“Self-supervised learning methods based on reconstructing images or video from corrupted versions have failed to produce good internal representations for downstream tasks like object recognition.”