Hypothesis & Implications
The matrix cells in thalamus which target L5st might be involved in dealing with multiple orientations when determining locations on objects.
This hypothesis predicts that head direction cells in the thalamus project to L5st. It suggests that the projection from matrix cells to L5st is the sole connection which only serves orientation invariance.
Jeff Hawkins wrote that L2 could represent the generic object and L3 could represent it in a particular state. Should grid cell or displacement cell modules include minicolumns?
L5st cells project to L2 and maybe not L3.
Perhaps orientation is treated the same as location, except it anchors the representation to the object itself rather than the feature. That could be how the brain converts different object states and different orientations into the same object. Path integrating between each state wouldn’t work, though.
I’ll try to summarize it.
Matrix cells low in the hierarchy target different layers in primates and rodents. This difference could be because of 2d-like processing early in the primate visual cortex. That would mean the projection from matrix cells to L5st is for handling orientation.
I think lower L6 CT cells are for attention to sets of possibilities. In rodents, those cells have a massive reciprocal connection with L5st, and this reciprocal connection doesn’t exist in primate visual cortex. That is consistent with solving orientation by anchoring L5st cells to the orientation-variant object.
One flaw in my reasoning is that, in primates, L2/3 replaces L5st in terms of connections with L6 CT cells. I’m not sure this L2/3 circuit is completely missing, though.
One potential source of hints at how the cortex handles orientation is differences between regions. In primates, it seems like V1 doesn’t deal with orientation. Representations lower in the hierarchy are 2d, meaning they only care about the images on the retinas, whereas higher regions represent 3d things. (My memory is vague about that and there are other explanations besides dealing with orientation, though.)
Primate cortical hierarchies are much larger than those of rodents, so primates can afford to devote regions to orientation-variant processing to represent fine details, but rodent cortices are too small to do that.
Thalamic matrix cells target the superficial layers of primate V1, whereas they target L1 and L5st in rodent barrel cortex*. A possible explanation is that L5st-targeting matrix cells are involved in dealing with orientation. I don’t know why barrel cortex doesn’t receive input from L2/3-targeting matrix cells.
*There seem to be matrix and core cells in primary and higher order nuclei, and core+matrix-projecting cells in higher order nuclei. Matrix/core and primary/higher order are defined here by the cortical layers and regions they target, not by the inputs to thalamus. I’m confident that this organization is close to the truth.
I also think matrix cells are for attention to or representation of sets of possibilities. Upper L6 cells target core cells and L4, whereas lower L6 cells target matrix cells and L2/3 (primates) or L5st (rodents). These two types of cells could correspond to two types of attention- one to sensory inputs and the other to sets of possibilities. This would mean L5st cells, lower L6 CT cells, and L5st-targeting matrix cells are involved in handling orientation and switching sets of possibilities. Since L5st cells are in primate V1, the only connection which solely serves orientation invariance might be from matrix cells to L5st cells.
I don’t anymore think that L6 CT cells can do path integration using facilitation, because facilitation only occurs between the first two spikes and then the synapses depress. Population level studies found facilitation over long trains, but that is probably because not every cell activates in response to every shock/laser pulse.
All or most resources are in my notes. https://docs.google.com/document/d/1vgfQAY3-uo9cDLKumHLuiuuQ06wyZhwk8Wy--6zS9oE