If the animal is awake and attending you have the counter-flowing stream arriving at the perceptual area doing what HTM calls the location signal. This additional signal is a confounding factor is trying to sort out what is flowing UP the sensory stream.
The H&W work and what has followed is useful in sorting out what is going up the hierarchy.
I’ll be leading a journal club on Crazy Quilting next week with the Numenta research team. I will record it and show you all, so stay tuned for that. In the meantime, here are my notes. I’m not promising they will make any sense.
Very good start. I am planning to hunt down the reference(s?) that mentions cells OTHER than pyramidal in mapping RFs. I can’t remember which paper(s) it was in so I will have to get lucky in search terms.
This is one of the problems in having 13K papers in the library, with many of them scans. Naming and key words are everything it finding anything. Unfortunately - I did not learn this until it had grown to a large mess.
I have some housekeeping to do; a lot of it.
@rhyolight - looking at your notes - how are you feeling about the possibility that the scope of a quilt dovetails to the local spread of activation fields?
Also - How are you feeling about the “well shuffled” part where the local features in the association area include a mix of local and distant sensing interspersed with other modalities? I can see how some of this would happen with level skipping between the maps and patches in each hierarchy.
It seems to me that these will have to be in the same local area to form an SDR that includes this information to do sensory fusion. If it is not done all in one space it will have to be done scattered over some map-2-map connections. Considering the known topology of the inter-map connections it seems likely that much of this fusion is done in the association areas.
I think of cell types as including things like which layer the cell is in and what other layers or brain areas it sends signals to. Pyramidal cells with different connection patterns can be intermingled, so I think they’re referring to what you are describing but not necessarily separate physical layers.
The way complex cells are defined seems to vary a lot (like different math or whatnot). They seem to mean anything which isn’t a simple cell. It all seems too vision-specific to me, so I would be wary of specializations.
The same cell type can be simple or complex depending on the visual submodality. In layer 6 of macaque V1, there are three types of corticothalamic cells which roughly correspond to three types of LGN cells (parvocellular, magnocellular, and koniocellular). The ones which correspond to parvocellular-like LGN cells are simple, whereas the other two kinds are complex. 
The barrel cortex seems like the best region for determining whether all regions have quilting patterns, which I understand to mean mapping onto the cortical sheet with discontinuities and generally what is or looks like disorganization.
Barrel cortex has a very orderly columnar organization in layer 4. There is one cortical column for each whisker, called a barrel in L4, and this columnar organization is present in some but not all other layers or sublayers. Things like whisker bend direction could be mapped though, which could have quilting patterns.
If you need a quilting pattern for the map of sensory patches/whiskers, it might have to be in the space between those columns, called septa in layer 4. Septa is driven by a different subnucleus of VPM than the barrels .
Whiskers are discontinuous, so each barrel column has a discontinuous border with its neighbors, except there are cells between those columns with less obvious columnar organization or no columns.
Are coordinate transformations relevant to quilting patterns?
I think the septa in barrel cortex and the koniocellular system in primate visual cortex are involved in coordinate transformations or sensory anchoring, involving dual-projecting L6 CT cells. When a rodent whisks (rhythmically moves its whiskers at e.g. 20 hz), there is a map of the whisked space in L2/3 , which is proof that primary cortex cares about location in some way. I only have preliminary evidence that septa is involved in this but a bunch of things point towards it.
You mention distance in your notes. One possibility is that septa determines distance by comparing info from different whiskers. Cells in the septa have larger receptive fields than those in the barrels, so they could compare whiskers. I’m pretty sure it has something to do with behavior or location though since cells vertically aligned with the septa and a couple related parts of the thalamus are closely associated with motor cortex.
I can list some sources if any of this is relevant and give sources for what I didn’t cite.
I don’t know exactly what you mean by this so maybe this isn’t relevant. (I should go find where it was mentioned on this forum elsewhere.)
Part of POm (higher order thalamus for barrel cortex) receives a direct sensory input (and input from L5 of barrel cortex), so I guess level skipping can include the sensory input. It’s a different sensory input (trigeminal nucleus) than barrel cortex L4 gets, with larger receptive fields.
I’m not so sure of that anymore. There are two types of quilts, the first is the major sensory field projection onto the cortical substrate. It is globally discontinuous, but locally continuous. This is a very different type of quilt than the “fingerprint” / “zebra stripe” quilts. I’m not sure these patterns encourage the spread of activations any more than a direct projection of their positions within the sensory space would.
This only happens in the “major sensory projection” described above, specifically in somatic cortex, but also to a lesser extend in visual cortex, as there is folding of the visual field along V1/V2 border and further discontinuities further up the hierarchy. And here I would not call the input “well shuffled” really. Most of the sensory area patches are in generally the right place. It just looks like a drunk threw them together… or that the nerve bundles just sortof fell together that way as they joined up toward the spinal column.
Maybe the more important function of this minor quilting (fingerprints, stripes) is to create local groups so that computations can occur. It reminds me of how mini-column neighborhoods are necessary during spatial pooling when there is topology in the structure. Cells need to be grouped somehow for local computations to start happening, and I think this discontinuity (especially where borders between slabs are sharp) might arise to allow it. (I don’t know how to describe the more continuous minor quilting, like orientation and RA/SA yet.)
