Numenta Numenta Research Meetings, 8 Nov 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Persistent activity theory of working memory

Description

Florian Fiebig will talk about the (struggling) persistent activity theory of working memory.

A

That's video 1970, okay,.

B

I think we're alive. I, don't know.

A

B

We're ready go for it all.

A

Right cool I'm, talking about the persistent activity, theory of working memory, and it has to do a little bit with the fact that I'm thinking activity and with respect to momenta, obviously I'm thinking about the activity in the cortical learning framework and I have some things to say about that. But I want to make sure that is understood where it's coming from. So that's why I'm gonna do like a quick dive into how people think about working memory and what we can still find in the textbooks, but what is changing rapidly and why so?

A

We've had indications that prefrontal cortex is engaged in working memory for a long time from lesion studies that show when we thought it is impaired, somehow that it's particularly working memory functions that suffer. One thing that we should maybe quickly talk about is what do I mean by working memory. So just the cognitive definition of it's like I, haven't studied cognitive size per se, so not sure if I can give a proper definition.

A

But the point is you want to disambiguate short-term memory from working memory in the sense that working memory is working memory because in it is operable, but you can work with it and it has 14 characteristics in the sense that it is. You know fast learning and low capacity, but it is that the key point is that you are keeping something active, so you can work on it because short-term memory just focuses on the aspect that you can.

C

A

But then you put it into working.

C

A

Least there, the cognitive theorists have made this argument that there's a difference to the working memory system, in the sense that so, for example, all the different modalities I have like dedicated subsystems. So, for example, you have a phonetic, you like a special circuit. That is just there for being.

A

You know phonetic content right, so you can think about a phrase or sentence without actually thinking about its contents. You can just repeat its phonemes.

C

A

Know we have such a provision to like it's called the visuospatial sketchpad. You can close your eyes in the room.

D

A

In order to be cognitively capable, these things have to be workable and not just offered information.

C

Oh yeah, moving memory in the brain is false. We don't move memories in the brain. Ever there were no atomic Imperial year. You can. Basically, we invoke the same status of activity under different conditions, so move this into working memory.

C

A

Of that process well.

C

A

Signatures of the process.

C

Exact imaging is another activity, meaning that point of that do them yeah. If you basically remember something is to take the state or activity of a bunch of neurons and being able to remote, it.

A

Right I think just it's tricky to take these words literally right. Well,.

E

There's memory then stored in storage and then, when we're sensing objects and stuff, we saw in volcanic activity. That corresponds, let's say to the set of objects that are possible. Given the least we have so far and that's kind of been some dynamic state and we can, in your work, write.

C

E

It in the sense that we can form a predictions based on in the next bit.

C

The memories oh yeah, dynamic.

E

State that's associated with what we're sending data and then we can use it to form predictions and expect.

C

You, the brain, operates I.

C

E

That's what he's saying is twenties defining work if I'm understand us the connection between short-term memory and working the long term I.

C

Don't know where you're going with this I've kind of come to the clues a long time ago that the only way you can store anything changes in synapse and sometimes those can be very permanent, like you're growing these announces, but often they have to be right, temporary and you're changing something about the synapses and but it's all, basically the same type of associative learning. It's just different mechanisms of July, and so anyway, I guess. I'd react very negative when people start making these distinctions about two types of memory and them running around.

C

A

Is clear, though, that the brain areas that are correlated with certain memory demands and well.

C

I suppose in fact, likely that's early areas.

C

Where certain areas are going to be more suited for certain of the performed very short-term synaptic changes in other areas.

C

And she was describing different. Such expressions in these pyramidal cells would be frontal, cortex versus other ones, which allowed me to do this sort of more rapid learning, but it's the same for distal parameter cells. It did their cells express different genes which allow them to be more suited for this very short term. I.

A

Don't plan to plan anything else, but the original thinking what working is really quite different, because, following those early indications that somehow working memory tasks tend to be in parent, if you want to think about in short-term memory, think of it in terms of short-term memory, tasks that require the ability to hold items in memory for some limited amount of time or combine items in in mind again.

A

Don't think that, literally and of literally combining right in a cognitive sentence, combining I tend to be impaired when prefrontal cortex is damaged, so people started, recording from prefrontal cortex and so some of the earliest people who did that work. People like your Kim Foster, GABA, Rakesh una Hoshi. You who've like written all these seminal papers about it and the early work kind of worked a little bit like this. So you had some behavioral tasks.

A

Some monkey was looking through like some glanced.

A

Oracles so like there's two sort of doors way in which your hand through and then there's two buckets and you lock these doors. Normally, you lift the screen, put a piece of Apple in one of the two buckets right on the for the left hand or the right hand. So like one or two doors, and then you close it again and you wait some delay periods and then you allow the monkey to reach through the doors. And obviously the point of this is that the monkey needs to see what happens.

A

I, see where the fruit got put, but the two doors lock each other. So you can only stick your hand through one of those. We need to make a choice, what you're gonna go for, and so you need to buffer that information right about what to do, but it for some amount of time.

