WEBVTT

00:00:03.497 --> 00:00:10.497
This is the Convergent Science Network podcast. Leading researchers in the domain

00:00:10.497 --> 00:00:16.777
of neuroscience, brain theory and technology are interviewed by Paul Verschure and Tony Prescott.

00:00:26.097 --> 00:00:29.797
Is paul for sure with a conversion science network and

00:00:29.797 --> 00:00:32.937
here i'm speaking with abehart fetz and

00:00:32.937 --> 00:00:35.637
ab spoke this morning in our

00:00:35.637 --> 00:00:41.457
in our summer school about what he called the self in the brain so but what

00:00:41.457 --> 00:00:46.377
what did you really have in mind when you described the self in the brain So

00:00:46.377 --> 00:00:52.817
there's a common experience of everybody knows that there's a self in their

00:00:52.817 --> 00:00:54.757
brain that is interacting with the world.

00:00:54.757 --> 00:01:00.837
And I brought it up as a model for understanding how the brain interacts with

00:01:00.837 --> 00:01:05.577
external devices, how the brain would interact with brain-computer interfaces,

00:01:05.857 --> 00:01:13.157
and also how the self would deal with recurrent brain-computer interfaces,

00:01:13.457 --> 00:01:15.637
that is, bidirectional brain-computer interfaces.

00:01:15.637 --> 00:01:18.777
Faces right because this is if you

00:01:18.777 --> 00:01:21.817
want this is also a metaphor right of how right how

00:01:21.817 --> 00:01:24.857
volitional control could be exerted over even

00:01:24.857 --> 00:01:29.577
single neurons in some sense yes because usually we we think about okay we are

00:01:29.577 --> 00:01:33.757
controlling the body we control the world given our goals and our wishes our

00:01:33.757 --> 00:01:39.937
intentions but what you have have specialized in uh with also really very exciting

00:01:39.937 --> 00:01:43.897
results is in some sense how this mapping, if you want,

00:01:43.977 --> 00:01:48.157
between a brain and the volitional control of the brain itself and external

00:01:48.157 --> 00:01:50.617
devices can be almost arbitrarily mapped, right?

00:01:50.857 --> 00:01:55.097
That's right. Your first experiments were in the 1960s in this domain.

00:01:55.477 --> 00:02:00.357
That's correct. So what was the key observation that really pushed this forward?

00:02:00.757 --> 00:02:05.357
So these experiments were done as a postdoctoral fellow.

00:02:05.457 --> 00:02:11.017
I came from MIT just having a degree in physics, which was totally useless for what I was going to do.

00:02:11.497 --> 00:02:16.037
But I needed to know about neurophysiology and behavior, and I thought it would

00:02:16.037 --> 00:02:23.837
be fun to put these two together and do experiments on training monkeys to volitionally

00:02:23.837 --> 00:02:26.017
control the neural activity in their brain.

00:02:26.357 --> 00:02:29.537
And it was a great deal of fun, and I wish I'd stuck with it, but...

00:02:30.493 --> 00:02:36.013
Because it turns out, in retrospect, that what I had looked at as biofeedback

00:02:36.013 --> 00:02:39.693
for these operant conditioning studies, which is the deflection of a meter arm

00:02:39.693 --> 00:02:41.953
controlled by brain activity,

00:02:42.113 --> 00:02:46.953
is sort of the paradigm that's now being used in brain-machine interfaces.

00:02:48.033 --> 00:02:52.673
And I wish I had had the connection between meter arm and prosthetic arm.

00:02:52.813 --> 00:02:54.293
My career would have been quite different.

00:02:54.453 --> 00:03:00.673
Right. It is a demonstration of your narrow vision that this meter arm is just

00:03:00.673 --> 00:03:02.473
feedback, not a prosthetic device.

00:03:02.873 --> 00:03:07.093
Right, because in those experiments, so I guess you were already exposed to

00:03:07.093 --> 00:03:09.693
notions like operant conditioning and so on at the time.

00:03:10.093 --> 00:03:13.793
Not so much. I mean, this work exposed me to that.

00:03:13.913 --> 00:03:21.453
I was working in a lab of Mitch Glickstein, and he had some very bright postdoctoral

00:03:21.453 --> 00:03:23.413
fellows who knew all about operant conditioning.

00:03:23.413 --> 00:03:30.033
And I also worked with a practitioner of the art, Dom Finocchio,

00:03:30.173 --> 00:03:36.333
who was really an expert on this, and taught me all the lingo and techniques of it.

00:03:36.433 --> 00:03:40.313
So basically, it was something learned as we went along.

00:03:40.533 --> 00:03:45.733
Right. But learned in the context of operantly conditioning neural activity instead of behavior.

00:03:45.733 --> 00:03:51.593
Okay, but I think this very first experiment that you published in 69,

00:03:51.913 --> 00:03:56.813
in science, in some sense set the tone for what followed.

00:03:57.613 --> 00:04:01.553
Because in some, what you tried to do there is say, okay, the monkey was watching

00:04:01.553 --> 00:04:04.493
a dial, but now it had to train.

00:04:04.613 --> 00:04:08.713
The monkey was rewarded for moving the dial in a certain direction,

00:04:08.873 --> 00:04:12.893
but the response that was being conditioned here was really neural activity.

00:04:13.273 --> 00:04:17.693
Exactly. So he was actually, the way I saw it, he was being rewarded for controlling

00:04:17.693 --> 00:04:25.133
the neural activity, and the dial was a help, a conditioned reinforcer, if anything,

00:04:25.413 --> 00:04:30.713
and helped the monkey to zero in on what it was that was going to get rewarded.

00:04:30.713 --> 00:04:38.113
And the monkey quickly learned, I'm anthropomorphizing, but his behavior was

00:04:38.113 --> 00:04:42.573
as if he had learned that the rightward deflection of this meter arm was going

00:04:42.573 --> 00:04:44.473
to be associated with applesauce reward,

00:04:44.773 --> 00:04:48.073
and he would very quickly learn to generate whatever.

00:04:49.110 --> 00:04:54.490
Neural activity or behavior would drive that meter arm to the right and get rewarded.

00:04:54.990 --> 00:04:59.730
But now these neurons, prior to the experiment, were these neurons in any way

00:04:59.730 --> 00:05:03.770
involved with similar kinds of movements in the world?

00:05:04.150 --> 00:05:09.190
Definitely. These were motor cortex neurons. So very large numbers of them were,

00:05:09.310 --> 00:05:14.430
a large fraction of the ones that we worked with were actually neurons that

00:05:14.430 --> 00:05:17.170
were involved in generating movements.

00:05:17.170 --> 00:05:22.570
And so, in fact, one of the original rationales for these studies was to determine

00:05:22.570 --> 00:05:26.570
what the movements were that were associated with particular cells. Right.

00:05:29.170 --> 00:05:34.190
So this gave you, I guess, the initial idea that you actually could change these

00:05:34.190 --> 00:05:38.590
response properties of neurons and map them differently to the world.

00:05:39.070 --> 00:05:42.370
So what was really the next step there?

00:05:42.510 --> 00:05:48.070
And when did this notion of volitional really come in in the development of this whole approach?

00:05:48.070 --> 00:05:54.710
Well, this concept of volitional control of neurons is simply just a way of

00:05:54.710 --> 00:06:02.470
saying that the monkey was generating these responses as if he were making a movement.

00:06:02.470 --> 00:06:09.470
And so what happened ironically is that as we were doing this work,

00:06:09.690 --> 00:06:15.990
the question came up as to whether or not the cells were really causally involved

00:06:15.990 --> 00:06:18.670
in generating the movements and the muscle activity.

00:06:19.390 --> 00:06:29.030
And so I got distracted by a causal question, which unfortunately got me off

00:06:29.030 --> 00:06:30.630
the track of operant conditioning.

00:06:30.630 --> 00:06:37.490
And that technique was to demonstrate causality by doing things like spike-triggered

00:06:37.490 --> 00:06:41.970
averaging of muscle activity, which is a way to determine that the cell actually

00:06:41.970 --> 00:06:43.610
had an output effect on muscles.

00:06:43.930 --> 00:06:50.610
And that was such a devilishly difficult procedure that that gobbled up my time

00:06:50.610 --> 00:06:52.750
for the next decade or two.

00:06:52.750 --> 00:06:57.270
And looking back, what I should have done is stuck with the more fun things,

00:06:57.370 --> 00:07:03.830
which was to see how far we could exploit this operant conditioning paradigm.

00:07:04.290 --> 00:07:08.810
Right. But then for me to understand the logic of that, why did you move to

00:07:08.810 --> 00:07:13.650
a paradigm where you really started to map neural response to muscle activity?

00:07:13.910 --> 00:07:17.930
What was the concept there? Well, the concept was to really,

00:07:18.190 --> 00:07:27.210
in this time, the activity of neurons was being recorded in monkeys generating

00:07:27.210 --> 00:07:32.630
behavior and correlating the changes in the neural activity to the changes in behavior.

00:07:33.610 --> 00:07:40.370
But everybody was frustrated about the lack of real evidence that the cells

00:07:40.370 --> 00:07:44.610
were causally related to the behavior. So this came out of...

00:07:45.942 --> 00:07:52.642
Questions like, how would we show that the cell actually has a causal effect on the muscle activity?

00:07:53.262 --> 00:08:01.182
And this, I think, was an issue that appealed to my physics background to get

00:08:01.182 --> 00:08:07.522
to the bottom of mechanisms and pursue answers to questions like that.

00:08:07.522 --> 00:08:10.402
And so that's what I did for the next couple of decades.

00:08:10.602 --> 00:08:21.122
We did this correlation method of confirming that cells had a real output effect on the muscles.

00:08:21.942 --> 00:08:28.042
And I think that actually did solve a lot of issues as to what the properties

00:08:28.042 --> 00:08:33.602
of these output cells was and how their properties differed from other motor

00:08:33.602 --> 00:08:35.882
cortex cells that didn't have this output on muscles.

00:08:35.882 --> 00:08:40.062
So I think it was – I'm being a little facetious saying it was a waste of time,

00:08:40.142 --> 00:08:46.502
but I think it was definitely a useful enterprise agenda to do this.

00:08:46.502 --> 00:08:50.442
But I do believe that even though it might not be at the core of your current

00:08:50.442 --> 00:08:52.462
interest, I mean, this mapping to muscle activity,

00:08:52.822 --> 00:08:57.122
there are some interesting aspects to the learning dynamics you already observed

00:08:57.122 --> 00:09:02.322
then that I think are also relevant for our current discussion about brain-computer interfaces.

00:09:02.322 --> 00:09:06.602
For instance, that the learning, also what you talked about this morning,

00:09:06.682 --> 00:09:09.962
already these early studies, I felt there was something very strange going on.

00:09:10.682 --> 00:09:15.582
So you train this neuron to, let's say, control a certain muscle based on reward

00:09:15.582 --> 00:09:19.602
or an aversive stimulus, depending on the conditions.

00:09:19.902 --> 00:09:22.482
But it's not necessarily these mappings are static.

00:09:22.802 --> 00:09:26.542
It seems that they're very transient in nature, right? So you induce a sort

00:09:26.542 --> 00:09:30.502
of responsibility because there's reward, but as soon as, let's say,

00:09:30.502 --> 00:09:34.902
the animal is moved into its own cage and it's outside of the experimental context,

00:09:35.162 --> 00:09:37.142
the mapping very quickly disappears.

00:09:37.662 --> 00:09:41.802
Well, I see it more like there is this mapping or this relationship,

00:09:42.222 --> 00:09:48.862
let's say, between a motor cortex cell and a set of muscles that it may be linked

00:09:48.862 --> 00:09:52.502
to functionally in the sense of co-varying with those muscles.

00:09:52.502 --> 00:10:01.242
When we reward the cell, we're basically looking at not just the activity of

00:10:01.242 --> 00:10:04.122
the cell, but also at the correlated muscle activity.

00:10:04.422 --> 00:10:07.982
And so that's a natural relationship that exists in the cell.

00:10:08.556 --> 00:10:11.696
Cage as well as in the experimental booth.

00:10:12.316 --> 00:10:17.596
And we're not actually rewarding a relationship there.

00:10:17.736 --> 00:10:23.336
We're basically rewarding one component of a circuit, which includes the cell and the muscles.

00:10:23.476 --> 00:10:29.016
So part of the issue is how necessary is that relationship?

00:10:29.216 --> 00:10:34.636
So one of the fun things we did was to operantly condition the dissociation

00:10:34.636 --> 00:10:38.016
of very consistent relationships.

00:10:38.256 --> 00:10:41.676
So there was a cell, for example, that always fired with the biceps.

