Shift-pessimistic active learning using robust bias-aware prediction

摘要: Existing approaches to active learning are generally optimistic about their certainty with respect data shift between labeled and unlabeled data. They assume that unknown datapoint labels follow the inductive biases of learner. As a result, most useful data-point labels—ones refute current biases— rarely solicited. We propose shift-pessimistic approach assumes worst-case conditional label distribution. This closely aligns model uncertainty generalization error, enabling more solicitation. investigate theoretical benefits this demonstrate its empirical advantages on probabilistic binary classification tasks.