You Don't Need Domain-Specific Data Augmentations When Scaling Self-Supervised Learning

All instantiations of this paradigm were trained using strong and well-established hand-crafted data augmentations, leading to the general belief that they are required for the proper training and performance of such models.