KOOPMAN OPERATOR THEORETIC METHODS FOR EFFICIENT TRAINING AND ANALYSIS OF DEEP NEURAL NETWORKS

Abstract

Machine learning, particularly deep neural networks (DNNs), have become a ubiquitous tool for researchers across many scientific disciplines. The past decade has seen a tremendous increase in the capability of DNNs, as models have become increasingly large and intricate. However, such improvements have come at the cost of requiring significant computational resources and time, as training these DNNs is highly complicated. The very nature of their complexity makes standard analytical tools impractical for the design and analysis of DNNs. This proposal leverages Koopman operator theory, a data-driven and mathematically rich dynamical systems framework, to develop novel approaches for understanding the behavior of DNNs during training, and to make that training and later deployment efficient.

Document Details

Document Type
DoD Grant Award
Publication Date
Apr 20, 2023
Source ID
FA95502210531

Entities

People

  • Maria Fonoberova

Organizations

  • Air Force Office of Scientific Research
  • United States Air Force

Tags

Fields of Study

  • Computer science

Readers

  • Neural Network Machine Learning.
  • Systems Analysis and Design
  • Wave Propagation and Nonlinear Chaotic Dynamics.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks