[00:00:03] Speaker A: When I was thinking about learning and memory, I really wanted to be able to find the right cells that were, you know, encoding the memory. I was really interested in being able to, you know, sift through all the neurons in a particular. Even in a small brain area, which could still be 100,000 neurons, right?
[00:00:23] Speaker B: Yeah.
[00:00:24] Speaker A: To be able to find the right ones. And that's when I made the FOSS GFP transgenic mouse. This was in 2000 actually, that I built the construct. And the idea was to use activity dependent gene expression, which we knew only happened in a tiny fraction of cells, to point me towards the cells that would be encoding a memory.
What we're doing is we're looking at synapses that are changing and using animal behavior to change those synapses. And then from that we can make some inference about what is it that the cortex is actually trying to compute.
I have to keep reminding myself that I cannot be an expert in the entire brain. And I'm going to focus on primary sensory cortex because in fact, we don't even know what it's doing. We don't know how it's transforming inputs and we don't know what that all that recurrent connectivity is doing.
So, you know, my new goal, like what would I feel like I did it is I want to understand what sensory cortex is computing.
[00:01:41] Speaker B: This is brain inspired, powered by the transmitter. Hey everyone, I'm Paul Middlebrooks and that was Alison Barth. So Alison runs the Barth lab at Carnegie Mellon University where they use learning experiments in mice to try to figure out how to how the cortex works. As you may know, the brain in general, but also the cortex itself is made up of a large variety of neuron cell types with different activity properties and so on. And Allison is down in the weeds in those neuronal cell types. She has the job of identifying those different cell types in sensory cortex and seeing how they change when animals learn to associate rewards with sensory stimulation.
So unlike many of the guests on this podcast in recent years, even who take a much more zoomed out view and look at how populations of neurons carry out some function. For example, Allison is happiest down at the cellular level.
So we talk about her work, why she prefers to work at that scale, and a variety of related topics. Thanks as always to my Patreon supporters. You can learn how to support the show at braininspired. Co. You can also learn more about Alison and her work in the show
[email protected] podcast243. I hope you enjoy our conversation When I began my little nerve, my little neuroscience journey, I was a technician before I went to graduate school.
And this is a long time ago, and I was taking out mouse early developing visual cortex. And the lab was all about whole cell patch clamping to individual cells and studying plasticity during learning and adding drugs and recording LFPs and seeing the responses of the synaptic strengths, essentially inferring the synaptic strength response over time through these protocols. And that was a long time ago. And that actually, it got me really excited about neuroscience. But I also thought, like, oh, my God, what? Like, how are we going to ever understand what's actually going on? Like, we're, we're, you know, we're patch clamping, you know, one cell at a time, and looking at the responses, and the brain's just like huge, like, really messy, complex thing. Then I went to graduate school and I was studying extracellular spiking, right? Put the electrode down and you just listen for the spikes. And. And I thought, oh, this is. And it was really exciting. And I thought, oh, my God, how are we going to understand, you know, listening to one cell at a time, one spiking cell at a time, all of the messy stuff? So, first of all, just how did you get interested in neuroscience back in the day? And then I want sort of this broad picture of my worries, which are still like, how are we going to understand it all?
[00:04:35] Speaker A: Well, you know, when I went to grad school, I actually had a very specific plan about what I was going to do.
And I was really interested in human evolution. And in fact, I actually applied to grad school in anthropology and sort of molecular evolution and anthropology.
And I got in. Right. It was. There was a project I was really interested in working on. I was at Harvard, it was. The woman's name was Mary Ellen Ruvalo. And the project was to try to look at genomic differences between chimps and humans. And I mean, in retrospect, it was a good thing that I didn't do that, but because the technology just wasn't good enough to really understand the differences that we might have found.
But I decided it would be better for me to get sort of a more rigorous degree, like in biology. And so then I ended up doing molecular and cell biology at Berkeley. And the person that I wanted to work with there was somebody who'd been looking at mitochondrial evolution, and unfortunately, he'd just been diagnosed with leukemia.
And then I kind of, you know, had like, some other ideas about who I wanted to work with. And Then one of those people, they moved. I mean, it was, you know, a lot of shuffle.
When I moved to Berkeley, I was staying in this house with a bunch of grad students as I was looking for a place to live. And it was right at the time when there was an anatomical discovery about differences in the brains of gay men versus straight men.
And this is, like, a really interesting group of people. And, you know, we'd sit around in this flat. I was crashing and talking about, you know, what does that mean? And sort of biological determinism.
And there were a couple neuroscientists there. A guy named Paul Pavlitis was one of the, you know, my hosts in this. This apartment. And I was just like, wow, you know, the brain's really interesting. So I. In fact, I just found a notebook that I had prepared when I was looking at grad schools and writing applications, and I listed a bunch of different disciplines.
And next to neuroscience, I wrote.
I was not interested in neuroscience, but this. This discovery. And then this, like, really interesting group of people were like, okay, this is a really. This would be a cool area to go into.
I ended up the lab that I joined was actually looking at DNA recombination in the brain. So this is a long story, and I'll just cut it short, but basically, you know, in the immune system, there's somatic DNA recombination.
This guy, Hitoshi Sakano, had discovered evidence for somatic DNA recombination in the brain. And, you know, we were just getting the sense that there were not that many genes, and yet there's so much diversity in neurons. And so gene recombination, like in the immune system, which is the capability of making, you know, millions and millions of different types of antibodies, maybe that kind of recombination would be.
Would be useful in the brain. So I joined this lab. It became kind of a problem because he was really an immunologist. And I got more and more interested in the brain. And eventually I moved to an olfaction lab, which, you know, the olfactory receptor genes were a really good candidate for somatic DNA recombination to sort of regulate, you know, a thousand different genes. Maybe you kind of swapped them into an active promoter site.
And. And so that's what I did for my PhD. I was a PhD student student with John Nye, and, you know, as many PhD students as happens to many PhD students, I'm sure yourself, you know, by the end of your PhD, you're like, wow, I'm so tired of doing These experiments. And I, I was doing in situ hybridizations, looking at gene expression and the olfactory system of the zebrafish.
And I got.
[00:08:21] Speaker B: Is this something where you would do the same thing over day in and day out and like getting, building up the data?
[00:08:25] Speaker A: Yeah, yeah, yeah. But, but the, you know, the experiments were like, you do the experiment and then you had to dip the slides in a photographic emulsion and then you had to wait for three months. So you really better get it right because three months later you were going to be just figuring out if you just made some fundamental error in the experimental design, and then you'd have to do it again and wait another three months for the answer, not only again,
[00:08:49] Speaker B: but you have to do it for like two weeks straight or more. You know, like. And then wait.
[00:08:53] Speaker A: Yeah, it was, it was really painful.
And so, and, and you know, just the dipping of slides, you know, I had to sit in the dark room, zero light, right? So you feel everything.
Dip these slides into emulsion and then let them dry and then pack them in little boxes and tape around them and wrap them in foil. And anyway, it was like, I was like, I'm never doing this again.
And I had taught a little discussion course, seminar course for Berkeley undergrads about learning and memory and cortical plasticity. And I really liked this whole framework of thinking about learning. First you could move from the molecular to the sort of behavioral, right? There was a through line in how you could think about this.
And we're just beginning to use molecular tools to manipulate neurons, to test specific hypotheses about how kind of low level phenomenon might be related to sort of higher level behavioral phenomenon. So that's, that's sort of how I got to learning in memory.
But, but because of my background in cell and molecular biology, it was really like about synapses, right? The synapses were kind of the workhorse of that, of those long lasting changes in brain function.
And so that's why I was, you know, looking at synaptic plasticity, sort of, you know, more cellular labs for my postdoc.
[00:10:17] Speaker B: That's kind of a hard pivot. I mean, so I think that's a really smart thing to do. Sorry, we're already off like on a tangent, but I find this stuff interesting.
I sort of made, I think, a mistake by doing kind of the same thing during my postdoc as I did during my PhD. And as you were saying, I was already feeling kind of burnt out in that arena and I knew that I needed it. Different skills, right? Essentially to just become a broader, skilled scientist. And it sounds like that's what you actually did.
[00:10:47] Speaker A: You know, I get bored easily, which is, you know, it's kind of a problem. But, you know, it's a personality trait. So, yeah, I didn't want to do the same thing. I chose something totally different. And I did watch some of my peers, you know, move to a new lab where they would, like, you know, do structural biology on a different protein. And, you know, in a way, they. It was a steep learning curve for me. You know, I hadn't done any animal surgery. I hadn't really worked with mice before, and so I had to, you know, learn how to do all those things.
But on the other hand, you know, I brought a lot of thinking about a whole different system and that expertise in that different system.
[00:11:25] Speaker B: Yeah.
[00:11:26] Speaker A: To a different problem. Which then is super fruitful.
[00:11:29] Speaker B: Yeah. Well, maybe fast forward and tell us broadly, I guess, how you would summarize what you do these days, because we'll get into this more. But you kind of came in, or here's how it goes in my head is that you had been interesting, interested in how learning works. And I could read a direct quote, you know, from your papers, because that's kind of shifted into an interest in using learning to study how the cortex works. I mean, is that a fair assessment?
[00:11:56] Speaker A: Yeah, yeah. I mean, you know, I think there's like, a sort of a clear story about, like, well, how was I going to approach this? And, you know, I was really convinced. And actually, it's related to what I was doing in olfaction. I remember when I was a postdoc, my advisor, Kevin Fox, asked me, you know, because he thought olfaction is so different from, you know.
[00:12:17] Speaker B: Yeah.
[00:12:18] Speaker A: Learning memory. How is that? Why would I even think to move in that direction?
And I was like, you know, in the olfactory system, there are a lot of different cell types. Right? There's the cinnamon salad, the apple sal cell, and the, you know, formaldehyde cell. Right.
