Avanish Srinivasan: Have you ever wondered how billions of neurons somehow work together to create thoughts, memories, decisions, and even consciousness in itself? Why does the brain sometimes behave in predictable patterns? Well, at other times it seems almost chaotic. How can math and computation help us understand something as complex as the human mind? These questions sit at the center of one of the fastest growing areas in neuroscience, computational neurodynamics. From the synchronized brain activity to neural networks, from decision making to disorders of the brain, scientists are beginning to uncover the hidden rules that govern how neurons interact over time and how these interactions give rise to behavior, cognition, and learning. Hello, I'm Avnish and this is The Early Perspective, a podcast where we explore how the brain learns, how it works, and how it changes over time. Especially how what that means for how we learn, think and live. Today we're going to be diving into one of the most fascinating topics in neuroscience today computational work and neurodynamics. An area of research that combines neuroscience, mathematics, physics, and even computer science to better understand how the brain functions as a complex system. When most people think about neuroscience, they picture brain scans or individual neurons firing. But computational neuroscience asks something deeper. How does Enormous networks of neurons coordinate activity. How does the brain process information so efficiently? And can understanding these dynamics help us better understand intelligence, behavior, and neurological disease? I'm incredibly excited to welcome today's guest, Dr. Brent Doiron. Dr. Doiron is the pr is a professor of neurobiology and statistics at the University of Chicago, where he also serves as the chair of the Committee on Computational Neuroscience. He originally studied physics and mathematics at the University of Ottawa before earning a PhD in physics. He then transitioned into theoretical neuroscience during his postdoc fellowship at the New York University. Over the course of his career from the University of Pittsburgh to the University of Chicago, his research is focused on understanding how large populations of neurons interact dynamically to process information in the brain. His work combines neuroscience, Mathematics, computation, and statistics to study topics like neural correlations, cortical networks, variability in brain activity, and how collective neurodynamics shape information processing. I'm really excited for this conversation to explore how neurodynamics can help us better understand the brain itself. Doctor Dorian, thank you so much for joining me today.
Brent Doiron: Thanks so much for having me. This is exciting. I I love reaching out to sort of younger audiences. Thanks for giving me the opportunity.
Avanish Srinivasan: Yeah, of course. Okay, so before we get straight into the conversation, I just want to go back to the beginning. Could you tell me a little bit more about your journey into neuroscience and computational research?
Brent Doiron: Sure. I started actually even as an undergrad at in nineteen, I think ninety-eight when w at the University of Ottawa, I got put on a project that the professor I was working with said was neuroscience. I had no idea about the brain, nothing like that. neuroscience at the computational side was sort of in its infancy at that time. So I started there and then I sort of got hooked. And then during and I absolutely a PhD with that same professor at the University of Ottawa, my PhD was in physics, but I I worked looking at neurocomputation from a physics point of view, specifically working on an area of neuroscience called neuroethology. Which is where they look at specialized organisms and look at neural computations that certain organisms have specialized. And the organism I was working on was the weakly electric fish. It's a classic neural prep where there are fish that produce electric fields that surround their body. And then when objects that have a different impedance than the water around the fish come in, then the field lines will morph and then on the surface of the skin of these animals. There's receptors that are able to see that. It's almost like having 3D vision, but it's in a fog because the electric fields decay quickly. the reason I think people studied this in the nervous system is there's sort of lots of control for this. We know how to manipulate electric fields. we can also re the these these animals produce electric fields when they're paralyzed. So you could study them in the sort of 80s and 90s. When people weren't really doing awake behaving, especially in fish, it's very difficult to do that. Now people are doing that. Anyway, that was how I started working on sort of mathematical modeling of that. And then during my postdoc, as you mentioned, then I sort of walked into sort of the broader neuroscience community at NYU and then found myself in mathematics departments through most of my career. but then now here I'm in a modified neuroscience department here and and statistics, but but I I'm I've been slowly over the years. becoming more and more of a neuroscientist. now that's how I think
Avanish Srinivasan: Yeah. Was there any specific moment that really pushed you into the field of neuroscience particularly was or was it just that project?
Brent Doiron: Well, I was I was interested in the nineties, people were really getting hot about complex systems in general. they've changed over the years, but in the eighties and nineties there was a lot of excitement about that. And one of the areas that I wanted to work on was actually social insect dynamics. So I wanted to look work on like how bees work as communities or ants work as communities, but I didn't have an option at University of Ottawa. And then sort of that's how I got convinced to say, Well, why don't you Try looking at neurons. Neurons are, you know, simple-ish agents. They're not actually that simple, but they work as a collective. and so then the mathematics, the kinds of mathematical techniques we use to study groups of neurons are similar to how we study sort of groups of you know different simple agents. And then that pushed me in the field. And then, you know, I got hooked, and I there was a there was a lot of excitement about comp neuro.
Avanish Srinivasan: Yeah.
Brent Doiron: in the sort of early two thousands and there continues to be today. So I felt fortunate to sort of jump on a bandwagon early.
Avanish Srinivasan: Yeah, I think it's very interesting that math and neuroscience really work together and complement each other in such a interesting way. I think a lot of people don't realize that they are so similar and they think that they're separate worlds. But I don't see we briefly touched on this, but what exactly is computational neuroscience? It combines the two, but how so?
Brent Doiron: Yeah, I think there's computational neuroscience is a broad term, and you know, it probably differs depending on who you're talking to. I think sort of there's two branches of this. One aspect of computational neuroscience is building new analysis tools given just this vast amount of data that just every year the amount of the brain that we can measure at one moment in time and the scales that we can measure that brain activity at, you know, change all the time. And we need to have quantitative tools to be able to catalog and analyze that data. And now the fact that people are really doing connectome dynamic connectomes where they're basically going in and measuring every possible connection in right now fly brains and and worms they work on. Those are massive data sets. So we need to have ways to distill or extract the salient aspects of these data sets. Mathematics is great at doing that and sort of machine learning which is sort of a branch of statistics here, really pushes that. That's one branch of computational neuroscience, helping people organize and interpret data. The other branch that I'm more invested in myself, in my research group, is modeling the nervous system. We know a lot about how single neurons act. We know how to model single neurons. actually, there's a Nobel Prize for this in the 60s from Hodgkin and Huxley. so having Electrical circuit models of neurons has a long tradition. And then so if we know how single neurons work and we know how synapses, the thing that connect neurons together, work, then it makes sense to start expanding our models to have groups of neurons that interact with one another. So these are more mathematical models of brain circuits, much more in the tradition of statistical physics, where we sort of have simplified, I don't want to say spherical cow, but you know, oversimplified models. And they're trying to ask what are the critical aspects of the neural activity or the connections between neurons that are needed to explain brain-wide neural activity or even behavior. So those are sort of circuit biological models. That's more the side I'm on. Of course, there's interactions between the two. We need to analyze big data sets to validate our models. So it's really silly to say these are separate areas, but there's a spectrum.
Avanish Srinivasan: Right. Mm-hmm. So then I know that you talked about Hodgkin and Huxley and that model, but then what exactly are these models telling us which are not I mean, at least which experiments can't say alone?
Brent Doiron: Well, so Hodgkin and Huxley, they were, they were actually physiologists. They were sort of they they was really some one of the earliest sort of gold standard models of of of neurons came from two physiologists that were you know very mathematically trained. So the Hotkin-Huxley model talks about how different ionic species, in particular potassium and sodium, how potassium and sodium interact with one another in terms of how they flow across a membrane. in order to produce the action potential, which is sort of the central currency of information in the brain, people call them spikes. So neurons produce these spikes. And Hodgkin and Huxley really figured out how these spikes happen. And it was really about an interaction between two different ionic species. Experimentally, you can isolate one species from another, but if you want to sort of understand how they both go together. That's experimentally very difficult to do. So what they did is they played this game where they would isolate these from one another, build a mathematical framework that when put together should produce an action potential. And then lo and behold, it did produce an action potential. The cool thing about what they they it was a it was sort of a complicated set of equations that now of course if you wanted to numerically solve these on your phone it would be trivial. It would be, you know, you know, milliseconds on your phone. But back then, Hodgkin and Huxley, they were doing these calculations in the late 40s, where if you wanted to do anything with a computer, you know, you had to be working with the military, the Manhattan Project, like there was no computers to really do these things. So Huxley, he was the the the student of the pair, he ended up doing this with a crank calculator.
Avanish Srinivasan: Yeah.
Brent Doiron: So he had like an accountants calculator with you know, and he and he did the numeric. So that took him weeks to do calculations that we would do now quickly. So these guys were super ahead of their time. but to answer your question about what this gains, we sat out we now know how different ionic species interact with one another to produce real activity, and that sort of gave the birth of pharmacology because we could manipulate. the the level of these ionic components with different drugs we can block components we can enhance components and sort of the birthplace of sort of pharmacology at the cellular and synaptic level was sort of at that time and that of course you know has a a history today pharmacology and and mental health have have a big component so I wouldn't necessarily say it all started with Hodgkin Huxley but understanding the the biology of neurons and how chemicals can affect that biology started with
Avanish Srinivasan: Right. So then you're saying that especially these models can provide us more insight that experiments alone can't just allow us to find, right?
