Summary

In this chapter, we reviewed explainability in AI/ML through an eight-dimension categorization. Although this is not necessarily a comprehensive or exhaustive overview, this does give us a big picture of who to explain to, different stages and scopes to explain, various kinds of input and output formats of the explanation, common ML problems and objectives types, and finally, different post-hoc explainability methods. We then provided two concrete exercises to explore the SHAP and Transformers Interpret toolboxes, which can provide perturbation and gradient-based feature attribution explanations for NLP text sentiment DL models.

This gives us a solid foundation for using explainability tools for DL models. However, given the active development of XAI, this is only the beginning of using XAI in DL models. Additional explainability toolboxes such as TruLens (https://github.com/truera/trulens), Alibi (https://github.com/SeldonIO/alibi), Microsoft Responsible AI Toolbox (https...