Margin error and generalization capabilities of multiclass discriminant systems
Abstract
This technical report exposes an extension of a lemma by Bartlett to the multiclass case. First, discussing the notion of margin and how its use for biclass discriminant models, we introduce the notion of margin for the multiclass case. Then, we develop our calculations and prove how to derive generalization error bounds for classifiers having multiple outputs. These bounds are expressed in terms of covering numbers. This report shows how to bound these quantities for the set of functions we consider and give an example for linear discriminant systems.