The brain represents and transforms external variables to accomplish goals. Representation and transformation are inherently noisy when performed by neurons. One way to extract a less noisy estimate of the encoded variable is by averaging over neural populations. Population coding is widely used by human observers to estimate the encoded variable, and has been interpreted as a model of how downstream brain areas may readout the variable, but leads to only modest improvements in estimation for the number of neurons involved.
Is there a better way?
I will show that the entorhinal grid cell code for animal location is capable of essentially exact removal of noise from noisy estimates of the animal's position, in contrast with averaging-based noise reduction.
Noise removal is an intrinsic self-checking process enabled by the peculiar structure of the grid code, and does not rely on the existence of external cues. I will describe how the grid cell code is computationally equivalent to minimum-distance separation codes. I will show how a simple neural network model of the EC-HPC-EC loop performs this noise removal for improvements in accuracy easily exceeding 10^4 compared to the best possible classical population code using the same number of neurons.
BACK
