I am still digesting this paper but three things jump out that I have to relate now.
Earlier work with stronger signals produced a more diffuse and geographically larger response. The important thing here is that the amplitude of the response is signaled by the size of the RF. This is important considering that the nerve response is essentially binary. It is not clear if this larger response uses the hex boundaries.
The response is in both L2 and L4, and different types of cells in each layer. I will have to think on this. It seems a very important factoid but I am not sure what conclusion to draw from this.
I have seen this before in V1 but I missed the potential implication at that time: Discontinuity in the RF maps. One of the puzzles in hierarchical processing is that we expect the data to be spread into larger areas as it ascends the chain and attributes this to fan-out of the connections between maps. The theoretical pictures of receptive field go to great lengths to show these neat topographical maps. Then you see the lab work and it is a crazy quilt of strips and patches - not really at all as the theory suggests. Topology is retained for small areas but not so much for larger expanses. WTF?
Maybe we are looking at this wrong - if the mapping between inputs in all the senses follows a pattern of crazy quilting at a larger scale this would go a long way towards supporting the spreading that we expect in a hierarchy. If the association areas work the way I think they do they need a mix of well-shuffled but topographically meaningful data. I thought that this shuffling happened on the path from the primary sensory cortical areas but it may be starting even earlier. I suspect that the degree of “crazy quilting” starts at about spatial extent where the degree of diffusion by the spread of connecting fibers leaves off; form follows function.
This feeds into your “brain in a black box” thing. We know that the brain does not do variables as we are used to doing in programming - all connections are fixed and the data paths have to deliver the data to the right place for processing. In this case, we expect that the association areas will connect sight and touch to from our personal space. The place where they come together is in the parietal association areas. We expect that our SDRs will be formed from this soup of meaningful dots of sensation. For this to happen we will have to have both close and far away dots of meaning available to form these representations, from all the senses that may be meaningfully combined. This hub (association area) is where the hex-grid code is formed to signal these higher-level representations to other hubs.
To anyone new to this community reading this thread… this term “crazy quilting” is just something we’ve been using here on HTM forum to refer to the strange ways evolution has found to project things onto cortical substrate. You’ll find no reference to “crazy quilts” or quilts of any kind in the mentioned neuroscience papers or in neuroscience jargon.
Re: #3 above, yeah exactly. This is one reason the “1000 Brains Theory” is appealing. It explains why continuous topology is unnecessary (even if local topologies are still obviously important). Hierarchies still emerge, but higher regions sample the same input as lower regions, but with different sampling properties. Models at every level are proposing an object or a collection of features as output to higher levels.
I think the “crazy quilting” effect on sensory before it gets to cortex is a way to normalize the data and present it to large swaths of cortical tissue so larger patterns can be discerned. It could be seen similar to the way one neuron’s RF is somewhat a random portion of its potential input.
Now we have to talk about what vs where, don’t we? We don’t have an answer to this problem, but my instinct is that there are two primary reference frames represented in cortical columns, one for allocentric representation of objects (the universe of space in your brain), and the other for egocentric representation of self within the current environment.
I think – and this totally speculative and uninformed – that the what and where pathways might also have this “crazy quilting” pattern across the entire cortex. I have not talked to the research team about this, mind you. But what reminded me of it was that there are borders between “slabs” of the crazy quilt… and sometimes the borders are not sharp, and there is some continuous gradient between the two parts. And I wonder if this could interface between what and where slabs…
Wouldn’t this just be one level up the hierarchy from the individual senses?
It does not make sense to me that the hub is localized to one place in the cortical sheet. From what I understood the hex-grid code was formed all over cortex?
I don’t think it is so much a matter of local function such as hex-grid coding but connectivity.
The connectivity of the hub areas are other hub areas where other areas are more involved with map-2-map connections.
If you look deeper into areas like the visual system you will see that there is a progression of several maps from V1 to the related association region. I have come to believe that what is happening is a continue extraction of more refined properties of the senses that end up being presented to the association regions. This provides more features for the SDR/hex-grids more to work with in forming patterns. Critters with more processing maps have more refined discrimination abilities by this process.
In the Mountcastle paper, it also shows the “crazy quilting” between slowly responding (SA) and rapidly responding (RA) slabs of neurons. It seems the boundaries between the quilting slabs is continuous and fuzzy, but the cortical column boundaries much sharper.
I would like to see the most recent Numenta paper fitted to this data.
These research finding have been in the back of my mind as I read the location and thousand brains concepts - I have real problems making the connections. One of the most troubling has been the “laminar differences” paper. There are other with much the same findings.
I see that Numenta has a certain view (that has been evolving) of how the layers work together for locations - I have seen many papers that just don’t fit any of the proposed models and some are directly contradictory. The interactions with the sub-cortical structures are very important and I just don’t see how these well researched finding fit in at all.
Perhaps all will fall into place as the theory is refined but I do have many unanswered questions.
Peter Schiller discusses these cell types in this video, around the 26 minute mark:
This video is from an MIT OpenCourseWare course called “Sensory Systems.”
He goes into a lot of detail which you may find helpful as a refresher/primer/review introduction to some of the material related to your discussion.
Perhaps this explains why I was offering this as interesting application of horizontal mutual excitation and inhibitory interconnections. I find it intriguing that with minor biologically plausible tuning you could use the same mechanism in different places to achieve differing functions. (simple cell/Gabor filter and hex-grid formation)