TEDxBrussels
Jeremy Howard: The wonderful and terrifying implications of computers that can learn
Filmed:
Readability: 4.5
2,532,971 views
What happens when we teach a computer how to learn? Technologist Jeremy Howard shares some surprising new developments in the fast-moving field of deep learning, a technique that can give computers the ability to learn Chinese, or to recognize objects in photos, or to help think through a medical diagnosis. (One deep learning tool, after watching hours of YouTube, taught itself the concept of "cats.") Get caught up on a field that will change the way the computers around you behave ... sooner than you probably think.
Jeremy Howard - Data scientist
Jeremy Howard imagines how advanced machine learning can improve our lives. Full bio
Jeremy Howard imagines how advanced machine learning can improve our lives. Full bio
Double-click the English transcript below to play the video.
00:12
It used to be that if you wanted
to get a computer to do something new,
to get a computer to do something new,
0
880
4013
00:16
you would have to program it.
1
4893
1554
00:18
Now, programming, for those of you here
that haven't done it yourself,
that haven't done it yourself,
2
6447
3411
00:21
requires laying out in excruciating detail
3
9858
3502
00:25
every single step that you want
the computer to do
the computer to do
4
13360
3367
00:28
in order to achieve your goal.
5
16727
2362
00:31
Now, if you want to do something
that you don't know how to do yourself,
that you don't know how to do yourself,
6
19089
3496
00:34
then this is going
to be a great challenge.
to be a great challenge.
7
22585
2063
00:36
So this was the challenge faced
by this man, Arthur Samuel.
by this man, Arthur Samuel.
8
24648
3483
00:40
In 1956, he wanted to get this computer
9
28131
4077
00:44
to be able to beat him at checkers.
10
32208
2340
00:46
How can you write a program,
11
34548
2040
00:48
lay out in excruciating detail,
how to be better than you at checkers?
how to be better than you at checkers?
12
36588
3806
00:52
So he came up with an idea:
13
40394
1722
00:54
he had the computer play
against itself thousands of times
against itself thousands of times
14
42116
3724
00:57
and learn how to play checkers.
15
45840
2524
01:00
And indeed it worked,
and in fact, by 1962,
and in fact, by 1962,
16
48364
3180
01:03
this computer had beaten
the Connecticut state champion.
the Connecticut state champion.
17
51544
4017
01:07
So Arthur Samuel was
the father of machine learning,
the father of machine learning,
18
55561
2973
01:10
and I have a great debt to him,
19
58534
1717
01:12
because I am a machine
learning practitioner.
learning practitioner.
20
60251
2763
01:15
I was the president of Kaggle,
21
63014
1465
01:16
a community of over 200,000
machine learning practictioners.
machine learning practictioners.
22
64479
3388
01:19
Kaggle puts up competitions
23
67867
2058
01:21
to try and get them to solve
previously unsolved problems,
previously unsolved problems,
24
69925
3708
01:25
and it's been successful
hundreds of times.
hundreds of times.
25
73633
3837
01:29
So from this vantage point,
I was able to find out
I was able to find out
26
77470
2470
01:31
a lot about what machine learning
can do in the past, can do today,
can do in the past, can do today,
27
79940
3950
01:35
and what it could do in the future.
28
83890
2362
01:38
Perhaps the first big success of
machine learning commercially was Google.
machine learning commercially was Google.
29
86252
4423
01:42
Google showed that it is
possible to find information
possible to find information
30
90675
3109
01:45
by using a computer algorithm,
31
93784
1752
01:47
and this algorithm is based
on machine learning.
on machine learning.
32
95536
2901
01:50
Since that time, there have been many
commercial successes of machine learning.
commercial successes of machine learning.
33
98437
3886
01:54
Companies like Amazon and Netflix
34
102323
1837
01:56
use machine learning to suggest
products that you might like to buy,
products that you might like to buy,
35
104160
3716
01:59
movies that you might like to watch.
36
107876
2020
02:01
Sometimes, it's almost creepy.
37
109896
1807
02:03
Companies like LinkedIn and Facebook
38
111703
1954
02:05
sometimes will tell you about
who your friends might be
who your friends might be
39
113657
2594
02:08
and you have no idea how it did it,
40
116251
1977
02:10
and this is because it's using
the power of machine learning.
the power of machine learning.
41
118228
2967
02:13
These are algorithms that have
learned how to do this from data
learned how to do this from data
42
121195
2957
02:16
rather than being programmed by hand.
43
124152
3247
02:19
This is also how IBM was successful
44
127399
2478
02:21
in getting Watson to beat
the two world champions at "Jeopardy,"
the two world champions at "Jeopardy,"
45
129877
3862
02:25
answering incredibly subtle
and complex questions like this one.
and complex questions like this one.
46
133739
3225
02:28
["The ancient 'Lion of Nimrud' went missing
from this city's national museum in 2003
(along with a lot of other stuff)"]
from this city's national museum in 2003
(along with a lot of other stuff)"]
47
136964
2835
02:31
This is also why we are now able
to see the first self-driving cars.
to see the first self-driving cars.
48
139799
3235
02:35
If you want to be able to tell
the difference between, say,
the difference between, say,
49
143034
2822
02:37
a tree and a pedestrian,
well, that's pretty important.
well, that's pretty important.
50
145856
2632
02:40
We don't know how to write
those programs by hand,
those programs by hand,
51
148488
2587
02:43
but with machine learning,
this is now possible.
this is now possible.
52
151075
2997
02:46
And in fact, this car has driven
over a million miles
over a million miles
53
154072
2608
02:48
without any accidents on regular roads.
54
156680
3506
02:52
So we now know that computers can learn,
55
160196
3914
02:56
and computers can learn to do things
56
164110
1900
02:58
that we actually sometimes
don't know how to do ourselves,
don't know how to do ourselves,
57
166010
2838
03:00
or maybe can do them better than us.
58
168848
2885
03:03
One of the most amazing examples
I've seen of machine learning
I've seen of machine learning
59
171733
4195
03:07
happened on a project that I ran at Kaggle
60
175928
2392
03:10
where a team run by a guy
called Geoffrey Hinton
called Geoffrey Hinton
61
178320
3591
03:13
from the University of Toronto
62
181911
1552
03:15
won a competition for
automatic drug discovery.
automatic drug discovery.
63
183463
2677
03:18
Now, what was extraordinary here
is not just that they beat
is not just that they beat
64
186140
2847
03:20
all of the algorithms developed by Merck
or the international academic community,
or the international academic community,
65
188987
4013
03:25
but nobody on the team had any background
in chemistry or biology or life sciences,
in chemistry or biology or life sciences,
66
193000
5061
03:30
and they did it in two weeks.
67
198061
2169
03:32
How did they do this?
68
200230
1381
03:34
They used an extraordinary algorithm
called deep learning.
called deep learning.
69
202421
2921
03:37
So important was this that in fact
the success was covered
the success was covered
70
205342
2949
03:40
in The New York Times in a front page
article a few weeks later.
article a few weeks later.
71
208291
3121
03:43
This is Geoffrey Hinton
here on the left-hand side.
here on the left-hand side.
72
211412
2735
03:46
Deep learning is an algorithm
inspired by how the human brain works,
inspired by how the human brain works,
73
214147
4341
03:50
and as a result it's an algorithm
74
218488
1812
03:52
which has no theoretical limitations
on what it can do.
on what it can do.
75
220300
3841
03:56
The more data you give it and the more
computation time you give it,
computation time you give it,
76
224141
2823
03:58
the better it gets.
77
226964
1312
04:00
The New York Times also
showed in this article
showed in this article
78
228276
2339
04:02
another extraordinary
result of deep learning
result of deep learning
79
230615
2242
04:04
which I'm going to show you now.
80
232857
2712
04:07
It shows that computers
can listen and understand.
can listen and understand.
81
235569
4941
04:12
(Video) Richard Rashid: Now, the last step
82
240510
2711
04:15
that I want to be able
to take in this process
to take in this process
83
243221
3025
04:18
is to actually speak to you in Chinese.
84
246246
4715
04:22
Now the key thing there is,
85
250961
2635
04:25
we've been able to take a large amount
of information from many Chinese speakers
of information from many Chinese speakers
86
253596
5002
04:30
and produce a text-to-speech system
87
258598
2530
04:33
that takes Chinese text
and converts it into Chinese language,
and converts it into Chinese language,
88
261128
4673
04:37
and then we've taken
an hour or so of my own voice
an hour or so of my own voice
89
265801
4128
04:41
and we've used that to modulate
90
269929
1891
04:43
the standard text-to-speech system
so that it would sound like me.
so that it would sound like me.
91
271820
4544
04:48
Again, the result's not perfect.
92
276364
2540
04:50
There are in fact quite a few errors.
93
278904
2648
04:53
(In Chinese)
94
281552
2484
04:56
(Applause)
95
284036
3367
05:01
There's much work to be done in this area.
