The wonderful and terrifying implications of computers that can learn

Well, it used to be that if you wanted
to get a computer to do something new,

you would have to program it.

Now, programming, for those of you here
that haven't done it yourself,

requires laying out in excruciating detail

every single step that you want
the computer to do

in order to achieve your goal.

Now, if you want to do something

that you don't know how to do yourself,

then this is going
to be a great challenge.

So this was the challenged faced
by this man, Arthur Samuel.

In 1956, he wanted to get this computer

to be able to beat him at checkers.

How can you write a program,

lay out in excruciating detail
how to be better than you at Checkers?

So he came up with an idea:

he had the computer play
against itself thousands of times

and learn how to play checkers.

And indeed it worked,
and in fact, by 1962,

this computer had beaten
the Connecticut state champion.

So Arthur Samuel was
the father of machine learning,

and I have a great debt to him,

because I am a machine
learning practitioner.

I was the president of Kaggle,

a community of over 200,000
machine learning practictioners.

Kaggle puts up competitions

to try and get them to solve
previously unsolved problems,

and it's been successful

hundreds of times.

So from this vantage point,
I was able to find out a lot

about what machine learning
can do in the past, can do today,

and what it could do in the future.

Perhaps the first big success
of machine learning commercially

was Google.

Google showed that it is
possible to find information

by using a computer algorithm,

and this algorithm is based
on machine learning.

Since that time, there has been many
commercial successes of machine learning.

Companies like Amazon and Netflix

use machine learning to suggest
products that you might like to buy,

movies that you might like to watch.

Sometimes, it's almost creepy.

Companies like LinkedIn and Facebook

sometimes will tell you about
who your friends might be

and you have no idea how it did it,

and this is because it's using
the power of machine learning.

These are algorithms that have
learned how to do this from data

rather than being programmed by hand.

This is also how IBM was successful

in getting Watson to beat
the two world champions at Jeopardy!,

answering incredibly subtle
and complex questions like this one.

This is also why we are now able
to see the first self-driving cars.

If you want to be able to tell
the difference between, say,

a tree and a pedestrian,

well, that's pretty important.

We don't know how to write
those programs by hand,

but with machine learning,
this is now possible.

And in fact, this car has driven

over a million miles without any accidents

on regular roads.

So we now know that computers can learn,

and computers can learn to do things

that we actually sometimes
don't know how to do ourselves,

or maybe can do them better than us.

One of the most amazing examples
I've seen of machine learning

happened on a project I ran at Kaggle

where a team run by a guy
called Geoffrey Hinton

from the University of Toronto

won a competition for
automatic drug discovery.

Now, what was extraordinary here
is not just that they beat

all of the algorithms developed by Merck

or the international academic community,

but nobody on the team had any background

in chemistry or biology or life sciences,

and they did it in two weeks.

How did they do this?

They used an extraordinary algorithm
called deep learning.

So important was this that in fact
the success was covered

in the New York Times in a front page
article a few weeks later.

This is Geoffrey Hinton
here on the left-hand side.

Deep learning is an algorithm
inspired by how the human brain works,

and as a result it's an algorithm

which has no theoretical limitations

on what it can do.

The more data you give it and the more
computation time you give it,

the better it gets.

The New York Times also
showed in this article

another extraordinary
result of deep learning

which I'm going to show you now.

It shows that computers can listen

and understand.

(Video) Speaker: Now, the last step

that I want to be able
to take in this process

is to actually speak to you in Chinese.

Now the key thing there is,

we've been able to take
a large amount of information

from many Chinese speakers

and produce a text-to-speech system

that takes Chinese text

and converts it into Chinese language,

and then we've taken
an hour or so of my own voice

and we've used that to modulate

the standard text-to-speech system

so that it would sound like me.

Again, the result's not perfect.

There are in fact quite a few errors.

(In Chinese)

(Applause)

There's much work to be done in this area.

(In Chinese)

(Applause)

Jeremy Howard: Well, that was at
a machine learning conference in China.

