What do you think the role of these oscillations are? Many believe they are used as attractor states for memory retrieval.
Interesting, I’ve just read on wikipedia that ‘liquid’ in ‘liquid state machine’ is used as the analogy similar to what the speaker used to describe travelling waves.
The word liquid in the name comes from the analogy drawn to dropping a stone into a still body of water or other liquid. The falling stone will generate ripples in the liquid. The input (motion of the falling stone) has been converted into a spatio-temporal pattern of liquid displacement (ripples).
Interesting coincidence, or did the speaker draw that analogy from ML to neuroscience?
I think they all started with a pebble in a pond and went from there.
I don’t know much about attractor states, so I don’t know if oscillations could be involved in memory retrieval.
I think oscillations are involved in coordinate transforms, and probably other things.
Interference between three oscillations with slightly different frequencies based on walking speed and direction has been proposed to contribute to forming grid cells. This is not proven but it makes sense, especially if walking speed is controlled by oscillation speed in the first place.
My current opinion is that L5 does not process sensory contents and is instead concerned with something like which cortical columns are receiving sensory input, the shape of that sensory input, attention, timing of sensory input (which was proposed by Jeff Hawkins), etc. I’m working on a hypothesis based on that for grid cell formation with propagating waves which solves some of my questions about layer 5.
Questions about L5
My hypothesis only suggests answers to some of these questions.
- Why does one L5 cell type have robust sensory responses with direct thalamic input (at least in some regions) and project up the hierarchy, suggesting a sensory role, but also supposedly acts as the only cortical motor output, suggesting a role in behavior?
- Why does it seem not driven by thalamus in some regions? Why does it supposedly mirror L2/3 as an output layer and why does it receive strong seemingly minicolumnar input from L2/3?
- Why does it have smaller short latency receptive fields and larger intracortically driven longer latency receptive fields?
- What is the role of burst firing, and does it even exist in normal function, or perhaps it exists in a less extreme form? Do L5 slender tufted cells really only fire much during behavior, and if so, why?
- Why are there contradictory results about the influence of the apical dendrite?
- Is there a third fundamental L5 pyramidal cell type and what is it like?
- Why are there contradictory results about the influence of input from a higher order thalamic nucleus, POm?
Generally, I think L5 processes details of how the sensory input begins, such as the sequence in which different parts of the fingertip make contact when poking a surface. Propagating waves are useful for detecting sensory onset dynamics because they occur at the level of the cortical sheet and are useful for precise timing.
Details of sensory input onset is another source of information besides the static sensory input after making contact with the object, but it is ignored by HTM.
Edit: Via another area, hippocampus can modulate walking speed. I’ve only read the abstract.
Theta oscillations regulate the speed of locomotion via a hippocampus to lateral septum pathway (Franziska Bender, Maria Gorbati, Marta Carus Cadavieco, Natalia Denisova, Xiaojie Gao, Constance Holman, Tatiana Korotkova, and Alexey Ponomarenko, 2015) https://www.nature.com/articles/ncomms9521
Scientists have proposed numerous possible roles for brain waves. A leading hypothesis holds that synchronous oscillations serve to “bind” information in different locations together as pertaining to the same “thing,” such as different features of a visual object (shape, color, movement, etcetera). A related idea is they facilitate the transfer of information among regions. But such hypotheses require brain waves to be synchronous, producing “standing” waves (analogous to two people swinging a jump rope up and down) rather than traveling waves (as in a crowd doing “the wave” at a sports event).
If the sub-cortical structures use a wave to coordinate the activities of an area of cortex is that somehow more profound than if it just blinked on and off like a Christmas tree bulb?
I could easily see the wave as an artifact of generating this alpha activity in the cortex.
Are there any neurological theories that use this wave-like pattern of activity to do anything useful?
It seems there’s a significant link to self-organizing-maps when it comes to waves. If neurons close to each other represent similar features then waves are a product of input stimulus. The gif below shows how a bar of light gradually changing orientation causes waves in visual cortex as neurons close to each other represent slightly different orientations.
That is, of course, if this is the same kind of waves they are referring to?
