Practical Neural Network Recipies in C++

72

Chapter 5

It is a profound testimony to the approximating power of

feedforward networks that such a close and smooth fit is able to be

obtained by using only two hidden neurons. The relatively large error

that occurs near

(0,

1)

is partially due to the nonlinearity of the output

neuron's activation function. When feedforward networks are used for

function approximation where serious noise problems are not expected,

the output neurons are usually given linear activation functions.

Alternatively, if it is difficult or impossible to change the activation

function due to program limitations, the data can be scaled to cover a

narrow interval