DISCOVERING PREDICTABLE CLASSIFICATIONS

Prediction problems are among the most common learning problems for ne ural networks (e.g., in the context of time series prediction, control , etc.). With many such problems, however, perfect prediction is inher ently impossible. For such cases we present novel unsupervised systems that learn to classify patterns such that the classifications are pre dictable while still being as specific as possible. The approach can b e related to the IMAX method of Becker and Hinton (1989) and Zemel and Hinton (1991). Experiments include a binary stereo task proposed by B ecker and Hinton, which can be solved more readily by our system.