Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs
This addresses the problem of efficiently training ensembles for deep learning practitioners, offering a method that is faster than existing approaches.
The paper shows that optima of deep neural network loss functions are connected by simple curves with nearly constant accuracy, and introduces Fast Geometric Ensembling (FGE) to train high-performing ensembles quickly, achieving improved performance over Snapshot Ensembles on CIFAR-10, CIFAR-100, and ImageNet.
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.