Definition of convergence

[Oblynx]

Right now convergence is defined as the activation not changing by >5%. However in the source material, convergence means "given a set of neurons from each consecutive activation, this set stops adding new neurons".

[Oblynx]

I implemented this "amortized convergence" in 57f05be
However, it doesn't apply cleanly to the HTM implementation, because the minicolumns are initially bursting, meaning that many more neurons are initially recruited, then gradually pruned from the representation. The set of previously activated neurons is therefore full from the beginning.

In this implementation, I used a rolling window of t=5 steps for which firing neurons to include in the convergence set. The resulting metric didn't stabilize much better than looking only at the last activation.

[Oblynx]

Alternative convergence metric based on ¹: stable firing rates for most neurons after the stimulus.
I can calculate firing rates over a short rolling window (eg t=5). The expected behavior is:

• assembly neurons should have consistently high firing rates
• non-assembly neurons should have consistently low firing rates

The convergence metric should capture the distinction between the 2 groups.

Footnotes

1. C. Gastaldi, T. Schwalger, E. D. Falco, R. Q. Quiroga, and W. Gerstner, “When shared concept cells support associations: Theory of overlapping memory engrams,” PLOS Computational Biology, vol. 17, no. 12, p. e1009691, Dec. 2021, doi: 10.1371/journal.pcbi.1009691. ↩

[Oblynx]

Could also try the similarity measure in equation (2) from reference [1].