Sensorimotor Object Recognition Using Cortical Grid Cells

Created on 2022-09-07T22:31:49-05:00

Return to the Index

This card pertains to a resource available on the internet.

This card can also be read via Gemini.

TODO see if the learning process in this paper can be used to identify using visual features

TODO come back and look in to the math for dealing with triangle grids

Existence of grid and place cells in the brain.

Combination of grid+place cells updated by interial movement and visual cues to identify current location.

Grid cells apply a triangular grid ("triangular lattice") and encode whichever dot is closest to the current position. There are only so many steps in every direction and eventually the grid repeats over again.

Quinn: in other papers they mention "grid modules" which is when multiple grid cells have the same grid size but are offset to represent sub-grid coordinates.

TODO I don't know what place cells do.

Sensation of touch seems to involve integrating position data along with feeling data from skin sensors.

Sensory features mesh together with spatial ones; familiar of one can inform the other.

Integrating Motion

Movement symbols were provided so the test network knows which direction has been taken. Then the position SDR is changed to correspond to the new input. The network is given a map encoded as symbols and movement directions.

Once trained it can be asked to find a position based on only a partial set of symbols and directions. The network was able to find a location in the learned map which is reached by the specified directions.

Grid models will keep outputs firing based on how unique the current path is. As more symbols are input then more outputs are suppressed showing only the ones that match the current path. As the path becomes more specific less positions are hot.