The thalamus in cognitive control and flexibility (1:19:21)
Date Posted:
October 5, 2018
Date Recorded:
June 27, 2018
Speaker(s):
Michael Halassa
All Captioned Videos CBMM Summer Lecture Series
Description:
Mike Halassa, MIT
MICHAEL HALASSA: I'm in this department. I run a lab on the fifth floor. The lab is interested in understanding the basic architectures that implement cognitive algorithms, something like attention, executive function, working memory. We want to understand what are the brain circuits that are involved? What are the architectural solutions that the brain has for these kinds of problems? And down to some implementatioal-- so we want understand the algorithm. How does the brain-- what kind of computational framework does the brain use to solve these kinds of problems? And what exactly are those computations instantiated in?
I'm going to talk to you today about the thalamus and its role in sort of gating cortical interactions as a solution for some of these flexible cognitive computations. And yeah, so that's basically thematically what the lab is interested in.
But practically speaking, I mean, the work relies on collecting data in animals. So the key approach that we have is using animals that are mechanistically tractable, meaning that you can access their circuits either through genetics or either through cell-specific labeling. Because at the end of the day, we want to build these models, and we want to have some way of being able to validate them in the actual animal.
So traditionally, we've used mice to get the kinds of neural data and behavioral data coregistered and to sort of get a sense of when the animal does task x that I'm going to talk to you about, what are the brain areas that get engaged? And how do they do what they do? And what kind of potential computational solutions can we glean from looking at that data, right?
So mice have been really, really good to us because we can get them to do complicated tasks that allow us to record activity and then go in, and based on the stories that we imagine those neural activity to be telling, test specific predictions through perturbations, like optical perturbations or drugs or whatever, right? OK.
And more recently, we've gotten to a point where we wanted to push the cognitive end of things more, and mice were sort of incapable of doing certain things that we wanted to test. So we've now moved into looking at tree shrews, which I can tell you a little bit about. Part of my talk will be about this particular animal model that's still genetically accessible, has all the rodent tools that make rodents a viable model in neuroscience but has primate-level cognition. So I can talk to you about that.
OK. Does anybody have any questions so far about anything that I said? Yeah.
AUDIENCE: How difficult is to get OKed to work with tree shrews. Like, how long have you been working with them?
MICHAEL HALASSA: So that's a good question. The tree shrews have been a model system, as far as I can tell, since-- I mean, the papers I'm aware of date back to the '70s. Actually, Mark Baer, who is here in the department, did his undergraduate honors thesis on the shrew visual system and that was in the '80s, maybe. So the lab that he was in used shrews to look at anatomy of the visual system.
And the reason why that was an attractive model is because it had a primate-like V1. So people thought it was a good model for studying vision. Right now, I would say, the biggest, the most-- the best known lab for using the shrew as a model in vision is David Fitzpatrick, who is, if somebody was saying they're from Florida Atlantic University, is that-- yeah? So that's right-- no?
AUDIENCE: Florida International--
MICHAEL HALASSA: International, oh, yeah, yeah. So there's a Florida-- what's that?
AUDIENCE: It's nearby.
MICHAEL HALASSA: Yeah, the Max Planck Institute-- so he's the director of the Max Planck. He uses tree shrews as a model for vision. But nobody's actually used them as a model for cognition. So we actually had to test some of these cognitive tasks in them to see how well they do, and they just blow rodents out of the water. I mean, they can display primate-like behavior in an animal that really is the size of of a rat and has a brain composition that looks like-- I mean, it's a bigger-- I'll show you. We'll talk about it when we get to it.
Any other question? Somebody else had a question.
AUDIENCE: No.
MICHAEL HALASSA: What's that? It was answered.
AUDIENCE: Oh, it was answered. OK.
MICHAEL HALASSA: All right. So should we start? Yeah? So what I'm going to tell you about today is the role of the thalamus in cognitive control and flexibility. But let me tell you a little bit about cognition. We sort of talked about it. But what I mean by cognition is not something that's particularly intellectual or hard or yeah, anything intellectual. What I mean by cognition is something that's not reflexive. Like, being able to take an input and generate an output that's not reflexive but rather is based on an internal model of the world. That's what I mean by cognition.
And that type of operation is happening in everyday scenario at multiple different levels, right? So I'm going to just give you a real-life example to give you the intuition of what I mean by having an internal model of the world, updating it, using it to map and input onto an output, right?
So imagine that you're basically walking in a forest, and you lay your eyes on this particular object. You're not really-- I know. It's gone. That is part of the trick, right? Because most of the time, you're moving your eyes multiple times a second. You're basically foveating or fixating on an object for just a few milliseconds, which is sort of what happened here.
So what are the kinds of stories that you start generating when you lay your eyes on something? You sort of think that you've seen something like this as part of that object, right? And if I had to ask you what is that? What is being shown on the screen? Does anybody know?
AUDIENCE: Unripe blueberries.
MICHAEL HALASSA: Green berries, right? But you said blueberries, right?
AUDIENCE: I said unripe blueberries.
MICHAEL HALASSA: Unripe blueberries. So that was not part of what was displayed. I can tell you that because I generated that image. But you immediately said that it looks like unripe blueberries because you remember blueberries from years of reinforcement learning, where you put these things in your mouth as a baby, and they're delicious. And that's a memory that stuck, right?
And see, the difference, what I mean by cognition is the thing that separates us from animals that display very reflexive behavior is the fact that we can form these hierarchical associations of different object categories. So the reason why you immediately said these are unripe blueberries is because you have an idea of seasons, right? And the seasons that hierarchy of ideas, where seasons is part of how objects are related to one another allows you to make an inference that these really are unripe blueberries.
And then this is the kind of inference that allows you to generate sort of a prediction. Are these things blueberries? And over time, what happens? You lay your eyes on something for 100 milliseconds. You're not sure what it is. Over time, you have this expectation. You have this inference. You have this update signal. And over time, what you're doing is you're adjusting your sensory filters in different parts of your brain to route information in a particular way so that something that was amorphous, a big green blob, becomes something that's more obviously a blueberry bush that you can go and interact with. Does that make sense?
And that's basically the key of what I mean when I say cognition. It is this process that's dynamic, that's based on some internal model, some memory and based on gating of different signals. So this dynamic updating of our perception will be key to how we interact with the world.
So the example that I always like to give is that frogs are really good at catching flies, but they don't really tell stories about the flies, right? They don't go and say, well, I'm going to hide behind that particular bush because this particular type of fly is going to be the plum, ripe, type of-- if you can call flies ripe or whatever-- this plum fly is going to be appearing behind this bush. I'm going to hide and wait for it. That's the kind of behavior of waiting, imagining, all these kinds of things are associated with higher cognition. And the reason is, is because we have a brain that can support all these kinds of nested hierarchical associations. So the goal of the-- No, no.
