Pattern Learning by Functional-Differential Neural Networks with Arbitrary Path Weights

Abstract

This paper proves the universal theorem on associative learning that culminates my 1967–1972 articles on this subject. The theorem is universal in the following sense. It says that if my associative learning laws were invented at a prescribed time during the evolutionary process, then they could be used to guarantee unbiased associative learning in essentially any later evolutionary specialization. That is, the laws are capable of learning arbitrary spatial patterns in arbitrarily many, simultaneously active sampling channels that are activated by arbitrary continuous data preprocessing in an essentially arbitrary anatomy. The learning of arbitrary space-time patterns is also guaranteed given modest requirements on the temporal regularity of stimulus sampling, as in avalanches and generalizations thereof.