00:00:12
It used to be that if you wanted
to get a computer to do something new,
00:00:16
you would have to program it.
00:00:18
Now, programming, for those of you here
that haven't done it yourself,
00:00:21
requires laying out in excruciating detail
00:00:25
every single step that you want
the computer to do
00:00:28
in order to achieve your goal.
00:00:31
Now, if you want to do something
that you don't know how to do yourself,
00:00:34
then this is going
to be a great challenge.
00:00:36
So this was the challenge faced
by this man, Arthur Samuel.
00:00:40
In 1956, he wanted to get this computer
00:00:44
to be able to beat him at checkers.
00:00:46
How can you write a program,
00:00:48
lay out in excruciating detail,
how to be better than you at checkers?
00:00:52
So he came up with an idea:
00:00:54
he had the computer play
against itself thousands of times
00:00:57
and learn how to play checkers.
00:01:00
And indeed it worked,
and in fact, by 1962,
00:01:03
this computer had beaten
the Connecticut state champion.
00:01:07
So Arthur Samuel was
the father of machine learning,
00:01:10
and I have a great debt to him,
00:01:12
because I am a machine
learning practitioner.
00:01:15
I was the president of Kaggle,
00:01:16
a community of over 200,000
machine learning practictioners.
00:01:19
Kaggle puts up competitions
00:01:21
to try and get them to solve
previously unsolved problems,
00:01:25
and it's been successful
hundreds of times.
00:01:29
So from this vantage point,
I was able to find out
00:01:31
a lot about what machine learning
can do in the past, can do today,
00:01:35
and what it could do in the future.
00:01:38
Perhaps the first big success of
machine learning commercially was Google.
00:01:42
Google showed that it is
possible to find information
00:01:45
by using a computer algorithm,
00:01:47
and this algorithm is based
on machine learning.
00:01:50
Since that time, there have been many
commercial successes of machine learning.
00:01:54
Companies like Amazon and Netflix
00:01:56
use machine learning to suggest
products that you might like to buy,
00:01:59
movies that you might like to watch.
00:02:01
Sometimes, it's almost creepy.
00:02:03
Companies like LinkedIn and Facebook
00:02:05
sometimes will tell you about
who your friends might be
00:02:08
and you have no idea how it did it,
00:02:10
and this is because it's using
the power of machine learning.
00:02:13
These are algorithms that have
learned how to do this from data
00:02:16
rather than being programmed by hand.
00:02:19
This is also how IBM was successful
00:02:21
in getting Watson to beat
the two world champions at "Jeopardy,"
00:02:25
answering incredibly subtle
and complex questions like this one.
00:02:28
["The ancient 'Lion of Nimrud' went missing
from this city's national museum in 2003
(along with a lot of other stuff)"]
00:02:31
This is also why we are now able
to see the first self-driving cars.
00:02:35
If you want to be able to tell
the difference between, say,
00:02:37
a tree and a pedestrian,
well, that's pretty important.
00:02:40
We don't know how to write
those programs by hand,
00:02:43
but with machine learning,
this is now possible.
00:02:46
And in fact, this car has driven
over a million miles
00:02:48
without any accidents on regular roads.
00:02:52
So we now know that computers can learn,
00:02:56
and computers can learn to do things
00:02:58
that we actually sometimes
don't know how to do ourselves,
00:03:00
or maybe can do them better than us.
00:03:03
One of the most amazing examples
I've seen of machine learning
00:03:07
happened on a project that I ran at Kaggle
00:03:10
where a team run by a guy
called Geoffrey Hinton
00:03:13
from the University of Toronto
00:03:15
won a competition for
automatic drug discovery.
00:03:18
Now, what was extraordinary here
is not just that they beat
00:03:20
all of the algorithms developed by Merck
or the international academic community,
00:03:25
but nobody on the team had any background
in chemistry or biology or life sciences,
00:03:30
and they did it in two weeks.
00:03:32
How did they do this?
00:03:34
They used an extraordinary algorithm
called deep learning.
00:03:37
So important was this that in fact
the success was covered
00:03:40
in The New York Times in a front page
article a few weeks later.
00:03:43
This is Geoffrey Hinton
here on the left-hand side.
00:03:46
Deep learning is an algorithm
inspired by how the human brain works,
00:03:50
and as a result it's an algorithm
00:03:52
which has no theoretical limitations
on what it can do.
00:03:56
The more data you give it and the more
computation time you give it,
00:03:58
the better it gets.
