CASTing Your Model: Learning to Localize Improves Self-Supervised Representations