A

And then you get to act on the task and then the record from prefrontal cortex and what they showed that there are these neurons that respond very strongly to this task, prefrontal cortex, which so the most typical ones right, the ones that people then I read about we're sort of these neurons for Hertz here, which then jumps up to some higher firing rate until such time as the item is to be retrieved, at which case it returns to lower height, there were different profiles seen so neurons with a lower firing activity that actually didn't respond to the cue, but was then active over the delay period, the newest that ramped up over the cue period, neurons that are only active during the cue period.

A

So you see all these interesting state changes both during the pew and at the signal when the monkey is allowed to act on that information. So the time scale for these things, you can see it down here right. This is five seconds, so every one of these black bars light is like a spike counting off different. So this is a single electrode is for the early days right when we didn't have big electrode arrays, and it's just counting the spikes over a time ban off, I, think 500, milliseconds and.

E

They've been probably didn't, say, one layer, no.

A

All the way too early for that right, they were just hoping to find reliable activity, and what they may quickly showed was that in fact, there's a strong correlation that you can prove between the ability to do the task to remember what to do correctly or to at least you know, obtain a food reward and the hiring in the delay period. So where you could show that there's these units that you can find where, when their activity.

C

A

The you can predict that the monkey is going to make a mistake and that holds up under you know some statistical analysis, and so the idea was okay.

A

The information is somehow in this activity, because we see this activity in the buffering period and when that activity is disrupted because it goes way back, you can be distracted or something. Then the monkey fails. So the information is in the activity that was sort of like the early earliest interpretation I.

C

Don't know I'm gonna to a thousand different things.

C

I make coffee and I put the back in the fridge.

C

I'm, not thinking about are tending to it, but I know where it is and and one could partially you could explain this as an attentional vacuums, not a memory mechanism right. The monkeys system.

C

We just counted by the idea that there's many new activity goes down and then the monkey thing up form as well, but that is country evidence base.

C

I, don't think people at, for you know in an hour and a half to nine make coffee the first cop in the next cop I. Don't think I think your alternate interpretations of this and.

A

So so this idea caught caught a bit of attention, of course, because it was a clear correlate of the of the memory or at least people followed it that way and a lot of theories to flock to this idea, because they knew about all these interesting and back to network. You know the earliest ones, the the whole field networks or you just have a couple of nodes right and then you can connect some of them and you can get reverberating activity. So the.

D

Idea being you.

A

Have like two different attractors embedded in network, and now you know one code. One is coding for the left box and one is called me for the right box and the idea is that when the Jew comes to switch on this attractor, then these neurons will activate each other and the activity will keep going around some groove aberrant, re-entrant activity. So people did this with excitatory nodes or some sort right.

A

Those are early days, it's not terribly but physically, and that came a little later when the people did spiking modems of this, and then they argued that well. The item is sort of kept in storage by the reverberating activity here and they are competing with another item pool. So there might even be you know later when people did.

A

Say get these balanced tia networks with embedded attractors, which you can trigger to so those nodes that information right to put that to put those representations in an active state, and then then you have a very clear case that well, when you disrupt that activity, the items gone, that's very nice and when, as long as that activity keeps going, you are buffering sort of that specific. Are you putting a lot of all the things that are encoded in this system?

A

With you know, embedded attractors, you are now preferring right, this particular state, but you are selecting it and so people build theoretical models around that. So let me show you an example.

A

So, for example, there's like meet paper from 2003 that was before the paradigm shift, and so what they did is they have this nice spiking neural networks right and then you have a couple of sort of attractors if you're loading into this, and so what you see is that the activity you shoots up from, like some baseline to some crazy high number, thirty Hertz or something.

A

During the queue and then when the queue subsides, the activity stays aloft for a while and in fact they even manage to put in multiple items. Here you can see that they can put in one two three four items before at the fifth one. The first.

C

Item starts dropping out yet.

A

This is an attractor, multiple item working and me.

C

A

It's an attractive perspective on working memory and it's inspired by this idea of persistent activity right, so arguing that it is the activity itself that offers the item. And now, if you want, multiple items were you're going to need to have multiple attackers that act together.

A

That creates some problem, there's in fact, some people even faster when the when the controversy get started, got like stuck on this idea, and so you see papers like this one rageous lee so blatant in its violations of biology where they made sort of a nice multi-item short-term memory mechanisms, and they wanted to do it with these attractor networks and they they do a lot of nice things. You might remember this why it's for this, like facilitation in this network.

A

These are like spike in your networks, and you have like these dedicated pools for different items, which that can then be loaded in and their share in inhibitory pool to balance the network, and you see these right. These are spiking neurons, so they have an expression for membrane voltage and that's different transmitters in here.

A

And whatnot, so it's quite a biophysical model, but then what they do with it are things that are not really seen in nature right, where the activity of these neurons jump from like zero Hertz to like a hundred Hertz. When there's the cue present and then it settles in some 60s and averts regime, where the items are then maintained.

A

That is not really what we saw in that data. Obviously, we've done since a lot.

