WEBVTT

00:00:03.517 --> 00:00:10.517
This is the Convergent Science Network podcast. Leading researchers in the domain

00:00:10.517 --> 00:00:16.797
of neuroscience, brain theory and technology are interviewed by Paul Verschoor and Tony Prescott.

00:00:19.917 --> 00:00:26.477
Okay, so I'm Paul Verschure with our Barcelona Brain Cognition Technology Summer

00:00:26.477 --> 00:00:32.177
School here with Hillel Kiel from Case Western University, a reserve university, sorry.

00:00:32.477 --> 00:00:35.957
Well, I pronounce it Hillel Kiel and I'm at Case Western Reserve University.

00:00:36.677 --> 00:00:40.757
Almost right. It's an amalgam between Case Institute and Western Reserve University

00:00:40.757 --> 00:00:43.717
and everybody gets confused by the name, so no problem.

00:00:43.917 --> 00:00:48.457
I'm not the only one. No, you're not the only one. Welcome to the Conversion

00:00:48.457 --> 00:00:49.537
Science Network podcast.

00:00:50.417 --> 00:00:53.057
And so you were speaking in our summer school.

00:00:55.297 --> 00:00:58.937
And for us, it was a great opportunity because you have been active in this

00:00:58.937 --> 00:01:04.257
whole domain of trying to understand biology at sort of a system level.

00:01:04.677 --> 00:01:09.857
And to also bring this together with a very much technology-oriented view and

00:01:09.857 --> 00:01:12.337
using robots as a way to test theory and so on.

00:01:12.357 --> 00:01:17.697
That's exactly where we want to be. and you are really one of the leading characters

00:01:17.697 --> 00:01:23.217
in this field but now so what I saw was really interesting,

00:01:25.515 --> 00:01:29.215
very, very impressive in your talk and very useful is you really try to bring

00:01:29.215 --> 00:01:30.275
us to this list of principles.

00:01:30.495 --> 00:01:33.515
So where you say, look, I look at these systems and here I'm really going to

00:01:33.515 --> 00:01:35.795
give you principles. Of course, we can argue then about these principles.

00:01:35.915 --> 00:01:37.675
And we did. That's probably what we will do again.

00:01:38.755 --> 00:01:41.655
But I think this is really where we want to go, right? So we want to extract these principles.

00:01:41.955 --> 00:01:46.535
And you started fairly simple with, well, of course, simple can still be complicated,

00:01:46.575 --> 00:01:52.495
but rather low scale, let's say, the small scale neuromuscular systems.

00:01:52.695 --> 00:01:57.855
That's right. So maybe we can start a discussion there that you can try to describe

00:01:57.855 --> 00:02:02.115
a little bit what are these key principles are that you found at that level

00:02:02.115 --> 00:02:03.155
of system organization.

00:02:03.615 --> 00:02:06.555
So let me step back and say something about the general principles thing.

00:02:06.595 --> 00:02:09.815
I didn't show this slide, but this is one that actually it's in one of the reviews

00:02:09.815 --> 00:02:14.915
I wrote in 1997 with my colleague Randy Beer called The Brain Has a Body.

00:02:15.175 --> 00:02:19.075
And that's a particularly influential reference. That's probably one of my few

00:02:19.075 --> 00:02:23.775
citation classics. If you go to Google Scholar, it's got, I don't know, over 400 citations.

00:02:23.955 --> 00:02:26.515
For a neurobiologist, that's pretty impressive.

00:02:27.175 --> 00:02:30.055
Unless, of course, you invented a new technique that everybody's using.

00:02:30.755 --> 00:02:37.295
And in that, what we presented was the idea that what evolution selects for

00:02:37.295 --> 00:02:40.015
are not bare brains or particular bodies.

00:02:40.215 --> 00:02:44.115
It's a coupled system of the brain, the body, and the environment.

00:02:44.395 --> 00:02:49.995
So brains are embedded and they are embodied within a particular body with a certain biomechanics.

00:02:50.395 --> 00:02:55.975
And then the agent as a whole, brain-body dynamic, has to work in an environment

00:02:55.975 --> 00:02:57.695
and has to function in that environment.

00:02:58.335 --> 00:03:02.115
So what I was essentially doing, sort of you keyed in correctly,

00:03:02.355 --> 00:03:03.675
we're looking for general principles.

00:03:03.915 --> 00:03:09.755
But in many cases, what's informing our search is that larger framework,

00:03:09.895 --> 00:03:12.315
this idea of the couple dynamical systems framework.

00:03:12.615 --> 00:03:17.095
And one way to look at it is to say, okay, many neurobiologists are,

00:03:17.295 --> 00:03:21.135
you'll excuse the expression, they're neurocentrists.

00:03:21.435 --> 00:03:24.335
For them, if it's not the nervous system, they're not interested. it.

00:03:24.375 --> 00:03:27.035
And they think that the most important thing might be discovering a new channel

00:03:27.035 --> 00:03:29.615
or a particular molecule and stuff like that.

00:03:29.735 --> 00:03:32.195
And again, I have a lot of respect for the reductionist approach.

00:03:33.392 --> 00:03:38.172
From the point of view of this larger framework, what evolution selects for is the entire package.

00:03:38.632 --> 00:03:42.992
And so a particular channel, well, if it's dysregulated, affects behavior,

00:03:43.112 --> 00:03:44.832
and then the animal dies, yes.

00:03:45.252 --> 00:03:48.152
But otherwise, it may just do, it may be one voice in the crowd.

00:03:49.392 --> 00:03:52.072
Biomechanics is something that many neurobiologists don't focus on,

00:03:52.152 --> 00:03:56.692
because what they find is if they can throw the body away and just get the brain

00:03:56.692 --> 00:03:59.252
in the dish, they can do all their experiments very quickly.

00:03:59.712 --> 00:04:03.772
So the question was, what is the role of the biomechanics?

00:04:03.792 --> 00:04:07.172
And in a sense, that was one of the themes that ran through several of the other

00:04:07.172 --> 00:04:12.792
issues in the talk, though I started with that and focused simply on an unusual

00:04:12.792 --> 00:04:14.772
biomechanical periphery.

00:04:14.872 --> 00:04:17.992
And that sort of helps sharpen people's focus.

00:04:18.192 --> 00:04:21.812
We're used to, if you're a roboticist, you're certainly used to thinking about

00:04:21.812 --> 00:04:26.592
actuators, you're thinking about joint torques, you're thinking about kinematics,

00:04:26.812 --> 00:04:31.272
you know about the problem of how ill-posed it is if you have excess degrees of freedom.

00:04:31.872 --> 00:04:35.212
But what happens if you're confronted with a completely different body plan?

00:04:35.392 --> 00:04:40.232
It's all soft, and it has potentially, you know, it can wrap itself in knots.

00:04:40.792 --> 00:04:45.992
So one of the things that makes that, I think, a nice place to start is it disorients

00:04:45.992 --> 00:04:49.092
people enough to get them thinking about why might this be, you know,

00:04:49.092 --> 00:04:50.732
how would you do this? How would you control this?

00:04:51.192 --> 00:04:55.412
I know Frank Grass is going to be talking about cephalopod control and the kinds

00:04:55.412 --> 00:04:56.772
of things that octopi can do.

00:04:56.952 --> 00:05:01.032
But even our tongues, or especially our tongues, are incredibly flexible devices.

00:05:01.572 --> 00:05:06.072
Our conversations being, we're using, I use the term muscular hydrostats,

00:05:06.192 --> 00:05:09.452
that was originated by Kieran Smith.

00:05:09.772 --> 00:05:14.392
And this is, this challenges your whole idea of how control works.

00:05:14.732 --> 00:05:17.752
And the interesting thing was, we initially thought, well, this is going to

00:05:17.752 --> 00:05:18.652
be incredibly complicated.

00:05:18.912 --> 00:05:20.292
What we discovered is you actually

00:05:20.292 --> 00:05:23.692
analyze the mechanics, there are some simplifications that emerge.

00:05:24.052 --> 00:05:28.632
And then that seemed to me a useful principle to enunciate. That, and again.

00:05:29.486 --> 00:05:35.206
I take your points very well, that that principle is a typical biology principle.

00:05:35.326 --> 00:05:37.686
It's not the same as a standard, say, physics principle.

00:05:38.206 --> 00:05:41.986
In a given context, when you talk about a specific behavior,

00:05:42.546 --> 00:05:46.906
if you define what the biomechanics does in that context and in that behavior,

00:05:47.206 --> 00:05:50.266
then that will simplify enormously the control questions.

00:05:50.606 --> 00:05:54.446
Right. So can you give some examples of this kind of simplification that you

00:05:54.446 --> 00:05:55.686
get through through to biomechanics.

00:05:56.106 --> 00:06:00.566
So one of the things I talked about in the particular model we did of a tongue.

00:06:01.106 --> 00:06:04.866
So you have this longitudinal muscle that runs down the center of the tongue,

00:06:04.906 --> 00:06:07.086
and you have a circumferential muscle that wraps around it.

00:06:07.106 --> 00:06:10.186
You could think of it like, if our listeners are trying to form an image,

00:06:10.326 --> 00:06:12.526
it might be helpful to think of a hot dog in a bun.

00:06:13.226 --> 00:06:17.806
Okay? So you have the central muscle, the longitudinal muscle is the hot dog,

00:06:17.826 --> 00:06:19.246
and the bun is the circumferential muscle.

00:06:19.606 --> 00:06:24.906
So if the central muscle contracts, what happens is the whole thing gets short

00:06:24.906 --> 00:06:27.126
and fat because the volume can't change.

00:06:27.406 --> 00:06:30.566
Now, what's fascinating is this. Again, I'm not going to go through this in

00:06:30.566 --> 00:06:33.126
any detail, but I showed a little bit of the math in my talk.

00:06:33.706 --> 00:06:38.086
When the tongue is long, it turns out it has a huge mechanical advantage relative

00:06:38.086 --> 00:06:39.326
to the circumferential muscle.

00:06:39.826 --> 00:06:44.366
What that means is for neural control, if I put just a pulse of activation in

00:06:44.366 --> 00:06:47.526
to the longitudinal muscle, it's going to shorten immediately.

00:06:47.966 --> 00:06:51.466
And the circumferential muscle, even if it's working very, very hard,

00:06:51.646 --> 00:06:56.386
will not be able to overcome those forces until the whole tongue has gotten pretty short.

00:06:56.926 --> 00:07:01.686
So it almost drops out of the control picture. In fact, if you didn't activate

00:07:01.686 --> 00:07:03.926
at all, there are passive forces that would stop it.

00:07:04.586 --> 00:07:08.786
So the real key issue is you might want to activate it because you want to extend

00:07:08.786 --> 00:07:10.446
the tongue again and get it going again.

00:07:10.606 --> 00:07:14.266
But if you didn't, if you just wanted a single lap, You could ignore the circumferential muscle.

00:07:14.666 --> 00:07:18.626
So this is something that would not be obvious to somebody unless they'd sat

00:07:18.626 --> 00:07:19.726
down and looked at the mechanics.

00:07:20.046 --> 00:07:23.306
Once you look at the mechanics, it becomes very obvious. And then you can look

00:07:23.306 --> 00:07:26.166
at the neural control that is actually used, that has evolved for it,

00:07:26.206 --> 00:07:28.246
and say, oh, that's why these features are there.

00:07:28.926 --> 00:07:33.566
But you could also argue that maybe this is, let's say, just a happy coincidence

00:07:33.566 --> 00:07:37.366
because this tongue has to fit in your mouth, for instance. Okay.

00:07:37.406 --> 00:07:42.146
So if this is really where building robots becomes the crucial thing,

00:07:42.226 --> 00:07:47.006
because it's from an evolutionary standpoint, there may be any number of possible

00:07:47.006 --> 00:07:51.126
historical accidents that led you to have the tongue in the particular shape it is.

00:07:51.426 --> 00:07:56.126
By building the robotic device, you can actually say, well, at least is it physically

00:07:56.126 --> 00:07:59.566
realistic to say that this phenomena happens in this way?

00:07:59.566 --> 00:08:03.506
And when you do that and you see that it does, as I emphasized in my talk,

00:08:03.606 --> 00:08:06.006
and I have to always emphasize this to anyone who's listening,

00:08:06.146 --> 00:08:08.466
it doesn't prove that it works that way in biology.

00:08:08.686 --> 00:08:12.646
You're raising the key question. There may be other contingencies, historical accidents.

00:08:13.876 --> 00:08:17.296
It provides a physical argument that at least it's possible.

00:08:17.856 --> 00:08:22.376
And then there are ways of testing that. Now, we can't run evolution over again,

00:08:22.496 --> 00:08:26.396
unless you deal with like bacteria, which have very, very short generation times.

