047 Interpretable Machine Learning - Christoph Molnar

Debates & Ethics

The debate over what makes a model interpretable is ongoing, with Keith Duggar and Christoph Molnar discussing the limitations of so-called interpretable models like linear models and decision trees. Keith argues that even these models become less interpretable as dimensionality increases, compressing complex structures into a few coefficients 1. Christoph agrees, noting that while decision trees can be simplified, they often sacrifice predictive accuracy for interpretability. The conversation also touches on the ethical implications of interpretability, with Tim Scarfe noting the importance of building tools that allow human intuition and ethics to guide machine learning models 2.

We have to build the tool set that you're talking about and continue this process of exploring how to better explain and interpret ML models so that human beings can have that oversight because it's the only thing that's going to give us comfort, really, as a society.

--- Keith Duggar

Christoph acknowledges that while interpretability and fairness are related, they are distinct fields, each requiring its own focus and expertise.

   

Interpretability Challenges

Interpreting machine learning models presents significant challenges, particularly when the interpretability model fails to accurately approximate the actual ML model. Tim Scarfe highlights the pitfalls of using interpretability methods that may lead to incorrect assumptions, such as the case with the COMPAS model in the US 3. Christoph Molnar emphasizes the issue of feature dependence, where correlated features complicate attribution and extrapolation, often breaking the model's predictions 4.

When you have feature dependence, it makes attribution and extrapolation problematic.

--- Christoph Molnar

He suggests that while conditional permutation schemes exist to maintain joint distributions, they can sometimes exacerbate the problem, highlighting the complexity of achieving true interpretability.

   

Philosophy & Simplicity

The philosophy of interpretability in machine learning involves balancing complexity with understanding. Christoph Molnar envisions a future where interpretability methods are standardized and integrated into engineering processes, making them accessible to all 5. Tim Scarfe advocates for starting with simple models and gradually increasing complexity only as needed, emphasizing the importance of understanding feature dependence to avoid misleading interpretations 6.

We should always seek to understand and simplify problems wherever we can.

--- Tim Scarfe

This approach aligns with the broader goal of making machine learning models not only more interpretable but also more reliable and trustworthy.