Physics-Inspired Interpretability Of Machine Learning Models