00:10:42.756 --> 00:10:49.656
But then when the monkey was rewarded for firing the cell without the muscle

00:10:49.656 --> 00:10:52.196
activity, he very quickly learned to dissociate them.

00:10:52.556 --> 00:10:58.276
So that was of interest and sort of flew in the face of the dogma that the motor

00:10:58.276 --> 00:11:04.416
cortex cells have a relatively stable and reliable relationship to the muscles.

00:11:04.696 --> 00:11:08.616
Right. So do you believe that there's a full dissociation there,

00:11:08.776 --> 00:11:11.316
or are there some constraints acting upon that system?

00:11:11.756 --> 00:11:18.636
Yeah, I think there's a full dissociation in the sense that one turns off while the other stays active.

00:11:19.696 --> 00:11:26.336
But there are various ways that that can be achieved by just changing the balance

00:11:26.336 --> 00:11:32.136
of inputs to the two to the cell and to the muscle so that they can be independently controlled,

00:11:33.016 --> 00:11:38.476
but do you believe i could for instance retrain my my the homunculus of my motor

00:11:38.476 --> 00:11:46.496
cortex to be inverted that's a big uh challenge if you're talking about the

00:11:46.496 --> 00:11:48.676
whole homunculus we're just We're just talking about a single cell.

00:11:48.776 --> 00:11:51.196
I know, I know, but I just tried to understand the boundaries, right?

00:11:51.356 --> 00:11:54.776
Yeah, right. So, for example, relating it to the homunculus,

00:11:54.996 --> 00:12:06.556
the question might be, can the cell in the hand area be related to movement of the foot?

00:12:07.916 --> 00:12:14.996
And it's probably pretty straightforward to reward the animal and succeed in

00:12:14.996 --> 00:12:19.056
getting him to co-activate the foot in the cell that was related to the hand.

00:12:19.276 --> 00:12:27.396
It doesn't change any functional relationship in the sense of circuitry.

00:12:27.516 --> 00:12:32.816
It just changes the pattern of activation, which is quite different.

00:12:32.816 --> 00:12:36.896
I mean, that pattern of activation is sort of a transient thing that's modifiable

00:12:36.896 --> 00:12:39.096
by these operant techniques.

00:12:40.226 --> 00:12:49.146
But the underlying circuitry is a little bit more fixed and anatomical, as it were.

00:12:49.546 --> 00:12:58.166
But the way these units, the cells and the muscles are activated can be quite flexible.

00:12:58.286 --> 00:13:02.706
And that's actually essentially what this operant conditioning paradigm demonstrates.

00:13:03.046 --> 00:13:07.046
Right. So then, in some of your saying, you could multiplex it.

00:13:07.046 --> 00:13:10.626
You could say, look, there's sort of one layer, one frame of reference that

00:13:10.626 --> 00:13:14.726
is maybe a bit more hardwired, but then on top of that, you could wire in almost

00:13:14.726 --> 00:13:18.086
an arbitrary response set, and they can coexist. exist?

00:13:18.366 --> 00:13:22.786
Well, the term wired for that second thing is a little too strong.

00:13:22.966 --> 00:13:28.726
I think I see wired as more having to do with anatomical connections and activation

00:13:28.726 --> 00:13:36.806
as being the pattern in which the activity is distributed through those hardwired circuits.

00:13:37.246 --> 00:13:43.746
And so the activity patterns can vary quite a bit depending on how the brain

00:13:43.746 --> 00:13:47.366
propagates or generates this activity.

00:13:48.106 --> 00:13:53.806
And it can activate the elements in a fixed circuit in a variable way.

00:13:53.966 --> 00:13:57.906
That's another way to put it. But still, also to do that. You're superimposing

00:13:57.906 --> 00:14:00.466
that. I mean, that's superimposed on the structure.

00:14:00.626 --> 00:14:03.986
That's correct. But still, to superimpose it, you do have to change some of

00:14:03.986 --> 00:14:06.946
the wires to implement that activity pattern.

00:14:07.066 --> 00:14:10.526
Well, I beg to differ. I don't know that you actually are changing wires so

00:14:10.526 --> 00:14:14.746
much as working with the wired system in new ways.

00:14:14.846 --> 00:14:20.586
You basically are propagating activity and ignoring some of the wires that exist

00:14:20.586 --> 00:14:29.406
in this process and activating or propagating activities through some other wires that exist.

00:14:29.406 --> 00:14:37.966
But the activity is sort of the waves on top of this relatively fixed structure.

00:14:38.226 --> 00:14:42.046
But you would agree that you're changing synapses to accomplish that?

00:14:42.566 --> 00:14:44.426
No. I'm sorry.

00:14:45.466 --> 00:14:48.726
But look, how do you get selectivity then in that system?

00:14:50.986 --> 00:14:58.766
That's a good question. You get selectivity by changing the pattern of excitation

00:14:58.766 --> 00:15:04.326
and inhibition that exists in the circuitry and activating cells in different ways.

00:15:04.446 --> 00:15:08.426
I mean, there's a very deep question you're asking is how the brain can use

00:15:08.426 --> 00:15:13.186
a fixed circuit, which I believe it has, in variable ways.

00:15:14.966 --> 00:15:18.746
Let's take one example. So one of these early experiments you described,

00:15:19.046 --> 00:15:22.206
you're mapping two neurons to a muscle.

00:15:23.401 --> 00:15:28.821
And then you showed that you could condition that neuron to go both up or down

00:15:28.821 --> 00:15:33.941
its response, dependent on how you were conditioning it. That's right.

00:15:34.741 --> 00:15:41.141
So now, given the reward contingency, you are modulating the response of this neuron.

00:15:41.341 --> 00:15:44.561
It can go in any direction you want, up or down.

00:15:45.021 --> 00:15:48.721
What's the substrate of that if it's not wires and it's not synapses?

00:15:48.721 --> 00:15:54.321
It is obviously the wires that connect to those cells.

00:15:54.601 --> 00:15:59.901
So we have these two cells that are neighboring cells that have very similar

00:15:59.901 --> 00:16:03.141
normal relationships to a joint.

00:16:03.141 --> 00:16:09.021
But when the monkey is rewarded for activating them independently,

00:16:09.341 --> 00:16:16.761
he figures a way to independently control the synaptic input to those cells

00:16:16.761 --> 00:16:20.861
to make one's activity go up and the other's activity go down.

00:16:21.061 --> 00:16:28.841
So it's a flexible way of activating the circuitry, but it's not actually modifying

00:16:28.841 --> 00:16:33.381
the physiological circuitry. the connections are there.

00:16:33.601 --> 00:16:39.021
So how I read what you're saying is when you say, look, there's like a fixed wired system.

00:16:39.761 --> 00:16:43.321
That's not really changing dramatically due to the conditioning.

00:16:43.501 --> 00:16:46.061
Correct. But there's sort of a modulation of that circuit.

00:16:46.341 --> 00:16:50.061
Exactly. Possibly expressed in, let's say, synaptic connectivity or whatever.

00:16:51.021 --> 00:16:55.281
Okay. Yeah? That might lead to the specificity of these changes.

00:16:55.461 --> 00:16:56.681
It's something along these lines.

00:16:57.961 --> 00:17:04.481
Yes. Okay. So there... Well, so I'm trying to get you to say what the real substrate

00:17:04.481 --> 00:17:05.721
is of these changes, right?

00:17:09.016 --> 00:17:13.396
Yeah, well, I wish I could give you an answer as to exactly how that happens,

00:17:13.456 --> 00:17:15.276
or even give you an analogy.

00:17:15.496 --> 00:17:22.376
I guess one thing that comes to mind is you've got railroad tracks that go all

00:17:22.376 --> 00:17:23.796
over the place. That's the structure.

00:17:23.876 --> 00:17:27.776
And you can have trains riding in different ways over those tracks.

00:17:29.296 --> 00:17:32.416
And one is structure and the other is activity.

00:17:32.896 --> 00:17:39.896
Right. So here, in terms of activation of neurons, it's based on a circuit.

00:17:40.016 --> 00:17:46.016
But the way that circuit is activated generates different patterns of activity.

00:17:46.416 --> 00:17:51.576
And if your question is a demand for an explanation of how that happens.

00:17:52.696 --> 00:17:58.276
I have to admit that I'm not totally sure other than wave my hands and say,

00:17:58.416 --> 00:18:02.476
this is all top-down, as it were.

00:18:02.476 --> 00:18:07.376
In other words, generated from within the brain in ways that are analogous to

00:18:07.376 --> 00:18:10.576
the way you can move different muscles volitionally, independently.

00:18:10.936 --> 00:18:16.756
It's pretty much the same thing because cell activity and muscles are similar.

00:18:17.116 --> 00:18:21.136
So, okay, now that we have sort of an understanding of the substrate that's

00:18:21.136 --> 00:18:24.776
implementing these changes, even though it's not completely clear. Right. Okay.

00:18:25.676 --> 00:18:30.536
You did equate that with a substrate that would support mental imagery.

00:18:30.936 --> 00:18:36.556
Yes. So what's the link there exactly? So the link is that the same cells that

00:18:36.556 --> 00:18:40.396
generate a movement are,

00:18:40.696 --> 00:18:46.376
in many cases, the same cells that are activated when you just imagine that movement.

00:18:46.376 --> 00:18:57.196
And this principle of relationship between natural activity in relation to visual stimulus, for example,

00:18:57.316 --> 00:19:01.976
and imagining that visual stimulus pertains to other areas.

00:19:02.016 --> 00:19:07.356
So in general, imagery uses a lot of the same neural substrate.

00:19:08.647 --> 00:19:15.207
But it doesn't actually activate that substrate enough to produce the movement.

00:19:15.327 --> 00:19:17.387
You don't act out everything you think about.

00:19:18.347 --> 00:19:24.187
So there's a switch that stops this imagery from being expressed.

00:19:25.707 --> 00:19:35.567
Right. So here we see this flexibility of the brain to sort of remodel itself

00:19:35.847 --> 00:19:41.027
and to change its mappings almost arbitrarily, but we looked only at output structures.

00:19:41.627 --> 00:19:48.247
So do you see the same kind of superposition of states also acting out between

00:19:48.247 --> 00:19:53.267
multiple modalities or motor systems and sensory systems?

00:19:53.587 --> 00:19:56.567
So are the mappings also arbitrary in that respect?

00:19:57.607 --> 00:20:01.687
If I understand you right, then the question is on the output side.

00:20:01.687 --> 00:20:07.687
Actually, the motor cortex is sort of on the output side with regard to controlling

00:20:07.687 --> 00:20:12.767
muscles because it's, at minimal,

00:20:12.887 --> 00:20:17.227
it's one synapse away from the motor neurons, but even several synapses away.

00:20:17.347 --> 00:20:23.147
It's functionally tied very closely to activation of motor neurons that contract

00:20:23.147 --> 00:20:30.027
muscles. And there the flexibility is quite clear when you probe,

00:20:30.267 --> 00:20:32.227
when you do experiments that probe that flexibility.

00:20:32.367 --> 00:20:37.767
For example, this operant conditioning paradigm is a way of probing how flexible

00:20:37.767 --> 00:20:39.627
the relationships actually are.

00:20:40.307 --> 00:20:42.387
Does that have anything to do with your question? Yeah, sure.

00:20:42.567 --> 00:20:49.067
So, but could I overlay, let's say, sensory responses over this motor core?

00:20:49.067 --> 00:20:53.867
Could I properly condition these motor neurons to respond to sensory stimuli?

00:20:56.527 --> 00:21:01.767
Well, first of all, the idea that you're conditioning the neurons is a little

00:21:01.767 --> 00:21:07.687
bit misleading because you're conditioning the animal to activate the motor neuron in a certain way.

00:21:07.687 --> 00:21:16.467
So you have to see it as being part of a distributed network of activity that

00:21:16.467 --> 00:21:19.987
is generated by the subject,

00:21:20.067 --> 00:21:23.247
the monkey or the human or whoever's in this experiment,

00:21:23.327 --> 00:21:26.527
and not attribute this activity.

00:21:28.128 --> 00:21:30.808
Um behavior specifically to the

00:21:30.808 --> 00:21:33.988
cell in isolation okay so that's an important

00:21:33.988 --> 00:21:38.128
distinction to understand how it's all being generated and

00:21:38.128 --> 00:21:42.708
what it all means now there was something about the superposition of sensory

00:21:42.708 --> 00:21:47.248
and motor that you were getting at in terms of arbitrary mapping of sensory

00:21:47.248 --> 00:21:53.128
input to motor output exactly um because earlier yeah well go ahead because

00:21:53.128 --> 00:21:54.908
Because earlier we just probed, let's say,

00:21:54.928 --> 00:21:57.948
the boundaries of remapping within a motor system.