96
289446
3576
05:05
(In Chinese)
97
293022
3645
05:08
(Applause)
98
296667
3433
05:13
Jeremy Howard: Well, that was at
a machine learning conference in China.
a machine learning conference in China.
99
301345
3399
05:16
It's not often, actually,
at academic conferences
at academic conferences
100
304744
2370
05:19
that you do hear spontaneous applause,
101
307114
1897
05:21
although of course sometimes
at TEDx conferences, feel free.
at TEDx conferences, feel free.
102
309011
3676
05:24
Everything you saw there
was happening with deep learning.
was happening with deep learning.
103
312687
2795
05:27
(Applause) Thank you.
104
315482
1525
05:29
The transcription in English
was deep learning.
was deep learning.
105
317007
2282
05:31
The translation to Chinese and the text
in the top right, deep learning,
in the top right, deep learning,
106
319289
3412
05:34
and the construction of the voice
was deep learning as well.
was deep learning as well.
107
322701
3307
05:38
So deep learning is
this extraordinary thing.
this extraordinary thing.
108
326008
3234
05:41
It's a single algorithm that
can seem to do almost anything,
can seem to do almost anything,
109
329242
3099
05:44
and I discovered that a year earlier,
it had also learned to see.
it had also learned to see.
110
332341
3111
05:47
In this obscure competition from Germany
111
335452
2176
05:49
called the German Traffic Sign
Recognition Benchmark,
Recognition Benchmark,
112
337628
2597
05:52
deep learning had learned
to recognize traffic signs like this one.
to recognize traffic signs like this one.
113
340225
3393
05:55
Not only could it
recognize the traffic signs
recognize the traffic signs
114
343618
2094
05:57
better than any other algorithm,
115
345712
1758
05:59
the leaderboard actually showed
it was better than people,
it was better than people,
116
347470
2719
06:02
about twice as good as people.
117
350189
1852
06:04
So by 2011, we had the first example
118
352041
1996
06:06
of computers that can see
better than people.
better than people.
119
354037
3405
06:09
Since that time, a lot has happened.
120
357442
2049
06:11
In 2012, Google announced that
they had a deep learning algorithm
they had a deep learning algorithm
121
359491
3514
06:15
watch YouTube videos
122
363005
1415
06:16
and crunched the data
on 16,000 computers for a month,
on 16,000 computers for a month,
123
364420
3437
06:19
and the computer independently learned
about concepts such as people and cats
about concepts such as people and cats
124
367857
4361
06:24
just by watching the videos.
125
372218
1809
06:26
This is much like the way
that humans learn.
that humans learn.
126
374027
2352
06:28
Humans don't learn
by being told what they see,
by being told what they see,
127
376379
2740
06:31
but by learning for themselves
what these things are.
what these things are.
128
379119
3331
06:34
Also in 2012, Geoffrey Hinton,
who we saw earlier,
who we saw earlier,
129
382450
3369
06:37
won the very popular ImageNet competition,
130
385819
2858
06:40
looking to try to figure out
from one and a half million images
from one and a half million images
131
388677
4141
06:44
what they're pictures of.
132
392818
1438
06:46
As of 2014, we're now down
to a six percent error rate
to a six percent error rate
133
394256
3533
06:49
in image recognition.
134
397789
1453
06:51
This is better than people, again.
135
399242
2026
06:53
So machines really are doing
an extraordinarily good job of this,
an extraordinarily good job of this,
136
401268
3769
06:57
and it is now being used in industry.
137
405037
2269
06:59
For example, Google announced last year
138
407306
3042
07:02
that they had mapped every single
location in France in two hours,
location in France in two hours,
139
410348
4585
07:06
and the way they did it was
that they fed street view images
that they fed street view images
140
414933
3447
07:10
into a deep learning algorithm
to recognize and read street numbers.
to recognize and read street numbers.
141
418380
4319
07:14
Imagine how long
it would have taken before:
it would have taken before:
142
422699
2220
07:16
dozens of people, many years.
143
424919
3355
07:20
This is also happening in China.
144
428274
1911
07:22
Baidu is kind of
the Chinese Google, I guess,
the Chinese Google, I guess,
145
430185
4036
07:26
and what you see here in the top left
146
434221
2283
07:28
is an example of a picture that I uploaded
to Baidu's deep learning system,
to Baidu's deep learning system,
147
436504
3974
07:32
and underneath you can see that the system
has understood what that picture is
has understood what that picture is
148
440478
3769
07:36
and found similar images.
