A possible homolog of presaccadic predictive remapping in the barrel cortex


I’ve written about [1] a lot. Sorry if I sound like a broken record. I keep coming back to it because it describes a map conversion which I think L5 slender tufted cells (L5 ST) help produce. In barrel cortex, L5 ST cells are linked to two things which at first seem like violations of feedforward/feedback hierarchical signalling in barrel cortex. If pre-movement predictive remapping is universal, it has implications for object recognition. I will explain why L5 ST cells likely perform the path integration described in the locations paper [2]. The results of two studies also suggest a mechanism for converting to allocentric location.

Papers and Interpretation

[3] and [4] found somewhat different results.

The first [3] compared responses to a pole during whisking before and after removing surround whiskers (whiskers besides the principal whisker, the one to which the neuron responds best). They found that intact surround whiskers tended to shift preferred pole positions towards the back of the arc which the principal whisker sweeps or further back*. This means that neurons responded to whiskers to which the bar was about to move (technically the whiskers are what move). This study also found a continuous map of the whisked space in L2/3.
*I assume the sensory input occurs only during the forward motion of the whiskers, for the purpose of the formation of the map.

Since they measured RFs in terms of the location of the pole, which is the kind of RFs they were looking for, the authors do not claim angle-based responses are generated by these mechanisms. Instead, they only claim that multi-whisker integration produces an organized map in L2/3. It is a map of the whisked space, not angle to the head, since it is relative to the middle of the arc whisked, so if the whisking midpoint is shifting forward, the map also shifts.

The other study [4] found and cited results which suggest an alternative explanation for the map of whisked space in L2/3. Neurons tend to prefer whisker deflection directions towards the surround whiskers to which they respond. They also tend to prefer surround whiskers which correspond to cortical columns which are nearby, so the map is continuous. This map is also much more organized in L2/3 than L4. Overall, this causes neurons to respond to whiskers in the same space to which the principal whisker deflects, which has been hypothesized to produce feature binding by hebbian learning.

Those results all seem consistent with the results in [3], but what is actually going on is different from what [3] concludes. During whisking in [3], when a whisker contacts a pole, it deflects backwards. That means the cells which fire will tend to be the ones responsive to whiskers towards the back.
Therefore, there isn’t exactly a map of the whisked space. Instead, when the object moves relative to the whiskers even when not whisking, neurons in L2/3 might shift their receptive fields to where the object is about to move. Locking on to the object like that could produce an allocentric representation, but this is at least an egocentric coordinate transformation.

This phenomenon is similar to presaccadic predictive remapping. Cells in barrel cortex likely respond to the inputs on the sensory patch to which the object is moving, or they respond to that sensory patch once the object reaches it. Likewise, cells in FEF respond to inputs on the sensory patch currently viewing what the cell’s RF will view after the upcoming saccade. When the sensor moves rather than the object, in both cases, cells shift their receptive fields towards where the movement will soon bring their normal RFs.

The information from those two studies is ambiguous, so exactly what happens in barrel cortex is ambiguous. It could be different from FEF if barrel cortex produces allocentric representations, assuming egocentric vs. allocentric is the right way of framing it. I might have made mistakes in interpretation because I don’t have a clear picture of what is going on. To simplify things, I’ll assume barrel cortex does pre-movement predictive remapping, since that is probably close enough to the truth for these conclusions.

Remapping is a lot like the path integration described in the locations paper [2]. It maintains the proper reference frame. Since cortical regions are basically the same in the what and where pathways, egocentric remapping probably shares most mechanisms with path integration.

I think L5 ST cells do path integration or participate in it because it explains a lot. L5 ST cells project broadly to L2/3, so they could produce the map found there. If L5 ST cells represent the possible locations and path integrate them each movement, which also represents possible objects, like in the locations paper, then it makes sense for them to project to L2/3. Since L5 ST and L2/3 would both represent possible objects, they are somewhat interchangeable. For example, higher order thalamus projects to L5 ST cells in rodent barrel cortex, whereas it targets L2/3 in primate visual cortex. A type of L6 CT cell targets L2/3 rather than L5 ST depending on the species.

L5 ST cells receive strong facilitating input from lower L6 CT cells. Those cells are thought to be grid cells if I’m not mistaken.

In barrel cortex, lower L6 CT cells seem to be outside barrel-like structures, so mostly in septal columns. Septal columns are much more linked with motor cortex than S2. L5 ST cells receive facilitating input from a higher order subnucleus with an early response component gated by motor cortex, and lower L6 CT cells project to that subnucleus. Those CT cells also target part of VPM which drives septal cells in L4, and that part of VPM receives M1 L6 even though it doesn’t project to M1. So far it seems similar in primate V1. Overall, this lower L6 CT/L5 ST system is probably tied to the opposite where/what pathway, which makes it appropriate for path integration.

L5 ST cells project to the striatum. That makes sense if they are involved in predictive remapping, since selecting behaviors requires prediction. Narrowing down possibilities is also useful for selecting behaviors.

[1] Surround Integration Organizes a Spatial Map during Active Sensation (Scott R. Pluta, Evan H. Lyall, Greg I. Telian, Elena Ryapolova-Webb, and Hillel Adesnik, 2017)

[2] Locations in the Neocortex: A Theory of Sensorimotor Object Recognition Using Cortical Grid Cells

[3] Surround Integration Organizes a Spatial Map during Active Sensation (Scott R. Pluta, Evan H. Lyall, Greg I. Telian, Elena Ryapolova-Webb, and Hillel Adesnik, 2017)

[4] A somatotopic map of vibrissa motion direction within a barrel column (Mark L. Andermann and Christopher I. Moore, 2006)