Global Convergence of Sobolev Training for Overparameterized Neural Networks

Jorio Cocola, Paul Hand

Sobolev loss is used when training a network to approximate the values and derivatives of a target function at a prescribed set of input points. Recent works have demonstrated its…