Adversarial Reweighting with $\alpha$-Power Maximization for Domain Adaptation