I am making some research (focusing on neuroscience, not HTM). It would help me, though, to know if there are any findings in HTM regarding how spatial pooler parameters influence learning speed.
In particular: did anyone work out how learning speed changes as a function of column dimensions (usual one is 1024)?
I saw plots (e.g. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5661005/figure/F4/) where column dimension is kept constant, number of objects in the set is the variable and learning speed or accuracy on the y-axis, but I didn’t see yet a graph where column dimension is on the x-axis and learning speed (time to achieve minimum accuracy) on the y-axis.
If anyone could point me in the direction of any work which had been made, I would be thankful.
In the human model, the learning rate is modulated by several factors external to the local pooler.
It is clear that the memory has multiple time scales, from imediate working memory through to lifetime.
Learning rates are also variable from slow straight Hebbian to one-shot.
At a bare minimum, the amygdala releases learning rate modification chemical messengers.
The “searchlight of attention” enhances activation patterns and could well serve to increase the local activity and indirectly enhance learning rate.
The memory consolidation phase (during sleep and dreaming) may boost the connections of recent learing.
I just read a paper that identified a messenger that was strongly connected to memory formation and the odd thing was that it seems to be involved in shrinking of connection strength. This does make sense if you think of learning as a competitive activity.
Perhaps more on any of this later if you are interested in this line of thought.
I agree that there are several factors external to the pooler, and that they are very important.
I am researching whether there are internal ones as well (without the pretense of them being more important than the external ones).