00:08:26.456 --> 00:08:28.456
They don't have tongues, so it'd be hard to do it with them.

00:08:29.076 --> 00:08:34.876
But you can try to create devices like robots, and then you can,

00:08:34.896 --> 00:08:40.296
either in simulation or in the actual device, you can play with the properties.

00:08:40.416 --> 00:08:46.896
So you can change where the key point of trade-off will be, change the relative

00:08:46.896 --> 00:08:50.296
mechanical advantage, depending on how you set up the materials and the actuation.

00:08:50.556 --> 00:08:54.496
And then you can see whether the control ideas continue to be relevant.

00:08:54.696 --> 00:08:57.216
And again, when we've done those kinds of things, usually in simulation,

00:08:57.356 --> 00:09:00.996
but also sometimes in robots, that has paid off very nicely. Right.

00:09:01.236 --> 00:09:06.796
But now, what was interesting with the tongue control is you showed us that

00:09:06.796 --> 00:09:12.816
when you try to, let's say, generalize the trivial interpretation of the control

00:09:12.816 --> 00:09:15.096
of the two muscles, you would run into difficulties.

00:09:15.096 --> 00:09:19.116
So one of the things the animal has to do, of course, is if it's lapping up

00:09:19.116 --> 00:09:23.716
like what it does insides of eggs, most animals have to worry.

00:09:24.376 --> 00:09:27.376
First of all, if they're hungry, they want to eat quickly. But there's also

00:09:27.376 --> 00:09:30.116
an evolutionary issue, which is during the time that they're feeding,

00:09:30.256 --> 00:09:34.376
they may be very good prey items and become someone else's lunch.

00:09:34.556 --> 00:09:38.596
So often you want to eat on the run and then literally get out of there.

00:09:38.736 --> 00:09:41.956
Or another animal may come and grab it away.

00:09:42.156 --> 00:09:46.436
So they need speed. And the problem is, what we found was if we simply tried

00:09:46.436 --> 00:09:51.096
to do a scale version, a faster scaled version of the inputs that had worked

00:09:51.096 --> 00:09:55.076
for the tongue lapping that we started with, they did not work properly.

00:09:55.236 --> 00:10:00.636
They were filtered by both the mechanics and by the low-pass filtering of the muscle.

00:10:00.796 --> 00:10:05.796
And so we had to jigger them in order to get the same frequency behavior at a faster time scale.

00:10:06.216 --> 00:10:10.496
What's interesting is when we then went and looked at some other data where

00:10:10.496 --> 00:10:13.956
people have looked not just at lapping, but other behaviors in,

00:10:14.016 --> 00:10:18.536
for example, locomotion, you will find if you look at the EMGs,

00:10:18.576 --> 00:10:20.816
they are simply not scaled versions of one another.

00:10:20.936 --> 00:10:24.416
You can't get away with doing that. Part of that, probably a large part of that,

00:10:24.516 --> 00:10:27.436
is the low-pass filtering. But the other part is the mechanics.

00:10:27.936 --> 00:10:32.736
And so it was very nice to realize that, oh, this is not just a problem for

00:10:32.736 --> 00:10:37.016
getting this model to work. This might be a general principle for where mechanics

00:10:37.016 --> 00:10:39.996
and neural control have to be kept in mind together.

00:10:40.356 --> 00:10:44.496
Right, exactly. And that also led you to one of your, if you want,

00:10:44.576 --> 00:10:49.136
principles, maybe more behavioral one in terms of timing is everything.

00:10:49.356 --> 00:10:50.416
Timing is everything, absolutely.

00:10:50.816 --> 00:10:56.076
And so there was a very interesting paper, I was actually asked to comment on

00:10:56.076 --> 00:11:01.996
it, that came out in Science about, I think, six months to nine months ago, about cat lappings.

00:11:02.659 --> 00:11:04.639
I don't know if you saw it, but it was actually on the cover of Science.

00:11:05.239 --> 00:11:09.499
And I guess because of the paper I published, they were interested in my comments

00:11:09.499 --> 00:11:10.599
on it. It was a very nice paper.

00:11:10.959 --> 00:11:15.799
And it turns out most people think about lapping as you immerse the tongue into

00:11:15.799 --> 00:11:19.399
the fluid, and then you essentially use it as a cup and withdraw.

00:11:19.679 --> 00:11:23.759
And that's probably how tupinambis does it. It coats its tongue with the fluid

00:11:23.759 --> 00:11:25.599
and then withdraws it and then scrapes it off.

00:11:26.519 --> 00:11:29.899
Cats don't do that. It turns out they do the very fast lapping,

00:11:29.939 --> 00:11:33.339
but the fluid doesn't go on top of their tongue. It's on the bottom.

00:11:33.499 --> 00:11:40.379
What they do is they rapidly create a column of fluid, which they then withdraw in.

00:11:40.819 --> 00:11:44.299
And so what happened was these were guys at MIT. It was fascinating.

00:11:44.419 --> 00:11:46.839
The guy in mechanical engineering, I guess, had his pet cat,

00:11:46.999 --> 00:11:49.319
noticed something interesting, took movies.

00:11:49.579 --> 00:11:53.479
They created this very interesting robot that created this fluid column.

00:11:53.479 --> 00:11:58.299
And they then did analysis on how fast you had to move and what the areas would

00:11:58.299 --> 00:12:00.919
have to be so that you could maintain the fluid column.

00:12:01.339 --> 00:12:05.259
And then they went and looked at the speed of lapping in a whole variety of

00:12:05.259 --> 00:12:10.079
different basically cat animals, animals from the cat species,

00:12:10.479 --> 00:12:13.739
lions and tigers and cats and things like that.

00:12:13.739 --> 00:12:17.799
And they found that if you looked at the scales, there were some very nice scaling

00:12:17.799 --> 00:12:22.519
laws that fit with their model of how you actually did lapping. Very, very beautiful.

00:12:23.159 --> 00:12:27.539
Again, I very much like the paper because here, once again, people think about

00:12:27.539 --> 00:12:29.839
the principles in terms of the mechanics and the physics.

00:12:30.259 --> 00:12:33.479
And then they realize that that's going to create constraints on what the nervous

00:12:33.479 --> 00:12:35.879
system has to do to solve the problem that way.

00:12:36.019 --> 00:12:41.939
Right. So then after that, you sort of now start to move to,

00:12:41.979 --> 00:12:46.839
if you want, a complexification of tongues, which is like peristaltic movements

00:12:46.839 --> 00:12:49.419
across multiple segments.

00:12:49.599 --> 00:12:53.119
This could be sort of multiple tongues glued together in some funky way.

00:12:54.559 --> 00:13:00.179
And from there, you moved on to the aplysia. Right. Right.

00:13:00.379 --> 00:13:05.899
So what are the key insights that you gained looking at these peristaltic movements?

00:13:06.419 --> 00:13:10.559
That you see in different kinds of animal species. So what you see is the thing

00:13:10.559 --> 00:13:14.739
that we had initially focused on, and the literature emphasizes that peristaltic

00:13:14.739 --> 00:13:16.219
involves segmental movement.

00:13:16.699 --> 00:13:21.259
And as I mentioned, Arne McNeil Alexander has this lovely book on animal locomotion.

00:13:21.259 --> 00:13:23.419
That's the name, and I highly recommend it.

00:13:24.319 --> 00:13:27.279
I'm not related to him in any way, but I highly recommend it.

00:13:27.339 --> 00:13:28.879
It's a great book. It's very accessible.

00:13:29.139 --> 00:13:34.199
And he has a whole beautiful chapter on some of the issues in the mechanics

00:13:34.199 --> 00:13:36.899
of locomotion, and he talks about peristaltic locomotion.

00:13:37.399 --> 00:13:40.779
And one of the things he emphasizes is that you have to get the mass moving

00:13:40.779 --> 00:13:43.259
and then you have to slow it down and decelerate it.

00:13:43.953 --> 00:13:46.693
And this is really the classic way that people think about it.

00:13:46.733 --> 00:13:49.433
And if you have segmental locomotion, that's how it works.

00:13:50.733 --> 00:13:57.113
But if you actually spend time staring at earthworms, so after the rain in Cleveland,

00:13:57.233 --> 00:13:58.553
and we get lots and lots of rain,

00:13:58.753 --> 00:14:04.973
the earthworms, which will otherwise drown because they need the oxygen,

00:14:05.153 --> 00:14:08.273
will come out and walk across the sidewalks.

00:14:08.273 --> 00:14:13.153
So you get an opportunity, if you care, to look at worms on a fairly regular basis.

00:14:13.413 --> 00:14:17.733
Now, I think most people either squish them or ignore them, but if you saw me,

00:14:17.833 --> 00:14:21.593
I would be bending down and actually staring at the worm for an extended period of time.

00:14:21.793 --> 00:14:24.633
And what you see is there are different regions in the body.

00:14:24.713 --> 00:14:28.833
It's not a simple movement, actually, but the animal does have this wave of

00:14:28.833 --> 00:14:31.733
contraction that goes from one end to the other. So on the one hand,

00:14:31.733 --> 00:14:32.593
there are clearly segments.

00:14:32.673 --> 00:14:37.193
You can see them, but you see this wave that's moving, and the wave is very smooth.

00:14:37.193 --> 00:14:40.073
It's not confined to one segment at

00:14:40.073 --> 00:14:42.853
a time so this led us to think about what would

00:14:42.853 --> 00:14:45.873
happen if we did this analysis at the

00:14:45.873 --> 00:14:48.533
differential level what if you took segments and

00:14:48.533 --> 00:14:53.173
took them down to the level of very very tiny elements and that analysis which

00:14:53.173 --> 00:14:57.713
again i hope will be coming out within the next few months ordinarily when people

00:14:57.713 --> 00:15:01.473
talk about peristaltic motion they think of it as rather energetically inefficient

00:15:01.473 --> 00:15:07.853
and very slow and this analysis showed that actually that's neither of those are necessarily true.

00:15:09.331 --> 00:15:15.211
If you set it up properly, you can actually keep the center of mass at a constant velocity.

00:15:15.571 --> 00:15:18.511
And of course, there are going to be frictional forces within the body.

00:15:18.611 --> 00:15:21.871
You don't have to depend on external friction to keep moving,

00:15:22.051 --> 00:15:27.431
which means that the energies are not due to losses due to frictional contact.

00:15:28.411 --> 00:15:31.851
And that means that you could probably sustain these movements if you could

00:15:31.851 --> 00:15:35.431
minimize internal friction for quite a long period of time.

00:15:35.591 --> 00:15:39.551
That was very interesting. And the robot I showed that is built on that principle

00:15:39.551 --> 00:15:42.331
moves quite, quite fast, as you saw in the video.

00:15:42.451 --> 00:15:46.131
Those earlier videos, at least one of the underwater one, we had to speed up

00:15:46.131 --> 00:15:47.911
because it was quite slow and it was boring to watch.

00:15:48.111 --> 00:15:51.031
Whereas this one, you had to actually walk fast to keep up with it.

00:15:51.251 --> 00:15:57.891
But this system, what we want to know, of course, what are then these biomechanical

00:15:57.891 --> 00:16:02.851
properties that actually make this an effective locomotion device?

00:16:03.331 --> 00:16:05.711
That's right. So what are the

00:16:05.711 --> 00:16:11.151
key biomechanics that make this task actually doable if you are a worm?

00:16:11.351 --> 00:16:14.331
So this would require, and this is not an area I have pursued,

00:16:14.451 --> 00:16:16.111
but I'll say a few things from what I've read.

00:16:16.251 --> 00:16:19.551
And if someone was going to be inspired by this to follow up,

00:16:19.651 --> 00:16:25.331
what I'd recommend doing is spending some time studying the details of the cross

00:16:25.331 --> 00:16:30.571
bridges and the actual mechanical arrangements within the musculature of the body wall of the worm.

00:16:30.571 --> 00:16:33.791
Because what people have shown in a variety of different settings,

00:16:33.951 --> 00:16:39.811
so I have a colleague, Kishin Nishikawa, who works on ballistic movements in frog tongues.

00:16:40.071 --> 00:16:43.071
They can actually, there are three different classes. In one class,

00:16:43.151 --> 00:16:46.291
the animal essentially throws the tongue out. That's a ballistic movement.

00:16:46.571 --> 00:16:49.671
And there's another class, for example, where they use a muscular hydrostatic

00:16:49.671 --> 00:16:51.831
property and they slowly protrude it.

00:16:52.211 --> 00:16:56.131
So here you have a tongue. You have species that are similar,

00:16:56.231 --> 00:17:00.151
but one is feeding on termites that are relatively slow and that are in crevices,

00:17:00.191 --> 00:17:04.271
and they use a muscular hydrostatic property, and one is going after insects

00:17:04.271 --> 00:17:07.171
that can fly away fast, and they use a ballistic inertial tongue.

