Skip to main content
Loading Explore guides…
Neural Dynamics of Cognition · Acedra
Home
Explore
Psychology
Neural Dynamics of Cognition
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
171 segments
Copy
Hide video
0:06
Good afternoon, everyone. Welcome to our webinar. My name is Earl Miller. I'm a professor in the Picower Institute for Learning and Memory and Department of Brain and Cognitive Sciences at MIT. And first, let me start by saying that I hope you are all safe and well. So in case you don't know me and what I do, our laboratory studies the neural dynamics of cognition. We are interested in cognitive functions like working memory, attention, decision-making, and so on.
0:40
To understand the neural mechanisms underlying cognition, we perform multiple-electrode recordings in non-human primates, rhesus macaques, performing cognitive-demanding tasks. And you'll see examples of these tasks throughout my talk. We use electrode arrays that allow us to look at local circuitry within brain areas, micro-circuitry if you will. And also, macro-circuitry, the interactions between brain areas that are distant or in separate parts of the brain.
1:15
And today I'll be talking about two types of measurements we can make from these micro-electrodes. And that is spiking activity, pulses of individual neurons, as well as local field potentials that are the average activity of lots of neurons underneath each recording electrode that gives insight into what populations of neurons are doing. And those LFPs are particularly useful for looking at network properties. Today's topic will be working memory.
1:50
Working memory is the temporary of holding information online to make it available for further processing. When you think about things, when you hold things in mind, when you make plans, when you move thoughts, when you bang thoughts around your skull, you're using working memory. But importantly, one of the take home messages from this talk as that working memory is not just short-term memory. Your brain has many examples-- several examples-- of short-term memory buffers.
2:21
What makes working memory special is volition. We have control over working memory. We decide what to think about. We decide when to act. We decide if we need further information. We decide when to make a decision. Volition is what makes working memory special, because it gives us executive control over thought and action. Rather than just be creatures that can just respond to the environment, we can interact with the environment in our own way. So this talk will be a brief survey of some older
2:54
results in my laboratory. And then a more in-depth look at new data and a new-- maybe not a new model of working memory, but an update to the standard model of working memory.
3:06
So when we talk about executive control and volition, executive control requires what is called top-down information. Top-down information-- bottom-up information is the sensory information coming into your brain. Top-down information is the knowledge your brain gleans from its interactions with the outside world. Top-down information is the rules of how our world works. What's appropriate in certain situations, inappropriate in certain situations.
3:32
What goals might be available out there, and what plans have been successful at achieving those goals in the past or similar goals in the past. So it's just our knowledge of how the world works. And a major function of the prefrontal cortex is to absorb and use top-down information to plan and execute goal-directed behavior. And also suppress counter-productive behavior. Essentially, your prefrontal cortex-- which is at the front of your head, behind your forehead, above your eyes--
3:59
is the brain's executive. It's not the only part of the brain involved in goal-directed behavior or executive control, but it is the main part of the brain. It's the main part of the brain whose function is to give us executive control over thought and action. So I'll begin with one of our earlier studies, one of the first studies showing a distillation of top-down information in cortex, in prefrontal cortex particularly. Then I'll move on to how this folds into our new view of working memory.
4:30
So another example of top-down information came from a study that our laboratory did in collaboration with Tommy Poggio's lab. David Freeman, a graduate student in my lab, and Max Risenhoover, a graduate student in Tommy's lab, spearheaded this work. And what we wanted to know is whether there was neurons in the brain-- whether the prefrontal cortex and other brain areas, whether they can really absorb top-down information in a way that distills
4:58
the essence of what's needed for a given goal-directed task and throw away irrelevant details. So what we did is we took three prototype cats, three prototype dogs, and we used a morphing algorithm that morphed or blends the shapes between these prototypes. So we looked across this line here. This cat is gradually becoming this dog. This dog is gradually becoming this cat. So we have a smoothly-varying shapespace here. But we trained the monkeys that anything that was more than 50%
5:32
cat was by definition a cat. Anything more than 50% dog was by definition a dog. And we could draw this line anywhere or even change its orientation and retrain the monkeys. You get the same effects. The point is, we have this smoothly-varying shapespace where these cat and dog shapes are smoothly changing. But a sharp dividing line separates one from the other. And the goal of this task was for the monkeys to figure out what makes a cat versus what makes a dog. And they did so, very well.
6:01
Now what this morphing algorithm allowed us to do is disentangle physical appearance from category membership. Category membership is the top-down information the monkey needed to solve this task, where the exact physical appearance was something that could be boiled away, thrown away, was unnecessary. And we could do this, because here's our boundary here, 50% boundary. You notice that this cat right here, right across the cat-dog boundary-- this cat looks very much like this dog.
6:28
So they look alike, but they're in different categories. At the same time, different members of the same category, like this house cat here and this mountain lion here, could look very different from one another. So we're looking for neurons that distill this top-down knowledge about the category membership of these shapes and throw away the irrelevant details about the exact physical appearance. And that's what we found. Here's an example of a single neuron in the prefrontal cortex.
6:55
What the tiles show are all the different cat stimuli we used for that day on the left. The right shows all the different dog stimuli we used on that day. And the color indicates how much that neuron activated, how many spikes that neuron gave off, to each of these different cat or dog stimuli. And as you can see, this neuron and many others like it-- this neuron gave lots of spikes, lots of impulses for all the different cat stimuli in a pretty similar fashion,
7:25
and gave fewer spikes or no spikes to all the different dog stimuli. So this neuron is neatly separating cats versus dogs. So it's sensitive to the category membership of these stimuli, but it is insensitive to what the stimuli looked like, because you get pretty much the same activity to all the different cats and all the different dogs. So again, category membership encoded in neurons in the prefrontal cortex. Irrelevant details about appearance, throw it away.
7:54
And of course, this neuron gave more activity to cats, less activity to dogs. Other neurons did the opposite. Now another example of this came from Andreas Nieder. When he was in my lab, he trained monkeys to recognize small numbers-- numbers one through five-- across a wide range of different-looking displays. Just a small example of this is on the bottom of the figure here. And what we found is that neurons in the prefrontal cortex tracked the numerical values of the displays.
8:24
So here's an example of a single neuron here. Here's spikes per second, how many impulses the neuron is giving off on the y-axis. Time is on the x-axis. These gray areas is with monkeys viewing one of these displays. And this particular neuron gave the most activity whenever a display had four things in it. And progressively, less activity for three or two or one things. Other neurons had different preferences for different number displays. But altogether, different neurons
8:51
in the prefrontal cortex could encode the numerical value of the displays. But they were insensitive to the exact appearance of the display. So the exact appearance didn't make a difference. All that mattered was what numerosity they were representing, which is exactly what we trained the monkeys to do. And finally, another even higher-order example comes from Joni Wallis. When she was in my lab, she trained monkeys to switch back and forth between two rules. One was a match rule.
9:20
Choose the picture that was the same as the picture you saw before. In this case, the monkey is seeing two pictures-- seeing a picture at the start of the behavioral trial. Then there's a short memory delay of one second or so while the monkey thinks about the picture. Then the monkey gets an alternative of two different pictures. If the monkey was following the match rule, it meant that it should choose the picture that was the same as it saw before.
9:45
Now if instead the monkey was following the non-match rule, that meant when the monkey was-- after it saw this picture and saw two other alternative pictures at the end, the monkey had to choose the picture that was different from the one it saw before. The one that was a non-match. So for this task, the monkey had to hold in working memory both the picture that it saw at the start of the trial and which rule it was currently following. And we cue that rule.
10:10
We can't just tell the monkey do match or non-match. We cue the monkey with a variety of sensory cues, some of which are shown here. Tones, shapes on the computer screen, drops of juice-- monkeys love apple juice-- that told the monkey whether to follow the match rule same rule, or the non-match different rule. And what we found is that many neurons in the prefrontal cortex tracked the rule the monkey was following and threw away all these irrelevant details.