AUDIENCE: Genomes seem like they're really kind of basic evolution, What do you mean by hierarchical?
MICHAEL HALASSA: Hierarchy of this case has to do with you have an object category-- so first of all, the hierarchy can be at different levels, you know? There is a hierarchy of form, right? You form a hierarchy. So you can have objects that are decomposed into their basic elements. That's a hierarchy of form, right?
And then you get to a point where you form the invariance of a blueberry to the low level sensory features like you know exactly what color they are or exactly what shape they are, the size. That's a level of hierarchy, right? That's an object category that's blueberry. That belongs in another category of fruit, for example.
And then the category of blueberries being-- green blueberries and blue blueberries being related to each other is another hierarchy that relates different object categories to one another. That's what I mean by hierarchy.
AUDIENCE: And now those categories are assumed to be in place in everyone's head? So right, that blueberries belong--
MICHAEL HALASSA: Well, you spend the first whatever years of your life building those categories, right? I mean, you're not born with these categories. That's the idea. The idea is that you build those categories over things like reinforcement learning. I mean, all of learning is to build these hierarchical representations. And the reason why you have a bigger brain is because you can build more hierarchies.
A monkey, for example-- part of the reason why you would appreciate art is you have the level of invariance over categories that you can relate something very abstract like a Picasso to the shapes of real human faces. Those are related by this meta level of all face categories that don't have to be realistic. And a monkey, for example, may not have that. I don't know if monkeys would appreciate art the same way that we would.
So these are the sort of like broad ideas that frame the kinds of stuff that we do in the lab, you know? So what we're trying to get after is what are these kinds of algorithms where you make an observation, compare it to memory, update your observation, and incorporate it in a selection mechanism, right? That's what we study in the lab.
And we think, just to give you the sort of-- so getting back to this idea of hierarchies, if we look at the cortex of a person, what we notice is that the representations in the cortex go from being pretty simple in the back of the cortex to being much more abstract, as you move from the back of the cortex to the front of the cortex. So as you get input from the visual world to V1, you generate these representations of edges and motion and basic form vision. And as you go down the ventral stream, for example, you start generating object categories and so on. And as you go to areas like the prefrontal cortex, then you're generating representations that are much, much more abstract, those of rules and categories and categories over categories and so on. So that's the idea.
See, one thing that I wanted to emphasize at this point is the brain is not only solving blueberry tasks, the sort of things that we end up being confronted with that have similar features but different details are vast. We're interacting with all kinds of objects in our environment from different modalities, et cetera. So there's got to be a mechanism for these cortical circuits to be reconfigured in a task-dependent way, right? And there's got to be a mechanism for these reconfigured circuits to talk to each other in the right way to solve a particular task.
And the architectural solution that the brain takes in order to solve these kinds of problems is not obvious. And one of the things that my lab has been focused on is to try to understand whether this area called the thalamus, which is centrally located, can play a role in configuring and gating cortical interactions in a task-relevant matter. And that's basically what I'm going to talk to you about today is evidence that the thalamus can, under certain conditions, play these kinds of roles. Make sense? That's the premise for the content of the discussion today
And some of you are neuroscience majors, so you've probably heard about the thalamus. Some of you may not have heard about the thalamus, right? Who's heard about the thalamus? OK. So everybody's heard about thalamus, to a large degree.
So if you were to tell me what does the thalamus do? What would you say?
AUDIENCE: It relays.
MICHAEL HALASSA: It relays. That is the common view of the thalamus. It's that it's a relay of information, sensory information, from the outside world to the cortex. That is wrong. And the reason why it's wrong is because it's based on how this particular region of the thalamus works, the lateral geniculate nucleus. So the LGN is the source of visual input from the retina to the primary visual cortex.
And yes, under certain conditions, the LGN can be formalized, the function of the LGN can be formalized as a relay. It gets topographical input from the retina, projects topographically to V1. And you can describe V1 receptive fields based on a convolution of spike rates coming from the LGN into a particular V1 neuron. So the V1 neuron, the representation of that cell can be described by a weighted sum of LGN inputs, and it makes perfect sense to call that a relay, OK?
But what I'm going to try to convince you today is that a very small subset of thalamic territories operate in this manner. Most of the thalamus, in your brains and mine, aren't even connected to sensory inputs. They receive most of their inputs from cortical areas, and they do something with it that cannot be described as relay, play much more interesting roles in cognitive computations and these cognitive algorithms than taking an input and outputting an output that gets just summed to form a representation, OK? So I will show you data that supports the view of dynamic reconfiguration of cortical networks in a task-relevant matter.
AUDIENCE: So if it were just a relay, then it could be replaced by just a bundle of axons, right? But it's performing some kind of function, right? So LGN is doing something to the input before it gets to V1.
MICHAEL HALASSA: Yeah. So there is some smoothing operation that happens in the LGN. There is some upsampling from the retina to the LGN, and there is some temporal filtering that can happen in the LGN. But it cannot be replaced by nothing, right? And there's also some gating happening at the LGN.
So, for example, if you're performing a task where multiple modalities are competing, like vision and audition, the LGN is a good area for you to sort of suppress one modality in favor of another if there are competition between the senses. So you can do that at the level of the LGN, but you can't do that if there's not an actual neuron there to do it to.
And we've studied processes like that, like gating of LGN input as a function. And the task that I'm going to tell you about in a few slides is actually solved that way. Basically, it can be solved at the level of gating input at the level of LGN. So it's not a passive relay. It's more of a dynamic filter. That's how I would think about it. But the output operation can be described as a relay. Does that makes sense? Because it's just taking away some of the input. But even that is not a universal thalamic output.
AUDIENCE: So us undergrads, we're taught our neuroscience class that's it's just a relay. But if you went on both neurosciences?
MICHAEL HALASSA: They would say the same thing. Yeah, yeah, yeah. That's why I have a job is because I'm saying no. It's doing something else. Yeah. Yeah. Well, I'll show you why. It was a discovery that was completely serendipitous. We weren't looking for this. But we saw something that basically told us this can't work like a relay because of x, and I will get to what x is. OK.
AUDIENCE: But isn't like the number of inputs to the thalamus greater than actually going out?
MICHAEL HALASSA: That's for LGN. That's for LGN. When people say the thalamus in undergraduate textbooks, they mean the LGN. They don't mean most of the thalamus, which actually, people haven't really studied as much. It's almost like a curse that has been placed because of how successful vision has been. So one of the inadvertent consequences of success. It's like, you just basically say, well the LGN looks like a relay, then that's probably how the thalamus works.
