WEBVTT

00:00:03.557 --> 00:00:10.517
This is the Convergent Science Network podcast. Leading researchers in the domain

00:00:10.517 --> 00:00:16.737
of neuroscience, brain theory and technology are interviewed by Paul Verschure and Tony Prescott.

00:00:19.097 --> 00:00:23.237
It's Paul Verschure with the Convergent Science Network podcast together with

00:00:23.237 --> 00:00:24.617
my colleague Tony Prescott.

00:00:25.517 --> 00:00:30.397
And we're here today with Randy Beer, here, one of the speakers of our BCBT summer school.

00:00:30.857 --> 00:00:37.357
And Randy, you have been now for quite a while pushing a very specific view

00:00:37.357 --> 00:00:41.337
on how we should understand brain behavior and environment.

00:00:41.677 --> 00:00:44.917
So what's that specific perspective you bring to bear on that?

00:00:45.157 --> 00:00:49.317
Well, I mean, part of it is just that we should look at brain,

00:00:49.517 --> 00:00:55.737
body, and environment as the unit of analysis, that we shouldn't just focus on neural activity.

00:00:55.737 --> 00:01:00.657
It shouldn't just focus on anatomical sort of connectivity of brains and so

00:01:00.657 --> 00:01:02.917
on, that we need to think of them in context. Mm-hmm.

00:01:03.677 --> 00:01:08.977
Beyond that, I guess I've primarily until very recently, which I'll talk about,

00:01:09.317 --> 00:01:16.957
pushed the idea that we need to think about this brain-body-environment system as coupled dynamics,

00:01:17.577 --> 00:01:21.897
as coupled dynamical systems, and we need to use tools from dynamical systems

00:01:21.897 --> 00:01:27.057
theory to try and understand the behavior that such things produce.

00:01:27.317 --> 00:01:31.337
But lately, what we've been doing and what I mentioned in my talk was I also

00:01:31.337 --> 00:01:36.737
think we We can bring other bodies of mathematics to bear on these systems.

00:01:36.777 --> 00:01:41.497
In particular, we've been exploring the use of information theory and then the

00:01:41.497 --> 00:01:43.757
relationship between dynamics and information.

00:01:43.857 --> 00:01:49.317
Right. But before we get down and dirty on that, I think you also,

00:01:49.397 --> 00:01:55.177
you could now argue if you say, well, the unit to analyze is actually brain, body, and environment.

00:01:55.697 --> 00:01:59.017
I could say, well, that's nice. Now you have just complexified the problem and

00:01:59.017 --> 00:02:03.977
you will still be forced to sort of decompose that or split it up in some way,

00:02:04.077 --> 00:02:06.437
right? So how does that not get you in trouble?

00:02:07.437 --> 00:02:13.377
Well, I mean, I guess that presupposes that splitting things up is a bad thing. Yeah.

00:02:15.145 --> 00:02:18.945
I think you always have to make some splits. When you talk about brain-body-environment

00:02:18.945 --> 00:02:24.065
system, you've taken a physical universe and you've carved it up into what's

00:02:24.065 --> 00:02:26.105
the brain and what's the body and what's the environment.

00:02:26.365 --> 00:02:31.265
So I don't necessarily have an opposition to making splits.

00:02:31.745 --> 00:02:36.885
But I think the focus when you're studying, say, the brain-body-environment

00:02:36.885 --> 00:02:40.605
system from a dynamical point of view is on the interaction between the components

00:02:40.605 --> 00:02:45.025
rather than thinking about each of the components in isolation and imagining

00:02:45.025 --> 00:02:49.845
that the total behavior is just some kind of a sum of the individual behaviors.

00:02:50.145 --> 00:02:54.885
Right. But then on top of that, you also made a pretty strong point,

00:02:54.925 --> 00:02:59.285
I thought, by saying, well, I don't necessarily want to advocate a single view

00:02:59.285 --> 00:03:02.085
on how we should understand these systems,

00:03:02.225 --> 00:03:06.605
and I certainly don't want to make any ontological commitments to any of these views.

00:03:06.745 --> 00:03:11.925
So why are you saying that? Why are you so strong? Well, I think there's been

00:03:11.925 --> 00:03:17.725
a lot of debate about, say, let's just focus on the brain because that's often

00:03:17.725 --> 00:03:18.665
what this debate involves.

00:03:18.885 --> 00:03:21.245
The brain is a computer.

00:03:21.645 --> 00:03:25.785
The brain is an information processing system. The brain is a dynamical system.

00:03:25.925 --> 00:03:29.725
The brain is a prediction machine. The brain is a complex network, and so on and so forth.

00:03:30.765 --> 00:03:35.085
What I tried to do in my talk was suggest that it might be more fruitful to

00:03:35.085 --> 00:03:37.885
reformulate that a little bit.

00:03:38.965 --> 00:03:43.025
So that's sort of an ontological claim into more of an epistemological claim,

00:03:43.185 --> 00:03:48.885
which is simply that underlying each of those positions, there's some body of

00:03:48.885 --> 00:03:53.545
mathematics, information theory, dynamical systems theory, Bayesian theory,

00:03:53.985 --> 00:03:55.865
formal language, formal theory of computation,

00:03:56.325 --> 00:03:58.965
which isn't intrinsically right or wrong.

00:03:59.105 --> 00:04:03.185
It's kind of a lens through which we can look at the operation of a system.

00:04:03.385 --> 00:04:08.725
And I think the focus should be on the utility of using these various lenses

00:04:08.725 --> 00:04:15.345
as a way to understand, to make predictions about, to get interesting insights

00:04:15.345 --> 00:04:17.705
into the systems that we look at.

00:04:17.725 --> 00:04:21.465
In that sense, dynamical systems theory isn't right or wrong as compared to

00:04:21.465 --> 00:04:25.105
information theory or Bayesian whatever. whatever, it's a matter of under what

00:04:25.105 --> 00:04:28.705
conditions are these different things useful. Now that's not to say that I don't want...

00:04:30.701 --> 00:04:34.281
A general theory. It's just to say that I don't think any of the things that

00:04:34.281 --> 00:04:38.341
have been put on the table are really formulated in a way that they are.

00:04:38.821 --> 00:04:41.481
They're offering such theories. Part of the thing, I was talking with Tony earlier,

00:04:41.761 --> 00:04:46.541
it's not clear how you would ever definitively test one of these things.

00:04:46.721 --> 00:04:51.681
What experiment would you do to disprove the statement that the brain is a dynamical system?

00:04:52.521 --> 00:04:56.841
What test would you do to disprove the idea that the brain makes predictions?

00:04:57.361 --> 00:05:03.281
But you could also argue that this is partially just comparing apples and oranges,

00:05:03.361 --> 00:05:07.621
because the examples you give, dynamical systems, information theory, are methods, right?

00:05:07.661 --> 00:05:12.481
And methods, you could say, are by definition neutral towards… That's the point.

00:05:12.481 --> 00:05:16.161
…unnatural status, right? That's the point. I mean, they're not theories.

00:05:16.341 --> 00:05:18.681
But I don't think these other things are really theories either.

00:05:18.881 --> 00:05:24.261
If anything, I think they're pre-theoretical sort of intuitions about how things

00:05:24.261 --> 00:05:28.701
must be in order to accomplish the behavior that the systems produce.

00:05:29.221 --> 00:05:33.601
But I think that begs the question, though, if these aren't general theories,

00:05:33.761 --> 00:05:36.401
what would a general theory look like in your view?

00:05:36.601 --> 00:05:42.141
I mean, how would you go about starting to build some parsimonious general description

00:05:42.141 --> 00:05:46.441
of, for instance, the brain? Yeah, so, I mean, I don't have one to offer.

00:05:46.781 --> 00:05:52.421
What I'm arguing for is a methodology to get there. And the methodology is not

00:05:52.421 --> 00:05:57.661
this discussion we've just had about different mathematical languages.

00:05:57.761 --> 00:06:02.621
The methodology is this sort of toy models approach, okay, that we need to start

00:06:02.621 --> 00:06:06.521
with simple enough models that raise these issues that we're trying to understand

00:06:06.521 --> 00:06:08.061
and build a theory for that.

00:06:08.960 --> 00:06:12.260
And then tried to incrementally extend it. Now, I'm not saying that's the only

00:06:12.260 --> 00:06:15.800
way to go, but if you look in science, I think historically fundamental theories

00:06:15.800 --> 00:06:20.460
have generally been built around toy models, which were then subsequently extended.

00:06:20.960 --> 00:06:24.160
I'm not saying you can't build sort of empirical theories in other ways,

00:06:24.260 --> 00:06:27.600
but fundamental theories in science have typically been built that way.

00:06:27.800 --> 00:06:32.320
So you're saying it's too early to say what a general theory might look like.

00:06:32.420 --> 00:06:34.620
Yes, I think that's a fair statement. Well, I'm not sure if I would agree with

00:06:34.620 --> 00:06:36.840
that. I mean, I mean, if you look at theories also in physics,

00:06:37.220 --> 00:06:42.920
people would basically be trying to explain natural phenomena that they would

00:06:42.920 --> 00:06:47.260
be confronted with, and they would find validity of the explanations in making

00:06:47.260 --> 00:06:48.760
predictions you can test, right?

00:06:48.820 --> 00:06:52.340
So it's not necessarily similar kind of toy systems as you study.

00:06:52.340 --> 00:06:55.740
What I'm talking about is, so for example, take motion.

00:06:56.000 --> 00:07:02.240
Okay, if you look at the world 350 years ago, motion is a really complicated thing.

00:07:02.640 --> 00:07:08.100
Birds fly, waves crash on the beach, rocks slide down the sides of mountains,

00:07:08.340 --> 00:07:09.920
lights move across the sky.

00:07:10.140 --> 00:07:15.140
What possible principles could underlie all of that?

00:07:15.300 --> 00:07:22.420
And the fact is that people didn't even recognize those as being the same things 350, 400 years ago.

00:07:22.600 --> 00:07:27.420
So if you started out trying to build a theory of these things like that,

00:07:27.600 --> 00:07:29.580
you would be in trouble. And of course, that's not how it developed.

00:07:30.280 --> 00:07:33.880
Galileo, for example, just to pick a starting point, looked,

00:07:34.000 --> 00:07:36.720
considered the motion of balls...

00:07:37.453 --> 00:07:41.893
Across frictionless planes. And he argued that if you thought about that,

00:07:42.133 --> 00:07:45.713
which by the way is quite different from what you actually see if you roll a

00:07:45.713 --> 00:07:49.233
ball across the floor, in particular, friction brings it to a stop.

00:07:49.413 --> 00:07:54.633
But what he realized was that that's a surface appearance, which is actually

00:07:54.633 --> 00:08:00.853
not what you would expect in this highly idealized case of an object moving

00:08:00.853 --> 00:08:02.713
frictionlessly across a plane.

00:08:02.993 --> 00:08:07.793
Ising models are an example of that, where you took a really simple model trying

00:08:07.793 --> 00:08:10.873
to get a handle on what phase transitions were fundamentally about.

00:08:11.533 --> 00:08:15.393
I mentioned in my talk there's a lot of interesting models in quantum gravity

00:08:15.393 --> 00:08:21.893
now which are in principle incapable of accounting for our actual universe,

00:08:22.093 --> 00:08:25.893
and yet they're designed as conceptual tools to kind of play with issues like

00:08:25.893 --> 00:08:29.453
what does it mean to quantize time, which is an issue that comes up a lot.

00:08:29.833 --> 00:08:33.113
This is a really important point because it really means, okay,

00:08:33.213 --> 00:08:37.413
what is the ontological status is in the end of the models that we try to study.

00:08:37.593 --> 00:08:42.573
And you're following a very, I think, a well-defined route where you're saying,

00:08:42.673 --> 00:08:46.813
well, let's not be over-enthusiastic about what we can really achieve today.

00:08:47.713 --> 00:08:52.393
Let's be systematic and focus first on our methods and sharpen these methods

00:08:52.393 --> 00:08:55.193
on test cases that we can sort of control and understand.

00:08:55.533 --> 00:09:00.713
And maybe out of that will come sufficient insight to start to think about theory.

00:09:00.993 --> 00:09:05.413
This is how I hear. Well, again, you keep using methods, but I don't think of

00:09:05.413 --> 00:09:07.793
them as methods. I think of them as mathematical languages.

00:09:09.713 --> 00:09:13.573
Evolutionary algorithms is a method. Dynamical systems theory is not a method.

00:09:13.693 --> 00:09:17.593
It's a mathematical language that you can apply to any system that fits a certain

00:09:17.593 --> 00:09:21.413
form, namely that you can have a state space, a time set, and an evolution operator.