149
444247
2236
07:38
The similar images actually
have similar backgrounds,
have similar backgrounds,
150
446483
2736
07:41
similar directions of the faces,
151
449219
1658
07:42
even some with their tongue out.
152
450877
1788
07:44
This is not clearly looking
at the text of a web page.
at the text of a web page.
153
452665
3030
07:47
All I uploaded was an image.
154
455695
1412
07:49
So we now have computers which
really understand what they see
really understand what they see
155
457107
4021
07:53
and can therefore search databases
156
461128
1624
07:54
of hundreds of millions
of images in real time.
of images in real time.
157
462752
3554
07:58
So what does it mean
now that computers can see?
now that computers can see?
158
466306
3230
08:01
Well, it's not just
that computers can see.
that computers can see.
159
469536
2017
08:03
In fact, deep learning
has done more than that.
has done more than that.
160
471553
2069
08:05
Complex, nuanced sentences like this one
161
473622
2948
08:08
are now understandable
with deep learning algorithms.
with deep learning algorithms.
162
476570
2824
08:11
As you can see here,
163
479394
1303
08:12
this Stanford-based system
showing the red dot at the top
showing the red dot at the top
164
480697
2768
08:15
has figured out that this sentence
is expressing negative sentiment.
is expressing negative sentiment.
165
483465
3919
08:19
Deep learning now in fact
is near human performance
is near human performance
166
487384
3406
08:22
at understanding what sentences are about
and what it is saying about those things.
and what it is saying about those things.
167
490802
5121
08:27
Also, deep learning has
been used to read Chinese,
been used to read Chinese,
168
495923
2728
08:30
again at about native
Chinese speaker level.
Chinese speaker level.
169
498651
3156
08:33
This algorithm developed
out of Switzerland
out of Switzerland
170
501807
2168
08:35
by people, none of whom speak
or understand any Chinese.
or understand any Chinese.
171
503975
3356
08:39
As I say, using deep learning
172
507331
2051
08:41
is about the best system
in the world for this,
in the world for this,
173
509382
2219
08:43
even compared to native
human understanding.
human understanding.
174
511601
5117
08:48
This is a system that we
put together at my company
put together at my company
175
516718
2964
08:51
which shows putting
all this stuff together.
all this stuff together.
176
519682
2046
08:53
These are pictures which
have no text attached,
have no text attached,
177
521728
2461
08:56
and as I'm typing in here sentences,
178
524189
2352
08:58
in real time it's understanding
these pictures
these pictures
179
526541
2969
09:01
and figuring out what they're about
180
529510
1679
09:03
and finding pictures that are similar
to the text that I'm writing.
to the text that I'm writing.
181
531189
3163
09:06
So you can see, it's actually
understanding my sentences
understanding my sentences
182
534352
2756
09:09
and actually understanding these pictures.
183
537108
2224
09:11
I know that you've seen
something like this on Google,
something like this on Google,
184
539332
2559
09:13
where you can type in things
and it will show you pictures,
and it will show you pictures,
185
541891
2775
09:16
but actually what it's doing is it's
searching the webpage for the text.
searching the webpage for the text.
186
544666
3424
09:20
This is very different from actually
understanding the images.
understanding the images.
187
548090
3001
09:23
This is something that computers
have only been able to do
have only been able to do
188
551091
2752
09:25
for the first time in the last few months.
189
553843
3248
09:29
So we can see now that computers
can not only see but they can also read,
can not only see but they can also read,
190
557091
4091
09:33
and, of course, we've shown that they
can understand what they hear.
can understand what they hear.
191
561182
3765
09:36
Perhaps not surprising now that
I'm going to tell you they can write.
I'm going to tell you they can write.
192
564947
3442
09:40
Here is some text that I generated
using a deep learning algorithm yesterday.
using a deep learning algorithm yesterday.
193
568389
4783
09:45
And here is some text that an algorithm
out of Stanford generated.
out of Stanford generated.
194
573172
3924
09:49
Each of these sentences was generated
195
577096
1764
09:50
by a deep learning algorithm
to describe each of those pictures.
to describe each of those pictures.
196
578860
4249
09:55
This algorithm before has never seen
a man in a black shirt playing a guitar.
a man in a black shirt playing a guitar.
197
583109
4472
09:59
It's seen a man before,
it's seen black before,
it's seen black before,
198
587581
2220
10:01
it's seen a guitar before,
199
589801
1599
10:03
but it has independently generated
this novel description of this picture.
this novel description of this picture.
