Representation Learning Using Adversarially-Contrastive Optimal Transport