0:00:17.816,0:00:21.325
Our world is changing in many ways

0:00:21.325,0:00:25.975
and one of the things which is going[br]to have a huge impact on our future

0:00:25.975,0:00:29.363
is artificial intelligence - AI,

0:00:29.363,0:00:32.953
bringing another industrial revolution.

0:00:33.627,0:00:39.504
Previous industrial revolutions[br]expanded human's mechanical power.

0:00:40.014,0:00:45.572
This new revolution,[br]this second machine age

0:00:45.572,0:00:50.122
is going to expand[br]our cognitive abilities,

0:00:50.122,0:00:52.102
our mental power.

0:00:52.782,0:00:57.177
Computers are not just going[br]to replace manual labor,

0:00:57.597,0:00:59.897
but also mental labor.

0:01:00.500,0:01:03.450
So, where do we stand today?

0:01:04.034,0:01:07.724
You may have heard[br]about what happened last March

0:01:07.724,0:01:11.776
when a machine learning system[br]called AlphaGo

0:01:11.776,0:01:17.708
used deep learning to beat[br]the world champion at the game of Go.

0:01:18.279,0:01:20.679
Go is an ancient Chinese game

0:01:20.679,0:01:24.159
which had been much more difficult[br]for computers to master

0:01:24.159,0:01:25.982
than the game of chess.

0:01:26.893,0:01:32.086
How did we succeed,[br]now, after decades of AI research?

0:01:33.068,0:01:36.698
AlphaGo was trained to play Go.

0:01:37.678,0:01:41.300
First, by watching over and over

0:01:41.814,0:01:46.894
tens of millions of moves made[br]by very strong human players.

0:01:47.746,0:01:52.496
Then, by playing against itself,[br]millions of games.

0:01:54.222,0:01:59.941
Machine Learning allows computers[br]to learn from examples.

0:02:00.465,0:02:02.575
To learn from data.

0:02:03.885,0:02:07.235
Machine learning[br]has turned out to be a key

0:02:07.235,0:02:11.635
to cram knowledge into computers.

0:02:12.174,0:02:14.066
And this is important

0:02:14.066,0:02:19.296
because knowledge[br]is what enables intelligence.

0:02:20.438,0:02:26.768
Putting knowledge into computers had been[br]a challenge for previous approaches to AI.

0:02:27.515,0:02:28.745
Why?

0:02:29.059,0:02:33.859
There are many things[br]which we know intuitively.

0:02:34.601,0:02:38.081
So we cannot communicate them verbally.

0:02:38.619,0:02:42.780
We do not have conscious access[br]to that intuitive knowledge.

0:02:43.270,0:02:46.690
How can we program computers[br]without knowledge?

0:02:47.664,0:02:49.114
What's the solution?

0:02:49.314,0:02:55.343
The solution is for machines to learn[br]that knowledge by themselves,

0:02:55.343,0:02:56.443
just as we do.

0:02:56.443,0:03:03.194
And this is important because knowledge[br]is what enables intelligence.

0:03:03.194,0:03:06.974
My mission has been[br]to contribute to discover

0:03:06.974,0:03:12.676
and understand principles[br]of intelligence through learning.

0:03:13.166,0:03:18.116
Whether animal, human or machine learning.

0:03:19.450,0:03:25.066
I and others believe that there are[br]a few key principles,

0:03:25.066,0:03:27.296
just like the law of physics.

0:03:27.885,0:03:32.745
Simple principles which could explain[br]our own intelligence

0:03:32.745,0:03:36.741
and help us build intelligent machines.

0:03:37.885,0:03:41.595
For example, think about the laws[br]of aerodynamics

0:03:41.595,0:03:48.036
which are general enough to explain[br]the flight of both, birds and planes.

0:03:49.146,0:03:55.381
Wouldn't it be amazing to discover[br]such simple but powerful principles

0:03:55.381,0:03:59.186
that would explain intelligence itself?

0:04:00.026,0:04:03.394
Well, we've made some progress.

