Learning to imagine: Using deep learning and memory-augmented neural networks to model cortical-hippocampal interaction during mental simulation

Abstract

Mental simulation is a ubiquitous feature of daily cognition. Lesion, behavioral, and neuroimaging studies have established strong parallels between imagining new experiences and episodic recall. However, current research provides a less thorough account of the differences between recall and simulation. Additionally, prominent verbal theories do not offer a computational-level understanding of how episodic recall is used in the service of mental simulation. In addition to servicing recall and mental simulation, episodic memory also assists in processing events by helping to fill in a predictive model of the ongoing situation. Given the deep similarities between mental simulation and event cognition, we built a model of mental simulation based on an existing memory-augmented neural network model of event processing. We designed a new objective function and task for the model. The model’s objective was to maximize reward by generating long sequences of events and avoiding uncertainty. We hypothesized that the model would learn a new, more permissive recall policy, allowing it to use recalled details from episodic memory to generate long sequences. Our model learned to generate longer sequences; however, it did not use details from episodic memory in simulated sequences and did not converge on optimal behavior. We detail follow-up experiments to understand why changing recall policy was insufficient and why details from episodic memory were not incorporated into simulated sequences. Broadly, our model helps lay the groundwork for a novel, computational-level account of the role of episodic memory in mental simulation and begins to elucidate the necessary conditions for successful simulations.