How do zero-shot, one-shot and few-shot learning differ?