00:17:07.471 --> 00:17:11.091
She's looked very carefully at the cross bridges and come up with some very

00:17:11.091 --> 00:17:16.691
interesting models for how there may be very important nonlinear ways in which

00:17:16.691 --> 00:17:19.911
you can extend muscles much further and also contract them much further,

00:17:20.011 --> 00:17:21.871
depending on how the cross bridges work.

00:17:22.091 --> 00:17:26.011
So I would spend time looking at the material properties and understanding that.

00:17:26.011 --> 00:17:30.011
And then if that was, as I would expect, was true,

00:17:30.111 --> 00:17:33.031
I would then talk to people who are doing biomimetics and see if I could build

00:17:33.031 --> 00:17:38.291
devices like that, perhaps initially macroscopic, but ultimately with advances

00:17:38.291 --> 00:17:42.031
in nanomaterials, it would be very exciting to try to get someone interested in that.

00:17:42.211 --> 00:17:46.631
And that would be the way I think you could, again, not prove that the biology

00:17:46.631 --> 00:17:50.751
works that way, but strongly suggest that that's the key physical property.

00:17:50.751 --> 00:17:56.771
But for instance, in the system you showed, which is this larger device that

00:17:56.771 --> 00:17:59.711
uses parasitic movements to move, I don't know exactly the dimension,

00:17:59.791 --> 00:18:03.191
it looked about like a meter long or something like that. So pretty, pretty big.

00:18:04.101 --> 00:18:08.921
Also there, you must have identified biomechanical contributions to its ability

00:18:08.921 --> 00:18:10.541
to actually generate these waves.

00:18:10.541 --> 00:18:14.341
The student who did this work, Alex Boxerbaum,

00:18:14.621 --> 00:18:18.301
spent some time mainly focused on the questions of how would you get,

00:18:18.461 --> 00:18:25.701
what material could he put together that would give him the kinds of changes

00:18:25.701 --> 00:18:31.241
in strain that were crucial for the movements that he was looking at.

00:18:31.241 --> 00:18:34.221
And he came up with the idea of essentially and

00:18:34.221 --> 00:18:38.661
again partially inspired by what you see in the physical arrangements in muscles

00:18:38.661 --> 00:18:44.861
in animals but this idea of actually weaving the um the cable in a helical fashion

00:18:44.861 --> 00:18:49.341
one in one direction and then another in another direction and then basically

00:18:49.341 --> 00:18:53.381
pinning the cables together so that they formed a whole series of rhomboids.

00:18:54.141 --> 00:18:59.421
And now by using these cables within them again this is his idea you could shorten

00:18:59.421 --> 00:19:02.961
or lengthen different regions of the cable and because it was now pinned,

00:19:03.201 --> 00:19:07.781
instead of just shortening, it would pull together and force the others apart,

00:19:08.061 --> 00:19:12.701
causing that expansion in that particular region. Very ingenious on his part.

00:19:12.941 --> 00:19:17.661
Right. Very ingenious. But the step to understanding how this relates to the

00:19:17.661 --> 00:19:21.881
neural control of such a biomechanical system. Yeah, so that's what we're working on right now.

00:19:22.001 --> 00:19:25.221
And actually one of the things we're doing, I talked a little bit about these

00:19:25.221 --> 00:19:27.861
stable hetero channels, which we'll come to next.

00:19:28.181 --> 00:19:30.941
And we actually just got funding from the National Science Foundation,

00:19:31.261 --> 00:19:32.801
myself and my colleague, Roger Quinn,

00:19:32.981 --> 00:19:39.581
to look at this, actually trying to put together more of the degrees of freedom

00:19:39.581 --> 00:19:44.961
of this device with this stable heteroclinic channel network idea to see if

00:19:44.961 --> 00:19:48.081
we can independently control different degrees of freedom,

00:19:48.201 --> 00:19:51.341
actuate and independently control them. I don't know if it's going to work.

00:19:52.061 --> 00:19:57.221
I have a history of saying, I think this is true, and then it does actually show itself true.

00:19:57.341 --> 00:20:00.741
And I have a history of encountering lots of skepticism when I make these claims.

00:20:01.061 --> 00:20:05.481
But I'm also, more often than not, I'm also wrong.

00:20:05.681 --> 00:20:10.101
But we can actuate individual parts of the mesh.

00:20:10.721 --> 00:20:15.181
And we can thus separately control and activate them. And I think we can get

00:20:15.181 --> 00:20:20.521
the thing to bend and lift up and actually form curves and possibly also negotiate

00:20:20.521 --> 00:20:21.761
fairly tortuous terrain.

00:20:21.981 --> 00:20:25.481
It may take us several years, but I think that's all within our grasp.

00:20:25.641 --> 00:20:31.641
Okay. But then do you see the controller that's behind that or the neural control

00:20:31.641 --> 00:20:38.821
as still reducing this complexity or roughly mapping one-to-one on the complexity

00:20:38.821 --> 00:20:41.561
of this morphology? Yeah, yeah, yeah. So I think the issue is this.

00:20:41.701 --> 00:20:47.281
If I'm trying to create peristaltic waves, unified waves, I actually,

00:20:47.441 --> 00:20:51.601
and again, in the robot we showed, that's a one degree of freedom device in terms of control.

00:20:52.407 --> 00:20:57.887
Because the cam basically allows you to pull on the different cables and change

00:20:57.887 --> 00:21:02.527
their length as it circles, and that's all you need to get the peristaltic waves.

00:21:02.727 --> 00:21:06.207
So what that suggests is that under certain circumstances, if you're creating

00:21:06.207 --> 00:21:12.327
waves, a collective contraction of specific elements in sequence will be all

00:21:12.327 --> 00:21:14.507
you need, and the whole thing will do what you want.

00:21:14.507 --> 00:21:19.687
Now, if you want precise, careful movements where you're exploring your way

00:21:19.687 --> 00:21:23.507
through obstacles and then you're curling around and perhaps pulling something,

00:21:24.507 --> 00:21:29.167
you're not going to be able to create this nice unified, let's actuate everything

00:21:29.167 --> 00:21:31.627
in a simple way. You're going to have to do much more complex things.

00:21:32.147 --> 00:21:36.207
I think the way I think about it is this, that when you don't need all the excess

00:21:36.207 --> 00:21:40.367
degrees of freedom, what much of the neural control does is to try to simplify

00:21:40.367 --> 00:21:43.827
down those degrees of freedom by appropriate collective activation.

00:21:44.547 --> 00:21:48.247
And when you do need them, they're there. And again, I think the nervous system

00:21:48.247 --> 00:21:52.547
may use some of the principles I was talking about to fractionate them out and

00:21:52.547 --> 00:21:53.827
pull out the things that it needs.

00:21:54.367 --> 00:21:59.227
So the answer, once again, is it will depend on the context in which the thing is being used.

00:21:59.347 --> 00:22:02.547
And what we're You're going to move towards a situation where,

00:22:02.707 --> 00:22:05.707
and again, this is very nice in terms of the controllers I was talking about,

00:22:05.847 --> 00:22:09.567
you can gang them together so that they do one thing all at once,

00:22:09.687 --> 00:22:15.007
but you can fractionate them if necessary so they can map onto much finer detail. Right.

00:22:16.127 --> 00:22:21.707
So it's something you were structuring also in this presentation from,

00:22:21.847 --> 00:22:27.427
let's say, fairly simple sensor motor systems or cellular motor systems to more complex ones.

00:22:27.427 --> 00:22:32.147
And so after the peristaltic movement, and then this nice proof of concept,

00:22:32.227 --> 00:22:34.827
if you want, which is a one meter long worm,

00:22:35.627 --> 00:22:40.887
you then brought it to the point that you say, look, but the use of these muscles,

00:22:41.007 --> 00:22:44.247
we should think about in terms of context. And this is an important principle.

00:22:44.587 --> 00:22:49.027
And to illustrate that principle, you used also a plesia grasping. Yes. Right?

00:22:49.207 --> 00:22:54.607
So what do you really mean when you say, okay, muscles are used in context?

00:22:54.927 --> 00:22:58.827
Okay. So the example I gave in the talk, let me talk about it a little bit.

00:22:59.267 --> 00:23:04.947
So it turns out that in our hip, there are certain muscles that act for either,

00:23:05.027 --> 00:23:10.147
I think it's moving forward or moving back, depending on where the rest of the hip is.

00:23:10.607 --> 00:23:18.847
In the arm, the brachioradialis is used for, if your palm is up,

00:23:18.907 --> 00:23:22.987
this crosses the joint and it's crucial for turning your palm down.

00:23:22.987 --> 00:23:28.487
Now, so if you think about the direction of forces, it's going to rotate in one direction.

00:23:28.907 --> 00:23:33.647
Now, once you have the arm down, the hand down, that same muscle plays a role

00:23:33.647 --> 00:23:36.027
in rotating the arm back in the opposite direction.

00:23:36.427 --> 00:23:39.767
So if you look in terms of the torque arm, it actually switches direction,

00:23:39.887 --> 00:23:45.427
and it can do so in part because of the new position it finds itself in.

00:23:46.052 --> 00:23:52.772
And several people, and this was some years ago, using cadaver studies and pulling

00:23:52.772 --> 00:23:59.112
on different muscles in the context of limbs, showed that as you move the limb through the workspace,

00:23:59.592 --> 00:24:04.652
what the muscles did was not a simple relationship and was very complex. flex.

00:24:04.892 --> 00:24:08.892
And this is actually, if you go back and look at some work that people had done

00:24:08.892 --> 00:24:13.132
back even in the 19th and very early 20th century, people did very careful work

00:24:13.132 --> 00:24:14.952
on, say, frog musculature.

00:24:15.232 --> 00:24:20.272
They showed that as you move through the workspace, what the muscles did changed.

00:24:20.732 --> 00:24:24.152
So this idea of a protractor, retractor, pronation, supination,

00:24:24.372 --> 00:24:26.412
these are all fine. You teach medical students that.

00:24:26.512 --> 00:24:30.412
It's very helpful. And in standard configurations, that's how it works.

00:24:30.412 --> 00:24:34.532
We found in the aplysius system, and I'm emphasizing the vertebrate examples

00:24:34.532 --> 00:24:36.972
to begin with to show that I think this is much more general,

00:24:37.132 --> 00:24:41.432
that as one part of the structure moved relative to other parts of the structure,

00:24:41.572 --> 00:24:44.412
the actual functional role of the muscle changed.

00:24:44.992 --> 00:24:49.752
So a muscle that was thought to push a grasper back could actually move it towards

00:24:49.752 --> 00:24:51.712
the jaws in the appropriate context.

00:24:52.412 --> 00:24:55.652
And the two points then that I made, which I thought were very crucial,

00:24:55.792 --> 00:24:59.032
was that what a muscle does is a function of its mechanical context.

00:24:59.032 --> 00:25:02.712
And the second one, which I didn't illustrate but could go on at great length

00:25:02.712 --> 00:25:07.852
because we've shown this, because muscles do that, when you are trying to do

00:25:07.852 --> 00:25:09.732
different behaviors or behavioral variants,

00:25:10.112 --> 00:25:15.192
what you do is you call upon a coalition of the relevant muscles in that particular

00:25:15.192 --> 00:25:17.712
mechanical context to make that happen.

00:25:17.712 --> 00:25:23.172
And some muscles can't be used in a particular context, and other muscles must

00:25:23.172 --> 00:25:26.592
be used, and other muscles can variably be used.

00:25:26.812 --> 00:25:30.772
So the general principle I stated, which I think is a broad truth,

00:25:31.092 --> 00:25:36.132
is that it's actual coalitions, changing coalitions of muscles are what give

00:25:36.132 --> 00:25:37.632
rise to multifunctionality.

00:25:37.832 --> 00:25:39.672
And I think that that's a general principle.

00:25:40.452 --> 00:25:46.572
Right, but it's interesting, right? Because now we went from this worm to the

00:25:46.572 --> 00:25:50.012
idea to test it on this grasping response of a plesia.

00:25:50.152 --> 00:25:54.052
Right. And in some sense, we get the complexification of the whole system. That's correct.

00:25:54.192 --> 00:25:57.852
Because the strength of your, let's say, your C.

00:25:57.872 --> 00:26:01.612
Elegans model was like, well, it's one degree of freedom that they can control

00:26:01.612 --> 00:26:03.732
with very minimal control. That's right. Same for your tongue.

00:26:03.952 --> 00:26:07.192
Right. But remember, let's talk about the a plesia grasper.