I don’t think that would be different if the waves are subcortically generated, but as I understand it, propagating waves can be generated by sensory input intracortically. The way I picture it, a sensory input to, say, the fingertip causes a bump of activity on the corresponding point on the cortical sheet, which includes interneurons in that point on the sheet, inhibiting adjacent cells including interneurons, disinhibiting slightly further pyramidal neurons and also directly exciting them. The pattern continues from there, so the wave spreads.
What interests me is that propagating waves probably are what generate longer latency responses to parts of the receptive field further from the thalamus-driven center, at least in barrel cortex L5. Based on what I know about L5, it cannot serve a purely behavioral role, but it also isn’t well suited for sensory processing. That contradiction left me without a clue about what L5 does for a while. The longer latency responses were the clue which led to ideas about what L5 does. So propagating waves kind of underlie most of my opinions right now.
Longer latency responses to stimuli further from the center of the receptive field suggest a type of sensory processing which is not concerned with things like texture or other fine details of an object’s surface, but instead concerned with positions of things in space or on the sensor over time. Besides the need to generate a location signal, there’s a lot of information available from the shape of the object’s surface on the sensor and how it evolves over precise time scales.
At some level, the brain must treat these two types of sensory information as separate things, just as it treats sensory input with an external cause differently from sensory input caused by behavior.
I was thinking more of the coordinated global activity like Terry Sejnowski is describing here at time index 16:00:
I think that this activation derived from sub-cortical structures.
I feel like you are describing a different activity more like what is shown at time index 8:00:
I suspect that this bias activity is due to lateral connections within the cortex. And yes - SOM does need to boost cells around the “winning” cell so this would tend to support that interpretation.
Another idea is that at different phases of the wave, different parts of the process are occurring. Perhaps there is an “inference” part of the wave and a “learning” part.
If you check out the “three visual streams” paper they have temporal predictive cells based exactly on that principle.
They break the wave into PLUS and MINUS phases where the plus phase is the upper layers forming an opinion about the “ground truth” of sensation and the minus phase (at the end) comparing a prediction in the lower layers to this ground truth.
The plumbing involves a pass through part of the pulvinar but that does not materially affect the basic mechanism of using timing of the wave to do temporal prediction.
BTW: This particular pulvinar based/predictive mechanism is part of the only plausible scheme that I have seen that accomplishes the long sought goal of a biologically plausible back-prop behavior. If you are interested in this topic you owe it to yourself to do the hard work of reading the paper and references. Some very good stuff going on there.
That second video is basically what I’m talking about, although it’s a bit different because what I’m talking about is partially suprathreshold, but it’s roughly on the same scale.
In barrel cortex, mostly it’s subthreshold, especially for cells in barrels corresponding to whiskers several whiskers away from the stimulated one, but cells still fire in other barrels than the stimulated whisker.
Maybe that’s a different phenomenon than found in the experiment described in the second part of the video you mentioned (8:00), or maybe it’s because V1 expects to normally have some visual input everywhere and if there isn’t something next to a feature, it knows there isn’t a feature there since it is already sort of touching that part of space with the retina.
Hippocampal theta frequency waves are probably subcortically generated, but why would whatever subcortical structures cause that oscillation send a travelling wave when it could cause the same oscillation simultaneously everywhere or starting from random points? It takes a little bit of time to travel, which might indicate functional importance if the time for the wave to travel is longer latency than lateral connections. There could also be a boring reason.
Local processing does not reach everywhere; I can see that a traveling wave could be a chain of local processes.
Perhaps shaped by the feedback from the reciprocally connected cortex.
I just ran into this nice paper on traveling waves.
Do check out the movies in the supplemental materials.
Bitking I have to say that nice paper sure got my attention! This video even shows vector pointers similar to the (instead upstream) yellow wave direction pointers shown in the upper right corner of the ID Lab 6.1:
I have been hoping that there would be a breakthrough like this. The abundant traveling wave action is exactly what I expected.
I’m still confident that HTM theory will work for temporal related learning that must occur at each place in the cortical map these (moving straw that locations in turn look through?) traveling waves travel across. Simplest way I can think of to proceed is add a visual cortex network that maps to retina locations where associated detectors start a wave in a given direction. I’m not sure what else would be needed to convert that information to places in space the navigation network requires but simple wave generating components would be there, and then need something like HTM theory to make a simple virtual critter come to life.
