Nonsmooth differential calculus for robustness in Machine Learning
Abstract
Deep learning and AI s rely on the use of automatic differentiation which can be considered as a clever way to orchestrate elementary learning mechanisms. In mathematical terms the recent theory of conservative gradients allowed to rigorously model autodiff algorithms as implemented on modern machine learning computer libraries. The goal of this proposal is to deepen this theory further to derive sensitivity and robustness tools for deep learning.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 20, 2023
- Source ID
- FA86552217012
Entities
People
- Jerome Bolte
Organizations
- Air Force Office of Scientific Research
- United States Air Force