10:36
So here's an example of a single neuron. Spikes per second, again, on the y-axis. Time on the x-axis. The monkey is seeing the first picture or sample out here. The monkey is seeing the two alternate pictures out here and making its decision. Here is the working memory delay in the middle. And here's an example of four different cuing conditions we used to cue. Two cues meant the same rule. Another two cues meant a different rule. And this neuron and others like it
11:00
are tracking which rule the monkey is following and not tracking what the cues were. So once again, we have an example of neurons in the prefrontal cortex, the brain's executive, that is distilling the top-down essence of what's needed to solve this task. It's boiling away the irrelevant details and just pulling out the information needed to solve-- the top-down information needed to solve the task. So the prefrontal cortex distills top-down information needed to solve the task.
11:30
And it disregards irrelevant details. So about 20 years ago when we were doing this work, we found something that was thought to be very curious at the time. We randomly selected neurons from the prefrontal cortex using electrode arrays. Just we drop our arrays in and we find whatever neurons we happen to find. And we found that a large proportion of neurons in the prefrontal cortex showed just the property we were looking for. 30% to 40% of randomly-selected neurons
11:57
in the prefrontal cortex showed task-related information, whether it was cats versus dogs or numbers or choose same versus choose different. Or any number of many other experiments we did. We're finding huge proportions of neurons in the prefrontal cortex that seem to be tuned to the task that we just trained the monkey to do. Now what does this mean? Why is this? Well, one possibility-- why is there such a huge representation of this task in the monkey's prefrontal cortex?
12:24
Almost half of the neurons. Well, one possibility is that the task's information is over-represented as a result of training. If you do something like train an animal to detect different frequency vibrations on their fingertip, give the animal enough training, the finger representation will expand in the somatosensory cortex. That was a possibility. But I don't think that really explains these results, because our monkeys had months of experience on these tasks,
12:55
whereas humans have years, even decades of experience doing the same thing. So the amount of experience they have in this task was not unusual. So that didn't seem like a likely explanation to me. Another possibility is the monkey can learn two to three things and their brains will fill up and they can't learn anything else, because they've used up all the neurons in the prefrontal cortex. No. I will say this. This was sarcastically suggested to me at a talk
13:22
about 20 years ago when I first presented this work. That's because years ago, systems of neuroscience had the paradigm that every neuron had one particular function. Each neuron has one specific function, like a gear in a big clock. And if we figure out what each neuron does, we can put the whole brain together. But what results like these suggested is something very, very different. Namely, that many cortical neurons don't have a single function. They are are multi-functional.
13:54
They do different things in different contexts and different tasks. They're like utility players on a baseball team. They can play one role or another, depending on the demands that are currently given to them. John Duncan and I called this adaptive coding. More recently, Stefano Fusi called this mixed selectivity. Now, again, this adaptive coding-- 20 years ago, this was thought of as a bit of a stretch. In fact, a senior investigator accused John Duncan
14:25
and I of turning the cortex into a bowl of porridge, by suggesting that neurons are multi-functional. But I'm happy to say that since then, this finding has been replicated over and over and over again. But more recently, what Stefano Fusi did is he examined-- he made computation models and examined the properties of mixed selectivity to show why it's important that the cortex has these multi-functional neurons. And what Stefano showed with his computational modeling,
14:54
using data from our laboratory, is that mixed selectivity adds computational horsepower. If you want to do high-level cognition, you need mixed selectivity. What it does is these multi-functional mixed selectivity neurons creates a high-dimensional representational space. A lot of information is intermingling in the same networks of neurons. So arbitrary information can be put together really suddenly to create new rules, new tasks, new whatever. And most of what we do on a day-to-day basis
15:24
are these arbitrary rules, like red means stop and green means go. We can pull together arbitrary bits of information and make a spontaneous new rule. And that's what mixed selectivity provides, is high-dimensional representational space where that can happen. Another thing it allows us is solving more complex problems. Stefano showed that when you have a network model where every neuron does one thing, it only can learn very simple problems.
15:52
It can't learn very complex, especially nonlinear-type problems. It also provides greater storage capacity. By using our neurons, neurons recorded from our laboratory, as a measure of how much information these networks of neurons can store, what Stefano showed is that if you have a neural network model where every neuron does one thing, it can learn a handful of tasks. Then it quickly runs out of storage space.
16:21
But when you start adding mixed selectivity, multi-functional neurons into the mix, at some point, around 30% to 40% of the neurons-- maybe not coincidentally-- 30% or 40% of neurons in their networks have these mixed selectivity properties. Then the network goes capacity unlimited. It can learn as many tasks as you can throw at it. And also it buys flexibility, because in this high-dimensional representational space, you can pull together different representations very quickly
16:51
and then pull together other representations. It buys you cognitive flexibility that allows you to learn things quickly, learn arbitrary rules quickly, because all the information is intermingling in the same network of neurons.
17:04
But this mixed selectivity, though, this multi-functionality of neurons raises a question. What mixed selectivity means is that neurons participate in multiple ensembles. If each individual neuron has different functions, it means that they participate-- an ensemble is a collection of neurons that represents a single thought, a single precept, a single memory. And the idea is that an ensemble of neurons is what underlies these things. But if neurons are multi-functional,
17:36
that means they participate in multiple ensembles. And if so, how do you select one ensemble from the underlying anatomy? What I learned in graduate school is that anatomy is destiny. If two neurons have a strong connection and one fires, the other will fire. So if anatomy is destiny, how do you select a single ensemble from these interacting multi-connected networks of neurons? So, for example here, we have two ensembles. Think y in pink, think x in purple.
18:08
And they share common elements, common neurons. If anatomy was everything, if I tried to think pink, think y, I would activate all the pink neurons. And also, I'd activate the neurons that are also part of the blue ensemble. And then both ensembles would be activated and we'd have a smushing of representations in the brain. So how do you do this? How do you pull out these single ensembles? Well, one proposal suggested by people like Wolf Singer, Pascal
18:38
Fries, Andreas Engel and others is that neural ensembles are formed by synchronized rhythms. Sure, you have this underlying anatomy. But the way you have a unique representation of, say, the pink ensemble, is that neurons in the pink ensemble synchronize their ongoing rhythms. Your brain is highly rhythmic. There's an obvious signal you can record from the brain. If you stick electrodes on your scalp, you get these squiggly lines. They're higher frequency when you're awake.
19:07
When you go to sleep, they shift to low frequencies. If they ever go flat, you're either dead or in a coma or deeply anesthetized. So your brain is constantly oscillating at frequencies everywhere from 1 Hertz to 100 Hertz or more. And the idea here is that neurons that form a single ensemble, they oscillate at the same frequency and in phase with one another. So all these pink neurons are oscillating together. And all these blue neurons are oscillating
19:33
at either a different frequency or a different phase, so they're out of sync with the pink neurons. That way, we can tell one ensemble from the other. So the idea here is that-- you've heard, maybe, from Donald Hebb, a famous neuroscientist from the '40s that neurons that fire together are wired together. It's how you form anatomy in the brain. We want to add something on top of that. Neurons that hum together temporarily wired together. This allows one ensemble to be selected
20:03
from overlapping anatomy, because you can pull something off from the overlapping anatomy just by synchronizing the constituent elements of a given ensemble. And what this buys you is cognitive flexibility. You can change your mode of thought. Learn new things. Change your mind from one moment to the next in a moment to moment basis, by shifting patterns of oscillatory resonance. If I want to change my mind, these patterns of oscillatory resonance shift around, and now I have a new thought.
20:30
I don't have to rewire my brain every time I change my mind. So the way to think about this is anatomy is not destiny in the brain. Anatomy is possibility. Anatomy is like the road and highway system. It says where traffic could go. Something else says where traffic does go from moment to moment. Spikes are the traffic. They carry information. And we think these patterns of oscillatory resonance are the road signs or traffic lights on the brain's anatomy or road and highway system.