00:21:58.188 --> 00:22:02.068
Yes. Right? And there was like, well, there might be some constraints on the

00:22:02.068 --> 00:22:04.908
system, but within those, you can sort of freely map things around.

00:22:05.308 --> 00:22:10.388
So then the obvious next question is to say like, okay, but what are then the

00:22:10.388 --> 00:22:13.268
boundaries with respect to some sort of sensory motor mapping?

00:22:13.508 --> 00:22:21.828
Okay. Okay, so that actually is implicitly addressed in these studies because

00:22:21.828 --> 00:22:28.228
the motor cortex cells that were involved in these conditioning experiments

00:22:28.228 --> 00:22:30.908
also respond to peripheral input.

00:22:31.048 --> 00:22:33.408
So they have a sensory response.

00:22:34.608 --> 00:22:38.148
For example, these two cells we were talking about that got dissociated had

00:22:38.148 --> 00:22:43.468
a nice drive from knee extension, if I remember correctly. They were both driven

00:22:43.468 --> 00:22:45.448
by knee extension. So that's a sensory input.

00:22:46.468 --> 00:22:52.088
And then the operant conditioning got the monkey to individually activate the

00:22:52.088 --> 00:22:54.308
cells without the knee moving at all.

00:22:54.448 --> 00:22:59.588
So in that sense, that answers your question that, yes, the mapping between

00:22:59.588 --> 00:23:05.248
the peripheral input to these cells and the motor output was dissociated by

00:23:05.248 --> 00:23:07.728
this paradigm of conditioning.

00:23:07.728 --> 00:23:13.048
And it was dissociated by virtue of the fact that we were requiring the monkey

00:23:13.048 --> 00:23:15.488
to demonstrate that he has...

00:23:16.650 --> 00:23:21.230
A central volitional input that activates a cell that's independent from the peripheral input.

00:23:22.010 --> 00:23:28.650
So in that sense, I guess you'd have to say that any cell that has multiple

00:23:28.650 --> 00:23:36.690
inputs that can be independently controlled are demonstrating this capacity

00:23:36.690 --> 00:23:40.730
or can demonstrate the capacity for dissociation of the maps.

00:23:41.210 --> 00:23:44.710
SL. Okay, but Tony? yeah

00:23:44.710 --> 00:23:47.710
i i uh talking about extending it in your

00:23:47.710 --> 00:23:50.750
talk you mentioned that this is a volitional control

00:23:50.750 --> 00:23:53.710
yes um and but what that

00:23:53.710 --> 00:23:56.830
made me think was well the animal or

00:23:56.830 --> 00:24:01.950
if it was in a person has to actively think and concentrate on making this movement

00:24:01.950 --> 00:24:06.570
and then they can make the movement and i'm wondering to what extent you think

00:24:06.570 --> 00:24:11.550
uh that will be available then for automation so that But rather than having

00:24:11.550 --> 00:24:14.150
to concentrate on moving your arm in the trajectory,

00:24:14.470 --> 00:24:19.330
once you've learned that, will it be accessible as an automatic movement in

00:24:19.330 --> 00:24:23.430
the same way that other kind of movements are ones that are made naturally?

00:24:23.870 --> 00:24:28.030
Oh, yes, I think so. I think the natural movements are a good model of what

00:24:28.030 --> 00:24:30.530
happens in this context as well.

00:24:30.710 --> 00:24:41.430
Right. And I'm not even certain that volitional control requires an amount of conscious guidance.

00:24:41.710 --> 00:24:43.710
I mean, I think it could be...

00:24:47.108 --> 00:24:50.508
Analogous to reaching for something without thinking too much about it.

00:24:50.748 --> 00:24:55.608
So, and if you take, for example, sort of a rhythmic pattern generation,

00:24:55.928 --> 00:24:56.988
so for instance, walking.

00:24:57.188 --> 00:25:02.168
So imagine you wanted to use your system to help somebody, a spinal patient,

00:25:02.988 --> 00:25:04.228
who were unable to use their legs.

00:25:04.768 --> 00:25:09.488
I mean, could you imagine that they would be able to, through their motor cortex,

00:25:09.788 --> 00:25:15.208
drive the spinal pattern generators for walking and modulate them in a way that

00:25:15.208 --> 00:25:19.108
people could walk relatively naturally, or would you imagine it would require

00:25:19.108 --> 00:25:24.548
a lot of concentration through your motor cortex to control your legs in stepping patterns?

00:25:24.948 --> 00:25:26.768
I mean, how do you see that developing? If you had a spinal cord injury,

00:25:26.968 --> 00:25:30.488
you mean? Or if you… If you had a spinal cord injury, and imagining that we

00:25:30.488 --> 00:25:36.348
could somehow wire the motor cortex directly into circuits that control spinal

00:25:36.348 --> 00:25:38.768
pattern generators for leg locomotion.

00:25:38.768 --> 00:25:40.848
Tony, I think we might get there.

00:25:41.068 --> 00:25:46.448
We will get there because what I would like Ab to do first is to explain to

00:25:46.448 --> 00:25:51.288
us more in detail how he has been wiring into muscles and spinal cords.

00:25:51.408 --> 00:25:54.228
Then we can address that question a bit more.

00:25:54.628 --> 00:25:57.728
Well, we can give him some time to come up with an answer. Yeah,

00:25:57.748 --> 00:26:00.628
I'm going to need time to think of an answer to that one.

00:26:00.628 --> 00:26:05.868
I want to make this transition out to the biofeedback, probably because this

00:26:05.868 --> 00:26:10.008
has dominated certainly the last years of your work.

00:26:10.668 --> 00:26:14.828
So why did you end up or how did you stumble into this whole biofeedback notion?

00:26:15.068 --> 00:26:18.808
What was the transition there? Why did you go for that topic?

00:26:19.308 --> 00:26:24.608
Well, like I was saying, I was a postdoc looking for a way to learn about recording

00:26:24.608 --> 00:26:28.348
neural activity in awake monkeys and operant conditioning.

00:26:28.348 --> 00:26:34.068
Motioning and ed everett said already um demonstrated

00:26:34.068 --> 00:26:37.168
the paradigm of training a monkey to make a movement and recording

00:26:37.168 --> 00:26:41.208
motor cortex cells in relation to movement so i thought i would turn that on

00:26:41.208 --> 00:26:46.828
its head and see whether you can train the monkey to activate motor cortex cells

00:26:46.828 --> 00:26:50.588
and see what sort of movements he made it was just a fun thing to do as a postdoc

00:26:50.588 --> 00:26:55.548
right but i think the first studies came out in the early 60s right in this domain,

00:26:56.068 --> 00:27:00.028
Yeah, you're talking about Everts. For instance, yeah. Right.

00:27:00.988 --> 00:27:08.988
But now I think you sort of revolutionized a lot of that work also by using

00:27:08.988 --> 00:27:10.888
this technology that you call Neurochip.

00:27:11.848 --> 00:27:15.268
Right. This is a very recent development. Right, exactly. It's Neurochip. Right.

00:27:16.468 --> 00:27:22.828
And the question is? Well, what's a Neurochip? Why was that such an important step in all this work?

00:27:24.188 --> 00:27:27.528
Good question. I don't know. I have to think about whether I can give you a

00:27:27.528 --> 00:27:31.908
rational sequence of where that came from. Uh.

00:27:37.194 --> 00:27:43.654
I think it came out of the blue because the Neurochip was a device that had

00:27:43.654 --> 00:27:45.714
been developed in Tom Daniels' lab.

00:27:47.634 --> 00:27:51.034
And Jaideep Mavuri, a graduate student in electrical engineering,

00:27:51.194 --> 00:27:59.034
was doing the programming and application of this to investigating the control

00:27:59.034 --> 00:28:00.954
system of a moth, Manduka.

00:28:01.274 --> 00:28:04.194
And I ran into Tom Daniels and he told me about this.

00:28:04.254 --> 00:28:08.554
He's very excited. And he got me excited, and I said, well, good God,

00:28:08.614 --> 00:28:13.114
if the moths can carry this thing, a monkey could do that even easier.

00:28:13.494 --> 00:28:16.614
And we can add a lot of other stuff to it.

00:28:16.714 --> 00:28:22.334
And so I started to think about this in the monkey.

00:28:22.494 --> 00:28:30.574
And actually, this brilliant postdoctoral fellow, Andy Jackson and Jai Deep, got together.

00:28:30.574 --> 00:28:35.594
This is a good example of the productive consequences of putting together two

00:28:35.594 --> 00:28:43.794
people that have complementary talents, and they worked away at making this actually work.

00:28:43.974 --> 00:28:50.454
So it's a serendipitous encounter with Tom Daniel, I guess.

00:28:50.454 --> 00:28:56.594
But Neurochip essentially allows you to have an autonomous integrated package

00:28:56.594 --> 00:29:01.434
sitting on the skull of the monkey to sort of train, condition,

00:29:01.834 --> 00:29:04.854
remap, or exchange signals with that system.

00:29:06.325 --> 00:29:11.225
So was that the key step there, that it would be completely wireless,

00:29:11.505 --> 00:29:13.325
autonomous, locally programmed?

00:29:13.805 --> 00:29:18.605
Yeah, that was all part of what the Neurochip could do. It could operate by

00:29:18.605 --> 00:29:20.065
itself with battery power.

00:29:21.445 --> 00:29:26.685
And then the other key development is development of wire electrodes.

00:29:26.785 --> 00:29:33.205
Again, Andy Jackson gets credit for these wires that were embedded in the motor

00:29:33.205 --> 00:29:39.865
cortex and could record neural activity even during free behavior,

00:29:40.065 --> 00:29:42.325
the monkey jumping around and so forth.

00:29:42.505 --> 00:29:48.545
This was not something that could have been anticipated, the stability of these recordings.

00:29:50.125 --> 00:29:56.845
Because normally when you record neurons in the brain, it's a very fragile business

00:29:56.845 --> 00:30:00.845
subject to artifacts, movement artifacts and stuff like that.

00:30:00.845 --> 00:30:09.205
So these embedded wires actually are pretty solidly embedded and allow the recordings

00:30:09.205 --> 00:30:11.605
to continue during days of free behavior.

00:30:11.805 --> 00:30:14.325
So that was another pretty crucial

00:30:14.325 --> 00:30:18.265
element to make this neurochip paradigm work over days of behavior.

00:30:18.765 --> 00:30:24.485
And then Andy thought that it would be great, Andy Jackson thought it would

00:30:24.485 --> 00:30:30.145
be great to test this Hebbian plasticity by doing spike-triggered stimulation.

00:30:31.125 --> 00:30:37.705
And it worked. Right. So they're extracellular recordings, presumably.

00:30:38.045 --> 00:30:43.085
Yes, they are. That they're making. And so presumably you're detecting several

00:30:43.085 --> 00:30:46.985
neurons there, and you're sorting them using spike sorting?

00:30:47.185 --> 00:30:52.685
No, this is actually typically the wires next to one neuron, so you don't have to...

00:30:54.686 --> 00:30:59.306
Do the separation. But the neurochip is programmable so that it can detect the

00:30:59.306 --> 00:31:03.006
waveform of a single cell even in the presence of other cells. That's right.

00:31:03.546 --> 00:31:06.166
And how many wires can you put in at a time?

00:31:07.126 --> 00:31:11.766
Well, I think the record now is Tim Lucas put in about 30 of these wires.

00:31:12.426 --> 00:31:17.106
It's a real tour de force, but it actually worked. And it has one neuron per wire?

00:31:18.226 --> 00:31:22.526
When you're lucky, yeah. That's right. Or if you're really lucky,

00:31:22.626 --> 00:31:24.466
then you can get more than one.

00:31:24.666 --> 00:31:29.826
But the neurochip that we've been working with up till now typically has up

00:31:29.826 --> 00:31:31.286
to three channels of input.

00:31:32.146 --> 00:31:37.806
But the next one that's coming down the pipeline is going to be really cool. Lots of channels.

00:31:38.126 --> 00:31:42.746
Right. And then outputs. Our outputs, where do the outputs end up?

00:31:42.866 --> 00:31:47.806
So the outputs are typically electrical stimuli like pulses that are triggered

00:31:47.806 --> 00:31:49.466
from the action potentials of the cell.

00:31:49.766 --> 00:31:55.626
And those are delivered in the motor cortex or or a spinal cord,

00:31:55.766 --> 00:32:04.106
or eventually in muscles when we have a large enough or a potent enough stimulator

00:32:04.106 --> 00:32:05.706
to activate the muscles.

