> The “grid units” that spontaneously emerged in the network were remarkably similar to what’s seen in animals’ brains, right down to the hexagonal grid.
Could someone with more experience in ML explain what this means? In what sense do NN cells have positions or geometry? What are the NN heat maps below the quote showing?
The neurons themselves don’t form grids. A map of the points where firing of a grid neuron in happens the real world forms a triangular/ hexagonal grid.
These neurons seem to have discovered what board gamers found out much later - hexagonal grids are better for calculating movement.
Could someone with more experience in ML explain what this means? In what sense do NN cells have positions or geometry? What are the NN heat maps below the quote showing?