Daniel Wolpert: The real reason for brains

341,413 views ・ 2011-11-03

TED


Please double-click on the English subtitles below to play the video.

00:15
I'm a neuroscientist.
0
15260
2000
00:17
And in neuroscience,
1
17260
2000
00:19
we have to deal with many difficult questions about the brain.
2
19260
3000
00:22
But I want to start with the easiest question
3
22260
2000
00:24
and the question you really should have all asked yourselves at some point in your life,
4
24260
3000
00:27
because it's a fundamental question
5
27260
2000
00:29
if we want to understand brain function.
6
29260
2000
00:31
And that is, why do we and other animals
7
31260
2000
00:33
have brains?
8
33260
2000
00:35
Not all species on our planet have brains,
9
35260
3000
00:38
so if we want to know what the brain is for,
10
38260
2000
00:40
let's think about why we evolved one.
11
40260
2000
00:42
Now you may reason that we have one
12
42260
2000
00:44
to perceive the world or to think,
13
44260
2000
00:46
and that's completely wrong.
14
46260
2000
00:48
If you think about this question for any length of time,
15
48260
3000
00:51
it's blindingly obvious why we have a brain.
16
51260
2000
00:53
We have a brain for one reason and one reason only,
17
53260
3000
00:56
and that's to produce adaptable and complex movements.
18
56260
3000
00:59
There is no other reason to have a brain.
19
59260
2000
01:01
Think about it.
20
61260
2000
01:03
Movement is the only way you have
21
63260
2000
01:05
of affecting the world around you.
22
65260
2000
01:07
Now that's not quite true. There's one other way, and that's through sweating.
23
67260
3000
01:10
But apart from that,
24
70260
2000
01:12
everything else goes through contractions of muscles.
25
72260
2000
01:14
So think about communication --
26
74260
2000
01:16
speech, gestures, writing, sign language --
27
76260
3000
01:19
they're all mediated through contractions of your muscles.
28
79260
3000
01:22
So it's really important to remember
29
82260
2000
01:24
that sensory, memory and cognitive processes are all important,
30
84260
4000
01:28
but they're only important
31
88260
2000
01:30
to either drive or suppress future movements.
32
90260
2000
01:32
There can be no evolutionary advantage
33
92260
2000
01:34
to laying down memories of childhood
34
94260
2000
01:36
or perceiving the color of a rose
35
96260
2000
01:38
if it doesn't affect the way you're going to move later in life.
36
98260
3000
01:41
Now for those who don't believe this argument,
37
101260
2000
01:43
we have trees and grass on our planet without the brain,
38
103260
2000
01:45
but the clinching evidence is this animal here --
39
105260
2000
01:47
the humble sea squirt.
40
107260
2000
01:49
Rudimentary animal, has a nervous system,
41
109260
3000
01:52
swims around in the ocean in its juvenile life.
42
112260
2000
01:54
And at some point of its life,
43
114260
2000
01:56
it implants on a rock.
44
116260
2000
01:58
And the first thing it does in implanting on that rock, which it never leaves,
45
118260
3000
02:01
is to digest its own brain and nervous system
46
121260
3000
02:04
for food.
47
124260
2000
02:06
So once you don't need to move,
48
126260
2000
02:08
you don't need the luxury of that brain.
49
128260
3000
02:11
And this animal is often taken
50
131260
2000
02:13
as an analogy to what happens at universities
51
133260
2000
02:15
when professors get tenure,
52
135260
2000
02:17
but that's a different subject.
53
137260
2000
02:19
(Applause)
54
139260
2000
02:21
So I am a movement chauvinist.
55
141260
3000
02:24
I believe movement is the most important function of the brain --
56
144260
2000
02:26
don't let anyone tell you that it's not true.
57
146260
2000
02:28
Now if movement is so important,
58
148260
2000
02:30
how well are we doing
59
150260
2000
02:32
understanding how the brain controls movement?
60
152260
2000
02:34
And the answer is we're doing extremely poorly; it's a very hard problem.
61
154260
2000
02:36
But we can look at how well we're doing
62
156260
2000
02:38
by thinking about how well we're doing building machines
63
158260
2000
02:40
which can do what humans can do.
