Sum of Ranked Range Loss for Supervised Learning