Layer 6 Notes Summary

Things Which Don't Fit a Review (from the previous version of this post)

Things Which Don’t Fit a Review

Links to google docs for notes

From most to least recent-

Layer 6: https://docs.google.com/document/d/1vgfQAY3-uo9cDLKumHLuiuuQ06wyZhwk8Wy--6zS9oE/edit?usp=sharing

L5 Notes 3: https://docs.google.com/document/d/1AvmHC85iZxWzDMJY5Oe0A-4IthvhzJFTA1NO4FVoUDo/edit?usp=sharing

L5 Notes 2: https://docs.google.com/document/d/1PLThMPgXVmZFP_ibe6uHaq1adev7N5DOtSybSkF-5tw/edit?usp=sharing

L5 Notes 1: https://docs.google.com/document/d/1ySuL-qf6hHSH7Iw8THLFMy21pV_q6KzRmsnv60dD480/edit?usp=sharing

Neuroscience Notes (hippocampus, EC, and L5/6 mainly but I’m not drawing on these since they’re old): https://docs.google.com/document/d/1pF0cWO0DueySSmmqHBUQNEjDyRVVSDIKU__DoQV8zZg/edit?usp=sharing

I’d just treat the last two as sources of sources because I was newer to neuroscience, and the first three will probably take a while to work through so maybe treat those as sources of sources too.

Terminology which is potentially confusing and hard to look up

Corticothalamic Cells: Cells which project from the cortex to the thalamus.

[name 1]o[name 2]al: From [name 1] to [name 2]. Examples: corticocortical, thalamocortical, and corticospinal. The names are sometimes confusing. Retinogeniculate: from the retina to LGN, which is the primary thalamus for vision. Corticotectal or corticocollicular: from the cortex to the superior colliculus. Corticocollicular can also mean to the inferior colliculus. Corticofugal: from the cortex to outside the cortex.

These get abbreviated often. I will use these abbreviations: CC (corticocortical), TC (thalamocortical), and CT (corticothalamic).

Driver or driving: roughly speaking, a signal which can cause cells to fire without other inputs. A more exact definition is based on the properties of the synapses involved on their locations on the receiving cell.

Primary thalamus: a thalamic nucleus whose inputs from the cortex typically cannot cause cells to fire if there are no other inputs. I use it to refer to primary sensory thalamus, which receives driving inputs from the sensory input, but it can also refer to some motor nuclei, which receive driving inputs from structures like the cerebellum.

Approach

I try to take a lot of notes and look for consistencies and contradictions to minimize misinformation and make inferences about what’s going on. The main goal is to eventually generate hints towards roles of cell types.

The main challenge with neuroscience is that very few things are solid facts. Basic things, like which cell types are directly driven to spike by the thalamus, are still disputed, so it’s hard to theorize about intelligence. But neuroscience also provides a solution to its massive ambiguity, which is its massive redundancy. There are loads of studies about something new and unique, but also plenty of studies about the same thing in different regions, species, and brain states, using different techniques. This redundancy helps determine what things are part of the fundamental circuit. More importantly, it’s all part of larger circuits, so a bunch of possible facts can point towards that circuit.

I don’t like the term circuit. It implies something like a computer circuit, with discrete steps. If you look at the same transistor in each instance of a repeated circuit, you can’t really assign it a role of its own. You can only say that it does something which is meaningful for the output of the whole circuit. The cortex isn’t like that. Each layer has a bunch of inputs from and outputs to other layers, so there’s no clear flow of information. Most layers also have inputs from and outputs to other cortical or subcortical areas. The point of the cortical circuit isn’t to take an input and spit out an output. Instead, each cell type plays certain roles.

Take layer 5 for example. If I understand correctly, L5 is thought to be for relative location at Numenta. All the signs seem to point to that role. It generates motor output (a change in location) but also exists in great numbers in non-motor regions, projects up the hierarchy, has wide receptive fields on the sensor (allowing converting to location in another spatial system), has lower selectivity to details than most other excitatory cell types, and communicates with L6 and itself through far-reaching connections. That’s probably one part of a broader role not so easily expressible in words, but it’s something to build on, unlike a rigid circuit. Other ideas about roles likely led to this one, like the idea that a major role of the cortex is to handle locations on objects, so roles help determine more roles.

Roles can be expressed with a brief thought, but neuroscience has a lot of ambiguity, so it takes time to arrive at that thought. I’m not an expert at neuroscience, but it took me two years of researching layer 5 and several hundred pages of notes to produce a hypothesis about what it does (sensory onset processing for grid fields on the sensor), and that hypothesis wasn’t very good. Still, I think this approach works. I read on this forum that L5 is thought to be for displacements and misunderstood, thinking displacements were changes in location rather than relative locations of parts of the object (I could still be wrong). Based on the idea of forming grid fields on the sensor’s surface, I thought it might be for location instead of changes in location. That’s a small change, but some of the evidence for this role was my evidence for the prior hypothesis. The point is, thinking about things in terms of roles allows flexibility and building up a framework. I have wasted a lot of time by not doing that.