Brent Doiron: That's right. So in today's world, it's cheap to do numerical analysis, much cheaper and easier than it is to do experimental work here. So you can use these models if they're well made and they're well constructed and well validated, then you can use them as a hypothesis generator for future experiments in a way that, you know, now you can do many, many, many numerical experiments cheaply and quickly, and then the sort of the the the most performing ones in the sense of manipulating neural activity might be the ones that then you would first try to do experiments to replicate.
Avanish Srinivasan: But now with the advent of better technology calculators and even better AI and machine learning, are these o models going more obsolete because we don't need at least the computational a at least the literal calculations which these models provided or at least let us, I guess, do it faster, right?
Brent Doiron: Mm-hmm. I think interpretability. I certainly think I can when you when we look at, for instance, the Hotchkin Huxley model, I can I can understand that solution in terms of ionic of ionic flow, in terms of sodium, potassium, in terms of of nonlinearities, in terms of of how channels that are holes in the membrane and how they act. This is a super interpretable, well-founded theory. If we're gonna start throwing AI
Avanish Srinivasan: Right.
Brent Doiron: And not asking AI to be interpretable, then I don't sure you might be able to predict things, but having an understanding of what's happening might be out the window. I think there has to be a good balance between these two. So you know, the classic modeling that, you know, physicists and and computational neuroscience at a later point do gives us a simple interpretability that we can talk about our models with one another. Whereas if it's just straight AI with no sense of of how things are happening,
Avanish Srinivasan: Right.
Brent Doiron: It's not as exciting to some of us.
Avanish Srinivasan: Right. The models are providing a level of understanding that AI alone can't probably do yet, at least. Right.
Brent Doiron: Not yet. And you know, it'll be a back and forth between I mean AI is of course making huge advances in pure mathematics because those problems are much, much more crystallized, right? Where there's very well established frameworks. So there pure mathematics is going through a bit of a renaissance right now because of that. I don't know if we're there yet in applied mathematics or applied physics or mathematical biology is not quite there yet. Who knows what the future holds? But yeah, definitely.
Avanish Srinivasan: Mm-hmm. Yeah. Mm-hmm. Yeah.
Brent Doiron: Saying I'm you're not gonna use AI in your research program is sort of foolish. I think a lot of us are adapting to to using AI as a tool or as a a blackboard to work off of. asking it to do everything and then being unaware of what happens is probably not the way to go. But saying you're not gonna use AI to study new problems is probably limiting yourself.
Avanish Srinivasan: Yeah, for sure. Mm-hmm. Mm-hmm, yeah. And I I know another thing that you study is neurodynamics, and that is one of the main topics of our episode. But of course this is a topic that most people would not even be aware of. What exactly does it mean when you say neural dynamics? What what is a scientist talking about?
Brent Doiron: Sure. So of course, you know, all of sort of cognition and responses and sensory inputs and movement, everything happens dynamically in time, right? So so the fact that neurons unfold and their their activity unfolds in time is nothing special. What I think what people talk about when they say neural dynamics that makes it interesting is that a lot of the movements that we make or the decisions that we make or memories that we store from short term or sensory inputs that we respond to. Their dimension, the effective number of variables of these is far less than the number of neurons that are actually enacting these computations. So this movement is much lower. There's you know millions, tens of millions of neurons in the motor cortex that are controlling movement. But yet I can describe my movement, you know, in a much, much lower dimensional space. You know, 20 dimensions, you know, ad joints that move in certain directions. So that's true for movement, for sensory inputs. The salient aspects of sensory inputs are can be low dimensional for tasks we want to do. Decision making is in principle, it's at its you know, most primitive sense is just a binary variable. Yes, left, right, no, yes, right? So these are very, very low dimensional things. So when people talk about neurodynamics, they want to say how what how do we go from the high-dimensional space of neural activity?
Avanish Srinivasan: Yeah. Yeah.
Brent Doiron: to the appropriate low dimensional space that governs movement, cognition, sensory re representation. So when people say neurodynamics, it's they're really thinking about what's the low dimensional space in which neural activity sh and its dynamics should be thought of unfolding in. Yep. So a lot of this deals with dimensionality reduction. That's an area of machine learning. which, you know, ultimately, you know, if people are interested in comp neural, one of the best things you can do is learn linear algorithms.
Avanish Srinivasan: Right. Yeah.
Brent Doiron: The linear algebra tells us how to think about spaces and how to think about subspaces in in those spaces. And that's it, those are that's exactly what neural dynamics does. What's the appropriate subspace that the important activity of the nervous system is happening in?
Avanish Srinivasan: Right. Right. Then if it's so multi dimensional, how is it how are scientists able to reduce it and at least into something interpretable in at least in what you were talking about? Yeah.
Brent Doiron: Sure. So do you can imagine, imagine sort of every neuron is an axis, right? So if I have if I have you know a million neurons, then I have a million-dimensional space. So I have a million axes, right? And then at any moment in time, the the activity of those million neurons is a vector in that space or a dot in that space. And then as activity changes, you get a trajectory. And that trajectory in principle lives in this super high-dimensional space, but when you
Avanish Srinivasan: Right.
Brent Doiron: Ask is this neural trajectory because neurons are interacting with one another, one neuron might be competitive with another neuron in the sense that when one neuron's active, it will suppress the activity of another neuron. So given that competition between the two, they're tethered to one another. So they're not independent neurons. When one's up, the other one's down, when one's down, the other one's up. So all of these interactions that are are happening through the the circuit in the nervous system. Is going to put all these constraints. So those trajectories in that million-dimensional space live actually in a much lower-dimensional subspace in there. And identifying that subspace and how that activity goes in that subspace is how we can interpret in a reasonable way how neural activity subtends decisions, which are low-dimensional, movements, which are low dimensional, representation, which is often low dimensional. Identifying that space is the first step at sort of asking. How is this collection of neurons really doing the job as a group?
Avanish Srinivasan: Right. So then these lower dimensions are just essentially interpretations of the relationships between the different neurons and understanding that can help you reduce it in a way.
Brent Doiron: That's I yet but from a mechanistic point of view, I definitely think that both how inputs drive a bunch of neurons as well as their interactions between each other, those are are signatures of that low-dimensional space. or that low dimensional space shows shows signatures of activity, I should say. It's you know, both sides of the same coin. but also how how it's much easier for some downstream center, some area.
Avanish Srinivasan: But
Brent Doiron: That this group of neurons is projecting to, it's much easier for that downstream center to interpret low-dimensional activity. So to build decoders, often we build decoders of the nervous system, but one area of the brain has to decode what's happening in another area of the brain, right? That's that's how the brain works. And having that sort of having a low-dimensional representation in one area makes it so much easier for another brain area to help.
Avanish Srinivasan: Yeah. Yeah.
Brent Doiron: understand what's happening there and to build a neural code for whatever aspect of of neural activity we want to talk about.
Avanish Srinivasan: Right. Can these patterns of neural activity tell us anything about behavior or cognition?
Brent Doiron: Sure, yeah. One you can you can well certainly one classic framework is if we have decision making, right? And an animal is integrating some evidence, it's trying to decide. I mean, often these these tasks are some some noisy signal, and the animal has to sort of detect what the signal is in this corrupted framework. So it takes some time. So they have to integrate evidence. And then at some point in time, they'll make a decision, you know, yes or no, left, right, whatever it is, right? So if you think in a low-dimensional framework, basically what you have is sort of two at what people call attractors. If there's a dynamical landscape there, then there's a point in that space that is the yes decision, and there's a point in that space that's the no decision, and there's a point in that space that separates the activity that will go yes or go no. That can all happen effectively. I've just described something that's a one-dimensional thing. It's a line, right? Maybe the yes is at the minor that yes is at one, no is at minus one, and the decision boundary is at zero, right? So I've collapsed decision making that happens in a very, very high neural-dimensional space onto just a one-dimensional line here. And that's that's the that's all you need to think about binary decisions.
Avanish Srinivasan: Right. Yeah, I think it's really fascinating, especially allowing this allows us to interpret exactly what's going on in the brain in a way that we can genuinely understand despite it being so complex and so multi dimensional, as you're talking about.
Brent Doiron: Exactly. It's how do you extract the salient aspects of neural activity to help you understand the thing you want, which is often cognition, right? so how do we do that? When when when tasks that we get animals to do or even humans to do end up being much lower dimensional than neuroactivity, then that it's really useful to try to understand neural activity in the dimension that's appropriate for the cognition.
Avanish Srinivasan: Yeah.
Brent Doiron: Not living in this high dimensional space that's just a spaghetti of neural activity. I don't know what to do with that. But if I can collapse into a line and I know that points on this line mean things, like yes or no, or maybe, then great. Then then I've learned a lot. And if I can do that rigorously through you know rigorous mathematics, then you know, then I'm confident that this is a reasonable path to take to take.
Avanish Srinivasan: Yeah, especially because if you have a lot of data points going all over the place, you can't really understand anything without it genuinely being reduced in in a way that you can see and visualize. Yeah, for sure. Yeah. And then just to pivot, since we did talk about AI and artificial intelligence in general, how are biological neural networks similar to artificial neural networks?
