One of these greebles is not like the others: Semi-supervised models for similarity structures

Abstract

When studying human concepts, artificial categories are often used that are very simple, with a chosen set of discrete features. Images such as “greebles”, animal-like three-dimensional figures, are a more realistic and interesting alternative artificial stimulus set for categorization experiments. However, especially for more complex stimuli, it is not obvious that participants’ perceived similarity structure will be as simple as the structure used to construct the stimuli. Therefore, it is safest to first empirically obtain a measure of the stimuli’s structure. We demonstrate the use of an odd one out task in order to model people’s mental representations of a set of greebles, an example set of moderately complex stimuli. We show that, although the greebles were constructed such that they can be classified into two sets of categories, these categories have a considerable amount of additional latent structure that can be extracted by similarity modeling.