Overview of Techniques for Rule Extraction From Neural Networks

Abstract

Neural networks have been a staple of the machine learning community since their inception decades ago. They have enjoyed a further surge in popularity as hardware and data limitations have allowed for the creation of deep neural networks, which have shown remarkable results in fields such as reinforcement learning and image classification, among others. However, the black box nature of neural networks presents a hurdle to their adoption, particularly in safety-critical domains such as medical diagnosis. A natural approach to explain the behavior of neural networks is the extraction of an interpretable set of rules which sufficiently mimic the behavior of the neural network. In this review, we provide an overview of the latest research in the field of rule extraction from neural networks. We also present a taxonomy for rule extraction techniques and highlight areas which we feel could be targeted for future research.