Irresistibly Attractive Fruitless Feature Dimensions

Although selective attention allocation has been suggested to be one of the most important processes implemented in the recent computational models of category learning (e.g., Kruschke, 1992), the models’ predictions on attention allocation have been virtually ignored by cognitive modeling researchers. Rather almost all modeling studies had focused solely on the models’ capabilities in reproducing observed learning curves or classification response patterns, and thus the models’ descriptive validities on their selective attention allocation processes have remained still mostly untested. Though they did not directly evaluate the descriptive validities of attention processes in computational models of category learning, Matsuka, Corter, and Markman (2004) conducted a set of simulation studies indicating that information on predicted patterns of attention allocation can be informative in differentiating models. In that study, three recent models of category learning, including ALCOVE (Kruschke, 1992) and SUSATIN (Love & Medin, 1998, Love, Medin, Gureckis, 2004), were compared in multiple aspects, including predicted attention allocation patterns. Matsuka et al. (2004) found that the models perform comparably in reproducing the observed classification response profiles but gave markedly different predictions on selective attention allocation patterns. The results of the simulation studies also indicated the models have different tendencies in their attention allocation patterns. ALCOVE has a tendency to be attracted to a feature dimension that consists of unique elements which help differentiate each exemplar. SUSTAIN on the other hand, has a tendency to be attracted to a feature dimension whose elements are less diverse or more homogeneous. If these tendencies are generalizable, then their descriptive validities may become questionable in some cases. For example, in order to categorize humans into either male or female, ALCOVE would pay good amount of attention to each person’s finger print (i.e., feature dimension consists of unique elements), and SUSTAIN would pay good amount of attention to the fact the they are humans (i.e., feature dimension consist of a constant element). These tendencies, however, in less extreme cases, might be realistic tendencies in humans’ “irrational” cognitive processes. In any case, we first should examine the models’ mechanisms that are associated with such tendencies in order to evaluate the models’ descriptive validities. Specifically, in the present study, we extend Matsuka et al. (2004) to investigate those tendencies in attention allocation associated with two prominent models of category learning, namely ALOCE (Kruschke, 1992) and SUSTAIN (Love et al, 1998, 2004) with a simulation study using an artificial stimulus set.