Unveiling the structure of the grid cell code in novel environments

Understanding how the brain builds maps of novel environments on behavioral timescales (seconds-minutes) is a fundamental question for neuroscience. Theoretical works posit a Hebbian plasticity mechanism for integrating novel landmarks [1]. Drosophila have been shown to use such a mechanism to integrate novel landmarks on timescales of minutes [2,3]. However, current evidence suggests that plasticity in the medial entorhinal cortex (MEC) operates on the timescale of days [4] — leaving open the question of how the mammalian brain maps novel environments on behavioral timescales. Here we investigate this question using Neuropixels recordings of hundreds of neurons in mice navigating diverse virtual reality (VR) environments. We find that MEC is capable of building stable, consistent maps of novel environments after just a single exposure (one-shot). Novel spectral analyses reveal that this is accomplished by visual landmarks driving a bump of neural activity onto stable periodic trajectories on a toroidal attractor. However, visual landmarks also distort the grid cell code so that distances in real space are no longer proportional to distances on the neural manifold. Nevertheless these distortions are predictable, leading to an underlying structure which persists across diverse environments. Remarkably, this structure is sufficiently rigid that a simple 2-D dynamical model allows us, for the first time, to predict grid cell responses in a novel environment before the animal enters that environment.
Together, these results point to a picture of MEC as learning through the dynamics of a fixed circuit, capable of one-shot mapping novel environments at the cost of susceptibility to distortions and spontaneous remapping. Despite these distortions, we show in behavioral assays that rapidly generated MEC maps enable mice to one- or few-shot learn to navigate to hidden rewards in novel environments. These results lend credence to the promise of “learning through dynamics” to subserve flexible behavior.