Biography: Dr Tianrui Li is a Professor, Dean of School of Computing and Artificial Intelligenc, and the Director of the Key Lab of Cloud Computing and Intelligent Technique of Sichuan Province, Southwest Jiaotong University, China. Since 2000, he has co-edited 10 books, 12 special issues of international journals, received 36 Chinese invention patents and published over 500 research papers with more than 32,000 citations in refereed journals (e.g., AI, IJCV, IEEE TPAMI, IEEE TKDE, IEEE TIFS, IEEE TPDS) and conferences (e.g., AAAI, ACL, CVPR, ICCV, ICDE, ICML, IJCAI, KDD, UbiComp, WWW). 7 papers were ESI Hot Papers and 26 papers were ESI Highly Cited Papers. He has been ranked among the world’s top 2% scientists and honored as a China Highly Cited Researcher. He serves as Editor-in-Chief of Human-Centric Intelligent Systems, Editor of Information Fusion, Associate Editor of ACM Transactions on Intelligent Systems and Technology, etc. He is the Fellow and Vice-President of IRSS, and the Secretary of ACM SIGKDD China Chapter.

Speech Title: Hierarchical Learning for 3D Visual Intelligence

Abstract: Human cognition naturally organizes visual information in hierarchical and multi-granular forms, providing inspiration for advancing 3D visual intelligence. To bridge the gap between semantic hierarchies and spatial multi-scale structures in visual understanding, this report systematically introduces a novel framework of ‘Hierarchical Learning’. The framework integrates three components: semantic modeling guided by cognitive semantic hierarchies, representation mechanisms driven by geometrical hierarchical structural priors, and perceptual stability ensured by hierarchical robust sampling - together forming a new framework for 3D visual understanding that aligns more closely with human multi-granular cognition. We are the first to establish a theoretical link between semantic consistency across hierarchical semantic layers and entropy maximization within each layer. Based on it, a novel Hierarchical Embedding Fusion Module (HEFM) and a hierarchical regularization term are proposed. To support datasets without pre-defined hierarchies, we propose a new label class sematic hierarchy generation algorithm based on vision-language models and design a quality evaluation metric. Finally, within this framework, to address real-world 3D point cloud corruptions, we also introduce a robust local-global balanced point cloud sampling protocol. Experimental results show that, compared to state-of-the-art methods, our approach significantly improves the performance of 3D semantic segmentation and robust classification, demonstrating strong potential in applications such as autonomous driving, urban planning, and digital twins.

Prof. Yang Chen

Southeast University, China