IJCAI 2023 Papers — Page 7
International Joint Conference on Artificial Intelligence · 639 papers
Towards Hierarchical Policy Learning for Conversational Recommendation with Hypergraph-based Reinforcement Learning
Sen Zhao (Huazhong University of Science and Technology), Zujie Wen (Ant Group)
Recommendation SystemRecurrent Neural NetworkGraph Neural NetworkTransformerReinforcement LearningTextSequential
🎯 What it does: Propose a Director-Actor Hierarchical Conversation Recommendation framework (DAHCR), which significantly improves the effectiveness of multi-round dialogue recommendations by separating inquiry and recommendation decisions, dynamically modeling user preferences with hypergraphs, and guiding weak supervision through intrinsic motivation.
Towards Incremental NER Data Augmentation via Syntactic-aware Insertion Transformer
Wenjun Ke (Southeast University), Rui Qi (China Life Property Casualty Insurance Company Limited)
RecognitionData SynthesisTransformerText
🎯 What it does: Propose SAINT, an insertive Transformer that leverages syntactic information to generate NER incremental data that maintains both semantic diversity and syntactic consistency;
Towards Long-delayed Sparsity: Learning a Better Transformer through Reward Redistribution
Tianchen Zhu (Beihang University), Jianxin Li (Beihang University)
OptimizationTransformerReinforcement LearningImageSequential
🎯 What it does: Propose an adaptive reward reallocation algorithm named DTRD to address the performance degradation of Decision Transformer in long-term delayed reward environments.
Towards Lossless Head Pruning through Automatic Peer Distillation for Language Models
Bingbing Li (University of Connecticut), Caiwen Ding (University of Connecticut)
Computational EfficiencyKnowledge DistillationTransformerText
🎯 What it does: A head pruning method based on differentiable attention head pruning and 'Peer Distillation' knowledge recovery is studied, which can significantly reduce the number of heads while maintaining or improving BERT performance.
Towards Robust Gan-Generated Image Detection: A Multi-View Completion Representation
Chi Liu (University of Technology Sydney), Wanlei Zhou (City University of Macau)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: Propose a framework based on multi-perspective completion and cross-perspective classification for detecting GAN-generated images.
Towards Robust Scene Text Image Super-resolution via Explicit Location Enhancement
Hang Guo (Tsinghua Shenzhen International Graduate School, Tsinghua University), Shu-Tao Xia (Tsinghua Shenzhen International Graduate School, Tsinghua University)
RecognitionSuper ResolutionTransformerVision Language ModelImage
🎯 What it does: Propose a scene text image super-resolution method named LEMMA, which distinguishes text from the background by leveraging explicit character position enhancement, bidirectional multimodal alignment, and adaptive fusion to improve recognition performance.
Towards Semantics- and Domain-Aware Adversarial Attacks
Jianping Zhang (Chinese University of Hong Kong), Michael R. Lyu (Chinese University of Hong Kong)
Domain AdaptationAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes a word-level adversarial attack method based on semantics and domain awareness, utilizing contrastive learning and domain pre-training to train language models for generating replacement words, and balancing attack success rate and sample quality through an iterative update framework.
Towards Sharp Analysis for Distributed Learning with Random Features
Jian Li (Chinese Academy of Sciences), Yong Liu (Renmin University of China)
OptimizationFederated Learning
🎯 What it does: This paper extends the optimal rates of distributed learning and random features to inaccessible scenarios, reducing the required number of random features through data-dependent generation strategies while leveraging additional unlabeled data to increase the allowable number of partitions.
TPS++: Attention-Enhanced Thin-Plate Spline for Scene Text Recognition
Tianlun Zheng (Fudan University), Yu-Gang Jiang (Fudan University)
RecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: Propose TPS++, an attention-enhanced Thin-Plate Spline (TPS) transformation for preprocessing in scene text recognition, improving traditional TPS's control point regression and content-agnostic parameter estimation.
