Thousand Brains Hangout with Jeff & Subutai


Follow-up from the hangout.

Factoid - Moser grid cells in the EC repeat across the entire map
Factoid - lateral connections are limited in scope to about a single macro column - about 300 to 1500 µm range with the 500 µm being average. I have papers that give solid backing of these numbers.
Factoid - we know from tachistoscope recognition tasks that the voting can’t take very long - about 1 x alpha cycle seems to be the time scale. This puts very real restrictions on lateral sharing of information through “telegraphing” by lateral spreading.

Q: where are the 1K-brain grid units getting the information to vote if the lateral connections do not span that very long distance?
Q: is Numenta saying that they are describing the underpinnings of the Moser grid cells? If the answer is yes then there must be an explanation of this long-distance property.
I need a biological substrate outlined that describes this behavior. I get the proposal that the columns somehow vote to support the 1K-brain model. They have to have something to vote on.

Example: In this paper Horizontal Synaptic Connections in Monkey Prefrontal Cortex the lateral connections from the L2/3 cells are given as an average about about 500 µm.

The dimensions of the EC map spans somewhere in excess of 1 cm.
The ratio between a column lateral span and the the extents of the EC yields somewhere between 20:1 to 30:1 tiling (as you move along the EC) that has to covered to do this voting.
An example of measuring the extents of the EC:


It’s obvious, any system can struggle at some level of the input destruction. But we are talking about principals here. We can see an object once and then recognize the same type of objects (not even the object itself).


Hi Bitking,

Lateral connections for most cortical cells are limited. The lateral connections of grid cells are also limited and the L6 cells that we propose are grid cell equivalents in the neocortex are noted for the narrowness of their spread. However, it is well known that in the neocortex there are two classes of cells that project long distances: L2/3 pyramidal cells and L5 cortical-cortical pyramidal cells. These cells send their axon into the white matter and project across the hemispheres and long distances with the same hemisphere. These long distant connections are noted in numerous papers and in review papers. Subutai and I attended a conference at CSHL in October and the nature of these long range connections was discussed among many anatomists.

The thousand brains theory states that only some cell layers make long distant connections. Grid cells would not be among them. In the columns and frameworks paper we argue that the cells representing an object should project long distances. We proposed that would be L2/3. I have now progressed further and believe that L2/3 are actually equivalent to place cells, which are stable over changes in orientation. And L5cc cells represent objects. Voting in both of these populations makes sense.

The time of exposure to a pattern can be very short, but the time to recognition will be longer. I recall it takes a few hundred milliseconds before a flashed visual input is recognized. This was measured by in vivo recordings in monkeys. This is plenty of time for the voting between columns to occur. The visual input can be much shorter and doesn’t have to be sustained.

I didn’t read this paper, but could it be they are describing the distance of connections from axon branches that don’t leave the local cortex? It is believed that all L2/3 cells send an axonal branch into the white matter. These are long distance.


I would suggest that you read the paper. The connection described are lateral branches of the axonal projections of L2/3 cells that reach other L2/3 cells. They strongly assert that the bulk of these connections are both long distance (500 µm) and remarkably regular in length.

This (and other papers) is the basis of my hex-grid coding proposal.


@bitking The paper you cited says this:

As shown in Figure 3, labeled axons running parallel to the pia could be followed for up to 2000 μm, even in slices that were thinner (100 μm) than those used for the electrophysiological recordings (400–500 μm).


In the present study, we demonstrate electrophysiologically that these long-distance axon collaterals (originating up to ∼2.0 mm from the recorded layer 3 pyramidal neuron) provide monosynaptic excitatory inputs to layer 3 pyramidal neurons.

~2mm? I don’t see the reference to 500um you mentioned.


Perhaps this could be helpful?
Taken from the same paper:

I did not see if they controlled for cells repeating the excitation and passing it along to longer distances. If this is indeed what is happening then there would be peaks in multiples of the primary distance. Note that this would be exactly what would happen with hex-grid formations.


