A Bayesian Test for Comparing Classifier Errors