In christophM/interpretable-ml-book: Interpretable machine learning

Rule Extraction and Model Extraction

Some methods, which are all only tested on single hidden layer networks. - KT method: Extract for each neuron a rule - Extract decision tree (Hinton) - CRED (2001, Sato and Tsukimoto) - Trepan

• Model tree distillation (for CNNs)

DeepRed

LIME

Types of rule extraction: Decompositional, Pedagogical, Eclectic Pedagogical is simply model-agnostic.

You can also look at different levels: Approximate the rules of a single neuron Approximate the whole network classification.

TODO: Link overview paper.

christophM/interpretable-ml-book documentation built on March 10, 2024, 10:34 a.m.