00:26:07.864 --> 00:26:11.164
If I showed it to you in all its glory, and if you come to Cleveland,

00:26:11.264 --> 00:26:14.104
I'll be happy to do so. Or if you invite me and bring me slugs,

00:26:14.124 --> 00:26:15.624
I'll dissect them out and show them.

00:26:16.684 --> 00:26:19.964
Very complicated. And what you will see is it's a soft tissue structure.

00:26:20.204 --> 00:26:24.744
And if I poke it or excite it, you'll see all this complex shimmying around

00:26:24.744 --> 00:26:26.164
that it's capable of doing.

00:26:26.684 --> 00:26:30.024
So even though there are constraints because of its physical nature,

00:26:30.384 --> 00:26:35.484
and it's not jello, but it is very flexible.

00:26:35.604 --> 00:26:38.484
And there are all sorts of different things that the musculature can do.

00:26:39.444 --> 00:26:43.884
In the analysis that I'm doing, we're really simplifying down to talk about

00:26:43.884 --> 00:26:48.924
essentially two degrees of freedom, which is opening and closing and protraction and retraction.

00:26:49.464 --> 00:26:53.504
So again, I didn't give a lecture on aplysia feeding. If you asked me to do

00:26:53.504 --> 00:26:54.824
that, I would have happily done so.

00:26:54.964 --> 00:26:58.204
So we've taken this whole complicated structure that has all sorts of different

00:26:58.204 --> 00:27:01.824
potential degrees of freedoms and maybe 15 or 20 muscles.

00:27:02.244 --> 00:27:05.864
And we actually think about it and we can show that in terms of the behavior,

00:27:06.324 --> 00:27:09.744
if you just focus on opening and closing and protraction and retraction,

00:27:10.044 --> 00:27:13.244
you can create all of the different functions I talk about, biting,

00:27:13.344 --> 00:27:14.004
swallowing, and rejection.

00:27:14.644 --> 00:27:19.984
By changing the timing, the duration, and the phasing of those key two components,

00:27:20.144 --> 00:27:21.384
you can get all the different behaviors.

00:27:21.824 --> 00:27:26.664
So in fact, we see that as an enormous simplification. In addition,

00:27:26.804 --> 00:27:29.124
and again, I didn't have time to talk about this,

00:27:30.028 --> 00:27:35.368
Because the musculature is fairly slow, we can actually have done some analysis

00:27:35.368 --> 00:27:36.708
of the mechanics in detail.

00:27:37.588 --> 00:27:41.048
We can define what we call neuromechanical equilibrium points.

00:27:41.368 --> 00:27:47.168
So for a given pattern of activation in the neural outputs, we can actually

00:27:47.168 --> 00:27:51.168
predict the trajectory that the musculature will take over the next several

00:27:51.168 --> 00:27:53.848
hundreds of milliseconds.

00:27:55.328 --> 00:27:58.968
And we can talk about that as the target the animal is aiming towards.

00:27:58.968 --> 00:28:00.548
You're dealing with very low

00:28:00.548 --> 00:28:05.088
mass, a fairly viscous system, so inertial forces are not in the picture.

00:28:05.288 --> 00:28:09.848
So it's elastic and viscous forces.

00:28:10.048 --> 00:28:14.528
So velocity-dependent forces, as well as position-dependent forces.

00:28:14.828 --> 00:28:20.648
So not so much mass-dependent forces. So what you're dealing with is a system

00:28:20.648 --> 00:28:24.328
where you can actually see smooth trajectories and you can talk about these

00:28:24.328 --> 00:28:28.908
equilibrium points defined by the neural control as well as by the mechanics.

00:28:29.268 --> 00:28:33.588
And that it provides also a huge simplification for thinking about where is

00:28:33.588 --> 00:28:39.208
the nervous system trying to push the whole, the properties of the system from one set to another.

00:28:39.468 --> 00:28:44.388
So even though it's a much higher dimensional system, this idea of understanding

00:28:44.388 --> 00:28:49.728
the mechanics properly and analyzing it properly and taking the muscular hydrostatic

00:28:49.728 --> 00:28:53.048
properties and using that to see where is the system going,

00:28:53.168 --> 00:28:55.348
how to analyze it, has led to simplification.

00:28:55.588 --> 00:29:00.508
But it is a more complicated system, so it doesn't simplify down quite as much

00:29:00.508 --> 00:29:02.548
as the original one I started with.

00:29:02.748 --> 00:29:07.088
Right. But the question that this raises is that, okay, but now what are the

00:29:07.088 --> 00:29:09.448
implications of neural control, right?

00:29:09.488 --> 00:29:13.708
And how does the biomechanics actually constrain that and make it solvable?

00:29:13.928 --> 00:29:19.008
Because in some sense, isn't this sort of a restatement of the Bernstein problem in some sense?

00:29:19.128 --> 00:29:22.128
Because you say, look, I have many different ways to perform a movement,

00:29:22.268 --> 00:29:25.168
so in that sense- Yeah, exactly. Well, he pointed out two different things.

00:29:25.168 --> 00:29:27.868
First of all, he talked about the invariance of the motor field.

00:29:28.660 --> 00:29:33.220
And I think this is still something we do not understand. I can learn to sign

00:29:33.220 --> 00:29:39.400
my name under a microscope or with a paintbrush against the side of a barn.

00:29:39.600 --> 00:29:43.100
And I'm going to be, in one case, using a very, very tiny set of muscles.

00:29:43.200 --> 00:29:45.200
In the other case, I'm going to be running with my entire body,

00:29:45.260 --> 00:29:49.820
and the signature will look the same. So he identified early on this notion

00:29:49.820 --> 00:29:55.120
of motor invariance that I think is still a very deep question that has not

00:29:55.120 --> 00:29:56.520
really been properly addressed.

00:29:56.800 --> 00:30:00.300
Part of the reason I'm so excited about the neuromechanical equilibrium points

00:30:00.300 --> 00:30:04.040
in the slow system, I know about equilibrium points, the fast system,

00:30:04.060 --> 00:30:05.400
and all the debates in that area.

00:30:05.540 --> 00:30:09.080
I am agnostic about that, and I've seen the debates pro and con.

00:30:09.340 --> 00:30:14.260
But in my system, which is much slower, where I think we can much more easily

00:30:14.260 --> 00:30:15.860
make the argument and then show

00:30:15.860 --> 00:30:20.700
data for it, I think this is an incredibly powerful organizing principle.

00:30:21.860 --> 00:30:26.060
The point that Bernstein was making, though, was he doesn't ever use the term

00:30:26.060 --> 00:30:27.320
cursive dimensionality.

00:30:27.480 --> 00:30:31.980
And he doesn't ever, in his book, at least as I've read it, say excess degrees

00:30:31.980 --> 00:30:33.060
of freedom are a problem.

00:30:33.420 --> 00:30:38.300
What he talks about is how they are sculpted into these biodynamic waves that

00:30:38.300 --> 00:30:43.500
give you this patterning And that when individuals go through the lifespan.

00:30:44.020 --> 00:30:50.140
You see these pattern movements that change over age, or if someone has a wooden

00:30:50.140 --> 00:30:54.620
leg that they have to use as a prosthetic, how they incorporate it into the biodynamic waves.

00:30:54.780 --> 00:31:01.000
He had this vision of an internal dynamics that was mapped onto those degrees

00:31:01.000 --> 00:31:06.740
of freedom and that provided some kind of unification and simplification of their control. crawl.

00:31:06.820 --> 00:31:10.020
So I've been very influenced by that, I think. And so if I'm restating something

00:31:10.020 --> 00:31:12.080
that Bernstein said, you've just complimented it.

00:31:12.540 --> 00:31:16.660
No, no, look, this is what it sounds like a little bit. But what I'm looking

00:31:16.660 --> 00:31:21.500
for is then, okay, we have all these possible combinations or different,

00:31:21.540 --> 00:31:24.600
if you want, now skeletal motor contexts in which a muscle can act.

00:31:24.960 --> 00:31:28.840
On the one end, you have your biomechanical constraints that will limit this in some sense.

00:31:28.960 --> 00:31:33.840
But now the question is, how does your neural control now tune itself to then

00:31:33.840 --> 00:31:35.840
these, let's say, valid combinations, right?

00:31:35.880 --> 00:31:40.500
This valid set of muscle combinations. So you're asking a very interesting question.

00:31:40.680 --> 00:31:44.640
So this actually would jump me to the end of the talk and the question of initial conditions.

00:31:45.060 --> 00:31:48.460
So let me spell this out. So I have a colleague at Case.

00:31:48.540 --> 00:31:51.700
Her name is Lynn Landmesser, and she's been studying the early development of

00:31:51.700 --> 00:31:54.560
the nervous system, especially the spinal cord, for many, many years.

00:31:54.960 --> 00:31:57.740
One of the things that she has discovered is that well before birth,

00:31:57.940 --> 00:32:03.320
there are spontaneous simultaneous patterns that start in the nervous system, in the spinal cord.

00:32:03.460 --> 00:32:11.180
And so prior to having wired up, this is before motor neurons make contact with the periphery.

00:32:12.720 --> 00:32:14.900
Regular dynamic patterns are established.

00:32:16.246 --> 00:32:22.266
Once contact is made, there are trophic factors that allow the size of the musculature

00:32:22.266 --> 00:32:27.166
to map back onto the nervous system and increase, for example, proliferation.

00:32:27.386 --> 00:32:30.426
So a larger periphery will generate more motor neurons.

00:32:31.106 --> 00:32:38.106
But also, not just in terms of numbers, in terms of the dynamics, once there is contact.

00:32:38.786 --> 00:32:44.866
Contact. Another investigator who's studied chicks before hatching,

00:32:44.866 --> 00:32:48.546
Anne Becky, has shown that there are spontaneous movements that,

00:32:48.646 --> 00:32:52.726
and so again, any mother will tell you about the fetal kicking and stuff like that.

00:32:52.826 --> 00:32:56.746
There is an ongoing coupling that's happening prior to birth,

00:32:56.926 --> 00:33:00.306
whereby the periphery is providing feedback, sensory inputs,

00:33:00.566 --> 00:33:05.946
and the nervous system is generating these spontaneous dynamic patterns,

00:33:06.166 --> 00:33:08.086
and they're shaping each other.

00:33:09.086 --> 00:33:12.266
And that's what you have to start with when you're born.

00:33:12.546 --> 00:33:15.986
So you're not starting with a tabula rasa. And an infant doesn't lie,

00:33:16.066 --> 00:33:18.926
they're passive and immobile. It's already moving things around.

00:33:19.226 --> 00:33:22.806
And then, I mean, this is like a major, major breakthrough for my kids,

00:33:22.886 --> 00:33:27.546
when you can finally get the thumb into the mouth reliably and keep it there. That's like amazing.

00:33:28.646 --> 00:33:34.546
So there are some built-ins like suckling, and then there's these more complex

00:33:34.546 --> 00:33:41.086
things that are built in and there's this dynamic continuous reshaping between the periphery,

00:33:41.086 --> 00:33:45.386
its experiences in the world and the nervous system. That's what's going on there.

00:33:45.526 --> 00:33:49.346
So what happens is this kind of jittering around, this play,

00:33:49.566 --> 00:33:54.586
this activation, this is the dynamic way you tune that system so the nervous

00:33:54.586 --> 00:33:58.586
system knows about the degrees of freedom that are out there and it takes advantage

00:33:58.586 --> 00:34:00.086
of as many of them as it needs to.

00:34:00.086 --> 00:34:03.746
The other thing you raised in part of your question to my talk,

00:34:03.786 --> 00:34:07.586
which I just have to talk about again, because I thought it was a very important

00:34:07.586 --> 00:34:10.166
issue, especially in higher organisms like ourselves,

00:34:10.526 --> 00:34:16.846
useful repetitive patterns are speeded up and they become part of our repertoire.

00:34:17.026 --> 00:34:21.126
They are like reflexes. And that's not something you see in all animals.

00:34:21.166 --> 00:34:24.586
Many, as I said, in the lower organisms, you have these fixed action patterns

00:34:24.586 --> 00:34:27.786
and those are pre-wired and they can't be changed.

00:34:28.326 --> 00:34:31.646
I didn't talk about this in detail, but a wasp that's going through nest building,

00:34:31.826 --> 00:34:35.246
if you interrupt it and then let it continue, it's just like a machine.

00:34:35.326 --> 00:34:37.706
It just goes back and continues doing exactly what it was doing.

00:34:37.906 --> 00:34:41.706
You would not see that in a primate. You would certainly not see that in a human.

00:34:42.848 --> 00:34:46.388
And yet, if there's something like our tennis swing or piano playing,

00:34:46.528 --> 00:34:48.248
that becomes an automatic behavior.

00:34:48.768 --> 00:34:53.908
So again, very important. So this tuning, shaping, and this dynamic view that

00:34:53.908 --> 00:34:57.628
I'm trying to argue is how the nervous system does it.

