Building and Breaking the Chain: A Model of Reward Prediction Error Integration and Segmentation of Memory.

Abstract

Prediction errors drive reinforcement learning and organize episodic memory into distinct contexts, but do these effects interact? Here, we review the roles of midbrain dopamine, the locus coeruleus, and the hippocampus in event cognition to propose and simulate the theoretical influence of two prediction error signals in integrating versus segmenting events in memory. We suggest that signed reward prediction errors can build mental models of reward environments, increasing the contextual similarity (integration) of experiences with stronger, more stable reward expectations. On the other hand, unsigned reward prediction errors can signal a new model of the environment, generating a contextual shift (segmentation) between experiences that crossed them. We moreover predicted that these differences in contextual similarity give rise to distinct patterns of temporal-order memory. We combined these ideas in a computational model to account for a seemingly paradoxical pattern of temporal-order memory where greater representational distance helps order memory within context but impairs it across contexts. We found that simulating signed reward prediction error integration and unsigned reward prediction error segmentation differentially enabled the model to perform associative chaining, which involved reactivating items between two tested probes to assist with sequential retrieval. In summary, our simulations provide a unifying explanation for the varied ways that neuromodulatory systems alter event cognition and memory.