Item

Abstract

Teaching involves a mixture of instruction, self-studying in the absence of a teacher and assessment that provides corrective feedback. Despite the mixture of supervised and unsupervised learning in the real-world, the literature on human learning has traditionally focused on one or the other, or reinforcement learning. By contrast, literature on semi-supervised learning is surprisingly recent, conflicting and sparse. Reports about the benefit of unsupervised information in category learning conflict across experimental designs, leading researchers to conclude that its effects on learning may be minimal at best. Here, we adopt a machine teaching approach to create a targeted test of the effects of unsupervised information in a simple categorization task. Taking two paradigmatic models of semi-supervised category learning (prototype and exemplar models), we infer that the sequential difficulty of the unsupervised items affects learning in the models due to self-reinforcement of beliefs. Critically, the models make opposite predictions as to whether unsupervised items should optimally be ordered from easy-to-hard or hard-to-easy, setting the stage for an empirical test. We find that hard-to-easy ordering leads to better task performance, consistent with the predictions of the prototype model. However, additional analyses show that the model predicts this result for reasons that are not consistent with human data, as it underestimates performance drops in the face of hard items. In sum, our machine teaching approach revealed novel evidence that ordering of unsupervised information affects category learning and highlighted shortcomings of existing semi-supervised categorization models. Future work will help understand semi-supervised learning principles and their connections with results on supervised easy-to-hard schedules and training set idealization. This has the potential to help improve teaching curricula.