This refers to the type of inter-areal connections between maps.
Look at figure 2 in this paper: https://www.cell.com/trends/cognitive-sciences/pdf/S1364-6613(13)00122-8.pdf
The drawing labeled “Combinatorial learning and emergence of semantic categories” is a perfect example of this level skipping concept. This happens all over the brain. The older ideas of hierarchy envisioned an orderly progression from senses to the association areas. Modern mapping techniques of tracts have moved on to connection matrices that show that the level of connections between two defined areas.
It is almost possible to say that everything is hooked to everything to some degree. Some areas more - some less. Several posts earlier in this same thread I posted links to papers that explore this same topic in great detail.
@rhyolight I don’t think I have done a very good job of explaining what I mean when I say “shuffling of receptive fields.”
I hope this diagram will show how topology is preserved but mixed with expanded scales of receptive fields.
I don’t know much about this topic, but I’m skeptical of connectivity matrices. Leveling skipping is definitely a thing, still. I just get annoyed with connectivity averages because I focus on one cell type at a time and connectivity is really important.
The cell types and subcellular target (e.g. proximal or distal basal or apical apex) matter. An orderly hierarchy and dense connectivity (in terms of which region targets which regions) aren’t contradictory. You could probably tell distal/proximal synapses apart by comparing results from retrograde tracing and responses evoked by the presynaptic cells, so connectivity matrices are pretty useful in that way.
For an unrealistic example of why an orderly hierarchy and non-hierarchical connections are compatible:
Think of a hierarchy of spatial poolers for a few different senses, each with a temporal memory. Predictive connections onto distal dendrites between each temporal memory don’t need to obey the hierarchy and can go between the hierarchies for different senses, and it is still a strict hierarchy in terms of minicolumn states.
I agree that this idea can be a bit disturbing on first exposure. It certainly was for me.
I have been reading through the results of mapping tracts using newer imaging technology and it really does seem as if there are more connections than were originally found using injection tracing.
On the functional side of things this does make some sense. In a computer program you can access a variable from any part of the program that needs a sample of whatever information it is conveying. The brain does not work that way - if information must be available it will have to be conveyed by projections from one place to another. For sensor fusion I can see a very real need for taps of various parts of different sensory streams to be connected at different levels of processing.
For each of three cortical areas I investigated (somatic, visual, auditory):
there is one primary sensory field projection:
transforms sensory input field onto physical cortical substrate
direct input from sense through thalamus
exists at birth
structure / topology does not change with learning
globally discontinuous, locally continuous
there are (many?) minor input property projections for each sensory input:
allows different input sources to express at same point in sensory field
does not exist at birth, emerges with learning
these typically look like stripes between binary variables
may have distinct (occular dominance) or fuzzy (SA/RA) borders
could emerge because of direct input source differences or because of the first level of pattern recognition in cortical processing is responding to specific low-level input patterns (like SA / RA)
low level input patterns
observed only in anesthatized animals in V1 (as far as I could find)
orientations of a line or direction of movement in V1
For #1 above, visual and auditory cortex projections are most similar. Both have mirrored continuous sensory space projections. The somatic cortex also has mirror representations, but the way it treats sensory field location seems different. It is not a continuous field, but a series of semi-discontinuous patches of space, each with local continuity.
For #2 above, when input is from different physical sources (like two eyes or two ears), there is a physical merge of neuron fibers into a striped pattern (for V1 you can even see the stripes in the LGN). Eyes and ears seem to project directly to grouped sections of neurons with sharp borders. The SA/RA stripes in somatic cortex are similar because there seems to be a small amount of cortical processing occurring in order to respond only to one stimulus or the other (note that SA/RA stripes don’t have the sharp borders of the binocular / binaural stripes). This is learned over time, I don’t really know how they emerge, but I also don’t think it matters too much as long as each input can project to the right location in the sensory space. For skin, there is not the problem of separated input sources (both RA/SA can be deduced from the same patch of skin, the neurons firing are physically close to each other in the sensory space, not in separate eyes/ears). Why there are even SA/RA stripes, I don’t understand.
For #3 above, I think this is an echo of sensory input into an unconscious cortex. Once we can monitor a waking cortex, these transformations will likely make a lot more sense wrt what is in the visual field. It is hard to say anything about these patterns except that there is likely some very low-level pattern recognition going on even in the receptive fields of the first level of input to the cortex.
Some of the early work on #2 “local” patterns it that they are learned on exposure to the environment. Deprived upbringing results in impairment past the plastic period; this is the Gabor filters that are formed by learning.
On the local stripes - these form stereo perception in the eye and phase/location signals in the ear.
Considering that these are strong location signals I would think that this is very important to the current grid thinking that is being developed at the Numenta mothership.
In any experiments where the animal is anesthetized (most all of them above), the object location signal doesn’t exist because it is generated in the cortex. As the eyes / ears focus on objects in the sensory field, the job of the cortical columns will be the same no matter. There is no experimental data on this, as far as I can tell.
Mark you are really helpful suggesting papers, hopefully I’m not taking advantage of you, but I have a specific request.
I would love to find evidence of striping in auditory cortex along the lines of the H&W orientation/direction topology. There is clearly a binaural projection of the input just like binocular visual input, and it follows that there may be a stripes for different low-level characteristics of sound like tone, attack, volume, etc?