Adversarial training for predictive tasks: theoretical analysis and limitations in the deterministic case