On Separability of Loss Functions, and Revisiting Discriminative Vs Generative Models