Metric Flows with Neural Networks

James Halverson, Fabian Ruehle

We develop a general theory of flows in the space of Riemannian metrics induced by neural network gradient descent. This is motivated in part by recent advances in approximating Ca…