报告人
陈宗昊
University College London
时间
2026年1月13日 星期二
下午 14:00-15:00
地点
602会议室
Abstract
Optimization of neural networks is notoriously challenging due to the highly non-convex nature of their parameter landscapes. In recent years, a mean-field perspective has emerged as a powerful framework for understanding neural network training in the large-width limit, where the training of network parameters can be described by a gradient flow over probability measures. This talk will first review recent developments in the mean-field analysis of neural network optimization for single-stage learning problems, such as standard supervised learning.
I will then present my recent work on extending the mean-field framework to multi-stage learning problems, with a particular focus on bilevel optimization in nonparametric instrumental variable (NPIV) regression.
Finally, I will discuss the limitations and open challenges of the mean-field perspective for neural network optimization, and outline several directions for future research.
Biography
陈宗昊目前在英国伦敦大学学院（University College London）基础人工智能中心攻读博士学位，师从 François-Xavier Briol 与 Arthur Gretton。他的研究兴趣聚焦于通过优化与泛化两大视角理解机器学习算法。目前的研究方向包括生成模型、蒙特卡洛方法以及因果推断。在攻读博士之前，他于清华大学电子工程系获得学士学位（2022），并曾获清华大学特等奖学金。
