“Vector-based navigation using grid-like representations in artificial agents” was published in Nature on 9 May 2018 by Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap, and colleagues at DeepMind and University College London. It is a notable case of artificial intelligence and neuroscience meeting in the middle.
In the mammalian brain, navigation depends in part on grid cells in the entorhinal cortex, neurons that fire in a regular hexagonal pattern as an animal moves through space. Their discovery by Edvard Moser and May-Britt Moser, building on John O’Keefe’s earlier work on place cells, earned the 2014 Nobel Prize in Physiology or Medicine. Grid cells are thought to support path integration: keeping track of position by accumulating self-motion.
The DeepMind team trained a recurrent neural network to do path integration from movement signals, the same job grid cells are believed to perform. Without being told to, the network developed internal units whose firing patterns closely resembled biological grid cells. When this trained network was combined with deep reinforcement learning, the resulting agent navigated efficiently through unfamiliar and changing virtual environments, even taking novel shortcuts, supporting the theory that grid-like codes enable vector-based navigation toward goals.
The result cut both ways. For AI, it produced an agent that navigated better than baselines lacking the grid-like representation. For neuroscience, it offered computational evidence that grid cells may emerge because they are an efficient solution to the navigation problem, not merely a quirk of biology - an example of an artificial network helping test a hypothesis about a real brain.