>often its unclear that doing so would end up with results that are significantly better than traditional approaches

If Deep Learning can approximate any function, does that mean that in some cases it ends up approximating the output of traditional approaches?