Specializations Help Understand the Universal Neocortex
L2/3 seems to vary more between species than layers 5 and 6, and it seems especially specialized in primates, so it is difficult tell what is universal. But specializations can clarify the universal cortical system. They serve relatively clear functions and are in context of the universal cortical system.
Rounded Structures Cause Problems
Compartments are often rounded. This can cause conflating adjacent compartments when they are treated like columns or similar. Sublayers or layers arguably can extend a bit higher or lower into spaces between rounded compartments.
There might be rounded structures in several layers or sublayers of V1 . I’ll still use terms like blob columns, but they probably have different widths at each depth.
V1 Sublayer Nomenclature
In most primates, what Brodmann called L4 is divided into L4A, L4B, L4Ca, and L4Cb. L4A and L4B are arguably part of L2/3 proper because they project to contralateral V1 . I’ll treat them as part of L2/3 proper, but L4 of mouse A1 projects contralaterally . Hassler’s nomenclature fits this better but I’m used to Brodmann’s nomenclature . L4A is quite thin and apparently missing in some species so I’ll mostly ignore it .
Projections from L2/3/4B to V2
In L2/3, cytochrome oxidase labels blobs, which are rounded structures with space between them. They are mainly in L3 but reach into L2 . Blobs are located at the borders between pinwheels, which are circular maps of orientation . They are also aligned with ocular dominance columns .
Often, L2/3 is only divided into blobs and interblobs (or more accurately, the single continuous interblob space). However, there is also a border section between blobs and interblobs.
These compartments are related to the dorsal and ventral streams. In V2, cytochrome oxidase labels three types of stripes in V2, which are either in the dorsal or ventral stream. Roughly speaking, in L2/3 proper, blob, interblob, and border columns respectively target thin, pale, and thick stripes [13, 16, 18]. As a result, blob and interblob columns are in the ventral stream, while border columns are in the dorsal stream. However, all three types of columns project to all three types of stripes. Cells projecting to thick stripes also target other stripes relatively often, at least cells in L2/3.
There are some differences between projections to the ventral and dorsal streams . L2/3 contains by far the largest populations projecting to the ventral stream stripes and superficial layers of the dorsal stream stripes. L4B has a large population projecting to deep layers of thick stripes. Whereas L4B is associated in that way with the dorsal stream, the deep layers of V1 are associated with the ventral stream. In any type of column, the deep layers of V1 (pooled) only target ventral stream stripes, and only in their L6. Lastly, of the layers discussed, only L2/3 targets the set of pale stripes which are medial to thick stripes.
It seems that L4B mainly targets the ventral stream with pyramidal cells and the dorsal stream with spiny stellate cells. In this sublayer, pyramidal cells are the main source of projections to V2, whereas stellate cells are the main source of projections to MT, which is in the dorsal stream . Twice as many stellate cells than pyramidal cells target thick stripes , so pyramidal cells mainly target V2 in ventral stream stripes.
Differences Between Blobs, Interblobs, and Borders
Lateral projections from L2/3 blobs are preferentially to other blobs and interblobs to interblobs . Patchy connections like these often indicate compartments or perhaps mapping something like orientation. Border cells do not have as long lateral projections and mainly form boutons in interblobs and also near their somata.
Blobs and interblobs have different response properties. Usually, cells in blobs are monocular and cells in interblobs are binocular , perhaps because blobs are aligned with ocular dominance columns. Additionally, cells in blobs prefer lower spatial frequencies (larger receptive fields), have lower orientation selectivity, and are more color selective than cells in interblobs . Cells in blobs are also more selective for contrast .
Most of the interlaminar connectivity studies I have read do not acknowledge border columns, which makes their results regarding blobs/interblobs difficult to interpret. Keep border columns in mind wherever I only mention blobs and interblobs.