It's not often, actually,
at academic conferences

that you do hear spontaneous applause,

although of course sometimes
at TEDx conferences, feel free.

Everything you saw there
was happening with deep learning.

(Applause) Thank you.

The transcription in English
was deep learning.

The translation to Chinese and the text
in the top right, deep learning,

and the construction of the voice
was deep learning as well.

So deep learning is
this extraordinary thing.

It's a single algorithm that
can seem to do almost anything,

and I discovered that a year earlier,

it had also learned to see.

In this obscure competition from Germany

called the German Traffic Sign Benchmark,

deep learning had learned
to recognize traffic signs like this one.

Not only could it
recognize the traffic signs

better than any other algorithm,

the leaderboard actually showed
it was better than people,

about twice as good as people.

So by 2011, we had the first example

of computers that can see

better than people.

Since that time, a lot has happened.

In 2012, Google announced that
they had a deep learning algorithm

watch YouTube videos

and crunched the data
on 16,000 computers for a month,

and the computer independently learned
about concepts such as people and cats

just by watching the videos.

This is much like the way
that humans learn.

Humans don't learn
by being told what they see,

but by learning for themselves
what these things are.

Also in 2012, Geoffrey Hinton,

who we saw earlier,

won the very popular ImageNet competition,

looking to figure out out from
one and a half million images

what they're pictures of.

As of 2014, we're now down
to a six percent error rate

in image recognition.

This is better than people, again.

So machines really are doing
an extraordinarily good job of this,

and it is now being used in industry.

For example, Google announced last year

that they had mapped every single
location in France in two hours,

and the way they did it is
that they fed street view images

into a deep learning algorithm

to recognize and read street numbers.

Imagine how long
it would have taken before:

dozens of people, many years.

This is also happening in China.

Baidu is kind of the Chinese Google,

I guess,

and what you see here in the top left

is an example of a picture that I uploaded
to Baidu's deep learning system,

and underneath you can see that the system
has understood what that picture is

and found similar images.

The similar images actually
have similar backgrounds,

similar directions of the faces,

even some with their tongue out.

This is not clearly looking
at the text of a web page.

All I uploaded was an image.

So we now have computers which
really understand what they see

and can therefore search databases

of hundreds of millions
of images in real time.

So what does it mean
now that computers can see?

Well it's not just that computers can see.

In fact, deep learning
has done more than that.

Complex, nuanced sentences like this one

are now understandable
with deep learning algorithms.

As you can see here,

this Stanford-based system
showing the red dot at the top

has figured out that this sentence
is expressing negative sentiment.

Deep learning now in fact

is near human performance

at understanding what sentences are about

and what it is saying about those things.

Also, deep learning has
been used to read Chinese,

again at about native
Chinese speaker level.

This algorithm developed
out of Switzerland

by people, none of whom speak
or understand any Chinese.

As I say, using deep learning

is about the best system
in the world for this,

even compared to native
human understanding.

This is a system that we
put together at my company

which shows putting
all this stuff together.

These are pictures which
have no text attached,

and as I'm typing in here sentences,

in real time it's understanding
these pictures

and figuring out what they're about

and finding pictures that are similar
to the text that I'm writing.

So you can see, it's actually
understanding my sentences

and actually understanding these pictures.

I know that you've seen
something like this on Google,

where you can type in things

and it will show you pictures,

but actually what it's doing is it's
searching the webpage for the text.

This is very different from actually
understanding the images.

This is something that computers
have only been able to do

for the first time in the last few months.

So we can see now that computers
can not only see but they can also read,

and of course we've shown that they
can understand what they hear.

Perhaps not surprising, now, that
I'm going to tell you they can write.

Here is some text that I generated
using a deep learning algorithm yesterday.

And here is some text that
an algorithm out Stanford generated.

Each of these sentences was generated

by a deep learning algorithm
to describe each of those pictures.

This algorithm before has never seen

a man in a black shirt playing a guitar.