00:09:21.713 --> 00:09:25.273
If you can describe those things, then you've got it. I don't think it's a contradiction, though.

00:09:25.433 --> 00:09:30.713
You described it as a lens, so it's something that you use on the outside.

00:09:30.793 --> 00:09:33.453
Yeah, it's a language. It's a language that you use to talk about something.

00:09:35.716 --> 00:09:40.436
I'm not saying that all modeling takes that form.

00:09:40.636 --> 00:09:43.576
It's one of the reasons why I emphasize this term toy models,

00:09:43.656 --> 00:09:45.256
which is what they're typically called in physics.

00:09:45.336 --> 00:09:49.056
Because I think it's a particular kind of model that in the behavioral and brain

00:09:49.056 --> 00:09:50.776
sciences most people don't realize.

00:09:51.656 --> 00:09:55.596
Typically what modeling means in the behavioral and brain sciences, biology in general.

00:09:55.796 --> 00:09:58.496
Wait, I really tried to characterize what you're doing in those terms.

00:09:58.496 --> 00:10:03.436
So how I read or how I listen to what you're saying, how I understand it is

00:10:03.436 --> 00:10:08.456
to say, well, let's first sharpen these languages that give me this lens to

00:10:08.456 --> 00:10:10.276
look at, if you want, nature.

00:10:11.156 --> 00:10:14.496
Once I've done that sufficiently and I've sharpened these languages,

00:10:14.796 --> 00:10:16.876
then we can start to worry about theory.

00:10:16.996 --> 00:10:19.436
But theory is not right now our highest priority.

00:10:20.136 --> 00:10:22.056
No, I don't think that's accurate.

00:10:23.176 --> 00:10:27.616
What I'm trying to put forward, for example, in the talk, I basically laid out

00:10:27.616 --> 00:10:30.536
what you might call, I didn't use the word, but we've used it elsewhere,

00:10:30.696 --> 00:10:36.976
kind of the information flow architecture of this relational categorization agent.

00:10:38.016 --> 00:10:43.696
That's something that one might imagine being the basis of an information processing

00:10:43.696 --> 00:10:48.136
theory that you could imagine building for lots of other agents.

00:10:48.136 --> 00:10:55.076
What is the pattern of flow of information over time across all the elements of the system?

00:10:56.236 --> 00:11:01.816
Okay, so I don't think, I understand why you're saying this,

00:11:01.876 --> 00:11:07.596
but I don't view myself as basically saying we should postpone theory and just focus on methods.

00:11:07.856 --> 00:11:12.156
I rather see myself as saying we should try to build theory around these toy

00:11:12.156 --> 00:11:15.916
models first and then try to extend it.

00:11:15.916 --> 00:11:20.076
But I guess what other people are saying, which is different from you,

00:11:20.176 --> 00:11:24.956
is, for instance, Gary Marcus was here earlier this week, and he was saying

00:11:24.956 --> 00:11:30.416
in quite a strong way that symbol processing, he was using the example of list processing,

00:11:30.656 --> 00:11:33.676
is a fundamental mechanism that the brain must use.

00:11:33.956 --> 00:11:38.496
So he's taking something that we've learned from symbolic AI,

00:11:38.736 --> 00:11:43.396
and he's saying this isn't just a tool for thinking about how the brain does computation.

00:11:43.396 --> 00:11:48.296
The brain really does computation in this kind of way, or the brain really does

00:11:48.296 --> 00:11:50.676
solve the problems it has in this sort of way.

00:11:50.796 --> 00:11:54.396
And I think in the past, I mean, you may have had a stronger position,

00:11:54.576 --> 00:11:56.856
for instance, in terms of dynamical systems.

00:11:57.196 --> 00:12:03.396
But are you saying now that all of these approaches which are using these tools

00:12:03.396 --> 00:12:04.996
or methods or lenses, whatever,

00:12:05.176 --> 00:12:09.976
not just to study the brain, but as metaphors for understanding the brain,

00:12:10.116 --> 00:12:12.936
you're saying that that's no longer a good way?

00:12:13.396 --> 00:12:16.696
I think the metaphors can be very misleading at this point.

00:12:18.576 --> 00:12:21.536
Like I said, the way I would say it is, how would you empirically,

00:12:21.696 --> 00:12:25.356
so I didn't hear the talk, I can't comment directly, but how would you empirically

00:12:25.356 --> 00:12:29.156
resolve the hypothesis that the brain is a dynamical system?

00:12:30.756 --> 00:12:34.936
It's just not. It's not. And the same thing is true with information processing

00:12:34.936 --> 00:12:37.376
or symbol processing or so on. I'm not sure.

00:12:37.496 --> 00:12:40.316
I don't see what experimental program would resolve that question.

00:12:40.856 --> 00:12:45.096
I guess you'd have to be more specific in terms of what kind of dynamic system

00:12:45.096 --> 00:12:46.136
it was. We can answer this, right?

00:12:46.196 --> 00:12:49.436
Because you could argue that people like Zhao Jingwang and,

00:12:50.016 --> 00:12:55.776
Gustavo Deco here, who actually are applying the dynamical systems view of,

00:12:55.896 --> 00:13:00.976
let's say, attractor landscapes to complex brain dynamics, trying to explain

00:13:00.976 --> 00:13:03.916
the properties of prefrontal cortex in particular,

00:13:04.276 --> 00:13:09.316
trying to explain behavior phenomena like attentional selection and so on.

00:13:09.416 --> 00:13:14.976
So I could argue that there's actually a pretty prominent movement in computational

00:13:14.976 --> 00:13:16.416
neuroscience doing that directly.

00:13:16.656 --> 00:13:23.056
But that sounds like exactly what I meant by using the lens of of dynamical systems theory, okay?

00:13:23.156 --> 00:13:26.496
I don't see how any of that is an ontological claim.

00:13:26.956 --> 00:13:31.616
It's saying the utility of the dynamical systems language is quite high for

00:13:31.616 --> 00:13:34.216
at least some subset of brain processing.

00:13:34.776 --> 00:13:38.176
Well, actually, that's been just the debate very much in the hippocampal literature.

00:13:39.076 --> 00:13:46.176
On whether you might find a memory shaped like in a tractor landscape in CA3

00:13:46.176 --> 00:13:47.616
of the hippocampus, right?

00:13:47.676 --> 00:13:51.476
And then people observe that if you smoothly change environments,

00:13:51.796 --> 00:13:56.136
you actually don't have very stable responses in that system.

00:13:56.236 --> 00:13:58.636
And then they would say, well, here, you see, these cannot be attractors,

00:13:58.656 --> 00:14:02.976
because if it's an attractor, you should jump into that same state if the environment

00:14:02.976 --> 00:14:04.956
is the same, right? Yeah.

00:14:05.910 --> 00:14:09.330
However, then it was shown that under continuous plasticity,

00:14:09.390 --> 00:14:12.610
actually, you might have attractor landscapes that also are reshaping themselves.

00:14:12.750 --> 00:14:16.750
So you don't necessarily always follow the same attractor. So I think in that

00:14:16.750 --> 00:14:20.290
domain, people really have been trying to be pretty literal in saying,

00:14:20.430 --> 00:14:27.010
well, this dynamic systems picture of memory is literally captured in that structure.

00:14:27.270 --> 00:14:32.890
And I hear you saying like, well, this might be a mistake. This might be an over-interpretation.

00:14:33.390 --> 00:14:39.090
I mean, whenever you look through a mathematical lens, it focuses you on certain kinds of questions.

00:14:39.570 --> 00:14:43.310
Looking through the lens of information theory does not lend itself to,

00:14:43.370 --> 00:14:44.690
for example, bifurcations.

00:14:45.350 --> 00:14:47.890
So if you look through the lens of dynamical systems theory,

00:14:47.970 --> 00:14:50.890
of course, the vocabulary that you're going to be bringing to bear are going

00:14:50.890 --> 00:14:53.770
to be things like attractors and bifurcations, but not just those.

00:14:53.930 --> 00:14:56.530
That was part of the point I made in my talk. Transient dynamics,

00:14:56.770 --> 00:15:00.470
which is sort of what you're referring to, is also to be expected.

00:15:01.610 --> 00:15:06.070
All the focus on attractors, I think, in dynamical systems theory over the years

00:15:06.070 --> 00:15:11.330
has been a bit of a special case because we know that when you take dynamical

00:15:11.330 --> 00:15:14.410
systems and drive them with lots of signals from an external environment,

00:15:14.690 --> 00:15:16.930
you're almost never going to be in true attractors.

00:15:16.970 --> 00:15:19.530
You're almost always going to be in transient structures.

00:15:19.870 --> 00:15:23.690
So again, I don't see anything inconsistent there. Using a particular language

00:15:23.690 --> 00:15:29.310
suggests or makes it more natural or gives you the tools to answer certain kinds

00:15:29.310 --> 00:15:30.510
of questions and not other ones.

00:15:30.550 --> 00:15:35.550
So it's not at all surprising that the language of dynamical systems applied

00:15:35.550 --> 00:15:41.090
to the system you're describing would suggest those kinds of avenues.

00:15:41.450 --> 00:15:48.710
But what I'm saying is the statement that the brain is a dynamical system is

00:15:48.710 --> 00:15:51.370
not a very theoretically interesting statement.

00:15:51.730 --> 00:15:53.610
It's kind of like saying that.

00:15:55.770 --> 00:15:59.110
Gravitation is a differential system, right?

00:15:59.210 --> 00:16:04.290
That's not the sort of theory, quote-unquote, that's put forward in physics.

00:16:04.510 --> 00:16:09.830
But that's the analogous statement, I think. Rather, the statement that you

00:16:09.830 --> 00:16:15.710
see in physics is something like gravitation involves an inward-verse-square

00:16:15.710 --> 00:16:17.670
law of attraction, say, for Newton.

00:16:17.850 --> 00:16:25.090
Or gravitation is a curvature of space-time, whose particular form I can describe

00:16:25.090 --> 00:16:26.570
using differential equations.

00:16:26.910 --> 00:16:30.690
That's the analogy I'm trying to make. It's sort of at the wrong level to make

00:16:30.690 --> 00:16:33.150
these grand sweeping statements.

00:16:34.110 --> 00:16:36.890
Take, for instance, the statement that the brain is a dynamical system.

00:16:37.170 --> 00:16:43.030
So, people have made that statement in order to point out what it's not.

00:16:43.310 --> 00:16:46.610
And for people like Esther Tellem, for instance, she would make that statement

00:16:46.610 --> 00:16:48.350
very strongly because she was trying to.

00:16:49.230 --> 00:16:54.070
Counteract a view, for instance, that knowledge is iconic, that there are representations

00:16:54.070 --> 00:17:00.190
that are maybe pre-specified in some way and that come to life when you do cognition. And she would.

00:17:01.067 --> 00:17:03.947
Presented a very powerful argument i think to say that

00:17:03.947 --> 00:17:07.547
look these representations don't have to pre-exist that

00:17:07.547 --> 00:17:11.307
if you take a dynamic systems perspective then in

00:17:11.307 --> 00:17:13.967
the uh when you engage in the

00:17:13.967 --> 00:17:18.887
behavior that the mind generates these representations in an emergent way even

00:17:18.887 --> 00:17:21.747
you might not even use the word representation many people would object to that

00:17:21.747 --> 00:17:26.747
so really it's a very strong position about what the brain is not perhaps as

00:17:26.747 --> 00:17:31.967
much as about what the brain is and And powerful explanations can follow, for instance,

00:17:32.207 --> 00:17:36.207
about the development of behavior, how we come from a system that can do very

00:17:36.207 --> 00:17:39.807
little to a system that is capable of this behavioral complexity.

00:17:40.267 --> 00:17:45.847
So within psychology, certainly, I think that's been a powerful approach.

00:17:46.207 --> 00:17:50.767
Maybe we shouldn't call it a theory, but a powerful approach for refining our

00:17:50.767 --> 00:17:53.067
ideas to think about what cognition is.

00:17:53.287 --> 00:17:57.967
Yeah, and again, I'm just repeating myself, but the fact is when you apply different

00:17:57.967 --> 00:18:01.527
lenses is to a system, they're going to suggest different things because they

00:18:01.527 --> 00:18:05.487
bring different vocabularies to bear, they bring different perspectives to bear.

00:18:05.687 --> 00:18:08.667
And if you take something like the lens of dynamical systems theory,

00:18:08.787 --> 00:18:12.647
which is this is exactly what Esther did and applied it to a kind of problem

00:18:12.647 --> 00:18:15.807
that people had never thought of in those terms before, it's going to be very

00:18:15.807 --> 00:18:17.947
productive, which is not, again,

00:18:18.147 --> 00:18:23.247
addressing the question of testing or what it means to say the brain is a dynamical system.

