Tackling High-Frequency Challenges: From Shallow to Multi-Layer Neural Networks

讲座题目：Tackling High-Frequency Challenges: From Shallow to Multi-Layer Neural Networks
讲座时间 Datetime: 2025年5月30日，周五，15:00 – 16:00
地点 Venue: 腾讯会议456-509-544
主持人 Host：张海樟 教授
报告人 Speaker: 张仕俊 助理教授
单位 Affiliation: 香港理工大学
报告摘要 Abstract:
This talk explores the limitations of shallow neural networks in addressing high-frequency challenges and introduces a novel solution through a multi-layer, multi-component neural network (MMNN) architecture. We demonstrate that shallow networks behave like low-pass filters, struggling with high-frequency components due to machine precision constraints and slow learning dynamics. The MMNN architecture overcomes these obstacles by efficiently decomposing complex functions, achieving substantial improvements in accuracy and computational efficiency. Numerical experiments validate the effectiveness of this approach in capturing fine details in oscillatory functions.
报告人简介：
张仕俊博士现任香港理工大学（PolyU）助理教授（终身教职轨），主要从事逼近理论与神经网络领域的研究，重点关注深度网络的表达能力、高频函数逼近及激活函数设计。其研究成果发表于 Journal of Machine LearningResearch、SIAM Journal on Mathematical Analysis、NeurIPS、ICML 等顶级 期刊与会议，总引用近千次（h-index 10）。