Networks of exponential neurons for multivariate function approximation
Shlomo Geva, Joaquin Sitte · 1991
A three-layer neural network, having a hidden layer of neurons with an exponential transfer function, capable of performing function approximation more accurately, and more economically, than a conventional multilayer perceptron (MLP) having neurons with a sigmoidal transfer function, is described. The network was trained by a variation of the standard backpropagation gradient-descent technique. The results of a difficult approximation problem, where a conventional MLP of similar size simply fails to perform within reasonable constraints on training time, are shown graphically.>