An interactive learning method where a human corrects the model's mistakes during training to fix distribution mismatch.