1
00:00:17,816 --> 00:00:21,325
Our world is changing in many ways

2
00:00:21,325 --> 00:00:25,975
and one of the things which is going
to have a huge impact on our future

3
00:00:25,975 --> 00:00:29,363
is artificial intelligence - AI,

4
00:00:29,363 --> 00:00:32,953
bringing another industrial revolution.

5
00:00:33,627 --> 00:00:39,504
Previous industrial revolutions
expanded human's mechanical power.

6
00:00:40,014 --> 00:00:45,572
This new revolution,
this second machine age

7
00:00:45,572 --> 00:00:50,122
is going to expand
our cognitive abilities,

8
00:00:50,122 --> 00:00:52,102
our mental power.

9
00:00:52,782 --> 00:00:57,177
Computers are not just going
to replace manual labor,

10
00:00:57,597 --> 00:00:59,897
but also mental labor.

11
00:01:00,500 --> 00:01:03,450
So, where do we stand today?

12
00:01:04,034 --> 00:01:07,724
You may have heard
about what happened last March

13
00:01:07,724 --> 00:01:11,776
when a machine learning system
called AlphaGo

14
00:01:11,776 --> 00:01:17,708
used deep learning to beat
the world champion at the game of Go.

15
00:01:18,279 --> 00:01:20,679
Go is an ancient Chinese game

16
00:01:20,679 --> 00:01:24,159
which had been much more difficult
for computers to master

17
00:01:24,159 --> 00:01:25,982
than the game of chess.

18
00:01:26,893 --> 00:01:32,086
How did we succeed,
now, after decades of AI research?

19
00:01:33,068 --> 00:01:36,698
AlphaGo was trained to play Go.

20
00:01:37,678 --> 00:01:41,300
First, by watching over and over

21
00:01:41,814 --> 00:01:46,894
tens of millions of moves made
by very strong human players.

22
00:01:47,746 --> 00:01:52,496
Then, by playing against itself,
millions of games.

23
00:01:54,222 --> 00:01:59,941
Machine Learning allows computers
to learn from examples.

24
00:02:00,465 --> 00:02:02,575
To learn from data.

25
00:02:03,885 --> 00:02:07,235
Machine learning
has turned out to be a key

26
00:02:07,235 --> 00:02:11,635
to cram knowledge into computers.

27
00:02:12,174 --> 00:02:14,066
And this is important

28
00:02:14,066 --> 00:02:19,296
because knowledge
is what enables intelligence.

29
00:02:20,438 --> 00:02:26,768
Putting knowledge into computers had been
a challenge for previous approaches to AI.

30
00:02:27,515 --> 00:02:28,745
Why?

31
00:02:29,059 --> 00:02:33,859
There are many things
which we know intuitively.

32
00:02:34,601 --> 00:02:38,081
So we cannot communicate them verbally.

33
00:02:38,619 --> 00:02:42,780
We do not have conscious access
to that intuitive knowledge.

34
00:02:43,270 --> 00:02:46,690
How can we program computers
without knowledge?

35
00:02:47,664 --> 00:02:49,114
What's the solution?

36
00:02:49,314 --> 00:02:55,343
The solution is for machines to learn
that knowledge by themselves,

37
00:02:55,343 --> 00:02:56,443
just as we do.

38
00:02:56,443 --> 00:03:03,194
And this is important because knowledge
is what enables intelligence.

39
00:03:03,194 --> 00:03:06,974
My mission has been
to contribute to discover

40
00:03:06,974 --> 00:03:12,676
and understand principles
of intelligence through learning.

41
00:03:13,166 --> 00:03:18,116
Whether animal, human or machine learning.

42
00:03:19,450 --> 00:03:25,066
I and others believe that there are
a few key principles,

43
00:03:25,066 --> 00:03:27,296
just like the law of physics.

44
00:03:27,885 --> 00:03:32,745
Simple principles which could explain
our own intelligence

45
00:03:32,745 --> 00:03:36,741
and help us build intelligent machines.

46
00:03:37,885 --> 00:03:41,595
For example, think about the laws
of aerodynamics

47
00:03:41,595 --> 00:03:48,036
which are general enough to explain
the flight of both, birds and planes.

48
00:03:49,146 --> 00:03:55,381
Wouldn't it be amazing to discover
such simple but powerful principles

49
00:03:55,381 --> 00:03:59,186
that would explain intelligence itself?

50
00:04:00,026 --> 00:04:03,394
Well, we've made some progress.

