Neuroscience: Artificial agent takes a short cut

A newly developed computer program has mammal-like navigational abilities, reports a Nature paper that also provides insights into the workings of the human brain.

Neural networks - computer systems that are modelled on the human brain - are capable of many impressive feats, such as object recognition, but they struggle when it comes to navigation. In the human brain, navigational skills are underpinned by specialized neurons called grid cells, which fire in a regular pattern across space to help animals keep track of their position. Andrea Banino, Dharshan Kumaran, Caswell Barry and colleagues did not set out to program grid-like cells in their computer model, but the phenomenon emerged after an artificial agent was trained to navigate unfamiliar virtual environments. As the agent tried to find its way from A to B in a maze, it became so competent that it started to take short cuts, just as a mammal would, and it outperformed a human expert who attempted the same task. Grid-like cells underpinned the agent’s notable capabilities, demonstrating that their role goes beyond providing a GPS-like position signal to planning direct routes between places.

Although grid cells have been the subject of extensive research, with their discovery leading to the award of the Nobel Prize for Physiology or Medicine in 2014, their computational functions have been difficult to fathom. This study helps to explain just how grid cells manage to encode spatial information, and offers support to the idea that grid cells play a vital role in vector-based navigation.