The wonderful and terrifying implications of computers that can learn
-
0:01 - 0:05It used to be that if you wanted
to get a computer to do something new, -
0:05 - 0:06you would have to program it.
-
0:06 - 0:10Now, programming, for those of you here
that haven't done it yourself, -
0:10 - 0:13requires laying out in excruciating detail
-
0:13 - 0:17every single step that you want
the computer to do -
0:17 - 0:19in order to achieve your goal.
-
0:19 - 0:23Now, if you want to do something
that you don't know how to do yourself, -
0:23 - 0:25then this is going
to be a great challenge. -
0:25 - 0:28So this was the challenge faced
by this man, Arthur Samuel. -
0:28 - 0:32In 1956, he wanted to get this computer
-
0:32 - 0:35to be able to beat him at checkers.
-
0:35 - 0:37How can you write a program,
-
0:37 - 0:40lay out in excruciating detail,
how to be better than you at checkers? -
0:40 - 0:42So he came up with an idea:
-
0:42 - 0:46he had the computer play
against itself thousands of times -
0:46 - 0:48and learn how to play checkers.
-
0:48 - 0:52And indeed it worked,
and in fact, by 1962, -
0:52 - 0:56this computer had beaten
the Connecticut state champion. -
0:56 - 0:59So Arthur Samuel was
the father of machine learning, -
0:59 - 1:00and I have a great debt to him,
-
1:00 - 1:03because I am a machine
learning practitioner. -
1:03 - 1:04I was the president of Kaggle,
-
1:04 - 1:08a community of over 200,000
machine learning practictioners. -
1:08 - 1:10Kaggle puts up competitions
-
1:10 - 1:14to try and get them to solve
previously unsolved problems, -
1:14 - 1:17and it's been successful
hundreds of times. -
1:17 - 1:20So from this vantage point,
I was able to find out -
1:20 - 1:24a lot about what machine learning
can do in the past, can do today, -
1:24 - 1:26and what it could do in the future.
-
1:26 - 1:31Perhaps the first big success of
machine learning commercially was Google. -
1:31 - 1:34Google showed that it is
possible to find information -
1:34 - 1:36by using a computer algorithm,
-
1:36 - 1:38and this algorithm is based
on machine learning. -
1:38 - 1:42Since that time, there have been many
commercial successes of machine learning. -
1:42 - 1:44Companies like Amazon and Netflix
-
1:44 - 1:48use machine learning to suggest
products that you might like to buy, -
1:48 - 1:50movies that you might like to watch.
-
1:50 - 1:52Sometimes, it's almost creepy.
-
1:52 - 1:54Companies like LinkedIn and Facebook
-
1:54 - 1:56sometimes will tell you about
who your friends might be -
1:56 - 1:58and you have no idea how it did it,
-
1:58 - 2:01and this is because it's using
the power of machine learning. -
2:01 - 2:04These are algorithms that have
learned how to do this from data -
2:04 - 2:07rather than being programmed by hand.
-
2:07 - 2:10This is also how IBM was successful
-
2:10 - 2:14in getting Watson to beat
the two world champions at "Jeopardy," -
2:14 - 2:17answering incredibly subtle
and complex questions like this one. -
2:17 - 2:20["The ancient 'Lion of Nimrud' went missing
from this city's national museum in 2003
(along with a lot of other stuff)"] -
2:20 - 2:23This is also why we are now able
to see the first self-driving cars. -
2:23 - 2:26If you want to be able to tell
the difference between, say, -
2:26 - 2:28a tree and a pedestrian,
well, that's pretty important. -
2:28 - 2:31We don't know how to write
those programs by hand, -
2:31 - 2:34but with machine learning,
this is now possible. -
2:34 - 2:37And in fact, this car has driven
over a million miles -
2:37 - 2:40without any accidents on regular roads.
