Multi-prong Framework Toward Quality-Assured AI Decision Making
G. Y. Hong, Alvis C. M. Fong · 2019
Can we trust decisions made by artificial intelligence (AI)? This is a question that is important not just for computer scientists, but also an increasingly broad spectrum of non-specialist users of AI. AI is already used for decision support in healthcare (e.g. diagnosis), business and commerce (e.g. hiring, loan applications), and so on. There is therefore an urgent need to address the question of quality assurance of AI decision making. Contemporary AI focuses on a connectionist approach with an emphasis on artificial neural networks (NN) that get deeper and deeper for better quantitative regression or classification performance. However, while NN and more generally machine learning (ML) algorithms perform well statistically, results are individually unreliable. When ML fails, it can fail miserably with unclear failure models due to a lack of interpretability of many of the results generated by contemporary AI. This paper addresses this critical and timely issue by proposing a multi-prong framework that brings together elements of good old AI and ensemble decision making for enhanced interpretability, and best practices in software quality assurance, toward ensuring a high standard of quality in AI decision making. The framework is illustrated with use case scenarios drawn from multiple domains to demonstrate its broad applicability.