Thoughts about sublayers

Physical sublayers are not important to intelligence, since they are just how the cell bodies are organized. What really matters is the cell types and their processing. However, determining cell types seems to be hard, so most studies use physical sublayers to help infer cell types. Physical sublayers are a tool for interpreting results. For example, in the barrel cortex, slender tufted cells are in L5a and thick tufted cells are in L5b, so the depth in the cortical sheet indicates the cell type.

It is important to get the sublayers and their cell types right early. Layer 5 seems to have a mostly unacknowledged third cell type, not limited to one region, and it took a long time before I learned about it. That made a lot of my notes ambiguous, which I’d like to avoid repeating.

Linking physical sublayers to cell types is not straightforward. L5a and L5b are swapped or blur together in some species and regions, and it has further subdivisions at least in motor cortex, for example. In layer 6, the connection between physical depth and cell type already seems more confusing than in L5.

Are primary CT cells only in L6A in rat V1 and barrel cortex?

The apparent depth ranges are actually quite similar in (8) and (10), based on differences in their defined widths of L6Bb and whether it was included in their usages of the term L6. That suggests they are both in L6Ba.

Their depths are unclear, however, based on (8 fig. 3) and (10 fig. 2). The study which claims they are only in L6A found a small fraction of them a bit past half the depth, and they excluded different fractions of L6 as L6Bb, so they found nearly identical ranges of depths. Since they both found basket cells limited to L6A, CT cells in barrel cortex are organized into cortical columns only in L6A (7), VPM is more easily identified as a target than POm (17), and the density below L6A is much lower in the study on barrel cortex, they are probably on in L6A in barrel cortex as that study claims. The study on V1 which found them in L6A and L6Ba found a similar density in both of those sublayers, but it classified cells based on whether the apical dendrite passed into L4, which is somewhat arbitrary and therefore could produce false clusters. Also, the basket cells were still found only in L6A in V1, so overall it is unclear whether they are in L6Ba in V1.

The anchoring effect

The putative TT cells are described by an old article that has been cited over a thousand times. It’s unfortunate a single article with potential flaws can have such a large influence. I worry that drawing on this article will just amplify widespread underlying opinions. Mostly this is an excuse to talk about something relevant to research. There are probably better examples of the anchoring effect in the literature.

The first study on a topic usually has a huge impact on the field. People form opinions based on the first thing they see. For me, that makes me not want to take notes on contradicting things, not because I don’t want to have those notes, but because it’s easier and feels like more progress to build on an existing framework of knowledge.

I find it’s a good idea to throw out the first couple things you hear or read about something. Even though I try to do that, what I read first still ends up biasing me often, and sometimes there’s not much choice because not much information is available. It’s an especially big problem for abstract phenomena which seem like big hints at roles of cell types.

It doesn’t just apply to opinions about the facts like cell types. Hypotheses about roles and mechanisms are perhaps even more strongly impacted. Right now, I am heavily biased by my first idea about what layer 5 does. It involves receptive fields effectively travelling across the sensor, so the long-reaching axons from L5 into L6 and the very long receptive fields in L6 make it hard not to think about L6 in terms of that idea. That’s a major problem if my idea is a bit off the mark or plain wrong. I’ve heard it’s a good idea to keep multiple hypotheses about the same thing at any given time, which could help.

I suppose the period between the first hypothesis and the second one is a good time to lock up that first hypothesis and pretend it doesn’t exist, at least for a bit. It’s too hard not to get excited, though.

Extra credits has a good youtube video about the anchoring effect.

An attempt to assign the CT classes to two responsivity classes

A couple of studies identified a group of cells with response to sensory inputs and a group without. Both describe ambiguous classes. The unresponsive type is probably primary CT, while the responsive type is probably dual-projecting CT.

The first study (or rather, series of similar studies) is only about subcortically projecting L6 cells in awake rabbits (3). Those cells are probably corticothalamic, but I only have access to the abstracts. That series of studies is about somatosensory and motor regions, but at least in V1, there are some cells called Meynert cells which project to the superior colliculus, unusually long distances within V1, and to other regions (12). More research is required, but Meynert cells are pretty unusual and perhaps evolutionarily derived from L5, since they are near the border (12). For now, I will assume the series of studies only found CT cells.