Tracking Different Ant Species: An Unsupervised Domain Adaptation Framework and a Dataset for Multi-object Tracking
Chamath Abeysinghe (Monash University), Bernd Meyer (Monash University)
Object TrackingDomain AdaptationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkVideo
🎯 What it does: Proposes an unsupervised domain adaptation multi-object tracking framework (DA-Tracker) that can transfer across different ant species and experimental environments, and first constructs a large-scale ant tracking dataset tailored for this task.
Tractable Diversity: Scalable Multiperspective Ontology Management via Standpoint EL
Lucía Gómez Álvarez (TU Dresden), Hannes Strass (TU Dresden)
Biomedical Data
🎯 What it does: Proposed Standpoint EL (S_EL) — a multi-perspective lightweight description logic that can uniformly represent and associate different perspectives or contexts while maintaining the PTIME inference complexity of EL;
Transferable Curricula through Difficulty Conditioned Generators
Sidney Tio (Singapore Management University), Pradeep Varakantham (Singapore Management University)
Data SynthesisRepresentation LearningReinforcement LearningTabular
🎯 What it does: Propose a trainable environment generation model PERM based on Item Response Theory (IRT), which generates curriculum aligned with the agent's capabilities in parameterized environments for reinforcement learning (RL) agents, and verifies its transferability.
Treewidth-Aware Complexity for Evaluating Epistemic Logic Programs
Jorge Fandinno (University of Nebraska Omaha), Markus Hecher (Massachusetts Institute of Technology)
🎯 What it does: This paper studies the complexity of the existence of world views under the treewidth parameter, providing a refined complexity analysis for three classes of Epistemic Logic Programs (ELPs): tight, head-cycle-free (HCF), and ι-tight.
Truthful Auctions for Automated Bidding in Online Advertising
Yidan Xing (Shanghai Jiao Tong University), Guihai Chen (Shanghai Jiao Tong University)
OptimizationFinance Related
🎯 What it does: This paper proposes a new online advertising automated bidding auction model that considers advertisers' budgets and return on investment (ROI) as private constraints, and analyzes the truthful conditions under this multi-dimensional setting.
Truthful Fair Mechanisms for Allocating Mixed Divisible and Indivisible Goods
Zihao Li (Nanyang Technological University), Biaoshuai Tao (Shanghai Jiao Tong University)
Optimization
🎯 What it does: This paper studies how to achieve both fairness (EFM) and truthfulness in the allocation of mixed divisible and indivisible goods.
U-Match: Two-view Correspondence Learning with Hierarchy-aware Local Context Aggregation
Zizhuo Li (Wuhan University), Jiayi Ma (Wuhan University)
Pose EstimationRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkImagePoint CloudBenchmark
🎯 What it does: Proposed U-Match, an attention graph neural network for two-view matching learning, which can implicitly capture local context at multiple levels and fuse global information;
Unbiased Gradient Boosting Decision Tree with Unbiased Feature Importance
Zheyu Zhang (Tsinghua University), Jian Li (Tsinghua University)
ClassificationExplainability and InterpretabilityHyperparameter SearchTabularBenchmark
🎯 What it does: Studied the dual bias in splitting gain estimation and splitting search algorithms in Gradient Boosted Decision Trees (GBDT), and proposed an unbiased gain feature importance metric and the UnbiasedGBM algorithm based on this.
Unbiased Risk Estimator to Multi-Labeled Complementary Label Learning
Yi Gao (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationOptimization
🎯 What it does: This paper proposes an unbiased risk estimator for multi-label complementary label learning (MLCLL) and provides an upper bound on the estimation error.
Uncovering the Largest Community in Social Networks at Scale
Shohei Matsugu (University of Tsukuba), Hiroaki Shiokawa (University of Tsukuba)
OptimizationComputational EfficiencyGraph
🎯 What it does: Proposed and implemented a new Branch-and-Merge (BnM) algorithm for precisely finding maximum k-plex in large-scale social networks.
Understanding the Generalization Ability of Deep Learning Algorithms: A Kernelized Rényi's Entropy Perspective
Yuxin Dong (Xi'an Jiaotong University), Chen Li (Xi'an Jiaotong University)
OptimizationExplainability and InterpretabilityImage
🎯 What it does: Propose a new information measure called kernel-based Rényi entropy, and use it to perform computable theoretical analysis of the generalization error in SGD and SGLD algorithms in deep learning.
