In depth study, the input data changes, but the depth of the training error of the same how to return a responsibility?