Pathways to L2/3/4B
Thalamocortical Pathways to L4C 
Koniocellular cells in LGN project to L1 and blobs
P LGN cells also target L4A in species which have it, but not in chimps and humans, whose L4A and L4B are modified [17, 20]. Some K LGN cells are in P LGN layers, which can cause confusion.
Some M LGN cells with different response properties only target upper L4Ca.
Lower L4Ca receives M and P LGN, and unlike the rest of L4C, it has good direction and orientation selectivity . This sublayer is thin.
Inputs to L4B
L4Ca targets L4B a lot more than it targets L2/3 .
L4B mainly projects to the ventral stream with pyramidal cells and the dorsal stream with spiny stellate cells, as discussed earlier. Both cell types receive input from L4Ca . Pyramidal cells besides the ones which target MT (a dorsal stream region) also receive input from L4Cb, but probably on their apical dendrites . It seems like both streams receive L4Ca in L4B, while only ventral stream L4B is modulated by L4Cb.
Inputs to L2/3
I need to research differences between sublayers of L2/3 more.
L4Cb targets L2/3 a lot more than it targets L4B .
L4Cb projects to L3 without any blob/interblob preference . This projection spreads out compared to some projections between layers with the same maps, possibly to switch from mapping the retina to mapping direction selectivity . Upper L4Ca targets blobs, whereas lower L4Ca targets interblobs .
L4Cb does not target L3B cells which project out of V1 . Corticocortical L3B cells receive more of their input from L3B, L4A, L4B, and L4Ca than local cells do, and their apical dendrites are tufted unlike most local cells .
L1/2/3a pooled receive no input from L4C or L4A, which are LGN recipient layers . Instead, they receive input from L3B and L4B, as well as deep layers. The projection from L3B spreads broadly, whereas the projection from L4B is columnar . The projection from L4B to L3B is also columnar, at least for L4B cells which project to thick stripes . Comparing numbers and amplitudes of EPSPs, it is possible that L4B cells in blob columns do not project in a columnar fashion .
Projections from L4B within V1
One study reconstructed cells in L4B which project to thick stripes . Some target L4B, the whole depth of L2/3, and L5. Some target L4B and L5. Some only target L5. A couple mainly target L6, one of which also targets L5.
Most of those projections seem macrocolumnar in lateral extent. Projections within L4B are transcolumnar. Some cells which just target L4B and L5 have transcolumnar axon arbors in both.
Ventral Stream Does Extra Things
It seems like the ventral stream performs much of the same processing as the dorsal stream, except it adds additional aspects of processing. Mostly, this is because it seems like there are more modulatory connections in the ventral stream, meaning connections which do not drive firing (such as distal dendrites or metabotropic). It also seems like the ventral stream processes dorsal stream-related information more so than vice-versa.
Deep layers of V1 only project to V2 in ventral stream stripes, and only in L6. L6 is generally modulatory.
L4B mainly targets the ventral stream with pyramidal cells and the dorsal stream with spiny stellate cells. Inputs to apical dendrites are mostly modulatory. This seems like putting magnocellular information in context of parvocellular information.
Ventral stream columns in L2/3 have longer lateral projections, which allows more interlaminar interactions, which are presumably modulatory. Whereas dorsal stream columns mainly target interblob columns of the ventral stream, which suggests the ventral stream receives more information associated with the other stream than vice-versa.
Cells in L2/3 which project to thick stripes also target other stripes relatively often, which suggests the ventral stream receives more information associated with the other stream than vice-versa.
L3B corticocortical cells only receive direct input from magnocellular L4C, which is more involved in the dorsal stream, so there is probably an extra processing step for the ventral stream.
Perhaps extra aspects of processing could help understand the ventral stream.
Speculation that Blobs are Higher Order
If higher order regions serve to recognize larger things like in HTM theory, then different levels of the cortical hierarchy will process different sensory inputs in some cases. That is because different aspects of the sensory input are on different scales. Blobs process different aspects of the sensory input than interblobs, and they have larger receptive fields, so they are arguably higher order.