64
160260
2000
02:42
Think about the game of chess.
65
162260
2000
02:44
How well are we doing determining what piece to move where?
66
164260
3000
02:47
If you pit Garry Kasparov here, when he's not in jail,
67
167260
3000
02:50
against IBM's Deep Blue,
68
170260
2000
02:52
well the answer is IBM's Deep Blue will occasionally win.
69
172260
3000
02:55
And I think if IBM's Deep Blue played anyone in this room, it would win every time.
70
175260
3000
02:58
That problem is solved.
71
178260
2000
03:00
What about the problem
72
180260
2000
03:02
of picking up a chess piece,
73
182260
2000
03:04
dexterously manipulating it and putting it back down on the board?
74
184260
3000
03:07
If you put a five year-old child's dexterity against the best robots of today,
75
187260
3000
03:10
the answer is simple:
76
190260
2000
03:12
the child wins easily.
77
192260
2000
03:14
There's no competition at all.
78
194260
2000
03:16
Now why is that top problem so easy
79
196260
2000
03:18
and the bottom problem so hard?
80
198260
2000
03:20
One reason is a very smart five year-old
81
200260
2000
03:22
could tell you the algorithm for that top problem --
82
202260
2000
03:24
look at all possible moves to the end of the game
83
204260
2000
03:26
and choose the one that makes you win.
84
206260
2000
03:28
So it's a very simple algorithm.
85
208260
2000
03:30
Now of course there are other moves,
86
210260
2000
03:32
but with vast computers we approximate
87
212260
2000
03:34
and come close to the optimal solution.
88
214260
2000
03:36
When it comes to being dexterous,
89
216260
2000
03:38
it's not even clear what the algorithm is you have to solve to be dexterous.
90
218260
2000
03:40
And we'll see you have to both perceive and act on the world,
91
220260
2000
03:42
which has a lot of problems.
92
222260
2000
03:44
But let me show you cutting-edge robotics.
93
224260
2000
03:46
Now a lot of robotics is very impressive,
94
226260
2000
03:48
but manipulation robotics is really just in the dark ages.
95
228260
3000
03:51
So this is the end of a Ph.D. project
96
231260
2000
03:53
from one of the best robotics institutes.
97
233260
2000
03:55
And the student has trained this robot
98
235260
2000
03:57
to pour this water into a glass.
99
237260
2000
03:59
It's a hard problem because the water sloshes about, but it can do it.
100
239260
3000
04:02
But it doesn't do it with anything like the agility of a human.
101
242260
3000
04:05
Now if you want this robot to do a different task,
102
245260
3000
04:08
that's another three-year Ph.D. program.
103
248260
3000
04:11
There is no generalization at all
104
251260
2000
04:13
from one task to another in robotics.
105
253260
2000
04:15
Now we can compare this
106
255260
2000
04:17
to cutting-edge human performance.
107
257260
2000
04:19
So what I'm going to show you is Emily Fox
108
259260
2000
04:21
winning the world record for cup stacking.
109
261260
3000
04:24
Now the Americans in the audience will know all about cup stacking.
110
264260
2000
04:26
It's a high school sport
111
266260
2000
04:28
where you have 12 cups you have to stack and unstack
112
268260
2000
04:30
against the clock in a prescribed order.
113
270260
2000
04:32
And this is her getting the world record in real time.
114
272260
3000
04:39
(Laughter)
115
279260
8000
04:47
(Applause)
116
287260
5000
04:52
And she's pretty happy.
117
292260
2000
04:54
We have no idea what is going on inside her brain when she does that,
118
294260
2000
04:56
and that's what we'd like to know.
119
296260
2000
04:58
So in my group, what we try to do
120
298260
2000
05:00
is reverse engineer how humans control movement.
121
300260
3000
05:03
And it sounds like an easy problem.
122
303260
2000
05:05
You send a command down, it causes muscles to contract.
123
305260
2000
05:07
Your arm or body moves,
124
307260
2000
05:09
and you get sensory feedback from vision, from skin, from muscles and so on.
125
309260
3000
05:12
The trouble is
126
312260
2000
05:14
these signals are not the beautiful signals you want them to be.