OK So yeah. So today, I'm going to tell you about this structure called the mediodorsal thalamus, which is one of the largest thalamic territories in the human brain. Actually, it shrinks in disorders like schizophrenia, for example. There's a whole story about the mediodorsal thalamus and schizophrenia out there in the literature. But it's hard to know exactly what to make of that story because it's unclear what computations the MD thalamus does, right?
So that's basically what my lab has been focused on is understanding the relationship between some of these abstract representations of the PFC that are involved in gating sensory inputs by controlling how information flows in other brain areas, generating these rules, these control signals, in an example like a blueberry bush or whatever. And what does the mediodorsal thalamus do? Does it do the same thing? Does it do something that the PFC cannot do? What exactly does-- yeah?
AUDIENCE: I do have another question. You mentioned that the MD thalamus is smaller in schizophrenic patients. Is this when they're having hallucinations or just on a normal--
MICHAEL HALASSA: So we can speculate once I talk about the things that the thalamus does. I think there's an interesting realm of speculation that we can engage in about, like, what are the findings that we have in animals tell us about the nature of the MD deficit that would lead to the symptoms of schizophrenia? I think there's a story there, but that's more speculative, but we can talk about that, OK?
So the blueberry bush example only works once, right, because once you see it, it's done. So how do we actually come up with tasks that would allow us to discover the neural underpinnings of a process like attentional selection or selecting one input over another or filtering, right? So this is the kind of stuff that we do in mice, come up with tasks that capture that cognitive operation.
And I'm going to-- yeah, so, yeah. So I went out of order a little bit. Anyway, so I'm going to tell you about these three things-- the interaction between the PFC and the MD. One has to do with-- the first two stories have to do with how the MD reconfigures the PFC representations of task rules. And the last one has to do with talking about how these circuits change form as you go from a mouse to a tree shrew, as you go from a rodent to a primate-like animal and with some parallels into how do we imagine it to work in a human and how do we imagine it to be relevant to public health, AI, a lot of different things?
So back to the original thought-- yeah, blueberry thing works once. We need a task that captures this operation of filtering inputs in a goal-directed manner. So we do that in mice, but I'm going to ask you, just humor me and try to solve this task that I'm going to present, OK? And that would just give you the intuition of what kind of operation we're asking the mice to do. And it's actually a hard task. So it starts off pretty easy, but it gets pretty complicated.
So all I would like you to do is look at the board. I'll get out of your way. And basically, just read every word that's presented on the screen, OK? That's it. Just read it. I got to hear it, otherwise this won't work.
AUDIENCE: Red, purple, red, green, yellow, blue, blue, yellow, orange.
OK. That was easy, right? So that was an easy task, right? So now, what I would like you to do is completely ignore what the word is, don't read it. Just report its color. OK.
AUDIENCE: Green purple, red, blue, yellow, green, red, yellow, blue, red.
MICHAEL HALASSA: OK. A little harder, right? A little harder but not impossible. OK. So now, what I would like you to do is when you see this symbol, report the color. When you see this symbol, report the word, right?
And these are going to vary on a trial-by-trial basis. So on each trial, there is going to be this random symbol that is going to appear. You understand what it means, right? But in each trial, you're going to do a different kind of operation, Right? OK.
AUDIENCE: Blue, [LAUGHTER] red, red, red, yellow, green, red, blue, blue, red.
MICHAEL HALASSA: OK. So that's hard, right? That's hard. So what happened there? Right. Well, there's some divided attention, right? You're basically splitting your resources between-- the objects that are being presented are identical, right? There's nothing changing. It's just a colored word. What I'm asking you to do is take this arbitrary symbol, turn it into a rule, and use that rule to filter the word, that object differently. One operation is to pull out the meaning or what the word says, another one, just pull out its color, right?
That's what we tried to do in mice. So it turns out that they can't read very well. So then what we do-- and you might think this is arbitrary, like this task is completely arbitrary that I showed you. But if you think about it, these symbols are arbitrary, right? They mean nothing outside of this particular context. But if you think about it a little bit more, language is equally arbitrary. These are completely arbitrary sounds, symbols that we all, depending on what language you speak, you take it as gospel. This is how it's supposed to be. But it's equally artificial, human-generated, culturally imposed, nothing special about that.
So that's part of the reason why we're-- and it's a very cognitive thing. When you're learning a language, it takes a lot of mental resources, right? And So we're trying to capture operations like that. Although it's artificial, it does teach us something fundamental about how the brain operates, that's the idea but in a controlled way.
So this is the task. Because mice don't read, we have to teach them something that they can do. And this is what we do. We have them do this two-alternative forced choice task, where on every trial, they're selecting between conflicting visual and auditory targets, right? So on any given trial, the mouse gets a configuration of target stimuli that looks like this one or like this one, OK?
The mouse initiates each trial. So we control its head and body position precisely. There's no tolerance. If the mouse pulls out or changes its head angle or whatever, the trial aborts. So it's almost like fixating the head of the animal. Otherwise, the animal is free to roam. But the animal initiates the trial. Something happens that I'll tell you about in a minute or in less than that, and the animal has to select between these conflicting visual and auditory target stimuli.
Depending on what the trial type is, if it's a visual trial, for example, the animal has to turn to the visual stimulus and go to the poke associated with it, stick its snout, it gets a reward. If it goes to the other one, it gets a time out, OK?
And what you should be asking me at this point is how does a mouse know whether this is a visual trial or an auditory trial? So at the beginning of each trial, we randomly present one of these two learned cues, which is either a high or a low-pass filtered white noise. High pass means attent to audition. Low pass means attent to vision. Or we can mix them up. It doesn't really matter but their meaning is learned, as arbitrary as the symbols that I showed you.
The animal gets this cue. We call these the cues. We call these the targets. Animal gets the cue, holds it in mind over a delayed period, while it's maintaining head and body position. So we make sure it doesn't move around, and then uses it presumably to map onto the appropriate target stimulus, similar to what you just did with the words and colors, OK?
So that's the basic task that we developed, and it was a miracle, actually, that it worked in mice because-- yeah?
AUDIENCE: So how do you-- so you're saying it either has to attend to visual or auditory, and they're like a mishmash?
MICHAEL HALASSA: How do I know that the animal is not solving one task and going pro-anti? Because they're not trained that way. So the animals are trained separately on the visual task and the auditory task. And at the end, for all of the recordings that I'm going to show you, they are in conflict. But for a lot of the non-recording, when we're just training the animals, on sessions in which we're not actually recording, or we are just training them between experiments, we randomize everything.