In the frameworks paper, you say they are corticocortical L6 cells. I assume that’s just a typo because L6 CT cells are the ones with reciprocal connections with L4. In case it isn’t, here is some information about those.

Do you mean a specific kind of L6 CC cell? In barrel cortex, some are more limited to barrel columns, but most are trans-columnar. They can have axons traveling multiple columns and I recall their dendritic arbors are wider than those of L6 CT cells.

The subplate remnant part of L6 (which is mostly a compact layer at the bottom of L6 in rodents and more so the underlying white matter in primates):

The rest of L6:

L6 CC -> L6 CT has long horizontal connectivity:

You talked about displacement cells and some problems with object compositionality. I think L5 TT cells are involved in object recognition in addition to displacements and object compositionality (but not necessarily different mechanisms or concepts). They seem really fit for option selection, like perceptual decisions. When an object is ambiguous, it might need to choose the best interpretation, even just to get rid of extra active cells that represent other objects but are only still on because of noise. Option selection is also useful for behavior and choosing what to attend.

Alongside option selection, I think they convert temporally imprecise representations into immediately relevant ones. This relates to your idea that L5 does precise timing. When it selects between objects, it says, what I’m looking at is object X. Whereas a union of possible object SDRs depends on what parts of the object it has already seen, so that is temporally imprecise in a sense.

The strange projections of L5 TT cells, to thalamus and motor structures, support this. They probably aid subcortical sensory processing, e.g. in the superficial layers of the superior colliculus, because some motor things, like maps of potential targets, are hard to distinguish from sensory things. They are also the only cells which drive the thalamus. Perhaps they solve some sort of “language barrier” between the cortex and subcortex. The cortex integrates information over time, like when the sensor explores an object. That could confuse subcortical structures, so perhaps L5 converts to something with immediate relevance. It could choose between options to say that the object is X, or to generate a precisely timed motor command from a behavioral plan signal. It could convert relatively stable activity into precisely tied transient signals with ramping activity or something like presaccadic predictive remapping. This also makes L5TT a good driver of the thalamus because it generates an item in the next hierarchical level’s stream of inputs. The thalamus needs a temporally precise driver because its burst/tonic mode firing is time-sensitive. This would mean every cortical region receives a direct subcortex-like (temporally precise) input from the thalamus, even the ones in primates without any direct transthalamic sensory inputs.
(I don’t mean the subcortex just cares about the present, just that the signal to it needs to be like that.)


That is fine, but I don’t see your argument against the Thousand Brains Theory based on the length of lateral connections in L2/3 pyramidal axons. I did a search and found at least one other paper talking about these long range connections. I don’t think your argument that lateral connections in the brain are insufficient to support TBT holds water. (If that is indeed what you are arguing?)


I took a quick look at this paper. I recognized it, had read it a long time ago. They are looking at slice preparations which would not include the connections I am referring to. The “long range horizontal” connections they are referring to are from L2/3 axons that don’t leave the local layer. They are not discussing the L2/3 axons collaterals that exit into the white matter and project to other regions and across the corpus callosum to the other hemisphere.


You are correct, sorry for the typo. The L6 CT cells are the ones I was referring to. I am not familiar with what they look like in barrel cortex. In Thomson and Lamay 2007 they describe these as “upright pyramidal cells with a well-developed apical dendritic tuft and a terminal axonal arbour in layer 4”. And “neither class of CT cells [L6a and L6b] has long horizontal axon collaterals in the infra-granular layers. The majority of the axonal branches turn up toward the pia close to the some.”


Once again I am coming off as attacking your work. I am not doing this at all. I am trying to fit it into a real world problem to validate it and have been unable to do so. I was hoping that you would understand the problem I am seeing and know how it can be worked with.

So far I have gotten from Numenta is a wonderful way to look at the computational approach from the predictive viewpoint; this has solved several theoretical problems I was having in state transitions from one stable neural pattern to another. Your view of the dendrites as an SDR seems to be a much more productive way of understanding what is happening.

