Background of the verification and validation of neural networks

Abstract

This book is an introduction to the methods and procedures that have proven to be successftil for the V&V of artificial neural networks used in safety-critical or mission-critical applications. There are two types of neural networks discussed in this book, fixed (non-adaptive) and dynamic (adaptive). Fixed neural networks are trained before deployment and do not change. Dynamic neural networks learn during deployment. Artificial neural networks stem from research of the human brain. Neural networks are made up of neurons (or nodes), connections, weights, and layers. The first development of experimental artificial neural networks dates back to 1943 with research conducted by W.S. McCulloch. Since then, neural networks have made their way into safety-critical and mission-critical applications, such as intelligent flight control. V&V is the formal process by which to test and examine a system for dependability, reliability, and assurance. Verification tries to answer the question "Is the product being built right?" Validation tries to answer the question "Is the right product being built?" The methods and procedures presented in this book have been evaluated on a real-world intelligent flight control application for the IPCS project at NASA DFRC. The IPCS project uses adaptive neural networks to control and stabilize an P-15 research aircraft in catastrophic conditions for pilot assistance. © 2006 Springer Science+Business Media, Inc.