00:18:23.247 --> 00:18:28.207
But there's sort of an intermediate issue then before we start to look at the

00:18:28.207 --> 00:18:29.787
specific test case that you analyzed.

00:18:31.267 --> 00:18:36.727
You cannot completely disconnect the lens you take and your ontological commitments

00:18:36.727 --> 00:18:39.467
because that lens, like if you take the lens of dynamical systems,

00:18:39.747 --> 00:18:43.767
it will bias the kind of data you will consider and what properties you consider.

00:18:44.007 --> 00:18:47.367
And in that sense, it will bias your interpretation of what these things really

00:18:47.367 --> 00:18:50.587
mean. That's all true except, I think, the statement about ontological commitments.

00:18:50.827 --> 00:18:55.987
Because if you take a lens provisionally, the way I'm suggesting,

00:18:56.187 --> 00:18:59.267
then you're not making ontological commitments, or at least you're not making

00:18:59.267 --> 00:19:00.787
absolute ontological commitments.

00:19:00.927 --> 00:19:05.047
You're basically saying, I think it might be useful to look at this system through

00:19:05.047 --> 00:19:06.507
the lens of dynamical systems theory.

00:19:06.647 --> 00:19:10.447
So for the purposes of doing that, I'm going to think of it as having this state

00:19:10.447 --> 00:19:12.887
space and so on and so forth. Right.

00:19:13.877 --> 00:19:16.817
If I'm doing that provisionally, if I'm just as likely to later pick up another

00:19:16.817 --> 00:19:19.117
lens, then I'm not making any fundamental commitment.

00:19:19.397 --> 00:19:23.377
Whereas some of these other things that you're suggesting are really fundamental commitments. Sure.

00:19:24.657 --> 00:19:29.477
But we cannot deny that there might be biases. Absolutely. Every lens has a

00:19:29.477 --> 00:19:31.137
bias. Every lens. This was the point.

00:19:31.377 --> 00:19:34.897
Which is why I think it's useful to emphasize that one should have a set of

00:19:34.897 --> 00:19:35.937
lenses in one's toolkit.

00:19:36.157 --> 00:19:38.877
Okay. And the second thing is, of course, you now say, look,

00:19:38.977 --> 00:19:44.697
I propose an approach where we in detail analyze toy systems,

00:19:44.817 --> 00:19:48.117
and this is the way forward because the toy systems are sort of more controllable.

00:19:49.297 --> 00:19:53.577
But of course, there's then this risk that if you choose your toy system incorrectly,

00:19:54.117 --> 00:19:57.697
that you're going to sort of dig the tunnel in the wrong direction, right?

00:19:57.757 --> 00:20:02.277
So how do you make sure that your toy systems has the constraints that help

00:20:02.277 --> 00:20:04.677
you to generalize towards a phenomenon you really want to capture?

00:20:04.677 --> 00:20:08.497
So basically any methodology has pluses and minuses.

00:20:08.517 --> 00:20:13.157
And in general, that is certainly a concern with the methodology that we're talking about.

00:20:13.297 --> 00:20:17.677
The way I'm trying to address it is, as I mentioned at the very end of my talk,

00:20:17.817 --> 00:20:21.877
I think we've honed some of these tools to the point where they're worth trying

00:20:21.877 --> 00:20:26.177
on a biological system, which is a very different sort of model that we're doing

00:20:26.177 --> 00:20:27.777
there. It's not a toy model.

00:20:28.257 --> 00:20:32.277
It's intended to be an empirically testable model. And so you have to engage

00:20:32.277 --> 00:20:35.357
in the prediction, experimental test, refinement kind of loop,

00:20:35.417 --> 00:20:38.257
which is not something that makes any sense for the toy models because they're

00:20:38.257 --> 00:20:39.857
not intended to making any predictions.

00:20:39.857 --> 00:20:45.637
And so, this is a big reason why I've been shifting to applying evolutionary

00:20:45.637 --> 00:20:51.357
algorithms and these dynamical and information theoretic analysis techniques to C.

00:20:51.437 --> 00:20:58.117
Elegans because I think it might be the kind of actual biological system where

00:20:58.117 --> 00:20:59.237
we can try these things out.

00:20:59.377 --> 00:21:01.997
Right. And it may fail in various ways.

00:21:02.197 --> 00:21:07.897
And so, that is ultimately the epistemological test, right? Are these useful or not?

00:21:08.077 --> 00:21:14.517
Exactly. But now, so the test case that you discussed was a population,

00:21:14.517 --> 00:21:16.257
if you want, of very simple agents.

00:21:17.235 --> 00:21:26.215
That might be static or active, and that we're supposed to detect a looming object.

00:21:28.695 --> 00:21:31.615
And you define this as a categorization task.

00:21:31.995 --> 00:21:34.815
Yeah, you can call it classification if you want to, but the point is it's not

00:21:34.815 --> 00:21:38.495
just detecting, it's actually making some discrimination about the relationship

00:21:38.495 --> 00:21:41.215
between the two objects it sees in sequence. Okay.

00:21:42.095 --> 00:21:46.575
So what's important here, there's, let's say, a simple environment that has

00:21:46.575 --> 00:21:48.635
certain properties, that's the looming stimulus, right?

00:21:49.055 --> 00:21:52.935
There's a simple embodied agent because it can move in space.

00:21:53.535 --> 00:22:00.215
And then there is a control system that, if you want, transforms properties

00:22:00.215 --> 00:22:03.535
of that stimulus into a reaction, if you want.

00:22:04.795 --> 00:22:09.535
And then the control system, this neural-like control system that you were in

00:22:09.535 --> 00:22:12.755
the end studying with either dynamical systems approaches or an information

00:22:12.755 --> 00:22:17.755
theoretic approach, you generated many, many, many exemplars of that using a

00:22:17.755 --> 00:22:19.035
genetic algorithm. Yes. Okay.

00:22:21.079 --> 00:22:24.179
So, then you compared, let's say, the static case versus the active case.

00:22:24.719 --> 00:22:28.339
And you also showed us in this analysis that if you analyzed,

00:22:28.419 --> 00:22:32.779
in particular, this neural controller of this agent, using either a dynamical

00:22:32.779 --> 00:22:35.199
systems perspective or an information theoretic perspective,

00:22:35.379 --> 00:22:38.559
you would have a very complementary dissection, if you want,

00:22:38.759 --> 00:22:40.999
of the functional properties of this agent.

00:22:42.839 --> 00:22:51.219
So, what does the dynamical systems lens on that system now exactly tell us? What do we learn from it?

00:22:52.039 --> 00:22:57.599
Well, so for example, one of the things that doesn't come up at all in the information

00:22:57.599 --> 00:23:01.979
theoretic analysis is how important the bifurcation idea is,

00:23:02.059 --> 00:23:05.039
that there's this discontinuous change in the system's response properties.

00:23:05.159 --> 00:23:07.219
That's just not a notion in information theory.

00:23:08.039 --> 00:23:12.419
Another one has to do with the role of the discontinuities that the sensors introduce.

00:23:12.639 --> 00:23:15.599
I mean, these again are not notions in information theory. You're just seeing

00:23:15.599 --> 00:23:20.399
these generalized correlations being developed and distributed around over time.

00:23:20.979 --> 00:23:27.399
On the other hand, one thing that's not so obvious from the dynamical systems

00:23:27.399 --> 00:23:31.519
analysis alone, because it deals with sort of the full state space of dynamics,

00:23:31.679 --> 00:23:36.119
is which particular combinations of elements are the most appropriate ones to

00:23:36.119 --> 00:23:39.719
be focusing on at any given point in time, which is something that the information

00:23:39.719 --> 00:23:43.579
theoretic analysis, I think, brought out much more cleanly than the dynamical

00:23:43.579 --> 00:23:45.459
analysis does. But as you say, they're complementary.

00:23:45.959 --> 00:23:52.239
So it's not like either is inconsistent with what the other one says,

00:23:52.479 --> 00:23:54.739
but they bring out different features of the system.

00:23:54.959 --> 00:23:58.719
And that's why I think it's really interesting to look for bridges between these

00:23:58.719 --> 00:24:02.359
different stories, which was the final part I wasn't able to get to in the talk. Okay.

00:24:02.799 --> 00:24:06.479
But now the variation across these populations that you analyzed,

00:24:06.619 --> 00:24:09.999
because in the end you actually analyzed one exemplar, I think,

00:24:10.019 --> 00:24:10.979
from this whole population.

00:24:11.139 --> 00:24:15.879
Yeah, we actually looked at about 10 of each, but only one in this great detail. Right. Okay.

00:24:17.542 --> 00:24:22.042
You generate these with genetic algorithms. Is that a key ingredient of this

00:24:22.042 --> 00:24:25.262
experiment, or it could have been any way in which you can generate a variable

00:24:25.262 --> 00:24:27.242
population? I think the latter, yeah.

00:24:27.362 --> 00:24:30.802
I think any stochastic optimization technique would have been fine,

00:24:30.862 --> 00:24:32.542
the way I'm using genetic algorithms.

00:24:32.942 --> 00:24:38.762
But I do think the stochastic feature is important, because otherwise you end up biasing things.

00:24:38.822 --> 00:24:43.262
So you tend to fall into the same architecture each time, and you're not really

00:24:43.262 --> 00:24:46.082
exploring the space of the possibilities in the same way. That's,

00:24:46.082 --> 00:24:47.322
I think, the important feature. Right.

00:24:47.542 --> 00:24:52.742
But then the fixed feature was the number of sensors that you had at the surface

00:24:52.742 --> 00:24:55.142
of this agent to detect the looming stimulus.

00:24:56.082 --> 00:25:01.062
Then you had three hidden units that would sort of transform the sensor state

00:25:01.062 --> 00:25:05.622
into a motor state, which was encoded by two units that could,

00:25:05.722 --> 00:25:08.102
for the active case, drive you to the left or to the right.

00:25:08.182 --> 00:25:11.502
Well, even in the passive case, they drive you. That's why I don't really like

00:25:11.502 --> 00:25:17.742
the static versus... Okay. In both cases, in the end, they have to move to express their decision.

00:25:17.882 --> 00:25:21.382
The difference is just whether or not they move only when the second object

00:25:21.382 --> 00:25:25.462
is falling or whether they move during the time that both objects are falling.

00:25:25.562 --> 00:25:27.622
Anyway, just a minor correction. No, no, this is good.

00:25:27.862 --> 00:25:33.582
So then the task itself was indeed, if you want, you had a training case and a test case, right?

00:25:34.522 --> 00:25:38.482
There was a probe trial second, and then there was an exposure trial.

00:25:38.802 --> 00:25:42.582
Yeah, so there's a cue and then a probe. Right. And then between those,

00:25:42.922 --> 00:25:47.462
the system could, in this network structure, maintain some sort of memory if

00:25:47.462 --> 00:25:49.382
its dynamics would allow that. Yes.

00:25:49.862 --> 00:25:56.262
Now, how many different network templates did you find that would give rise

00:25:56.262 --> 00:25:58.742
to the same kind of dynamical landscape?

00:25:59.162 --> 00:26:00.722
Well, so it depends what you mean by the same.

00:26:01.902 --> 00:26:06.282
You're going to tell me. Every single run comes up with different parameters.

00:26:07.142 --> 00:26:10.802
So the problem is interestingly rich enough that it's not like there's just

00:26:10.802 --> 00:26:14.062
a unique solution. And that's, I think, very, very typical.

00:26:18.109 --> 00:26:22.749
So the details of every, say, dynamical analysis or every informational analysis would be different.

00:26:23.089 --> 00:26:29.709
But the general idea of having some transient manifold evolving through the

00:26:29.709 --> 00:26:33.329
state space of the entire system, not just the inner neurons,

00:26:33.429 --> 00:26:37.709
but that's what I focused on for the passive agent, applies across the board.

00:26:38.729 --> 00:26:44.689
The geometry of that particular manifold at the end of the Q stage is very interesting.

00:26:44.689 --> 00:26:53.069
And also which dimensions of the state space that manifold actually transits.

00:26:53.209 --> 00:26:59.109
So if you think about it, in the passive agent, it's basically the interneuronal

00:26:59.109 --> 00:27:00.269
state space that's important.

00:27:00.549 --> 00:27:04.489
And as you pointed out, some of them are more important than others for that particular agent.

00:27:05.109 --> 00:27:09.289
In the active agents, one way of thinking about it is, I didn't show a dynamical

00:27:09.289 --> 00:27:11.289
analysis of an active agent, but you could do it.

