Learning Bregman Divergences with Application to Robustness