BI 185 Eric Yttri: Orchestrating Behavior

[00:00:03] Eric: A big open question will be one of the goals of this naturalistic behavior field moving forward is how continuous is behavior, both behavior itself as well as its neural representation?

You the general idea is that motor cortex is saying what to do and stridem is saying how to do it.

And all this is making it sound like I have thought out my future dreams of being.

[00:00:45] Paul: I had no idea we’re going to talk about this.

[00:00:48] Eric: Me either.

[00:00:52] Paul: This is brain inspired. I’m Paul, and today my guest is actually my current employer. As some of you know, I recently got back into the research world, and in particular, I work in Eric Itry’s lab at Carnegie Mellon University. Eric’s lab studies the relationship between various kinds of behaviors and the neural activity in a few brain areas known to be involved in enacting and shaping those behaviors. Namely, those brain areas are the motor cortex and basal ganglia. And to study that, he uses tools like optogenetics and neuronal recordings and stimulations in mice while they perform certain tasks, or in my case, while they freely behave, wandering around an enclosed space. So in this episode, we talk about how Eric got to where he is now, how and why the motor cortex and basal ganglia are still mysteries, despite lots of theories and experimental work. We talk about Eric’s work on trying to solve those mysteries using both trained tasks and more naturalistic behavior. We discuss a little bit the challenge of even answering the question, what is a behavior? Which you would think would be a simple question, but it’s not. And we discuss lots more. You can find relevant links in the show notes at Braininspired co podcast 185, where you can also learn how to support this podcast on Patreon. If you’re feeling generous, you may notice if you’re watching, I am not in my usual studio location, and that’s because I’m at a conference, and now I must go mingle at the conference opening reception. So wish me mingling luck. I wish you luck in your studies and your endeavors, and I’m glad you’re here to mentally mingle with Eric.

Like me, you have a non human primate, neurophysiology, cognition background, very similar, actually, you know, reaching. I didn’t never do reaching, but I was all sakads, eye movements, and you’ve published some in eye movements.

And unlike me, you’ve stayed in academia, essentially, and I went away, and now I’m. You’re back into the fold, groveling back, yeah, and you took a chance on having me in your lab, but I’ve never asked you how your own personal take on your trajectory from those non human primate, that non human primate world to. And you did some rat work, and now mouse work. Was this like a conscious decision or is that just kind of how it flowed or what?

[00:03:30] Eric: Well, so the rat work was in my undergrad, it was okay.

[00:03:37] Paul: Yeah.

[00:03:41] Eric: So I had a great education working with primates, and Larry Snyder was the world’s best advisor.

But I thought a lot of the questions were getting sort of rehashed over and over again. And particularly for the right questions, doing monkey work is absolutely vital. And in some ways, for some questions, you can only do it with non human primates.

And then there was the entire toolbox, including optogenetics, that was opening up for mice.

And so for my postdoc, I interviewed at a couple of mouse labs, a couple of monkey labs.

[00:04:30] Paul: Oh, you did? So you had that door open still.

[00:04:33] Eric: I had that door open and was just looking for good questions. And so what I ended up doing, when I moved to Genelia research campus, they had were in the process of doing some work, training mice how to do a reaching task with a little joystick, a PlayStation three controller with a coat hanger shoved in the end. And I thought, oh, great, I get to continue doing the reaching and what I think the non human primate stuff, both the education and the sort of approach, looks really in depth at behavior at a level that I think for rodents is coming online. But for years it was just lever presses.

And for reaching your sakads, you look at movement time, you look at trajectory and the variance of those trajectories and hypometria and all these things, these important kinematics that now are becoming much more popular in the field. But when I started my postdoc twelve years ago, were fairly uncommon for rodents. But, yeah, initially I was trying to train mice to do a center out reaching task with four different directions.

It did not work, but it’s good to be ambitious sometimes. So the idea was kind of keep the reaching thing going and ask some more detailed questions.

Thanks to the trade offs, the benefits of the model.

[00:06:24] Paul: Yeah.

One of the things that I want to ask from that is, so non human primates, the reason that we study them is because there’s as close as we can get to our brains and our cognitive abilities while still being ethically okay.

In this day and age, these tasks are being adapted for things like rodents, mice, but they have different brains and they. And they have different cognitive abilities. It’s not like a direct translation. And I know that there’s a lot of excitement in studying. So on the one hand, there’s a lot of people saying, well, mice, they’re a lot smarter than we first thought or than we gave them credit for, and we can study these cognitive functions in mice. Right? And on the other hand, there’s pushback saying, yeah, but you’re not necessarily studying the same cognitive functions. They have different brains. The brains aren’t perfect homologues. From an area, let’s say motor cortex, from a mouse to motor cortex of a non human primate, where do you stand in that? And have you sort of changed your mind at all through your studies?

[00:07:40] Eric: It’s interesting you went that way with the question, because my first response was, you said, oh, we study monkeys because they’re really close to humans.

I’ve seen plenty of great human work on the area that I was studying in vital cortex, and they have two area lips and five different divisions of prital reach region. And monkey brains are as similar to human brains, roughly as mice are, in that there’s some general principles. Prior cortex is one thing, a precipitous cortex is going to be doing another, but there’s a whole lot of differences. And so if you hadn’t gone the direction you went, I would have actually said, we studied, or at least I studied monkey behavior and monkey neurophysiology because of what a monkey can do, both cognitively and in terms of its behaviors. If you want to study in depth attention, or certainly sakads or skilled reaching, you either need a task that is made accessible to a rodent, or for that matter, or a monkey, or you just can’t do it. And so for any question, you pick the appropriate model system. And so I’m not trying. When I was studying monkeys, I wasn’t trying to understand the human brain.

There are some ideas. For instance, parietal hemi neglect.

I was injecting muskamol and inactivating parietal cortex. And some of the things we were looking at were related to that, the prital hemine neglect in humans. But I was interested in monkeys. And now that I’m studying rodents, I hope that I’m uncovering some fundamental principles that apply to at least mammalian brains, if not vertebrate brains, very broadly, glossing over the fine points.

But when I study a mouse, I’m studying a mouse. And not just that, I’m keenly aware as we’ve moved into more and more naturalistic behaviors. I’m studying a super inbred mouse that typically is raised in a prison population.

And each of even within the inbred lines, there’s plenty of nuances there. So I try and keep. I guess the word is not humble, but I know about the mouse that I’m studying and the neurons that I’m studying, and I try and keep it limited to that. If others can take those ideas and run with them, or we look at other people’s data and applies more broadly. Fantastic.

[00:10:56] Paul: But is that why you got into neuroscience? So I got into neuroscience because you’ve been talking about brains and comparative across species. Right. But then there’s the cognitive functions, comparative aspect across species. And I got into neuroscience because I wanted to understand consciousness, et cetera, et cetera. Highfalutin stuff. Right. And understand higher cognition. And then one way to see my trajectory is just getting, like, lower, lower level questions.

One of the things that you said, and we’ll talk about naturalistic behavior, is there’s still questions like, we don’t really even know what motor cortex does, not that it has, like, one function. Right. But you get in, you think, well, I can study brains to study, in your case, like attention, intention in parietal areas, for example. And these are cognitive things that you can’t necessarily study in a mouse yet. Maybe they have great. Maybe they’re a great model for attention. But you’re talking about brains. Do you think the same way about cognition, or have your interests changed? And do you think that mice are a good animal model for that?

[00:12:08] Eric: So, I guess to come at it from my origins, even as a high schooler, I was lucky enough to know that I wanted the job I have today.

