BI 221 Ann Kennedy: Theory Beneath the Cortical Surface

[00:00:03] Speaker A: Once I got to this subcortical area, I realized that there really was a lot of opportunity for a theorist to make a contribution and kind of connect some of these older models and ways of thinking about instincts and behavior from the behavior first perspective and connecting them to modern systems neuroscience methods. I think that there's a clear importance of neuromodulators and neuropeptides in brain function, almost like a context signal, something that reshapes the dynamics of a neural network to do something different depending on the condition that this signaling molecule is reflecting. She doesn't feel pain, but she also reports that she's never felt angry or afraid. She's just like super chill. It seems like it's very much suppressed in her, maybe because this mutation is reducing the ability of her hypothalamus to kind of rev up its activity and produce persistent firing and persistent motivational states. [00:01:18] Speaker B: This is brain Inspired, Powered by the transmitter I am Paul. Welcome to Brain Inspired. Ann Kennedy is Associate professor at the Scripps Research Institute, and she runs the Laboratory for Theoretical Neuroscience and Behavior. Among other things, Anne has been studying how processes that are important for life, like survival, threat response, motivation, pain, how those processes are mediated through subcortical brain areas like the hypothalamus. She also pays attention to the time course those life processes require, which has led her to consider the expression of things like proteins that help shape neural activity throughout the brain so that we can behave appropriately in those different contexts like threat and pain and so on. So you'll hear us talk about how this is still a pretty open field in theoretical neuroscience, unlike the historically heavy use of theory in popular brain areas throughout the cortex, and the historically narrow focus on spikes or action potentials as the only game in town, basically when it comes to neural computation. We discussed that and I link in the show notes to a commentary piece Anne wrote in which she argues for both top down and bottom up theoretical approaches. I also link to her papers about the early evolution of nervous systems, how heterogeneity, AKA diversity of neurons, is an advantage for neural computations, and we discuss a KAGGLE competition that she and her team developed to benchmark automated behavioral labels of behaving organisms, the goal of which is to produce consistent behavioral labels across different labs, which use different recording setups and systems and serve as a common tool for different researchers to use and to aggregate data into bigger and better data sets. So you can find those show notes and learn how to support this podcast to get full episodes, the full archive and more at Braininspired co. Thank you to my Patreon supporters and thank you as always to the Transmitter for your support. Here's Anne and I. So I saw you. I've been aware of your work for some time now, but then I saw you give a talk a couple months ago at the Georgia Tech. What was it called? Interface Neuro Workshop Conference. And I think you started off your talk by saying how out of place you felt at a conference where they were talking about brain implants and modulating brain activity, et cetera. If I remember correctly, you said you felt a little out of place giving a talk in that environment. Why was that? [00:04:07] Speaker A: It's not what my lab mainly focuses on. The organizers were some friends and I. I guess I felt a little bit. I'm not exactly hearing paralysis or human disease the way that they are. I think it's a fascinating community, but yeah, different background, different set of goals. I guess a common thread in thinking about how the brain modulates itself versus how we might modulate it as device developers. [00:04:37] Speaker B: Oh, okay. Yeah, that's interesting. We'll get more into that. I mean, one of my takeaways, and I was. I don't know why, like I said, I was aware of your work, but maybe the one or a few of the threads of your work that most relate to my own work. So I was surprised when it's. When it seemed to me that you expressed an interest in life processes, which is kind of the antithesis of the computational perspective that dominates neuroscience. But now I think I had that wrong, that it is from a computational viewpoint. So I was struggling to figure out, like, what do I really want to know from you? That question has changed over the course of, you know, a couple weeks in leading up to this, but is it, are you interested in life processes or are you interested in looking at different levels of cognitive function at a systems view? What. What is it that. How am I. How am I saying those things. Things wrong? [00:05:41] Speaker A: I mean, I think one of the nice things about theory is you can kind of touch on a lot of different topics. But yeah, life processes. And I guess one of my guiding impulses has been to look for things that are kind of outside the beaten path of theoretical neuroscience. Subcortical structures, areas that we haven't really modeled to the same extent that we have, say, cortical processes for motor control and sensory processing. So that led me to life processes, I guess, survival behaviors, how the brain modulates our behavior in response to our needs. Something I started into just because it felt under explored by theorists and Then kind of discovered that there's actually a lot of people thinking about this. It's just they're not necessarily systems neuroscientists and that it's only been fairly recently that the parts of the brain that are involved in survival behaviors keeping us alive have been accessible to the tools of systems neuroscience, to single cell resolution recording, to stimulation and manipulation. So once I got to this subcortical area, I realized that there really was a lot of opportunity for a theorist to make a contribution and kind of connect some of these older models and ways of thinking about instincts and behavior from the behavior first perspective and connecting them to modern systems neuroscience methods. [00:07:18] Speaker B: Yeah. In some sense. Well, first of all, I was going to. You ruined it. Because I was going to make a joke that if. You know what I've learned in my many years studying neuroscience is that the cortex is the only interesting part of the brain and everything else is so, so boring. Right. [00:07:34] Speaker A: It's decorative. It's just the protective outer shell that keeps the middle part. No, well, yeah, yeah, but you did. [00:07:42] Speaker B: Say that you're, you're interested in these parts of the brain that are dedicated to more survival and motivation in these life processes. But the cortex is also. It's just not that that's how it's been traditionally studied. I suppose that's true. [00:07:55] Speaker A: Right. I mean, cortex isn't evolved alongside these deep subcortical structures and it interacts with them and provides information to them and enriches the things that they can do. I'd say you can do a lot without a cortex. I was. My favorite study that I like to mention is a work done by a friend who was a grad student in Marcus Meister's lab when I was a postdoc at Caltech. The Meister lab got hold of this mouse line that the, the driver line expressed in all of the neural progenitors that went to, went on to form the cells of cortex and the cells of hippocampus. And you could express a toxin in these progenitor cells and just completely remove the cortex and hippocampus from the developing brain. The mice came out just fine. They got like a little smaller looking head, they're a little more anxious and more aggressive, but they can do a lot of mouse stuff and they've got nothing of those more, more recent structures. [00:08:54] Speaker B: Like what kind of stuff can they go. Go ahead, sorry. [00:08:58] Speaker A: Yeah, they can grooming, they can groom, they can mate, they can fight, they can escape from threats. They, they don't learn especially well. But a lot of the core things that you have to get right to survive and reproduce. They get right because that's, that's the stuff that's most ancient to our behavior and that's, that's the stuff that is in these subcortical structures. And cortex informs that. It gives you plasticity to those behaviors and context dependence. But you, you need those things to, to survive. [00:09:34] Speaker B: Well, okay, so since you mentioned that they don't learn, I mean, so I was going to say these subcortical structures are often what we associated with these more innate behaviors. And when you say innate, that tends, you tend to think of, you come out of the womb or the egg pre programmed with these things. However, there are people who like, who say that, well, there are people who study learning. Well, there are people who claim based on their own research that these innate capabilities actually are learned. And there is learning in these subcortical structures. [00:10:08] Speaker A: Oh, there absolutely is. Yeah. I think it's, I like the language of canalization that you're kind of, you've built a brain that is predisposed to producing certain outputs. Like it still has to refine its activity via not necessarily like supervised learning, like I'm teaching you how to eat or how to fight, but like there's some experience dependent and activity dependent wiring up of these structures. But they tend to all converge on the same wiring and not like at the cellular level, but the same repertoire of behaviors, the same drives, the same sensory motor transforms across animals. [00:10:50] Speaker B: So those are deep canals, deep attractors. Yes, in the dynamical systems viewpoint. [00:10:56] Speaker A: Yeah. This comes out of Waddington's landscapes. This idea that you have this kind of formation of. It's, it's funny when he wrote about this, he wrote this book the Strategy of the genes in 1957 and he's, he's a biologist, he's not a mathematician. So he calls these basins of attraction creodes, not attractors. And he describes it entirely through this sort of intuitive way of thinking that genes can kind of push cells along particular developmental trajectories and if you perturb them away from those trajectories, they'll recover to them. So it's an attractor. He just does and didn't know that language in that area of math. But I think that idea is very much present both in development of cells and also in wiring of the brain. That you have kind of attractor states of circuits that predispose animals to forming particular patterns of behavior. [00:11:59] Speaker B: So you come from cortical work largely. I mean, I know that Larry Abbott does a lot of different work with a lot of different brain structures. But you know, coming out of that vein of research. [00:12:12] Speaker A: Well, my, my graduate work was in electric fish. [00:12:15] Speaker B: Oh, okay, okay. [00:12:17] Speaker A: So it's kind of impossible to get scooped in electric fish work. [00:12:21] Speaker B: See, that's what you were going for with the subcortical structures. Right? The anti scoop career. Right. [00:12:27] Speaker A: I had worked in electric fish and like you can know the whole literature, you can know all the labs. And I had done a little bit of work on Drosophila mushroom body and I was like, no, this is too crowded. There's too many people working on the same problems. And I saw David Anderson give a talk at Cosine. This is like 2011, 2012 on Dayu Lin's work. Optogenetically evoking aggression behavior in by stimulating neurons in VMHvl. And it just felt so wildly different from anything anybody else was talking about at Cosine. I mean, this is like you're hitting a part of the brain with an optogenetic hammer. This is very artificial, but you're getting a complex visually guided behavior. [00:13:06] Speaker B: Right. [00:13:07] Speaker A: And it just felt like everything else felt like it was kind of missing this very clear thing that brains are supposed to be doing, producing these complex behaviors. And at