21:03
So anatomy is potential in the brain. It's latent. It says what thoughts you could have. What thoughts you actually do have depend on these patterns of synchronized oscillatory resonance. That's the idea. When Tim Buschman was in my lab, he decided to test this directly by training monkeys to switch back and forth between two different rules. So you saw the match, non-match rule earlier. These are different types of rules. The monkey saw either a vertical or horizontal bar and the bar
21:35
could be red or blue. But we trained the monkey at different trials to pay attention to either the color of the stimulus, red or blue, or pay attention to the orientation, vertical or horizontal, and make a response based on either the color or the orientation. So the monkey is switching back and forth between these two different rules. And what Tim found is shown summarized here in this slide. This shows a representation of the monkey's lateral prefrontal cortex.
22:02
Arcuate sulcus here, principal sulcus. The lateral prefrontal cortex is like right here on the side of your head. On the monkey, this is posterior. That's anterior. The monkey's eye would be around here, looking forward. And each of these recordings-- each of these circles is a different recording site where we had an micro-electrode in the monkey's lateral prefrontal cortex. Now what Tim examined-- he examined both LFPs and spikes. But for this analysis, he examined LFP synchrony
22:29
between different recording sites. Again, LFPs-- local field potentials-- are these mass actions of millions of neurons near each recording electrode. And what Tim found was that different patterns of different recording sites showed synchronized LFP rhythms for one rule or the other rule. Now we looked across all frequencies from 1 Hertz to 100 Hertz for these sort of synchronized patterns that corresponded to the rules. And we found them only in the beta range, 12 to 30 Hertz.
23:05
Higher frequency, like 35 Hertz, 40 Hertz and above, gamma, nothing was there. I mean, there was oscillations going on there. But they didn't form these synchronized patterns. Only in the beta range, 12 to 30 Hertz, did we find that different recording sites synchronized their rhythms differently for one rule versus the other. So this slide in left here shows each line connects recording sites that shows significant increases in LFP synchrony in the beta range for the color rule.
23:42
On the right here are the lines connecting the pairs of recording sites that showed an increase in beta-band LFP synchrony whenever the monkey was following the orientation. And here we see on the macro level kind of what I was talking about the single neuron level, and that is mixed selectivity. There's a lot of overlap in these patterns. Many of the recording sites participate in both the color rule network and the orientation rule network. Or if you will, the color ensemble or the orientation
24:12
ensemble. Many of the recording sites participate in both. But there's unique patterns of synchrony for the color rule versus the orientation rule. So it does seem that synchronizing rhythms-- in this case, synchronizing beta rhythms in the prefrontal cortex-- seemed to form representations, ensembles of information that represented which rule the monkey was currently following.
24:41
So once Tim made this discovery, we then looked and we did further experiments. We looked into old data. And we found these effects virtually in every data set we looked at. So Evan Antzoulatos, when he was in the lab, he trained monkeys to recognize spatial categories. A stimulus appeared somewhere on a computer screen and the monkey had to tell us whether the stimulus was up versus down. And we changed the definition of what was up and down on a trial by trial basis,
25:10
so the monkey couldn't just simply memorize where the stimulus appeared. The monkey had to make a categorical judgment, up versus down. And what Evan found was that there was different patterns of these synchronized ensembles, beta ensembles, both within the prefrontal cortex and also between the prefrontal cortex and the posterior parietal cortex. Posterior parietal cortex is another higher order corollary involved in spatial cognition, among other things.
25:38
And we found that these patterns of these beta-band ensembles were evident both in the prefrontal cortex-- like one pattern for up, another pattern for down. And also between the prefrontal cortex and posterior parietal cortex. One pattern between those areas for the category up. Another pattern of beta synchrony for the category down. We also looked back over our old shape category, our cat versus dog data, and found the exact same thing. Different patterns of beta synchrony
26:11
across different recording sites. In this case, I'm talking all about LFPs here. Different patterns of a synchronized beta activity between LFPs in the prefrontal cortex. One pattern for cats, another pattern for dogs. So it's shifting patterns of oscillatory beta-band resonance in the prefrontal cortex, depending on whether the animal is making a cat judgment or dog judgment. In all the experiments I've told you so far, we've trained our animals and then recorded in the brain
26:45
to see what we found in their brains. But Evan Antzoulatos did another experiment where we trained monkeys to learn new categories online while we examined their brains. And what he did is he trained monkeys to learn by trial and error two different categories, category A and B. These categories were formed by dot patterns. Here's one example of one category prototype here on the left, another category prototype on the right. We made two new categories each day. And the monkeys had to learn.
27:16
And we showed the monkeys exemplars of these categories, distortions of these prototypes. And the monkey had to learn what makes category A versus what makes Category B by trial and error. And what Even found is that as the-- first, the animal is just guessing. Just guesses A or B, A or B, doesn't really know. But after a while, when the animal gets enough experience, they start picking up on the category membership. And what Evan found is that as the animal was learning
27:42
these category memberships for these dot patterns, you could actually watch these beta-band ensembles forming in the prefrontal cortex and forming between the prefrontal cortex and a subcortical area important for learning, the striatum. So they're not just there in the cortex. We happened to find them. We could actually watch them forming as the animal learns. And they form these beta-band rhythmic ensembles, actually formed in parallel with the monkey acquiring category information.
28:14
OK, now that we've talked about that-- about top-down information-- let's get to today's topic, working memory. The temporary holding of information online to make it available for further processing. So how do you hold thoughts in working memory? Working memory isn't just holding thoughts. It's also, like I said, volition. It's also manipulation of thoughts. But its basic function is holding thoughts in mind online so you can use them. The classic model for the past 50 years or so
28:45
has been persistent spiking of neurons. Here's a couple classic examples from Pat Goldman-Rakic. In this experiment here on the left, the monkey is looking at this dot. And you put a cue somewhere on the computer screen. The monkey's job is to remember the location of that cue, while keeping its eye still. Then after a short multi-second memory delay, the monkey tells you, oh, the cue was on the upper left. So the monkey's holding the cue in working memory. This shows a single trial example.
29:15
Here's what you see from one neuron averaged across multiple trials. Because in a lot of neurophysiology, you have to average across multiple trials in order to increase your signal-to-noise ratio. So what happens typically is there's some baseline level of activity spiking from these neurons. And then you show the animal a cue it has to hold in working memory. That cue activates neurons in the prefrontal cortex and other cortical areas. And then after the cue goes away,
29:42
the neuron maintains some elevated level of spiking activity during the delay, until the monkey makes his response and everything goes back down to baseline. So the idea here is a very straightforward and appealing model, because what it suggests is that you activate a neural ensemble when there's a sensory input. You look at a cue, it activates an ensemble. Then an ensemble simply remains in active state after the cue goes away. And therefore, it's available for processing.
30:12
However, more recent work has shown under close examination, and especially in real time-- i.e. in single trials where you're not averaging across multiple trials-- that neurons, many neurons in the prefrontal cortex and other cortical areas involved in the working memory, don't show purely persistent spiking. A few neurons do show persistent spiking. But the bulk of the population show very sparse and bursty activity during the memory delay.
30:39
You often get this illusion of this persistent activity during the delay, because you're averaging across multiple trials. On some trials, the animal fired bit-bat-boom. Other trials, the neuron may fire bing-bang-boom. So they're firing at different times in the trial. And you get this illusion of persistence by averaging across all these different instances. But in single trials, the neurons actually-- most of them show very sparse activity.