51
00:04:04,384 --> 00:04:10,857
My collaborators and I have contributed
in recent years in a revolution in AI

52
00:04:11,777 --> 00:04:16,397
with our research on neural networks
and deep learning,

53
00:04:16,397 --> 00:04:20,946
an approach to machine learning
which is inspired by the brain.

54
00:04:22,041 --> 00:04:25,243
It started with speech recognition

55
00:04:25,243 --> 00:04:29,963
on your phones,
with neural networks since 2012.

56
00:04:30,977 --> 00:04:35,647
Shortly after, came a breakthrough
in computer vision.

57
00:04:36,680 --> 00:04:43,087
Computers can now do a pretty good job
of recognizing the content of images.

58
00:04:43,674 --> 00:04:50,049
In fact, they approach human performance
on some benchmarks over the last 5 years.

59
00:04:50,711 --> 00:04:54,721
A computer can now get
an intuitive understanding

60
00:04:54,721 --> 00:04:58,191
of the visual appearance of a Go-board

61
00:04:58,191 --> 00:05:01,763
that is comparable to that
of the best human players.

62
00:05:01,763 --> 00:05:03,454
More recently,

63
00:05:03,454 --> 00:05:06,584
following some discoveries made in my lab,

64
00:05:06,584 --> 00:05:11,408
deep learning has been used to translate
from one language to another

65
00:05:11,414 --> 00:05:14,441
and you are going to start seeing
this in Google translate.

66
00:05:15,191 --> 00:05:18,192
This is expanding the computer's ability

67
00:05:18,192 --> 00:05:22,532
to understand and generate
natural language.

68
00:05:23,550 --> 00:05:25,517
But don't be fooled.

69
00:05:25,517 --> 00:05:30,048
We are still very, very far from a machine

70
00:05:30,048 --> 00:05:34,033
that would be as able as humans

71
00:05:34,033 --> 00:05:37,593
to learn to master
many aspects of our world.

72
00:05:38,541 --> 00:05:41,237
So, let's take an example.

73
00:05:41,637 --> 00:05:46,787
Even a two year old child
is able to learn things

74
00:05:46,787 --> 00:05:50,657
in a way that computers
are not able to do right now.

75
00:05:51,767 --> 00:05:56,169
A two year old child actually
masters intuitive physics.

76
00:05:56,968 --> 00:06:01,908
She knows when she drops a ball
that it is going to fall down.

77
00:06:02,493 --> 00:06:06,093
When she spills some liquids
she expects the resulting mess.

78
00:06:06,586 --> 00:06:09,516
Her parents do not need to teach her

79
00:06:09,516 --> 00:06:12,980
about Newton's laws
or differential equations.

80
00:06:13,840 --> 00:06:20,200
She discovers all these things by herself
in a unsupervised way.

81
00:06:21,352 --> 00:06:27,712
Unsupervised learning actually remains
one of the key challenges for AI.

82
00:06:28,184 --> 00:06:33,014
And it may take several more decades
of fundamental research

83
00:06:33,014 --> 00:06:34,674
to crack that knot.

84
00:06:34,674 --> 00:06:40,895
Unsupervised learning is actually trying
to discover representations of the data.

85
00:06:41,729 --> 00:06:43,779
Let me show you and example.

86
00:06:44,364 --> 00:06:49,346
Consider a page on the screen
that you're seeing with your eyes

87
00:06:49,346 --> 00:06:54,196
or that the computer is seeing
as an image, a bunch of pixels.

88
00:06:54,993 --> 00:07:00,113
In order to answer a question
about the content of the image

89
00:07:00,863 --> 00:07:05,211
you need to understand
its high-level meaning.

90
00:07:05,674 --> 00:07:10,821
This high level meaning corresponds
to the highest level of representation

91
00:07:10,821 --> 00:07:12,321
in your brain.

92
00:07:12,906 --> 00:07:18,308
Low down, you have
the individual meaning of words

93
00:07:19,188 --> 00:07:23,798
and even lower down, you have characters
which make up the words.

94
00:07:24,810 --> 00:07:27,677
Those characters could be
rendered in different ways

95
00:07:27,677 --> 00:07:30,879
with different strokes
that make up the characters.

96
00:07:31,559 --> 00:07:34,839
And those strokes are made up of edges


97
00:07:34,839 --> 00:07:37,284
and those edges are made up of pixels.

98
00:07:37,284 --> 00:07:40,454
So these are different
levels of representation.

