Efficient Decompositional Rule Extraction for Deep Neural Networks

  • 2021-11-24 16:54:10
  • Mateo Espinosa Zarlenga, Zohreh Shams, Mateja Jamnik
  • 0

Abstract

In recent years, there has been significant work on increasing bothinterpretability and debuggability of a Deep Neural Network (DNN) by extractinga rule-based model that approximates its decision boundary. Nevertheless,current DNN rule extraction methods that consider a DNN's latent space whenextracting rules, known as decompositional algorithms, are either restricted tosingle-layer DNNs or intractable as the size of the DNN or data grows. In thispaper, we address these limitations by introducing ECLAIRE, a novelpolynomial-time rule extraction algorithm capable of scaling to both large DNNarchitectures and large training datasets. We evaluate ECLAIRE on a widevariety of tasks, ranging from breast cancer prognosis to particle detection,and show that it consistently extracts more accurate and comprehensible rulesets than the current state-of-the-art methods while using orders of magnitudeless computational resources. We make all of our methods available, including arule set visualisation interface, through the open-source REMIX library(https://github.com/mateoespinosa/remix).

 

Quick Read (beta)

loading the full paper ...