C

Of what was wrong well,.

A

So, okay, right and just that major.

C

A

Rates are way so there's a couple of problems with sort of this this perspective, so I should start with the things that are most easy to see right so in the first attacks right 1973. The biggest problem with this with this paradigm is that you are confounding the memory with motor response, meaning when the Marquee sees where the Apple is put right.

A

You no longer needs to maintain a memory for what he has seen and memory in the in the sense that we think of memory in the abstract sense, but instead what he needs to store is the motor commands on which arm to move, and all he needs to do now is withhold that movement.

A

So you have a preparatory motor response that.

C

Would be the system with my argument for attention you're attending to the.

A

People other tasks, because it turns out that when you do these big motor tasks, we're like the monkeys have to move on. So what you get you know artifacts, and so they wanted to have better control. A lot is happening here, so they trained monkeys to they're an adult. This is the oculomotor of delayed response tasks, which became very famous because it creates data that is very nicely. Persistent. I'll talk a little bit.

A

Why so you make the monkey is dead, stare at dot, and you show some some symbol: I, don't know, let's say triangle in a certain position and then that stimulus goes away and the monkey is not allowed to move the eye from the fixation dot right. But if he, you know, keeps focused there, then after some ten seconds or something she gets like a signal or some sorts light.

C

E

On or like the the signal.

A

Here, changes across or something and then he's allowed to move his eyes in a saccade to whatever location the the symbol was, and then you can do this at various locations. Of course right and you can show that the monkey can't execute this task. So you get rid of the whole arm movement thing, but you still have the problem that the moment when the symbol appears the monkey can forget all about that symbol. All he needs to remember now is that I'm gonna need to look right as soon as I'm allowed to, but.

E

I'm not gonna do that yet because could you do an experiment.

A

Right and so people started getting concerned about this problem right.

C

So this is the problem of the.

E

You know you, might you.

B

E

A

E

Afterwards, that tells you what you can add store the actual right. Much more so.

A

People have come up with slightly better.

C

You still need to.

A

Remember the location- and you know yeah hoping to know that like too much discussion with us, so we can actually get through this.

E

A

First things that put it here right, I was concerned about what what is what are these neurons that are switching on their firing actually doing? What are they actually encoding? Other encoding, this base all encoding, the the visual memory other encoding, the motor command to be executed like what is it actually, that is being buffered there is that attention right is that inhibition of the motor command to be executed like trying to you know not move, and so you can.

A

You can modify this Oakland oculomotor delayed response task right where you have to move to some cue after some delay period by instead saying well, you don't you're not moving to where the cue appears, but you have to move to.

A

You know ninety degrees perpendicular to, and then you can do an interesting analysis in what these neurons that are responding are actually selecting for, and it turns out, they are in fact, actually not so much selecting for the memory they are selecting for the place where given a move to so in these in these spike raster plots or like these different trials right.

A

What you see is that you get high spike rates for a preferred direction, like you see all the different possible directions where you put to a saccade to I turned I movement and in fact these neurons have a preferred direction, and so that is disconcerting right, because we are apparently confounding confounding. These ideas.

A

It's essentially the same thing that you would do with the arm: you're not allowed to move the arm until you get the go signal, like the doors are not know, you're allowed to reach, except it's a much more control paradigm, because they're, you know they're head fakes and they calm and stable and they're not doing much else, and you know if they can't do anything else with right. So you today reductive paradigm notice, one of the things that people are obviously concerned, so the the biggest problem.

A

None was like this obviously a risk for electrode selection bias when you only have one electrode and that you stick it in that, there's some risk that you are biased towards neurons that are nicely be portable and have strong firing rates. So that is obviously something the influence of the motor planning. How do you disentangle the two? The best way is to actually do a delayed match to Santa tassel or, like some kind of company toilet review.

A

There was some concern about the averaging within and across trials and I'm, going to get to that now, um which is this this problem of when when they did when they found these notes, that was so nicely responsive. You did many trials and you also had these large bins and the question is: are these neurons actually that persistently active as well? These theoretical models that had like this without reading activity, often at high rates right, were suggesting I had people from physics were very you.

D

Know, eager and studying all the estate transitions.

A

With different factors and was like a cool.

D

Thing to play with.

A

But the problem is like, as soon as people started, looking a bit deeper under other hoods, so this is probably the paper that started like the big big revolt on on the theory of working memory. It was this review by who took a bunch of different databases from all kinds of different working memory. Studies see all these authors, I mentioned before right, Forster from the Hashi goldman rakish, and they took a couple different databases and we analyzed all of those without all the averaging, without selecting particularly nice electrodes that you can show on a plot.

A

But instead they looked at all the recordings and what they found is that a so they try to classify these neurons into different types. My neurons that are delay activated. So they have. You know an activity. Delay notice that stabili delay activated so weather delay go the activity goes up and stays for a while delay, inhibited.

