Checklist for Validating Trustworthy AI

Abstract

In the recent years AI technologies have been improved and utilized in the various real-world fields such as life, economy, finance, transportation, game, etc. Especially, the deep learning, one of the learning-based machine learning methods, has shown remarkable performance improvement in a broad variety of studies. As the widespread use of AI systems including the deep learning, however, the issue of reliability of AI-based systems has recently emerged. In the case of the many AI systems, they use a huge amount of data to train their models as well as the system is very complex for humans to comprehend. Hence, humans cannot be able to understand AI systems and have no confidence in the results they generate. Furthermore, it can cause various problems such as unexplained system error or uncontrollable system behavior when using AI systems in the real-world, and even can lead to very serious situations in sensitive services such as aviation, medical care, and security. In this paper, we examine a checklist to improve a reliability of AI system. Specifically, we introduce considerations with regard to the life cycle of an AI system.