Picking up on some of the BIG IDEAS in brain research, which was wonderfully chaotic when we last discussed in December under the hashtag #BrainIdeasCountdown, e.g. neuromatch.social/@NicoleCRust

Here's an attempt to fill in some blanks, and let's flip the hashtag: #BigBrainIdeas. I'll focus on the notion that there are facts, ideas and then there are "Big Ideas" and I'll focus on the last one. Please join in!

I'd argue that one of the most influential Big Ideas about the brain in the latter half of the 20th century is the is the notion that:

The neocortex of the brain is made up of a generic functional element that is repeated again and again and from this repetition, all of cortical function emerges

I'm talking about the cortical column, first described by Vernon Mountcastle in 1957. The unit contains ~10K neurons and humans have ~25 million of them. The rapid evolution of humans is proposed to have followed from a rapid expansion of cortex that happened because it has this repetitive crystalline structure. The gist behind the "functional" bit is that each unit always does the same generic computation, and the different functions of different brain areas result from the different inputs that these units receive. @TrackingActions very nicely summarizes the ideas here: nature.com/articles/s41583-022

So what does this generic functional unit do? Proposals vary. One idea, also reflected in deep convolutional neural networks, is that it does two(ish) things: selectivity and invariance, stacked repetitively to support things like recognizing objects. Other proposals suggest that the brain is a prediction machine and each unit contributes a little bit to those predictions in a manner that relies not just on feedforward connectivity, but also feedback. Some proposals suggest that the function of the unit varies along a gradient as a consequence of biophysical properties like receptor expression: nature.com/articles/s41583-020.

Among brain researchers, this Big Idea is polarizing - obvious to some and misguided to others. Where are you in terms of your 'buy in' with this big idea?

@NicoleCRust @TrackingActions @cogneurophys I am not sure that this idea has actually been very influential on the grand scale of neurocog theories. The modern theoretical approach seems to revolve around understanding how brain areas connect in networks to solve problems, and I can't see how generic computation would inspire/drive this perspective.

@bwyble @TrackingActions @cogneurophys
The idea spawned maps like this, which have been highly influential for neurocog theories, no?

@NicoleCRust @TrackingActions @cogneurophys I thought that diagram was the result of neuroanatomy studies. Why would a generalized function theory lead to a highly specific wiring diagram like this?

@bwyble
Yes, but ... Felleman & Van Essen defined the hierarchical levels of this diagram according to the cannonical microcircuit rule: L4 receives input; L2/3=feedforward output; L4/5=feedback output.

@DrYohanJohn @NicoleCRust I agree with Yohan, I don't think the microcircuit is crucial there, rather it's observing that there are laminar patterns that are more or less ubiquitous.

@bwyble @DrYohanJohn
Interesting! To me, those ideas could not be more connected.

@NicoleCRust @bwyble @DrYohanJohn

shameless plug, read my CONB on this topic with Adam K, for my more edited, fleshed out opinion!
sciencedirect.com/science/arti

(just insert "microcircuit motif" whereever you read cell-type, Adam K. and I disagree about which phrase is the better one :p)

I get the argument about confusing implementation and computation, but I agree with with @NicoleCRust that the idea of canonical cortical computations has been super influential, especially / at least in vision (which is all of computational neuro anyway, amiright?)

I think the idea of simple, repeated computation is kind of necessary / permissive for a certain types of "grand unified theory" that are very intuitively appealing exactly because they squash together computation and implementation into one little thing that is intuitively understandable in words. The fact that these "theories" are conceptually "small" down to implementation mean that people understand them, they catch on, they drive research.

Now whether this is positive or not, is an open debate. But I think there's no way we can say that ideas like predictive processing, backprop, divisive normalization, and maybe even convolution havent been wildly influential in the field.

To state this in a slightly more aggressive way than I feel: I do think the idea of a canonical microcircuit is very useful, because it's studiable! If we just start off by assuming everything is brain soup it's very easy to just give up and assume we'll never understand implementation. I rather start off with the assumption that there exist architectural motifs that matter, and try to take that as a hypothesis from which to start, than just admit defeat.

So maybe the answer is, it depends on the level of explanation you are lookign for. If you dont care baout multilevel understanding then "how brain regions connect in networks" black box type of understanding is enough. I personally think it's only step 1, and then understanding more details of implementation is step 2. Furthermore, I've come to think that the implementation detaiuls will probably constrain and help us understand the higher level.

@achristensen56 @NicoleCRust @bwyble @DrYohanJohn

The fact that, when, say, sight is lost early, other sensory modalities can successfully invade areas they wouldn't necessarily be processed, in suggests standardisation of function/microanatomy across large parts of the brain.