B

A

Ramping neuron said I'm non-responsive, and they tried to classify all of them by some metric and, of course, the first problem was that it turns out that these neurons, that are somehow selective, are already in a minority. The most prominent ones are the ones that are really not selective for the memory itself out of different databases. So looking this is both looking in varietal cortex, as well as a prefrontal cortex, and this is already using more advanced tasks. This delayed match to sample tasks, so they already get rid of the.

D

A

Problem, but it turns out that these neurons that people wanted to talk about a stable delay activated no, it's actually really rare. It might be just like some two neurons out of 960 recorded and, and the other thing was that these Delta frequencies right so the jump up and activity is actually quite low. So here you see like some baseline frequency and then the Delta serve water. If they are delay activated, how much higher is the activity during the delay?

A

Well, you can find a couple of neurons that have a very nice jump, but you can see. This is a if you look at the whole data. That's it's a gobbled mess right, so sort of strengthening, like this fear that there might be you know, selection bias and they they went further than that. So.

A

So they went, they really went deep down, I'm gonna like jump to there and actually well. Maybe this is worthwhile right so because it turns out when you look at the many trial averages of a neuron that you found to be nicely stable, it delay activated. It looks very nice if you look at the black line here right. This mean I'm assume. You know a little bit.

A

That's good right. This mean the black line is very nicely low at baseline and then during the delay activity it jumps off by a couple, but when you plug in the individual trials, you know something.

A

C

A

So obviously, I'm curious to see where there's oscillations behind, because the model that I'm using do some oscillations of different type actually, but the main point is Cirrus ever fall, so one is, there seems to be a problem with multi trial, averaging right that you only get this nice delay jump if you average over multiple try, a three don't really explain the individual time.

A

The second problem is that there are some periods during this practically always there's always some period in the delay activated part where you're supposed to fire fire firing, wait where the activity actually drops below the baseline, meaning. If that is your mechanism for in reverberation, to hold that activity, that's where your activity, you have your items drop out, but it doesn't write like these monkeys are still capable of doing this task and then the third problem is you see that there is very sharp changes where from one time bin to the next right note.

A

These are 500 millisecond time bills, which are the binning that people use their government records and foster like these early investigators.

A

The the problem is they're changing like rapidly right from beam to beam. They're super calm, suggesting that the temporal dynamics yeah. What is behind that actually not faster, then then. These time builds allow for. Maybe to accurately see this you need like hundred millisecond time, Vince or even sharper.

A

So if you really want to understand what is happening, you need you need to get ray of the multi trial averaging. You need to get rid of the intro trial, averaging these big bins right and maybe then you would have a chance of like you know, actually looking under the hood. What is the actual mechanism because clearly the memories? You know like this argument that the theorists were they're making on that broad argument from the average data that it's reverberating activity is quite shaken. You look at these data right. What.

E

Do you think is the right time bid and.

A

Since I think there's a role for gamma oscillation, the right time, bins for me would be something like because then you have, they can resolve 100 Hertz oscillations.

A

No, but you would still get the slower component to see by component, so it still has about nothing. So let me just jump to the confusion of this review document.

A

A couple of these things I'm going to pick out from here, so one is that experimentally observed data do not uniformly conform to the characteristics that we ascribe to it right. This idea that there's some no spontaneous firing rate and some some large excitatory frequency change from baseline to the memory period right. That is what these theoretical models were saying in some crazy cases, arguing about I just found out rages, but you know there were more.

A

There were models that were capable of doing this in a more modest, sensible way, and then this idea that there's a stable, highly structured network like a sub Network which codes for you know these particular is particular memorandum. And then you would have some by stable switching dynamics. So I mean a lot of analysis on that study from a physical, but the data doesn't really show that right. First in the that have all of these properties, but because you need random selectivity right rather than selectivity for a place right for the motor command right.

A

If you actually want to be memory specific, so you need some way to disentangle with Petric use or delayed master sample tasks. You want low baseline frequency. You want not change and find a frequency from baseline to delay. You want reliable, intertribal, behavior right and you want a stable structure, delay period firing and then it turns out that hold true, for you know like some two out of nine hundred cells or something in the in these databases.

A

So that is not automatic, then the mean average firing increased from baseline to delay is like six seven Hertz and the parietal cortex, and about 1.8 Hertz in in prefrontal cortex database, which had a delayed neshty sample task. Sdr is.

A

Some spatial delay response, I, think, right and, of course, substantial variability right. So less than three percent of the responses in the prefrontal database were actually sort of performing too much of this.

A

So the representation of a particular stimulus with the single, stable Network state may be both unfeasible and physiologically undesirable right, so that reflecting the fact that these cells tend to remap that the code is actually you stable. There are some cells that might be stable. One of the trials.

A

But the question is really: what seems to be happening is that, even if there were behind this and some vibrating activity, it seems to me like that these things are remapping all the time and then the question is not as that work. So more modern analysis, essentially by me, some nice type of memory that actually in the coding space.

A

If you recall from hundreds of neurons there, you find both a dynamic component and the stable component, so you can do some analysis and show that there's some neurons that actually acquire of a couple of trials, a very strong selectivity for very specific Miranda. So those are the nice ones that we like, but on top of that, there's a dynamic response which which spans the whole subspace, which means that the activity is not really stable in some fixed point, which is what these you know, early attractor models would suggest or.

