An Introduction to Framework Adaptations for Additional Assurance of a Deep Neural Network Within Naval Test and Evaluation

Abstract

The complexity of modern warfare has rapidly outmatched the capacity of a human brain to accomplish the required tasks of a defined mission set. Task-shedding mundane tasks would prove immensely beneficial, freeing the warfighter to solve more complex issues; however, most tasks that a human might find menial, and shed-worthy, prove vastly abstract for a computer to solve. Advances in Deep Neural Network technology have demonstrated extensive applications as of late. As DNNs become more capable of accomplishing increasingly complex tasks, and the processors to run those neural nets continue to decrease in size, incorporation of DNN technology into legacy and next-generation aerial Department of Defense platforms has become eminently useful and advantageous. The assimilation of DNN-based systems using traditional testing methods and frameworks to produce artifacts in support of platform certification within Naval Airworthiness, however, proves prohibitive from a cost and time perspective, is not factored for agile development, and would provide an incomplete understanding of the capabilities and limitations of a neural network. The framework presented in this paper provides updated methodologies and considerations for the testing and evaluation and assurance of neural networks in support of the Naval Test and Evaluation process.