0:04:04.384,0:04:10.857
My collaborators and I have contributed[br]in recent years in a revolution in AI

0:04:11.777,0:04:16.397
with our research on neural networks[br]and deep learning,

0:04:16.397,0:04:20.946
an approach to machine learning[br]which is inspired by the brain.

0:04:22.041,0:04:25.243
It started with speech recognition

0:04:25.243,0:04:29.963
on your phones,[br]with neural networks since 2012.

0:04:30.977,0:04:35.647
Shortly after, came a breakthrough[br]in computer vision.

0:04:36.680,0:04:43.087
Computers can now do a pretty good job[br]of recognizing the content of images.

0:04:43.674,0:04:50.049
In fact, they approach human performance[br]on some benchmarks over the last 5 years.

0:04:50.711,0:04:54.721
A computer can now get[br]an intuitive understanding

0:04:54.721,0:04:58.191
of the visual appearance of a Go-board

0:04:58.191,0:05:01.763
that is comparable to that[br]of the best human players.

0:05:01.763,0:05:03.454
More recently,

0:05:03.454,0:05:06.584
following some discoveries made in my lab,

0:05:06.584,0:05:11.408
deep learning has been used to translate[br]from one language to another

0:05:11.414,0:05:14.441
and you are going to start seeing[br]this in Google translate.

0:05:15.191,0:05:18.192
This is expanding the computer's ability

0:05:18.192,0:05:22.532
to understand and generate[br]natural language.

0:05:23.550,0:05:25.517
But don't be fooled.

0:05:25.517,0:05:30.048
We are still very, very far from a machine

0:05:30.048,0:05:34.033
that would be as able as humans

0:05:34.033,0:05:37.593
to learn to master[br]many aspects of our world.

0:05:38.541,0:05:41.237
So, let's take an example.

0:05:41.637,0:05:46.787
Even a two year old child[br]is able to learn things

0:05:46.787,0:05:50.657
in a way that computers[br]are not able to do right now.

0:05:51.767,0:05:56.169
A two year old child actually[br]masters intuitive physics.

0:05:56.968,0:06:01.908
She knows when she drops a ball[br]that it is going to fall down.

0:06:02.493,0:06:06.093
When she spills some liquids[br]she expects the resulting mess.

0:06:06.586,0:06:09.516
Her parents do not need to teach her

0:06:09.516,0:06:12.980
about Newton's laws[br]or differential equations.

0:06:13.840,0:06:20.200
She discovers all these things by herself[br]in a unsupervised way.

0:06:21.352,0:06:27.712
Unsupervised learning actually remains[br]one of the key challenges for AI.

0:06:28.184,0:06:33.014
And it may take several more decades[br]of fundamental research

0:06:33.014,0:06:34.674
to crack that knot.

0:06:34.674,0:06:40.895
Unsupervised learning is actually trying[br]to discover representations of the data.

0:06:41.729,0:06:43.779
Let me show you and example.

0:06:44.364,0:06:49.346
Consider a page on the screen[br]that you're seeing with your eyes

0:06:49.346,0:06:54.196
or that the computer is seeing[br]as an image, a bunch of pixels.

0:06:54.993,0:07:00.113
In order to answer a question[br]about the content of the image

0:07:00.863,0:07:05.211
you need to understand[br]its high-level meaning.

0:07:05.674,0:07:10.821
This high level meaning corresponds[br]to the highest level of representation

0:07:10.821,0:07:12.321
in your brain.

0:07:12.906,0:07:18.308
Low down, you have[br]the individual meaning of words

0:07:19.188,0:07:23.798
and even lower down, you have characters[br]which make up the words.

0:07:24.810,0:07:27.677
Those characters could be[br]rendered in different ways

0:07:27.677,0:07:30.879
with different strokes[br]that make up the characters.

0:07:31.559,0:07:34.839
And those strokes are made up of edges[br]

0:07:34.839,0:07:37.284
and those edges are made up of pixels.

0:07:37.284,0:07:40.454
So these are different[br]levels of representation.