Blobs receive input from K LGN cells, which also target L1. Higher order nuclei target L1 of lower order regions and usually also other layers besides L4. That is similar to the projections of K LGN cells. I think matrix cells like K LGN cells are just relays for higher order raw sensory information.
I’m not sure how hierarchy can serve both scale and abstraction (e.g. object composition), so maybe scale is not exactly the same as hierarchical level. That could mean inputs to L1 are dictated by scale rather than hierarchical level. However, the septal domain of barrel cortex is higher order in a literal sense as far as I can tell (driven by a higher order part of the thalamus), so it is not unreasonable for a single region to have compartments at two levels of the cortical hierarchy.
Axons of L2/3 cells ramify in L2/3 and L5a/b . L2/3 cells directly evoke EPSPs in excitatory cells of all layers except L4, in the same column and contralateral barrel cortex . However, cells deeper in L2/3 have more boutons in L4, perhaps targeting interneurons . Callosal cells are mainly in L3 and L5 in barrel cortex, although they exist in all layers . In A1, callosal cells are mainly in L3, L4, and upper L5, densest in L4 .
Besides the same region, they also target higher order regions at least in L2/3 and L5 [2, 27]. Corticocortical cells in L2/3 sometimes also target other telencephalic structures, such as the striatum and amygdala [2, 27]. However, L5 intratelencephalic cells target the striatum more .
There are no clear sublayers in rodent L2/3, and properties probably vary in a continuous manner with depth . Sizes of proximal dendritic arbors might not vary continuously with depth, though . These are all glutamate uncaging studies.
More superficial cells in A1 receive a larger portion of their input from L2/3 and less from L4, on average . In barrel cortex, this pattern is not apparent in a single barrel column, but deeper L2/3 receives more input from L4 of neighboring columns . Broad input from L4 is like the change is what is being mapped between L4 and L2/3 in primates. In primates, upper L2/3 maps orientation like L4B (which is the lowest part of L2/3 proper) , so the map in upper L2/3 of rodents might derive from lower L2/3 rather than L4.
In A1, L2/3 is a much larger input than L4 , whereas L4 is a much stronger input than L2/3 in barrel cortex . Lower L2/3 receives as much thalamocortical input as L4 does in A1, which could explain the difference , although lower L2/3 also receives at least some thalamocortical input in barrel cortex . Another major difference is that deeper cells receive more input from upper L5 in A1 , whereas more superficial cells receive much more input from upper L5 in barrel cortex . Like L4, upper L5 is a much stronger source of input in barrel cortex than in A1. In both regions, L2/3 is the main source of inhibitory input.
In barrel cortex, cells deeper in L2/3 target the same column’s L4 and L5b more . Cells more superficial in L2/3 target L2/3 of neighboring columns more .
Cells adjacent to L1 are different from other cells in L2 of mouse temporal cortex .
Maps in L2/3
Like in primates, the map of the sensor is not as smooth in L2/3 as L4, or L3b/4 of A1 to be more exact, which is the main target of the thalamus in A1 . This likely reflects a change in what is being mapped, like the change from a continuous map of the retina to a map of orientation in primates. For example, A1 is tonotopic in L2/3 only at scales greater than 300 microns, and there is preferentially patchy connectivity at a scale of 300 microns . That seems similar to patchy connections in primates, which reflect maps and/or compartments.
In mouse barrel cortex, L4 maps the whiskers more smoothly than L2/3, whereas L2/3 maps something closer to physical space more smoothly than L4. This will take some explaining and maybe some technical knowledge, but I think it’s worth it because the mechanism for determining location (or similar) is basically the same one Jeff Hawkins mentioned in visual cortex.
One study found a map of the space swept by the whiskers during whisking, produced by combining inputs from multiple whiskers . The researchers placed a vertical pole at various positions in the whisked space, along the rostrocaudal axis. Before and after trimming all whiskers except the principal whisker*, they measured responses at each position using calcium imaging in L2/3 and L4, and spikes in L5. Comparing the responses (not temporally precise) indicated how surround whiskers* shift the preferred location of the bar.