00:04:00
The New York Times also
showed in this article
00:04:02
another extraordinary
result of deep learning
00:04:04
which I'm going to show you now.
00:04:07
It shows that computers
can listen and understand.
00:04:12
(Video) Richard Rashid: Now, the last step
00:04:15
that I want to be able
to take in this process
00:04:18
is to actually speak to you in Chinese.
00:04:22
Now the key thing there is,
00:04:25
we've been able to take a large amount
of information from many Chinese speakers
00:04:30
and produce a text-to-speech system
00:04:33
that takes Chinese text
and converts it into Chinese language,
00:04:37
and then we've taken
an hour or so of my own voice
00:04:41
and we've used that to modulate
00:04:43
the standard text-to-speech system
so that it would sound like me.
00:04:48
Again, the result's not perfect.
00:04:50
There are in fact quite a few errors.
00:04:53
(In Chinese)
00:04:56
(Applause)
00:05:01
There's much work to be done in this area.
00:05:05
(In Chinese)
00:05:08
(Applause)
00:05:13
Jeremy Howard: Well, that was at
a machine learning conference in China.
00:05:16
It's not often, actually,
at academic conferences
00:05:19
that you do hear spontaneous applause,
00:05:21
although of course sometimes
at TEDx conferences, feel free.
00:05:24
Everything you saw there
was happening with deep learning.
00:05:27
(Applause) Thank you.
00:05:29
The transcription in English
was deep learning.
00:05:31
The translation to Chinese and the text
in the top right, deep learning,
00:05:34
and the construction of the voice
was deep learning as well.
00:05:38
So deep learning is
this extraordinary thing.
00:05:41
It's a single algorithm that
can seem to do almost anything,
00:05:44
and I discovered that a year earlier,
it had also learned to see.
00:05:47
In this obscure competition from Germany
00:05:49
called the German Traffic Sign
Recognition Benchmark,
00:05:52
deep learning had learned
to recognize traffic signs like this one.
00:05:55
Not only could it
recognize the traffic signs
00:05:57
better than any other algorithm,
00:05:59
the leaderboard actually showed
it was better than people,
00:06:02
about twice as good as people.
00:06:04
So by 2011, we had the first example
00:06:06
of computers that can see
better than people.
00:06:09
Since that time, a lot has happened.
00:06:11
In 2012, Google announced that
they had a deep learning algorithm
00:06:15
watch YouTube videos
00:06:16
and crunched the data
on 16,000 computers for a month,
00:06:19
and the computer independently learned
about concepts such as people and cats
00:06:24
just by watching the videos.
00:06:26
This is much like the way
that humans learn.
00:06:28
Humans don't learn
by being told what they see,
00:06:31
but by learning for themselves
what these things are.
00:06:34
Also in 2012, Geoffrey Hinton,
who we saw earlier,
00:06:37
won the very popular ImageNet competition,
00:06:40
looking to try to figure out
from one and a half million images
00:06:44
what they're pictures of.
00:06:46
As of 2014, we're now down
to a six percent error rate
00:06:49
in image recognition.
00:06:51
This is better than people, again.
00:06:53
So machines really are doing
an extraordinarily good job of this,
00:06:57
and it is now being used in industry.
00:06:59
For example, Google announced last year
00:07:02
that they had mapped every single
location in France in two hours,
00:07:06
and the way they did it was
that they fed street view images
00:07:10
into a deep learning algorithm
to recognize and read street numbers.
00:07:14
Imagine how long
it would have taken before:
00:07:16
dozens of people, many years.
00:07:20
This is also happening in China.
00:07:22
Baidu is kind of
the Chinese Google, I guess,
00:07:26
and what you see here in the top left
00:07:28
is an example of a picture that I uploaded
to Baidu's deep learning system,
00:07:32
and underneath you can see that the system
has understood what that picture is
00:07:36
and found similar images.
00:07:38
The similar images actually
have similar backgrounds,
00:07:41
similar directions of the faces,
00:07:42
even some with their tongue out.
00:07:44
This is not clearly looking
at the text of a web page.
00:07:47
All I uploaded was an image.
00:07:49
So we now have computers which
really understand what they see
00:07:53
and can therefore search databases
00:07:54
of hundreds of millions
of images in real time.
00:07:58
So what does it mean
now that computers can see?
00:08:01
Well, it's not just
that computers can see.