Brent Doiron: That's right. Absolutely. I'm a simple guy. I need things to be simple. Well, so neural neur I mean, one in part, you know, AI and its infancy really started from thinking about about you know binary neural agents. Actually some of that happened here at U Chicago with with Pitts, McCulloch and Pitts. They were two researchers that were sort of really influenced by Neumann about how to think about sort of the first, you know, model of the nervous system, which is binary agents, right? And now these were the earliest stories of of of neural networks. But so how are they similar? Is that it's also this very high-dimensional space. There's lots of agents in an in an AN artificial neural network. There's lots of agents. they interact with often complex wirings, so often random matrices, or if it's a learned system, as they often do, it's still a pretty complicated interaction matrix. That sort of devoid of symmetry. And symmetry is one thing that physicists use all the time. And sometimes, you know, neural networks, ANNs aren't as symmetric as we'd like, to be able to, you know, take whatever statistical mechanics tool we have in our toolbox, not going to apply there. Right. So that's in one way how ANNs are similar to to actual brains is heterogeneity and complexity of the interactions. One way they're really different. Is is that the nervous system, both individual neurons and synapses that link neurons are very, very unreliable. There's a lot of noise in the nervous system. Synapses, you know, if I if you were you know an engineer and you went to your boss with a model of a biological brain, you just get fired. It just the neurons are just really bad at responding to things. Sometimes they respond, sometimes they don't, sometimes they respond a lot.
Avanish Srinivasan: Not reliable? Yeah.
Brent Doiron: synapses fail all the time, you know, the warranty on on these things is really low. So but yet yet the brain does amazing things. The brain does, you know, you know very it's able to do inference at levels that sort of are beyond what what machines can do often. so how does it work with all of how does we how do biological brains do such amazing feats of cognition? With such you know poor parts is a super interesting question, right? And maybe the variability we have there can be used to do computation, or maybe because computation is done in these low-dimensional spaces, they can organize these spaces to be orthogonal or perpendicular to noise axes. So if you have an if you have a lot of noise that goes in one dimension of the system. The best way you can do computation is often make your computation to be perpendicular to that axis of noise. And then who cares if it's noisy? It's not it's not affecting the low dimensional computation. Yeah.
Avanish Srinivasan: Yeah. Is that one of the biggest questions being trying I mean, at least that scientists are trying to answer right now, or have we come close to at least answering in what way the the brains has so much prowess compared to AI at least in handling these, as you put it, poor quality parts and getting better results?
Brent Doiron: Yeah, that's a great it's a great question. It's it's often task dependent. I mean, certainly A A and N's can do things that better than we can do, right? So there's certain kinds of computations, and that's why we use them, right? But it's it's an open question. I think it's a little ill-posed in the sense that you know it the answer to that question, which is sort of general, is probably gonna be dependent on the task, the animal, the brain area that.
Avanish Srinivasan: Right.
Brent Doiron: We we talk about. There are certain areas of the nervous system where reliability is premium. So they certainly do this. So there's certain certain areas where that the variability I was talking about is something that is minimized. In cortex and hippocampus, the variability is present. Some of that variability may be the fact that on every trial, you know, when an animal decides to do something is different than every trial. So if you can organize and compare. Your trials against the right time axis, maybe some of that variability will go away. People have looked at that, and that's that's in part true. So part of it may be explaining the variability. I was talking about maybe it's not completely due to the fact that synapses and neurons are sort of unreliable. They are, they're unreliable, but maybe some of this is our inability as researchers to sort of control all of the nuisance variables. that are out there that make individual trials different. So as we get better at understanding how to probe the nervous system and to control unwanted aspects of variability, then we can hone in on what the true variability of the nervous system is. That's probably where some people are putting their their efforts now.
Avanish Srinivasan: Right. And is reducing this at least the unknown and these other extra variables still similar to just reducing the dimensions or is it just a completely different process?
Brent Doiron: That that's that's what that's doing is sort of reducing the uncontrolled variance. So for instance, in I'll give you a concrete example. In the vi when you're looking at activity of neurons in the visual system, right? So there you're recording in V1 or V2, these are area the visual system's a pathway. When you're recording, let's just say V V1, you'll see that neurons on a trial-to-trial basis, even if you give the exact same stimulus on a trial-to-trial basis. They will fire different numbers of axe potentials and at different times, right? But you can reduce that variance significantly if you train the animal to focus its eye at the middle of a screen. If on every trial you let the animal six or human to saccade over the image, and that saccade pathway is different on every trial. Even though the image is the same, the sampling of that image is different in every trial. So that that variability you see in the nervous system isn't due to neurons and synapses. It's due to this to the way the animal's sampling that that that image. So if you can say, no, no, don't move your eyes, you know, focus. You can you can do that. You can train animals to just look at a spot. And then there, a lot of that variability I was just talking about goes away, but not all of it. There's a lot of a lot of variability that remains. So and then when you can sort of get rid of all of the sort of behavioral uncontrolled variability that we can then sort of looking at the residual variability that might be neural is is that dissection is something that people are looking at.
Avanish Srinivasan: Right. So there's just a lot of variables, a lot of factors that just go into even understanding this.
Brent Doiron: I mean behavior's super variable and especially when you're doing it with, you know, you know, monkeys or mice. Mice in particular. You c you cannot really train a mouse to stare necessarily. You can train monkeys, but you can't train a mouse to do it. Mice are just gonna look where they're gonna look and that's it. You can't you know, what are you gonna do? And flies are gonna be even worse. Yeah.
Avanish Srinivasan: And then of course in some and many ways actually the human brain happens to be even more efficient than AI because of a lot of what we talked about. So when we're studying the human brain, why is it important to study groups of neurons versus just a singular neuron itself?
Brent Doiron: That's a really good question. I think the early earlier stuff in the 90s, from Stanford professor Bill Newsom and Michael Shadlin and Tony Moffshin and Bill Newsom's graduate student, Ken Britton, they looked at trying to decide what the motion of a visual field was by recording from neurons in an area of the brain that is sensitive to motion here, and they were able to. Sort of decode the motion of activity just by sort of looking at a single neuron. So they were, they were, and and to predict the animal's decision, these animals were doing a task that was you know, what to decide whether motion was going one way or another way. Here, and this again, one of these noisy signals, so it's not it's not so straightforward on how to do it. And they were able to sort of mimic the decision making of the animal by just looking at a single neuron. So Back in those days, a single neuron was enough. Now there, the task was low-dimensional. And probably the reason they could do that is that the single neuron was getting inputs from a bunch of other neurons. So you're just looking at a readout of somebody that's looking at lots of neurons. So it's not that that single neuron was doing all the work, it's that it was working with its buddy neurons, and we could read out that group collective activity from one neuron's behavior here. So when we started recording from more than one neuron, then we could start doing more understanding more complicated tasks that you couldn't really extract from just binary single neurons. So movement in particular, I think trying to understand how movement works, that really sort of the Krishna Shinoy, who was also another Stanford professor, his him and his sort of legacy of postdocs and graduate students that are now faculty at amazing places. They started to start thinking about how collective groups of neurons could control complex movements, which couldn't really be understood from just a single neuron there. So, so really looking at so I think even when you look at a single neuron, you're looking at how the collective group of neurons are acting, and you're just reading it out from one neuron. So it's never the case, other than in super simple situations where a single neuron does all the job.
Avanish Srinivasan: Right. Mm-hmm. So then especially in cases where you're involved with higher dimensional things, you s you need to study a group of neurons to at least try to understand and analyze exactly what's going on. Because a single neuron doesn't have enough information.
Brent Doiron: Yeah, that's right. When you had very, very simple tasks, you could project that group activity through a single neuron's output and understand things. But the more complicated the tasks are, you're not going to be able to look at a single neuron. We're always looking at subgroups of neurons. I mean, you know, even though there's millions and millions of neurons, we can the best labs now can record a thousand neurons at once, maybe. So still like, you know, orders of magnitude, smaller numbers of neurons that are engaged. And we yet we can really reconstruct animal behavior and animal decisions quite faithfully when you have a thousand neurons. So that's super fortunate. I think neuroscience would have been, you know, deadness tracks. If you had to record from from 90% of the neurons in the brain to have a hope of understanding what the brain is doing, then then we'd never we'd never do anything.
Avanish Srinivasan: Mm-hmm. Right.
Brent Doiron: Yeah. So but the fact that you can get a sense of what the nervous system is doing by looking at a small group of neurons, because though that group of neurons is talking to the guys you're not looking at and and and reading them out. And then now we can read out their activity. and that that that activity is a low-dimensional framework, that's the secret sauce that allows us to have a hope of understanding cognition, but only recording less than one percent. Of the neurons available. Yeah.
Avanish Srinivasan: Yeah. Are all activities and behaviors a result of neurons interacting together or is there anything which is just a single neuron performing its task, for example?
Brent Doiron: Well, in simple organisms like escape movements in fish, there's this like mother cell that is a single neuron and it's if it's driven, it'll get a a escape reflex in an animal. That's an example where a single neuron will be important. so in invertebrate systems there might be examples like that, but I think sort of, you know, how doesn't how how doesn't even a neuron know how does a neuron know What's going on if you have a neuron that's controlling movement, even in the simplest organism, really, other than the the mouth's system, if you have a single cell that controls movement, it's getting input from a sensory neuron that's looking at the sensory world. So you really need like at least two neurons there, one to look at the world outside and one to move muscles in response to that world, right? So you really do need in that scenario, to you know, two cells and
Avanish Srinivasan: Right.