C

In today's talk, are you just talking about I'm.

A

Talking mostly about the problems of this, yes, maybe we can get to solutions so so this is sort of going towards like the motor planning problem, they're averaging within and across trials. Since the Big Shot II review, the electro selection bias arrays.

C

A

Being just recording everything which is nowadays better because now it's been just putting this whole databases out there, so people can do all this post analysis.

A

There's a theoretical argument from energy efficiency to be made when your model relies on reverberant activity of very high-flying ways like we and some tips right that keep these going at a very high rate. Then your burning ship tons of energy. Why do that? If you can buffer the information? Also in some small little protein change- or you know like a little change at a synapse phosphorylate, some receptors, and so that's just more sort of like a theoretical argument, then there's a problem of not robust maintenance.

A

The problem with these models is that how do we explain that there are, even even while it's true that there's a correlation between the delay activation and the behavioral performance so and the delay activation drops that tends to be correlated with failures in the working memory task that has been shown across all kinds of different things, but that only holds two kind of in the average. So people have looked at this now even in humans. Marx dogs have been doing bad work in Oxford, showing that the activity signatures of these different memory items very nice.

A

You can decode them, they stick out, but when you, when you distract people making this attention intervention, if you give them an intermittent task, if you mix multiple items and the sort of two tasks that are actually more like.

C

A

For the long-term memory, people were bothered by the two, so they argued that the stuff that sticks around for a long time- you're- probably gonna capital representation or something, but they would work items.

C

A

Anyway, so the problem with this is that these networks are not really good at robustly maintaining items in the sense that, while they have knives on during the practice which are strong without.

B

A

But the problem is: how do you explain that activity can disappear for some time and then finally, there's a problem with multi-item working memory? The problem is, if there's one dominant attractor right, then you can encode one item and it's a little bit similar to the things. The other problems with motor planning at the ethic of selection bias is the simpler tasks where you have to encode only one item and you can prepare for your action, and you know when you pick the right electrode you get these very nice.

A

You know strong activations and delivery that goes back to baseline as soon as you. You know, if executed on the task- and you no longer need to hold on to that information, but these things all attenuate when you drop any of these requirements. When you take away the motor pre-planning, you know you that a delay activity is no longer as strong when you analyzed all of the electrodes right. Suddenly, the data is not a strong. When you do multiple items suddenly the delay activity tends to much attenuated. That has to be a lot of pauses.

A

You know silence in between. So all of these things that move things closer to sort of actual behavioral demands on working memory rather than a very controlled experimental paradigm. They all diminish the power of this whole persistent activity argument.

A

So a couple few more things to show.

A

So we talked about this big important review paper, which really turned turned a lot of heads that mean people have been trying to build synthetic models to explain all these different shapes of all the neurons that are not just stable delay activated so neurons. That would show me see if I can find these actually match them here. Yeah actually did this, so they tried to take all these different categories that sharpie and identified in the terms of well.

A

These are the very nice ones, the ones that jump from baseline to some high delay activity, but then there's also ones that actually delay inhibited and they're ones that are delay, ramping ones that are, you know like decaying throughout, and can you build a network that actually has classes of all these different kinds of neurons and with a lot of tuning and a lot of trickery, you can get there, but that doesn't really teach you much in terms of you know how the you know line might work that way.

A

So people have been doing reviews on this, making similar arguments to those that I made from my dissertation. These problems that I identified actually all exactly from my dissertation and are writing review papers about this now right? So, yes, maybe not! That is the tumid way of saying it's. I have.

E

Culture I have.

A

Some simply thing with that: it's the the problem is, of course, is that all the people who've proposed this and all the people that have build a career on these nice attractor models and studying the dynamics to tell them that they have been not necessarily mislead, but that some early selection biases and some early over interpretation and a bit of unlucky sort of early conditions of how we got started on this.

A

Let us bit down a bit of a gong path right, and so they are trying not to offend everybody when they write a thing back when they write, maybe not.

A

Yeah you know, I can and so I have a lot harsher things to say about any of this.

E

A

Marcus and I met at cognitive news and societies yearly meeting right, downtown San, Francisco milah was presenting some of the data that the day, if we call it with mafia to race, with a multi-item delayed national sampled, not so getting rid of the motor planning getting rid of the electro selection bias analyzing individual trials, getting rid of the averaging with very fine temporal resolution, millisecond resolution across all four equals getting rid of that problem and doing a multi-item memory task getting rid of that problem and so trying to address all of these concerns at the same time.

A

So it's you know a lot more ambitious and then what they see is that the most informative electrodes, the ones that you know are best at predicting past performance and that, where with the theoretical decoder, can read out the most information and can predict what memory item was what many items were actually shown to the monkey? Those tend to be one that are not necessarily, even you know, persistent in the sense they tend to become a person.