99
00:07:41,079 --> 00:07:44,236
But the pixels are not
sufficient by themselves

100
00:07:44,236 --> 00:07:46,584
to make sense of the image,

101
00:07:46,584 --> 00:07:51,904
to answer a high level question
about the content of the page.

102
00:07:52,932 --> 00:07:57,594
Your brain actually has
these different levels of representation

103
00:07:57,594 --> 00:08:02,291
starting with neurons
in the first visual area of cortex - V1,

104
00:08:02,291 --> 00:08:04,596
which recognizes edges.

105
00:08:04,596 --> 00:08:09,334
And then, neurons in the second
visual area of cortex - V2,

106
00:08:09,334 --> 00:08:12,800
which recognizes strokes and small shapes.

107
00:08:12,800 --> 00:08:17,060
Higher up, you have neurons
which detect parts of objects

108
00:08:17,060 --> 00:08:19,992
and then objects and full scenes.

109
00:08:21,182 --> 00:08:24,757
Neural networks,
when they're trained with images,

110
00:08:24,757 --> 00:08:28,860
can actually discover these types
of levels of representation

111
00:08:28,860 --> 00:08:32,778
that match pretty well
what we observe in the brain.

112
00:08:33,638 --> 00:08:38,798
Both, biological neural networks,
which are what you have in your brain

113
00:08:38,804 --> 00:08:42,828
and the deep neural networks
that we train on our machines

114
00:08:42,845 --> 00:08:48,075
can learn to transform from one level
of representation to the next,

115
00:08:48,369 --> 00:08:53,299
with the high levels corresponding
to more abstract notions.

116
00:08:53,299 --> 00:08:57,562
For example the abstract notion
of the character A

117
00:08:57,562 --> 00:09:00,891
can be rendered in many different ways
at the lowest levels

118
00:09:00,891 --> 00:09:03,887
as many different configurations of pixels

119
00:09:03,887 --> 00:09:09,097
depending on the position,
rotation, font and so on.

120
00:09:10,445 --> 00:09:15,815
So, how do we learn
these high levels of representations?

121
00:09:16,962 --> 00:09:20,681
One thing that has been
very successful up to now

122
00:09:20,681 --> 00:09:22,853
in the applications of deep learning,

123
00:09:22,855 --> 00:09:25,985
is what we call supervised learning.

124
00:09:26,297 --> 00:09:31,588
With supervised learning, the computer
needs to be taken by the hand

125
00:09:31,594 --> 00:09:35,467
and humans have to tell the computer
the answer to many questions.

126
00:09:35,467 --> 00:09:41,420
For example, on millions and millions
of images, humans have to tell the machine

127
00:09:41,420 --> 00:09:44,271
well... for this image, it is a cat.

128
00:09:44,273 --> 00:09:47,095
For this image, it is a dog.

129
00:09:47,095 --> 00:09:49,585
For this image, it is a laptop.

130
00:09:49,605 --> 00:09:55,595
For this image, it is a keyboard,
And so on, and so on millions of times.

131
00:09:56,066 --> 00:10:01,026
This is very painful and we use
crowdsourcing to manage to do that.

132
00:10:01,461 --> 00:10:03,396
Although, this is very powerful

133
00:10:03,416 --> 00:10:06,269
and we are able to solve
many interesting problems,

134
00:10:06,269 --> 00:10:08,313
humans are much stronger

135
00:10:08,313 --> 00:10:12,076
and they can learn over many more
different aspects of the world

136
00:10:12,076 --> 00:10:13,809
in a much more autonomous way,

137
00:10:13,809 --> 00:10:17,609
just as we've seen with the child
learning about intuitive physics.

138
00:10:17,625 --> 00:10:23,739
Unsupervised learning could also help us
deal with self-driving cars.

139
00:10:24,567 --> 00:10:26,097
Let me explain what I mean:

140
00:10:26,097 --> 00:10:31,835
Unsupervised learning allows computers
to project themselves into the future

141
00:10:31,835 --> 00:10:37,205
to generate plausible futures
conditioned on the current situation.

142
00:10:38,369 --> 00:10:42,899
And that allows computers to reason
and to plan ahead.

143
00:10:43,450 --> 00:10:47,985
Even for circumstances
they have not been trained on.

144
00:10:48,751 --> 00:10:50,441
This is important

145
00:10:50,441 --> 00:10:53,951
because if we use supervised learning
we would have to tell the computers

146
00:10:53,951 --> 00:10:57,395
about all the circumstances
where the car could be

147
00:10:57,395 --> 00:11:01,375
and how humans
would react in that situation.

