The wonderful and terrifying implications of computers that can learn | Jeremy Howard

598,999 views ใƒป 2014-12-16

TED


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

ืžืชืจื’ื: Shir Ben Asher Kestin ืžื‘ืงืจ: Zeeva Livshitz
00:12
It used to be that if you wanted to get a computer to do something new,
0
12880
4013
ืคืขื, ืื ืจืฆื™ืช ืœื’ืจื•ื ืœืžื—ืฉื‘ ืœืขืฉื•ืช ืžืฉื”ื• ื—ื“ืฉ
00:16
you would have to program it.
1
16893
1554
ื”ื™ื™ืช ืฆืจื™ืš ืœืชื›ื ืช ืื•ืชื•.
00:18
Now, programming, for those of you here that haven't done it yourself,
2
18447
3411
ื›ืขืช, ืชื›ื ื•ืช, ืœืžื™ ืžื›ื ืฉืœื ืขืฉื• ื–ืืช ื‘ืขืฆืžื,
00:21
requires laying out in excruciating detail
3
21858
3502
ื“ื•ืจืฉ ืœืคืจื•ืฉ ื‘ืคื™ืจื•ื˜ ืžื™ื™ื’ืข
ื›ืœ ืฆืขื“ ื•ืฆืขื“ ืฉืืชื” ืจื•ืฆื” ืฉื”ืžื—ืฉื‘ ื™ืขืฉื”
00:25
every single step that you want the computer to do
4
25360
3367
00:28
in order to achieve your goal.
5
28727
2362
ืขืœ ืžื ืช ืœื”ืฉื™ื’ ืืช ื”ืžื˜ืจื” ืฉืœืš.
ื›ืขืช, ืื ืืชื” ืจื•ืฆื” ืœืขืฉื•ืช ืžืฉื”ื• ืฉืืชื” ืœื ื™ื•ื“ืข ืื™ืš ืœืขืฉื•ืช ื‘ืขืฆืžืš,
00:31
Now, if you want to do something that you don't know how to do yourself,
6
31089
3496
00:34
then this is going to be a great challenge.
7
34585
2063
ืื– ื–ื” ื”ื•ืœืš ืœื”ื™ื•ืช ืืชื’ืจ ืจืฆื™ื ื™.
00:36
So this was the challenge faced by this man, Arthur Samuel.
8
36648
3483
ืื– ื–ื” ื”ื™ื” ื”ืืชื’ืจ ืฉืขืžื“ ื‘ืคื ื™ ื”ืื™ืฉ ื”ื–ื”, ืืจืชื•ืจ ืกืžื•ืืœ.
ื‘-1956, ื”ื•ื ืจืฆื” ืœื’ืจื•ื ืœืžื—ืฉื‘ ื”ื–ื”
00:40
In 1956, he wanted to get this computer
9
40131
4077
ืœื”ื™ื•ืช ืžืกื•ื’ืœ ืœื”ื‘ื™ืก ืื•ืชื• ื‘ื“ืžืงื”.
00:44
to be able to beat him at checkers.
10
44208
2340
00:46
How can you write a program,
11
46548
2040
ืื™ืš ืืชื” ื™ื›ื•ืœ ืœื›ืชื•ื‘ ืชื•ื›ื ื”,
00:48
lay out in excruciating detail, how to be better than you at checkers?
12
48588
3806
ืœืคืจื•ืฉ ื‘ืคื™ืจื•ื˜ ืžื™ื™ื’ืข, ืื™ืš ืœื”ื™ื•ืช ื™ื•ืชืจ ื˜ื•ื‘ ืžืžืš ื‘ื“ืžืงื”?
00:52
So he came up with an idea:
13
52394
1722
ืื– ื”ื•ื ื”ื’ื” ืจืขื™ื•ืŸ:
ื”ื•ื ื ืชืŸ ืœืžื—ืฉื‘ ืœืฉื—ืง ื ื’ื“ ืขืฆืžื• ืืœืคื™ ืคืขืžื™ื
00:54
he had the computer play against itself thousands of times
14
54116
3724
00:57
and learn how to play checkers.
15
57840
2524
ื•ืœืœืžื•ื“ ืื™ืš ืœืฉื—ืง ื“ืžืงื”.
01:00
And indeed it worked, and in fact, by 1962,
16
60364
3180
ื•ืื›ืŸ ื–ื” ืขื‘ื“, ื•ืœืžืขืฉื” ืขื“ 1962,
01:03
this computer had beaten the Connecticut state champion.
17
63544
4017
ื”ืžื—ืฉื‘ ื”ื–ื” ื ื™ืฆื— ืืช ื”ืืœื•ืฃ ืฉืœ ืงื•ื ื˜ื™ืงื˜.
01:07
So Arthur Samuel was the father of machine learning,
18
67561
2973
ืื– ืืจืชื•ืจ ืกืžื•ืืœ ื”ื™ื” ืื‘ื™ ื”ืœืžื™ื“ื” ื”ื—ื™ืฉื•ื‘ื™ืช.
01:10
and I have a great debt to him,
19
70534
1717
ื•ืื ื™ ื—ื‘ ืœื• ื—ื•ื‘ ืขื ืง.
ื‘ื’ืœืœ ืฉืื ื™ ืขื•ืกืง ื‘ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช.
01:12
because I am a machine learning practitioner.
20
72251
2763
ื”ื™ื™ืชื™ ื”ื ืฉื™ื ืฉืœ ืงืื’ืœ,
01:15
I was the president of Kaggle,
21
75014
1465
01:16
a community of over 200,000 machine learning practictioners.
22
76479
3388
ืงื”ื™ืœื” ืฉืžื•ื ื” ืžืขืœ 200,000 ืื ืฉื™ื ืฉืขื•ืกืงื™ื ื‘ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช.
01:19
Kaggle puts up competitions
23
79867
2058
ืงืื’ืœ ืžืืจื’ื ืช ืชื—ืจื•ื™ื•ืช
01:21
to try and get them to solve previously unsolved problems,
24
81925
3708
ืœื ืกื•ืช ื•ืœื’ืจื•ื ืœื”ื ืœืคืชื•ืจ ื‘ืขื™ื•ืช ืฉืœื ื ืคืชืจื• ืขื“ ื›ื”,
01:25
and it's been successful hundreds of times.
25
85633
3837
ื•ื–ื• ื”ื™ื™ืชื” ื”ืฆืœื—ื” ืžืื•ืช ืคืขืžื™ื.
01:29
So from this vantage point, I was able to find out
26
89470
2470
ืื– ืžืชืฆืคื™ืช ืžื ืงื•ื“ืช ื™ืชืจื•ืŸ ื–ื•, ื”ืฆืœื—ืชื™ ืœื’ืœื•ืช
01:31
a lot about what machine learning can do in the past, can do today,
27
91940
3950
ื”ืจื‘ื” ืขืœ ืžื” ืฉืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช ื™ื›ื•ืœื” ืœืขืฉื•ืช ื‘ืขื‘ืจ, ืœืขืฉื•ืช ื”ื™ื•ื,
01:35
and what it could do in the future.
28
95890
2362
ื•ืžื” ื”ื™ื ื™ื›ื•ืœื” ืœืขืฉื•ืช ื‘ืขืชื™ื“.
ืื•ืœื™ ื”ื”ืฆืœื—ื” ื”ื’ื“ื•ืœื” ื”ืจืืฉื•ื ื” ืฉืœ ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช ื‘ืื•ืคืŸ ืžืกื—ืจื™ ื”ื™ื™ืชื” ื’ื•ื’ืœ.
01:38
Perhaps the first big success of machine learning commercially was Google.
29
98252
4423
01:42
Google showed that it is possible to find information
30
102675
3109
ื’ื•ื’ืœ ื”ืจืืชื” ืฉื–ื” ืืคืฉืจื™ ืœืžืฆื•ื ืžื™ื“ืข
01:45
by using a computer algorithm,
31
105784
1752
ืขืœ ื™ื“ื™ ืฉื™ืžื•ืฉ ื‘ืืœื’ื•ืจื™ืชื ืžืžื•ื—ืฉื‘,
01:47
and this algorithm is based on machine learning.
32
107536
2901
ื•ื”ืืœื’ื•ืจื™ืชื ื”ื–ื” ืžื‘ื•ืกืก ืขืœ ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช.
01:50
Since that time, there have been many commercial successes of machine learning.
33
110437
3886
ืžืื–, ื”ื™ื• ื”ืจื‘ื” ื”ืฆืœื—ื•ืช ืžืกื—ืจื™ื•ืช ืฉืœ ืœืžื™ื“ื” ืžืžื•ื—ืฉื‘ืช.
01:54
Companies like Amazon and Netflix
34
114323
1837
ื—ื‘ืจื•ืช ื›ืžื• ืืžื–ื•ืŸ ื•ื ื˜ืคืœื™ืงืก
01:56
use machine learning to suggest products that you might like to buy,
35
116160
3716
ืžืฉืชืžืฉื•ืช ื‘ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช ื›ื“ื™ ืœื”ืฆื™ืข ืžื•ืฆืจื™ื ืฉืื•ืœื™ ืชืจืฆื” ืœืงื ื•ืช,
01:59
movies that you might like to watch.
36
119876
2020
ืกืจื˜ื™ื ืฉืื•ืœื™ ืชืจืฆื” ืœืจืื•ืช.
02:01
Sometimes, it's almost creepy.
37
121896
1807
ืœืคืขืžื™ื, ื–ื” ื›ืžืขื˜ ืžืคื—ื™ื“.
02:03
Companies like LinkedIn and Facebook
38
123703
1954
ื—ื‘ืจื•ืช ื›ืžื• ืœื™ื ืงื“ืื™ืŸ ื•ืคื™ื™ืกื‘ื•ืง
02:05
sometimes will tell you about who your friends might be
39
125657
2594
ืœืคืขืžื™ื ืื•ืžืจื•ืช ืœืš ืžื™ ืขืฉื•ื™ื™ื ืœื”ื™ื•ืช ื”ื—ื‘ืจื™ื ืฉืœืš
02:08
and you have no idea how it did it,
40
128251
1977
ื•ืื™ืŸ ืœืš ืžื•ืฉื’ ืื™ืš ื”ืŸ ืขืฉื• ืืช ื–ื”,
02:10
and this is because it's using the power of machine learning.
41
130228
2967
ื•ื–ื” ื‘ื’ืœืœ ืฉื”ืŸ ืžืฉืชืžืฉื•ืช ื‘ื›ื•ื—ื” ืฉืœ ื”ืœืžื™ื“ื” ื”ื—ื™ืฉื•ื‘ื™ืช.
02:13
These are algorithms that have learned how to do this from data
42
133195
2957
ืืœื• ื”ื ืืœื’ื•ืจื™ืชืžื™ื ืฉืœืžื“ื• ืื™ืš ืœืขืฉื•ืช ื–ืืช ืžื ืชื•ื ื™ื
02:16
rather than being programmed by hand.
43
136152
3247
ื‘ืžืงื•ื ืœื”ื™ื•ืช ืžืชื•ื›ื ืชื•ืช ืœืขืฉื•ืช ื–ืืช ื‘ืื•ืคืŸ ื™ื“ื ื™.
02:19
This is also how IBM was successful
44
139399
2478
ื›ืš ื’ื IBM ื”ืฆืœื™ื—ื”
02:21
in getting Watson to beat the two world champions at "Jeopardy,"
45
141877
3862
ืœื’ืจื•ื ืœื•ื•ื˜ืกื•ืŸ ืœื”ื‘ื™ืก ืืช ืฉื ื™ ืืœื•ืคื™ ื”ืขื•ืœื ื‘"ื’'ืคืจื“ื™",
02:25
answering incredibly subtle and complex questions like this one.
46
145739
3225
ื•ืœืขื ื•ืช ืขืœ ืฉืืœื•ืช ืขื“ื™ื ื•ืช ื•ืžื•ืจื›ื‘ื•ืช ืœื”ืคืœื™ื ื›ืžื• ื–ืืช:
02:28
["The ancient 'Lion of Nimrud' went missing from this city's national museum in 2003 (along with a lot of other stuff)"]
47
148964
2835
["ื”'ืืจื™ื” ืฉืœ ื ืžืจื•ื“' ื”ืขืชื™ืง ื ืขืœื ืžื”ืžื•ื–ื™ืื•ืŸ ื”ืœืื•ืžื™ ืฉืœ ื”ืขื™ืจ ื”ื–ืืช ื‘-2003 (ื™ื—ื“ ืขื ืขื•ื“ ื”ืจื‘ื” ื“ื‘ืจื™ื)"]
02:31
This is also why we are now able to see the first self-driving cars.
48
151799
3235
ื–ื• ื’ื ื”ืกื™ื‘ื” ื‘ื’ืœืœื” ืื ื• ืขื›ืฉื™ื• ืจื•ืื™ื ืืช ื”ืžื›ื•ื ื™ื•ืช ื”ืจืืฉื•ื ื•ืช ืฉื ื•ื”ื’ื•ืช ืœื‘ื“.
02:35
If you want to be able to tell the difference between, say,
49
155034
2822
ืื ืืชื ืจื•ืฆื™ื ืœื”ื™ื•ืช ืžืกื•ื’ืœื™ื ืœื”ื‘ื—ื™ืŸ ื‘ื™ืŸ, ืœืžืฉืœ,
02:37
a tree and a pedestrian, well, that's pretty important.
50
157856
2632
ืขืฅ ื•ื”ื•ืœืš ืจื’ืœ, ื•ื‘ื›ืŸ, ื–ื” ื“ื™ ื—ืฉื•ื‘.
02:40
We don't know how to write those programs by hand,
51
160488
2587
ืื ื—ื ื• ืœื ื™ื•ื“ืขื™ื ืื™ืš ืœื›ืชื•ื‘ ืืช ื”ืชื•ื›ื ื•ืช ื”ืœืœื• ื‘ืื•ืคืŸ ื™ื“ื ื™,
02:43
but with machine learning, this is now possible.
52
163075
2997
ืืš ืขื ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช, ื–ื” ืืคืฉืจื™ ื›ืขืช.
02:46
And in fact, this car has driven over a million miles
53
166072
2608
ื•ืœืžืขืฉื”, ื”ืžื›ื•ื ื™ืช ื”ื–ื• ื ืกืขื” ืžืขืœ ืžื™ืœื™ื•ืŸ ืžื™ื™ืœื™ื
02:48
without any accidents on regular roads.
54
168680
3506
ืœืœื ืืฃ ืชืื•ื ื” ื‘ื›ื‘ื™ืฉื™ื ืจื’ื™ืœื™ื.
02:52
So we now know that computers can learn,
55
172196
3914
ืื– ืื ื—ื ื• ื™ื•ื“ืขื™ื ืฉืžื—ืฉื‘ื™ื ื™ื›ื•ืœื™ื ืœืœืžื•ื“,
02:56
and computers can learn to do things
56
176110
1900
ื•ืžื—ืฉื‘ื™ื ื™ื›ื•ืœื™ื ืœืœืžื•ื“ ืœืขืฉื•ืช ื“ื‘ืจื™ื
02:58
that we actually sometimes don't know how to do ourselves,
57
178010
2838
ืฉืื ื—ื ื• ืœืžืขืฉื” ืœืคืขืžื™ื ืœื ื™ื•ื“ืขื™ื ืœืขืฉื•ืช ื‘ืขืฆืžื ื•,
03:00
or maybe can do them better than us.
58
180848
2885
ืื• ืื•ืœื™ ื™ื›ื•ืœื™ื ืœืขืฉื•ืช ืื•ืชื ื™ื•ืชืจ ื˜ื•ื‘ ืžืื™ืชื ื•.
03:03
One of the most amazing examples I've seen of machine learning
59
183733
4195
ืื—ืช ืžื”ื“ื•ื’ืžืื•ืช ื”ืžื“ื”ื™ืžื•ืช ื‘ื™ื•ืชืจ ืฉืจืื™ืชื™ ืœืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช
03:07
happened on a project that I ran at Kaggle
60
187928
2392
ืงืจืชื” ื‘ืคืจื•ื™ื™ืงื˜ ืฉื”ื•ื‘ืœืชื™ ื‘ืงืื’ืœ
03:10
where a team run by a guy called Geoffrey Hinton
61
190320
3591
ืฉื‘ื• ืฆื•ื•ืช ืฉื”ื•ื‘ืœ ืขืœ ื™ื“ื™ ื‘ื—ื•ืจ ื‘ืฉื ื’'ืคืจื™ ื”ื™ื ื˜ื•ืŸ
03:13
from the University of Toronto
62
193911
1552
ืžืื•ื ื™ื‘ืจืกื™ื˜ืช ื˜ื•ืจื•ื ื˜ื•
03:15
won a competition for automatic drug discovery.
63
195463
2677
ื–ื›ื” ื‘ืชื—ืจื•ืช ืขืœ ืคื™ืชื•ื— ืชืจื•ืคื•ืช ืื•ื˜ื•ืžื˜ื™.
03:18
Now, what was extraordinary here is not just that they beat
64
198140
2847
ื›ืขืช, ืžื” ืฉืžื“ื”ื™ื ืคื” ื–ื” ืœื ืจืง ืฉื”ื ื ื™ืฆื—ื•
03:20
all of the algorithms developed by Merck or the international academic community,
65
200987
4013
ืืช ื›ืœ ื”ืืœื’ื•ืจื™ืชืžื™ื ืฉืคื•ืชื—ื• ืขืœ ื™ื“ื™ ืžืจืง ืื• ืงื”ื™ืœืช ื”ืืงื“ืžื™ื” ื”ืขื•ืœืžื™ืช,
03:25
but nobody on the team had any background in chemistry or biology or life sciences,
66
205000
5061
ืืœื ืฉืœืืฃ ืื—ื“ ื‘ืฆื•ื•ืช ืœื ื”ื™ื” ืจืงืข ื‘ื›ื™ืžื™ื”, ื‘ื™ื•ืœื•ื’ื™ื” ืื• ืžื“ืขื™ ื”ื—ื™ื™ื,
03:30
and they did it in two weeks.
67
210061
2169
ื•ื”ื ืขืฉื• ื–ืืช ื‘ืฉื‘ื•ืขื™ื™ื.
03:32
How did they do this?
68
212230
1381
ืื™ืš ื”ื ืขืฉื• ื–ืืช?
03:34
They used an extraordinary algorithm called deep learning.
69
214421
2921
ื”ื ื”ืฉืชืžืฉื• ื‘ืืœื’ื•ืจื™ืชื ื™ื•ืฆื ื“ื•ืคืŸ ืฉื ืงืจื ืœืžื™ื“ื” ืขืžื•ืงื”.
03:37
So important was this that in fact the success was covered
70
217342
2949
ื–ื” ื”ื™ื” ื›ืœ ื›ืš ื—ืฉื•ื‘ ืฉื”ื”ืฆืœื—ื” ืกื•ืงืจื”
03:40
in The New York Times in a front page article a few weeks later.
71
220291
3121
ื‘ื ื™ื• ื™ื•ืจืง ื˜ื™ื™ืžืก ื‘ืžืืžืจ ื‘ืขืžื•ื“ ื”ืจืืฉื™ ื›ืžื” ืฉื‘ื•ืขื•ืช ืœืื—ืจ ืžื›ืŸ.
03:43
This is Geoffrey Hinton here on the left-hand side.
72
223412
2735
ื–ื” ื’'ืคืจื™ ื”ื™ื ื˜ื•ืŸ ื›ืืŸ ืžืฉืžืืœ.
03:46
Deep learning is an algorithm inspired by how the human brain works,
73
226147
4341
ืœืžื™ื“ื” ืขืžื•ืงื” ื”ื•ื ืืœื’ื•ืจื™ืชื ืฉืฉื•ืื‘ ื”ืฉืจืื” ืžื”ืžื•ื— ื”ืื ื•ืฉื™,
03:50
and as a result it's an algorithm
74
230488
1812
ื•ื›ืชื•ืฆืื” ืžื›ืš ื”ื•ื ืืœื’ื•ืจื™ืชื
03:52
which has no theoretical limitations on what it can do.
75
232300
3841
ืฉืื™ืŸ ืœื• ืžื’ื‘ืœื•ืช ืชื™ืื•ืจื˜ื™ื•ืช ืขืœ ืžื” ืฉื”ื•ื ืžืกื•ื’ืœ ืœืขืฉื•ืช.
03:56
The more data you give it and the more computation time you give it,
76
236141
2823
ื›ื›ืœ ืฉื ื•ืชื ื™ื ืœื• ื™ื•ืชืจ ื ืชื•ื ื™ื ื•ื–ืžืŸ ื—ื™ืฉื•ื‘,
03:58
the better it gets.
77
238964
1312
ื›ืš ื”ื•ื ืžืฉืชืคืจ.
04:00
The New York Times also showed in this article
78
240276
2339
ื”ื ื™ื• ื™ื•ืจืง ื˜ื™ื™ืžืก ื’ื ื”ืจืื” ื‘ืžืืžืจ ื”ื–ื”
04:02
another extraordinary result of deep learning
79
242615
2242
ืชื•ืฆืื” ืžื“ื”ื™ืžื” ื ื•ืกืคืช ืฉืœ ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช
04:04
which I'm going to show you now.
80
244857
2712
ืฉืื ื™ ื”ื•ืœืš ืœื”ืจืื•ืช ืœื›ื ื›ืขืช.
04:07
It shows that computers can listen and understand.
81
247569
4941
ื”ื ืžืจืื™ื ืฉืžื—ืฉื‘ื™ื ื™ื›ื•ืœื™ื ืœื”ืงืฉื™ื‘ ื•ืœื”ื‘ื™ืŸ.
04:12
(Video) Richard Rashid: Now, the last step
82
252510
2711
(ื•ื™ื“ืื•) ืจื™ืฆ'ืจื“ ืจืืฉื™ื“: ืขื›ืฉื™ื•, ื”ืฆืขื“ ื”ืื—ืจื•ืŸ
04:15
that I want to be able to take in this process
83
255221
3025
ืฉืื ื™ ืจื•ืฆื” ืœื”ื™ื•ืช ืžืกื•ื’ืœ ืœื‘ืฆืข ื‘ืชื”ืœื™ืš ื”ื–ื”
04:18
is to actually speak to you in Chinese.
84
258246
4715
ื”ื•ื ืœืžืขืฉื” ืœื“ื‘ืจ ืืœื™ื›ื ื‘ืกื™ื ื™ืช.
04:22
Now the key thing there is,
85
262961
2635
ืขื›ืฉื™ื• ื”ืžืคืชื— ืคื”,
04:25
we've been able to take a large amount of information from many Chinese speakers
86
265596
5002
ื”ื•ื ืฉื”ืฆืœื—ื ื• ืœืงื—ืช ื›ืžื•ืช ื’ื“ื•ืœื” ืฉืœ ืžื™ื“ืข ืžื”ืจื‘ื” ื“ื•ื‘ืจื™ ืกื™ื ื™ืช
04:30
and produce a text-to-speech system
87
270598
2530
ื•ืœื™ื™ืฆืจ ืžืขืจื›ืช ืฉืœ ื˜ืงืกื˜-ืœื“ื™ื‘ื•ืจ
04:33
that takes Chinese text and converts it into Chinese language,
88
273128
4673
ืฉืœื•ืงื—ืช ื˜ืงืกื˜ ื‘ืกื™ื ื™ืช ื•ื”ื•ืคื›ืช ืื•ืชื• ืœืฉืคื” ืกื™ื ื™ืช,
04:37
and then we've taken an hour or so of my own voice
89
277801
4128
ื•ืื– ืœืงื—ื ื• ื›ืฉืขื” ืฉืœ ื”ืงื•ืœ ืฉืœื™
04:41
and we've used that to modulate
90
281929
1891
ื•ื”ืฉืชืžืฉื ื• ื‘ื–ื” ื›ื“ื™ ืœื”ืชืื™ื
04:43
the standard text-to-speech system so that it would sound like me.
91
283820
4544
ืืช ื”ืžืขืจื›ืช ื”ืกื˜ื ื“ืจื˜ื™ืช ืฉืœ ื˜ืงืกื˜-ืœื“ื™ื‘ื•ืจ ื›ืš ืฉื”ื™ื ืชื™ืฉืžืข ื›ืžื•ื ื™.
04:48
Again, the result's not perfect.
92
288364
2540
ืฉื•ื‘, ื”ืชื•ืฆืื” ืœื ืžื•ืฉืœืžืช.
04:50
There are in fact quite a few errors.
93
290904
2648
ื™ืฉ ืœืžืขืฉื” ื“ื™ ื”ืจื‘ื” ืฉื’ื™ืื•ืช.
04:53
(In Chinese)
94
293552
2484
(ื‘ืกื™ื ื™ืช)
04:56
(Applause)
95
296036
3367
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
05:01
There's much work to be done in this area.
96
301446
3576
ื™ืฉ ื”ืจื‘ื” ืขื‘ื•ื“ื” ืœืขืฉื•ืช ื‘ืชื—ื•ื ื”ื–ื”.
05:05
(In Chinese)
97
305022
3645
(ืกื™ื ื™ืช)
05:08
(Applause)
98
308667
3433
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
05:13
Jeremy Howard: Well, that was at a machine learning conference in China.
99
313345
3399
(ื’'ืจืžื™ ื”ื•ื•ืืจื“:) ื•ื‘ื›ืŸ, ื–ื” ื”ื™ื” ื‘ื›ื ืก ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช ื‘ืกื™ืŸ.
05:16
It's not often, actually, at academic conferences
100
316744
2370
ื–ื” ื“ื™ ื ื“ื™ืจ, ืœืžืขืฉื”, ื‘ื›ื ืกื™ื ืืงื“ืžืื™ื
05:19
that you do hear spontaneous applause,
101
319114
1897
ืœืฉืžื•ืข ืžื—ื™ืื•ืช ื›ืคื™ื™ื ืกืคื•ื ื˜ื ื™ื•ืช,
05:21
although of course sometimes at TEDx conferences, feel free.
102
321011
3676
ืœืžืจื•ืช ืฉื›ืžื•ื‘ืŸ ืœืคืขืžื™ื ื‘ื›ื ืกื™ื ืฉืœ TEDx, ื”ืจื’ื™ืฉื• ื—ื•ืคืฉื™.
05:24
Everything you saw there was happening with deep learning.
103
324687
2795
ื›ืœ ืžื” ืฉืจืื™ืชื ืฉื ืงืจื” ืขื ืœืžื™ื“ื” ืขืžื•ืงื”.
05:27
(Applause) Thank you.
104
327482
1525
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื) ืชื•ื“ื”.
05:29
The transcription in English was deep learning.
105
329007
2282
ื”ืชืžืœื•ืœ ืœืื ื’ืœื™ืช ื”ื™ื” ืœืžื™ื“ื” ืขืžื•ืงื”.
05:31
The translation to Chinese and the text in the top right, deep learning,
106
331289
3412
ื”ืชืจื’ื•ื ืœืกื™ื ื™ืช ื•ื”ื˜ืงืกื˜ ืžื™ืžื™ืŸ ืœืžืขืœื”- ืœืžื™ื“ื” ืขืžื•ืงื”.
05:34
and the construction of the voice was deep learning as well.
107
334701
3307
ื•ื”ื‘ื ื™ื” ืฉืœ ื”ืงื•ืœ ื”ื™ืชื” ืœืžื™ื“ื” ืขืžื•ืงื” ื’ื ื›ืŸ.
05:38
So deep learning is this extraordinary thing.
108
338008
3234
ืื– ืœืžื™ื“ื” ืขืžื•ืงื” ื”ื™ื ื“ื‘ืจ ืžื“ื”ื™ื.
05:41
It's a single algorithm that can seem to do almost anything,
109
341242
3099
ื–ื” ืืœื’ื•ืจื™ืชื ื™ื—ื™ื“ ืฉื™ื›ื•ืœ ืœืขืฉื•ืช ื›ืžืขื˜ ื”ื›ืœ.
05:44
and I discovered that a year earlier, it had also learned to see.
110
344341
3111
ื•ื’ื™ืœื™ืชื™ ืฉืฉื ื” ืงื•ื“ื ืœื›ืŸ, ื”ื•ื ื’ื ืœืžื“ ืœืจืื•ืช.
05:47
In this obscure competition from Germany
111
347452
2176
ื‘ืชื—ืจื•ืช ืœื ืžื•ื›ืจืช ื‘ื’ืจืžื ื™ื”
05:49
called the German Traffic Sign Recognition Benchmark,
112
349628
2597
ื‘ืฉื ื”ื‘ื•ื—ืŸ ื”ื’ืจืžื ื™ ืœื‘ื™ืฆื•ืขื™ ื–ื™ื”ื•ื™ ืฉืœื˜ื™ ืชื ื•ืขื”
05:52
deep learning had learned to recognize traffic signs like this one.
113
352225
3393
ืœืžื™ื“ื” ืขืžื•ืงื” ืœืžื“ื” ืœื–ื”ื•ืช ืฉืœื˜ื™ ืชื ื•ืขื” ื›ืžื• ื–ื”
05:55
Not only could it recognize the traffic signs
114
355618
2094
ื”ื™ื ืœื ืจืง ื–ื™ื”ืชื” ืืช ื”ืฉืœื˜ื™ื
05:57
better than any other algorithm,
115
357712
1758
ื™ื•ืชืจ ื˜ื•ื‘ ืžื›ืœ ืืœื’ื•ืจื™ืชื ืื—ืจ,
05:59
the leaderboard actually showed it was better than people,
116
359470
2719
ืœื•ื— ื”ืชื•ืฆืื•ืช ืžืžืฉ ื”ืจืื” ืฉื”ื™ื ื™ื•ืชืจ ื˜ื•ื‘ื” ืžืื ืฉื™ื,
06:02
about twice as good as people.
117
362189
1852
ื‘ืขืจืš ืคื™ 2 ื™ื•ืชืจ ื˜ื•ื‘ ืžืื ืฉื™ื.
06:04
So by 2011, we had the first example
118
364041
1996
ืื– ืขื“ 2011, ื”ื™ืชื” ืœื ื• ืืช ื”ื“ื•ื’ืžื” ื”ืจืืฉื•ื ื”
06:06
of computers that can see better than people.
119
366037
3405
ืœืžื—ืฉื‘ื™ื ืฉื™ื›ื•ืœื™ื ืœืจืื•ืช ื™ื•ืชืจ ื˜ื•ื‘ ืžืื ืฉื™ื.
06:09
Since that time, a lot has happened.
120
369442
2049
ืžืื–, ื”ืจื‘ื” ื”ืชืจื—ืฉ.
06:11
In 2012, Google announced that they had a deep learning algorithm
121
371491
3514
ื‘-2012, ื’ื•ื’ืœ ื”ื›ืจื™ื–ื” ืฉื”ื™ื ื ืชื ื” ืœืืœื’ื•ืจื™ืชื ืœืžื™ื“ื” ืขืžื•ืงื”
06:15
watch YouTube videos
122
375005
1415
ืœืฆืคื•ืช ื‘ืกืจื˜ื•ื ื™ ื™ื•ื˜ื™ื•ื‘
06:16
and crunched the data on 16,000 computers for a month,
123
376420
3437
ื•ื”ื ืขื‘ื“ื• ืขืœ ื”ื ืชื•ื ื™ื ืขืœ 16,000 ืžื—ืฉื‘ื™ื ืœืžืฉืš ื—ื•ื“ืฉ,
06:19
and the computer independently learned about concepts such as people and cats
124
379857
4361
ื•ื”ืžื—ืฉื‘ ืœืžื“ ื‘ืื•ืคืŸ ืขืฆืžืื™ ืขืœ ืžื•ืฉื’ื™ื ื›ืžื• ืื ืฉื™ื ื•ื—ืชื•ืœื™ื
06:24
just by watching the videos.
125
384218
1809
ืจืง ืžืฆืคื™ื™ื” ื‘ืกืจื˜ื•ื ื™ื.
06:26
This is much like the way that humans learn.
126
386027
2352
ื–ื” ืžืื•ื“ ื“ื•ืžื” ืœื“ืจืš ื‘ื” ืื ืฉื™ื ืœื•ืžื“ื™ื.
06:28
Humans don't learn by being told what they see,
127
388379
2740
ื‘ื ื™ ืื“ื ืœื ืœื•ืžื“ื™ื ืขืœ ื™ื“ื™ ื›ืš ืฉืื•ืžืจื™ื ืœื”ื ืžื” ื”ื ืจื•ืื™ื,
06:31
but by learning for themselves what these things are.
128
391119
3331
ืืœื ืขืœ ื™ื“ื™ ืœืžื™ื“ื” ืขืฆืžืื™ืช ืฉืœ ืžื”ื ืื•ืชื ื“ื‘ืจื™ื.
06:34
Also in 2012, Geoffrey Hinton, who we saw earlier,
129
394450
3369
ื’ื ื‘-2012, ื’'ืคืจื™ ื”ื™ื ื˜ื•ืŸ, ืฉืจืื™ื ื• ืงื•ื“ื,
06:37
won the very popular ImageNet competition,
130
397819
2858
ื ื™ืฆื— ื‘ืชื—ืจื•ืช ืื™ืžื’'ื ื˜ ื”ืžืื•ื“ ืคื•ืคื•ืœืจื™ืช,
06:40
looking to try to figure out from one and a half million images
131
400677
4141
ืฉื‘ื™ืงืฉื” ืœื ืกื•ืช ืœื’ืœื•ืช ืขืœ 1.5 ืžื™ืœื™ื•ืŸ ืชืžื•ื ื•ืช
06:44
what they're pictures of.
132
404818
1438
ืชืžื•ื ื•ืช ืฉืœ ืžื” ื”ืŸ.
06:46
As of 2014, we're now down to a six percent error rate
133
406256
3533
ื ื›ื•ืŸ ืœ-2014, ื™ืจื“ื ื• ืœืฉื™ืขื•ืจ ืฉื’ื™ืื•ืช ืฉืœ 6%
06:49
in image recognition.
134
409789
1453
ื‘ื–ื™ื”ื•ื™ ืชืžื•ื ื”.
06:51
This is better than people, again.
135
411242
2026
ื–ื” ื™ื•ืชืจ ื˜ื•ื‘ ืžืื ืฉื™ื, ืฉื•ื‘.
06:53
So machines really are doing an extraordinarily good job of this,
136
413268
3769
ืื– ืžื›ื•ื ื•ืช ื‘ืืžืช ืขื•ืฉื•ืช ืขื‘ื•ื“ื” ืžืฆื•ื™ื™ื ืช ื‘ื›ืš,
06:57
and it is now being used in industry.
137
417037
2269
ื•ื–ื” ืขื›ืฉื™ื• ื ื›ื ืก ืœืฉื™ืžื•ืฉ ื‘ืชืขืฉื™ื™ื”.
06:59
For example, Google announced last year
138
419306
3042
ืœืžืฉืœ, ื’ื•ื’ืœ ื”ื›ืจื™ื–ื” ื‘ืฉื ื” ืฉืขื‘ืจื”
07:02
that they had mapped every single location in France in two hours,
139
422348
4585
ืฉื”ื ืžื™ืคื• ื›ืœ ืžื™ืงื•ื ื•ืžื™ืงื•ื ื‘ืฆืจืคืช ืชื•ืš ืฉืขืชื™ื™ื,
07:06
and the way they did it was that they fed street view images
140
426933
3447
ื•ื”ื“ืจืš ื‘ื” ื”ื ืขืฉื• ื–ืืช ื”ื™ืชื” ื”ื–ื ืช ืชืžื•ื ื•ืช ืจื—ื•ื‘
07:10
into a deep learning algorithm to recognize and read street numbers.
141
430380
4319
ืœืืœื’ื•ืจื™ืชื ืœืžื™ื“ื” ืขืžื•ืงื” ืฉื™ื–ื”ื” ื•ื™ืงืจื ืžืกืคืจื™ ืจื—ื•ื‘ื•ืช.
07:14
Imagine how long it would have taken before:
142
434699
2220
ื“ืžื™ื™ื ื• ื›ืžื” ื–ืžืŸ ื–ื” ื”ื™ื” ืœื•ืงื— ื‘ืขื‘ืจ:
07:16
dozens of people, many years.
143
436919
3355
ืขืฉืจื•ืช ืื ืฉื™ื, ืฉื ื™ื ืจื‘ื•ืช.
07:20
This is also happening in China.
144
440274
1911
ื–ื” ื’ื ืงื•ืจื” ื‘ืกื™ืŸ.
07:22
Baidu is kind of the Chinese Google, I guess,
145
442185
4036
ื‘ืื™ื“ื• ื”ื•ื ืกื•ื’ ืฉืœ ื”ื’ื•ื’ืœ ื”ืกื™ื ื™, ืื ื™ ืžื ื™ื—.
07:26
and what you see here in the top left
146
446221
2283
ื•ืžื” ืฉืจื•ืื™ื ื›ืืŸ ื‘ืฉืžืืœ ืœืžืขืœื”
07:28
is an example of a picture that I uploaded to Baidu's deep learning system,
147
448504
3974
ื”ื•ื ื“ื•ื’ืžื” ืœืชืžื•ื ื” ืฉืื ื™ ื”ืขืœื™ืชื™ ืœืžืขืจื›ืช ื”ืœืžื™ื“ื” ื”ืขืžื•ืงื” ืฉืœ ื‘ืื™ื“ื•,
07:32
and underneath you can see that the system has understood what that picture is
148
452478
3769
ื•ืžืชื—ืช ืืคืฉืจ ืœืจืื•ืช ืฉื”ืžืขืจื›ืช ื”ื‘ื™ื ื” ืžื” ื”ืชืžื•ื ื”
07:36
and found similar images.
149
456247
2236
ื•ืžืฆืื” ืชืžื•ื ื•ืช ื“ื•ืžื•ืช.
07:38
The similar images actually have similar backgrounds,
150
458483
2736
ืœืชืžื•ื ื•ืช ื”ื“ื•ืžื•ืช ืœืžืขืฉื” ื™ืฉ ืจืงืขื™ื ื“ื•ืžื™ื,
07:41
similar directions of the faces,
151
461219
1658
ื›ื™ื•ื•ื ื™ ืคืจืฆื•ืคื™ื ื“ื•ืžื™ื,
07:42
even some with their tongue out.
152
462877
1788
ืืคื™ืœื• ื›ืžื” ืขื ื”ืœืฉื•ืŸ ื‘ื—ื•ืฅ.
07:44
This is not clearly looking at the text of a web page.
153
464665
3030
ื–ื” ืœื ืœื”ืกืชื›ืœ ืขืœ ื˜ืงืกื˜ ืฉืœ ื“ืฃ ืื™ื ื˜ืจื ื˜.
07:47
All I uploaded was an image.
154
467695
1412
ื›ืœ ืžื” ืฉืื ื™ ื”ืขืœื™ืชื™ ื”ื™ื” ืชืžื•ื ื”.
07:49
So we now have computers which really understand what they see
155
469107
4021
ืื– ื›ืขืช ื™ืฉ ืœื ื• ืžื—ืฉื‘ื™ื ืฉื‘ืืžืช ืžื‘ื™ื ื™ื ืžื” ื”ื ืจื•ืื™ื
07:53
and can therefore search databases
156
473128
1624
ื•ืœื›ืŸ ื™ื›ื•ืœื™ื ืœื—ืคืฉ ื‘ืžืื’ืจื™ ืžื™ื“ืข
07:54
of hundreds of millions of images in real time.
157
474752
3554
ืฉืœ ืžืื•ืช ืžื™ืœื™ื•ื ื™ ืชืžื•ื ื•ืช ื‘ื–ืžืŸ ืืžืช.
07:58
So what does it mean now that computers can see?
158
478306
3230
ืื– ืžื” ื–ื” ืื•ืžืจ ืขื›ืฉื™ื• ืฉืžื—ืฉื‘ื™ื ื™ื›ื•ืœื™ื ืœืจืื•ืช?
08:01
Well, it's not just that computers can see.
159
481536
2017
ื•ื‘ื›ืŸ, ื–ื” ืœื ืจืง ืฉืžื—ืฉื‘ื™ื ื™ื›ื•ืœื™ื ืœืจืื•ืช.
08:03
In fact, deep learning has done more than that.
160
483553
2069
ืœืžืขืฉื”, ืœืžื™ื“ื” ืขืžื•ืงื” ืขืฉืชื” ื™ื•ืชืจ ืžื›ืš.
08:05
Complex, nuanced sentences like this one
161
485622
2948
ืžืฉืคื˜ื™ื ืžื•ืจื›ื‘ื™ื ืขื ื ื™ื•ืื ืกื™ื, ื›ืžื• ื–ื”
08:08
are now understandable with deep learning algorithms.
162
488570
2824
ื”ื ื›ืขืช ืžื•ื‘ื ื™ื ื‘ืขื–ืจืช ืืœื’ื•ืจื™ืชืžื™ื ืฉืœ ืœืžื™ื“ื” ืขืžื•ืงื”.
08:11
As you can see here,
163
491394
1303
ื›ืžื• ืฉืืชื ืจื•ืื™ื ืคื”,
08:12
this Stanford-based system showing the red dot at the top
164
492697
2768
ื”ืžืขืจื›ืช ื”ื–ื• ืฉืžื‘ื•ืกืกืช ื‘ืกื˜ื ืคื•ืจื“ ืฉืžืจืื” ืืช ื”ื ืงื•ื“ื” ื”ืื“ื•ืžื” ืœืžืขืœื”
08:15
has figured out that this sentence is expressing negative sentiment.
165
495465
3919
ื”ื‘ื™ื ื” ืฉื”ืžืฉืคื˜ ื”ื–ื” ืžื‘ื˜ื ืจื’ืฉ ืฉืœื™ืœื™.
08:19
Deep learning now in fact is near human performance
166
499384
3406
ืœืžื™ื“ื” ืขืžื•ืงื” ื”ื™ื ื›ืขืช ืœืžืขืฉื” ื›ืžืขื˜ ื‘ืจืžื” ืฉืœ ื‘ืŸ ืื“ื
08:22
at understanding what sentences are about and what it is saying about those things.
167
502802
5121
ื‘ืœื”ื‘ื™ืŸ ืขืœ ืžื” ืžืฉืคื˜ื™ื ืžื“ื‘ืจื™ื ื•ืžื” ื”ื ืื•ืžืจื™ื ืขืœ ืื•ืชื ื“ื‘ืจื™ื.
08:27
Also, deep learning has been used to read Chinese,
168
507923
2728
ื‘ื ื•ืกืฃ, ืขืฉื• ืฉื™ืžื•ืฉ ื‘ืœืžื™ื“ื” ืขืžื•ืงื” ืœืงืจื™ืืช ืกื™ื ื™ืช,
08:30
again at about native Chinese speaker level.
169
510651
3156
ืฉื•ื‘, ื‘ืขืจืš ื‘ืจืžื” ืฉืœ ื“ื•ื‘ืจ ืกื™ื ื™ืช ืžืœื™ื“ื”.
08:33
This algorithm developed out of Switzerland
170
513807
2168
ื”ืืœื’ื•ืจื™ืชื ื”ื–ื” ืฉืคื•ืชื— ื‘ืฉื•ื•ื™ืฅ
08:35
by people, none of whom speak or understand any Chinese.
171
515975
3356
ืขืœ ื™ื“ื™ ืื ืฉื™ื ืฉืืฃ ืื—ื“ ืžื”ื ืœื ื“ื•ื‘ืจ ืื• ืžื‘ื™ืŸ ืกื™ื ื™ืช.
08:39
As I say, using deep learning
172
519331
2051
ื›ืžื• ืฉืืžืจืชื™, ืฉื™ืžื•ืฉ ื‘ืœืžื™ื“ื” ืขืžื•ืงื”
08:41
is about the best system in the world for this,
173
521382
2219
ื”ื™ื ื‘ืขืจืš ื”ืžืขืจื›ืช ื”ื›ื™ ื˜ื•ื‘ื” ื‘ืขื•ืœื ืœืฉื ื›ืš,
08:43
even compared to native human understanding.
174
523601
5117
ืืคื™ืœื• ื‘ื”ืฉื•ื•ืื” ืœื”ื‘ื ื” ืื ื•ืฉื™ืช.
08:48
This is a system that we put together at my company
175
528718
2964
ื–ื•ื”ื™ ืžืขืจื›ืช ืื•ืชื” ืื ื• ื‘ื•ื ื™ื ื‘ื—ื‘ืจื” ืฉืœื™
08:51
which shows putting all this stuff together.
176
531682
2046
ืฉืžืจืื” ืื™ืš ืžื—ื‘ืจื™ื ืืช ื›ืœ ื”ื“ื‘ืจื™ื ื”ืืœื”.
08:53
These are pictures which have no text attached,
177
533728
2461
ืืœื• ืชืžื•ื ื•ืช ืœื”ืŸ ืœื ืžืฆื•ืจืฃ ื˜ืงืกื˜,
08:56
and as I'm typing in here sentences,
178
536189
2352
ื•ื›ืฉืื ื™ ืžืงืœื™ื“ ืืช ื”ืžืฉืคื˜ื™ื ื”ืืœื” ื›ืืŸ,
08:58
in real time it's understanding these pictures
179
538541
2969
ื‘ื–ืžืŸ ืืžืช ื–ื” ืžื‘ื™ืŸ ืืช ื”ืชืžื•ื ื•ืช ื”ืืœื•
09:01
and figuring out what they're about
180
541510
1679
ื•ืžื‘ื™ืŸ ืขืœ ืžื” ื”ืŸ
09:03
and finding pictures that are similar to the text that I'm writing.
181
543189
3163
ื•ืžื‘ื™ืŸ ืฉื”ืชืžื•ื ื•ืช ื“ื•ืžื•ืช ืœื˜ืงืกื˜ ืฉืื ื™ ื›ื•ืชื‘.
09:06
So you can see, it's actually understanding my sentences
182
546352
2756
ืื– ืืชื ืจื•ืื™ื, ื–ื” ืœืžืขืฉื” ืžื‘ื™ืŸ ืืช ื”ืžืฉืคื˜ื™ื ืฉืœื™
09:09
and actually understanding these pictures.
183
549108
2224
ื•ืžืžืฉ ืžื‘ื™ืŸ ืืช ื”ืชืžื•ื ื•ืช ื”ืœืœื•.
09:11
I know that you've seen something like this on Google,
184
551332
2559
ืื ื™ ื™ื•ื“ืข ืฉืจืื™ืชื ืžืฉื”ื• ื“ื•ืžื” ื‘ื’ื•ื’ืœ,
09:13
where you can type in things and it will show you pictures,
185
553891
2775
ืฉื‘ื• ืืคืฉืจ ืœื”ืงืœื™ื“ ื“ื‘ืจื™ื ื•ื–ื” ืžืจืื” ืœืš ืชืžื•ื ื•ืช,
09:16
but actually what it's doing is it's searching the webpage for the text.
186
556666
3424
ืื‘ืœ ืœืžืขืฉื” ืžื” ืฉื–ื” ืขื•ืฉื” ื–ื” ืœื—ืคืฉ ื‘ื“ืฃ ื”ืื™ื ื˜ืจื ื˜ ืื—ืจ ื˜ืงืกื˜.
09:20
This is very different from actually understanding the images.
187
560090
3001
ื–ื” ืžืื•ื“ ืฉื•ื ื” ืžืืฉืจ ืžืžืฉ ืœื”ื‘ื™ืŸ ืืช ื”ืชืžื•ื ื•ืช.
09:23
This is something that computers have only been able to do
188
563091
2752
ื–ื” ืžืฉื”ื• ืฉืžื—ืฉื‘ื™ื ื”ื™ื• ืžืกื•ื’ืœื™ื ืœืขืฉื•ืช
09:25
for the first time in the last few months.
189
565843
3248
ื‘ืคืขื ื”ืจืืฉื•ื ื” ืจืง ื‘ื—ื•ื“ืฉื™ื ื”ืื—ืจื•ื ื™ื.
09:29
So we can see now that computers can not only see but they can also read,
190
569091
4091
ืื– ืื ื• ืจื•ืื™ื ืขื›ืฉื™ื• ืฉืžื—ืฉื‘ื™ื ื™ื›ื•ืœื™ื ืœื ืจืง ืœืจืื•ืช ืืœื ื’ื ืœืงืจื•ื,
09:33
and, of course, we've shown that they can understand what they hear.
191
573182
3765
ื•ื›ืžื•ื‘ืŸ, ื”ืจืื™ื ื• ืฉื”ื ื™ื›ื•ืœื™ื ืœื”ื‘ื™ืŸ ืืช ืžื” ืฉื”ื ืฉื•ืžืขื™ื.
09:36
Perhaps not surprising now that I'm going to tell you they can write.
192
576947
3442
ืื•ืœื™ ื–ื” ืœื ืžืคืชื™ืข ื›ืขืช ืฉืื ื™ ืขื•ืžื“ ืœืกืคืจ ืœื›ื ืฉื”ื ื™ื›ื•ืœื™ื ืœื›ืชื•ื‘.
09:40
Here is some text that I generated using a deep learning algorithm yesterday.
193
580389
4783
ื”ื ื” ื˜ืงืกื˜ ืฉื™ื™ืฆืจืชื™ ืขืœ ื™ื“ื™ ืืœื’ื•ืจื™ืชื ืœืžื™ื“ื” ืขืžื•ืงื” ืืชืžื•ืœ.
09:45
And here is some text that an algorithm out of Stanford generated.
194
585172
3924
ื•ื”ื ื” ื˜ืงืกื˜ ืฉืืœื’ื•ืจื™ืชื ืžืกื˜ื ืคื•ืจื“ ื™ื™ืฆืจ.
09:49
Each of these sentences was generated
195
589096
1764
ื›ืœ ืื—ื“ ืžื”ืžืฉืคื˜ื™ื ื”ืœืœื• ื ื•ืฆืจ
09:50
by a deep learning algorithm to describe each of those pictures.
196
590860
4249
ืขืœ ื™ื’ื™ ืืœื’ื•ืจื™ืชื ืœืžื™ื“ื” ืขืžื•ืงื” ืขืœ ืžื ืช ืœืชืืจ ื›ืœ ืื—ืช ืžื”ืชืžื•ื ื•ืช ื”ืืœื•.
09:55
This algorithm before has never seen a man in a black shirt playing a guitar.
197
595109
4472
ื”ืืœื’ื•ืจื™ืชื ื”ื–ื” ืœื ืจืื” ืœืคื ื™ ื›ืŸ ืžืขื•ืœื ืื™ืฉ ื‘ื—ื•ืœืฆื” ืฉื—ื•ืจื” ืฉืžื ื’ืŸ ื‘ื’ื™ื˜ืจื”.
09:59
It's seen a man before, it's seen black before,
198
599581
2220
ื”ื•ื ืจืื” ืื“ื ื‘ืขื‘ืจ, ื”ื•ื ืจืื” ืฉื—ื•ืจ ื‘ืขื‘ืจ,
10:01
it's seen a guitar before,
199
601801
1599
ื”ื•ื ืจืื” ื’ื™ื˜ืจื” ื‘ืขื‘ืจ,
10:03
but it has independently generated this novel description of this picture.
200
603400
4294
ืืš ื”ื•ื ื™ื™ืฆืจ ื‘ืื•ืคืŸ ืขืฆืžืื™ ืืช ื”ืชื™ืื•ืจ ื”ื—ื“ืฉ ืฉืœ ื”ืชืžื•ื ื”.
10:07
We're still not quite at human performance here, but we're close.
201
607694
3502
ืื ื—ื ื• ืขื“ื™ื™ืŸ ืœื ื‘ื“ื™ื•ืง ื‘ืจืžื” ืื ื•ืฉื™ืช ื›ืืŸ, ืืš ืื ื—ื ื• ืงืจื•ื‘ื™ื.
10:11
In tests, humans prefer the computer-generated caption
202
611196
4068
ื‘ืžื‘ื—ื ื™ื, ื‘ื ื™ ืื“ื ืžืขื“ื™ืคื™ื ืืช ื”ื›ื•ืชืจืช ื”ืžื™ื•ืฆืจืช ืขืœ ื™ื“ื™ ืžื—ืฉื‘
10:15
one out of four times.
203
615264
1527
ืื—ืช ืœื›ืœ ืืจื‘ืข ืคืขืžื™ื.
10:16
Now this system is now only two weeks old,
204
616791
2064
ืขื›ืฉื™ื• ื”ืžืขืจื›ืช ื”ื–ื• ื”ื™ื ืจืง ื‘ืช ืฉื‘ื•ืขื™ื™ื,
10:18
so probably within the next year,
205
618855
1846
ืื– ืกื‘ื™ืจ ืฉืžืชื™ืฉื”ื• ื‘ืฉื ื” ื”ื‘ืื”,
10:20
the computer algorithm will be well past human performance
206
620701
2801
ื”ืืœื’ื•ืจื™ืชื ื”ืžืžื•ื—ืฉื‘ ื™ืขื‘ื•ืจ ื‘ื”ืจื‘ื” ืืช ื”ื™ื›ื•ืœื•ืช ื”ืื ื•ืฉื™ื•ืช
10:23
at the rate things are going.
207
623502
1862
ื‘ืงืฆื‘ ืฉื‘ื• ื“ื‘ืจื™ื ืžืชืงื“ืžื™ื.
10:25
So computers can also write.
208
625364
3049
ืื– ืžื—ืฉื‘ื™ื ื’ื ื™ื›ื•ืœื™ื ืœื›ืชื•ื‘.
10:28
So we put all this together and it leads to very exciting opportunities.
209
628413
3475
ืื ืœื•ืงื—ื™ื ืืช ื›ืœ ื–ื” ื™ื—ื“ ื–ื” ืžื•ื‘ื™ืœ ืœืืคืฉืจื•ื™ื•ืช ืžืจื’ืฉื•ืช.
10:31
For example, in medicine,
210
631888
1492
ืœืžืฉืœ, ื‘ืจืคื•ืื”,
10:33
a team in Boston announced that they had discovered
211
633380
2525
ืฆื•ื•ืช ื‘ื‘ื•ืกื˜ื•ืŸ ื”ื›ืจื™ื– ืฉื”ื ื’ื™ืœื•
10:35
dozens of new clinically relevant features
212
635905
2949
ืขืฉืจื•ืช ืžืืคื™ื™ื ื™ื ืจืœื•ื•ื ื˜ื™ื™ื ืงืœื™ื ื™ืช
10:38
of tumors which help doctors make a prognosis of a cancer.
213
638854
4266
ืฉืœ ื’ื™ื“ื•ืœื™ื, ืฉื™ื›ื•ืœื™ื ืœืขื–ื•ืจ ืœืจื•ืคืื™ื ืœืงื‘ืœ ืคืจื•ื’ื ื•ื–ื” ืฉืœ ื”ืกืจื˜ืŸ.
10:44
Very similarly, in Stanford,
214
644220
2296
ื‘ืื•ืคืŸ ื“ื•ืžื”, ื‘ืกื˜ื ืคื•ืจื“,
10:46
a group there announced that, looking at tissues under magnification,
215
646516
3663
ืงื‘ื•ืฆื” ื”ื›ืจื™ื–ื” ื›ื™, ืœืื—ืจ ืฉื”ืกืชื›ืœื• ืขืœ ืจืงืžื•ืช ื‘ื”ื’ื“ืœื”,
10:50
they've developed a machine learning-based system
216
650179
2381
ื”ื ืคื™ืชื—ื• ืžืขืจื›ืช ืžื‘ื•ืกืกืช ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช
10:52
which in fact is better than human pathologists
217
652560
2582
ืฉื”ื™ื ืœืžืขืฉื” ื˜ื•ื‘ื” ื™ื•ืชืจ ืžืคืชื•ืœื•ื’ื™ื ืื ื•ืฉื™ื™ื
10:55
at predicting survival rates for cancer sufferers.
218
655142
4377
ื‘ื—ื™ื–ื•ื™ ืกื™ื›ื•ื™ื™ ื”ื™ืฉืจื“ื•ืช ืขื‘ื•ืจ ื”ืกื•ื‘ืœื™ื ืžืกืจื˜ืŸ.
10:59
In both of these cases, not only were the predictions more accurate,
219
659519
3245
ื‘ืฉื ื™ ื”ืžืงืจื™ื ื”ืืœื”, ืœื ืจืง ืฉื”ืชื—ื–ื™ื•ืช ื”ื™ื• ื™ื•ืชืจ ืžื“ื•ื™ื™ืงื•ืช,
11:02
but they generated new insightful science.
220
662764
2502
ืืœื ืฉื”ืŸ ื’ื ื”ืคื™ืงื• ืžื“ืข ื—ื“ืฉ ื‘ืขืœ ืชื•ื‘ื ื•ืช.
11:05
In the radiology case,
221
665276
1505
ื‘ืžืงืจื” ืฉืœ ื”ืจื“ื™ื•ืœื•ื’ื™ื”,
11:06
they were new clinical indicators that humans can understand.
222
666781
3095
ื”ื™ื• ืื™ื ื“ื™ืงื˜ื•ืจื™ื ืงืœื™ื ื™ื™ื ื—ื“ืฉื™ื ืฉื‘ื ื™ ืื“ื ื™ื›ื•ืœื™ื ืœื”ื‘ื™ืŸ.
11:09
In this pathology case,
223
669876
1792
ื‘ืžืงืจื” ื”ื–ื” ืฉืœ ื”ืคืชื•ืœื•ื’ื™ื”,
11:11
the computer system actually discovered that the cells around the cancer
224
671668
4500
ื”ืžืขืจื›ืช ื”ืžืžื•ื—ืฉื‘ืช ื’ื™ืœืชื” ืœืžืขืฉื” ืฉื”ืชืื™ื ืฉืกื‘ื™ื‘ ื”ืกืจื˜ืŸ
11:16
are as important as the cancer cells themselves
225
676168
3340
ื”ื ื—ืฉื•ื‘ื™ื ื›ืžื• ืชืื™ ื”ืกืจื˜ืŸ ืขืฆืžื
11:19
in making a diagnosis.
226
679508
1752
ืœืื‘ื—ื ื”.
11:21
This is the opposite of what pathologists had been taught for decades.
227
681260
5361
ื–ื” ื”ื”ื™ืคืš ืžืžื” ืฉืคืชื•ืœื•ื’ื™ื ืœืžื“ื• ื‘ืžืฉืš ืขืฉื•ืจื™ื.
11:26
In each of those two cases, they were systems developed
228
686621
3292
ื‘ื›ืœ ืื—ื“ ืžืฉื ื™ ื”ืžืงืจื™ื ื”ืœืœื•, ืืœื• ื”ื™ื• ืžืขืจื›ื•ืช ืฉืคื•ืชื—ื•
11:29
by a combination of medical experts and machine learning experts,
229
689913
3621
ืขืœ ื™ื“ื™ ืฉื™ืœื•ื‘ ืฉืœ ืžื•ืžื—ื™ื ืจืคื•ืื™ื™ื ื•ืžื•ืžื—ื™ ืœืžื™ื“ื” ื—ื™ืฉื•ื‘ื™ืช,
11:33
but as of last year, we're now beyond that too.
230
693534
2741
ืืš ื”ื—ืœ ืžื”ืฉื ื” ืฉืขื‘ืจื”, ืื ื—ื ื• ืžืขื‘ืจ ื’ื ืœื›ืš.
11:36
This is an example of identifying cancerous areas
231
696275
3549
ื”ื ื” ื“ื•ื’ืžื” ืœื–ื™ื”ื•ื™ ืื™ื–ื•ืจื™ื ืกืจื˜ื ื™ื™ื
11:39
of human tissue under a microscope.
232
699824
2530
ื‘ืจืงืžื” ืื ื•ืฉื™ืช ืชื—ืช ืžื™ืงืจื•ืกืงื•ืค.
11:42
The system being shown here can identify those areas more accurately,
233
702354
4613
ื”ืžืขืจื›ืช ืฉืืชื ืจื•ืื™ื ื›ืืŸ ื™ื›ื•ืœื” ืœื–ื”ื•ืช ืืช ืื•ืชื ืื™ื–ื•ืจื™ื ื‘ืื•ืคื• ื™ื•ืชืจ ืžื“ื•ื™ื™ืง,
11:46
or about as accurately, as human pathologists,
234
706967
2775
ืื• ื‘ืขืจืš ืžื“ื•ื™ื™ืง ื‘ืื•ืชื” ืžื™ื“ื”, ื›ืžื• ืคืชื•ืœื•ื’ ืื ื•ืฉื™,
11:49
but was built entirely with deep learning using no medical expertise
235
709742
3392
ืืš ื ื‘ื ืชื” ืืš ื•ืจืง ืขืœ ื™ื“ื™ ืœืžื™ื“ื” ืขืžื•ืงื”, ืœืœื ืžื•ืžื—ื™ื•ืช ืจืคื•ืื™ืช
11:53
by people who have no background in the field.
236
713134
2526
ืขืœ ื™ื“ื™ ืื ืฉื™ื ืฉืื™ืŸ ืœื”ื ืจืงืข ื‘ืชื—ื•ื ื›ืœืœ.
11:56
Similarly, here, this neuron segmentation.
237
716730
2555
ื‘ืื•ืคืŸ ื“ื•ืžื”, ื›ืืŸ, ื”ืกื’ืžื ื˜ืฆื™ื” ืฉืœ ื”ื ื•ื™ืจื•ืŸ.
11:59
We can now segment neurons about as accurately as humans can,
238
719285
3668
ืื ื• ื™ื›ื•ืœื™ื ื›ืขืช ืœื—ืœืง ื ื•ื™ืจื•ื ื™ื ืœืกื’ืžื ื˜ื™ื ื›ืžืขื˜ ื‘ืื•ืคืŸ ืžื“ื•ื™ื™ืง ื›ืžื• ื‘ื ื™ ืื“ื,
12:02
but this system was developed with deep learning
239
722953
2717
ืืš ื”ืžืขืจื›ืช ื”ื–ื• ืคื•ืชื—ื” ืขืœ ื™ื“ื™ ืœืžื™ื“ื” ืขืžื•ืงื”
12:05
using people with no previous background in medicine.
240
725670
3251
ื•ืื ืฉื™ื ืœืœื ืจืงืข ื‘ืจืคื•ืื”.
12:08
So myself, as somebody with no previous background in medicine,
241
728921
3227
ืื– ืื ื™, ื›ืžื™ืฉื”ื• ืœืœื ืจืงืข ื‘ืจืคื•ืื”,
12:12
I seem to be entirely well qualified to start a new medical company,
242
732148
3727
ื ืจืื” ืฉืื ื™ ืžืกืคื™ืง ืžืฆื•ื™ื™ื“ ืœื”ืงื™ื ื—ื‘ืจื” ืจืคื•ืื™ืช ื—ื“ืฉื”.
12:15
which I did.
243
735875
2146
ื•ื–ื” ืžื” ืฉืขืฉื™ืชื™.
12:18
I was kind of terrified of doing it,
244
738021
1740
ื”ื™ื™ืชื™ ื“ื™ ืžื‘ื•ืขืช ืžื›ืš,
12:19
but the theory seemed to suggest that it ought to be possible
245
739761
2889
ืื‘ืœ ื”ืชื™ืื•ืจื™ื” ื”ืฆื™ืขื” ืฉื–ื” ืฆืจื™ืš ืœื”ื™ื•ืช ืืคืฉืจื™
12:22
to do very useful medicine using just these data analytic techniques.
246
742650
5492
ืœืขืฉื•ืช ืจืคื•ืื” ืฉื™ืžื•ืฉื™ืช ื‘ืฉื™ืžื•ืฉ ื‘ื˜ื›ื ื™ืงื•ืช ืื ืœื™ื–ืช ืžื™ื“ืข ื‘ืœื‘ื“.
12:28
And thankfully, the feedback has been fantastic,
247
748142
2480
ื•ืœืžืจื‘ื” ื”ืžื–ืœ, ื”ืžืฉื•ื‘ ื”ื™ื” ืคื ื˜ืกื ื˜ื™,
12:30
not just from the media but from the medical community,
248
750622
2356
ืœื ืจืง ืžื”ืชืงืฉื•ืจืช ืืœื ื’ื ืžื”ืงื”ื™ืœื” ื”ืจืคื•ืื™ืช,
12:32
who have been very supportive.
249
752978
2344
ืฉื”ื™ื• ืžืื•ื“ ืชื•ืžื›ื™ื.
12:35
The theory is that we can take the middle part of the medical process
250
755322
4149
ื”ืชืื•ืจื™ื” ื”ื™ื ืฉืื ื—ื ื• ื™ื›ื•ืœื™ื ืœืงื—ืช ืืช ื—ืœืง ื”ื‘ื™ื ื™ื™ื ืฉืœ ื”ืชื”ืœื™ืš ื”ืจืคื•ืื™
12:39
and turn that into data analysis as much as possible,
251
759471
2893
ื•ืœื”ืคื•ืš ืื•ืชื• ืœืื ืœื™ื–ืช ืžื™ื“ืข ื›ื›ืœ ื”ื ื™ืชืŸ,
12:42
leaving doctors to do what they're best at.
252
762364
3065
ื•ืœืืคืฉืจ ืœืจื•ืคืื™ื ืœืขืฉื•ืช ืืช ืžื” ืฉื”ื ื”ื›ื™ ื˜ื•ื‘ื™ื ื‘ื•.
12:45
I want to give you an example.
253
765429
1602
ืื ื™ ืจื•ืฆื” ืœืชืช ืœื›ื ื“ื•ื’ืžื”.
12:47
It now takes us about 15 minutes to generate a new medical diagnostic test
254
767031
4944
ื›ื™ื•ื ื–ื” ืœื•ืงื— 15 ื“ืงื•ืช ืœื™ื™ืฆืจ ืžื‘ื—ืŸ ื“ื™ืื’ื ื•ืกื˜ื™ ื—ื“ืฉ
12:51
and I'll show you that in real time now,
255
771975
1954
ื•ืื ื™ ืืจืื” ืœื›ื ื–ืืช ื‘ื–ืžืŸ ืืžืช ืขื›ืฉื™ื•,
12:53
but I've compressed it down to three minutes by cutting some pieces out.
256
773929
3487
ืื‘ืœ ื“ื—ืกืชื™ ืืช ื–ื” ืœ-3 ื“ืงื•ืช ืขืœ ื™ื“ื™ ื”ื•ืฆืืช ื›ืžื” ื—ืœืงื™ื.
12:57
Rather than showing you creating a medical diagnostic test,
257
777416
3061
ื‘ืžืงื•ื ืœื”ืจืื•ืช ืœื›ื ื™ืฆื™ืจืช ืžื‘ื—ืŸ ื“ื™ืื’ื ื•ืกื˜ื™ ืจืคื•ืื™ ื—ื“ืฉ,
13:00
I'm going to show you a diagnostic test of car images,
258
780477
3369
ืื ื™ ื”ื•ืœืš ืœื”ืจืื•ืช ืœื›ื ืžื‘ื—ืŸ ื“ื™ืื’ื ื•ืกื˜ื™ ืฉืœ ืชืžื•ื ื•ืช ืฉืœ ืžื›ื•ื ื™ื•ืช,
13:03
because that's something we can all understand.
259
783846
2222
ื‘ื’ืœืœ ืฉื–ื” ืžืฉื”ื• ืฉื›ื•ืœื ื• ื™ื›ื•ืœื™ื ืœื”ื‘ื™ืŸ.
13:06
So here we're starting with about 1.5 million car images,
260
786068
3201
ืื– ื”ื ื” ืื ื—ื ื• ืžืชื—ื™ืœื™ื ืขื ื‘ืขืจืš 1.5 ืžื™ืœื™ื•ืŸ ืชืžื•ื ื•ืช ืฉืœ ืžื›ื•ื ื™ื•ืช,
13:09
and I want to create something that can split them into the angle
261
789269
3206
ื•ืื ื™ ืจื•ืฆื” ืœื™ืฆื•ืจ ืžืฉื”ื• ืฉื™ื›ื•ืœ ืœืคืฆืœ ืื•ืชื ืœื–ื•ื™ืช
13:12
of the photo that's being taken.
262
792475
2223
ืฉื‘ื” ื”ืชืžื•ื ื” ืฆื•ืœืžื”.
13:14
So these images are entirely unlabeled, so I have to start from scratch.
263
794698
3888
ืื– ื”ืชืžื•ื ื•ืช ื”ืืœื” ื”ืŸ ืœื’ืžืจื™ ืœื ืžืชื•ื™ื™ื’ื•ืช, ืื– ืื ื™ ืฆืจื™ืš ืœื”ืชื—ื™ืœ ืžืืคืก.
13:18
With our deep learning algorithm,
264
798586
1865
ื‘ืขื–ืจืช ืืœื’ื•ืจื™ืชื ื”ืœืžื™ื“ื” ื”ืขืžื•ืงื” ืฉืœื ื•,
13:20
it can automatically identify areas of structure in these images.
265
800451
3707
ื”ื•ื ื™ื›ื•ืœ ืœื–ื”ื•ืช ืื•ื˜ื•ืžื˜ื™ืช ืื™ื–ื•ืจื™ื ืฉืœ ืžื‘ื ื” ื‘ืชืžื•ื ื•ืช ืืœื•.
13:24
So the nice thing is that the human and the computer can now work together.
266
804158
3620
ืื– ื”ื“ื‘ืจ ื”ื ื—ืžื“ ื”ื•ื ืฉื”ืื“ื ื•ื”ืžื—ืฉื‘ ื™ื›ื•ืœื™ื ืœืขื‘ื•ื“ ืขืชื” ื™ื—ื“.
13:27
So the human, as you can see here,
267
807778
2178
ืื– ื”ืื“ื, ื›ืžื• ืฉืืชื ืจื•ืื™ื ื›ืืŸ,
13:29
is telling the computer about areas of interest
268
809956
2675
ืื•ืžืจ ืœืžื—ืฉื‘ ืขืœ ืื™ื–ื•ืจื™ ืขื ื™ื™ืŸ
13:32
which it wants the computer then to try and use to improve its algorithm.
269
812631
4650
ืฉื”ื•ื ืจื•ืฆื” ืฉื”ืžื—ืฉื‘ ื™ื ืกื” ืœื”ืฉืชืžืฉ ื‘ื”ื ืœืฉื™ืคื•ืจ ื”ืืœื’ื•ืจื™ืชื.
13:37
Now, these deep learning systems actually are in 16,000-dimensional space,
270
817281
4296
ื›ืขืช, ืžืขืจื›ื•ืช ื”ืœืžื™ื“ื” ื”ืขืžื•ืงื” ื”ืœืœื• ื”ืŸ ืœืžืขืฉื” ื‘ื—ืœืœ ื‘ืขืœ 16,000 ืžื™ืžื“ื™ื,
13:41
so you can see here the computer rotating this through that space,
271
821577
3432
ืื– ืืชื ืจื•ืื™ื ื›ืืŸ ืฉื”ืžื—ืฉื‘ ืžืกื•ื‘ื‘ ื–ืืช ื“ืจืš ื”ืžืจื—ื‘ ื”ื–ื”,
13:45
trying to find new areas of structure.
272
825009
1992
ื•ืžื ืกื” ืœืžืฆื•ื ืื™ื–ื•ืจื™ื ื—ื“ืฉื™ื ืฉืœ ืžื‘ื ื”.
13:47
And when it does so successfully,
273
827001
1781
ื•ื›ืฉื”ื•ื ืžืฆืœื™ื—,
13:48
the human who is driving it can then point out the areas that are interesting.
274
828782
4004
ื”ืื“ื ืฉื ื•ื”ื’ ื‘ื• ื™ื›ื•ืœ ืœืฆื™ื™ืŸ ืืช ื”ืื™ื–ื•ืจื™ื ืฉืžืขื ื™ื™ื ื™ื.
13:52
So here, the computer has successfully found areas,
275
832786
2422
ืื– ื”ื ื”, ื”ืžื—ืฉื‘ ืžืฆื ื‘ื”ืฆืœื—ื” ืื™ื–ื•ืจื™ื,
13:55
for example, angles.
276
835208
2562
ืœื“ื•ื’ืžื”, ื–ื•ื•ื™ื•ืช.
13:57
So as we go through this process,
277
837770
1606
ืื– ื›ืฉืื ื—ื ื• ืžืชืงื“ืžื™ื ื‘ืชื”ืœื™ืš ื”ื–ื”,
13:59
we're gradually telling the computer more and more
278
839376
2340
ืื ื—ื ื• ื‘ื”ื“ืจื’ื” ืื•ืžืจื™ื ืœืžื—ืฉื‘ ืขื•ื“ ื•ืขื•ื“
14:01
about the kinds of structures we're looking for.
279
841716
2428
ืื•ื“ื•ืช ืกื•ื’ื™ ื”ืžื‘ื ื™ื ืฉืื ื—ื ื• ืžื—ืคืฉื™ื.
14:04
You can imagine in a diagnostic test
280
844144
1772
ืืชื ื™ื›ื•ืœื™ื ืœื“ืžื™ื™ืŸ ื–ืืช ื‘ืžื‘ื—ืŸ ื“ื™ืื’ื ื•ืกื˜ื™
14:05
this would be a pathologist identifying areas of pathosis, for example,
281
845916
3350
ื–ื” ื™ื”ื™ื” ืคืชื•ืœื•ื’ ืฉืžื–ื”ื” ืื™ื–ื•ืจื™ ืžื—ืœื”, ืœื“ื•ื’ืžื”,
14:09
or a radiologist indicating potentially troublesome nodules.
282
849266
5026
ืื• ืจื“ื™ื•ืœื•ื’ ืฉืžืฆื‘ื™ืข ืขืœ ืงืฉืจื™ื•ืช ืฉื™ื›ื•ืœื•ืช ืœื”ื™ื•ืช ื‘ืขื™ื™ืชื™ื•ืช.
14:14
And sometimes it can be difficult for the algorithm.
283
854292
2559
ื•ืœืคืขืžื™ื ื–ื” ื™ื›ื•ืœ ืœื”ื™ื•ืช ืงืฉื” ืœืืœื’ื•ืจื™ืชื.
14:16
In this case, it got kind of confused.
284
856851
1964
ื‘ืžืงืจื” ื”ื–ื”, ื”ื•ื ื“ื™ ื”ืชื‘ืœื‘ืœ.
14:18
The fronts and the backs of the cars are all mixed up.
285
858815
2550
ื”ื—ืœืงื™ื ื”ืงื“ืžื™ื™ื ื•ื”ืื—ื•ืจื™ื™ื ืฉืœ ื”ืžื›ื•ื ื™ื•ืช ืžืขื•ืจื‘ื‘ื™ื.
14:21
So here we have to be a bit more careful,
286
861365
2072
ืื– ื›ืืŸ ืขืœื™ื ื• ืœื”ื™ื•ืช ื˜ื™ืคื” ื™ื•ืชืจ ื–ื”ื™ืจื™ื,
14:23
manually selecting these fronts as opposed to the backs,
287
863437
3232
ื•ืœื‘ื—ื•ืจ ื‘ืื•ืคืŸ ื™ื“ื ื™ ืืช ื”ื—ืœืงื™ื ื”ืงื“ืžื™ื™ื ื‘ื ื™ื’ื•ื“ ืœืืœื• ื”ืื—ื•ืจื™ื™ื,
14:26
then telling the computer that this is a type of group
288
866669
5506
ื•ืื– ืœื”ื’ื™ื“ ืœืžื—ืฉื‘ ืฉื–ื• ืกื•ื’ ืฉืœ ืงื‘ื•ืฆื”
14:32
that we're interested in.
289
872175
1348
ืฉืื ื—ื ื• ืžืขื•ื ื™ื ื™ื ื‘ื”.
14:33
So we do that for a while, we skip over a little bit,
290
873523
2677
ืื– ืื ื—ื ื• ืขื•ืฉื™ื ื–ืืช ืœืžืฉืš ื–ืžืŸ ืžื”, ืžื“ืœื’ื™ื ืขืœ ืงืฆืช,
14:36
and then we train the machine learning algorithm
291
876200
2246
ื•ืื– ืื ื—ื ื• ืžืืžื ื™ื ืืช ืืœื’ื•ืจื™ืชื ื”ืœืžื™ื“ื” ื”ื—ื™ืฉื•ื‘ื™ืช
14:38
based on these couple of hundred things,
292
878446
1974
ื‘ื”ืชื‘ืกืก ืขืœ ืื•ืชื 200 ื“ื‘ืจื™ื,
14:40
and we hope that it's gotten a lot better.
293
880420
2025
ื•ืžืงื•ื•ื™ื ืฉื”ื•ื ื”ืฉืชืคืจ ื‘ื”ืจื‘ื”.
14:42
You can see, it's now started to fade some of these pictures out,
294
882445
3073
ืืชื ืจื•ืื™ื, ื”ื•ื ื”ืชื—ื™ืœ ืœื”ื“ื”ื•ืช ื›ืžื” ืžื”ืชืžื•ื ื•ืช ื”ืืœื•,
14:45
showing us that it already is recognizing how to understand some of these itself.
295
885518
4708
ื•ืžืจืื” ืœื ื• ืฉื”ื•ื ื›ื‘ืจ ืžื–ื”ื” ืื™ืš ืœื”ื‘ื™ืŸ ื›ืžื” ืžืืœื” ื‘ืขืฆืžื•.
14:50
We can then use this concept of similar images,
296
890226
2902
ืื– ืื ื—ื ื• ื™ื›ื•ืœื™ื ืœื”ืฉืชืžืฉ ื‘ืงื•ื ืกืคื˜ ื”ื–ื” ืฉืœ ืชืžื•ื ื•ืช ื“ื•ืžื•ืช,
14:53
and using similar images, you can now see,
297
893128
2094
ื•ืขืœ ื™ื“ื™ ืฉื™ืžื•ืฉ ื‘ืชืžื•ื ื•ืช ื“ื•ืžื•ืช, ืืชื ื™ื›ื•ืœื™ื ืขื›ืฉื™ื• ืœืจืื•ืช,
14:55
the computer at this point is able to entirely find just the fronts of cars.
298
895222
4019
ืฉื”ืžื—ืฉื‘ ื‘ื ืงื•ื“ื” ื–ื• ืžืกื•ื’ืœ ืœืžืฆื•ื ืจืง ืืช ื”ื—ื–ื™ืช ืฉืœ ืžื›ื•ื ื™ื•ืช.
14:59
So at this point, the human can tell the computer,
299
899241
2948
ืื– ื‘ื ืงื•ื“ื” ื–ื•, ื”ืื“ื ื™ื›ื•ืœ ืœื•ืžืจ ืœืžื—ืฉื‘,
15:02
okay, yes, you've done a good job of that.
300
902189
2293
ืื•ืงื™ื™, ื›ืŸ, ืขืฉื™ืช ืขื‘ื•ื“ื” ื˜ื•ื‘ื”.
15:05
Sometimes, of course, even at this point
301
905652
2185
ืœืคืขืžื™ื, ื›ืžื•ื‘ืŸ, ืืคื™ืœื• ื‘ื ืงื•ื“ื” ื–ื•
15:07
it's still difficult to separate out groups.
302
907837
3674
ื–ื” ืขื“ื™ื™ืŸ ืงืฉื” ืœื”ืคืจื™ื“ ื‘ื™ืŸ ืงื‘ื•ืฆื•ืช.
15:11
In this case, even after we let the computer try to rotate this for a while,
303
911511
3884
ื‘ืžืงืจื” ื”ื–ื”, ืืคื™ืœื• ืื—ืจื™ ืฉืื ื—ื ื• ื ื•ืชื ื™ื ืœืžื—ืฉื‘ ืœื ืกื•ืช ืœืกื•ื‘ื‘ ื–ืืช ืœืžืฉืš ื–ืžืŸ ืžื”,
15:15
we still find that the left sides and the right sides pictures
304
915399
3345
ืื ื—ื ื• ืขื“ื™ื™ืŸ ืžื•ืฆืื™ื ืฉื”ืชืžื•ื ื•ืช ืฉืœ ื”ืฆื“ื“ื™ื ื”ืฉืžืืœื™ื™ื™ื ื•ื”ื™ืžื ื™ื™ื
15:18
are all mixed up together.
305
918744
1478
ืžืขื•ืจื‘ื‘ื•ืช ื–ื• ื‘ื–ื•.
15:20
So we can again give the computer some hints,
306
920222
2140
ืื– ืื ื—ื ื• ืฉื•ื‘ ื™ื›ื•ืœื™ื ืœืชืช ืœืžื—ืฉื‘ ื›ืžื” ืจืžื–ื™ื,
15:22
and we say, okay, try and find a projection that separates out
307
922362
2976
ื•ืœื”ื’ื™ื“, ืื•ืงื™ื™, ื ืกื” ืœืžืฆื•ื ืชื›ื ื™ืช ืฉืžืคืจื™ื“ื”
15:25
the left sides and the right sides as much as possible
308
925338
2607
ืืช ืฆื“ ื™ืžื™ืŸ ืžืฆื“ ืฉืžืืœ ืขื“ ื›ืžื” ืฉื ื™ืชืŸ
15:27
using this deep learning algorithm.
309
927945
2122
ืขืœ ื™ื“ื™ ืฉื™ืžื•ืฉ ื‘ืืœื’ื•ืจื™ืชื ื”ืœืžื™ื“ื” ื”ืขืžื•ืงื” ื”ื–ื”.
15:30
And giving it that hint -- ah, okay, it's been successful.
310
930067
2942
ื•ืื—ืจื™ ืฉื ืชื ื• ืœื• ืืช ื”ืจืžื– ื”ื–ื”-- ืื”, ืื•ืงื™ื™, ื”ื•ื ื”ืฆืœื™ื—.
15:33
It's managed to find a way of thinking about these objects
311
933009
2882
ื”ื•ื ื”ืฆืœื™ื— ืœืžืฆื•ื ื“ืจืš ืœื—ืฉื•ื‘ ืขืœ ื”ืื•ื‘ื™ื™ืงื˜ื™ื ื”ืืœื”
15:35
that's separated out these together.
312
935891
2380
ืฉื”ืคืจื™ื“ื• ืืช ื–ื” ื™ื—ื“.
15:38
So you get the idea here.
313
938271
2438
ืื– ืืชื ืžื‘ื™ื ื™ื ืืช ื”ืจืขื™ื•ืŸ ืคื”.
15:40
This is a case not where the human is being replaced by a computer,
314
940709
8197
ื–ื” ืœื ืžืงืจื” ืฉื‘ื• ื”ืื“ื ืžื•ื—ืœืฃ ืขืœ ื™ื“ื™ ื”ืžื—ืฉื‘,
15:48
but where they're working together.
315
948906
2640
ืืœื ืฉื”ื ืขื•ื‘ื“ื™ื ื™ื—ื“.
15:51
What we're doing here is we're replacing something that used to take a team
316
951546
3550
ืžื” ืฉืื ื—ื ื• ืขื•ืฉื™ื ื›ืืŸ ื”ื•ื ืœื”ื—ืœื™ืฃ ืžืฉื”ื• ืฉื”ื™ื” ืœื•ืงื— ืœืฆื•ื•ืช
15:55
of five or six people about seven years
317
955096
2002
ืฉืœ ื—ืžื™ืฉื” ืื• ืฉื™ืฉื” ืื ืฉื™ื ื‘ืขืจืš ืฉื‘ืข ืฉื ื™ื
15:57
and replacing it with something that takes 15 minutes
318
957098
2605
ื•ืžื—ืœื™ืคื™ื ืื•ืชื• ื‘ืžืฉื”ื• ืฉืœื•ืงื— 15 ื“ืงื•ืช
15:59
for one person acting alone.
319
959703
2505
ืœืื“ื ืื—ื“ ืฉืคื•ืขืœ ืœื‘ื“.
16:02
So this process takes about four or five iterations.
320
962208
3950
ืื– ื”ืชื”ืœื™ืš ื”ื–ื” ืœื•ืงื— ื‘ืขืจืš ืืจื‘ืข ืื• ื—ืžืฉ ื—ื–ืจื•ืช.
16:06
You can see we now have 62 percent
321
966158
1859
ืืชื ืจื•ืื™ื ืขื›ืฉื™ื• ืฉื™ืฉ ืœื ื• 62 ืื—ื•ื–
16:08
of our 1.5 million images classified correctly.
322
968017
2959
ืž-1.5 ืžื™ืœื™ื•ืŸ ื”ืชืžื•ื ื•ืช ืฉืœื ื• ืžืกื•ื•ื’ื•ืช ื ื›ื•ื ื”.
16:10
And at this point, we can start to quite quickly
323
970976
2472
ื‘ื ืงื•ื“ื” ื–ื•, ืื ื—ื ื• ื™ื›ื•ืœื™ื ืœื”ืชื—ื™ืœ ื“ื™ ื‘ืžื”ื™ืจื•ืช
16:13
grab whole big sections,
324
973448
1297
ืœืงื—ืช ื—ืœืงื ื’ื“ื•ืœื™ื ื‘ืžืœื•ืื,
16:14
check through them to make sure that there's no mistakes.
325
974745
2919
ืœืขื‘ื•ืจ ืขืœื™ื”ื ื•ืœื•ื•ื“ื ืฉืื™ืŸ ื˜ืขื•ื™ื•ืช.
16:17
Where there are mistakes, we can let the computer know about them.
326
977664
3952
ื”ื™ื›ืŸ ืฉื™ืฉื ืŸ ื˜ืขื•ื™ื•ืช, ืื ื• ื™ื›ื•ืœื™ื ืœื™ื™ื“ืข ืืช ื”ืžื—ืฉื‘ ืขืœื™ื”ืŸ.
16:21
And using this kind of process for each of the different groups,
327
981616
3045
ื•ืขืœ ื™ื“ื™ ืฉื™ืžื•ืฉ ื‘ืกื•ื’ ื–ื” ืฉืœ ืชื”ืœื™ืš ืขื‘ื•ืจ ื›ืœ ืื—ืช ืžื”ืงื‘ื•ืฆื•ืช ื”ืฉื•ื ื•ืช,
16:24
we are now up to an 80 percent success rate
328
984661
2487
ืื ื—ื ื• ืขื›ืฉื™ื• ืžื’ื™ืขื™ื ืœ-80 ืื—ื•ื–ื™ ื”ืฆืœื—ื”
16:27
in classifying the 1.5 million images.
329
987148
2415
ื‘ืกื™ื•ื•ื’ ืžื™ืœื™ื•ืŸ ื•ื—ืฆื™ ื”ืชืžื•ื ื•ืช.
16:29
And at this point, it's just a case
330
989563
2078
ื•ื‘ื ืงื•ื“ื” ื–ื•, ื–ื”ื• ืจืง ืžืงืจื”
16:31
of finding the small number that aren't classified correctly,
331
991641
3579
ืฉืœ ืžืฆื™ืืช ื”ืžืกืคืจ ื”ืงื˜ืŸ ืฉืœื ืžืกื•ื•ื’ ื›ืจืื•ื™,
16:35
and trying to understand why.
332
995220
2888
ื•ืœื ืกื•ืช ืœื”ื‘ื™ืŸ ืœืžื”.
16:38
And using that approach,
333
998108
1743
ืขืœ ื™ื“ื™ ืฉื™ืžื•ืฉ ื‘ื’ื™ืฉื” ื–ื•,
16:39
by 15 minutes we get to 97 percent classification rates.
334
999851
4121
ืชื•ืš 15 ื“ืงื•ืช ืื ื• ืžื’ื™ืขื™ื ืœ-97 ืื—ื•ื–ื™ ืกื™ื•ื•ื’.
16:43
So this kind of technique could allow us to fix a major problem,
335
1003972
4600
ืื– ืกื•ื’ ื›ื–ื” ืฉืœ ื˜ื›ื ื™ืงื” ื™ื›ื•ืœ ืœืืคืฉืจ ืœื ื• ืœืชืงืŸ ื‘ืขื™ื” ื’ื“ื•ืœื”,
16:48
which is that there's a lack of medical expertise in the world.
336
1008578
3036
ืฉื”ื™ื ืฉื™ืฉ ื—ื•ืกืจ ื‘ืžื•ืžื—ื™ื•ืช ืจืคื•ืื™ืช ื‘ืขื•ืœื.
16:51
The World Economic Forum says that there's between a 10x and a 20x
337
1011614
3489
ืคื•ืจื•ื ื”ื›ืœื›ืœื” ื”ืขื•ืœืžื™ ืื•ืžืจ ืฉื™ืฉ ื‘ื™ืŸ 10X ื•-20X
16:55
shortage of physicians in the developing world,
338
1015103
2624
ืฉืœ ืžื—ืกื•ืจ ื‘ืจื•ืคืื™ื ื‘ืขื•ืœื ื”ืžืชืคืชื—,
16:57
and it would take about 300 years
339
1017727
2113
ื•ืฉื–ื” ื™ื™ืงื— ื‘ืขืจืš 300 ืฉื ื”
16:59
to train enough people to fix that problem.
340
1019840
2894
ืœืืžืŸ ืžืกืคื™ืง ืื ืฉื™ื ืœืชืงืŸ ืืช ื”ื‘ืขื™ื” ื”ื–ื•.
17:02
So imagine if we can help enhance their efficiency
341
1022734
2885
ืื– ื“ืžื™ื™ื ื• ืื ื”ื™ื™ื ื• ื™ื›ื•ืœื™ื ืœืขื–ื•ืจ ืœื”ื’ื‘ื™ืจ ืืช ื”ื™ืขื™ืœื•ืช ืฉืœื”ื
17:05
using these deep learning approaches?
342
1025619
2839
ืขืœ ื™ื“ื™ ืฉื™ืžื•ืฉ ื‘ื’ื™ืฉืช ื”ืœืžื™ื“ื” ื”ืขืžื•ืงื” ื”ื–ื•?
17:08
So I'm very excited about the opportunities.
343
1028458
2232
ืื– ืื ื™ ืžืื•ื“ ืžืชืจื’ืฉ ืžื”ืืคืฉืจื•ื™ื•ืช.
17:10
I'm also concerned about the problems.
344
1030690
2589
ืื ื™ ื’ื ืžืื•ื“ ืžื•ื“ืื’ ืžื”ื‘ืขื™ื•ืช.
17:13
The problem here is that every area in blue on this map
345
1033279
3124
ื”ื‘ืขื™ื” ื›ืืŸ ื”ื™ื ืฉื›ืœ ืื™ื–ื•ืจ ื‘ื›ื—ื•ืœ ืขืœ ื”ืžืคื”
17:16
is somewhere where services are over 80 percent of employment.
346
1036403
3769
ื”ื•ื ืžืงื•ื ืฉื‘ื• ืฉื™ืจื•ืชื™ื ื”ื ืžืขืœ ืœ-80 ืื—ื•ื– ืžื”ืชืขืกื•ืงื”.
17:20
What are services?
347
1040172
1787
ืžื” ื–ื” ืฉื™ืจื•ืชื™ื?
17:21
These are services.
348
1041959
1514
ืืœื• ืฉื™ืจื•ืชื™ื.
17:23
These are also the exact things that computers have just learned how to do.
349
1043473
4154
ืืœื• ื”ื ื’ื ื‘ื“ื™ื•ืง ื”ื“ื‘ืจื™ื ืฉืžื—ืฉื‘ื™ื ื‘ื“ื™ื•ืง ืœืžื“ื• ืื™ืš ืœืขืฉื•ืช.
17:27
So 80 percent of the world's employment in the developed world
350
1047627
3804
ืื– 80 ืื—ื•ื– ืžื”ืชืขืกื•ืงื” ื”ืขื•ืœืžื™ืช ื‘ืขื•ืœื ื”ืžืคื•ืชื—
17:31
is stuff that computers have just learned how to do.
351
1051431
2532
ื”ื ื“ื‘ืจื™ื ืฉืžื—ืฉื‘ื™ื ื‘ื“ื™ื•ืง ืœืžื“ื• ืื™ืš ืœืขืฉื•ืช.
17:33
What does that mean?
352
1053963
1440
ืžื” ื–ื” ืื•ืžืจ?
17:35
Well, it'll be fine. They'll be replaced by other jobs.
353
1055403
2583
ื•ื‘ื›ืŸ, ื–ื” ื™ื”ื™ื” ื‘ืกื“ืจ. ื”ืŸ ื™ื•ื—ืœืคื• ื‘ืขื‘ื•ื“ื•ืช ืื—ืจื•ืช.
17:37
For example, there will be more jobs for data scientists.
354
1057986
2707
ืœืžืฉืœ, ื™ื”ื™ื• ื™ื•ืชืจ ืขื‘ื•ื“ื•ืช ืขื‘ื•ืจ ืžื“ืขื ื™ ื ืชื•ื ื™ื.
17:40
Well, not really.
355
1060693
817
ื•ื‘ื›ืŸ, ืœื ื‘ืืžืช.
17:41
It doesn't take data scientists very long to build these things.
356
1061510
3118
ืœื ืœื•ืงื— ืœืžื“ืขื ื™ื ื–ืžืŸ ืจื‘ ืœื‘ื ื•ืช ืืช ื”ื“ื‘ืจื™ื ื”ืœืœื•.
17:44
For example, these four algorithms were all built by the same guy.
357
1064628
3252
ืœืžืฉืœ, ืืจื‘ืขืช ื”ืืœื’ื•ืจื™ืชืžื™ื ื”ืืœื” ื ื‘ื ื• ื›ื•ืœื ืขืœ ื™ื“ื™ ืื•ืชื• ืื“ื.
17:47
So if you think, oh, it's all happened before,
358
1067880
2438
ืื– ืื ืืชื ื—ื•ืฉื‘ื™ื, ื”ื•, ื›ืœ ื–ื” ื›ื‘ืจ ืงืจื” ื‘ืขื‘ืจ,
17:50
we've seen the results in the past of when new things come along
359
1070318
3808
ืจืื™ื ื• ืืช ื”ืชื•ืฆืื•ืช ื‘ืขื‘ืจ ื›ืืฉืจ ื“ื‘ืจื™ื ื—ื“ืฉื™ื ื”ื’ื™ืขื•
17:54
and they get replaced by new jobs,
360
1074126
2252
ื•ื”ื ืžื•ื—ืœืคื™ื ื‘ืขื‘ื•ื“ื•ืช ื—ื“ืฉื•ืช,
17:56
what are these new jobs going to be?
361
1076378
2116
ืื‘ืœ ืžื” ื”ืขื‘ื•ื“ื•ืช ื”ื—ื“ืฉื•ืช ื”ืœืœื• ื”ื•ืœื›ื•ืช ืœื”ื™ื•ืช?
17:58
It's very hard for us to estimate this,
362
1078494
1871
ื–ื” ืžืื•ื“ ืงืฉื” ื‘ืฉื‘ื™ืœื ื• ืœื”ืขืจื™ืš ื–ืืช,
18:00
because human performance grows at this gradual rate,
363
1080365
2739
ื‘ื’ืœืœ ืฉื”ื‘ื™ืฆื•ืขื™ื ื”ืื ื•ืฉื™ื™ื ื’ื“ืœื™ื ื‘ืงืฆื‘ ื”ื”ื“ืจื’ืชื™ ื”ื–ื”,
18:03
but we now have a system, deep learning,
364
1083104
2562
ืื‘ืœ ื™ืฉ ืœื ื• ืขื›ืฉื™ื• ืžืขืจื›ืช, ืœืžื™ื“ื” ืขืžื•ืงื”,
18:05
that we know actually grows in capability exponentially.
365
1085666
3227
ืฉืื ื—ื ื• ื™ื•ื“ืขื™ื ืฉืžืžืฉ ื’ื“ืœื” ื‘ื™ื›ื•ืœื•ืช ืฉืœื” ื‘ืื•ืคืŸ ืืงืกืคื•ื ื ืฆื™ืืœื™.
18:08
And we're here.
366
1088893
1605
ื•ืื ื—ื ื• ื›ืืŸ.
18:10
So currently, we see the things around us
367
1090498
2061
ืื– ื›ืจื’ืข, ืื ื—ื ื• ืจื•ืื™ื ืืช ื”ื“ื‘ืจื™ื ืฉืกื‘ื™ื‘ื ื•
18:12
and we say, "Oh, computers are still pretty dumb." Right?
368
1092559
2676
ื•ืื•ืžืจื™ื, "ื”ื•, ืžื—ืฉื‘ื™ื ื”ื ืขื“ื™ื™ืŸ ื“ื™ ื˜ื™ืคืฉื™ื." ื ื›ื•ืŸ?
18:15
But in five years' time, computers will be off this chart.
369
1095235
3429
ืื‘ืœ ืชื•ืš ื—ืžืฉ ืฉื ื™ื, ืžื—ืฉื‘ื™ื ื™ื”ื™ื• ืžื—ื•ืฅ ืœื˜ื‘ืœื” ื”ื–ื•.
18:18
So we need to be starting to think about this capability right now.
370
1098664
3865
ืื– ืื ื—ื ื• ืฆืจื™ื›ื™ื ืœื”ืชื—ื™ืœ ืœื—ืฉื•ื‘ ืขืœ ื”ื™ื›ื•ืœืช ื”ื–ื• ื›ื‘ืจ ืขื›ืฉื™ื•.
18:22
We have seen this once before, of course.
371
1102529
2050
ืจืื™ื ื• ืืช ื–ื” ืงื•ืจื” ื‘ืขื‘ืจ, ื›ืžื•ื‘ืŸ.
18:24
In the Industrial Revolution,
372
1104579
1387
ื‘ืžื”ืคื›ื” ื”ืชืขืฉื™ื™ืชื™ืช,
18:25
we saw a step change in capability thanks to engines.
373
1105966
2851
ืจืื™ื ื• ืฉื™ื ื•ื™ ืžืฉืžืขื•ืชื™ ื‘ื™ื›ื•ืœื•ืช ื”ื•ื“ื•ืช ืœืžื ื•ืขื™ื.
18:29
The thing is, though, that after a while, things flattened out.
374
1109667
3138
ื”ืขื ื™ื™ืŸ ื”ื•ื ืฉืื—ืจื™ ื–ืžืŸ ืžื”, ื”ืขื ื™ื™ื ื™ื ื”ืฉืชื˜ื—ื•.
18:32
There was social disruption,
375
1112805
1702
ื”ื™ื™ืชื” ื”ืคืจืขื” ื—ื‘ืจืชื™ืช,
18:34
but once engines were used to generate power in all the situations,
376
1114507
3439
ืื‘ืœ ื‘ืจื’ืข ืฉื”ืฉืชืžืฉื• ื‘ืžื ื•ืขื™ื ืขืœ ืžื ืช ืœื™ื™ืฆืจ ื—ืฉืžืœ ื‘ื›ืœ ื”ืžืฆื‘ื™ื,
18:37
things really settled down.
377
1117946
2354
ื“ื‘ืจื™ื ื”ืชื—ื™ืœื• ืœื”ื™ืจื’ืข.
18:40
The Machine Learning Revolution
378
1120300
1473
ืžื”ืคื™ื›ืช ื”ืœืžื™ื“ื” ื”ื—ื™ืฉื•ื‘ื™ืช
18:41
is going to be very different from the Industrial Revolution,
379
1121773
2909
ื”ื•ืœื›ืช ืœื”ื™ื•ืช ืฉื•ื ื” ืžืื•ื“ ืžื”ืžื”ืคื™ื›ื” ื”ืชืขืฉื™ื™ืชื™ืช,
18:44
because the Machine Learning Revolution, it never settles down.
380
1124682
2950
ื›ื™ื•ื•ืŸ ืฉืžื”ืคื™ื›ืช ื”ืœืžื™ื“ื” ื”ื—ื™ืฉื•ื‘ื™ืช, ื”ื™ื ืœืขื•ืœื ืœื ื ืจื’ืขืช.
18:47
The better computers get at intellectual activities,
381
1127632
2982
ื›ื›ืœ ืฉืžื—ืฉื‘ื™ื ื™ื”ื™ื• ื™ื•ืชืจ ื˜ื•ื‘ื™ื ื‘ืคืขื™ืœื•ื™ืช ืื™ื ื˜ืœืงื˜ื•ืืœื™ื•ืช,
18:50
the more they can build better computers to be better at intellectual capabilities,
382
1130614
4248
ื›ืš ื”ื ื™ื›ื•ืœื™ื ืœื‘ื ื•ืช ืžื—ืฉื‘ื™ื ื˜ื•ื‘ื™ื ื™ื•ืชืจ ืฉื™ื”ื™ื• ื˜ื•ื‘ื™ื ื™ื•ืชืจ ื‘ื™ื›ื•ืœื•ืช ืื™ื ื˜ืœืงื˜ื•ืืœื™ื•ืช,
18:54
so this is going to be a kind of change
383
1134862
1908
ืื– ื–ื” ื”ื•ืœืš ืœื”ื™ื•ืช ืกื•ื’ ืฉืœ ืฉื™ื ื•ื™
18:56
that the world has actually never experienced before,
384
1136770
2478
ืฉื”ืขื•ืœื ืœืžืขืฉื” ืœื ื—ื•ื•ื” ืžืขื•ืœื,
18:59
so your previous understanding of what's possible is different.
385
1139248
3306
ืื– ื”ื”ื‘ื ื” ื”ืงื•ื“ืžืช ืฉืœื›ื ืฉืœ ืžื” ืืคืฉืจื™, ื”ื™ื ืฉื•ื ื”.
19:02
This is already impacting us.
386
1142974
1780
ื–ื” ื›ื‘ืจ ืžื›ื” ื‘ื ื•.
19:04
In the last 25 years, as capital productivity has increased,
387
1144754
3630
ื‘-25 ื”ืฉื ื™ื ื”ืื—ืจื•ื ื•ืช, ื›ืฉืคืจื™ื•ืŸ ื”ื”ื•ืŸ ื”ื•ืœืš ื•ื’ื“ืœ,
19:08
labor productivity has been flat, in fact even a little bit down.
388
1148400
4188
ืคืจื™ื•ืŸ ื”ืขื‘ื•ื“ื” ื ืฉืืจ ืฉื˜ื•ื—, ืœืžืขืฉื” ืืคื™ืœื• ืžืขื˜ ื™ืจื“.
19:13
So I want us to start having this discussion now.
389
1153408
2741
ืื– ืื ื™ ืจื•ืฆื” ืฉืื ื—ื ื• ื ืชื—ื™ืœ ืœื ื”ืœ ืืช ื”ื“ื™ื•ืŸ ื”ื–ื” ื›ืขืช.
19:16
I know that when I often tell people about this situation,
390
1156149
3027
ืื ื™ ื™ื•ื“ืข ืฉื”ืจื‘ื” ืคืขืžื™ื ื›ืฉืื ื™ ืžืกืคืจ ืœืื ืฉื™ื ืขืœ ื”ืžืฆื‘ ื”ื–ื”,
19:19
people can be quite dismissive.
391
1159176
1490
ืื ืฉื™ื ื™ื›ื•ืœื™ื ืœื‘ื˜ืœ ืื•ืชื™.
19:20
Well, computers can't really think,
392
1160666
1673
ื•ื‘ื›ืŸ, ืžื—ืฉื‘ื™ื ืœื ื™ื›ื•ืœื™ื ืžืžืฉ ืœื—ืฉื•ื‘,
19:22
they don't emote, they don't understand poetry,
393
1162339
3028
ื”ื ืœื ืžื‘ื™ืขื™ื ืจื’ืฉื•ืช, ื”ื ืœื ืžื‘ื™ื ื™ื ืฉื™ืจื”,
19:25
we don't really understand how they work.
394
1165367
2521
ืื ื—ื ื• ืœื ืžืžืฉ ืžื‘ื™ื ื™ื ืื™ืš ื”ื ืขื•ื‘ื“ื™ื.
19:27
So what?
395
1167888
1486
ืื– ืžื”?
19:29
Computers right now can do the things
396
1169374
1804
ืžื—ืฉื‘ื™ื ื›ืจื’ืข ื™ื›ื•ืœื™ื ืœืขืฉื•ืช ืืช ื”ื“ื‘ืจื™ื
19:31
that humans spend most of their time being paid to do,
397
1171178
2719
ืฉื‘ื ื™ ืื“ื ืžืงื‘ืœื™ื ืชืฉืœื•ื ืœืขืฉื•ืช ื‘ืžืฉืš ืจื•ื‘ ื—ื™ื™ื”ื,
19:33
so now's the time to start thinking
398
1173897
1731
ืื– ืขื›ืฉื™ื• ื”ื–ืžืŸ ืœื”ืชื—ื™ืœ ืœื—ืฉื•ื‘
19:35
about how we're going to adjust our social structures and economic structures
399
1175628
4387
ืขืœ ืื™ืš ืื ื—ื ื• ื”ื•ืœื›ื™ื ืœื”ืชืื™ื ืืช ื”ืžื‘ื ื™ื ื”ื—ื‘ืจืชื™ื™ื ื•ื”ื›ืœื›ืœื™ื™ื ืฉืœื ื•
19:40
to be aware of this new reality.
400
1180015
1840
ื›ืš ืฉื™ื”ื™ื• ืžื•ื“ืขื™ื ืœืžืฆื™ืื•ืช ื”ื—ื“ืฉื” ื”ื–ื•.
19:41
Thank you.
401
1181855
1533
ืชื•ื“ื”.
19:43
(Applause)
402
1183388
802
(ื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