00:08:03
In fact, deep learning
has done more than that.
00:08:05
Complex, nuanced sentences like this one
00:08:08
are now understandable
with deep learning algorithms.
00:08:11
As you can see here,
00:08:12
this Stanford-based system
showing the red dot at the top
00:08:15
has figured out that this sentence
is expressing negative sentiment.
00:08:19
Deep learning now in fact
is near human performance
00:08:22
at understanding what sentences are about
and what it is saying about those things.
00:08:27
Also, deep learning has
been used to read Chinese,
00:08:30
again at about native
Chinese speaker level.
00:08:33
This algorithm developed
out of Switzerland
00:08:35
by people, none of whom speak
or understand any Chinese.
00:08:39
As I say, using deep learning
00:08:41
is about the best system
in the world for this,
00:08:43
even compared to native
human understanding.
00:08:48
This is a system that we
put together at my company
00:08:51
which shows putting
all this stuff together.
00:08:53
These are pictures which
have no text attached,
00:08:56
and as I'm typing in here sentences,
00:08:58
in real time it's understanding
these pictures
00:09:01
and figuring out what they're about
00:09:03
and finding pictures that are similar
to the text that I'm writing.
00:09:06
So you can see, it's actually
understanding my sentences
00:09:09
and actually understanding these pictures.
00:09:11
I know that you've seen
something like this on Google,
00:09:13
where you can type in things
and it will show you pictures,
00:09:16
but actually what it's doing is it's
searching the webpage for the text.
00:09:20
This is very different from actually
understanding the images.
00:09:23
This is something that computers
have only been able to do
00:09:25
for the first time in the last few months.
00:09:29
So we can see now that computers
can not only see but they can also read,
00:09:33
and, of course, we've shown that they
can understand what they hear.
00:09:36
Perhaps not surprising now that
I'm going to tell you they can write.
00:09:40
Here is some text that I generated
using a deep learning algorithm yesterday.
00:09:45
And here is some text that an algorithm
out of Stanford generated.
00:09:49
Each of these sentences was generated
00:09:50
by a deep learning algorithm
to describe each of those pictures.
00:09:55
This algorithm before has never seen
a man in a black shirt playing a guitar.
00:09:59
It's seen a man before,
it's seen black before,
00:10:01
it's seen a guitar before,
00:10:03
but it has independently generated
this novel description of this picture.
00:10:07
We're still not quite at human
performance here, but we're close.
00:10:11
In tests, humans prefer
the computer-generated caption
00:10:15
one out of four times.
00:10:16
Now this system is now only two weeks old,
00:10:18
so probably within the next year,
00:10:20
the computer algorithm will be
well past human performance
00:10:23
at the rate things are going.
00:10:25
So computers can also write.
00:10:28
So we put all this together and it leads
to very exciting opportunities.
00:10:31
For example, in medicine,
00:10:33
a team in Boston announced
that they had discovered
00:10:35
dozens of new clinically relevant features
00:10:38
of tumors which help doctors
make a prognosis of a cancer.
00:10:44
Very similarly, in Stanford,
00:10:46
a group there announced that,
looking at tissues under magnification,
00:10:50
they've developed
a machine learning-based system
00:10:52
which in fact is better
than human pathologists
00:10:55
at predicting survival rates
for cancer sufferers.
00:10:59
In both of these cases, not only
were the predictions more accurate,
00:11:02
but they generated new insightful science.
00:11:05
In the radiology case,
00:11:06
they were new clinical indicators
that humans can understand.
00:11:09
In this pathology case,
00:11:11
the computer system actually discovered
that the cells around the cancer
00:11:16
are as important as
the cancer cells themselves
00:11:19
in making a diagnosis.
00:11:21
This is the opposite of what pathologists
had been taught for decades.
00:11:26
In each of those two cases,
they were systems developed
00:11:29
by a combination of medical experts
and machine learning experts,
00:11:33
but as of last year,
we're now beyond that too.
00:11:36
This is an example of
identifying cancerous areas
00:11:39
of human tissue under a microscope.
00:11:42
The system being shown here
can identify those areas more accurately,
00:11:46
or about as accurately,
as human pathologists,
00:11:49
but was built entirely with deep learning
using no medical expertise
00:11:53
by people who have
no background in the field.
00:11:56
Similarly, here, this neuron segmentation.
00:11:59
We can now segment neurons
about as accurately as humans can,
00:12:02
but this system was developed
with deep learning
00:12:05
using people with no previous
background in medicine.
