Addressing Covariate Shift in Active Learning with Adversarial Prediction

Active learning approaches used in practice are generally optimistic about their certainty with respect to data shift between labeled and unlabeled data. They assume that unknown datapoint labels follow the inductive biases of the active learner. As a result, the most useful datapoint labels— ones that refute current inductive biases—are rarely solicited. We propose an adversarial approach to active learning that assumes the worstcase about the unknown conditional label distribution under covariate shift. This closely aligns model uncertainty and expected error with generalization error, enabling more useful label solicitation. We investigate the benefits of this approach on classification tasks.