A

So when they look in the frequency bands, they find that the electrodes that have like these gamma bursts tend to be the most informative ones, and these gamma bursts. Are they happen? Sort of like it's past intervals? You see them like. You know like a second apart this one here at 900, milliseconds and then 1,600.

A

You know like half a second later this another brief burst and, of course that might explain these sharp transitions that you saw in the multi in the big bins right, if, in fact, behind the persistent activity behind this graph right with a Q and then the elevated activity and the low the low step between those like, let's say, four Hertz. If, in fact, what is happening is not that these nodes are jumping for Hertz donar or not, they are jumping to come up.

A

We can see you know like some 16 months or something but they're doing so for baby from out of time. So you have these repeated brief, burst events instead buffering the information, then you will see in the average a persistent increase, but you have lots of period of silence in between information.

A

Is not in the reverberating activity, but it's probably in some synaptic change, and so that's why people are now digging out all these synaptic working memory models again and trying to build models that can can't do this I'm very proud that I am one of the very few people in the field who can explain the exact shape of these gamma bursts in terms of the number of spikes in terms of the different frequency components of it in terms of the number of bursts per second in the delay period.

A

So my model matches a lot of the detailed electrophysiological like the characteristics of this kind of a memory maintenance mechanism, but people are advocating for a new type of synaptic mechanisms now they're revisiting, for example, the monkey lo pepper, which I showed last time which I've got each population burst. No, you know really all that we listed because actually there's multiple spikes and then on one, but it's at least getting to a signature like this, where the most information is actually in the units that are intermittently active at very high rates.

A

That doesn't obviously explain what all the other neurons are doing, but it is interesting that the most informative neurons, the ones that contained the most information about the actual memory item rather than a motor command attention about. You know the number of items right are actually the ones that behave like this, and so so. That is why I'm interested in these new models, which I guess I'm, going to talk about a little bit.

A

Mikkel Lucas presented a model that can explain some of this, but it's unfortunately not capable of sort of learning these structures that would give rise to that so they're likely encoded in his model as I have a way of generating them using a fast associative learning rule, which is why I have an interest in fast associative learning rules, and it's interesting that you know dendritic computation might play a role there so that it's broadening my own.

D

A

A little bit no mental, strong focus on dendritic computation and certainly helping me with that, and so that's pretty much what I wanted to present today when I give this. You know long arch from from the 70s these earliest trials, where we try to understand working on the hotter weather, hits run, build early models all the way working through all the big different problems that this has trying to address. The one I want to shift the paradigm, many on the old textbooks bill.

A

If you meet the new science textbooks one like three four years ago that still talked about this is an activity clock. Did you look at the working memory? Section that's going to go away. I am convinced. This is an actual paradigm, and this, maybe not is a very, is a way of being very careful.

A

Oh I know it means so you mean if there were multiple gamma was happening, what relationships they have to each other? That.

C

A

A really good question- I I'm, not sure I mean you know like Miller lab. They would definitely have data on that. It would be interesting with respect to my bond model. His mind model predicts that if you're gonna have these neurons that are like wired together to lose, these long bursts that they're actually gonna be synchronous inside the mini column.

A

So the gamma is coherent across the neurons of the mini column that share the same selectivity, but they're gonna be slightly out of phase with the select this cycle of activation here is going to be the same, but the individual gamma cycles here might be shifted a little bit.

D

A

That's at least the argument that there's some underlying oscillation here and when these do occur. They occur on top of theta. So the original Loompas model had this observation that either you're gonna get sort of lesson. Low frequency B towers here, which are not information carrying sides of those in macaque or.

A

There's actual information.

D

A

So, for example, you can show that the number of these gamma bursts increase with load. So the more items you put in the more of these bursts again and obviously there's a limit to how many bursts you can have so the slacking of 3 per seconds or something which roughly matches my model in.

C

Simulations or you.

A

Work in both the simulation, as well as.

C

We have, when you say blow it in as many think what is it needle loaded so.

A

So these monkeys are shown several symbols right and they have to maintain them.

D

A

I, don't really use that much these delays and mostly counters from the connection delays in the network, is a substantial particular when you're thinking about the connections that, because what the working memory need really needs to do is hold on to sensory. You know representation. So if you have some lasting representation of visual items, inferior temporal cortex, then you're going to need to sort of control that item to stay active right.

A

We're indexing that memory of sorts and that top-down control it's very easily tractable with when you have electrodes and bridle areas, temporal areas- and you can see those delays- you can see prefrontal cortex, leading or lagging the areas that are responsible for the modality of the memory stimulus.

A

So there's actual, you know top down bottom up dynamics which are interesting to study and they impose a lot of phase shifts and and delays, and that.

E

The working memory most of the literature- that's focused on prefrontal, cortex and parietal areas sign up like that, but if we think about the court of law as having all of the different, you know capabilities that might be. You know useful anywhere.

E

You know I guess the question is: are there correlates of working memory that would that people have shown in primary sensory areas or in growing.

A

Up in immediate areas.

E

Writings like that, or is this something that's really very special to you- know prefrontal cortex or something.

