Mostly because the loss function space is not well understood, we need to do some kind of descent