Practical Neural Network Recipies in C++

Designing

the

Training Set

249

Stratified Training Data

There is one situation in which an especially large number of training

samples should be used. This is when the data is highly stratified.

The presence of many subclasses in the training data necessitates

larger training sets because communal random patterns are much

more likely. Suppose that the characteristics of a subclass place it in

a position in data space that is fairly isolated from other subclasses.

Then we must rely on many training samples from this subclass to

swamp out random patterns common to members of this subclass. If

we have only five samples from this subclass, and