00:27:11.289 --> 00:27:16.129
But it's not the extent of that manifold in the internal state space,

00:27:16.429 --> 00:27:21.269
it's actually its extent in the body position part of the state space,

00:27:21.389 --> 00:27:24.329
which after all is just part of the total state space of the system.

00:27:24.569 --> 00:27:29.749
So those kinds of considerations and these issues about how does the pattern

00:27:29.749 --> 00:27:34.329
of information, how does the information flow through the system and so on.

00:27:34.329 --> 00:27:40.589
What I was after, Randy, is to say, look, you generate let's say a thousand exemplars. Right.

00:27:40.709 --> 00:27:44.449
I don't know how big the pool really was. Was it 1,000? No, it's smaller than that. Okay, whatever.

00:27:44.649 --> 00:27:48.569
Dozens. Okay. But then I would expect, but maybe I'm wrong here,

00:27:48.869 --> 00:27:55.789
that at least if you would analyze from a dynamical systems perspective,

00:27:56.069 --> 00:28:02.409
the whole population of exemplars you have, you would expect some prototypical

00:28:02.409 --> 00:28:05.429
forms of dynamics to emerge.

00:28:05.429 --> 00:28:10.909
And you characterized some of that, let's say the manifolds, but that to me.

00:28:11.574 --> 00:28:14.734
Would more suggest, okay, where do we look for these invariants?

00:28:15.154 --> 00:28:18.914
So can you be more specific about the invariant patterns you might see?

00:28:19.014 --> 00:28:21.774
Like for instance, in the example, you showed that we indeed saw one neuron

00:28:21.774 --> 00:28:24.494
doing most of the job. I mean, it's my characterization.

00:28:24.934 --> 00:28:30.614
Was that like a pattern emerging in, let's say, 30% of your exemplars?

00:28:30.854 --> 00:28:35.074
I don't have the statistics for it. But I mean, in general, there's no reason

00:28:35.074 --> 00:28:37.014
that it would be one neuron over the other.

00:28:37.054 --> 00:28:39.714
There's no reason in general why it couldn't be multiple neurons.

00:28:39.714 --> 00:28:41.814
And you do see examples of all of that.

00:28:41.834 --> 00:28:46.114
Why I'm asking you this and why this is important to me is that if we want to

00:28:46.114 --> 00:28:52.094
understand biological systems or brains or what have you, should we focus on

00:28:52.094 --> 00:28:54.754
the invariant patterns across exemplars?

00:28:54.814 --> 00:28:59.594
This is now a guiding principle like the retina projects to the thalamus.

00:28:59.694 --> 00:29:01.494
Okay. And we see that in all of them.

00:29:01.714 --> 00:29:05.174
So this is, of course, it's a completely different kind of description.

00:29:05.174 --> 00:29:11.194
So the question is, in this dynamical picture, if we start to complexify and

00:29:11.194 --> 00:29:15.174
say, oh no, every individual is very individual, there are no general features.

00:29:15.694 --> 00:29:18.894
Then it might not give us a lot of leverage to understand how these things work.

00:29:18.894 --> 00:29:20.114
I understand exactly what you're saying.

00:29:20.214 --> 00:29:23.874
And that's why I'm saying that if you're looking for generalities in these agents,

00:29:24.074 --> 00:29:29.154
you need to look at the level of these transient manifolds and the way in which they're transformed.

00:29:29.454 --> 00:29:31.414
You need to look at the way they're split.

00:29:31.754 --> 00:29:34.374
That's an important feature that you show up over and over again.

00:29:34.494 --> 00:29:39.854
When you drop the second agent, this curve of states, when you drop the second object,

00:29:40.094 --> 00:29:45.074
the curve of states gets spread into a sheet of states and then bifurcation

00:29:45.074 --> 00:29:47.234
slice that sheet into a decision.

00:29:47.234 --> 00:29:49.994
That is true across the board. Okay.

00:29:50.334 --> 00:29:54.974
So one thing that I'm not clear about is that you seem to be doing two things

00:29:54.974 --> 00:29:56.194
here with these toy systems.

00:29:56.334 --> 00:30:02.174
One is to, first of all, refine and sharpen your lenses and say,

00:30:02.254 --> 00:30:04.834
how are these lenses useful for understanding these model systems?

00:30:05.654 --> 00:30:09.694
And the second thing is maybe going towards a,

00:30:11.304 --> 00:30:15.444
a set of theories about these kinds of systems, and ones that you might build

00:30:15.444 --> 00:30:20.704
out of these, where perhaps you might start to develop a kind of taxonomy of

00:30:20.704 --> 00:30:25.484
simple machines, or machines with these kinds of simple neural circuits.

00:30:26.304 --> 00:30:27.884
I mean, is that right? Is that a goal?

00:30:29.764 --> 00:30:33.844
I don't know about taxonomy specifically, but… Well, a theory about these kinds

00:30:33.844 --> 00:30:35.644
of machines… Yes, that's what I'm trying to get at.

00:30:35.764 --> 00:30:40.644
And again, the point is that we're going to be testing these same ideas in the

00:30:40.644 --> 00:30:42.264
C. elegans models that we're building now.

00:30:42.384 --> 00:30:46.384
So I mean, the idea we didn't, so you're right, part of it has to do with tool development.

00:30:46.524 --> 00:30:49.944
Some of the information theoretic tools we're using didn't exist before we started

00:30:49.944 --> 00:30:52.664
trying to do an information theoretic analysis of these agents.

00:30:52.884 --> 00:30:55.524
Some of them did, some of them were developed specifically to do that.

00:30:55.744 --> 00:31:00.784
Now that we have the tools, we can apply them to other systems like the C. elegans system.

00:31:01.024 --> 00:31:04.464
So part of it is tool development, but also part of it is that as you look through

00:31:04.464 --> 00:31:11.884
these different lenses lenses and start to formulate explanations of what you see through that lens,

00:31:12.024 --> 00:31:17.004
those become tentative at least frameworks for trying to actually build theories.

00:31:17.304 --> 00:31:20.144
So the thing I mentioned, if the information theory is there's now,

00:31:20.264 --> 00:31:23.224
we now have this notion of sort of an information flow architecture.

00:31:23.584 --> 00:31:26.104
Okay, which transcends the detailed

00:31:26.104 --> 00:31:31.764
elements and the actual quantitative information about there's .37,

00:31:32.024 --> 00:31:36.044
you know, general mutual information between these two elements,

00:31:36.244 --> 00:31:40.464
but it leads us to focus on certain kinds of what's the overall pattern of flow through the system?

00:31:40.584 --> 00:31:42.624
What are the appropriate variables

00:31:42.624 --> 00:31:46.044
to look at? What are the appropriate informational quantities and so on?

00:31:46.144 --> 00:31:49.764
That's something we can apply to other systems like the C. elegans system.

00:31:49.944 --> 00:31:54.424
Is that something that's going towards a reduced description of the machine? Possibly.

00:31:54.644 --> 00:31:59.904
I think it's an idea at this point. I mean, and we're going to be testing it,

00:31:59.924 --> 00:32:01.324
as I said, in the C. elegans system.

00:32:02.113 --> 00:32:04.393
But now there's something interesting about the C. elegans example,

00:32:04.613 --> 00:32:09.773
that C. elegans, as you also pointed out yourself, we have 302 identified neurons.

00:32:10.213 --> 00:32:13.453
They are connected in a rather heterogeneous way.

00:32:14.473 --> 00:32:17.273
As are most nervous systems, actually, when you look at them.

00:32:17.273 --> 00:32:20.793
No, but I mean, what I mean with that is that they don't really implement any

00:32:20.793 --> 00:32:22.293
kind of parallelism, really.

00:32:22.413 --> 00:32:24.533
It's more… I have no idea what you mean by that.

00:32:24.853 --> 00:32:29.613
It's actually a highly recurrent network. No, but what I mean with that is not

00:32:29.613 --> 00:32:35.813
that we have, let's say, basic circuit templates that are replicated in a parallel fashion, right?

00:32:35.873 --> 00:32:38.933
It's a very complicated thing. Yeah, there are no cortices in C.

00:32:38.973 --> 00:32:40.013
Elegans. For instance, right?

00:32:40.253 --> 00:32:43.353
But neither is there a cerebellum or basal ganglia and so on, right?

00:32:43.473 --> 00:32:48.653
So you have a bunch of cells that are complex in themselves and that have,

00:32:48.813 --> 00:32:51.893
let's say, heterogeneous asymmetric kinds of interactions that we don't fully

00:32:51.893 --> 00:32:54.573
understand, that we want to understand. Yeah, I mean, there's a lot of symmetry too.

00:32:54.653 --> 00:32:57.213
There's bilateral symmetry, for example. Sure, absolutely.

00:32:57.833 --> 00:33:04.793
But what I'm driving at is to say that already at this sort of topological nature.

00:33:05.333 --> 00:33:09.573
The C. elegans nervous system might be rather different from the agent you have been using.

00:33:09.813 --> 00:33:13.033
And it sort of illustrates this point I made earlier. Like if you go for a toy

00:33:13.033 --> 00:33:17.813
system, it should, of course, be a segue into this natural system you want to understand.

00:33:17.813 --> 00:33:21.113
That and now i could say well the agent you studied definitely

00:33:21.113 --> 00:33:24.133
has a parallelism in this in its organization

00:33:24.133 --> 00:33:29.593
that you have a bunch of of uniform receptors that only vary in their placement

00:33:29.593 --> 00:33:32.933
on the periphery the same holds for the internet three interneurons they're

00:33:32.933 --> 00:33:36.593
in some sense identical but they vary in how they're interconnected whatever

00:33:36.593 --> 00:33:39.793
but there's a symmetry there again and the same for these output units so this

00:33:39.793 --> 00:33:43.253
kind of parallelism and symmetry you will not find in the C. Elegans brain.

00:33:43.613 --> 00:33:48.493
So has it then made it right? So why would I therefore believe that the lens

00:33:48.493 --> 00:33:55.213
you have sort of sharpened on this artificial agent would help you to get access to the C.

00:33:55.213 --> 00:33:59.313
Elegans system? You don't need to believe it because we're trying it. Sure, I know.

00:33:59.613 --> 00:34:06.833
The result of that will either support or not the idea.

00:34:07.053 --> 00:34:11.833
I mean, it's not something one has a belief in, except maybe in whether or not

00:34:11.833 --> 00:34:13.893
one spends time pursuing this direction.

00:34:14.153 --> 00:34:17.113
But in the end, we're going to find out whether some of the things I just described

00:34:17.113 --> 00:34:19.393
to you in our dynamical and information

00:34:19.393 --> 00:34:24.153
analysis of the relational categorization agent carries over to C.

00:34:24.213 --> 00:34:26.393
Elegans. So what have you found so far?

00:34:32.415 --> 00:34:37.635
The only thing I can say right now is we've done an information theoretic analysis,

00:34:38.415 --> 00:34:41.235
which isn't published yet, on the C.

00:34:41.275 --> 00:34:46.155
Elegans circuits and the tools that we developed and the ideas and especially

00:34:46.155 --> 00:34:49.555
this notion of information flow architecture turned out to be extremely powerful

00:34:49.555 --> 00:34:52.555
in characterizing what's going on.

00:34:52.695 --> 00:34:54.295
Because even in C. elegans, so

00:34:54.295 --> 00:34:56.895
even in biological systems, you have a tremendous amount of variability.

00:34:58.075 --> 00:35:01.535
There are a number of famous examples of this which I could get into,

00:35:01.635 --> 00:35:03.115
but I'm not sure we want to spend the time.

00:35:03.675 --> 00:35:08.695
There's every reason to expect that the variability that you see in these toy

00:35:08.695 --> 00:35:13.435
models using evolutionary algorithms is in fact a desirable feature of them

00:35:13.435 --> 00:35:17.475
because it's actually more reflective of what you see in biology than this idea

00:35:17.475 --> 00:35:20.675
that there's the one true brain or the one true whatever.

00:35:20.935 --> 00:35:25.215
But it's interesting, right? Because traditionally, I think initially you worked

00:35:25.215 --> 00:35:26.815
a lot on a dynamical systems perspective.

00:35:26.915 --> 00:35:29.955
Yes, I did. And I think the information theoretic one is a bit more recent as

00:35:29.955 --> 00:35:31.495
far as I understand. Yes, oh, absolutely it is, yeah.

00:35:31.595 --> 00:35:34.995
But it's interesting now with C. elegans, your first success,

00:35:35.115 --> 00:35:37.095
if you want, has been with the information theoretic measure.

00:35:37.295 --> 00:35:40.875
Well, that's not entirely true. I mean, so we published about a year and a half

00:35:40.875 --> 00:35:46.335
ago our first paper on this model, and that was basically much more dynamical than it is now.