There's really very discrete receptive fields for hundreds or thousands of different odors. And so it was very easy for me to think about the brain also being functionally subdivided into specialized cells that had very distinct roles. Okay. And so when I was thinking about learning and memory, I really wanted to be able to find the right cells that were, you know, encoding the memory.
And so as a postdoc, when I was a postdoc with Rob Melenka, I was really interested in being able to, you know, sift through all the neurons in A particular. Even in a small brain area, which could still be 100,000 neurons. Right?
[00:13:14] Speaker B: Yeah.
[00:13:14] Speaker A: To be able to find the right ones. And that's when I made the FOS GFP transgenic mouse. This was in 2000, actually, that I built the construct. And the idea was to use activity dependent gene expression, which we knew only happened in a tiny fraction of cells, to point me towards the cells that would be encoding a memory. Okay. So that. And that actually was the first time. And I wanted to do that so I could patch those cells. I wanted to make that leap between here's some synapses that have changed, here's some cells that are the target for that, and then here's some behavior, here's some memory, here's some representation.
[00:13:56] Speaker B: Then you have a full story.
[00:13:58] Speaker A: Yeah, then I had a full story. I wanted the whole, you know, like from the molecular changes at the synapse to the behavioral change in the animal.
[00:14:08] Speaker B: So I mean, we. There's a lot of different things I could ask right now, but I mean, how has your.
So that's a very kind of ball and not ball and chain like billiard ball slash domino. Very mechanistic. I mean, are you still. Do you still see the brain as functioning in that way? Because I've been in your talks, I've seen the matrix of cell types and connectivity profiles and stuff, and it's overwhelming. And I just think like, well, how can you even make sense of that in a reductionistic, mechanistic sort of story? Because that's not a paragraph, that's a tome.
[00:14:43] Speaker A: Yeah, well, okay, so one sort of like change in my mental framework about this problem is that I am. And I don't want to offend anybody here, but I just. I mean, everybody has their own sort of interest filter, you know, like some people really need to know what the phosphorylation site is or how the protein, you know, sort of undergoes some structural transformation.
And right now I'm just not that interested in the molecules that are doing those things. I know there are molecules and I just went to a Gordon conference about synapses where there was a lot of molecular discussion. I'm glad to know about.
Helps my imagination.
But I don't necessarily need to figure that out for the synapses that I'm looking at.
But I do think that synaptic plasticity is a great way to point us towards the right circuits, the right brain areas that are doing some interesting computation.
So all of a sudden I'm not studying how do calcium permeable Ampa receptors mediate calcium signaling at the synapse to augment strength. No, I don't think I need to know that. What we're doing is we're looking at synapses that are changing and using animal behavior to change those synapses. And then from that we can make some inference about what is it that the cortex is actually trying to compute. So synaptic plasticity now is like a fact, Right? It's a method to then focus our hypothesis about what this circuit is trying to accomplish.
[00:16:25] Speaker B: Well, yeah, so I was going to ask you, like, at what level of abstraction you're most comfortable with, and you were just speaking to that. But because it's fine to abstract away from the molecular processes going on and the membranes of the neurons, et cetera, and you don't need to worry about that. So you can kind of back off on that. So you're kind of in this level of. You're still sifting through and looking for the engram type cells or like the important cells among this swath of confusion and complexity.
[00:16:53] Speaker A: Yeah, but let me just say, I mean, you know, we're wrong all the time. And I, I think.
[00:16:59] Speaker B: Speak for yourself.
[00:17:00] Speaker A: Okay, I'm wrong all the time. And all I can say is, you know, I think it's really important to have a strong hypothesis and then to be able to say, no, that was not right. So, you know, I, I should be proud of being wrong all the time because if I kept being right, then I, I think I'm not pushing hard enough on, you know, what I, what my mental model is.
I mean, I'm not sure that there are engram cells in sensory cortex.
And I think that idea, that, and it was reasonable to look there.
[00:17:31] Speaker B: I tried to find them.
[00:17:32] Speaker A: Yeah, I tried to find them, but I think we're seeing other things going on there.
There must be engram cells, right? I mean, the brain clearly is encoding things, and it's not that every neuron is encoding things, so it's a fraction of them, it's a subset. So they must be somewhere. I'm just not sure that they're in sensory cortex.
[00:17:55] Speaker B: Well, let's, let's orient the listeners to, you know, where you are, what the animal is, and kind of what you're doing to, to undergird the rest of our conversation. So.
[00:18:06] Speaker A: Yeah, and I'm glad you sort of started off with like this whole long trajectory of like, how I became interested in this and, you know, why I ended up here.
So, you know, when we made the fosgp transgenic mouse. The goal was to identify activated cells and then to see that synapses onto those cells were more plastic than synapses onto other cells. And so, you know, we had a really specific protocol in mind. We were going to look in sensory cortex where we knew there was a lot of plasticity and we were going to then identify those particular neurons and record from them in sensory cortex.
At the time, it wasn't really learning related plasticity and it's hard to remember. But you know, 20 years ago, everybody's just trying to see that something like LTP was actually happening in the brain.
[00:18:55] Speaker B: That's what I began. That's where I, as a tech people kind of felt like, yes, it is happening, but we have to sort of prove it and how long does it last and is it important for development? So yeah, yeah, yeah, that was 20 years ago.
[00:19:08] Speaker A: Yeah, right. And I mean it took a while to find evidence that something like LTP was really, you know, the same thing that was going on in the brain during normal behavior.
So, so we knew that if you altered sensory input, you could change receptive field properties of neurons.
And so that was a perfectly reasonable way to look for an engram. But that's not really a memory, right? I mean if I, you know, have some binocular deprivation or if I remove all but one whisker in the mouse and I'm looking in the whisker representation in the barrel cortex, you know, is that memory or is that just some sort of dynamic reorganization which is not necessarily the same as ah, when I see this, I want to do this. Or you know, it's not causal inference for sure, which is like, that's really interesting. Right?
It is plasticity, but I'm not exactly sure that it was the kind of memory anyway that I wanted to study.
So we spent a lot of time manipulating sensory inputs to drive this plasticity. Right. That gets at the goal of relating some behavioral kind of animal experience to then synaptic changes.
But after a while it just wasn't really very satisfying. Right. Because I really wanted to study learning memory.
I actually, because there's. So yeah, I'm interested. Obviously, like this is, you know, my work and so I'm really, you know, I, I could talk about it for hours. But I think you asked me more specific question and I think I've gone off on like a.
[00:20:41] Speaker B: No, no, no, I'll.
What I was wanting to lead up to is just to orient the listeners like you know, where the, the sort of ground truth stuff like where you're recording why you're recording there. And then like, what?
Even like, you know, sort of the, the overarching sort of experimental protocol that you're doing to try to find answers. Yeah.
[00:21:02] Speaker A: Okay. So when I started off, we were using whole cell patch clamp recording in acute brain slices.
And you know, for somebody who's not a neuroscientist, the idea that you can take a tiny little tissue biopsy and the neurons behave like, you know, kind of neurons. Right. They, they're pretty normal. I mean, for a while, at least for hours, though, is amazing. And when I try to present this to my biology department and I'm like, here are my exciting results. And they're like, wait, the tissue's still alive? And you're like, yeah, yeah, yeah, it's fine. But anyway, so the thing about that preparation is that it gives you extraordinary control. You really know what you're looking at. Right. So, you know, you can control the temperature, you control the ion concentrations, you can wash on drugs and wash them off again.
So you have really great resolution about what you're looking at. But of course, you don't have animal behavior and you also don't have any long distance circuitry where you have a really local circuit.
[00:22:01] Speaker B: Yeah.
[00:22:01] Speaker A: So.
[00:22:02] Speaker B: And in fact, you've messed a lot of, you potentially have messed a lot of stuff up because we know there are long range connections and take those away and then, I don't know, can you even say that you have a healthy preparation even though the neurons are responding, et cetera? Yeah.
[00:22:15] Speaker A: On the other hand, you know, I mean, individuals are super heterogeneous. Right. So I don't have brain state. The slice isn't sleeping, the slice isn't angry, the slice. Slice isn't hungry. Right. So you can really, you know, you control all of these variables and when you find something, you're like, you nailed it. It's really there. Right.
But on the other hand, like, you don't know how it's functioning while the animal is behaving. And actually, I'm going to make a. Maybe it's a controversial point, or maybe everybody's kind of coming to this conclusion. You know, people are really enamored with recording from neurons in awake behaving animals. And I get it. It's really exciting to think that you have a little window into what an animal might be experiencing or thinking. Right. It's so exciting.
But I'm just going to make the argument that we have learned less from that than we might have hoped. Right. There's okay, this is actually maybe a little too close to home for you, Paul, so I apologize.
[00:23:15] Speaker B: You're not, I cannot be offended, so, but my listeners can, I'm sure.
[00:23:20] Speaker A: I, I, I mean, you know, we're recording from thousands and thousands of neurons, right. And, and we do see things like we can decode, you know, in different brain areas and say, oh, there's a signal there.
But it's, it's amazing how confusing the data have actually been.
And so, I mean, you know, in the slice, actually, even though it's totally outside of the behaving animal, and we're really looking at, you know, kind of a static represent a static change in some circuit. Right. If it lasts by the time you make this little tissue biopsy, then, you know, it's, it's certainly. They are in vivo. Right. But, you know, it, there's something really solid about those observations, and I find it for me anyway, that, you know, we do see big changes, even in fixed tissue, even in acute brain slices.
And we're moving now to the awake behaving animal. But I'm glad we didn't start off that way because I actually think it has been less informative than everybody hoped. I mean, everybody hoped if we could only record the activity of every neuron, like all at once, will understand everything.
[00:24:38] Speaker B: That's a hope. Yeah, I don't know that that was because, you know, that used to be a favorite question of mine. And I've heard many other people say, like what? You know, how many neurons would you want to record? Like, what if we could record all the neurons? You know? And everyone says, yes, I hope that would help. And, but then what would you do with it? And everyone says, I don't know.
