Scalable Multilevel Training of Large Neural Networks

Colin Ponce | 19-ERD-019

Executive Summary

To address a common data flow problem in the field of machine learning, we will develop multilevel methods for increasing the efficiency of neural network training processes that readily scale up to accommodate large-scale problems. Applicable across many domains, this research supports scientists and analysts in fields that employ neural networks to analyze simulation results (nuclear stockpile stewardship, high-energy-density physics) and multimodal data (counterterrorism, bioscience) as a basis for their predictions.