Practical Neural Network Recipies in C++

Designing

Feedforward Network Architectures

185

and follow the algorithm until that performance level is obtained. At

this time, do not validate the network. Save it and then demand

better performance. Obviously, we will immediately add a hidden

neuron. A new training test set may then have to be merged into the

training set. The algorithm is run until either we get sick of dealing

with a huge training set, or our new expectation is met. As long as the

fimdamental assumption of the quality of the test sets is met, our

actions are legal. Our performance goal is limited only by our ability

to collect new data and by our computational resources ...