Adaptive Adversarial Multi-task Representation Learning