Learning to Pivot with Adversarial Networks