Numenta Numenta Research Meetings, 28 Oct 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Tsodyks-Markram model of STP (part 2)

Description

Moving forward on the STP model to show the first application (beyond merely matching electrophysiology), a working memory model by Mongillo,Barak &Tsodyks (https://science.sciencemag.org/content/319/5869/1543). Free PDF access through here: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.352.9618&rep=rep1&type=pdf

A

You guys ready.

A

You only have an hour: I should get going. Okay, let's go.

B

Okay, cool maybe said he go alright, so I'm gonna try to keep this. You know couldn't.

C

B

Good thing is already laid some groundwork in this first time we talked about facilitation in terms of the biochemicals and how to understand facilitation, what it actually does in terms of releasing increasing the release probabilities for synaptic vesicles to make signups use more powerful after just a couple of bursts, a couple of spikes right then, last time I talked about this model here, which is very nice account of synaptic depression. I, just like can quickly go over it again, just to refresh your memory right.

B

The idea is that there is some some reserve pool of synaptic resources captured by some variable X right and that whenever there is a spike at time, delta T right. So this is a is this simple function, it's one at the time and then you're gonna use some fraction of the available resources that gets taken out of the result and add it into sort of the active portion portion of transmitter.

B

This active portion of transmitter, however, the case with some time constant that is typically the synaptic, referred to as the synaptic time constant. So this can be something for foremost through middle cell transmissions. That is relatively short time. Constant, like five milliseconds for AMPA receptors can be bit slower for like NBA receptors, 100 milliseconds, but it's relatively fast compared to the recovery operation of them inactivated the transmitter rate, so this inactivates with center console. So this tournament gets obstructed here and shows up.

D

B

Then this term is inactive. Transmitter gets recovered into the reserve, pool up there with some slower time concept like 500 milliseconds, or something for typical pyramidal cells and super granular layers.

B

I think the numbers are a similar for layer five, but that of course, is just a depression part, and so the one thing that you can now because it's slow, because it's slow to put things back in the reserve. We talked that there's actually multiple mechanisms at work right, because one is the receptor to the transmitter we have take where it actually gets. You know put into the cells again and then put into mystical x', but another part of this recovery time.

B

Constant is also the movement of vesicles from a reserve pool into many releasable vesicles and all of those are modulated by a calcium signals in various ways. So the by molecular machinery talked about the wave first time. It's quite intricate. At least I alluded to a lot of the details. The time I'm not here to talk about that because not relevant to what I have in mind folks, because what I actually want to get you is what all of this is useful for.

B

So so the last complication then to take this model and to incorporate facilitation. Is this? We use, excuse the somewhat weird number culture: it's not my fault, it's their original, sonex and mark on papers and they're. Both you know euros of mind, swap and try to mostly use the notation so where you now say this use fraction of the percentage of transmitter that is used when there is a spike right.

B

This is now not a constant anymore, so before I said like that number, maybe is something like a constant twenty percent, so that, once you sub have your first like submitted, there's only eighty percent.

B

We saw slip which means that the second spike does only you know one point, one point: six seven and the third spike does it less unless so that, after four five spikes, the synapse is suppressed, but this number is now changing because you might have facilitation to, and so in order to unify these models, they said: okay, let's drill out this number and actually make this time-dependent as well.

B

They do note that by adding a I, don't really know, I think they should have added a hat or a bar or something, but in any case you now describe this as a temporally dependent variable. So actually the facilitation now is not a constant number, but it changes right. The syllable facilitation goes up and then we spike we're sort of the un-- facilitator part can increase by some. So this is now the real facilitation time constant this facilitation factor. So this might be something like 0.05 or so.

C

B

5% or something that your use of transmitter goes up every time you have a spike. So that's the facilitation model, but it's incorporated into a model that can model synaptic, depression and obviously there's it gets complicated. Now, because you have both of these mechanisms at work and there can be spike trains which are fast or slow, but you could do some edge cases.

