AI That Connects the Digital and Physical Worlds | Anima Anandkumar | TED

51,924 views ・ 2024-07-15

TED


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

00:04
I grew up with parents who are engineers.
0
4459
3462
00:07
They were among the first to bring computerized manufacturing
1
7921
3545
00:11
to my hometown in India.
2
11508
1793
00:13
Growing up as a young girl,
3
13635
1710
00:15
I remember being fascinated
4
15345
2252
00:17
how these computer programs didn't just reside within a computer,
5
17639
4463
00:22
but touched the physical world
6
22102
2043
00:24
and produced these beautiful and precise metal parts.
7
24145
4046
00:28
Over the last two decades, as I pursued AI research,
8
28984
4212
00:33
this memory continued to inspire me
9
33196
2836
00:36
to connect the physical
10
36032
1794
00:37
and digital worlds together.
11
37826
1918
00:40
I am working on AI that transforms the way we do science and engineering.
12
40370
5464
00:46
Scientific research and engineering design
13
46459
3337
00:49
currently involves a lot of trial and error.
14
49796
2628
00:53
Many long hours are spent in the lab doing experiments.
15
53091
4129
00:57
So it's not just the great ideas that propel science forward.
16
57846
4337
01:02
You need these experiments to validate findings
17
62183
3629
01:05
and spark new ideas.
18
65812
2127
01:08
How can language models help here?
19
68565
2419
01:11
What if I ask ChatGPT to come up with a better design of an aircraft wing,
20
71359
5839
01:17
or a drone that flies on a turbulent wind?
21
77240
3087
01:20
It may suggest something.
22
80702
1668
01:22
It may even draw something.
23
82370
2127
01:24
But how do we know this is any good?
24
84539
2377
01:27
We don't.
25
87542
1126
01:29
Language models hallucinate because they have no physical grounding.
26
89085
5214
01:34
While language models may help generate new ideas,
27
94299
4046
01:38
they cannot attack the hard part of science
28
98345
3128
01:41
which is simulating the necessary physics
29
101473
4421
01:45
to replace the Nab experiments.
30
105935
2795
01:49
In order to model scientific and physical phenomena,
31
109272
3337
01:52
text alone is not sufficient.
32
112651
2627
01:55
To get to AI with universal physical understanding,
33
115862
4797
02:00
we need to train it on the data of the world we observe.
34
120659
5005
02:06
And not just that, also its hidden details.
35
126247
3754
02:10
From the intricacies of quantum chemistry
36
130335
3045
02:13
that happen at the smallest level
37
133421
2628
02:16
to molecules and proteins that influence how all biological processes work,
38
136091
6089
02:22
to ocean currents and clouds that happen at planetary scales and beyond,
39
142222
5714
02:27
we need AI that can capture these whole range of physical phenomena.
40
147977
6507
02:34
We need AI that can really zoom into the fine details
41
154984
4630
02:39
in order to simulate these phenomena accurately.
42
159614
4171
02:43
To capture the cloud movements,
43
163785
2336
02:46
and predict how clouds move and change in our atmosphere,
44
166162
4171
02:50
we need to be able to zoom into the fine details
45
170375
3503
02:53
of the turbulent fluid flow.
46
173920
1835
02:56
Standard deep learning uses a fixed number of pixels.
47
176673
4129
03:01
So if you zoom in, it gets blurry
48
181261
2752
03:04
and not all the details are captured.
49
184013
2837
03:06
We invented an AI technology called neural operators
50
186891
4505
03:11
that represents the data as continuous functions or shapes,
51
191396
4713
03:16
and allows us to zoom in indefinitely to any resolution or scale.
52
196151
5589
03:22
Neural operators allow us to train on data
53
202365
4171
03:26
at multiple scales or resolutions.
54
206578
2544
03:29
And also allows us to incorporate
55
209789
2086
03:31
the knowledge of mathematical equations
56
211916
2837
03:34
to fill in the finer details
57
214753
2210
03:37
when only limited resolution data is available.
58
217005
3670
03:41
Such learning at multiple scales is essential for scientific understanding
59
221050
6257
03:47
and neural operators enable this.
60
227307
3086
03:51
With neural operators,
61
231186
1543
03:52
we can simulate physical phenomena such as fluid dynamics
62
232771
4129
03:56
as much as a million times faster than traditional simulations.
63
236941
4922
04:02
Last year, we used neural operators to invent a better medical catheter.
64
242781
5130
04:08
A medical catheter is a tube that draws fluids out of the human body.
65
248203
5338
04:13
Unfortunately, the bacteria tend to swim upstream against the fluid flow
66
253583
4880
04:18
and infect the human.
67
258505
1668
04:20
In fact, annually there is more than half a million cases
68
260215
4796
04:25
of such healthcare-related infections,
69
265011
2836
04:27
and this is one of the leading causes.
70
267889
2419
04:30
Last year, we used neural operators to change the inside of the catheter
71
270767
5672
04:36
from smooth to ridged.
72
276439
2711
04:39
With ridges, now we have vortices created as the fluid flows,
73
279150
6090
04:45
and we can hope to stop the bacteria from swimming upstream
74
285240
3670
04:48
because of these vortices.
75
288952
2085
04:51
But to get this correct,
76
291496
2044
04:53
we need the shape of the ridges to be exactly right.
77
293581
3754
04:57
In the past, this would have been done by trial and error.
78
297836
4087
05:02
