Towards Human-Compatible AI for Well-being by Integrating Physiological Viewpoint With Machine Learning Viewpoint

This paper focuses on “human-compatible AI” which aligns with human values and remains under human control to pre-vent unintended and harmful consequences, and discusses it to develop human-compatible AI for well-being. For this is-sue, this paper proposes the human-compatible AI for a sleep as one of the human-compatible AI for well-being, which is designed to have the functions of (1) checking how the esti-mated sleep stage (corresponding to suggestions to users) fol-lows the biological rhythms which determine their health conditions (corresponding to human values) and (2) modify-ing the estimated sleep stage according to their biological rhythms. To investigate an importance of the proposed ap-proach, this paper applies it into the sleep stage estimation based on the acceleration sensor data. Through the human subject experiment, the following implications have been re-vealed: (1) it is dangerous to simply employ machine learning (i.e., Random Forest in this research) for the sleep stage esti-mation because the sleep stage is artificially estimated without following the ultradian rhythm which are generally found in humans; and (2) it is important to integrate the physiological characteristic (i.e., the ultradian rhythm) with machine learning for the sleep stage estimation because such an inte-gration can estimate the sleep stage that follow the ultradian rhythm.