
242
Chapter 14
ate network, which gets rid of three-quarters of what is left. The
remaining small chunk goes to a fairly complex network, which finds
all micro features plus just a few false alarms. The final network is
extremely slow, but it can be counted on to separate correctly the
sheep from the goats. Four or more levels in the hierarchy are not at
all uncommon in practice.
Factorable Problems
Sometimes a large problem can be systematically broken down into
several smaller problems. When this is so, it may be appropriate to
use separate neural networks for each part. The decision of whether
or not to do this should be based on practica ...