self-supervised contrastive representation learning has proved incredibly
successful in the vision and natural language domains, enabling
state-of-the-art performance with orders of magnitude less labeled data.
However, such methods are domain-specific and little has been done to lever