A Model of Natural Computation based on Recurrent Neural Networks and Reciprocal Images

D.S. Greer (USA)


Natural neural systems, connectionism, and imageunderstanding


A new design that connects together many small feed forward neural networks to form a large, recurrent, image association processor is presented. The resulting recurrent neural network uses local image information in order to form global associations. Just as a single bit of information and its complement hold each other in place in an SR flip-flop, two arbitrary reciprocal images can hold each other in place in an image processor. The new, Ψ-MAP design consists of a pair of two-dimensional processor grids. Each processing element has only locally connections, but in unison, they form an orderly interconnection pattern analogous to the recurrent bit connections in an SR flip-flop. The central nervous system contains many topological maps that mathematically resemble images. Consequently, algorithms that can store and recall image associations are interesting from both a scientific and an engineering point of view. An array of these association processors is similar to the collection of Brodmann areas in the neocortex. We also present a new, physiologically realistic mechanism for controlling the processor array.

Important Links:

Go Back