the time I was like, oh, nobody must be working on that. It's going to be a fun open area of exploration. Turns out people, people are working on that. Good number of people, but yeah, probably. [00:13:28] Speaker B: For a long time too. [00:13:29] Speaker A: Yeah, yeah. [00:13:31] Speaker B: One thing I was going to ask you because you've written about evolution and studying early nervous systems and pre nervous systems. So did that interest come from the early graduate work that you were doing or how did you come to that? [00:13:46] Speaker A: So I guess before grad school I was briefly a technician in a stem cell lab. I did a lot of co transfection of transcription factors into a cell line and looking at how one transcription factor affected expression of genes involved in a like a dendrocyte progenitor cell differentiation. Totally random, but that like I hadn't realized until then. This is going to sound dumb, that different cells in the body express different genes and that. And I really fell in love before I ever worked on neuroscience with developmental biology and from developmental biology to evo devo. So this study of development from an evolutionary perspective, I never really worked in this space. But just the set of problems that you think about in those fields of like how you form an organism from a single cell, how you evolve a new behavior or evolve a new function, the ways that evolution has taken parts of a body plan and fine tuned them and adapted them over time. That was just something going back to high school that I was really in love with. And so it was in the background for a while through grad school. But it's something that's really stayed with me. [00:15:05] Speaker B: Okay, so then maybe I had it backwards. Maybe the thing to ask you is how you went from that to then studying like the cognition in the dynamical systems regime in naturalistically behaving and otherwise organisms. [00:15:22] Speaker A: Yeah. So when I, when I was applying for grad schools, like, I also thought neuroscience was cool. I think a lot of people just have this general interest in how their brains work. When I was applying for grad schools, I interviewed at Columbia just because one of my professors mentioned it on like a list of schools I should look into. And Larry Abbott was meeting with the grad students and he showed them what eventually was force. This work with David Sicilo, training recurrent neural networks to produce complex behaviors and, well, complex time series. The example he had was like this. They had like motion capture data from a person who was walking and they trained the model to reproduce that, like, walking behavior. [00:16:07] Speaker B: Yeah. [00:16:08] Speaker A: And that was just mind blowing and felt like I wanted to know how that worked and how, how you could have these distributed networks of neurons do these complex things. [00:16:23] Speaker B: So, but, okay, so did you immediately, I promise we're going to get to like your specific work, but so did you immediately connect the. That. So a recurrent neural network is all about dynamical systems. Right. Because there's recurrence and there's dynamics and that's how they compute. Whether it's like a liquid state machine, you know, or you're using the force algorithm to train an rnn. And when you. So when you saw that, for example, Larry Abbott giving that demonstration, did you immediately connect that dynamical systems view to the cellular and subcellular levels and to, like, things like subcortical processes? Because you're finding in these subcortical areas that they have these same kinds of dynamical systems, shapes and attractors, et cetera. [00:17:11] Speaker A: Oh, I mean, in grad school, not at all. Like, I didn't know anything about subcortical structures until I started my postdoc. I entered the lab. [00:17:20] Speaker B: I forgot. I forgot. You were dumb. Right? [00:17:21] Speaker A: I was dumb. I mean, it took me a while to even get like VMHVL down. Right. And like, say, the right acronym. After I joined the Anderson Lab, I had learned about dynamical systems in undergrad and saw this as a cool application. But I think the narrative that was in theory at the time was individual neurons have really short membrane time constants. They can't do things that require working memory and complex motor control and things that require you to really integrate information over long periods of time. And rest of our computing was the way past that. And that was really compelling and exciting to me. When I started grad school and I was, for the first year or two I was working a lot on just sort of making, trying to learn more about what a recurrent neural network can do and how it does it. There was this, I guess I, it wasn't very connected to data. So I was always kind of struggling with. Is what I'm doing really telling us anything about the brain or am I just studying this dynamical system that is nice and easy to study? [00:18:34] Speaker B: Oh, because what you're trying to do is like a little tic tac toe problem or like some sort of binary on off thing to train it to. [00:18:42] Speaker A: Do some abstract sort of work reduced thing. And I was barely even training, I was like looking at memory capacities of these things. Like how long into the past can you reconstruct the input to these networks? And I think what I eventually settled on some years later was that yes, you can get long time scales out of recurrence of a neural, of a firing rate neural network. But I think biology probably solves the problem other time in most of the time in other ways. It has a variety of time scales that it has access to through the time constants of molecular signaling processes, through circulating hormones in the blood, through, through gene expression timescales. That maybe by focusing on reservoir computing it was really cool, but maybe by focusing on understanding this cool thing we were missing how biology was actually solving the problems. [00:19:36] Speaker B: Was it cool and yet somewhat narrow and brittle? Like because you had to sort of tweak it just right to get it to function. Right. My understanding of reservoir computing is that it actually has a pretty high capacity for expression. And that's the whole point. Like, so resmob computing, right? You, you randomly assign weights to all, between all the parameters and if you have like enough units, you can then train it with like a linear readout to do lots of what some people think are interesting. Lots of things. Yeah, whether they're interesting or not, lots of things. But, but so you spend a lot of. And people spend a lot of time like training these things and trying to figure out what parameters worked. And so was the problem that. So am I wrong? Are they not highly expressive or are they highly expressive? But if you set them up just the right way and then if you tweak them, then they aren't or. [00:20:28] Speaker A: I think we've gotten better at training them stably. I have a postdoc who does this a lot. He's fitting some RNN models to some data right now. And with the right constraints and tricks, I think that you can get past this like fiddly feeling of trying to train neural networks to do things. I think that methods have just improved a lot over the past 10, 15 years for this kind of model fitting. [00:20:54] Speaker B: Sure. [00:20:56] Speaker A: It was more just that I wasn't sure. It bothered me that I wasn't studying brains. I was studying this sort of toy model system. That's really just something that everybody has to figure out when they're in grad school. Right. Is do you care about methods? Do you care about understanding math and dynamics? Do you care about machine learning? For me, it was really that being able to point back to the biology and point to a system that I was helping to understand that made me feel like I was contributing to the field. [00:21:32] Speaker B: I was going to ask you this later, but since you brought up machine learning and it sounds like you're interested in the actual biology, biological brain and processes, where are you on the neuro AI hype train these days? [00:21:48] Speaker A: I always have to like stop and try to think about how that's defined. [00:21:52] Speaker B: It's not really well, but nothing is. [00:21:55] Speaker A: What is it like improving AI by looking for additional principles of computation in the brain? [00:22:03] Speaker B: That and the reverse. Right. It's the beautiful, the virtuous circle of AI and neuroscience. [00:22:09] Speaker A: Yeah. But I think most people who say they work on neuro AI, it's not like I'm using AI tools to understand my neural data. It's also studying the brain will help us build better networks. [00:22:20] Speaker B: Sure. [00:22:21] Speaker A: I mean, I will say, I mean I think the family of networks and methods that people in AI use is pretty different from where the brain solve problems. And I think you can make a compelling argument for maybe it's good to try to understand how brains solve problems, not just so that we can help fix people when brains break down, when things aren't working properly, but also to adapt those principles for our own computational tools. Nervous systems have this capacity to do things locally instead of having sort of centralized ascending hierarchy of control. That I think is interesting and compelling and that we don't really fully know how to leverage in machine learning and AI systems. I think that there's obviously an energy efficient argument, efficiency argument to be made for understanding how Brains compute things and how that's different from the way that we compute things in a massive language model running on our GPUs. So I think there's a good argument to be made for studying the brain as a way to push forward our engineering of artificial systems. At least I've tried to make that argument. For me, understanding the brain is really the goal, but I do think like there's, there've got to be tricks that the brain is using that these networks aren't and whether that's just a substrate thing the brain uses. Neurons communicate via big array of different signaling molecules that have intrinsic time constants. Like these might be things that we can't really implement in a, in a computer, but I think they're interesting to understand. [00:24:24] Speaker B: Are you going to put money on that? [00:24:29] Speaker A: No, I. [00:24:30] Speaker B: Okay, I don't have you. Another time maybe when you've done a little more research. But on the other hand, it's not like dynamical systems came out of nowhere. But you mentioned David Sicilo, who was working with Larry Abbott doing the reservoir computing and the force algorithm that you mentioned earlier, and his modeling work with Monte is one of the early works that people point to as a proof of principle that if you study machine learning or artificial neural networks, it actually gives rise to principles that then you can transfer over to help understand brains. Right. The dynamics. Right. In this. [00:25:10] Speaker A: Oh for sure. [00:25:11] Speaker B: In this recurrent neural network trained to do a simple kind of context dependent decision making task. And you study the dynamics of the artificial network and you actually learn something potentially about how, how brains do it. Even though brains are doing all sorts of different timescale signaling and using all sorts of different molecules and. Yeah, scales, et cetera. So. So you must appreciate that aspect of it, I suppose. [00:25:39] Speaker A: Yeah. Okay. So training a network on a task and then looking at how it's learned to solve that task and trying to use that as a way to interpret neural activity. [00:25:47] Speaker B: Sure. [00:25:47] Speaker A: Yeah. No, I think that that is a big aspect of how we interpret neural activity. Like when we have recordings of neurons from some complex behavior, we need some in to how we're going to interpret those recordings. And I think the dynamical systems framework is a really powerful one and that it has helped us to deal with the fact that a lot of our handcrafted ways of interpreting neurons, things