Conference on Cognitive Computational Neuroscience

We investigate the role of cognitive maps and the hippocampal-entorhinal architecture in mental navigation (MNAV) by building a neural network model. The model uses a continuous-time recurrent neural network (CTRNN) for action decisions and a hippocampal-entorhinal model network, MESH (Memory network with Scaffold and Heteroassociation), for encoding and learning maps. The model is trained on a navigation-to-sample (NTS) task and tested on NTS in a MNAV setting (no sensory feedback) in five different environments (image sequences). The CTRNN with MESH tackles MNAV by reconstructing the next image via path integration and vastly outperforms the CTRNN alone in both tasks, showing better generalization to unseen pairs within each environment and faster adaptation to new environments. The study demonstrates the importance of hippocampal cognitive maps in enabling data-efficient and generalizable learning in the brain.