Reproduce “On the Relationship Between Interpretability and Explainability in Machine Learning” on OpenResearch

About this paper

Benjamin Leblanc, Pascal Germain

This conceptual paper argues that interpretability and explainability in machine learning are complementary, rather than substitutable, concepts. It details the individual shortcomings of each approach and demonstrates how combining them in "explained interpretable predictors" mitigates these flaws, advocating for a new paradigm in trustworthy AI.

Key insights

• Explaining black-box models is fundamentally limited by issues such as blind trust, inherent inaccuracy due to simplification, misalignment of understanding, and vulnerability to adversarial attacks.
• Inherently interpretable models, while transparent, can provide incomplete information, are subject to individual interpretation, can be harder to train, and may not alone ensure critical outcomes like fairness or robustness.
• When combined, interpretability provides an objective ground for verifying explanations, while explainability offers objective metrics and completeness that enhance the insights from interpretable models.

Summary by alphaXiv.