A Three-Layer Configural Cue Model of Category Learning Rates

Abstract

The relative difficulty of six different category structures revealed in Shepard, Hovland, and lenkins’ 1961 classic category learning paradigm [18] is predicted using simple channel capacity calculations. This approach is subsequently used to inform the design of a three-layer connectionist network based on Gluck and Bower’s configural cue model of category learning [3]. The extra layer of nodes in this model consists of intermediate or bottleneck nodes which lie between each spatial group of nodes, representing particular cues and cue configurations, and each category label node. The weights in these nodes learn to approximate the correlation between the output of the’ space’ and the target output. The model, using two free parameters, shows a superior fit to the human data than the configural cue model and its variants evaluated by Nosofsky, Gluck, Palmeri, McKinley, and Glauthier [16] in their replication of the Shepard et al. experiment [18]. The reason for this and the applicability of the model to other category learning paradigms is discussed.