinary differential equations (ODEs), which can online update model parameters as the feature space changes over time and ensure convergence to the same optimal solution as a completely new training;
Where Will Players Move Next? Dynamic Graphs and Hierarchical Fusion for Movement Forecasting in Badminton
Kai-Shiang Chang (National Yang Ming Chiao Tung University), Wen-Chih Peng (National Yang Ming Chiao Tung University)
Graph Neural NetworkGraphTime SeriesSequential
🎯 What it does: Proposes a motion prediction task that predicts the type and location of a player's next shot based on badminton rally records.
Which Shortcut Solution Do Question Answering Models Prefer to Learn?
Kazutoshi Shinoda (University of Tokyo), Akiko Aizawa (National Institute of Informatics)
TransformerSupervised Fine-TuningText
🎯 What it does: This paper systematically evaluates and quantifies the shortcut learning tendencies and learnability of question-answering models under biased training sets through behavioral experiments, loss surface visualization, and MDL calculations.
Why Capsule Neural Networks Do Not Scale: Challenging the Dynamic Parse-Tree Assumption
Matthias Mitterreiter (Friedrich Schiller University Jena), Sören Laue (Technical University Kaiserslautern)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically evaluates the performance of the original Capsule Neural Network (CapsNet) proposed by Sabour et al. in terms of scalability, routing dynamics, viewpoint invariance, and capsule activation through extensive experiments and theoretical analysis, and questions its core concept—the implementation of the parse-tree;
Wiener Graph Deconvolutional Network Improves Graph Self-Supervised Learning
Jiashun Cheng (Hong Kong University of Science and Technology), Fugee Tsung (Hong Kong University of Science and Technology)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a Wiener frequency domain filter-based Graph Deconvolutional Network (WGDN) for the decoder in graph self-supervised learning, enhancing the quality of node and graph-level representations.
WIERT: Web Information Extraction via Render Tree
Zimeng Li (Beihang University), Daxin Jiang (Microsoft)
TransformerLarge Language ModelText
🎯 What it does: A Web information extraction method based on the rendering tree, called WIERT, is proposed.
Win-Win: A Privacy-Preserving Federated Framework for Dual-Target Cross-Domain Recommendation
Gaode Chen (Institute of Information Engineering, Chinese Academy of Sciences), Yu Zheng (JD Technology)
Recommendation SystemFederated LearningSafty and PrivacyTabular
🎯 What it does: Designed the P2FCDR framework to achieve federated learning and local differential privacy for dual-target cross-domain recommendation.
WLD-Reg: A Data-Dependent Within-Layer Diversity Regularizer
Firas Laakom (Tampere University), Moncef Gabbouj (Tampere University)
OptimizationImage
🎯 What it does: A data-driven intra-layer diversity regularization method is proposed, encouraging the activation diversity of neurons within the same layer.
WSiP: Wave Superposition Inspired Pooling for Dynamic Interactions-Aware Trajectory Prediction
Renzhi Wang (Central South University), Xiang Wang (National University of Defense Technology)
Autonomous DrivingExplainability and InterpretabilityRecurrent Neural NetworkGenerative Adversarial NetworkTime SeriesSequential
🎯 What it does: A wave-pooling method based on wave superposition is proposed and embedded into an encoding-decoding framework WSiP for predicting vehicle trajectories on highways.
XClusters: Explainability-First Clustering
Hyunseung Hwang (Korea Advanced Institute of Science and Technology), Steven Euijong Whang (Korea Advanced Institute of Science and Technology)
OptimizationExplainability and InterpretabilityTime SeriesSequentialFinance Related
🎯 What it does: Proposes the XClusters framework, which jointly optimizes clustering error and decision tree interpretability (number of tree nodes), while dynamically adjusting parameters during the clustering process.
XRand: Differentially Private Defense against Explanation-Guided Attacks
Truc Nguyen (University of Florida), My T. Thai (New Jersey Institute of Technology)
Safty and PrivacyExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a two-stage local differential privacy (LDP) mechanism named XRAND, which randomizes interpretative information in a machine learning as a service (MLaaS) environment to defend against explanation-based backdoor attacks.
Yet Another Traffic Classifier: A Masked Autoencoder Based Traffic Transformer with Multi-Level Flow Representation
Ruijie Zhao (Shanghai Jiao Tong University), Zhi Xue (Shanghai Jiao Tong University)
ClassificationRepresentation LearningTransformerAuto EncoderTabular
🎯 What it does: Developed a traffic Transformer based on MAE (YaTC) that classifies network traffic using multi-layer flow representation.
YOLOV: Making Still Image Object Detectors Great at Video Object Detection
Yuheng Shi (Tianjin University), Xiaojie Guo (TuSimple)
Object DetectionConvolutional Neural NetworkImageVideo
🎯 What it does: In the video object detection task, a one-stage detector based on YOLOX selects high-confidence boxes through post-processing and aggregates the features of these boxes to improve detection accuracy.
Zero-Cost Operation Scoring in Differentiable Architecture Search
Lichuan Xiang (University of Warwick), Hongkai Wen (University of Warwick)
Neural Architecture SearchImageBenchmark
🎯 What it does: This paper proposes a zero-cost perturbation-based operation scoring method called Zero-Cost-PT, aimed at improving local operation selection in differentiable neural architecture search.
Zero-Knowledge Proofs for Classical Planning Problems
Augusto B. Corrêa (University of Basel), Remo Christen (University of Basel)
Safty and PrivacyComputational Efficiency
🎯 What it does: A zero-knowledge proof protocol is proposed for classical planning tasks with polynomial-length plans, proving the existence of plans without revealing any plan details.
Zero-Shot Assistance in Sequential Decision Problems
Sebastiaan De Peuter (Aalto University), Samuel Kaski (University of Manchester)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: This study investigates a method for solving new sequential decision-making problems with the assistance of an advisor under zero-shot assistance.
Zero-Shot Cross-Lingual Event Argument Extraction with Language-Oriented Prefix-Tuning
Pengfei Cao (National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences), Jun Zhao (National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies zero-shot cross-lingual event argument extraction and proposes the LAPIN model.
Zero-Shot Face-Based Voice Conversion: Bottleneck-Free Speech Disentanglement in the Real-World Scenario
Shao-En Weng (National Yang Ming Chiao Tung University), Wen-Huang Cheng (National Yang Ming Chiao Tung University)
GenerationData SynthesisGenerative Adversarial NetworkVideoAudio
🎯 What it does: A bottleneck-free facial-to-speech conversion model (SP-FaceVC) is proposed, which extracts speech content features through low-pass lifespan processing and achieves zero-shot cross-modal speech synthesis using average facial embeddings, reparameterization, and multi-scale discriminators.
Zero-Shot Linear Combinations of Grounded Social Interactions with Linear Social MDPs
Ravi Tejwani (Massachusetts Institute of Technology), Andrei Barbu (Massachusetts Institute of Technology)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningVideo
🎯 What it does: A Linear Social MDP model is proposed, which can generate and combine various social interactions among robots in a zero-sample manner.
Zero-Shot Rumor Detection with Propagation Structure via Prompt Learning
Hongzhan Lin (Hong Kong Baptist University), Ruifang Liu (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A zero-shot rumor detection framework RPL is proposed, combining propagation structure with prompt learning.
Zero-Shot Slot Filling with Slot-Prefix Prompting and Attention Relationship Descriptor
Qiaoyang Luo (University of Adelaide), Lingqiao Liu (University of Adelaide)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A zero-shot slot filling method based on Slot-Prefix prompts and attention relationship descriptors is proposed, utilizing a pre-trained language model to achieve single-stage sequence labeling.