Minimum Grid Cells for Place Cell Activity in Eleven Enclosures

Abstract. The goal of this work was to uncover the minimum number of grid cells that can re-create the firing patterns of place cells in eleven different enclosures. This is useful for understanding firing patterns of place cells to design a very compact neurally-inspired spatial navigation system in hardware. For the project, we took place cell firing data from the Moser lab. Of the 342 place cells recorded, we selected one of the place cells firing in a single enclosure, one in six and one in all eleven enclosures. In MATLAB, we created grid cell firing patterns for 11250 grid cells. Connection weights between the place and grid cells were learned using machine learning- gradient descent algorithm. In all three cases, the place fields could be learned from at least 36 grid cells with three spatial firing frequencies, 2 directions for each frequency and 6 different phases for each direction. Including fewer phases reduced precision in deciphering the place. Weights learned were normally distributed with a wider spread and multimodal distribution for rooms with uneven, larger or multiple firing fields and narrower distribution otherwise. In rooms with no firing, all the weights varied a little around zero.