Brent Doiron: lots of invertebrates, worms, flies, all that stuff. You know, they have a collection of neurons much smaller than we have here and it's much more stereotyped. But having just a bunch of neurons that don't interact with one another, the only time I can say that that's probably the case, I guess, is in the cochlea, one one area. So if we the way we understand sound. is we have a bunch of cells in our sort of subcortical areas in the inner inner ear, the cochlea, there, where individual neurons there are tuned to specific frequencies of sound. So a an acoustic sound has a rich component of frequencies. And individual neurons in in an array of neurons or a line what people call a line of neurons, a lateral line here, then when individual sounds come in, low frequency sounds will make one neuron fire, high frequency sounds will make another neuron fire. Those neurons don't talk to each other. But in order to understand the complex acoustic sound, somebody has to look at all of them.
Avanish Srinivasan: Of course.
Brent Doiron: Yeah. So though the neurons don't talk, but as a group they are responsible for coding a part of the sensory scene. And then as a collective, somebody looking at them all can get a sense of the whole sensory scene.
Avanish Srinivasan: Mm. Yeah. Especially it I guess most people would not realize that any simple action involves so many neurons and so many different systems. It it's not just of course the processing. You need to process the visual information and then the and you need to actually take action, the muscles. And there's so many so many different parts of that system that people just fail to account for most times.
Brent Doiron: Yeah, and I think a lot of a lot of what we've thought about in the nervous system is that these jobs are compartmentalized. There's parts of the nervous system that's responsible for vision. There's parts responsible for auditory responses, somatosensory responses, movement, decision makings. So we think that there is, you know, this compartmentalization of neural function. That that idea has been challenged a little bit. That's a fun thing in the last sort of 10 years. When you start recording from Mice in particular, you'll see activity in visual areas of mice that really are connected to the motor movements of these animals here. So probably what's happening, or we think what's happening, is motor signals in motor areas not don't just project to muscles to move muscles, but they project to sensory areas to tell that sensory area, hey, we're moving, right? And maybe you should change the way that you're interpreting this visual input. If a mouse, for instance, is stationary versus running, the visual field is very different. When you're running, you know, objects are coming at you all the time, really quickly. When you're stationary, I and you're not afraid of that. When you're running and objects are coming, you just have to move out of the way of a stationary object. But if you're sitting still and objects coming at you, you get us get scared because something's moving towards you, right? So there has to be A change of interpretation or context depending on whether you're running or whether you're you're you're just passively experiencing the world. And there's it's maybe reasonable that motor signals might go down and influence sensory areas, visual areas for that. So, but you tend to see then that that computation isn't just happening isolated in one area, it makes its decision, goes to another area, that area makes its decision.
Avanish Srinivasan: Right.
Brent Doiron: There's this interconnectedness between different brain areas. And looking at a complex task will probably have representation in lots of different brain areas at once. And sort of understanding how we should tweeze those representations apart is sort of a new area that because we can look at more brain regions at once than we could have sort of 20 years ago.
Avanish Srinivasan: Right, yeah. A lot almost all the parts of the brain are working together once and they're all like systems which are interconnected, right?
Brent Doiron: Yep, yep. So it's it's where do we live? We have an interconnected system, but yet components in that system do have different computational roles. You we don't want to think of the brain as just like a soup where all neurons are all over the place and it doesn't matter, they're all the same, right? So that's how this differs a lot from statistical mechanics, right? Where there you often have very similar particles interacting with one another. So in the nervous system.
Avanish Srinivasan: Right.
Brent Doiron: you know, how different neurons in different brain areas wire together, what are their biological features? These are all sort of differences. So we think there's compartmentalization or specialization of activity in certain brain areas, but they have to interact with one another in order to perform interesting computations. So how do you navigate, you know, that that compartmentalization and specialization with the interconnectedness required to do, you know, real interesting computations. That's a that's a challenge.
Avanish Srinivasan: Yeah, for sure. That's really fascinating. And I also want to pivot into another major question in neuroscience, of course, that is disorders of the brain. How can computational neuroscience help us better understand these neurological or psychiatric disorders?
Brent Doiron: Mm-hmm. So that's a great question. And you know, computation there's an area of psychiatry called computational psychiatry. That's that's been around for about 15 years now. There's pushes there. I will say, sort of, I guess a lot of disease states, of a sort of a zeroth order model. Zeroth order model means a model that's like the base-level model that clearly needs to be built upon. But a lot of neurological disorders that we see come from an imbalance. Between the amount of excitation, the amount of excitement that neurons have interacting with one another, from the amount of inhibition. So the ability for one neuron to stop another neuron's activity. And we see that a lot of disease states can be under can be characterized, not understood necessarily, but diagnosed in part with an imbalance between excitation and inhibition. Epilepsy, for instance, is sort of a lack of inhibition. So when you have lots of neurons that just want to excite each other. Then everybody gets excited, all the neurons fire, and that's a seizure. That's what a seizure is. When you have uncontrolled massive activity of a bunch of neurons, then you're in a seizure. So inhibition or the ability for neurons to stop one another is super important. And when that imbalance happens, we see similar things in schizophrenia. We see even some signatures of excitation and inhibition balance in autism. of course, the challenge is epilepsy and schizophrenia and autism aren't even remotely the same at the behavioral level, right? So there is a signature of those diseases that that are the same, excitation inhibition. But how exactly that imbalance manifests itself in sort of the phenotypes, the behavioral phenotypes of these of these diseases is an open question. But computational neuroscience can help us understand what it means.
Avanish Srinivasan: Course.
Brent Doiron: For excitation and inhibition through models of the nervous system. How when you dysregulate that excitation and inhibition, what happens? Right? And then maybe how can you recover that if you have an abnormal circuit? What's the best way to push it into a regime where it acts like a normal circuit?
Avanish Srinivasan: Right. Mm-hmm. Yeah. Yeah, for sure. I I think that's a really interesting way to think about it because a lot of these disorders are related to the inhibition and excitation as you're saying, especially the seizures. So are these disorders also connected to abnormal neural dynamics?
Brent Doiron: Sure. So if you have often what can happen is if you have sort of unregulated interactions amongst excitatory cells that aren't aren't stopped by inhibition, what sometimes will happen there is the dynamics of the system become can become very low dimensional. A seizure is really like a one dimension everybody goes up, everybody goes down. It's this huge thing. So that's a super low dimensional dynamic. So it's possible that in that framework you don't have the necessary space to be able to do interesting computations because it's been hijacked by this ultra low dimensional dynamic that is not in the right direction to do computation. So that's a possibility there. That's a very simplistic model for that. But yeah, if you want to, if you by identifying the right low dimensional spaces that cognition should operate in, and then asking when we have an abnormal circuit.
Avanish Srinivasan: Mm-hmm. Yeah.
Brent Doiron: perhaps because there's you know too much interaction that's putting too much coordination among cells. How how is that destroying our computational model of the of a cognitive brain working properly in this low dimensional space?
Avanish Srinivasan: Right. Yeah. Definitely. I I I totally get that. Okay, so before we wrap up, I just want to do a quick rapid fire round of questions. Just a couple of series of questions. I just want you to give quick, short answers to each of them. Okay. So what do you think is one of the biggest unanswered questions in computational neuroscience right now?
Brent Doiron: Sure. All right. Why the brain has so many cell classes? So many different kinds of neurons. Why?
Avanish Srinivasan: What what technologies are the most exciting to you right now?
Brent Doiron: Right now, for me, it's looking at voltage sensitive dyes and activators. So being able to image individual neurons at the membrane potential level. This is exciting.
Avanish Srinivasan: Right. And if somebody else was trying to enter this field, if a student was trying enter this field, what advice would you give to them?
Brent Doiron: Linear algebra. Linear algebra is the I mean that's that's people I find don't walk into neuroscience knowing linear algebra. They come in knowing a lot about calculus. Calculus is great, but nobody's gonna get you to do some, you know, spherical integral. but linear algebra is the tool that we use to understand dimensions.
Avanish Srinivasan: Right, okay. And what do think is a mindset that would help people succeed in research in general?
Brent Doiron: So in general, sort of you know, be be focused on a on a on a problem, right? And to and be able to sort of for a brief period of time focus in on on something and ignore other aspects. that's how you kind of get deep into something. And when you're done that problem, then step away and ask what's the next problem, and it's okay to pivot in those scenarios. So there's periods of time where you have to be singularly focused, and there's periods of time where you have to take your head out of the sand.
Avanish Srinivasan: Zoom out. Mm-hmm.
Brent Doiron: And look at the world and say, what's the next thing I want to do? And having having that flip-flop between those very different ways of looking at the world's important.
Avanish Srinivasan: Right, okay. So last question. If somebody listening could only remember one thing from this episode, what would do you want that to be?
Brent Doiron: That that understanding how the brain works requires you to record from subsets of the nervous system. And that's enough because the way we think computation is happening lives in low dimensional spaces because computation is about coordination between activity. And it's not just an ideal gas model of the neuron of the brain where individual neurons are just doing whatever they want and they're independent of one another. No, it's the coordinations that make computation in possible and the the signature of that coordination is lower dimensional activity than is possible to have.
Avanish Srinivasan: Right. Yeah, there's definitely lots of synchrony and coordination that people ignore often.
