Neural network design: methodology

Summary form only given. The design of a neural network for the processing and identification of temporal signal patterns can be described concisely using the mathematics of stochastic processes. Considered in this work is the basic mathematical methodology for the design of two-dimensional neural networks specific to signal processing. A two-dimensional neural network is considered to be any neural network that may be topologically deformed into an array which satisfies the following three criteria. First, all nodes that have the temporal signal pattern as at least one input are aligned on the same edge of the array. Second, all nodes with the required output signal are similarly aligned along the contrasting edge. Third, the two remaining edges of the two-dimensional array consist only of nodes that are connected to nodes interior to the network. The input signal of interest is a temporally dependent sequence of measured values of a real, bounded, and continuous function. The output signal is required to have a unique maximum, and the curvature at the maximum shall be very small in comparison to the maximum value.<>