Practical Neural Network Recipies in C++

362

Chapter

20

Some applications cannot be adequately served by providing nothing

more than an ambiguous output-neuron activation

level.

Suppose that

we must design a neural network that resides in a fighter plane,

examining the radar return signature of distant objects. The purpose

of our network is to help the pilot judge whether or not a blip on the

screen is an aircraft. She would be dismayed if the network simply

reported that a blip that just appeared had activated the network's

output neuron to a 73-percent activation level! It would be far more

informative for the network to report that there is a 92-percent chance

that the blip is