It's seen a man before,
it's seen black before,

it's seen a guitar before,

but it has independently generated
this novel description of this picture.

We're still not quite at human
performance here, but we're close.

In tests, humans prefer
the computer-generated caption

one out of four times.

Now this system is now only two weeks old,

so probably within the next year,
the computer algorithm will be

well past human performance

at the rate things are going.

So computers can also write.

So we put all this together and it leads
to very exciting opportunities.

For example, in medicine,

a team in Boston announced
that they had discovered

dozens of new clinically-relevant features

of tumors which help doctors

make a prognosis of a cancer.

Very similarly, in Stanford,

a group there announced that,
looking at tissues under magnification,

they've developed a machine
learning based system

which in fact is better
than human pathologists

at predicting survival rates
for cancer sufferers.

In both of these cases, not only
were the predictions more accurate,

but they generated new insightful science.

In the radiology case,

they were new clinical indicators
that humans can understand.

In this pathology case,

the computer system actually discovered

that the cells around the cancer

are as important as
the cancer cells themselves

in making a diagnosis.

This is the opposite of what
pathologists had been taught

for decades.

In each of those two cases,
they were systems developed

by a combination of medical experts

and machine learning experts,

but as of last year,
we're now beyond that too.

This is an example of
identifying cancerous areas

of human tissue under a microscope.

The system being shown here

can identify those areas more accurately,

or about as accurately,
as human pathologists,

but was built entirely with deep learning

using no medical expertise

by people who have
no background in the field.

Similarly, here, this neuron segmentation.

We can now segment neurons
about as accurately as humans can,

but this system was developed
with deep learning using people

with no previous background in medicine.

So myself, as somebody with
no previous background in medicine,

I seem to be entirely well qualified

to start a new medical company,

which I did.

I was kind of terrified of doing it,

but the theory seemed to suggest
that it ought to be possible

to do very useful medicine

using just these data analytic techniques.

And thankfully, the feedback
has been fantastic,

not just from the media
but from the medical community,

who have been very supportive.

The theory is that we can take
the middle part of the medical process

and turn that into data analysis
as much as possible,

leaving doctors to do
what they're best at.

I want to give you an example.

It now takes us about 15 minutes

to generate a new medical diagnostic test

and I'll show you that in real time now,

but I've compressed it
down to three minutes

by cutting some pieces out.

Rather than showing you
creating a medical diagnostic test,

I'm going to show you a diagnostic test

of car images,

because that's something
we can all understand.

So here we're starting with about
one and a half million car images,

and I want to create something
that can split them into the angle

of the photo that's being taken.

So these images are entirely unlabeled,

so I have to start from scratch.

With our deep learning algorithm,

it can automatically identify
areas of structure in these images.

So the nice thing is that the human

and the computer can now work together.

So the human, as you can see here,

is telling the computer
about areas of interest

which it wants the computer then
to try and use to improve its algorithm.

Now these deep learning systems actually

are in 16,000 dimensional space,

you can see here the computer
rotating this through that space,

trying to find new areas of structure.

And when it does so successfully,

the human who is driving it can then

point out the areas that are interesting.

So here, the computer has
successfully found areas,

for example, angles.

So as we go through this project,

we're gradually telling
the computer more and more

about the kinds of structures
we're looking for.

You can imagine in a diagnostic test

this would be a pathologist identifying
areas of pathosis, for example,

or a radiologist indicating

potentially troublesome nodules.

And sometimes it can be
difficult for the algorithm.

In this case, it's got kind of confused.

The fronts and the backs
of the cars are all mixed up.

So here we have to be a bit more careful,

manually selecting these fronts
as opposed to the backs,

then telling the computer

that this is a type of group

that we're interested in.

So we do that for a while,
we skip over it a little bit,

and then we train the
machine learning algorithm

based on these couple of hundred things,

and we hope that it's gotten a lot better.

You can see, it's now started to fade
some of these pictures out,

showing us that it already is recognizing

how to understand some of these itself.

