Walk me through how you would design a self-supervised learning objective for a dataset where the underlying ground truth is partially unknown.
