We have a new preprint available on arXiv 1908.11843: Partitioned integrators for thermodynamic parameterization of neural networks

In the paper we propose a new way to train (parameterize) neural networks for classification problems. We employ “sampling” algorithms, which rely on discretized stochastic differential equations (SDEs), to train neural networks and show that these “thermodynamic parameterization” methods can be faster, robuster and more accurate than standard optimizers, such as SGD and ADAM, for classification problems with complicated loss landscapes.

This is joint work with Ben Leimkuhler and Charlie Matthews.