Design a version of the catheter,
79
302507
2210
05:04
build it out, take it to the lab,
80
304759
3170
05:07
observe a hypothesis if something went wrong,
81
307971
3336
05:11
rinse and repeat and redesign again.
82
311307
2795
05:14
But instead, we taught AI the behavior of the fluid flow inside the tube,
83
314561
6631
05:21
and with it, our neural operator model was able to directly propose
84
321234
5047
05:26
an optimized design.
85
326322
1835
05:28
We 3D-printed the design only once to verify that it worked.
86
328157
5381
05:33
In the video, you're seeing our catheter being tested in the lab.
87
333580
4546
05:38
The bacteria are not able to swim upstream,
88
338167
2753
05:40
are instead being pushed out with the fluid flow.
89
340962
3879
05:44
In fact, we measured the reduction in bacterial contamination
90
344883
5005
05:49
by more than 100-fold.
91
349929
2127
05:52
So in this case, the neural operators were specialized to understand
92
352515
4338
05:56
fluid flow in a tube.
93
356853
1626
05:58
What other applications can AI tackle
94
358938
3837
06:02
and help us solve such pressing problems?
95
362775
3671
06:06
Can deep learning beat numerical weather models?
96
366487
3879
06:10
A group of leading weather scientists asked this question in February 2021,
97
370783
6549
06:17
in a "Royal Society" publication.
98
377332
2127
06:20
They felt that AI was still in its infancy,
99
380293
3420
06:23
and that a number of fundamental breakthroughs would be needed
100
383713
3670
06:27
for AI to become competitive with traditional weather models,
101
387383
4463
06:31
and that would take years or even decades.
102
391846
3087
06:34
Exactly a year later,
103
394933
2210
06:37
we released FourCastNet.
104
397143
2127
06:39
Using neural operators,
105
399270
1919
06:41
we built the first fully AI-based weather model
106
401189
4713
06:45
that is high resolution
107
405944
1751
06:47
and is tens of thousands of times faster than traditional weather models.
108
407737
5005
06:52
What used to take a big supercomputer
109
412784
3170
06:55
can now run on a gaming PC that you may have at home.
110
415995
4505
07:01
This model is also running
111
421042
1877
07:02
at the European Centre for Medium-Range Weather Forecasting,
112
422961
3837
07:06
one of the premier weather agencies of the world.
113
426839
3379
07:10
And our AI model is not just tens of thousands of times faster
114
430218
4880
07:15
than traditional models.
115
435139
1669
07:16
It's also more accurate in many cases.
116
436849
3170
07:20
On September 16 last year,
117
440019
2920
07:22
Hurricane Lee hit the coast of Nova Scotia, Canada.
118
442939
4212
07:27
A full ten days earlier,
119
447193
2127
07:29
our FourCastNet model correctly predicted
120
449320
3420
07:32
that the hurricane would make landfall,
121
452740
2628
07:35
but the traditional weather model
122
455368
1835
07:37
predicted the hurricane would skip the coast.
123
457245
2544
07:39
Only five days later, on September 11,
124
459831
2669
07:42
did the traditional weather model correct its forecast to predict landfall.
125
462542
4546
07:47
Extreme weather events such as Hurricane Lee
126
467463
3295
07:50
will only increase further unless we take action
127
470758
4296
07:55
on climate change.
128
475096
1543
07:56
Such as finding new, clean sources of energy.
129
476639
3545
08:00
Nuclear fusion is one of them.
130
480727
2460
08:03
But unfortunately, there are still big challenges with it.
131
483980
3378
08:07
The fusion reactor heats up the plasma
132
487400
2711
08:10
to extremely high temperatures to get fusion started.
133
490153
3795
08:14
And sometimes this hot plasma can escape confinement
134
494574
4421
08:18
and can damage the reactor.
135
498995
2127
08:21
We train neural operators to simulate and predict
136
501122
3712
08:24
the evolution of plasma inside the reactor.
137
504876
3795
08:28
And with it,
138
508713
1209
08:29
we can use this to predict disruptions before they occur
139
509922
4463
08:34
and take corrective action in the real world.
140
514385
3128
08:37
We are enabling the possibility of nuclear fusion
141
517555
4171
08:41
becoming a reality.
142
521726
2169
08:43
So neural operators and AI broadly
143
523936
3921
08:47
are enabling us to tackle hard scientific challenges
144
527899
4087
08:52
such as climate change and nuclear fusion.
145
532028
3420
08:55
To me, this is just the beginning.
146
535490
2586
08:58
So far, these AI models are limited to the narrow domains they're trained on.
147
538785
6047
09:05
What if you had an AI model
148
545583
3170
09:08
that could solve all and any scientific problem?
149
548753
3712
09:12
From designing better drones, aircrafts, rockets,
150
552507
4713
09:17
and even better drugs and medical devices?
151
557220
3128
09:20
Such an AI model would greatly benefit humanity.
152
560765
3837
09:25
This is what we are working on.
153
565228
2252
09:27
We are building a generalist AI model with emergent capabilities
154
567480
5547
09:33
that can simulate any physical phenomena
155
573027
2795
09:35
and generate novel designs that were previously out of reach.
156
575863
4255
09:40
This is how we scale up neural operators
157
580743
3337
09:44
to enable general intelligence with universal physical understanding.
158
584080
4671
09:49
Thank you.
159
589335
1210
09:50
(Applause)
160
590545
2627
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