Conference on Cognitive Computational Neuroscience

Human learning occurs under uncertainty. Uncertainty is heterogeneous, however, with different forms of uncertainty exerting distinct influences on learning. For example, one can be uncertain about what to do to maximize outcomes (“policy uncertainty”), while also being uncertain about general world knowledge (“epistemic uncertainty”). During learning one might need to strike a balance between competing motivations, such as choosing to maximize rewards while also wanting to gather more knowledge about the world that may be useful for the future. Learning about social partners highlights the advantage of adaptively leveraging both policy and epistemic forms of uncertainty. Here we developed a novel eyetracking method to capture trial-level estimates of epistemic and policy uncertainty to better characterize how uncertainty is resolved during social learning. Using empirically-derived uncertainty metrics based on eye patterns, and a Bayesian reinforcement learning model, we show that humans flexibly switch between relying on policy and epistemic uncertainty to learn in dynamic social environments.