Bing Brunton on the Connectome, Neural Computation, and How the Brain Controls the Body

Bing Brunton

Thanks, Sean. I'm glad to be here.

Bing Brunton

You went right for it. Some of the confusion around connectomes is exactly what it is, because people use that word in different ways, and the terminology does matter here. The rough idea is that the brain is composed of cells, because it's an organ like every other organ in your body. The cells work by electrical activity, and they talk to each other through electricity. Unlike an anonymous net of cells just passing messages forwards and backwards, the cells have specific identities. Some of them have specific jobs, and they also have specific localisations — some cells are found in particular parts of the brain and nervous system and some are not. So there's essentially a wiring diagram of the brain. Think about building a really big, complicated building — a skyscraper. You'd have a wiring diagram, literally an engineering diagram: this is where the transformers are, I'm going to flip this switch and it's going to turn on these lights over here. The connectome, roughly speaking, is that for all the cells and their connections in the brain.

Bing Brunton

Now, the difficulty comes in how you define the units. Do you want a connectome at the scale of individual cells and how they connect to other individual cells? That's one way people have used the term. But there are also what we call meso-scale connectomes. In particular, there are certain animals that are so big — humans, for example, or even smaller rodents — where we technologically don't have the capability of getting the cell-by-cell connectome. We just can't do it. Some people think we should, some think it's impossible, some think even if we could have it, it's useless. When you hear about the human connectome, it's not at the scale of cells and how they connect. It's mostly brain areas and how the brain areas connect to each other. So people use the term to mean an area-by-area connectome as well.

Bing Brunton

There's a lot of coarse-graining, and people don't agree on how they use that term. It's like the whole -omics thing in biology. Every word that ends in -omics — genome, proteome, transcriptome — is supposed to mean a comprehensive map thereof. People usually agree that if I tell you, 'Hey Sean, I got a genome of a new spider that I found,' you'd expect that genome to be at the resolution of the base pairs, the A, C, G and Ts. You have that expectation. If I gave you something else, you'd say, 'That's not a genome, I don't know what this is.' We don't have that in connectomes. We don't quite agree on the scale of description — whether you need every single neuron in that spider for it to be the connectome of a spider.

Bing Brunton

There are other cells, and they're clearly important. The rough estimate, in my understanding, is that half the cells in your brain are not neurons. Our word for not-neurons is just glia, which doesn't mean anything except that it's the word for it. People used to think they're just there like custodial staff, but that's trivialising. They do a lot more than that. They're involved in all kinds of vital functions and they have their own dynamics. Understanding all the other cells in your brain, what they do, and what they do in concert with the neurons, is a really exciting emergent field in neuroscience.

Bing Brunton

Oh, for sure. There's definitely stuff in the extracellular space. All I meant was that all of life as we know it is made out of cells. We can quibble about viruses later, but living organisms are composed of cells. One of the lessons we're going to keep bumping into is that biology is messy. Things are squishy and interconnected and complex. A macroscopic organism is pretty much a matter of teamwork between different kinds of cells, but also between cells and non-cell substances. Take your bones, your skeleton. A lot of its material properties come from the calcium matrix and other stuff. But it's also an intricate, meshy structure that has blood vessels all inside it, because it needs to be vascularised — otherwise it's going to die. It needs sugar, it needs oxygen to stay alive. Even something you think of as structural is not like a stainless steel beam in a building. It's alive, in a way that only cells can keep it alive.

Bing Brunton

We think about that quite a bit. It's really a great fundamental question in biology: how organisms are able to recover from injury and repair themselves — or sometimes not.

Bing Brunton

We're going to try. It's going to be fun.

Bing Brunton

It is, yeah. The fine-grain connectome — we can call it the cellular-level one — would be a big old matrix listing every single neuron and how it connects to every other neuron. From a computational perspective, because I am a computationalist, by the time it gets to me it's that gigantic connectivity matrix. It has structure, it's sparse, it's not at all random. And it's not symmetric either. Neurons talk to one another, but they don't necessarily listen.