*Principal whisker means the whisker which produces the strongest response for the recorded cell. It often corresponds to the barrel column in which the cell resides.
*Surround whiskers here means whiskers other than the principal whisker.
They only looked for the map of the whisked space in L2/3, but it probably exists in all layers. In L4, cells often required contact with the principal whisker to respond, but surround whiskers still impact L4. They reduce the response to more rostral pole positions and increase the response to more caudal positions. The impact of surround whiskers is similar in L2/3, except they do not reduce the response to more rostral positions much, and they mainly increase the response to more caudal positions. It is difficult to compare L5 because they recorded spikes rather than using calcium imaging, and the response was increased by surround whiskers regardless of pole position. However, like in L2/3 and L4, surround whiskers increased the response to more caudal positions more.
They hypothesized how this surround integration produces the map of whisked space. The first whisker to make contact evokes the largest response because rostral whiskers have not made contact, and those whiskers would reduce the response. The response of the cells which prefer the first whisker is further enhanced by contact with more caudal whiskers. More and more caudal whiskers evoke smaller and smaller responses because there are more and more rostral whiskers to inhibit their evoked responses and fewer and fewer caudal whiskers to enhance them. This produces a smooth map of the position in the whisked space. It might be smooth because multiple whiskers make contact at a time, or perhaps phase in the whisking cycle plays a role. Phase modulates L2/3 cells which project to M1, which is discussed in a later section.
Another study suggests a similar but more generally applicable mechanism . In each barrel column, there is a map of whisker deflection direction preference, more organized in L2/3 than in L4. Cells respond to deflection of the corresponding whisker in a particular direction, but they also respond to deflection of certain other whiskers, preferentially. Specifically, the whiskers towards which the corresponding whisker deflects in the preferred direction also evoke responses. As a result, cells might respond to particular locations in the space above the whiskers.
Let’s think about the other study in context of . During whisking, when a whisker hits the pole, it deflects backwards*. Deflecting backwards is towards more caudal whiskers, so the cells which respond are the ones which also respond to more caudal whiskers. More rostral whiskers might inhibit them because they are in the exact opposite direction of the preferred one. So the results in  can be explained by preferring deflection in a particular direction and deflection of whiskers in that direction. However, this mechanism is more generalized.
*I’m ignoring the fact that whiskers would deflect forwards when whiskers move backwards during the whisking cycle. Maybe that part of the whisking cycle is less important for sensory processing.
Let’s say the pole is at an angle, so whiskers deflect backwards and up during whisking. L2/3 would map the position of the pole along the axis orthogonal to the pole. I recall Jeff Hawkins saying he thinks part of location processing in V1 is preferring bars at certain orientations and movement of those bars in directions orthogonal to the preferred orientation. It seems like a very similar thing is true of barrel cortex.
Sublaminar Sensory Responses Reflect What and Where Streams
Sensory Responses in General
A small subpopulation produces the vast majority of spikes during the sensory response . Subthreshold responses are not rare, however . Instead of a lack of excitatory input, the sparsity is caused by inhibitory input and hyperpolarized resting potentials [24, 4]. Similarly, L4 targets L2/3 pyramidal cells a lot, but it also targets L2/3 interneurons .
Sensory Responses in A1
When awake, upper L2/3 of mouse A1 has much lower tone-evoked and spontaneous firing rates than lower L2/3 . In an anesthesia study, IPSPs are more broadly tuned than EPSPs in upper L2/3, whereas they have similar bandwidths in lower L2/3 as well as in L4 .
A couple other studies which used two other anesthetics found stronger sensory responses in upper L2/3, which is the opposite of the awake state [22, 28].
Overall, short latency sensory responses in upper L2/3 are more strongly or more broadly inhibited than in lower L2/3.
