ABSTRACT: In this article I show how three underlying principles of several models --- including the (generalized) context model (Medin & Schaffer, 1978; Nosofsky 1986) and standard back propagation networks (Rumelhart, Hinton & Williams, 1986) --- have been synthesized into the ALCOVE model (Kruschke, 1992). I illustrate the importance of each principle with data from human category-learning experiments. Whereas models that incorporate the three principles can fit the human data from various experiments reasonably well, several other models that lack one or more of the principles fail to capture human performance.
I discuss the geneology of ALCOVE and its relations to standard back propagation, radial basis function networks, and the generalized context model, among others. Each of the models implements one or more of the principles of error-driven learning, dimensional attention shifts, and quasi-local representation.