-
2:40 - 2:44So we now know that computers can learn,
-
2:44 - 2:46and computers can learn to do things
-
2:46 - 2:49that we actually sometimes
don't know how to do ourselves, -
2:49 - 2:52or maybe can do them better than us.
-
2:52 - 2:56One of the most amazing examples
I've seen of machine learning -
2:56 - 2:58happened on a project that I ran at Kaggle
-
2:58 - 3:02where a team run by a guy
called Geoffrey Hinton -
3:02 - 3:03from the University of Toronto
-
3:03 - 3:06won a competition for
automatic drug discovery. -
3:06 - 3:09Now, what was extraordinary here
is not just that they beat -
3:09 - 3:13all of the algorithms developed by Merck
or the international academic community, -
3:13 - 3:18but nobody on the team had any background
in chemistry or biology or life sciences, -
3:18 - 3:20and they did it in two weeks.
-
3:20 - 3:22How did they do this?
-
3:22 - 3:25They used an extraordinary algorithm
called deep learning. -
3:25 - 3:28So important was this that in fact
the success was covered -
3:28 - 3:31in The New York Times in a front page
article a few weeks later. -
3:31 - 3:34This is Geoffrey Hinton
here on the left-hand side. -
3:34 - 3:38Deep learning is an algorithm
inspired by how the human brain works, -
3:38 - 3:40and as a result it's an algorithm
-
3:40 - 3:44which has no theoretical limitations
on what it can do. -
3:44 - 3:47The more data you give it and the more
computation time you give it, -
3:47 - 3:48the better it gets.
-
3:48 - 3:51The New York Times also
showed in this article -
3:51 - 3:53another extraordinary
result of deep learning -
3:53 - 3:56which I'm going to show you now.
-
3:56 - 4:01It shows that computers
can listen and understand. -
4:01 - 4:03(Video) Richard Rashid: Now, the last step
-
4:03 - 4:06that I want to be able
to take in this process -
4:06 - 4:11is to actually speak to you in Chinese.
-
4:11 - 4:14Now the key thing there is,
-
4:14 - 4:19we've been able to take a large amount
of information from many Chinese speakers -
4:19 - 4:21and produce a text-to-speech system
-
4:21 - 4:26that takes Chinese text
and converts it into Chinese language, -
4:26 - 4:30and then we've taken
an hour or so of my own voice -
4:30 - 4:32and we've used that to modulate
-
4:32 - 4:36the standard text-to-speech system
so that it would sound like me. -
4:36 - 4:39Again, the result's not perfect.
-
4:39 - 4:42There are in fact quite a few errors.
-
4:42 - 4:44(In Chinese)
-
4:44 - 4:47(Applause)
-
4:49 - 4:53There's much work to be done in this area.
-
4:53 - 4:57(In Chinese)
-
4:57 - 5:00(Applause)
-
5:01 - 5:05Jeremy Howard: Well, that was at
a machine learning conference in China. -
5:05 - 5:07It's not often, actually,
at academic conferences -
5:07 - 5:09that you do hear spontaneous applause,
-
5:09 - 5:13although of course sometimes
at TEDx conferences, feel free. -
5:13 - 5:15Everything you saw there
was happening with deep learning. -
5:15 - 5:17(Applause) Thank you.
-
5:17 - 5:19The transcription in English
was deep learning. -
5:19 - 5:23The translation to Chinese and the text
in the top right, deep learning, -
5:23 - 5:26and the construction of the voice
was deep learning as well. -
5:26 - 5:29So deep learning is
this extraordinary thing. -
5:29 - 5:32It's a single algorithm that
can seem to do almost anything, -
5:32 - 5:35and I discovered that a year earlier,
it had also learned to see. -
5:35 - 5:38In this obscure competition from Germany
-
5:38 - 5:40called the German Traffic Sign
Recognition Benchmark, -
5:40 - 5:44deep learning had learned
to recognize traffic signs like this one. -
5:44 - 5:46Not only could it
recognize the traffic signs -
5:46 - 5:47better than any other algorithm,
-
5:47 - 5:50the leaderboard actually showed
it was better than people, -
5:50 - 5:52about twice as good as people.
