Are loss functions all the same in statistical learning?