Bing Brunton

There's a lot of extra information in it. The giant connectivity matrix is definitely part of it, but it's nowhere near all the information we get out of this technology. The identity of the cells matters. You've probably heard of dopamine, serotonin — there are dopamine cells, serotonin cells, and if they both fire an action potential, those messages are completely different. The other thing that really matters is how the messages are received. It's very context-dependent, like language. If you say exactly the same thing to two different people, depending on your relationship with them, they can hear very different messages.

Bing Brunton

That is entirely correct. The identities of the cells matter, and there are lots of other detailed biophysical properties of each cell that clearly matter, though we don't know by how much. What I try to tell my graduate students is this: say I'm a civil engineer trying to build a building, and I need materials to hold up the roof. The beam is made out of atoms, and I know that down there somewhere there's quantum mechanics — but we're not solving Schrödinger's equation in order to design a roof. That's way too much. It's super interesting, but you don't need it for the task of building a roof. That's sort of where we are. I know I don't need every single detail about the biophysical parameters of these cells. They get really funky — crazy non-linear, super special, almost impossible to measure. People will spend an entire PhD measuring one cell. But do we need it for these very holistic models of the entire animal nervous system? Probably not. Where do we stop? It's hard to say. I know I don't need every single detail, but I don't know which of them are actually crucial.

Bing Brunton

I think that's fair. If you think of it computationally in terms of maps: if you can define exactly what a neuron's inputs are and what its outputs are, you can infer some function that maps the inputs to the outputs. And one of the computational clues, in order to run these simulations, is that you don't need every single detail of how that map is implemented to approximate its function.

Bing Brunton

I'll admit I was skeptical — the audience can't see, but I'm raising my hand. This whole thing started when I was in grad school and I first heard about really large efforts to produce connectome data sets, 15 or 20 years ago. I was skeptical on a couple of fronts. Skeptical that it was even going to work at all — could we actually reconstruct one of these things at sufficient scale, which involves running a transmission electron microscope for months straight, making zero mistakes? And then further skeptical that even if somebody handed it to you magically tomorrow, what would you do with it? How could you make sense of this giant spaghetti monster? It's only pretty recently that some of the work my lab has been doing with collaborators has started to convince me that we might actually be able to do this.

Bing Brunton

The reason lots of people were skeptical — there were essays written maybe 10 or 15 years ago by people in the field, including Eve Marder and Cori Bargmann — is that they knew there were so many other details not observable by the connectome. Information about all the channels, the biophysical properties of some of these cells — we can't get them from the connectome. Nobody ever thought we could. So the disagreement was whether the stuff you can measure, effectively these connectivity matrices, is sufficient to teach us something, versus the other logical extreme: that it's utterly useless because you actually need all the other stuff. There's a giant continuum of opinions. My current opinion is swaying a little closer to the view that we can actually do something useful with this data set.

Bing Brunton

The first one we got was about 30 years ago — a full connectivity matrix of the C. elegans nematode worm. It's not an earthworm, the kind you see attempting to cross the sidewalk. They're much smaller nematode worms, about a millimetre long, and they live in the soil. If you scooped up any soil in your garden and looked under a microscope, you'd very likely see them. This particular species has been studied a lot in molecular biology because they breed really quickly, so we have tons of tools. They have about a thousand cells and about 300 neurons. The connectivity matrix of those 300-ish neurons was mapped out decades ago. And one of the first things people in connectomics always bring up is that we've had the connectome of C. elegans for so long, and yet we still don't understand it. There are good technical reasons why the worm is actually really difficult to understand from a connectomics perspective.

Bing Brunton

The one that came out much more recently, in the last year or two, is a couple of efforts by giant collaborative teams — I was not involved, I was cheering them on from the sidelines — to map the full connectivity matrix of a Drosophila fruit fly. This is the kind of fruit fly that infests my kitchen at the end of every summer, buzzing around any rotten fruit or a pile of compost. These little guys are more like three millimetres long, the size of a grain of rice, and the entirety of their nervous system is more like the size of a sesame seed. They're small enough that it's been possible to reconstruct the entirety of their brain and nervous system.

Bing Brunton

We have a brain in our heads and also a spinal cord — that constitutes our central nervous system. Insects have an analogous structure: a central brain inside the head, going down the neck just like ours, and then, instead of a spinal cord, invertebrates have a ventral nerve cord. It's remarkably similar in structure and organisation to our spinal cord, but instead of being on their back, it's on their belly side — that's why it's called ventral. That whole thing has been mapped out. There are two data sets, one male and one female fruit fly, published only in the last half a year or so.