-
5:52 - 5:54So by 2011, we had the first example
-
5:54 - 5:57of computers that can see
better than people. -
5:57 - 5:59Since that time, a lot has happened.
-
5:59 - 6:03In 2012, Google announced that
they had a deep learning algorithm -
6:03 - 6:04watch YouTube videos
-
6:04 - 6:08and crunched the data
on 16,000 computers for a month, -
6:08 - 6:12and the computer independently learned
about concepts such as people and cats -
6:12 - 6:14just by watching the videos.
-
6:14 - 6:16This is much like the way
that humans learn. -
6:16 - 6:19Humans don't learn
by being told what they see, -
6:19 - 6:22but by learning for themselves
what these things are. -
6:22 - 6:26Also in 2012, Geoffrey Hinton,
who we saw earlier, -
6:26 - 6:29won the very popular ImageNet competition,
-
6:29 - 6:33looking to try to figure out
from one and a half million images -
6:33 - 6:34what they're pictures of.
-
6:34 - 6:38As of 2014, we're now down
to a six percent error rate -
6:38 - 6:39in image recognition.
-
6:39 - 6:41This is better than people, again.
-
6:41 - 6:45So machines really are doing
an extraordinarily good job of this, -
6:45 - 6:47and it is now being used in industry.
-
6:47 - 6:50For example, Google announced last year
-
6:50 - 6:55that they had mapped every single
location in France in two hours, -
6:55 - 6:58and the way they did it was
that they fed street view images -
6:58 - 7:03into a deep learning algorithm
to recognize and read street numbers. -
7:03 - 7:05Imagine how long
it would have taken before: -
7:05 - 7:08dozens of people, many years.
-
7:08 - 7:10This is also happening in China.
-
7:10 - 7:14Baidu is kind of
the Chinese Google, I guess, -
7:14 - 7:17and what you see here in the top left
-
7:17 - 7:20is an example of a picture that I uploaded
to Baidu's deep learning system, -
7:20 - 7:24and underneath you can see that the system
has understood what that picture is -
7:24 - 7:26and found similar images.
-
7:26 - 7:29The similar images actually
have similar backgrounds, -
7:29 - 7:31similar directions of the faces,
-
7:31 - 7:33even some with their tongue out.
-
7:33 - 7:36This is not clearly looking
at the text of a web page. -
7:36 - 7:37All I uploaded was an image.
-
7:37 - 7:41So we now have computers which
really understand what they see -
7:41 - 7:43and can therefore search databases
-
7:43 - 7:46of hundreds of millions
of images in real time. -
7:46 - 7:50So what does it mean
now that computers can see? -
7:50 - 7:52Well, it's not just
that computers can see. -
7:52 - 7:54In fact, deep learning
has done more than that. -
7:54 - 7:57Complex, nuanced sentences like this one
-
7:57 - 7:59are now understandable
with deep learning algorithms. -
7:59 - 8:01As you can see here,
-
8:01 - 8:03this Stanford-based system
showing the red dot at the top -
8:03 - 8:07has figured out that this sentence
is expressing negative sentiment. -
8:07 - 8:11Deep learning now in fact
is near human performance -
8:11 - 8:16at understanding what sentences are about
and what it is saying about those things. -
8:16 - 8:19Also, deep learning has
been used to read Chinese, -
8:19 - 8:22again at about native
Chinese speaker level. -
8:22 - 8:24This algorithm developed
out of Switzerland -
8:24 - 8:27by people, none of whom speak
or understand any Chinese. -
8:27 - 8:29As I say, using deep learning
-
8:29 - 8:32is about the best system
in the world for this, -
8:32 - 8:37even compared to native
human understanding. -
8:37 - 8:40This is a system that we
put together at my company -
8:40 - 8:42which shows putting
all this stuff together. -
8:42 - 8:44These are pictures which
have no text attached, -
8:44 - 8:47and as I'm typing in here sentences,
-
8:47 - 8:50in real time it's understanding
these pictures -
8:50 - 8:51and figuring out what they're about
-
8:51 - 8:54and finding pictures that are similar
to the text that I'm writing. -
8:54 - 8:57So you can see, it's actually
understanding my sentences -
8:57 - 8:59and actually understanding these pictures.
