KISS My ML Assurance

For some aircraft systems, the use of machine learning (ML) technologies such as neural networks (NN) may be a desirable choice for implementing a software component. In this paper we advocate a "Keep it simple stupid" (KISS) approach to evaluation and approval of ML-based components as part of the aircraft certification process. We show that the use of ML technologies to implement all or part of a software component doesn’t change most of the traditional (meaning non-ML) development and assurance activities. However, there are some aspects of traditional activities that present challenges or do not provide the expected assurance, necessitating a different approach. We conclude by applying our approach to an example avionics system to illustrate the concepts.