Cut your Losses with Squentropy

Like Hui, Mikhail Belkin, Stephen Wright

Nearly all practical neural models for classification are trained using cross-entropy loss. Yet this ubiquitous choice is supported by little theoretical or empirical evidence. Rec…