Weaving it All Together - A Methodology for the Verification and Validation of Adaptive Neural Networks

The allure of artificial intelligence systems is their ability to adapt beyond their initial design allowing accommodation of changing environmental conditions, autonomous decision and control, and emergent behavior. However, the standard methods of software assurance fail to address the intelligent attributes of these systems. If such systems are ever to be adopted for use in complex and high assurance applications, a standard needs to be developed for a rigorous, comprehensive, and practical approach to the verification and validation (V&V) of learning systems. Based upon our experiences with online learning neural networks in safety- and mission-critical systems, we propose a comprehensive methodology for neural networks that takes independent strands of V&V research related to adaptive systems and weaves them through each phase of the development life cycle. The activities proposed complement current IEEE Standard 1012-1998 V&V activities, result from diverse research efforts, and carefully address the adaptive nature of the system at each stage of its development. The methodology, while containing guidance specific for neural networks, has been written in broad terms and some parts are applicable to learning systems in general.