Title: Segmenting experience into generalizable predictive knowledge
Euan Prentis, doctoral student in the Bakkour Lab, Department of Psychology, University of Chicago
Abstract: Human experience unfolds gradually over time. To make effective decisions, it is therefore necessary to predict outcomes that may occur at distant points in the future. By learning which events generally follow from one another – a process termed predictive learning – humans can infer which actions will bring us to the best futures, and effectively arbitrate between choice options. A challenge of decision making in the real world is that events are complex, composed of numerous changing features. Predictive learning must be generalized across events as features change. The present research probes how generalizable predictive representations are learned. Using a combination of computational modelling and eye tracking, we demonstrate that successful generalization may be achieved by learning at the level of features (feature-based learning) rather than events (conjunctive learning). Particularly, an inductive bias that segments learning into semantic categories may promote more accurate feature-based learning, and better choice.