[00:12:22] Paul: What the hell?

[00:12:22] Eric: Way back when. Yeah, I missed out on the whole switching your major three times.

So, yeah, maybe even as early as late middle school, I loved music, philosophy, and psychology and biology, and so combined philosophy, psychology, and biology into a field that used to be called psychobiology, now neuroscience.

And so it’s always been about the brain and particularly behavior. For me, if we can help treat diseases in humans, fantastic.

Definitely a plus.

And we do go into some models of OCD and Parkinson’s in order to benefit mankind, and not just with knowledge, the little bits that we do. But, yeah, I’ve always been at it for the brain, and with that, I don’t even know.

Cognition carries a lot of weight to it.

And I make the comment that I’m super jealous of people studying sensory systems, particularly primary sensory courses or on out towards the periphery, because they have such great control. And the motor system, basal ganglia, motor cortex, we’re closer to the output the people who study all the wonderful mess in between, my hats off to them.

[00:14:14] Paul: You envy v. One people?

[00:14:16] Eric: Yes, I envy v. One people, incredibly, for all the control that they get access to, but for the stuff between sensory motor, the cognition side, that’s hard to do, and some people do it fantastically, but there’s a real lack of roadmap there.

Looking in from the inside, I kind of feel foolish studying cortical stridal circuits in a mammal when we can’t even explain all the ins and outs of a nematode nervous system or a fly.

[00:15:08] Paul: So why don’t you do that? Why don’t you go to nematodes?

[00:15:13] Eric: I only recently got associate. Give me a little bit of time if I had to jump. And actually, the first experiment in the lab was I got some slime molds and was just going to do some computational modeling on their performance. But, yeah, if I jump ship from mice, I would go, slime mold or octopus, just because that’s so different.

[00:15:41] Paul: Those are two completely different paths.

[00:15:45] Eric: They are both organisms that are capable of. I mean, maybe this is coming back at cognition. They have a surprisingly complex and adept cognitive ability and comparatively far fewer moving parts.

Slime molds are just a single cell, and octopuses are.

It’s somewhat of a distributed circuit, but not nearly the neuron count of, say, even a mouse brain.

But, yeah, if I had to do that, I think I’d get wild and go with one or both of those.

[00:16:35] Paul: Is that something that you think you could split your. Like, could you expand your lab and have a half slime mold, half mouse, or third and third and third? Like, could you do that? Would you?

[00:16:48] Eric: Certainly not third too much.

I mean, in the first six months, while we’re waiting to get things going, I did have the thought of growing slime mold in petri dishes in my desk.

[00:17:06] Paul: You had them, right?

[00:17:07] Eric: I had them.

[00:17:08] Paul: So what happened? You haven’t told me this.

[00:17:12] Eric: I take care of mice and my daughter much better than I happen to take care of slime molds.

[00:17:21] Paul: Enough said. Got you. Yeah, maybe your future is an octopi, then.

[00:17:27] Eric: We’Ll see.

But in terms of some of these questions, we have two unbanded and probabilistic decision making tasks going on in the lab.

It might be fun to see if the models that we and our collaborators build, if we can get the same sort of things out of slime molds. And there’s some good evidence that, to a large extent, you can.

I don’t see that being 50 50, but if an enterprising, probably more master’s or undergraduate student said, hey, I want to play around with some slime molds and ask some cool questions.

I would probably giggle with delight.

[00:18:20] Paul: Well, see, I think you’re. So you’re saying setting myself up, yeah, undergrad or master’s. Well, I think you’re undershooting the dilemmas that will arise, because everything will be harder than you think it is, and it’ll snowball and turn into a bigger, bigger project, I would imagine.

[00:18:39] Eric: And I think that may be part of it. Master’s students are in the lab no more than two years, and the typical undergrad, at least once they get going, is probably two to three years max. So if it’s a terrible idea, there’s a much shorter clock.

[00:18:57] Paul: Okay, fail fast. Right. And break things.

[00:19:01] Eric: Yeah, you don’t know your try unless you break things every now and then.

[00:19:06] Paul: All right, well, so we haven’t talked any real science. Oh, you know what? Let me just interject, because since we’re on the topic of animal models, and maybe we’ll just kind of get this out of the way, one of the things that I’ve been interested in the past few years is this recent push to use artificial neural networks to study brain dynamics and parts and functions and stuff. And it struck me in this conversation, this would be a relevant time to ask you your thoughts on, instead of using animals, using networks. Right. And what we can glean from artificial networks. And I’m going to guess that you’re not terribly interested in that approach by itself anyway.

[00:19:55] Eric: I tend to think of computational models, AI, deep learning methods as best used as part of a circle, a cycle. We ask questions of the biological brain, we find things out, and then either as a screen or to do.

[00:20:23] Paul: A.

[00:20:24] Eric: Screen, because it would be hard to do all these theoretical experiments in the biological brain, or just difficult generally because of some feature of the question itself, try things out in an in silica model, whatever the case may be, and then use those results to make predictions to come back and ask a more intelligent and pointed question. And several groups are taking this approach. Again, there’s lots of important questions out there.

As someone who has always been interested in the brain, I’m going to be focused on the brain, and I know there are better solutions now to this issue, but I know a common complaint of artificial neural networks is you’re placing one complex system that you don’t understand with another complex system you don’t understand.

That is not wholly true anymore. And I think there’s some really cool things we can do with that. But I would still think that we would be, at least in terms of the questions that I’m most excited about were best served if we asked the hard questions, technically hard questions of the insilica brain and go from that neural network and then use those results to ask better questions of the biological neural network, if we can do that.

[00:22:21] Paul: Also, not to harp on the octopus and the slime mold, I think, in terms of mammalian brains. Right. And modeling mammalian brains, especially with deep learning models.

And I guess what I’m trying to ask is, what question are you interested in?

And maybe they’re orthogonal. Right, because you said you’re interested in the brain, but then which brain? Right. Because an octopus brain is very different than a mammalian brain, and then there’s a slime mold and stuff, and it wouldn’t necessarily make sense to model.

I guess you could.

How important would it be to set up? Okay, so let me back up. One of the successes of deep learning models is that the architecture doesn’t matter as much. Right. You just give it a task and then it learns, and then, lo and behold, it has some dynamics that are similar to dynamics in whatever brain area you’re studying and whatever cognitive task you’re studying.

[00:23:22] Eric: And hopefully, universality holds up and everything works out great. Yeah.

[00:23:27] Paul: But then if you probed an octopus brain in the same task, I would imagine you would get a different answer. I mean, I don’t know. I guess you’d have to do it.

[00:23:37] Eric: Maybe, and sort of circling back to asking the right question in the right way.

My goal should probably be define more to brain and behavior like things, doing things. That’s why I’ve said on the.

Despite all the benefits, away from the sensory side.

So how does an octopus figure out which arm to reach out for a tasty crab? And how does it get the arm out there? How does it grasp it? How does it know where it is in space and figure out which of the arms to use? How does it decide how to open up a jar, even though it’s never seen a jar before?

And so much of the allure, there would be the particular differences if I just wanted to understand, how do I get an effector from point a to point b?

Unless I was a soft body roboticist, there was no way I would approach an octopus. And all this is making it sound like I have thought out my future dreams of being.

[00:25:04] Paul: I had no idea. We’re going to talk about this.

[00:25:07] Eric: Me either.

Yeah. Making it seem like I have thought of this out much better.

Than just a few random conversations with.

[00:25:22] Paul: Sorry, I’m sorry, we’re going down.