B

So, for example, when the two spikes that you are transmitting are really far apart, then you can cancel out, for example, the fast time dynamics here, and you can assume that there is no remaining facilitation or depression from a previous bike and you can simplify something so there so that you can get simple. You can resolve that differential equation with the simple exponential.

E

B

Which then depends on the time between the two spikes and the facilitation time? Constant? So just to give you some ideas, sort of Phyllis, said facilitation. Time constant, obviously depends a lot on these calcium dynamics that act on increasing the release of transmitter.

B

I mentioned when I talked about this two stalks ago, that there's fast mechanisms of F, 1 and F 2 facilitation, which have time constants of like 10 to 20, milliseconds and like 100, milliseconds and there's slower process, is called augmentation, which might be on the scale of light like like a second or two and then there's also so-called post titanic' potentiation was also a short term effect and that decays after you know a minute and so for for modeling any particular neuron you're going to need to fit those parameters and just to give you some typical parameters for typical pyramidal cells.

B

That I have been working with I'm going to give you some numbers here, so keeping my pressure time constant. The time for recovery may be a second that time constant here is on the order of five millisecond for millisiemens milliseconds for their ample receptors. That facilitation time constant, is to the key larger, so.

B

Unfitted to account for pyramidal cell behaviors, and so this is a relatively slow process. The facilitation tends to stick around for longer. Then it takes the neuron to recover resources to make the synapse effective again. That turns out to be a really useful thing and to give you some quick intuition, though I'll first turn to this plot here, so that is from the sonics Markram 98 paper. What you see on top here is a depressing signups, but you're measuring the postsynaptic potential. So what is observed as a deflection of the membrane potential.

B

I presume this is I presume. This is a somatic jump. I mean this is. This is also probably simulated, but it was. This is how it looks, but that's some nice extra metal coatings that actually look like a typical, so.

F

F

Exactly right so.

B

So what you see here is some moral madness depressing. So you see that the first fight is very big.

B

It deflects almost a full Mabel disobey to clear, very large, deflection on the postsynaptic side and then the second spike is already quite a bit quite it it's smaller and the third one, it's even smaller on the fourth, you know that the knurl is essentially exhausted right and then, if you keep this bike trained up, the pre-synaptic spike train, there's never enough time to recover those resources, so that synapse will not regain its strength until it's actually silent for a while. You see the opposite in.

F

B

Was something to say.

B

F

B

Like what did also says.

F

B

F

Have a facilitation facilitated.

B

Signups, so one thing you already notice here is that with a bit of a different scale here right, so this is not co2, long believe all the co2 so because of the first dilatation now the sentence builds up strength so over the course of some six seven spikes it grows to like five or six times the original strength it is I can quickly Apple that, of course, Marcus original intuition. When we talked about this last time that indeed facilitating silence is tend to be originally weaker and then become very powerful and indeed very strong.

B

Sign of disease tend to be faster pressing because, of course, when you're, a strong sign, ups you're, probably also using a lot of vesicles. So so there's a powerful reuptake.

F

Blower situation.

B

A smaller correlation- and it does make.

E

Some me to defense.

E

You can reframe the siltation and depression to be depression. Depressing synapses are ones with a very low threshold, and facilitating synapses are ones that have a higher behind that there's a lot of oh yeah, you saying the data that might be.

F

The right answer, so that makes sense me I.

B

Didn't either but I looked it up after we talked about it. I just wanted to mention it. So it's generally true. Then, of course you know 30 40 percent of the cases that don't fit. That story.

D

C

F

B

Of cases probably.

F

B

So it is true that you can often find like the specific microcircuit pathways, typically have a known characterization, so certain pathways are known to be faster, pressing.

A

B

Adapting look so, for example, like basket cells from their fire. Yes,.

F

Very lives are very common, just.

F

Right you started low, is all been increasing it, but.

D

It's not obvious.

