n two simulation studies, we compare the attention learning predictions of three well-known adaptive network models of category learning: ALCOVE, RASHNL, and SUSTAIN. The simulation studies use novel stimulus structures designed to explore the effects of predictor diagnosticity and independence, and differentiate the models regarding their tendencies to learn simple rules versus exemplar-based representations for categories. An interesting phenomenon is described in which the models (especially SUSTAIN and RASHNL) learn to attend to a completely nondiagnostic constant dimension.