Review of Research in the Field of Developing Methods to Extract Rules From Artificial Neural Networks

Abstract

Abstract: A large-scale review and analysis of the existing methods and approaches to extract rules from artificial neural networks, including deep learning neural networks, is carried out. A wide range of methods and approaches to extract rules and related approaches to develop explainable artificial intelligence (AI) systems are considered. The taxonomy and several directions in studies of explainable neural networks related to the extraction of rules from neural networks, which allow the user to get an idea of how the neural network uses the input data, and also, using rules, to reveal the hidden relationships of the input data and the results found, are explored. This review focuses on the relationship of the most common rule-based explanation systems in AI with the most powerful machine learning algorithms using neural networks. In addition to rule extraction, other methods of constructing explainable AI systems are considered based on the construction of special modules that interpret each step of changing the neural network’s weights. A comprehensive analysis of the existing research makes it possible to draw conclusions about the appropriateness of using certain approaches. The results of the analysis will allow us to get a detailed picture of the state of research in this area and create our own applications based on neural networks, the results of which can be studied in detail and their reliability evaluated. The development of such systems is necessary for the development of the digital economy in Russia and the creation of applications that allow making responsible and explainable management decisions in critical areas of the national economy.