Bing Brunton

The brain has 150,000, and the ventral nerve cord has an additional 22,000. A much bigger matrix than our little C. elegans. And the important thing about the size, paradoxically, is that it's actually a little bit easier to understand from the connectivity matrix. The reason C. elegans has been so hard to understand is that it took us a while to figure out, as a community, that they do a lot of computation not using that kind of connectivity matrix. There's a ton of chemical communication — they're constantly squirting out neurotransmitters and other chemicals at each other. There's a lot of mechanical computation too: it's a squishy thing that crawls around in a soil matrix, with a lot of mechanical stretching and reflex loops that are mechanically coupled with its body. So the way they function as an animal takes advantage of lots of other computational properties — chemical and mechanical, in addition to neural. The fact that we had the neural connectivity matrix just wasn't quite good enough to understand what they do.

Bing Brunton

By contrast, our hope is that the fruit fly's connectivity matrix is more directly helpful, because it's a little bigger, it has jointed limbs just like humans do, and it has enough cells that there are actual cell types — not every cell is its own little snowflake. In C. elegans the neurons are so unspecialised that single cells have multiple sensory modalities feeding into them. Where we have a visual system with cells that detect photons, and an olfactory system with cells that detect smells, the worm has single cells doing several jobs, because it's so tiny and compressed. They've had to multiplex.

Bing Brunton

Right. Do you know the longest cell in your body? It's about as tall as you are. You have cells — the same kind we were talking about in the ventral nerve cord, but in your spinal cord — that are responsible for how you know you stubbed your toe. One end is at your big toe, and the other end goes all the way up to your brain stem.

Bing Brunton

You can do that — there are other cells involved and you can hand off the message — but this is the normal architecture. The advantage of having one cell do it is that you can do it really fast, and if you stub your toe, your brain really wants to know about it very quickly. This is how you don't fall over. Your body does that without your thinking about it. If you're hiking and you kick a rock, you don't fall over, and you don't want to waste your precious time thinking about how not to fall over. You want to keep having that conversation about number theory with your buddy.

Bing Brunton

That's right. The slightly longer story is that this is a long-time collaboration with a friend of mine, John Tuthill. John is a fly experimentalist — his lab does neurophysiology, studying the ventral nerve cord, the sensors coming in and the motor control going out. We've been collaborating for about a decade and co-advised a series of graduate students and postdocs doing theory and modelling. John's also been really involved in these connectomics efforts, so a lot of what I know that's not wrong is because I learned it from John. The stuff that's wrong, I made up — I take responsibility.

Bing Brunton

A couple of years ago, John and I were taking a walk, and we had a brand new PhD student thinking about joining our labs. We were wondering what to have them do. John said, we almost have a ventral nerve cord connectome — it's almost ready, they were cleaning it up and curating it — what if we just simulated it? And I said, that's never going to work, let me tell you all the ways this is not going to work. So I told him all the ways — the biophysics, all the parameters we don't know, tons of stuff. But by the end of the walk we'd come to, well, let's try it anyway. We don't lose anything. Let's give it a good old grad school try. I do think half the secret to succeeding in graduate school is listening to your advisor tell you it won't work and distinguishing when they're right from when they're wrong. Our student, Sarah, listened to us, said okay, and went off and wrote some code.

Bing Brunton

Long story short, it took a couple of years, but we kept at it, partly because the preliminary stuff was interesting. What ended up happening is that we went after a question biologists and neuroscientists have been asking for over a hundred years: how does the nervous system generate rhythms from non-rhythms? To give you context on why this matters — all animal movements are rhythmic. Not just animals; even bacteria move by spinning their flagellum. Walking, running, swimming, slithering, crawling — basically all locomotion is rhythmic. So your nervous system needs some way of generating the instructions for your muscles to move in a circle. That's fundamental. Ever since the 1910s, the first experiments demonstrated that these rhythms are not generated by reflex only — your central nervous system, somewhere in your brain and spinal cord, is capable of generating these cycles. But we didn't know exactly where, which cells did it, or how.

