Policing functions for machine learning systems
Thai Son Hoang, Naoto Sato, Tomoyuki Myosin, Michael J. Butler, Yuichiroh Nakagawa, Hideto Ogawa · ePrints Soton (University of Southampton) · 2018
Machine learning (ML) systems typically involve complex decision making mechanisms while lack clear and concise specifications. Demonstrating the quality of ML systems therefore is a challenging task. We propose an approach combining formal methods and metamorphic testing for improving the quality of ML systems. In particular, our framework enables the possibility of developing policing functions for runtime monitoring ML systems based on metamorphic relations.