Prof. Jianguo Zhang, Southern University of Science and Technology, China
Biography: Jianguo Zhang is currently a Professor in Department of Computer Science and Engineering, Southern University of Science and Technology. Previously, he was a Reader in Computing, the Head of Internationalisation with School of Science and Engineering, University of Dundee, UK; a lecturer with Electronics, Electrical Engineering and Computer Science at Queen’s University Belfast, UK; a researcher with Department of Computer Science at Queen Mary University of London, the Lear group of INRIA Rhône-Alpes in France, School of Electrical and Electronic Engineering at Nanyang Technological University of Singapore. He received a PhD in National Lab of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China, 2002. He has considerable experience in machine learning and image processing with sustained interests in multimodal visual recognition, and medical image analysis. His team maintains a strong record of winning international challenges of visual recognition in both natural and medical images. He has published widely in top journal and conferences including IEEE Trans. On Pattern Analysis and Machine Intelligence, International Journal of Computer Vision, Medical Image Analysis, IEEE Trans. Image Processing, CVPR, ICCV, AAAI, IJCAI, ICML, NeurIPS, ECCV, MICCAI etc. His first-author most-cited paper has received over 2.7K citations (Google Scholar), with an overall citation of 12.9K+. He was the general chair of the Chinese Conference on Pattern Recognition and Computer Vision (PRCV) 2022.
Speech Title: Image Semantic Segmentation with Missing Annotations in Different Scenarios
Abstract: Models of Deep Neural Networks usually require a large amount of annotated data. However, acquiring such annotations is tedious and expensive, especially for image semantic segmentation, where annotations at pixel-level are required. Therefore, how to build/adapt a high-performing deep model with limited/missing annotations during training or testing (inference) is challenging in semantic segmentation. In this talk, I will introduce our recent work on semantic segmentations with missing annotations in different scenarios. 1) for unsupervised image semantic segmentation (UISS), we propose the core properties of how to build a pixelwise UISS model, in which we present a connection and comparison between UISS and image-wise representation learning. Based on this, we proposed a robust network called Semantic Attention Network (SAN), in which a new module Semantic Attention (SEAT) is proposed to generate pixel-wise and semantic features dynamically. Experimental results on multiple semantic segmentation benchmarks show that our unsupervised segmentation framework specializes in catching semantic representations, which outperforms all the existing unpretrained and even several pretrained methods tested. 2) For test-time adaptation, we introduce a new perspective from the view of feature alignment and uniformity. A test-time self-distillation strategy is designed to ensure the uniformity and a memorized spatial local clustering strategy is introduced for the test-time feature alignment. Results show the efficacy of our method in different tasks.
Prof. Lifang Wu, Beijing University of Technology, China
Biography: 毋立芳，教授，博士生导师，北京工业大学信息学部智能媒体计算研究所所长。1991年、1994年、2003年分别于北京工业大学获得工学学士、工学硕士、工学博士学位。1994年至今在北京工业大学工作，2005年9月-12月英国伯明翰大学访问学者，2009年10月-2010年5月美国纽约州立大学布法罗分校访问学者。近年来承担科技部重点专项课题、国家自然科学基金、北京市科技计划项目等20余项，在IEEETAFFC、IEEE TCSVT、PR等发表论文100余篇，获授权发明专利40余项，获PRCV2021 最佳论文提名奖，ICCV2021 HTCVW 最佳论文。获北京市科学奖励2项，相关学会的科学技术奖励3项，入选2022年度首都最美巾帼奋斗者，2020年中国电子学会优秀科技工作者称号。中国图象图形学学会（CSIG）理事，中国计算机学会（CCF）杰出会员，CCF计算机视觉专委会常务委员兼副秘书长，CSIG视觉大数据专委会常务委员。《信号处理》、《中国科技论文信息卷》、《中国图象图形学报》编委，参与组织PRCV2019、NCIG2020、ChinaMM2020、PRCV2021、PRCV2022、ChinaMM2023等。
Speech Title: Knowledge Aided Relation Inference for Group Activity Recognition
Abstract: Group activity recognition involves classifying diverse individual actions and constructing complex relations. Most existing methods utilize only the appearance features, and failed to recognize the potential benefits of introducing knowledge to assist the task. Is there what kind of knowledge in group activity recognition? How to use such knowledge? In this talk, I will Introduce our two related works. The first one tries to generates activity-specific features for different group activities through the introducing of labeled semantic information. it is effective to weakly supervised group activity recognition task. And the second one explores concretizing the knowledge into the action distribution in different group activities and enhancing relation inference for GAR using the concretized knowledge.
Prof. Ning Zhong, Maebashi Institute of Technology, Japan
Biography: Ning Zhong received the Ph.D. degree from the University of Tokyo. He currently holds positions as the chairman of the Web Intelligence Consortium (WIC, wi-consortium.org), professor emeritus and visiting professor, and previously served as a professor in the Department of Life Science and Informatics at Maebashi Institute of Technology, Japan. Dr. Zhong's research interests focus on Web Intelligence (WI), Brain Informatics (BI), Machine Learning, Data Mining, Intelligent Health Technologies, and Intelligent Systems. In 2000 and 2004, Dr. Zhong and his colleagues introduced WI and BI as new research directions, respectively. By conducting interdisciplinary research and driving industrial innovation in the fusion of web intelligence and brain informatics, Dr. Zhong has pioneered the frontier research domain of brain-machine intelligence based on brain big data. Dr. Zhong serves as the founding editor-in-chief of the Web Intelligence journal (IOS Press), the editor-in-chief of the Brain Informatics journal (Springer Nature). Dr. Zhong is a foreign fellow of the Engineering Academy of Japan (EAJ).
Speech Title: Web Intelligence (WI3.0) Empowering the Future Intelligent Society
Abstract: This report reviews the development of Web Intelligence (WI3.0), introduces the current wave of generative AI sparked by ChatGPT and its implications, and outlines the characteristics and insights of Japan's version of Society 5.0. It looks ahead to the unprecedented new opportunities and challenges that the development of a human-centered intelligent society will bring, revealing the novel technologies and models of intelligent services in the connected world driven by the highly integrated synergy of digital and physical spaces, the convergence of human-machine-things interactions, and the emergence of innovative industries and new business formats in the context of a service-oriented society.
Prof. Aijun An, York University, Canada
Biography: Aijun An Department of Electrical Engineering and Computer Science, York University, Toronto, ON M3J 1P3, Canada (aan@cse.yorku.ca). Dr. An received a Ph.D. degree in computer science from the University of Regina, Regina, SK, Canada. She is a Professor with the Department of Electrical Engineering and Computer Science, York University, Toronto, ON, Canada. She has authored and coauthored extensively in various well-respected journals and conferences on data mining, databases, machine learning, and NLP.
Speech Title: Question Generation from Documents: From Rule-Based Approaches to the Use of LLMs
Abstract: Question Generation (QG) from text has increasingly gained interest due to its utility in various applications, such as educational reading comprehension assessments, data augmentation for training question-answering systems, and response generation in conversational systems. In collaboration with an industry partner, we have been exploring automatic question generation from documents for specialized question-answering systems. In this talk, I will present the diverse methods we have utilized and developed for question generation, including rule-based, sequence-to-sequence, semantic role label-based sequence-to-sequence, and LLM-based approaches. I will also discuss experimental results that compare these methods, highlighting the advantages and drawbacks of each. If time permits, I will briefly introduce two related pieces of work on document chunking and image description generation.