F

Here, by the way, look at this in the first spike on the facilitating ones about 1 millivolt mine is about 1 millivolt. Therefore, this site doesn't show that right that doesn't show that, but that's not about logic. According that's assuming it.

F

B

It doesn't make some sense in the sense that if you have a very powerful synapse, you're gonna have a lot of leasable vesicles, so they're readily releasable pool of of w dock. This, like this protein, that's in synaptic wall, they're right where the vesicle doctor, which makes it really really scible upon, like a calcium confirmation signal which you get because as calcium influx when there's the presynaptic spike, which n actually released its vesicle. And so, if you have lots of releases the vesicles, you would deplete the available pool.

B

They have asked, which means that powerful sign up sees. They often have more powerful. We uptake mechanisms to so that, so they have a big engine, but it turns out very quickly. I.

F

B

That makes no scientific inquiry very difficult, because typically, the way that people measure the strength of signups is it does have a slope in on the signups with one individual spikes so that it can measure IP STIs and then they average a lot of those. But what you actually want to do is you want to mimic more biological?

B

You know patterns of communication which might be domestic come up or you have like a short packet of like high spikes and web succession at a high rate right, which actually get would you give which get which get signals right differently? Then, if you transmitted five spikes, you know.

F

B

Way, I understand it: I mean that both exist I, don't know exact numbers, but it seems to me that there's a balance, some sort and it's a little bit dynamic too, because of course it depends on you know how they sign it.

F

Right that could basically flip the way, I think about learning right.

B

Which is exactly why I'm talking about this.

F

F

The mark of being able to put.

E

Memories in it that wasn't I was going for something even simpler than that. I was going to just saying that you can.

F

F

B

Is a big idea and in fact it is right now at the core of the break-in paradigm shift in how we think about working memory. One of the reasons why so I started thinking a lot about working memory is because it allows us to explain a lot of activity, petechia makak, working memory, performance and mattresses of recordings, and it is really neat mechanism to get activity silent, working memory, because then you can offer information in the facilitated silences yeah. So this was exactly the case.

F

Minutes exactly was passing and.

F

And if I saw try to market III read because these are these are people typical working memory, but these memories.

F

B

Because it's the times gets an interesting question right: this bridges from spiking dynamics on the on the scale of milliseconds, all the way to buffering memories for a couple of seconds, maybe half a minute right, but it doesn't explain. You know that you have such a key memory for my.

F

B

Yeah, it depends on these time constants right there, people feel mentation with like 6 seconds. If you have a big pool, you know you can at the top yeah okay.

F

B

So what I really wanted to get to is this, which is, which is proposes an account of working memory on the basis of this through the job possum.

D

B

That is from 2008 I think these theories have become Lully. You know sort of alive in part because a colleague of mine who make a long list who went to MIT to join Miller lab, who proposed a bit more extravagant version of this, the lot more cortical micro circuitry, which actually generates very specific characteristic markers for that kind of working memory activity, which then was subsequently confirmed in macaque multi-item working memory, tasks that Miller there.

A

B

A

A big deal out of it.

B

So he's calling it what a memory tree point. Oh that's how we originally met when I sort of talked a little bit about that at the CNS, that the count of neuroscience Society see a meeting last year, and so so this was sort of one of the first papers to push the idea. It's not terribly intricate here yet, but you will but I want to show sensually we'll start with this. So this already shows almost everything that talk about.

B

So one thing is you now have this X variable that reserve pool right and with that gets depleted with a respite by this. You utilization factor right, so how much of the sign of these resources gets taken out when the spike has to be transmitted? How many? What percentage of release the listicles right get taken out, and so that changes the changes, uses up the transmitter reserve and we get to recover with its depression hang outside. But now this you number is again itself dynamic so because it there might be some facilitation, dynamic.

E

B

Right so these equations look a little bit similar to those questions. I hope you recognize that so when you now and you have some pre post pair right and you want to transmit some spike train. What we see is that when you transmit a couple of spikes there, this facilitation that gets build up right. So this number goes up to you know. Almost one like it converges on one is infinitely fast spike train that has to be transmitted and, of course, the synaptic resources deplete at the same time.

