The thinking that led to these ideas / credit to dmac
This is based on dmac’s idea in Should grid cell or displacement cell modules include minicolumns?
I wrote some ideas in that thread, but I’ve done some more thinking and that was stream of consciousness writing so it takes a while to get to the point.
I’m not focusing on the exact mechanisms of path integration, but let’s say it really does need to use minicolumns for location, with grid cells being the minicolumn states. That seems likely because otherwise, it doesn’t seem like it would have enough union capacity. One of the ideas in the recent papers is narrowing down possible current locations, which requires unions of SDRs. Let’s also say grid cells are in layer 6.
There are only three excitatory cell types in L6 as far as I know. Corticothalamic cells aren’t a good fit because they modulate the thalamus likely for attention, which grid cells probably don’t do. If they are grid cells, they can’t do much with minicolumns because they don’t synapse on other excitatory cells much in L6. Corticoclaustral cells are few in numbers so they probably can’t represent unions. That only leaves corticocortical cells.
These ideas were based on dmac’s ideas in other ways but I’m trying not to focus on exact mechanisms.
These are the roles I think each cell type in L6 serves. There are subtypes and these aren’t meant to be too specific.
Most generally, they do something related to surprise, maybe nearly the same thing the thalamus does. I think they switch sets of possibilities during surprise. They could do more specific things, like representing those possibilities.
They produce path integration signals within the cortex and do prediction evidence accumulation for the thalamus. For narrowing down sets of possibilities which change over time (like location), those two things are not contradictory.
Corticoclaustral and Intracortical
Object Recognition in General
To represent the positioning of a feature on an object, it needs to know that feature’s allocentric orientation, not just its location, by which I mean translation. This might mean there are two parallel systems, one for location and one for orientation.
When it is narrowing down sets of possibilities, it might encounter something which fits none of those possibilities. When that happens, it either hasn’t learned the object or it has moved from one object to another. In that case, it needs to add a bunch of possibilities and maybe remove the previously represented ones.
That’s similar to the layer which narrows down possible locations of the sensory patch. Each movement, it needs to completely change those possible locations by path integration. My point will make more sense when I describe ideas about L6 CC cells.
Thalamus Burst/Tonic Mode
This section is just to make a comparison with L6 CC cells, but I want to disagree with the common opinion that the thalamus is a gate.
I agree with interpretations that burst/tonic mode in the thalamus is for surprise. If it doesn’t expect something, the thalamocortical cell is in burst mode, so if it activates it will send a stronger signal to the cortex. I think it still can choose to attend things or not, based on the size of the depolarizing signal, which I think of as a prediction signal even when it is just predicting that the thing it is looking at will still be there during the next time step.
To gate things, I don’t think it just activates the TRN. The TRN itself has burst and tonic mode. The bursts are much longer than those of TC cells, so they’re supposedly suited for activating GABAB receptors. That means when the cortex depolarizes TRN cells for a little while, it might actually reduce the inhibition of TC cells. Also, although the cortex sends initially stronger signals to TRN cells than TC cells, those synapses are small and weaker than the signals from TC cells to TRN cells. TRN cells have gap junctions and inhibit each-other, which complicates things further. Also, burst and tonic mode are on a continuum.
All of this is really weird if the thalamus is simply a gate controlled by the cortex. If that were true, it wouldn’t have burst and tonic mode. I think the thalamus alone filters for new inputs, and with the cortex evolved on top, it filters for things which the cortex didn’t expect, sort of.
If the cortex has no representation of a possibility (temporal memory predictions, possible objects, or whatever else), the TC cells will burst if it senses that thing. If it has a little representation, it isn’t attending that thing but it knows it might be there or is there, so the thalamus filters it out. If it has a lot of representation, it is attending it. Since the signal to the thalamus is strong, even though the cells are in tonic mode, the signals are enhanced because the depolarization is large. (By amount of representation, I mean the number of cortical cells being used to represent the thing or their firing rates).
L6 Corticocortical Cells Switch and Maintain Sets of Possibilities
These cells are comparable to TC cells. They have initially elevated firing rates during current injection and I assume during normal functioning too. Like TC cells, they respond most strongly to the sensory stimulus if they are predictively depolarized a lot or not at all, which are like attending it or not expecting it. If it attends something, they could cause the set of possibilities to remain active, and if is is surprised, they could generate a new set of possibilities.
They could also do other things and they could represent those possibilities.
There are multiple layers representing sets of possibilities (one for objects, one for locations, and maybe one for displacements in location, one for orientations, and one for displacements in orientations).
There are several types of L6 CC cells, seemingly all with initially elevated firing rates. During surprise, one L6 CC type could switch the set of possible locations, one could switch the set of possible objects, and so forth.
They receive stronger input from the thalamus than L6 CT cells, although maybe just because L6 CT cells fire sparsely. However, CT cells are often completely unresponsive to simple sensory stimuli, so it’s more likely that L6 CT cells fire a lot in some circumstances.
This could have evolved before all the pieces of the location system were in place because representing sets of possibilities is useful for other things.
Cells Downstream of L6 CT Cells do Path Integration
L6 CT cells have facilitating outputs, and that facilitation decays slowly. That means as the movement progresses or for each new displacement, their signal will be stronger, which could do path integration.
In visual cortex, they receive a head rotation signal, which inhibits or excites them proportionally to the speed of rotating. They can path integrate rotation postsynaptically. They also activate basket cells, including with facilitation in other layers but not L6 basket cells if I recall correctly, which could convert each different signal size into a different representation for each different orientation displacement.
If they cause path integration for orientation and location separately, one of the downstream targets would probably be L4 and the other would probably be L5 slender tufted cells, based on the subclasses I’ve read about. Slender tufted cells seem weird to me (they project to the striatum but are also CC and target all the other layers possibly), and orientation also seems hard, so maybe those problems solve each other.
Sending a path integration signal seems like it contradicts sending a predictive signal to the thalamus for surprise. But a facilitating predictive signal could allow evidence accumulation. Each input which supports the prediction causes the predictive signal to be stronger, switching TC cells more and more towards tonic mode. T-type calcium channels and facilitation both grow and decay over somewhat long timescales, I think, so they can both do evidence accumulation. Evidence accumulation and path integration are both applicable to narrowing down the current possible locations, so they aren’t necessarily contradictory.