Good generalization performance of pattern classiers is
achieved when the capacity of the classication function
is matched to the size of the training set Classiers with
a large number of adjustable parameters and therefore
large capacity likely learn the training set without error