Novel unsupervised classification method

Assume we are given a set of pairs of patterns. We know that both patterns of each pair belong to the same class. We do not know in advance, however, anything about the nature of the classes, which features are characteristic for each class, how many classes there are, and which patterns belong tho which class. We present a novel unsupervised neural system that learns without a teacher to create distributed representations of classes such that patterns belonging to the same class represented by the same activation pattern while patterns belonging to different classes are represented by different activation patterns. The approach can be related to the IMAX method of Hinton, Becker and Zemel (1989, 1991). Experiments include a stereo task proposed by Becker and Hinton, which can be solved more readily by our system.