Measuring Stiffness in Residual Neural Networks

Abstract

In this work, we define the concept of stiffness for residual neural networks (ResNets) relying on the fact that ResNets can be viewed as a discretization of an underlying neural ordinary differential equation (NODE). We then propose several metrics for the stiffness of a ResNet. We compare these measures numerically by examining their evolution over the course of training a ResNet on several test problems. We find that stiffness tends to increase as a result of training, and suggest the developed stiffness metrics can be used as training penalties, providing a novel means of regularization for ResNets.

Publication
RAMSES: Reduced order models; Approximation theory; Machine learning; Surrogates, Emulators and Simulators

Related