Yeah, so,
[00:25:00] Speaker A: so I mean, you know, you have to kind of walk into something with a hypothesis. And I think the, you know, recording from lots of cells can. There are all sorts of caveats with that. Right. Because there may be cells that are relatively silent but very important, and you also have trouble differentiating, you know, different nodes of the network that might be computationally super important.
So we were working in the slice, we could see all sorts of interesting things happening in the slice, which means that they were happening in the animal.
[00:25:35] Speaker B: So best. I'm glad that people like you are doing this kind of work. Right. I think of you as sort of, and people who are doing this kind of work is sort of in the trenches doing this still very difficult work, you know, that can't be automated. You know, much of it cannot be automated. We'll talk about your automated behavioral training system.
But what would the, what would the, what's the grand goal? What would, what would make your career be the best it could have been? Right.
You would find three different cell types that change their synaptic properties in different ways and can tell a story among the 30,000 or whatever. Or like, what would that look like? What, what does success look like?
[00:26:22] Speaker A: Yeah, I mean, you know, it's funny, I was at a family sort of birthday party with, you know, various people who I hadn't met before, like my cousin's parents, right. And so this guy is like 80 years old and you know, he, we're making conversation. He was very nice, gentleman and he's, he was a judge and he sort of got right to the point. I appreciate it. And he's like, so how's your work going to win the Nobel Prize?
[00:26:49] Speaker B: Oh God. I don't mean that.
[00:26:51] Speaker A: Hold on. I was like, fair enough. Right, let's see if I can tackle that question. Okay, so I mean, obviously, you know, you want your work to be important, right? You want, you're headed towards some sort of goal. Initially I wanted to, you know, find the end groom and. But I told you, I've just. We've been wrong about so many things. I just think that the idea that, you know, some activated cells will represent the memory in primary sensory cortex, I think it's, it's probably totally wrong.
So I mean, my focus is like narrower and narrower and narrower because, you know, maybe the memory is just not in sensory cortex. Fine. And I have to keep reminding myself that I cannot be an expert in the entire brain and I'm going to focus on primary sensory cortex because in fact, we don't even know what it's doing. We don't know how it's transforming inputs and we don't know what that all that recurrent connectivity is doing.
So, you know, my new goal, like, what would I feel like I did it is I want to understand what sensory cortex is computing.
[00:28:03] Speaker B: Right. But okay, so then given what you're doing now, like, would that.
Given the high diversity of neuron cell types. Right. So everything, you know, people do in the molecular world is like you target a specific cell type and then it expresses some protein that you can then like record when it's active and it.
What?
[00:28:25] Speaker A: I mean, sort of.
[00:28:26] Speaker B: Yeah, okay, this is very broad picture, right? Yeah, I mean, there's lots of other techniques as well.
But anyway, so then, then you're looking at a specific cell type and, and sort of my point Was there are many, many cell types. And I. I do want to ask you, like, how many different cell types do we need to care about? Like what? Because this is, like, the level of abstraction that I'm comfortable leaving. You know, I'm stepping back just for my own interests and sanity. Right.
But some, you know, so you find, like, three different cell types, and maybe it's not the engram, but you want to be able to tell a story about this particular cell type and how it's related to associative learning or what it's doing in the cortical column.
[00:29:16] Speaker A: Okay. I guess what I'm looking for is a signal.
I'm looking for cells that are.
That care about something that's relevant to learning.
And I am interested in how they're computing what they're computing.
But I actually think that we haven't carefully defined what might be computed, even in primary sensory cortex.
[00:29:43] Speaker B: I will say it's unfair. You're talking to someone who doesn't think it computes anything. But that's a different philosophical sort of. Yeah.
[00:29:49] Speaker A: You know, but I mean, one of the things that we discovered, you know, over the past few years is that sensory cortex is really sensitive to contingency.
It's really sensitive to causal inference and. And to sort of errors. Right. And so we can do things like manipulate the. The predictability of stimulus A with outcome B.
And sensory cortex is, like, super sensitive to it. It's responding to that. It's adapting to that.
And so, I mean, why? Right.
Because I think that that's important to sensory cortex. Right. I mean, I'm pretty sure if I was recording from the hypothalamus, I wouldn't see that neurons in the hypothalamus are sensitive to stimulus reward contingencies. I don't think neurons in the hypothalamus are building a hypothesis or testing a hypothesis about the accuracy of that stimulus in predicting an outcome. Right. And I think sensory cortex is. Right. We see signal in discrete cell types. So part of the reason I'm so interested in cell types is because, you know, I had done some unit recording when I was a postdoc with Kevin Fox, and you basically put your electrode in and you just hope to find a neuron that has a response, and then you can classify that, you know, monitor that cell's response to some sensory input, and, you know, you can get some spikes. But there's all sorts of caveats about that. If a unit doesn't respond, you don't record from it. Right, Right. So you're already biased towards responding from units that are active, right?
[00:31:30] Speaker B: Well, yeah. And even worse, like in my case, you know, if, if the, if a, you got a good neuron, you can hear some spikes, but if it doesn't sound like it's related to what's going on, you bypass it. Because you want a neuron that might be tuned property of the stimulus or task that you've trained your organism on.
[00:31:48] Speaker A: I mean, I remember one of my colleagues, a monkey neurophysiologist, talking about the tenure cell. And you know, they were doing some experiments and they had a hypothesis and they found a neuron and it did exactly what they had hypothesized.
[00:32:01] Speaker B: Yeah, right.
[00:32:02] Speaker A: So the tenure cell. And you know, I, I, I certainly believe that that neuron was behaving in that way and then that there must be some brain computation that drives activity of that neuron in that predictable way. So like thumbs up. Right? That's great, but that's not really like what the network is computing. I mean, it's a product of that computation. But I mean, don't we need to account for, you know, all the neurons in that network? Right. Not just the one that tells you that, you know, it has some sensory responses.
[00:32:34] Speaker B: Well, that's sort of my question for you. Like in terms of thinking, let's just be super simple and say it's like one cell type that you're going to end up like really focusing in, on or whatever as being important in the story of how the cortex works or something like that. Then it leaves open the question what are the other, what do we have to 20,000 cell types now?
[00:32:53] Speaker A: Okay, so I mean, first of all, as somebody who came from a very reductionist background, I do think it matters that you know what you're looking at.
And so I'm just looking for signal.
And right now we're basically, it's like Hubel and Weasel, we're mapping receptive fields. But in this case, the receptive field is not some feature, some sensory feature. The receptive field is.
Can the animal come up with an account, a causal account for some input and some outcome. Okay. And so the receptive field for this neuron is what is the internal hypothesis? The internal model of the, of, of the animal. Right. Of the brain.
So I'm using cell type specific measurements to help me understand what this might be computing. I care about the cell types because I don't want to get confused by looking at 100 cells that are doing other things or not spiking enough to be able to reveal any computation or are, you know, have, have some response property that is, you know, maybe it is a feature specific response property, but is only, you know, barely modulated by something like stimulus and reward or stimulus and outcome contingencies.
So it's just a way of making my question really precise.
[00:34:24] Speaker B: Let's. Okay, so I hate to do it again, but we're going to back up for a second because I want to just ask you about your burgeoning love for predictive processing accounts. But, but I want, but I want people to have in mind like what's actually going on like these days with, you know, your experimental setup essentially. So it's a mouse, it's the, there's whisker, air puffs and rewards, et cetera, contingencies of the task. So you don't have to go into the details, but just sort of broadly what's going on with the pairing of air puffs and, you know.
[00:34:54] Speaker A: Sure.
So, okay, but just a little bit of history here. You know, we move from like, let's see how synapses can be changed from like passive sensory experience to actually trying to have a learning paradigm. So what we did was super simple. We had a sensory stimulus. In our case, it was a multi whisker stimulus delivered to one side of the animal's face. And that was followed then by a water reward.
And we, you know, the intention was at the beginning really to be able to do this in a super comprehensive way. So we needed to be able to train the animals in a high throughput, what manner? Right. So that was going to be home cage training. No water deprivation, no other cues. Let the animal learn this. And that's really, you know, what they do all the time. You know, the mouse in my kitchen is trying to figure out, you know, where there was something to eat. And it figured out that the chocolate is on the shelf in the pantry. Right. You know, it's all about, you know, causal inference and then, you know, sort of driving behavior.
So we set this up and we discovered that we could manipulate how that stimulus could predict the outcome, the reward, and then that would actually change the way that synapses in the brain were modified.
So we were really pleased with this. We two conditions. One is where the stimulus predicts the reward 100% of the time. And the other condition is where the stimulus doesn't predict the reward, that sometimes there's a stimulus and no reward, sometimes there's a reward and no stimulus and all combinations of that. And we would look at a particular input pathway and we could see that it was changing when the stimulus was predicting the reward with high accuracy. And it wasn't changing when the stimulus. Same number of stimuli. Right. But when that predictive power was absent.
Okay. So now what we've been doing is across lots of different cell types across.
And even more kind of stimulus, reward contingency, manipulations, mapping how different neurons are changing.
[00:37:02] Speaker B: Okay, that's great. And I don't know if you want to say something about. Well, you already did.
You said something about the automated training environment.
[00:37:11] Speaker A: Yeah. Let me just say, though, I mean, you know, as a.
Maybe this is because as a grad student, I was doing this incredibly cumbersome technique that would take months for any results to emerge. You know, I really.
And I think this is a good thing for scientists, but, you know, I'm impatient, and I want to figure out how to do something faster. And so the automated training was a way for us to move away from, you know, like, someone going to the cage every day and then, you know, training the animal and then putting it back. And. And I wanted to do it hands off, and I wanted to be able to have people in my lab spend their time testing our hypotheses, not training mice. So at this point, we've probably trained 1500 mice and of many different strains and looking at many different pathways. But it's been super informative, and we never would have been able to learn what we are learning if we had one or two animals in each group.
