Generalization Gradients for Categorization of Polygon Patterns

Abstract

Two polygons at each of 3 levels of association value (AV) were used to produce 72 variants with 2 different units of perturbation (dissimilarity to the original) and a varying number of perturbations (1 to 6) per form. More correct categorizations were observed for patterns with the smaller unit of perturbation and also for those with the highest level of AV (p < .01). Generalization gradients evaluating the frequency of correct categorizations against an increasing number of perturbations were sharpest for the larger unit of distortion (p < .01), and no systematic differences in gradient due to association value were evident. Results were interpreted as stressing the importance of association value and constraint metrics for categorization tasks.