Bing Brunton

We understand it in general, and if we have time I'd love to come back to it, because I love dynamical systems, and it connects to our work on the connectome. In the intervening hundred years, lots of people have studied these circuits. The ability of your nervous system to generate rhythms isn't only important for locomotion — it's also important for breathing. You inhale, exhale, inhale, exhale. You can control it, but if you don't think about it, it just happens. That's generated by what we call a central pattern generator, a CPG circuit. Digestion is cyclic too — you have to churn the stuff in your digestive system, a sequence of muscle contractions that gets the food down. The most studied CPG circuit is actually in the crab digestive system. There are a couple of adorable little neurons responsible for churning what's in the crab stomach. And we know which neurons are in charge — this is the work of Eve Marder, who has studied it for decades. That system is so extraordinarily well understood that, only slightly starkly, people say the crab digestive circuit is the only neural circuit we actually understand in all of neurobiology. It's a bit of an exaggeration, but it's not untrue either.

Bing Brunton

That's probably the plainest definition. Roboticists love the CPG — a lot of modern robotics is built on these oscillator equations. I talk to roboticists who have no idea about the neurobiology of central pattern generators, because for them it doesn't matter. Just write an equation, it goes around in a circle. It doesn't really matter how it's implemented by cells; you just care that there exists a thing that goes in a circle. But we didn't know what the actual cells and their connections were, in an actual nervous system, that generated these rhythms for any animal that walks. Nobody had ever found what the cells are, what their names are, how they work.

Bing Brunton

We had an opportunity. It's not that we were smarter than everyone else who'd worked on it — we just had the complete connectivity map of the ventral nerve cord of a fruit fly. We figured, whatever it is, it's got to be in there somewhere. Instead of building it up from individual components I can do experiments on, we took the reductionist approach: it's in here somewhere. We got it down to a network of about 4,000 cells.

Bing Brunton

Precisely. First we made it more manageable by focusing on only the two front legs. Insects have six legs; we got rid of the other four. Then we got rid of all the parts of the nervous system that don't control the two front legs. That's how we got to about 4,000. Then we simulated that, and we demonstrated that those 4,000 neurons could generate a cycle — motor rhythms that actuate the muscles that move the leg. And this is no reinforcement learning, no machine learning, no deep learning at all. We're just doing brute-force numerical simulations of this giant connectivity matrix. You write a lot of code and run it a million times, and we can get these rhythms to come out. Then we asked: now that it's in here somewhere, let's try to reduce it. Let's cut away one cell at a time. Do I need this one? No. Do I need this one? No. You keep going until you've thrown away as many cells as you possibly can without losing the rhythm, and what's left over is the minimal circuit.

Bing Brunton

Not so. This is something my friend and colleague Michael Dickinson is fond of saying: insect wings are actually novel limbs. For every other flying animal — bird wings are modified arms, bat wings are modified arms — the wings used to be not-wings. Not so of insect wings. They're not modified legs. There are theories about how they evolved, but they're actually novel structures. It's not that they had eight pairs of legs and two became wings. These are genuinely new things. And this is reflected in the nervous system. Just as there are parts of your spine that go to the left leg, the right leg, the trunk, they have parts of their ventral nerve cord that go to each of the six legs — little balls that stick out, a bit bigger because they have more cells. And they have the same thing for the wings: little clumps of cells corresponding to the wings.

Bing Brunton

It worked great. In the pruning study I described — where we took a functioning system that generated CPG-like rhythms and started cutting away everything that didn't seem necessary — I remember sitting in this office with Sarah and John the day we decided to give it a try. We started with 4,000 cells. I told Sarah, just give it a try; if you get it down to a few dozen cells, if that's the minimum circuit, I'd be ecstatic. That would be a really cool result. She went off and did it. The answer was three.

Bing Brunton

Three cells. That's the minimum you need. And they have names — we know who they are in the fly nervous system.

Bing Brunton

That's a great question, because I actually have no idea what their real names are. Their names are known, and their lineages are known, so we sort of know where they came from — but the names are a series of letters and numbers and I can't remember them. John knows their names. We gave them pet names, though, of course. I told you earlier the cells have identities, and it matters what type they are. Two of the cells are excitatory — they make other cells more excited — and one is inhibitory. So they're called E1 and E2, and the last one is I1. And they're connected in a very understandable architecture motif that explains why this tiny circuit can generate cycles.

