047 Interpretable Machine Learning - Christoph Molnar

Uncertainty

In the realm of interpretable machine learning, quantifying uncertainty is crucial for enhancing model reliability. Christoph Molnar emphasizes the need for statistical rigor, pointing out that current interpretability methods often overlook uncertainty, leading to potential pitfalls 1. He suggests that incorporating distributional assumptions and confidence intervals, akin to traditional statistics, could improve the robustness of model explanations 1.

It's better to have not only just one number or one explanation, but also have the distribution to do of this explanation or this number and to quantify what uncertainty is behind computing this number.

--- Christoph Molnar

Tim Scarfe and Keith Duggar discuss the inherent probabilistic nature of models, highlighting the importance of not discarding valuable information during simplification 2.

   

Challenges

Applying statistical rigor to machine learning models presents significant challenges, often leading to complex mathematical explanations that may not enhance interpretability. Keith Duggar questions whether creating complex math models to explain other complex models truly advances understanding 3. Christoph Molnar warns of the risk of p-hacking in interpretability, a practice prevalent in natural sciences that could soon affect machine learning if not addressed 4.

You have something you don't understand, and you explain it with something you don't understand.

--- Christoph Molnar

Tim Scarfe4.