00:32:05.946 --> 00:32:10.446
So the first experiments, or one of the first, was creating what you called

00:32:10.446 --> 00:32:14.926
artificial connections between cortex and spinal cord, if I'm correct, or that muscle.

00:32:15.186 --> 00:32:19.546
Or even between cortex and cortex. So the very first experiments of Annie Jackson

00:32:19.546 --> 00:32:24.066
and Jadeep Mavuri was connecting motor cortex sites.

00:32:24.206 --> 00:32:30.046
So these were two separate sites in the precentral gyrus that could be functionally

00:32:30.046 --> 00:32:31.946
connected by this artificial loop.

00:32:32.726 --> 00:32:39.026
More recently, we've been working with artificial connections from motor cortex

00:32:39.026 --> 00:32:44.046
cells to spinal cord, and that works very nicely too.

00:32:44.226 --> 00:32:50.086
It's a little more distance, and the spinal cord is a little more challenging in the sense of stable.

00:32:51.463 --> 00:32:54.483
Implantation of the electrodes, but it can be made to work.

00:32:54.903 --> 00:32:58.043
But then one of the first experiments, I think, was one of these artificial

00:32:58.043 --> 00:33:02.903
connections from motor cortex to muscles in the wrist, if I'm correct.

00:33:03.683 --> 00:33:08.583
Yes. Through which the monkey, so the monkey had to control these muscles and

00:33:08.583 --> 00:33:13.303
then using that response, it could control a cursor on a computer screen.

00:33:13.663 --> 00:33:14.563
Well, this is the way it worked.

00:33:14.963 --> 00:33:21.223
So this wasn't actually done with a neurochip, it was actually done with rack-mounted instrumentation.

00:33:21.463 --> 00:33:26.723
But later on, the same thing was done with the neurochip.

00:33:27.583 --> 00:33:33.983
But basically, the idea was to take the activity of a cell recorded in the motor

00:33:33.983 --> 00:33:39.683
cortex and deliver stimuli into the muscle, or more accurately,

00:33:39.843 --> 00:33:41.123
the nerves that go to the muscle.

00:33:41.123 --> 00:33:46.703
And so the cell was directly controlling the stimulation of muscles,

00:33:46.803 --> 00:33:52.883
which produced twitches that allowed the monkey to succeed in playing his video

00:33:52.883 --> 00:33:54.783
game, which was to get forces into,

00:33:54.983 --> 00:33:58.503
cursors that represent forces, into targets.

00:33:58.923 --> 00:34:02.123
Okay, so it had to generate these forces with these muscles,

00:34:02.263 --> 00:34:06.143
or how exactly? Yes, that's right. The muscles generated the forces.

00:34:06.443 --> 00:34:14.383
And so this is work of Chet Moritz and the promoter that showed that if a monkey

00:34:14.383 --> 00:34:20.623
is trained to get a cursor into targets with normal muscle activity,

00:34:20.823 --> 00:34:24.383
and then you block the nerves to the muscles, paralyze them,

00:34:24.483 --> 00:34:31.263
then you can bridge this lost connection with direct connection from the cell

00:34:31.263 --> 00:34:33.463
to electrical stimulation of the muscles.

00:34:33.723 --> 00:34:38.583
And the monkey will quickly learn to drive the cell to stimulate the muscle

00:34:38.583 --> 00:34:41.243
to get the cursor into the target. Right. Okay?

00:34:41.583 --> 00:34:46.183
You were saying in your talk that actually directly stimulating the muscle creates

00:34:46.183 --> 00:34:49.783
this problem of recruiting the largest small cells in the wrong order.

00:34:50.003 --> 00:34:52.683
That's right. It's not a natural way to activate the muscle.

00:34:52.823 --> 00:34:57.623
So there's a natural recruitment order from what they're called small motor

00:34:57.623 --> 00:35:03.283
neurons which have low thresholds and low forces to large motor neurons that

00:35:03.283 --> 00:35:07.683
have more twitch tension but adapt very quickly.

00:35:07.923 --> 00:35:14.923
And electrically stimulating this produces recruitment of the large before the

00:35:14.923 --> 00:35:21.363
small, and more natural ways of activating the muscle recruits the small before the large.

00:35:22.443 --> 00:35:28.623
And the small motor units have much longer, they can be activated much longer.

00:35:28.623 --> 00:35:32.043
So it's much nicer to do it the natural way.

00:35:33.864 --> 00:35:40.144
That having been said, it's possible to work around that and have the monkey

00:35:40.144 --> 00:35:43.524
generate these movements through activating things electrically.

00:35:43.744 --> 00:35:48.544
But now, in some sense, in this experiment, you have introduced a bias by using

00:35:48.544 --> 00:35:53.684
muscles that in the life of the monkey are also used for moving objects.

00:35:54.164 --> 00:35:57.864
Yes. Oh, definitely. Right. They're just temporarily disconnected through this

00:35:57.864 --> 00:35:58.924
nerve block. Exactly right. Right.

00:35:58.964 --> 00:36:04.864
So you're reactivating the subset of muscles that would have been used under

00:36:04.864 --> 00:36:08.824
normal conditions as well to move an object in the world. Definitely. Absolutely.

00:36:09.264 --> 00:36:14.164
But now, would you imagine that you could also have acquired this mapping,

00:36:14.344 --> 00:36:23.324
could have twitched muscles just anywhere else in the body controlling that same cursor? Yes.

00:36:23.984 --> 00:36:28.084
So if you had the cursor controlled by the contraction of any other muscle,

00:36:28.244 --> 00:36:30.824
the experiment would be essentially the same.

00:36:31.304 --> 00:36:38.864
So one of the factors here is that the cells don't necessarily have to be related

00:36:38.864 --> 00:36:40.604
to the muscle that you're using.

00:36:40.604 --> 00:36:44.424
The monkey just simply needs to learn how

00:36:44.424 --> 00:36:47.244
to activate that cell which is similar to

00:36:47.244 --> 00:36:50.424
the operant conditioning paradigm and then once

00:36:50.424 --> 00:36:54.144
the monkey gets control of cell activity then

00:36:54.144 --> 00:36:59.844
you link that cell activity to stimulation of any arbitrary muscle and if the

00:36:59.844 --> 00:37:06.264
contraction of that muscle gets linked to the cursor then you're home free this

00:37:06.264 --> 00:37:11.024
is going to work i think but it would be The acquisition would be equally fast

00:37:11.024 --> 00:37:13.424
or it might take longer if it's, let's say,

00:37:13.444 --> 00:37:16.344
your calf or something like that.

00:37:16.584 --> 00:37:23.044
Yeah, I think it would be just as fast because I'm thinking that the speed of

00:37:23.044 --> 00:37:29.084
acquisition is more a function of gaining control of the cell than it is of

00:37:29.084 --> 00:37:30.764
the muscle, the nature of the muscle.

00:37:30.764 --> 00:37:35.564
So once you've picked a muscle and you've connected it to the cursor,

00:37:35.824 --> 00:37:41.304
it doesn't matter where it is physically so much, there might be some muscle

00:37:41.304 --> 00:37:42.544
properties that would...

00:37:43.264 --> 00:37:45.204
And if it's smooth muscle, would it also work?

00:37:46.064 --> 00:37:50.884
Yeah, well, good question. I'm not sure. I'd have to make sure.

00:37:54.390 --> 00:37:56.450
I think so. I would say so, but

00:37:56.450 --> 00:37:58.990
I'm sort of guessing. It might be an interesting experiment to perform.

00:37:59.350 --> 00:38:05.050
Definitely. Oh, yeah. I guess one of the issues there is how much the,

00:38:05.210 --> 00:38:11.070
going back to this point of it being volitional, how much is the monkey aware

00:38:11.070 --> 00:38:17.790
that it's using a muscle or a set of neurons that command a particular muscle to control it?

00:38:17.790 --> 00:38:23.750
So I'm imagining, okay, so if this is my arm and I just have to imagine moving

00:38:23.750 --> 00:38:27.930
my arm and it operates through your system to control those muscles,

00:38:28.130 --> 00:38:35.570
that's much more natural than if I have to imagine moving my leg and as a result my arm moves.

00:38:35.770 --> 00:38:40.490
So I can see that there might be some mapping conflicts there which wouldn't

00:38:40.490 --> 00:38:45.770
exist if you can actually target the original set of neurons, motor neurons.

00:38:47.790 --> 00:38:51.250
Or do you think I'm over-intellectualizing it? I'm thinking this through.

00:38:51.350 --> 00:38:55.510
And I'm guessing that as far as the monkey is concerned, the idea is to get

00:38:55.510 --> 00:38:58.110
that cursor into the target.

00:38:58.230 --> 00:39:01.350
And once he gets the cursor into the target with cell activity,

00:39:01.690 --> 00:39:07.810
the rest of the stuff is not something that he's cognitively concerned about.

00:39:07.950 --> 00:39:15.070
I think he's just experiencing a sort of nonlinear relationship between the

00:39:15.070 --> 00:39:20.510
cell activity and the cursor. and it's non-linear because of the recruitment

00:39:20.510 --> 00:39:22.110
of the muscle electrically.

00:39:22.190 --> 00:39:26.910
But which muscle it is, I don't think the monkey really… I guess in the case

00:39:26.910 --> 00:39:31.330
of the cursor, there isn't anything that I could naturally do to move a cursor

00:39:31.330 --> 00:39:33.330
without moving a hand or something.

00:39:33.510 --> 00:39:36.550
But in the case where you're controlling your arm….

00:39:37.753 --> 00:39:41.853
Then it would make sense, presumably, to find the arm area and see if you could

00:39:41.853 --> 00:39:45.913
connect that to the arm muscles rather than use another area?

00:39:46.173 --> 00:39:50.553
Well, you'd think so. I mean, that's the agenda of the decoding group.

00:39:50.633 --> 00:39:57.053
And they approach it from that point of view is to find cells that are naturally

00:39:57.053 --> 00:40:03.293
related to the limb that you want to control and work with that.

00:40:03.293 --> 00:40:09.513
But one of the nice things about this study is that it demonstrated that's not necessary.

00:40:09.673 --> 00:40:16.733
You can actually work with any motor cortex cell that the monkey can control

00:40:16.733 --> 00:40:18.353
and link it to the muscle.

00:40:18.913 --> 00:40:23.833
And pretty much any motor cortex cell can be volitionally controlled.

00:40:24.393 --> 00:40:29.953
But now you also mentioned that the cell tuning itself does not predict the

00:40:29.953 --> 00:40:31.133
ability to control the cursor.

00:40:31.493 --> 00:40:37.213
Exactly. But in addition, you also said that you could triple the number of

00:40:37.213 --> 00:40:40.893
neurons that can be recruited to now control the factor.

00:40:41.173 --> 00:40:44.613
So what does that exactly mean? So what that means is in that experiment,

00:40:44.873 --> 00:40:52.053
about two-thirds of the cells that were used in the study were not really showing

00:40:52.053 --> 00:40:57.673
directional tuning in relation to wrist movements, torques around the wrist.

00:40:57.673 --> 00:41:05.053
And all of those cells could be volitionally controlled and linked to a flexor

00:41:05.053 --> 00:41:09.853
extensor musculature and the monkey would very quickly transition from control

00:41:09.853 --> 00:41:15.813
of the cell to stimulating those muscles and generating the cursor movement.

00:41:16.393 --> 00:41:24.193
And in that sense it liberates the paradigm from the necessity of finding cells related to the wrist,

00:41:25.613 --> 00:41:32.373
and this two-thirds is a number that pertains to the sample in this particular sample.

00:41:33.991 --> 00:41:38.611
But actually, the number is much larger than two-thirds.

00:41:39.431 --> 00:41:44.711
I mean, when you start considering cells in other areas, leg area,

00:41:45.091 --> 00:41:48.371
non-motor cortex, pre-motor cortex, who knows?

00:41:48.431 --> 00:41:53.911
It's to be determined how many different cortical areas you can demonstrate

00:41:53.911 --> 00:41:54.911
this volitional control.

00:41:54.911 --> 00:42:02.631
Then you've increased the space of, let's say,

00:42:02.691 --> 00:42:08.711
source cells or the number of cells that you could recruit into this sort of

00:42:08.711 --> 00:42:18.131
a paradigm enormously as opposed to if you had to find cells that were related to the limb.

00:42:18.131 --> 00:42:23.671
And this becomes a crucial issue in the case of stroke, where the area that

00:42:23.671 --> 00:42:29.031
might control the hand muscle is lost, and you don't have any cells that are related to the hand.