Bing Brunton

To be more precise: we believe the three neurons are sufficient to generate the rhythm. They're not sufficient to actually control the walking — there are dozens of individual muscles that need to be coordinated in the legs. There are many more muscles than there are degrees of freedom in a limb, so actually controlling them to do something coordinated and not clumsy is more complicated. But our hypothesis is that these three neurons generate the basic rhythm, and then there are other cells involved to make it actually walk. That's the lesson we keep learning: there's a lot of teamwork in biology, a lot of responsibility shared among different subcommittees. I don't feel the nervous system is wasting cells. I don't think biology is wasteful. There's redundancy, and there's a good reason for the nervous system to be redundant in case it gets injured, but I don't think we have cells for no reason.

Bing Brunton

Because neurons are some of the most expensive cells to maintain in your body, my hypothesis would be that if a cell is genuinely not necessary, the body would find a way for it not to be there over a longer time frame. It's actually more plausible to have vestigial organs in the body than vestigial neurons. Those vestigial organs usually didn't go away for a reason — they got stuck, basically, in the way evolution works.

Bing Brunton

To learn about all the other stuff, our vision — with tons of collaborators, because this is a giant team effort — is to actually embody the nervous system. To take the connectome in all its glory and put it inside a body where it belonged all along. Not a mechanical body but a simulated one, more like a video game body. My son's been playing Red Dead Redemption, riding a little horse around a virtual environment — clop, clop, clop. But that's just the animation; it doesn't matter whether it's biomechanically realistic. We want to do that, but actually have it be biologically interpretable and physically realistic. It would be a physics engine — modelling F equals ma.

Bing Brunton

It's in progress, and I'm really excited about it. This is a bit superlative, but I've rarely in my career felt so much conviction that something is the right thing to do. It's so obvious to me that the brain does not live in a jar. It always controlled a body — a specific body, with these limbs and muscles and joints and sensors — in order to move around the world and eat and collect information and do all the things animals do. So it's obvious to me that we need to understand the brain and nervous system in the context of the body it interacts with to produce the behaviours the animal actually does. That's the grand overall vision. It's early days, but I'm really excited about it.

Bing Brunton

For sure — it's super easy to hijack the nervous system of a fly. Part of the reason we work in flies is that it was the genetic organism of choice for a long time, so our ability to hijack every aspect of its nervous system through gene engineering — putting proteins in it, shining lasers at it — already exists. That's the reason we work in flies: the wealth of knowledge accumulated over decades. We know so much more about their everything than a spider, for example. We're neuroscientists, we love lasers, so there are a lot of lasers going on. We shine lasers at them and make them do things — when they're walking, flying, trying to sniff.

Bing Brunton

Oh, it is. We love lasers. I think we might love lasers slightly more than physicists do, because we just play with them.

Bing Brunton

It is a guess at the moment. We do need to do the validation experiments — to corroborate our predictions by doing experiments on these actual neurons. For technical reasons that's ongoing; we haven't done it yet. So this is still a very strong hypothesis in my mind. We have good reason to make this guess, but it's still a guess until we can confirm it biologically. One of the things that's cool about this result is that, as a computational modelling person, I've spent the majority of my career fitting data — somebody has an observation, and I write equations and code to recapitulate it. This is one of the few instances where the model actually came before the experiments. We were agnostic going in. We had this giant data set, we simulated it, and we made a prediction of things we didn't know before.

Bing Brunton

Part of this result I haven't mentioned: we haven't quite got to the three cells we predicted to be the core CPG circuit, but we made predictions about other parts of the nervous system. There's one pathway that comes down from the neck — a cell that, in our model, has a name, and I do actually know this one. It has four letters and numbers. Our model said, if you zap it with a laser, it should make the leg tap, back and forth. Nobody had ever studied this neuron before. But somebody had made a cell line — a fly we could order that had the correct proteins in it — so we ordered it, grew it, cut its head off, glued it to a stick, and shone a laser at it. And it tapped its leg. It was an actual model-driven prediction. Nobody had any idea what this neuron did. Most cells in the nervous system are like that: we know it's a neuron, we have a nomenclature, we sort of know where it came from, but we don't know what it does.