[00:38:05] Speaker B: Mm. Just because the statistical power wouldn't be high enough.
[00:38:08] Speaker A: Yeah. Right. Because there's heterogeneity across animals. Yeah.
[00:38:12] Speaker B: So, okay, how does cortex work?
Have you figured it out?
[00:38:15] Speaker A: Yeah.
Well, okay, let me just say I was looking for the engram I was hoping to find.
[00:38:23] Speaker B: Sorry, I'll just say, like, because we've used the term. So engram is the idea that a particular set of cells are sort of hardwired, that, like. That the. The memory trace is. Is within. Like, whenever that group of cells is active, that is the trace of a particular memory in the brain. And those. Those are the cells. That is the engram. Okay. Just wanted to define it. Yeah.
[00:38:50] Speaker A: That you could ablate them and then the memory would be gone. Right. Sort of like eternal sunshine of the spotless mind. Right. Like we could control whether you could remember or not or severance. Right. Like, we could turn them off, and you just have. It's the. It's gone white. Right.
So, you know, in. I think what I, like, want to convey is that I don't know if that's there in sensory cortex. And you know, it was reasonable to look, we know there's lots of plasticity there, but I just am not.
I mean, we know certain pathways are augmented. We know, and we can even find those very specific input and target specific synapses. But those synaptic changes generally don't last that long.
And when you're imaging or recording in vivo, you also don't see really long term changes.
And what people have said is that, oh yeah, but you just had the wrong cells. If you could find the right cells, you'll see long lasting changes that correspond to, to some memory.
And that might be true. You can't argue from the absence of evidence. Right. It doesn't mean they're not there. But we thought these FOSS expressing cells would be the right way to tackle it.
And it just doesn't seem like they are the ones that capture this, like specially capture the synaptic change. In fact, inputs onto those cells often get weaker.
So I don't, you know, I don't know. I mean, I only can say from the data that we've collected, you know, what is a reasonable hypothesis. But right now I think that sensory cortex is not the place for us to look for those cells that encode a very specific memory. Right.
[00:40:37] Speaker B: So yeah, yeah. So back to my unfair question. And so how does it work? How does cortex work? That's right. Like how do you, how do you go from what you just said that there are lasting changes, but they are in layer 2, 3. I don't think you said layer 2, 3, but I think that's where you were talking about.
But, and they do care, but they're a very small proportion of the population. I'm sorry to like keep hammering on about this, but I want to get like, like I wouldn't know how to think about how cortex broadly works based on that very particular finding.
[00:41:14] Speaker A: Right. Okay. So you mentioned this at the beginning that you know, I've been interested in learning and you know, we were going to study it in somatosensory cortex and you know, I was looking for where's the trace, where's the memory, you know, where is learning happening in sensory cortex? And I kind of leaving that behind a little bit because that data just are not really encouraging. I, we haven't found that kind of representation of learning in a discrete set of cells. And I feel like what we're doing now is we're using learning as a way to probe what sensory cortex does. So that is like a totally different framework. Now we're using learning. Learning is the. Is. Is the, you know, that the hammer by which is the tool. Yeah. That's like how we're kind of opening up the box because we see that sensory cortex really cares. Okay.
So, you know, one of the nice things about working in a more reduced prep is that you can kind of look at whatever you want to look at. Right. We're not restricted to this layer or to cells that are really common so that we can get enough, you know, neurons to be able to draw a strong conclusion. And so we can look across the whole cortical column from layer one.
Right. All the way down to the very bottom of the column in layer six, we can look at which neurons are changing and we see very discrete things happening. Almost everything that has been really significant is happening in superficial layers in layer two, three.
Layer one, maybe in layer two, three. Yeah.
[00:42:52] Speaker B: What does that tell you?
[00:42:56] Speaker A: It's interesting because I think that layer 2, 3 is its own sort of computational unit.
And I think that, you know, layer four is a. Layer five is kind of an output. And I think that layer five is maybe just reporting what layer two, three is doing. Okay. I don't want to. I don't want to have this set in stone because like I said, you know, we're wrong about everything all the time. But, you know, we just don't see super long lasting changes in layer five neurons. And I think layer five is sending a signal somewhere else. And actually we don't see a lot of changes, a lot of durable changes in layer four either. So layer four is often seen as the input layer of the cortex. You know, there's some modifications to that hypothesis, but sensory information arise in layer four, and in somatosensory cortex, fast, it's like six to nine milliseconds. You can get a spike in nine milliseconds. Right.
[00:43:52] Speaker B: So those are direct projections that not going through the thalamus.
[00:43:55] Speaker A: Yeah, no, they go through the thalamus. Right. But like, you have three milliseconds, you know, in the trigeminal, you have six milliseconds in the thalamus, you have nine milliseconds in layer four. Right. There's not a lot of processing. Right. Those first spikes are just telling you there was something in the world and the brain is now receiving that. Okay.
And that information also goes to layer two, three. It also sort of goes to layer five. But I don't think there's a lot of plasticity in layer 4, not long lasting plasticity. And we don't see long lasting plasticity in layer 5, but we do see it in superficial layers.
And so that suggests to me that layer 2, 3 may be sort of an associative network that is calculating things in a way that is important for delivering that signal then to downstream rain areas.
[00:44:48] Speaker B: And so then layer 5 would be doing very minimal, sort of, it's sort of collecting all the information to send along.
[00:44:54] Speaker A: Yeah. So I said I'm wrong about everything, but it's a strong hypothesis and it's testable. Right. So, yeah, I'm going to say layer five. I think it's just reporting what layer two, three did. But the brains of the operation are in layer 2, 3.
[00:45:09] Speaker B: The brains of the operation.
Is that an intentional joke? Yeah. Okay. Okay, good.
But aren't you, I believe you've become more and more interested in the predictive processing account of brains in general, but also, I guess, early sensory cortex.
[00:45:28] Speaker A: Yeah, I mean, and some of this has to do with time that I spent at UCL in London and it was during the pandemic, it was my sabbatical and so I had a lot of time to think and it was fine. It was fine. I had a lot of good conversations with Sonia Hofer and Tom Mercek's Logel and Hampstead Heath, you know, and Adam Packer, you know, walking around six feet apart talking about science.
But in a way that that framework, it revises the sort of feature representation framework that, you know, I think people who study sensory cortex have been kind of wedded to for the past 50 years. Right. And rightly so. There definitely are feature representations in sensory cortex, but there's a lot of cells that don't have feature representations.
And then, you know, when people look for plasticity of a given feature during learning, you know, it's just hard to see it. It's not really obvious.
And so, you know, like, you know, I don't want to be stupid about this. Like, if we don't see it, maybe we should look at something else.
And that's why the predictive processing framework has been interesting to me because it kind of gives me another way to consider what might be represented in sensory cortex.
[00:46:57] Speaker B: So couldn't it be both? Right. You do have some stimulus selectivity and you could also have a predictive processing account and they could sort of be meshed in the same cortical column. But then what would the answer be to what the cortical column is doing or what cortex is doing? Right. If it's a little bit of This, a little bit of that. How do we tell that story then?
[00:47:18] Speaker A: Okay, so I, I think that, you know, one of the problems in engineered systems, like if you have a sensor, an engineered system, is just how much data you collect. So let's imagine you have a security camera in your convenience store, right? And the camera's always on, right? But that's a lot of data and you're going to overwrite it, right? Because you can't store all the data indefinitely. Right. And in fact, our brains are receiving like enormous amounts of sensory inputs all the time, right? It's a lot, and energetically, it's a lot to be encoding and storing all that. So one thing I like about the predictive processing framework is it's a compression issue, right? You can basically get rid of a lot of stuff when it doesn't seem relevant, right.
And so the predictive processing doesn't really say that. But, but I, I like the account that you're really going to be sensitive to things that are change, changing, right? So I mean, this is what we see in the frog, right? The frog is basically, or, or, you know, remember in Jurassic Park? I mean, maybe that's too old and most people haven't even seen that anymore. But right, if, if T. Rex is coming at you and you're totally still, right.
It doesn't see you, right. It's sensitive to movement, right? Because that's something that's changing.
And, and it's interesting when things are changing because it might be delicious, right? You might want to eat it.
[00:48:45] Speaker B: It works for cats too, by the way.
[00:48:47] Speaker A: Oh, is that right still?
[00:48:48] Speaker B: Yeah, for me it has, yeah.
[00:48:50] Speaker A: Okay. I don't have a cat. See, I, I, this is a great thing about having pets and children really is you, you get to learn a lot about how the brain works.
[00:48:57] Speaker B: It doesn't work for children. They see you all the time.
[00:49:01] Speaker A: Yeah, exactly. Right. Because they're so smart.
But I think this is a really important challenge for neural networks is how do you keep the stuff that you need and then just not store all the other stuff, right.
And I think for people who have sensory processing deficits, that's why being in an overstimulating environment is overwhelming. It's paralyzing, right? So our brains are trying to get rid of information, right? To just pay attention to the things that you don't know. And I'll just point out, right, like we, it's kind of enjoyable to be in a new place where you notice all sorts of new things. So for example, you know, when you Go to a new country.
I was talking about this with my husband the other day. So he joined me before we were married. I was a postdoc in the UK for a little bit. And he came to, he came to Great Britain and like, there were so many things that were different, right? Like, you know, he walked into our apartment and the refrigerator was the size of a teacup. And he's like, my God, do this. Right? And, and so, and he remembers, like yesterday we were talking about this. He remembers those things that were surprising, right? And so, you know, those are the kinds of, like what, what we like about travel is everything looks new again. You're like, oh, wow. They do it that way, you know, like, oh, they're, you know, the construction workers vests are pink, not yellow. You know, you, you sort of just notice these different things.
And, and that's a great way, that's a great strategy for a brain, right? Notice the things that are different. Don't pay attention to the things that stay the same.
Okay. So I think that is a really compelling account of what sensory cortex might be doing is filtering out all the stuff that you already know about, that you've already seen.
