debugging a machine learning algorithm is already hard without this approximation method. I can see adding another layer could make it extremely harder.