results with SP w numActiveColumnsPerInhArea

[breznak]

now the params are directly comparable to Numenta's.

[breznak]

FYI @steinroe this branch enables the old/new reintroduced param numActiveColumnsPerInhArea.

Note: it needs the specific branch from htm.core.

• currently it underperforms our "tuned" params with localAreaDensity option
• do you have somewhere what are the exact params for NumentaTM, Numenta detectors?

[psteinroe]

See htm-community/NAB/pull/28. Still underperforms though. I would propose we try to

• optimise on these too
• use the fake spatial anomaly detector Numenta used on both, the localAreaDensity version as well as the numActiveColumnsPerInhArea one to see if we can reach the NAB score numenta does.

What do you think?

EDIT: Or maybe go back to check wether NAB is working correctly?

[breznak]

use the fake spatial anomaly detector Numenta used on both

yes, let's do that first. That way we're as close to numenta as possible.

• it just seems the numActiveCols is not a silver bullet (atleast w/o any work)
• so we can't (again) say tuning params for either would bring us better scores
• still need to decide if one is "biologically" better
• :+1: for tuning both sets of params

[breznak]

Should we merge this branch into fixing_...? Both params can now be used. Currently localAreaDensity > numActiveCols

[psteinroe]

I would propose to merge it and then creating a new branch for adding the fake spatial anomaly detector, as its probably easier to work with 2 branches instead of 4 when running the optimisations.

[breznak]

a new branch for adding the fake spatial anomaly detector

Let's add that just as a commit as well, fake spAnomaly can be controlled with a toggle in the detector

[psteinroe]

are you working on adding the fake spAnomaly? otherwise I would do it 👍

[breznak]

are you working on adding the fake spAnomaly? otherwise I would do it

please go ahead. for a couple of hours I'll need to do sth else