So how do, how do you do that?
Yeah, right. And I mean, this comes down now to like, you know, being able to make predictions. And it also comes down to surprise. I think surprise is fascinating.
Right. And it comes like, it's related to humor. Right. Why is something funny? It's because you're, it's surprising, right. You're not expecting it. And I think we get kind of a thrill. We get a little jolt of something when something is close to our expectations but doesn't meet our expectations.
[00:51:20] Speaker B: But then. So how do you.
It seems like if you are going to. So you need to essentially forget about the stuff that doesn't matter.
Is it a paradox then that that's the stuff that you need to predict? Well, so you don't notice it. Right. So the idea of the predictive brain is, right, if what's coming in is like novel and it doesn't match the prediction of your generative predictive brain, then it encodes the error and that's what gets passed up and that's what you eventually see. But to do that, you actually have to predict the thing that you have that you don't remember. Right?
[00:51:59] Speaker A: Right. So there must be some sort of short term buffer, right, where you're subtracting that off. And I don't understand that at all. I mean, people have been really thoughtful about how you could Use like short term synaptic plasticity. I don't know if that's what it is.
And the experiments we're doing can't really address that. So I think I've kind of just put that on the shelf for a minute. Because a key thing about being a scientist is to figure out the question that you can answer, but then to not get waylaid by all the things you can't answer. But you're right, that's really important.
So you have to have some representation of, right. Some maybe some ongoing signal that might not be changing or is changing really slowly so that you can, you know, basically cancel it out. Right.
[00:52:50] Speaker B: Okay. So you discover some important cells that are changing in layer two, three. And layer two, three is the brains of the coracle column and layer five is the reporter cell. And you could, you've really characterized, well, you know, let's say 10 different types of neurons and their responses to the hammer that you're hitting them with via learning. Right. And you have that like all.
So then do we need to also characterize the 50 other cell types? Like what, where is the satisficing line of the account?
[00:53:24] Speaker A: Right.
[00:53:25] Speaker B: That do we need to understand in a story fashion, in a reductionist, mechanical fashion fashion, what each of those neurons are contributing to and what they're not contributing to, or when they're active, etc.
[00:53:39] Speaker A: I would like to, and I think that we will get insight by doing that. But in the meantime, by using learning as a probe for cortical function, we've discovered some really interesting things.
And that is in particular that the cortex is very sensitive to the predictive accuracy of sensory input to some outcome.
And Right. Just like Hubel and Wiesel were like, oh my God, visual cortex neurons, they like oriented lines.
I am like absolutely thrilled to see that this one neuron subtype that is rare, right. But it really cares about whether that stimulus is predicting some outcome. That means it gets information about that. And it also means it's not even like online. We see some modulation, Right. But it undergoes long lasting modifications when that stimulus and reward are paired. When the stimulus is predictive of the reward and it doesn't do that when they're not contingent.
[00:54:50] Speaker B: But so that eventually that story needs to be fit with the story of all the other cell types. Right. It's just a daunting task. Like one of the reasons why. So going back to your.
It wasn't a criticism necessarily. You know, you're saying, well, I don't know what we've really learned if we've learned as much as we thought we would learn by recording more neurons or something like that, right? But one of the things that we have learned is that if you are on board with more is different and complexity and using a dynamical systems theory kind of approach to these things, right? You record all the neurons, you don't get to choose which ones you're recording. So you have all those neurons that are barely active, but might be really important in certain circumstances. You take all of that, you reduce it down via your favorite dimensionality reduction technique, and all of a sudden you see these dynamical trajectories that you wouldn't see by recording single neurons, and that allows you to abstract away from the single neurons themselves. Right? Now, I know there's a lot of controversy on what that buys you and what it doesn't buy you, but a lot of people are comfortable at that level of abstraction. And, but, but you're.
And then I go, you know, I visited Carol Colby a couple months ago around Christmas, you know, and she is still on the, like, I'm a single neuron person. It has to be single neurons because they're the ones that are doing the things, right? So the single neuron doctrine is still very much in effect, right. So I guess I'm trying to suss out, like, more, you know, like, if I were you, I would. It would feel daunting to me to, like, know, like. Well, I have to, like, keep in mind what this layer, 2, 3 neuron is doing. And I have a strong feel, you know, evidence that it's doing this particular thing. But then I have to fit it with all the other neurons.
[00:56:35] Speaker A: Yeah, well, I mean, you know, so I think most scientists, you know, go through some, you know, this is your job, right? You're a professional scientist, right? You start your lab and, you know, you have some good ideas, your boots are on the ground, you've been. Just been a postdoc, and then, you know, like, you kind of like, establish yourself. And for me at least, I needed a good problem. That was really hard. Okay. Because you're going to spend some decades working on that. So for me, this is a good problem and it's really hard. So I like that. Okay.
[00:57:06] Speaker B: Yeah.
[00:57:07] Speaker A: So, I mean, the way as. As reductionist, the way that I would approach this is by, you know, like, trying to ask the right question and breaking it into small components.
So we found an inhibitory neuron subtype that behaves in a particular way.
First of all, it undergoes very specific kinds of synaptic Modifications Reliably. Reliably. Significantly, compared to other neurons. Right. So it's a molecular subtype that is a somatostatin expressing neuron that has a sort of Martinotti type, but only in layer 2, 3. And the synaptic inputs, the excitatory and synaptic inputs onto the cell go down by about 25%.
Okay, so super specific.
[00:57:56] Speaker B: So that tells me that's a big percentage.
[00:57:58] Speaker A: Enormous. Yeah, enormous. So that tells me that this cell is involved in some changing computation that is occurring when animals are learning. And this synaptic plasticity doesn't happen when the stimulus and rewards are not contingent. Okay. So it's really about when there is an accurate prediction of stimulus leading to reward. Right. When that coupling is really strong. These neurons are changing their anatomical plasticity, but also their activity changes in vivo, which means their computational function should be changing as well. What is their computational function? Well, they have an axon in layer one. And so it's quite conceivable that they are filtering information, feedback information or higher order information.
And so that is, you know, again, it's a tiny hypothesis, Paul. I'm not explaining what the whole column is doing, except for I am saying, you know, I don't think maybe it's happening in layer four, and I'm not sure it's happening so much.
It might be happening in layer five. I mean, some people have found some evidence for this, but it's definitely happening in these layer 2, 3 inhibitory neurons. And it's influencing, we think, the activity of their axons in layer one.
So tiny questions, but they will lead to a more comprehensive account about how that changes, you know, columnar processing.
[00:59:31] Speaker B: It must have been like, also kind of thrilling to think like that, just from a career kind of standpoint and a not getting bored kind of standpoint like that, you know, that there's so much work to be done and, you know, what kind of work needs to be done. And so you're, you're filled up for the foreseeable future in terms of like, what, what is fun and important for you to be doing. Right?
[00:59:57] Speaker A: Yeah. Although, you know, I mean, you also have to be sort of flexible enough. Like, you know, we've been marching through a bunch of different cell types, but recently it was unpublished. But this is super exciting, you know, so these somatostatin neurons, broadly as a population, not just this one small group of them, they are really sensitive to this contingency operation.
And we can see in individual animals that when the stimulus is not predictive of the reward. We can see some of these somatostatin neurons develop responses to the non contingency, right? To the absence of the stimulants, for example.
And it's so interesting because it's a small population and you really have to look at it at the level of the individual animal. Because if you average across all animals, you know, you need to like see that in this animal, you, you think it's developing this belief and based upon its, you know, licking behavior, its behavioral responses. And then you can see this neuron changing as that belief, you know, that in quote, belief seems to be generated.
And so I mean, you know, I'm pretty opportunistic. Like, that's super cool, right?
I mean, especially when you have a neuron that responds to the absence of a sensory stimulus. Like, the one thing I can say for sure is it's not responding to a stimulus because there wasn't one. Right?
That's a clean experimental preparation. And yet we see it generating some responses that's internal, that's coming from inside, right? So I'm not, you know, so like devoted to understanding every single cell type that I can't just like put that on hold for a minute and say, gosh, where's that coming from? That's super cool.
[01:01:49] Speaker B: Yeah, well that's, that's the fun part about science too, is that, that, what is the phrase, huh? That's funny. That's what leads to like the best kinds of questions in science in general or most fun, is there? We were joking earlier about, you know, if we could record every single neuron, that would solve all of our problems. It hasn't solved all of our problems at all. You know, so, but, but so you, you know, people used to think like, well, I, I just need to record more neurons. What's holding me back is I can't record more neurons like simultaneously so I can see the population or something. So what? Is there anything that is like clearly holding you back right now? Like if, if you could just get over this one obstacle or get this new technology or something like that?
[01:02:34] Speaker A: Um, I mean, you know, technically speaking, in the lab, the things that hold us back are like, you know, my laser is on the fritz and the surgery to, you know, build a window into the brains for us to monitor the activity of these neurons is, Is complicated. Right? People have to learn how to do it and there's a lot of things that can go wrong.
But you know, I mean, of course we also have to train the animals. But I, I think, you know, one of the things about learning and we see the animals learn really fast. I mean the animals can learn in a session, right? So a few hundred trials, we can see their behavior changing and then that behavior evolves. And there's a lot of different, you know, neural changes that progress over time.
It's annoying that we can't collect more data faster. Right. But it's really like, can we do the surgery accurately and is it going to be perfect every time the data analysis? You know, I mean, I'm far enough away from the analytical, you know, like sequence that you need to move through that. You know, to me it just seems like, you know, you push a button
[01:03:44] Speaker B: and you know, grad student button. Is that the button you push?
[01:03:48] Speaker A: Yeah, yeah. I mean, so that, that does sort of take some time. But I am a firm believer that having more animals is better because there's so much heterogeneity across animals.