00:42:29.031 --> 00:42:35.771
But you'd have the possibility of going to cells in another area that would

00:42:35.771 --> 00:42:43.391
normally have involved control of something else, but it's in an area that's still viable,

00:42:43.531 --> 00:42:49.551
and those cells can be then used to control stimulation of hand muscles. Right.

00:42:50.631 --> 00:42:56.451
So most of your experiments, as I understand it, then you're recording a train

00:42:56.451 --> 00:43:01.191
of spikes from a motor neuron and motor cortex, and you're relaying that exact

00:43:01.191 --> 00:43:04.451
train to the target with some latency.

00:43:04.651 --> 00:43:08.711
Is that right? Well, first of all, I think the motor neurons,

00:43:08.891 --> 00:43:13.131
strictly speaking, are the cells in the spinal cord that connect directly to the muscle.

00:43:13.171 --> 00:43:19.331
Sorry, I mean the cortical. And so motor cortex cells are cells that may or

00:43:19.331 --> 00:43:21.271
may not project to the spinal cord. Yeah, exactly.

00:43:21.571 --> 00:43:30.631
But you can take their activity and use it to control the stimulation of the muscle.

00:43:30.711 --> 00:43:35.931
I think that's… So they're cortical projection neurons. But what I found surprising

00:43:35.931 --> 00:43:42.031
was that you get a hit with just about every neuron that you find,

00:43:42.211 --> 00:43:47.131
and presumably in some of these areas of cortex there are interneurons that

00:43:47.131 --> 00:43:52.631
could you be hitting, or do you know if you're always targeting projection neurons?

00:43:52.631 --> 00:43:56.151
No, no, we don't know, and it doesn't matter, I don't think.

00:43:56.711 --> 00:44:02.751
I'm pretty sure on the basis of how many cells in the motor cortex have been

00:44:02.751 --> 00:44:07.951
successfully controlled, I would say that some of them are projection,

00:44:08.271 --> 00:44:10.491
some of them are probably not projection.

00:44:10.491 --> 00:44:16.811
Injection, there's a bias toward getting cells with very large action potentials

00:44:16.811 --> 00:44:22.631
that are reliably isolated over long periods of time, so that would mean we're biasing toward.

00:44:23.503 --> 00:44:30.763
Layer five pyramidal neurons. Pyramidal now meaning the morphology of the cell is called pyramidal.

00:44:31.103 --> 00:44:34.523
But that doesn't mean they go into the pyramidal tract of the spinal cord.

00:44:34.643 --> 00:44:36.103
It just means that that's the shape.

00:44:36.403 --> 00:44:41.623
And so they generate these nice big action potentials that are pretty stable.

00:44:41.763 --> 00:44:45.283
But where they project, we don't know.

00:44:45.383 --> 00:44:51.443
And frankly, don't worry too much about because because the key is to use their

00:44:51.443 --> 00:44:57.183
activity and use the ability of the animal to control that activity in a useful

00:44:57.183 --> 00:44:59.463
way. But it does seem surprising.

00:44:59.723 --> 00:45:02.683
I mean, if you think about cortical circuits, you don't think about all the

00:45:02.683 --> 00:45:03.863
neurons being the same there.

00:45:03.923 --> 00:45:09.063
And you think about interneurons having some modulatory role related to projection

00:45:09.063 --> 00:45:13.523
neurons, and therefore possibly producing very different kinds of signals and

00:45:13.523 --> 00:45:15.823
being used to very different kinds of inputs.

00:45:15.823 --> 00:45:19.003
Outputs so whereas i can understand maybe a projection neuron

00:45:19.003 --> 00:45:22.703
which is targeting a motor neuron you could

00:45:22.703 --> 00:45:25.983
you could take the output of that and direct it down

00:45:25.983 --> 00:45:28.763
to the the downstream muscle and that makes some

00:45:28.763 --> 00:45:32.843
sense but if it's an interneuron you're having to really the brain is having

00:45:32.843 --> 00:45:35.763
to really reprogram that interneuron to do something it's never done before

00:45:35.763 --> 00:45:41.923
well i think you're over uh projecting function into cell types right first

00:45:41.923 --> 00:45:46.463
of all uh your idea of how the The interneuron would modulate this activity,

00:45:46.583 --> 00:45:51.263
also require some degree of flexibility of the way they're recruited.

00:45:52.283 --> 00:45:59.043
So even interneurons could be as easily volitionally controllable, I would think.

00:46:00.503 --> 00:46:04.923
There's another issue which has to do with the morphology of most of these interneurons

00:46:04.923 --> 00:46:08.243
are sort of spherically,

00:46:09.323 --> 00:46:14.603
as spherical dendrites and closed fields and so they're less likely to be actually

00:46:14.603 --> 00:46:20.563
recorded by these tungsten electrodes But there's another aspect of course that

00:46:20.563 --> 00:46:22.283
we might not want to interpret.

00:46:23.404 --> 00:46:28.224
A circuit too literal in terms of a single cell. I mean, these neurons are embedded

00:46:28.224 --> 00:46:29.824
in quite a dense volume of cells.

00:46:30.084 --> 00:46:33.764
And maybe what you're training here is the response of this whole volume.

00:46:33.964 --> 00:46:38.324
And activity within the volume will be highly correlated, whether that's layer

00:46:38.324 --> 00:46:42.444
5 pyramid or an interneuron or a layer 4 stellate cell.

00:46:42.644 --> 00:46:46.304
Exactly. They will be tightly coupled in their responses. Is that how you think about it?

00:46:46.364 --> 00:46:52.544
I do, yes. Yes, I think that we're talking about a large distributed group of

00:46:52.544 --> 00:46:58.264
interconnected neurons of which we sample one or maybe two,

00:46:58.444 --> 00:47:03.024
but they're all going to be co-activated more or less.

00:47:03.524 --> 00:47:06.724
Do you have any physiological evidence for that or is there any?

00:47:06.724 --> 00:47:12.704
Well, when you do record neighboring cells, originally when we recorded more

00:47:12.704 --> 00:47:18.284
than one and conditioned one, often the neighboring cell was also modulated, but not always.

00:47:18.864 --> 00:47:24.284
So, it depends on where the other cells are physically located.

00:47:24.704 --> 00:47:28.844
Even though they're synaptically interconnected, they don't necessarily have to be neighbors.

00:47:29.344 --> 00:47:35.004
And my view of it is that this is a fairly distributed population of cells.

00:47:35.444 --> 00:47:40.804
Are you intending to measure this directly or you think that's not so much of a problem at this time?

00:47:53.130 --> 00:47:58.050
Maybe even three in some cases. And the question is, what are the other cells doing?

00:47:58.230 --> 00:48:02.670
So we'll have an answer once all that torrent of data gets analyzed.

00:48:03.030 --> 00:48:07.430
Exactly. So now, but the next step here was that you sort of went straight.

00:48:07.590 --> 00:48:11.750
So you bypassed out a muscle because then the next step became,

00:48:11.870 --> 00:48:14.450
look, maybe we can just wire it straight into the spinal cord.

00:48:14.990 --> 00:48:18.370
That became the next exercise. I mean, the last experiment we discussed was

00:48:18.370 --> 00:48:21.070
really like moving the cursor with muscle twitches.

00:48:21.130 --> 00:48:25.270
Right. But now the next step was spinal cord. Spinal cord. Why was that an important target?

00:48:25.510 --> 00:48:33.610
So the spinal cord is important because it recruits the motor units more naturally, as we just discussed.

00:48:33.870 --> 00:48:43.630
And it also, spinal stimuli tend to recruit synergistic groups of muscles together.

00:48:43.770 --> 00:48:51.070
So that's what you ultimately want. If you try to achieve that by muscle stimulation,

00:48:51.250 --> 00:48:55.030
you'd have to implant a lot of muscles and learn how to co-activate them.

00:48:55.190 --> 00:49:02.730
Whereas in the spinal cord, you're at a place where you can recruit them sort of naturally as a group.

00:49:03.170 --> 00:49:06.810
But that would suggest that you would know how the spinal cord is organized

00:49:06.810 --> 00:49:13.870
and how activity in spinal motor neurons maps to coherent movement patterns.

00:49:14.210 --> 00:49:21.290
Is that known? No, it's an empirical issue. So we have done the mapping of the spinal cord.

00:49:22.270 --> 00:49:28.770
That is to say, drive an electrode systematically through the spinal cord and measure the output.

00:49:29.490 --> 00:49:35.410
And what the take-home message from those experiments is that you can't really

00:49:35.410 --> 00:49:36.970
predict what the output is.

00:49:37.050 --> 00:49:41.210
It's an empirical issue. And the reason you can't predict is because what you're

00:49:41.210 --> 00:49:45.350
stimulating is lots of fibers, and fibers are intertwined.

00:49:46.397 --> 00:49:50.917
Tangled in unpredictable ways as a function of location in the spinal cord.

00:49:51.917 --> 00:49:59.817
So ultimately, if you want to get a spinal site where you get contraction of

00:49:59.817 --> 00:50:03.517
a particular muscle, you're going to have to hunt for it. And then when you

00:50:03.517 --> 00:50:05.157
find it, you're going to have to hang on to it.

00:50:05.457 --> 00:50:11.017
So it's not something that you can reliably predict beforehand.

00:50:11.837 --> 00:50:16.877
So the mapping, another way to put it is the mapping of the cord isn't as clean

00:50:16.877 --> 00:50:23.297
cut as the mapping of output effects in motor cortex for example. Right.

00:50:23.937 --> 00:50:29.657
But then the neural response that you mapped onto spinal cord was response in

00:50:29.657 --> 00:50:31.597
the gamma range. That's about 40 hertz type.

00:50:31.877 --> 00:50:37.757
Oh, there's one study you're talking about. So there you were able to have the

00:50:37.757 --> 00:50:41.977
monkey modulate 40 hertz responses in its motor cortex, I assume.

00:50:42.997 --> 00:50:45.837
To then move a cursor along one dimension of movement.

00:50:46.397 --> 00:50:49.537
Right yeah um but now do you

00:50:49.537 --> 00:50:52.337
think you can move that much further could you go also again

00:50:52.337 --> 00:50:55.497
from two single cells driving spinal cord absolutely

00:50:55.497 --> 00:50:58.237
that's we've done that we have the data we just

00:50:58.237 --> 00:51:02.377
need to analyze it okay so there are no limitations there in terms of the spinal

00:51:02.377 --> 00:51:07.457
control you can get i think there probably are uh limitations in the number

00:51:07.457 --> 00:51:13.357
of independent outputs you can expect to find and how practical it is to go

00:51:13.357 --> 00:51:15.517
searching for the ones you need.

00:51:16.037 --> 00:51:19.837
So there are going to be challenges to make this, for example,

00:51:20.057 --> 00:51:22.177
a useful prosthetic agenda.

00:51:22.617 --> 00:51:25.137
That's of course the obvious target, right, where you would say,

00:51:25.217 --> 00:51:29.837
well, we can sort of, in case of spinal lesions, we can directly bypass the

00:51:29.837 --> 00:51:34.537
lesion, go straight from M1 into spinal cord and move the legs or something

00:51:34.537 --> 00:51:35.857
like this. Is that a feasible outlook?

00:51:36.437 --> 00:51:38.457
Eventually, I think it will be there.

00:51:41.337 --> 00:51:45.337
So, there's another thing that we're working on that might relate to that,

00:51:45.377 --> 00:51:48.037
having to do with stimulating,

00:51:48.257 --> 00:51:53.477
not intraspinally, which is an evasive technology, and the spinal cord likes

00:51:53.477 --> 00:51:57.097
to eventually reject these electrodes that are in there.

00:51:57.875 --> 00:52:02.615
So, the thing we're going to test next is whether we can get similar effects

00:52:02.615 --> 00:52:07.735
by stimulating the surface of the spinal cord instead of doing it intraspinally.

00:52:07.855 --> 00:52:14.615
So, the question is whether we'll be able to find enough differentiation of

00:52:14.615 --> 00:52:20.435
the output effects that can be evoked from the surface of the spinal cord or

00:52:20.435 --> 00:52:26.675
whether we actually need to poke these wires into the cord and search out.

00:52:26.675 --> 00:52:30.115
But would the monkey be able to also correct now this mapping for itself?

00:52:30.215 --> 00:52:33.255
Imagine we go from motor cortex into spinal cord.

00:52:33.495 --> 00:52:37.595
We want to have, let's say, a certain walking gait. Imagine that's our target.

00:52:37.955 --> 00:52:42.455
But initially we get some strange twitches because we're not placed correctly in the spinal cord.

