Learning to Approximate a Bregman Divergence