Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs
arXiv (Cornell University)2018
Citations Over Time
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.
Related Papers
- → Maximal global snapshot with concurrent initiators(2002)33 cited
- → A Data Snapshot Approach for Making Real-Time Predictions in Basketball(2018)9 cited
- → Multi-time Snapshot based Off-Grid DOA Estimation of Sparse Array Antennas using MFOCUSS Algorithm(2020)3 cited
- → Two-dimension Direction Finding Using Single-snapshot Data(2006)1 cited