Error rates on an unseen test dataset for clouds of randomly initialized runs, comparing large single networks against a two-step process. The contour of each cloud is outlined and non-dominated solutions are highlighted