00:34:57.668 --> 00:35:00.828
I don't see that as necessarily working through some kind of internal representation.

00:35:00.828 --> 00:35:06.488
I see this as this trial and error playing around, very good initial dynamics,

00:35:06.728 --> 00:35:11.228
feedback from the environment shaping the connectivity, the local connectivity,

00:35:11.528 --> 00:35:16.108
and generating an effective dynamic model.

00:35:16.588 --> 00:35:20.188
That's what's exploited subsequently. So as if the nervous system is freezing

00:35:20.188 --> 00:35:23.308
its degrees of freedom to control the periphery. When it needs to,

00:35:23.428 --> 00:35:26.888
and then unfreezing them when it doesn't. And that's, I think,

00:35:26.908 --> 00:35:28.388
a really critical insight.

00:35:28.508 --> 00:35:32.448
So that sometimes things look absolutely seamless and incredibly easy,

00:35:32.548 --> 00:35:36.728
but anyone, you say you do sports, so you can remember the difference.

00:35:36.728 --> 00:35:40.268
I mean, this is something that was very striking to me. Watching someone do the crawl.

00:35:40.808 --> 00:35:44.568
When you watch someone who's learning how to do the crawl, you see all the effort

00:35:44.568 --> 00:35:49.868
and you can see all these inefficiencies and the person is doing the movements.

00:35:50.268 --> 00:35:54.068
Then you look at someone who's a trained swimmer and it's this seamless,

00:35:54.148 --> 00:35:56.888
beautiful motion. I mean, it's just gorgeous.

00:35:57.328 --> 00:36:02.028
And there's one motion after another follows in this absolutely seamless way.

00:36:02.168 --> 00:36:05.528
There's a real beauty to it. That coordination,

00:36:05.748 --> 00:36:12.428
that choreography is something that reduces the unnecessary degrees of freedom

00:36:12.428 --> 00:36:15.568
and allows the person to do it incredibly effectively and efficiently.

00:36:15.828 --> 00:36:17.548
Again, I think it's a tuning of dynamics.

00:36:18.408 --> 00:36:24.928
But then we should try to understand a little bit what these key organizing principles are.

00:36:25.068 --> 00:36:30.788
Exactly. And for that, you start to look, this is of control,

00:36:30.908 --> 00:36:35.348
you tied very much to also ideas about, let's say, attractor dynamics and how

00:36:35.348 --> 00:36:36.588
attractor dynamics are regulated.

00:36:36.828 --> 00:36:40.248
So how is that giving us insight in organizational principles?

00:36:40.368 --> 00:36:44.748
Okay, so again, the notion of attractor dynamics and why I'm attracted to attractor

00:36:44.748 --> 00:36:46.208
dynamics is the following.

00:36:46.208 --> 00:36:51.308
First of all, although these are qualitative dynamics in many cases,

00:36:51.548 --> 00:36:55.268
there are both, in lower dimensions, there are very, very rigorous mathematical

00:36:55.268 --> 00:36:59.428
things you can do, and in higher dimensions, very interesting numerical things you can do.

00:36:59.926 --> 00:37:02.546
And what's, again, interesting is that you don't need to have representations

00:37:02.546 --> 00:37:08.666
internally to get very complex behavior, and you get smoothness and robustness

00:37:08.666 --> 00:37:12.506
and the ability to handle noise and perturbations essentially for free.

00:37:12.666 --> 00:37:18.446
You don't have to put in all of that. You don't have to plan for the combinational

00:37:18.446 --> 00:37:22.226
complexity explosion that's going to happen if you have to deal with all the

00:37:22.226 --> 00:37:23.346
possible different cases.

00:37:23.526 --> 00:37:29.006
That comes for free. So attractor dynamics is inherently robust to perturbation.

00:37:29.046 --> 00:37:30.666
That's why we call it an attractor.

00:37:31.366 --> 00:37:37.806
And it has a lot of architecture and it has a lot of tools that allow you to

00:37:37.806 --> 00:37:42.046
set things up and to try to map them onto things like nervous systems.

00:37:42.186 --> 00:37:46.626
So I didn't have time to talk about this in my talk, but Rabinovich has shown,

00:37:46.666 --> 00:37:49.946
and we're starting to do some of this, that some of the ideas that I talked

00:37:49.946 --> 00:37:55.606
about from attractor dynamics can be mapped very naturally onto known neural architectures.

00:37:55.726 --> 00:37:59.886
And what that means then is that doesn't guarantee that you'll know that neuron

00:37:59.886 --> 00:38:02.546
A makes this connection to neuron B.

00:38:02.906 --> 00:38:08.066
But it will say that if this hypothesis is correct, this class of neurons should

00:38:08.066 --> 00:38:11.966
be tightly coupled to one another, and they should be activated together.

00:38:12.546 --> 00:38:17.366
They should inhibit these other groups in such a way that they can't take over

00:38:17.366 --> 00:38:18.726
while this one is activated.

00:38:19.246 --> 00:38:23.066
And then there There should be appropriate, you'll excuse the term,

00:38:23.086 --> 00:38:28.926
recurrent loops such that you can destabilize that pattern of activity and allow

00:38:28.926 --> 00:38:32.526
another one now to generate a burst, be active,

00:38:32.706 --> 00:38:36.906
also have its own internal coupling and keep the other parts silent for that period of time.

00:38:37.386 --> 00:38:42.106
So what's very nice is as I see these descriptions and I look at the models,

00:38:42.186 --> 00:38:45.166
and again, the math is something I can follow so I can actually understand what

00:38:45.166 --> 00:38:46.846
they're talking about. I immediately

00:38:46.846 --> 00:38:49.286
start thinking in terms of the neural circuitry that we're studying.

00:38:49.366 --> 00:38:51.926
And I have some ideas for experiments that I would like to try,

00:38:52.086 --> 00:38:54.966
which would test whether these ideas are of any value.

00:38:55.246 --> 00:38:58.346
And again, as I stressed in my talk, when it comes to theory,

00:38:58.446 --> 00:38:59.566
I'm an absolute agnostic.

00:39:00.160 --> 00:39:03.300
I'm, I do not, I'm not a believer in the sense that, you know,

00:39:03.320 --> 00:39:04.860
this theory is right. That theory is wrong.

00:39:04.960 --> 00:39:08.820
If a piece of a theory will work, I'm great. That's great. Let me take it and

00:39:08.820 --> 00:39:10.480
I'll use it. If it doesn't work. Okay.

00:39:10.660 --> 00:39:13.360
So we'll come up with a better theory or use a different theory.

00:39:13.580 --> 00:39:15.000
But there are two things about the attractor networks.

00:39:16.580 --> 00:39:21.780
So on the one hand, it could also possibly mislead you because maybe in a tractor

00:39:21.780 --> 00:39:27.240
that, that you observe at some, at some level of the state space of this organism,

00:39:27.260 --> 00:39:29.840
because it's up to you on what level you describe it. Right, right.

00:39:30.160 --> 00:39:34.740
Right. These attractors might be some mixture of neural states and biomechanics. That's right.

00:39:34.940 --> 00:39:39.960
And in some sense, you want to tease these factors apart.

00:39:40.360 --> 00:39:42.600
Well, that's the question. But now it's starting to become, because as you said

00:39:42.600 --> 00:39:46.540
yourself, you then interpret your attractor in terms of some recurrent neural structure.

00:39:46.720 --> 00:39:49.620
Okay. So the interesting point that you're raising is this, and this is something

00:39:49.620 --> 00:39:52.940
that came up in the second talk, which we may get to or not.

00:39:53.060 --> 00:39:58.900
I just found that really inspiring. But the issue is that when we look and do

00:39:58.900 --> 00:40:04.380
folk psychology or folk philosophy or anything like that, we tend to carve nature up in certain ways.

00:40:04.720 --> 00:40:08.920
So the elements are earth, air, fire, and water, right?

00:40:09.020 --> 00:40:13.380
Because that's what we see. but then as science progresses we begin to discover

00:40:13.380 --> 00:40:16.920
that that's actually not how nature is carved up at all but we that doesn't

00:40:16.920 --> 00:40:19.540
lead us to abandon the notion that there might be elements.

00:40:20.080 --> 00:40:26.020
It just requires us to rigorously refine what we mean by that and then come

00:40:26.020 --> 00:40:29.040
up with good operational definitions for what is an element what's a molecule

00:40:29.040 --> 00:40:34.860
and what's what's not similarly it seems to me that the formulation of the neuromechanical

00:40:34.860 --> 00:40:38.140
equilibrium point hypothesis orthosis that, again, I didn't talk about in my talk,

00:40:38.240 --> 00:40:42.660
but it's something we're actively working with, may or may not carve nature

00:40:42.660 --> 00:40:44.060
at its joints. We don't know.

00:40:44.380 --> 00:40:48.060
But it's an organizing principle that allows us to look at the bursts of activity

00:40:48.060 --> 00:40:51.640
that come out of the nervous system and the musculature that's active at a particular

00:40:51.640 --> 00:40:56.640
time and say, ah, we can make a good prediction as to where this is going to go next.

00:40:56.860 --> 00:41:00.820
Now, one thing that illustrates the likelihood that this is going to work is the following.

00:41:01.900 --> 00:41:05.840
And again, this is very qualitative, and I immediately admit that.

00:41:06.020 --> 00:41:11.080
But my students set up the animals and they have them instrumented so that they're

00:41:11.080 --> 00:41:14.520
recording from that key muscle I talked about that pushes the grasper forward

00:41:14.520 --> 00:41:15.500
in these three motor neurons.

00:41:15.720 --> 00:41:19.960
I walk in and I don't look at the animal. I look at the recordings. I say, oh, it's biting.

00:41:20.440 --> 00:41:22.380
And my student says, yeah, you're right.

00:41:23.485 --> 00:41:31.045
Now, again, I'm using my visual cortex as a stand-in for some other analyzer.

00:41:31.105 --> 00:41:34.225
But what I'm suggesting is that if we can begin,

00:41:34.425 --> 00:41:39.885
as I and my students have begun, to recognize certain motor patterns as they're

00:41:39.885 --> 00:41:43.585
starting to unfold and make a prediction as to what's going to happen,

00:41:43.725 --> 00:41:48.645
that suggests that these ways of thinking about it and conceptualizing what's

00:41:48.645 --> 00:41:52.725
going on may be very powerful for understanding the organization of the system.

00:41:53.225 --> 00:41:57.925
They may be artificial impositions that we've created that have nothing to do with nature.

00:41:58.645 --> 00:42:02.965
At least being aware of that is the basis for going in and probing the system

00:42:02.965 --> 00:42:04.185
to try to show that you're wrong.

00:42:04.345 --> 00:42:07.025
And so that's one of the things that we make a lot of effort to do.

00:42:07.365 --> 00:42:13.345
So if the systems actually do have a mix of the neural and mechanical,

00:42:13.525 --> 00:42:17.825
and that is actually what's going on, then if we find in the nervous system

00:42:17.825 --> 00:42:21.065
a tractor dynamics which mix those those two appropriately.

00:42:21.565 --> 00:42:25.865
Again, it's not a proof. It's very suggestive. The key point that you're getting

00:42:25.865 --> 00:42:29.665
at, which I take very strongly and is, again, I'm hoping to press this forward,

00:42:29.785 --> 00:42:34.685
depends on resources and depends on students, would be to disrupt the system

00:42:34.685 --> 00:42:41.085
in predictive ways and show that if I take out a particular key interneuron

00:42:41.085 --> 00:42:44.845
that is supposed to instantiate what I'm arguing is one of these attractors,

00:42:44.905 --> 00:42:47.025
and I prematurely turn it off,

00:42:47.105 --> 00:42:49.865
or I turn on another, other, which ordinarily would be inhibited.

00:42:50.085 --> 00:42:53.945
I can predict what that perturbation is likely to do based on the attractor

00:42:53.945 --> 00:42:56.825
dynamics and show that the behavioral changes are similar.

00:42:56.985 --> 00:43:02.965
Now, that's still not a complete proof, but it's a much stronger sense that,

00:43:03.005 --> 00:43:08.185
yes, the way I manipulate the system and modify it is reflecting what's going on.

00:43:08.245 --> 00:43:11.285
It's not just a useful summary for the way I think about it.

00:43:11.345 --> 00:43:14.225
It might actually be how the system itself is built. Right.

00:43:14.385 --> 00:43:17.365
But this is also an important point here because in

00:43:17.365 --> 00:43:20.385
your in your modeling in the end uh

00:43:20.385 --> 00:43:25.825
of these of these tracker dynamics um what we look at in the end was some sort

00:43:25.825 --> 00:43:29.765
of recurrently coupled network where you both had sort of self-excitation and

00:43:29.765 --> 00:43:34.965
then full lateral excitation right in the network and then the obvious question

00:43:34.965 --> 00:43:39.945
that then i also pose is of course well but but why would i believe that that

00:43:40.045 --> 00:43:42.585
reflects, let's say, the nervous system of a plesia.

