Unifying Non-Maximum Likelihood Learning Objectives with Minimum KL Contraction