HAMILTON-JACOBI PDE METHODS TO ACCELERATE TRAINING IN DEEP LEARNING
Abstract
A major challenge for the deep learning research community is to better understand generalization. Generalization means that the trained DNN performs effectively on relevant data outside the training set. Generalization as currently understood, is not mathematically tractable, since it involves a combination of data, network architecture, and regularization, along with training methods. Our approach facilitates generalization by widening the local minima at various scales via PDE methods, in particular the Hamilton-Jacobi (HJ) equation. This reduces classification error (trapping in non-global minima) and speeds-up convergence.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 30, 2018
- Source ID
- FA95501810167
Entities
People
- Stanley Osher
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California, Los Angeles