00:43:42.945 --> 00:43:48.185
And for that, don't we need to impose a few more constraints that relate to

00:43:48.185 --> 00:43:49.185
the sociology of the anatomy?

00:43:49.445 --> 00:43:51.685
And again, again, I didn't have time to talk about this. There's been,

00:43:51.865 --> 00:43:54.225
as I mentioned, now about...

00:43:56.488 --> 00:43:59.128
Nearly it's three and a half decades of work i

00:43:59.128 --> 00:44:01.988
mean irving started focusing on implicit feeding his first

00:44:01.988 --> 00:44:04.768
publication is 1974 but he'd started

00:44:04.768 --> 00:44:10.288
doing that a year or two beforehand so it may be nearly um four decades actually

00:44:10.288 --> 00:44:19.088
and as i said we have about 200 or so of the elements now each of the in the

00:44:19.088 --> 00:44:22.608
controller for the um feeding apparatus is what's called the buckle ganglia

00:44:22.608 --> 00:44:23.348
it's an an organization.

00:44:23.628 --> 00:44:28.748
It's a collection of nerve cells. They're two paired ganglia.

00:44:28.828 --> 00:44:31.128
And in those, there are about 1,000 nerve cells each.

00:44:31.288 --> 00:44:34.068
About 150 of them are motor neurons.

00:44:34.468 --> 00:44:37.908
Some tens of them are key interneurons, and the rest of them are sensory neurons.

00:44:38.648 --> 00:44:41.808
We know many of the motor neurons. We know some of the interneurons.

00:44:41.888 --> 00:44:44.628
We know relatively few, but at least a few of the sensory neurons.

00:44:45.088 --> 00:44:49.848
And we've mapped a lot of the connections. So when I talk about recurrent neural networks, works.

00:44:49.908 --> 00:44:53.428
We can show that this neuron actually not only inhibits this,

00:44:53.468 --> 00:44:58.248
it excites that, and that one comes back and after a delay can inhibit that one or excite that one.

00:44:58.328 --> 00:45:04.648
So we see huge numbers of loops that are based on binary recordings from the nerve cells.

00:45:05.248 --> 00:45:09.908
But still, in these loops, you would have very characteristic transduction delays

00:45:09.908 --> 00:45:14.228
that you might not capture in these models because these models in that sense are more uniform.

00:45:14.548 --> 00:45:19.388
Well, the interesting issue though is remember that for the real nervous system,

00:45:19.468 --> 00:45:23.048
it still has to deal with the low-pass filtering of the musculature, the real musculature.

00:45:23.208 --> 00:45:26.508
So what's going to happen is some of the fast phenomena,

00:45:26.768 --> 00:45:30.848
which are important for the dynamics of the nervous system unfolding in detail,

00:45:31.108 --> 00:45:36.588
are going to be lost when you push them through the periphery because the periphery

00:45:36.588 --> 00:45:40.088
is only going to be getting a low-pass filtered version of that.

00:45:40.288 --> 00:45:44.908
So if our models capture the low-pass filtering properties reasonably well,

00:45:44.908 --> 00:45:49.348
Well, even if not all of the details of the dynamics at higher,

00:45:49.568 --> 00:45:54.128
at faster time scales are captured, we may nevertheless capture some aspects

00:45:54.128 --> 00:45:57.548
of the dynamics of the nervous system that matter for the periphery.

00:45:57.608 --> 00:46:01.088
Sure, but then you still have to show that in terms of the time constants within

00:46:01.088 --> 00:46:04.308
that system, they sort of match that low pass shielded version of the periphery.

00:46:04.388 --> 00:46:08.728
And in fact, what we do see, for example, the B31, B32 interneurons,

00:46:08.788 --> 00:46:12.308
which I very briefly mentioned and have been characterized. So back in 96,

00:46:12.608 --> 00:46:15.588
this was a collaboration between my lab and Avi Susswein.

00:46:15.668 --> 00:46:18.388
This was published in Journal of Neurophysiology, two companion papers.

00:46:18.908 --> 00:46:24.728
These neurons generate plateau-like potentials. So when they get turned on,

00:46:24.788 --> 00:46:26.768
they actually fire for an extended period of time.

00:46:27.248 --> 00:46:32.688
And their axons go out to the periphery, and they activate the grasper's protractor

00:46:32.688 --> 00:46:34.188
muscle, the I2 muscle. Awesome.

00:46:34.368 --> 00:46:39.848
So the time course of their activation is very well tuned to the muscle.

00:46:40.008 --> 00:46:41.668
And in fact, we did a paper in 1999.

00:46:42.308 --> 00:46:44.248
This was a collaboration with Pat Crago.

00:46:45.118 --> 00:46:49.878
And a student that he and I shared, we actually did a detailed biomechanical

00:46:49.878 --> 00:46:54.758
study with a servomotor to look at what happens as you stimulate the nerve and

00:46:54.758 --> 00:46:58.878
looked at the force frequency, length, tension, and force velocity properties of the muscle.

00:46:58.938 --> 00:47:02.278
And we actually built a model of it based on the neural inputs.

00:47:02.378 --> 00:47:05.518
And we show that if you took EMG recordings, you could actually get movements

00:47:05.518 --> 00:47:09.238
out of that model that look very similar to the ones that you actually measured.

00:47:09.638 --> 00:47:12.898
So again, when I talk about that kind of transduction stuff,

00:47:13.098 --> 00:47:15.618
it's not based on just, well, I think it's okay.

00:47:15.718 --> 00:47:18.238
We actually have done a lot of that hard work. Not as much as,

00:47:18.278 --> 00:47:23.218
I mean, it's a sensitive point for me, for any biologist, because there's always more that I can do.

00:47:23.358 --> 00:47:26.078
And so immediately you're going to get me on the defensive correctly,

00:47:26.338 --> 00:47:30.338
because talking to my fellow biologists, they're going to demand even more,

00:47:30.398 --> 00:47:31.658
and I demand that of myself.

00:47:32.118 --> 00:47:38.178
But we have gone quite a far ways to argue that the things that we're seeing are right.

00:47:38.318 --> 00:47:40.978
So for example, one of the things we found, which is quite interesting,

00:47:41.178 --> 00:47:48.538
if you look at the force-frequency relationship, if the neuron fires at 6 hertz

00:47:48.538 --> 00:47:51.418
or 5 hertz, very little happens in the muscle.

00:47:51.998 --> 00:47:57.118
If it goes above 10 hertz for a period of about 200 or 300 milliseconds,

00:47:57.318 --> 00:47:58.318
force begins to develop.

00:47:58.778 --> 00:48:01.918
And that was just something we found from actually doing the stimulation.

00:48:02.218 --> 00:48:06.918
Now you look in the animal or you look on the muscle and you look at its activation

00:48:06.918 --> 00:48:12.018
pattern, and lo and behold, When protraction is occurring, it goes on very intensely

00:48:12.018 --> 00:48:14.918
above that frequency for that period of time at least.

00:48:15.558 --> 00:48:20.398
And what we also showed was that if you increase the duration of the activation,

00:48:20.618 --> 00:48:22.518
you can increase the force the muscle generates.

00:48:22.698 --> 00:48:27.158
It's very sensitive to that. So yeah, we are very sensitive to those issues

00:48:27.158 --> 00:48:29.498
and our system is sensitive to those issues.

00:48:29.698 --> 00:48:33.018
Right, but you do agree that that's sort of right now in the future, right?

00:48:33.038 --> 00:48:36.198
To match these ideas about… Well, what I'm saying is that for this one particular

00:48:36.198 --> 00:48:39.498
example, this one muscle and this particular set of interneurons,

00:48:39.558 --> 00:48:41.538
we actually have done the matching and it matches well.

00:48:41.658 --> 00:48:45.138
Right. So does that tell me that the rest of it matches? Of course not.

00:48:45.158 --> 00:48:46.438
I take your point very strongly.

00:48:46.558 --> 00:48:51.418
But what it does is it encourages me to say, hmm, we might really be on the right track.

00:48:51.718 --> 00:48:56.078
Right. So would you claim that the plesia brain is also like a liquid state machine now?

00:48:56.978 --> 00:48:58.258
No, I don't think I would.

00:48:59.298 --> 00:49:04.598
Okay. Nor would I claim that it's a quantum computer. All right. Good. That's progress.

00:49:08.947 --> 00:49:12.887
The last type of experiments that you were describing were these simulations

00:49:12.887 --> 00:49:17.607
of, let's say, sort of a very hybrid kind of model.

00:49:17.847 --> 00:49:20.587
Yes, the artificial insect. Exactly, yeah, different bits and pieces.

00:49:20.667 --> 00:49:24.047
How many insects did you, or different species did you combine there?

00:49:24.147 --> 00:49:28.527
We used probably five or six different stories that people had developed over

00:49:28.527 --> 00:49:30.467
the years and put them together.

00:49:30.707 --> 00:49:33.447
I mean, one of the things that was amusing to me, people had warned me about

00:49:33.447 --> 00:49:37.227
this, but you can have 30 or 40 years of hard neurobiological work,

00:49:37.227 --> 00:49:41.347
And one model can eat all that up in about a month or two. It's amazing.

00:49:41.627 --> 00:49:44.327
It's so scary. And then the person's coming back to you and say,

00:49:44.447 --> 00:49:45.507
well, what should we do here?

00:49:45.767 --> 00:49:48.967
I said, well, they haven't done the measurements yet. So what am I supposed to do?

00:49:49.227 --> 00:49:54.027
Now, Randy wasn't like that. He actually mastered the literature and he had good ideas himself.

00:49:55.087 --> 00:49:59.827
But the parts of this that we were the least satisfied with were ones where

00:49:59.827 --> 00:50:01.987
we had to actually make things up.

00:50:03.067 --> 00:50:07.427
And to the extent that we could use dynamics and connectivity,

00:50:07.947 --> 00:50:11.127
that actually emerged from the literature but had never been looked at this

00:50:11.127 --> 00:50:15.947
way we were both much happier and it was very interesting because no one had

00:50:15.947 --> 00:50:20.767
to at the time we did this try to actually number one take all these different

00:50:20.767 --> 00:50:25.287
things and put them into a model but number two try to create an entire functioning.

00:50:26.587 --> 00:50:31.027
Agent that could actually get around and do stuff I mean this is quite some

00:50:31.027 --> 00:50:35.207
time back already Yeah, we did this in the late 80s, and that review article

00:50:35.207 --> 00:50:38.287
that came out in American Scientist was in 1990.

00:50:38.627 --> 00:50:43.187
Right. So this was quite a while. So Rodney Brooks was at that time just starting

00:50:43.187 --> 00:50:44.807
to do some of his interesting work.

00:50:44.907 --> 00:50:48.367
And it was very interesting to me because I think he was reasonably successful

00:50:48.367 --> 00:50:55.147
when he was down to the level of doing these, you know, Genghis and these other robotic insects.

00:50:55.207 --> 00:50:59.667
When he went up to cog and things like that, the progress slowed down very substantially.

00:50:59.987 --> 00:51:01.827
Absolutely. No, this is clear.

00:51:03.307 --> 00:51:08.027
But then, so also there, so earlier we discussed the issue of validation and

00:51:08.027 --> 00:51:09.807
let's say the neuronal level.

00:51:09.967 --> 00:51:12.947
And then with these experiments, we had to make this issue of validation of the behavioral level.

00:51:13.087 --> 00:51:17.527
Yes. Because at what point, when can you really say, look, behaviorally this is valid, right?

00:51:17.547 --> 00:51:21.607
Is it all about, okay, as long as it survives my simulated environment,

00:51:21.947 --> 00:51:25.707
or should we look at very, let's say, as long as it displays the same behavioral

00:51:25.707 --> 00:51:27.807
patterns I would observe in the animal on these conditions?

00:51:27.807 --> 00:51:30.747
This goes back, and again, at the beginning, very beginning of my talk,

00:51:30.887 --> 00:51:33.327
I talked a little bit about the role that modeling plays.

00:51:33.487 --> 00:51:37.327
And I didn't specify this, but let me sort of clarify this because it's something

00:51:37.327 --> 00:51:40.267
I've given a lot of thought to, and I like that question very, very much.

00:51:40.978 --> 00:51:45.598
There are two really different ways of thinking about what a model or a theory can do.

00:51:46.218 --> 00:51:52.558
One is a quantitative predictor of what an actual system will do next.

00:51:52.838 --> 00:51:58.638
And the other is almost a kind of applied philosophical exploration of the space of possibilities.

