Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs

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 a simple polygonal chain with only one bend, 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 CIFAR10 and CIFAR-100, using state-of-the-art deep residual networks. On ImageNet we improve the top-1 error-rate of a pre-trained ResNet by 0.56% by running FGE for just 5 epochs.