Functional mapping of desired signals for improved performance of fully dynamic, feedforward supervised neural networks

A new method is presented of functional mapping of the desired signal used for the training of dynamic, feedforward, supervised neural networks. There are two basic ideas explored in the paper. First is the idea of real time training. The second idea is to pass the desired signal through the same number and form of nonlinearities as the data encounters as it passes from the input layer through intermediate layers to the output. This is compared with the traditional method of training at the output layer with a desired signal that is linear with the input signal. The comparison are made on a low order feedforward backpropagation networks. The outputs of the input and intermediate layers are passed through logistic nonlinearities. Thus, in the functional mapping cases, the desired signal is passed through a logistic function twice