@strangetruther @achristensen56 @NicoleCRust @bwyble @DrYohanJohn

A simple argument for some sort of canonical cortical computation is: "cerebral cortex ... processes ... diverse tasks with what appears to be a remarkably uniform, primarily six-layer architecture, albeit with significant differences in details across species and cortical areas [1,2,3􏰩,4–10,11􏰩,12–14]. ... This has long suggested the idea that a piece of six-layer cortex with a surface area on the order of a square millimeter constitutes a fundamental cortical ‘processing unit’ [e.g. 16,17].The cortex varies in surface area by a factor of 10000 across a set of 37 mammalian species, while thickness (the distance across the layers) varies only by a factor of 10 over the same species [18], suggesting that the most salient evolutionary change in cortex has been enormous multiplication of the number of ‘units’ [e.g. 14]."
The last two references are:
14. Rakic P: Confusing cortical columns. Proc Natl Acad Sci U S A 2008, 105:12099-12100.
18. Hofman MA: On the evolution and geometry of the brain in mammals. Prog Neurobiol 1989, 32:137-158.

This is from the 1st paragraph of a Current Opinion review paper that I wrote that Nicole cited further up in the thread (pubmed.ncbi.nlm.nih.gov/268680)

More generally I think there are (at least) 4 mammalian (and, except for cortex, vertebrate) brain structures that each clearly have repeating architecture, and that -- at least as studied in primates -- communicate pretty intimately with one another: cortex, thalamus, basal ganglia, and cerebellum. They communicate with specificity, eg a given piece of cortex communicates with given thalamic nuclei and given regions of basal ganglia and cerebellum, which communicate with one another, e.g. Boston & Strick nature.com/articles/s41583-018. These specific ctx/BG/cerebellum interactions cover at least posterior parietal through frontal cortex, and perhaps higher sensory cortices as well, i.e. they cover all sorts of cognitive processing, not just motor processing which is the traditional function assigned to BG and cerebellum. So it's not just the enormous multiplication of cortical "units" (with diversification, i.e. the spectrum Y.J. referred to) , but also the corresponding multiplication of their partner thalamic, basal ganglia, and cerebellar "units" that suggest some fundamental computional operation, albeit again with diversification.

You don't see this sort of thing in the brainstem. Different bespoke nuclei or other sorts of neural units each do different pieces of different computations. In contrast, the existence of these structures with repeated modular subunits with roughly repeating architectures (despite much variability and diversification), and with specific patterns of interconnections between their modules, as well as their enormous growth in numbers of modules at least in mammalian evolution, all just scream out that some sort of computational motif is being repeated (with variations on the repeated units, much as multiple copies of a gene provide a substrate for evolution into multiple variants -- and occasionally quite new structures). That wouldn't happen by accident.

Kenji Doya long ago postulated that cortex is for associative learning, BG for reinforcement learning, and cerebellum for error-correcting learning. That still sounds like a decent 0th-order take. And, I'll add my speculation, one function of thalamus -- not all that it is doing -- is to take any modality of information whatsoever and convert it into a language that cortex understands, using a roughly uniform architecture with roughly uniform biophysics across all these different modalities of information.

@kendmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

Thanks for these thoughts Ken, very nicely put.

I agree with everything you say here but, while replication of cortical tissue occurs and this suggests some kind of repeated computational motif, this doesn't necessarily suggest that cortical columns are the fundamental unit of replication. Replication could iterate over much larger or smaller units, down perhaps to just individual neurons, could it not?

#cortex #neuroscience

@bwyble @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

Well, the question is, what is the level of the unit that you see repeated? The fact that individual neurons are repeated would not account for any of the specificity of the repeated architectures one sees across each of the four structures. Based on about a 1sq mm chunk of V1 containing all preferred orientations from both eyes at a given retinotopic position, with the next chunk over (e.g., about 1mm away) having half-overlapping receptive fields, Hubel and Wiesel suggested that size of chunk as a cortical processing unit (the size varies across species from 1/2 mm to 1.5 mm or so, but order of magnitude 1 mm). That is at least roughly consistent with dense connectivity in cortex extending horizontally maybe 200-300 microns from any given point, along with sparser connectivity over longer distances. Comparable structures have been seen in many other cortical areas. So that is what I think of as the unit that repeats. Others have focused on a 25 x 25 micron or so "minicolumn", or the area spanned by the neurons spawned from a given radial glial cell. That also repeats, but personally seems to me too small to be a computational unit, i.e. to do a self-contained computation.

