Grid cells - Caswell Barry, UCL

Key Takeaways

grid cells are type of brain cell and first of all they were discovered in rodents but now we know they exist in mammals in general they seem to be very important for sort of measuring out space our good analogy is that they're like the grid lines on a map except on a map we've used square grid lines and the brains using hexagonal grid lines and you have small scale grid lines and larger scale grid lines originally people used to think they were very important force that we call path integration which is basically updates where you think you are based on how you're moving so I know where I am right now and if I move it 10 meters then I end up somewhere over there but actually we're moving beyond that now while we still think path integration is important focusing on the fact that we believe they function like a map they sort of relate one place to another they also seem to be very important for planning routes so if you wouldn't know the shortest route between here and somewhere else that you've visited then the grid lines would provide you with something called a vector based navigation and allow you to calculate the direct line distance with the recent paper we were trying to do several things so first of all we were trying to test this idea of vector based navigation so our grid cells really the substrate and machinery that allows you to sort of calculate these direct line routes between places but equally we're trying to do think a little bit more now we were trying to see if you could take what we know about how the mammalian brain works and incorporates it into artificial agents to improve the way that they work in this case the way they navigate first of all we were able to show that if we trained an agent or a network to paths integrate then we found that actually developed grid cells so this is remarkable convergence of form in that they looked like red cells in the mammalian brain they had the same hexagonal pattern with more tightly tessellated fields and then we were able to take that network and transfer into a larger network that was now controlling an agent navigating around and that agent was then able to navigate flexibly and demanding virtual environments and so for example one of the tests that we wanted to see whether to do is whether it could take shortcuts we had eight environments that would change while the agent was in there so it might find one long route to the goal but then a new door would open up presenting a shortcut so the agent with grid cells could utilize that shot go straight through to the goal so this is important for several reasons we've confirmed that grid cells are important for database navigations we've shown that we can use an artificial network to conduct neuroscience experiments but also we've shown that incorporating what we know about how the brain solves a problem into an artificial network can provide that network with abilities that wouldn't have otherwise have had [Music]