148
00:11:02,451 --> 00:11:06,191
How did I learn to avoid
dangerous driving behavior?

149
00:11:07,276 --> 00:11:10,791
Did I have to die
a thousand times in an accident?

150
00:11:10,793 --> 00:11:12,106
(Laughter)

151
00:11:12,106 --> 00:11:14,606
Well, that's the way we train
machines right now.

152
00:11:15,175 --> 00:11:18,340
So, it's not going to fly
or at least not to drive.

153
00:11:18,340 --> 00:11:19,928
(Laughter)

154
00:11:21,288 --> 00:11:25,657
So, what we need is to train our models

155
00:11:25,657 --> 00:11:31,294
to be able to generate plausible images
or plausible futures, be creative.

156
00:11:31,294 --> 00:11:33,934
And we are making progress with that.

157
00:11:33,934 --> 00:11:37,457
So, we're training
these deep neural networks

158
00:11:37,463 --> 00:11:40,818
to go from high-level meaning to pixels

159
00:11:40,818 --> 00:11:43,298
rather than from pixels
to high level meaning,

160
00:11:43,307 --> 00:11:46,787
going into the other direction
through the levels of representation.

161
00:11:46,787 --> 00:11:50,461
And this way, the computer
can generate images

162
00:11:51,191 --> 00:11:55,072
that are new images different
from what the computer has seen

163
00:11:55,072 --> 00:11:56,488
while it was trained,

164
00:11:57,018 --> 00:12:00,369
but are plausible and look like natural images.

165
00:12:01,888 --> 00:12:06,332
We can also use these models
to dream up strange,

166
00:12:06,342 --> 00:12:09,492
sometimes scary images,

167
00:12:09,492 --> 00:12:11,795
just like our dreams and nightmares.

168
00:12:12,682 --> 00:12:16,847
Here's some images
that were synthesized by the computer

169
00:12:16,847 --> 00:12:19,826
using these deep charted models.

170
00:12:19,826 --> 00:12:21,651
They look like natural images

171
00:12:21,651 --> 00:12:24,551
but if you look closely,
you will see they are different

172
00:12:25,458 --> 00:12:28,697
and they're still missing
some of the important details

173
00:12:28,697 --> 00:12:31,063
that we would recognize as natural.

174
00:12:31,995 --> 00:12:33,951
About 10 years ago,

175
00:12:33,951 --> 00:12:38,921
unsupervised learning has been
a key to the breakthrough

176
00:12:38,921 --> 00:12:42,443
that we obtained
discovering deep learning.

177
00:12:44,140 --> 00:12:48,055
This was happening in just few labs,
including mine at the time

178
00:12:48,055 --> 00:12:51,455
at a time when neural networks
were not popular.

179
00:12:51,455 --> 00:12:55,217
They were almost abandoned
by the scientific community.

180
00:12:56,394 --> 00:12:58,935
Now, things have changed a lot.

181
00:12:58,935 --> 00:13:01,375
It has become a very hard field.

182
00:13:01,384 --> 00:13:06,933
There are now hundreds of students
every year applying for graduate studies

183
00:13:06,954 --> 00:13:09,784
at my lab with my collaborators.

184
00:13:11,010 --> 00:13:16,630
Montreal has become
the largest academic concentration

185
00:13:16,637 --> 00:13:19,387
of deep learning researchers in the world.

186
00:13:20,182 --> 00:13:26,115
We just received a huge
research grant of 94 million dollars

187
00:13:26,127 --> 00:13:29,797
to push the boundaries
of AI and data science

188
00:13:29,797 --> 00:13:36,067
and also to transfer technology of deep
learning and data science to the industry.

189
00:13:37,249 --> 00:13:43,791
Business people stimulated by all this
are creating start-ups, industrial labs,

190
00:13:43,791 --> 00:13:46,914
many of which near the universities.

191
00:13:48,543 --> 00:13:49,625
For example,

192
00:13:49,625 --> 00:13:54,733
just a few weeks ago, we announced
the launch of a start-up factory

193
00:13:54,733 --> 00:13:56,507
called 'Element AI'

194
00:13:56,507 --> 00:13:59,605
which is going to focus
on the deep learning applications.

195
00:14:01,562 --> 00:14:05,722
There are just not enough
deep learning experts.