00:51:59.078 --> 00:52:04.598
Those are often much less respected, but they can be very important because

00:52:04.598 --> 00:52:09.418
they can say, look, we didn't even realize that you could create from these

00:52:09.418 --> 00:52:10.938
various ideas that are out in the literature,

00:52:11.158 --> 00:52:15.838
a device that could actually survive in this very simplified environment for

00:52:15.838 --> 00:52:22.358
an extended period of time, that already changes how people couple with the process of modeling.

00:52:22.498 --> 00:52:25.678
That gets people who would ordinarily scoff and say, models,

00:52:25.858 --> 00:52:30.418
waste of time, to say, wow, you could really do that?

00:52:30.798 --> 00:52:34.618
I would love to do something like that. And then my immediate response when

00:52:34.618 --> 00:52:37.118
someone comes to me and does that is your question.

00:52:37.238 --> 00:52:40.058
What is your question? What do you want to do with it?

00:52:40.198 --> 00:52:44.198
Are you interested in trying to capture a behavioral phenomenon here?

00:52:44.378 --> 00:52:48.218
Are you interested in a level crossing model? You want to test that some underlying

00:52:48.218 --> 00:52:51.738
mechanism actually generates the macroscopic thing you're seeing?

00:52:52.018 --> 00:52:54.658
Or are you actually trying to make a very quantitative prediction?

00:52:54.998 --> 00:52:58.498
Or are you just an engineer who would like to take some key ideas from biology

00:52:58.498 --> 00:53:02.278
and run with them and make it into a methodology that may no longer have made

00:53:02.278 --> 00:53:06.418
contact with the biology, but could be used in a general way? What's your interest?

00:53:07.038 --> 00:53:11.098
That's the next question. And then the validation, and I gave you a little bit

00:53:11.098 --> 00:53:16.338
of this, comes from how useful is it for those particular very different applications?

00:53:16.738 --> 00:53:21.358
So for an engineer, to the extent that, I mean, as a neuroscientist,

00:53:21.398 --> 00:53:25.778
I think that feed forward neural networks with backprop is a pathetic caricature

00:53:25.778 --> 00:53:28.998
of the complexity and the richness of the dynamics of real nervous systems,

00:53:29.078 --> 00:53:32.338
but as a way of generating interesting mappings,

00:53:32.578 --> 00:53:34.238
it's fascinating.

00:53:34.238 --> 00:53:38.258
If, on the other hand, vector array machines can do just as good a job or faster,

00:53:38.438 --> 00:53:41.058
then they're going to take over. And I have no problems with that.

00:53:41.318 --> 00:53:44.718
But then if I say, well, understanding the nervous system isn't that useful,

00:53:44.838 --> 00:53:46.318
see what happened with neural networks.

00:53:46.698 --> 00:53:51.418
I say, well, that was because the engineers took the part that they could actually

00:53:51.418 --> 00:53:55.658
controllably work with and claim that that was the whole thing.

00:53:55.738 --> 00:53:59.798
And if it wasn't, then maybe they missed some of the parts that matter, like the dynamic.

00:54:00.158 --> 00:54:06.618
Right. But the two issues you highlighted here is on the one hand to be able

00:54:06.618 --> 00:54:10.498
to predict for the system itself what it would do next in a certain situation. Right.

00:54:11.874 --> 00:54:17.434
But what you seem to exclude in that summary is to predict back into the empirical domain.

00:54:17.654 --> 00:54:20.654
So are you excluding this? No, not at all. Not at all. Because again,

00:54:20.774 --> 00:54:23.614
I talked a little bit about level crossing models before you came,

00:54:23.734 --> 00:54:25.054
and I want to emphasize this.

00:54:25.554 --> 00:54:29.314
And I showed that again when I showed in the Eplizy example.

00:54:29.754 --> 00:54:34.074
We had this hypothesis about how this muscle could do essentially two different

00:54:34.074 --> 00:54:35.914
things, push the grasper forward or backwards.

00:54:36.574 --> 00:54:39.154
And building a physical instantiation said, yes, it could.

00:54:39.934 --> 00:54:43.914
It didn't prove it, but it said, yes, it could. Moreover, it said,

00:54:43.994 --> 00:54:46.634
if that's true, it might have to be activated.

00:54:46.834 --> 00:54:49.954
Again, there was this sequence of activations that we had to use.

00:54:50.974 --> 00:54:54.434
That suggested something for what you might see in the animal,

00:54:54.534 --> 00:54:55.934
which could be tested. it.

00:54:55.954 --> 00:54:59.374
And in fact, we've been spending time looking at the different domains.

00:54:59.494 --> 00:55:01.614
It turns out that jaw muscle actually has different domains,

00:55:01.754 --> 00:55:04.634
and there are different motor neurons addressed to the different domains,

00:55:04.834 --> 00:55:09.694
anterior and posterior, and their activation patterns change in time depending

00:55:09.694 --> 00:55:11.194
on what behavior the animal's doing.

00:55:11.414 --> 00:55:15.534
So some of the things that the robot was doing was actually suggesting something

00:55:15.534 --> 00:55:19.514
about what you might see at this broad level if you actually look carefully

00:55:19.514 --> 00:55:23.074
at the biological system.

00:55:23.154 --> 00:55:27.154
So I love when that happens, and I'm very much aware that that has to happen

00:55:27.154 --> 00:55:30.094
for it to validate it as useful for the biology.

00:55:30.434 --> 00:55:35.774
Okay. But then for the different models you described, which of these comes

00:55:35.774 --> 00:55:39.154
the closest to actually satisfying all these different levels of constraints?

00:55:39.254 --> 00:55:43.514
Well, I think the argument I was making about stable heteroclinic channels and

00:55:43.514 --> 00:55:46.234
the possibility of mapping that onto neural architectures.

00:55:47.550 --> 00:55:50.130
The reason I'm so excited about that, of course, is we haven't done all the

00:55:50.130 --> 00:55:53.230
hard tests. So I can pretend in my mind that it all works.

00:55:54.310 --> 00:55:59.050
But my sense is that that could actually be a way of breaking open these problems,

00:55:59.230 --> 00:56:04.130
and especially in systems where you are trying to generate a stable pattern

00:56:04.130 --> 00:56:06.970
of activation of some number of elements.

00:56:07.050 --> 00:56:13.510
Because, again, the low-pass filtering of the periphery is an enormous opportunity to simplify.

00:56:13.510 --> 00:56:17.370
Amplify, because what it means is that really, really fast transition changes

00:56:17.370 --> 00:56:21.830
are going to all be filtered out, so that you're going to have to maintain state

00:56:21.830 --> 00:56:25.630
in the nervous system for long enough for that message to get out and have an

00:56:25.630 --> 00:56:26.910
impact on the periphery.

00:56:27.190 --> 00:56:29.550
And that means that some neurons are going to go into saturation,

00:56:29.890 --> 00:56:33.650
others are going to be off, and that's going to happen for hundreds of milliseconds.

00:56:34.490 --> 00:56:37.930
So that then means that you can define, it's a little artificial,

00:56:38.110 --> 00:56:39.770
but you can define stable patterns

00:56:39.770 --> 00:56:43.030
in the nervous system and look for the mechanisms that generate them.

00:56:43.310 --> 00:56:46.610
And that's the thing that we would then be looking for. And that would be corresponding

00:56:46.610 --> 00:56:49.270
to these various attractors that we've talked about.

00:56:49.410 --> 00:56:52.250
So that's, I think, kind of exciting. Yes, absolutely.

00:56:52.670 --> 00:56:59.230
So in the end, when you also try to make this step towards engineering,

00:56:59.610 --> 00:57:04.170
you highlighted four principles that you felt, look, if you want to talk about biomimetic,

00:57:04.850 --> 00:57:09.170
engineering or biologically based engineering, then what you have to think about

00:57:09.170 --> 00:57:12.090
is evolution, learning, development, and initial condition. Yes.

00:57:12.350 --> 00:57:17.810
Right. So why are these now the four key principles you would like to generalize into engineering?

00:57:17.970 --> 00:57:20.590
All right. So let me go through them. And that's a wonderful question.

00:57:23.010 --> 00:57:23.890
Engineering, design.

00:57:25.422 --> 00:57:33.182
Um, and then there's manufacturing, then there's control, and then there's history.

00:57:33.742 --> 00:57:38.162
I think those are the four things. So if we look at, uh, cars,

00:57:38.382 --> 00:57:43.162
so there's a, in Cleveland, there's the auto aviation, uh, museum because Cleveland

00:57:43.162 --> 00:57:46.702
was actually, people forget this, one of the birthplaces of various different

00:57:46.702 --> 00:57:48.142
kinds of automobile companies. journeys.

00:57:49.022 --> 00:57:56.402
And the initial version of cars is really a horseless buggy.

00:57:56.782 --> 00:57:59.682
They look exactly like buggies did in the late 19th century,

00:57:59.782 --> 00:58:02.102
and there's a motor in front. And that's what they look like.

00:58:02.202 --> 00:58:04.662
They don't look at all like what we think of as cars.

00:58:04.842 --> 00:58:08.142
It took some time for people to free themselves from that.

00:58:08.602 --> 00:58:14.422
Now, to some extent, engineers, because of the desire to create disruptive technologies

00:58:14.422 --> 00:58:18.762
are very into this idea of coming up with something completely new.

00:58:18.882 --> 00:58:23.122
So one of the things you're doing to take notes is an iPad, which I think is

00:58:23.122 --> 00:58:24.942
a beautiful example of disruptive technology.

00:58:25.042 --> 00:58:29.662
People had come up with various tablet architectures and none of them had this

00:58:29.662 --> 00:58:34.482
sort of seamless integration and this way of doing media and this way of really

00:58:34.482 --> 00:58:37.702
just beautifully, elegantly putting all the different things together.

00:58:37.922 --> 00:58:41.182
And so this is displacing and it's defining a whole new market.

00:58:41.482 --> 00:58:43.662
So there's an enormous press for that in engineering.

00:58:44.702 --> 00:58:49.382
In biological systems, history matters a great deal, because as I said,

00:58:49.542 --> 00:58:52.862
when an organism is born, it has to be up and running instantly.

00:58:53.302 --> 00:58:56.762
And so initial conditions become absolutely crucial.

00:58:57.422 --> 00:59:01.382
If you start off with an organism that more or less has to be programmed from

00:59:01.382 --> 00:59:03.102
birth in very complex ways,

00:59:03.382 --> 00:59:07.662
then you had better build around it a social structure that allows it to be

00:59:07.662 --> 00:59:13.362
taken care of in its helpless state for years and years until it finally is

00:59:13.362 --> 00:59:14.322
released into the world.

00:59:14.882 --> 00:59:19.122
Human infants fall into that category. But if you're an insect,

00:59:19.222 --> 00:59:21.942
you don't have that luxury. You get up out of the egg and you walk away.

00:59:23.541 --> 00:59:26.701
And when you molt, you go someplace, you're still for a while,

00:59:26.781 --> 00:59:31.481
and then you discard your old body and walk away from it and immediately start doing things.

00:59:31.601 --> 00:59:36.141
You don't have the luxury to relearn how to do various things. It just has to work.

00:59:37.021 --> 00:59:41.161
Okay? So that's why it's stressed initial conditions. And that initial conditions

00:59:41.161 --> 00:59:45.421
also, I wanted to, that really dovetails with all the other things.

00:59:45.641 --> 00:59:49.461
What the tendency among engineers, and I completely understand that having done

00:59:49.461 --> 00:59:52.721
some engineering myself, you want to come up with this really brilliant idea.

00:59:52.721 --> 00:59:54.721
And then you want it to be a design principle.

00:59:54.961 --> 00:59:59.541
And so you have your hammer and everything is going to be a nail, okay?

00:59:59.641 --> 01:00:03.181
So if you are a control theorist, then everything is going to put in the rubric

01:00:03.181 --> 01:00:06.321
of control theory. If you're an information theorist, everything can be stated

01:00:06.321 --> 01:00:07.501
in terms of information theory.

01:00:07.681 --> 01:00:12.001
If you realize the power of Kalman filters, then everything is going to be done

01:00:12.001 --> 01:00:14.241
using a Kalman filter, et cetera, et cetera, et cetera.

01:00:14.881 --> 01:00:21.641
But biological systems, the initial conditions is just an example of how much history matters.

01:00:22.321 --> 01:00:26.881
Evolution, which is not a design process of the sort that engineers are used

01:00:26.881 --> 01:00:29.901
to, is one that builds on what already exists.

01:00:30.201 --> 01:00:33.361
It's canalized by all the things that have happened beforehand.

01:00:34.241 --> 01:00:38.901
And it basically is a proof of principle system for design.

