Validation and Verification of Embedded Neural Systems

Abstract

In a safety critical environment every element of a system must be dependable. Although neural network technology could be applied to solve various problems in safety critical environments, there is currently no established method to validate and verify systems embedding neural network technology. This deficiency is mainly due to the fact that neural network technology corresponds to a very different way of viewing software, and that validation and verification methodologies developed for conventional software do not take into account these differences. This thesis is aimed at showing that even if at first sight a neural network function may look like a very compact ‘black box’, it is in fact not true and, therefore, the validation and the verification of neural systems are possible. Firstly, a short presentation of the business context in which this project has been placed is given. Secondly, an overview of neural network technology, which includes in particular a description of the main principles on which it is based, is presented. Thirdly, a set of good practice rules to develop a successful neural network application is suggested. From these good practice rules, guidelines to assess a neural network system have been derived and tested against two case studies. The main findings are discussed. Finally, a set of research issues and suggestions for future research are provided. This document is necessarily limited in its scope. Indeed, there is a considerable variety of models related to neural network technology, and each of them has its own characteristics. A general framework to validate and verify all these models would probably be too general and, therefore, might be unable to detect some significant mistakes specific to the model considered. Similarly, we cannot consider each model individually. We have, therefore, chosen to restrict the scope of this document to the 2 main types of models currently used, namely, the radial basis function network and the multi-layer perceptron. We have excluded, for instance, unsupervised techniques and recurrent neural networks that raise notoriously difficult problems which could not be addressed within the time available. However, whatever the model used, they are related in that they are data-driven. Consequently, issues relative to the data discussed in this thesis are central and generally applicable to each of these models.