Relevance of a-posteriori probability errors in bayesian classification.