288 — Vector-based navigation using grid-like representations in artificial agents

Banino, Barry, & Kumaran et al (10.1038/s41586-018-0102-6)

Read on 04 June 2018
#neuroscience  #deep-learning  #machine-learning  #neural-network  #entorhinal-cortex  #place-area  #grid-cells  #place-cells  #navigation  #group:google 

This research marries biological spatial navigation (#navigation) and the role of entorhinal grid cells (#grid-cell) with machine-learning approaches: In these experiments, a simulated agent moves through a simulated environment, receiving orientation and location cues (similar to how visual cortex might transmit information to downstream cortical targets in the brain). The trained network exhibited areas of hexagonally periodic units that responded heavily to spatially distributed information.

In other words, it appears that grid cells naturally arose when the neural network was trained to understand movement and location in an analogous way to how the mammalian brain understands movement and location.

The researchers retrained the network several times over in order to perturb the network and verify that these units arose in a robust manner, rather than arising stochastically or coincidentally.

Fascinatingly, this grid-cell organization also enabled the agent to accomodate changes in its environment. When ‘doors’ were closed (to represent temporarily blocked passages), the agent learned to avoid those paths and look for alternate routes. And when doors were opened, the agent learned to use the closest passage in order to most efficiently arrive at its goal.

This is particularly exciting because it means that hexagonal grid layouts — like those found in cortex — are not some biological coincidence, but a real, significant, meaningful property of encoding space and relative distances in an easily accessible, distributed manner.