31:04
And I'll show you an example of that in a couple of minutes. So this led to some updates to the classic model. One of the early updates came from Pat Goldman-Rakic, where one of her last papers, she showed that in the prefrontal cortex, spiking induces temporary increase in synaptic weight changes, lasting about a second or so. And in her paper with Wang, they concluded that working memory wasn't just persistent spiking. It was spiking plus these temporary increases
31:37
in synaptic weights that are induced by the spiking. So you have spiking that changes synaptic weights. Then you have more spiking, constantly refreshing these synaptic weight changes. More recently, Mark Stokes suggested something similar with the Activity-Silent model, where he had the neurons in the brain involved in working memory. They activate. They go silent. They become active again. It isn't just pure persistent activity. It's a pure maintenance of some sensory input.
32:07
And there's also the Dynamic Attractor model of Mikael Lundqvist, which I'll tell you about in a moment. But the point of all these models is that, from Pat Goldman-Rakic on up, that spiking is not doing all the work. It's helped along by temporary impressions that spiking leaves in networks. And this saves energy. Your brain uses a huge amount of metabolic energy. And spiking takes a huge amount of metabolic energy. So this idea of persistent spiking never really sat well with me.
32:38
Evolution likes to conserve energy. Why would you use working memory all the time? Why would your brain evolve a system for maintaining information and having a function you use all time, by this constant persistent firing that's using up all this energy? By having the spiking happening less frequently and the synaptic weight changes sort of carrying the memories between spikings, you save a lot of energy. But also, it has functional consequences. It makes working memories more robust.
33:07
Memories held by persistent activity-- you can model persistent activity by what's called attractor states. And these models have shown that memories held by persistent activity alone are easily disrupted. Some distraction comes along. Or you want to add some information to working memory. Any new input disrupts the attractor state and changes the memory. And also, for the same reason, it's hard to hold more than one memory at a time, because the memories-- similar attractor
33:37
states tend to smush together. And you get chaos in the brain. So these temporary synaptic weight changes solve these problems by limiting the amount of time working memory spends in these vulnerable attractor states. If attractor states when the neurons are active is a bad state to be in, by having these synaptic weight changes, you limit the amount of time that the neurons are in this-- why the memories are in this disruptible form. So I'm going to tell you about the Dynamic Attractor
34:07
model of Mikael Lundqvist. But first of all, background information. Mikael's model also depends on spiking and then silence and synaptic weight changes. And Mikael published this paper in Journal of Cognitive Neuroscience in 2011. I happened to be the editor who was handing the paper. And I thought, wow, this is a pure network model. No data, just a computational model that Mikael built from biophysics on up. And I remember when this paper came across my desk,
34:38
I thought, wow, this is a new way of looking at something I've thought about for a couple of decades. And after the paper was accepted, Mikael sends me an email that says, dear Dr. Miller, I know you know about my paper. It shows that everything you thought about working memory is wrong. Can I come to your lab and use your data to prove you wrong? And I was kind of charmed by this, because I like to think I have a healthy disrespect for dogma. So I said sure.
35:06
But I'm not telling you this to tell you how charming Mikael is. And he is charming. I'm telling you this because this Dynamic Attractor Network model wasn't a model we built to fit around existing results. Mikael built the model from computational and biophysical principles, nuts and bolts on up. And then after the model existed, we then went and tested it in the data. And we found that the data supported the model. So here's how it works. Neurons coding the same stimulus are recurrently
35:39
connected in the local clusters. Clusters form assemblies. Another way of saying ensembles. Clusters of neurons, local connections, and more wide-ranging connections form ensembles. Competing clusters share feedback in, addition. In other words, different ensembles, different representations in the brain, are all trying to shut each other up. This is familiar because Pat Goldman-Rakic's original model of working memory had this competition between ensembles.
36:07
And it's also featured in Bob Desimone and John Duncan's bias competition model of attention. All these ensembles in your brain are all competing with each other for representation. Now in Mikael's model, default network state is oscillations in the beta range.
36:25
Now I'll point out that there's no such thing as a default network state in the brain. Your brain's always doing something. But for this purposes, default means no sensory input. So in no sensory input, your cortical neurons are bubbling along in the beta range. Occasionally, the membrane potentials due to the oscillations are crossing the spike threshold, so neurons spike at some low level. But then you have a sensory input. That influx of energy from the sensory input kicks the rhythms
36:57
up to a higher frequency-- gamma, above 35 Hertz-- for a very short burst. Then all the other ensembles shut it down. And this burst of spike, this gamma and this burst of spiking, causes the synthetic weight changes that carry the working memories. So that's the basic model. The way we tested the model is shown here in this slide. I'm not going to walk you through all the details. The important thing is the monkey is holding either two or three pictures-- in this case, colored squares--
37:27
in working memory. That's it. Just holding two or three. I'll show you examples of trials where the monkey was holding two things in working memory. But that's all there was to the task. So what we found in the prefrontal cortex is shown here. This is a example of just a single trial, not averaging, just a single trial. It shows the LFP power-- power of the local field potentials-- at different frequencies. Power is in color. Red is high power, blue is low power. Frequency on the y-axis.
37:57
And here's time during the trial. So S1 is when the monkey saw the first pictures, stimulus one. There was a short gap in time. Then the monkey saw the second picture, stimulus two. Then there was this longer delay. And the monkey had to tell us what pictures it saw at the end of the delay. And what we found is that in individual trials, there was these bursts of gamma taking place during stimulus encoding, as you see here. The stimulus appears and there's a burst of gamma, burst of gamma here.
38:25
And then a burst of gamma during the memory delay. These gamma bursts were very short in duration, less than an average of three or four cycles. And they're very narrow in frequency. And they bounce up and down in frequency. These bursts of gamma co-occurred with the spiking that carried information about the working memories. The stimuli the monkey was holding in working memory. In fact, some recording sites in the prefrontal cortex had spikes and these gamma bursts co-associated,
38:55
whereas other recording sites in the prefrontal cortex had spikes but no gamma bursts. But only the gamma burst sites-- only the sites that had this co-occurrence of gamma bursting and spiking are the ones that carried working memory information. Other sites had lots of spiking going on, but they didn't carry any information. And they had no gamma bursting. So that fits the model that you had encoding of working memory by these bursts of gamma and associated spiking. Now what this shows here--
39:23
this makes this point here how these gamma bursts are associated with information content in working memory. What we've plotted here is not spiking, but information in spiking on the y-axis, versus time on the x-axis. And here's the two stimuli the monkey's holding in working memory. The first one appears and the other one appears. Here's the memory delay. The monkey's tested out here. The red shows the information about these two stimuli in spiking at sites that had the gamma bursting.
39:54
Gamma-modulated sites, we call them. And blue shows the information in spiking about these two stimuli at sites that had spiking, but no gamma bursting. And as we see, there's information, plenty of information about what the monkey is holding in working memory at the gamma bursting sites, but virtually no information at the sites that did not show gamma bursting. Now it wasn't just spiking that was doing it, because the average spike rate between these two
40:22
different types of sites was nearly identical. The difference between one or the other was the information. If there was gamma bursting, there was information. No gamma bursting, no information. Now another thing is that these gamma bursts occurred mainly during stimulus encoding, and also spontaneously during the memory delay. They alternated with bursts of beta. So down here in the lower frequencies, you see there's these two-- in this trial-- two beta bursts.
40:51
They also were very near on frequency, short in duration. And they alternated with the gamma bursting. So whenever there was bursts of gamma, there tended to be no beta. Gamma being, again, high frequency, beta, low frequency. When there was bursts of gamma, there was no beta. And if there were bursts of beta, there was no gamma. So the working memory seemed to be this alternation between gamma and spiking carrying information, alternating with these bursts of beta.
41:21
And we think this may help explain volition, how we control what we think about. Now why is that? Well, one reason is shown here. This shows now on the right side, not an individual trial, but an average of gamma or beta burst rate across trials. And again, this is the burst rate on the y-axis, time on the x-axis. Here's the first stimulus. There's the second stimulus. Here's the working memory delay. And as you can see, the gamma and beta bursting are like mirror images of one another.