200
591400
4294
10:07
We're still not quite at human
performance here, but we're close.
performance here, but we're close.
201
595694
3502
10:11
In tests, humans prefer
the computer-generated caption
the computer-generated caption
202
599196
4068
10:15
one out of four times.
203
603264
1527
10:16
Now this system is now only two weeks old,
204
604791
2064
10:18
so probably within the next year,
205
606855
1846
10:20
the computer algorithm will be
well past human performance
well past human performance
206
608701
2801
10:23
at the rate things are going.
207
611502
1862
10:25
So computers can also write.
208
613364
3049
10:28
So we put all this together and it leads
to very exciting opportunities.
to very exciting opportunities.
209
616413
3475
10:31
For example, in medicine,
210
619888
1492
10:33
a team in Boston announced
that they had discovered
that they had discovered
211
621380
2525
10:35
dozens of new clinically relevant features
212
623905
2949
10:38
of tumors which help doctors
make a prognosis of a cancer.
make a prognosis of a cancer.
213
626854
4266
10:44
Very similarly, in Stanford,
214
632220
2296
10:46
a group there announced that,
looking at tissues under magnification,
looking at tissues under magnification,
215
634516
3663
10:50
they've developed
a machine learning-based system
a machine learning-based system
216
638179
2381
10:52
which in fact is better
than human pathologists
than human pathologists
217
640560
2582
10:55
at predicting survival rates
for cancer sufferers.
for cancer sufferers.
218
643142
4377
10:59
In both of these cases, not only
were the predictions more accurate,
were the predictions more accurate,
219
647519
3245
11:02
but they generated new insightful science.
220
650764
2502
11:05
In the radiology case,
221
653276
1505
11:06
they were new clinical indicators
that humans can understand.
that humans can understand.
222
654781
3095
11:09
In this pathology case,
223
657876
1792
11:11
the computer system actually discovered
that the cells around the cancer
that the cells around the cancer
224
659668
4500
11:16
are as important as
the cancer cells themselves
the cancer cells themselves
225
664168
3340
11:19
in making a diagnosis.
226
667508
1752
11:21
This is the opposite of what pathologists
had been taught for decades.
had been taught for decades.
227
669260
5361
11:26
In each of those two cases,
they were systems developed
they were systems developed
228
674621
3292
11:29
by a combination of medical experts
and machine learning experts,
and machine learning experts,
229
677913
3621
11:33
but as of last year,
we're now beyond that too.
we're now beyond that too.
230
681534
2741
11:36
This is an example of
identifying cancerous areas
identifying cancerous areas
231
684275
3549
11:39
of human tissue under a microscope.
232
687824
2530
11:42
The system being shown here
can identify those areas more accurately,
can identify those areas more accurately,
233
690354
4613
11:46
or about as accurately,
as human pathologists,
as human pathologists,
234
694967
2775
11:49
but was built entirely with deep learning
using no medical expertise
using no medical expertise
235
697742
3392
11:53
by people who have
no background in the field.
no background in the field.
236
701134
2526
11:56
Similarly, here, this neuron segmentation.
237
704730
2555
11:59
We can now segment neurons
about as accurately as humans can,
about as accurately as humans can,
238
707285
3668
12:02
but this system was developed
with deep learning
with deep learning
239
710953
2717
12:05
using people with no previous
background in medicine.
background in medicine.
240
713670
3251
12:08
So myself, as somebody with
no previous background in medicine,
no previous background in medicine,
241
716921
3227
12:12
I seem to be entirely well qualified
to start a new medical company,
to start a new medical company,
242
720148
3727
12:15
which I did.
243
723875
2146
12:18
I was kind of terrified of doing it,
244
726021
1740
12:19
but the theory seemed to suggest
that it ought to be possible
that it ought to be possible
245
727761
2889
12:22
to do very useful medicine
using just these data analytic techniques.
using just these data analytic techniques.
246
730650
5492
12:28
And thankfully, the feedback
has been fantastic,
has been fantastic,
247
736142
2480
12:30
not just from the media
but from the medical community,
but from the medical community,
248
738622
2356
12:32
who have been very supportive.
249
740978
2344
12:35
The theory is that we can take
the middle part of the medical process
the middle part of the medical process
250
743322
4149
12:39
and turn that into data analysis
as much as possible,
as much as possible,
251
747471
2893
12:42
leaving doctors to do
what they're best at.
what they're best at.
252
750364
3065
12:45
I want to give you an example.
