报告题目:Learning with norm-based neural networks: model capacity and computational-statistical gaps
报 告 人:刘方辉 博士
报告人所在单位:University of Warwick, UK
报告日期:2024年9月4日
报告时间:14:30-15:30
报告地点:光华楼东主楼 2001
报告摘要:
In this talk, I will discuss some fundamental questions in modern machine learning:
-What is the suitable model capacity of over-parameterized models?
-What is the suitable function space for feature learning?
-Which function can be learned by two-layer neural networks, statistical and/or computational efficiently?
-What is the computational-statistical gap behind this?
My talk will partly answer the above questions, both theoretically and empirically.