Learning Theory of Deep Neural Networks
Sep 09,2021 Activity
报告主题：
Learning Theory of Deep Neural Networks
报告嘉宾：
林绍波教授（西安交通大学）
报告时间：
9月9日（星期四）10:00-11:00
参与方式：
加入 Zoom 会议
https://cuhk-edu-cn.zoom.com.cn/j/99970274588?pwd=Tk0wUnlwSk1LWWlSWXo5UUQzWTRuQT09
会议号：999 7027 4588
密码：681114
报告简介:
Deep learning has attracted enormous research activities in the past decade. In this talk, we focus on theoretically demonstrating the power of depth, the role of massive data and the existence of a perfect global minima of ERM on deep ReLU nets. In particular, we derive optimal error bounds for running ERM on deep ReLU nets for numerous types of data. Our main approach is a novel deepening technique that constructs exact interpolation networks for any deep nets without affecting their generalization capability.
