G.M. Megson
University of Newcastle upon Tyne. 1993
A new type of multi-layer neural network based on sets of trigonometric polynomials is introduced. It is shown that the weights of the network can be derived directly from the coefficients of the discrete analog form of the fourier least squares polynomial for the set generated from the output patterns, and a so-called separator function produced from input patterns of the "training" set. As a consequence the need for a protracted training period is removed and replaced by a more systematic analysis of output functions. The method is illustrated on a collection of examples. Some general properties and design principles regarding network size and generalization abilities are also considered.