41:55
Gamma goes up during stimulus encoding. It drops down at the start of the memory delay, then gradually builds up over the course of the memory delay as the animal's getting ready to read information out of working memory. And beta does the opposite. When gamma goes up, beta goes down. And when beta goes up, gamma goes down. So there are two opposite signals. They're anti-correlated. And this suggested to us a mechanism for volitional control over working memory.
42:22
If gamma and associated spiking is actually carrying the working memories, well, we can turn on and off working memory storage by changing alpha-beta. Or changing-- I told you this was a beta, but you may notice that these frequencies dip a little into the alpha range, too. Around 10 Hertz is around the border between alpha and beta. Most of what I'm talking about is in the beta range, sometimes a little bit alpha. But these borderlines in different frequency bands
42:48
are a little bit arbitrary, so I'll just refer to them as beta. So if gamma is carrying the working memories-- gamma and the spiking associated with it is carrying the working memories, and beta shows a mirror opposite of gamma, maybe we can regulate working memory storage by regulating beta. If we want to encode things into working memory, we turn down beta, so gamma can be expressed and information can be encoded in working memory. And if we want to clear out working memory, we turn up beta.
43:20
That should drive gamma down and drive away the spiking that's carrying the working memories. And working memory is cleared out. So gamma is the contents of working memory, what you're actually holding in working memory. And beta may be an inhibitory control signal that regulates storage of information in working memory.
43:40
So I mentioned that gamma is encoding the actual information held in working memory. What is this beta doing? Well, for the first half of my talk, you learned that top-down information is expressed by unique patterns of beta coherence in the prefrontal cortex and other cortical areas. We've looked for these examples of these rhythmic ensembles in the brain for different types and a wide variety of top-down information, like rules and different types of categories.
44:10
And we always find it in this beta range. And importantly, it isn't just that-- it isn't beta power that's carrying the top-down information. It's patterns of beta. Different patterns, different beta rhythmic ensembles. So what that means is these different patterns can target-- if they're carrying top-down information from the front of the brain to the back of the brain, because they're patterns, they can target specific representations in posterior cortex.
44:40
So in other words, beta-gamma interactions may provide a means for top-down knowledge carried by beta-- carried by this patterned beta synchrony-- to regulate the processing of bottom-up information, sensory information, carried by gamma. So in order to look more carefully at this, Andre Bastos-- who's currently in the lab-- and Roman Lunes-- they decided to look at the micro-circuitry in different cortical columns that support these beta-gamma interactions. Now a standard recording electrode
45:12
has a recording surface just on the very tip where we record individual neurons or local LFPs. Andre and Roman used these laminar probes that have multiple contacts along the shaft. So you can record from all layers of cortex simultaneously. And what they found is illustrated here. First of all, what they found is they found that the spiking that's carrying working memories-- the contents of working memory-- is mainly taking place in the superficial layers of cortex.
45:43
So here's the cortical layer here, the distance along the vertical axis of the cortex. The dotted line is on layer four, where information first comes into each cortical column. We could determine where layer four was, the middle layer, by doing a current syn-source analysis. You show a stimulus, you see where the information first flows into the column. And that's where layer four is. So above layer four are the superficial layers, below layer four are the deep layers.
46:14
And this is multi-unit activity expressed in color, with the red being lots of activity, blue being suppression relative to baseline, and greenish being no activity relative to baseline. And the monkey was performing a working memory task here where it saw a cue and then had to hold the cue in memory over this working memory delay. And most of the spiking that was carrying these working memories were in the superficial layers of cortex. Now that made sense to us, because the superficial layers
46:42
of cortex are the feed-forward layers of cortex. They're the layers of cortex that are carrying information from the back of the brain to the front of brain. Sensory information from the back of the brain up through the brain. So given that the monkey is holding this recent sensory cue in working memory, it makes sense that the sensory cue would mainly be represented in the superficial layers of cortex-- the feed-forward layers that are feeding sensory information from the back of the brain
47:11
to the front of the brain. Whereas by contrast, the deep layers of cortex are the feedback layers. They're carrying information from the front of the brain to the back of the brain. Top-down information from the front of the brain to the back of the brain. So spiking, carrying the sensory information the monkey is holding in working memory in the superficial layers. When they looked across for beta and gamma power, what they found was that there was-- beta and gamma power in the LFPs were
47:39
evident in both the superficial and deep layers of cortex. But gamma was stronger in the superficial layers of cortex, whereas beta was stronger in the deep layers of cortex. And that made sense, because again, gamma is associated with bottom-up sensory information the monkey is holding in working memory. Gamma should be-- given that it's associated with bottom-up information, it should be in the superficial or feed-forward layers of cortex. Remember, I mentioned that beta is associated
48:06
with top-down information. It's supposed to be carrying this top-down information from the front of the brain to the back of the brain. It makes sense there would be more beta in the deep layers of cortex, the feedback layers of cortex. They then did a Granger's causality analysis, which allows you to look within each frequency band at which signals are influencing one another. Granger's causality is a measure of how much one signal is pushing another one along.
48:29
And what they found is that deep-layer beta had a big influence on superficial-layer beta, but not the other way around. It's a one-way interaction where deep-layer beta was regulating superficial-layer beta. But there was no interaction from superficial to deep layers. So one way from deep layer to superficial layer interaction. And gamma was pretty much not really pushing around anything at all. So here's our Granger's causality measure and frequency.
48:56
A lot of the red shows the deep to superficial layer influence. And the blue, superficial to deep. And all the influence is deep beta influencing superficial-layer beta. Now if you look across frequency bands now, here's the correlation between beta power and gamma power in the superficial versus deep layers of cortex. On the x-axis, here's the layer providing beta, superficial versus deep layers. On the y-axis is the layer providing gamma, where we're recording gamma from,
49:34
deep or superficial layers. And the red is the correlation. Red is positive correlation. Blue is a negative correlation. And what this shows is this negative correlation between beta and gamma I showed you earlier is taking place between cortical layers. Whenever deep-layer beta is high, superficial-layer gamma is low. Whenever superficial-layer gamma is high, deep-layer beta is low. So the same anti-correlation, but now playing across different cortical layers.
50:06
So what this suggested is that beta carrying top-down information may play an indirect inhibitory role by regulating superficial-layer beta. And thereby regulating gamma may inhibit superficial-layer gamma and thereby regulate or gate the information that makes it into working memory and is maintained in working memory. So top-down information in beta having an inhibitory influence on the gamma in superficial layers that are feeding forward sensory information.
50:38
So in order to explore this further, we decided to look at a more complex working memory task, where there's multiple decision points and operations. We wanted to know whether this beta signal is really acting like a control signal. So I'm going to just briefly say, the easy stuff that happened in the beginning the trial and the end of the trial before I move on to what's unique about this particular task. So first of all the task. The monkey starts. It's a sequence matching task.
51:11
The monkey saw two pictures in a row. There was a working memory delay. Then the monkey saw a test sequence. If the test sequence was the same one as the monkey saw earlier, the monkey pulls a lever to say yes, that's a match. It's the same sequence I saw a few seconds ago. But if the test sequence was a non-match-- it could be different pictures. Or the same pictures in the wrong order. Or one picture's right, the other picture's wrong. That meant the whole sequence was a non-match.
51:38
And the monkey held the lever, failed to pull a lever to tell us no, that's a non-match. So very quickly, at the beginning of the trial, every working memory task has a beginning and end. So I'm just going to go through and summarize this quickly. In the beginning of the trial, the monkey needed to encode information in working memory. What happened and what should happen is beta should go down, beta power goes down. Gamma power goes up, then spikes start carrying the working memories.