It's just that for the particular-- so because randomization introduces two more conditions, and we're limited by the number of trials that we can get on each session from an animal, it's hard to randomize everything on every single session. So for the recording sessions, they're in conflict. For everything else, they're randomized. Does that makes sense?
And we've done a lot of controlled experiments to make sure that the animal is not using-- it's actually harder for mice to do a pro-anti strategy because that's basically what the alternative for this is. And the pro-anti strategy gives predictions that are less compatible with the data than a sensory selection strategy. That's a great question, though. I mean, it's an important question.
Any other questions? Does everybody get this question, what it means? I mean, to put simply, what your question is, a blind mouse could potentially solve this, right? But it turns out that they don't because of the way that we train them. It's a good thing to be worried about. All right?
So I was just telling you that it was a miracle that, for me at least, that mice could learn this type of behavior because traditionally, mice have been used with things like the water mains. I don't know if people know what that is, and T mazes and things that the behavioral clump is pretty loose, so they have multiple interpretations.
But we spent quite a bit of time just making sure that mice do the tasks the way we want them to do it, and they can do this at a high level not using these low-level strategies. Because if you let mice solve tasks with low-level strategies, they will. So we were very happy about it and-- yeah?
AUDIENCE: Are the mice permitted to make any kind of motor movements?
MICHAEL HALASSA: During the-- yeah. So yeah, yeah, I mean, we don't-- so they're permitted to do that. We've done high-speed videography recording during the delay period of the task. We don't see anything in particular. We don't see them doing things like that. Nothing is obvious. Yeah.
AUDIENCE: Have you tried lengthening the delay period?
MICHAEL HALASSA: Yeah.
AUDIENCE: How far out can you go with that?
MICHAEL HALASSA: Five seconds. Yeah. Yeah we've done-- yeah?
AUDIENCE: I really don't know what the visual capacity of the mice is, but have you tried using levers or using different forms to them so that you can just change their location, or that would be really outside the idea?
MICHAEL HALASSA: What do you mean?
AUDIENCE: The visual selection, is that like-- because I don't know if I'm seeing a lever there.
MICHAEL HALASSA: There's no lever. This is just an LED flash. There's no touchscreen. So the target stimuli is an auditory sweep or a light flash, 100-millisecond just white light flash.
AUDIENCE: So how long does it take you to train an animal from novice to expert?
MICHAEL HALASSA: 12 to 16 weeks. Yeah. Yeah, it takes a long time. Yeah, this is a hard task to train animals on.
AUDIENCE: I expected something way worse.
MICHAEL HALASSA: Oh. Well, yeah. So that's the basic task. I'll show you. We basically nest this in a hierarchy of rules too, and that takes half of the life of the animal. Basically, by the time that they're performing, they're geriatric. It's a geriatric population of mice that we have running the task, but I mean, you know it works fine.
So the reason why we wanted to do this in mice-- does anybody have an intuition for that why? Why bother?
AUDIENCE: For optogenetic stuff.
MICHAEL HALASSA: Like what? So give me your dream optogenetic manipulation here in this particular--
AUDIENCE: Like silence the part of the thalamus you think is--
MICHAEL HALASSA: Who cares about the thalamus? Like, there's tons of things to silence first, right? Like, for example, this question about what strategy is the animal using, right? It's very hard to do that with behavior. You can do that with behavior. But a really cool way to do it is say, OK, if the animal is solving this using a sensory selection strategy, then if I silence the visual cortex or the auditory cortex in the same animal, I should have visual defects or auditory defects, right? Right.
But if the animals are doing a pro-anti strategy, then some animals will have a V1 effect, some animals will have an A1 effect, something like that. And you can do that, and it turns out that mice. That's part of the sort of reason why we think this is sensory selection is in each animal, you suppress V1, A1, you get effects on both that are sort of selection. They're equivalent to dimming the light or lowering the amplitude of the sound. You can get parametric changes in behavioral performance as a function of V1 and activation that look very similar to dimming the light.
So those are the kinds of things that are hard to do in a prep that doesn't have optogenetic access, for example. So we know that the selection component of the task can actually be solved at the level of the thalamus, at the level of LGN, for example. Where, if we turn off the geniculate during this delay period or during the selection period, the animal basically has deficits on the visual part of the task.
And we know that if we silence V1, we only get an effect during the sensory selection, during the target presentation. And the other thing that is actually really cool is we can now ask, where does the executive part of the task take place? Where does this configuration rule as a cue being turned into rule held in mind, where is that mind, right? What is holding onto that memory of the cue generating an instruction to control the sensory selection?
And the idea there is that it's in the prefrontal cortex, right? So we can actually confirm that, do trial-by-trial random optogenetic activation of the PFC in the mouse. And what we see is that animals can no longer select between conflicting visual and auditory stimuli. They basically guess. They just adopt a guessing strategy, and that's basically as if the animal missed the cue, right?
So then, that basically allows us to study PFC-dependent behavior in the mouse, which is really cool, right? So we can simply go in and stick these multielectrode probes into the PFC of the mouse-- that's the PFC of the mouse-- and ask, what do the neurons care about, right? Because now we're starting to construct the circuit diagram of how the task is solved. We know that the PFC is involved, but we don't know how it's involved, right?
So we stick electrodes in the PFC, and we're recording-- presumably, the most of the phenomenology that's interesting is going to be in the delay period because that's when the optogenetic suppression shows a behavioral effect.
And this is what we find. So for those of you who may not be used to looking at these kinds of plots, these are four plots from a single PFC cell that shows task-relevant modulation, task-relevant activity. These are called Rasters, Raster Plots, and these are called peristimulus time histograms, PSTHs.
So let's look at the Rasters. I've separated the trial. So in the Rasters, each line is a trial and each tick is an action potential that's generated from this cell. Zero is the time at which the cues are presented, and in this particular experiment, 500 millisecond is when the targets are represented.
So what do we see here? When we separate the trials into attend to division, attend to audition, we see that this cell seems to show an increase in spike rate at this precise moment during the delay, right? And you can see that in the PSTH, which is just basically the average, just counting these spikes and smoothing them and getting confidence intervals, basically, over the estimated spike rate. And what you basically see is that there's a peak here, not a peak here, right? Everybody sees that?
So that's one cell that seems to care about this particular moment in time, 300 millisecond, all right? But if you look at a population of PFC cells, what you see is that they tile the entire delay period, right? And there are two PFC populations, one that seems to care about attend to vision, and one seems to care about attend to audition.