127
314260
2000
05:16
So one thing that makes controlling movement difficult
128
316260
2000
05:18
is, for example, sensory feedback is extremely noisy.
129
318260
3000
05:21
Now by noise, I do not mean sound.
130
321260
3000
05:24
We use it in the engineering and neuroscience sense
131
324260
2000
05:26
meaning a random noise corrupting a signal.
132
326260
2000
05:28
So the old days before digital radio when you were tuning in your radio
133
328260
3000
05:31
and you heard "crrcckkk" on the station you wanted to hear,
134
331260
2000
05:33
that was the noise.
135
333260
2000
05:35
But more generally, this noise is something that corrupts the signal.
136
335260
3000
05:38
So for example, if you put your hand under a table
137
338260
2000
05:40
and try to localize it with your other hand,
138
340260
2000
05:42
you can be off by several centimeters
139
342260
2000
05:44
due to the noise in sensory feedback.
140
344260
2000
05:46
Similarly, when you put motor output on movement output,
141
346260
2000
05:48
it's extremely noisy.
142
348260
2000
05:50
Forget about trying to hit the bull's eye in darts,
143
350260
2000
05:52
just aim for the same spot over and over again.
144
352260
2000
05:54
You have a huge spread due to movement variability.
145
354260
3000
05:57
And more than that, the outside world, or task,
146
357260
2000
05:59
is both ambiguous and variable.
147
359260
2000
06:01
The teapot could be full, it could be empty.
148
361260
2000
06:03
It changes over time.
149
363260
2000
06:05
So we work in a whole sensory movement task soup of noise.
150
365260
4000
06:09
Now this noise is so great
151
369260
2000
06:11
that society places a huge premium
152
371260
2000
06:13
on those of us who can reduce the consequences of noise.
153
373260
3000
06:16
So if you're lucky enough to be able to knock a small white ball
154
376260
3000
06:19
into a hole several hundred yards away using a long metal stick,
155
379260
3000
06:22
our society will be willing to reward you
156
382260
2000
06:24
with hundreds of millions of dollars.
157
384260
3000
06:27
Now what I want to convince you of
158
387260
2000
06:29
is the brain also goes through a lot of effort
159
389260
2000
06:31
to reduce the negative consequences
160
391260
2000
06:33
of this sort of noise and variability.
161
393260
2000
06:35
And to do that, I'm going to tell you about a framework
162
395260
2000
06:37
which is very popular in statistics and machine learning of the last 50 years
163
397260
3000
06:40
called Bayesian decision theory.
164
400260
2000
06:42
And it's more recently a unifying way
165
402260
3000
06:45
to think about how the brain deals with uncertainty.
166
405260
3000
06:48
And the fundamental idea is you want to make inferences and then take actions.
167
408260
3000
06:51
So let's think about the inference.
168
411260
2000
06:53
You want to generate beliefs about the world.
169
413260
2000
06:55
So what are beliefs?
170
415260
2000
06:57
Beliefs could be: where are my arms in space?
171
417260
2000
06:59
Am I looking at a cat or a fox?
172
419260
2000
07:01
But we're going to represent beliefs with probabilities.
173
421260
3000
07:04
So we're going to represent a belief
174
424260
2000
07:06
with a number between zero and one --
175
426260
2000
07:08
zero meaning I don't believe it at all, one means I'm absolutely certain.
176
428260
3000
07:11
And numbers in between give you the gray levels of uncertainty.
177
431260
3000
07:14
And the key idea to Bayesian inference
178
434260
2000
07:16
is you have two sources of information
179
436260
2000
07:18
from which to make your inference.
180
438260
2000
07:20
You have data,
181
440260
2000
07:22
and data in neuroscience is sensory input.
182
442260
2000
07:24
So I have sensory input, which I can take in to make beliefs.
183
444260
3000
07:27
But there's another source of information, and that's effectively prior knowledge.
184
447260
3000
07:30
You accumulate knowledge throughout your life in memories.
185
450260
3000
07:33
And the point about Bayesian decision theory
186
453260
2000
07:35
is it gives you the mathematics
187
455260
2000
07:37
of the optimal way to combine
188
457260
2000
07:39
your prior knowledge with your sensory evidence
189
459260
2000
07:41
to generate new beliefs.