52:04
Then at the end of the trial, the monkey needs to clear out working memory. What should happen and what does happen is beta goes up-- the opposite of what happened at the beginning of the trial. Beta goes up, gamma goes down, and spikes stopped carrying the working memories. So that result is in this paper. You're welcome to read it. But let me get on to what's interesting about this particular task. What is interesting what this task is there's a crucial decision point here
52:31
at the start of this test sequence. Once the monkey sees this first test stimulus here, it has information about whether the subsequent sequence is going to be a match or non-match. Now the monkey is not allowed to make a behavioral response until the very end of the trial, but this decision point is occurring seconds before the monkey makes his behavioral response. So we're looking at a pure decision, independent of behavioral response. So what happens in this decision point?
52:58
Let's say the first test stimulus is a non-match. Well, if the first test is a non-match, the whole sequence is wrong. If the first one's wrong, everything's wrong. The second one's going to be wrong. It doesn't make a difference what the second one was, the sequence is wrong. So what the monkey can do, even though it's trained not to respond to the very end of the trial, what the monkey can do is just forget about it. Just clear out working memory. The monkey knows the first one's a non-match,
53:23
so even though I have to wait to respond, I don't really need to pay attention anymore. I can just clear out working memory. So what happens? Now we're looking at the sequence between-- the sequence at the end of the trial. Here's the first test stimulus here, between these lines. Here's a second test stimulus here, when the monkey has to make its behavioral response here at the end of the trial. And this is the delay between the two test sequences. So if the first test is a non-match,
53:50
the monkey knows that in working memory, the task is more or less over. And the monkey can clear out working memory. So what happens? What happens is that there's a big increase in beta, right after the first test stimulus appears and the monkey knows it's a non-match. And that drives gamma down. So here's a nice strong gamma response to the first test stimulus. The monkey knows it's a non-match, clears working memory out. Beta goes up, drives gamma down.
54:18
And now the gamma response to the second test stimulus is now blunted, because it's no longer behaviorally relevant. And then what should happen is spikes should stop carrying working memories. And that's exactly what happens. It drives information in spiking. The increase in beta and decrease in gamma drives spiking down to near zero. There's a small registering of the second test stimulus here, but now neurons have pretty much stopped carrying working memories.
54:45
Now what about when the first test stimulus appears, what if that first test object is a match? Well, if the first object is a match, the monkey can't forget about that first test object. He has to now start thinking about the test object, because of the second test object, the one he now has to judge. So it has to switch the contents of working memory. I've already judged this stimulus here. The first one's right. Now I've got to think ahead to the next stimulus.
55:11
The monkey is switching the contents of working memory from one stimulus to another. What happens here? Well, what happens is something similar but distinctly different from what you saw earlier. Here's again our sequence of the two test stimuli out here at the end of the trial. Now there's a short latency, but short duration increase in beta. Earlier, when the monkey can clear out working memory, this increase in beta bridged this entire memory delay and drove gamma down and drove spiking down.
55:43
But now, the monkey has to switch the contents of working memory. There's an increase in beta, which we think clears out the first stimulus in working memory. But then beta quickly drops down before the second test stimulus appears. And this should allow gamma to re-express, to reemerge. That's exactly what happens. So now when beta is being driven down, now gamma can reemerge. And there's a nice robust gamma response to the second test stimulus, which
56:11
is allowed because beta has been driven down. And what should happen is that spikes should start carrying information. This should allow spikes to start expressing information about the second test object the monkey is anticipating. And that's exactly what happens. So these beta-gamma interactions, the beta really does act like a control signal that's not only opening up working memory and clearing it out, but allows you to switch the contents of working memory.
56:39
Now another thing us behavioral neurophysiologists do is we look at error trials. If a given neurophysiological effect we think is behaviorally relevant, if it's relevant for the performance of this task for example, it should go haywire when the monkey makes an error. Now in this particular task, the monkey was making a lever response at the end of the trial. All the monkey was saying, yes or no, go or no go, match or non-match. We don't know why the monkey made an error.
57:10
Did it make an error because it got the first stimulus wrong? The second stimulus wrong? Did it mix the two up? We don't know. But by using the beta-gamma dynamics on correct trials, we could use that as a barometer to see what was going on in incorrect trials. So we used deviations in the beta-gamma dynamics on correct trials to interpret error trials. And that told us why the monkey was making errors. We could tell why exactly the monkey made an error.
57:37
Did the monkey get the first stimulus wrong, second stimulus wrong? Did it misjudge this, misjudge that? We could tell that by looking at these interactions, these beta-gamma dynamics. And we couldn't tell that from spiking alone. Spiking would tell us where the monkey would make an error, but wouldn't tell us why the monkey made an error. Only with these beta-gamma dynamics could we tell exactly what the error was. So to summarize, this is our model.
58:03
Top-down information is carried from the front of the brain to the back of the brain by these patterns of beta coherence in deep layers of cortex. Couple of interesting things about deep layers of cortex. The thalamus forms loops of deep-layer cortex. And the thalamus is interesting in this case because we learned from people like our own Mike Halassa that the anterior thalamic nuclei are critical for carrying feedback information in the cortex. Not feed-forward, but feedback information.
58:33
And we learned from work like Yuri Solomon in Sabine Castro's laboratory that one of the ways a thalamus does this is by providing lower-frequency oscillations like beta to the cortex. So this deep-layer beta carrying top-down information regulates superficial-layer beta. It's weaker in the superficial layers, but it's there. So deep-layer beta regulates the expression of superficial-layer beta. Superficial-layer beta is anti-correlated with gamma. So it regulates gamma.
59:02
And in turn, that regulates the spiking that actually carries working memories. So again, we don't know yet from these experiments why volition happens, what volition is. But we believe this is a canonical cortical circuit by which top-down information can regulate the expression and flow of bottom-up information in the brain. How top-down can regulate bottom-up. Now I've talked about this in terms of working memory and prefrontal cortex. But since then, we've done experiments all over cortex.
59:35
And we're finding these same interactions everywhere we look. So we think it's a general mechanism by which the cortex uses top-down to control bottom-up. And in fact, Andre Bastos in the lab has just completed a series of experiments where he shows that these patterns of top-down beta can actually target specific representations in the visual cortex. The animal's going to predict a particular stimulus, while the beta effects are strongest at the places in sensory cortex, in visual cortex, that
1:00:08
represent that stimulus. So that's it. Thank you for listening. Thank you for joining in our webinar. I also want to thank all the people in my laboratory who do all the really hard work in conducting these experiments. Here is the laboratory here and a few friends. I'd also like to thank our funders, of course. And I'm happy to-- and thank you for your attention. And I'm happy to take any questions now. Thank you.
1:00:41
OK, thank you. Anybody that is going to be called on to ask questions, we will unmute you on our end. But you do also need to unmute on your end. Our first question comes from Pegah. Hello, can you hear me now? Yes, I can. OK, thank you very much for the great talk. It was great. I cannot help but remember all the literature that exists about alpha, mainly in the posterior areas as sort of like a functional oscillation gating sensory information. It kind of-- like what beta does here
1:01:22
in the context of working memory reminds me of what alpha does for sensory gating, probably. I wanted to know your idea about this. Well, that's exactly right. In fact, we recently completed a series of experiments where we recorded from all over the cortex-- from the sensory cortex up to prefrontal cortex. And alpha is doing the same thing in the visual cortex that beta's doing in prefrontal cortex. In fact, as you move from visual cortex up to prefrontal cortex,
1:01:51
alpha gradually changes into beta. So it's the same mechanism. Just there's a gradient from posterior to anterior. OK, thank you. Thank you.
1:02:04
Thank you. Our next question comes from Wayne Wu. Hi there. Can you hear me? Yeah. That was really interesting. Thank you. And the work you've done is really great as well. The claim about volition I'm trying to grapple with. The first part of the talk, the beta was associated with, say, task rules, I guess, that the monkey has to keep track of. And that sort of has a clear link to volition in the sense that it's regulating processing to perform a task. But I was a little bit--
1:02:40
I was trying to figure out how to really map the bond to the last set of experiments that you were talking about, where beta and gamma are anti-correlated. So in those experiments, you want beta to basically be a mechanism of clearing out working memory, as required in the task. And I guess the one slight disconnect is just to understand what that might have to do with task rules. Right. Well, the rules of the task are things that allow the animal to gate information out of working memory.
