Robust and Fast Training via Per-Sample Clipping

Davide Nobile, Philipp Grohs

We propose a robust gradient estimator based on per-sample gradient clipping and analyze its properties both theoretically and empirically. We show that the resulting method, per-s…