Machine-Precision Neural Networks for Multiscale Dynamics

Explore a detailed technical talk by Stanford University's Assistant Professor Ching-Yao Lai discussing breakthrough developments in deep learning precision for scientific applications. Learn how multi-stage neural networks can achieve machine-precision accuracy through an innovative training process that divides learning into distinct stages, with each subsequent network optimizing residual errors from previous stages. Discover how this approach effectively addresses the longstanding challenge of prediction error reduction in neural networks, particularly for regression problems and physics-informed neural networks, enabling near double-floating point precision within finite iterations. Gain insights into overcoming spectral bias in multiscale problems and understand the practical implications for scientific computing applications. The presentation draws from Dr. Lai's extensive research background in physics, mechanical engineering, and geophysics, supported by her recent recognitions including the 2023 Google Research Scholar Award and 2024 Sloan Research Fellowship.