0:07:41.079,0:07:44.236
But the pixels are not[br]sufficient by themselves

0:07:44.236,0:07:46.584
to make sense of the image,

0:07:46.584,0:07:51.904
to answer a high level question[br]about the content of the page.

0:07:52.932,0:07:57.594
Your brain actually has[br]these different levels of representation

0:07:57.594,0:08:02.291
starting with neurons[br]in the first visual area of cortex - V1,

0:08:02.291,0:08:04.596
which recognizes edges.

0:08:04.596,0:08:09.334
And then, neurons in the second[br]visual area of cortex - V2,

0:08:09.334,0:08:12.800
which recognizes strokes and small shapes.

0:08:12.800,0:08:17.060
Higher up, you have neurons[br]which detect parts of objects

0:08:17.060,0:08:19.992
and then objects and full scenes.

0:08:21.182,0:08:24.757
Neural networks,[br]when they're trained with images,

0:08:24.757,0:08:28.860
can actually discover these types[br]of levels of representation

0:08:28.860,0:08:32.778
that match pretty well[br]what we observe in the brain.

0:08:33.638,0:08:38.798
Both, biological neural networks,[br]which are what you have in your brain

0:08:38.804,0:08:42.828
and the deep neural networks[br]that we train on our machines

0:08:42.845,0:08:48.075
can learn to transform from one level[br]of representation to the next,

0:08:48.369,0:08:53.299
with the high levels corresponding[br]to more abstract notions.

0:08:53.299,0:08:57.562
For example the abstract notion[br]of the character A

0:08:57.562,0:09:00.891
can be rendered in many different ways[br]at the lowest levels

0:09:00.891,0:09:03.887
as many different configurations of pixels

0:09:03.887,0:09:09.097
depending on the position,[br]rotation, font and so on.

0:09:10.445,0:09:15.815
So, how do we learn[br]these high levels of representations?

0:09:16.962,0:09:20.681
One thing that has been[br]very successful up to now

0:09:20.681,0:09:22.853
in the applications of deep learning,

0:09:22.855,0:09:25.985
is what we call supervised learning.

0:09:26.297,0:09:31.588
With supervised learning, the computer[br]needs to be taken by the hand

0:09:31.594,0:09:35.467
and humans have to tell the computer[br]the answer to many questions.

0:09:35.467,0:09:41.420
For example, on millions and millions[br]of images, humans have to tell the machine

0:09:41.420,0:09:44.271
well... for this image, it is a cat.

0:09:44.273,0:09:47.095
For this image, it is a dog.

0:09:47.095,0:09:49.585
For this image, it is a laptop.

0:09:49.605,0:09:55.595
For this image, it is a keyboard,[br]And so on, and so on millions of times.

0:09:56.066,0:10:01.026
This is very painful and we use[br]crowdsourcing to manage to do that.

0:10:01.461,0:10:03.396
Although, this is very powerful

0:10:03.416,0:10:06.269
and we are able to solve[br]many interesting problems,

0:10:06.269,0:10:08.313
humans are much stronger

0:10:08.313,0:10:12.076
and they can learn over many more[br]different aspects of the world

0:10:12.076,0:10:13.809
in a much more autonomous way,

0:10:13.809,0:10:17.609
just as we've seen with the child[br]learning about intuitive physics.

0:10:17.625,0:10:23.739
Unsupervised learning could also help us[br]deal with self-driving cars.

0:10:24.567,0:10:26.097
Let me explain what I mean:

0:10:26.097,0:10:31.835
Unsupervised learning allows computers[br]to project themselves into the future

0:10:31.835,0:10:37.205
to generate plausible futures[br]conditioned on the current situation.

0:10:38.369,0:10:42.899
And that allows computers to reason[br]and to plan ahead.

0:10:43.450,0:10:47.985
Even for circumstances[br]they have not been trained on.

0:10:48.751,0:10:50.441
This is important

0:10:50.441,0:10:53.951
because if we use supervised learning[br]we would have to tell the computers

0:10:53.951,0:10:57.395
about all the circumstances[br]where the car could be

0:10:57.395,0:11:01.375
and how humans[br]would react in that situation.

