A Bayesian Expected Error Reduction Approach To Active Learning