Pattern classification performance was examined in relation to the configuration of the similarity space yielded by multidimensional scaling. Subjects were asked to discriminate between four schematic face stimuli used as category prototypes and then sort the other stimuli as well as the prototype stimuli into one of four categories as quickly as possible. Each prototype face was located approximately on each dimensional axis in a similarity space having one of the dimensional properties in moderate degree. The reaction time data showed that extreme stimuli for which the dimensional properties are clearer than that of their prototype were classified significantly faster than their prototypes. The results showed that, instead of classifying patterns in terms of their similarity to the prototype, subjects abstracted the dimensional properties which characterize the prototype patterns and sorted each pattern based on the knowledge of these abstracted properties.