Active Learning for Probabilistic Hypotheses Using the Maximum Gibbs Error Criterion