Most models of categorization describe categorization in two steps. First, a distance between an object and a category is extracted from a psychological similarity space. Secondly, computations on these distances determine the probability of assigning a new object to a category. In these models, categorization depends on the similarity between stimulus representations. This study examines the reverse influence of categorization on the representation of stimuli. New, artificial 2-D shapes (which looked a bit like chromosomes) were created by combining two separable shape dimensions: aspect ratio and curvature. In the first experiment, subjects completed a categorization task of one hour, based on a one-dimensional criterion (e.g. curvature as relevant) and ignoring the variability on the second dimension (e.g. aspect ratio as irrelevant). A subsequent same-different task showed improved discriminability for the relevant shape dimension (e.g. d’ increase from 2 to 3). Additional experiments will investigate how this learning effect for aspect ratio or curvature generalizes towards other stimulus sets and how a 45°-rotation for the relevant and the irrelevant dimension in the same stimulus space (making the dimensions integral instead of separable) will affect this learning effect.