190
461260
2000
07:43
And I've put the formula up there.
191
463260
2000
07:45
I'm not going to explain what that formula is, but it's very beautiful.
192
465260
2000
07:47
And it has real beauty and real explanatory power.
193
467260
3000
07:50
And what it really says, and what you want to estimate,
194
470260
2000
07:52
is the probability of different beliefs
195
472260
2000
07:54
given your sensory input.
196
474260
2000
07:56
So let me give you an intuitive example.
197
476260
2000
07:58
Imagine you're learning to play tennis
198
478260
3000
08:01
and you want to decide where the ball is going to bounce
199
481260
2000
08:03
as it comes over the net towards you.
200
483260
2000
08:05
There are two sources of information
201
485260
2000
08:07
Bayes' rule tells you.
202
487260
2000
08:09
There's sensory evidence -- you can use visual information auditory information,
203
489260
3000
08:12
and that might tell you it's going to land in that red spot.
204
492260
3000
08:15
But you know that your senses are not perfect,
205
495260
3000
08:18
and therefore there's some variability of where it's going to land
206
498260
2000
08:20
shown by that cloud of red,
207
500260
2000
08:22
representing numbers between 0.5 and maybe 0.1.
208
502260
3000
08:26
That information is available in the current shot,
209
506260
2000
08:28
but there's another source of information
210
508260
2000
08:30
not available on the current shot,
211
510260
2000
08:32
but only available by repeated experience in the game of tennis,
212
512260
3000
08:35
and that's that the ball doesn't bounce
213
515260
2000
08:37
with equal probability over the court during the match.
214
517260
2000
08:39
If you're playing against a very good opponent,
215
519260
2000
08:41
they may distribute it in that green area,
216
521260
2000
08:43
which is the prior distribution,
217
523260
2000
08:45
making it hard for you to return.
218
525260
2000
08:47
Now both these sources of information carry important information.
219
527260
2000
08:49
And what Bayes' rule says
220
529260
2000
08:51
is that I should multiply the numbers on the red by the numbers on the green
221
531260
3000
08:54
to get the numbers of the yellow, which have the ellipses,
222
534260
3000
08:57
and that's my belief.
223
537260
2000
08:59
So it's the optimal way of combining information.
224
539260
3000
09:02
Now I wouldn't tell you all this if it wasn't that a few years ago,
225
542260
2000
09:04
we showed this is exactly what people do
226
544260
2000
09:06
when they learn new movement skills.
227
546260
2000
09:08
And what it means
228
548260
2000
09:10
is we really are Bayesian inference machines.
229
550260
2000
09:12
As we go around, we learn about statistics of the world and lay that down,
230
552260
4000
09:16
but we also learn
231
556260
2000
09:18
about how noisy our own sensory apparatus is,
232
558260
2000
09:20
and then combine those
233
560260
2000
09:22
in a real Bayesian way.
234
562260
2000
09:24
Now a key part to the Bayesian is this part of the formula.
235
564260
3000
09:27
And what this part really says
236
567260
2000
09:29
is I have to predict the probability
237
569260
2000
09:31
of different sensory feedbacks
238
571260
2000
09:33
given my beliefs.
239
573260
2000
09:35
So that really means I have to make predictions of the future.
240
575260
3000
09:38
And I want to convince you the brain does make predictions
241
578260
2000
09:40
of the sensory feedback it's going to get.
242
580260
2000
09:42
And moreover, it profoundly changes your perceptions
243
582260
2000
09:44
by what you do.
244
584260
2000
09:46
And to do that, I'll tell you
245
586260
2000
09:48
about how the brain deals with sensory input.
246
588260
2000
09:50
So you send a command out,
247
590260
3000
09:53
you get sensory feedback back,
248
593260
2000
09:55
and that transformation is governed
249
595260
2000
09:57
by the physics of your body and your sensory apparatus.
250
597260
3000
10:00
But you can imagine looking inside the brain.
251
600260
2000
10:02
And here's inside the brain.
252
602260
2000
10:04
You might have a little predictor, a neural simulator,
253
604260
2000
10:06
of the physics of your body and your senses.
