All your loss are belong to Bayes Christian Walder 1 2 Richard Nock

In this paper, we rely on a broader view of proper composite losses and a recent construct from information geometry, source functions, whose fitting alleviates constraints faced by canonical links.