Brent Doiron: And
Avanish Srinivasan: Doctor Doran, thank you so much for joining me.
Brent Doiron: it's and and that's that's that coordination is loady. That's a it's a l the signature of that is lower dimensions than is possible to have. And that makes doing my job a lot easier. I love low dimensional things because I can understand them. High dimensional things is very hard.
Avanish Srinivasan: Yeah, for sure. Thank you so much for joining me today and helping really just make this complex technical field feel so approachable and exciting. I think
Brent Doiron: Great. Well happy to do that. Happy to do that. Thanks so much for having me. It's lot of fun.
Avanish Srinivasan: Yeah, of course. Computational neuroscience isn't just biological, it's really dynamic and interactive. I think we really found that out from today's episode. For listeners, if this episode changes the way that you think about the brain or spark curiosity about neuroscience, AI, or just computational in general, share it with somebody you think would enjoy it. I'm Avniche, this is the Early Perspective, and I'll see you in the next episode.
Brent Doiron: Thanks so much for having me. This is exciting. I I love reaching out to sort of younger audiences. Thanks for giving me the opportunity.
Avanish Srinivasan: Yeah, of course. Okay, so before we get straight into the conversation, I just want to go back to the beginning. Could you tell me a little bit more about your journey into neuroscience and computational research?
Brent Doiron: Sure. I started actually even as an undergrad at in nineteen, I think ninety-eight when w at the University of Ottawa, I got put on a project that the professor I was working with said was neuroscience. I had no idea about the brain, nothing like that. neuroscience at the computational side was sort of in its infancy at that time. So I started there and then I sort of got hooked. And then during and I absolutely a PhD with that same professor at the University of Ottawa, my PhD was in physics, but I I worked looking at neurocomputation from a physics point of view, specifically working on an area of neuroscience called neuroethology. Which is where they look at specialized organisms and look at neural computations that certain organisms have specialized. And the organism I was working on was the weakly electric fish. It's a classic neural prep where there are fish that produce electric fields that surround their body. And then when objects that have a different impedance than the water around the fish come in, then the field lines will morph and then on the surface of the skin of these animals. There's receptors that are able to see that. It's almost like having 3D vision, but it's in a fog because the electric fields decay quickly. the reason I think people studied this in the nervous system is there's sort of lots of control for this. We know how to manipulate electric fields. we can also re the these these animals produce electric fields when they're paralyzed. So you could study them in the sort of 80s and 90s. When people weren't really doing awake behaving, especially in fish, it's very difficult to do that. Now people are doing that. Anyway, that was how I started working on sort of mathematical modeling of that. And then during my postdoc, as you mentioned, then I sort of walked into sort of the broader neuroscience community at NYU and then found myself in mathematics departments through most of my career. but then now here I'm in a modified neuroscience department here and and statistics, but but I I'm I've been slowly over the years. becoming more and more of a neuroscientist. now that's how I think
Avanish Srinivasan: Yeah. Was there any specific moment that really pushed you into the field of neuroscience particularly was or was it just that project?
Brent Doiron: Well, I was I was interested in the nineties, people were really getting hot about complex systems in general. they've changed over the years, but in the eighties and nineties there was a lot of excitement about that. And one of the areas that I wanted to work on was actually social insect dynamics. So I wanted to look work on like how bees work as communities or ants work as communities, but I didn't have an option at University of Ottawa. And then sort of that's how I got convinced to say, Well, why don't you Try looking at neurons. Neurons are, you know, simple-ish agents. They're not actually that simple, but they work as a collective. and so then the mathematics, the kinds of mathematical techniques we use to study groups of neurons are similar to how we study sort of groups of you know different simple agents. And then that pushed me in the field. And then, you know, I got hooked, and I there was a there was a lot of excitement about comp neuro.
Avanish Srinivasan: Yeah.
Brent Doiron: in the sort of early two thousands and there continues to be today. So I felt fortunate to sort of jump on a bandwagon early.
Avanish Srinivasan: Yeah, I think it's very interesting that math and neuroscience really work together and complement each other in such a interesting way. I think a lot of people don't realize that they are so similar and they think that they're separate worlds. But I don't see we briefly touched on this, but what exactly is computational neuroscience? It combines the two, but how so?
Brent Doiron: Yeah, I think there's computational neuroscience is a broad term, and you know, it probably differs depending on who you're talking to. I think sort of there's two branches of this. One aspect of computational neuroscience is building new analysis tools given just this vast amount of data that just every year the amount of the brain that we can measure at one moment in time and the scales that we can measure that brain activity at, you know, change all the time. And we need to have quantitative tools to be able to catalog and analyze that data. And now the fact that people are really doing connectome dynamic connectomes where they're basically going in and measuring every possible connection in right now fly brains and and worms they work on. Those are massive data sets. So we need to have ways to distill or extract the salient aspects of these data sets. Mathematics is great at doing that and sort of machine learning which is sort of a branch of statistics here, really pushes that. That's one branch of computational neuroscience, helping people organize and interpret data. The other branch that I'm more invested in myself, in my research group, is modeling the nervous system. We know a lot about how single neurons act. We know how to model single neurons. actually, there's a Nobel Prize for this in the 60s from Hodgkin and Huxley. so having Electrical circuit models of neurons has a long tradition. And then so if we know how single neurons work and we know how synapses, the thing that connect neurons together, work, then it makes sense to start expanding our models to have groups of neurons that interact with one another. So these are more mathematical models of brain circuits, much more in the tradition of statistical physics, where we sort of have simplified, I don't want to say spherical cow, but you know, oversimplified models. And they're trying to ask what are the critical aspects of the neural activity or the connections between neurons that are needed to explain brain-wide neural activity or even behavior. So those are sort of circuit biological models. That's more the side I'm on. Of course, there's interactions between the two. We need to analyze big data sets to validate our models. So it's really silly to say these are separate areas, but there's a spectrum.
Avanish Srinivasan: Right. Mm-hmm. So then I know that you talked about Hodgkin and Huxley and that model, but then what exactly are these models telling us which are not I mean, at least which experiments can't say alone?
Brent Doiron: Well, so Hodgkin and Huxley, they were, they were actually physiologists. They were sort of they they was really some one of the earliest sort of gold standard models of of of neurons came from two physiologists that were you know very mathematically trained. So the Hotkin-Huxley model talks about how different ionic species, in particular potassium and sodium, how potassium and sodium interact with one another in terms of how they flow across a membrane. in order to produce the action potential, which is sort of the central currency of information in the brain, people call them spikes. So neurons produce these spikes. And Hodgkin and Huxley really figured out how these spikes happen. And it was really about an interaction between two different ionic species. Experimentally, you can isolate one species from another, but if you want to sort of understand how they both go together. That's experimentally very difficult to do. So what they did is they played this game where they would isolate these from one another, build a mathematical framework that when put together should produce an action potential. And then lo and behold, it did produce an action potential. The cool thing about what they they it was a it was sort of a complicated set of equations that now of course if you wanted to numerically solve these on your phone it would be trivial. It would be, you know, you know, milliseconds on your phone. But back then, Hodgkin and Huxley, they were doing these calculations in the late 40s, where if you wanted to do anything with a computer, you know, you had to be working with the military, the Manhattan Project, like there was no computers to really do these things. So Huxley, he was the the the student of the pair, he ended up doing this with a crank calculator.
Avanish Srinivasan: Yeah.
Brent Doiron: So he had like an accountants calculator with you know, and he and he did the numeric. So that took him weeks to do calculations that we would do now quickly. So these guys were super ahead of their time. but to answer your question about what this gains, we sat out we now know how different ionic species interact with one another to produce real activity, and that sort of gave the birth of pharmacology because we could manipulate. the the level of these ionic components with different drugs we can block components we can enhance components and sort of the birthplace of sort of pharmacology at the cellular and synaptic level was sort of at that time and that of course you know has a a history today pharmacology and and mental health have have a big component so I wouldn't necessarily say it all started with Hodgkin Huxley but understanding the the biology of neurons and how chemicals can affect that biology started with
Avanish Srinivasan: Right. So then you're saying that especially these models can provide us more insight that experiments alone can't just allow us to find, right?
Brent Doiron: That's right. So in today's world, it's cheap to do numerical analysis, much cheaper and easier than it is to do experimental work here. So you can use these models if they're well made and they're well constructed and well validated, then you can use them as a hypothesis generator for future experiments in a way that, you know, now you can do many, many, many numerical experiments cheaply and quickly, and then the sort of the the the most performing ones in the sense of manipulating neural activity might be the ones that then you would first try to do experiments to replicate.
Avanish Srinivasan: But now with the advent of better technology calculators and even better AI and machine learning, are these o models going more obsolete because we don't need at least the computational a at least the literal calculations which these models provided or at least let us, I guess, do it faster, right?
Brent Doiron: Mm-hmm. I think interpretability. I certainly think I can when you when we look at, for instance, the Hotchkin Huxley model, I can I can understand that solution in terms of ionic of ionic flow, in terms of sodium, potassium, in terms of of nonlinearities, in terms of of how channels that are holes in the membrane and how they act. This is a super interpretable, well-founded theory. If we're gonna start throwing AI
Avanish Srinivasan: Right.
