Feb. 17, 11:45, Wei Ji Ma. Optimality and Probabilistic Computation in Visual Categorization

Chris Baker

Brain and Cognitive Sciences, MIT

Bayesian social inference: Modeling human reasoning about beliefs, desires, goals, and social relations

I will present a computational framework for understanding
human Theory of Mind (ToM): our conception of others' mental states
and their relation to the world and behavior. ToM supports behavioral
prediction, but also mental state inference: attribution of the
beliefs, desires, knowledge and other thoughts that explain observed
behavior. It also allows us to leverage observations of others'
actions to learn about the world. At the heart of the proposed
framework is an intuitive theory of intentional agency: a causal model
of how agents' beliefs, desires and goals interact with situations to
produce behavior. This intuitive theory of (inter)action is formalized
using partially observable Markov decision processes and Markov games.
Social reasoning is cast as Bayesian inference over models of
intentional action, reconstructing the mental states that give rise to
behavior. I will describe several behavioral experiments which
presented human subjects with trajectories of agents moving in simple
spatial contexts and collected human inferences about agents' beliefs,
desires, goals, and social intentions toward other agents. For
example, in an experiment inspired by classic false-belief tests of
theory of mind, people performed joint inference of agents' beliefs
and desires, given their actions in a partially observable
environment. In another experiment, people learned the location of
hidden food sources by observing the trajectory of an agent with a
known preference function. In these experiments, theory-based Bayesian
models predict human social inferences substantially better than
simpler variants or alternative models based on the analysis of
low-level motion features.