B

So in our facilitation and depression going on at the same time, but I want you to notice here already is that the depression recovers faster, then the facilitation decays away, because the facilitation has a time constant of I think in this papers, 1.5 seconds, whereas the depression recovers with 500.

B

So that means that when you now have a period of silence and you after the rest want to transmit the signal spike, you now get the benefit of having recovered the depression, but still having a facilitated finance. The resulting epsp will be a lot larger than it was at the very first single spike transmission, so you now have short term potentiated this science. This is an effect that will not stick around forever, but as long as you keep the network reactivating, it will stick around. This turns out to be very useful mechanism.

B

You have embedded attractors so I'm, not sure how many of you are familiar with attractor theories of neocortex, but the idea is quite simply that when you have selective subpopulations, let's say I don't know. Well, you have all these many columns that are selected for a specific orientation that they are more likely than not more likely than the average connected with strong excitatory silences, but which is a heavy uncorrelated firing finding which I take as a sign that there might be heavy.

D

Logging rules in.

B

The brain, because these connections into people, somehow they're, not genetically pre-programmed, but so the idea is now that you have clusters or distributed representations that fire together to form a bunch of. So in this case you see four different, embedded attractors here, where you have different selective cells, which are wired together with strong connections like the bold lines, it's a balanced network, so there's also inhibitory neurons and an awesome twenty percent or something maybe fifty percent I, don't know how to balance it right.

E

B

So the inhibition is nonspecific, it's just random right in more detailed models. Like the cortex model said you know, we don't now lately, there's a lot more accounting for the basic uppers and the reach of those cells. But this is a very abstract yeah I'm, all right, excitatory integer.

F

B

It's it's a general right, so it's a network, dynamic study of sort and they have a bunch of embedded attractors and now. The key characteristics that you want to go for in working memory is the idea that there's many many long term items thousands of them embedded in the network. But you now want the ability to not just hold one active that is so reasonably easy with an attractor that just keeps going. That was the very, very early working memory work which led to this idea of perpetual activity might persistent activity. Your processes of working.

A

B

When you think of something you know that activities which is on and it stays done until you get to retrieve the task after some period and that match very nicely on to monkey recordings, but these are very trivialized tasks. I don't have much with real working memory, cousin we're working memory. You're, not you know staring at a table. You get more than one item.

F

B

Extra members have figured out how to help you through this. Now it's multiple items and distractors and in between tasks right and non-sale, in queues and whatnot, so they're, the experimental techniques have gotten more complicated and as soon as we did that the whole story about that there's always persistent activity kind of broke apart.

B

Instead, you got a lot more sort of like these intermittent activations first and then you had long stretches of silence so, like mark Stokes, a Knoxville is talking about this activity, silent working memory, where he like has to start to toss that silence the population that was supposedly active now for a long time and then there's the Saleen queue and suddenly over the working memory is alive again. Where did it go in the meantime, if it was really activity silent it?

B

Some of you one answer to that needs to be in the synapse ease if it's not in the neural activity. So what about the sign of these can change on the order of a couple seconds? Well, you did your costume, because the buffers for like five seconds or something so that's the potential book mechanism. If we're capable of refreshing the memory- and that is exactly what they did here- so they have a mint Network.

B

The simplest case is here right, so you just have this balanced network you're, seeing these cells here I think so there are a couple of cells here in black for specific population and specific green cells, I think and so there's some noisy activity where the network is kind of in balance and irregular spiking bubbling along from random background inputs, and then they provide a strong stimulus here. So they stimulate all the black neurons right and they make them fire at. You know like some higher rate, like 20-30 percent Saudi, Hertz I, don't know, maybe 65.

E

I don't know, the number here depends.