[00:25:23] Eric: No, this is kind of cool. I hope somebody reaches out. And I actually chaired a session at SFN this past year, and there was a really cool octopus talk there.

So, yeah, maybe it’s in the cards.

We’ll see.

[00:25:43] Paul: All right, let’s get back to what you spend most of your time thinking about. What you mentioned that something was not the question that you’re most interested in or the questions that you’re most interested in. Do you have a question that you’re most interested in? I mean, that’s an unfair question itself. I know.

[00:26:04] Eric: I think the overall extremely high level question is trying to understand how circuits produce the appropriate behavior. How do we transform thought into action?

And so this encompasses decision making and sometimes value estimation, motor control.

So for the very broad question, I will give the very broad answer, but I think it’s still appropriate in terms of what gets me excited and that all these things, one way or another, fall underneath it.

[00:26:47] Paul: So, circuits, we study. You study motor cortex and basal ganglia these days mostly.

And of course, there’s a motor cortex, basal ganglia, loops, there’s corticoscerebellum, loops and stuff. So these things are connected in loops. So it is a circuit. But of course, there’s a lot of divergence and convergence within these circuits.

Of course, cerebellum itself is still a mystery.

[00:27:17] Eric: And it’s half the neurons in your whole brain.

[00:27:20] Paul: It’s more than half.

[00:27:20] Eric: Ridiculous.

[00:27:21] Paul: 70%, isn’t it?

[00:27:24] Eric: Yeah. Depends on the species.

[00:27:25] Paul: What about the octopus, Eric?

[00:27:27] Eric: Yeah, pretty sure it doesn’t have a cerebellum, but show what I know.

[00:27:34] Paul: Yeah, that’s right.

I always think of that whenever there’s an estimate for the number of total neurons. Like, we have so many neurons, I think. Yeah. And they’re all in the cerebellum, which is an interesting basket.

[00:27:45] Eric: Cells, right?

No, it’s the tiny ones that send up the parallel fiber.

[00:27:52] Paul: Yes. Stellates.

[00:27:53] Eric: No, someone’s going to.

[00:27:55] Paul: Scientists. We need to know this.

[00:27:56] Eric: We got to edit this out because someone yell at us.

[00:28:00] Paul: Lots of neurons.

[00:28:01] Eric: Lots of neurons in the cerebellum.

[00:28:03] Paul: Well, that just goes to show you how neglected the cerebellum is, too, right?

[00:28:08] Eric: Oh, yeah.

[00:28:09] Paul: Although plenty of people study it, and a lot of people are saying these days, I’m hearing, especially from philosophers, for some reason, they think that the cerebellum is coming back into popularity. I don’t.

[00:28:20] Eric: It’s. David Maher is thought to be a philosopher. He was a neuroscientist. I mean, philosophers or whoever will claim him. He’s a computational scientist.

Yeah. I mean, thinking of Mars, three levels of analysis and all that, that comes out of largely cerebellum, at least as far as I understand.

[00:28:44] Paul: Well, he had one of the early models for the cerebellum as well, which was wrapped up in his Mars levels analyses. But I mentioned the cerebellum because it’s still a mystery. But so know if you open a neuroscience textbook, it’ll tell you motor cortex commands, motor behaviors, and basal ganglia, depending on which textbook you open, either is a gate which allows things to happen, or it inhibits certain actions and on and on. But we really don’t know. I mean, we know so much. And both of these brain areas, motor cortex and basal ganglia, are still mysteries. I know that you have your own, would you call them, hypotheses about frameworks, hypotheses about what they do? Why don’t we start there? Why don’t you share your ideas about what motor cortex and basal ganglia are doing, and then if you can just kind of pit it against a couple other ideas that have been studied.

[00:29:49] Eric: Sure.

So, I think the foundation of most, if not all, models of basal ganglia function and cortical stridal dynamics, including the cortical basic anglithalamic loop, are largely based on the work of Roger Albin, Delong and Crooker and others in the mid to late 80s who realized that certain diseases that has affected certain populations within the basal ganga either led to a hyperkinetic state moving more or a hypokinetic state moving less. And chief among these would be Parkinson’s disease.

And so that bedrock serves, or that serves as the bedrock for most of the existing models right now, where myself, and this is not anything that I invented, I’d say the truest to that model, and one that’s been championed by John Crackhouer, David Robbie, Rob Turner, who doesn’t get nearly enough credit for the brilliant work he does, is that the basic angular serves as a modulator of vigor.

You can think of vigor as force over time, is one way to think of it. There are other interpretations thereof. But modulating the kinematics of an action that is selected upstream, presumably with motor cortex and its friends, and that’s made sense to us. In probably the biggest paper we’ve had so far, I stimulated either the direct or indirect pathway of the stritum, and these are two populations of opponent neurons in the stritum of mammals. As well as most vertebrates.

And the idea there that what we found is that we could change the speed of a mouse reaching using one of these joysticks.

That if we simulate, on fast reaches, this direct pathway population, we would get more fast reaches. We wouldn’t necessarily speed up that reach, but we’d bias the animal in its performance.

And if we stimulate slow reaches, we got more slow reaches. The opposite was true for this indirect pathway population.

The chief alternate hypothesis, which is most often taught in textbooks, or sometimes just put right next to it as if it’s the same thing, is that when it comes to the motor side of things, again, that the direct and indirect pathways are pro movement and anti movement. And this is where the gating idea that you introduced comes about. That cortex and its friends would have several different possible actions that could be performed at any given time.

One researcher theorized it to be around 4000, and that through these parallel action channels, the stridem selects, the direct pathway selects which one behavior to do, and the indirect pathway selects the 3999 to suppress.

[00:33:57] Paul: Where did this number come from? Do we know?

[00:34:01] Eric: It was based on a computational model of just what’s possible.

So, not necessarily that the brain necessarily is doing it, but you suppress a whole lot of things that you don’t want, and you facilitate and promote the action that you do.

And some of the leading evidence, let me back up.

This was very attractive from a computational standpoint, and much of the support for this comes in the form of computational models, where the connectivity insider architecture lends itself well to an action selection approach.

But if the evidence for it, which there’s plenty of it, probably the best, is Lex Kravitz, who’s at Washu, work that he did while he was at UCSF.

He stimulated the director indirect pathway just as I did, but left the laser on for 1020, 30 seconds.

And in this case, those neurons that were optogenically stimulated, if it was direct pathway stimulation, the mouse would start walking after a few seconds and start and walk.

And if you stimulate the indirect pathway, the sort of stop pathway, presumably because you were gating off all action, the animal would stop locomoting and wouldn’t start again.

This is fine, except not that any optogenetic stimulation is particularly physiological, but these neurons are only active very infrequently. Most of them have baseline firing rates, at least in a mouse of under 1. They fire very specifically around movement onset, at least in trained behaviors, as well as locomotion.

That’s just sort of odd that you need to stimulate for these longer periods of time. Also the fact that it was limited to locomotion. If you stimulate one side, it’s as if you’re applying either the gas or the brake to the contralateral side of the body, but not the contralateral side of the body. It’s the turn. But in order to turn left or right, you need both your left and your right limbs and your head and all the postural muscles of the torso.