00:52:42.875 --> 00:52:48.695
Do you think there's a possibility that it's the motor commands coming in from

00:52:48.695 --> 00:52:54.275
the motor cortex that induce that error and that can be remapped evolutionally?

00:52:54.295 --> 00:53:01.395
Yes. Is that what you expect? To some extent, that's true, but the big factor

00:53:01.395 --> 00:53:05.375
is whether your output effects from the spinal cord, which are sort of a given

00:53:05.375 --> 00:53:07.635
basis functions, as it were,

00:53:07.795 --> 00:53:10.775
include the movements that you want to generate.

00:53:10.975 --> 00:53:16.815
If they don't include it, there's no way that cortical control is going to be

00:53:16.815 --> 00:53:20.535
able to generate something that the stimulation doesn't produce.

00:53:20.535 --> 00:53:28.475
But if you have it in your repertoire, then my prediction is that the brain,

00:53:28.655 --> 00:53:32.535
given sufficient time to learn to optimize that stimulation,

00:53:32.875 --> 00:53:38.035
can recruit those outputs to functionally useful ends.

00:53:38.415 --> 00:53:42.335
So you have great confidence in the plastic capabilities of brains. I do.

00:53:42.635 --> 00:53:45.475
Actually, yeah, that's right. It's an article of faith. We'll see.

00:53:45.675 --> 00:53:47.475
I mean, there may be other complications.

00:53:47.735 --> 00:53:52.735
I don't know. You know, things like you don't want to stimulate in the spinal

00:53:52.735 --> 00:53:58.055
cord dorsal horn too much because that's pretty aversive and you need to be in the right place.

00:53:58.315 --> 00:54:03.515
And so there's issues like that. Sure. But you have a backup plan,

00:54:03.615 --> 00:54:05.875
I guess, which is to go directly to the target muscles.

00:54:07.175 --> 00:54:09.195
Well, yes. If you can't go through the spinal cord. That's right.

00:54:09.195 --> 00:54:13.915
But as we said, the target muscles directly have this recruitment problem.

00:54:14.075 --> 00:54:19.635
Yeah, yeah. And the number of muscles that you can activate directly,

00:54:19.955 --> 00:54:22.255
each one requires its own set of electrodes.

00:54:23.195 --> 00:54:28.075
So that gets to be a lot of wires. I guess if you can understand more about

00:54:28.075 --> 00:54:33.235
how the spinal cord is doing the recruitment, and then you could modulate your

00:54:33.235 --> 00:54:37.715
cortical signal to be more like a spinal cord signal, is that?

00:54:39.088 --> 00:54:44.648
Yeah, well... Then you're getting more towards the decoding, recoding guys, I guess.

00:54:44.648 --> 00:54:47.408
I'm thinking more in terms of

00:54:47.408 --> 00:54:50.348
the brain being able

00:54:50.348 --> 00:54:59.048
to recruit the motor cortex cells that control the stimulation in a way that

00:54:59.048 --> 00:55:07.648
makes that output from that stimulation be not practical for the movements that

00:55:07.648 --> 00:55:09.428
the subject wants to attain.

00:55:09.428 --> 00:55:18.468
And so, that whole thing involves a learning process that could take days and

00:55:18.468 --> 00:55:22.128
could be supported by this sort of an implant.

00:55:22.368 --> 00:55:27.568
That's part of this idea, the advantage of an implant is that it provides the

00:55:27.568 --> 00:55:31.888
subject plenty of time to optimize this control.

00:55:32.248 --> 00:55:39.508
So, that's an important… But now, in some sense, there must be a minimum command

00:55:39.508 --> 00:55:44.628
set that has to go down from motor cortex into spinal cord to get coherent movement.

00:55:45.008 --> 00:55:50.968
That's right. You need as many independent outputs as you want to independently

00:55:50.968 --> 00:55:53.848
control the stimulus outputs.

00:55:54.328 --> 00:55:58.328
So, in the healthy brain, what would that be?

00:55:58.548 --> 00:56:04.748
So, how many, let's say, I want to induce a walking gait. so how many signals

00:56:04.748 --> 00:56:08.868
do have to come out of my motor cortex to induce that or to control that.

00:56:10.380 --> 00:56:14.500
Well, that's a good question.

00:56:14.660 --> 00:56:21.060
So the answer to that question depends on how many components of the gate you want to control.

00:56:21.300 --> 00:56:26.460
So you can use computers to help along a little bit to generate patterns of activity.

00:56:27.440 --> 00:56:32.620
And then in the extreme, you might imagine you just have one cell that turns it on and off.

00:56:33.420 --> 00:56:37.900
On the other hand, on the other extreme, you want outputs that control every

00:56:37.900 --> 00:56:43.840
component of the gate. And that's a much more formidable challenge and probably

00:56:43.840 --> 00:56:46.200
unlikely to be the way that we want to go.

00:56:46.640 --> 00:56:52.980
So I think you're going to wind up with some combination of pre-programmed patterns

00:56:52.980 --> 00:56:56.680
that are controlled by a smaller set of cells.

00:56:57.020 --> 00:56:59.320
So getting back to your question, how many?

00:56:59.760 --> 00:57:04.320
I'm guessing it would be useful to have at least a half a dozen cells.

00:57:05.420 --> 00:57:08.420
Because it relates a little bit to this issue of bandwidth

00:57:08.420 --> 00:57:11.340
and capacity right and now in some

00:57:11.340 --> 00:57:14.400
sense we can get away with a very limited bandwidth

00:57:14.400 --> 00:57:18.620
because we only control very few degrees of freedom we move cursors over one

00:57:18.620 --> 00:57:24.180
or two dimensions and so on right so so do you see this as as a as a bottleneck

00:57:24.180 --> 00:57:28.840
as a real limitation for technology or do you think it will just scale up to

00:57:28.840 --> 00:57:32.540
any set of degrees of freedom. The independent control?

00:57:33.140 --> 00:57:34.840
It's a good question. I don't know.

00:57:36.040 --> 00:57:42.080
I don't think it's unlimited because there are ultimately relationships between these cells.

00:57:42.160 --> 00:57:48.440
They're not totally independent, so they're probably going to be going to run

00:57:48.440 --> 00:57:53.060
into the fact that it's not quite so easy to do them all individually. Right.

00:57:54.401 --> 00:58:00.441
It is the cautious answer. My intuition is that, again, given enough time,

00:58:00.801 --> 00:58:06.001
I bet you the brain is going to demonstrate that it can handle a goodly number

00:58:06.001 --> 00:58:08.681
of degrees of freedom. But you have to give it enough time.

00:58:09.001 --> 00:58:14.941
Right. So then another important aspect you touched upon was learning.

00:58:15.321 --> 00:58:19.961
So you really also have configured your system that's in a very controlled condition.

00:58:19.961 --> 00:58:24.561
You can look at different learning paradigms, different forms of Hebbian learning.

00:58:25.601 --> 00:58:27.481
So what were the key things that you have learned?

00:58:29.401 --> 00:58:38.161
Well, with regard to Hebbian learning, we've found that you can induce the synaptic

00:58:38.161 --> 00:58:42.361
plasticity that is mediated by Hebbian mechanisms.

00:58:42.361 --> 00:58:48.561
That is to say, you can strengthen the connections between neurons,

00:58:48.821 --> 00:58:57.481
and in this motor cortex experiment where stimulation and recording were both done in motor cortex,

00:58:57.741 --> 00:59:06.861
those changes lasted a surprisingly long time, 10 days in one case, and it never reverted.

00:59:06.981 --> 00:59:11.821
But in other experiments, it hasn't lasted that long. So another lesson is you

00:59:11.821 --> 00:59:17.421
can make those changes, but don't count on them staying around in the absence

00:59:17.421 --> 00:59:18.821
of continued conditioning.

00:59:19.301 --> 00:59:24.281
But now the first experiment in motor cortex, basically you use your neurochip

00:59:24.281 --> 00:59:32.801
to impose certain correlation patterns between neurons with the idea that,

00:59:32.861 --> 00:59:34.901
okay, what fires together wires together.

00:59:34.901 --> 00:59:39.521
So if you force them to be synchronized in their response, you assume this experiment,

00:59:39.581 --> 00:59:40.701
they also would wire together.

00:59:41.081 --> 00:59:44.441
That's right. And then the idea would be the response properties would become more similar.

00:59:46.841 --> 00:59:47.721
Well, yeah.

00:59:49.198 --> 00:59:53.718
The connections are strengthened. We don't know really all that much about the response properties.

00:59:54.478 --> 00:59:59.198
All we know is that the connections that were mediating the output effects by

00:59:59.198 --> 01:00:02.238
stimulating those sites had changed.

01:00:02.538 --> 01:00:08.858
And what that meant for the way these cells were active is an interesting question

01:00:08.858 --> 01:00:13.218
for the future, but not one that we could… But wait, but your measure,

01:00:13.298 --> 01:00:19.278
your performance measure was how the cell's response correlated with a movement

01:00:19.278 --> 01:00:20.738
pattern, a movement of the monkey.

01:00:21.498 --> 01:00:28.318
So, no, the measure in the motor cortex study was the movements that were generated

01:00:28.318 --> 01:00:34.558
by electrically stimulating the sites in the cortex that were recorded from

01:00:34.558 --> 01:00:37.298
and that were stimulated, and then there was a control site.

01:00:37.298 --> 01:00:44.798
And so the measure was really a somewhat artificial measure of what you could

01:00:44.798 --> 01:00:47.278
evoke with a train of stimuli from these sites.

01:00:47.758 --> 01:00:55.498
But those outputs changed in a way that is most easily understood as strengthening

01:00:55.498 --> 01:00:59.258
a connection between the recording site and the stimulation site.

01:00:59.418 --> 01:01:05.478
The outputs were consistently in a direction that could be explained by strengthening those connections.

01:01:05.478 --> 01:01:12.198
Now, the question of what that meant for the way cells in these two areas are

01:01:12.198 --> 01:01:17.498
correlated, whether that was changing or not, is not something that we measured,

01:01:17.738 --> 01:01:19.558
although it would be great to be able to do that.

01:01:19.558 --> 01:01:24.178
But I thought your data were pointing to this because you stimulated the cell

01:01:24.178 --> 01:01:29.098
or the site where you had been imposing a correlation with the recording site.

01:01:29.418 --> 01:01:33.998
Then you looked at what movement it correlated later on with individual stimulation.

01:01:34.498 --> 01:01:39.318
And you saw that it was sort of a similar movement as in the recording site

01:01:39.318 --> 01:01:42.098
and dissimilar from the control site.

01:01:42.438 --> 01:01:45.438
So you just have to remember that the movements we're talking about are movements

01:01:45.438 --> 01:01:49.398
that are evoked by electrical stimulation, not movements that were generated by the monkey.

01:01:49.558 --> 01:01:53.238
Right, sure, absolutely. So I think there's a lot to be.

01:01:54.464 --> 01:02:00.784
Pursued here in terms of what the change in these synaptic connections means

01:02:00.784 --> 01:02:04.604
for change in functional interactions between these sites.

01:02:04.764 --> 01:02:09.004
So that would be a great follow-up experiment.

01:02:09.424 --> 01:02:13.424
But then the other thing that you emphasized very strongly in these experiments

01:02:13.424 --> 01:02:18.784
on plasticity was that it's not just any form of Hebbian learning that is driven

01:02:18.784 --> 01:02:22.604
by correlation, but it really follows these ideas of spike-time dependent learning.

01:02:22.804 --> 01:02:28.764
Yes. Where you depress or potentiate the synapse given the latency between pre-

01:02:28.764 --> 01:02:29.724
and post-synaptic activity.

01:02:30.024 --> 01:02:35.344
Exactly. Right. So that means at very short latencies, you would be depressing.

01:02:35.524 --> 01:02:38.384
And at longer latencies, you would be potentiating.

01:02:39.004 --> 01:02:44.304
And at some point, you don't do anything. So what was the exact observation there?

01:02:44.304 --> 01:02:51.504
So the ultimate mechanism is that if you have a presynaptic input that's activated

01:02:51.504 --> 01:02:54.044
relative to a postsynaptic cell activation,

01:02:55.164 --> 01:03:01.764
the strengthening of the connection is dependent on the time between the presynaptic

01:03:01.764 --> 01:03:02.684
and postsynaptic input.

01:03:02.904 --> 01:03:08.424
So if that presynaptic input comes within 50 milliseconds of the postsynaptic

01:03:08.424 --> 01:03:12.564
activation, You can strengthen the connection, and that's consistent with what

01:03:12.564 --> 01:03:15.744
we saw in the motor cortex study and the corticospinal study.