B

On time axis well, the the blacks offset is not agreements as far as I can see so and what you get then is that obviously, the facilitation of the stand-up scenes that are involved in that it shoots up and they depress so that afterwards, you you have sort of the network bubbling along again and when your background input is strong enough, then there will be a time point where the depression, like this exit available resources, has recovered close enough to one again that we sign up sees with a residual facilitation, are strong enough to kick off that population again.

B

So there's a population burst where then the originally reactivated population, without a specific cue, reactivates itself. After some time.

C

It's a spontaneous.

B

Reactivation and, of course,.

F

B

Right and then you can do that with multiple, so so so you, you start off first with this idea. If the background activation is strong enough, you get these repeated activations likes. You get a network that is self refreshing, its memory content, because after this initial stimulation, which stimulus which was external, the network will have sort of depressed the necessary synapses. They will recover strength, they would still have facilitation, so the sign-up sheets are stronger than normal, but they can't quite fire yet because of depression, and then they will reactivate.

F

So that's the second so.

B

One two three, so the second sweethearts reactivation so three times the second, you get a get a population person.

C

B

A simulation like and I will, eventually you know. One of these talks in the future. I'll show you the actual macaque monkey data, not quite because it gets the activity looks different as soon as you have a micro circuitry and about explained.

F

B

In these equations, what can I do with.

F

B

All right and obviously what you can do with sufficiently strong background. You don't get this justice reactivation first, but then you might actually get individual neurons that are part of the attractor gets a bubble along right. So you never get sort of again these these spikes of activation button state you get persistently elevated activity. The reason why they show this plot is because that matched the very early investigations of one item delayed national.

F

B

Which means that you yeah it means that in this silent period here some of these items is I will be strong enough to find again, but not the whole attractor, and so, as a result, you get sort of a temporal decoupling and instead of these discreet bursts. Now you get an elevated activity.

F

B

You can you're gonna equilibria here right between the you see that this stays horizontal. The depression and facilitation are both flat. Now, yes, right what some modulation of some soil right, it's right! Is it understanding what is not.

C

B

Depression, so there you see the rapid drop off here at the end of the population versus way that pushes the not the amount of available resources for firing right but can be. You know, used at every spike Goffstown and that puts makes the synapse is that a part of this attractor weak enough such that attractor.

C

Cannot stay active.

F

B

You mean why I stopped at the red arrow, okay, I thought you meant what happened but I guess we changed the background. Excitation.

F

B

Turn the network.

C

B

Order to have a self reactivating network with a self reactivating subpopulation, you need to have some kind of randomly correlated spiking that normally doesn't trigger an attractor unless the the member signup sees of that attractor are facilitated, in which case you will get a bubbling along of that attractor until such time that you lower the activation of that whole network threshold, where that cannot occur,.

B

F

B

The turning of opportunity out there big burst yes right, and you will see something related to this in the much more detail you.

F

B

To see the different states and then it can be in right and the cool thing is there is the state where the network can buffer some memory in two different ways in intermittent bursts or intermittently, increased fine.

C

There is, does that correlate.

B

B

F

B

F

B

I think what people mean is something like self-excitation, and that is always in these attractor networks, because they are strongly connected. So that means, when some part of the attractor is active, it will recruit the rest of it, which means that activity can go around and some signups. This might be depressed. But you know the network would keep bubbling along either like going intimately silent and recovering, or by actually bubbling along the different units which are then taking turns to fire, which gives you the steady-state solution.

F

B

Here's a really cool thing you can do now and with this network you can. Let me just talk about this one, so you now have two attractors in the networks so sells to zero to 80 upon one attractor and sells 80 to 160 I'm part of a different in the track. Now you can stimulate the first item, so loading it into working memory. If you will and what you will then see.

B

Is that because you get this, you know it's a play of facilitation compression you're, going to have the memory item automatically being retained by the network, but every time it is about to be forgotten. The repression is decayed, so so low that it can reactivate itself again, and so it will refresh its own value and keep itself alive.

