Keynote Speakers

Title: Engineering and reverse-engineering theories of mind for human and cooperative AI agents

Abstract:  I will briefly present the computational and cognitive science foundations of engineering systems that attempt to simulate and scale human minds' intuitive theories of other minds, broadly known under the heading of "Bayesian Inverse Planning".  Some of these ideas will have been introduced in Rebecca Saxe's morning keynote, in joint work we have done with the goals of scientific theory-building and theory-testing in cognitive science: Can we build a machine-implementable model of how human minds reason about others' mental states that is quantitatively predictive and explanatory, and that is robust and general enough to be the basis for scientifically grounded technologies and interventions that improve our lives?  I will then discuss how, in an AI context, this work could provide foundations for building machines that think, learn and act productively and cooperatively with people, in ways that we humans have come to expect in our interactions with other humans -- both trusted friends and people we are just meeting for the first time.  I will highlight recent progress based on advances in probabilistic programming languages, language models, and their integration, that together offer the potential to scale Bayesian approaches to inverse planning from the very simple settings where these ideas have been worked out in cognitive science to much more complex, real-world settings and deeper recursive forms of reasoning.  I will show briefly how this work provides the basis for cognitive science accounts (and potentially scalable AI technologies) that capture many aspects of human social intelligence, such as communication, storytelling, norm formation and adaptation, social evaluation ,and moral judgment and decision-making.