Graham Harper
Sandia National Laboratories
Biography
Graham Harper is a staff member at Sandia National Laboratories in the scientific machine learning department. He has expertise in finite element methods, domain decomposition, linear solvers, multigrid, earth system modeling, and machine learning. He has developed high-performance software for several libraries, including Trilinos, deal.II, and Neko. In his free time, Graham enjoys 3D printing, electronics repair, and gardening.
SRP Project Title
Multilevel Machine Learning for Material Modeling
Topical Areas
Applied Computer Science; Applied Mathematics; High Performance Computing; Materials Engineering
Abstract
Multilevel methods are a class of algorithms which accelerate solutions of large problems by solving a sequence of smaller problems. Recent developments have shown multilevel methods may be used to train large neural networks by instead training a sequence of smaller networks. The benefits of these approaches are derived from the comparatively lower costs of training smaller networks. This is particularly effective for Kolmogorov-Arnold Networks (KANs), which are a type of neural network which learn individual activation functions. This also allows for development of mathematical theory related to polynomial approximation of functions. These theories and algorithms may be used to accelerate training of neural networks for a variety of different applications, such as material modeling engineering applications.
Desired Skills
Mathematics theory Machine learning software and theory Computer science and programming Engineering and materials
Lightning Talk Title
Multilevel Training for Machine Learning
Keywords
Mathematics; Multilevel Methods; Machine Learning; Neural Networks; Kolmogorov-Arnold Networks; Computer Science; Material Modeling; Engineering;