The relevance of labels in semi-supervised learning depends on category structure

Wai Keen VongUniversity of Adelaide
Amy PerforsUniversity of Adelaide
Daniel NavarroUniversity of Adelaide

Abstract

The study of semi-supervised category learning has shown mixed results on how people jointly use labeled and unlabeled information when learning categories. Here we investigate the possibility that people are sensitive to the value of both labeled and unlabeled items, and that this depends on the structure of the underlying categories. We use an unconstrained free-sorting categorization experiment with a mixture of both labeled and unlabeled stimuli. The results showed that when the distribution of stimuli involved distinct clusters, participants preferred to use the same strategies to sort the stimuli regardless of whether they were given any additional category label information. However, when the stimuli distribution was ambiguous, the sorting strategies people used were strongly influenced by the labeled information given. We capture performance in both cases with an extension to Anderson’s Rational Model that does not know the exact number of category labels in advance.

Files

The relevance of labels in semi-supervised learning depends on category structure (120 KB)



Back to Table of Contents