Learning Stochastic Majority Votes by Minimizing a PAC-Bayes Generalization Bound