Item

Abstract

This thesis focusses on characterising how unsupervised training affects learning in humans and models in the context of simple semi-supervised category learning tasks. Despite a mixture of supervised and unsupervised learning in the real-world, the literature on semi-supervised learning is surprisingly recent with conflicting results about whether unsupervised exposure benefits learning. We contribute to this literature by providing novel evidence that unsupervised training can help or hurt performance depending on the order in which different sorts and qualities of information are provided and on the alignment between the learner’s internal representations with the true category structure, and by providing insights into computational accounts thereof. We first employ a machine teaching approach to analyse the operation of three existing semi-supervised models in detail, providing new insights into their workings and revealing important directions for improvement. We then extract these models’ predictions about optimal semi-supervised sequences, and thereby test them against empirical data that we collect to assess whether and how the benefit of unsupervised training depends on the order of unsupervised trials. Subjects performed better at test when trained on a hard-to-easy unsupervised training schedule than on an easy-to-hard schedule, posing problems for all the models. We next argue that the benefit of unsupervised training should depend on the alignment between the experimenter-defined task and the internal representations of the learner. Finally, a second empirical study supports this hypothesis, showing that unsupervised training improves performance only when task and internal representations are well-aligned, but diminishes performance when misaligned. Since internal representations are not routinely measured, this result also provides a possible explanation as to why the literature on semi-supervised learning has previously reported conflicting evidence. In sum, this work provides new insights into the subtle effects of unsupervised category learning in both humans and existing models of semi-supervised category learning.