[01:04:02] Speaker B: So is that what I was going to ask you about is the role of like theory, if, if lack of good theory is holding you back, you know, at all? Because I know that you, you can do the experimental setup, you have your, your training paradigm and you can see the changes. But to make sense of those changes, we talked about predictive processing, which is a, a pretty strong theory. And there's like, you know, pretty good evidence that there's something going on along those lines. How to interpret what that is. Like, you know, there's a lot of, there's controversy and you know, it's not like settled. Ah, the brain is a predictive machine, even though in some sense it has to be anticipatory, et cetera. But so, so the, among like the theoretical stories and, or developing your sort of own theory about what's happening. I mean, is that, what role does that play in your scientific endeavors?
[01:04:53] Speaker A: I mean, I, I would love to have better theories.
I mean, you know, I, I've been reading a lot of psychology papers recently and it's so interesting how different disciplines have sort of very different conventions. So. Yeah, I mean, psychologists write a lot of papers and they cite each other all the time, you know, so I mean, different fields have different, like I said, conventions. Right. But I mean the citation index for many sort of psychological studies, if you look at some, you know, eminent investigator and I mean, they're all like in the hundreds, like their age.
[01:05:30] Speaker B: They're citing themselves a lot too.
[01:05:32] Speaker A: Yeah, it's really interesting. So they're quite self referential and then they have these sort of strong hypotheses, but they're not really grounded in what the neurons could potentially do.
[01:05:43] Speaker B: Right.
[01:05:44] Speaker A: So that's fine. I mean, I'm interested in what the neurons are going to do. Right. At the end of the day, it's neurons that are doing the thing. So we should understand what the neurons are doing.
But they do give me some hypotheses to think about. But, you know, I think at the end of the day, I mean, I just want to see it with my own eyes. You know, I just don't believe anything. And, and I'm an experimentalist. I want to be there, I want to see it because I, I, I want to, and I, and then I want to tweak it in just a little way. I would love to, you know, collaborate more with theoreticians. I, I, and I'm interested in, in what people, you know, the models that people are building. Unfortunately, I find a lot of it to be very inaccessible. Right. So, you know, there'll be a paper by people who I know or, you know, I have, you know, interacted with them at meetings before, and I'm like, I just, I'm not, I can't follow it. And so it's hard for me to implement, you know, some of their ideas or to, you know, instantiate their ideas in a, in a behaving animal with like, a specific cell type, because they're pretty far removed from it.
[01:06:52] Speaker B: Right.
[01:06:52] Speaker A: So, I mean, we need to be able to talk to each other.
And I think for a lot of people, a lot of theorists. Right, maybe kind of like what you were alluding to is like, please don't tell me about another dozen cell types. I just, I don't want to think about it. It's too much. And I can't incorporate it into my model. Right. And then they're, and then they're pretty happy with the dynamical systems model because, you know, that you, you can, you can extract that sort of trajectory.
I mean, I was having this conversation with David Brock a few months ago.
He's like, it's all dynamical trajectories.
And I'm like, you know, really? And then I was talking to Byron Yu, and he's like, no, no, no, you know, I mean, we can do much better than that. I mean, I'll say. Actually, I think that a lot of these trajectories represent neuromodulation.
And then, you know, we should be understanding what is controlling the neuromodulator.
And I think that the cortical computation in sensory cortex is actually driving the
[01:07:57] Speaker B: neuromodulation was driving the neuromodulation to then drive the changes in the cortical column. So it's a circular causality. Yeah, yeah. Well, that's the problem is like, you have circular causality. It's what William Wimsat called the causal thicket, where all things are affecting each other and there's not even like a real hierarchy. It's not like a clean, linear hierarchical situation. It is more like what's called a heterarchy, where things are. No one's in charge. They're all affecting each other at different spatio, temporal scales, neuromodulators, spiking units, molecular turnover, et cetera, et cetera. So, yeah, I don't so even like the modeling approach. Like, I think that's why people like David and many, many others are sympathetic to this sort. Like, okay, let's really abstract out because we're going to be able to say something about the shape or the dynamics and not worry about the details. And I think I'm sympathetic to that as well, but I think it's because I'm not that bright and I can't hang on to too many things at one time, you know?
[01:09:01] Speaker A: Well, I mean, so, you know, a lot of this kind of kicks the cam, right? Like, so you see this dynamical trajectory change.
But why? Right. And so, you know, for example, I think some of that trajectory change is a state change, right. Like you have some neuromodulatory milieu that then shifts all the, the activity of all these neurons in a particular direction.
[01:09:28] Speaker B: Sure.
[01:09:28] Speaker A: But then you have to say, well, what activated that neuromodulatory system?
[01:09:34] Speaker B: Do you, do you. That. That's the. In a reductionist mechanistic approach, you do.
[01:09:39] Speaker A: But if you abandon that as an extra thing, Right. I mean, who controls the controller? Right. If, like norepinephrine release is what is causing this kind of shift to some new network, then don't we need to understand what controls norepinephrine release?
[01:09:54] Speaker B: Yeah, but this is when the circular causality comes in and you don't have this domino kind of explanatory chain.
[01:10:01] Speaker A: I wouldn't give up so easily. First of all, I think that, you know, there's a temporal aspect to it, right? Like what turns on, let's say it's norepinephrine. What turns on norepinephrine. That's at a given time. Right. And the causal interaction means that what turns it on has to happen before.
Right. So, I mean, I do think that there is some computation in sensory Cortex that the outcome of which is to activate some of these neuromodules.
[01:10:31] Speaker B: Yeah, that would be cool. To make that.
To see that happening and give evidence for that.
But so, okay, so modern AI, so this podcast, it's less and less so, but it's supposed to. I think the tagline is still like, where AI and neuroscience converge or something like that. I need to change the tagline, but I'm sure AI has changed in the way that your lab does things in a sort of.
In an analysis way. But. But deep learning models are still being used now it's transformers now the brain is a transformer. Right.
Still being used in the same kind of way to. To look at the dynamics. And even people like Jim DeCarlo looking at the activity of single units versus single neurons. Right. In the visual.
What pathway? The. The ventral visual stream.
On the other hand, like, you know, well, before leading anymore, I mean, do you have thoughts like, what, what about, like, is AI teaching us anything about the brain? Should we be studying AI to learn about the brain?
[01:11:41] Speaker A: You know, about 10 years ago, I went to a cosine workshop. And.
Was it 10 years ago? Maybe it was. It was a while ago.
And it was about sort of AI and, you know, cortical computations. I think it was cortical computations, actually. And, and so I was in the same session with, you know, various, you know, computational sort of AI people and giving a talk.
[01:12:10] Speaker B: You were giving a talk in the session?
[01:12:12] Speaker A: Yeah, I mean, I was invited to be part of the session. Right. So. So we were back to back with other people. This probably had something to do with some work. I had been collaborating with theoretical biologist Saket Nablaka, and we'd been looking at pruning rates and network optimization, but it doesn't happen so much anymore.
It's not easy to engage with the AI people. And I think in part because it's working pretty well, even though it's very energy intensive.
And also, you know, there's just. It's very special. The biology is very specialized.
And so the people who are interested in maybe, you know, better AI just don't want to have to think about somatostatin CALP2 expressing Martinotti cells. It's. It's too much. It's too granular.
But I would argue, you know, what. What makes me really excited are new ideas. Right? New ideas about how things might work that I hadn't considered before.
And I think the granular kind of analysis is what helps us appreciate that.
But, you know, instead of coming Closer together. I feel like we're further apart and it might just meet me like maybe, you know, my work is not as easily accessible or has diverged, but I see it as being super related.
Right. I mean just in terms of energy expenditure.
[01:13:41] Speaker B: Sure.
[01:13:41] Speaker A: The, the, you know, the wet. The meat does it much better than the silicon circuits, right. And, and that's why, you know, something like, you know, how do you compress streams of information? Right? We, we nailed it, right.
Our biological systems have figure that out and you know, the, the Intellica systems haven't.
[01:14:03] Speaker B: Yeah, I mean there's this kind of, I've mentioned this plenty of times, but it, to me it seems like neuroscientists, neuros neuroscience writ large is saying like AI, you need more brain like stuff. And AI is like we don't care, you know. And, and I don't know that a lot of people come back to this energy efficiency issue that how energy inefficient AI is. I don't know that the AI folks care about that either.
[01:14:32] Speaker A: They don't care because I don't, I don't think they care. I mean, and you know that, that's fine, right? But let's keep it, let's keep in mind, right? You know, in ancient, you know, Central Middle America civilizations, right, they actually had wheels like children's toys. They, they found children's toys that have wheels. Like you could wheel around your little, you know, goat or something.
But they didn't use them for transporting things long distances.
You know. Why? I mean, I guess they had something that was kind of working okay for them and you know, they just hadn't sort of decided that they were going to, you know, make that leap. But the wheel was there, which is fascinating, right? So it may be that, you know, AI, we're not energy limited right now, but it's expensive. And I mean, you know, everyone's been so excited about AI and we're, you know, the cost of it is not really apparent to us right now because you know, we're, we're kind of using these models that have built and it's, you know, there's just a lot of enthusiasm for it. But the more we use it, the more important the cost is going to be. And you know, I mean, you know, the data centers didn't exist, you know, five or ten years ago, right. But as they begin to draw more and more energy and you know, cooling requirements, things like that, then you know, people will become more interested in this.
[01:15:57] Speaker B: You're not worried about AI being conscious.
[01:16:03] Speaker A: I guess, you know, I don't know how I would define consciousness.
I mean, at the end of the day, I was talking to some undergrads the other day, two guys, and they're. They're really great kids. They're. One of them had been in my class.
I teach an introductory neuroscience class. And one of them, he came back for a reunion weekend, and another one was still in my class, and I sort of took him out for coffee. I thought they should meet each other. And they were like, you know, both of them had a sort of CS background, and they were like, yeah, you know, I think the goal here, you know, of life. Right. Is information optimization.
And I was like, oh, my God, you guys. No.
