Practical Neural Network Recipies in C++

Probabilistic

Neural Networks

219

In summary, it must be emphasized that the author does not

mean to discourage careful experimentation to determine better

optimization criteria. Certainly the simple counting method given here

is primitive. All that is intended is a stern warning that coming up

with something better will not be a trivial task.

Bayesian Confidence Measures

In most practical classification problems, there is a possibility that

unknown samples may not be drawn from any of the trained classes.

A submarine sonar classifier may have been trained

to

identify several

types of subs, but no attempt was made to include whales, boulders ...