00:12:08
So myself, as somebody with
no previous background in medicine,
00:12:12
I seem to be entirely well qualified
to start a new medical company,
00:12:15
which I did.
00:12:18
I was kind of terrified of doing it,
00:12:19
but the theory seemed to suggest
that it ought to be possible
00:12:22
to do very useful medicine
using just these data analytic techniques.
00:12:28
And thankfully, the feedback
has been fantastic,
00:12:30
not just from the media
but from the medical community,
00:12:32
who have been very supportive.
00:12:35
The theory is that we can take
the middle part of the medical process
00:12:39
and turn that into data analysis
as much as possible,
00:12:42
leaving doctors to do
what they're best at.
00:12:45
I want to give you an example.
00:12:47
It now takes us about 15 minutes
to generate a new medical diagnostic test
00:12:51
and I'll show you that in real time now,
00:12:53
but I've compressed it down to
three minutes by cutting some pieces out.
00:12:57
Rather than showing you
creating a medical diagnostic test,
00:13:00
I'm going to show you
a diagnostic test of car images,
00:13:03
because that's something
we can all understand.
00:13:06
So here we're starting with
about 1.5 million car images,
00:13:09
and I want to create something
that can split them into the angle
00:13:12
of the photo that's being taken.
00:13:14
So these images are entirely unlabeled,
so I have to start from scratch.
00:13:18
With our deep learning algorithm,
00:13:20
it can automatically identify
areas of structure in these images.
00:13:24
So the nice thing is that the human
and the computer can now work together.
00:13:27
So the human, as you can see here,
00:13:29
is telling the computer
about areas of interest
00:13:32
which it wants the computer then
to try and use to improve its algorithm.
00:13:37
Now, these deep learning systems actually
are in 16,000-dimensional space,
00:13:41
so you can see here the computer
rotating this through that space,
00:13:45
trying to find new areas of structure.
00:13:47
And when it does so successfully,
00:13:48
the human who is driving it can then
point out the areas that are interesting.
00:13:52
So here, the computer has
successfully found areas,
00:13:55
for example, angles.
00:13:57
So as we go through this process,
00:13:59
we're gradually telling
the computer more and more
00:14:01
about the kinds of structures
we're looking for.
00:14:04
You can imagine in a diagnostic test
00:14:05
this would be a pathologist identifying
areas of pathosis, for example,
00:14:09
or a radiologist indicating
potentially troublesome nodules.
00:14:14
And sometimes it can be
difficult for the algorithm.
00:14:16
In this case, it got kind of confused.
00:14:18
The fronts and the backs
of the cars are all mixed up.
00:14:21
So here we have to be a bit more careful,
00:14:23
manually selecting these fronts
as opposed to the backs,
00:14:26
then telling the computer
that this is a type of group
00:14:32
that we're interested in.
00:14:33
So we do that for a while,
we skip over a little bit,
00:14:36
and then we train the
machine learning algorithm
00:14:38
based on these couple of hundred things,
00:14:40
and we hope that it's gotten a lot better.
00:14:42
You can see, it's now started to fade
some of these pictures out,
00:14:45
showing us that it already is recognizing
how to understand some of these itself.
00:14:50
We can then use this concept
of similar images,
00:14:53
and using similar images, you can now see,
00:14:55
the computer at this point is able to
entirely find just the fronts of cars.
00:14:59
So at this point, the human
can tell the computer,
00:15:02
okay, yes, you've done
a good job of that.
00:15:05
Sometimes, of course, even at this point
00:15:07
it's still difficult
to separate out groups.
00:15:11
In this case, even after we let the
computer try to rotate this for a while,
00:15:15
we still find that the left sides
and the right sides pictures
00:15:18
are all mixed up together.
00:15:20
So we can again give
the computer some hints,
00:15:22
and we say, okay, try and find
a projection that separates out
00:15:25
the left sides and the right sides
as much as possible
00:15:27
using this deep learning algorithm.
00:15:30
And giving it that hint --
ah, okay, it's been successful.
00:15:33
It's managed to find a way
of thinking about these objects
00:15:35
that's separated out these together.
00:15:38
So you get the idea here.
00:15:40
This is a case not where the human
is being replaced by a computer,
00:15:48
but where they're working together.
00:15:51
What we're doing here is we're replacing
something that used to take a team
00:15:55
of five or six people about seven years
00:15:57
and replacing it with something
that takes 15 minutes
00:15:59
for one person acting alone.