-
8:59 - 9:02I know that you've seen
something like this on Google, -
9:02 - 9:05where you can type in things
and it will show you pictures, -
9:05 - 9:08but actually what it's doing is it's
searching the webpage for the text. -
9:08 - 9:11This is very different from actually
understanding the images. -
9:11 - 9:14This is something that computers
have only been able to do -
9:14 - 9:17for the first time in the last few months.
-
9:17 - 9:21So we can see now that computers
can not only see but they can also read, -
9:21 - 9:25and, of course, we've shown that they
can understand what they hear. -
9:25 - 9:28Perhaps not surprising now that
I'm going to tell you they can write. -
9:28 - 9:33Here is some text that I generated
using a deep learning algorithm yesterday. -
9:33 - 9:37And here is some text that an algorithm
out of Stanford generated. -
9:37 - 9:39Each of these sentences was generated
-
9:39 - 9:43by a deep learning algorithm
to describe each of those pictures. -
9:43 - 9:48This algorithm before has never seen
a man in a black shirt playing a guitar. -
9:48 - 9:50It's seen a man before,
it's seen black before, -
9:50 - 9:51it's seen a guitar before,
-
9:51 - 9:56but it has independently generated
this novel description of this picture. -
9:56 - 9:59We're still not quite at human
performance here, but we're close. -
9:59 - 10:03In tests, humans prefer
the computer-generated caption -
10:03 - 10:05one out of four times.
-
10:05 - 10:07Now this system is now only two weeks old,
-
10:07 - 10:09so probably within the next year,
-
10:09 - 10:12the computer algorithm will be
well past human performance -
10:12 - 10:13at the rate things are going.
-
10:13 - 10:16So computers can also write.
-
10:16 - 10:20So we put all this together and it leads
to very exciting opportunities. -
10:20 - 10:21For example, in medicine,
-
10:21 - 10:24a team in Boston announced
that they had discovered -
10:24 - 10:27dozens of new clinically relevant features
-
10:27 - 10:31of tumors which help doctors
make a prognosis of a cancer. -
10:32 - 10:35Very similarly, in Stanford,
-
10:35 - 10:38a group there announced that,
looking at tissues under magnification, -
10:38 - 10:41they've developed
a machine learning-based system -
10:41 - 10:43which in fact is better
than human pathologists -
10:43 - 10:48at predicting survival rates
for cancer sufferers. -
10:48 - 10:51In both of these cases, not only
were the predictions more accurate, -
10:51 - 10:53but they generated new insightful science.
-
10:53 - 10:55In the radiology case,
-
10:55 - 10:58they were new clinical indicators
that humans can understand. -
10:58 - 11:00In this pathology case,
-
11:00 - 11:04the computer system actually discovered
that the cells around the cancer -
11:04 - 11:08are as important as
the cancer cells themselves -
11:08 - 11:09in making a diagnosis.
-
11:09 - 11:15This is the opposite of what pathologists
had been taught for decades. -
11:15 - 11:18In each of those two cases,
they were systems developed -
11:18 - 11:22by a combination of medical experts
and machine learning experts, -
11:22 - 11:24but as of last year,
we're now beyond that too. -
11:24 - 11:28This is an example of
identifying cancerous areas -
11:28 - 11:30of human tissue under a microscope.
-
11:30 - 11:35The system being shown here
can identify those areas more accurately, -
11:35 - 11:38or about as accurately,
as human pathologists, -
11:38 - 11:41but was built entirely with deep learning
using no medical expertise -
11:41 - 11:44by people who have
no background in the field. -
11:45 - 11:47Similarly, here, this neuron segmentation.
-
11:47 - 11:51We can now segment neurons
about as accurately as humans can, -
11:51 - 11:54but this system was developed
with deep learning -
11:54 - 11:57using people with no previous
background in medicine. -
11:57 - 12:00So myself, as somebody with
no previous background in medicine, -
12:00 - 12:04I seem to be entirely well qualified
to start a new medical company, -
12:04 - 12:06which I did.
-
12:06 - 12:08I was kind of terrified of doing it,
-
12:08 - 12:11but the theory seemed to suggest
that it ought to be possible -
12:11 - 12:16to do very useful medicine
using just these data analytic techniques. -
12:16 - 12:19And thankfully, the feedback
has been fantastic, -
12:19 - 12:21not just from the media
but from the medical community, -
12:21 - 12:23who have been very supportive.
-
12:23 - 12:27The theory is that we can take
the middle part of the medical process -
12:27 - 12:30and turn that into data analysis
as much as possible, -
12:30 - 12:33leaving doctors to do
what they're best at. -
12:33 - 12:35I want to give you an example.
-
12:35 - 12:40It now takes us about 15 minutes
to generate a new medical diagnostic test -
12:40 - 12:42and I'll show you that in real time now,
-
12:42 - 12:45but I've compressed it down to
three minutes by cutting some pieces out. -
12:45 - 12:48Rather than showing you
creating a medical diagnostic test, -
12:48 - 12:52I'm going to show you
a diagnostic test of car images, -
12:52 - 12:54because that's something
we can all understand. -
12:54 - 12:57So here we're starting with
about 1.5 million car images, -
12:57 - 13:00and I want to create something
that can split them into the angle -
13:00 - 13:03of the photo that's being taken.
-
13:03 - 13:07So these images are entirely unlabeled,
so I have to start from scratch. -
13:07 - 13:08With our deep learning algorithm,
-
13:08 - 13:12it can automatically identify
areas of structure in these images. -
13:12 - 13:16So the nice thing is that the human
and the computer can now work together. -
13:16 - 13:18So the human, as you can see here,
-
13:18 - 13:21is telling the computer
about areas of interest -
13:21 - 13:25which it wants the computer then
to try and use to improve its algorithm. -
13:25 - 13:30Now, these deep learning systems actually
are in 16,000-dimensional space, -
13:30 - 13:33so you can see here the computer
rotating this through that space, -
13:33 - 13:35trying to find new areas of structure.
-
13:35 - 13:37And when it does so successfully,
-
13:37 - 13:41the human who is driving it can then
point out the areas that are interesting. -
13:41 - 13:43So here, the computer has
successfully found areas, -
13:43 - 13:46for example, angles.
-
13:46 - 13:47So as we go through this process,
-
13:47 - 13:50we're gradually telling
the computer more and more -
13:50 - 13:52about the kinds of structures
we're looking for. -
13:52 - 13:54You can imagine in a diagnostic test
-
13:54 - 13:57this would be a pathologist identifying
areas of pathosis, for example, -
13:57 - 14:02or a radiologist indicating
potentially troublesome nodules. -
14:02 - 14:05And sometimes it can be
difficult for the algorithm. -
14:05 - 14:07In this case, it got kind of confused.
-
14:07 - 14:09The fronts and the backs
of the cars are all mixed up. -
14:09 - 14:11So here we have to be a bit more careful,
-
14:11 - 14:15manually selecting these fronts
as opposed to the backs, -
14:15 - 14:20then telling the computer
that this is a type of group -
14:20 - 14:22that we're interested in.
-
14:22 - 14:24So we do that for a while,
we skip over a little bit, -
14:24 - 14:26and then we train the
machine learning algorithm -
14:26 - 14:28based on these couple of hundred things,
-
14:28 - 14:30and we hope that it's gotten a lot better.
-
14:30 - 14:34You can see, it's now started to fade
some of these pictures out, -
14:34 - 14:38showing us that it already is recognizing
how to understand some of these itself. -
14:38 - 14:41We can then use this concept
of similar images, -
14:41 - 14:43and using similar images, you can now see,
-
14:43 - 14:47the computer at this point is able to
entirely find just the fronts of cars. -
14:47 - 14:50So at this point, the human
can tell the computer, -
14:50 - 14:52okay, yes, you've done
a good job of that. -
14:54 - 14:56Sometimes, of course, even at this point
-
14:56 - 15:00it's still difficult
to separate out groups. -
15:00 - 15:03In this case, even after we let the
computer try to rotate this for a while, -
15:03 - 15:07we still find that the left sides
and the right sides pictures -
15:07 - 15:08are all mixed up together.
-
15:08 - 15:10So we can again give
the computer some hints, -
15:10 - 15:13and we say, okay, try and find
a projection that separates out -
15:13 - 15:16the left sides and the right sides
as much as possible -
15:16 - 15:18using this deep learning algorithm.
-
15:18 - 15:21And giving it that hint --
ah, okay, it's been successful. -
15:21 - 15:24It's managed to find a way
of thinking about these objects -
15:24 - 15:26that's separated out these together.
-
15:26 - 15:29So you get the idea here.
-
15:29 - 15:37This is a case not where the human
is being replaced by a computer, -
15:37 - 15:40but where they're working together.
-
15:40 - 15:43What we're doing here is we're replacing
something that used to take a team -
15:43 - 15:45of five or six people about seven years
-
15:45 - 15:48and replacing it with something
that takes 15 minutes -
15:48 - 15:50for one person acting alone.
-
15:50 - 15:54So this process takes about
four or five iterations. -
15:54 - 15:56You can see we now have 62 percent
-
15:56 - 15:59of our 1.5 million images
classified correctly. -
15:59 - 16:01And at this point, we
can start to quite quickly -
16:01 - 16:03grab whole big sections,
-
16:03 - 16:06check through them to make sure
that there's no mistakes. -
16:06 - 16:10Where there are mistakes, we can
let the computer know about them. -
16:10 - 16:13And using this kind of process
for each of the different groups, -
16:13 - 16:15we are now up to
an 80 percent success rate -
16:15 - 16:18in classifying the 1.5 million images.
-
16:18 - 16:20And at this point, it's just a case
-
16:20 - 16:23of finding the small number
that aren't classified correctly, -
16:23 - 16:26and trying to understand why.
-
16:26 - 16:28And using that approach,
-
16:28 - 16:32by 15 minutes we get
to 97 percent classification rates. -
16:32 - 16:37So this kind of technique
could allow us to fix a major problem, -
16:37 - 16:40which is that there's a lack
of medical expertise in the world. -
16:40 - 16:43The World Economic Forum says
that there's between a 10x and a 20x -
16:43 - 16:46shortage of physicians
in the developing world, -
16:46 - 16:48and it would take about 300 years
-
16:48 - 16:51to train enough people
to fix that problem. -
16:51 - 16:54So imagine if we can help
enhance their efficiency -
16:54 - 16:56using these deep learning approaches?
-
16:56 - 16:59So I'm very excited
about the opportunities. -
16:59 - 17:01I'm also concerned about the problems.
-
17:01 - 17:04The problem here is that
every area in blue on this map -
17:04 - 17:08is somewhere where services
are over 80 percent of employment. -
17:08 - 17:10What are services?
-
17:10 - 17:11These are services.
-
17:11 - 17:16These are also the exact things that
computers have just learned how to do. -
17:16 - 17:19So 80 percent of the world's employment
in the developed world -
17:19 - 17:22is stuff that computers
have just learned how to do. -
17:22 - 17:23What does that mean?
