Loss Surfaces, Mode Connectivity, And Fast Ensembling Of DNNs

Authors:
Timur Garipov Moscow State University
Pavel Izmailov Cornell University
Dmitrii Podoprikhin XTX markets
Dmitry Vetrov Higher School of Economics, Samsung AI Center, Moscow
Andrew Wilson Cornell University

Introduction:

The loss functions of deep neural networks are complex and their geometric properties are not well understood.The authors introduce a training procedure to discover these high-accuracy pathways between modes.

Abstract:

The loss functions of deep neural networks are complex and their geometric properties are not well understood. We show that the optima of these complex loss functions are in fact connected by simple curves, over which training and test accuracy are nearly constant. We introduce a training procedure to discover these high-accuracy pathways between modes. Inspired by this new geometric insight, we also propose a new ensembling method entitled Fast Geometric Ensembling (FGE). Using FGE we can train high-performing ensembles in the time required to train a single model. We achieve improved performance compared to the recent state-of-the-art Snapshot Ensembles, on CIFAR-10, CIFAR-100, and ImageNet.

You may want to know: