If your data distribution shifts, use self-learning