ABSTRACT
Introduction
Two factors may guide concept learning: catpuring predictive structure in input, and forming groupings which can be categorized by a simple criteria. Experiments from unsupervised learning suggest that there is a strong bias to sort things based on their similarity, similarity being property covariation. These sorts provide maximum predictive utility (Billman & Knutson, 1996; Anderson, 1991). Work on free sorting of examples has found that people frequently sort based on the values of a single dimension (Ahn & Medin, 1992). We find evidence that people will use both biases under different conditions.
