Demystifying statistical learning based on efficient influence functions