So if you just look at these things and do linear decoding on them, just PCA or any kind of decoding, you will basically be able to figure out what the rule is. Yeah?
AUDIENCE: What proportion?
MICHAEL HALASSA: 20%. I should have said that. I'm sorry. These are all regular spiking cells, presumably, putative excitatory cells. OK?
So we think we have an idea of how the mouse is keeping this rule in mind based on these sequences of neural activity in the PFC. So the idea is that this is a sequential structure, where literally, these neurons are talking to each other. You have an initial stimulus-induced response in the PFC. Then there is like a domino, but that domino is cue-specific.
So there is one domino that tells us that the animal is keeping this attend to audition. Or at this point, we don't know, right? At this particular point, we don't know if this is a representation of the cue or of the meaning of the cue, right? We don't know if it's a rule or if it's a sensory representation, right?
But what we do know is that it is consistent with the synaptic chain model, similar to what has been described in the songbird for encoding a song sequence, where neurons actually talk to each other in a feedforward mode manner. And the reason why we know that is twofold. One is when we look at neurons in the PFC and look for putative synaptic interactions using the spike time cross-correlations, we see that those kinds of phenomena are much more common if the neurons are actually task modulated and if they are coding the same cue and if they have overlapping temporal fields. So what that suggests is that the cells that have putative synaptic interactions are the ones that are being recruited into the task and form this feedforward chain structure, right?
The other thing that we can do is because it's a mouse, we can actually do very short optogenetic perturbations and ask, if we perturb PFC activity at any moment during the delay, does subsequent activity fall apart? If it's a domino, if we freeze one of the one of the pieces and don't let it go through, then would the other pieces stand still? And they do. So we don't see this modulation of activity if we perturb the sequential structure at any moment in the delay.
AUDIENCE: If you extend the delay period is it cyclic like this?
MICHAEL HALASSA: Yeah. So what happens is when we have a very long delay period, a neuron that gets recruited once gets, after a second, recruited again. So it looks like it's going through a--
AUDIENCE: Does the period of that change?
MICHAEL HALASSA: A second. Oh, no, it doesn't. It doesn't. That's fixed. Even though we vary the delay period on a trial-by-trial basis for example, and that doesn't change. The precision by which these cells spike is pretty fixed.
AUDIENCE: Does it expect how long the delay period is?
MICHAEL HALASSA: So we've had tasks in which the animal that's fixed-delay period tasks in which we actually vary the delay period on very, very broad timescales, like a second or two seconds. The animal doesn't know which one it is. We don't really see an effect on this phenomenology at all. There doesn't seem to be a temporal expectation component.
AUDIENCE: So these are two different groups of neurons in the PFC?
MICHAEL HALASSA: Yeah.
AUDIENCE: How are they organized?
MICHAEL HALASSA: So nothing obvious, nothing obvious that we can-- we get these neurons when we randomly sample with these electrodes. These are adjustable multielectrode drives. I don't know if anybody's in Matt Wilson's lab. No? But these are classical tetro drives. Oh, in the Harnett Lab, you guys have something like that. Jacob Voight's has this flex drive. Yeah. So, yeah. So nothing obvious.
AUDIENCE: OK Also, Also, my other question is when you're doing the perturbations, are you manipulating the projection or are you manipulating the [INAUDIBLE]?
MICHAEL HALASSA: Everything. We're looking at everything. So we basically turn on inhibition, for example. Just turn on inhibitory neurons with channelrhodopsin at a small period of time and then look at the consequences. So yeah, yeah.
Yeah. so basically, this is what we think happens. And so thinking about this domino analogy for a little bit, say, OK, well, is it really sufficient for the animal to just hear this cue? And all of a sudden you're going to have this domino in the PFC. Is that the only ingredient that exists? The animal plus cue equals sequences?
And the answer is no because we've done this experiment-- this was an early experiment just to see what the nature of this representation is. Is it dependent on the task, right? Does the animal have to be engaged in the task in order to get these sequences? And the answer is yes because if we actually look at-- if we present the cues inside the task, we get this task-relevant activity. If we represent them outside of the task, we don't see anything. So the animal's got to be engaged in the task.
And then that begs the question of, OK, so what does the engagement in the task do? Does it recruit more brain areas? And one of those brain areas that we imagined would be recruited was the mediodorsal thalamus because it interacts strongly with the PFC. And it turns out, yes. If we actually inactivate the mediodorsal thalamus, those sequences fall apart. We don't really see any dominoes. The MD has to be recruited in order for these sequences to emerge.
So then that begged the question, what does the mediodorsal thalamus do, since it's the strongest source of excitatory input that is reciprocally connected to the PFC? And this was the experiment that convinced me that the thalamus is not a relay because this is what we saw, right?
This is a single MD neuron, MD thalamus neuron, that showed these similar bumps to the PFC but almost no neuron that we've ever looked at in the mouse thalamus cared about whether it was an auditory trial or a visual trial, right? It had these bumps, but these bumps seem to track something else other than what condition, what cue category am I in?
So then, when we started cooking up circuit diagrams-- and these inputs are completely dependent-- these responses are completely dependent on inputs from the PFC. So if you silence the PFC, those thalamic responses go away, right? And there's some asymmetry between what happens in the PFC and the MD. So we're pretty confident that the way these signals are generated is by single MD neurons having convergent input from the PFC. So it doesn't really matter whether this neuron is-- this neuron gets inputs from both the visual and the auditory networks in the PFC, and that's how it computes and generates these responses.
So then it makes no sense that this would be the source of cue information to the PFC because actually, when we silence the MD during the cue presentation, PFC responses are fine. It's only when we silence it during the delay period. If we eliminate the delay period, the MD is not required, right?
So what basically that told us is that the cue information comes to the PFC. There is a domino that's dependent on cells talking to each other, and the thalamus is generating a signal that's allowing those neurons to talk to each other, not providing an input that gets-- you do a weighted sum on to compute the receptive field of the cortical neurons. Does that make sense?
So the operation must be different because why would you have a non-specific representation being somehow magically decoded in the PFC. There's not an obvious way in which you can tell the difference between these two responses. Yeah.
AUDIENCE: So is a neuron like this, at once being driven by those two individual rule neurons in the PFC and also projecting back to that?
MICHAEL HALASSA: Yes, yes. That's the idea. The idea is that this is how it works. So that it's projecting back and it's maintaining the activity over time. And what we didn't know is exactly what that operation was. Like, is it just basically that it's increasing the excitability in the PFC? That would be the most obvious thing, where what you have is the activity is subthreshold somehow. And then you get this extra juice from the MD, and it brings neurons closer to threshold so the domino can move, right? That would be the sort of idea that we can think about.