253
753429
1602
12:47
It now takes us about 15 minutes
to generate a new medical diagnostic test
to generate a new medical diagnostic test
254
755031
4944
12:51
and I'll show you that in real time now,
255
759975
1954
12:53
but I've compressed it down to
three minutes by cutting some pieces out.
three minutes by cutting some pieces out.
256
761929
3487
12:57
Rather than showing you
creating a medical diagnostic test,
creating a medical diagnostic test,
257
765416
3061
13:00
I'm going to show you
a diagnostic test of car images,
a diagnostic test of car images,
258
768477
3369
13:03
because that's something
we can all understand.
we can all understand.
259
771846
2222
13:06
So here we're starting with
about 1.5 million car images,
about 1.5 million car images,
260
774068
3201
13:09
and I want to create something
that can split them into the angle
that can split them into the angle
261
777269
3206
13:12
of the photo that's being taken.
262
780475
2223
13:14
So these images are entirely unlabeled,
so I have to start from scratch.
so I have to start from scratch.
263
782698
3888
13:18
With our deep learning algorithm,
264
786586
1865
13:20
it can automatically identify
areas of structure in these images.
areas of structure in these images.
265
788451
3707
13:24
So the nice thing is that the human
and the computer can now work together.
and the computer can now work together.
266
792158
3620
13:27
So the human, as you can see here,
267
795778
2178
13:29
is telling the computer
about areas of interest
about areas of interest
268
797956
2675
13:32
which it wants the computer then
to try and use to improve its algorithm.
to try and use to improve its algorithm.
269
800631
4650
13:37
Now, these deep learning systems actually
are in 16,000-dimensional space,
are in 16,000-dimensional space,
270
805281
4296
13:41
so you can see here the computer
rotating this through that space,
rotating this through that space,
271
809577
3432
13:45
trying to find new areas of structure.
272
813009
1992
13:47
And when it does so successfully,
273
815001
1781
13:48
the human who is driving it can then
point out the areas that are interesting.
point out the areas that are interesting.
274
816782
4004
13:52
So here, the computer has
successfully found areas,
successfully found areas,
275
820786
2422
13:55
for example, angles.
276
823208
2562
13:57
So as we go through this process,
277
825770
1606
13:59
we're gradually telling
the computer more and more
the computer more and more
278
827376
2340
14:01
about the kinds of structures
we're looking for.
we're looking for.
279
829716
2428
14:04
You can imagine in a diagnostic test
280
832144
1772
14:05
this would be a pathologist identifying
areas of pathosis, for example,
areas of pathosis, for example,
281
833916
3350
14:09
or a radiologist indicating
potentially troublesome nodules.
potentially troublesome nodules.
282
837266
5026
14:14
And sometimes it can be
difficult for the algorithm.
difficult for the algorithm.
283
842292
2559
14:16
In this case, it got kind of confused.
284
844851
1964
14:18
The fronts and the backs
of the cars are all mixed up.
of the cars are all mixed up.
285
846815
2550
14:21
So here we have to be a bit more careful,
286
849365
2072
14:23
manually selecting these fronts
as opposed to the backs,
as opposed to the backs,
287
851437
3232
14:26
then telling the computer
that this is a type of group
that this is a type of group
288
854669
5506
14:32
that we're interested in.
289
860175
1348
14:33
So we do that for a while,
we skip over a little bit,
we skip over a little bit,
290
861523
2677
14:36
and then we train the
machine learning algorithm
machine learning algorithm
291
864200
2246
14:38
based on these couple of hundred things,
292
866446
1974
14:40
and we hope that it's gotten a lot better.
293
868420
2025
14:42
You can see, it's now started to fade
some of these pictures out,
some of these pictures out,
294
870445
3073
14:45
showing us that it already is recognizing
how to understand some of these itself.
how to understand some of these itself.
295
873518
4708
14:50
We can then use this concept
of similar images,
of similar images,
296
878226
2902
14:53
and using similar images, you can now see,
297
881128
2094
14:55
the computer at this point is able to
entirely find just the fronts of cars.
entirely find just the fronts of cars.
298
883222
4019
14:59
So at this point, the human
can tell the computer,
can tell the computer,
299
887241
2948
15:02
okay, yes, you've done
a good job of that.
a good job of that.
300
890189
2293
15:05
Sometimes, of course, even at this point
301
893652
2185
15:07
it's still difficult
to separate out groups.
to separate out groups.
302
895837
3674
15:11
In this case, even after we let the
computer try to rotate this for a while,
computer try to rotate this for a while,
303
899511
3884
15:15
we still find that the left sides
and the right sides pictures
and the right sides pictures
304
903399
3345
15:18
are all mixed up together.
