Self-Supervised Representation Learning as Mutual Information Maximization