254
606260
2000
10:08
So as you send a movement command down,
255
608260
2000
10:10
you tap a copy of that off
256
610260
2000
10:12
and run it into your neural simulator
257
612260
2000
10:14
to anticipate the sensory consequences of your actions.
258
614260
4000
10:18
So as I shake this ketchup bottle,
259
618260
2000
10:20
I get some true sensory feedback as the function of time in the bottom row.
260
620260
3000
10:23
And if I've got a good predictor, it predicts the same thing.
261
623260
3000
10:26
Well why would I bother doing that?
262
626260
2000
10:28
I'm going to get the same feedback anyway.
263
628260
2000
10:30
Well there's good reasons.
264
630260
2000
10:32
Imagine, as I shake the ketchup bottle,
265
632260
2000
10:34
someone very kindly comes up to me and taps it on the back for me.
266
634260
3000
10:37
Now I get an extra source of sensory information
267
637260
2000
10:39
due to that external act.
268
639260
2000
10:41
So I get two sources.
269
641260
2000
10:43
I get you tapping on it, and I get me shaking it,
270
643260
3000
10:46
but from my senses' point of view,
271
646260
2000
10:48
that is combined together into one source of information.
272
648260
3000
10:51
Now there's good reason to believe
273
651260
2000
10:53
that you would want to be able to distinguish external events from internal events.
274
653260
3000
10:56
Because external events are actually much more behaviorally relevant
275
656260
3000
10:59
than feeling everything that's going on inside my body.
276
659260
3000
11:02
So one way to reconstruct that
277
662260
2000
11:04
is to compare the prediction --
278
664260
2000
11:06
which is only based on your movement commands --
279
666260
2000
11:08
with the reality.
280
668260
2000
11:10
Any discrepancy should hopefully be external.
281
670260
3000
11:13
So as I go around the world,
282
673260
2000
11:15
I'm making predictions of what I should get, subtracting them off.
283
675260
3000
11:18
Everything left over is external to me.
284
678260
2000
11:20
What evidence is there for this?
285
680260
2000
11:22
Well there's one very clear example
286
682260
2000
11:24
where a sensation generated by myself feels very different
287
684260
2000
11:26
then if generated by another person.
288
686260
2000
11:28
And so we decided the most obvious place to start
289
688260
2000
11:30
was with tickling.
290
690260
2000
11:32
It's been known for a long time, you can't tickle yourself
291
692260
2000
11:34
as well as other people can.
292
694260
2000
11:36
But it hasn't really been shown, it's because you have a neural simulator,
293
696260
3000
11:39
simulating your own body
294
699260
2000
11:41
and subtracting off that sense.
295
701260
2000
11:43
So we can bring the experiments of the 21st century
296
703260
3000
11:46
by applying robotic technologies to this problem.
297
706260
3000
11:49
And in effect, what we have is some sort of stick in one hand attached to a robot,
298
709260
3000
11:52
and they're going to move that back and forward.
299
712260
2000
11:54
And then we're going to track that with a computer
300
714260
2000
11:56
and use it to control another robot,
301
716260
2000
11:58
which is going to tickle their palm with another stick.
302
718260
2000
12:00
And then we're going to ask them to rate a bunch of things
303
720260
2000
12:02
including ticklishness.
304
722260
2000
12:04
I'll show you just one part of our study.
305
724260
2000
12:06
And here I've taken away the robots,
306
726260
2000
12:08
but basically people move with their right arm sinusoidally back and forward.
307
728260
3000
12:11
And we replay that to the other hand with a time delay.
308
731260
3000
12:14
Either no time delay,
309
734260
2000
12:16
in which case light would just tickle your palm,
310
736260
2000
12:18
or with a time delay of two-tenths of three-tenths of a second.
311
738260
4000
12:22
So the important point here
312
742260
2000
12:24
is the right hand always does the same things -- sinusoidal movement.
313
744260
3000
12:27
The left hand always is the same and puts sinusoidal tickle.
314
747260
3000
12:30
All we're playing with is a tempo causality.
315
750260
2000
12:32
And as we go from naught to 0.1 second,
316
752260
2000
12:34
it becomes more ticklish.
