Towards Automated Circuit Discovery for Mechanistic Interpretability