Interpreting and improving deep-learning models with reality checks