I think features are coded as locations. For example, colors are continuous values, so they need non-categorical coding. To me, it makes more sense to use the same circuits for abstract concepts this way.
I’m not sure, but it seems like L2/3 and L5tt are the only layers with lateral inhibition via SOM+ cells. L2/3 and L5tt are respectively thought to be for locations and displacements, so SOM+ cells could reflect processing positioning.
There are a couple ways SOM+ cells relate to positioning. Inputs to SOM+ cells facilitate reliably, not just between the first two EPSPs, so they could perform path integration. VIP+ cells are modulated by behavior and mainly inhibit SOM+ and PV+ cells, so SOM+ cells seem involved in a circuit related to self-movement.
SOM+ cells don’t respond right away during the sensory response because their inputs facilitate. In , PV+ cells receive feedforward inputs and control the sparsity of the initial sensory response, whereas SOM+ cells might control the sparsity during the rest of the sensory response, although they target apical dendrites. Since feedforward inhibition is simpler than facilitating lateral inhibition of apical tufts, maybe the initial sensory response represents the raw feature, e.g. minicolumnar activation, and then the rest of the sensory response converts to the location.
L2/3 seems to do some sort of map conversion from raw sensory input to feature [3, 4, 5]. The map in L4 is of the whiskers in barrel cortex, whereas it is closer to a map of the whisked space in L2/3. It might also receive a subcortically produced map, since there are shifting RFs in the thalamic nucleus VPM (I assume they didn’t just record the two thin higher order subnuclei). But maps are generally less organized in L2/3 than in L4, at least when measured in the coordinate system of the sensory surface.
If the same layer represents features and then locations, maybe features are directly converted to locations based in part on the feature itself.
Some of this is based on what I recall from 2 years ago, so I apologize if any facts are wrong. I don’t want to put much time into this because I just started studying L2/3 and I don’t want any functional ideas stuck in my head too early.
 Surround Integration Organizes a Spatial Map during Active Sensation
 A somatotopic map of vibrissa motion direction within a barrel column
 Dynamic and distributed properties of many-neuron ensembles in the ventral posterior medial thalamus of awake rats