Learning to Predict Trustworthiness with Steep Slope Loss