0:11:02.451,0:11:06.191
How did I learn to avoid[br]dangerous driving behavior?

0:11:07.276,0:11:10.791
Did I have to die[br]a thousand times in an accident?

0:11:10.793,0:11:12.106
(Laughter)

0:11:12.106,0:11:14.606
Well, that's the way we train[br]machines right now.

0:11:15.175,0:11:18.340
So, it's not going to fly[br]or at least not to drive.

0:11:18.340,0:11:19.928
(Laughter)

0:11:21.288,0:11:25.657
So, what we need is to train our models

0:11:25.657,0:11:31.294
to be able to generate plausible images[br]or plausible futures, be creative.

0:11:31.294,0:11:33.934
And we are making progress with that.

0:11:33.934,0:11:37.457
So, we're training[br]these deep neural networks

0:11:37.463,0:11:40.818
to go from high-level meaning to pixels

0:11:40.818,0:11:43.298
rather than from pixels[br]to high level meaning,

0:11:43.307,0:11:46.787
going into the other direction[br]through the levels of representation.

0:11:46.787,0:11:50.461
And this way, the computer[br]can generate images

0:11:51.191,0:11:55.072
that are new images different[br]from what the computer has seen

0:11:55.072,0:11:56.488
while it was trained,

0:11:57.018,0:12:00.369
but are plausible and look like natural images.

0:12:01.888,0:12:06.332
We can also use these models[br]to dream up strange,

0:12:06.342,0:12:09.492
sometimes scary images,

0:12:09.492,0:12:11.795
just like our dreams and nightmares.

0:12:12.682,0:12:16.847
Here's some images[br]that were synthesized by the computer

0:12:16.847,0:12:19.826
using these deep charted models.

0:12:19.826,0:12:21.651
They look like natural images

0:12:21.651,0:12:24.551
but if you look closely,[br]you will see they are different

0:12:25.458,0:12:28.697
and they're still missing[br]some of the important details

0:12:28.697,0:12:31.063
that we would recognize as natural.

0:12:31.995,0:12:33.951
About 10 years ago,

0:12:33.951,0:12:38.921
unsupervised learning has been[br]a key to the breakthrough

0:12:38.921,0:12:42.443
that we obtained[br]discovering deep learning.

0:12:44.140,0:12:48.055
This was happening in just few labs,[br]including mine at the time

0:12:48.055,0:12:51.455
at a time when neural networks[br]were not popular.

0:12:51.455,0:12:55.217
They were almost abandoned[br]by the scientific community.

0:12:56.394,0:12:58.935
Now, things have changed a lot.

0:12:58.935,0:13:01.375
It has become a very hot field.

0:13:01.384,0:13:06.933
There are now hundreds of students[br]every year applying for graduate studies

0:13:06.954,0:13:09.784
at my lab with my collaborators.

0:13:11.010,0:13:16.630
Montreal has become[br]the largest academic concentration

0:13:16.637,0:13:19.387
of deep learning researchers in the world.

0:13:20.182,0:13:26.115
We just received a huge[br]research grant of 94 million dollars

0:13:26.127,0:13:29.797
to push the boundaries[br]of AI and data science

0:13:29.797,0:13:36.067
and also to transfer technology of deep[br]learning and data science to the industry.

0:13:37.249,0:13:43.791
Business people stimulated by all this[br]are creating start-ups, industrial labs,

0:13:43.791,0:13:46.914
many of which near the universities.

0:13:48.543,0:13:49.625
For example,

0:13:49.625,0:13:54.733
just a few weeks ago, we announced[br]the launch of a start-up factory

0:13:54.733,0:13:56.507
called 'Element AI'

0:13:56.507,0:13:59.605
which is going to focus[br]on the deep learning applications.

0:14:01.562,0:14:05.722
There are just not enough[br]deep learning experts.