317
754260
2000
12:36
As you go from 0.1 to 0.2,
318
756260
2000
12:38
it becomes more ticklish at the end.
319
758260
2000
12:40
And by 0.2 of a second,
320
760260
2000
12:42
it's equivalently ticklish
321
762260
2000
12:44
to the robot that just tickled you without you doing anything.
322
764260
2000
12:46
So whatever is responsible for this cancellation
323
766260
2000
12:48
is extremely tightly coupled with tempo causality.
324
768260
3000
12:51
And based on this illustration, we really convinced ourselves in the field
325
771260
3000
12:54
that the brain's making precise predictions
326
774260
2000
12:56
and subtracting them off from the sensations.
327
776260
3000
12:59
Now I have to admit, these are the worst studies my lab has ever run.
328
779260
3000
13:02
Because the tickle sensation on the palm comes and goes,
329
782260
2000
13:04
you need large numbers of subjects
330
784260
2000
13:06
with these stars making them significant.
331
786260
2000
13:08
So we were looking for a much more objective way
332
788260
2000
13:10
to assess this phenomena.
333
790260
2000
13:12
And in the intervening years I had two daughters.
334
792260
2000
13:14
And one thing you notice about children in backseats of cars on long journeys,
335
794260
3000
13:17
they get into fights --
336
797260
2000
13:19
which started with one of them doing something to the other, the other retaliating.
337
799260
3000
13:22
It quickly escalates.
338
802260
2000
13:24
And children tend to get into fights which escalate in terms of force.
339
804260
3000
13:27
Now when I screamed at my children to stop,
340
807260
2000
13:29
sometimes they would both say to me
341
809260
2000
13:31
the other person hit them harder.
342
811260
3000
13:34
Now I happen to know my children don't lie,
343
814260
2000
13:36
so I thought, as a neuroscientist,
344
816260
2000
13:38
it was important how I could explain
345
818260
2000
13:40
how they were telling inconsistent truths.
346
820260
2000
13:42
And we hypothesize based on the tickling study
347
822260
2000
13:44
that when one child hits another,
348
824260
2000
13:46
they generate the movement command.
349
826260
2000
13:48
They predict the sensory consequences and subtract it off.
350
828260
3000
13:51
So they actually think they've hit the person less hard than they have --
351
831260
2000
13:53
rather like the tickling.
352
833260
2000
13:55
Whereas the passive recipient
353
835260
2000
13:57
doesn't make the prediction, feels the full blow.
354
837260
2000
13:59
So if they retaliate with the same force,
355
839260
2000
14:01
the first person will think it's been escalated.
356
841260
2000
14:03
So we decided to test this in the lab.
357
843260
2000
14:05
(Laughter)
358
845260
3000
14:08
Now we don't work with children, we don't work with hitting,
359
848260
2000
14:10
but the concept is identical.
360
850260
2000
14:12
We bring in two adults. We tell them they're going to play a game.
361
852260
3000
14:15
And so here's player one and player two sitting opposite to each other.
362
855260
2000
14:17
And the game is very simple.
363
857260
2000
14:19
We started with a motor
364
859260
2000
14:21
with a little lever, a little force transfuser.
365
861260
2000
14:23
And we use this motor to apply force down to player one's fingers
366
863260
2000
14:25
for three seconds and then it stops.
367
865260
3000
14:28
And that player's been told, remember the experience of that force
368
868260
3000
14:31
and use your other finger
369
871260
2000
14:33
to apply the same force
370
873260
2000
14:35
down to the other subject's finger through a force transfuser -- and they do that.
371
875260
3000
14:38
And player two's been told, remember the experience of that force.
372
878260
3000
14:41
Use your other hand to apply the force back down.
373
881260
3000
14:44
And so they take it in turns
374
884260
2000
14:46
to apply the force they've just experienced back and forward.
375
886260
2000
14:48
But critically,
376
888260
2000
14:50
they're briefed about the rules of the game in separate rooms.
377
890260
3000
14:53
So they don't know the rules the other person's playing by.
378
893260
2000
14:55
And what we've measured
379
895260
2000
14:57
is the force as a function of terms.
