Embedded Ensembles: infinite width limit and operating regimes

摘要

A memory efficient approach to ensembling neural networks is to share most weights among the ensembled models by means of a single reference network. We refer to this strategy as