like tuning curves, just don't capture much of the variance of neural firing rates. So I, I definitely agree that a dynamical systems framework is a good way to make sense of, of neural data. [00:26:28] Speaker B: But you go ahead I was going. [00:26:30] Speaker A: To say, I don't know how far you can push it because you're not necessarily going to find in your artificial network a solution to a problem that is the same solution the brain uses. Like it's a thing you can try and sometimes it works, but sometimes it doesn't. And maybe that's because you didn't have the right representation of the task. Maybe the way that information is reaching a network, how it's processed, by the time you get to the computation, you're trying to understand. In the case of subcortical structures, we're very much abstracted away from the visual field of the animal, we think, and the rich sensory environment. We're dealing more with slow motivational states and things that are hard to predict. So coming up with like, if I took a neural network and I trained it to forage, when there's a predator around the night, attack it. Like, I feel like it's going to take some effort for a virtual agent trained on these kinds of survival related tasks to really have, behave. Have dynamics to its state that you can map onto brain structures. [00:27:52] Speaker B: Well, you just said that those subcortical structures have slower dynamics, but under threat you have to have fast dynamics. Right. If your flight kicks in, it's fast. And that's subcortical as well. [00:28:04] Speaker A: Right, right. So for threat spec specifically, you need to be able to respond quickly to it, but you also need persistence. So if you see a predator and you run and now you don't see it anymore, you need to remember that you just saw a predator and keep hiding for a while. So there's not necessarily slow, but there's persistence to the dynamics of these internal states that are important in a lot of cases. [00:28:27] Speaker B: All right, so let's rewind. Not rewind, but let's go back before brains existed, which you've written about recently. And then maybe we'll work up to the early evolutionary versions of brains, which is like the subcortical processes. Right. Which then became more elaborated. And then we grew this cortex which you referred to as what, the extra. [00:28:50] Speaker A: Stuff, outer shell, the refinement. [00:28:54] Speaker B: It's like a turtle shell, it's there for protection. [00:28:58] Speaker A: Yeah, exactly. Totally unnecessary. [00:29:02] Speaker B: Yeah. This is a quote, actually, very short quote, but it's nice. From the paper that you wrote about the evolution of nervous systems. Let me see, I have the title right here of the paper. Dynamics of Neural Activity and Early Nervous System Evolution. Much of what we think about as neural arose before the nervous system. This is what I was getting at in terms of Your thinking and where it came from as viewing everything at this systems dynamical level. Because a lot of people who get into, like, theoretical neuroscience, there's a population doctrine now. Right. We're beyond single neuron tuning curves. And so we're kind of going more and more abstract. But then. And people resist, like studying ion channels and neurochemical signaling, but you're kind of going back down to that level, but viewing it from that kind of systems perspective. I think that that's what I was trying to get at earlier. So this is kind of along that same line. [00:29:59] Speaker A: Yeah. And I'd say that this is. There's a growing few of us that think that. That are interested in these sort of molecular and subcortical aspects to neural computation. So a big part of the inspiration for that line in that paper was Romain Brett has this great paper on the behavioral repertoire of paramecium, which he calls a swimming neuron. So it has sensory systems, it has motor effectors. It can produce complex behaviors, exploratory behaviors. It can bump into things and adjust its course like it does all of these things. And it's a single cell. And within a single cell, everything is chemical signaling between your sensors and your effectors. But the parts that are there within the cell are the same as the same parts that are present in a larger organism. I mean, the motor effectors have changed. We don't have flagella. We've moved to having contractile tissue. But the sensory molecules are touch sensors and photo sensors. Those have been around for a long time. [00:31:06] Speaker B: The same molecules or the same functions with different molecules? [00:31:11] Speaker A: To a large extent. [00:31:14] Speaker B: Not that it really matters, but I'm curious. [00:31:17] Speaker A: A lot of the same channels and molecules predate nervous systems. A lot of the ion channels are there. Mechanosensors and photo sensors are pretty similar between. What is it? Doesn't channelrhodopsin come from cyanobacteria? [00:31:36] Speaker B: That sounds right. [00:31:38] Speaker A: Man, I should have googled some of this before coming on this podcast. [00:31:41] Speaker B: Say it, and then if it's wrong, they'll take it out. [00:31:45] Speaker A: Oh, man. Okay. Channelrhodopsin clearly came from cyanobacteria. [00:31:51] Speaker B: Clearly. [00:31:52] Speaker A: Piezo receptors for mechanosensation are found in. They're found in plants, they're found in single cell organisms. The pieces are all there because they served a function before they were used in neurons. [00:32:04] Speaker B: Yeah, it's not like evolution was like, hey, oh yeah, we need nervous sensors now. [00:32:08] Speaker A: I need sensors now. [00:32:09] Speaker B: I need to invent some new stuff. Yeah, yeah, yeah. [00:32:12] Speaker A: So there Were these parts that were found in cells, single celled organisms that became co opted to do some, I guess, the same thing in a different structure. Once you got to the point of having a nervous system. And there's like a couple folks who really think about this, like how do you go from what is the evolutionary process of going from a single cell to a multicellular organism? How do you go from every cell being identical and being in a colony to having more specialization? And from specialization, how do you develop organs and nervous systems and contractile tissue? It's really a fascinating area to explore and something where we have a pretty decent understanding of what that process could have looked like once you get to. I mean, there's a couple of different theories about nervous system origin, whether it was controlling ciliated motion or contractile tissue, or whether it was primarily secretory and modulating activity of other cells. [00:33:11] Speaker B: Those are two ways. But you expound upon a third way, which is sort of the spontaneous endogenous activity geared towards maintaining like an internal state. Am I saying that right? [00:33:25] Speaker A: Yeah. So that was something that we pushed in that paper because we were looking at this through the lens of jellyfish, which have a lot of endogenous activity. And endogenous periodic activity is pretty easy to generate from a lot of dynamical systems. All you need is some coupled, some feedback coupling with a separation of time scales of your interacting components, like one that's fast and one that's slow. [00:33:51] Speaker B: Does one need to be excitatory one inhibitory as well, like for the. [00:33:55] Speaker A: I think you can do it with a purely inhibitory system. I think that's how it works in the pyloric rhythm, I guess. [00:34:02] Speaker B: Well, probably. I guess it depends on the time scales of the. Just the signaling? [00:34:05] Speaker A: Yeah, yeah. It's the separation, having separate time scales of your interacting components that's most important. But it's very easy to get and possibly a useful thing to have. Periodic activity shows up in a lot of parts of the body. It's involved in swallowing and digestion, swimming if you're a jellyfish. Circadian rhythms, possibly periodic signals in release of hormones and signals like phasoconstriction. There's like interesting bursting of insulin releasing cells. A variety of systems that really have these kind of oscillators in them. One of the arguments for that is that if you need to convey a signal from one area to another, you can do it in the amplitude modulation way. You can send like a fixed level of a substance of a ligand to your downstream receptor or you can do it in a frequency modulating way. And if you do amplitude modulation and you have any noise in your signal, it's impossible to distinguish that noise from like the, the amplitude you're trying to convey. But if you're communicating with pulses, having little white noise on top of that isn't going to interfere with your sensing of the frequency of those pulses. [00:35:23] Speaker B: Okay. [00:35:24] Speaker A: Oscillators were possibly an, a feature of early nervous systems and are useful for a variety of biological functions. [00:35:31] Speaker B: So yeah, I mean, there's all this talk about what are brains for? They're for movement, they're for perception, they're for survival. [00:35:39] Speaker A: Control of development. [00:35:41] Speaker B: Control of development. [00:35:42] Speaker A: Something that controls the timing of hormone release, something that drives you through. [00:35:49] Speaker B: Well, you're talking about the oscillations themselves, right? [00:35:52] Speaker A: Oh, no. Just like a metamorphosis when you decide to pupate and emerge into your adult form. These are things that are controlled by the nervous system, Right? Puberty is controlled by the nervous system. It does a lot of things beyond behavior, right? [00:36:07] Speaker B: Well, right. Okay. That's the anti. Brains are for behavior. But. Well, so all of these are just so stories. But there's an argument to be made that brains are for controlling like the development and internal milieu and signaling. And the thing that you point to is that it doesn't take much for oscillations to occur and that oscillations are a great way to signal these things. So I'm going to. So just from that. So oscillations are important. Endogenous, spontaneous activity, important before nervous systems. Now you have, these days, people like Earl Miller saying cognition is rhythm. You know, it's all oscillations, like on up. Right. And there is this recent paper looking at like, networks. Cortical, of course, cortical networks. I don't know if subcortical stuff was in there too. Using fmri, finding that there are these transitions from different sub networks and that it's rhythm and it's rhythmic, that it flows through these transition states, like common transition states on a cycle in oscillation. So is it oscillations all the way up? Is cognition rhythm? Is that your bet or is that just one perspective? [00:37:24] Speaker A: I, I've never worked in that, that space of like, looking at oscillations in, in human data or in vertebrate data. So I, I'd have to, I, I often struggle with, like, how do you ground the oscillations that you're pulling out of your EEG signal and like, relate that to what neurons are doing? [00:37:44] Speaker B: Right. [00:37:45] Speaker A: I Do think that it's worth thinking about the frequency component of neural signals? For example, I was talking to Lynn Chan about her imaging of serotonergic signaling recently, and she was mentioning that if you look at, she develops these optic sensors for extracellular neuromodulatory and peptidergic signals, and in this case serotonin. She was looking at different frequency components of the serotonin signal. And you really do see that there's like, there's high frequency fluctuations in concentration of serotonin and then there's a intermediate and low. And if you look at the families of receptors for serotonin, there are some that have a certain sort of low pass filtering of their response. They need to be able to bind to serotonin for a certain amount of time to really kick off a downstream signaling cascade. So it's, it's something my, my postdoc and I are talking about is like, is separation of signals in frequency space