Unifying Core-Guided and Implicit Hitting Set Based Optimization
Hannes Ihalainen (University of Helsinki), Matti Järvisalo
Optimization
🎯 What it does: Proposes a unified framework UNIMAXSAT that abstracts the two types of maximum satisfiability (MaxSAT) algorithms, core-guided (CG) and implicit hitting set (IHS), into a single algorithmic structure, and provides a correctness proof for this framework; based on this, a new variant ABSTCG is implemented and its effectiveness is demonstrated through practical evaluations.
Universal Adaptive Data Augmentation
Xiaogang Xu (Zhejiang Lab), Hengshuang Zhao (University of Hong Kong)
ClassificationObject DetectionSegmentationImage
🎯 What it does: Proposed and implemented a unified method (UADA) that adaptively updates data augmentation parameters during training based on the target model's gradient.
Unreliable Partial Label Learning with Recursive Separation
Yu Shi (Southeast University), Xin Geng (Southeast University)
ClassificationData-Centric LearningImageText
🎯 What it does: This paper proposes a two-stage framework for learning with partially unreliable labels called UPLLRS. It first uses adaptive recursive separation to divide the data into reliable and unreliable subsets. Then, it performs label disambiguation on the reliable subset and employs semi-supervised learning on the unreliable subset to enhance model performance.
Unveiling Concepts Learned by a World-Class Chess-Playing Agent
Aðalsteinn Pálsson (Reykjavik University), Yngvi Björnsson (Reykjavik University)
Explainability and InterpretabilityRepresentation LearningSequential
🎯 What it does: By conducting concept probing and global explanation on the Stockfish NNUE evaluation model, revealing its internally learned chess concepts and comparing them with traditional handcrafted evaluation models.
VecoCare: Visit Sequences-Clinical Notes Joint Learning for Diagnosis Prediction in Healthcare Data
Yongxin Xu (Key Laboratory of High Confidence Software Technologies Ministry of Education), Bing Xie (Key Laboratory of High Confidence Software Technologies Ministry of Education)
TransformerLarge Language ModelContrastive LearningTextSequentialElectronic Health Records
🎯 What it does: This paper proposes a joint learning method called VecoCare to fuse access sequences and clinical notes in medical data, thereby improving the accuracy of diagnostic predictions.
VGOS: Voxel Grid Optimization for View Synthesis from Sparse Inputs
Jiakai Sun (Zhejiang University), Wei Xing (Zhejiang University)
GenerationNeural Radiance FieldImage
🎯 What it does: Achieve fast view synthesis for sparse view inputs by directly optimizing voxel grids, with a training time of only 3-5 minutes;
Video Diffusion Models with Local-Global Context Guidance
Siyuan Yang (Tsinghua University), You He (Tsinghua University)
GenerationConvolutional Neural NetworkTransformerDiffusion modelVideo
🎯 What it does: Propose a locally-global context-guided autoregressive video diffusion model (LGC-VD) for video prediction, interpolation, and unconditional generation.
Video Frame Interpolation with Densely Queried Bilateral Correlation
Chang Zhou (Nanjing University), Gangshan Wu (Nanjing University)
GenerationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: This work proposes an inter-frame interpolation method based on dense bidirectional correlation, achieving high-quality video frame interpolation by modeling correlation between two frames at a single high resolution, combined with motion generation and refinement modules, as well as a lightweight synthesis network.
Video Object Segmentation in Panoptic Wild Scenes
Yuanyou Xu (Zhejiang University), Yi Yang (Zhejiang University)
Object TrackingSegmentationTransformerImageVideoBenchmark
🎯 What it does: This paper studies semi-supervised video object segmentation in panoramic outdoor scenes, proposing the VIPOSeg benchmark dataset and the PAOT model.
Violin: Virtual Overbridge Linking for Enhancing Semi-supervised Learning on Graphs with Limited Labels
Siyue Xie (Chinese University of Hong Kong), Wing Cheong Lau (Chinese University of Hong Kong)
ClassificationRepresentation LearningData-Centric LearningGraph Neural NetworkGraph
🎯 What it does: Designed the Violin framework, which utilizes learning-generated virtual bridges (VOs) to associate unlabelled nodes with labelled nodes, thereby expanding the receptive field of GNNs and enhancing their representation capability.