00:35:46.335 --> 00:35:47.935
On the C. elegans model? On the C. elegans model, yeah. Yeah,

00:35:48.015 --> 00:35:51.255
it's the more recent work that we're doing that's looked at the information

00:35:51.255 --> 00:35:54.955
flow architecture idea in the context of C-elegance.

00:35:54.995 --> 00:35:59.255
What's interesting with the C-elegance is that we do know the full connectivity of the circuit.

00:35:59.475 --> 00:36:03.855
Yes, but we know almost nothing about the biophysics and neurophysiology underlying it. Well, exactly.

00:36:04.335 --> 00:36:10.055
But we can, with this relatively simple model, say, okay, given the connectivity,

00:36:10.515 --> 00:36:13.575
how much more can we infer via these approaches?

00:36:13.755 --> 00:36:18.135
So it's a good model system in which to do that. and then we can think,

00:36:18.175 --> 00:36:23.115
well, if we knew the full connectome of the human brain, what could we learn from that?

00:36:23.355 --> 00:36:28.255
So in that way, that's part of your strategy. Oh, absolutely.

00:36:28.455 --> 00:36:31.415
Yeah. I mean, one of the things that we looked at in this previous paper that's

00:36:31.415 --> 00:36:32.475
been published for a while.

00:36:34.170 --> 00:36:41.250
There's tremendous variability when we evolved 100 chemotaxis models for C.

00:36:41.290 --> 00:36:46.270
Elegans, and there's tremendous variability in the parameters that you get,

00:36:46.370 --> 00:36:49.770
even though behaviorally they're almost identical in terms of their… So again,

00:36:49.830 --> 00:36:53.750
it just shows you that this happens even with those constraints.

00:36:54.010 --> 00:36:59.190
They're just not sufficient constraints. But what's intriguing in this work

00:36:59.190 --> 00:37:04.250
that's currently in press is that it looks like if you, when you look at the

00:37:04.250 --> 00:37:06.510
information architecture of these,

00:37:06.750 --> 00:37:11.050
I think about 70 some of them turn out to be really high performing of the 100.

00:37:11.230 --> 00:37:14.450
So some fail and some don't do very, you know, they don't fail,

00:37:14.590 --> 00:37:16.210
but they don't do as well as possible.

00:37:16.450 --> 00:37:22.530
So maybe the subset of 70 of them all share almost identical information architectures,

00:37:22.570 --> 00:37:25.530
even though they have tremendous variability at the parametric level.

00:37:25.890 --> 00:37:29.550
So when you're looking at C. elegans, you have the connectivity,

00:37:29.790 --> 00:37:30.970
but you have all these other gaps.

00:37:31.170 --> 00:37:36.690
Then the methodology that you have with your toy models, which is not exhaustively,

00:37:36.790 --> 00:37:41.050
but to fairly comprehensively explore the space of possible models. That's exactly right.

00:37:41.890 --> 00:37:45.170
You can't do that anymore. So you can't explore the whole of the design space,

00:37:45.410 --> 00:37:51.250
can you? Well, I mean, you can never explore the whole, but that's exactly the,

00:37:51.290 --> 00:37:54.550
I guess I don't understand the question, that's exactly the strategy is that

00:37:54.550 --> 00:37:56.550
we take what's known about C.

00:37:56.610 --> 00:38:02.170
Elegans, in our case, namely the connectivity and the behavior and some physiology

00:38:02.170 --> 00:38:05.710
to do with the sensors, actually. That's actually been pretty well worked out.

00:38:05.790 --> 00:38:08.530
And we constrain our evolutionary algorithm by that information,

00:38:08.610 --> 00:38:11.530
and we use the evolutionary algorithm to fill in the remaining details,

00:38:11.750 --> 00:38:15.950
namely the sort of neurophysiological parameters, which are unknown.

00:38:16.490 --> 00:38:20.090
And if you do that once, it's not very interesting because that's just a solution.

00:38:20.310 --> 00:38:23.610
But if you do it many, many times, then you start to be able to talk about the

00:38:23.610 --> 00:38:26.650
ensemble of solutions, and that has interesting structure in it.

00:38:26.790 --> 00:38:30.290
But, I mean, the number of unknowns, I imagine, is quite high.

00:38:30.450 --> 00:38:36.810
Yes. And even with your genetic algorithms, you cannot exhaustively explore

00:38:36.810 --> 00:38:38.490
that space. No, no. Again, exhaustive is not possible.

00:38:38.590 --> 00:38:43.010
So you're sampling a larger space. You're sampling a space and you're looking for patterns in that.

00:38:43.110 --> 00:38:46.930
So if every single individual is completely different at all levels of analysis,

00:38:47.270 --> 00:38:48.870
then you've learned nothing, right?

00:38:48.990 --> 00:38:53.270
But what's interesting is they can vary quite a bit at the level of the individual

00:38:53.270 --> 00:38:57.090
neurophysiological parameters, and yet you start to see patterns when you look,

00:38:57.110 --> 00:39:00.350
say, dynamically or information theoretically across the ensemble.

00:39:00.850 --> 00:39:02.770
Okay, but the question, I guess, is that...

00:39:04.175 --> 00:39:10.835
With C. elegans, it's a test case for your methodology. So how big can the space

00:39:10.835 --> 00:39:16.075
of unknowns be before this strategy of trying to explore it with GAs is going to break down?

00:39:16.175 --> 00:39:19.275
Yeah. You won't see the very big. I mean, I can't give you a magic number.

00:39:19.375 --> 00:39:22.995
Right. In part, that has to do with how much computation you can throw at it,

00:39:23.035 --> 00:39:25.595
which particular optimization techniques you use.

00:39:25.735 --> 00:39:28.035
And it has, of course, a lot to do with the system you're studying,

00:39:28.135 --> 00:39:33.315
which a priori we don't know how much structure or not might be in that system.

00:39:33.315 --> 00:39:37.235
But then also, with your analysis tools, you're maybe going to want to automate

00:39:37.235 --> 00:39:40.255
some of the analysis beyond sort of… Possibly.

00:39:40.395 --> 00:39:44.315
I mean, I think… So the biggest systems we've tended to do in our toy models

00:39:44.315 --> 00:39:46.175
have been sort of 30 neurons or less.

00:39:46.275 --> 00:39:49.195
And so I think C. elegans is actually a good next step.

00:39:49.315 --> 00:39:52.295
We're talking about an order of magnitude if you think about the entire nervous

00:39:52.295 --> 00:39:54.635
system rather than just the bits we're looking at so far.

00:39:54.795 --> 00:39:59.335
So it strikes me as the right scale to be going from 30 to 300 rather than,

00:39:59.335 --> 00:40:02.415
say, 30 to 300 billion or something.

00:40:02.415 --> 00:40:08.135
But have you learned from your toy models ways of doing analyses with dynamic

00:40:08.135 --> 00:40:11.155
systems and information theory that you could say automate it?

00:40:11.295 --> 00:40:15.095
Well, I mean, so many, many years ago I built a system called Dynamica that

00:40:15.095 --> 00:40:18.815
basically let you automate some aspects of dynamical analysis,

00:40:18.915 --> 00:40:19.995
and we've always used that.

00:40:20.675 --> 00:40:24.975
We don't yet have such a system for information theoretic analysis,

00:40:25.135 --> 00:40:28.895
but certainly as you gain experience with it, it's easier and easier to write

00:40:28.895 --> 00:40:32.035
code that sort of modularizes bits and pieces of it.

00:40:32.035 --> 00:40:36.175
And then you can sort of throw that at a big complicated system and just let

00:40:36.175 --> 00:40:38.595
it grind away for a few days and give you the results.

00:40:38.715 --> 00:40:41.455
So I think that's part of building the tools, but it's an ongoing process.

00:40:41.695 --> 00:40:43.095
But part of the tools could be

00:40:43.095 --> 00:40:46.055
to say, look, this is the sort of thing I'm looking for from the analysis.

00:40:46.375 --> 00:40:48.855
And then, you know, there's sort of a meta-analysis.

00:40:49.555 --> 00:40:53.515
Yeah, I mean, ideally that would be the case. I think for the most part.

00:40:54.642 --> 00:40:58.642
There's no sort of magic wand to analysis. It's a creative activity.

00:40:58.862 --> 00:41:04.262
And so at least present tends to involve people engaging with a set of tools

00:41:04.262 --> 00:41:09.782
with some system and using their own intuitions and thought processes to sort

00:41:09.782 --> 00:41:12.642
of guide whether or not you could ultimately automate the whole thing.

00:41:12.762 --> 00:41:19.722
I have no idea, possibly. But now for the C. elegans, could you give us an idea of the unknowns?

00:41:19.842 --> 00:41:24.502
That means of the key parameters that you think you have to have a handle on

00:41:24.502 --> 00:41:25.622
to understand that system.

00:41:25.942 --> 00:41:30.042
How many are actually identified and how many are actually partially known and

00:41:30.042 --> 00:41:31.342
how many are completely missing?

00:41:31.662 --> 00:41:34.262
That's not entirely a well-defined question. So here's what we know.

00:41:34.342 --> 00:41:39.502
There are about 8,000 connections among those 302 neurons.

00:41:40.962 --> 00:41:43.162
We know very little about the

00:41:43.162 --> 00:41:46.442
neurophysiology, actually electrical properties of each of the neurons.

00:41:46.622 --> 00:41:49.082
One of the things that we know is that they don't seem to spike.

00:41:50.082 --> 00:41:54.282
So the kind of model neurons that we've been using I think are actually fair

00:41:54.282 --> 00:41:55.762
representations of them.

00:41:56.142 --> 00:42:00.182
There are, however, nonlinear response characteristics to the neurons,

00:42:00.422 --> 00:42:03.442
most of which have never been characterized except at the sensory level.

00:42:04.142 --> 00:42:07.742
We don't know the signs of, as far as I know, any of the connections,

00:42:08.022 --> 00:42:10.902
let alone the magnitudes.

00:42:11.791 --> 00:42:15.431
So we're mostly talking about, so I guess this is what I'm saying.

00:42:15.511 --> 00:42:19.631
If you fix a neural model like the one we're using, then I can calculate a number

00:42:19.631 --> 00:42:21.011
of parameters we're talking about.

00:42:21.151 --> 00:42:25.271
But you can't really fix a neural model either because to use the model like

00:42:25.271 --> 00:42:27.231
we're using is based on the knowledge that C.

00:42:27.251 --> 00:42:31.011
Elegans neurons don't spike, that they have certain sort of nonlinear characteristics

00:42:31.011 --> 00:42:33.391
in synaptic transmission, and so on.

00:42:33.631 --> 00:42:37.011
But those details are likely to become

00:42:37.011 --> 00:42:40.971
richer as you actually start to do neurophysiology on the individual cell.

00:42:40.971 --> 00:42:46.691
So, like I said, we know there are current channels, active voltage-gated and

00:42:46.691 --> 00:42:49.131
chemically-gated current channels in C. elegans neurons.

00:42:49.311 --> 00:42:51.691
So, they're not just passive membrane. Okay.

00:42:52.771 --> 00:42:56.491
Once you start doing voltage clamp-like dissections of those things,

00:42:56.611 --> 00:42:59.951
you're going to be building more complicated models of the neurons and the number

00:42:59.951 --> 00:43:01.151
of parameters would be going up.

00:43:01.211 --> 00:43:04.251
So, it will be a more incremental approach, you're saying? It's always got to be incremental.

00:43:04.391 --> 00:43:07.351
I mean, even in something that's quote-unquote as simple as C.

00:43:07.351 --> 00:43:11.231
Elegans is still quite complicated. And you've got to make a cut somewhere.

00:43:11.411 --> 00:43:17.131
We're doing the same thing with the body. I mean, you know, as this last talk

00:43:17.131 --> 00:43:21.371
pointed out, if you just look at the somatic cells in CL, and it's not the germline

00:43:21.371 --> 00:43:25.751
cells, there are about 1,000, just under 1,000 cells in the entire body.

00:43:26.151 --> 00:43:30.191
Okay. We're not modeling at that level at all. So we, for example,

00:43:30.251 --> 00:43:33.711
model muscle at a much higher level than the individual muscle cells.

00:43:34.411 --> 00:43:38.191
Okay. So you're always making cuts, and they're always provisional because as

00:43:38.191 --> 00:43:43.111
your model progresses and as it engages with experiments, it's going to need to be refined.

00:43:43.491 --> 00:43:48.611
Right, but it's also to follow up on this whole question of how many of the

00:43:48.611 --> 00:43:50.851
gaps can you fill in with the genetic algorithms, right?