B

It's a little bit like a RAM refresh right where, where the the residual strength of the network is used to update the strength of the network such that it never goes away, and then you can even have an intermittent period here of noise. So this is like anchoring your activity right where, like both items, get a little bit of activation, but it doesn't matter because the facilitation yeah right this one- he stays high enough- that when the depression has recovered again, there's another computer bus.

B

So the network keeps itself alive even over a bridge of a full second of sort of interruption, yeah right, it can still buffer the memory a bit but yeah. This isn't about.

B

Exactly so, that obviously depends on the facilitation, a constant of the presentation. Time constant is five seconds.

C

B

So here they use like 2.7 seconds right. I know that was when the other item sold it here. So then they activate another item and the cool thing is that doesn't delete the old memory instead, what is now happening is that these these attractors are taking turns to activate the reasons why they cannot be active at the same time is because there's competition right, there is an inhibitory pool which means that when mana tractor kicks off, there's elevated inhibition now in the network, in addition to the excitation going around, but that excitation is specific.

B

The inhibition is an specific meaning. When this first I can be reactivate, it's going to shut down this population and when this item then can activate is going to shut down the other population, but that inhibition does not delete the facilitation. Nor does it stop the recovery of the depression. So that means that you now have multiple items taking charm reactivating you get a multiplexing of the working memory content in time, which, interestingly enough matches conscious experience and is a neat way to explain the very tight.

B

Right, so you so you have a limited. Somehow you have a vague.

F

B

What I meant with conscious experience is more that you are consciously aware of one thing.

B

That doesn't match yes, but then also matches is the cognitive thing. Oh.

F

B

And this is not a theory of consciousness unconscious I just wanted to point out that it was very nice that it provides a natural limit on working memory capacity, which is actually quite low, because these different items are not a gun and now I'm going to compete for foreground, and they need to reactivate in a time window that is compatible with facilitation.

B

So if the facilitation time is like two seconds, then you're going to need to get an activation kind of you know at least one you know the other second or something or else your facilitation will be so low. That.

F

Between continue between those two representations, one is like what what is.

F

And so just tell me two things at once: right: it's not working like we have Nia, Slee and.

F

B

So it's a very rich mix and it could be useful for a lot of different things. One of the things that it is useful for is sort of addressing working memory and that's why I'm in the in the game of these types of models that you essentially kicked off by this paper and subsequent reviews of working memory activity that showed that it's not quite as persistent as people.

F

B

It can do a thousand different things and I actually want to propose it for some things for HTML, because it I think also provides neat solution to the problem. I think that model has with these big unions, which is just massive amounts of activity.

B

If you map that onto some neural activity space, but there's a nice way to bring that down without destroying the Union by multiplexing in time, will it make anything like this, which is very straightforward and easy to implement, but I will get to that when we get to more detail to quadrakill models sometime in the future, I think just having shown this right, this idea that you can buffer items keep in mind that you might have a network with a thousand embedded items.

B

This is just about picking out a couple of them that you're now going to keep bubbling along and active right in an elevated state of some sort which are prioritized for further computation, and so that's what this is useful. What this does not explain and that's in part. What my work then is trying to address is how these attractors can be embedded in the network dynamically using short-term plasticity, so that the problem is a little bit that everything you can think of in this kind of a working memory system is already in the network.

B

So you cannot like these attractors are already pre-loaded right and, of course, they make some argument about. Ok, these networks somehow get constructed by a long term, potentiation.

B

Yes- and that was exactly the argument that I'm making they don't talk about how that would work, because the problem is, of course, is that the facilitation is a presynaptic mechanism and as a result of that, it's non associative. So it's not specific to the connection. So if you facilitate synapse right or put differently, if you it, like all the outgoing connections of a neuron that has now spiked, are more powerful, not.

F

Just ones that connect to the neurons.

F

B

To me too, but but still when you embed attractors already in this network, then then this works, but you cannot dynamic you. You have no way of associating.

