Abstract
With a better understanding of the loss surfaces for multilayer networks, wecan build more robust and accurate training procedures. Recently it wasdiscovered that independently trained SGD solutions can be connected alongone-dimensional paths of near-constant training loss. In this paper, we showthat there are mode-connecting simplicial complexes that form multi-dimensionalmanifolds of low loss, connecting many independently trained models. Inspiredby this discovery, we show how to efficiently build simplicial complexes forfast ensembling, outperforming independently trained deep ensembles inaccuracy, calibration, and robustness to dataset shift. Notably, our approachonly requires a few training epochs to discover a low-loss simplex, startingfrom a pre-trained solution. Code is available athttps://github.com/g-benton/loss-surface-simplexes.