Right. I mean, is that really what you're looking for? And I think, you know, maybe if you're 22 years old, even doing CS, like, yeah, you want to optimize information, you'll make better decisions. It'll be, you know, you'll move faster through things. But come on. Right? Is that really what gives us pleasure in the world is better information?
[01:17:03] Speaker B: So maybe certain types of people. Certain types of people, yes.
[01:17:07] Speaker A: But, I mean, I'm sad for them. Right. Because there's a lot of other things that are beautiful or fun or emotionally rewarding or, you know, just feel good. Right. That don't have to deal with perfect information. Right.
So, yeah, I suppose, you know, we might get to the point where, you know, AI consciousness is important, but, I mean, would it be an agent like myself that I should take into account?
No. No, I don't think so.
[01:17:37] Speaker B: Yeah. You could turn it off.
[01:17:38] Speaker A: Yeah, I can turn it off. Right. And I mean, you know, I guess maybe I have a hierarchy of, you know.
[01:17:46] Speaker B: Do you swat a mosquito?
[01:17:48] Speaker A: I do. Right. And I work on mice, you know, So I sort of decided, you know, that a child is more important than a mouse. Right.
[01:17:56] Speaker B: So, yeah, I mean, the reason why I brought that up, first of all, it was kind of a gotcha sort of question because it's, you know, sort of in the ethos out there, and I think it's absurd. But the reason why I brought it up is because you mentioning energy efficiency, I used to.
I'm coming around to appreciating that.
What I've. What I've become more and more interested in is something along the lines of biological naturalism, wherein the materials. So there's this idea of computational functionalism that if we just build it in a computer, as long as it's the same function, it'll be conscious. It'll and that's the same thing, right? And have agency and all that stuff.
[01:18:35] Speaker A: Stuff.
[01:18:35] Speaker B: But I've become more and more interested in the idea of what's called organizational closure. There's work I'm putting a workshop on in the fall. You should come to. You'll get announced soon. But related to all this, but it incorporates energy efficiency in terms of the way that the system is closed, is an open system, but it sort of has a circular causality that we're talking about. And it's almost like it has to be built. Has to be built, has to grow and live in that particular kind of organization for it to have that kind of energy efficiency, but also for it to perform all the functions that we love, that give us agency and meaning and all of that jazz. And so anyway, that's a long winded way of saying, I guess I'm becoming more interested in energy efficiency as being important, not because like, oh, we need to make AI more energy efficient, but as an actual key part that you cannot tell the story of our living cognition without including an energy efficiency component.
[01:19:40] Speaker A: Yeah, Yeah, I think 100%. I mean, you know, we've had three and a half billion years of kind of optimization and selection, right?
[01:19:50] Speaker B: Yeah, we nailed it over three and a half billion years ago.
[01:19:53] Speaker A: Over three and a half. And I'm, you know, I'm sure, you know, we did, we, we could nail it, you know, in a different way. Right. You know, maybe even in a better way. But, but this happens to be where we ended up, right? So. Yeah, absolutely. Energy is a part of that. You know, the other, the other important thing for me about I, So, you know, I'm at an undergrad institution, right. And I love talking to undergrads. And you know, there are these.
Our department encourages us to sort of run a kind of journal club for undergrads. And that's kind of fun. You know, I use it as a way of talking about something that I'm like things I'm thinking about. And I kind of want to put it together and sort of get some feedback about.
[01:20:33] Speaker B: Cool.
[01:20:33] Speaker A: Yeah. And I was, I, I sort of did one evening discussion about how the brain is not a computer.
And you know, that was like my premise. And then I was kind of thinking of all the different ways that the brain is different from a computer. Right. So one is this kind of, you know, energy issue, right. Like it's powering itself self. Right. It's kind of part of what it is doing is driving behavior that will then, you know, enable it to Continue functioning.
Anyway, I made a bunch of arguments, and some of them were kind of silly and some of them were, you know, profound. I mean, one is, right? Like the instructions for every brain are in every single cell, right? And the hardware is the software, right? I mean, that's sort of, you know, and that it can repair itself, right? Which, you know, if you think about engineered networks, they don't repair themselves. I mean, I can imagine in 100 years, we're going to have a whole bunch of defunct data centers where something broke. We don't know what it is. And it was cheaper to build another one, right?
[01:21:38] Speaker B: Oh, maybe, yeah.
[01:21:40] Speaker A: I mean, you know, that's like the natural outcome of things, right? So, like this obsolescence. But, you know, it's. It's really hard to fix things. You know, the brain is masterful at fixing itself, right?
[01:21:52] Speaker B: I mean, you know, as I grow older, I'm having that thing where like, oh, I don't know what I did to my shoulder, but all of a sudden it's bad for like four months or something, you know? But I've marveled more and more, you know, even you cut your thumb with a knife or something, a couple days it's healed. Like, it's. It's a. It's like a. And you don't have to do anything. It's a miracle. It's such a beautiful thing.
[01:22:12] Speaker A: Yeah, yeah. I mean, you know, and that's sort of all across, right? That's your body, that's your brain, right? But it's this, like, constantly adapting system which, you know, we're, we're. We're. You kind of see it in AI networks, but, you know, not exactly. And the repair stuff, it's just not trivial, right? You know, like your vacuum cleaner breaks, right? Your toaster breaks, right? I mean, the circuit board on my oven breaks, and then I'm waiting for three weeks to get a new circuit board, right? It's not trivial. The sort of repair and kind of, you know, breakage and repair cycle. So I'm interested, though. And like. So you study the cortex, right? I mean, you've been recording lots of neurons. Like, what do you think the cortex is trying to compute?
[01:22:56] Speaker B: Oh, I don't think it's trying to compute at all. I think it's like. So the way I'm thinking about it right now is, like, it would be weird.
Well, let me ask you a question.
And instead of putting my own opinion into it, but like, when you go to a symphony, would you say that the orchestra is computing the Symphony?
[01:23:17] Speaker A: Well, the orchestra has some instructions and they're producing an output.
[01:23:23] Speaker B: Well, okay, when you write a symphony, are you computing that symphony? I guess the way I think about it more is like, I think actually Eric used this term orchestrate, and it. I think of things more in terms of providing constraints and shaping. And instead of when I think of computing, I actually mean that I like to keep the term reserved for like Turing computing, where you have an algorithm which is a step by step sequence to get to a desired computational result.
I can never tell when a biologist uses computing because it's ubiquitous. Do they sort of sometimes mean just like whatever the brain is doing? I'm going to call that computing and that means computing to me is like a trivial thing and then it becomes like meaningless and you can call anything computing. Like there are terms like information representation. All of these terms have those same sorts of problems. So I think of the cortex. You'll see some of my results in a GHBS in a couple months. But anyway, I think of it as in the context of, of my task and the data that I have as sort of shaping and coming online in a graded degree as needed. And because it's kind of floating around in the right place. But we know if you take it out, it's not going to change a lot of the behaviors that it's supposed to be necessary for. But if you then ask the animal to engage in the task more or to start to like use behaviors in a skillful way, then it becomes like more sharply tuned, properties change. So I think it's like being engaged more and. But I think of it as this graded sort of thing instead of like some computer like crunching along and like spitting out the computation. Right.
[01:25:08] Speaker A: So I mean, you know, the, the brain has a synaptic network. So unlike an engineered network, right. Where you build some connection and you could turn the connection on or off, but I mean, generally what you build is now there, right. So the brain can change the weights of those connections and it can also dynamically do it. Right. To sort of silence something or reveal something, right?
[01:25:31] Speaker B: Yeah.
[01:25:32] Speaker A: So that gives it some functionality that, you know, is, is, could be, could be really useful. And I mean, it's there.
So I guess that it's likely to have some advantages. Right. And so, you know, when you say, well, there's some, some, you know, kind of there is information there and you can then, you know, sort of elevate that, you know, information, right. You make it more available or make it less available. Yeah, I can see sort of neural mechanisms for that.
I mean, but when I say computation, like, I guess I mean, I could just say what is the transformation, right. From spikes? Right. I often show, you know, my students, like, here's a spike, right. And it moves from one brain area to another brain area. Right. You're just transferring spikes.
So what are the principles for that transfer of spikes?
That's a pretty spike centric world view, you know, because some people might say that, you know, there's like this effective coupling.
[01:26:34] Speaker B: Sure.
[01:26:34] Speaker A: And that, you know, there's all of this kind of waves of, you know, field potentials and that's possible. But I, I mean, if we just think about spikes, I want to know what's the transformation? You get four spikes in, you get four spikes out. Right. Do you change the delay? Do you change the interval? Right. That's, that's sort of what I mean by, by computation.
[01:26:54] Speaker B: Right.
[01:26:54] Speaker A: And I think I could find that
[01:26:57] Speaker B: you can, well, you can find analogs of computations anywhere you look if you have a high enough degree of allowance for noise and, or if it, if you squint hard enough.
[01:27:09] Speaker A: Right.
[01:27:09] Speaker B: But it, but like, if you.
[01:27:11] Speaker A: Well, that's pretty damning.
[01:27:14] Speaker B: Well, we all do, right. I mean, it's confirmation bias, essentially. But if, you know, if you abandon even the idea that you just described, which is a totally fine idea, and sort of the classic idea is like, well, all right, I'm going to send four spikes and next time I'm going to send them a little bit slower. And that's a different computation.
But if you come at it, and the problem is, I don't think we have the right conceptual engineering yet or vocabulary or clarity in our vocabulary to, to think in these terms because it's a very engineering, computer science way to think about it. I'm sending four bits.
[01:27:48] Speaker A: I know, I know, I know. And I'm a prisoner, right. Of the 21st century in how I think about information transfer. Right. Because it really could be that you don't need three spikes if four spikes would be fine. Right. And that, you know, maybe glia are the one that are really integrating these signals, you know, to drive some long lasting, you know, change that supports, you know, some signal processing. I, I don't. But I'm still talking about signal processing. Right.