196
00:14:06,355 --> 00:14:10,677
So, they are getting paid crazy salaries,

197
00:14:11,027 --> 00:14:17,212
and many of my former academic colleagues
have accepted generous deals

198
00:14:17,228 --> 00:14:20,518
from companies to work in industrial labs.

199
00:14:21,081 --> 00:14:25,010
I, for myself, have chosen
to stay in university,

200
00:14:25,010 --> 00:14:27,166
to work for the public good,

201
00:14:27,166 --> 00:14:28,886
to work with students,

202
00:14:28,902 --> 00:14:30,592
to remain independent.

203
00:14:30,596 --> 00:14:34,836
To guide the next generation
of deep learning experts.

204
00:14:35,294 --> 00:14:41,024
One thing that we are doing
beyond commercial value

205
00:14:41,024 --> 00:14:44,654
is thinking about the social
implications of AI.

206
00:14:45,881 --> 00:14:50,026
Many of us are now starting
to turn our eyes

207
00:14:50,026 --> 00:14:55,986
towards social value added
applications, like health.

208
00:14:56,457 --> 00:14:58,956
We think that we can use deep learning

209
00:14:58,956 --> 00:15:02,696
to improve treatment
with personalized medicine.

210
00:15:03,956 --> 00:15:05,671
I believe that in the future,

211
00:15:05,671 --> 00:15:10,361
as we collect more data from millions
and billions people around the earth,

212
00:15:10,361 --> 00:15:13,856
we will be able to provide medical advice

213
00:15:13,856 --> 00:15:17,246
to billions of people
who don't have access to it right now.

214
00:15:17,601 --> 00:15:22,724
And we can imagine many other
applications for social value of AI.

215
00:15:23,140 --> 00:15:26,238
For example, something
that will come out of our research

216
00:15:26,238 --> 00:15:28,582
on natural language understanding

217
00:15:29,328 --> 00:15:31,199
is providing all kinds of services

218
00:15:31,199 --> 00:15:34,059
like legal services,
to those who can't afford them.

219
00:15:34,512 --> 00:15:37,342
We are now turning our eyes

220
00:15:37,342 --> 00:15:41,132
also towards the social implications
of AI in my community.

221
00:15:41,690 --> 00:15:44,801
But it's not just for experts
to think about this.

222
00:15:46,026 --> 00:15:49,936
I believe that beyond the math
and the jargon,

223
00:15:49,936 --> 00:15:53,102
ordinary people can get the sense

224
00:15:53,138 --> 00:15:55,901
of what goes on under the hood

225
00:15:55,901 --> 00:16:01,191
enough to participate
in the important decisions

226
00:16:01,191 --> 00:16:06,547
that will take place, in the next
few years and decades about AI.

227
00:16:07,580 --> 00:16:09,280
So please,

228
00:16:09,930 --> 00:16:16,230
set aside your fees and give yourself
some space to learn about it.

229
00:16:17,842 --> 00:16:22,532
My collaborators and I have written
several introductory papers

230
00:16:22,542 --> 00:16:25,376
and a book entitled "Deep Learning"

231
00:16:25,376 --> 00:16:29,619
to help students and engineers
jump into this exciting field.

232
00:16:30,659 --> 00:16:35,784
There are also many online resources:
softwares, tutorials, videos..

233
00:16:36,310 --> 00:16:41,210
and many undergraduate students
are learning a lot of this

234
00:16:41,210 --> 00:16:44,548
about research in deep learning
by themselves,

235
00:16:44,548 --> 00:16:47,835
to later join the ranks of labs like mine.

236
00:16:49,370 --> 00:16:55,170
Ai is going to have a profound
impact on our society.

237
00:16:56,652 --> 00:17:01,672
So, it's important to ask:
How are we going to use it?

238
00:17:03,368 --> 00:17:07,896
Immense positives may come
along with negatives

239
00:17:07,896 --> 00:17:10,166
such as military use

240
00:17:10,797 --> 00:17:15,357
or rapid disruptive changes
in the job market.

241
00:17:15,948 --> 00:17:21,629
To make sure the collective choices
that will be made about AI

242
00:17:21,629 --> 00:17:23,074
in the next few years,

243
00:17:23,074 --> 00:17:25,144
will be for the benefit of all,

244
00:17:25,144 --> 00:17:28,557
every citizen should take an active role

245
00:17:28,557 --> 00:17:32,911
in defining how AI will shape our future.

246
00:17:33,871 --> 00:17:34,891
Thank you.

247
00:17:35,065 --> 00:17:39,395
(Applause)