Exploring model-based versus model-free pupillometry correlates to reinforcement learning parameters

AbstractWhile many recent studies have successfully used reinforcement learning (RL) frameworks to explain large portions of variance within neurobiological and decision-making datasets, the relatability of such models to the true mechanisms and dynamics underlying human learning, cognition, and behavior is arguably still quite limited--in part due to the exclusion of well-defined mechanisms controlling the dynamics of sensory-model updating (particularly during exploratory behavior) and sensory-model extraction (for use of exploitative behavior) processes. In an attempt to mend this gap, the current study investigates the diameter of the pupil as a potential signature of both ongoing sensory-model updating and sensory-model extraction processes. With the use of a hybrid Q-learning model, these hypothesized correlates are found to account for discrepancies in pupil diameter between model-based and model-free learning strategies during exploratory and exploitative behavior, and simultaneously frame human learning experience as a dynamic interplay between sensory-model updating and recollection processes.

Return to previous page