Explore a 49-minute lecture where Ching-Yao Lai discusses the development of multi-stage neural networks that can achieve machine precision in scientific applications. Learn how these networks divide training into stages, with each new network optimized to fit residues from previous stages, allowing prediction errors to approach double-floating point precision within finite iterations. Discover how this breakthrough addresses longstanding accuracy limitations in neural network training and tackles spectral bias in multiscale problems, including applications in finding blow-up solutions in fluid dynamics. This presentation was part of the Simons Collaboration on Wave Turbulence Annual Meeting 2024.
Machine-Precision Neural Networks for Multiscale Dynamics
Overview
Syllabus
Ching-Yao Lai: Machine-Precision Neural Networks for Multiscale Dynamics (December 6, 2024)
Taught by
Simons Foundation
Related Courses
-
Machine-Precision Neural Networks for Multiscale Dynamics
-
Low Precision Inference and Training for Deep Neural Networks
-
Stochastic Learning Dynamics and Generalization in Neural Networks
-
Introduction to Neural Networks - Lecture 25
-
Practical Issues with Neural Networks in Machine Learning - Lecture 27
-
Training Overparametrized Neural Networks: Early Alignment Phenomenon and Simplicity Bias