380
897260
2000
14:59
And if we look at what we start with,
381
899260
2000
15:01
a quarter of a Newton there, a number of turns,
382
901260
2000
15:03
perfect would be that red line.
383
903260
2000
15:05
And what we see in all pairs of subjects is this --
384
905260
3000
15:08
a 70 percent escalation in force
385
908260
2000
15:10
on each go.
386
910260
2000
15:12
So it really suggests, when you're doing this --
387
912260
2000
15:14
based on this study and others we've done --
388
914260
2000
15:16
that the brain is canceling the sensory consequences
389
916260
2000
15:18
and underestimating the force it's producing.
390
918260
2000
15:20
So it re-shows the brain makes predictions
391
920260
2000
15:22
and fundamentally changes the precepts.
392
922260
3000
15:25
So we've made inferences, we've done predictions,
393
925260
3000
15:28
now we have to generate actions.
394
928260
2000
15:30
And what Bayes' rule says is, given my beliefs,
395
930260
2000
15:32
the action should in some sense be optimal.
396
932260
2000
15:34
But we've got a problem.
397
934260
2000
15:36
Tasks are symbolic -- I want to drink, I want to dance --
398
936260
3000
15:39
but the movement system has to contract 600 muscles
399
939260
2000
15:41
in a particular sequence.
400
941260
2000
15:43
And there's a big gap
401
943260
2000
15:45
between the task and the movement system.
402
945260
2000
15:47
So it could be bridged in infinitely many different ways.
403
947260
2000
15:49
So think about just a point to point movement.
404
949260
2000
15:51
I could choose these two paths
405
951260
2000
15:53
out of an infinite number of paths.
406
953260
2000
15:55
Having chosen a particular path,
407
955260
2000
15:57
I can hold my hand on that path
408
957260
2000
15:59
as infinitely many different joint configurations.
409
959260
2000
16:01
And I can hold my arm in a particular joint configuration
410
961260
2000
16:03
either very stiff or very relaxed.
411
963260
2000
16:05
So I have a huge amount of choice to make.
412
965260
3000
16:08
Now it turns out, we are extremely stereotypical.
413
968260
3000
16:11
We all move the same way pretty much.
414
971260
3000
16:14
And so it turns out we're so stereotypical,
415
974260
2000
16:16
our brains have got dedicated neural circuitry
416
976260
2000
16:18
to decode this stereotyping.
417
978260
2000
16:20
So if I take some dots
418
980260
2000
16:22
and set them in motion with biological motion,
419
982260
3000
16:25
your brain's circuitry would understand instantly what's going on.
420
985260
3000
16:28
Now this is a bunch of dots moving.
421
988260
2000
16:30
You will know what this person is doing,
422
990260
3000
16:33
whether happy, sad, old, young -- a huge amount of information.
423
993260
3000
16:36
If these dots were cars going on a racing circuit,
424
996260
2000
16:38
you would have absolutely no idea what's going on.
425
998260
3000
16:41
So why is it
426
1001260
2000
16:43
that we move the particular ways we do?
427
1003260
2000
16:45
Well let's think about what really happens.
428
1005260
2000
16:47
Maybe we don't all quite move the same way.
429
1007260
3000
16:50
Maybe there's variation in the population.
430
1010260
2000
16:52
And maybe those who move better than others
431
1012260
2000
16:54
have got more chance of getting their children into the next generation.
432
1014260
2000
16:56
So in evolutionary scales, movements get better.
433
1016260
3000
16:59
And perhaps in life, movements get better through learning.
434
1019260
3000
17:02
So what is it about a movement which is good or bad?
435
1022260
2000
17:04
Imagine I want to intercept this ball.
436
1024260
2000
17:06
Here are two possible paths to that ball.
437
1026260
3000
17:09
Well if I choose the left-hand path,
438
1029260
2000
17:11
I can work out the forces required
439
1031260
2000
17:13
in one of my muscles as a function of time.
440
1033260
2000
17:15
But there's noise added to this.
441
1035260
2000
17:17
So what I actually get, based on this lovely, smooth, desired force,
442
1037260
3000
17:20
is a very noisy version.
443
1040260
2000
17:22
So if I pick the same command through many times,
444
1042260
3000
17:25
I will get a different noisy version each time, because noise changes each time.