So that just sort of struck me as OD. And so that was a lot of the impetus for the very brief and very selective stimulation that we did. So if we stimulated on every reach or on a subset of reaches, we observed no effect with these under half second pulse light to sort of just try and nudge the system, add a few more spikes when the neuron would normally be active, we found no effect on that. And since then, we’ve managed to expand this to starting and stopping. So in work that we should be submitting soon, we looked and saw if we stimulate the direct pathway every time a mouse stops locomoting in the open field, the suggestion of the gating pro and anti kinetic action selection side would be that you should, just as lex demonstrated quite well, if you stimulate that mouse, should stimulate direct pathway, that mouse should take off walking, start moving again. Yeah, what we found was Germans, we, the opposite mouse will stay pretty much seated for minutes at a time. If we stimulate every time, the animal stops moving. So looks a lot more like reinforcement. And along a kinetic axis of, in this case, locomotor speed, an indirect pathway has the opposite effect. The animal just more or less won’t stop walking for the 20 minutes in which stimulation can occur. I think what’s really critical and that I think is an informative point is causality is the wrong term. But while we stimulate sometimes, even though the direct pathway stimulation biased the animal to be stopped walking, to not be walking, sometimes we get up and walk in the middle of it.

So we’re biasing the performance, we’re not selecting the action, but you’re stimulating, you’re.

[00:39:54] Paul: Trying to nudge, you’re not doing electroshock therapy.

[00:39:59] Eric: We’re not, at least as far as we can tell. So that might be part of it. But I think the interpretation that I think may be more appropriate with our causal manipulations and in what we see with our chlorive measures like recording, is that these populations of stridental neurons acts more as a game modulator.

So whatever signal, and this also has great benefits with the other motor or non motor roles ascribed to the area if you’re just a game modulator. And actually, interesting, you brought up cerebellum. Cerebellum and its wonderful repeated architecture serves as a computational unit that, in theory, can be applied to any input that it receives, and it’s going to do its ltd learning wonderfulness on whatever the input is.

[00:41:08] Paul: If there’s any part of the brain that you could almost consider in isolation, the cerebellum would be it. Right. Is it granule cells? Is that what.

[00:41:16] Eric: Yes, granule cells. That’s the little one.

[00:41:19] Paul: Yeah. Whereas the brain is like so intertwined and there’s so much different cross talk and recursiveness that it’s just messier. But the cerebellum is clean in that respect. Ish.

[00:41:33] Eric: Yeah. And so our idea with this other subcortical, not that it is any cleaner in the connectivity sense, is that if it serves as a game modulator and if you know some basics of optimal control theory, if you have a push and a pull on any feature that you might want to control, it gives you a lot better control to have that push and pull than just the push. Some people say, oh, but I can drive my car with just the gas. If you needed to drive at 34.3, would want to have a foot on both the gas and the brake to control that.

And so if these neurons are acting as a gain modulator of whatever the incoming signal is, that applied in a certain fashion. Getting back to how you ask your questions, that can look like action selection because of turning that volume knob, as it were, doing that game modulation depending on the inputs, and that’s viewed through a very specific lens when looking at vigor, it’s more obvious there that it would be doing that. And then plenty of people look at the basal ganglia and the stritum in terms of its cognitive or auditory processing or all sorts of things.

To get back at sort of these, what is the brain doing? And are there any sort of fundamental properties of neural circuits?

It seems like you could copy and paste this computational unit of, we’ve called it a history dependent gain modulator. So it’s getting feedback information from dopaminergic cells and all sorts of convergent inputs from all over the brain. It would be a really good system, and our experimental evidence suggests it’s a great way to pair descending inputs, feedback from your dopamine system, and to alter, to tweak those descending inputs, and either have it be a feedback system or in many organisms, a majority of the output of the basic angulus feed forward down to brainstem nuclei, superior caliculus, et cetera.

For mammals, those same nuclei are getting convergent inputs from these cortical areas as well.

Yeah, there’s a beautiful paper that Josh Dubman wrote with John Cracker. We also had one with the word history dependent gain in the title, sort of pushing this idea.

But, yeah, that’s where we stand.

I still have a job, so it’s certainly not figured out one way or another and we’re going after these things. But, yeah, the general idea is that motor cortex, not necessarily m one specifically, but motor cortex is saying what to do and Strytom is saying how to do it.

And I think about my monkey work quite a bit, because even in a visual sensory motor area like lip or pride or reach region, I knew 150 milliseconds or 300 milliseconds ahead of time if that monkey was going to Sakad left or right, use its left or right arm to reach.

And so just the timelines of the action selection model have never quite made sense to me.

[00:45:58] Paul: You can have a gain modulator that acts as an action selector in the extremes, right.

This goes back to what does brain area x do? Well, it’s not like one function, because under those two alternative hypotheses, it can actually be both.

[00:46:17] Eric: Absolutely.

There’s some reasons why.

And I said view through a specific light, it can look like action selection.

I think there’s some other more nuanced work that suggests that. That actual action selection might not be.

[00:46:43] Paul: What’S going on, but actual action selection as opposed to gain modulation.

[00:46:50] Eric: Action selection, yeah, action, something like that.

It could certainly look like it. But if it’s performing a selection between many possible actions.

[00:47:11] Paul: Right, the way it’s written up in the textbooks, that’s the actual action selection you’re talking about.

[00:47:17] Eric: That’s the better way to get at that. I think maybe we go back and edit that one too.

But, yeah, it can look like action selection, but I think for selecting between dozens of possible actions, the evidence just isn’t there. And we have a few lines of evidence that suggest that that.

But that’s also not the case.

But it’s still an open question.

In fact, writing a perspective piece right now with several dozen authors on just sort of trying to come to terms with these competing hypotheses and some of the original, I talked about the bedrock theory on hypo and hyperkinetic modalities.

Some of the people who came up with those models have said, guys, you’ve been using the same model since the mid 80s.

Maybe you should get creative and move on to something else. But if you read a basal ganglia or a corticosteroidal paper, you’re almost required to cite albin and Delong. Brilliant guys.

But, yeah, there is some doctrine in there to deal with part of what.

[00:49:07] Paul: You’Ve been talking about, especially with basal ganglia, and this holds for motor cortex as well. But your experiments on basal ganglia also showed this.

And basal ganglia is also wrapped up with the story of learning. Right. Very involved in learning.

And part of your stimulation experiments showed that, that gain modulation is learned and lasts for some time. Yes. Either whichever way you stimulate for the faster or for the slower, it lasts for some time.

[00:49:40] Eric: It followed the policy.

[00:49:41] Paul: Right.

[00:49:42] Eric: Which is policy. Learning is a beautiful form of learning.

[00:49:47] Paul: Yeah. So I’m not sure if right now is the right time to pivot to. So this is all in the context of training animals to perform tasks, to move these joysticks. Right. And then you do some causal manipulations, some stimulation, then you’re recording neurons, and the organisms have to learn the task. They have to put effort into the task, et cetera.

And so that’s a large part of your history and current research is studying these tasks and how learning is affected and the kinematics and what’s happening at the neural level.

And you’ve taken kind of a recent turn, as has a lot of the field, into, quote, unquote naturalistic. Neuroechological. Is that the term?

[00:50:37] Eric: Ethological?

[00:50:39] Paul: Oh, I say etho. Is it etho?

[00:50:42] Eric: You said echo.

[00:50:44] Paul: Did I say ecological? Okay.

[00:50:48] Eric: We got invited to study elephants in the nature reserve, but we’ve not taken them off on that. Yeah, that one.

[00:50:59] Paul: Do you say neuroethological or neuroethological?

[00:51:06] Eric: It’s neuro to me, which is the last person you should trust, especially as your employer.

I say neuroethology, but it’s neuroethological. Okay.

[00:51:22] Paul: Anyway, the turn toward naturalistic behaviors. Neuroethology.

[00:51:29] Eric: Yeah.

