Voting on movement

In this Numenta video Jeff Hawkins on Object Modeling in the Thousand Brains Theory (Part Two) - September 9, 2021 - YouTube the folks touch on movement voting. Each column has sends out a motor command, but if lots of columns are all doing this at once, how do you decide which motor command to follow? It was theorized that their might be voting of some kind to decide which motor commands to send out.

I was wondering if the motor output command would be one that reduces the number of possible objects that something could be by the biggest factor. That is, in the world of sight, the eye saccade that would reduce the most amount of ambiguity.
In goal oriented behaviour it would be different, as you would want the moment that gets you the closest to your goal.

Just random thoughts.

2 Likes

I think there-s a rather competitional decision process than voting.
Like a team in which each player has its own local, limited set of competences and situation awareness - about the goal, current state and what that individual should do - which is inherently different different from other teammate (because they have different inputs and outputs) and each decides about their own course of actions which coordinated with every other teammate actions would optimally achieve the goal.

This coordinated team process may be confused with voting when seen from the outside. Specially when you can not see/comprehend exactly where each signal out of millions comes from and where it goes to.

Here-s what I believe is problematic with voting:

  • Voting is a slow, cumbersome and imperfect process.
  • most often it requires several alternative options upon which action to take otherwise why vote about the only action that can possibly be done?
  • Which begs the question, where the proposals of alternative actions come from?
  • it requires equal, accurate awarness of both input space (current state) and potential actions space among all voters and also equivalent levels of competence otherwise chances to reach a majority in real time are quite slim. Look at any animal fighting with an adversary, there is little time to vote about what each muscle fiber signal should be every few milliseconds.
  • it might be useful for more slow processes like planning or learning from previous experience but even then a coordinated team of players each with its own special skill might be more effective.

This sounds like a perfect fit to a theory about (at least a subset of) epilepsies… when the brain fails to decide which motor command to follow, so the muscles just follow all?

It seems that the voting on action happens in subcortical structures.

From the linked thread:

“Action plans from cortex are transmitted to basal ganglia where they compete with each other, weighted by info on value of predicted outcomes, relayed in a loop through thalamus, back to cortex, until a winner is settled on.”

“Signals from basal ganglia to motor command centers in midbrain then “release” the winning action, while keeping all others inhibited.”

6 Likes

You may be giving the columns too much agency. They are not sitting around waiting for a consensus before making a final decision. I tend to think of the voting process more like clocks on a wall gradually syncing up (except much faster). They are continuously processing input and context while generating output. Cross-column voting provides a mechanism for them to quickly align their representations and their outputs to maximize their cumulative influence on the downstream regions.

3 Likes

That would make sense. From an evolutionary point of view these are capabilities that should predate higher structures, and from a timing point of view speed matters.

1 Like

Perhaps voting works upstream as well as downstream ? If I take several columns involved in different senses(touch, sight, sound etc) and they’re all picking up signals consistent with the appearance of a dog, then the voting process could run up the hierarchy to activate a column which is the invariant representation of a dog. (I hadn’t thought of this as voting, but the mechanism seems similar).
The mechanics of voting presumably come down to enough cross-column axons firing together, in order to reach the activation threshold for their target (or have I misunderstood how voting works in practice ?).

Frankly, that doesn’t sound like voting at all. Alternative explanations compete with each other until ambiguity is somehow solved.

In most competitions you have an obvious and clear winner who knocks out the opponent. When I’m not sure I see a dog or cat I look for cues. If it meows, the dog is suddenly knocked out of the guessing game.

If there were voting I would consider stuff like “95% of the cats are about same size while 85% of dogs are not so final vote is it has to be a cat”.

No, we don’t do that. We work with certainties when we-re not certain then a more slow, complex processing happens (pondering) which might end with a voting, or split decision if you want.

I’d rather consider that maybe each “column” seeks and develops a narrow niche of particular competences, becomes a specialist upon certain feature/property/thing and that thing’s possible relationships with other features (by strengthening connections with their respective “columns”), and its opinion is weighted with its “expertise score” built upon the record of previous correct/failed predictions.

PS the “every column has a complete model of the cup” (and mouse, grass, number Pi, coronavirus and anything we can think of) and they reach some consensus knowing all what they-re looking at makes little sense to me.

I’m not sure if voting is a good word for that.

I am interpreting that to sound more like some sort of attractor that “latches” onto an action program.

To expand on this, here are some more articles:

Essentially, there are two pathways through the Basal Ganglia. One pathway processes good / rewarding things and the other processes bad / punishing things. The two pathways compete against each other and whichever pathway wins determines whether you take action or hold still.

If there is voting on movement, I think we’d be dealing with a circuit like this:

  1. Motor output cells don’t do any voting, and they aren’t simply motor output cells (displacements? representation shifts?)
  2. The cells voting project to striatum and to cortex in a wide-reaching manner. (They don’t project to other subcortical structures, just striatum).
  3. The voting cells project to motor output cells, which project back but mostly just through thalamus. This is probably a fairly strong loop, with the motor output cells perhaps initiating proximal inputs to the voting cells.
  4. Again, they’re not simply motor output cells. They’re somewhat mysterious.
Details.

(General)
Motor output cells are L5tt cells (thick-tufted layer 5 cells). The cells which I think are the voting cells are L5st cells (slender-tufted).

(1)
L5tt cells don’t seem like they’d vote because they don’t project much to other cortical regions. They’re not like L2/3, L5st, and maybe L6cc, which are good candidates for global voting. They could still do low-range voting, e.g. voting between columns in a single cortical region or maybe a few regions.

L5tt cells aren’t simply motor output cells. They’re directly activated by the thalamus, they’re much like other components of the cortical circuit (lots of them, minicolumns, etc.), and they’re the only way for cortex to pass info through thalamic attention mechanisms.

(2)
L5st cells project to the striatum and have long-range corticocortical projections. They’re fit for reward-related voting.

L5tt cells (the motor output) very likely don’t just send motor commands (e.g. they conflate motor / sensory and they’re not so different from other cortical cells). So I think they don’t simply vote on movements.

(4)
Why do L5tt cells fire besides to generate movements? Numenta theories have suggested they’re displacements (between locations). I think they’re for something more general, but still shifts in representations, so a superset of displacements.

That’d be useful. It might be broader, since there’s more than object recognition, at least sequences (“what song am I listening to? I’d better pay attention to certain things and turn my head or whatnot.”).

I’d imagine a single cortical column isn’t enough for generating intelligent plans, so they might still need voting.
Also, kinda like how voting reduces ambiguity of the object, location, or whatnot, there’s ambiguity in the form of a blame game. L5tt cells don’t initially know what actions they cause, nor do cells sending signals to L5tt cells. Maybe voting could help with that blame game, although I’m not sure how.