Biography

Matthias Rolf is Senior Lecturer in Computing and Robotics at Oxford Brookes University, UK since 2016. He previously was Specially Appointed Researcher and Assistant Professor at Osaka University, Japan  (2013-2016) and research assistant at Bielefeld University (2008-2013) where he obtained his PhD (with highest honors) in 2012. His research interest spans robotics, machine learning, software engineering, and cognitive science, with a particular focus on developmental robotics. His work on “goal babbling” as an efficient way of sensorimotor learning lead to several research awards and international news coverage. His recent research focuses on machine learning approaches to agents’ autonomous learning of own goals as well as ethical AI.

Abstract

From social interaction to ethical AI: a developmental roadmap

AI and robot ethics have recently gained a lot of attention because adaptive machines are increasingly involved in ethically sensitive scenarios and cause incidents of public outcry. Much of the debate has been focused on achieving highest moral standards in handling ethical dilemmas on which not even humans can agree, which indicates that the wrong questions are being asked. While traditionally engineered artifacts, including AI, require the designer to ensure ethical compliance, learning machines that change through interaction with people after their deployment can not be vetted in just the same way. I will argue that in order to progress on this issue, we need to look at it strictly through the lens of what behavior seems socially acceptable, rather than idealistically ethical. Machines would then need to determine what behavior is compliant with social and moral norms, and therefore be receptive to social feedback from people. I will discuss a roadmap of computational and experimental questions to
address the development of socially acceptable machines, and emphasize the need for social reward mechanisms and learning architectures that integrate these while reaching beyond limitations of traditional reinforcement-learning agents. I suggest to use the metaphor of “needs” to bridge rewards and higher level abstractions such as goals for both communication and action generation in a social context. We then suggest a series of experimental questions and possible platforms and paradigms to guide future research in the area.