But more generally, I'd want to characterize the repeated units of each of the four structures that are connected, that talk to one another -- a thalamo-cortical-basal-ganglia-cerebellar unit. I suspect the cortical part is something like Hubel & Wiesel's square mm, but maybe it would be more like a cortical area, I'm really not sure.

@kendmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

Yes, I think that's the hard/interesting question about the scale of replication. Perhaps one can make a fairly strong inference about V1 from observations of orientation sensitivity and also ocular dominance, but that might be specific to V1. In my view, the brain likes to specialize as much as it likes to repeat, and the two demands play off against each other.

What I'd like to see to be more confident of repeating columns is evidence somewhere else, like tonotopy, or striping in the motor system. That might exist but I've never heard of it.

@bwyble @kendmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

Mixed selectivity multifunctional neurons seem to be common in cortex. That argues for a lot of repetition and not a lot of strict specialization.

@ekmiller @kendmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

They do, but there's also a lot of ways that neural tissue specializes as well, especially when one looks at areas like the hippocampus.

I've always been a splitter rather than a lumper though, and this has affected my perspective on everything neuro.

@bwyble @kendmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

I'm more of a splitter when it comes to the brain's infrastructure. But anatomy is like the road-and-highway system. It says where traffic *could* go, not where it flows from moment to moment.

Cortical function, IMHO, will be about emergent properties and highly interactive dynamics that have less respective for boundaries and splits.

That is not to say that infrastructure isn't important. It's foundational

@ekmiller @bwyble @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

Brad, re evidence of repeating columns outside V1: you see it in V2, MT, V4, IT. More generally, a sign of periodic organization of preferred features is the patchy arrangement of long-range (>1mm) horizontal projections, which in V1 are patchy because they're targeting other patches of similar preferred orientation. My memory of old anatomy papers on higher areas of monkey cortex are that, wherever you look in primate cortex, you see such patchiness of long-range connections. I've forgotten what papers those were though, quite possibly from Jenny Lund, or maybe Kathy Rockland, or someone else ...

@kendmiller @ekmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

Awesome. Such patchiness would be exactly the kind of evidence that I would find convincing. I wonder if it is present in PFC et al.

@bwyble @ekmiller @strangetruther @achristensen56 @NicoleCRust @DrYohanJohn

My vague memory is that it was seen in frontal areas. Not a trustworthy memory though. One of us should dig up those old papers. @DrYohanJohn does anatomy, maybe he could identify the papers, or ask Helen Barbas if she could?

I shared our recent perspective paper somewhere else in this thread (I think).

frontiersin.org/articles/10.33

Helen is somewhat skeptical of the column as a canonical structure: M1 is strikingly different from V1, and parts of ACC (such as A25) have almost no granular layer 4. But 'vertical' and 'horizontal' are good organizing principles for local and long-range connectivity.

@kendmiller @bwyble @ekmiller @strangetruther @achristensen56 @NicoleCRust

Follow

@DrYohanJohn @bwyble @ekmiller @strangetruther @achristensen56 @NicoleCRust

But do you know about the structure of within-area long-range horizontal connections in motor and frontal cortex? My memory is that these projections are periodic/patchy, but I am not remembering precisely what the paper(s) were or what areas they studied.

@DrYohanJohn @bwyble @ekmiller @strangetruther @achristensen56 @NicoleCRust

OK here's one example, from Jenny Lund and others, onlinelibrary.wiley.com/doi/ab

The did both anterograde and retrograde tracing from injections in prefrontal areas 9 and 46. Both intrinsic (within-area) projections and "associational" (to other pre-frontal areas) projections formed periodic patches or stripes, with a repeat spacing of about 780 microns for intrinsic connections; no significant differences were seen between structure/distribution of intrinsic vs associational projections, but they seem to have less data on associational and don't specifically state a repeat distance for them.

I attach an example figure. The figure legend reads "Fig. 6. Brightfield photomicrograph of four discrete intrinsic clusters of CtB-positive neurons (arrows) [KM: CTB is a retrograde tracer; they also did the same experiments with the same results with an anterograde tracer] distributed along the dorsal bank of the principal sulcus (PSI in area 46. All clusters arise from a single injection site near the area 9/46 border, which is located immediately to the left (medial) of the left edge of the micrograph. Clusters of anterogradely-labeled axons are also apparent coincident with neuron clusters. Scalebar = 200microns".

@DrYohanJohn @bwyble @ekmiller @strangetruther @achristensen56 @NicoleCRust

How do you describe that pattern? Periodic but with a weaker continuous pattern across layer 6??

Sign in to participate in the conversation
Qoto Mastodon

QOTO: Question Others to Teach Ourselves
An inclusive, Academic Freedom, instance
All cultures welcome.
Hate speech and harassment strictly forbidden.