a thing that we should be thinking about in terms of how it can influence communication between barren areas. Also, maybe development of neural networks. If you have plasticity rules that function on certain timescales, they let certain frequency ranges of activity shape synaptic weights and they filter out other ranges of activity. I don't know if different brain areas use certain frequency bands to communicate with each other. It's just not an area where I've ever worked. But I think that it's an interest. This temporal component of neural activity is a thing to keep in mind. [00:39:31] Speaker B: Yeah. Okay. So this is not to bring us back to AI, but this is a glaring omission in any AI system is the temporal aspect of. There's no timing in an AI system. Right, right. [00:39:47] Speaker A: It's not a problem that they. Unless you're doing robotics. And closed loop control is not something that most of these systems that we think about have to deal with. [00:39:57] Speaker B: Yeah, right. But, but in, in biological systems, it is timing all the way up and all the way down. I mean, it is, it's, it's crucial. Just. You're just listening. The development, all of the, you have to time the fight or flight. I mean, it's all, it's all timing. So I'm on a kick. I'm betraying this a little bit. So I, I keep coming back to like, God, what are the differences between artificial systems and natural biological systems? What are the differences that make a difference? And dynamics writ large is one. So to say that you can point to dynamics, quote unquote, in like recurrent neural networks that you. That are artificial machine learning systems. But what I mean by dynamics more so is like timing of things. So everything is very sensitive in biological process processes, Whether it's oscillations or not. Everything is very sensitive to timing. So that's kind of what I was getting at. [00:40:54] Speaker A: Yeah, the timing is something you can't really separate from biological computation the way you can from computation in a deep neural network. Yeah, absolutely. I think there was this field for a while of asynchronous computing, Computing without a clock signal where all of your components have to kind of do things on their own time frames, without any kind of central controller that's keeping track of the progress of computation and how you could make algorithms that work in such an asynchronous system. I think it's. That's a problem that the brain has to deal with. This fact that everything is done locally. You don't have global teaching signals for learning rules necessarily. You don't have global clocks that are deciding when to query information from one area, when to pass information to another area. Time is part of the embodiment of biological neural networks. I guess the other. This is kind of changing track a little bit. But the other thing that's very clearly essential in biological. To the function of biological neural networks and something I'm very hung up on is just the diversity of signaling molecules and signaling pathways between neurons. [00:42:07] Speaker B: Oh, okay. Well, yeah, so I was going to bring up the. So timing was one thing. The other thing that you focus on in that paper about nervous system evolution is the endogenous nature of the spontaneous activity, Oscillations or timing or. Not that it is like an endogenous, like endogenously produced. So I just wanted to highlight that. I don't know if you want to comment on that as well, but yeah. [00:42:30] Speaker A: That was largely in response to a lot of the literature on nervous system evolution. When it thought about behavior, it was very much stimulus response as the framework, as opposed to kind of ongoing dynamics of a neural system and more of a continuous control problem. And so we're trying to just speak to that part of the literature and say it's not just something bumps into you and you respond to it, but there's constant action of the nervous system on the body that's taking place in organisms. And we should take that into account when we're thinking about what the early nervous systems looked like. [00:43:12] Speaker B: You brought up the heterogeneity of the parts. So there's the famous. In my little complexity discussion group, we're going to get to this more is different paper. I had to be careful when I say that because your calling card is different is more. You don't seem that pleased with that little turn of phrase. But it's good advertising, right? [00:43:36] Speaker A: Yeah, yeah. It just popped into my head when I was writing. Yeah. I think this goes back to this dissatisfaction I had when I was first learning about reservoir computing, that we were relying on firing rates to do everything, and that this didn't feel like what biological systems were doing. I remember when I was a grad student early on, asking some of the experimentalists at. At Columbia what neuromodulators were and, like, what they did and how many of them they were. [00:44:09] Speaker B: And it was like, what'd they say? Oh, they just raised their head. [00:44:12] Speaker A: I mean, a lot. It was just this big unknown when I was in grad school. Like, I knew there were some of them out there. I didn't know what they did. [00:44:20] Speaker B: They seem to have seemed kind of inconvenient in my. Yeah, just from my growing up in neuroscience. It's like, well, we'll get to them one day. But the important thing is the neurons and how often they're firing. [00:44:32] Speaker A: Right. And the spikes. And I think it's a bit of a drunk in the line post thing that the spikes is what you can measure. And so let's describe everything in terms of spikes and call it a day. [00:44:41] Speaker B: Right. [00:44:42] Speaker A: There's been this push for foundation models of the brain lately where you just record all the spikes and all the conditions and then you're done. But the spike is telling you that someone is talking, it's not telling you what they're saying. And it's. It's something that I started to become aware of when I was a postdoc and trying to get the hang of hypothalamus as pretty different from how I'd grown up thinking about neurons. And the more I read about it, the more kind of fascinating opportunity for theory. I think there is there that there's so many ways for neurons to communicate with each other. You really have not one weight matrix, but many based on these different communication channels. I think that there's a clear importance of neuromodulators and neuropeptides in brain function that we've found interesting effects of these in sort of providing a. Almost like a context signal, something that reshapes the dynamics of a neural network to do something different depending on the condition that this signaling molecule is reflecting. I'm thinking here, for example, I've done some work recently on Neuropeptide Y, which is released by neurons in the arcuate nucleus. When animals are food restricted, these arcuate nucleus neurons send projections to different parts of the brain and start to release npy. When you're food restricted, there's a projection to the medial preoptic area that reduces fertility and delays onset of puberty. This is sort of saying if you're starving, it's not a good idea to get pregnant. And so maybe wait for a little bit. [00:46:35] Speaker B: This is in mice that we're talking. [00:46:37] Speaker A: This is in mice? [00:46:37] Speaker B: Yeah, yeah. [00:46:39] Speaker A: When mice are pregnant, they. I forget if it's like double or triple their calorie intake. And then when they're lactating, they lose 30% of the calcium in their bones just to produce milk. So like it's a. [00:46:51] Speaker B: This is a neuropeptide Y, this is. [00:46:54] Speaker A: Pregnancy and child puppery that your body undergoes these like huge physical demands. [00:47:00] Speaker B: Yeah, sure. [00:47:00] Speaker A: And if you're starving, you're not gonna do so hot if you get pregnant. [00:47:05] Speaker B: Right. [00:47:06] Speaker A: So NPY is kind of trying to reduce the likelihood of that. And then my collaborator Nick Bentley found that there's a separate set of NPY releasing neurons that project to parabrachial nucleus and shut out a lot of chronic pain signals. So inflammatory pain, sort of long lasting nerve pain. Animals that are food restricted don't respond to these pain signals the way they do if they're fed. You can give a mouse a formalin injection in its hind paw and get this big inflammatory response in the paw. But if the mouse is starving, it just completely ignores it and behaves as if it's not experiencing the pain. [00:47:45] Speaker B: Because the story there would be that it needs to find food and other things are important then. [00:47:49] Speaker A: Yeah, if you're starving, if you're in pain, you go back to your nest and lick your wounds and recuperate. But if you're starving and you do that, you're going to die. So you need to kind of turn down this thing and go out and do this other thing that's more important for survival. So NPY seems to play this role in just kind of preferentially biasing the animal away from some actions and towards other actions. There's a boatload of these things. My collaborator Moriel Zolikowski showed that chronic social isolation drives this brain and body wide upregulation of the neuropeptide tachykinin 2. Mice have signs of TAC2 overexpression in inhibitory interneurons and cortex. Cortex does something in subcortical regions in the spleen and the gonad. And she found that if she took a. And these isolated mice are also a lot more anxious and a lot more aggressive towards other mice. If she takes a group housed mouse and she over expresses Tac 2 in different areas, she can get just the anxiety phenotype or just the aggression phenotype. And if she blocks Tac2 overexpression in an isolated mouse, it behaves as if it's group housed. So this is sort of this, I don't know if you want to say inductive bias or contingency plan that's kind of built into the genome that's saying if you encounter certain circumstances, turn on production of this thing and turning on production of that thing somehow magically manages to change the dynamics of fast timescale computation to alter behavior. So that is really fascinating to me is how you can make a brain that effectively modulates itself so that it doesn't just do one thing, it can do many different things without at the same time breaking. Like you don't suddenly forget how to recognize objects. When you're socially isolated, a lot of stuff is fine, but then some things are very different. [00:49:55] Speaker B: So if you get your leg bit off by a shark when you're surfing, you can still see the surfboard to try to maintain your buoyancy. And you're not too worried about your leg being gone. You're focused on getting or maybe not. [00:50:07] Speaker A: Feeling the pain that you do when you get to the beach. You're very worried when you get to the beach. You're in panic mode. You've got adrenaline pumping and you're acting, but you're not prioritizing like obsessing over the pain. [00:50:20] Speaker B: But you said, okay, so the way, like the classic way. Right. Is of causality. So what I want to ask you about is causality. It sounds like you're naming these peptides and when they're generated, they have these massive effects on our cognition while still leaving some parts of our cognition intact. Right. But if you. Pain is absent when NPY is expressed. Right. At a certain amount, I think. I'm not sure if I'm mapping the names onto the function here, but yeah, that sounds like the system is fragile even though you just said it's robust. But how do you think of it causally? I mean, do we want to say that like that is a mechanistic account of function or is it you use the word that the peptides had a role in. Do we need to bring it back from strong Billiard ball causation to more of a context dependent, softer causality story. How do you think about causality in that way? [00:51:20] Speaker A: Yeah, causality is tricky. Okay, yeah, maybe I've been a bit billiard ball in the way of framing it and thinking about it. I mean, we do have results from