They can beat any human at the Chinese game ‘Go’ without any human help, and match thousands of faces on photographs in just seconds, but until now algorithms have not mastered something that might seem one of the simplest tasks: spatial awareness.
Researchers at DeepMind, owned by Google parent company Alphabet, have managed to mimic the complex way our brains navigate the spaces around us using artificial intelligence. And it’s helping us to understand brains in a way we couldn’t before.
While it seems like second nature, what goes on in your brain when you walk down a familiar street – for example, navigating past obstacles, remembering shops and deciding when to cross the road or how to take a shortcut – is surprisingly complex.
Lots of brain cells play a part in working out where an animal is. For example place cells, which fire when an organism occupies a particular position, and head-direction cells, which signal the animals' head orientation.
Another crucial kind of neuron responsible for spatial navigation is called a grid cell. Grid cells are a third kind of neuron, found in the region of the brain involved in spatial learning, memories and knowledge of general facts.
Unlike place and head-direction cells, grid cells are arranged in a perfectly hexagonal shape, and they fire as an animal moves around an environment, tracking its progress as it moves. It is thought the regular shape of the connections between the cells help animals to replicate the objects and space around them, and how they are linked together, like a map inside your head.
They are constantly being updated. When the animal enters a new place, the same grid cells will start to fire all over again, working out the new environment.
Grid cells were discovered in 2005 when May-Britt Moser and Edvard Moser were recording brain signals from rats as they moved around, and noticed a surprisingly regular pattern. “The incredible discovery of grid cells showed that the brain creates maps of places by overlaying a spatial grid, something that would be very helpful in knowing where we are - providing us with something akin to a GPS signal,” says Dharshan Kumaran, research scientist at DeepMind.
But exactly how grid cells work remains a mystery. One question that remains is whether or not they help animals perform something called vector-based navigation - the ability to calculate the direction and distance between two points to select the shortest route, and to find shortcuts in a route.
Now, the brains behind the AI that beat one of the world's best players at the ancient board game of Go have brought us a step closer to answering this question.
“Everything that we think, remember, and feel is encoded in our brains in some way,” says Andrea Banino, a research scientist at DeepMind. “To understand this, we must learn how to look at sets of neurons, measure their activity, and relate this to our behaviours. However, this is very hard to do in biological brains.”
Instead, DeepMind hopes to answer questions about how the brain works, using neural networks, algorithms inspired by the way neurons work in the brain. The results are reported in the journal Nature.
The team, in collaboration with UCL scientists, developed a type of neural network which remembers the output from the previous step and uses it as input for the next. This is called a recurrent network. They input the speed, direction and head-tilt of a real rat that moved around a real maze, to simulate what happens in a rat’s brain when it moves around.
When the virtual rat was trained to find its position, the team found grid-like patterns of activity started to form in the network. Some nodes were being used more than others. These are similar to the patterns shown in grid cells in the brain of animals moving around.
After the network was trained, the next question was whether or not the neural network's version of grid cells could allow it to find the best way between two points, and it turns out they can.
“The human brain is the only existing proof we have that the sort of general intelligence we’re trying to build is even possible, so it makes sense to look to neuroscience as a source of inspiration for new types of algorithms,” says Demis Hassabis, co-founder and CEO, DeepMind.
“But we believe that this inspiration should be a two-way street, with insights also flowing back from AI research to shed light on open questions in neuroscience. This work is a good example of that: by building an artificial agent that was capable of navigating in a complex environment, we underscored the importance and extended our understanding of biological grid cells in mammalian navigation.”
The results show that artificial intelligence can be used not only to replicate the brain, but understand it better. For example, in this case it has been proven that grid cells can be used for vector-based navigation, meaning they are not useful just for working out where we are, but for calculating the best route between two points.
"Making deep-learning systems more intelligible to human reasoning is an exciting challenge for the future," say Francesco Savelli and James Knierim, from the Zanvyl Krieger Mind/Brain Institute, at Johns Hopkins University, Baltimore. The pair wrote an accompanying comment article in Nature, and were not involved in the original study.
The researchers at DeepMind hope the study will be used as inspiration for studying other parts of the brain using artificial intelligence, including the parts that understand sound or control limbs.
“This has proven to be extremely hard with traditional neuroscience so, in the future, if we could improve these artificial models, we could potentially use them to understand other brain functionalities,” says Banino. “This would be a giant step toward the future of brain understanding.”
This article was originally published by WIRED UK
