A recent breakthrough in deep learning theory shows that the training of over-parameterized deep neural networks (DNNs) can be characterized by the neural tangent kernel (NTK). However, existing optimization and generalization guarantees for deep neural networks (DNNs) typically require a network width larger than a high degree polynomial of the training sample size $n$ and the inverse of the target accuracy $epsilon^{-1}$. In this talk, I will discuss how this over-parameterization condition can be improved to more practical settings. Specifically, I will first explain why over-parameterized DNNs can be optimized to zero training error in the NTK regime, and then show what kind of functions can be learnt by DNNs with small test errors. Under standard NTK-type assumptions, these optimization and generalization guarantees hold with network width polylogarithmic in $n$ and $epsilon^{-1}$.
30 Apr 2020
3:00pm - 4:00pm
Where
https://hkust.zoom.com.cn/j/5616960008
Speakers/Performers
Dr. Yuan CAO
University of California at Los Angeles
Organizer(S)
Department of Mathematics
Contact/Enquiries
mathseminar@ust.hk
Payment Details
Audience
Alumni, Faculty and Staff, PG Students, UG Students
Language(s)
English
Other Events
14 Jul 2025
Seminar, Lecture, Talk
IAS / School of Science Joint Lecture - Boron Clusters
Abstract The study of carbon clusters led to the discoveries of fullerenes, carbon nanotubes, and graphene. Are there other elements that can form similar nanostructures? To answer this questio...
15 May 2025
Seminar, Lecture, Talk
IAS / School of Science Joint Lecture - Laser Spectroscopy of Computable Atoms and Molecules with Unprecedented Accuracy
Abstract Precision spectroscopy of the hydrogen atom, a fundamental two-body system, has been instrumental in shaping quantum mechanics. Today, advances in theory and experiment allow us to ext...