we propose a new variational bound for estimating mutual information and show the strength of our estimator in large-scale self-supervised representation learning through MI maximization.