Rethinking Data Augmentation: Self-Supervision and Self-Distillation