Spatial Pooler - Simple and minimal implementation of the Spatial Pooler by Numenta

If you want a deep dive into the theory behind the Spatial Pooler:

HTM School: https://invidio.us/watch?v=R5UoFNtv5AU

https://invidio.us/watch?v=rHvjykCIrZM

Numenta Homepage(s): https://numenta.com/ https://numenta.org/

Blog post: https://numenta.org/blog/2016/02/19/mathmatical-formalization-of-htm-spatial-pooler/

How to use this

If you are in this Folder:

from spatial_pooler import SpatialPooler, overlap
import numpy as np

sp = SpatialPooler(50,100)
output  = sp.run(np.random.randn(50) > 0)

print(output)

The most important method in the SpatialPooler class is SpatialPooler.run(). It takes a numpy bool array, that represents a SDR and returns another numpy bool array that is the output of the SP.

The SP always learns.

More examples and documentation in the .ipynb files in the examples folders.

The Architecure and meaning of State -> Potential Pool -> Permanence -> Activation -> State'

As you may have noted, the description of this Repository is State -> Potential Pool -> Permanence -> Activation -> State', but what does this mean? If your are more into SP Theory, you may noted that this is a picture of the SP Algorithm. First there is a State, a SDR that is also the input to the SP. And then for each column in the SP, the Algorithm first ignores those parts of the state that are not in the potential pool, then those for which the permanence is less than a threshhold. After these two extractions, the algorithm counts the active parts of the state left and tests again whether that count is greater than a threshhold. If so, the collumn upon which these steps where executed is activated and the output is activated.