A

Right so typically people make the argument that you can read out the identity of working memory items from early sensory cortex, but you can do so because of it back protections, not because the items are.

E

Obtained, that's not the level of question I'm asking okay, it's not like in these tasks. Do they find things just are there other tasks? If you think about a generic definition of working that you wouldn't expect these tasks, that's artificial working learning be one, but there might be another tasks that are more appropriate than both working that we 81. You see what I mean yeah.

A

It's a different type.

E

Of question, so you know it's like it's kind of this idea of having.

E

It's more like this definition that you have a dynamic. That's.

E

A very generic definition that could very well happen anyone, but you would need different experiments to show that it's.

A

Just generally true that the areas where the, where you find the units that are most multimodal and task specific and have the strongest decoder and within these experiments.

E

It is experiment all.

A

Right, okay, I've been doing a couple, different kinds of sdrs and odrs different motor response task, and it tends to be true that in the higher order areas like this, you find a lot of units that are very good for decoding. You know the tasks, the rules they're, the items that are currently being maintained.

E

Yeah I think of a couple of experiments that are very short. If you watch, you know they signatures. So you know very stuff.

C

You know there's some things. I didn't like about these experiments in the very beginning and analyzing stop my theme of fundamentally flawed from the story. It is human normal. The data window I mean, for example, they talked about. This is what they do. This is their job is to do working memory. Is it's like if something complex, big part of your brain is following the world, and this is a tiny part of it, which is consistent with this idea that we are really depending here, momentary attentional mechanisms.

C

Not one is when you look at all these other cells. They.

C

Know that's one of your bothers me just 20 minute. Talking about this is memory I, just don't think of it as memory memories, always in the synapses. These are working states of the brain, so it's much more likely attentional state. What are you really coming to memories? I find that all kinds of ISM memory of right and I argued that me for what I think is the big big thing that occurred in my mind in the last few years. Nothing some Tyler, not that funny that the working states of the system active neurons are cycles.

C

We have always known that these different cycles that are going on, but the idea that there's multiplex 30 cycles and it's a we and we first came out to work I mean in the hippocampus, but that's like a huge idea right, there's multiple things happening in time and they're, not just multiple items. I'm a prize, the hippocampal progression of things that are going on there try presenting, doesn't need to accrue from time through space and whack, and so this is there's a there's, a big overlapping new idea.

C

That's come about I, don't know how many years it's been that that, yes, you got this memory stored in synapses, whether its permanent stem exchanges, genetic is countable temporary, changing you and then you have an activity state that the activity stays themselves have elapsed. One accomplished life that we just and that's like a big big idea that we.

A

Need to look under the hood, you can't just.

C

A

Go with spikes, you need to know, what's actually happening.

D

A

Voltages and for me not just some way.

C

Well, the kind of thing you think about is known for their asset, that's which state about your thought, but in reality there was a sequence of these running or happening baby rap and.

C

It's just like there isn't a single state of the bringing the activity of the brain or spike in the two states and the rate rapid form, and that can use like the big discovery. You know that's like for Beijing that.

A

One of the cool things that has happened recently that I get very excited about work by Mark Stoops and both strongly emphasizes this whole argument. That activity is, you know like it's like just a emergent phenomenon of their of the connectivity really happens, and these interesting working memory tasks where you show that when you have these periods, like you show some item which can't be decoded by some means, and then you show that that decoder goes back to baseline, so that and kind of disappears.

A

But then, when you show in on specific stimulus, so they literally just so like like a strong grating pattern or like what Nathan does the TMS pulse of the brain, like literally just language, is blasted with something when a tally. On specific? What you see is that almost like impulse response in engineering where you test a system by putting on the input.

A

But the cool thing is, it doesn't work because after these unspecific pulses after just like in strong, you know, contrast stimuli which I don't have orientation selectivity year for this campus and suddenly the items that you showed earlier, the working memory items they pull out, they show up again, and so so. That obviously is an argument that the information is still there and actually you can make it visible by having lots of activity and.

C

Stable states of this circuit in the brain and sure give a pulse again, remember one being when they stick a pole, there was a neurosurgeon music or something okay. That's the same thing right.

C

A

Yes, it's just making the broad argument that look. The information is indeed not in the activity actually.

C

A

Know for working memory, people you have to.

C

C

A

D

On you should bye-bye.

C

D

C

Question I think your body I think are worth walking an assumption. Every quarter column should be doing it and I think the point that.

C

A

So I only wanted to point out ever wants to follow this at my dissertation has extended part on this talking about the persistent activity theory in the controversy around it ways to think about synaptic working memory.

E

A

Think Matt already linked out it's a couple of times, but they can also just search for the for the title of course, and let me just scroll up that gets obviously.

A

That's my doctoral thesis.

C

You know there's this, you remember er, maybe you can put a about it. They give me mug eases task. If this is like this is what monies do you know like reality? Shuts off force constrain your own caste and we asked him that I knew it's. My tiny.

A

Home match, you know they were.

C

Totally on that time,.

C

C

C

C

Where they are but but.

