SGD with Large Step Sizes Learns Sparse Features.

Maksym Andriushchenko, Aditya Vardhan Varre, Loucas Pillaud-Vivien, Nicolas Flammarion

Browse the full ICML paper archive.