00:43:51.031 --> 00:43:53.591
Right, right. For instance, also with these kinds of brains,

00:43:53.851 --> 00:43:57.211
you have to think about the kinds of neurotransmission they use,

00:43:57.371 --> 00:44:01.651
which is also to learn with all sorts of strange peptides we barely understand and so on.

00:44:01.811 --> 00:44:07.211
So how do you get a handle on that? Well, Ian, you get a handle on it by not

00:44:07.211 --> 00:44:11.791
wringing one's hands about how complicated life is, but diving in.

00:44:11.931 --> 00:44:13.471
Right. You have to make some cuts.

00:44:13.651 --> 00:44:17.851
And this is why I'm a big advocate of models being involved very,

00:44:17.951 --> 00:44:19.471
very early in the process.

00:44:19.511 --> 00:44:22.891
As I mentioned in my talk, I really don't like the idea that we don't have enough

00:44:22.891 --> 00:44:24.311
data yet to start to model.

00:44:24.471 --> 00:44:30.451
I think it's much better to jump in with a model and don't be overly enamored of your model.

00:44:30.511 --> 00:44:33.831
Models are almost always wrong. In fact, they're basically always wrong,

00:44:33.931 --> 00:44:38.811
but some models are wrong in interesting ways that focus you on what additional

00:44:38.811 --> 00:44:41.131
things you might need to know and so on.

00:44:41.231 --> 00:44:44.431
And I think that's what's important about it. But going back to your question,

00:44:44.551 --> 00:44:46.571
I mean, I don't know how to put a number on it.

00:44:46.631 --> 00:44:49.631
Certainly, that's a problem with optimization techniques.

00:44:49.791 --> 00:44:55.071
You can't just have 8,000, let alone, say, 800,000 or something free parameters

00:44:55.071 --> 00:44:59.571
and expect to get any kind of a useful sample running it a few thousand times

00:44:59.571 --> 00:45:01.711
or something like that. But, um...

00:45:03.249 --> 00:45:06.309
But rather than try to solve that problem in general, like I said,

00:45:06.329 --> 00:45:08.629
we jump in with a particular level of abstraction,

00:45:08.889 --> 00:45:12.449
and we put a stake down, and then we can start looking at making predictions,

00:45:12.589 --> 00:45:15.829
some of which may be verified, many of which, of course, are going to be knocked

00:45:15.829 --> 00:45:19.609
down, and use that to start refining. Where do we need to put the effort?

00:45:19.749 --> 00:45:22.869
What things can we assume? And the other thing is data is always coming along.

00:45:23.349 --> 00:45:26.749
There are new techniques now that are going to make neurophysiology and C.

00:45:26.769 --> 00:45:31.529
Elegans easier than it's been for decades. And that's going to start providing

00:45:31.529 --> 00:45:34.949
information that we don't have to use a genetic algorithm to fill in anymore. more.

00:45:35.009 --> 00:45:39.369
I guess I'm just worried on the scaling problem that in your toy model system

00:45:39.369 --> 00:45:44.929
you say well okay let's evolve these systems that can match the behavioral criteria

00:45:44.929 --> 00:45:49.169
and then we'll apply information theory dynamic systems theory to look at them

00:45:49.169 --> 00:45:53.869
and you then really focus in on I think you said 10 models and one in real detail.

00:45:54.129 --> 00:46:00.529
Now with C. elegans maybe you have the same constraint from behavior but maybe

00:46:00.529 --> 00:46:03.529
there'll be such a huge number of models that can potentially fit that,

00:46:03.649 --> 00:46:06.709
and then when you try to apply your lenses to that, how are you going to...

00:46:06.709 --> 00:46:10.429
So what we've done in C. elegans is exactly the same as what we did here.

00:46:10.569 --> 00:46:16.029
So we look at the population, we study in detail, usually the best couple,

00:46:16.269 --> 00:46:20.529
and then we start asking how much of what we learned in those in-depth studies

00:46:20.529 --> 00:46:22.429
generalize to a broader set.

00:46:23.071 --> 00:46:26.991
And what you start finding is that many of the details of what you discovered

00:46:26.991 --> 00:46:31.231
in that analysis don't generalize, but there's some level of description that

00:46:31.231 --> 00:46:36.651
resulted from your analysis, which starts to, you see, recurring over multiple additional ones.

00:46:36.831 --> 00:46:39.911
And so, I mean, that's the only way I know how to do it. Without automated techniques,

00:46:40.111 --> 00:46:42.291
you can't do the whole population, right?

00:46:42.651 --> 00:46:46.651
And you have to analyze some in detail because a priori, you don't know where

00:46:46.651 --> 00:46:48.751
the right line is, where the cut is.

00:46:48.891 --> 00:46:51.811
So you've got to go through one in detail or a few in detail and then you start.

00:46:52.551 --> 00:46:56.231
In some sense, you do internal hypothesis testing just with the model itself.

00:46:56.711 --> 00:46:58.711
Well, okay, looking at that detail,

00:46:58.951 --> 00:47:02.951
I would guess that that connection is always going to be inhibitory.

00:47:03.091 --> 00:47:06.731
Oops, here's three examples that that's not true. Okay, that's too low a level.

00:47:06.831 --> 00:47:09.131
Let's try a slightly higher level of description.

00:47:09.451 --> 00:47:12.411
And so you start doing some testing there until you find, oh,

00:47:12.451 --> 00:47:16.091
at this level of description that I can make, it actually seems to start to

00:47:16.091 --> 00:47:19.031
generalize across a significant fraction of the population.

00:47:19.491 --> 00:47:23.371
So now you also introduced us to a number of information theoretic measures

00:47:23.371 --> 00:47:29.511
that you used as a complement to the dynamical systems lens on these synthetic agents.

00:47:29.771 --> 00:47:35.131
So they're the obvious mutual information measures, but you also introduced a few new ones.

00:47:35.551 --> 00:47:39.331
So what are these measures and why did you feel you had to introduce them?

00:47:39.671 --> 00:47:44.631
Well, I mean, the biggest problem I think is that mutual information,

00:47:44.631 --> 00:47:47.871
a sort of real workhorse as far as I'm concerned of information theory,

00:47:48.071 --> 00:47:50.831
is an average over many, many different things.

00:47:51.111 --> 00:47:57.271
And if you want to understand more details about the information flow in a system,

00:47:57.431 --> 00:48:01.011
you need to start to unroll that average, start looking at some of the things

00:48:01.011 --> 00:48:02.291
that are normally being averaged over.

00:48:02.431 --> 00:48:05.631
So two of the things I talked about was you can unroll over time,

00:48:05.791 --> 00:48:10.211
so you can start to see that some component carries information at one point

00:48:10.211 --> 00:48:13.991
in time that's quite different than the information it carries at another point in time.

00:48:14.631 --> 00:48:18.131
And that's part of what information flow is all about. How does these patterns

00:48:18.131 --> 00:48:21.571
of information across the system change over time?

00:48:21.791 --> 00:48:26.831
And you also find that if you unroll over the stimulus values,

00:48:27.051 --> 00:48:32.251
the actual possible outcomes of experiments on your variable that you're exploring,

00:48:32.411 --> 00:48:36.151
you start to see patterns there too, that some elements may carry much less

00:48:36.151 --> 00:48:39.311
information about some range of the stimulus than they do other ranges.

00:48:39.531 --> 00:48:41.031
So those are the ideas there.

00:48:41.911 --> 00:48:44.731
And that's not really us. Those are measures that are in the literature,

00:48:44.791 --> 00:48:48.171
which just haven't traditionally been applied in the way that we're applying them.

00:48:48.531 --> 00:48:52.111
Then you get into dynamic measures, and that's a much more complicated area.

00:48:52.211 --> 00:48:54.531
You want to be able to talk about things like….

00:48:56.056 --> 00:49:00.916
When elements gain or lose information and how you quantify gain or loss and

00:49:00.916 --> 00:49:04.336
how information is transferred from one element to another.

00:49:04.576 --> 00:49:08.536
And part of the problem, the whole metaphor of flow for information is a little

00:49:08.536 --> 00:49:13.616
misleading because information isn't conserved like mass is or something like that in a liquid.

00:49:13.816 --> 00:49:17.576
So you can't just talk about, well, it went down here, it went up there,

00:49:17.696 --> 00:49:19.696
therefore there was some kind of a transfer.

00:49:19.856 --> 00:49:26.876
You need to be more complicated And part of it involves characterizing information

00:49:26.876 --> 00:49:30.836
between more than two variables, so-called multivariate mutual information.

00:49:31.016 --> 00:49:35.736
And that, as I mentioned in my talk, is a very, very active area of research

00:49:35.736 --> 00:49:40.676
with a lot of different ideas about what properties a mutual information measure ought to have.

00:49:40.676 --> 00:49:46.136
Um, and the particular measures we took are based on our particular way of looking

00:49:46.136 --> 00:49:48.956
at that, which, which itself is somewhat complicated.

00:49:49.016 --> 00:49:52.096
There are several different levels of, of things we've put forward,

00:49:52.276 --> 00:49:56.156
some of which are fairly uncontroversial and some of which are much more controversial.

00:49:56.156 --> 00:50:00.556
And so the main point there is that I guess the lesson I would take from that

00:50:00.556 --> 00:50:04.296
is that in trying to understand these simple toy agents.

00:50:04.576 --> 00:50:08.916
We ended up pushing the tools of information theory in a particular way,

00:50:09.016 --> 00:50:13.856
which I think is probably going to be important for applying those tools, for example, to C.

00:50:13.916 --> 00:50:18.276
Elegans. Okay, so even if there were no other benefit to have been gained from

00:50:18.276 --> 00:50:22.256
analyzing these toy models, and I believe that there is additional benefit that's

00:50:22.256 --> 00:50:26.096
been gained, that tool development, I think, is actually very, very important.

00:50:26.496 --> 00:50:31.036
But then you also introduced measures of information that started to include time.

00:50:31.476 --> 00:50:35.696
So does that mean you actually start to merge the dynamical systems view with

00:50:35.696 --> 00:50:38.076
the information theoretic view? It's not a matter of merging it.

00:50:38.116 --> 00:50:43.636
There's actually a significant subset of work in information theory that's interested

00:50:43.636 --> 00:50:44.736
in dynamic information.

00:50:45.116 --> 00:50:49.516
What I think, again, the way I like to look at it is you have these two distinct

00:50:49.516 --> 00:50:53.416
lenses, and it's interesting to ask what's the relationship between the two.

00:50:53.416 --> 00:50:56.016
So, your intuition is exactly right.

00:50:56.176 --> 00:51:01.916
If information is changing over time, and dynamics is about changes in state over time,

00:51:02.096 --> 00:51:05.976
then one would think that there's some interesting bridges to be built there,

00:51:06.096 --> 00:51:09.416
and that's exactly what we've tried to do in the part of the talk I wasn't able

00:51:09.416 --> 00:51:12.416
to get to, is show you in some detail how,

00:51:12.556 --> 00:51:16.676
for example, the structure of these manifolds of states that show up as being

00:51:16.676 --> 00:51:21.976
so important when you look at the dynamical systems lens, can be exactly related

00:51:21.976 --> 00:51:27.156
to the information measures that show up when you use information theory.

00:51:28.177 --> 00:51:32.437
But now you could also imagine that you would like to bring it back to causal

00:51:32.437 --> 00:51:34.337
interaction in your network, right?

00:51:34.397 --> 00:51:39.657
And the dynamical systems view or the information theoretic view doesn't necessarily

00:51:39.657 --> 00:51:40.637
give you that automatically.

00:51:40.997 --> 00:51:44.777
It doesn't necessarily, no. I mean, given that we're doing our dynamical analysis

00:51:44.777 --> 00:51:51.137
at the level of the fundamental states of the system, it actually ends up being a causal story.

00:51:51.357 --> 00:51:55.317
But you're right that dynamical systems theory itself isn't intrinsically causal

00:51:55.317 --> 00:52:00.217
or not, because you can look at collective variables where it's not directly causal anymore.

00:52:00.217 --> 00:52:03.997
So the question is, do you feel that you have to now insert a third lens that

00:52:03.997 --> 00:52:08.017
gets you a more systematic handle on the causal interaction?

00:52:08.197 --> 00:52:12.037
I don't view that as a different… Again, remember, for me the lenses are the mathematical tool.

00:52:12.217 --> 00:52:16.577
So for example, you could apply the dynamical systems lens at multiple levels.

00:52:16.897 --> 00:52:23.157
And in fact, I think the way causality works in science is… So if you look at

00:52:23.157 --> 00:52:26.157
a system at one level, it's always ever going to be a description.

00:52:26.937 --> 00:52:30.597
You look at a system at two levels and you look at bridging the lower level

00:52:30.597 --> 00:52:33.217
and the higher level, then you're talking about causality.