Brent Doiron: And not asking AI to be interpretable, then I don't sure you might be able to predict things, but having an understanding of what's happening might be out the window. I think there has to be a good balance between these two. So you know, the classic modeling that, you know, physicists and and computational neuroscience at a later point do gives us a simple interpretability that we can talk about our models with one another. Whereas if it's just straight AI with no sense of of how things are happening,
Avanish Srinivasan: Right.
Brent Doiron: It's not as exciting to some of us.
Avanish Srinivasan: Right. The models are providing a level of understanding that AI alone can't probably do yet, at least. Right.
Brent Doiron: Not yet. And you know, it'll be a back and forth between I mean AI is of course making huge advances in pure mathematics because those problems are much, much more crystallized, right? Where there's very well established frameworks. So there pure mathematics is going through a bit of a renaissance right now because of that. I don't know if we're there yet in applied mathematics or applied physics or mathematical biology is not quite there yet. Who knows what the future holds? But yeah, definitely.
Avanish Srinivasan: Mm-hmm. Yeah. Mm-hmm. Yeah.
Brent Doiron: Saying I'm you're not gonna use AI in your research program is sort of foolish. I think a lot of us are adapting to to using AI as a tool or as a a blackboard to work off of. asking it to do everything and then being unaware of what happens is probably not the way to go. But saying you're not gonna use AI to study new problems is probably limiting yourself.
Avanish Srinivasan: Yeah, for sure. Mm-hmm. Mm-hmm, yeah. And I I know another thing that you study is neurodynamics, and that is one of the main topics of our episode. But of course this is a topic that most people would not even be aware of. What exactly does it mean when you say neural dynamics? What what is a scientist talking about?
Brent Doiron: Sure. So of course, you know, all of sort of cognition and responses and sensory inputs and movement, everything happens dynamically in time, right? So so the fact that neurons unfold and their their activity unfolds in time is nothing special. What I think what people talk about when they say neural dynamics that makes it interesting is that a lot of the movements that we make or the decisions that we make or memories that we store from short term or sensory inputs that we respond to. Their dimension, the effective number of variables of these is far less than the number of neurons that are actually enacting these computations. So this movement is much lower. There's you know millions, tens of millions of neurons in the motor cortex that are controlling movement. But yet I can describe my movement, you know, in a much, much lower dimensional space. You know, 20 dimensions, you know, ad joints that move in certain directions. So that's true for movement, for sensory inputs. The salient aspects of sensory inputs are can be low dimensional for tasks we want to do. Decision making is in principle, it's at its you know, most primitive sense is just a binary variable. Yes, left, right, no, yes, right? So these are very, very low dimensional things. So when people talk about neurodynamics, they want to say how what how do we go from the high-dimensional space of neural activity?
Avanish Srinivasan: Yeah. Yeah.
Brent Doiron: to the appropriate low dimensional space that governs movement, cognition, sensory re representation. So when people say neurodynamics, it's they're really thinking about what's the low dimensional space in which neural activity sh and its dynamics should be thought of unfolding in. Yep. So a lot of this deals with dimensionality reduction. That's an area of machine learning. which, you know, ultimately, you know, if people are interested in comp neural, one of the best things you can do is learn linear algorithms.
Avanish Srinivasan: Right. Yeah.
Brent Doiron: The linear algebra tells us how to think about spaces and how to think about subspaces in in those spaces. And that's it, those are that's exactly what neural dynamics does. What's the appropriate subspace that the important activity of the nervous system is happening in?
Avanish Srinivasan: Right. Right. Then if it's so multi dimensional, how is it how are scientists able to reduce it and at least into something interpretable in at least in what you were talking about? Yeah.
Brent Doiron: Sure. So do you can imagine, imagine sort of every neuron is an axis, right? So if I have if I have you know a million neurons, then I have a million-dimensional space. So I have a million axes, right? And then at any moment in time, the the activity of those million neurons is a vector in that space or a dot in that space. And then as activity changes, you get a trajectory. And that trajectory in principle lives in this super high-dimensional space, but when you
Avanish Srinivasan: Right.
Brent Doiron: Ask is this neural trajectory because neurons are interacting with one another, one neuron might be competitive with another neuron in the sense that when one neuron's active, it will suppress the activity of another neuron. So given that competition between the two, they're tethered to one another. So they're not independent neurons. When one's up, the other one's down, when one's down, the other one's up. So all of these interactions that are are happening through the the circuit in the nervous system. Is going to put all these constraints. So those trajectories in that million-dimensional space live actually in a much lower-dimensional subspace in there. And identifying that subspace and how that activity goes in that subspace is how we can interpret in a reasonable way how neural activity subtends decisions, which are low-dimensional, movements, which are low dimensional, representation, which is often low dimensional. Identifying that space is the first step at sort of asking. How is this collection of neurons really doing the job as a group?
Avanish Srinivasan: Right. So then these lower dimensions are just essentially interpretations of the relationships between the different neurons and understanding that can help you reduce it in a way.
Brent Doiron: That's I yet but from a mechanistic point of view, I definitely think that both how inputs drive a bunch of neurons as well as their interactions between each other, those are are signatures of that low-dimensional space. or that low dimensional space shows shows signatures of activity, I should say. It's you know, both sides of the same coin. but also how how it's much easier for some downstream center, some area.
Avanish Srinivasan: But
Brent Doiron: That this group of neurons is projecting to, it's much easier for that downstream center to interpret low-dimensional activity. So to build decoders, often we build decoders of the nervous system, but one area of the brain has to decode what's happening in another area of the brain, right? That's that's how the brain works. And having that sort of having a low-dimensional representation in one area makes it so much easier for another brain area to help.
Avanish Srinivasan: Yeah. Yeah.
Brent Doiron: understand what's happening there and to build a neural code for whatever aspect of of neural activity we want to talk about.
Avanish Srinivasan: Right. Can these patterns of neural activity tell us anything about behavior or cognition?
Brent Doiron: Sure, yeah. One you can you can well certainly one classic framework is if we have decision making, right? And an animal is integrating some evidence, it's trying to decide. I mean, often these these tasks are some some noisy signal, and the animal has to sort of detect what the signal is in this corrupted framework. So it takes some time. So they have to integrate evidence. And then at some point in time, they'll make a decision, you know, yes or no, left, right, whatever it is, right? So if you think in a low-dimensional framework, basically what you have is sort of two at what people call attractors. If there's a dynamical landscape there, then there's a point in that space that is the yes decision, and there's a point in that space that's the no decision, and there's a point in that space that separates the activity that will go yes or go no. That can all happen effectively. I've just described something that's a one-dimensional thing. It's a line, right? Maybe the yes is at the minor that yes is at one, no is at minus one, and the decision boundary is at zero, right? So I've collapsed decision making that happens in a very, very high neural-dimensional space onto just a one-dimensional line here. And that's that's the that's all you need to think about binary decisions.
Avanish Srinivasan: Right. Yeah, I think it's really fascinating, especially allowing this allows us to interpret exactly what's going on in the brain in a way that we can genuinely understand despite it being so complex and so multi dimensional, as you're talking about.
Brent Doiron: Exactly. It's how do you extract the salient aspects of neural activity to help you understand the thing you want, which is often cognition, right? so how do we do that? When when when tasks that we get animals to do or even humans to do end up being much lower dimensional than neuroactivity, then that it's really useful to try to understand neural activity in the dimension that's appropriate for the cognition.
Avanish Srinivasan: Yeah.
Brent Doiron: Not living in this high dimensional space that's just a spaghetti of neural activity. I don't know what to do with that. But if I can collapse into a line and I know that points on this line mean things, like yes or no, or maybe, then great. Then then I've learned a lot. And if I can do that rigorously through you know rigorous mathematics, then you know, then I'm confident that this is a reasonable path to take to take.
Avanish Srinivasan: Yeah, especially because if you have a lot of data points going all over the place, you can't really understand anything without it genuinely being reduced in in a way that you can see and visualize. Yeah, for sure. Yeah. And then just to pivot, since we did talk about AI and artificial intelligence in general, how are biological neural networks similar to artificial neural networks?
Brent Doiron: That's right. Absolutely. I'm a simple guy. I need things to be simple. Well, so neural neur I mean, one in part, you know, AI and its infancy really started from thinking about about you know binary neural agents. Actually some of that happened here at U Chicago with with Pitts, McCulloch and Pitts. They were two researchers that were sort of really influenced by Neumann about how to think about sort of the first, you know, model of the nervous system, which is binary agents, right? And now these were the earliest stories of of of neural networks. But so how are they similar? Is that it's also this very high-dimensional space. There's lots of agents in an in an AN artificial neural network. There's lots of agents. they interact with often complex wirings, so often random matrices, or if it's a learned system, as they often do, it's still a pretty complicated interaction matrix. That sort of devoid of symmetry. And symmetry is one thing that physicists use all the time. And sometimes, you know, neural networks, ANNs aren't as symmetric as we'd like, to be able to, you know, take whatever statistical mechanics tool we have in our toolbox, not going to apply there. Right. So that's in one way how ANNs are similar to to actual brains is heterogeneity and complexity of the interactions. One way they're really different. Is is that the nervous system, both individual neurons and synapses that link neurons are very, very unreliable. There's a lot of noise in the nervous system. Synapses, you know, if I if you were you know an engineer and you went to your boss with a model of a biological brain, you just get fired. It just the neurons are just really bad at responding to things. Sometimes they respond, sometimes they don't, sometimes they respond a lot.
