Chapter 9: Fundamentals of Deep Learning Explainability

Explainability is providing selective human-understandable explanations for a decision provided by an automated system. In the context of this book, during the full life cycle of deep learning (DL) development, explainability should be emphasized as a first-class artifact, along with the other three pillars: data, code, and model. This is because different stakeholders and regulators, model developers, and final consumers of the model output may have different needs to understand how the data is used and why the model produces certain predictions or classifications. Without such understanding, it will be difficult to gain the trust of the consumers of the model output or to diagnose what could have gone wrong when model output results drift. This also means that explainability tools should be employed not only for explaining prediction results from a deployed model in production or during offline experimentation, but also for...