Trusting and understanding integrated gradients

First off, the integrated gradient technique is available off-the-shelf in a few open source libraries such as shap and captum. We can utilize the method from the library without reimplementing the method. However, obtaining explanations without understanding the technicalities behind the technique can reduce trust in the explanations. If you don’t trust the technique, the explanation results themselves hardly mean anything. Subsequently, if you don’t explain your predictions, the prediction results themselves hardly mean anything! In this section, we will dive into what integrated gradients do so that you can trust what integrated gradients can explain to you.

Integrated gradients add a few extra components to the basic gradient-based feature importance-based explanations. These components are geared toward satisfying a few critical axioms that make or break the reliability of an explanation method. These axioms are...