
Functional
Link
Networks
227
raw input data. However, the nonlinear transfer function of the
output neuron makes this network much more tolerant of chaotic noise
contaminating the measured data, as compared to ordinary regression.
Yet we retain the advantage of practicable validation, as a functional
link network is nothing but ordinary regression followed by a simple
activation function whose behavior is well known.
The entire preceding discussion has been based on the premise
that the purpose of the input functions is to eliminate the hidden
layer. However, it must be pointed out that sometimes it is still
beneficial to use a hidden layer ...