On the Computational Role of Predictive Grid Cells

The computational role of grid cells in medial entorhinal cortex (MEC) has been the subject of intensive investigation over the past two decades, with dominant theories attributing them to supporting path integration and serving as a basis for predictive coding. Recent electrophysiological recordings of MEC in rodents have found a subpopulation of neurons that have grid-like tuning, not with respect to where the animal currently is, but where the animal will be. These ``predictive grid cells'', which are distinct from the subpopulation of grid cells, appear consistent with the hypothesis that MEC supports predictive coding. However, given that existing predictive coding theories have been used to understand grid cells, it remains unclear what the functional significance of predictive grid cells is. To begin to answer the question of what computational role predictive grid cells play, we first analyze an independent, previously recorded, MEC dataset, finding robust evidence of predictive grid coding. This suggests that predictive grid cells are a general feature of MEC. We then interrogate the emergent representations of recurrent neural networks (RNNs) trained to path-integrate. We find that these RNNs, which have previously been shown to develop units with grid-like representations, also develop units with predictive grid-like representations, even though the RNNs are not trained with a predictive objective. We validate the predictive grid tuning of these units by finding that they exhibit significant path-dependence, even when controlling for the same allocentric spatial position. In addition, we find that band-like units in the RNN, which have representation analogous to band cells in MEC, exhibit greater predictive grid tuning than expected by chance. Preliminary evidence shows that ablating predictive grid units in the RNN leads to changes in the autocorrelation of standard grid unit activity, particularly at longer time-scales. Collectively, our results challenge the idea that predictive grid cells must necessarily be used for explicit prediction and planning, and instead may be involved in shaping the population level activity of grid cells.