
Designing
the
Training
Set
251
Balancing the Classes
At least some consideration should always be given to the choice of
how many members of each class or natural group should be present
in the training set. Some neural network models, such as the
probabilistic neural net, can implicitly adjust for unbalanced training
sets.
But others, such as the multilayer feedforward net, cannot do so
without special training procedures. When a network learns by
minimizing the mean error across the entire training set, the propor-
tional representation in the training set can have a profound influence
on the network's performance. If a particular subclas