perturbation experiments. For example, with the NPY story. If you just, if you take the neurons in parabrachial nucleus, they express the receptor for npy and you just inhibit those neurons artificially, you get the same effect as if the animal is food restricted. So you can go in and kind of apply a perturbation or if you block npy, you block the hunger induced suppression of pain responses. So we can manipulate these systems and we can ask, do they really produce the effect that we think they're doing? And that's a lot of the hard part of experimental neuroscience is like really convincing yourself that this correlation between NPY release and a change in pain coping behavior is a causal link. [00:52:25] Speaker B: Yeah, yeah. This makes me also think our neurosciences turn toward naturalistic behaviors. And what you're describing are naturalistic type behaviors. And yet in a truly ecologically fully naturalistic setting, mouse might be pregnant and hungry and running from a hawk. And you know all of these. So there's lots of different contexts going on. Whereas in the lab, you're like, well, we're gonna. It's in a. [00:52:56] Speaker A: Do one thing at a time. [00:52:57] Speaker B: We're gonna do one thing at a time. So it is reductive in that respect, I suppose. [00:53:01] Speaker A: Yeah. That was something that I think I mentioned in my commentary recently, is that in one sense you're studying a behavior that's high dimensional. The mouse can do anything. It's behaving spontaneously in another sense, in one of these essays. So like if you're studying aggression, you have a mouse in its home cage, you plonk another mouse in there, you watch them for 10 minutes and then you stop. In a sense, that's a very low dimensional thing to do. You're looking at the animal in a particular state with a single type of stimulus in a particular environment. So the dimensionality of the thing that you're studying really depends on where in the brain you're studying it. Like if I was studying motor control, I'd see a lot of different types of actions during that resident intruder essay. If you're studying motivational states, you've got a pretty simple single axis of variation that you're looking at. And I've started to see work looking at how these motivational states interact with each other. [00:54:03] Speaker B: Individually within an individual brain interact with each other. [00:54:07] Speaker A: So I guess the work with Nick would be an interaction of hunger and pain, or pain really is. It's not nociception. It really is a motivational and affective thing. We have a follow up project where we're looking at interactions between hunger and predator threat, which is something that other folks have looked at. Interactions between hunger and aggression. So people have started to look at these pairwise interactions to really go to the full world of an animal and like all of the states that it can be in really would require. [00:54:40] Speaker B: That's impossible. [00:54:42] Speaker A: Yeah. It would take careful experimental design to collect a data set that's really going. [00:54:48] Speaker B: To be usable to answer that question. [00:54:54] Speaker A: To study those things. [00:54:55] Speaker B: Totally ecologically valid sense. Yeah, but it's at the same time it's, you know, like if you have a, the perfect model of a cat is a cat. Right. So I, you know, you're always abstracting. What's that? [00:55:07] Speaker A: Preferably the same cat. [00:55:09] Speaker B: Preferably the same cat. But you know, with models or with, with theory you're always abstracting. So there's never going to be the perfect. You don't just want to reproduce everything you have to abstract and to say something about the, the process that you're interested in. So I'm not sure that you would want to reproduce the full umwelt of the, of the affordances for the animal. [00:55:31] Speaker A: Yeah, I think in practice you move one system at a time, one project at a time and try to explain it and understand it and then make some predictions about how it might interact with some other system. Like it's, it's an iterative process. [00:55:51] Speaker B: Okay. So when you were describing the action of the neuropeptides and their downstream effects, that word causality is linked with effects. And I just said effects. Anyway, you started talking about, you mentioned in there how these sort of control signals and the subcortical processes are sort of context dependent. Right. And sort of can sculpt or shape the way that neural activity happens in other brain areas. [00:56:23] Speaker A: Yeah. [00:56:23] Speaker B: One way of thinking about that. So you know, there's the whole manifold push. Everything's a manifold these days in low dimensional population studies of neural activity. And the word subspace has become more popular now. And so therefore I think of everything where you can have like one population of neurons and a bunch of different subspaces, which means that like different pools of neurons are acting in different ways depending on what kind of internal dynamics they have and what kind of input they're receiving. So do you think of it in terms of like, subspaces of, you know, orthogonal population activity, different manifolds in the state of possible manifolds, or do you think of it as just like a gain, sort of increasing the firing rates of everything, decreasing, that sort of thing? [00:57:15] Speaker A: Yeah. So I think this is where there's room for theory is just figuring out what is the space of things that you can do when you're modulating activity of a pool of neurons. If you're dumping on a neuropeptide, are you just. There's a set of cells that's responsive now that wasn't responsive previously. So you kind of just recruited new cells into your computing population. Are you changing the excitability of everybody? It's something that I think theory can really contribute to. We've done a little bit of work in this space thinking about heterogeneous neural populations and how adjusting the heterogeneity of the excitability of neurons can change the way they compute as a reservoir. So my postdoc, Richard Guest worked on this. We added a pretty simple form of heterogeneity to neurons, which is just. We said, we have a population of spiking cells, and rather than all having the same threshold, we have a distribution of spike thresholds. Some cells are a little bit easier to excite, some cells are a little bit harder to excite. And then we asked, as you change the width of this distribution, how does it change the dynamics of the population? We initially did this thinking about just variation within a cell type and the fact that if you look at a given population of cells, they're not all the same. There's some variance to them. But I guess we talked to Luca Mazucato, who's been thinking about a very similar model for understanding the effects of acetylcholine on cortical dynamics. And Luca pointed out, which I agreed completely, that this distribution of spike thresholds could really be changing on the fly depending on availability of some ligand. Like if you dump on. If you dump acetylcholine onto cortex, some cells get depolarized and some cells get hyperpolarized. The more you jump on, the more this distribution broadens. So you can really change this, the width of the spike threshold distribution on the fly using a neuromodulatory signal. In Richard's work, what he did is he took the spiking network of heterogeneous neurons and derived a set of a mean field model that described just the firing rate of the population and Its behavior. With this reduced mean field model, he can perform bifurcation analysis and look at the different computational regimes that you can push the system into, depending on how much input you're feeding it and depending on the degree of spike threshold heterogeneity of the population. And it was really fascinating. Like, you can produce pretty profound effects on just the transfer function of your neural population. By adjusting spike threshold heterogeneity across the pool of neurons in a pool of excitatory cells, you can linearize their responses. This makes it makes activity higher dimensional and better for things like function generation, but it reduces this nice bistable regime that homogeneous network has that's really useful for things like working memory. [01:00:31] Speaker B: So in a super high heterogeneous regime, you push it toward a linear regime. [01:00:40] Speaker A: Yeah, a linear input output transfer function. Transfer function, it becomes higher dimensional, can be entrained by a wider range of frequencies. It has richer repertoire of. [01:00:56] Speaker B: Oh yeah, well, capacity maybe. Right. But, but isn't it like, if you push it too far, then you can't do anything because it's too high dimensional? Am I reading that right? [01:01:06] Speaker A: I mean, yeah, if you push it too far, you have neurons that are spiking all the time and neurons that are silent. So you kind of get into this extreme regime where your spike thresholds are so crazy that you just lose cells in your effective computational population. But there's a pretty decent range of distribution widths that can give you very different behaviors. [01:01:31] Speaker B: So it's not a one sweet spot, it's a range of. It's a sweet range. [01:01:37] Speaker A: Yeah, I mean, it's just, it's another knob that you can turn to change what your system does. [01:01:43] Speaker B: Okay. [01:01:43] Speaker A: And if you do this in the inhibitory cells, it's very different. If your inhibitory neurons are all the same spike threshold, are all carbon copies of each other, you very often fall into these bursting regimes of the network where you get sort of like epileptiform activity, making the inhibitory neurons more heterogeneous, unmasks the bifurcation structure of the excitatory cells. So if you're doing this on the fly, you can do maybe a sort of gating of information where by changing how variable your excitatory cells, your inhibitory cells are some, if you make them more heterogeneous now the transfer function of your excitatory cells shines through and then you switch to make them more homogeneous and you'd lose that transfer function and you fall back into the bursting regime. So it's these means of controlling Computation in neural population without fine tuning of synaptic weights and supervised learning. It's just this thing that you could do out of the box. [01:02:46] Speaker B: I mean, I was thinking about subcortical processes as dials, knobs, things like that. So this reminds me of Max Schein's work from a few years ago as well, where he essentially, he and Michael Breakspear, whom will be actually tonight, I released that episode, who were on my podcast recently, but they controlled the gain essentially, kind of, I guess the spiking threshold essentially in their model of populations of neurons. And found, you know, there are these scale free dynamics and there's like these sweet spots. And so it reminds me a little bit of your work. In his case, he was mapping it onto the ascending arousal system. Right. And so was it acetylchol? No, it's noradrenaline that like, if you turn up or down the spiking threshold of neurons with the neuroadrenaline projections or whatever, you get that same sort of sweet range. Not the same sort of, but a sweet range within which there's lots that you can do high capacity and, and you can think of the ascending arousal system as a control knob in that sense. Are all these subcortical processes, do you think of them as control knobs? You know, how do you think of them? Is this a cybernetics kind of view of subcortical processes? [01:04:02] Speaker A: Yeah, I mean, I think it's. It's an open question of how. How high dimensional these control knobs are. [01:04:12] Speaker B: How high dimensional are they? [01:04:15] Speaker A: I mean, we've got a lot of these, got a lot of channels to communicate with. [01:04:19] Speaker B: Yeah. [01:04:22] Speaker A: Although a fairly small number of ways that those channels can influence postsynaptic cells. We have hundreds of G protein coupled receptors. But then