1:03:15
The monkey knows, oh, this first stimulus I'm seeing, I should remember it. Once the trial is over, I need to forget it, because another trial is coming along. And it's going to be another stimulus I have to hold in mind. So that's all implicit in the rules. In that particular task, we didn't vary the rules. So we didn't see beta changing with different task rules. Neurophysiology is a bit like radar. You only get a signal out when you put a signal in. And in order to see top-down beta changing
1:03:43
with different task rules, you have to vary the rules. We did that in other experiments, just not in this one. So the idea is that the beta is carrying this task rule information, allowing all this expression of the rules and following the rules at different points in the trial. But in that experiment, we didn't vary the rules. So we didn't see it, but you see it's there from other experiments we've done. So you might see in a different experiment that if you did
1:04:10
vary the rules, then you'd have beta correlated in the right ways with whatever the task-relevant stimulus is at that time. Oh exactly. And you saw kind of a-- the last experiment I talked about, you sort of saw something like that going on, where the animal knows from the experience of the task and the rules that if the first test stimulus is non-match boom done. Working memory is over. And you see that beta going up in response. So it isn't exactly the representation of rules,
1:04:40
because we're not varying the rule in that case. But we do see the beta carrying forth an implementation of the rule. Thank you. That's great.
1:04:52
Thank you. Our next question comes from Gal Raz. Thanks so much for your talk. Thank you. My question is about exactly this interpretation of clearing out by this beta oscillation, because you also showed a surge in beta between the two samples. So you showed first two samples, and you have high gamma on the first sample. Then beta increases and then high gamma again for the second sample. But in that case, we don't want to clear out the first sample, right?
1:05:22
So I'm wondering if you can say more about how you interpret the surge in beta in that case. I think that's where the synaptic weight changes come in. So we first encode information and work memory gamma goes up. Then it drops down and beta goes back up. But then there's a synaptic weight change that allows the memories to be carried through, even though gamma and spiking is low and beta is high. But the important thing about these synaptic weight changes is they last--
1:05:47
and this is in the Goldman-Rakic experiment-- last like under a second, based on calcium dynamics. So if you don't refresh the synaptic weight changes with more spiking every once in a while, they'll naturally fade away. So the idea there is that even though after the first stimulus appears and gamma is going down and beta is going up, the memory is still there in latent form in the synaptic weight changes, waiting to be refreshed once the memory delay ensues. Gotcha. Thank you so much.
1:06:19
Thank you. Thank you. Our next question comes from Meng Wang. Hello. Hello. Can you hear me? Yes. Hi. Thanks for your talk. So I have a question about the paper, I think it's from Tim Buschman, 2012 neurons. So that's where you show a beta organized network, right? Yes. According to the color of orientation. So did you also analyze why in resting states? Because from our research on the beta network in the primary sensory cortex, there is also a beta organized network even in the resting state.
1:07:23
Yeah, and also, if you are in a behavioral task like a sensation or something, the beta organized network was even reduced. Sure. I mean, the important thing to remember is that when a trial's over and the monkeys-- or whatever human or animals-- are waiting between trials, the brain doesn't just turn off. It's doing all sorts of other things. It's thinking about other things. And they're going to be driven by these kind of dynamics. Now since we're not studying those--
1:07:57
we don't have control over what the monkeys or animal or human is thinking between trials, we can't tell what the beta dynamics are corresponding to. But presumably, they're operating the same between trials as they are within a trial. Except that within a trial, we can tie it to rules and content, because we know what the animal should be doing. So yes, it sort of reminds me of a-- it's a bit reminiscent of default state network kind of things, where you get lots and lots of default state
1:08:29
activity-- default network activity in the brain-- when a subject's lying in a scanner doing nothing. Then you start the trial, and all of a sudden, activity goes down. Well, the reason why that happens is because when the subject is lying in a scanner or the monkey's sitting in a chair or something and not doing the task, the monkey is thinking about a dozen things. Or the human's thinking about a dozen things, like why the hell am I lying in the scanner? How much am I being paid for this?
1:08:54
And stuff like that. But then the trial starts and now you're narrowing the focus down to one thing. So there'll be a corresponding narrowing of activity once the brain focuses from whatever thoughts are popping in their head into a single thought. So yeah, you do see this high beta during the inter-trial interval. That's because the brain is doing its own thing in that case. And it drops down to a different level and then varies along that different level
1:09:24
in the trial, because now the animal's thinking about one thing at a time. OK, next question comes from Mastwals. Hi, thank you. My question was regarding the learning of the task. It seems like the goal for learning these categories and rules is the reward that the dead animal would learn by. And that may involve the neuromodulation and dopamine or other neuromodulators. I was wondering-- I was curious about what your thoughts were on how those neuromodulators would affect
1:10:03
the formation of the categories and rules and the beta patterns in the frontal cortex. Oh sure. Yeah, we think that since these are goal-directed tasks and the animal is working for a reward, we assume that-- and there's lots of evidence to support-- that this is driven by dopamine dynamics acting not only in the frontal cortex, but also places like the striatum. Especially in the striatum, where dopaminergic neurons-- the reward neurons-- are hammering the striatum, almost
1:10:32
like a gating role for plasticity in the striatum. And then the striatum can feed patterns up to the prefrontal cortex, which can then do further elaboration and learning. But yeah, this will all be driven by those same dopamine dynamics that drive other tasks where we're actually varying things like reward value. We're not changing reward value, so we're not seeing the dopamine changing. But we assume that this is-- you can't do everything in one experiment or one set of experiments.
1:11:00
But yeah, this will all be driven by those kind of dynamics. Do you think that dopamine has any direct role in actual formation of those beta patterns? Well, I think it plays a role in changing synaptic weights. In that sense, yes, it will. I mean, that's what the animal-- the animal's learning these-- there's two things going on here. There's the temporary changes in synaptic weights that are happening when you're holding a particular stimulus in working memory.
1:11:31
And those are not going to be dopamine-driven. They're going to be just results of spiking and temporary impression. But then if you want to learn something for the long term, then you have to change synaptic weights more permanently. And that's where the dopamine is going to come in for that. OK, thank you very much. Sure, thank you. Thank you. Our next question comes from Cooper Penner. Hi, Dr. Miller, appreciate your time and your decades of work. Did you get the--
1:11:58
My question is about your network model of accrual learning. And I guess, in my own mind, I'm just having trouble balancing all those interconnected nodes-- interconnected beta nodes-- and the discretized nature of beta. You show it as these individual events. Do you see it as just simultaneous individual moments of beta that come up at the same time? How do you think about it? Because for me, it's difficult if I'm not considering it as a continuous rhythm.
1:12:32
Locally, the activity of beta and gamma bursts tend to be synchronized. By locally, what I mean is within the vicinity of a recording electrode. But it is. You move further away from the recording sites, they're not synchronized. The beta-- you're filling space with these beta and gamma bursts. You're killing time with these beta-gamma bursts across wide ranges of cortex. But locally, they happen in a very synchronized way, as you'd expect, because we're seeing this in the LFPs.
1:13:03
If they weren't happening in a synchronized way, we wouldn't be able to detect them. What's interesting about if you look at things like-- if you look across things like the average low frequency, like beta or even lower frequency, theta, you see cognition actually waxing and waning in a rhythmic way. So one example of this is Tim Buschman also did an experiment in the lab where he had the animal keep its eyes still, but search a visual display for something.
