报告人：许超 教授（浙江大学）

报告时间：2026年11月06日 16：05-17：00

报告地点：数学楼301

摘要： Deep learning excels at function approximation and outperforms conventional methods in learning-based control tasks such as robot planning and complex physical system modeling, which require solving large-scale optimization problems. Though gradient optimizers like Adam train neural networks effectively, their generalization and convergence dynamics are not fully understood. This talk treats deep learning optimization as continuous-time dynamical systems and leverages control and physical theories to analyze and design optimizers. It interprets momentum optimizers as particles moving on loss landscapes and explains core learning properties including approximation, convergence and generalization from a dynamical perspective. This work offers new insights to develop more efficient optimizers. Experimental results on fluid-driven robotic locomotion and autonomous robotics applications are also presented.

报告人简介： 许超，博士、教授、博导，浙江大学湖州研究院院长，FAST实验室创始人，PIE实验室（π引擎）创始人，入选浙江省“万人计划”领军人才，荣获2024年国际基础科学大会（ICBS）前沿科学奖。从事“AI控制论与机器人学”交叉领域研究，在《Science Robotics》《Nature Machine Intelligence》《Nature Communications》《Nature Biomedical Engineering》等刊发表系列成果，入选爱思唯尔全球前2%顶尖科学家榜单；创办期刊IET Cyber-Systems & Robotics（《IET·智能系统与机器人》，ESCI检索）任执行总主编，并担任2025年IROS大会组织主席。牵头研发的集群机器人入选“奋进新时代”国家成就展并被国家博物馆永久收藏，多项科研成果被央视《走进重点实验室》栏目集中报道；相关成果被科技日报、光明日报、央视2套、浙江省新闻联播、泰晤士报等国内外媒体多次报导。