TRAINING NETS OF HARDWARE REALIZABLE SIGMA-PI UNITS

Learning convergence is demonstrated for networks of nodes which are defined by a population of values at the vertices of the n-dimensional hypercube. These are functionally equivalent to higher order nodes or sigma-pi units but have the potential to be implemented in readily available memory components. The cube based structure also offers insight into the process of generalization. Three algorithms are discussed: reward-penalty, back propagation, and a new one based on system identification in control theory. It is shown that reward-penalty style techniques emerge as a special case of system identification.