00:52:33.597 --> 00:52:37.597
And we, as you noticed in the talk, typically look at two levels.

00:52:37.597 --> 00:52:42.237
We have the level of behavior, which is our high level, and we have the level

00:52:42.237 --> 00:52:46.717
of, say, the mechanical and the neural states that are involved,

00:52:46.937 --> 00:52:51.657
which for us is the causal level because there is no lower physics below that.

00:52:51.817 --> 00:52:56.937
And so I think we're always talking about causal explanations, but you're right.

00:52:56.977 --> 00:53:00.317
It's not intrinsic to a given mathematical lens. I agree with that. Mm-hmm.

00:53:01.359 --> 00:53:07.279
So then when you were summarizing, so these were the examples that you used

00:53:07.279 --> 00:53:13.479
to introduce this approach, what I called methods, but you called it language and lenses.

00:53:15.259 --> 00:53:18.619
But in some sense, if you then sort of step back a little bit and say,

00:53:18.659 --> 00:53:23.779
okay, but what have we achieved with this over the last 25 years about, I would think, no?

00:53:23.919 --> 00:53:28.639
Maybe you're in this domain for a while. You had quite a list of phenomena that

00:53:28.639 --> 00:53:31.999
you were alluding to or that you were saying, look, we have a handle on these

00:53:31.999 --> 00:53:33.899
phenomena in some way with this approach, right?

00:53:33.979 --> 00:53:39.139
And this was starting in Camel, Texas, but then you talked also about attention,

00:53:39.579 --> 00:53:43.859
about a minimally cognitive agent, you talked about learning, right?

00:53:43.939 --> 00:53:47.959
So is it, what makes, for instance, the agent we have just discussed,

00:53:48.219 --> 00:53:50.779
what makes that agent minimally cognitive?

00:53:51.219 --> 00:53:54.939
Okay, that's a good question. I never got a chance to respond to with Tony.

00:53:56.099 --> 00:53:59.459
So, to explain it, I have to give you just a tiny little bit of history.

00:53:59.879 --> 00:54:03.399
And that is that when I first started doing this work, as you may know,

00:54:03.639 --> 00:54:05.859
the first examples were locomotion. Right.

00:54:05.979 --> 00:54:09.819
That was what I... And there is when I focused on the importance of the feedback

00:54:09.819 --> 00:54:13.019
through the body, and this whole brain-body-environment idea was originally

00:54:13.019 --> 00:54:15.399
coming from looking at motor control, like locomotion.

00:54:16.519 --> 00:54:21.359
And I and others that were doing this kind of work at the time felt that the

00:54:21.359 --> 00:54:24.919
kinds of lessons that we were learning about brain-body environment systems

00:54:24.919 --> 00:54:28.419
by studying these motor control tasks had more general applicability.

00:54:28.859 --> 00:54:33.739
But the problem is if you go to, for example, a cognitive science audience and

00:54:33.739 --> 00:54:36.199
talk about walking, they're not impressed.

00:54:39.479 --> 00:54:42.419
Different people mean different things by minimally cognitive.

00:54:42.719 --> 00:54:46.839
The way I defined minimally cognitive in the first paper where I used that term

00:54:46.839 --> 00:54:49.739
was it It is basically behavior.

00:54:50.459 --> 00:54:55.279
It's the simplest behavior that raises genuinely cognitive issues.

00:54:56.079 --> 00:55:00.999
And how I evaluate that is very simple. If you present walking to an audience

00:55:00.999 --> 00:55:03.399
of cognitive scientists, nobody cares.

00:55:03.619 --> 00:55:08.879
You present categorization or a selective attention or something like that,

00:55:09.059 --> 00:55:09.799
a referential communication.

00:55:09.979 --> 00:55:13.199
These are all things we've looked at to an audience of cognitive science people.

00:55:13.379 --> 00:55:17.259
They sit up and they're interested in what you have to say. So I'm not taking

00:55:17.259 --> 00:55:22.879
any theoretical position in the notion of minimally cognitive as to what is or isn't cognitive.

00:55:23.079 --> 00:55:26.739
It's simply saying there's a certain level of sophistication of the behavior

00:55:26.739 --> 00:55:31.639
that has to be recognized as being cognitively interesting by the cognitive

00:55:31.639 --> 00:55:37.359
science community so that we can engage them with these issues about brain-body-environment

00:55:37.359 --> 00:55:39.399
systems and the role of dynamics and so on.

00:55:40.429 --> 00:55:45.529
So I took it to mean that your agent has a minimal memory, and as a result,

00:55:45.549 --> 00:55:49.329
it can maintain information over time, and that gives it then,

00:55:49.429 --> 00:55:51.389
let's say, a cognitive state.

00:55:51.709 --> 00:55:56.169
Yeah. So, I mean, if for you, memory is one of those key trigger issues,

00:55:56.389 --> 00:55:57.769
then absolutely, that's fine.

00:55:57.889 --> 00:56:01.489
No, no, I think that… But some of the tasks we've looked at that I call minimally

00:56:01.489 --> 00:56:04.149
cognitive don't have memory associated with them.

00:56:04.149 --> 00:56:07.229
Okay, so what makes me feel a bit uneasy is that you're saying,

00:56:07.349 --> 00:56:12.609
look, I'm using a label that's not necessarily exactly accurate for what I do,

00:56:12.829 --> 00:56:18.149
like minimal cognitive, but it helps me to communicate what I do to a certain community, right?

00:56:18.329 --> 00:56:23.289
Well, in my perspective, cognition has a very specific definition that's almost

00:56:23.289 --> 00:56:26.769
domain independent, or it has to be, otherwise we're not making progress in science.

00:56:27.149 --> 00:56:31.309
And there, cognition is always tied to some forms of knowledge, right?

00:56:31.349 --> 00:56:36.049
This is also how it is defined. So, that would mean there is some aspect of

00:56:36.049 --> 00:56:39.729
knowledge acquisition, retention and expression that makes a system cognitive.

00:56:42.929 --> 00:56:46.309
I felt that you're more loose about this. Oh, I absolutely am.

00:56:46.409 --> 00:56:47.369
I'm trying to explain it.

00:56:48.769 --> 00:56:52.989
It's not as simple as you just said. If you go and ask a room full of cognitive scientists,

00:56:53.629 --> 00:56:57.129
actually let's do it this way, if you separated them so they're not all in the

00:56:57.129 --> 00:57:00.569
same room and asked them individually how they would characterize what is and

00:57:00.569 --> 00:57:04.309
isn't cognitive, you would likely get as many answers as there are people that you ask.

00:57:04.369 --> 00:57:09.309
And that's why, I mean, like everything else I've been saying, I think it's premature.

00:57:09.709 --> 00:57:13.069
I don't think we have an accepted definition of what cognitive is.

00:57:13.189 --> 00:57:18.709
But what we do have is a field that studies cognition, and they have a set of

00:57:18.709 --> 00:57:23.569
intuitions about what is or isn't sufficiently cognitive to be worth study by that field.

00:57:23.809 --> 00:57:29.069
And all I'm saying is that some of our toy models, if they're going to have

00:57:29.069 --> 00:57:33.209
anything to say to cognitive science as a field needs to engage those intuitions.

00:57:33.209 --> 00:57:37.209
So if for you it's memory, then some of the things I listed under that list

00:57:37.209 --> 00:57:40.129
of minimally cognitive behavior probably wouldn't count for you,

00:57:40.189 --> 00:57:41.849
but others would, and that's fine.

00:57:42.069 --> 00:57:46.249
I have no problem with that. This is something that I would object to because

00:57:46.249 --> 00:57:48.649
I think it's really important that as a field.

00:57:49.666 --> 00:57:54.906
Do try to converge on definitions because otherwise we cannot phrase our hypothesis in a coherent way.

00:57:55.066 --> 00:57:58.306
I wouldn't disagree with that need, but I'm just, what I'm trying to do is tell

00:57:58.306 --> 00:58:02.106
you is realistically the state of the field at the moment, I don't think you would disagree.

00:58:02.246 --> 00:58:07.866
So I've actually been a little unhappy that other people have used the word

00:58:07.866 --> 00:58:11.306
minimally cognitive to mean something more than that,

00:58:11.346 --> 00:58:17.446
that somehow there is a line that's been put forward that where something above

00:58:17.446 --> 00:58:20.146
that line is cognitive and something below that line isn't.

00:58:20.906 --> 00:58:25.426
I just didn't mean it by that term. Others may have a definition.

00:58:25.646 --> 00:58:29.246
I'm aware of a number of different sort of fundamental definitions,

00:58:29.286 --> 00:58:32.406
like one that comes out of Maturana and Varela's work about what's cognitive,

00:58:32.546 --> 00:58:34.566
which is another area of work that I'm involved in.

00:58:35.386 --> 00:58:42.546
But I don't feel a need to subscribe or not to such an issue in order to say

00:58:42.546 --> 00:58:45.966
that, look, some of the brain-body environment models that we're developing

00:58:45.966 --> 00:58:48.466
are engaging issues of interest in cognitive science.

00:58:48.846 --> 00:58:53.326
I think, I mean, so in your list you had, I think, items like working memory,

00:58:53.406 --> 00:58:55.246
or was it short-term memory? Short-term memory, yeah.

00:58:55.746 --> 00:59:00.606
Each of those is a very specific task.

00:59:01.006 --> 00:59:06.186
Yeah, but actually that maps onto something which, to you, is not a direct analog

00:59:06.186 --> 00:59:11.726
of those phenomenon the way we might think of them in mammals.

00:59:11.726 --> 00:59:15.666
Yeah, how could they be with a dozen neurons or something like that?

00:59:15.706 --> 00:59:19.846
But on the other hand, you defined a couple of, I think, relatively new terms,

00:59:19.926 --> 00:59:24.266
like information offloading, information self-structuring. Those are both terms from the literature.

00:59:24.546 --> 00:59:31.106
Are those in the same category of things that are perhaps high-level cognitive structures?

00:59:31.466 --> 00:59:34.486
No. Okay, so they're in a different category. Could you explain the difference?

00:59:34.686 --> 00:59:37.486
Well, I didn't list those under what you're talking about. I know.

00:59:37.566 --> 00:59:40.766
They appeared in the talk. So why are those different? So, um.

00:59:42.122 --> 00:59:46.982
Those terms have been developed in information theory, or more correctly,

00:59:47.082 --> 00:59:51.802
in the application of information theory to, let's say, animals.

00:59:51.862 --> 00:59:55.482
I'm trying to be very general about it, rather than just necessarily cognitive systems.

00:59:55.742 --> 01:00:01.082
So information offloading is a term where you take information that's inside

01:00:01.082 --> 01:00:04.602
the system and you put it outside the system for a while,

01:00:04.682 --> 01:00:09.542
and then you later re-interact with that offloaded information to sort of bring

01:00:09.542 --> 01:00:11.462
it back into the operation of the system.

01:00:12.122 --> 01:00:16.882
I can't tell you who defined it originally, but one of the earlier papers that

01:00:16.882 --> 01:00:21.622
I read was by Olaf Sporns and Max Longorella, where that term was used.

01:00:21.702 --> 01:00:23.362
Information self-structuring is similar.

01:00:23.722 --> 01:00:29.602
In that case, it's that you move your body around so as to elicit information

01:00:29.602 --> 01:00:33.642
from the environment that's not necessarily there passively for you to pick

01:00:33.642 --> 01:00:36.002
up on. So those aren't tasks.

01:00:36.922 --> 01:00:41.442
If you look at a textbook in cognitive science, you're not going to find a section

01:00:41.442 --> 01:00:42.982
on information offloading.

01:00:43.002 --> 01:00:47.582
You will find a section on short-term memory and language, let's say,

01:00:47.622 --> 01:00:52.662
rather than communication, and relational categorization. Those are things you'll

01:00:52.662 --> 01:00:55.382
find in a textbook. So they're just two different sorts of terms.

01:00:55.882 --> 01:01:00.282
I'm not entirely sure because you then alluded to examples of how people offload

01:01:00.282 --> 01:01:03.002
information, for instance, by writing things down. No, in fact,

01:01:03.002 --> 01:01:08.382
that's very interesting. I mean, uh, it's a more recent, it's a more recent, uh.

01:01:09.885 --> 01:01:14.405
Component of the vocabulary of cognitive science to talk about things like offloading.

01:01:14.985 --> 01:01:18.845
So maybe eventually there will be a chapter in a cognitive science or cognitive

01:01:18.845 --> 01:01:20.545
psychology book about such things.

01:01:20.765 --> 01:01:25.285
There are certainly people like David Kirsch that have been arguing how we organize

01:01:25.285 --> 01:01:28.785
our environments is a really key component of our cognitive processes.

