Overview
The
following sections outline some approaches to predicting and estimating generalization
ability, either a priori from static parameters such as network size, or after
observing training performance. The problem of estimating the true performance
of a prediction system trained on a limited data set is a basic statistics problem
so it is not surprising that many of these methods are direct applications of
statistical techniques. Techniques mentioned here are covered in more depth in
[44], [389], [317].