Learning on spaces of distributions: from recommender systems to generative adversarial networks