0:14:06.355,0:14:10.677
So, they are getting paid crazy salaries,

0:14:11.027,0:14:17.212
and many of my former academic colleagues[br]have accepted generous deals

0:14:17.228,0:14:20.518
from companies to work in industrial labs.

0:14:21.081,0:14:25.010
I, for myself, have chosen[br]to stay in university,

0:14:25.010,0:14:27.166
to work for the public good,

0:14:27.166,0:14:28.886
to work with students,

0:14:28.902,0:14:30.592
to remain independent.

0:14:30.596,0:14:34.836
To guide the next generation[br]of deep learning experts.

0:14:35.294,0:14:41.024
One thing that we are doing[br]beyond commercial value

0:14:41.024,0:14:44.654
is thinking about the social[br]implications of AI.

0:14:45.881,0:14:50.026
Many of us are now starting[br]to turn our eyes

0:14:50.026,0:14:55.986
towards social value added[br]applications, like health.

0:14:56.457,0:14:58.956
We think that we can use deep learning

0:14:58.956,0:15:02.696
to improve treatment[br]with personalized medicine.

0:15:03.956,0:15:05.671
I believe that in the future,

0:15:05.671,0:15:10.361
as we collect more data from millions[br]and billions people around the earth,

0:15:10.361,0:15:13.856
we will be able to provide medical advice

0:15:13.856,0:15:17.246
to billions of people[br]who don't have access to it right now.

0:15:17.601,0:15:22.724
And we can imagine many other[br]applications for social value of AI.

0:15:23.140,0:15:26.238
For example, something[br]that will come out of our research

0:15:26.238,0:15:28.582
on natural language understanding

0:15:29.328,0:15:31.199
is providing all kinds of services

0:15:31.199,0:15:34.059
like legal services,[br]to those who can't afford them.

0:15:34.512,0:15:37.342
We are now turning our eyes

0:15:37.342,0:15:41.132
also towards the social implications[br]of AI in my community.

0:15:41.690,0:15:44.801
But it's not just for experts[br]to think about this.

0:15:46.026,0:15:49.936
I believe that beyond the math[br]and the jargon,

0:15:49.936,0:15:53.102
ordinary people can get the sense

0:15:53.138,0:15:55.901
of what goes on under the hood

0:15:55.901,0:16:01.191
enough to participate[br]in the important decisions

0:16:01.191,0:16:06.547
that will take place, in the next[br]few years and decades about AI.

0:16:07.580,0:16:09.280
So please,

0:16:09.930,0:16:16.230
set aside your fees and give yourself[br]some space to learn about it.

0:16:17.842,0:16:22.532
My collaborators and I have written[br]several introductory papers

0:16:22.542,0:16:25.376
and a book entitled "Deep Learning"

0:16:25.376,0:16:29.619
to help students and engineers[br]jump into this exciting field.

0:16:30.659,0:16:35.784
There are also many online resources:[br]softwares, tutorials, videos..

0:16:36.310,0:16:41.210
and many undergraduate students[br]are learning a lot of this

0:16:41.210,0:16:44.548
about research in deep learning[br]by themselves,

0:16:44.548,0:16:47.835
to later join the ranks of labs like mine.

0:16:49.370,0:16:55.170
Ai is going to have a profound[br]impact on our society.

0:16:56.652,0:17:01.672
So, it's important to ask:[br]How are we going to use it?

0:17:03.368,0:17:07.896
Immense positives may come[br]along with negatives

0:17:07.896,0:17:10.166
such as military use

0:17:10.797,0:17:15.357
or rapid disruptive changes[br]in the job market.

0:17:15.948,0:17:21.629
To make sure the collective choices[br]that will be made about AI

0:17:21.629,0:17:23.074
in the next few years,

0:17:23.074,0:17:25.144
will be for the benefit of all,

0:17:25.144,0:17:28.557
every citizen should take an active role

0:17:28.557,0:17:32.911
in defining how AI will shape our future.

0:17:33.871,0:17:34.891
Thank you.

0:17:35.065,0:17:39.395
(Applause)