Common pitfalls in using AI in high-risk domains

Abstract

Deep learning, AI and machine learning are emerging as important tools e.g., to segment, classify and detect pathologies medical diagnostics. Powerful and easy to use frameworks for machine learning have increased the accessibility of these methods. At the same time this removes the need for machine learning experts with deep insights in the limitations of the technology and places the power of AI in the hands of domain experts. Although such high-risk domains are very diverse, the challenges using machine learning are related. This paper will discuss insights from application of machine learning in domains as diverse as medical image analysis, and conditionbased monitoring in the Norwegian Oil and gas industry. We then discuss how these insights can be applied when analyzing hyperspectral data from human skin.