[00:51:30] Paul: Well, I wanted to say neuroethological behaviors again, but neuroethology encompasses, it’s redundant to say behaviors again. Neuroethology behaviors in there. Right.

In the case of what you’ve done thus far, in your case, it’s letting a mouse just kind of explore a space, giving it. There’s no task. Right. There’s no reward to earn. There’s no skill to learn. So these are kind of like innate kinds of behaviors and just recording neurons. Right. In motor cortex and in basal ganglia and seeing how these areas and the access of what you call the access of m one basal ganglia motor basal ganglia access.

Seeing what’s happening in these brain areas during these naturalistic behaviors, I don’t think I’ve ever asked you what you expected to find in the first place.

[00:52:31] Eric: So I don’t know that I had a real expectation of what I wanted to find.

The motivation behind it really is what we were just talking about of if this is, if the basic angle do action selection, or if cortex says, what action to do.

The fact that our behavioral space for actions is to press a lever or not, or to move a joystick to turn left and right, which is sort of baked into the problem of if you simulate for 20 seconds, and the only behavioral output in that paradigm is it walks more.

It takes a little bit, but it walks more.

[00:53:32] Paul: That’s something. Not to denigrate that. That’s something.

[00:53:34] Eric: It’s huge.

And at the time, that was twelve years ago.

That was monumental. Still is.

But part of the motivation to get into more ethological behaviors and more, I’d say our focus is more on the spontaneous side. The untrained part of it is because now the action space that you are selecting from is one much bigger. It’s not turn left or turn right. It also is unencumbered by the potential caveats of training or overtraining or fact that most mice don’t see a joystick as they’re walking through the woods.

That was a big motivation to try and figure out what these circuits are doing is to let nature make a lot more of the decisions. So that opens up our behavioral parameter space, which may open up our neural dimensionality.

And we have some initial evidence that that is the case.

Not as big as I thought, but that sort of remains to be nailed down.

[00:55:10] Paul: I’ll interject here and just say, just as a point of, I guess, education, and we talked about a lot about it on the podcast. But modern neuroscience is kind of enamored with these dimensionality reduction techniques that work especially well in constrained laboratory tasks, like reaching where there is a very specific movement that an animal will make in various directions. Right. And then you take a high dimensional data recording, let’s say thousands of neurons, and you can use dimensionality reduction techniques. And then instead of reading out the thousand dimensional neural activity, you reduce it down to a few important dimensions based on the variance of how all these neurons are covarying with each other and et cetera. And then you can plot out the trajectory of this low dimensional trajectory, and there are distinct trajectories for, let’s say, different arm reach directions. Right. And this has been kind of a powerful way to consider especially what’s happening in motor cortex and in motor areas in general.

And why did I interject that? I interjected that because.

Right. So I highlighted that that’s been very effective, especially in these constrained tasks. And so it’s an open question of whether, in these unconstrained, kind of naturalistic tasks that you’re going for now, whether that same kind of dimensionality reduction approach will be as effective or what will come of it, essentially.

[00:56:50] Eric: Right. And we know that when it comes to just simple, in the most basic and maybe abstract form of neural encoding, we know lots of neurons behave, have firing rates and responses like lots of other neurons, and this redundancy of representation almost has to happen. No matter how you think the brain is organizing itself, we are losing and gaining tons of neurons all the time as adults. It’s more on the losing side, but we aren’t fundamentally different because of that. And so the idea of saying, hey, can we take these thousand neurons that are, some of them are more like each other than others, they’re all their own, maybe special snowflakes, but we can get a composite distilled down summary out of these neurons, I think has to be the case no matter what. But you’re right, I think when it comes to the dimensionality of representations as well as the complexity of these dimensions, people like to talk about manifolds, which is sort of the shape that these neural trajectories take on.

We don’t know about the complexity of how many dimensions, how many of these. If you think of each dimension as a feature that may be repeated or shared to some extent between several neurons, we don’t know how many there are and how complex that is sort of laid out.

And that’s one of the reasons that we’re getting into this, because if you train and train and train an animal to do a task, and they only have one thing to worry about, it’s reasonable from a strictly hand waving perspective to think, well, we might not be getting the whole picture. The brain is sort of in this, I don’t want to say default state. That term is taken in this simplistic state of just needing to do whatever you’re going to do. And we only hear from active neurons anyway. So there may be plenty of other neurons that are representing other dimensions of behavior or whatever else is going on for the brain that you’ll just never see in this format. I’d say one other part is that over training itself, besides the fact that these are typically much, are very simple things, the overtraining part of it, the practice and practice and practice may have, and almost always goal directed, I think, is a factor that is rarely brought up, may have profound changes on how we interpret those signals. And it’s been known for years that if you put your electrode in the brain early on in training, and this is across species, across many tasks, and even if you try and match the kinematics of individual trials as best you can early on training, you do not get the pretty response profiles that you associate with well trained tasks at a single neuron or population level.

Are they there? Maybe something has to be going on, but it’s not very obvious what’s going on.

[01:01:23] Paul: It’s not like, let’s say it’s a reaching task. It’s not like you were unable to reach before. It’s just that you hadn’t been reaching in that context, under that goal directedness, in that specific location, that you need to be within a few millimeters of a certain area to get a reward. So there’s a lot going on, and you’re really shaping the brain, in that case, to reach in a certain manner. So it’s like you’re creating something to study it. So you’re not studying reaching or behavior reaching behavior in general. You’re studying reaching behavior in a very specified, goal oriented task under reduced conditions, et cetera. Right.

[01:02:07] Eric: Yeah.

[01:02:09] Paul: I don’t mean to pick on reaching in particular.

[01:02:11] Eric: Yeah, that task is still being done in our lab, and there’s some great work. But just to make my perspective clear, you said that a little bit sort of tongue in cheek, that, oh, this is way too abstracted. And there’s all these potential problems. Once again, the right approach for the right question is fantastic. And so we have a senior grad student, Mark Nicholas, who is getting after this.

What does motor cortex do and what does strytum do? And lesion? Motor cortex in a reaching task. And lo and behold, without m one, if you look immediately after and in the days after that lesion happens, mouse can’t do that task. It can still.

So in the second part of the paper, we actually look at a mouse walking down a hallway. It walks pretty well until it has to make a decision. And Sylvia Arbor and several other groups have similar effects to this.

But in this case, again, a trained artificial task can be very useful. In this case, we’re trying to complement both sides of it, of the well trained and the spontaneous. When we have the animal walking down the hall, there’s no reward.

But I will say you listed off the different goals of you have to be within a centimeter of the target or the reward.

All these facets that you just named a minute ago, and several more are really important to us and why we are going after these more naturalistic, spontaneous behaviors and applying algorithms to try and give the same structure that we would have with our trained trials.

And we’ve had a fair amount of success with that.

But looking at if we pull that away and we’ll soon be starting to sort of pepper in some of those caveats that you mentioned of, oh, but what if this is a for reward? Or what if this is a fairly repeated, within a certain context behavior, trying to build a bridge? But you’re right in a sense, although I don’t think things will be night and day different.

There’s a lot of opportunity for learning what the brain does in its spontaneously generating behavior state.

And again, you were focusing on learning before. We’re constantly learning. Sometimes we’re focused on learning or not. But if we can’t learn a particular behavior, learn how to interact with something, our fitness goes down and we’re not going to pass on our genes.

So this is, the tasks certainly have a place, have their place in neuroscience and in ethology itself, because it’s, it’s not that sure.

[01:06:09] Paul: Yeah.