305
906744
1478
15:20
So we can again give
the computer some hints,
the computer some hints,
306
908222
2140
15:22
and we say, okay, try and find
a projection that separates out
a projection that separates out
307
910362
2976
15:25
the left sides and the right sides
as much as possible
as much as possible
308
913338
2607
15:27
using this deep learning algorithm.
309
915945
2122
15:30
And giving it that hint --
ah, okay, it's been successful.
ah, okay, it's been successful.
310
918067
2942
15:33
It's managed to find a way
of thinking about these objects
of thinking about these objects
311
921009
2882
15:35
that's separated out these together.
312
923891
2380
15:38
So you get the idea here.
313
926271
2438
15:40
This is a case not where the human
is being replaced by a computer,
is being replaced by a computer,
314
928709
8197
15:48
but where they're working together.
315
936906
2640
15:51
What we're doing here is we're replacing
something that used to take a team
something that used to take a team
316
939546
3550
15:55
of five or six people about seven years
317
943096
2002
15:57
and replacing it with something
that takes 15 minutes
that takes 15 minutes
318
945098
2605
15:59
for one person acting alone.
319
947703
2505
16:02
So this process takes about
four or five iterations.
four or five iterations.
320
950208
3950
16:06
You can see we now have 62 percent
321
954158
1859
16:08
of our 1.5 million images
classified correctly.
classified correctly.
322
956017
2959
16:10
And at this point, we
can start to quite quickly
can start to quite quickly
323
958976
2472
16:13
grab whole big sections,
324
961448
1297
16:14
check through them to make sure
that there's no mistakes.
that there's no mistakes.
325
962745
2919
16:17
Where there are mistakes, we can
let the computer know about them.
let the computer know about them.
326
965664
3952
16:21
And using this kind of process
for each of the different groups,
for each of the different groups,
327
969616
3045
16:24
we are now up to
an 80 percent success rate
an 80 percent success rate
328
972661
2487
16:27
in classifying the 1.5 million images.
329
975148
2415
16:29
And at this point, it's just a case
330
977563
2078
16:31
of finding the small number
that aren't classified correctly,
that aren't classified correctly,
331
979641
3579
16:35
and trying to understand why.
332
983220
2888
16:38
And using that approach,
333
986108
1743
16:39
by 15 minutes we get
to 97 percent classification rates.
to 97 percent classification rates.
334
987851
4121
16:43
So this kind of technique
could allow us to fix a major problem,
could allow us to fix a major problem,
335
991972
4600
16:48
which is that there's a lack
of medical expertise in the world.
of medical expertise in the world.
336
996578
3036
16:51
The World Economic Forum says
that there's between a 10x and a 20x
that there's between a 10x and a 20x
337
999614
3489
16:55
shortage of physicians
in the developing world,
in the developing world,
338
1003103
2624
16:57
and it would take about 300 years
339
1005727
2113
16:59
to train enough people
to fix that problem.
to fix that problem.
340
1007840
2894
17:02
So imagine if we can help
enhance their efficiency
enhance their efficiency
341
1010734
2885
17:05
using these deep learning approaches?
342
1013619
2839
17:08
So I'm very excited
about the opportunities.
about the opportunities.
343
1016458
2232
17:10
I'm also concerned about the problems.
344
1018690
2589
17:13
The problem here is that
every area in blue on this map
every area in blue on this map
345
1021279
3124
17:16
is somewhere where services
are over 80 percent of employment.
are over 80 percent of employment.
346
1024403
3769
17:20
What are services?
347
1028172
1787
17:21
These are services.
348
1029959
1514
17:23
These are also the exact things that
computers have just learned how to do.
computers have just learned how to do.
349
1031473
4154
17:27
So 80 percent of the world's employment
in the developed world
in the developed world
350
1035627
3804
17:31
is stuff that computers
have just learned how to do.
have just learned how to do.
351
1039431
2532
17:33
What does that mean?
352
1041963
1440
17:35
Well, it'll be fine.
They'll be replaced by other jobs.
They'll be replaced by other jobs.
353
1043403
2583
17:37
For example, there will be
more jobs for data scientists.
more jobs for data scientists.
354
1045986
2707
17:40
Well, not really.
355
1048693
817
17:41
It doesn't take data scientists
very long to build these things.
very long to build these things.
356
1049510
3118
17:44
For example, these four algorithms
were all built by the same guy.
were all built by the same guy.
