An interpretable ensemble method for deep representation learning