445
1045260
3000
17:28
So what I can show you here
446
1048260
2000
17:30
is how the variability of the movement will evolve
447
1050260
2000
17:32
if I choose that way.
448
1052260
2000
17:34
If I choose a different way of moving -- on the right for example --
449
1054260
3000
17:37
then I'll have a different command, different noise,
450
1057260
2000
17:39
playing through a noisy system, very complicated.
451
1059260
3000
17:42
All we can be sure of is the variability will be different.
452
1062260
3000
17:45
If I move in this particular way,
453
1065260
2000
17:47
I end up with a smaller variability across many movements.
454
1067260
3000
17:50
So if I have to choose between those two,
455
1070260
2000
17:52
I would choose the right one because it's less variable.
456
1072260
2000
17:54
And the fundamental idea
457
1074260
2000
17:56
is you want to plan your movements
458
1076260
2000
17:58
so as to minimize the negative consequence of the noise.
459
1078260
3000
18:01
And one intuition to get
460
1081260
2000
18:03
is actually the amount of noise or variability I show here
461
1083260
2000
18:05
gets bigger as the force gets bigger.
462
1085260
2000
18:07
So you want to avoid big forces as one principle.
463
1087260
3000
18:10
So we've shown that using this,
464
1090260
2000
18:12
we can explain a huge amount of data --
465
1092260
2000
18:14
that exactly people are going about their lives planning movements
466
1094260
3000
18:17
so as to minimize negative consequences of noise.
467
1097260
3000
18:20
So I hope I've convinced you the brain is there
468
1100260
2000
18:22
and evolved to control movement.
469
1102260
2000
18:24
And it's an intellectual challenge to understand how we do that.
470
1104260
3000
18:27
But it's also relevant
471
1107260
2000
18:29
for disease and rehabilitation.
472
1109260
2000
18:31
There are many diseases which effect movement.
473
1111260
3000
18:34
And hopefully if we understand how we control movement,
474
1114260
2000
18:36
we can apply that to robotic technology.
475
1116260
2000
18:38
And finally, I want to remind you,
476
1118260
2000
18:40
when you see animals do what look like very simple tasks,
477
1120260
2000
18:42
the actual complexity of what is going on inside their brain
478
1122260
2000
18:44
is really quite dramatic.
479
1124260
2000
18:46
Thank you very much.
480
1126260
2000
18:48
(Applause)
481
1128260
8000
18:56
Chris Anderson: Quick question for you, Dan.
482
1136260
2000
18:58
So you're a movement -- (DW: Chauvinist.) -- chauvinist.
483
1138260
4000
19:02
Does that mean that you think that the other things we think our brains are about --
484
1142260
3000
19:05
the dreaming, the yearning, the falling in love and all these things --
485
1145260
3000
19:08
are a kind of side show, an accident?
486
1148260
3000
19:11
DW: No, no, actually I think they're all important
487
1151260
2000
19:13
to drive the right movement behavior to get reproduction in the end.
488
1153260
3000
19:16
So I think people who study sensation or memory
489
1156260
3000
19:19
without realizing why you're laying down memories of childhood.
490
1159260
2000
19:21
The fact that we forget most of our childhood, for example,
491
1161260
3000
19:24
is probably fine, because it doesn't effect our movements later in life.
492
1164260
3000
19:27
You only need to store things which are really going to effect movement.
493
1167260
3000
19:30
CA: So you think that people thinking about the brain, and consciousness generally,
494
1170260
3000
19:33
could get real insight
495
1173260
2000
19:35
by saying, where does movement play in this game?
496
1175260
2000
19:37
DW: So people have found out for example
497
1177260
2000
19:39
that studying vision in the absence of realizing why you have vision
498
1179260
2000
19:41
is a mistake.
499
1181260
2000
19:43
You have to study vision with the realization
500
1183260
2000
19:45
of how the movement system is going to use vision.
501
1185260
2000
19:47
And it uses it very differently once you think about it that way.
502
1187260
2000
19:49
CA: Well that was quite fascinating. Thank you very much indeed.
503
1189260
3000
19:52
(Applause)
504
1192260
2000
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7