This series of studies did not explicitly describe two groups, but it comes close. It found that a bit less than half of the examined L6 cells respond to sensory stimulation with thalamus-driven spikes, and a similar percentage in another region had subthreshold responses. That suggests there is a group of CT cells, similar in numbers to the other group rather, which receive little net excitation to sensory stimuli.

The other study only reconstructed the unresponsive cells, and they were corticothalamic (2). It might have included CC cells, but it found twice as many unresponsive cells than normal cells. They selected regular spiking cells and CC cells have phasic responses, so they probably mostly used CT cells. The reconstructions appear incomplete, especially their tufts, but they still show unresponsive cells with apical dendrites reaching lower L4, which are probably dual-projecting CT cells. The responsive cells have apical dendrites only reaching lower L5, and both appear to be normal pyramidal cells, so they are probably dual-projecting CT cells rather than primary CT cells or CC cells.

The responsive cells in this study had excitation and inhibition at the cell’s receptive field, which drove spikes. The unresponsive cells did not have spiking receptive fields, but tonal receptive fields were identified based on local field potential and EPSP components of responses. Their lack of responses was caused by inhibition arriving a couple of milliseconds earlier than their excitation, inhibiting spontaneous activity rather than causing spiking. If the unresponsive cells are primary CT cells, this is consistent with them being in L6A, which is where all basket cells are located in L6 (8, 10).

There are a few differences between the studies which suggest these are not the right classes, all of which are from the study which used anesthesia rather than awake animals. The use of anesthesia probably explains all of these differences. The similar numbers of responsive and unresponsive cells in (3) suggests two classes, but (2) found twice as many silent cells. Anesthesia reduces sensory responses, especially complex ones not directly driven by the thalamus, which probably explains why they found few normal cells. Different regions could also have different numbers of each cell type (compare barrel cortex and V1 in 8 and 10). Another difference was spontaneous activity. The studies on awake rabbits found less spontaneous activity in unresponsive cells, whereas the other study found more spontaneous activity in unresponsive cells. Since sensory input causes reduced spontaneous spiking via interneurons, and since anesthetic probably reduces spontaneous activity of interneurons, which was quite fast in (3), this difference is attributable to the different brain states. The last difference was whether they concluded that the responsive cells had monosynaptic responses to the thalamus. The study which concluded they have monosynaptic input based that on latencies, but interneurons have longer latency sensory responses in L6 than L5 and L4 (3), which probably caused confusion.

Source (2) suggests that the unresponsive cells would respond with more complex sensory stimuli than a constant tone. That does not appear to be the case, depending on how complex the stimulation must be (3), but there must be some situations in which they respond. Perhaps they simply have extremely sparse activity, consistent with them having a role in synchronizing thalamic cells which are responding to the same feature. That means dual-projecting CT cells have another role, but it must be related since their projections to the thalamus have similar properties.

Interpretation of extreme sparsity: L6 has massive capacity for connections and responses, mostly unused

If L6 CT cells synchronize thalamic cells to separate firing times for each feature, then there are many potential responses needed, but fewer will end up used. This isn’t the case for simple stimuli like lines at particular orientations, but with just a bit more complexity, the possible numbers of those features becomes massive. Therefore, L6 needs the potential to form many different synapses, but only uses a small fraction.

The long horizontal axons in L6 from L5 cells have many apparent contacts, which are close enough together to suggest that they form synapses, but few of them actually do form synapses (6). In HTM terms, these connections involve spatially large pools of potential inputs, with few of them over the permanence threshold.

There are many CT cells in L6 which do not respond to any simple stimuli. Perhaps these really do not respond at all and are reserved for things which have not yet been learned. Cells which do not correspond to any feature and therefore do not yet synchronize thalamic cells for any feature might be needed since highly selective responses are required.

This role is compatible with burst and tonic modes in the thalamus. Bursting thalamic cells could draw attention to a place in space, rather than features since they are not put into time buckets for each feature and last longer than single spikes, and then CT feedback could change the attention to what is there. There are two forms of attention, to place (both egocentric and allocentric) and to identity. L6 could contribute to both, and the reticular thalamic nucleus could be involved in selecting locations to respond since it can put cells into burst mode. This mechanism could relate to the unresponsive cells which are inhibited slightly before they are excited, perhaps showing timing sensitivity.

I still need to research the thalamus.

Arbors are often misleading but can be useful

Dendritic and axonal arbors define where inputs can arrive from. If an arbor is restricted to a certain layer, it probably sends signals to or receives from that layer. It is also possible that dendrites or axons from another layer or sublayer come into that layer, where they connect, so arbors alone can be misleading. The long horizontal axons from L6 CC cells (7) and L5 cells (6) are long but each axon synapses somewhat sparsely, for example, leading to weaker overall connections between cell types than expected. Still, arbors can hint at roles when taken in the context of other hints at roles. In HTM terms, those long reaching axons have many potential synapses but few are above permanence threshold. Which roles could involve that?