But it turns out-- I can go through the data, but maybe that's not the most interesting thing to tell you about since I'm way over time. Maybe I won't go through the data. Maybe I'll just tell you that we've looked at this hypothesis, and it turns out it's not true, right?
It turns out that if you reactivate the MD thalamus, what happens is that the excitability in the PFC is reduced, not increased, right? What we drive mostly are inhibitory neurons in the PFC but what we see is that the temporal coordination between neurons is increased by measures like Granger causality or cross-correlations or whatever.
So then after all of that, we basically did this experiment and to directly test the intuition that activating the mediodorsal thalamus enhances these sequences by just enhancing the functional connections between the neurons and the PFC. And the idea here is that we're generating these functional connections through optogenetic stimulation.
So we're stimulating one part of the PFC, recording from another part of the PFC, right? So we generated optogenetic responses. And then we do that in the presence or absence of concurrent MD stimulation. So here's what we see. When we activate the MD, there's not much of an effect on the cortex. Basically, this is a typical example.
When we activate intracortically, we get this evoked response. But when we combine the thalamic stimulation with intracortical one, we get this nonlinear amplification. And that's basically the evidence for the thalamus-enhancing functional connections in the cortex
This is not a generic thing. If you look at the LGN, for example, and V1, you activate the LGN. You put a pulse in the LGN, you get a pulse out in cortex, which is expected. That's what the LGN is doing. It's relaying a signal. And then you combine it with intra-V1 stimulation and the responses are sublinear. So we think this is a special kind of relationship between the cortex and the thalamus for these higher order associated regions of the brain, right? A thalamic region that computes some signal based on cortical input and then uses it to maintain that representation.
But what I haven't told you, and I'll tell you about that really quickly, and then we can just sort of have an open discussion maybe. This is not important. Not important.
So the interesting thing is to understand what is the nature of this thalamic representation? Which we had not done before because the experiment that we designed could not tell us the difference between whether the animal was keeping in mind a cue or a rule, which are very different things. So in order to tell the difference, we had to introduce a new cue set. So remember, there is a high-pass noise, low-pass noise that tells the animal attend to vision, attend to audition.
Now what we did, we introduced another cue set for the same animal, where it has to do the task based on two visual cues, right? So rather than just doing the task based high-pass, low-pass, you can also do it based on UV and green lights. And the way we did this experiment is we trained the animals to do all of these things. But then the animal does it in blocks. So it can play the visual game followed by an auditory game.
In the session, we have a block of trials that are visual or auditory. Or we can actually mix and match. It doesn't really matter. The animal does it fine no matter how we construct these artificial blocks. And it does it better than if we randomize everything, which makes sense.
So imagine the original example of the colored word, we had four cues rather than two. And then I randomized the cues on every single trial or just pick two and randomize those two. It'll be easier to solve the task if I generate these artificial blocks. And that's basically what the animals do.
So anyway, so this experiment having the same neuron recorded under conditions of high-pass, low-pass, UV-green, allowed us to tell the difference, is the neuron responding to the cues themselves, one of the four cues, or to the meaning? Which would be I would respond to both high-pass and UV light just the same way. I don't care. That's some level of invariance or abstraction, right?
So we found both in the PFC. So there are some neurons in the PFC that respond to one cueing condition out of the four, and there are some neurons in the PFC that actually don't care what you cue the animal with, they respond to the meaning of the cue. And that was really cool, right?
In a mouse, you can actually find evidence for abstractions, which is really cool. So that was cool. And what we think the PFC of the mouse is constructed to be is to be hierarchical, right? You have cue responses that feed into these meaning responses or cue invariant responses.
And we can get to those kinds of inferences. Do you guys know what generalized linear models are? So we can take spike trains of these neurons that are recorded simultaneously and try to construct these connectivity measures between neurons based on-- explain the variance of one neuron population by others. And that way, we can build these-- we can extract these-- what are they called-- these coupling filters between neurons, right?
We can basically explain variance of the spiking of one neuron based on the spiking of other neurons in the population, and we can draw these directional arrows. And what those directional arrows look like is that these cue responses feed into the meaning responses, right? So that makes sense.
But one thing that-- like I was mentioning, is that even in a completely experimentally-controlled context condition, a queuing context, if we separate the cues into-- if we randomize over UV-green and then switch the animal to high-pass, low-pass, it does much better than if we randomize everything on every single trial. That's basically how it looks. Even if we combine UV and low-pass, for example, in a block, the animal does much better, which makes sense, right?
But what happens is that if you construct these arbitrary cueing blocks or cueing contexts, there's a switching cost. So what happens is that let's say the animal is doing this UV-green, you switch to high-pass, low-pass. The animal basically for about 10 trials gets confused, right? It has to recalibrate, and then goes back to doing the task.
And the intuition for that is imagine that you're driving in a city and then you're regulating your driving based on traffic lights. Now, you transition into the suburbs where they're all traffic signs, stop signs. And if you don't recalibrate, what happens is that you run a few stop signs. And you're like, oh, man, I just totally missed that. And that happens, right? I mean, that sort of makes intuitive sense. So that's what we think. It's just that mice don't learn to anticipate that, which is nice because that gives us some measure of the switching cost, OK?
So what that suggests is that even this arbitrary context that we've introduced to the animal is being kept track of because it has a performance advantage. Somebody--
AUDIENCE: Is there any change in activity of the meaning selective neurons during this?
MICHAEL HALASSA: So they go away, and then they come back after the animal-- that's a great question. But then this begs the question, who's keeping track of the context? And this is what we think the thalamus does. So remember those nonspecific responses that we saw in this HP-LP task? Those are actually context responses. We just had no way of telling them apart because there was not a different context, OK? So there are these high-pass, low-pass cells that don't respond in the UV-green condition, and the opposite is true.
Even if you construct these heteromodal blocks, we still see cells responding to that in the thalamus. So the thalamus basically captures whatever the statistical regularity of cue presentation is. It's keeping track of that larger sort of task variable that's built up across multiple trials. There are multiple-- the way the task is being computed is over different time scales. One that has to do with by trial cue presentation itself, hold the working memory. But then there's the running tab of what is, over the last 10 trials, what are the kinds of cues that have been presented? And if there is some predictability to it, it becomes a representation, and the thalamus is keeping track of that representation.