F

Associations we're not.

F

B

So right, let's say you wanted to link to items right. Let's say you want to associate this I can witness icon right to go back to the more theoretical model.

D

Here right, this.

B

Network has no mechanism for doing that. You have a very nice mechanism for picking out a couple of those attractors and keeping them online, but you cannot form new associations and you cannot embed an attractor that it's not already in there.

B

But again, this is a very nice maintenance mechanism and it's based on a very simple and very robust neuroscientific finding, namely that of facilitation profession, which are powerful criminal cell effects. Now, of course, cells that are schon. None of those signs I think I mentioned this last time so like in a big, powerful basket cells right. Have this big, like you know, sign-up sees that can, like you know, latch onto the certain cells and that big ipsp, so they suppress other neurons from firing and at very high rates.

B

They're fired, like you know: 70 80, Hertz no problem, and they don't seem to be depressing, but they're not facilitating either they just you know.

B

Exactly they have nonspecific, and so that also explains why these cells don't show a tuning selectivity when you probe them in cortex right. So you can measure what a neuron responds to by the so-called receptive field.

F

Distinction between facilitating.

F

B

If you, if it was just about maintain balance in the network, for example, and providing fast feedback inhibition as like contrast to enhance okay when I click on like it is the more sorts, then you don't need to be specific and the fact all you want to be like really powerful and reliable in terms of synaptic transmission, and you want to be fast, which is what is shown again again and all these called her models like the S&P, a paper that I studied when I look at the synaptic, you lace at the absolute shortest for the you know, column and positive basket cells to pyramidal cells.

B

F

The gym brief on probably.

B

Water, more distal ones, probably have more interesting than Alex. We.

F

B

It is, it gets confusing quickly, so it's complicated.

B

So what might be happening in the real brain right is that you get these intermittent Sereno reactivations, but the way that we typically investigate when we do these when we do these working memory tasks used to do them, is a lot of multi trial averaging and a lot of intro trial averaging so you're, taking big time, lengths of like 500, milliseconds and taking many tries, and when you do that it looks like the activity of let's say the second one goes up here and then it stays up with some fluctuation.

B

You know, but it's up and then, as soon as you end your trial. When you retrieve the item- and someone goes down so the item switches off right- if you think of like this, this example here so when you average many of these trials, these bursts are not always going to be the exact same spots and when your time bins are 500 milliseconds. So, like this width here right like the early working memory, human data Goodman wreckage.

B

Looking for the amazing pioneering work showing you know, the elevation of activity for specific working memory item matched exactly to the memory retention delay where the activity goes up and then goes down, but the problem is a lot of averaging.

B

So what you don't see is the more hidden embedded dynamics which might mean that actually, that population can be silent for long stretch of I, like what you see here, where this item doesn't activate for a second and you might actually have multiple items, and you know when you do multi item working memory, which they do. You might have multiple items active at the same time or, like you, know, multiplexing in time like what you see here and so to return to what you originally said.

B

Yes, the point of it is that when people did all this averaging, they concluded that the mechanism must look very much like this, that somehow there's a persistent elevation of activity where the neurons keep firing and the reason why there is memory is because they reassign each other. That's what they call with the operation right, and that is how the memory stays alive. The problem is that when you intermittently silence those neurons, the working memory is still not dead still there, so it must be somewhere else.

F

F

B

They they say that in this model it dies of the 2.7 or 3 seconds or something so.

C

It can breach a second because.

B

The facilitation you do see that you hear the blue curve right sloping down down down, there's a point where it reaches the baseline here and at that point it will definitely be forgotten. I. Just extrapolating looking at this picture yeah, you know, maybe here like two and a half seconds later. Also it will be a baseline again.

B

Doesn't always this in the brain when you're.

F

B

Averaging and trial, averaging sort of the two things that have obfuscated a lot of the much more intricate fast dynamics which are essentially getting we explore now that people have better tools and also been reawaken to the idea that well, these early models they're fine for explaining some things and they certainly matched what they saw an activity in macaques or by well these men, you try as an average, but the problem is that it led people down a bit of a garden path.

