A Self-Supervised Attribution Method for Explaining Neural Networks