Coherently Remapping Toroidal Cells But Not Grid Cells are Responsible for Path Integration in Virtual Agents

Abstract

When animals encounter a new environment, their cells with spatially modulated activity such as place cells and grid cells remap. Grid cells are particular in their remapping in that populations coherently remap as observed in spacing, orientation and phase changes. This process points to generality in grid cell activity that forms a standard computation across environments, which many speculate is path integration. Recently, normative artificial neural network models have shown that path integration and grid-cell-like activity can be obtained in recurrent neural networks (RNN) trained to navigate in a simulated two-dimensional environment. Remarkably, the emergent spatial profile of these grid-like cells is similar to biological cell responses in that they set up a toroidal structure. However, these models have not yet been tested in multiple environments relative to biological remapping. In this work, we extend the RNN normative model to multiple environments, modelled with place cell global remapping. Through this perturbation, we show (i) that the model can navigate multiple environments, (ii) that putative grid cells remain biologically plausible, (iii) a mechanism of how grid cells may remap, and (iv) a discussion on generalisation with inertial navigation in the grid and place cell network.

Remapping in a recurrent neural network model of navigation and context inference

“…Other work has studied remapping in trained artificial networks performing navigation. (45), (49) Unlike our results, these papers typically consider remapping across different physical environments. Whittengton et al (44) propose a normative model and a neural circuit that supports non-spatial remapping, which is perhaps most similar to the task constraints we studied.…”

NeuroAI - A strategic opportunity for Norway and Europe

“…There is a plethora of examples of such fruitful collaboration beyond the classical examples reported above. To mention a few more, the mechanism of replay in biological [7,8,9] and artificial neural networks [10,11,12], dopamine [13,14,15,16] and temporal difference learning [16,17], cortical columns [18] and "single-brain" models [19,20,21], space and time navigation in the brain [22,23,24] and the emergence of grid cells in artificial neural networks [25,26,27,28]. Hopefully, to be continued.…”