Active Learning with Expected Error Reduction