1:13:36
And then when it found it, it can move its eye to that. Sort of like paying attention to somebody, but we're not really looking at them. That's what that was like. It was like a where's Waldo thing, but you have to wait to move your eye until you find Waldo. In that case, we see that-- you look at the average beta across the frontal eye-fields, which are in the posterior part of the prefrontal cortex. You see that the animal's moving of its mind's eye,
1:14:00
searching for this Waldo, is happening in beta rhythms. And it's synchronized with the beta rhythms in the frontal eye-fields. So cognition, even at a higher level, is rhythmic in this fashion. And the exact frequency can change from task to task. For example, this was being captured in a series of experiments where the monkey also did a where's Waldo experiment. But now the monkey was allowed to move its eye around the computer screen to find Waldo. In that case, we found the same sort of thing
1:14:32
that Tim Buschman found. We only found lower frequencies, theta. Now why the shift to lower frequencies? The reason is you don't move your eyes four times a second. So these low frequencies do seem to act like a control signal, not only controlling when working memories are expressed, but they also control things like visible search and attention and actual eye movement. And if you have people do a sustained attention task, looking for a slight change in a stimulus
1:15:02
on a computer screen, you'll see their ability to detect that change waxing and waning in different frequency bands. Usually in the beta to alpha to theta range. So I hope that answers your question. Yeah, no, it does. Thank you. And we see something similar in our data in deeper structures. Awesome. I was happy to hear your talk. Thanks great. Thank you. Our next question is from Carolina Ferreira. Hi Carolina. Unmute. There you go. Yeah hi.
1:15:31
So great talk, thank you for sharing your work with us. I'm a clinician, so my question is a bit more clinically-oriented. So working memory dysfunction is like a common prodrome symptom in patients with schizophrenia. Do you think alterations in the person's gamma, alpha, beta, would seemingly could be associated with this? Sure. So Andre Bastos in the lab is working on extending this beta-gamma working memory model to predictive coding.
1:16:00
And the idea is your brain's constantly making predictions about what it expects to see next. And the idea is that the beta is carrying these predictions and suppressing gamma representations back in visual cortex that represent-- recording sites in the visual cortex that represent the stimulus the animal's expecting. And that's how predictive coding works. You want to suppress things you predict that are coming, because they're not very informative. You knew they were coming.
1:16:30
Now why do I mention predictive coding? Because it's been proposed that deficits in predictive coding can explain a lot of neuropsychiatric disorders, like autism. In autism, every stimulus is a new stimulus, a violation. The beta predictive mechanism is somehow weak. Or the gamma is overwhelming it. And as a result, you don't suppress anything. And everything gets past 40. You have gamma for everything. Schizophrenia has also been proposed to also-- I mean, predictive coding dysfunctions
1:17:01
are also thought to underlie predictive coding, maybe in the opposite direction. They're thought to underlie schizophrenia, sorry, maybe in the opposite where you have too much prediction and it kind of overwhelms reality. I'm not a clinician, but I can see ties to these different things. Thank you. Thank you. Our next question comes from Harish Gunasekaran. Hello. Hello. Hi. Hi professor. It was a great talk. I just have one question. I was quite intrigued by your induction analogy
1:17:35
between the anatomy with highways and the spikes with the traffic and the rhythms as the signboard or the traffic light. Right? That's right. So my question is where is the source of this kind of oscillations in the first place? In the first place, what is the origin of these oscillations. Is it localized to one particular micro-circuit? Or is it an emergence from the interactions of multiple circuits? Well, we think it's an emergent property of interactions both locally and--
1:18:08
depending on excitatory and inhibitory balance, interactions between neurons. But it can be shaped by things, which is how the brain can then co-opt it for use in top-down transmission of information. How the brain manages to get volitional control over its own thought and actions in that way, we could talk about interactions in cortex and basal ganglia and the thalamus and recursive loops that allow-- I mean, closed anatomical loops that allow recursive processing.
1:18:36
Throw dopamine in there to tell you what's important to learn. Build on that in a recursive fashion. And somehow, you get the actual volitional signal. That's a tough question. If I had the answers to all that, we would have the brain largely figured out. We just think that this cortical circuit that I talked about is the infrastructure by which this stuff acts. OK, thank you. Thank you, professor. Thank you. Next question is from is Jang Jin Kim. Thanks for the wonderful talk.
1:19:10
And I have two questions. First, your beta-gamma alternation seems really, really synchronized. But I guess that comes from the monkey's already knowing the task. My question is whether those beta-gamma alternation changes across the line. For example, before the monkey fully learned the task, does beta-gamma alternation-- is it less synchronized, compared to like the full learning? Yeah. Training animals-- recording like this during learning is a little bit-- it's not as straightforward
1:19:50
as you might think. Because the way you train animals is that 90% of training, you're sort of giving them hints. Giving them the answer, kind of. You're helping them along. And it's only in the very last 10% of training where you remove all the training wheels and now the animal's really doing the task. So there's a lot of things going on during a standard training regimen that makes these kind of questions difficult to ask. What we have done is do things like--
1:20:17
like Evan's done a lot of this in the laboratory where he picks a particular thing-- like you train the animal in the basic task. But the monkey has to learn new things all the time, like new categories. And that's when-- I didn't show the data, but in those experiments, you see these beta circuits ensembles actually forming as the animal learns something over a three or four-hour session. Now we need to re-examine that data for the anti-correlation
1:20:48
with gamma, but I predict it'll be there, because we're seeing that everywhere we look. I see. Yeah wonderful. One more question, please. So have you analyzed any of the phases of sequence between the spike and your beta and gamma? Oh yeah. Yes, we do. We have. So for example, Tim Buschman, in his experiment where he looked at pay attention to color versus pay attention to orientation, we saw a beta color rule ensemble versus a beta orientation rule ensemble.
1:21:19
There was phase synchrony between the spiking and these LFP rhythms. The beta rhythms. And what's interesting is that the beta ensembles are forming the rules. The rule of pay attention to color or pay attention to orientation. But the spikes that were actually carrying the sensory information-- not the rule, but the actual color red or the color blue or the representation of vertical versus horizontal-- those sensory neurons phase synchronized to the appropriate rhythmic rule ensemble.
1:21:54
Wonderful. Yeah, thank you. They were trickling down into the single neuron level. So yes, you do see those kind of phase coupling. OK, thank you. Thank you. Thank you. We have time for one last question and that will be from Stan. Hi Earl, this is from Stan Goldin. Hi Stan. Great talk. And what's puzzling me, and fascinating me is the following. I'm trying to picture if you were at a particular point in the cortex as one of these waveforms, beta or gamma,
1:22:28
travels from layer to layer, many of these neurons have quite a bit of dendritic branching. And so when a waveform approaches a distal dendrite, it may be very much out of phase with what's going on in a spiking cell body. And there's a lot of physiology suggesting that some neurons direct their dendrites towards cell bodies versus dendrites. So rather than considering the unit of activity involved in memory, would it be worthwhile to-- and I imagine you probably have reflected on would
1:23:10
regions of dendrites be-- if you somehow could record individual activity from dendritic branches in real time, like you're doing spiking. What would you see and what would be the implications of it? Well, we can't really do those experiments. Because with this neurophysiological approach, we're generally going to be recording from the largest thing there is, which is cell bodies. I know that. This is a thought experiment. Oh, a thought experiment. I'm going to speculate.
1:23:41
Well, one of the things we're looking at is how these rhythms, whether they form waves that travel. And preliminary data from the laboratory suggests that they do. And that has all sorts of implications for spike timing-dependent plasticity. So it isn't like it's happening everywhere simultaneously. These are waves that travel across the cortex and can travel from dendrites to cell bodies. So that's a whole other dimension we have to get into. Because these rhythms aren't just happening--
1:24:09
they're not just running in place. They form a wave structure. And we have to understand all that better. I don't have any answers for you, other than we're doing work on that now. And I hope to have an answer soon. That's great. Thank you very much. Thank you. Thank you very much. That comes to the end of our question and answer period. Thank you, Professor Miller, for your talk. Oh, please call me Earl. Thank you. Bye.
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat