How we teach computers to understand pictures | Fei Fei Li

1,159,394 views ใƒป 2015-03-23

TED


ืื ื ืœื—ืฅ ืคืขืžื™ื™ื ืขืœ ื”ื›ืชื•ื‘ื™ื•ืช ื‘ืื ื’ืœื™ืช ืœืžื˜ื” ื›ื“ื™ ืœื”ืคืขื™ืœ ืืช ื”ืกืจื˜ื•ืŸ.

ืžืชืจื’ื: hila scherba ืžื‘ืงืจ: Ido Dekkers
00:14
Let me show you something.
0
14366
3738
ื‘ื•ืื• ืื ื™ ืืจืื” ืœื›ื ืžืฉื”ื•.
00:18
(Video) Girl: Okay, that's a cat sitting in a bed.
1
18104
4156
(ื•ื™ื“ืื•) ื™ืœื“ื”: "ืื•ืงื™ื™, ื–ื” ื—ืชื•ืœ ืฉื™ื•ืฉื‘ ืขืœ ืžื™ื˜ื”.
00:22
The boy is petting the elephant.
2
22260
4040
ื”ื™ืœื“ ืžืœื˜ืฃ ืืช ื”ืคื™ืœ.
00:26
Those are people that are going on an airplane.
3
26300
4354
ืืœื” ืื ืฉื™ื ืฉืขื•ืœื™ื ืขืœ ืžื˜ื•ืก.
00:30
That's a big airplane.
4
30654
2810
ื–ื” ืžื˜ื•ืก ื’ื“ื•ืœ."
00:33
Fei-Fei Li: This is a three-year-old child
5
33464
2206
ื–ื• ื™ืœื“ื” ื‘ืช ืฉืœื•ืฉ ืฉื ื™ื
00:35
describing what she sees in a series of photos.
6
35670
3679
ืžืชืืจืช ืžื” ื”ื™ื ืจื•ืื” ื‘ืกื“ืจืช ืชืžื•ื ื•ืช.
00:39
She might still have a lot to learn about this world,
7
39349
2845
ืื•ืœื™ ื™ืฉ ืœื” ืขื•ื“ ื”ืจื‘ื” ืœืœืžื•ื“ ืขืœ ื”ืขื•ืœื,
00:42
but she's already an expert at one very important task:
8
42194
4549
ืื‘ืœ ื”ื™ื ื›ื‘ืจ ืžื•ืžื—ื™ืช ื‘ืžืฉื™ืžื” ืื—ืช ืžืื•ื“ ื—ืฉื•ื‘ื”:
00:46
to make sense of what she sees.
9
46743
2846
ืœื”ื‘ื™ืŸ ืžื” ื”ื™ื ืจื•ืื”.
00:50
Our society is more technologically advanced than ever.
10
50229
4226
ื”ื—ื‘ืจื” ืฉืœื ื• ื”ื™ื ื™ื•ืชืจ ืžืชืงื“ืžืช ื˜ื›ื ื•ืœื•ื’ื™ืช ืžืื™ ืคืขื.
00:54
We send people to the moon, we make phones that talk to us
11
54455
3629
ืื ื—ื ื• ืฉื•ืœื—ื™ื ืื ืฉื™ื ืœื™ืจื—, ืื ื—ื ื• ืžื™ื™ืฆืจื™ื ื˜ืœืคื•ื ื™ื ืฉืžื“ื‘ืจื™ื ืืœื™ื ื•
00:58
or customize radio stations that can play only music we like.
12
58084
4946
ืื• ืžืชืื™ืžื™ื ืื™ืฉื™ืช ืชื—ื ื•ืช ืจื“ื™ื• ืฉื™ื ื’ื ื• ืจืง ืžื•ืกื™ืงื” ืฉืื ื—ื ื• ืื•ื”ื‘ื™ื.
01:03
Yet, our most advanced machines and computers
13
63030
4055
ื•ืขื“ื™ื™ืŸ, ื”ืžื›ื•ื ื•ืช ื•ื”ืžื—ืฉื‘ื™ื ื”ืžืชืงื“ืžื™ื ื‘ื™ื•ืชืจ ืฉืœื ื•
01:07
still struggle at this task.
14
67085
2903
ืขื“ื™ื™ืŸ ืžืชืงืฉื™ื ื‘ืžืฉื™ืžื” ื”ื–ื•.
01:09
So I'm here today to give you a progress report
15
69988
3459
ืื– ืื ื™ ืคื” ื”ื™ื•ื ื›ื“ื™ ืœืชืช ืœื›ื ื“ื•"ื— ื”ืชืงื“ืžื•ืช
01:13
on the latest advances in our research in computer vision,
16
73447
4047
ืขืœ ื”ื”ืชืคืชื—ื•ื™ื•ืช ื”ืื—ืจื•ื ื•ืช ื‘ืžื—ืงืจ ืฉืœื ื• ืขืœ ืจืื™ื™ืช ืžื—ืฉื‘,
01:17
one of the most frontier and potentially revolutionary
17
77494
4161
ืื—ื“ ืžื”ื˜ื›ื ื•ืœื•ื’ื™ื•ืช ื”ื—ืœื•ืฆื™ื•ืช ื•ื”ืžื”ืคื›ื ื™ื•ืช ื‘ื™ื•ืชืจ
01:21
technologies in computer science.
18
81655
3206
ื‘ืžื“ืขื™ ื”ืžื—ืฉื‘.
01:24
Yes, we have prototyped cars that can drive by themselves,
19
84861
4551
ื ื›ื•ืŸ, ื™ืฉ ืœื ื• ืื‘ื˜ื™ืคื•ืก ืฉืœ ืžื›ื•ื ื™ื•ืช ืฉื™ื›ื•ืœื•ืช ืœื ืกื•ืข ื‘ืขืฆืžืŸ,
01:29
but without smart vision, they cannot really tell the difference
20
89412
3853
ืื‘ืœ ื‘ืœื™ ืจืื™ื™ื” ื—ื›ืžื”, ื”ืŸ ืœื ื™ื›ื•ืœื•ืช ื‘ืืžืช ืœื”ื‘ื“ื™ืœ
01:33
between a crumpled paper bag on the road, which can be run over,
21
93265
3970
ื‘ื™ืŸ ืฉืงื™ืช ื ื™ื™ืจ ืžืงื•ืคืœืช ืขืœ ื”ื›ื‘ื™ืฉ, ืฉืืคืฉืจ ืœื ืกื•ืข ืขืœื™ื”,
01:37
and a rock that size, which should be avoided.
22
97235
3340
ื•ืื‘ืŸ ื‘ื’ื•ื“ืœ ื”ื–ื”, ืฉืฆืจื™ืš ืœื”ื™ืžื ืข ืžืžื ื”.
01:41
We have made fabulous megapixel cameras,
23
101415
3390
ื™ื™ืฆืจื ื• ืžืฆืœืžื•ืช ืžื’ื”ืคื™ืงืกืœ ืžื“ื”ื™ืžื•ืช,
01:44
but we have not delivered sight to the blind.
24
104805
3135
ืื‘ืœ ืœื ื”ืฆืœื—ื ื• ืœื”ื‘ื™ื ืจืื™ื™ื” ืœืขื™ื•ื•ืจื™ื.
01:48
Drones can fly over massive land,
25
108420
3305
ืžื–ืœ"ื˜ื™ื ื™ื›ื•ืœื™ื ืœื˜ื•ืก ืžืขืœ ืฉื˜ื— ืขืฆื•ื,
01:51
but don't have enough vision technology
26
111725
2134
ืื‘ืœ ื—ืกืจื™ ื˜ื›ื ื•ืœื•ื’ื™ื™ืช ืจืื™ื™ื” ืžืกืคืงืช
01:53
to help us to track the changes of the rainforests.
27
113859
3461
ื‘ืฉื‘ื™ืœ ืœืขื–ื•ืจ ืœื ื• ืœืขืงื•ื‘ ืื—ืจื™ ื”ืฉื™ื ื•ื™ื™ื ื‘ื™ืขืจื•ืช ื”ื’ืฉื.
01:57
Security cameras are everywhere,
28
117320
2950
ืžืฆืœืžื•ืช ืื‘ื˜ื—ื” ื ืžืฆืื•ืช ื‘ื›ืœ ืžืงื•ื,
02:00
but they do not alert us when a child is drowning in a swimming pool.
29
120270
5067
ืื‘ืœ ื”ืŸ ืœื ืžืชืจื™ืขื•ืช ื›ืฉื™ืœื“ ื˜ื•ื‘ืข ื‘ื‘ืจื™ื›ื”.
02:06
Photos and videos are becoming an integral part of global life.
30
126167
5595
ืชืžื•ื ื•ืช ื•ืกืจื˜ื•ื ื™ื ื”ื•ืคื›ื™ื ืœื—ืœืง ื‘ืœืชื™ ื ืคืจื“ ืžื”ื—ื™ื™ื ื”ื’ืœื•ื‘ืœื™ื™ื.
02:11
They're being generated at a pace that's far beyond what any human,
31
131762
4087
ื”ื ื ื•ืฆืจื™ื ื‘ืงืฆื‘ ืฉื”ื•ื ืžืขืœ ืœื›ืœ ืžื” ืฉื›ืœ ืื“ื,
02:15
or teams of humans, could hope to view,
32
135849
2783
ืื• ืงื‘ื•ืฆื•ืช ืฉืœ ืื ืฉื™ื, ื™ื›ื•ืœื™ื ืœืงื•ื•ืช ืœืฆืคื•ืช ื‘ื”ื,
02:18
and you and I are contributing to that at this TED.
33
138632
3921
ื•ืืชื ื•ืื ื™ ืชื•ืจืžื™ื ืœื–ื” ื‘ืฉื™ื—ืช TED ื”ื–ื•.
02:22
Yet our most advanced software is still struggling at understanding
34
142553
5232
ืื‘ืœ ื”ืชื•ื›ื ื” ื”ืžืชืงื“ืžืช ื‘ื™ื•ืชืจ ืฉืœื ื• ืขื“ื™ื™ืŸ ืžืชืžื•ื“ื“ืช ื‘ืœื”ื‘ื™ืŸ
02:27
and managing this enormous content.
35
147785
3876
ื•ืœื ื”ืœ ืืช ื”ืชื•ื›ืŸ ื”ืขืฆื•ื ื”ื–ื”.
02:31
So in other words, collectively as a society,
36
151661
5272
ืื– ื‘ืžื™ืœื™ื ืื—ืจื•ืช, ื‘ืžืฉื•ืชืฃ ื›ื—ื‘ืจื”,
02:36
we're very much blind,
37
156933
1746
ืื ื—ื ื• ืžืื•ื“ ืขื™ื•ื•ืจื™ื,
02:38
because our smartest machines are still blind.
38
158679
3387
ื›ื™ ื”ืžื›ื•ื ื•ืช ื”ื—ื›ืžื•ืช ื‘ื™ื•ืชืจ ืฉืœื ื• ืขื“ื™ื™ืŸ ืขื™ื•ื•ืจื•ืช.
02:43
"Why is this so hard?" you may ask.
39
163526
2926
ืืชื ื™ื›ื•ืœื™ื ืœืฉืื•ืœ - "ืœืžื” ื–ื” ื›ืœ ื›ืš ืงืฉื”?"
02:46
Cameras can take pictures like this one
40
166452
2693
ืžืฆืœืžื•ืช ื™ื›ื•ืœื•ืช ืœืงื—ืช ืชืžื•ื ื•ืช ื›ืžื• ื–ื•,
02:49
by converting lights into a two-dimensional array of numbers
41
169145
3994
ืขืœ ื™ื“ื™ ื”ืžืจืช ืื•ืจื•ืช ืœืฉื˜ื— ื“ื• ืžื™ืžื“ื™ ืฉืœ ืžืกืคืจื™ื,
02:53
known as pixels,
42
173139
1650
ื”ื™ื“ื•ืขื™ื ื›ืคื™ืงืกืœื™ื,
02:54
but these are just lifeless numbers.
43
174789
2251
ืื‘ืœ ืืœื• ืจืง ืžืกืคืจื™ื ื—ืกืจื™ ื—ื™ื™ื.
02:57
They do not carry meaning in themselves.
44
177040
3111
ื”ื ืœื ื ื•ืฉืื™ื ืื™ื–ื•ืฉื”ื™ ืžืฉืžืขื•ืช ื‘ืขืฆืžื.
03:00
Just like to hear is not the same as to listen,
45
180151
4343
ื›ืžื• ืฉืœืฉืžื•ืข ื–ื” ืœื ืื•ืชื• ื“ื‘ืจ ื›ืžื• ืœื”ืงืฉื™ื‘,
03:04
to take pictures is not the same as to see,
46
184494
4040
ืœืงื—ืช ืชืžื•ื ื•ืช ื–ื” ืœื ืื•ืชื• ื“ื‘ืจ ื›ืžื• ืœืจืื•ืช,
03:08
and by seeing, we really mean understanding.
47
188534
3829
ื•ื‘ืœืจืื•ืช, ืื ื—ื ื• ืœืžืขืฉื” ืžืชื›ื•ื•ื ื™ื ืœืœื”ื‘ื™ืŸ.
03:13
In fact, it took Mother Nature 540 million years of hard work
48
193293
6177
ืœืžืขืฉื”, ื–ื” ืœืงื— ืœืื™ืžื ื˜ื‘ืข 540 ืžื™ืœื™ื•ืŸ ืฉื ื™ื ืฉืœ ืขื‘ื•ื“ื” ืงืฉื”
03:19
to do this task,
49
199470
1973
ืœืขืฉื•ืช ืืช ื”ืžืฉื™ืžื” ื”ื–ื•,
03:21
and much of that effort
50
201443
1881
ื•ื”ืจื‘ื” ืžื”ืžืืžืฅ ื”ื–ื”
03:23
went into developing the visual processing apparatus of our brains,
51
203324
5271
ื”ืœืš ืขืœ ืคื™ืชื•ื— ืžื ื’ื ื•ืŸ ื”ืขื™ื‘ื•ื“ ื”ื—ื–ื•ืชื™ ืฉืœ ื”ืžื•ื— ืฉืœื ื•,
03:28
not the eyes themselves.
52
208595
2647
ืœื ื”ืขื™ื ื™ื™ื ืขืฆืžืŸ.
03:31
So vision begins with the eyes,
53
211242
2747
ืื– ืจืื™ื™ื” ืžืชื—ื™ืœื” ื‘ืขื™ื ื™ื™ื,
03:33
but it truly takes place in the brain.
54
213989
3518
ืื‘ืœ ื‘ืืžืช ืžืชืจื—ืฉืช ื‘ืžื•ื—.
03:38
So for 15 years now, starting from my Ph.D. at Caltech
55
218287
5060
ืื– ื‘ืžืฉืš 15 ืฉื ื™ื ื”ืื—ืจื•ื ื•ืช, ืžื”ื“ื•ืงื˜ื•ืจื˜ ืฉืœื™ ื‘ืžื›ื•ืŸ ื”ื˜ื›ื ื•ืœื•ื’ื™ ืฉืœ ืงืœื™ืคื•ืจื ื™ื”
03:43
and then leading Stanford's Vision Lab,
56
223347
2926
ื•ืื– ื”ื•ื‘ืœืช ืžืขื‘ื“ืช ื”ืจืื™ื™ื” ื‘ืกื˜ื ืคื•ืจื“,
03:46
I've been working with my mentors, collaborators and students
57
226273
4396
ืื ื™ ืขื•ื‘ื“ืช ืขื ื”ืžื•ืจื™ื ื”ืจื•ื—ื ื™ื™ื ืฉืœื™, ืžืฉืชืคื™ ืคืขื•ืœื” ื•ืกื˜ื•ื“ื ื˜ื™ื,
03:50
to teach computers to see.
58
230669
2889
ืœืœืžื“ ืžื—ืฉื‘ื™ื ืœืจืื•ืช.
03:54
Our research field is called computer vision and machine learning.
59
234658
3294
ืชื—ื•ื ื”ืžื—ืงืจ ืฉืœื ื• ื ืงืจื - ืจืื™ื™ื” ืžืžื•ื—ืฉื‘ืช ื•ืœืžื™ื“ืช ืžื›ื•ื ื”.
03:57
It's part of the general field of artificial intelligence.
60
237952
3878
ื–ื” ื—ืœืง ืžืชื—ื•ื ื›ืœืœื™ ื™ื•ืชืจ ืฉืœ ืื™ื ื˜ืœื’ื ืฆื™ื” ืžืœืื›ื•ืชื™ืช.
04:03
So ultimately, we want to teach the machines to see just like we do:
61
243000
5493
ืื– ื‘ืกื•ืคื• ืฉืœ ื“ื‘ืจ, ืื ื—ื ื• ืจื•ืฆื™ื ืœืœืžื“ ืืช ื”ืžื›ื•ื ื•ืช ืœืจืื•ืช ื›ืžื• ืฉืื ื—ื ื• ืจื•ืื™ื:
04:08
naming objects, identifying people, inferring 3D geometry of things,
62
248493
5387
ืœื ืงื•ื‘ ื‘ืฉืžื•ืช ืฉืœ ืื•ื‘ื™ื™ืงื˜ื™ื, ืœื–ื”ื•ืช ืื ืฉื™ื, ืœื”ืกื™ืง ื’ื™ืื•ืžื˜ืจื™ืช ืชืœืช ืžื™ืžื“ื™ืช ืฉืœ ื“ื‘ืจื™ื,
04:13
understanding relations, emotions, actions and intentions.
63
253880
5688
ืœื”ื‘ื™ืŸ ืงืฉืจื™ื, ืจื’ืฉื•ืช, ืคืขื•ืœื•ืช ื•ื›ื•ื•ื ื•ืช.
04:19
You and I weave together entire stories of people, places and things
64
259568
6153
ืืชื ื•ืื ื™ ืจื•ืงืžื™ื ื‘ื™ื—ื“ ืกื™ืคื•ืจื™ื ืฉืœืžื™ื ืฉืœ ืื ืฉื™ื, ืžืงื•ืžื•ืช ื•ื“ื‘ืจื™ื
04:25
the moment we lay our gaze on them.
65
265721
2164
ื‘ืจื’ืข ืฉืื ื—ื ื• ืžื ื™ื—ื™ื ืขืœื™ื”ื ืืช ื”ืžื‘ื˜ ืฉืœื ื•.
04:28
The first step towards this goal is to teach a computer to see objects,
66
268955
5583
ื”ืฆืขื“ ื”ืจืืฉื•ืŸ ืœื›ื™ื•ื•ืŸ ื”ืžื˜ืจื” ื”ื–ื• ื”ื•ื ืœืœืžื“ ืžื—ืฉื‘ ืœืจืื•ืช ื—ืคืฆื™ื,
04:34
the building block of the visual world.
67
274538
3368
ืื‘ืŸ ื”ื‘ื ื™ื™ืŸ ืฉืœ ื”ืขื•ืœื ื”ื—ื–ื•ืชื™.
04:37
In its simplest terms, imagine this teaching process
68
277906
4434
ื‘ืžื•ื ื—ื™ื ื”ื›ื™ ืคืฉื•ื˜ื™ื, ื“ืžื™ื™ื ื• ืืช ืชื”ืœื™ืš ื”ืœืžื™ื“ื” ื”ื–ื”
04:42
as showing the computers some training images
69
282340
2995
ื›ืœื”ืจืื•ืช ืœืžื—ืฉื‘ื™ื ื›ืžื” ืชืžื•ื ื•ืช ืื™ืžื•ืŸ ืฉืœ ืื•ื‘ื™ื™ืงื˜ ืžืกื•ื™ื,
04:45
of a particular object, let's say cats,
70
285335
3321
ื‘ื•ืื• ื ืืžืจ ื—ืชื•ืœื™ื,
04:48
and designing a model that learns from these training images.
71
288656
4737
ื•ืขื™ืฆื•ื‘ ืžื•ื“ืœ ืฉื™ืœืžื“ ืžืชืžื•ื ื•ืช ื”ืื™ืžื•ืŸ ื”ืืœื•.
04:53
How hard can this be?
72
293393
2044
ื›ืžื” ืงืฉื” ื–ื” ื›ื‘ืจ ื™ื›ื•ืœ ืœื”ื™ื•ืช?
04:55
After all, a cat is just a collection of shapes and colors,
73
295437
4052
ืื—ืจื™ ื”ื›ืœ, ื—ืชื•ืœ ื”ื•ื ืคืฉื•ื˜ ืื•ืกืฃ ืฉืœ ืฆื•ืจื•ืช ื•ืฆื‘ืขื™ื,
04:59
and this is what we did in the early days of object modeling.
74
299489
4086
ื•ื–ื” ืžื” ืฉืื ื—ื ื• ืขืฉื™ื ื• ื‘ื™ืžื™ื ื”ืจืืฉื•ื ื™ื ืฉืœ ืฉื™ืžื•ืฉ ื›ืžื•ื“ืœ ืชื™ืื•ืจื˜ื™ ื‘ืื•ื‘ื™ื™ืงื˜ื™ื.
05:03
We'd tell the computer algorithm in a mathematical language
75
303575
3622
ื”ื™ื™ื ื• ืื•ืžืจื™ื ืœืืœื’ื•ืจื™ืชื ืฉืœ ื”ืžื—ืฉื‘ ื‘ืฉืคื” ืžืชืžื˜ื™ืช
05:07
that a cat has a round face, a chubby body,
76
307197
3343
ืฉื”ืคื ื™ื ืฉืœ ื—ืชื•ืœ ื”ืŸ ืขื’ื•ืœื•ืช, ื’ื•ืฃ ืฉืžื ืžืŸ,
05:10
two pointy ears, and a long tail,
77
310540
2299
ืฉื ื™ ืื•ื–ื ื™ื™ื ืžื—ื•ื“ื“ื•ืช, ื–ื ื‘ ืืจื•ืš,
05:12
and that looked all fine.
78
312839
1410
ื•ื–ื” ื”ื™ื” ื ืจืื” ื‘ืกื“ืจ ื’ืžื•ืจ.
05:14
But what about this cat?
79
314859
2113
ืื‘ืœ ืžื” ืขื ื”ื—ืชื•ืœ ื”ื–ื”?
05:16
(Laughter)
80
316972
1091
(ืฆื—ื•ืง)
05:18
It's all curled up.
81
318063
1626
ื”ื•ื ื›ื•ืœื• ืžื›ื•ืจื‘ืœ.
05:19
Now you have to add another shape and viewpoint to the object model.
82
319689
4719
ืขื›ืฉื™ื• ืฆืจื™ืš ืœื”ื•ืกื™ืฃ ืขื•ื“ ืฆื•ืจื” ื•ื ืงื•ื“ืช ืžื‘ื˜ ืœืžื•ื“ืœ ื”ืื•ื‘ื™ื™ืงื˜.
05:24
But what if cats are hidden?
83
324408
1715
ืื‘ืœ ืžื” ืื ื—ืชื•ืœื™ื ืžืชื—ื‘ืื™ื?
05:27
What about these silly cats?
84
327143
2219
ืžื” ืขื ื”ื—ืชื•ืœื™ื ื”ืžื˜ื•ืคืฉื™ื ื”ืืœื•?
05:31
Now you get my point.
85
331112
2417
ืขื›ืฉื™ื• ืืชื ืžืชื—ื™ืœื™ื ืœื”ื‘ื™ืŸ ืืช ื”ื ืงื•ื“ื” ืฉืœื™.
05:33
Even something as simple as a household pet
86
333529
3367
ืืคื™ืœื• ืžืฉื”ื• ืคืฉื•ื˜ ื›ืžื• ื—ื™ื™ืช ืžื—ืžื“ ื‘ื™ืชื™ืช
05:36
can present an infinite number of variations to the object model,
87
336896
4504
ื™ื›ื•ืœ ืœื”ืฆื™ื’ ืื™ื ืกื•ืฃ ืฆื•ืจื•ืช ืœืžื•ื“ืœ ืฉืœ ืื•ื‘ื™ื™ืงื˜,
05:41
and that's just one object.
88
341400
2233
ื•ื–ื” ืจืง ืื•ื‘ื™ื™ืงื˜ ืื—ื“.
05:44
So about eight years ago,
89
344573
2492
ืื– ืœืคื ื™ 8 ืฉื ื™ื ื‘ืขืจืš,
05:47
a very simple and profound observation changed my thinking.
90
347065
5030
ื”ืกืชื›ืœื•ืช ืžืื•ื“ ืคืฉื•ื˜ื” ื•ืžืขืžื™ืงื” ืฉื™ื ืชื” ืืช ื”ืžื—ืฉื‘ื” ืฉืœื™.
05:53
No one tells a child how to see,
91
353425
2685
ืืฃ ืื—ื“ ืœื ืื•ืžืจ ืœื™ืœื“ ืื™ืš ืœืจืื•ืช,
05:56
especially in the early years.
92
356110
2261
ื‘ืžื™ื•ื—ื“ ืœื ื‘ืฉื ื™ื ื”ืžื•ืงื“ืžื•ืช.
05:58
They learn this through real-world experiences and examples.
93
358371
5000
ื”ื ืœื•ืžื“ื™ื ื“ืจืš ื”ื ืกื™ื•ืŸ ื‘ืขื•ืœื ื”ืืžื™ืชื™ ื•ื“ื•ื’ืžืื•ืช.
06:03
If you consider a child's eyes
94
363371
2740
ืื ืชืชื™ื™ื—ืกื• ืœืขื™ื ื™ื™ื ืฉืœ ื™ืœื“
06:06
as a pair of biological cameras,
95
366111
2554
ื›ื–ื•ื’ ืžืฆืœืžื•ืช ื‘ื™ื•ืœื•ื’ื™ื•ืช
06:08
they take one picture about every 200 milliseconds,
96
368665
4180
ื”ืŸ ืœื•ืงื—ื•ืช ืชืžื•ื ื” ืื—ืช ื‘ืขืจืš ื›ืœ 200 ืืœืคื™ื•ืช ื”ืฉื ื™ื™ื”,
06:12
the average time an eye movement is made.
97
372845
3134
ื”ื–ืžืŸ ื”ืžืžื•ืฆืข ืฉืœ ืชื ื•ืขืช ืขื™ืŸ.
06:15
So by age three, a child would have seen hundreds of millions of pictures
98
375979
5550
ืื– ืขื“ ื’ื™ืœ ืฉืœื•ืฉ, ื™ืœื“ ื™ืจืื” ืžืื•ืช ืžื™ืœื™ื•ื ื™ ืชืžื•ื ื•ืช
06:21
of the real world.
99
381529
1834
ืฉืœ ื”ืขื•ืœื ื”ืืžื™ืชื™.
06:23
That's a lot of training examples.
100
383363
2280
ื–ื” ื”ืจื‘ื” ื“ื•ื’ืžืื•ืช ืื™ืžื•ืŸ.
06:26
So instead of focusing solely on better and better algorithms,
101
386383
5989
ืื– ื‘ืžืงื•ื ืœื”ืชืจื›ื– ืืš ื•ืจืง ืขืœ ืืœื’ื•ืจื™ืชืžื™ื ื˜ื•ื‘ื™ื ื™ื•ืชืจ ื•ื™ื•ืชืจ,
06:32
my insight was to give the algorithms the kind of training data
102
392372
5272
ื”ืชื•ื‘ื ื” ืฉืœื™ ื”ื™ืชื” ืœืชืช ืœืืœื’ื•ืจื™ืชืžื™ื ืืช ืกื•ื’ ืžื™ื“ืข ื”ืื™ืžื•ืŸ
06:37
that a child was given through experiences
103
397644
3319
ืฉื™ืœื“ ืžืงื‘ืœ ื“ืจืš ื ื™ืกื™ื•ืŸ
06:40
in both quantity and quality.
104
400963
3878
ื’ื ื‘ื›ืžื•ืช ื•ื’ื ื‘ืื™ื›ื•ืช.
06:44
Once we know this,
105
404841
1858
ื‘ืจื’ืข ืฉืื ื—ื ื• ื™ื•ื“ืขื™ื ืืช ื–ื”,
06:46
we knew we needed to collect a data set
106
406699
2971
ื™ื“ืขื ื• ืฉืื ื—ื ื• ืฆืจื™ื›ื™ื ืœืืกื•ืฃ ืžืขืจื›ืช ืžื™ื“ืข
06:49
that has far more images than we have ever had before,
107
409670
4459
ืฉื™ืฉ ื‘ื” ื”ืจื‘ื” ื™ื•ืชืจ ืชืžื•ื ื•ืช ืžืžื” ืฉื”ื™ื• ืœื ื• ืื™ ืคืขื,
06:54
perhaps thousands of times more,
108
414129
2577
ืื•ืœื™ ืคื™ ื›ืžื” ืืœืคื™ื ื™ื•ืชืจ,
06:56
and together with Professor Kai Li at Princeton University,
109
416706
4111
ื•ื™ื—ื“ ืขื ืคืจื•ืคืกื•ืจ ืงืื™ ืœื™ ืžืื•ื ื™ื‘ืจืกื™ื˜ืช ืคืจื™ื ืกื˜ื•ืŸ,
07:00
we launched the ImageNet project in 2007.
110
420817
4752
ื”ืชื—ืœื ื• ืืช ืคืจื•ื™ื™ืงื˜ ืื™ืžื’'ื ื˜ ื‘-2007.
07:05
Luckily, we didn't have to mount a camera on our head
111
425569
3838
ืœืžื–ืœื ื•, ืœื ื”ื™ื™ื ื• ืฆืจื™ื›ื™ื ืœืฉื™ื ืžืฆืœืžื” ืขืœ ืจืืฉื™ื ื•
07:09
and wait for many years.
112
429407
1764
ื•ืœื—ื›ื•ืช ื”ืจื‘ื” ืฉื ื™ื.
07:11
We went to the Internet,
113
431171
1463
ื ื™ื’ืฉื ื• ืœืื™ื ื˜ืจื ื˜,
07:12
the biggest treasure trove of pictures that humans have ever created.
114
432634
4436
ื”ืื•ืฆืจ ื”ื’ื“ื•ืœ ื‘ื™ื•ืชืจ ืฉืœ ืชืžื•ื ื•ืช ืฉื”ืื“ื ื™ืฆืจ ืื™ ืคืขื.
07:17
We downloaded nearly a billion images
115
437070
3041
ื”ื•ืจื“ื ื• ื›ืžืขื˜ ืžื™ืœื™ืืจื“ ืชืžื•ื ื•ืช
07:20
and used crowdsourcing technology like the Amazon Mechanical Turk platform
116
440111
5880
ื•ื”ืฉืชืžืฉื ื• ื‘ืคืขื™ืœื•ืช ืฉืœ ื”ืฆื™ื‘ื•ืจ ื”ืจื—ื‘ ื›ืžื• ื”ืคืœื˜ืคื•ืจืžืช ื”ืžื›ื ื™ืงืœ ื˜ื•ืจืง ืฉืœ ืืžื–ื•ืŸ
07:25
to help us to label these images.
117
445991
2339
07:28
At its peak, ImageNet was one of the biggest employers
118
448330
4900
ื‘ืฉื™ืื•, ืื™ืžื’'ื ื˜ ื”ื™ื” ืื—ื“ ื”ืžืขืกื™ืงื™ื ื”ื’ื“ื•ืœื™ื
07:33
of the Amazon Mechanical Turk workers:
119
453230
2996
ืฉืœ ืคืœื˜ืคื•ืจืžืช ื˜ื•ืจืง ืฉืœ ืืžื–ื•ืŸ:
07:36
together, almost 50,000 workers
120
456226
3854
ื‘ื™ื—ื“, ื›ืžืขื˜ 50,000 ืขื•ื‘ื“ื™ื
07:40
from 167 countries around the world
121
460080
4040
ืž-167 ืžื“ื™ื ื•ืช ืžืกื‘ื™ื‘ ืœืขื•ืœื
07:44
helped us to clean, sort and label
122
464120
3947
ืขื–ืจื• ืœื ื• ืœื ืงื•ืช, ืœืกื“ืจ ื•ืœืชื™ื™ื’
07:48
nearly a billion candidate images.
123
468067
3575
ื›ืžืขื˜ ืžื™ืœื™ืืจื“ ืชืžื•ื ื•ืช ืžื•ืขืžื“ื•ืช.
07:52
That was how much effort it took
124
472612
2653
ื–ื” ื›ืžื” ืžืืžืฅ ื ื“ืจืฉ
07:55
to capture even a fraction of the imagery
125
475265
3900
ื›ื“ื™ ืœืœื›ื•ื“ ืฉื‘ืจื™ืจ ืžื™ื›ื•ืœืช ื”ื“ื™ืžื•ื™
07:59
a child's mind takes in in the early developmental years.
126
479165
4171
ืฉืžื•ื— ืฉืœ ื™ืœื“ ืžืกื•ื’ืœ ืœืขืฉื•ืช ื‘ืฉื ื•ืช ื”ื”ืชืคืชื—ื•ืช ื”ืžื•ืงื“ืžื•ืช.
08:04
In hindsight, this idea of using big data
127
484148
3902
ื‘ื“ื™ืขื‘ื“, ื”ืจืขื™ื•ืŸ ืœื”ืฉืชืžืฉ ื‘ื‘ื™ื’ ื“ืื˜ื”
08:08
to train computer algorithms may seem obvious now,
128
488050
4550
ื‘ืฉื‘ื™ืœ ืœืืžืŸ ืืœื’ื•ืจื™ืชื ืฉืœ ืžื—ืฉื‘ ื ืจืื” ืื•ืœื™ ื‘ืจื•ืจ ืขื›ืฉื™ื•,
08:12
but back in 2007, it was not so obvious.
129
492600
4110
ืื‘ืœ ื‘-2007, ื–ื” ืœื ื”ื™ื” ื›ื–ื” ื‘ืจื•ืจ.
08:16
We were fairly alone on this journey for quite a while.
130
496710
3878
ื”ื™ื™ื ื• ื™ื—ืกื™ืช ืœื‘ื“ ื‘ืžืกืข ื”ื–ื” ืœืžืฉืš ื–ืžืŸ ืœื ืงืฆืจ.
08:20
Some very friendly colleagues advised me to do something more useful for my tenure,
131
500588
5003
ื›ืžื” ืขืžื™ืชื™ื ื™ื“ื™ื“ื•ืชื™ื™ื ื”ืฆื™ืขื• ืœื™ ืœืขืฉื•ืช ืžืฉื”ื• ืฉื™ืžื•ืฉื™ ื™ื•ืชืจ ื‘ืฉื‘ื™ืœ ื”ืงื‘ื™ืขื•ืช ืฉืœื™,
08:25
and we were constantly struggling for research funding.
132
505591
4342
ื•ื”ื™ื™ื ื• ื ืื‘ืงื™ื ื›ืœ ื”ื–ืžืŸ ืขืœ ืชืงืฆื™ื‘ื™ ืžื—ืงืจ.
08:29
Once, I even joked to my graduate students
133
509933
2485
ืคืขื ืื—ืช, ื”ืชื‘ื“ื—ืชื™ ืขื ื”ืกื˜ื•ื“ื ื˜ื™ื ืฉืœื™ ืœืชื•ืืจ ืฉื ื™
08:32
that I would just reopen my dry cleaner's shop to fund ImageNet.
134
512418
4063
ืฉืื ื™ ืคืฉื•ื˜ ืืคืชื— ืžื—ื“ืฉ ืืช ื”ื—ื ื•ืช ืœื ื™ืงื•ื™ ื™ื‘ืฉ ืฉืœื™ ื›ื“ื™ ืœืžืžืŸ ืืช ืื™ืžื’'ื ื˜.
08:36
After all, that's how I funded my college years.
135
516481
4761
ืื—ืจื™ ื”ื›ืœ, ื›ื›ื” ืžื™ืžื ืชื™ ืืช ืฉื ื•ืช ื”ืœื™ืžื•ื“ื™ื ืฉืœื™.
08:41
So we carried on.
136
521242
1856
ืื– ื”ืžืฉื›ื ื•.
08:43
In 2009, the ImageNet project delivered
137
523098
3715
ื‘-2009, ืคืจื•ื™ื™ืงื˜ ืื™ืžื’'ื ื˜ ืกื™ืคืง
08:46
a database of 15 million images
138
526813
4042
ื‘ืกื™ืก ื ืชื•ื ื™ื ืฉืœ 15 ืžื™ืœื™ื•ืŸ ืชืžื•ื ื•ืช
08:50
across 22,000 classes of objects and things
139
530855
4805
ืœืจื•ื—ื‘ 22,000 ืกื•ื’ื™ ืื•ื‘ื™ื™ืงื˜ื™ื ื•ื“ื‘ืจื™ื
08:55
organized by everyday English words.
140
535660
3320
ืžืื•ืจื’ื ื™ื ืœืคื™ ืฉืคื” ืื ื’ืœื™ืช ื™ื•ืžื™ื•ืžื™ืช.
08:58
In both quantity and quality,
141
538980
2926
ื’ื ื‘ื›ืžื•ืช ื•ื’ื ื‘ืื™ื›ื•ืช,
09:01
this was an unprecedented scale.
142
541906
2972
ื–ื” ื”ื™ื” ืงื ื” ืžื™ื“ื” ื—ืกืจ ืชืงื“ื™ื.
09:04
As an example, in the case of cats,
143
544878
3461
ื›ื“ื•ื’ืžื, ื‘ืžืงืจื” ืฉืœ ื—ืชื•ืœื™ื,
09:08
we have more than 62,000 cats
144
548339
2809
ื™ืฉ ืœื ื• ื™ื•ืชืจ ืž-62,000 ื—ืชื•ืœื™ื
09:11
of all kinds of looks and poses
145
551148
4110
ื‘ื›ืœ ืžื™ื ื™ ืžืจืื•ืช ื•ืชื ื•ื—ื•ืช
09:15
and across all species of domestic and wild cats.
146
555258
5223
ื•ืขืœ ืคื ื™ ื›ืœ ื”ืžื™ื ื™ื ืฉืœ ื—ืชื•ืœื™ื ื‘ื™ืชื™ื™ื ื•ืคืจืื™ื™ื.
09:20
We were thrilled to have put together ImageNet,
147
560481
3344
ื”ื™ื™ื ื• ื ืจื’ืฉื™ื ืœื”ืจื›ื™ื‘ ืืช ืื™ืžื’'ื ื˜,
09:23
and we wanted the whole research world to benefit from it,
148
563825
3738
ื•ืจืฆื™ื ื• ืฉื›ืœ ืขื•ืœื ื”ืžื—ืงืจ ื™ืจื•ื•ื™ื— ืžืžื ื•,
09:27
so in the TED fashion, we opened up the entire data set
149
567563
4041
ืื– ื‘ืจื•ื— TED, ืคืชื—ื ื• ืืช ื›ืœ ืžืขืจื›ืช ื”ื ืชื•ื ื™ื ืฉืœื ื•
09:31
to the worldwide research community for free.
150
571604
3592
ืœืงื”ื™ืœืช ื”ืžื—ืงืจ ื‘ืจื—ื‘ื™ ื”ืขื•ืœื ื‘ื—ื™ื ื.
09:36
(Applause)
151
576636
4000
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
09:41
Now that we have the data to nourish our computer brain,
152
581416
4538
ืขื›ืฉื™ื• ื›ืฉื™ืฉ ืœื ื• ืืช ื”ื ืชื•ื ื™ื ืœื”ื–ื™ืŸ ื‘ืžื•ื— ื”ืžืžื•ื—ืฉื‘ ืฉืœื ื•,
09:45
we're ready to come back to the algorithms themselves.
153
585954
3737
ืื ื—ื ื• ืžื•ื›ื ื™ื ืœื—ื–ื•ืจ ืœืืœื’ื•ืจื™ืชืžื™ื ืขืฆืžื.
09:49
As it turned out, the wealth of information provided by ImageNet
154
589691
5178
ื›ืคื™ ืฉื”ืชื‘ืจืจ, ืขื•ืฉืจ ื”ืžื™ื“ืข ืฉืกื•ืคืง ืขืœ ื™ื“ื™ ืื™ืžื’'ื ื˜
09:54
was a perfect match to a particular class of machine learning algorithms
155
594869
4806
ื”ื™ื” ื”ืชืืžื” ืžื•ืฉืœืžืช ืœืกื•ื’ ืžืกื•ื™ื™ื ืฉืœ ืืœื’ื•ืจื™ืชืžื™ื ืœืœืžื™ื“ืช ืžื›ื•ื ื”
09:59
called convolutional neural network,
156
599675
2415
ื”ื ืงืจืื™ื ืจืฉืช ืขืฆื‘ื™ื ืžื•ืจื›ื‘ืช,
10:02
pioneered by Kunihiko Fukushima, Geoff Hinton, and Yann LeCun
157
602090
5248
ืฉืคืจืฆื• ื“ืจืš ืขืœ ื™ื“ื™ ืงื•ื ื™ื”ื™ืงื• ืคื•ืงื•ืฉื™ืžื”, ื’'ืฃ ื”ื™ื ื˜ื•ืŸ ื•ื™ืืŸ ืœื”-ืงื•ืŸ
10:07
back in the 1970s and '80s.
158
607338
3645
ื‘ืฉื ื•ืช ื”-70 ื•ื”-80.
10:10
Just like the brain consists of billions of highly connected neurons,
159
610983
5619
ื‘ื“ื™ื•ืง ื›ืžื• ืฉื”ืžื•ื— ืžื›ื™ืœ ืžื™ืœื™ืืจื“ื™ ื ื•ื™ืจื•ื ื™ื ื”ืžื—ื•ื‘ืจื™ื ื”ื™ื˜ื‘,
10:16
a basic operating unit in a neural network
160
616602
3854
ื™ื—ื™ื“ืช ื”ืคืขืœื” ื‘ืกื™ืกื™ืช ื‘ืจืฉืช ื”ื ื•ื™ืจืืœื™ืช
10:20
is a neuron-like node.
161
620456
2415
ื”ื™ื ื”ืฆื•ืžืช ืžื ืชื‘ ื”ืžื™ื“ืข ื“ืžื•ื™ ื ื•ื™ืจื•ืŸ.
10:22
It takes input from other nodes
162
622871
2554
ื”ื•ื ืžืงื‘ืœ ืžื™ื“ืข ืžืฆืžืชื™ื ืื—ืจื™ื
10:25
and sends output to others.
163
625425
2718
ื•ืฉื•ืœื— ืื•ืชื ืœืื—ืจื™ื.
10:28
Moreover, these hundreds of thousands or even millions of nodes
164
628143
4713
ื™ืชืจื” ืžื–ื•, ืžืื•ืช ืืœืคื™ ืื• ืื•ืœื™ ืืคื™ืœื• ืžื™ืœื™ื•ื ื™ ื”ืฆืžืชื™ื
10:32
are organized in hierarchical layers,
165
632856
3227
ืžืื•ืจื’ื ื™ื ื‘ืฉื›ื‘ื•ืช ื”ื™ืจืจื›ื™ื•ืช,
10:36
also similar to the brain.
166
636083
2554
ื’ื ื›ืŸ ื‘ื“ื•ืžื” ืœืžื•ื—.
10:38
In a typical neural network we use to train our object recognition model,
167
638637
4783
ื‘ืจืฉืช ื ื•ื™ืจืืœื™ืช ื˜ื™ืคื•ืกื™ืช ืื ื• ืžืฉืชืžืฉื™ื ื›ื“ื™ ืœืืžืŸ ืืช ื”ืžื•ื“ืœ ื–ื™ื”ื•ื™ ื”ืื•ื‘ื™ื™ืงื˜ื™ื ืฉืœื ื•,
10:43
it has 24 million nodes,
168
643420
3181
ื™ืฉ ื‘ื• 24 ืžื™ืœื™ื•ืŸ ืฆืžืชื™ื,
10:46
140 million parameters,
169
646601
3297
140 ืžื™ืœื™ื•ืŸ ืžืฉืชื ื™ื,
10:49
and 15 billion connections.
170
649898
2763
ื•-15 ืžื™ืœื™ืืจื“ ืงืฉืจื™ื.
10:52
That's an enormous model.
171
652661
2415
ื–ื” ืžื•ื“ืœ ืขื ืง.
10:55
Powered by the massive data from ImageNet
172
655076
3901
ื”ืžื•ื ืข ืขืœ ื™ื“ื™ ืžื™ื“ืข ื ืชื•ื ื™ื ืขืฆื•ื ืžืื™ืžื’'ื ื˜
10:58
and the modern CPUs and GPUs to train such a humongous model,
173
658977
5433
ื•ื™ื—ื™ื“ื•ืช ื”ืขื™ื‘ื•ื“ ื”ืžืจื›ื–ื™ื•ืช ื•ื”ืžืขื‘ื“ื™ื ื”ื’ืจืคื™ื™ื ืœืื™ืžื•ืŸ ืžื•ื“ืœ ื›ื–ื” ื›ื‘ื™ืจ,
11:04
the convolutional neural network
174
664410
2369
ื”ืจืฉืช ื”ื ื•ื™ืจืืœื™ืช ื”ืžื•ืจื›ื‘ืช
11:06
blossomed in a way that no one expected.
175
666779
3436
ืคืจื—ื” ื‘ืฆื•ืจื” ืฉืืฃ ืื—ื“ ืœื ืฆื™ืคื” ืœื”.
11:10
It became the winning architecture
176
670215
2508
ื–ื” ื”ืคืš ืœื”ื™ื•ืช ื”ืืจื›ื™ื˜ืงื˜ื•ืจื” ื”ืžื ืฆื—ืช
11:12
to generate exciting new results in object recognition.
177
672723
5340
ื‘ื™ืฆื™ืจืช ืชื•ืฆืื•ืช ื—ื“ืฉื•ืช ื•ืžืจื’ืฉื•ืช ื‘ื–ื™ื”ื•ื™ ืื•ื‘ื™ื™ืงื˜ื™ื.
11:18
This is a computer telling us
178
678063
2810
ื–ื” ืžื—ืฉื‘ ืฉืื•ืžืจ ืœื ื•
11:20
this picture contains a cat
179
680873
2300
ืฉื”ืชืžื•ื ื” ื”ื–ื• ืžื›ื™ืœื” ื—ืชื•ืœ
11:23
and where the cat is.
180
683173
1903
ื•ืื™ืคื” ื ืžืฆื ื”ื—ืชื•ืœ.
11:25
Of course there are more things than cats,
181
685076
2112
ื›ืžื•ื‘ืŸ ืฉื™ืฉ ื™ื•ืชืจ ื“ื‘ืจื™ื ืžื—ืชื•ืœื™ื,
11:27
so here's a computer algorithm telling us
182
687188
2438
ืื– ื”ื ื” ืืœื’ื•ืจื™ืชื ืฉืœ ืžื—ืฉื‘ ืื•ืžืจ ืœื ื•
11:29
the picture contains a boy and a teddy bear;
183
689626
3274
ืฉื”ืชืžื•ื ื” ืžื›ื™ืœื” ื™ืœื“ ื•ื‘ื•ื‘ืช ื“ื•ื‘ื™;
11:32
a dog, a person, and a small kite in the background;
184
692900
4366
ื›ืœื‘, ืื“ื, ื•ืขืคื™ืคื•ืŸ ืงื˜ืŸ ื‘ืจืงืข;
11:37
or a picture of very busy things
185
697266
3135
ืื• ืชืžื•ื ื” ืฉืœ ื“ื‘ืจื™ื ืžืื•ื“ ืขืกื•ืงื™ื
11:40
like a man, a skateboard, railings, a lampost, and so on.
186
700401
4644
ื›ืžื• ืื™ืฉ, ืกืงื™ื™ื˜ื‘ื•ืจื“, ืžืขืงื•ืช, ืขืžื•ื“ ืชืื•ืจื” ื•ื›ืŸ ื”ืœืื”.
11:45
Sometimes, when the computer is not so confident about what it sees,
187
705045
5293
ืœืคืขืžื™ื, ื›ืฉื”ืžื—ืฉื‘ ืœื ื‘ื˜ื•ื— ืœื’ืžืจื™ ื‘ืžื” ืฉื”ื•ื ืจื•ืื”,
11:51
we have taught it to be smart enough
188
711498
2276
ืœื™ืžื“ื ื• ืื•ืชื• ืœื”ื™ื•ืช ื—ื›ื ืžืกืคื™ืง
11:53
to give us a safe answer instead of committing too much,
189
713774
3878
ื›ื“ื™ ืœืชืช ืœื ื• ืชืฉื•ื‘ื” ื‘ื˜ื•ื—ื” ื‘ืžืงื•ื ืœื”ืชื—ื™ื™ื‘ ื™ื•ืชืจ ืžื“ื™,
11:57
just like we would do,
190
717652
2811
ื‘ื“ื™ื•ืง ื›ืžื• ืฉืื ื—ื ื• ื”ื™ื™ื ื• ืขื•ืฉื™ื,
12:00
but other times our computer algorithm is remarkable at telling us
191
720463
4666
ืื‘ืœ ื‘ืคืขืžื™ื ืื—ืจื•ืช ื”ืืœื’ื•ืจื™ืชื ื”ืžืžื•ื—ืฉื‘ ืฉืœื ื• ืžืฆื•ื™ื™ืŸ ื‘ืœื”ื’ื™ื“ ืœื ื•
12:05
what exactly the objects are,
192
725129
2253
ืžื” ื‘ื“ื™ื•ืง ื”ื ื”ืื•ื‘ื™ื™ืงื˜ื™ื,
12:07
like the make, model, year of the cars.
193
727382
3436
ื›ืžื• ื”ื™ืฆืจืŸ, ื”ืžื•ื“ืœ ื•ื”ืฉื ื” ืฉืœ ืžื›ื•ื ื™ื•ืช.
12:10
We applied this algorithm to millions of Google Street View images
194
730818
5386
ื™ื™ืฉืžื ื• ืืช ื”ืืœื’ื•ืจื™ืชื ื”ื–ื” ืœืžื™ืœื™ื•ื ื™ ืชืžื•ื ื•ืช ืฉืœ ืžืคืช ื”ืจื—ื•ื‘ื•ืช ืฉืœ ื’ื•ื’ืœ
12:16
across hundreds of American cities,
195
736204
3135
ื‘ืžืื•ืช ืขืจื™ื ืืžืจื™ืงื ื™ื•ืช,
12:19
and we have learned something really interesting:
196
739339
2926
ื•ืœืžื“ื ื• ืžืฉื”ื• ืžืื•ื“ ืžืขื ื™ื™ืŸ:
12:22
first, it confirmed our common wisdom
197
742265
3320
ื“ื‘ืจ ืจืืฉื•ืŸ, ื–ื” ืื™ืžืช ืืช ื”ื™ื“ืข ื”ื ืคื•ืฅ
12:25
that car prices correlate very well
198
745585
3290
ืฉืžื—ื™ืจื™ ืžื›ื•ื ื™ื•ืช ื ืžืฆืื™ื ื‘ืงืฉืจ ื™ืฉื™ืจ
12:28
with household incomes.
199
748875
2345
ืœื”ื›ื ืกื•ืช ืžืฉืง ื”ื‘ื™ืช.
12:31
But surprisingly, car prices also correlate well
200
751220
4527
ืื‘ืœ ื‘ืื•ืคืŸ ืžืคืชื™ืข, ืžื—ื™ืจื™ ื”ืžื›ื•ื ื™ื•ืช ื ืžืฆืื™ื ื‘ืงืฉืจ ื™ืฉื™ืจ
12:35
with crime rates in cities,
201
755747
2300
ืขื ืจืžืช ื”ืคืฉืข ื‘ืขืจื™ื,
12:39
or voting patterns by zip codes.
202
759007
3963
ืื• ืชื‘ื ื™ืช ื”ืฆื‘ืขื•ืช ืขืœ ืคื™ ืžื™ืงื•ื“ื™ื.
12:44
So wait a minute. Is that it?
203
764060
2206
ืื– ืจื’ืข. ื–ื” ื”ื›ืœ?
12:46
Has the computer already matched or even surpassed human capabilities?
204
766266
5153
ื”ืื ื”ืžื—ืฉื‘ ื”ืฉื•ื•ื” ืื• ืืคื™ืœื• ืขืงืฃ ืืช ื”ื™ื›ื•ืœื•ืช ื”ืื ื•ืฉื™ื•ืช?
12:51
Not so fast.
205
771419
2138
ืœื ื›ืœ ื›ืš ืžื”ืจ.
12:53
So far, we have just taught the computer to see objects.
206
773557
4923
ืขื“ ืขื›ืฉื™ื•, ืื ื—ื ื• ืจืง ืœื™ืžื“ื ื• ืืช ื”ืžื—ืฉื‘ ืœืจืื•ืช ืื•ื‘ื™ื™ืงื˜ื™ื.
12:58
This is like a small child learning to utter a few nouns.
207
778480
4644
ื–ื” ื›ืžื• ืฉื™ืœื“ ืงื˜ืŸ ืœื•ืžื“ ืœื‘ื˜ื ืžืกืคืจ ืฉืžื•ืช ืขืฆื.
13:03
It's an incredible accomplishment,
208
783124
2670
ื–ื” ื”ื™ืฉื’ ืžื“ื”ื™ื,
13:05
but it's only the first step.
209
785794
2460
ืื‘ืœ ื–ื” ืจืง ื”ืฆืขื“ ื”ืจืืฉื•ืŸ.
13:08
Soon, another developmental milestone will be hit,
210
788254
3762
ื‘ืงืจื•ื‘, ืขื•ื“ ืื‘ืŸ ื“ืจืš ื”ืชืคืชื—ื•ืชื™ืช ืชื•ืฉื’,
13:12
and children begin to communicate in sentences.
211
792016
3461
ื•ื™ืœื“ื™ื ืžืชื—ื™ืœื™ื ืœืชืงืฉืจ ื‘ืžืฉืคื˜ื™ื.
13:15
So instead of saying this is a cat in the picture,
212
795477
4224
ืื– ื‘ืžืงื•ื ืœื”ื’ื™ื“ - ื–ื” ื—ืชื•ืœ ื‘ืชืžื•ื ื”,
13:19
you already heard the little girl telling us this is a cat lying on a bed.
213
799701
5202
ื›ื‘ืจ ืฉืžืขืชื ืืช ื”ื™ืœื“ื” ื”ืงื˜ื ื” ืื•ืžืจืช ืœื ื• ืฉื–ื” ื—ืชื•ืœ ืฉืฉื•ื›ื‘ ืขืœ ืžื™ื˜ื”.
13:24
So to teach a computer to see a picture and generate sentences,
214
804903
5595
ืื– ืœืœืžื“ ืžื—ืฉื‘ ืœืจืื•ืช ืชืžื•ื ื” ื•ืœื™ื™ืฆืจ ืžืฉืคื˜ื™ื,
13:30
the marriage between big data and machine learning algorithm
215
810498
3948
ื”ื ื™ืฉื•ืื™ื ื‘ื™ืŸ ื‘ื™ื’ ื“ืื˜ื” ืœืืœื’ื•ืจื™ืชื ืœื™ืžื•ื“ ืžื›ื•ื ื”
13:34
has to take another step.
216
814446
2275
ืฆืจื™ื›ื™ื ืœืงื—ืช ืขื•ื“ ืฆืขื“.
13:36
Now, the computer has to learn from both pictures
217
816721
4156
ืขื›ืฉื™ื•, ื”ืžื—ืฉื‘ ืฆืจื™ืš ืœืœืžื•ื“ ืžืฉืชื™ ื”ืชืžื•ื ื•ืช
13:40
as well as natural language sentences
218
820877
2856
ื›ืžื• ื’ื ืžืžืฉืคื˜ื™ื ื˜ื‘ืขื™ื™ื ื‘ืฉืคื”
13:43
generated by humans.
219
823733
3322
ืฉื ื•ืฆืจื™ื ืขืœ ื™ื“ื™ ื‘ื ื™ ืื“ื.
13:47
Just like the brain integrates vision and language,
220
827055
3853
ื‘ื“ื™ื•ืง ื›ืžื• ืฉื”ืžื•ื— ืžื™ื™ืฆืจ ืจืื™ื™ื” ื•ืฉืคื”,
13:50
we developed a model that connects parts of visual things
221
830908
5201
ืื ื—ื ื• ืคื™ืชื—ื ื• ืžื•ื“ืœ ืฉืžืงืฉืจ ื—ืœืงื™ื ืฉืœ ื“ื‘ืจื™ื ื•ื™ื–ื•ืืœื™ื
13:56
like visual snippets
222
836109
1904
ื›ืžื• ืžืงื˜ืขื™ื ืงืฆืจื™ื
13:58
with words and phrases in sentences.
223
838013
4203
ืขื ืžื™ืœื™ื ื•ื‘ื™ื˜ื•ื™ื™ื ื‘ืžืฉืคื˜ื™ื.
14:02
About four months ago,
224
842216
2763
ืœืคื ื™ ืืจื‘ืขื” ื—ื•ื“ืฉื™ื ื‘ืขืจืš,
14:04
we finally tied all this together
225
844979
2647
ืกื•ืฃ ืกื•ืฃ ืงืฉืจื ื• ืืช ื›ืœ ื–ื” ื‘ื™ื—ื“
14:07
and produced one of the first computer vision models
226
847626
3784
ื•ื™ืฆืจื ื• ืืช ืื—ื“ ืžืžื•ื“ืœื™ ื”ืจืื™ื™ื” ื”ืžืžื•ื—ืฉื‘ืช ื”ืจืืฉื•ื ื™ื
14:11
that is capable of generating a human-like sentence
227
851410
3994
ืฉืžืกื•ื’ืœื™ื ืœื™ื™ืฆืจ ืžืฉืคื˜ ื‘ื“ื•ืžื” ืœืื“ื
14:15
when it sees a picture for the first time.
228
855404
3506
ื›ืืฉืจ ื”ื•ื ืจื•ืื” ืชืžื•ื ื” ื‘ืคืขื ื”ืจืืฉื•ื ื”.
14:18
Now, I'm ready to show you what the computer says
229
858910
4644
ืขื›ืฉื™ื•, ืื ื™ ืžื•ื›ื ื” ืœื”ืจืื•ืช ืœื›ื ืžื” ื”ืžื—ืฉื‘ ืื•ืžืจ
14:23
when it sees the picture
230
863554
1975
ื›ืฉื”ื•ื ืจื•ืื” ืืช ื”ืชืžื•ื ื”
14:25
that the little girl saw at the beginning of this talk.
231
865529
3830
ืฉื”ื™ืœื“ื” ื”ืงื˜ื ื” ืจืืชื” ื‘ืชื—ื™ืœืช ื”ืฉื™ื—ื” ื”ื–ื•.
14:31
(Video) Computer: A man is standing next to an elephant.
232
871519
3344
(ื•ื™ื“ืื•) ืžื—ืฉื‘: ืื™ืฉ ืขื•ืžื“ ืœื™ื“ ืคื™ืœ.
14:36
A large airplane sitting on top of an airport runway.
233
876393
3634
ืžื˜ื•ืก ื’ื“ื•ืœ ืขื•ืžื“ ืขืœ ืžืกืœื•ืœ ื˜ื™ืกื”.
14:41
FFL: Of course, we're still working hard to improve our algorithms,
234
881057
4212
ืคื™ื™ ืคื™ื™: ื›ืžื•ื‘ืŸ, ืื ื—ื ื• ืขื“ื™ื™ืŸ ืขื•ื‘ื“ื™ื ืงืฉื” ื›ื“ื™ ืœืฉืคืจ ืืช ื”ืืœื’ื•ืจื™ืชืžื™ื ืฉืœื ื•,
14:45
and it still has a lot to learn.
235
885269
2596
ื•ืขื“ื™ื™ืŸ ื™ืฉ ืœื• ื”ืจื‘ื” ืœืœืžื•ื“.
14:47
(Applause)
236
887865
2291
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
14:51
And the computer still makes mistakes.
237
891556
3321
ื•ื”ืžื—ืฉื‘ ืขื“ื™ื™ืŸ ืขื•ืฉื” ื˜ืขื•ื™ื•ืช.
14:54
(Video) Computer: A cat lying on a bed in a blanket.
238
894877
3391
(ื•ื™ื“ืื•) ืžื—ืฉื‘: ื—ืชื•ืœ ืฉื•ื›ื‘ ืขืœ ืžื™ื˜ื” ื‘ืฉืžื™ื›ื”.
14:58
FFL: So of course, when it sees too many cats,
239
898268
2553
ืคื™ื™ ืคื™ื™: ืื– ื›ืžื•ื‘ืŸ, ื›ืฉื”ื•ื ืจื•ืื” ื™ื•ืชืจ ืžื“ื™ ื—ืชื•ืœื™ื,
15:00
it thinks everything might look like a cat.
240
900821
2926
ื”ื•ื ื—ื•ืฉื‘ ืฉื”ื›ืœ ื™ื›ื•ืœ ืœื”ื™ืจืื•ืช ื›ืžื• ื—ืชื•ืœ.
15:05
(Video) Computer: A young boy is holding a baseball bat.
241
905317
2864
(ื•ื™ื“ืื•) ืžื—ืฉื‘: ื™ืœื“ ืฆืขื™ืจ ืžื—ื–ื™ืง ืืœืช ื‘ื™ื™ืกื‘ื•ืœ.
15:08
(Laughter)
242
908181
1765
(ืฆื—ื•ืง)
15:09
FFL: Or, if it hasn't seen a toothbrush, it confuses it with a baseball bat.
243
909946
4583
ืคื™ื™ ืคื™ื™: ืื• ืื ื”ื•ื ืœื ืจืื” ืžื‘ืจืฉืช ืฉื™ื ื™ื™ื, ื”ื•ื ืžื‘ืœื‘ืœ ืืช ื–ื” ืขื ืืœืช ื‘ื™ื™ืกื‘ื•ืœ.
15:15
(Video) Computer: A man riding a horse down a street next to a building.
244
915309
3434
(ื•ื™ื“ืื•) ืžื—ืฉื‘: ืื™ืฉ ืจื•ื›ื‘ ืขืœ ืกื•ืก ื‘ืžื•ืจื“ ื”ืจื—ื•ื‘ ืœื™ื“ ื‘ื ื™ื™ืŸ.
15:18
(Laughter)
245
918743
2023
(ืฆื—ื•ืง)
15:20
FFL: We haven't taught Art 101 to the computers.
246
920766
3552
ืคื™ื™ ืคื™ื™: ืœื ืœื™ืžื“ื ื• ืืช ื”ืžื—ืฉื‘ื™ื ืžื‘ื•ื ืœืื•ืžื ื•ืช.
15:25
(Video) Computer: A zebra standing in a field of grass.
247
925768
2884
(ื•ื™ื“ืื•) ืžื—ืฉื‘: ื–ื‘ืจื” ืขื•ืžื“ืช ื‘ืฉื“ื” ืขืฉื‘.
15:28
FFL: And it hasn't learned to appreciate the stunning beauty of nature
248
928652
3367
ืคื™ื™ ืคื™ื™: ื•ื”ื•ื ืœื ืœืžื“ ืœื”ืขืจื™ืš ืืช ื”ื™ื•ืคื™ ื”ืžื“ื”ื™ื ืฉืœ ื”ื˜ื‘ืข
15:32
like you and I do.
249
932019
2438
ื›ืžื•ื ื™ ื•ื›ืžื•ื›ื.
15:34
So it has been a long journey.
250
934457
2832
ืื– ืขื“ื™ื™ืŸ ื™ืฉ ืœื• ื“ืจืš ืืจื•ื›ื”.
15:37
To get from age zero to three was hard.
251
937289
4226
ืœื”ื’ื™ืข ืžื’ื™ืœ ืืคืก ืœืฉืœื•ืฉ ื”ื™ื” ืงืฉื”.
15:41
The real challenge is to go from three to 13 and far beyond.
252
941515
5596
ื”ืืชื’ืจ ื”ืืžื™ืชื™ ื”ื•ื ืœื”ื’ื™ืข ืžืฉืœื•ืฉ ืœืฉืœื•ืฉ ืขืฉืจื” ื•ืžืขื‘ืจ ืœื–ื”.
15:47
Let me remind you with this picture of the boy and the cake again.
253
947111
4365
ืชืจืฉื• ืœื™ ืœื”ื–ื›ื™ืจ ืœื›ื ืขื ื”ืชืžื•ื ื” ืฉืœ ื”ื™ืœื“ ื•ื”ืขื•ื’ื” ืฉื•ื‘.
15:51
So far, we have taught the computer to see objects
254
951476
4064
ืขื“ ืขื›ืฉื™ื•, ืœื™ืžื“ื ื• ืืช ื”ืžื—ืฉื‘ ืœืจืื•ืช ืื•ื‘ื™ื™ืงื˜ื™ื
15:55
or even tell us a simple story when seeing a picture.
255
955540
4458
ืื• ืืคื™ืœื• ืœืกืคืจ ืœื ื• ืกื™ืคื•ืจ ืคืฉื•ื˜ ื›ืฉื”ื•ื ืจื•ืื” ืชืžื•ื ื”.
15:59
(Video) Computer: A person sitting at a table with a cake.
256
959998
3576
(ื•ื™ื“ืื•) ืžื—ืฉื‘: ืื“ื ื™ื•ืฉื‘ ืœื™ื“ ืฉื•ืœื—ืŸ ืขื ืขื•ื’ื”.
16:03
FFL: But there's so much more to this picture
257
963574
2630
ืคื™ื™ ืคื™ื™: ืื‘ืœ ื™ืฉ ืขื•ื“ ื›ืœ ื›ืš ื”ืจื‘ื” ื‘ืชืžื•ื ื” ื”ื–ื•
16:06
than just a person and a cake.