Your news was well timed. For me this winter became a dinosaur tracksite experiment marathon. The work is now mostly over due to spring weather making it too warm for (as in how road potholes are made) freeze-thaw cycle fracking to a depth of a foot or two, of a mostly above ground 20x40 foot block of bedrock. Work is planned for this summer in material with paper thin layers that are otherwise impossible to separate. Somewhere in them is at least one scientifically valuable layer needed by researchers who study here, but the scraps they saw came from an already weather worn area where we were lucky to find what we did.
I sense that neuroscience is finally close to discovering the fundamental principles of how our mind/body works. The need to include traveling waves and a predictive learning mechanism again for us makes the challenge much the same as successfully hybridizing HTM and the wave system I have been experimenting with.
Do you plan on trying to model these waves? Anyone?
I have this old post that discusses the possible origin of these waves. It is my suspicion that these arise in the thalamus and are injected into the cortex in L4. See the post for references and more detail:
You mention a larger model and I agree that HTM fits in this framework. I have outlined what I think that framework is here:
Some oscillations are sent from higher order thalamus to primary cortex in L2/3 or possibly L5a (slender tufted cells). I don’t know if they are relevant to what you are talking about.
Unique properties of high-order thalamic inputs versus cortical inputs to primary somatosensory cortex
In L2/3 pyramidal cells of barrel cortex, POm (higher order thalamus) triggers a fast depolarization (e.g. 75 ms) then inhibition then a depolarization lasting ~750 ms. The long depolarization has oscillations at 10-20 hz and occurs for the weakest stimulation of POm which triggers the fast depolarization.
The authors say the frequency is similar to sleep spindles, which are generated by interactions between thalamocortical cells and thalamic reticular nucleus cells. The depolarization is small but pushed 6x more cells over firing threshold during sensory responses. I think that probably involves disinhibition of apical dendrites (see the next article) so the depolarization is probably not so small, just in a separate compartment than the soma.
Based on their discussion, it seems like an attention signal.
The repetitive/long lasting input from POm seems to allow LTP of input from L4 to L2/3 without L2/3 cells firing.
Higher-Order Thalamocortical Inputs Gate Synaptic Long-Term Potentiation via Disinhibition
The final version has a paywall, but not the preprint.
Rhythmic whisker stimulation or rhythmic paired L4 + POm stimulation (8 hz) causes LTP of input from L4 to L2/3, but not of input from POm to L2/3. LTP occurs without the postsynaptic cell firing.
This plasticity requires POm to activate NMDARs on distal apical dendrites of L2/3 cells and it requires POm to activate VIP+ interneurons, which inhibit somatostatin+ interneurons, overall disinhibiting those distal apical dendrites.
There are multiple frequencies in the cortex that seem to be doing different things.
The alpha rate seems to be the rate that L2/3 uses to communicate between maps and the L2/3 & L4 layers have to work together to come up with an agreement on what to send.
L2/3 may be working at a much higher rate (gamma) to do local voting on hex-grid formation. (or macro-column voting if that is your thing)
I don’t recall seeing any oscillator circuits in the cortex but the lower and mid-brain is full of then. Having the thalamus gate those into the correct parts of cortex as needed seems plausible.
I recall your earlier thoughts and was not sure whether reproducing what the paper and videos show was a problem on your end, of the cortical sheet. Excellent!
My interests have been so way at the tip of the other end (where there is no longer the uniform 6 layer structure and ultimately connects to motors) I could only hope the paper was good news for you and others too.
I’m also excited by the greater than expected progress of this project:
I was expecting this level to be reached in another year or so. If this progress keeps up then by the end of this summer the only thing I might only need to add is a moving invisible shock zone arena environment and plug the critters rudimentary brain into that. Well done!
I have to wnder if there is a path from the lower brain structures, through th ethalamus, to the cortex.
I have noticed that there are patterns very similar to the overall cortex structure as shown in the clips above.
Also found here if you tube is blocked for you:
Following up on this loose end: where does this smell pattern end up in the cortex?