Grid cells in the brain and AI deep reinforcement learning behave in similar ways
Artificial intelligence learns to navigate in a way that resembles the working of a human brain, research by AI specialists and neuroscientists in London has shown.
The electronic components of their “artificial agent” show an activity pattern remarkably similar to the firing of specialist neurons that have evolved to help animals find their way around the world.
Scientists at DeepMind, Google ’s UK-based AI company, and University College London released their findings in the journal Nature.
Although the project has no immediate applications, they said, the results add important insights into both artificial and biological intelligence.
“It makes sense to look to neuroscience as a source of inspiration for new types of [AI] algorithms,” said Demis Hassabis, DeepMind chief executive. “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.”
At the heart of the project is the discovery in 2005 of specialist neurons called grid cells, which fire in a hexagonal pattern as animals explore their environment. These cells generate a system of coordinates in the brain, similar to hexagonal grid lines on a map, allowing for GPS-like positioning and navigation.
The DeepMind-UCL project aimed to investigate the computational functions of grid cells — how they enable the brain to calculate the distance and direction to a desired destination — which has remained a mystery in neuroscience.
The researchers built a computer network that simulated the movements of rodents navigating through simple mazes, using an AI technique called deep reinforcement learning. They found that patterns of activity very similar to biological grid cells “spontaneously emerged within the network, providing a striking convergence with the neural activity patterns observed in foraging mammals”.
Francesco Savelli and James Knierim, neuroscientists at Johns Hopkins University, Maryland, commented in Nature: “The emergence of grid-like units is an impressive example of deep learning doing what it does best: inventing an original, often unpredicted internal representation to help solve a task.”
Caswell Barry, UCL neuroscientist, said: “This agent performed at a super-human level, exceeding the ability of a professional game player, and exhibited the type of flexible navigation normally associated with animals, taking novel routes and shortcuts when they became available.”
Following the project’s success, Mr Barry expects AI to be used to test other ideas for how the brain works — for example how it perceives sound or moves limbs. “In future such networks may well provide a new way for scientists to conduct ‘experiments’, suggesting new theories and even replacing some of the work that is currently conducted in animals,” he said.