[01:06:09] Eric: This animal is created as an adult with everything it’s ever going to do, and it’s never going to update according to the chaos and conflict in its environment. Yeah. So it’s all important.

[01:06:24] Paul: One of the things that you said to me maybe months ago, I don’t remember if you said this the first time.

[01:06:30] Eric: We haven’t been allowed that long.

[01:06:31] Paul: Well, it’s been a few months, but anyway, it has stuck with me, is that, it’s a very simple point, but it really stuck with me.

You and I right now are behaving when you’re sitting down doing nothing. You’re behaving.

[01:06:49] Eric: I’m doing so many things when I’m sitting down doing nothing.

[01:06:52] Paul: You are not. Everybody is. You are.

We know the octopus is doing nothing when it’s sitting. But. Yeah, no, I think that’s one of the challenges. Right. Is, for example, what is a behavior, which is a very simple question that we don’t have a great answer to, but it’s been studied philosophically for a long time, et cetera. But just the fact that you and I right now are behaving and presumably, presumably our brains have something to do with that. Behavior. Let’s say that there’s a correlation between our brain activity and our behaviors. Right now, at least.

[01:07:31] Eric: Let’s hope so.

[01:07:32] Paul: Yeah, let’s hope so. And so when you have a mouse just kind of wandering around and exploring naturalistically, behaving, a lot of the times, it’s kind of sitting there and might kind of turn a little bit. And there are all these nuances that are just that neuroscience, for decades has sought to get rid of so that we can control exactly what we can measure, and so that we’re measuring a repeated behavior over and over. So then we can average the trials together and look at the average neural response and then compare conditions. And if these 30 trials in this condition, then we can compare it to 30 trials in that condition. But then all of a sudden, you have to kind of throw that out the window, which I think comes with challenges and is exciting.

[01:08:20] Eric: No, ma’am.

Yeah.

So you hit on a couple really important points there.

We are constantly in the middle of doing a behavior, and even the behaviors that we think we understand, it can be very complex to get back to reaching again in order to extend your hand out or to pull it back towards you.

We always think, oh, the shoulder and bicep, and you’re engaging all of your trunk muscles and the opposite side.

[01:08:59] Paul: That’s why my core is so strong. I’m reaching a lot. That’s how I work out my core.

[01:09:06] Eric: Your core is probably far better than mine. And I teach systems neuroscience to the undergrads here. And I love the example in the book that if you’re to pull a door open, the first muscles to engage are your calf muscles, because they’ve got to be able to pull you backwards so that you can pull against the door.

[01:09:29] Paul: Otherwise, you really just got to slam into the door.

[01:09:32] Eric: I’ve done that, too. More often, it’s the pushing, and the door fails to move because I don’t know how to read.

There’s a Gary Lawson, one of my favorite. Oh, yeah. School for the gifted.

[01:09:45] Paul: School of the gifted, yeah.

Anyway, to flesh that out, the door says pull, I think. And he’s pushing on the door, and then above the door, there’s a sign that says school for the gifted, something like that.

[01:09:59] Eric: Common sense ain’t so common.

Right. We’re constantly in a behavior which does make some of these tasks a little bit curious, that we think of the behavior as starting at point x, and then once they get to the target, it’s done. Same thing with learning.

Many people in the lab have probably heard dozens if not hundreds of times. That the fact that we’re always updating our internal models, we’re always learning. And that I get why some papers will say, oh, learning is here and performance is here.

There’s going to be different sort of times when there’s more focus, but it’s more likely.

How accurate or how much do you have to update your model?

I think you’re constantly tweaking and updating, and there’s good evidence for this.

There’s a lot going on.

And in terms of all of those nuances, we do our best. So a few years ago, an insanely gifted graduate student, Alex Su, we came out with a platform called b side, spelled B Soyd, which all sounds the same if you say it with a Brooklyn accent. I’m not going to do it on this show.

I know better than that.

But you can do it if you’d like to.

[01:11:45] Paul: I’m all right, thanks.

I really wanted to hear you do. I would butcher it.

[01:11:50] Eric: I did just a hair of it.

[01:11:52] Paul: Brooklyn. Okay.

[01:11:53] Eric: Yeah, I think.

Anyways, that takes pose estimation data. For instance, deep cut sleep, deep lab cut open pose.

[01:12:07] Paul: These are modern tools.

You’ll do it better than I do.

[01:12:14] Eric: Computer vision algorithms that reduce the thousands of pixels into a video down to key points that are essentially motion capture suits on an individual, but does so with modern computer vision techniques.

And so we can take those data, which are very informative, much more informative than the average pixel in a video.

And we look at the spatial temporal pattern of those movements. And anyone who has seen a motion capture suit, you can tell when someone is walking or jumping up and down just by the twelve points that you see on the body. And so b side is an unsupervised learning algorithm that identifies spatial temporal patterns in those points and adds interpretability to the pose estimation data by extracting those spatial temporal patterns that we would call behavior.

The code itself will work on anything where you need pattern recognition. So we’ve used that to get at trying to identify behaviors.

[01:13:37] Paul: And what it ends up spitting out is essentially categorical variables that correspond to these kind of repeated motifs that you might consider locomotion or turning left, turning right, things like that. That it groups, that it clusters into these common motifs. Right.

[01:13:55] Eric: It only knows it as group two, group twelve. But it turns out quite nicely that they tend to map on quite well to established ethological behaviors like locomotion. We actually got, there’s sort of three canonical types of grooming a mouse will do of sort of the face, the bigger motion of the head. If you think of a cat or a dog sort of pulling its ear down to clean, that’s the head groom and then turning back to groom the body.

We didn’t tell it anything, and that’s sort of the beauty of an unsupervised approach.

Whatever you pull out has some measure of robustness to it, or believability, in a sense. And we readily get those three different groups of behaviors. So then our toolkit, we actually have a paper that just came out in nature methods for a supervised approach. So we call it a side. We’re staying with the record dad joke of the a side and b side, where you might find a deep cut.

[01:15:13] Paul: That’s a reach. I like it. I like it’s a reach.

[01:15:15] Eric: That’s the basics behind it. Our emblem, actually, for a side is a record with an a on it.

I was at a workshop recently, and someone said we should get. What was it?

Get a version that’s specific for fish and call it seaside.

Or we’ve got. Yeah, that one was. I was both very proud and a little bit disappointed that I didn’t come up with it myself.

[01:15:46] Paul: Right. Yeah. Thinking of D now, but d side might actually happen.

[01:15:51] Eric: So currently, a side is both this unsupervised b side as well as a side. And you can decide, sorry.

Where you get into the behavioral segmentation side of things. But for more nuanced or specific behaviors, that supervised eye can be really important. And both of them are powerful tools. Again, pick the right approach for the right question.

But we’ve been using that to try to get us some sort of handle on these naturalistic, spontaneous behaviors.

And the fact that it maps on to what pathologists have been sort of honed in on for decades makes me feel better about the results.

And that serves as a jumping off point for then. Oh, if we subdivide into longer or shorter brooming or rearing or locomotive bouts or faster or slower or other dimensions, that we can make things more similar, we can get into those nuances. But we have a starting point. And something that in any talk I give on this, that I make a big point out of, is you mentioned before, we don’t know what motor cortex is doing. I think we have a better idea now than we ever have. But especially for these naturalistic behaviors, we didn’t have to have any sort of specific representation.

