I’m trying to figure out HOW, IF and WHAT kind of sequence similarity does the TPooler “preserve”.
From what I understand the TP is “glorified” UNION and is not clear when different similar sequences are encoded/merged by TP.
Do they preserve similarity once encoded as SDR ?
But then the question arises what similarity ? Lets take for example the following sequences :
ABC : ABX, AXC, BAC, CAB, XBC, ...
which one is the most similar to ABC ? which one are equally similar ?
In practice it is often the Minimum Edit Distance that is used !
Is this a feasible similarity measure ? If not what , why not ? Which one, then ?
The TP goal is to be unique and stable, but should also preserve similarity ?
The same way SP preserve SDR overlap, TP should preserve MED/? as TP-SDR overlap, right ?
As far as I know, there is no official implementation of a “temporal pooler”, just experimental research code. If you are talking about the union pooler logic in the research repo, I’m not sure anyone is actually working on this anymore.
is there other document about it, except this one : https://github.com/numenta/htmresearch/wiki/Overview-of-the-Temporal-Pooler
Especially interested in the voting algorithm between multiple TP’s ?
The closest thing I can think of is this:
lateralInput is where something like voting would occur.
That’s a really good point that somehow escaped me when I was testing a TP model… but I wonder how to incorporate it. Since the TP reads the cell activity directly, cells in different columns and cells in the same column appear equally unique to the TP. Perhaps appending an SDR of the column activity to the cell activity in the TP’s input would provide some similarity measure?