01:00:39.461 --> 01:00:42.761
If you leave offspring, you're good.

01:00:42.881 --> 01:00:46.601
And if you don't, you disappear from the gene pool and you become extinct and that doesn't work.

01:00:46.701 --> 01:00:49.661
And it doesn't mean that there was anything wrong. Many of the solutions that

01:00:49.661 --> 01:00:52.941
dinosaurs came up with, if we had a way of recreating a dinosaur,

01:00:53.141 --> 01:00:56.661
we might find the things that they're actually much better at than organisms

01:00:56.661 --> 01:00:58.641
that live in our world now.

01:00:59.081 --> 01:01:01.261
It wasn't a statement that they were somehow inadequate.

01:01:02.323 --> 01:01:05.303
Changes in their environment at that time were ones they couldn't tolerate,

01:01:05.403 --> 01:01:09.103
so they went extinct. And that could happen to us just as easily as it happened to the dinosaurs.

01:01:09.423 --> 01:01:15.743
But engineers might say that actually right now in their practice already,

01:01:15.823 --> 01:01:18.143
they have included things like learning and initial conditions.

01:01:18.323 --> 01:01:23.203
Yes, they've started to. But the realities are that the way they do it is very

01:01:23.203 --> 01:01:26.763
different from the way that biology still does, largely.

01:01:26.763 --> 01:01:31.923
And again, this goes back to when I talked earlier about the wiring up of the

01:01:31.923 --> 01:01:33.683
nervous system, the early nervous system.

01:01:34.043 --> 01:01:38.623
So what happens? So you have, over evolutionary time, you have a genetic code

01:01:38.623 --> 01:01:41.663
that then gives rise to an exponentially

01:01:41.663 --> 01:01:46.923
cascading process that takes one fertilized cell into, I mentioned,

01:01:47.063 --> 01:01:49.503
trillions of cells that are precisely organized.

01:01:50.143 --> 01:01:53.583
Organized, and the plasticity that allowed the development to unfold because

01:01:53.583 --> 01:01:57.003
there are all these local rules that allow the system to wire itself up,

01:01:57.083 --> 01:01:59.043
their gradients, their local cues.

01:01:59.283 --> 01:02:04.323
And so this whole system is not based on some central organizer that looks from

01:02:04.323 --> 01:02:06.503
the top and says, you go over here and you go over there.

01:02:06.763 --> 01:02:12.363
The system is building these cues as it's unfolding and generating new cues

01:02:12.363 --> 01:02:14.183
that help the next stage of unfold.

01:02:14.383 --> 01:02:18.083
It's really amazing. And people have really not wrapped their minds around it,

01:02:18.143 --> 01:02:22.123
either in biology fully or in modeling or in engineering.

01:02:22.243 --> 01:02:25.263
And I see that as a total frontier that we could do amazing things with.

01:02:25.443 --> 01:02:29.603
And that's also to which you apply, let's say, your Swiss army knife metaphor, right?

01:02:30.023 --> 01:02:33.763
Precisely. What we would do, and I have no complaints about this.

01:02:33.843 --> 01:02:37.203
I write code because I program, and I loved it. I mean, it's enormous fun.

01:02:37.363 --> 01:02:40.743
You are the god of your own world. You can make anything happen.

01:02:41.103 --> 01:02:44.903
But if you're a good coder, you are trained or you learn very quickly.

01:02:45.043 --> 01:02:46.943
You don't want self-modifying code.

01:02:47.183 --> 01:02:51.823
That's a bad thing. And you don't want to use other people's old code because you can't figure it out.

01:02:51.883 --> 01:02:55.523
And you want to make it modular and you want to give functions that have well-defined

01:02:55.523 --> 01:02:57.463
functionalities. You want to have clean interfaces.

01:02:57.783 --> 01:03:03.243
I mean, the Unix operating system, the pipes, forks, setting up processes,

01:03:03.583 --> 01:03:06.203
killing the processes, it's beautiful.

01:03:07.382 --> 01:03:11.302
It's beautiful. So again, I actually, unlike many biologists,

01:03:11.462 --> 01:03:13.262
I have immersed myself in how

01:03:13.262 --> 01:03:16.582
engineers do things, and I've built and used those tools, and I love them.

01:03:17.462 --> 01:03:21.382
But when I look at the biological systems, I don't see anything like that going on inside.

01:03:21.942 --> 01:03:26.442
And so what that tells me is, wait, if I really want to do this,

01:03:26.462 --> 01:03:31.402
certainly a digital computer can emulate anything, but the architecture is so

01:03:31.402 --> 01:03:33.862
different that it's not going to fall naturally on it.

01:03:33.862 --> 01:03:37.442
And I'm going to be fighting with it, and I'm going to run into all these computational

01:03:37.442 --> 01:03:41.762
bottlenecks because I don't even have an architecture that's massively parallel,

01:03:41.942 --> 01:03:46.022
asynchronous, and even though it's much slower, can do all of this stuff just

01:03:46.022 --> 01:03:48.042
naturally. And that's what I have in the nervous system.

01:03:48.842 --> 01:03:52.342
So what's been fun for me, but it's also sometimes frustrating working with

01:03:52.342 --> 01:03:58.682
engineers, is trying to open their eyes to the fact that really there is this

01:03:58.682 --> 01:04:01.222
other technology and there's another way of thinking about it.

01:04:01.222 --> 01:04:05.162
So why engineers do what they do is something that completely makes sense to

01:04:05.162 --> 01:04:07.342
me, and I don't want to encourage them not to do that.

01:04:07.482 --> 01:04:11.882
I want them to explore, take forays into, be willing to take risks,

01:04:11.962 --> 01:04:15.302
and try thinking about things really in a different way.

01:04:15.662 --> 01:04:19.602
And that's where immersing yourself in the biology and spending time with someone

01:04:19.602 --> 01:04:24.142
who, like myself, really cares about talking to engineers, knows math,

01:04:24.362 --> 01:04:28.622
actually enjoys these kinds of conversations, is a valuable exercise. size.

01:04:28.782 --> 01:04:32.002
One of the things I've done in the past with other engineers have approached me.

01:04:32.062 --> 01:04:35.742
So I will sit down with, for example, molecular biology of the cell and work

01:04:35.742 --> 01:04:39.442
through the chapter on development with someone who is interested in that.

01:04:39.542 --> 01:04:43.602
And it blows their mind. But having a guide, it really is like, you know...

01:04:44.220 --> 01:04:47.280
I think of the Aeneid. You really need to have your Virgil. You need to have

01:04:47.280 --> 01:04:52.640
someone with a lamp guiding you through all the acronyms and all this sort of stuff.

01:04:52.840 --> 01:04:58.280
Okay. Excellent. Without that, it becomes almost impossible to understand what's going on.

01:04:58.520 --> 01:05:02.000
So now to get to the finish line. Yes. Two questions.

01:05:02.480 --> 01:05:07.640
So if I now want to exercise this sort of Bernstein invariance of motor control,

01:05:07.880 --> 01:05:10.520
I'm painting the wall of the campus here.

01:05:10.520 --> 01:05:17.520
There's one law that we should apply to a biometric understanding of biology and technology.

01:05:18.040 --> 01:05:23.400
So this is the Hillel-Cheels law. Right, right. What's the one law?

01:05:24.540 --> 01:05:28.680
Must have got it right there. Pay attention to the biology. Okay.

01:05:28.760 --> 01:05:30.220
Pay attention to biology. And then?

01:05:31.700 --> 01:05:34.960
But I would say to the biology. That's the law for the engineers.

01:05:35.160 --> 01:05:40.260
Right. And from the point of view for the biologists, I paint on the other wall. Yeah.

01:05:40.520 --> 01:05:47.000
Um, focus on the principles because the difficulty you have when you come to

01:05:47.000 --> 01:05:50.980
most standard biology talks, even I have difficulty with this because I think

01:05:50.980 --> 01:05:54.320
differently than many biologists. I like math. I like abstraction.

01:05:54.700 --> 01:05:59.420
I'm not focused about, you know, one detail after another. They just bury you in details.

01:05:59.600 --> 01:06:02.300
They love the details. They can't get enough of the details.

01:06:02.300 --> 01:06:05.800
So you're sitting there going, they've got all these names, and they've got

01:06:05.800 --> 01:06:09.520
all these structures, and they have all these details. What matters, please?

01:06:10.040 --> 01:06:13.480
And the answer is, part of the reason it's so difficult to answer that is they

01:06:13.480 --> 01:06:15.860
don't know for sure because it may depend on context.

01:06:16.560 --> 01:06:21.780
But if you're focused on trying to articulate principles, if you think about

01:06:21.780 --> 01:06:25.260
the principles as a biologist, and if you pay attention to the biology as an

01:06:25.260 --> 01:06:28.480
engineer, I think you could make enormous advances in the next decade.

01:06:29.628 --> 01:06:34.688
Decade or two. Okay, but now if I have less patients in a whole decade and I

01:06:34.688 --> 01:06:36.828
want to go visit you five years from now.

01:06:36.988 --> 01:06:41.208
Yes. And I want to say, okay, in 2011, you made this one prediction.

01:06:41.288 --> 01:06:43.888
Today I'm checking whether it actually came out or not. Right,

01:06:43.888 --> 01:06:47.128
right, right. What's this one prediction that you're most enthusiastic about today?

01:06:47.548 --> 01:06:51.248
I would go with the stable heteroclinic channels as something that I hope within

01:06:51.248 --> 01:06:55.868
five years we have some more evidence that they might be relevant to how the system works.

01:06:56.248 --> 01:06:59.408
That would be the one I would go for in terms of my personal work.

01:06:59.628 --> 01:07:02.968
Because that's something I have control over and I see the experiments I would do.

01:07:03.068 --> 01:07:08.388
In terms of fields as a whole, there I have less, I'm less sanguine.

01:07:08.528 --> 01:07:13.028
There are rare places where people are trying to get biologists to actually

01:07:13.028 --> 01:07:15.828
do good biology, to talk to engineers who do good engineering.

01:07:16.148 --> 01:07:20.048
There are a few places. I would hope in five years a few of them have had some

01:07:20.048 --> 01:07:25.588
hits and have gotten other groups to say, oh, wait, you mean a collaboration

01:07:25.588 --> 01:07:28.568
between a biologist and engineer is not just that we get money together,

01:07:28.808 --> 01:07:33.108
we talk once every six months or a year, and then we get more money together?

01:07:33.288 --> 01:07:35.588
You mean we actually have to talk to each other on a daily basis?

01:07:36.108 --> 01:07:39.508
And I would like to see more of that happening, because if we saw more of that

01:07:39.508 --> 01:07:43.088
happening, some of the things I was pointing to, like focusing on development,

01:07:43.268 --> 01:07:48.108
like trying to come up with plasticity that affects global dynamics,

01:07:48.388 --> 01:07:52.988
really, that's going to take an interdisciplinary, cross-disciplinary effort.

01:07:52.988 --> 01:07:56.788
It's going to take the minds, the best minds of people who really are working.

01:07:57.448 --> 01:08:01.828
They're in the trenches, in the biological systems, and they're in the trenches

01:08:01.828 --> 01:08:07.428
building engineered systems, really working together and coming back and forth with each other.

01:08:07.428 --> 01:08:09.868
And again, I've been doing this for years with different colleagues,

01:08:09.908 --> 01:08:13.568
and it's been enormous fun. But it's the reason I've accomplished what I've accomplished.

01:08:14.848 --> 01:08:18.868
That constant willingness to go out of your immediate comfort area and to start

01:08:18.868 --> 01:08:22.468
to talk someone else's language and to see the world through their eyes and

01:08:22.468 --> 01:08:26.188
to recognize that that's the way you have to try to take what you're learning

01:08:26.188 --> 01:08:28.108
and help them see it that way.

01:08:28.768 --> 01:08:31.588
That's, I think, where the most productive things are going to happen over the

01:08:31.588 --> 01:08:32.588
next five years. Excellent.

01:08:32.768 --> 01:08:35.928
So, Lucille, Chiel, thank you very much for this conversation.

01:08:35.928 --> 01:08:39.108
It's a pleasure. Pleasure. Thank you. Sure.

01:08:40.148 --> 01:08:45.948
The CSN Podcast was produced by the Convergent Science Network of Biometrics

01:08:45.948 --> 01:08:52.348
and Biohybrid Systems, a project funded by the European Sevens Research Framework Program.

01:08:53.928 --> 01:08:59.208
For more interviews, recorded lectures, or upcoming conferences in the field

01:08:59.208 --> 01:09:05.488
of biometrics and biohybrid systems, go to csnnetwork.eu.

01:09:06.068 --> 01:09:07.628
And thank you for listening.

01:09:06.160 --> 01:09:13.680
Music.