And in fact, what we found is throughout motor cortex, dorsal stritum, and to some extent, even in ventral stritum, although the reasons why we see it in ventral stritum, I think, are different. We have a robust response that’s specific to each type of behavior and certainly in our deeper layers of motor cortex and dorsal stritum. With just a very simple, simple decoder, we can predict which of 16 behaviors a mouse is doing from just spike counts at a rate of around 60%.

With an f one adjusted score, which is one over 16, is a lot smaller than 60%.

So there’s something there.

It’s not 100%. If it was 100%, I would know we were wrong, in part because that’s just not how biology works.

[01:18:51] Paul: What if it was 103%?

[01:18:54] Eric: Then that’s not how math works.

But yeah, it gives us some new abilities to ask questions about what is action? What is behavior?

What in a naturalistic state that evolution is more often pushed these animals towards?

How are these things represented? How are the decisions made?

How does learning and adapting of these dimensions of representation change?

And certainly, thanks to tools like neuropixels and other recording modalities, we can start asking not just what does a handful of neurons within one area, but how do multiple areas interact?

And so we have complex, or I’d say a rich account of behavior and a rich account of neuroactivity to try and get that abstract thought to action. How is that transformation made?

[01:20:19] Paul: I wish you were busy this afternoon.

We had taken a few minutes. I could have shown you these trajectories in GPFA. It’ll be tomorrow. Tomorrow. We’ll do great.

[01:20:35] Eric: You’ll just have to publish it.

[01:20:37] Paul: Sure, you bet.

[01:20:38] Eric: Real soon and discuss it on here.

[01:20:40] Paul: I’ll come on the brain inspired podcast and talk about it.

It’s trucking.

[01:20:48] Eric: Nepotism at its best.

[01:20:50] Paul: Yes. Narcissism at its best. That would be.

[01:20:53] Eric: Oh, yeah, I guess for you, I.

[01:20:55] Paul: Would interview myself, right? Yeah.

[01:20:57] Eric: And I will say, if anyone’s watching the video, I’m going to undo this. Paul specifically requested that I did, and.

[01:21:06] Paul: It was a half joke. Okay, so you actually wear that when you teach?

[01:21:11] Eric: I put on a bow tie for teaching. I’ve loved bow ties for years. And it’s a nice opportunity because I’ll admit one thing I love about academia, at least more on the science side than the engineering side. I can come to work in sweatpants, and that’s what I normally do and do.

[01:21:32] Paul: Yeah, that’s right.

[01:21:33] Eric: Which?

Bow ties and sweatpants. Although I do have a somewhat eclectic fashion style.

[01:21:40] Paul: You do? I can vouch for that.

[01:21:42] Eric: I know better than bow tie and sweatpants.

I’ve learned over the years well, thank.

[01:21:49] Paul: You for wearing it. I appreciate it that you didn’t. I did it in half jest. And you actually came through for me, so I appreciate it.

[01:21:57] Eric: You got to support the people in your lab.

[01:22:01] Paul: That’s right.

So while you were talking about some of the challenges and some of the tools that we have and how b side can categorize these things, and it turns out that it matches very well the categories that b side comes up with matches very well to what has been described in the past by human know, by ethologists.

And I was reflecting on our earlier conversation about artificial networks and using them to study things in cognitive neuroscience. And this is a case where, like, behavior is a continuous, ongoing thing, and we’re just talking about the syllables and motifs. Right. And so in a sense, what we’re doing is saying, well, how can we apply the old way of doing things to this new kind of continuous, flowing behavior? Right. So we want to break it down into things that we can chunks that we could discrete chunks that we can study. Right. And say, when does it handles. Yeah. When does it start? When does it stop?

And in a sense, then you could use, like.

So what I was thinking is that an artificial network, well, you just turn it off and come back tomorrow. And you can’t do that with behavior. Right. Which is just a continuous. I mean, you can study it tomorrow.

[01:23:19] Eric: I mean, comas exist, but you’re not.

[01:23:22] Paul: Behaving during a coma.

What about sleep?

[01:23:29] Eric: Sleep is a funny one.

[01:23:33] Paul: Anyway, before we get on, move on. Yeah. So I do want to come back to that, but it just struck me that.

[01:23:39] Eric: Have you had Emory Brown on the podcast?

[01:23:41] Paul: I have not, but he’s on the list.

[01:23:44] Eric: Yeah. Ask him that question.

[01:23:46] Paul: Okay. Whether sleep is a behavior is behaving. Yeah.

[01:23:50] Eric: And the multiple consciousnesses thereof.

[01:23:53] Paul: Right? Yeah. His specialty. And coma.

[01:23:58] Eric: I will say b site can pull out.

No, we are not approved for an Iacirk for that.

It does pull out different poses.

And we’ve noticed that there’s different frequency band during sleep.

We have noticed different frequency bands related to those poses. And I believe there’s been a paper or two that through other means, I forgot if it was an automated method or not, has some correlation there.

Not the answer you’re looking for of behavior, but b side categories may apply to different stages of sleep.

But anyways, you were going to ask something far more.

[01:24:56] Paul: No, I was just noting that you can turn computers off, essentially. Right. So using an artificial network, which you run a pass through, and yes, you can study the dynamics which has been when you study recurrent neural networks, to compare them to the dynamics of decision making in brains, you can compare the dynamics. But anyway. Yeah, anyway, I was reflecting on the fact that artificial networks are static entities and living organisms are ever changing, ever behaving constantly. Continuous behavior, neural activity, flow. And it’s just a fundamental difference, though.

[01:25:36] Eric: The turning it off. I don’t see as that much of a difference. I’d say it’s more the initialization, because also certainly an ANn or a deep net can continue to learn or run and change and do whatever it’s going to do given the opportunity, know provide new information to.

[01:26:06] Paul: But I think I was thinking mostly in terms of it’s going to spit out a discrete thing and then it’s going to spit out another discrete thing. And that’s kind of what we’re getting at by.

So there’s a tension in me about applying that kind of approach to studying something that I see as continuous.

[01:26:26] Eric: And that is a big, open question that I think that will be one of the goals of this naturalistic behavior field moving forward is how continuous is behavior, both behavior itself as well as its neural representation?

Because although we can identify motifs for these individual naturalistic behaviors pretty well, surprisingly well. So I think we’ve made a decent foothold there.

It’s going to be messy. It almost has to be messy.

And I think better word for messy, I think, would be complexity. And it may be that in some areas, spinal cord, it’s not going to be a lot of messiness.

There’ll still be plenty of complexity, and there’s constantly just incredible work being done on that side of things.

Good coverage, good recovery. Yeah, I know what to do, but that’s going to be a lot closer to the muscle, and it kind of has to be a more straightforward something closer to a linear operator. Whereas you get into motor cortex or these cognitive areas, where I’m sure if we plunk down electrode into anterior cingulate or pfc. Ofc. Prow cortex, name your favorite area, I bet you we can still find a decoder for behavior. And who knows, v one.

I wouldn’t be entirely surprised if we can do well above chance from v one. It’s not the first experiment I’m going to do, but in just getting at the complexity and the sort of fidelity to the discontinuous idea of you’re doing behavior a and then you’re doing behavior b, we can get a sense of how these are orchestrated and how that information is handed off from one area to another. And the computational processes that are going on in this paper that’s in the works.

Not only do we show that we can get a robust response from individual neurons for some amount of neurons, for all these areas, motor cortex is much more promiscuous in what it shows a tuning for in terms of the behaviors.

So the average neuron in motor cortex, particularly layer five, will respond to four, five, six, even eight of the 16 behaviors that we’re keying in on.