what happens within the cell is fairly restricted. So it's this interesting. I think that the reason for this is that it gives you specificity. [01:04:43] Speaker B: Can you elaborate on that? You have a bunch of different types of receptors on the membrane of a cell that's receiving. [01:04:49] Speaker A: Yeah. [01:04:50] Speaker B: Input from some sort of subcortical structure. G protein coupled receptors. But then. So that seems like a high diversity. But then what you're saying is like internally there's a, there's a bottleneck sort of. Of what can happen. [01:05:01] Speaker A: Yeah, yeah. So inside the cell, G protein coupled receptor, you have the receptor part which is specific to its ligand or not. Well, has a set of ligands that it can bind. And then inside the cell you have the G alpha, G beta, G gamma subunits. And when the receptor gets Bound, these subunits separate and go off and do their thing. Inside the cell, there's way more of the receptor types than there are subunit types. And so once you activate the subunits, mostly they all converge on modulating the amount of cyclic amp inside of the cell. [01:05:32] Speaker B: Why do you want that dimensionality reduction? I'm sorry to interrupt. [01:05:35] Speaker A: Sorry. [01:05:36] Speaker B: I'm sorry to interrupt. Why do you want that dimensionality reduction? [01:05:40] Speaker A: It's a good question. It's still an open question. How much of a dimensionality reduction it is. Is cyclic amp like a pool within the cell, or are there compartments where you have cyclic amp in one part versus another and they can talk to, they can function independently of each other. It's the same thing with, like everything in biology communicates through calcium. So how do you use calcium to modulate action potentials and transcription and everything else or the compartments? What is that? That's always bothered me that you have this capacity, this, that everything kind of just runs through this bottleneck of calcium ions inside neuron, inside the cells. In terms of the diversity of receptors, I think that there's this concept in evolutionary biology called Ono's dilemma, which is if you have a gene that does X and you need it to do something else, this is. Need is like implying some agency that's not there. But you have a gene that does X, you'd like it to also do Y. It's hard to mutate it to do Y without breaking X. And so what you do is you make a copy of the gene and you mutate the copy. Maybe you express it in a different part of the brain. Maybe you change its binding affinity for a ligand or change it to prefer a different ligand entirely. [01:07:00] Speaker B: Evolution does this, you're saying, or this. [01:07:03] Speaker A: Is within lifetime evolution. [01:07:05] Speaker B: Evolution, okay. [01:07:07] Speaker A: So it is a way to take something that is useful, like a cell surface receptor, and increase kind of the range of things you can hook that receptor up to. So you're recruiting the same downstream signaling pathway, but now you can do it in a different subset of cells. By turning on this receptor, you can do it with a different threshold for activation or a different timescale of activation. And so we see dozens of receptors for serotonin or acetylcholine or at least handfuls of receptors for a lot of neuropeptides. That could be a way of achieving specificity, of making sure that the right subset of cells is listening to those signals. But then what you do once you get the signal is pretty consistent and conserved. So that's my hope, is that it feels like a big biological mess. But then the range of things that you're doing to modulate your cells is fairly small. [01:08:09] Speaker B: It is a big biological mess. I mean, it's mind spinning. So I'm glad people like you are working on it. I mean, I don't remember. My mind can only contain three acronyms. After that, I'm done. And so the world of, you know, neuropeptides is. I don't even. I don't. Wouldn't even know how to begin. [01:08:29] Speaker A: Well, but the hope is it doesn't matter. Like really what it is is just which entries of your weight matrix are non zero. And then like the cell can communicate with this subset of cells. Once it communicates, the thing those cells do is fairly constrained. So it's. It's again, it's like an. It would be a way of baking in connectivity to the brain without needing to learn it through experience. If I have one set of cells that releases a certain molecule and another set of cells that can listen to that molecule, I can form interaction between those two populations without needing to guide an axon to a dendrite in a very specific way. [01:09:14] Speaker B: Right. Okay. [01:09:15] Speaker A: And if I make the expression of the signaling molecule sensitive to animal state, to food restriction or social isolation or stress, then I could have some sort of the. The languages in evolutionary biology is phenotypic plasticity. So you have phenotypes that are contingency plans, like circuits and patterns of interaction in the brain that only become available when you need them. [01:09:42] Speaker B: Okay. On. Yeah, on the fly. So are these types of things that we're talking about. This is what you're partially pointing to in the commentary that you wrote about theoretical neuroscience having room to grow. So again, just going back. Traditionally, theoretical neuroscience has conceived of brains as these sort of symbolic computational processes doing tasks. Right. But you're really bringing it into a. The naturalistic world now, like the rest of the field is pushing it into these ecologically valid kinds of behaviors, et cetera. My microphone just fell down. Hold on one sec. [01:10:26] Speaker A: First. [01:10:30] Speaker B: But also you're looking at. Across levels and down into the subcellular processes and the way that these subcellular networks of communication and signaling. So is that where you see, see, theoretical neuroscience has room to grow? [01:10:47] Speaker A: I think it's one of the places. Yeah. That there's just a lot of room for theory and hypothesis generation here. How do you want these signaling systems to be organized? Do you want like a. The arcuate nucleus? Seems like it has this kind of hub and spoke deal going on. Where you send different projections to different targets that you want to modulate by what's the arcuate? That's where your hunger activated neurons are, the ones that release neuropeptide Y. Okay, so it's one part in the brain that broadcasts the signal to many areas. Is that how you always want to do it? If you look at the Tachycanine 2 system, it seems very different. The cells that express Tac2 in isolated animals are like all over the place. Like all of your inhibitory interneurons are producing it. So there's no one hub that's sending out a signal. It's genes that turn on in different sets of cells, maybe under different circumstances. So there's room to just think about what principles could be shaping these systems, how you would best use them if you wanted to use them to make your neural networks more flexible and expressive. And also how do you use them in a way that doesn't break what the networks are supposed to be doing in the first place? [01:12:05] Speaker B: You mentioned the word principles. Are they going to be the same sort of sets of. So I hesitate to return to dynamical systems view because I talk about it a lot. Right. But you can apply the dynamical systems to like any network of interacting parts. And that applies to traditionally what we do in the neurosciences, like the spiking neural networks, but it also applies to molecular networks, to subcellular protein interaction networks, et cetera. I mean, are the same principles going to apply in those systems like. Or will. There'll be a new set of principles depending on the level that we're looking at. [01:12:47] Speaker A: Yeah, I think we can. If we take these, like very molecular subcellular things, we can still think about principles of using them on a circuit scale. So for example, in my. My work that's impressed with Nick Bentley, this hunger suppression of pain. [01:13:07] Speaker B: Well, it's. I thought you. Isn't it accepted now? Is that. [01:13:10] Speaker A: Yes. [01:13:10] Speaker B: You mentioned congratulations officially, by the way. [01:13:13] Speaker A: Yeah, excited. So you have this release of neuropeptide Y when you're hungry and it's blocking pain. We. You can ask computationally, what is that? What is happening here? Should we think about this as. The first way that we approached this was like, oh, you're. When you're responding to pain, you're kind of like expending effort to, in the case of the mice, just like freeze or lick your wounds. And you're doing that instead of doing other things like foraging. And so if you're hungry, maybe you have a part of your brain that is sort of like comparing how hungry you are to how much pain you have and making a decision based on these relative levels. And when you're hungry, like that hunger bar is higher. So that could be one way of thinking about this signal is going into some comparator that's deciding what to do. [01:14:06] Speaker B: Do you think it's that simple? [01:14:08] Speaker A: Well, that was, that was what we were hoping we'd find going into it. But you can imagine it working other ways too. Like you have some motor control part of your brain that's saying, I'm in pain, I need to lick my wounds and stop foraging. And hunger could be just shutting off the output of that. Saying other stuff is more important. Don't activate these motor programs right now. You've got other stuff to do also seems like it makes sense like a stop signal. Yeah. And then the other thing that we thought about is you could be blocking the pain signal from getting to the brain in the first place. So NPY could be shutting out the pain signals so that if you have your policy, this abstract distributed the process by which the brain decides what to do next. If you just don't let the brain know that the pain is there in the first place, then it does other stuff. We built a little toy RL agent that had a pain state and an effort state and a policy. And we said which of these things should hunger be modulating? And the only thing that really matched this reduction in the pain coping behavior was blocking the input to the system. The other things like because the persistent, the hunger and because the pain input was so persistent, if you like raised your threshold to start responding to it, you'd like wait another couple seconds and then you'd start responding. Or if you increase the cost or blocked the action, you just start spamming that action until you got one through. So gating the input was a much easier way to control what our system did compared to controlling the policy or the output. And it also makes different predictions about what the animal is feeling. Because if we took a model and we blocked the motor output, it was actually in more pain than in the. [01:15:59] Speaker B: Sated case as assessed by behavioral observation, Right? [01:16:04] Speaker A: Well, not by behavior, like the behavior was the same, but if you look at the internal variables of pain and efforts that were used in making a decision, your pain would get higher before you would rise to a higher point before you're producing the actions. And the same was true if we messed with the policy itself. We're basically saying you have to be in more pain before you act. So if you see an action, it's actually communicating more underlying pain. Whereas if you're gating out input, then your underlying pain state is actually lower. And when we look, when Nick looked at parabrachial nucleus neurons, there were these cells that in a sated mouse had this like long lasted, lasting, persistent activity during the inflammatory pain state that were just gone in the food restricted animals. So it really seems like this is the percept of pain maybe. Yeah, we didn't write that in the paper. [01:16:56] Speaker B: I thought it was C fibers. Isn't it supposed to be C fibers? [01:16:59] Speaker A: Nociception. Nociception and pain are