Diving into the shallows: a computational perspective on large-scale shallow learning