Artificial Intelligence Taught To Think Like A Mammal Finds Its Way Through A Maze

The neural networks not only figured out the maze, but also began taking shortcuts. AlexLMX/Shutterstock

Josh Davis 09 May 2018, 18:00

For the first time, scientists have developed an artificial intelligence that can navigate in a way similar to that seen in mammals.

The extraordinary research, published in Nature, not only improves our understanding of these advanced computer programs but may also help shed light on how we ourselves navigate the world.

“Navigating a route between two places might seem like a simple, everyday task, but – strange as it might sound – nobody knows quite how we do this,” DeepMind’s Dharshan Kumaran said in a statement. While this navigational flexibility might seem natural and instinctive, the underlying process is actually incredibly complex.

To understand what is going on here, you first have to understand what grid cells are.

When an animal maps a room for the first time, there are certain nerve cells in the entorhinal cortex of the brain that fire in a regular pattern. When the firing of these cells are mapped in the brain, it creates a grid-like, hexagonal pattern – thus these cells became known as "grid cells". It has been theorized that these cells help animals determine spatial distances and take more direct routes.

The experiments with artificial agents produced grid-like patterns strikingly similar to biological grid cells in foraging mammals. Banio et al. 2018

The grid cells also coordinate with “place cells”, which have been found to fire when an animal reaches certain points in an environment. Together with the grid cells, they help the brain construct maps and enable what we term a “sense of place”. The discovery of both place cells and grid cells (and their role in our internal GPS) was awarded the Nobel Prize in Physiology and Medicine in 2014.

The researchers wanted to test whether or not these grid cells genuinely do help determine spatial distances, but rather than turning to biological models, they opted instead to use artificial neural networks. While these have been able to do some incredible things – such as recognize individual objects – they have tended to struggle with navigating spaces.

“We created an artificial system, which reproduced important features of how the brain support navigation (i.e. the set of neurons), then we used this system to test a hypothesis about how this is translated into how we navigate the world (i.e. the behavior),” explains Andrea Banino, a research scientist at DeepMind, in a statement.

Incredibly, when they set the program to work its way through unfamiliar virtual environments, a familiar pattern started to emerge. They found grid-like representations spontaneously appearing in the network that looked strikingly familiar to how grid cells fire in mammalian brains.

As the artificial agent started figuring out how to get from A to B in mazes, it not only worked it out (beating humans in the process), it also started taking shortcuts – just as an animal would.

Comments

If you liked this story, you'll love these

This website uses cookies

This website uses cookies to improve user experience. By continuing to use our website you consent to all cookies in accordance with our cookie policy.