Stochastic Re-weighted Gradient Descent via Distributionally Robust Optimization