Title: Towards human-like social intelligence in machines
Abstract: Virtually all areas of uniquely-human intelligence, from moral reasoning to language understanding, rely on social cognition. Characterizing its computational structure is therefore a central challenge in cognitive science and critical towards engineering human-like AI. In this talk I will present empirical and computational work showing that humans make sense of each other’s behavior through a form of inverse reinforcement learning (IRL). I will show how IRL-based inferences guide social reasoning, even in young children and infants, supporting a wide range of human social behavior. Finally, I will also discuss some ethical questions that arise when considering the role of this technology in potential surveillance applications.