On visual self-supervision and its effect on model robustness