And if we mess with the thalamus during the transition, the transition becomes longer. And why does that happen? Well, it turns out that there are two different kinds of thalamic cells, one that goes back and maintains the sequence like we've sort of inferred before. And there's a different type of thalamic cell, which has a slightly different response, that suppresses the opposite context. So as the mouse transitions from doing context one to context two, the cells encoding context one are suppressed. And if we suppress the thalamus optogenetically, that suppression is diminished.
So what happens is-- what we imagine happens, you have these competing representations that compete over meaning, right? Am I in a city? What should I do with my foot? Should I hit the brakes or not? And that instruction is, in this particular case, is it attend to vision, attend to audition? So those kinds of connections are under thalamic control, right?
So basically, the thalamus captures the context over a history of 10 trials or so, uses that to suppress pretty much everything that's not task-relevant. As you switch to a different kind of context, those get activated and you suppress the original context. And we've trained recurrent networks on a similar task and added this sort of contextual node to see whether there are performance advantages. And the performance advantages come out of the decodability on the meaning cells, or the output, basically. It's easier to decode the output if you suppress-- if you basically just make these cues context-selective, if their gating is by context. It's just easier if you just lower the noise on the decoder.
And the other thing that was cool actually, that is not obvious from this formulation, that we've tested in the model, is in a recurrent network or in a deep network or in any kind of even a feedforward deep network, if you're requiring the network to do multiple different tasks, and then you're optimizing the weights based on whatever the task being done at any particular moment in time, then what happens is you go from task one. You've optimized the weights for task one. You go to task two, you optimize the weights there. Task one is gone, right? You go back to task one, task one, there's not much of-- the weights are changed now. They're basically optimized for task two.
So we think this is a solution for things like catastrophic forgetting, right? Where you tag the cells that are context-specific. You suppress them, not only to make the decoding of the other task easier but also to protect the associated weights from being changed by the new learning, right?
And that's basically what we find in the model. Sorry. That's basically what we find in the model. We have a PFC model compared to a PFC-MD model. The PFC-only model doesn't do well on multiple switches because of these changes in weights, right? And it turns out that if we do this in the animal, it works exactly the same. So if we ask the animal to do multiple switches, then as the animal switches back to the original context, it does it faster because the weights are preserved.
We suppress the thalamus in the middle only during the cueing period, which we know doesn't impact sustained activity at all. And what happens is, basically it's as if the animal forgot what task it was doing. And there's a dose-dependent response. It depends on the number of trials that we inactivate the thalamus here that we get a behavior detriment. So basically, that's why sort of the mouse and the RNNs are a good match because we can do things pretty easily in both. And that's why using the mouse for these kinds of studies is pretty cool. Yeah?
AUDIENCE: So these tasks aren't necessarily that difficult, at least, from RNN point. But you're seeing these dynamics come out that are similar to what you're seeing in the mouse brain. Do you think that says anything about brains or how you can solve these tasks?
MICHAEL HALASSA: Well, it tells us about switching, right? It tells us about why would you bother having a thalamus, for example, both amplify and suppress context relevant representations in the PFC? Why would you need a solution like that? And what that tells you is that, well, the brain is not solving this one task. It's solving millions of tasks that we're not explicitly controlling.
So the animal-- yes, this is not a particularly hard task. You can get a network to do that very easily with 100 neurons, maybe less, right? But what's hard to do is get a network to solve this task plus everything else that the mouse solves. So for the mouse, everything is a context, right? It's like the original blueberry example. That's a particular context. By being in a forest, there's a context. Looking at a tree, that's a tree-- you know what I mean?
So capturing those hierarchies is what these same networks do, all right? So the intuition that we're trying to get is what are the gating operations, what are switching operations that the brain uses? And maybe get some principles of designing networks that can switch better or not forget. Yeah.
AUDIENCE: So in your RNN models, do you actually build in explicit, context-encoding neurons?
MICHAEL HALASSA: Yes. It's not learned. It's not learned. So we don't know what the learning rule for context is or what the larger circuit that learns the context beyond the thalamus. It's just asking the question of what's the advantage of a contextual representation on switching? But yeah, but I'm super interested in figuring out what the actual learning rule for context or capturing these multiple temporal hierarchies. Yeah.
AUDIENCE: What would be the advantage of having one network or model that can just do two tasks? Why not just have two different networks that specialize in that one task?
MICHAEL HALASSA: Because you run out of networks after-- you mean what is the advantage of animals doing that?
AUDIENCE: No. In the [INAUDIBLE].
MICHAEL HALASSA: Oh, yeah, yeah, yeah. Well, so I don't know. So there's multiple answers to that question. One is that to do pretty complicated-- efficiency. One is efficiency. With less neurons you can do more. Power consumption-- you can do a lot more with less if you have the right architectural features. You know what I mean?
I don't know. Somebody was talking about this the other day. For like a Tesla GPU, the GPU that goes into Tesla, it's consuming a ridiculous amount of energy. It's just consuming, I don't know, a household worth of energy over, I don't know, a week. Normally, you would spend this much electricity over a year. You're spending it in a week.
But with a brain, you're doing so much more than that for a fraction of the cost. So that would be an advantage, just figuring out what are these low-power, more efficient ways of solving the tasks?
Ultimately, if you want to build something that looks like a human, you want to pack it in a small-- I don't know. If that's the goal of AI is AGI, then learning how the brain solves these problems and what the clever tricks that it's come up with would be cool.
But in a more abstract infinite resource world, if the answer is purely computational, that's harder to answer. I don't know if there's a genuine sort of non-plumbing answer to this. You see what I mean? The brain solves that because it has these constraints and it had the particular evolutionary history. And the thalamus was there and used it in a particular way. If you wanted to design something from scratch, would you go down that route or would you go down optimizing RNNs the way we are just doing it now? I don't know. That's a good question.
AUDIENCE: So this might come from my experience in trying to the two kinds of models. But you said that you were looking for a network optimization. Do you mean kind of AI, or do you mean the networks that are based off of firing rates and connectivity [INAUDIBLE], the networks that actually just aim to reproduce neurological signals? Or do you mean the networks, like, actually classified as abstractive?
MICHAEL HALASSA: Where? In what context are you asking the question? Is there something I said in the beginning?
AUDIENCE: I thought you said that you tested something in the model that was kind of like in the organism in like, a neural-based model where there's firing.
MICHAEL HALASSA: Yeah. These are just sigmoid neurons. There's nothing biological about them. That's just a recurrent network with these sigmoid nodes. There's no excitation/inhibition. Everything is a-- there's no Dale's law. Yeah, it's a purely computational exercise.
So the firing rates that come out of that or the sort of the activity profiles aren't necessarily anything that we see in the data. So the point of that model wasn't to fit the data, if that's what you're asking. No. It was just to get some computational insight into why would you have something like that?