-
17:23 - 17:26Well, it'll be fine.
They'll be replaced by other jobs. -
17:26 - 17:29For example, there will be
more jobs for data scientists. -
17:29 - 17:30Well, not really.
-
17:30 - 17:33It doesn't take data scientists
very long to build these things. -
17:33 - 17:36For example, these four algorithms
were all built by the same guy. -
17:36 - 17:38So if you think, oh,
it's all happened before, -
17:38 - 17:42we've seen the results in the past
of when new things come along -
17:42 - 17:44and they get replaced by new jobs,
-
17:44 - 17:46what are these new jobs going to be?
-
17:46 - 17:48It's very hard for us to estimate this,
-
17:48 - 17:51because human performance
grows at this gradual rate, -
17:51 - 17:54but we now have a system, deep learning,
-
17:54 - 17:57that we know actually grows
in capability exponentially. -
17:57 - 17:58And we're here.
-
17:58 - 18:01So currently, we see the things around us
-
18:01 - 18:03and we say, "Oh, computers
are still pretty dumb." Right? -
18:03 - 18:07But in five years' time,
computers will be off this chart. -
18:07 - 18:11So we need to be starting to think
about this capability right now. -
18:11 - 18:13We have seen this once before, of course.
-
18:13 - 18:14In the Industrial Revolution,
-
18:14 - 18:17we saw a step change
in capability thanks to engines. -
18:18 - 18:21The thing is, though,
that after a while, things flattened out. -
18:21 - 18:23There was social disruption,
-
18:23 - 18:26but once engines were used
to generate power in all the situations, -
18:26 - 18:28things really settled down.
-
18:28 - 18:30The Machine Learning Revolution
-
18:30 - 18:33is going to be very different
from the Industrial Revolution, -
18:33 - 18:36because the Machine Learning Revolution,
it never settles down. -
18:36 - 18:39The better computers get
at intellectual activities, -
18:39 - 18:43the more they can build better computers
to be better at intellectual capabilities, -
18:43 - 18:45so this is going to be a kind of change
-
18:45 - 18:47that the world has actually
never experienced before, -
18:47 - 18:51so your previous understanding
of what's possible is different. -
18:51 - 18:53This is already impacting us.
-
18:53 - 18:56In the last 25 years,
as capital productivity has increased, -
18:56 - 19:01labor productivity has been flat,
in fact even a little bit down. -
19:01 - 19:04So I want us to start
having this discussion now. -
19:04 - 19:07I know that when I often tell people
about this situation, -
19:07 - 19:09people can be quite dismissive.
-
19:09 - 19:10Well, computers can't really think,
-
19:10 - 19:13they don't emote,
they don't understand poetry, -
19:13 - 19:16we don't really understand how they work.
-
19:16 - 19:17So what?
-
19:17 - 19:19Computers right now can do the things
-
19:19 - 19:22that humans spend most
of their time being paid to do, -
19:22 - 19:24so now's the time to start thinking
-
19:24 - 19:28about how we're going to adjust our
social structures and economic structures -
19:28 - 19:30to be aware of this new reality.
-
19:30 - 19:31Thank you.
-
19:31 - 19:32(Applause)
- Title:
- The wonderful and terrifying implications of computers that can learn
- Speaker:
- Jeremy Howard
- Description:
-
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.
- Video Language:
- English
- Team:
- closed TED
- Project:
- TEDTalks
- Duration:
- 19:45
Morton Bast edited English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Morton Bast edited English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Morton Bast edited English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Morton Bast edited English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Morton Bast approved English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Madeleine Aronson edited English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Madeleine Aronson edited English subtitles for The wonderful and terrifying implications of computers that can learn | ||
Madeleine Aronson accepted English subtitles for The wonderful and terrifying implications of computers that can learn |