Structural generalization and continual learning enabled by factorized entorhinal-hippocampal memory and entorhinal-parietal action circuits

Flexible behavior requires generalizable memory and learning. For example, we rapidly learn to commute in new cities by reusing our knowledge of Euclidean two-dimensional space and structures like roundabouts and subway systems without forgetting how to get to a favorite restaurant back home. Yet we lack a detailed understanding of how the brain uses existing knowledge to generalize while retaining the memory of specific past experiences. To address this gap, we combine behavioral measurements, neural recordings, and computational modeling in an abstract sequential image navigation task to study three forms of generalization: mnemonic generalization, from visual to mental navigation; transitive generalization, from trained to novel routes; and structural generalization, from familiar to new environments. In contrast to monkeys and humans, recurrent neural networks failed at all generalizations. We found that a structured entorhinal-hippocampal memory model, which provides a content-independent metric scaffold based on grid cells for storing experience, coupled to a policy recurrent network, succeeds at all three. The content-independent scaffold enables mnemonic and transitive generalization through path integration and facilitates structural generalization by allowing reuse of a previously learned action policy network. Moreover, the scaffold's high combinatorial capacity permits continual learning without catastrophic forgetting. We recorded neural activity from the entorhinal cortex and posterior parietal cortex of two monkeys performing the task and found two distinct computations across the neural population. Modularizing an entorhinal and parietal action policy network to separately track distance and initiate actions captured the distinct population dynamics and improved model performance. Finally, we added a reinforcement learning module to the network that enabled it to learn an appropriate scale factor to align the grid periodicity with the environmental temporal structure. Our findings reveal that an architecture which factorizes invariant metric representations from rapid sensory associations and a transferable policy learns, generalizes, and remembers like the brain.