The Structure of Perceptual Categories

When presented with a small set of sample objects, human observers have the striking capacity to induce a more general class. Generalization can even proceed from a single object ("one-shot categorization"). The inference is apparently guided by the principle that a good categorical hypothesis is one in which the observed object would be a typcal, "non-accidental," or generic example; this idea is formalized here as the Genericity Constraint. In the theory proposed here, each categorical hypothesis is a "generative model," a sequence of transformations by which the object is interpreted as having been created; objects are considered to be in the same category if they were created by the same set of operations. The set of all available category models can be explicitly enumerated in a lattice, an explicit structure that partially orders the models by their degree of regularity or genericity-more abstract models are higher in the lattice, and more regular or constrained models are lower. The Genericity Constraint dictates that among all the models on the lattice that apply, the observer should choose the one in which the observed object is generic, which is simply the lowest in the partial order. A series of experiments are reported in which subjects are asked to generalize from simple figures. The results corroborate the role of the lattice and the Genericity Constraint in subjects' interpretations