Training Machine Learning Models by Regularizing their Explanations