A novel neural network for temporal pattern identification with applications to control systems

A new neural architecture that can recognize and recall temporal sequences by a self-organizing method requiring little learning effort is presented. The general principle is based on a network of nodes made from groups of neurons. These nodes perform simple recognition and recall of state transitions and are arranged hierarchically so that short subsequences of an observed temporal sequence can easily be learned and then reconstructed. This hierarchy involves extensive interconnections between layers of nodes. The learning process changes the strengths of these interconnections. An application of such a network, involving the implementation of a process controller for a chemical system, is given