杨晓松教授学术报告:Understanding the black box of Deep learning from geometric perspective

发布时间:2025-04-24 浏览次数:10

报告题目:Understanding the black box of Deep learning from geometric perspective

报告时间:2025年4月24日14:10-16:30

报告地点:雁山理二608

摘要:Deep learning has firmly established itself as a leading branch of artificial intelligence, delivering outstanding performance across a broad range of applications. Despite this success, deploying deep neural networks often entails a time-consuming search for suitable architectures and hyperparameters—particularly in real-world settings. This talk approaches the problem from a geometric standpoint, aiming to illuminate the inner workings of neural networks—the workhorses of modern AI. After revisiting several foundational mathematical questions, we examine how the geometric structure of data manifolds in input space can inform the design of efficient network architectures. By bridging these two perspectives through topological and differential-geometric ideas--illustrated with concise toy examples--the talk reveals how geometry can guide principled network design. We conclude with tentative directions for future research on deep learning from a geometric perspective.

报告人简介:杨晓松, 自学完成本科教育,1995年进入中国科技大学攻读博士学位,从事微分拓扑和微分几何学习和研究,于1998年获中国科技大学基础数学博士学位。先后在重庆邮电大学,厦门大学信息科学与技术学院工作,控制理论与控制工程博士生导师。2004年起为华中科技大学特聘教授, 曾担任该校控制理论与控制工程博士生导师及电路与系统博士生导师。现为华中科技大学二级教授,基础数学博士生导师;运筹学与控制论博士生导师。曾任第八届中国自动化学会控制理论专业委员会委员、中国工业与应用数学学会理事。2001年获国务院政府特殊津贴。拓扑学和微分几何、动力系统的混沌理论及其应用、行走动力学、生态数学、机器人运动规划等领域做出多项工作,在Chaos, DCDS, IJBC, JMB、 Nonlineariy, Sys.Cont.Lett.,Topol.Appl. 等相关国际著名刊物上发表SCI收录文百余篇,出版专著3部。2023年和2024年入选全球前2%顶尖科学家榜单,并且连续每年入选 Elsevier “中国高被引学者”榜单。