Robust Learning under Distributional Shifts