01:03:17.204 --> 01:03:21.504
It turns out that spike timing dependent plasticity function,

01:03:21.764 --> 01:03:27.584
which I've just described in the sense of increasing the strength of connection,

01:03:27.904 --> 01:03:33.424
actually goes in the other direction of predicting a decrease in the strength

01:03:33.424 --> 01:03:39.704
of the connection if the postsynaptic cell is activated prior to the presynaptic input.

01:03:39.704 --> 01:03:46.004
So that's the non-causal way of activating these cells. In other words...

01:03:46.830 --> 01:03:52.870
The postsynaptic cell would probably be driven by the presynaptic input in the

01:03:52.870 --> 01:03:55.130
normal causal way. That would strengthen connections.

01:03:55.510 --> 01:04:02.070
But if you do it the other way around, then the bidirectional spike timing dependent

01:04:02.070 --> 01:04:08.810
plasticity function, if I can use that mouthful of words, is actually predicting a decrease.

01:04:09.090 --> 01:04:16.210
And this is the thing that we could demonstrate in the corticospinal system

01:04:16.390 --> 01:04:23.610
could be achieved when the delay between the spike and the spinal stimulus was zero.

01:04:23.970 --> 01:04:29.390
When the spinal stimulus was delivered as fast as possible after the cortical spike,

01:04:29.650 --> 01:04:34.410
it actually activated the postsynaptic cells in the spinal cord prior to the

01:04:34.410 --> 01:04:41.870
arrival of the corticospinal volley, and that resulted in a decrease in the connection.

01:04:42.110 --> 01:04:49.450
But that's a part of this curve that is only possible to explore through this

01:04:49.450 --> 01:04:56.290
means in situations where you have a long conduction time from the recording to the stimulated site.

01:04:56.630 --> 01:05:00.150
But then there's something interesting about this, right? Because what you found,

01:05:00.290 --> 01:05:04.590
this learning window you found, sort of shows the strongest potentiation about

01:05:04.590 --> 01:05:09.510
40 milliseconds and it showed a depression about 10 millisecond latency.

01:05:10.030 --> 01:05:16.310
So this is telling you something about the causal structure that these circuits try to maintain, right?

01:05:16.410 --> 01:05:22.050
Because they're saying, look, okay, if presynaptic activity is closer to 10

01:05:22.050 --> 01:05:25.370
milliseconds to my post-synaptic activity, then I'm not going to learn this.

01:05:25.450 --> 01:05:28.070
Then there's something wrong in the causal structure I'm dealing with.

01:05:28.290 --> 01:05:32.510
So is there anything special about these latencies you find in that spinal circuit

01:05:32.510 --> 01:05:35.710
with respect to the behavior that it controls?

01:05:36.670 --> 01:05:39.610
The leap from this spike timing

01:05:39.610 --> 01:05:43.970
dependent plasticity to behavior is an order

01:05:43.970 --> 01:05:51.590
of magnitude difference in the number of cells that are involved in generating

01:05:51.590 --> 01:06:00.070
the behavior in relation to the number of cells that are involved in demonstrating this plasticity.

01:06:00.070 --> 01:06:06.050
So ultimately, one would expect a relationship to behavior, but it doesn't work.

01:06:10.128 --> 01:06:17.888
Appear that our Hebbian plasticity studies are strong enough to mediate observable

01:06:17.888 --> 01:06:18.788
changes in the behavior.

01:06:19.028 --> 01:06:22.388
They're really small things, but they're nevertheless,

01:06:24.508 --> 01:06:30.748
cellular mechanisms that underlie more broad changes in behavior, I think.

01:06:30.908 --> 01:06:36.628
Right, but the first thing that's interesting, of course, is that you can induce

01:06:36.628 --> 01:06:40.728
this kind of Hebbian type plasticity using spike-time dependent learning in

01:06:40.728 --> 01:06:44.228
a spinal cord circuit that I find very surprising, right?

01:06:44.248 --> 01:06:48.128
Because it means you can also reshape, remodel these spinal circuits dependent

01:06:48.128 --> 01:06:52.508
on the kinds of pre-imposed signals that they pinch on it, which should also

01:06:52.508 --> 01:06:55.988
give us hope in terms of rewiring spinal circuits, interestingly enough.

01:06:56.548 --> 01:06:59.728
But then the other thing is, of course, that maybe these very short latencies

01:06:59.728 --> 01:07:03.608
you want to maintain because you go from, let's say, your motor cortex.

01:07:04.828 --> 01:07:09.768
Towards the periphery, projections from the spinal circuit into the skeletal

01:07:09.768 --> 01:07:12.888
muscle system and its feedback will occur in a certain time window.

01:07:13.768 --> 01:07:17.548
And these must be aligned correctly, right? Control signals from the cortex

01:07:17.548 --> 01:07:22.248
should precede any type of activity you receive from the periphery.

01:07:22.268 --> 01:07:25.868
So you can imagine that these latencies that seem now critical in this learning

01:07:25.868 --> 01:07:29.708
window that you found map to the latencies you might find in this transduction

01:07:29.708 --> 01:07:33.828
of control signals to the skeletal muscle system that you in the end are controlling.

01:07:34.008 --> 01:07:36.048
That's a little bit what I was after.

01:07:37.108 --> 01:07:42.188
Yeah, the changes that you see, are any of those long-lasting?

01:07:42.368 --> 01:07:46.208
I mean, can we think of it as remodeling the circuit or is it something that

01:07:46.208 --> 01:07:50.048
goes away? Well, that was one of the interesting observations in the study of

01:07:50.048 --> 01:07:53.168
the corticospinal plasticity is how long does it last?

01:07:53.588 --> 01:07:59.428
So in some cases, it lasted for at least two days after the end of the conditioning.

01:07:59.428 --> 01:08:02.628
And possibly even longer. We just didn't measure it.

01:08:02.928 --> 01:08:09.928
In other cases, in the majority probably, it dropped after the end of the conditioning

01:08:09.928 --> 01:08:12.688
in the next couple of days. And so...

01:08:14.432 --> 01:08:18.912
That raises the question of how one would make these changes more long-lasting.

01:08:19.032 --> 01:08:21.792
It's possible that you could use other interventions.

01:08:22.812 --> 01:08:26.292
Well, first of all, you could just do this conditioning much longer than we

01:08:26.292 --> 01:08:28.252
did. We just did it for a couple days.

01:08:29.132 --> 01:08:33.012
This is Yukio. We is actually Yukio Nishimura and Steve Promutter.

01:08:33.712 --> 01:08:36.352
And they did it for a couple days. delays.

01:08:38.752 --> 01:08:45.292
If you did it much longer, it's conceivable that this plasticity could lead

01:08:45.292 --> 01:08:48.752
to structural changes that would be more permanent.

01:08:49.472 --> 01:08:58.492
Number one, amount of conditioning. Number two, you could use neural modulators to change the strength.

01:08:58.492 --> 01:09:04.812
Like BDNF, for example, could be something that could make these changes last longer.

01:09:05.932 --> 01:09:08.352
And third, you could do things like polarization.

01:09:09.512 --> 01:09:13.972
DC polarization could also make these things last longer.

01:09:14.852 --> 01:09:20.472
But now, the next step in the process, the last experiment you were discussing,

01:09:20.692 --> 01:09:29.552
where you really try to now modulate or change reward systems themselves in the brain, right?

01:09:29.612 --> 01:09:33.172
Where you try to train towards nucleus accumbens, for instance.

01:09:33.892 --> 01:09:38.512
So what was the experiment there? So the experiment there was to use the neurochip

01:09:38.512 --> 01:09:41.132
to record activity somewhere.

01:09:41.132 --> 01:09:48.812
Somewhere, in our case it started with muscles, and then deliver stimuli that

01:09:48.812 --> 01:09:51.872
were driven by that activity in a rewarding site.

01:09:52.272 --> 01:09:58.292
So this is an intracranial site where electrical stimulation is behaviorally

01:09:58.292 --> 01:10:05.932
rewarding in the sense that the monkey will make movements to generate stimulation.

01:10:05.932 --> 01:10:11.352
And so in this paradigm with the neurochip,

01:10:11.412 --> 01:10:18.412
the idea was to close that loop from muscle activity to intracranial stimulation

01:10:18.412 --> 01:10:21.092
of the reinforcement site, nucleus accumbens.

01:10:21.112 --> 01:10:26.832
And the monkey very quickly learned to do what could track that muscle and amazingly

01:10:26.832 --> 01:10:33.352
did it for quite a long period of alternating on and off periods. Would it not rest?

01:10:34.012 --> 01:10:37.112
If you would leave it, would it do it forever? Well,

01:10:37.332 --> 01:10:42.832
we took it out to 20 hours, and the monkey was still doing pretty well at the

01:10:42.832 --> 01:10:48.952
end of 20 hours, although I think the data shows that control dropped somewhere

01:10:48.952 --> 01:10:50.552
in between, probably at night.

01:10:50.992 --> 01:10:53.892
But did you also observe that the monkey did not eat or drink?

01:10:53.952 --> 01:10:57.352
It was just self-stimulating? No, I wouldn't say that.

01:10:57.452 --> 01:11:02.092
I think it's not as potent as in the rat.

01:11:02.312 --> 01:11:07.752
I think those are fairly dramatic demonstrations in the rat that that's all

01:11:07.752 --> 01:11:10.992
they'll do until they die. But monkeys are...

01:11:12.686 --> 01:11:16.126
It's not that compelling a stimulation.

01:11:16.386 --> 01:11:21.006
He can drop it. I wish it were more compelling because we're now trying to obviously

01:11:21.006 --> 01:11:23.786
do this with neural activity, neural patterns.

01:11:23.906 --> 01:11:29.086
And it can work, but the results have been not quite as robust yet.

01:11:29.346 --> 01:11:33.246
But now you mentioned that you actually started the Lulis experiments because

01:11:33.246 --> 01:11:38.146
you wanted to get a handle on temporal coding, which I found a rather surprising step.

01:11:38.526 --> 01:11:43.286
Oh, yeah. Very interesting. Right. So that's a speculation that if this could

01:11:43.286 --> 01:11:45.366
all work, this is way out here now.

01:11:46.106 --> 01:11:51.706
So if this could all work, this is a paradigm that could be used to test the

01:11:51.706 --> 01:11:52.926
idea of temporal coding.

01:11:53.086 --> 01:11:59.546
So what we mean by that is, first of all, temporal coding is the assumption

01:11:59.546 --> 01:12:03.606
that information is coded in a temporal sequence of spikes,

01:12:04.286 --> 01:12:10.246
as opposed to rate coding, which says that information is coded in the average firing rate of cells.

01:12:11.606 --> 01:12:19.366
So, we know that rate coding works in the sensory and the motor system,

01:12:19.506 --> 01:12:23.586
and probably in a lot of the association areas.

01:12:24.326 --> 01:12:29.706
The question of whether temporal coding is actually being exploited by the brain

01:12:29.706 --> 01:12:32.526
is an interesting question, because if it were true,

01:12:32.746 --> 01:12:40.126
it would increase enormously the bandwidth, as it were, for neural computation.

01:12:40.746 --> 01:12:46.946
It would mean that you could use this additional dimension of time of spikes,

01:12:47.206 --> 01:12:51.366
in the ultimate case, for information coding.

01:12:51.786 --> 01:12:56.186
So that's a great idea. It's very enticing. It's, as I said,

01:12:56.186 --> 01:12:58.286
in dire need of experimental support. port.

01:12:59.386 --> 01:13:06.366
And the idea, the fantasy here is that if we implemented this neurochip to reward

01:13:06.366 --> 01:13:11.166
patterns of activity, temporal patterns, we could test whether those temporal

01:13:11.166 --> 01:13:13.606
patterns were really part of a volitionally controllable,

01:13:14.386 --> 01:13:15.486
repertoire of behavior.

01:13:15.826 --> 01:13:18.846
Right, exactly. That would be very exciting. It would be great.

01:13:19.186 --> 01:13:21.006
I should live that long, yeah.

01:13:21.766 --> 01:13:25.906
So then, so where do you see the future of this going?

01:13:26.226 --> 01:13:30.966
I mean, this is all really it's amazing work it really shows also the incredible

01:13:30.966 --> 01:13:35.586
plasticity of the brain and of course it raises all sorts of questions about

01:13:35.586 --> 01:13:40.406
application and neuroprosthetics and all sorts of amazing things and also something

01:13:40.406 --> 01:13:42.026
we are confronted in the field but rather,

01:13:43.046 --> 01:13:46.966
you know amazing demonstrations what might be possible.

