A detailed derivation of the relationship between generalization error and ambiguity in regression ensembles

Abstract

In this technical report we will show the complete sequence of steps for the derivation of the equation of the Ensemble-Error E = E - A , introduced in the paper by Krogh and Vedelsby [1], that describes the error E of an ensemble of networks related to the average error E of the single network and the Ambiguity A of the ensemble, which in turn is a measure of the ?disagreement? among the networks.