Vision Language Navigation with Knowledge-driven Environmental Dreamer
Fengda Zhu (Monash University), Xiaodan Liang (Sun Yat-sen University)
GenerationData SynthesisVision Language ModelAuto EncoderGenerative Adversarial NetworkImageTextBenchmark
🎯 What it does: Propose a knowledge-driven environment dream model (KED) that generates unseen house scenes with consistent texture and structure without increasing annotations, and uses it to augment vision-language navigation (VLN) data;
ViT-CX: Causal Explanation of Vision Transformers
Weiyan Xie (Hong Kong University of Science and Technology), Nevin L. Zhang (Hong Kong University of Science and Technology)
ClassificationExplainability and InterpretabilityTransformerImageBenchmark
🎯 What it does: This paper proposes a novel ViT-CX method for explaining Vision Transformer (ViT) predictions in image classification.
ViT-P3DE∗: Vision Transformer Based Multi-Camera Instance Association with Pseudo 3D Position Embeddings
Minseok Seo (Seoul National University), Xuan Truong Nguyen (Seoul National University)
Object TrackingTransformerImage
🎯 What it does: Studied a multi-camera instance association framework based on Vision Transformer, and proposed two enhancement methods: learnable pseudo 3D position embedding (P3DE) and joint patch generation (JPG).
Voice Guard: Protecting Voice Privacy with Strong and Imperceptible Adversarial Perturbation in the Time Domain
Jingyang Li (Wuhan University), Shengshan Hu (Huazhong University of Science and Technology)
Safty and PrivacyAudio
🎯 What it does: Propose Voice Guard, which actively defends against zero-shot voice conversion (VC) attacks by applying adversarial perturbations in the time domain to protect speaker privacy;
VS-Boost: Boosting Visual-Semantic Association for Generalized Zero-Shot Learning
Xiaofan Li (Xiamen University), Jianping Fan (Lenovo Research)
ClassificationConvolutional Neural NetworkVision Language ModelGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: Proposed the VS-Boost method, combining visual-semantic associations to enhance generalization in zero-shot learning.
WBFlow: Few-shot White Balance for sRGB Images via Reversible Neural Flows
Chunxiao Li (Beijing University of Posts and Telecommunications), Anlong Ming (Beijing University of Posts and Telecommunications)
RestorationConvolutional Neural NetworkFlow-based ModelImage
🎯 What it does: Developed a reversible neural flow network called WBFlow for white balancing sRGB images, achieving fast generalization across multiple cameras through few-shot learning.
Why Rumors Spread Fast in Social Networks, and How to Stop It
Ahad N. Zehmakan (Australian National University), Sajjad Hesamipour Khelejan (Trinity College Dublin)
Anomaly DetectionGraph
🎯 What it does: This paper proposes a rumor propagation model integrating IC, Push-Pull, and SIR, incorporating trust weights based on Jaccard similarity and a time-decaying forgetting mechanism. It analyzes the impact of graph structure (expansiveness and community connectivity) on propagation speed and range from both theoretical and experimental perspectives, while designing and evaluating six countermeasures to suppress rumors.
WiCo: Win-win Cooperation of Bottom-up and Top-down Referring Image Segmentation
Zesen Cheng (Peking University), Jie Chen (Peking University)
SegmentationTransformerVision Language ModelScore-based ModelMultimodality
🎯 What it does: This paper proposes the WiCo framework, which combines bottom-up and top-down reference image segmentation methods, achieving performance improvement through feature interaction and Gaussian scoring integration.
XFormer: Fast and Accurate Monocular 3D Body Capture
Lihui Qian (Huya Inc), Cheng-Bin Jin (Huya Inc)
Pose EstimationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerImageGraph
🎯 What it does: Proposes XFormer, a monocular real-time 3D human shape capture framework based on a dual-branch architecture (keypoint branch and image branch) and cross-modal Transformer.