We can then use this concept

of similar images,

and using similar images,
you can now see the computer

at this point is able to entirely
find just the fronts of cars.

So at this point, the human
can tell the computer,

"Okay, yes, you've done
a good job of that."

Sometimes, of course, even at this point

it's still difficult
to separate out groups.

In this case, even after we

let the computer try
to rotate this for a while,

we still find that the left sides
and the right sides pictures

are all mixed up together.

So we can again give
the computer some hints,

and we say, "Okay, try and find
a projection that separates out

the left sides and the right sides
as much as possible

using this deep learning algorithm."

And giving it that hint, ah,

it's been successful.

It's managed to find a way
of thinking about these objects

that's separated out these together.

So you get the idea here.

This is a case not where the human
is being replaced by a computer,

but where they're working together.

What we're doing here is we're replacing
something that used to take a team

of five or six people about seven years

and replacing it with something
that takes 15 minutes

for one person acting alone.

So this process takes about
four or five iterations.

You can see we now have 62 percent
of our one and a half million images

classified correctly.

And at this point, we
can start to quite quickly

grab whole big sections,

check through them to make sure
that there's no mistakes.

Where there are mistakes, we can
let the computer know about them.

And using this kind of process
for each of the different groups,

we are now up to
an 80 percent success rate

in classing the 1.5 million images.

And at this point, it's just a case

of finding the small number

that aren't classified correctly,

and trying to understand why.

And using that approach,

by 15 minutes we get
to 97 percent classification rates.

So this kind of technique
could allow us to fix

a major problem,

which is that there's a lack
of medical expertise in the world.

The World Economic Forum says
that there's between a 10x and a 20x

shortage of physicians
in the developing world,

and it would take about 300 years

to train enough people

to fix that problem.

So imagine if we can help
enhance their efficiency

using these deep learning approaches?

So I'm very excited
about the opportunities.

I'm also concerned about the problems.

The problem here is that
every area in blue on this map

is somewhere where services
are over 80 percent of employment.

What are services?

These are services.

These are also the exact things that
computers have just learned how to do.

So 80 percent of the world's employment

in the developed world is stuff that
computers have just learned how to do.

What does that mean? Well, it'll be fine.

They'll'll be replaced by other jobs.

For example, there will be
more jobs for data scientists.

Well, not really.

It doesn't take data scientists very long
to build these things.

For example, these four algorithms
were all built by the same guy.

So if you think, oh,
it's all happened before,

we've seen the results in the past
of when new things come along

and they get replaced by new jobs,

what are these new jobs going to be?

It's very hard for us to estimate this,

because human performance
grows at this gradual rate,

but we now have a system, deep learning,

that we know actually grows
in capability exponentially.

And we're here.

So currently, we see the things around us

and we see, "Oh, computers
are still pretty dumb." Right?

But in five years' time,

computers will be off this chart.

So we need to be starting to think
about this capability right now.

We have seen this once before, of course.

In the Industrial Revolution,

we saw a step change
in capability thanks to engines.

The thing is, though,

that after a while, things flattened out.

There was social disruption, but once
engines were used to generate power

and all the situations,

things really settled down.

The Machine Learning Revolution
is going to be very different

from the Industrial Revolution,

because the Machine Learning Revolution,
it never settles down.

The better computers get
at intellectual activities,

the more they can build better computers

to be better at intellectual capabilities,

so this is going to be a kind of change

that the world has actually
never experienced before,

so your previous understanding
of what's possible is different.

This is already impacting us.

In the last 25 years,

as capital productivity has increased,

labor productivity has been flat,

in fact even a little bit down.

So I want us to start
having this discussion now.

I know that when I often tell people
about this situation,

people can be quite dismissive.

Well, computers can't really think,

they don't emote,
they don't understand poetry,

we don't really understand how they work.

So what?

Computers right now can do the things

that humans spend most
of their time being paid to do,

so now's the time to start thinking

about how we're going to adjust our
social structures and economic structures

to be aware of this new reality.

Thank you.

(Applause)