Large Scale Adversarial Representation Learning