What use can the brain make of the massive flow of sensory information that occurs without any associated rewards or punishments? This question is reviewed in the light of connectionist models of unsupervised learning and some older ideas, namely the cognitive maps and working models of Tolman and Craik, and the idea that redundancy is important for understanding perception (Attneave 1954), the physiology of sensory pathways (Barlow 1959), and pattern recognition (Watanabe 1960). It is argued that (1) The redundancy of sensory messages provides the knowledge incorporated in the maps or models. (2) Some of this knowledge can be obtained by observations of mean, variance, and covariance of sensory messages, and perhaps also by a method called “minimum entropy coding.” (3) Such knowledge may be incorporated in a model of “what usually happens” with which incoming messages are automatically compared, enabling unexpected discrepancies to be immediately identified. (4) Knowledge of the sort incorporated into such a filter is a necessary prerequisite of ordinary learning, and a representation whose elements are independent makes it possible to form associations with logical functions of the elements, not just with the elements themselves.

To determine whether a particular sensory event is a reliable predictor of reward or punishment it is necessary to know the prior probability of that event. If the variables of a sensory representation normally occur independently of each other, then it is possible to derive the prior probability of any logical function of the variables from the prior probabilities of the individual variables, without any additional knowledge; hence such a representation enormously enlarges the scope of definable events that can be searched for reliable predictors. Finding a Minimum Entropy Code is a possible method of forming such a representation, and methods for doing this are explored in this paper. The main results are (1) to show how to find such a code when the probabilities of the input states form a geometric progression, as is shown to be nearly true for keyboard characters in normal text; (2) to show how a Minimum Entropy Code can be approximated by repeatedly recoding pairs, triples, etc. of an original 7-bit code for keyboard characters; (3) to prove that in some cases enlarging the capacity of the output channel can lower the entropy.