Last week scientists at DeepMind, the AI research arm of Google, alongside collaborators at University College London published a paper in Nature which aims to explain how a stubbornly mysterious population of neurons called grid cells help us navigate the world.
In taking on the grid cell question, researchers at the AI firm are jumping into some representationally fascinating, theoretically baffling, Nobel-Prize-serious neuroscience.
Representationally fascinating because it’s surprisingly straightforward to tell what a grid cell’s firing stands for. When Edward and May-Britt Moser let loose the rats in 2005 and recorded from parts of the brain hypothesized to support navigation, they found a regularly spaced, hexagonal lattice of firing fields that tracked the animal’s position as it moved through the environment, not unlike grid lines on a map (here’s a solid animation).
To understand how weird and cool this is, consider that neuroscientists have made an entire subfield – referred to with slight sci-fi verve as neural decoding – out of taking videos of the brain and then using highfalutin statistics to decide what the animal was doing at that moment. But with grid cells one can very nearly eye up the firing patterns and point to the actual location of the animal when the recording was made. No fancy math required – usually the exception rather than the rule.
Yet it remains theoretically baffling because, though they know that this map is there, scientists still aren’t sure exactly how the animal is using it. For example, its not clear how the animal might plan a novel path from its current location to some rewarding morsel of food or water, nor how the grid emerges with potentially incomplete knowledge of the environment.
Serious debate about the exact function of grid cells in navigation rages to this day. The Mosers won the Nobel Prize in Physiology for their discovery in 2014.
Now DeepMind, led by software engineer Andrea Banino, is jumping into the debate. Their attempt to illuminate the inner workings of the brain is a bit of a departure for the AI outfit, whose main contributions have mostly focused on building algorithms good enough to shellac anything with a biological brain in a litany of increasingly complicated games. Famously, the computer program AlphaGo beat 18-time Go world champ Lee Sedol in 2016, but DeepMind also spearheaded work that helped give machines the edge over humans in poker, DOTA, and a large portion of all atari games currently in existence.
The hope is that researchers can look at these algorithms as approximations of the strategies employed by the brain. So, for example, an artificial rat made up of artificial neurons trained to intelligently navigate a simulated environment might give some insight into how it’s done in things with genuine nervous systems.
This is the crux of the Nature paper. The researchers input speed, direction, and head orientation – all the information a biological rat would have access to – and trained a web of neural networks to locate itself in a simulated environment. Once this virtual rat was trained, the team found a familiar pattern of activity had emerged in a layer of the network, one that was regularly spaced and hexagonal, closely emulating biological grid cells. Essentially, DeepMind’s algorithm and evolution converged on the same solution for representing space.
Which leaves the problem of getting around. This is where DeepMind’s gaming background has real import. The basic gimmick of many old atari games is moving so as to avoid hazardous objects (e.g. technicolored ghosts) and collect rewarding objects (e.g. yellow dots, animated fruit). The authors recycled an older network from this gaming research, which teaches the simulated rat to navigate purely based on feedback from environmental rewards, and glued it onto the grid cell network. The whole architecture looks really Goldberg-esque at this point, but the result is that once the self-locating grid cells are also reinforced by the presence of digital food pellets, the virtual rat learns to use the general representation of the space to draw a route through grid points to one of these rewards. Crucially, the computer rat can also map out novel shortcuts and adapt to changes in the maze-like environments.
To validate this finding, the researchers performed ‘lesioning’ studies which cut the grid cells out of the virtual rats. This, it turns out, was totally debilitating. The rats lost their ability to make smart alterations in their chosen route when faced with shifts in the maze, suggesting that grid cells are at the core of intelligent navigation.
DeepMind also sees this project as a proof of concept for the idea that Artificial Intelligence can inform the study of biological intelligence and vice versa. The logic goes like this: in designing an AI to do a task, we know there exists at least one good engineering solution, the one that the brain uses, and in studying how the brain does a particular intelligent task we can often emulate a solution with an AI made up of artificial neurons. AI and neuroscience, in this picture, are two sides of the same super-detailed, hyper-embellished coin. The authors hope that their results with grid cells in simulated rats can impress this picture upon the two communities and inspire closer ties in future work.
Monkey Bars 