Question about Sparsity in Local Inhibition

question
inhibition
sparsity

#1

Hi,
I also have problem with the sparsity in Local Inhibition. Because I’ve read the code of nupic in github: https://github.com/numenta/nupic/blob/master/src/nupic/algorithms/spatial_pooler.py

Inside this, the targetDensity is calculated based on the inhibition Radius. While, the Inhibition radius is base on the average span. For example, when I have 6464 columns -> the sparsity would be 82 (0.02 6464) columns activated. However, after calculating, the density for each Neighborhood is calculated 0.5 . That leads to the sparsity in whole region is nearly 0.564*64 columns activated. I have some confusion about this point that can we actually control the sparsity in the local Inhibition.


Ideal density of SDRs
#2

You can’t strictly control the global sparsity with nupic local inhibition implementation as you realized. You need an alternative approach such as this.