different. [01:17:02] Speaker B: Yeah. Yeah. Okay. Yeah. [01:17:04] Speaker A: But then you could say, hey, I'm going to develop a new drug and I'm going to look at this population of neurons to see how much pain my animal's in. I don't have to watch it or I don't have to prick it in the paw to see if it's in pain or not. So it actually has huge implications for the field of chronic pain treatments because now you have a neural basis of the state that before we had to really do something to the animal to read out. And a person who's feeling chronic pain, it's not that you pin prick them and they say ouch. They're feeling it even if they're not communicating it. [01:17:37] Speaker B: Yeah. This is the point of the podcast where we take a break for a commercial and I sell NPY to the tech bro menosphere. The other thing that reminds me of is that similar cutting it off at the pass in the psychological literature of habit formation and changing habits. The way to effectively change habits is so you have the cue, you have your response to the cue and then you have the reward, which is the high that you feel or whatever, if it's drug related, etc. There's that loop or whatever. And if you take the cue away, that's like a super effective, effective way. You know, don't go to the, don't go under the bridge for the heroin or whatever. Just stay away from the bridge. So that's the same sort of principle, I suppose, at the neuro level. [01:18:25] Speaker A: Yeah. And it seems like it's a possibly a general thing. Like there's other cases where we see a change in animal behavior that's been linked to a change in sensory processing. For example, Catherine Dulac has this beautiful work on pup directed aggression in male mice. Male mice are typically infanticidal. If they encounter a mouse pup, they'll attack it. But if they encounter that mouse pup at A time that is like one mouse pregnancy. After they've mated so that it could be their pup, there's a shift in their behavior, and they show more parenting behavior as opposed to just attacking the pup. And the shift in behavior is linked to a change in the processing of sensory cues from the pheromonal cues. So, like, when I went into this, I was thinking, oh, the. The hypothalamus is this big policy. It's weighing these things off against each other and making a decision about what to do at a given moment. And there's magic and fanciness happening there. It could be that really the best way to change your behavior isn't to change the interaction of these drives or these. These need states is to change the way you're sensing your environment. [01:19:37] Speaker B: Elaborate on that. Just so I just. So it's like, clear in my head, it's not the interactions of the drive states within the hypothalamus. It is the way that you're reacting to the sensory states. [01:19:47] Speaker A: Yeah. At least that's how it worked out for this hunger and pain story, is that it was easier to change the behavior of our little RL agent by changing its input as opposed to, like, trying to manipulate the policy or the output. [01:20:00] Speaker B: Okay, wait. Actually, before we move on, I didn't ask you this. And the listeners need to know, what is the hypothalamus known for? What is it traditionally? What do we think that the hypothalamus does? [01:20:13] Speaker A: Right. There is a whole bunch of things. There's people who study it for its role in growth and development and puberty and control of hormone release. There are people who look at it for its role in feeding and metabolism. And then there are parts of hypothalamus that are involved in things like predator defense, reproductive behavior, aggression behavior. So in general, it's very deep in the brain. It's this set of interconnected nuclei. They express all sorts of different signaling molecules and receptors, and they seem to be involved in regulation and control of survival behaviors. [01:21:02] Speaker B: Okay, so that's what I was going to get at is I want to get your thoughts on the connection between these subcortical processes, the things that you study, these survival behaviors, pain motivation, hunger, these things that are traditionally thought of as sort of these basal level cognition, almost. Not even cognition. It's almost like this is just. This will take over if you need it. But what's beautiful about humans and some mammals, you know, is that we can ponder. We can. That's what our cortex is for. Right. So we have this, the beautiful thinking processes and we can imagine scenarios and simulate in our minds and we have models of the world, etc. How do you. And yet these, these lower subcortical. Lower subcortical processes, subcortical structures are so powerful to our behavior and they're affecting on our ongoing behavior. [01:21:53] Speaker A: Yeah. [01:21:54] Speaker B: And if, if I get my leg, my leg bit off by a shark, I will see the surfboard that my visual cortex is still. My visual perception is still working, but I probably won't be writing poetry in my head in that instance. Right. So these higher cognitive functions aren't really happening in general. So how do you think of that kind of cognition across the spectrum of what we consider cognition? Sorry, that was long winded. [01:22:18] Speaker A: I mean, it's almost a sort of like working memory for your needs. Like, it keeps track of like, how hungry am I, how thirsty am I, how stressed am I? Have I been in a fight recently? Are there other mice people around that I am stressed out about or want to mate with or want to interact with? [01:22:41] Speaker B: Did you say mice people? [01:22:42] Speaker A: Mice or people? [01:22:43] Speaker B: Mice or people? [01:22:44] Speaker A: I don't know. It informs our actions, but it doesn't command them. Right. [01:22:53] Speaker B: You can override them, you mean? [01:22:55] Speaker A: Yeah, there's this really cool case. There's this woman in Scotland who. She's one of these people who doesn't feel pain. She has a mutation in the gene for fatty acid amine hydrolase. Hydroxylase, something. She doesn't feel pain, but she also reports that she's never felt angry or afraid. She's just like super chill. And this is a gene that's expressed in hypothalamus. And so like she's still a person, but she doesn't have this sort of like persistence of these drives. Is like aggressive and anxiety related drives the way that you or I do. It seems like it's very much suppressed in her. Maybe because this mutation is reducing the ability of her hypothalamus to kind of rev up its activity and produce persistent firing and persistent motivational states. That's my hypothesis. [01:23:56] Speaker B: Well, did she. Otherwise she seems like completely. I mean, she's not some excellent poet. Right? Or some. I mean, is she like otherwise normal IQ, etc. Cognitive? [01:24:06] Speaker A: Yep. Just very chill. She gave birth and was like. That was weird. Like she. She can recognize when people are in pain or are distressed. But it's like a cognitive thing for her. There's no like gut feeling. Yeah. So the. It's emotions that are affected as opposed to like thinking and actions. [01:24:28] Speaker B: The whole embodied cognition Movement, the four E's. I can never remember which naming the E's. I can never remember all of their names at once. But we're going to call them the four E's. But it's. The push is that our cognition is connected to our bodies. Right. And we shouldn't think of our brains as like separate from the bodies. We shouldn't think of our brains as manifesting our cognition because everything we do is part of all the same system connected to our bodies. And the things that you study seem to be close to what's going on in our bodies, which are signaling these brain areas to spit out these peptides that affect our ongoing behavior and cognition. So where are you in the brain? Body dichotomy? Non dichotomy. [01:25:14] Speaker A: Is it a dichotomy? Do I have to pick one? [01:25:16] Speaker B: No, no. Well, that's what I'm saying is maybe it's a non dichotomy. The 4E people would say that there's no separation, right? [01:25:22] Speaker A: Yeah, I'll say the hypothalamus is in a kind of privileged position. It's. I forget the name of the thing. But there's like a part of the brain where the blood brain barrier is reduced and you get like the brain is able to sense signals in the blood. Hypothalamus is like sitting right there on that. [01:25:40] Speaker B: Oh, I didn't know that. [01:25:41] Speaker A: So it picks up a lot of signals from the blood that don't make it to other parts. Brain, the of. Of the brain that is. I think it's involved in, especially in things like hunger and thirst and sensing your nutritional need state, but also lets you sense circulating hormones and other things. Yeah, Brain body interactions is a very hot topic right now. And I think it's a really fascinating one. That's again, not a place where we've had a big neurothery presence in the past. Yeah. Going back to what I was saying earlier about reservoir computing, the body is an excellent way to have memories, to have long time scales that can influence neural computation. You have circulating signals that represent the state of the body, available nutrients, signals about past physical activity that can signal from the body to the brain certain need states. I think the role of hypothalamus to some extent is sensing those signals and inferring need states and using that to then broadcast to other parts of the brain what I need to do or what, what I should be prioritizing right now. Should I be paying attention to food smells? Should I be paying attention to other conspecifics around me. [01:27:02] Speaker B: Part of what you've been doing also. So just kind of changing topics as we get close to the end here is the modeling of behavior. So tell me more about what you're doing in that vein. And you're releasing a data set, if I understand correctly. [01:27:17] Speaker A: Yeah. Pretty soon. I first got involved in this just because my postdoctoral lab wanted to automate scoring of mouse behavior. Because when you're studying social behavior to interpret neural activity, you need to know what the mice are doing. And when I started, that was done by very patient postdocs or technicians going through and watching videos of mice interact and just manually labeling frame by frame what they're up to. Yeah, and I think that's still the case in a lot of labs. So like initially it was just making life easier for the study of behavior. And we worked for a while on computer vision systems for pose estimation and supervised behavior classification. It's still a thing that we're interested from that making life for scientists easier perspective. But I think there's also just things to be learned from studying the dynamics of animal behavior in a more quantitative computational way. And that's obviously a field that's really blown up over the past 10 years. [01:28:24] Speaker B: Yeah, there's all sorts of annotation, automated annotation tools right now. [01:28:29] Speaker A: Yeah. So we a couple years back started running into this problem where everybody who is publishing a paper on behavior classification or unsupervised behavior segmentation, they make their own in house data set and they show that their method works really well on that data set and nobody can really evaluate how well it's doing. [01:28:49] Speaker B: Sure. [01:28:50] Speaker A: And so my collaborator at Caltech, Pietro Perona, had this long history of putting out benchmark data sets for the fields of computer vision and machine learning. His former trainee, Fei Fei Li, put out imagenet, which clearly made a huge difference in the field of computer vision. So we wanted to do something similar, put out benchmark data sets for the field of behavioral neuroscience. We've run a series of multi agent behavior challenges where teams would be tasked with solving certain computational problems in mouse behavioral data sets. The first of these was just the same thing that we did in our mouse behavior paper. We published the tracking data set of interacting mice and we said make us classifiers that can detect when animals