What and Where Streams in Barrel Cortex Seem Sublaminar
In barrel cortex, S2-projecting (S2p) and M1-projecting cells are in the what and where streams . More S2p cells have sensory responses during texture discrimination, whereas more M1p cells respond during location discrimination .
It seems like upper L2/3 cells are S2p and lower L2/3 cells are M1p (but see [25 fig.1]). Correlations with depth are quite imprecise in L2/3 of rodents, so this is difficult to rule out. M1p cells  and lower L2/3 cells  have shorter latency, larger, faster EPSPs during sensory responses than S2p cells and upper L2/3 cells. Similarly, during repetitive active touch, PSPs can summate more in S2-projecting cells and their responses depress less over time . That seems more suited for producing the map of the whisked space discussed earlier, or for averaging across more whisking parameters.
Sensory responses in the where stream seem more straightforward than in the what stream. Like in A1, lower L2/3 receives direct input from primary thalamus . It probably maps the sensor more closely than upper L2/3 like it does in A1. In contrast, upper L2/3 probably receives more complex inhibitory inputs, like it seems to in A1, because there is a band of inhibitory and/or disinhibitory interneurons there [11, 24]. Sensory responses in lower L2/3 mimic the thalamus more, whereas sensory responses in upper L2/3 are further removed and under more complex inhibitory control.
M1p cells do not simply map the sensor, though. They are modulated by whisking phase , which could help produce the smooth map of the whisked space in L2/3. Lower L2/3 receives more input from L4 of neighboring barrel columns than upper L2/3 does, so it is suited for changing what is mapped. M1p cells could be the ones which produce the map of space above the whisker pad by preferring the barrel column’s corresponding whisker deflecting towards the preferred adjacent whisker. They might produce a map of the space sensed by the whiskers whether whisking or not whisking.
M1p cells respond to more whiskers than S2p cells do , consistent with a change in what is being mapped, and upper L2/3 receives less input from L4 of neighboring columns, so S2p cells might simply map the whiskers. However, S2p cells are modulated by whisking amplitude or possibly something more temporally precise  other than whisking phase . Perhaps S2p cells integrate location-related information more slowly and individually for each whisker.
L2/3 cells which mainly target S2 do not target M1/2, whereas those which mainly target M1/2 sometimes have some axon in S2 . That seems similar to primates, where the ventral stream seems to process some dorsal stream-related information but not vice-versa.
The interneurons in L2/3 express 5HT3AR (ionotropic serotonin receptor), somatostatin (SOM), or parvalbumin (PV). PV cells have much faster sensory responses than SOM cells, without much orientation selectivity. SOM cells are inhibited by sensory stimuli. Also, whereas membrane potential fluctuations are correlated in pyramidal, PV, and 5HT3AR cells, they are smaller and negatively correlated in SOM cells. Similarly, SOM cells are inhibited by stimuli on average, and they have similar orientation selectivity as pyramidal cells. They summate a very large visual receptive field, and inputs from pyramidal cells facilitate enough for a single pyramidal cell to activate them. It seems strange that they have sharp orientation selectivity because they receive input from 30% of pyramidal neurons and do not receive input from L4, so their orientation selectivity probably results from inhibitory inputs. They target distal dendrites, at least apical tufts.
5HT3AR cells have two types. Some are neurogliaform, which use volume transmission, and some are VIP bipolar cells, which likely inhibit other interneurons. They adapt and express nicotinergic receptors.
Relative to quiet wakefulness, whisking does not change PV membrane potentials, it depolarizes pyramidal cells by a few mV, it strongly depolarizes 5HT3AR cells, and it strongly inhibits SOM cells. Firing rates in the thalamus are higher during whisking, and these effects can be mimicked or blocked by activating or inhibiting the thalamus. Despite the changes in membrane potentials, sensory responses in somatosensory and auditory cortex are reduced during behavior. Increased thalamic firing rates might cause that reduction via synaptic depression. However, sensory responses in visual cortex increase during behavior. Activating higher order cortex might also mimic at least some of these effects .