357
1052628
3252
17:47
So if you think, oh,
it's all happened before,
it's all happened before,
358
1055880
2438
17:50
we've seen the results in the past
of when new things come along
of when new things come along
359
1058318
3808
17:54
and they get replaced by new jobs,
360
1062126
2252
17:56
what are these new jobs going to be?
361
1064378
2116
17:58
It's very hard for us to estimate this,
362
1066494
1871
18:00
because human performance
grows at this gradual rate,
grows at this gradual rate,
363
1068365
2739
18:03
but we now have a system, deep learning,
364
1071104
2562
18:05
that we know actually grows
in capability exponentially.
in capability exponentially.
365
1073666
3227
18:08
And we're here.
366
1076893
1605
18:10
So currently, we see the things around us
367
1078498
2061
18:12
and we say, "Oh, computers
are still pretty dumb." Right?
are still pretty dumb." Right?
368
1080559
2676
18:15
But in five years' time,
computers will be off this chart.
computers will be off this chart.
369
1083235
3429
18:18
So we need to be starting to think
about this capability right now.
about this capability right now.
370
1086664
3865
18:22
We have seen this once before, of course.
371
1090529
2050
18:24
In the Industrial Revolution,
372
1092579
1387
18:25
we saw a step change
in capability thanks to engines.
in capability thanks to engines.
373
1093966
2851
18:29
The thing is, though,
that after a while, things flattened out.
that after a while, things flattened out.
374
1097667
3138
18:32
There was social disruption,
375
1100805
1702
18:34
but once engines were used
to generate power in all the situations,
to generate power in all the situations,
376
1102507
3439
18:37
things really settled down.
377
1105946
2354
18:40
The Machine Learning Revolution
378
1108300
1473
18:41
is going to be very different
from the Industrial Revolution,
from the Industrial Revolution,
379
1109773
2909
18:44
because the Machine Learning Revolution,
it never settles down.
it never settles down.
380
1112682
2950
18:47
The better computers get
at intellectual activities,
at intellectual activities,
381
1115632
2982
18:50
the more they can build better computers
to be better at intellectual capabilities,
to be better at intellectual capabilities,
382
1118614
4248
18:54
so this is going to be a kind of change
383
1122862
1908
18:56
that the world has actually
never experienced before,
never experienced before,
384
1124770
2478
18:59
so your previous understanding
of what's possible is different.
of what's possible is different.
385
1127248
3306
19:02
This is already impacting us.
386
1130974
1780
19:04
In the last 25 years,
as capital productivity has increased,
as capital productivity has increased,
387
1132754
3630
19:08
labor productivity has been flat,
in fact even a little bit down.
in fact even a little bit down.
388
1136400
4188
19:13
So I want us to start
having this discussion now.
having this discussion now.
389
1141408
2741
19:16
I know that when I often tell people
about this situation,
about this situation,
390
1144149
3027
19:19
people can be quite dismissive.
391
1147176
1490
19:20
Well, computers can't really think,
392
1148666
1673
19:22
they don't emote,
they don't understand poetry,
they don't understand poetry,
393
1150339
3028
19:25
we don't really understand how they work.
394
1153367
2521
19:27
So what?
395
1155888
1486
19:29
Computers right now can do the things
396
1157374
1804
19:31
that humans spend most
of their time being paid to do,
of their time being paid to do,
397
1159178
2719
19:33
so now's the time to start thinking
398
1161897
1731
19:35
about how we're going to adjust our
social structures and economic structures
social structures and economic structures
399
1163628
4387
19:40
to be aware of this new reality.
400
1168015
1840
19:41
Thank you.
401
1169855
1533
19:43
(Applause)
402
1171388
802
ABOUT THE SPEAKER
Jeremy Howard - Data scientistJeremy Howard imagines how advanced machine learning can improve our lives.
Why you should listen
Jeremy Howard is the CEO of Enlitic, an advanced machine learning company in San Francisco. Previously, he was the president and chief scientist at Kaggle, a community and competition platform of over 200,000 data scientists. Howard is a faculty member at Singularity University, where he teaches data science. He is also a Young Global Leader with the World Economic Forum, and spoke at the World Economic Forum Annual Meeting 2014 on "Jobs for the Machines."
Howard advised Khosla Ventures as their Data Strategist, identifying opportunities for investing in data-driven startups and mentoring portfolio companies to build data-driven businesses. He was the founding CEO of two successful Australian startups, FastMail and Optimal Decisions Group.
Jeremy Howard | Speaker | TED.com