Extending a classic categorization study by Posner and Keele (1968), participants learned categories created from distortions of familiar, recognizable dot patterns, and then made pairwise similarity judgments of the patterns. Multidimensional scaling analyses revealed that the category "prototypes" were better characterized as ideal points, extremes in the psychological space, rather than as central tendencies. An exemplar-based model, the generalized context model (Nosofsky, 1986), accounted for categorization data both when prototypes were presented during training and when they were not: Apparently, even under conditions in which the prototypes were familiar and recognizable, people still seemed to rely on exemplar information during categorization. In addition, ideal point representations were found for novel prototype patterns displaying a symmetry structure in a second experiment. Implications of these results for the nature of category representations and category learning processes were discussed.