Making AI Models Intelligible to Experts (Interpretability)

Abstract

Artificial intelligence (AI) faces a critical “black box” problem, especially in safety-sensitive domains like healthcare and finance, where transparency is essential. This chapter provides a comprehensive exploration of interpretability, distinguishing it from post-hoc explainability. It covers methods from classical linear models and decision trees to modern techniques like GAMs, EBMs, and deep learning interpretation (saliency maps, Grad-CAM, LRP). The discussion extends to inherently interpretable architectures such as prototype-based, neuro-symbolic, and causal models, and situates interpretability within governance, using case studies and tools like Model Cards. We conclude that interpretability is a multidisciplinary imperative for building powerful, transparent, and human-aligned AI.

Ähnliche Arbeiten