Self-supervised learning from images: No negative pairs, no cluster-balancing.