Machine Learning Explainability: Inherently Interpretable Models (Stanford Seminar)

Added:

The Necessity of Post-Hoc Explainability and the Limits of Inherent Interpretability

While inherently interpretable models offer transparency, a major counterpoint argues that they are insufficient for highly complex, unstructured data like image, video, and natural language, where deep black-box models excel. Proponents of this view argue that forcing models to be inherently interpretable often requires sacrificing predictive accuracy and performance. Furthermore, critics point out that 'inherent interpretability' is often an illusion; as data dimensionality scales, even simple models like decision trees or Generalized Additive Models (GAMs) become too large and complex for human comprehension. Consequently, this perspective advocates for the continued development of sophisticated post-hoc explanation methods (such as SHAP, LIME, or feature attribution) applied to high-performing black-box models, arguing that utility should not be compromised for simplistic design.

Fundamental supervised learning concepts, including the distinction between classification and regression tasks.
Familiarity with standard baseline models, particularly linear regression, logistic regression, and decision trees.
Basic statistical concepts, including Generalized Linear Models (GLMs) and the bias-variance tradeoff.
Understanding of the core motivation for machine learning interpretability and the limitations of 'black-box' models.
Post-hoc explainability techniques for complex models, such as LIME, SHAP, and saliency maps.
Methods for quantitatively and qualitatively evaluating the quality, robustness, and faithfulness of explanations.
The role of interpretability in AI fairness, bias detection, and compliance with algorithmic regulatory frameworks (e.g., GDPR).
Domain-specific deployment of interpretable models in high-stakes fields like clinical healthcare, finance, and criminal justice.
19.3K views197likes44:57@stanfordonlineOriginal Release: 2022-11-03

Inherently interpretable machine learning models are designed to be understandable by construction, rather than requiring post-hoc explanation. Key approaches include: (1) Rule-based models like Bayesian Rule Lists (2016) that generate if-else-if rule lists using generative Bayesian processes with Metropolis-Hastings optimization, and Interpretable Decision Sets that optimize for sparsity, distinctness, precision, recall, and class coverage; (2) Risk scores that assign point values to conditions and sum them to predict risk levels, formulated as mixed integer programs; (3) Generalized Additive Models (GAMs) that model outcomes as additive combinations of shape functions for each input variable, with GAM² extending to pairwise interactions; (4) Prototype-based models that identify representative instances from data to classify new points; and (5) Attention layers in deep learning that highlight which input features influence each output decision, providing interpretability through explicit attention weights.