A system trained with self-supervised learning on images needs only 2-3 examples of a new object (like an elephant) to recognize it, making it more efficient than supervised learning which requires thousands of labeled examples—this is closer to how humans learn.