Since I have come to understand what you are proposing with HTM I have reworked much of my earlier work and find elements of HTM to be a powerful tool-set.

I have been working on a pet project that shows that the standard HTM model can be applied to solving the visual palimpsest problem. That is the one where subsequent visual fixations all lay one upon the other and yet we somehow combine them into a stable and unique representational code. The works of Calvin meshes wonderfully with HTM to yield an elegant solution to this problem. I will post it as soon as I finish the illustrative images and tighten up the text.

I am working in Uganda at the moment and don’t have access to my library and notes so I offered the only topical paper I happen to have anywhere on the forum. I have several other papers that support the 500 um lateral connections in my L2/3 assertion - just not with me at the moment. And yes, the stable output pattern between maps is also projected laterally within a map layer to work with other L2/3 neurons to form a local representation. This is exactly the connections that form the Calvin hex-grid tiles.

This is not an attack - I agree with your assertion that the cortical computation should be the same over all of the cortex; I just don’t see how to do this and am asking you how to make it do this task

As far as my questions about TBT model - I am trying hard to apply it to this problem and have been unable to see how to make it fit. The issue is the combination of a dozen or so sequential image snapshots formed by saccades passing through the map hierarchy and don’t see how they avoid turning that into globs of activation that can’t be separated into distinct codes; one activation response to an object in a column looks just like a different one.

To be clear on what I would recognize as a solution - stacks of different input visual images projecting to the same V1 cortex should end up making distinctively separable patterns that are spread out over the EC structures. This requires processing in both time and space. A winning answer should end up looking like the patterns that Moser has observed.

I am offering that if you take the property of relatively fixed long-range lateral connections to spread out the activation it may be a way to make the TBT model spread out over this space but to be truthful - I have not been able to make that work to produce the grid activation patterns and was hoping that you have solved this problem in some way.


Yes, they look basically exactly like that in the barrel cortex. Their tuft and axon aren’t always in L4, but usually, and they’re the same besides that.

In rats, at least barrel cortex and V1, some target L5a, or lower L2/3. They seem to target wherever the thalamus targets, and the same group targets the source of that signal. For example, the L6 CT cells in barrel cortex which target L5a also target the septa in L4 and POm. POm targets L5a and the septa. The tuft is also in the same layer.

I’m not sure this is normal. Those L6 CT cells in barrel cortex are in the septal domain, which is a higher order compartment or region, unlike the embedded barrel columns. I’m writing an opinion article about that. Other primary regions in rodents seem to have higher order components too, but much smaller. I think that supports the thousand brains theory of intelligence. If the septal domain isn’t completely separate from the embedded barrel columns, it means hierarchical level is comparable to submodalities (scale of receptive fields) or else it wouldn’t make sense for them to merge. In rodents, some hierarchical levels might share some layers somewhat, for efficiency.


Jeff, these local lateral connects are EXACTLY the ones I am talking about. I see these as integral to forming stable separable activation patterns in a single map in the cortex.

I am working to show how and why these connections are important to the HTM model. I understand that you are busy enough that you don’t have time to chase down every crackpot supposition that HTM fits into someone’s pet scheme.

I imagine that you have distractions flying at you from all directions and have to budget your time and attention carefully to keep developing your work. At this point my work with the Calvin tiles fits into the class of those distractions.

I will put up an explanatory post that clearly shows why these lateral connections are important to the HTM model as soon as I get the images and text to the point where it is clear and readable. Until then please stand by.


Sorry Mark, I thought you were providing evidence against the TBT, so I was defending it. If you are just pointing out avenues it can be taken for research at a macro level (higher than we are currently investigating), we welcome any effort or input from the community. All research in these areas are valuable, and I commend folks like @bitking @Paul_Lamb and others who’ve contributed their ideas back to the community. That being said, we continue to find the most important questions still within the granularity of one cortical column and its internal mechanisms.