are sniffing, mating or fighting. And when you have enough training data, you can do about as well as another human being. At this point, we released a second data set, maybe 22, which was aimed at unsupervised behavior analysis and representation learning. So the way that we. It's. It's hard to evaluate how good an unsupervised model is. You get some segmentation of your behavior trajectories and you're like, well, cool. [01:30:14] Speaker B: So what we said without like hand scoring the whole thing, you mean? Or. [01:30:18] Speaker A: Yeah. Like, is the gold. Yeah. Is this representation of behavior useful? Like, if you did a bad job on unsupervised segmentation, you might get clusters which correspond to where is the mouse in the cage. [01:30:31] Speaker B: Because you didn't control position, which you don't care about. You care about what it's doing no matter where it is. [01:30:36] Speaker A: Yes, exactly. [01:30:38] Speaker B: At least in some experimental setups, sometimes you do care. [01:30:40] Speaker A: Yeah, yeah. So what is useful is tough to pin down. [01:30:45] Speaker B: Yeah. [01:30:47] Speaker A: So we put together this data set where we had tracking data from mice and flies and beetle interaction data set. And we had like a gauntlet of tasks that you could do on that data, like detect when flies are receiving optogenetic stimulation, classify the sex of the flies, detect when the flies are doing wing threats. So we kept those tasks hidden and we said to the teams, give us a representation of what the animals are doing. And then our scoring of that representation was, if we train a simple linear classifier for our hidden tasks, how well does it do given your representation? So that was maybe 22. And that that data set is now out there, the pose datasets, and then also just raw videos of animals. So this year we're trying to tackle this problem of generalizability of behavior classifiers. So if you're a mouse social neuroscience lab and you, you're studying sniffing behavior, you train a classifier to detect when two mice are sniffing each other. If you give that classifier to another lab and they run it in their setup, it's not going to work. It's going to completely break because the frame rate of the camera might be different, the placement of the camera is different, lighting is different. Yeah, yeah. Is out of distribution. And so we emailed a bunch of different social neuroscience labs, and 15 of them contributed data sets of top view videos of socially interacting mice, most of them with pose estimates. Some of them we tracked themselves and all tracked ourselves. And then all of them with manual frame by frame annotation for some behavior of interest. And across the whole set of labs, I think we have like 38 different behaviors that people care about that they've annotated. [01:32:41] Speaker B: And then you've used this to train a model and therefore generalize across labs. [01:32:45] Speaker A: But the goal of the competition is to say how well can you train a model and generalize across labs? Because in practice my lab has focused more on neural dynamics and less on machine learning tools. But we want to make these datasets available. So this will launch on Kaggle in a week or two. [01:33:04] Speaker B: Oh, okay. And it will have launched. You should send me the link. So it will have launched by the time this comes out then hopefully. [01:33:12] Speaker A: It's been a slow process. We'll see. We're hoping sometime this month of September. So anyone who's interested can work on their own method to detect each lab's behaviors of interest from their pose data. And we're hoping. Not sure, but probably if you train a model that can kind of fuse data from different labs and learn the internal dynamics of mouse behavior in a way that's invariant to how a particular lab's camera is positioned and which body parts they track, that you'll get better performance in recognizing these social behaviors that, that we want to detect. So we put together this data set for a pretty cut and dry supervised behavior classification problem. But we're also really interested in if you can really build lab invariant representations of dynamics of mouse behavior, what can you learn about how animals make decisions and what information informs their decisions during social interactions? Because if you have pose estimation, you have a decent proxy for the animal's sensory environment. I mean, it's not perfect, but you know where their head is. You can kind of predict what they're seeing. And so you can use this to say like, given what my animal is experiencing, its location and its social context, can I predict what it's going to do next? And we've, we've done this in single animal data sets. We're looking to move into this. Not published yet, but some work in progress. But we're really interested in this problem of forecasting and using forecasting as a way of understanding the sensory motor transformations that are shaping our social behaviors and the context of our histories that shape our, our social actions. [01:35:07] Speaker B: So what is forecasting in this case? What does that mean? [01:35:10] Speaker A: So forecasting is like given my animal, given its recent history, predict just where it's going to go next or what it's going to do next. [01:35:19] Speaker B: Yeah, but in terms in behavior world, that's called forecasting. [01:35:23] Speaker A: Yep. And it comes up all over the place in behavior world. It's one of the big limiting factors for self driving cars is being able to forecast what pedestrians and other cars are going to be doing. It's useful for say, human robot interactions. If you have a robot that's interacting with A person that needs to be able to predict how a person is going to approach it and interact with it. So it exists in a broad range of applications in machine learning and robotics. And we are interested in forecasting because we think that building a not necessarily biological, but a generative model that predicts how an animal is going to behave will tell us something about the sensory information that informs those decisions. And also what on what time scales, the experiences and histories of the animal matter for predicting what it's going to do next. [01:36:17] Speaker B: And you think that this is going to be possible based on like the slower time scales of the states, from subcortical processes or just from purely like thinking behaviorally? Why do you think this is going to work? [01:36:31] Speaker A: Well, I guess question is how well it will work. And where we fall short is just one thing to ask. How deterministic is mouse behavior? If you see where a mouse is, can you predict what it's going to do next? From talking to experimentalists in my old lab, it seems like you can tell when a mouse is getting ready to attack. Kind of become a lot more dirty. They're rattling their tail, their actions are changing. So you can probably infer state from moment to moment actions. But also if you know the history of an animal, that might influence your prediction of what it's going to do next. History both what it's done in the past couple minutes and history in terms of the sex and the strain and the experimental history of the mouse, which is information metadata that we have from the labs that provided these videos. [01:37:21] Speaker B: So that'll be available, that data set, the. The cowboy competition will be. Will have been released by the time this comes out. So I'll point people to that. What's. It seems like you're firing on all cylinders, obviously. What is. What's holding you back? Is it. Is there something on the theoretical side studying, like, you know, are there tools that's holding. What is holding you back on the theory side? What is holding you specifically back on the theory side? In the subcortical subcellular signaling domain? [01:38:02] Speaker A: I think it's maybe more data to constrain our thinking. [01:38:09] Speaker B: More data. We need more data. This is the big data day. [01:38:12] Speaker A: Cop out answer. [01:38:13] Speaker B: Yeah, no, I wasn't judging. I just. I was surprised. [01:38:21] Speaker A: We see what a given population of cells in a given brain area is doing in whatever context the experimenter decided to study them. But we don't know what those cells are doing in every other context. So it's hard to. It's just, it's very early days still, for charting what these cells respond to, there's been in cortical neuroscience, as mesoscale imaging has become possible, we've come to realize that these functions that we think are localized to specific brain areas, they actually have maybe not computations everywhere, but signatures of those functions are present everywhere. Choice. You can decode from many parts of cortex. Actions. You can decode from many parts of cortex. So being able to look at the cortex as a whole during a task has really changed the way we think about local cortical computation. And subcortically, you can't do mesoscale imaging. You have to look at one population at a time and you have to look at one behavior at a time. So I'm really curious in. I would love to see sort of a broader range of contexts in which a given set of cells is recorded. Basically higher dimensional data. [01:39:40] Speaker B: Higher. Higher even. Higher dimensional. Yeah. Okay. That's a good way to. [01:39:43] Speaker A: Not number of neurons, but like the different modalities. [01:39:47] Speaker B: Different. Yeah. Yeah. Okay, gotcha. [01:39:50] Speaker A: Because there's just a lot of unknowns. [01:39:51] Speaker B: A lot. Yeah. All right. There are a lot of unknowns. And we started talking about how one of the reasons you're studying what you're studying is because you saw it as an opportunity where the field was, didn't feel crowded. And some people thrive in different. People thrive in different environments. Research type question environments. But if you were starting over these days. So what I'm going to ask you is like, what advice you would give to someone who was beginning their career, like what they should get into. But in some sense in this, like in the theoretical neuroscience commentary that you've written, you're sort of defining that space and saying, well, here's a good place to go to. But I don't want to answer for you or put words in your mouth. So what would you do if you were thinking about getting into theoretical neuroscience these days? [01:40:43] Speaker A: Yeah, I think there's which problems you work on and then there's also which skills you develop. And I think the skills have changed a lot from when I was a graduate student. Fitting models is no longer like a manual adjusting parameters until your model does what you want it to do and call it a day. We have tools for simulation based inference that let you really see the whole space of performances of a model and explore the capacity of a system for degeneracy and redundancy in solutions. So I think just building your skill set is the most important thing you can do as a graduate student. [01:41:28] Speaker B: Do you mean the breadth of skills. What do you mean? Skill set in what? [01:41:34] Speaker A: Building your technical skills? Yeah, I mean, you don't have to do everything, and I think it's easy to fall into a mindset where you just have to know more math than everyone around you and learn more math than everyone around you. Like that. You never get to the applications. I mean, you have to strike a compromise. If you specialize in the biology too early, you might miss out on some of the computational skills you need to address biological questions that interest you. I don't. I think, yeah, it's hard to give general advice because it really depends on where someone's coming from. [01:42:20] Speaker B: Sure. Okay. Anne, thank you so much for spending so much time with me. I'm happy for you that you have found what seems to be a huge space of possibilities for your future endeavors. So I appreciate your time. [01:42:33] Speaker A: Thank you so much. [01:42:41] Speaker B: Brain Inspired Brain brainstried 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.