Supplementary Materials: Improving Deep Learning Interpretability by Saliency Guided Training