[01:28:18] Speaker B: I mean, it's hard, it's impossible to not. I mean, you, you have a set of vocabulary that you have to deal with, so you either invent new terms or steal them from other fields or whatever.
[01:28:27] Speaker A: I, I mean, I guess what I would say is There are spikes, right. There are synapses, and they do change.
So it's hard for me to believe that that is just epiphenomenon. Okay.
Either the synaptic changes or the spikes. I think that they must be doing something to drive my behavior. Right. I think that they must be.
[01:28:49] Speaker B: I think that's a safe bet.
[01:28:50] Speaker A: I think so. I don't. I mean, you know.
Right. Maybe. Maybe it's about my immortal soul that's really animating me. But I think there's a good case for spikes. I think there's a good case for synaptic plasticity as well.
Right. But to go from that then to behavior. And I mean, not like this kind of behavior, Right. Where like, okay, I did it. Right. But a complex, motivated behavior. Like, I'm going to log in and talk to Paul this morning. Right. Like, you know, that we still don't really understand that.
[01:29:22] Speaker B: Yeah, yeah. It's a miracle.
[01:29:25] Speaker A: You know, you have to find the questions that you can actually answer. Right. And so I keep, you know, we become interested in like orbit frontal cortex and I keep like making myself focus in on sensory cortex. Right. Like ask a question where you can find an answer.
[01:29:47] Speaker B: Oh, Allison, it sounds like you're a good scientist. I'm not sure I approve here.
[01:29:54] Speaker A: Well, I mean, I've been encouraging. You know, I was. I sometimes talk to my students about what we know about the leech.
You know, like we know about sensory motor transformations in the leech. It's very satisfying. Right.
So maybe we should be studying the leech.
[01:30:09] Speaker B: Well, that's what a lot of people end up doing. Like I had Romain Brett on who's theoretical neuroscientist, but he's ended up studying paramecium, single celled, that are like neurons. A lot of people, you know, neuroscientists, are trying to model C. Elegans, you know, because we know it's connectome. So a lot of people are going to like, we. It's impossible right now with what we have conceptually, even technologically, to study the human brain. So let's go to like a very simple organism. And I understand that. And I also see it as sort of.
It's sad in one way because it's like kind of giving up on something, but it's also allowing you to actually ask and answer questions.
[01:30:49] Speaker A: I mean, do you regret working in a more complex system? You know, as a graduate student, you know, recording from single units and doing these complex behaviors? You think that?
[01:30:59] Speaker B: Well, I don't. I don't I still find it fascinating. And it's also like even my project now. We'll talk about this some other time, you know, but like, so in some sense it's comforting to know that it's too complex to answer anything substantial. And we're still grappling for like, what even would be a good question to ask.
And I'm on board with a multi perspectival approach where, you know, you're saying
[01:31:27] Speaker A: now this is what you're grappling with.
[01:31:29] Speaker B: Yeah, yeah, yeah. Sorry, was that the. I thought that was.
[01:31:32] Speaker A: Or you know, what you did as a graduate student. Right. Because we all hope that we're doing things that are kind of moving things forward. Right.
[01:31:39] Speaker B: I'm actually proud of. Yeah, I'm proud of my graduate work, but I was pretty naive when I came in. I knew nothing about like single cell activity or decision making, which is what I ended up kind of studying. I did metacognition, quote unquote, in monkeys. So I asked like this ridiculously high level cognitive. Cognitive question as high as I could. I mean, because I was so naive and I'm actually proud of it, even though I don't know if it moved anything forward. But I think it was solid science with all the caveats of studying higher cognitive cognition in non human primates. And there's a lot wrong with it, but I stand by the science. So
[01:32:15] Speaker A: I mean, you know, the cerebral cortex. So it may be that studying C. Elegans or paramecium. Right. Is just in an explanatory sense is more satisfying. Right. Because fewer moving parts and you really can then make predictions about how the activity or changes in one node of the network will influence then some outcome. But you know, the cortex just seems so fascinating to me because it, whatever it's doing with so useful in the mammalian brain that, you know, duplicating these modules. Yeah. Was. Was useful. Right. And so it does seem like there's sort of a general principle there.
You know, it like the wheel. Right. I mean, it's so useful to, you know, be able to have this kind of geometric structure that spreads force in such. In a particular way.
And I think the cortex is like that. It's like the wheel. Right. Like. And we don't understand why it's so useful. Right. So.
[01:33:14] Speaker B: Right, yeah, yeah, I totally agree. Yeah.
[01:33:17] Speaker A: Even though. Yeah. I mean, you know, the C Elegance people might come up with a better explanation than I ever will.
I still am, you know, kind of a sucker for this brain area.
[01:33:28] Speaker B: That's good. That's A pretty good place to end it. Except I want to ask you if there's something that you're excited about now that we haven't talked about. I mean, I mean, you talked a little bit about some results that had not been published yet, but anything we didn't talk about that you're excited about?
[01:33:45] Speaker A: Well, you know, I mentioned that we've been using learning as a way to probe cortical function.
And, you know, we have this training apparatus where we can manipulate sort of different stimulus reward contingencies. So I'm really excited about this right now because I'm interested in learning generally and I'm interested in why reinforcement is powerful.
And so we recently tried a kind of new set of training conditions where we have the stimulus and reward that are coupled, but then the additional trials that the animal carries out just have a reward, no stimulus.
And so there's a couple really interesting things about this. I mean, first, it's interesting from like a Montessori perspective, you know, like, how do you teach children? Right. You know, how do you deliver feedback?
But for the animals, like, even though four out of five trials have the stimulus, reward, pairing the one out of five trials that just have a reward obliterate learning.
[01:34:46] Speaker B: That's. That's crazy. Yeah.
[01:34:49] Speaker A: So. So the animal learns to disregard the stimulus when it gets these extra rewards
[01:34:55] Speaker B: because it know, well, we can anthropomorphize all we want, but, but the idea would be like it knows it's going to get reward either way.
[01:35:01] Speaker A: So it basically there's a, a neural process for discounting things that are not reliably related to the outcome. Right.
[01:35:12] Speaker B: But in this case, the stimulus is still reliable. No, no, you take that away because sometimes you don't pay a reward with the stimulus. So there's not an association.
[01:35:19] Speaker A: Right, right, right. And, and what's so interesting is, you know, four out of five trials, right. Are stimulus plus reward. And so if you think about this in a really, you know, kind of mechanical way, right. And you think of. So we're monitoring, by the way, plasticity at the level of synapses, input and target specific synapses. And if you think about that, right, four out of five trials, right. They're ratcheting up synaptic strength.
And then you have this reward only experience and it's so powerful, it, it like actually eliminates the synaptic plasticity that we see. So I think that is super interesting. Why those reward only trials can undo all of this synaptic plasticity and they also change behavior Right. They make the animal disregard the stimulus.
And we'd be really interested in what are the.
Why is this 1/5 of the trials? Why is it so much more powerful?
And so we're interested in the neurochemical basis for that. And I think we have some good candidates. So I'm really excited about that.
[01:36:25] Speaker B: What, what time scale does it happen on?
[01:36:28] Speaker A: A day or two, like over.
[01:36:31] Speaker B: So over the course of a day you can see the buildup of the strength. And then within the same day, or the. Or then you take away the pairing or just, just deliver reward without the stimulus. And then the next day you see.
[01:36:44] Speaker A: Right.
[01:36:44] Speaker B: Obliterated.
[01:36:45] Speaker A: The assay that we're using is, you know, we're monitoring synapses in fixed tissue. Right. So it's definitely there in the living animal. And so we see higher order thalamus synapses onto layer 2, 3 pyramidal cells. They get larger when the animal is learning.
Just in the early part of learning. Right. At two days of training, they get bigger, detectively bigger. But when you add these reward only trials for two days of training, so that's a few hundred trials. Like, let's say two days of training is probably 600 trials. And so maybe a hundred of those trials have reward only signal, but that synaptic potentiation, that increase in synapse size, that those higher organic inputs onto layer 2, three pyramidal cells is obliterated.
[01:37:35] Speaker B: Yeah, it's cool.
[01:37:37] Speaker A: Yeah. And that just tells you that the brain is a machine for trying to assess causal interactions and it is looking for reliable causal interactions. It's hypothesis testing all the time. Like, I think this. Oh, yeah, I got it. I think this. Yeah, I got it. Wait, that doesn't make. That didn't happen. That doesn't make sense. Right. And it's actually undoing some of that plasticity.
[01:38:09] Speaker B: All right. It sounds like that's going to be occupying you for some time to come. The why. So cool stuff. Thank you, Alison. I'm glad that we finally did this.
[01:38:19] Speaker A: Yeah, me too. It was fun to talk with you. We have lots more to talk about. We'll have to. And in fact, you know, I had David, Barack and a couple people over for dinner a few months ago. I'd love to have you and him talk about this. Maybe we get Byron to come too and talk about neural trajectories and what they show and what they don't show. Because if Byron is skeptical, Right. Then, I mean, he's like Mr. Trajectory. Right, right, right, right.
[01:38:47] Speaker B: It would be great to have a conversation about that. I know David actually, and I mean he's been on the podcast and we've interacted multiple times. So. So he and I are, are. We think similarly about these things. Although I, I think I have a lot more skepticism. But we shall see. Yeah, okay, that's good.
[01:39:03] Speaker A: Anyway, it was a pleasure. Thank you.
[01:39:12] Speaker B: Brain Inspired is powered by the Transmitter, an online publication that aims to deliver useful information, insights and tools to build bridges across neuroscience and advance research. Visit thetransmitter.org to explore the latest neuroscience news and perspectives written by journalists and scientists. If you value Brain Inspired, support it through Patreon to access full length episodes, join our Discord community and even influence who I invite to the podcast. Go to BrainInspired Co to learn more. The music you hear is a little slow jazzy blues performed by my friend Kyle Donovan. Thank you for your support. See you next time.
[01:39:56] Speaker A: Sam.
