
Evaluating Performance of Neural Networks
351
ROC (Receiver Operating Characteristic) Curves
A common use for neural networks is detecting the presence of a
condition. The network is trained to produce a low activation level if
the condition is absent, and a high level if it is present. Often, for lack
of any apparently better
choice,
a threshold of half-activation is chosen
for making the decision. If an unknown case activates the output
neuron to at least half of full activation, it is decided that the condition
is present. Otherwise, conclude that it is absent. Some of the
techniques previously described in this chapter can be used
to
evaluat ...