Avanish Srinivasan: Not reliable? Yeah.
Brent Doiron: synapses fail all the time, you know, the warranty on on these things is really low. So but yet yet the brain does amazing things. The brain does, you know, you know very it's able to do inference at levels that sort of are beyond what what machines can do often. so how does it work with all of how does we how do biological brains do such amazing feats of cognition? With such you know poor parts is a super interesting question, right? And maybe the variability we have there can be used to do computation, or maybe because computation is done in these low-dimensional spaces, they can organize these spaces to be orthogonal or perpendicular to noise axes. So if you have an if you have a lot of noise that goes in one dimension of the system. The best way you can do computation is often make your computation to be perpendicular to that axis of noise. And then who cares if it's noisy? It's not it's not affecting the low dimensional computation. Yeah.
Avanish Srinivasan: Yeah. Is that one of the biggest questions being trying I mean, at least that scientists are trying to answer right now, or have we come close to at least answering in what way the the brains has so much prowess compared to AI at least in handling these, as you put it, poor quality parts and getting better results?
Brent Doiron: Yeah, that's a great it's a great question. It's it's often task dependent. I mean, certainly A A and N's can do things that better than we can do, right? So there's certain kinds of computations, and that's why we use them, right? But it's it's an open question. I think it's a little ill-posed in the sense that you know it the answer to that question, which is sort of general, is probably gonna be dependent on the task, the animal, the brain area that.
Avanish Srinivasan: Right.
Brent Doiron: We we talk about. There are certain areas of the nervous system where reliability is premium. So they certainly do this. So there's certain certain areas where that the variability I was talking about is something that is minimized. In cortex and hippocampus, the variability is present. Some of that variability may be the fact that on every trial, you know, when an animal decides to do something is different than every trial. So if you can organize and compare. Your trials against the right time axis, maybe some of that variability will go away. People have looked at that, and that's that's in part true. So part of it may be explaining the variability. I was talking about maybe it's not completely due to the fact that synapses and neurons are sort of unreliable. They are, they're unreliable, but maybe some of this is our inability as researchers to sort of control all of the nuisance variables. that are out there that make individual trials different. So as we get better at understanding how to probe the nervous system and to control unwanted aspects of variability, then we can hone in on what the true variability of the nervous system is. That's probably where some people are putting their their efforts now.
Avanish Srinivasan: Right. And is reducing this at least the unknown and these other extra variables still similar to just reducing the dimensions or is it just a completely different process?
Brent Doiron: That that's that's what that's doing is sort of reducing the uncontrolled variance. So for instance, in I'll give you a concrete example. In the vi when you're looking at activity of neurons in the visual system, right? So there you're recording in V1 or V2, these are area the visual system's a pathway. When you're recording, let's just say V V1, you'll see that neurons on a trial-to-trial basis, even if you give the exact same stimulus on a trial-to-trial basis. They will fire different numbers of axe potentials and at different times, right? But you can reduce that variance significantly if you train the animal to focus its eye at the middle of a screen. If on every trial you let the animal six or human to saccade over the image, and that saccade pathway is different on every trial. Even though the image is the same, the sampling of that image is different in every trial. So that that variability you see in the nervous system isn't due to neurons and synapses. It's due to this to the way the animal's sampling that that that image. So if you can say, no, no, don't move your eyes, you know, focus. You can you can do that. You can train animals to just look at a spot. And then there, a lot of that variability I was just talking about goes away, but not all of it. There's a lot of a lot of variability that remains. So and then when you can sort of get rid of all of the sort of behavioral uncontrolled variability that we can then sort of looking at the residual variability that might be neural is is that dissection is something that people are looking at.
Avanish Srinivasan: Right. So there's just a lot of variables, a lot of factors that just go into even understanding this.
Brent Doiron: I mean behavior's super variable and especially when you're doing it with, you know, you know, monkeys or mice. Mice in particular. You c you cannot really train a mouse to stare necessarily. You can train monkeys, but you can't train a mouse to do it. Mice are just gonna look where they're gonna look and that's it. You can't you know, what are you gonna do? And flies are gonna be even worse. Yeah.
Avanish Srinivasan: And then of course in some and many ways actually the human brain happens to be even more efficient than AI because of a lot of what we talked about. So when we're studying the human brain, why is it important to study groups of neurons versus just a singular neuron itself?
Brent Doiron: That's a really good question. I think the early earlier stuff in the 90s, from Stanford professor Bill Newsom and Michael Shadlin and Tony Moffshin and Bill Newsom's graduate student, Ken Britton, they looked at trying to decide what the motion of a visual field was by recording from neurons in an area of the brain that is sensitive to motion here, and they were able to. Sort of decode the motion of activity just by sort of looking at a single neuron. So they were, they were, and and to predict the animal's decision, these animals were doing a task that was you know, what to decide whether motion was going one way or another way. Here, and this again, one of these noisy signals, so it's not it's not so straightforward on how to do it. And they were able to sort of mimic the decision making of the animal by just looking at a single neuron. So Back in those days, a single neuron was enough. Now there, the task was low-dimensional. And probably the reason they could do that is that the single neuron was getting inputs from a bunch of other neurons. So you're just looking at a readout of somebody that's looking at lots of neurons. So it's not that that single neuron was doing all the work, it's that it was working with its buddy neurons, and we could read out that group collective activity from one neuron's behavior here. So when we started recording from more than one neuron, then we could start doing more understanding more complicated tasks that you couldn't really extract from just binary single neurons. So movement in particular, I think trying to understand how movement works, that really sort of the Krishna Shinoy, who was also another Stanford professor, his him and his sort of legacy of postdocs and graduate students that are now faculty at amazing places. They started to start thinking about how collective groups of neurons could control complex movements, which couldn't really be understood from just a single neuron there. So, so really looking at so I think even when you look at a single neuron, you're looking at how the collective group of neurons are acting, and you're just reading it out from one neuron. So it's never the case, other than in super simple situations where a single neuron does all the job.
Avanish Srinivasan: Right. Mm-hmm. So then especially in cases where you're involved with higher dimensional things, you s you need to study a group of neurons to at least try to understand and analyze exactly what's going on. Because a single neuron doesn't have enough information.
Brent Doiron: Yeah, that's right. When you had very, very simple tasks, you could project that group activity through a single neuron's output and understand things. But the more complicated the tasks are, you're not going to be able to look at a single neuron. We're always looking at subgroups of neurons. I mean, you know, even though there's millions and millions of neurons, we can the best labs now can record a thousand neurons at once, maybe. So still like, you know, orders of magnitude, smaller numbers of neurons that are engaged. And we yet we can really reconstruct animal behavior and animal decisions quite faithfully when you have a thousand neurons. So that's super fortunate. I think neuroscience would have been, you know, deadness tracks. If you had to record from from 90% of the neurons in the brain to have a hope of understanding what the brain is doing, then then we'd never we'd never do anything.
Avanish Srinivasan: Mm-hmm. Right.
Brent Doiron: Yeah. So but the fact that you can get a sense of what the nervous system is doing by looking at a small group of neurons, because though that group of neurons is talking to the guys you're not looking at and and and reading them out. And then now we can read out their activity. and that that that activity is a low-dimensional framework, that's the secret sauce that allows us to have a hope of understanding cognition, but only recording less than one percent. Of the neurons available. Yeah.
Avanish Srinivasan: Yeah. Are all activities and behaviors a result of neurons interacting together or is there anything which is just a single neuron performing its task, for example?
Brent Doiron: Well, in simple organisms like escape movements in fish, there's this like mother cell that is a single neuron and it's if it's driven, it'll get a a escape reflex in an animal. That's an example where a single neuron will be important. so in invertebrate systems there might be examples like that, but I think sort of, you know, how doesn't how how doesn't even a neuron know how does a neuron know What's going on if you have a neuron that's controlling movement, even in the simplest organism, really, other than the the mouth's system, if you have a single cell that controls movement, it's getting input from a sensory neuron that's looking at the sensory world. So you really need like at least two neurons there, one to look at the world outside and one to move muscles in response to that world, right? So you really do need in that scenario, to you know, two cells and
Avanish Srinivasan: Right.
Brent Doiron: lots of invertebrates, worms, flies, all that stuff. You know, they have a collection of neurons much smaller than we have here and it's much more stereotyped. But having just a bunch of neurons that don't interact with one another, the only time I can say that that's probably the case, I guess, is in the cochlea, one one area. So if we the way we understand sound. is we have a bunch of cells in our sort of subcortical areas in the inner inner ear, the cochlea, there, where individual neurons there are tuned to specific frequencies of sound. So a an acoustic sound has a rich component of frequencies. And individual neurons in in an array of neurons or a line what people call a line of neurons, a lateral line here, then when individual sounds come in, low frequency sounds will make one neuron fire, high frequency sounds will make another neuron fire. Those neurons don't talk to each other. But in order to understand the complex acoustic sound, somebody has to look at all of them.
Avanish Srinivasan: Of course.
Brent Doiron: Yeah. So though the neurons don't talk, but as a group they are responsible for coding a part of the sensory scene. And then as a collective, somebody looking at them all can get a sense of the whole sensory scene.