B

The problem is that, because people said activity goes up, it's up for the memory and then it goes down and when it goes down in between, you know that memory is calm and, like all these correlations between the firing rate, elevated and memory content that led people to this idea that oh, there is actually a reverberating activity so there. So there is some circuitry that is firing all the time and that's where the early attractor models of working memory came from.

B

But of course, the problem with these networks is if the information is really in the ongoing activity and nowhere else, then, as soon as you silence, those the information would be gone and unrecoverable, but that is not matched by biology, so mechanisms that account, for you, know different synaptic mechanisms at different time. Scales can help address that problem and actually neatly address. Some experiment was, as I will talk about, that.

C

B

I mean I've not seen this one applied to to visual education, just in part because I don't do v1 and visual processing in general, I mean so much focus on dorsal, lateral, prefrontal, cortex and the dustman can work earlier. So for me, it's really about higher-order areas and integrated representation.

F

So certainly, some cellular dynamics.

F

So it could be, things are happening, but this is not trying to keep normal things. You go in.

F

You see, there's one thing and so I wouldn't think this explains that in a very direct sense, I think there's is literature that talks about that specific problem, so Commission that probably some papers about that that talk about what they I, don't think I. Don't think this explains that it generally could. It could be similar sort of typing in terms, and maybe some of the similar.

C

Thing I've noticed rugby is we're talking about that. We're not consciously aware this multiplexing, you see it as a more or less continuous, stated well.

B

You can do it on purpose right if you're playing memory at least I, do that when I play memory, why don't somebody else is doing a turn? I'm thinking, okay, I, like the house, stereo Tigers down Wallace there right, I'm, gonna, flip open this one now so I actually do that as a conscious to.

C

How something is holding different mechanisms might have to tutor step I'm, just thinking what are reasons. This is occurring at a subconscious level that we're not perceptible we're not for.

F

Some we're not, we don't receive the fusion either.

F

You're looking at something and you put input your brain changes visually to change the second milliseconds and you're, not aware of those changes, so there's more to it than that. In fact, so you know we proposal it solution to that which is pooling. Cells are stable, while the adult is telling reason for changing.

F

The details right and so that there's a lot of things going on this one, that's hundreds of seconds and you're, not aware of it. So the problem there's not much to share answer to this when we say were not aware, are some changes. Things are slow, like things are really long right. This doesn't say that we wear of it is just say that this mechanism does exist in some form. We don't have a sensitivity to it right so I'm suggesting that the real reason we're not the conscious of these changes.

F

That you're only conscious with a vote day and you're, not conscious movement of the eyes.

F

So I don't think the general fusion issue.

F

Discussion, you might think about the sense of fusion problem or the binding problem potential shooter problem. It was a problem how it is. We are not aware that our eyes are moving and how it is that I perceive this table in my hands and that's a much slower process.

F

This is really the way our brain keeps multiple things going at the same time. It's an alternative. The Union properties I think we need to talk about.

F

C

Visually, the effect of it is higher than the gamma ray gamma ray is the previous diagram. She showed basically gauge, so it was kind of the driving mechanism.

B

Defer that question to when we get to a detail actually like explains, come up first in detail and makes a very concise document about it and I think we should probably wrap this up, but any notes of time yeah you know. Obviously, if you serve questions, we can still talk about it, but I think I also did what what I wanted to do here with this integration model. I think it's a nice step to so show what then happened after this in maybe one or two more coming talks.

B

We can take our time with that, but I hope you see the continuity. How do you get from synaptic mechanism to something that is now starting to become cognitive, which is what I'm saying excited about? So thank you very much.

F

A

Thanks for in I'm, your offline.

A

Okay, thank you. Everybody for watching I, really appreciate it. I am going to end this stream. Now that was florian. Pubic you'll see more of him.