Learning Neural Network Subspaces

    Mitchell Wortsman 1 Maxwell Horton 2 Carlos Guestrin 2 Ali Farhadi 2 Mohammad Rastegari 2

            Abstract
   Recent observations have advanced our under-
   standing of the neural network optimization land-
   scape, revealing the existence of (1) paths of
   high accuracy containing diverse solutions and
   (2) wider minima offering improved performance.
   Previous methods observing diverse paths require
   multiple training runs. In contrast we aim ...