Whereas when we look at dorsal stritum, we see much more sparsity in its representation. Talking two or three behaviors, these market differences, which are sort of related to this idea of how continuous or discontinuous behavior truly is, I think are important because at the end of the day, we got to do whatever we’re going to do.

But there’s a whole lot of fudge room in there for far more complex things. We are not automatons.

And so this is also part of my difficulty with the action selection idea of the idea that discrete action channels, although that model has plenty of attractive things in terms of a theoretical computation, that’s fun to play with the math, as well as a useful framework through which to view brain and behavior, how you would actually form an action channel, and the discreteness that that implies.

I’m not saying that my half baked ideas are any better, but that has always rubbed me the wrong way.

[01:31:47] Paul: Fair enough.

[01:31:48] Eric: So, yeah, we’ll see just how discontinuous behavior is, and almost certainly different behaviors will occupy different parts of this complexity spectrum.

We’re actually recently struck up collaboration with Jose Mora, who’s one of the fathers of recent signal processing work, and I’m excited to have that. And really, anyone who is capable of asking some of these questions that I think are like any question in neuroscience, there’s going to be some assumptions, but there’s a lot of openness and unknowns that I think we can make some real headway on.

I imagine more and more in our lab, we’ll focus on this naturalistic side, but it’s still asking the question of how these circuits locally and in aggregate lead to the behavior of interest.

It’s going to be exciting for a while, and not just us. So many cool people doing it.

[01:33:24] Paul: Yeah, there’s a lot of people doing it, that’s for sure. It’s going to be exciting for a while.

[01:33:29] Eric: And then I’ll move to octopus.

[01:33:31] Paul: Then you’ll move to octopus.

Probably every scientist says this in the midst of their quote unquote time, but it seems like we’re in an exciting time because computationally, we’re way more powerful, all of us, in the past ten years, and it’s just continuing to increase. We can record more neurons than ever.

We have more tools, optogenetics, right? We can track neurons and stimulate them and watch them.

So, on the other hand, it opens, like, challenges of how to use the tools correctly, how to ask the right questions with the tools. Right. And it’s like a confluence of all these things. And now the move into naturalistic behavior. So there’s a lot to look at, right, or neuroethological behavior. There’s a lot to look at with these new tools. I did that without a hitch.

And earlier, I asked you if there’s anything you missed about the old days, right. Working with non human primates, with recording bus neurons, you’d go down with a single electrode, usually, and just listen for the sounds like it’s doing what you wanted it to be doing, and you record it, and there’s something wonderfully simple about that. And things are not as simple anymore. Everything’s a low dimensional manifold. From recording your high dimensional data, et cetera, et cetera. And there’s a lot more statistical analyses available.

Does a part of you wish that you were a neuroscientist 100 years ago? Or if you had to choose, would you go back and be a neuroscientist 50 years ago, or would you be right now? Or would you be 50 years from now if you had to transport yourself or not?

[01:35:17] Eric: As a curious person, 100 years from now is ideal, because we’ll know a whole lot more.

I don’t care if I’m right or wrong.

[01:35:28] Paul: I love that about you, by the way.

[01:35:33] Eric: I think it’s a critical part about science. But as it was phrased to me by one person, someone disproved.

It was during grad school, Greg D’Angeles, while he was still at washoo.

It says, someone cared enough to prove you wrong, and it’s still reading your papers. But no, I want to know what the answers are as a curious person.

So, yeah, I’d pick the future because it’ll take a while. And I think you mentioned our experimental tools being very robust.

We have Carson and others, sorry, recording tens of thousands of neurons with calcium imaging and Allen Institute and all the Josh Siegel and dozens of others with all their electrodes. Nick Simon.

So we have tons of, or even just with the single neuropixels probe. We have tons and tons of neurons and constantly more complex behaviors, trained and otherwise.

The annoying question that I like to ask people is if you could record action potentials from every neuro in the brain, how would that change your science?

[01:37:07] Paul: Classic annoying question.

[01:37:09] Eric: Oh yeah.

Typically the best answer I get is I wouldn’t have to do as many experiments because I would have my n for neurons would be very big.

[01:37:25] Paul: You’d have the only data set you needed, which is great.

[01:37:29] Eric: But yeah, I will say that I hope and I’m optimistic that in the next few years we will have more tools to deal with the volumes of data that you pointed out which are made possible by new experimental tools. But I think the answer that I was about to say is, as a neuroscientist.

Yeah. I would prefer 50 years ago to 100 years from now.

[01:38:03] Paul: Oh, that’s interesting.

[01:38:05] Eric: In part because of what I find really exciting. If you look, 50 years ago would have been, oh, that’s a scary thought.

1974.

[01:38:19] Paul: I know it’s already like late in the game in terms of a lot of things that are like the classic things that have already been discovered, essentially.

[01:38:26] Eric: Yeah.

I still think 2005 was six years ago, but that’s a different topic. Yes.

Think about all the titans.

Mount Castle, Everett Georgiapoulos.

These are mainly motor people. Eccles speaking of cerebellum, although most of his stuff was a bit earlier.

You’ve got all the great stuff in v one still sort of happening though, I guess. Still a little bit earlier. Yeah, you have all these pillars of findings and sort of in a very small way. And also to be clear, I am not saying that I am on par with eccles, but the sort of questions that all of them were going after are what really would excite me.

I just showed video of Hubo and weasel with their light beam.

That’s sort of the classic. They move it across the screen and you can hear each action potential from a neuron. And v one. I got so many weird looks setting up my lab. I said that I need every amplifier to come with a speaker so I can listen in on a neuron if I want to.

[01:40:15] Paul: You’re speaking to the choir. You’re preaching.

[01:40:17] Eric: Yeah, it’s a holdover from platinum iridium single unit electrodes in a monkey, but it is so satisfying.

I’m glad to have my multichannel electrodes.

[01:40:33] Paul: But true, no one hears very many. What would you listen to?

[01:40:39] Eric: Yeah, hear the tune it into just one of 384 channels, which you can do. We’ve done it in lab.

But the sort of questions that they were after and the way they went about it, there is something I don’t think it’s romanticizing it, but I think the questions and the approaches that I think are a whole lot of fun are one that’s kind of like what we have for naturalistic behavior.

As you said, it’s exploratory. We don’t necessarily know up from down, and that there are likely, hopefully, knock on wood, some low hanging principles, low hanging fruit of principles that we may be able to get at that years and years in the future.

There’s still the new principles to be found, but they’re more niche. You’ve moved down the dendrograms of your sub sub subfield more and more.

So there’s something sort of exciting to me, the creative approaches, really the same single unit recording that we’re talking about, it was still in the infancy 50 years ago, which is something that we don’t necessarily, we sort of take it for granted or saying, what is prior cortex doing? We’ve had some of this in a clinical sense, know Broadman areas, et cetera.

But yeah, guys like Mount Castle just marching around and saying, this is what I see. It’s a very cursory, I know, but.

[01:42:56] Paul: There’S something so nice about being able to do that. All right, Eric, this has been weird and fantastic.

[01:43:04] Eric: We’re not that.

[01:43:06] Paul: No, no, you didn’t make it weird. It’s just I’m about to say two things to you that I’ve never said to a previous guest. One, I’ll see you tomorrow, and two, I’ll bring the octopus.

[01:43:23] Eric: If it’s to eat. Don’t. I’m allergic. Please don’t.

[01:43:27] Paul: All right, thanks for doing.

[01:43:28] Eric: I hear it’s delicious. But yeah, you twisted my arm, but this was a lot of fun. And hopefully someone else thinks this is intelligible and maybe even interesting.