Towards A Rigorous Science of Interpretable Machine Learning

By Finale Doshi-Velez et al.

Published on March 2, 2017

Read the original document by opening this link in a new tab.

Table of Contents

1. What is Interpretability? De nition
2. Why interpretability?
3. How? A Taxonomy of Interpretability Evaluation
3.1 Application-grounded Evaluation: Real humans, real tasks
3.2 Human-grounded Metrics: Real humans, simpli ed tasks
3.3 Functionally-grounded Evaluation: No humans, proxy tasks
4. Open Problems in the Science of Interpretability, Theory and Practice

Summary

This paper discusses the need for interpretability in machine learning systems, especially in complex applications. It explores the definition of interpretability and its importance in various scenarios such as scientific understanding, safety, ethics, and mismatched objectives. The paper proposes a taxonomy for evaluating interpretability, including application-grounded, human-grounded, and functionally-grounded approaches. It also highlights open problems in the science of interpretability and suggests a data-driven approach to discover factors of interpretability.