Testing distributional assumptions of learning algorithms