Human cognitive neuroscience seminar

Speaker: Tia Gong (University of Edinburgh)

Title: Active causal structure learning in continuous time

Abstract: Most work on causal cognition has focused on learning and reasoning about contingency data of the sort collected across multiple observations or experiments. However, this is just the tip of the causal cognition iceberg. One of the more general problems lurking beneath is that of learning latent causal structure connecting one's actions with events as they unfold in continuous time. In this project, we examine how people actively learn about causal structure in a continuous-time setting, focusing on when and where within a system of interest people perform interventions, and how this shapes their learning. Across two behavioural experiments, we find that participants' interventions and accuracy patterns are well predicted by models that strike a balance between maximizing expected information and minimizing expected inferential complexity. That is, participants timed and targeted their interventions in a manner that produced simple and easy to learn from causal dynamics, and learn better when information is high while complexity is low. We discuss how the bounded online learning setting challenges existing normative accounts of active learning, and argue for the important role of metacognitive awareness of one's inferential limitations.