Improving ML Safety with Partial Specifications

Abstract

Advanced autonomy features of vehicles are typically difficult or impossible to specify precisely and this has led to the rise of machine learning (ML) from examples as an alternative implementation approach to traditional programming. Developing software without specifications sacrifices the ability to effectively verify the software yet this is a key component of safety assurance. In this paper, we suggest that while complete specifications may not be possible, partial specifications typically are and these could be used with ML to strengthen safety assurance. We review the types of partial specifications that are applicable for these problems and discuss the places in the ML development workflow that they could be used to improve the safety of ML-based components.