Bing Brunton

I sure hope so. Part of it is this idea of the embodied brain. Drawing analogies between our virtual models — the nervous system plus the biomechanics of the body — we and some other people have been calling them digital twins, a word we're borrowing from industrial engineering. The digital twins in industry are of things like airplanes and cities. There's a digital twin of the city of Singapore. It's a simulation — it doesn't have every single lightbulb, but it has many of the important parts, including the city's morphology and connectivity, and it's hooked up to real-time sensors so planners can predict disaster response or shift traffic patterns to relieve congestion. In close analogy, the thing we're thinking of building would be a digital twin of a behaving animal: a simulation of the nervous system, the interfaces between the nervous system and the body, situated in a virtual reality environment capable of interacting with things — surfaces that aren't flat, other agents that can even touch each other.

Bing Brunton

If we develop these simulations in close collaboration with our experimental collaborators, we should be able to predict what happens in parts of these circuits that are hard to predict otherwise. The whole thing has mad feedback and recurrence. One thing I've learned about humans and our ability to reason through rational thought is that we're really terrible at reasoning through feedback circuits and recurrence. We can follow a path A to B to C to D. But as soon as D goes back to B, and C goes back to A, our intuition for what's going to happen is really poor. That's one of the arguments for why we need these complicated computational models — we can't do it, but computers can.

Bing Brunton

Yes, so many — probably more than not. We need to be really careful. This is the idea behind the digital Sphinx paper we wrote a couple of weeks ago. I was starting to see a lot of work and conversation in the field — including by my own lab and our collaborators — where, because the whole thing is so overwhelming and we know we can't measure everything, we make a lot of assumptions. One thing we can measure with a lot of fidelity and relatively easily is just the behaviour output of the animal. We can get cameras, track what it's doing, see how it moves its legs, where it points its head — we have really good computer vision. So a lot of people are basically saying, this is the grounding: if we get a model that behaves like the animal, matching what the animal was observed to do with a camera, then surely we've got something right. Lots of people are doing this carefully, but what I was afraid of was that some people were starting to do it in a not-careful way.

Bing Brunton

In particular, there was some stuff coming out on social media by startup companies trying to fundraise, and our friends looked at it and said, you're overselling this — you're not doing what you said you did. What they said they did was upload a fly brain. That was the headline: they've uploaded a brain. It went viral and got a lot of attention, not just among non-scientists — I had chemist friends telling me, I heard they uploaded a fly brain, that sounds really cool. So, as an exercise, I was sitting around with one of my postdocs, Elliott Abe, and I said, this is bananas, this is not the right way of doing it. But explaining why took a lot of technical words — you have to understand reinforcement learning, the architecture of the nervous system. So we asked: what's the logical extreme of what they're effectively doing? They're not even really using the fly brain connectome. This could be anything — a random matrix. It might as well be a C. elegans worm matrix.

Bing Brunton

They simulated a portion of the fly brain. And crucially — since we spent so much time talking about the ventral nerve cord and how it controls leg movements — they did not simulate the ventral nerve cord. Even though their animation definitely had little legs moving around; that was the animation part. So Elliott and I thought, what if we upload a worm brain and train it to control the fly body? We can totally do this. He and another graduate student did it. They downloaded the C. elegans worm connectome — all 300 cells in its glory — and popped it into a reinforcement learning algorithm we've been using for other things, to control a biomechanically realistic fruit fly body walking around in the physics engine, imitating 3D kinematics of flies. And it works. It runs around just like we know flies do.

Bing Brunton

If you use enough deep learning and train it with good enough data, it is perfectly possible to get a worm brain to control a fly body. What we're learning from all this is that it's silliness. If you use that much deep learning, allowing all these parameters to change in ways that aren't obvious, the fact that you have a connectome and a hyper-realistic biomechanical body doesn't mean anything. It's not biologically meaningful. You can get behaviour fidelity without any biological fidelity.

Bing Brunton

We even tried a little. The worm connectome also has motor neurons — the ones that would control muscles — and we wired that population up to the fly body actuators, just for fun. But that's where the deep learning comes in: an artificial neural network does the mapping between the motor neurons and how it produces torque in the body, and we trained that. It's kind of just showing I can run the same software on a Mac or a PC.