01:13:48.479 --> 01:13:52.779
So where do you see this really go in this field? Well, I think the exciting

01:13:52.779 --> 01:13:58.499
prospect, looking in the long run, decades ahead, or do you want to talk just

01:13:58.499 --> 01:13:59.719
about the next couple of years?

01:13:59.919 --> 01:14:04.439
Both. Well, next couple of years will be modest increments.

01:14:04.839 --> 01:14:09.839
Well, we'll have many more channels in the next neurochip, so it'll be more than modest, I think.

01:14:09.839 --> 01:14:17.659
But in the long run, the exciting thing for me is to imagine that we would get

01:14:17.659 --> 01:14:26.279
to the point where the brain could directly access the computational power of an implant, let's say.

01:14:26.279 --> 01:14:35.479
And instead of interacting with our iPhone computers through normal sensory

01:14:35.479 --> 01:14:40.899
channels, we would have that computer chip available for direct interaction.

01:14:40.899 --> 01:14:47.619
So this is total science fiction today because the major hurdle that I can see

01:14:47.619 --> 01:14:53.639
is the problem of tapping the right parts of the brain to connect to the computer

01:14:53.639 --> 01:14:56.899
directly in both directions.

01:14:56.899 --> 01:15:03.039
Probably more seriously in the direction of the stimulation being delivered

01:15:03.039 --> 01:15:09.159
back into the brain in a way that the brain can decode and use.

01:15:09.439 --> 01:15:14.999
Because electrical stimulation is a very crude way of activating a mess of cells

01:15:14.999 --> 01:15:16.539
with different functions.

01:15:17.159 --> 01:15:26.779
And so what we need is a much more specific way of delivering the feedback from

01:15:26.779 --> 01:15:27.959
the computer into the brain.

01:15:28.259 --> 01:15:32.339
Right now, the best way to do it is through normal sensory channels,

01:15:32.379 --> 01:15:39.079
through tactile input or maybe stimulating the...

01:15:39.079 --> 01:15:44.139
But what you envision there is also that you can create completely new capabilities to a brain.

01:15:44.819 --> 01:15:49.859
Exactly. Yeah, yeah. Yeah, right. So the big fun thing to think about is what

01:15:49.859 --> 01:15:58.239
would you do to complement the computational power of the brain with the computational

01:15:58.239 --> 01:16:00.059
power of the computer chip?

01:16:01.339 --> 01:16:07.099
And, well, of course, the science fiction movies and writers have anticipated all this.

01:16:07.199 --> 01:16:12.279
And so we have to look to what they can imagine could happen.

01:16:12.279 --> 01:16:26.639
And I think there are some serious problems to direct transfer and bidirectional ways of information.

01:16:26.899 --> 01:16:31.979
I think right now, the most efficient way to do it is through normal sensory motor channels.

01:16:32.219 --> 01:16:36.839
But in the future, if we want to fantasize about where things could go,

01:16:37.519 --> 01:16:41.199
it would be interesting to see how you could uh.

01:16:42.412 --> 01:16:47.712
Implant a chip that the brain could access the computational power of directly.

01:16:48.812 --> 01:16:55.872
So right now, even the Google Glass is a device that makes this interaction,

01:16:56.852 --> 01:17:05.092
makes the computer small enough and the interaction easy enough that it could

01:17:05.092 --> 01:17:07.932
operate, but it's still operating through normal channels.

01:17:08.272 --> 01:17:12.492
Right. So the question is, and it's a serious question as to whether this is

01:17:12.492 --> 01:17:14.212
actually going to be possible.

01:17:14.372 --> 01:17:20.052
The question is whether you can bypass the normal channels and get useful functional

01:17:20.052 --> 01:17:25.952
interactions by directly stimulating and directly recording from the brain.

01:17:27.032 --> 01:17:32.072
I'm guessing that's a pretty much open question. And if I were to bet real money

01:17:32.072 --> 01:17:36.932
on it, I would say it's 50-50, maybe even.

01:17:36.932 --> 01:17:43.972
And maybe my imagination isn't robust enough to imagine this actually working in a practical way.

01:17:46.052 --> 01:17:50.652
So that's a kind of cognitive prosthesis, a sort of chip that helps you think better.

01:17:51.052 --> 01:17:55.672
And in terms of sort of a physical prosthetic, like a hand or an arm.

01:17:56.712 --> 01:18:01.232
How's the work you're going to do going to impact in that perhaps more near term?

01:18:01.232 --> 01:18:08.552
Well, there are other people that are doing more practical work on the more

01:18:08.552 --> 01:18:14.672
standard brain-machine interface, that is to say, an output to a robotic arm.

01:18:15.192 --> 01:18:20.372
But they're also now working toward a bidirectional device where it's not just

01:18:20.372 --> 01:18:27.272
output, but also there's some feedback from the, let's say, it's a prosthetic arm.

01:18:27.272 --> 01:18:35.012
And so you want to know what the joint positions are and what the forces,

01:18:35.152 --> 01:18:39.972
contact forces are, and you want to feed that information back somehow.

01:18:40.152 --> 01:18:44.012
So that does create a recurrent loop through this process.

01:18:44.978 --> 01:18:48.378
External machine this this artificial arm

01:18:48.378 --> 01:18:51.098
and that's where i

01:18:51.098 --> 01:18:53.738
think we'll be seeing progress because a lot

01:18:53.738 --> 01:18:59.278
of people are hammering away at it and right yeah so now so so we started out

01:18:59.278 --> 01:19:04.458
with your issue a metaphor of the self in the brain yes and with that you meant

01:19:04.458 --> 01:19:08.138
that that there are systems in the brain that can sort of control other systems

01:19:08.138 --> 01:19:12.078
in the brain right right so where Where is that self-system?

01:19:12.378 --> 01:19:17.158
Wonderful question. I wish I knew. It's somewhere behind my eyes.

01:19:17.598 --> 01:19:22.478
No, I'm just kidding. It's somewhere in there. It's distributed, but it's very robust.

01:19:23.018 --> 01:19:31.078
And I think it's a key element of the brain that needs further investigation.

01:19:32.018 --> 01:19:35.798
I think a lot of… But the key thing is you really see some distinction here.

01:19:35.798 --> 01:19:40.278
There's really a controller, some central system that is, again,

01:19:40.298 --> 01:19:45.118
controlling then these parts of the brain you are measuring from. Yes. Distinct systems.

01:19:45.618 --> 01:19:50.878
That's right. And it's merged anatomically, really distinct. Oh, not necessarily.

01:19:50.958 --> 01:19:59.818
I'm guessing that this controller includes this peripheral component of output,

01:19:59.958 --> 01:20:09.658
because it actually has to modify its control of that output if you have situations like I described,

01:20:09.658 --> 01:20:15.598
where you change the relationship between the output that you get and the output

01:20:15.598 --> 01:20:17.118
that you want to generate.

01:20:18.058 --> 01:20:29.138
Then the volitional controller part of this brain is adapting to the new contingencies.

01:20:29.438 --> 01:20:35.798
But that adaptation obviously requires the peripheral elements as well as the central ones.

01:20:35.798 --> 01:20:40.958
But, you know, in a sense, functionally, you can imagine that there's a separate

01:20:40.958 --> 01:20:48.578
question of why you want to get from A to B and why you want to adapt your ability to get from A to B.

01:20:48.578 --> 01:20:56.718
I'm thinking now of this Moussa Valdi and Shadmir study where A and B were points

01:20:56.718 --> 01:21:04.298
on a plane and getting from A to B required overcoming some strange force field.

01:21:04.298 --> 01:21:09.478
You still have the volitional controller part intending to get from A to B.

01:21:10.758 --> 01:21:16.158
And the mechanisms that adapt to this artificial force field that you have to

01:21:16.158 --> 01:21:20.778
overcome to get there as smoothly as possible,

01:21:21.438 --> 01:21:26.118
that's all conventionally considered part of the motor system.

01:21:27.078 --> 01:21:28.298
But I'm guessing that...

01:21:30.133 --> 01:21:34.133
That it could also be usefully considered as part of the volitional controller.

01:21:34.373 --> 01:21:45.273
So it's a semantic quibble. But is that self in the brain beyond the conditioning?

01:21:45.453 --> 01:21:49.593
Is there a core system that cannot in itself be conditioned following these

01:21:49.593 --> 01:21:50.753
methods you have been describing?

01:21:51.413 --> 01:21:53.813
That's a good question. That is immune to this manipulation?

01:21:54.693 --> 01:22:00.653
I don't know. What do you think? I don't think so. Right. it's really part of...

01:22:01.953 --> 01:22:04.813
I think it's more integrated. I don't think there's a clear distinction between

01:22:04.813 --> 01:22:07.493
the self and then the rest of the brain. I think it's much more integrated.

01:22:09.913 --> 01:22:13.093
So then to finish up, I have two questions.

01:22:13.193 --> 01:22:18.213
So look, we were talking earlier, right? So your first experiments in the 60s

01:22:18.213 --> 01:22:21.553
as a physicist, so you're in this field for really a long time.

01:22:22.213 --> 01:22:28.073
You have accomplished an amazing amount of work, gained a lot of insight in the brain, but then...

01:22:29.033 --> 01:22:35.133
Given all this experience, what is Ebb's law that we should adhere to studying the brain?

01:22:36.673 --> 01:22:43.453
Don't get caught up in conceptual preconceptions about what part of the brain does what.

01:22:43.873 --> 01:22:53.393
I've developed a more empirical approach to investigating the brain.

01:22:53.393 --> 01:22:57.793
This would be my advice to people getting into the field.

01:22:58.813 --> 01:23:05.713
So you can have preconceived ideas about this part of the brain does this and

01:23:05.713 --> 01:23:11.873
that part does that and design hypothesis-driven experiments based on that.

01:23:11.993 --> 01:23:21.193
But usually that often winds up being problematic and unsuccessful and boring in some cases.

01:23:21.193 --> 01:23:28.573
And so I think my take-home lesson of my life is that you have to just go in there and do it.

01:23:28.693 --> 01:23:36.053
So, for example, this question of what happens if you connect motor cortex to

01:23:36.053 --> 01:23:37.113
spinal cord stimulation.

01:23:38.693 --> 01:23:43.313
When I was first proposed, the NIH study section was appalled that this was

01:23:43.313 --> 01:23:49.193
way too complicated to be worth funding,

01:23:49.933 --> 01:23:53.873
because we don't know what the outcome is and where's our hypothesis?

01:23:54.233 --> 01:23:58.333
Fishing expedition, blah, blah, all those things. Well, geez,

01:23:58.513 --> 01:24:03.353
I mean, that's how you make progress is to actually do some new stuff and not

01:24:03.353 --> 01:24:06.753
be constrained by all your preconceptions as to how it would work.

01:24:07.413 --> 01:24:09.593
So I basically had to reformulate,

01:24:10.661 --> 01:24:16.101
motivation for that experiment to say we want to know what the behavioral adaptation

01:24:16.101 --> 01:24:20.141
is to having this artificial connection in parallel with the biological connection

01:24:20.141 --> 01:24:23.621
and we don't know what the outcome is but the fact that the monkey controls

01:24:23.621 --> 01:24:26.301
both the artificial and the biological connections,

01:24:26.861 --> 01:24:30.781
means that it's a meaningful questions to see how he integrates the two right

01:24:30.781 --> 01:24:35.621
and the study section like that better so this is all good yeah but anyway don't

01:24:35.621 --> 01:24:39.901
localize that's the key thing right And the last question, so five years from

01:24:39.901 --> 01:24:41.821
now, we're going to come visit you there in Washington.

01:24:42.241 --> 01:24:47.421
Okay. And we're going to say, look, you gave us this prediction in 2013,

01:24:47.641 --> 01:24:51.921
and today we're going to come check whether it was true or false.

01:24:52.181 --> 01:24:55.681
So what's this one prediction you would like to commit yourself to today?

01:24:56.501 --> 01:24:59.841
That we still have a lot of interesting questions to investigate.

01:25:02.501 --> 01:25:07.621
That's too easy. I know. Oh, well. All right, Epp.

01:25:07.641 --> 01:25:10.801
Thank you very much for this conversation. You bet. Thank you. Thank you.

01:25:12.880 --> 01:25:18.480
Music.

01:25:18.101 --> 01:25:23.941
The CSN Podcast was produced by the Convergent Science Network of Biometrics

01:25:23.941 --> 01:25:30.361
and Biohybrid Systems, a project funded by the European 7th Research Framework Program.

01:25:31.841 --> 01:25:37.181
For more interviews, recorded lectures, or upcoming conferences in the field

01:25:37.181 --> 01:25:43.201
of biometrics and biohybrid systems, go to csnnetwork.com.

01:25:43.920 --> 01:25:51.600
Music.