00:16:02
So this process takes about
four or five iterations.
00:16:06
You can see we now have 62 percent
00:16:08
of our 1.5 million images
classified correctly.
00:16:10
And at this point, we
can start to quite quickly
00:16:13
grab whole big sections,
00:16:14
check through them to make sure
that there's no mistakes.
00:16:17
Where there are mistakes, we can
let the computer know about them.
00:16:21
And using this kind of process
for each of the different groups,
00:16:24
we are now up to
an 80 percent success rate
00:16:27
in classifying the 1.5 million images.
00:16:29
And at this point, it's just a case
00:16:31
of finding the small number
that aren't classified correctly,
00:16:35
and trying to understand why.
00:16:38
And using that approach,
00:16:39
by 15 minutes we get
to 97 percent classification rates.
00:16:43
So this kind of technique
could allow us to fix a major problem,
00:16:48
which is that there's a lack
of medical expertise in the world.
00:16:51
The World Economic Forum says
that there's between a 10x and a 20x
00:16:55
shortage of physicians
in the developing world,
00:16:57
and it would take about 300 years
00:16:59
to train enough people
to fix that problem.
00:17:02
So imagine if we can help
enhance their efficiency
00:17:05
using these deep learning approaches?
00:17:08
So I'm very excited
about the opportunities.
00:17:10
I'm also concerned about the problems.
00:17:13
The problem here is that
every area in blue on this map
00:17:16
is somewhere where services
are over 80 percent of employment.
00:17:20
What are services?
00:17:21
These are services.
00:17:23
These are also the exact things that
computers have just learned how to do.
00:17:27
So 80 percent of the world's employment
in the developed world
00:17:31
is stuff that computers
have just learned how to do.
00:17:33
What does that mean?
00:17:35
Well, it'll be fine.
They'll be replaced by other jobs.
00:17:37
For example, there will be
more jobs for data scientists.
00:17:40
Well, not really.
00:17:41
It doesn't take data scientists
very long to build these things.
00:17:44
For example, these four algorithms
were all built by the same guy.
00:17:47
So if you think, oh,
it's all happened before,
00:17:50
we've seen the results in the past
of when new things come along
00:17:54
and they get replaced by new jobs,
00:17:56
what are these new jobs going to be?
00:17:58
It's very hard for us to estimate this,
00:18:00
because human performance
grows at this gradual rate,
00:18:03
but we now have a system, deep learning,
00:18:05
that we know actually grows
in capability exponentially.
00:18:08
And we're here.
00:18:10
So currently, we see the things around us
00:18:12
and we say, "Oh, computers
are still pretty dumb." Right?
00:18:15
But in five years' time,
computers will be off this chart.
00:18:18
So we need to be starting to think
about this capability right now.
00:18:22
We have seen this once before, of course.
00:18:24
In the Industrial Revolution,
00:18:25
we saw a step change
in capability thanks to engines.
00:18:29
The thing is, though,
that after a while, things flattened out.
00:18:32
There was social disruption,
00:18:34
but once engines were used
to generate power in all the situations,
00:18:37
things really settled down.
00:18:40
The Machine Learning Revolution
00:18:41
is going to be very different
from the Industrial Revolution,
00:18:44
because the Machine Learning Revolution,
it never settles down.
00:18:47
The better computers get
at intellectual activities,
00:18:50
the more they can build better computers
to be better at intellectual capabilities,
00:18:54
so this is going to be a kind of change
00:18:56
that the world has actually
never experienced before,
00:18:59
so your previous understanding
of what's possible is different.
00:19:02
This is already impacting us.
00:19:04
In the last 25 years,
as capital productivity has increased,
00:19:08
labor productivity has been flat,
in fact even a little bit down.
00:19:13
So I want us to start
having this discussion now.
00:19:16
I know that when I often tell people
about this situation,
00:19:19
people can be quite dismissive.
00:19:20
Well, computers can't really think,
00:19:22
they don't emote,
they don't understand poetry,
00:19:25
we don't really understand how they work.
00:19:27
So what?
00:19:29
Computers right now can do the things
00:19:31
that humans spend most
of their time being paid to do,
00:19:33
so now's the time to start thinking
00:19:35
about how we're going to adjust our
social structures and economic structures
00:19:40
to be aware of this new reality.
00:19:41
Thank you.
00:19:43
(Applause)