And the prediction was well, actually can preserve the weights if you suppress the output of some of these cue neurons in the network. And that's true. The details are actually very different in the network and in the animal, but the broad computational benefit is similar.
AUDIENCE: I may be misremembering, but in the Schmitt paper, I believe you activated in the thalamus, and you saw behavioral points increase. So I was wondering one of my question there was why, if you have that capacity, do you not see that-- since you're overtraining the animals, in some sense-- in the second study that you're showing that's unpublished?
MICHAEL HALASSA: It gets worse. Yeah, it gets worse. Yeah, the behavior actually doesn't-- if you optimize task one, you get the switching is much harder, and then the performance doesn't get better.
AUDIENCE: So would you claim that's why you don't see it increase?
MICHAEL HALASSA: Yeah. I mean, yeah. It's hard to make these arguments, but yes. I mean, the idea is you keep things at a particular level because you're not just optimizing that task. You're optimizing so many tasks. Yeah.
In general, these tasks are extremely contrived, right? I mean and the reason why is because they just provide us with some access to the actual computation. But they're not telling us what the gamut of all computations that the animal does are, what that space looks like. They don't tell us that. That's a different scientific question.
It's sort of like how Hodgkin Huxley used a voltage clamp to discover what the dynamics underlying an action potential. No neuron lives in a voltage clamp. All these are artificial, but they give you a sense of how things are constructed.
So yeah. Any questions or thoughts?
AUDIENCE: I just have a quick one. So what you said earlier about schizophrenia, isn't there like a rule-changing PT neuro test?
MICHAEL HALASSA: Wisconsin card-sorting task. Yes. Yes.
AUDIENCE: That was what you were referring to?
MICHAEL HALASSA: Oh, no, I was talking about actual MD thalamus changes of connectivity, both structural and functional, between the MD thalamus--
AUDIENCE: And that would be like, kind of, analogous almost like [INAUDIBLE] of the [INAUDIBLE] versus that task having a smaller--
MICHAEL HALASSA: Maybe. That would be a potential idea. I don't know if people have looked in a particular task and done like correlations between some thalamocortical measure and task performance. That would be the data that's required to make the inference, to test that inference. I don't know if people have done that. I doubt it, but yeah, that that would be a really cool-- actually, maybe. There's a couple of papers that have come out over the last couple of years on this topic that may have data of that flavor.
AUDIENCE: I guess, I'm still curious to know why your manipulations were projection-specific? What is the logic behind assigning them all? Activating usually turns off the entire PFC.
MICHAEL HALASSA: What do you mean?
AUDIENCE: You activated the entire PFC?
MICHAEL HALASSA: No, no, no. I mean, so it depends on the manipulation. I mean, there are manipulations that are projection-specific, like inactivating the terminals of PFC and MD or inactivating the terminals of MD and PFC. So we've done those types of experiments. That was the one experiment that we were talking about before, the chain experiment, where we've inactivated the PFC or the--
AUDIENCE: Like the SSFO one.
MICHAEL HALASSA: Oh, the SSFO experiment. You're asking, why am I making inferences about the MD acting directly on the PFC rather than through some intermediate structure? Is that the question?
AUDIENCE: Well, my question was like kind of a pretty basic question. I think you injected a virus in the region and then you activated in the same region. So in a way, there are other regions where the same region projects to. So would you be like protruding?
MICHAEL HALASSA: It's possible, right. It's possible, yeah. It's possible. We can't rule that out.
AUDIENCE: Have you seen anything like syn-chains in the RNN?
MICHAEL HALASSA: Oh, in the RNN, in the model? So we haven't optimized the RNN to do this.
AUDIENCE: Did it fit to neural data?
MICHAEL HALASSA: No, not at all. It's just fit to the task. It's just got to perform the task.
AUDIENCE: So is there a waiting period for the RNN?
MICHAEL HALASSA: Yeah, yeah, yeah. There's a delay period, yeah.
AUDIENCE: Is there a difference or any kind of categorical difference in the behavior dependent on the rule in the RNN?
MICHAEL HALASSA: No, the performance was equivalent on both the auditory and visual rule. So about the sequence stuff--
AUDIENCE: Yeah. Yeah, I guess I wanted to understand if there's some kind of sequence in the RNN that's different between the different tasks.
MICHAEL HALASSA: So we don't actually see sequences in-- not only in the RNN. We don't see sequences outside of mice. So in the shrew, which looks like this and has a brain that's pretty large. So it's a rat-sized animal, but its brain is a lot larger than a rat, right?
We ask it to do actually a slightly more complicated task, where it's not only-- it's basically-- that goes into mitigating the-- I don't know who asked the question. Did you ask the question about the conflict? Yeah, the conflict was-- yeah, so in a task like this, for example, yes, the stimuli are in conflict. But the animal is separately reporting whether it's doing a visual selection or an auditory selection, right? So it cannot simply go to the opposite target on an auditory trial. So that allows us to basically disassociate the sources of error that the animal makes.
And shrews do this much, much, much better than mice. I mean, mice sort of guess on the rule a lot. Shrews sort of know what they're doing. And what we see in the neural data is, yes, in the PFC-- the PFC cares about meaning. The MD cares about context. That's what we get from decoding. But we don't see these sparse sequences at all. What we see are these bumps, basically, much more dynamic representations over the delay period that look much more like what people have seen in primates.
AUDIENCE: Like in the motor cortex.
MICHAEL HALASSA: Like in motor cortex Like the Churchland stuff. Or in Merdad's lab, for example, that's what we see. So yeah.
AUDIENCE: I'm just trying to learn as much-- so dynamic can mean that its broader and more invariant or dynamic means--
MICHAEL HALASSA: Dynamic means it's changing more over time. Yeah. So that's what we see in the shrew, and we don't think that-- we never see the sparse synfire chain-like thing. So I don't know if there is like a ton to be learned from focusing on that. It's possible that it's interesting, and we can learn a lot of comparative stuff by looking at shrews and mice and how they've come up to solve the task in different ways. And maybe this is a way to solve more things or hold more items in working memory because you can slot them into these different-- I don't know.
I don't know. It's just hand waving at this point, but it's cool. I mean, we've got a lot of things going on in the lab that I'm really excited about. I hope that I've sort of relayed some of that excitement to you guys today.
have an interactive transcript feature enabled, which appears below the video when playing. Viewers can search for keywords in the video or click on any word in the transcript to jump to that point in the video. When searching, a dark bar with white vertical lines appears below the video frame. Each white line is an occurrence of the searched term and can be clicked on to jump to that spot in the video.
LEARNING