01:01:29.245 --> 01:01:34.005
So I mean, there's precedent for that, but it's just not the traditional set

01:01:34.005 --> 01:01:36.345
of topics you would see in cognitive science.

01:01:36.345 --> 01:01:41.005
But I mean, but on that basis, maybe there's an argument you could make that

01:01:41.005 --> 01:01:45.085
let's redefine some of these other concepts that have been knocking around cognitive

01:01:45.085 --> 01:01:48.085
science based on the things we can see in your- Sure.

01:01:48.185 --> 01:01:51.165
Well, in fact, I mean, if you look at, so each of the things I listed on that

01:01:51.165 --> 01:01:54.285
final slide, there's at least one paper, usually several papers on.

01:01:54.405 --> 01:01:58.665
I mean, each one of those makes a specific set of arguments about how the results

01:01:58.665 --> 01:02:02.585
of analyzing that agent suggest we should look at this process very differently

01:02:02.585 --> 01:02:06.285
than it's traditionally been looked at. I mean, that's kind of the modus operandi

01:02:06.285 --> 01:02:08.825
of the research program, is to keep doing that.

01:02:08.945 --> 01:02:12.865
So the stuff on learning that you mentioned, for example, specifically looks

01:02:12.865 --> 01:02:16.225
at an issue related to learning without synaptic plasticity.

01:02:16.485 --> 01:02:20.185
That's sort of the argument there that you, well, I won't get into the details,

01:02:20.285 --> 01:02:24.725
but I mean, each one of them has exactly that feature that perhaps… So maybe

01:02:24.725 --> 01:02:25.805
this is a challenge, right?

01:02:25.865 --> 01:02:30.205
I'm supposed to say like, well, I just relabel and I hope that people aren't

01:02:30.205 --> 01:02:31.385
unhappy with what I'm saying.

01:02:31.385 --> 01:02:35.965
Maybe it is really about redefining, but this is, I think, also Tony's invitation

01:02:35.965 --> 01:02:41.125
here. Yeah, I mean, in the end, so now you're going to make me record a strong

01:02:41.125 --> 01:02:42.125
statement here. Yes, please.

01:02:42.265 --> 01:02:46.745
But I mean, in the end, I'm not sure that cognition or cognitive is a useful,

01:02:46.945 --> 01:02:48.305
scientifically useful term. Mm-hmm.

01:02:48.766 --> 01:02:52.966
I think it tries to slice up things that ultimately they're just various kinds

01:02:52.966 --> 01:02:56.926
of behavior that, for example, to your memory point, are more or less driven

01:02:56.926 --> 01:02:59.526
by internal dynamics versus external dynamics.

01:02:59.926 --> 01:03:05.106
So this is one of the reasons why I may be acting as if I'm being coy about

01:03:05.106 --> 01:03:07.586
it, but I fundamentally, I'm just not sure it's a useful term,

01:03:07.606 --> 01:03:10.226
just like I'm not sure representation is really a useful term.

01:03:10.406 --> 01:03:13.246
So rather than arguing about it, which is something I used to do,

01:03:13.426 --> 01:03:19.246
I just don't use it and I find I've lost nothing by removing that term from

01:03:19.246 --> 01:03:22.506
my vocabulary and talking just about internal state and so on.

01:03:22.506 --> 01:03:25.146
I see your point, but I do believe to make progress in the field,

01:03:25.266 --> 01:03:28.406
certainly about building up a psychology or cognitive science,

01:03:28.746 --> 01:03:32.186
we do have to get clarity on the constructs we're going to use because they

01:03:32.186 --> 01:03:33.206
have to drive our theory.

01:03:33.286 --> 01:03:37.046
That's certainly true, but that doesn't necessarily mean that you have to clarify cognition.

01:03:37.346 --> 01:03:39.606
Absolutely. If, for example, that's not necessarily the right construct.

01:03:39.606 --> 01:03:43.966
Maybe we should forget about it and replace it with better defined constructs.

01:03:43.966 --> 01:03:47.346
The other thing is, if it is a useful distinction, I think eventually it will

01:03:47.346 --> 01:03:49.206
become clear how we ought to define it.

01:03:49.266 --> 01:03:53.186
But trying to define it before we really know what we're talking about just seems premature to me.

01:03:53.246 --> 01:03:57.846
It's kind of like defining mass before you have any notion of Newton's second

01:03:57.846 --> 01:04:01.546
law, which, by the way, the notion of mass there is completely different than

01:04:01.546 --> 01:04:02.526
it is in general relativity.

01:04:02.846 --> 01:04:07.146
I mean, so it itself was a changing, and mass is a much simpler concept than cognition. Right.

01:04:07.386 --> 01:04:13.406
So, Randy, now you have met over the last, let's say, 25, 30 years,

01:04:13.526 --> 01:04:16.826
being active and really pushing a very specific agenda with great success.

01:04:17.806 --> 01:04:20.126
Also influencing many people in their work.

01:04:21.806 --> 01:04:27.846
What would be Randy's law that people should adhere to to make progress in understanding

01:04:27.846 --> 01:04:30.186
brain, body, and environment?

01:04:31.246 --> 01:04:35.106
Yeah, so I don't know. I don't have a particular law that I can think of.

01:04:35.106 --> 01:04:36.926
What I would suggest is that...

01:04:40.421 --> 01:04:44.721
I think these days, the fact that so many people take for granted the idea of

01:04:44.721 --> 01:04:48.581
brain-body environment systems is just the proper unit of analysis has come

01:04:48.581 --> 01:04:51.441
out of the line of work that I was a part of, at least.

01:04:51.521 --> 01:04:54.961
Certainly many other people have pushed that line too, but I think it's much

01:04:54.961 --> 01:04:58.501
harder to argue with that view than it was when some of us were first starting

01:04:58.501 --> 01:05:01.401
out kind of as voices in the wilderness.

01:05:01.401 --> 01:05:06.981
I hope something else that's come out of the work I've done is just how important

01:05:06.981 --> 01:05:12.041
dynamical systems theory and the concepts of dynamical systems are as at least part of the toolbox.

01:05:12.361 --> 01:05:18.061
That also was something that was very, very, very controversial when we started out.

01:05:18.601 --> 01:05:24.781
I hope, this is one thing I think that's still in progress, that this notion

01:05:24.781 --> 01:05:28.961
of the toy models have a really important role to play in the development of

01:05:28.961 --> 01:05:31.981
theory in the behavioral and brain sciences.

01:05:32.101 --> 01:05:37.441
That's a somewhat more unusual position, and it's one I've been pushing for a while.

01:05:37.541 --> 01:05:40.961
But I hope that eventually that will sink in as, again, not the only way to

01:05:40.961 --> 01:05:43.261
proceed, but an important component of theory.

01:05:43.361 --> 01:05:48.261
And most recently, I hope that if we reformulate some of the key concepts in

01:05:48.261 --> 01:05:51.761
information theory, that it'll actually turn out to be another really useful

01:05:51.761 --> 01:05:54.621
tool that's not in competition with something like dynamical systems,

01:05:54.721 --> 01:05:57.161
but is a great augmentation to it.

01:05:57.261 --> 01:06:00.621
Which by the way, I still get a lot of, so I give the talk like this at a group

01:06:00.621 --> 01:06:04.661
that's full of dynamical systems people, and they just can't believe the words

01:06:04.661 --> 01:06:06.461
that are coming out of my mouth, right?

01:06:07.421 --> 01:06:11.501
So although that may be obvious for some people, for others that are clinging

01:06:11.501 --> 01:06:17.521
to, you know, the problem I think is conflating information theory as a body

01:06:17.521 --> 01:06:20.261
of mathematics with information

01:06:20.581 --> 01:06:23.521
processing as a notion that the brain pushes symbols around.

01:06:23.621 --> 01:06:25.801
And that's why I'm so insistent about this lens distinction.

01:06:26.161 --> 01:06:30.781
It's only because I want to make sure that people understand I'm talking about

01:06:30.781 --> 01:06:32.381
the lens and not the sort of,

01:06:32.733 --> 01:06:38.313
of framework that people were previously associating with information processing.

01:06:38.453 --> 01:06:42.733
But it would mean Randy's law would be like, don't worry about taking a minority view.

01:06:43.133 --> 01:06:45.453
Oh, I see. Okay. Would it be fair to say?

01:06:45.933 --> 01:06:49.733
I don't know. I mean, it can be dangerous too, right? Sure, absolutely.

01:06:49.993 --> 01:06:55.133
But look where it got you. I mean, but for me, science

01:06:55.133 --> 01:06:57.953
is most interesting at

01:06:57.953 --> 01:07:00.893
sort of the boundaries and the limits of what you're doing

01:07:00.893 --> 01:07:04.033
and so it's it's not necessarily good career

01:07:04.033 --> 01:07:07.513
advice for a student sometimes that you should sort of jump

01:07:07.513 --> 01:07:10.173
to a boundary and start pushing really hard and loudly but it

01:07:10.173 --> 01:07:12.813
certainly worked pretty well for me and i'm not sure i

01:07:12.813 --> 01:07:15.713
could have done it any other way it's sort of a personality thing i sort

01:07:15.713 --> 01:07:20.033
of normal science i guess doesn't engage me as much as sort of really pushing

01:07:20.033 --> 01:07:24.653
at conceptual boundaries right and things so now uh tony likes traveling and

01:07:24.653 --> 01:07:28.033
soon easy jets will fly to the states as well including indiana so we can send

01:07:28.033 --> 01:07:31.513
it for a little money to your lab and we're going to do that four years from

01:07:31.513 --> 01:07:35.053
now so Tony can test a prediction you're going to make today.

01:07:35.333 --> 01:07:41.113
So what specific prediction could you make now that you will see tested and

01:07:41.113 --> 01:07:45.493
validated four years from now and maybe rejected, but a specific prediction?

01:07:47.713 --> 01:07:51.553
Well, so, oh, you mean predictions about the field or predictions about some

01:07:51.553 --> 01:07:54.313
experimental... Science, what you do, the work you do, like the C.

01:07:54.353 --> 01:07:57.973
Elegans work or any other system you're working on, really specific scientific

01:07:57.973 --> 01:07:59.973
prediction that you will see tested?

01:08:00.633 --> 01:08:05.933
Well, so I mean, the problem is I don't want to talk about unpublished work.

01:08:06.073 --> 01:08:09.373
And so all I can talk about is work we've published. And it's actually most

01:08:09.373 --> 01:08:12.013
of our predictions were actually successfully tested in that.

01:08:12.013 --> 01:08:13.713
It could be more general than that, perhaps.

01:08:15.368 --> 01:08:20.448
What the field of C. elegans. Well, so I mean. No, in four years time. Anything Tony can test.

01:08:20.728 --> 01:08:25.608
Well, more general is easier. I mean, I really think it's within reach.

01:08:25.928 --> 01:08:29.388
Four years is short. I would maybe think it's decade long or so,

01:08:29.448 --> 01:08:34.728
but I really think it's within reach to imagine having a complete neural mechanical

01:08:34.728 --> 01:08:37.908
behavioral model of C. elegans. That's not- You said 10 years?

01:08:38.208 --> 01:08:43.128
Yeah. Okay. I think 10 years is pushing it.

01:08:43.148 --> 01:08:45.708
Now that's not, I don't know if that's a prediction or not, but I mean,

01:08:45.728 --> 01:08:49.028
even a year ago, I might not have said that because it was only over the past

01:08:49.028 --> 01:08:52.388
year that we had these new optical techniques that can image brain activity

01:08:52.388 --> 01:08:55.888
of a whole animal at the level of individual cells.

01:08:55.988 --> 01:08:58.468
That I think is ultimately going to be really, really important.

01:08:58.668 --> 01:09:02.628
But it's going to take some time for that to move into the scientific toolkit

01:09:02.628 --> 01:09:06.988
of the experimentalists that study the system and therefore can better constrain the modelers.

01:09:07.108 --> 01:09:10.428
Great. Well, Randy Beer, thank you very much for this conversation. All right. Thank you.

01:09:12.388 --> 01:09:17.668
Well, that was fun. Was it? Yeah. Good. The CSN podcast was produced by the

01:09:17.668 --> 01:09:21.568
Convergent Science Network of Biometrics and Biohybrid Systems,

01:09:22.028 --> 01:09:27.228
a project funded by the European Sevens Research Framework Programme.

01:09:28.488 --> 01:09:33.768
For more interviews, recorded lectures or upcoming conferences in the field

01:09:33.768 --> 01:09:40.028
of biometrics and biohybrid systems, go to csnnetwork.eu.

01:09:40.348 --> 01:09:42.188
And thank you for listening.

01:09:40.880 --> 01:09:47.600
Music.