Receptive fields can be misleading

A receptive field is just a place to which a cell can respond. That doesn’t mean it always responds to something in the receptive field. Cells in the same minicolumn can have the same receptive fields, but HTM’s temporal memory shows that not all of them necessarily respond. Taking receptive fields to be straightforward can hide a lot of what’s actually going on.

It might seem like very long receptive fields are simply that, but there is probably more to it, especially since the inputs which drive those receptive fields are from L5 slender tufted cells (they’re corticostriatal and corticocortical), each of which controls a separate part of the receptive field.

For example, if I ignore my anchoring bias, perhaps they serve to generate scanning RFs. Slender tufted cells have long latency responses, so they have wide ranging latencies, so the place on the sensor determines the latency of the long RF cell’s response. That in turn could modulate the thalamus to put each feature in a separate time bucket. For the L6 long RF cells, this process is literally scanning RFs, and for the thalamic cells, this process is scanning over features.

Why dual-projecting CT cells might be the ones with very long receptive fields

Dual-projecting CT cells have stretched axon arbors in the primary thalamus (and perhaps higher order thalamus) and in L5a. Their apical dendrites are also in L5a, where slender tufted cells are located. Slender tufted cells are probably the source of the long-reaching axons from L5 to L6 (4). In barrel cortex, slender tufted cells have smaller receptive fields than thick tufted cells (24), consistent with the putative thick tufted cells in V1 having responses which are less selective for location (4). Since slender tufted cells are presynaptic to thick tufted cells but much less often the other way around, thereby being better suited to have the less advanced receptive fields, and since the L5 cells which do not target L6 as much project to the superior colliculus in one study, the cells with long lateral axons in L6 are slender tufted cells.

For these stretched axon arbors to be involved in feature detection, they must be able to orient along the axis of any feature. Depending on what constitutes a feature and how it is mapped in the primary thalamic nucleus for whiskers, dual-projecting CT cells might not meet this requirement. Their arbors are not stretched in random directions. Instead, their axon arbors in L5a and in primary thalamus are stretched along axes corresponding to the row axis on the whisker pad (10). Assuming the arbors are one cortical column or TColumn wide, then these arbors correspond to rows of whiskers (or sections of those rows), arranged from the front of the head to the back. When the rat or mouse whisks, it moves its whiskers rhythmically back and forth, so whiskers along the same row move through overlapping areas of space. That makes this potential problem more ambiguous since it might identify features as sequences of whisker contacts, or something similar. In that case, it might not need arbors wider than a row to bind features by synchronizing thalamic cells. One potential counterargument is that receptive fields in barrel cortex tend to be stretched along the row axis, perhaps simply because they have lower acuity along that axis, so the stretched axon arbors simply match that stretching. However, if the feature binding characteristic is true for primary CT cells, it is probably also true for dual-projecting CT cells. Also, one study found a map of the whisked space (25). That means their receptive fields could actually be more circular, but in the whisked space rather than in terms of which whisker contacts the surface. Since the whiskers move through much of scanned space, many along the whisker axis will contact a surface at the same place in whisked space, giving the impression of stretched receptive fields. Perhaps the stretched axon arbors serve a role in mapping the space being moved through. However, it is worth researching whether primary CT cells actually could synchronize thalamic cells which correspond to any arbitrary feature.

Facilitation in TC Connections

This section is not specifically about layer 6.

Just as L6 corticothalamic cells facilitate, some thalamocortical cells also facilitate in the same way. Input from primary thalamus to a primary region does not do so, but the input to the primary region from higher order thalamus facilitates (20). Like L6 CT cells, this facilitation continues after many spikes and the synapses involve metabotropic receptors (20). Also like signal from CT cells to the reticular thalamic nucleus (17), these metabotropic signals are probably sent to both excitatory cells and inhibitory cells (21-23). L6 CT cells seem to control the thalamus similarly to how higher order thalamus controls primary cortex, at least broadly speaking.

CT Signals to the Thalamus

It is important to keep in mind that CT cells project to the thalamus. I haven’t done much research on that in a long time so this is based on memory. All types seem to target thalamic cells distally with facilitating synapses. They control firing mode, switching cells from burst mode (when thalamic relay cells are not excited much, they respond to sensory inputs with bursts) to tonic mode (they respond with single spikes). However, CT cells also project to the reticular thalamic nucleus (sources conflict on whether or not both CT types project to the RTN, perhaps demonstrating a third CT type), which can put thalamic cells in burst mode.

3 Likes