ETM suggestion: Use apical segments when determining the learning cell

[mrcslws]

Relevant code: https://github.com/numenta/nupic.core/blob/8a846d1d897440bc6ca1e46120520da166bc551a/src/nupic/experimental/ExtendedTemporalMemory.cpp#L616

I didn't change the algorithm at all, I just put all the logic for "learning cell" selection in one place.

When a column is bursting, we choose a cell to learn on. The current selection process is:

• if the column has matching basal segments
  • use the cell with the best matching basal segment
• else use the least used cell.

We should start also considering apical segments. Here's a trivial addition:

• if the column has matching basal segments
  • use the cell with the best matching basal segment
• if the column has matching apical segments
  • use the cell with the best matching apical segment
• else use the least used cell.

We could go further and give preference to a cell that has both a matching apical and basal segment. Or maybe a different change makes sense. (e.g. allowing multiple learning cells per segment in some cases?)

I haven't thought through the effects on experiments. It just seems pretty clear that this code is incomplete.

[floybix]

The comportex logic selects a learning cell jointly with its learning segments (zero or one each of distal & apical segments). This is not just for bursting, it covers all cases including where one or more segments are fully or partially matching -- i.e. meeting stimulus or learning thresholds.

This is by no means a complete solution and has not really been tested on anything except my "sequence replay" example. But food for thought perhaps.

Here goes

• if there is exactly one fully matching basal segment in the column, select that cell and segment. The apical segment in this case is the best matching one on that cell.
• if there are multiple fully matching basal segments in the column, select from those by the best matching apical segment if any (choosing randomly between multiple fully matching apical).
  • if no apical segments match, choose randomly between the basal segments.
• if there are partially matching basal segments, select the best matching cell & segment. The apical segment is the best matching one on that cell.
• if there are no partially matching basal segments, select by the best matching apical segment if any...
  • however, this apical-only matching only happens with some probability ("apical-bias-frac") - if it always happened then we could get identical representations of the same thing in different sequence contexts, assuming a common pooled context. By doing it with a probability we can generate overlapping but distinct representations. https://github.com/htm-community/comportex/pull/38
• fall back to random cell selection from the active cells in the column.

Code: https://github.com/htm-community/comportex/blob/863b68af4815d086803d2bf52676df5f699c51da/src/org/nfrac/comportex/cells.cljc#L498