C

C

A

C

There's no way but I would.

C

A

These cuffs don't mean anything to them.

C

I'm freezing away I'm crazy that either unnatural in the sense that they have no reference, training that we would typically associate.

C

I always knew there was a distinction. Biking ever put my finger on, but now.

B

I'm collecting.

C

One of the thing is it's kind of a reference and how that reference, a like nothing. Did you really see if you remember where I put it there, you said: okay, the milk was gonna, have some number between 1 and 50.

A

If I can take it a little bit differently right, M is a common problem in neuroscience. This idea that we are working with you say it like naive.

E

A

The the the problem is that we are either over training these animals on a specific task, so they so they learn to execute the task, but don't have sort of automating everything, or we insist on this idea that they must be naive animals and they don't know anything about and they're completely out of context and it's an artificial environment that is not natural to them, where they have no lead. No reference frame right, and so that is a common problem in your scientific investigation would like to hold that balance and typically, don't think about it.

A

They just do whatever is experimentally, convenient.

C

Okay, because you know we aren't limited by what we do mentally I, don't have it objection to these experiments. I. Think my objection is the interpretation you know they say you know that we from Pacoima these papers is responsible.

C

For all the neurons, do this crazy.

C

I think the experiments very much in these moments, and so we should all these papers may sound like they know what that was going on. Is they say all of it? How less you make it? We know all these things you want. So what can we extract? It is like you said, and striking out those fight training. That's the same thing in you. Okay,.

D

C

Not that they're we're trying to make all the spike trains match the stands, but when I do this dance I have a real respite trying okay night, you know that would be the right interpretation of that and I'm. Just touching on four points, because yeah opening.

E

Really interested to see how these kind of mechanisms in.

B

D

E

Like so one of the.

A

Problems that I have with you.

E

A

Unions, it's just like these massive activity that that would correspond to, if you take it literal as actual neural activity right, and is there a way to get around that, for example by you, know, multiplexing my multiplexing hypothesis.

C

Into the voting mechanism, fifteen college- well, fine! So let's say: okay, not requiring two unions. Everybody can so resolved at the.

D

Same time well,.

C

We shall all that what if we, both with what can do time with them, I don't waste your time can I still get it to work, because how do I know that you know if you're doing ABC I might be doing BCA and so actually lined up? Follow? Yes,.

A

C

How good is there a way to get around that problem? I think if you could somehow all these guys can multiplex through time and have multiple representations going sequentially if I can still get the voting to work.

E

C

E

It's possible because if you allow longer time than than five more seconds.

A

You know seconds.

E

Or something like that, then you can imagine the way we one way we thought about been voting. Is these active dendrites that you know fine trigger these patterns laterally and then, when you actually get input the depolarization.

E

E

C

D

C

Keep the Union idea mechanism underlying for kms info on, but I don't yeah.

A

Because you won't, you miss off, they have been apostles of space it because we know that we're capable of that- and we know that that was quick. So.

E

It's really interesting in the short-term plasticity mechanisms.

E

A

I guess it's gonna be a couple conversations but with Jeff, but the whole reason why I want to put all of this out there and talk about these things is because I watch your votes, not a few. You know possible into this conversation. I know that we all get busy with other things, but you know you care very much for the spam work and if we can advance a little bit or like branch off interesting idea that.

C

Would be brain is very it'd, be a very important thing if we could explain the the physics of the Union properties in its and its operations in a in a framework of framework of oscillations and I also want to wrap in these. This idea, I have about oscillations relating to scaling and that's important too so, there's a whole I kind of hinted it well, because a whole world of things opened up when I realized that you got this. These oscillations in there and multiplexing through them, and so I wrote a theory about that.

C

It seems being a really important next step, I'm still talking about unions as if they were as if they were just working. The way I talked about them, but under the covers it works. Ok,.

A

Awesome I'm very excited for the idea and I'm glad you.

D

A

Have this thing.

B

Everybody thank you for for watching today, I appreciate you, I gotta leave this meeting thanks for joining. We do these research meetings pretty much every week. I think we're trying just to do two a week right now. These are our our real research meetings, so that may not make sense. Sometimes we just kind of have to start a lot of a livestream I'm so glad that there are 20 people online watching appreciate that. Please give the video a like.

B

If you don't mind, it helps I want as many people as possible to watch these and just the fact that there's 20 is outstanding, because this is like pretty detailed, neuroscience stuff. You know: you've got PhD of neuroscience talking to you about very intricate details about heavy learning when types of memory so interesting content right. So thanks for watching come back and watch us later, I will have another stream, I, think Monday and so I will see you then, and there will be a podcast coming out.

B

If you like, these series of talks with Florian I have a podcast coming out with him. We talked for 40 minutes or so I. Think, and should be up at the end of the month. I just have to do all the editing and everything so you'll see that soon eric falco charles I was talking to someone johnson doug rose nick k. Mark brown thanks all you guys for jumping in on this Matthew and take care. I have a wonderful Friday and a wonderful weekend. I am signing out talk to you later.