258
966204
2270
ืžืืฉืจ ืจืง ืื“ื ื•ืขื•ื’ื”.
16:08
What the computer doesn't see is that this is a special Italian cake
259
968474
4467
ืžื” ืฉื”ืžื—ืฉื‘ ืœื ืจื•ืื” ื–ื” ืฉื–ื• ืขื•ื’ื” ืื™ื˜ืœืงื™ืช ืžื™ื•ื—ื“ืช
16:12
that's only served during Easter time.
260
972941
3217
ืฉืžื•ื’ืฉืช ืจืง ื‘ื—ื’ ื”ืคืกื—ื.
16:16
The boy is wearing his favorite t-shirt
261
976158
3205
ื”ื™ืœื“ ืœื•ื‘ืฉ ืืช ื”ื—ื•ืœืฆื” ื”ืื”ื•ื‘ื” ืขืœื™ื•
16:19
given to him as a gift by his father after a trip to Sydney,
262
979363
3970
ืฉื ื™ืชื ื” ืœื• ืขืœ ื™ื“ื™ ืื‘ื™ื• ืื—ืจื™ ื˜ื™ื•ืœ ื‘ืกื™ื“ื ื™,
16:23
and you and I can all tell how happy he is
263
983333
3808
ื•ืืชื ื•ืื ื™ ื™ื›ื•ื™ื ืœื”ื’ื™ื“ ื›ืžื” ืžืื•ืฉืจ ื”ื•ื
16:27
and what's exactly on his mind at that moment.
264
987141
3203
ื•ืžื” ื‘ื“ื™ื•ืง ื”ื•ื ื—ื•ืฉื‘ ื‘ืจื’ืข ื”ื–ื”.
16:31
This is my son Leo.
265
991214
3125
ื–ื” ื‘ื ื™ ืœื™ืื•.
16:34
On my quest for visual intelligence,
266
994339
2624
ื‘ืžืกืข ืฉืœื™ ืœืื™ื ื˜ืœื’ื ืฆื™ื” ื—ื–ื•ืชื™ืช,
16:36
I think of Leo constantly
267
996963
2391
ืื ื™ ื›ืœ ื”ื–ืžืŸ ื—ื•ืฉื‘ืช ืขืœ ืœื™ืื•
16:39
and the future world he will live in.
268
999354
2903
ื•ืขืœ ื”ืขื•ืœื ื”ืขืชื™ื“ื™ ื‘ื• ื”ื•ื ื™ื—ื™ื”.
16:42
When machines can see,
269
1002257
2021
ื›ืฉืžื›ื•ื ื•ืช ื™ื›ื•ืœื•ืช ืœืจืื•ืช,
16:44
doctors and nurses will have extra pairs of tireless eyes
270
1004278
4712
ืœืจื•ืคืื™ื ื•ืื—ื™ื•ืช ื™ื”ื™ื” ืขื•ื“ ื–ื•ื’ ืขื™ื ื™ื™ื ืฉืœื ืžืชืขื™ื™ืคื•ืช
16:48
to help them to diagnose and take care of patients.
271
1008990
4092
ื›ื“ื™ ืœืขื–ื•ืจ ืœื”ื ืœืื‘ื—ืŸ ื•ืœื“ืื•ื’ ืœืžื˜ื•ืคืœื™ื.
16:53
Cars will run smarter and safer on the road.
272
1013082
4383
ืžื›ื•ื ื™ื•ืช ื™ื ื•ืขื• ื‘ืฆื•ืจื” ื—ื›ืžื” ื™ื•ืชืจ ื•ื‘ื˜ื•ื—ื” ื™ื•ืชืจ ื‘ื“ืจื›ื™ื.
16:57
Robots, not just humans,
273
1017465
2694
ืจื•ื‘ื•ื˜ื™ื, ืœื ืจืง ื‘ื ื™ ืื“ื,
17:00
will help us to brave the disaster zones to save the trapped and wounded.
274
1020159
4849
ื™ืขื–ืจื• ืœื ื• ืœืขืžื•ื“ ื‘ื’ื‘ื•ืจื” ื‘ืื–ื•ืจื™ ืืกื•ืŸ ื•ืœื”ืฆื™ืœ ืืช ื”ืœื›ื•ื“ื™ื ื•ื”ืคืฆื•ืขื™ื.
17:05
We will discover new species, better materials,
275
1025798
3796
ืื ื—ื ื• ื ื’ืœื” ืžื™ื ื™ื ื—ื“ืฉื™ื, ื—ื•ืžืจื™ื ื˜ื•ื‘ื™ื ื™ื•ืชืจ,
17:09
and explore unseen frontiers with the help of the machines.
276
1029594
4509
ื•ื ื—ืงื•ืจ ื’ื‘ื•ืœื•ืช ื—ื“ืฉื™ื ืขื ืขื–ืจื” ืฉืœ ื”ืžื›ื•ื ื•ืช.
17:15
Little by little, we're giving sight to the machines.
277
1035113
4167
ืœืื˜ ืœืื˜, ืื ื—ื ื• ื ื•ืชื ื™ื ื™ื›ื•ืœืช ืจืื™ื™ื” ืœืžื›ื•ื ื•ืช.
17:19
First, we teach them to see.
278
1039280
2798
ืงื•ื“ื ื›ืœ, ืื ื—ื ื• ืžืœืžื“ื™ื ืื•ืชื ืœืจืื•ืช.
17:22
Then, they help us to see better.
279
1042078
2763
ืœืื—ืจ ืžื›ืŸ, ื”ื ื™ืขื–ืจื• ืœื ื• ืœืจืื•ืช ื˜ื•ื‘ ื™ื•ืชืจ.
17:24
For the first time, human eyes won't be the only ones
280
1044841
4165
ื‘ืคืขื ื”ืจืืฉื•ื ื”, ื”ืขื™ื ื™ื™ื ื”ืื ื•ืฉื™ื•ืช ืœื ื™ื”ื™ื• ื”ืขื™ื ื™ื™ื ื”ื™ื—ื™ื“ื•ืช
17:29
pondering and exploring our world.
281
1049006
2934
ืฉืžื”ืจื”ืจื•ืช ื•ื—ื•ืงืจื•ืช ืืช ื”ืขื•ืœื.
17:31
We will not only use the machines for their intelligence,
282
1051940
3460
ืื ื—ื ื• ืœื ืจืง ื ืฉืชืžืฉ ื‘ืžื›ื•ื ื•ืช ื‘ืฉื‘ื™ืœ ื”ืื™ื ื˜ืœื™ื’ื ืฆื™ื” ืฉืœื”ื,
17:35
we will also collaborate with them in ways that we cannot even imagine.
283
1055400
6179
ืื ื—ื ื• ื’ื ื ืฉืชืฃ ืื™ืชืŸ ืคืขื•ืœื” ื‘ื“ืจื›ื™ื ืฉืื ื—ื ื• ืืคื™ืœื• ืœื ื™ื›ื•ืœื™ื ืœื“ืžื™ื™ืŸ.
17:41
This is my quest:
284
1061579
2161
ื–ื” ื”ืžืกืข ืฉืœื™:
17:43
to give computers visual intelligence
285
1063740
2712
ืœืชืช ืœืžื—ืฉื‘ื™ื ืชื‘ื•ื ื” ื—ื–ื•ืชื™ืช
17:46
and to create a better future for Leo and for the world.
286
1066452
5131
ื•ืœื™ืฆื•ืจ ืขืชื™ื“ ื˜ื•ื‘ ื™ื•ืชืจ ื‘ืฉื‘ื™ืœ ืœื™ืื• ื•ื‘ืฉื‘ื™ืœ ื”ืขื•ืœื.
17:51
Thank you.
287
1071583
1811
ืชื•ื“ื” ืจื‘ื”.
17:53
(Applause)
288
1073394
3785
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
ืขืœ ืืชืจ ื–ื”

ืืชืจ ื–ื” ื™ืฆื™ื’ ื‘ืคื ื™ื›ื ืกืจื˜ื•ื ื™ YouTube ื”ืžื•ืขื™ืœื™ื ืœืœื™ืžื•ื“ ืื ื’ืœื™ืช. ืชื•ื›ืœื• ืœืจืื•ืช ืฉื™ืขื•ืจื™ ืื ื’ืœื™ืช ื”ืžื•ืขื‘ืจื™ื ืขืœ ื™ื“ื™ ืžื•ืจื™ื ืžื”ืฉื•ืจื” ื”ืจืืฉื•ื ื” ืžืจื—ื‘ื™ ื”ืขื•ืœื. ืœื—ืฅ ืคืขืžื™ื™ื ืขืœ ื”ื›ืชื•ื‘ื™ื•ืช ื‘ืื ื’ืœื™ืช ื”ืžื•ืฆื’ื•ืช ื‘ื›ืœ ื“ืฃ ื•ื™ื“ืื• ื›ื“ื™ ืœื”ืคืขื™ืœ ืืช ื”ืกืจื˜ื•ืŸ ืžืฉื. ื”ื›ืชื•ื‘ื™ื•ืช ื’ื•ืœืœื•ืช ื‘ืกื ื›ืจื•ืŸ ืขื ื”ืคืขืœืช ื”ื•ื•ื™ื“ืื•. ืื ื™ืฉ ืœืš ื”ืขืจื•ืช ืื• ื‘ืงืฉื•ืช, ืื ื ืฆื•ืจ ืื™ืชื ื• ืงืฉืจ ื‘ืืžืฆืขื•ืช ื˜ื•ืคืก ื™ืฆื™ืจืช ืงืฉืจ ื–ื”.

https://forms.gle/WvT1wiN1qDtmnspy7