Bing Brunton

It's an emulator, exactly. If you don't care about meaningful interfaces, you can get lots of things to plug together. HDMI matters — the real trick to getting something meaningful out of it is to actually engineer those interfaces.

Bing Brunton

I'll state this a little more strongly than I actually believe, for the sake of conversation. We have no examples we all agree on of agents that are intelligent and conscious except the ones that are embodied. The other ones may or may not be intelligent and conscious, but we don't agree. The only ones we actually agree on are embodied agents. Furthermore, all nervous systems evolved, starting about 500 million years ago, to control a body, to sense from the environment and respond in order to move around the world and seek food and mates. Everything we think of as reasoning, as consciousness — all of that machinery, all of those capabilities — evolved on top of the neural computations required for sensorimotor control, for sensing from the environment and moving the body around. So my guess is that understanding the platform, the substrate on which all those other capabilities were built, would be important for understanding the stuff above, as well as a strong constraint on how it could possibly have got there.

Bing Brunton

I am a scientist.

Bing Brunton

That I actually don't feel that strongly about. I don't think it's important to understand the details of how biology implements something to build an artificial system that takes advantage of some of those insights. The field of bio-inspired engineering is full of examples where just the concept that biology might do something was sufficient to inspire a perfectly good solution, without understanding the details. Bird flight is the first one everyone thinks of. We knew birds can fly, so we were inspired to fly. It turned out that imitating the way birds fly was an utter failure — we threw that out the window and started over with fixed-wing aircraft. The details of how a bird actually flies are fascinating and we continue to study them, but understanding them was not necessary for us to fly. Same with the central pattern generators — the roboticists just took the observation that there must be something inside your spinal cord generating these rhythms. They didn't need to know exactly how it worked. So I'm not sure we need to know the details of biological intelligence to get to artificial intelligence.

Bing Brunton

However — the concept that it may be necessary for that intelligent system to be actually embedded in something interactive, something with a physics. It doesn't have to be our physics, but maybe it should have some rules and constraints. Some notion of energy, some notion of conservation laws — not just limitless everything. Maybe that's important. This part I'm speculating on. But I don't think it's necessary to understand how biology works to get artificial intelligence.

Bing Brunton

I think that's absolutely true. I think the biology of consciousness is difficult. There's the psychology and the philosophy of consciousness, but getting at the biological basis of consciousness is pretty difficult, as far as I understand.

Bing Brunton

To be fair, David Chalmers always said the easy problem is hard.

Bing Brunton

In biology, there are no easy problems. Every time you come up with a dichotomy — surely it must be A versus B — 30 years later it's both. It's both, or C.

Bing Brunton

I sure hope so. One of the most interesting applications I think about a lot, in building our embodied models — the brain connected to a body — is exactly that interface between the nervous system and the musculoskeletal system, of which there are tons of pathologies and dysfunctions that are pretty terrible when they happen to you. Think about spinal injury, which affects both your nervous system and your neuromuscular control. Or if I have a bum ankle on one side, I start limping — that's a different neural strategy, and over time I might adopt a different gait. A bit longer, and the legs on one side of my body get stronger. So it's adapting at different time scales. Understanding those interactions — how your nervous system controls movement in compensation to injury, and how it attempts to repair it, or if you have an amputation you can't repair the amputation, but you can repair the function — all of those interactions are very important and poorly understood, because they're holistic. We just haven't had the tools.

Bing Brunton

That's one of the things we hope to do in building these embodied models: understand not just whether it walks, but what happens after it walks. What happens when we break it in different ways? How does it compensate? What is the role of plasticity — of growing new muscles versus new tendons versus new neural pathways? If we understand that, perhaps we'll have new clues about how to design therapeutics to help that process work better. Your physiotherapist does a lot of crazy stuff, but not all physiotherapists agree on how to rehabilitate after an injury. Is there guidance we might come up with by understanding the interactions between the nervous system and the musculoskeletal system a little better?

Bing Brunton

Biology is complicated. That's why it's the science of the 21st century, I guess. It's squishy and it's fun. I feel so privileged to be able to do it — to spend my time hanging out with my friends talking about brains and biomechanics. Sometimes I wake up and I can't believe this is a real job.

Bing Brunton

Thank you, Sean. This was super fun.