Avanish Srinivasan: Mm. Yeah. Especially it I guess most people would not realize that any simple action involves so many neurons and so many different systems. It it's not just of course the processing. You need to process the visual information and then the and you need to actually take action, the muscles. And there's so many so many different parts of that system that people just fail to account for most times.
Brent Doiron: Yeah, and I think a lot of a lot of what we've thought about in the nervous system is that these jobs are compartmentalized. There's parts of the nervous system that's responsible for vision. There's parts responsible for auditory responses, somatosensory responses, movement, decision makings. So we think that there is, you know, this compartmentalization of neural function. That that idea has been challenged a little bit. That's a fun thing in the last sort of 10 years. When you start recording from Mice in particular, you'll see activity in visual areas of mice that really are connected to the motor movements of these animals here. So probably what's happening, or we think what's happening, is motor signals in motor areas not don't just project to muscles to move muscles, but they project to sensory areas to tell that sensory area, hey, we're moving, right? And maybe you should change the way that you're interpreting this visual input. If a mouse, for instance, is stationary versus running, the visual field is very different. When you're running, you know, objects are coming at you all the time, really quickly. When you're stationary, I and you're not afraid of that. When you're running and objects are coming, you just have to move out of the way of a stationary object. But if you're sitting still and objects coming at you, you get us get scared because something's moving towards you, right? So there has to be A change of interpretation or context depending on whether you're running or whether you're you're you're just passively experiencing the world. And there's it's maybe reasonable that motor signals might go down and influence sensory areas, visual areas for that. So, but you tend to see then that that computation isn't just happening isolated in one area, it makes its decision, goes to another area, that area makes its decision.
Avanish Srinivasan: Right.
Brent Doiron: There's this interconnectedness between different brain areas. And looking at a complex task will probably have representation in lots of different brain areas at once. And sort of understanding how we should tweeze those representations apart is sort of a new area that because we can look at more brain regions at once than we could have sort of 20 years ago.
Avanish Srinivasan: Right, yeah. A lot almost all the parts of the brain are working together once and they're all like systems which are interconnected, right?
Brent Doiron: Yep, yep. So it's it's where do we live? We have an interconnected system, but yet components in that system do have different computational roles. You we don't want to think of the brain as just like a soup where all neurons are all over the place and it doesn't matter, they're all the same, right? So that's how this differs a lot from statistical mechanics, right? Where there you often have very similar particles interacting with one another. So in the nervous system.
Avanish Srinivasan: Right.
Brent Doiron: you know, how different neurons in different brain areas wire together, what are their biological features? These are all sort of differences. So we think there's compartmentalization or specialization of activity in certain brain areas, but they have to interact with one another in order to perform interesting computations. So how do you navigate, you know, that that compartmentalization and specialization with the interconnectedness required to do, you know, real interesting computations. That's a that's a challenge.
Avanish Srinivasan: Yeah, for sure. That's really fascinating. And I also want to pivot into another major question in neuroscience, of course, that is disorders of the brain. How can computational neuroscience help us better understand these neurological or psychiatric disorders?
Brent Doiron: Mm-hmm. So that's a great question. And you know, computation there's an area of psychiatry called computational psychiatry. That's that's been around for about 15 years now. There's pushes there. I will say, sort of, I guess a lot of disease states, of a sort of a zeroth order model. Zeroth order model means a model that's like the base-level model that clearly needs to be built upon. But a lot of neurological disorders that we see come from an imbalance. Between the amount of excitation, the amount of excitement that neurons have interacting with one another, from the amount of inhibition. So the ability for one neuron to stop another neuron's activity. And we see that a lot of disease states can be under can be characterized, not understood necessarily, but diagnosed in part with an imbalance between excitation and inhibition. Epilepsy, for instance, is sort of a lack of inhibition. So when you have lots of neurons that just want to excite each other. Then everybody gets excited, all the neurons fire, and that's a seizure. That's what a seizure is. When you have uncontrolled massive activity of a bunch of neurons, then you're in a seizure. So inhibition or the ability for neurons to stop one another is super important. And when that imbalance happens, we see similar things in schizophrenia. We see even some signatures of excitation and inhibition balance in autism. of course, the challenge is epilepsy and schizophrenia and autism aren't even remotely the same at the behavioral level, right? So there is a signature of those diseases that that are the same, excitation inhibition. But how exactly that imbalance manifests itself in sort of the phenotypes, the behavioral phenotypes of these of these diseases is an open question. But computational neuroscience can help us understand what it means.
Avanish Srinivasan: Course.
Brent Doiron: For excitation and inhibition through models of the nervous system. How when you dysregulate that excitation and inhibition, what happens? Right? And then maybe how can you recover that if you have an abnormal circuit? What's the best way to push it into a regime where it acts like a normal circuit?
Avanish Srinivasan: Right. Mm-hmm. Yeah. Yeah, for sure. I I think that's a really interesting way to think about it because a lot of these disorders are related to the inhibition and excitation as you're saying, especially the seizures. So are these disorders also connected to abnormal neural dynamics?
Brent Doiron: Sure. So if you have often what can happen is if you have sort of unregulated interactions amongst excitatory cells that aren't aren't stopped by inhibition, what sometimes will happen there is the dynamics of the system become can become very low dimensional. A seizure is really like a one dimension everybody goes up, everybody goes down. It's this huge thing. So that's a super low dimensional dynamic. So it's possible that in that framework you don't have the necessary space to be able to do interesting computations because it's been hijacked by this ultra low dimensional dynamic that is not in the right direction to do computation. So that's a possibility there. That's a very simplistic model for that. But yeah, if you want to, if you by identifying the right low dimensional spaces that cognition should operate in, and then asking when we have an abnormal circuit.
Avanish Srinivasan: Mm-hmm. Yeah.
Brent Doiron: perhaps because there's you know too much interaction that's putting too much coordination among cells. How how is that destroying our computational model of the of a cognitive brain working properly in this low dimensional space?
Avanish Srinivasan: Right. Yeah. Definitely. I I I totally get that. Okay, so before we wrap up, I just want to do a quick rapid fire round of questions. Just a couple of series of questions. I just want you to give quick, short answers to each of them. Okay. So what do you think is one of the biggest unanswered questions in computational neuroscience right now?
Brent Doiron: Sure. All right. Why the brain has so many cell classes? So many different kinds of neurons. Why?
Avanish Srinivasan: What what technologies are the most exciting to you right now?
Brent Doiron: Right now, for me, it's looking at voltage sensitive dyes and activators. So being able to image individual neurons at the membrane potential level. This is exciting.
Avanish Srinivasan: Right. And if somebody else was trying to enter this field, if a student was trying enter this field, what advice would you give to them?
Brent Doiron: Linear algebra. Linear algebra is the I mean that's that's people I find don't walk into neuroscience knowing linear algebra. They come in knowing a lot about calculus. Calculus is great, but nobody's gonna get you to do some, you know, spherical integral. but linear algebra is the tool that we use to understand dimensions.
Avanish Srinivasan: Right, okay. And what do think is a mindset that would help people succeed in research in general?
Brent Doiron: So in general, sort of you know, be be focused on a on a on a problem, right? And to and be able to sort of for a brief period of time focus in on on something and ignore other aspects. that's how you kind of get deep into something. And when you're done that problem, then step away and ask what's the next problem, and it's okay to pivot in those scenarios. So there's periods of time where you have to be singularly focused, and there's periods of time where you have to take your head out of the sand.
Avanish Srinivasan: Zoom out. Mm-hmm.
Brent Doiron: And look at the world and say, what's the next thing I want to do? And having having that flip-flop between those very different ways of looking at the world's important.
Avanish Srinivasan: Right, okay. So last question. If somebody listening could only remember one thing from this episode, what would do you want that to be?
Brent Doiron: That that understanding how the brain works requires you to record from subsets of the nervous system. And that's enough because the way we think computation is happening lives in low dimensional spaces because computation is about coordination between activity. And it's not just an ideal gas model of the neuron of the brain where individual neurons are just doing whatever they want and they're independent of one another. No, it's the coordinations that make computation in possible and the the signature of that coordination is lower dimensional activity than is possible to have.
Avanish Srinivasan: Right. Yeah, there's definitely lots of synchrony and coordination that people ignore often.
Brent Doiron: And
Avanish Srinivasan: Doctor Doran, thank you so much for joining me.
Brent Doiron: it's and and that's that's that coordination is loady. That's a it's a l the signature of that is lower dimensions than is possible to have. And that makes doing my job a lot easier. I love low dimensional things because I can understand them. High dimensional things is very hard.
Avanish Srinivasan: Yeah, for sure. Thank you so much for joining me today and helping really just make this complex technical field feel so approachable and exciting. I think
Brent Doiron: Great. Well happy to do that. Happy to do that. Thanks so much for having me. It's lot of fun.
Avanish Srinivasan: Yeah, of course. Computational neuroscience isn't just biological, it's really dynamic and interactive. I think we really found that out from today's episode. For listeners, if this episode changes the way that you think about the brain or spark curiosity about neuroscience, AI, or just computational in general, share it with somebody you think would enjoy it. I'm Avniche, this is the Early Perspective, and I'll see you in the next episode.
