#Theory locations update#

By activating random cells in each module and then updating activity with movement, the network can generate a set of object-specific location representations.

#Theory locations update#

Each module independently uses motor input to update its active grid cells, and the population activity across modules can represent an enormous number of possible locations. Similar to entorhinal grid cells, these cortical grid cells are arranged into multiple modules. This mechanism relies on grid-cell-like neurons that represent the location of sensor patches (for example, the tip of a finger) relative to objects. With this missing piece filled in, we present a neural network model that learns to recognize static objects, receiving only a sensorimotor sequence as input. This new paper provides such a neural mechanism. However that paper did not provide a neural mechanism for computing object-centric locations. It then learns objects by learning sets of sensory features at locations. Our October 2017 paper, “ A Theory of How Columns in the Neocortex Enable Learning the Structure of the World,” proposed that the cortex processes a sensorimotor sequence by converting it into a sequence of sensory features at object-centric locations. The neocortex aggregates information obtained via sensation and movement, but the underlying neural mechanisms are poorly understood. Marcus Lewis, Scott Purdy, Subutai Ahmad, & Jeff Hawkinsįrontiers in Neural Circuits 13, 22.