Concept Learning and Flexible Weighting

We previously introduced an exemplar model, named GCM-ISW, that exploits a highly exible weighting scheme. Our simulations showed that it records faster learning rates and higher asymptotic accuracies on several artiicial categorization tasks than models with more limited abilities to warp input spaces. This paper extends our previous work; it describes experimental results that suggest human subjects also invoke such highly exible schemes. In particular, our model provides signiicantly better ts than models with less exibility, and we hypothesize that humans selectively weight attributes depending on an item's location in the input space. We need more exible models of concept learning Many theories of human concept learning posit that concepts are represented by prototypes (Reed, 1972) or exemplars (Medin & Schaaer, 1978). Prototype models represent concepts by the \best example" or \central tendency" of the concept. 1 A new item belongs in a category C if it is relatively similar to C's prototype. Prototype models are relatively innexible; they discard a great deal of information that people use during concept learning (e.g., the number of ex-emplars in a concept (Homa & Cultice, 1984), the variability of features (Fried & Holyoak, 1984), correlations between features (Medin et al., 1982), and the particular exemplars used (Whittlesea, 1987)). Exemplar models instead represent concepts by their individual exemplars; a new item is assigned to 1 Other summary information may also be stored by more advanced prototype models; our concerns primarily target problems with \pure" prototype models. More accurately, we are interested in supporting the learning behavior displayed by the advanced exemplar models described in Section 3 regardless of the models' representation for categories (Barsalou, 1989). a category C if it is relatively similar to C's known exemplars. Exemplar representations are far more exible than prototype representations since they retain sensitivity to all of the information listed above. This exibility often translates to increased catego-rization accuracy. For example, unlike prototype models, humans and exemplar models can learn some non-linearly separable categories as easily as linearly separable categories (Medin & Schwanennugel, 1981). This capability is not limited to at learning archi-tectures; several researchers capture this exibility in radial basis networks (e.g., Kruschke, 1992; Hurwitz, 1991). While existing exemplar models are more exible than prototype models, they are still not suuciently exible. We argue that people represent categories not only with category exemplars, but also with a set of speciic weights associated with each exem-plar's (or set …